Datasets:

neuralwork

/

arxiver

like 344

Follow

neuralwork 54

2305.18249

Evolution of QPOs in GX 339-4 and EXO 1846-031 with Insight-HXMT and NICER

We conduct a spectral and timing analysis of GX 339-4 and EXO 1846-031 with the aim of studying the evolution of Type-C QPOs with spectral parameters. The high cadence data from Insight-HXMT and NICER allow us to track them. Type-C QPOs appear at the end of low-hard state and/or hard-intermediate state. The results reveal that the QPO frequency is closely related to the inner disk radius and mass accretion rate in the two sources. Such a correlation is nicely consistent with the dynamic frequency model.

Zuobin Zhang, Honghui Liu, Divya Rawat, Cosimo Bambi, Ranjeev Misra, Pengju Wang, Long Ji, Shu Zhang, Shuangnan Zhang

2023-05-29T17:15:50Z

http://arxiv.org/abs/2305.18249v2

# Evolution of QPOs in Gx 339\(-\)4 and Exo 1846\(-\)031 with _Insight_-Hxmt and _Nicer_ ###### Abstract We conduct a spectral and timing analysis of GX 339\(-\)4 and EXO 1846\(-\)031 with the aim of studying the evolution of Type-C QPOs with spectral parameters. The high cadence data from _Insight_-HXMT and _NICER_ allow us to track them. Type-C QPOs appear at the end of low-hard state and/or hard-intermediate state. The results reveal that the QPO frequency is closely related to the inner disk radius and mass accretion rate in the two sources. Such a correlation is nicely consistent with the dynamic frequency model. Subject headings:High energy astrophysics; X-ray astronomy; Low mass X-ray Binary; Stellar mass black holes ## 1. Introduction Quasi-periodic oscillations (QPOs) refers to narrow peaks structure in the power density spectrum (PDS) that are commonly observed in X-ray binaries (XRBs) (van der Klis, 2005). In black hole systems, QPOs are mainly split into low-frequency QPOs (LFQPOs, centroid frequency \(0.1-30\) Hz), and high frequency QPOs (HFQPOs, centroid frequency \(\geq 60\) Hz) (Belloni, 2010). Samimi et al. (1979) reported the'sporadic quasi-periodic behaviour' in the light curve of GX 339\(-\)4, and Motch et al. (1983) reported the first rigorous detection of QPOs for the same source. It was immediately recognized that the QPOs would have been a powerful tool to study the accretion process around black holes. Over the last forty years, especially after the launch of _RXTE_, we have accumulated a lot of knowledge about QPOs. Using a truncated power-law to fit the broadband noise in PDS, and a Lorentz function with the center frequency of \(\nu_{\rm QPO}\) to fit the low-frequency QPOs, Wijnands et al. (1999) found that there is an obvious positive correlation between the truncation frequency \(\nu_{\rm b}\) and the frequency of the low-frequency QPOs \(\nu_{\rm LF}\). Psaltis et al. (1999) reported that there is also a good positive correlation between the frequency of low-frequency QPOs and the frequency of the broadband noise (or high-frequency QPOs) in low mass XRBs, including black holes and neutron stars. We have observed QPOs in most black hole XRBs, and realized that low frequency QPOs can be divided into three types: Type-A, -B, and -C QPOs, based on quality factor, noise type, fractional rms, and phase delay (e.g., Wijnands et al., 1999; Sobczak et al., 2000; Casella et al., 2005; Motta et al., 2011). Different types QPOs occupy different regions on the hardness-intensity diagram, as well as obviously distribute in different areas on the center frequency and rms plots (e.g., Motta et al., 2011). The phenomenon of rapid transition between different types of QPOs has been found in some sources, and the time scale of this phenomenon can be very short (10 s) (e.g. Homan et al., 2020). In this work, we only focus on the Type-C QPOs. Type-C QPOs appear in the early stage of the outburst, particularly in the hard-intermediate state and the end of low-hard state. The centroid frequency varies from a few mHz to \(\sim 10\) Hz, and is tightly correlated with the spectral state. Vignarca et al. (2003) reported a positive correlation between the centroid frequency and the photon index \(\Gamma\). Motta et al. (2011) found Type-C QPOs trace out a well-defined track, and the centroid frequency obviously correlate with the corona flux and disk flux. The dependence of the QPO frequency and photon energy was illustrated by Qu et al. (2010). In addition to the phenomenological study of QPOs, many studies has been done on the theoretical explanation behind it. Most theoretical models explain QPO phenomenon through the following two different mechanisms: instabilities of the corona-disk system (e.g., Titarchuk & Fiorito, 2004; Mastichiadis et al., 2022; Varniere et al., 2012) or the geometrical effects of general relativity (e.g., Stella & Vietri, 1998; Ingram et al., 2009). Titarchuk & Fiorito (2004) introduced a transition layer in corona-disk system that can explain the QPO phenomenon in XRBs. The disk-corona natural frequency model was proposed by Mastichiadis et al. (2022), and they argued that type-C QPOs arise from the interaction of the hot corona with the cold accretion disk. Varniere et al. (2012) suggested that LFQPOs could result from the relativistic accretion-ejection instability (AEI). The geometrical effects model mainly refers to the precession of the corona region. This model interprets the QPOs as a Lense-Thirring precession of the innermost region of the accretion disk (e.g., Stella and Vietri, 1998; Ingram et al., 2009). In recent years, more and more observations have been analyzed to test these models. However, a unified model that can explain all QPO behaviors has not been found yet. Recently, Misra et al. (2020) identified the QPO frequency of GRS 1915+105 as the relativistic dynamic frequency of a truncated accretion disk with _AstroSat_ observations of that source. The authors found a strong correlation between the QPO frequency divided by the accretion rate and the inner disk radius. The correlation is actually consistent with the prediction of dynamic frequency under the assumption of a standard relativistic accretion model (Novikov and Thorne, 1973). Liu et al. (2021) extended the relation to cover a wider range of variations, and confirmed the high spin nature of the black hole in GRS 1915+105 with the data of _Insight_-HXMT (dubbed HXMT; Zhang et al., 2014). We note that GRS 1915+105 is a persistent source with particular properties (Belloni et al., 2000). We would like to test if this relation holds for other sources different from GRS 1915+105, and we notice that there are two appropriate sources, GX 339\(-\)4 and EXO 1846\(-\)031, in the archive. The XRB transient GX 339\(-\)4 is a typical low mass X-ray binary (LMXB) discovered in 1973 (Markert et al., 1973). It goes into bright outburst every a few years and all four X-ray states typically seen in XRBs have been detected in this system (e.g., Miyamoto et al., 1995; Homan and Belloni, 2005; Plant et al., 2014). GX 339\(-\)4 is located at 8-12 kpc with a black hole mass of 4-11 M\({}_{\bigodot}\)(Zdziarski et al., 2019). Strong relativistic reflection signatures have been found in this source in the hard and soft states (e.g., Garcia et al., 2015; Miller et al., 2004; Liu et al., 2022). Previous studies have found that the black hole in GX 339\(-\)4 has a very high spin (\(a_{*}\sim 0.95\), Garcia et al., 2015; Parker et al., 2016). The inclination angle of the accretion disk should have an intermediate value (Furst et al., 2015; Parker et al., 2016). Motta et al. (2011) systematically studied the properties and the behaviour of QPOs, as a function of the integrated broad-band variability and the spectral parameters. The authors suggested that the frequencies of all QPOs (including Type-C QPOs) correlate with the disk flux. EXO 1846\(-\)031 was discovered by the European X-ray Observatory Satellite (_EXOSAT_) when it went into outburst in April 1985 and is considered a LMXB (Parmar et al., 1993; Draghis et al., 2020). _CGRO_/BATSE detected a second outburst in 1994 (Zhang et al., 1994). Since then, the source was in a quiescent state for 25 years. EXO 1846\(-\)031 had a new outburst in 2019, which was monitored by X-ray missions (e.g., _MAXI_/GSC; HXMT; _NuSTAR_) and radio missions (e.g., _MeerKAT_; _AMI-LA_). _NuSTAR_ conducted a high-quality observation on August 3, 2019 with a 22.2 ks exposure time. Draghis et al. (2020) reported strong relativistic reflection features with the extremely sensitive _NuSTAR_ spectra, and argued that the source is a black hole with a nearly maximal spin parameter (\(a_{*}=0.997\)) at disk inclination of \(\theta=73^{\circ}\). EXO 1846\(-\)031 is located at 2.4-7.5 kpc according to the previous studies on X-ray and radio data (Parmar et al., 1993; Williams et al., 2022), with a black hole mass of \(\sim 9\) M\({}_{\bigodot}\)(Draghis et al., 2020; Williams et al., 2022). Liu et al. (2021) reported the observational results from a detailed timing analysis of EXO 1846\(-\)031 2019 outburst with the observations of HXMT and _NICER_. In this work, we focus on the latest HXMT and _NICER_ observations of GX 339\(-\)4 and EXO 1846\(-\)031, and present a detailed temporal and spectral analysis. The paper is organized as follows. Sec. 2 presents the observational data reduction. The spectral timing analysis is reported in Sec. 3. We discuss the results and report our conclusions in Sec. 4 and Sec. 5, respectively. ## 2. Observations and Data Reduction ### Data selection Starting from February 2021, GX 339\(-\)4 went into a new outburst that lasted for a few months. Fig. 1 shows the long-term light curve in the 2-20 keV band and the corresponding hardness ratio observed with _MAXI_ GSC. The hardness is defined as the ratio between the count rates at 4-10 keV and 2-4 keV. HXMT and _NICER_ extensively observed the 2021 outburst of the source. We went through all available HXMT and _NICER_ data and picked out those observations that show Type-C QPO signatures. The selected observations analyzed in this work are marked in the light curve of GX 339\(-\)4 in Fig. 1. Information about these observations is listed in Tab. A1. The 2019 outburst of EXO 1846\(-\)031 was first detected by _MAXI_/GSC on 2019 July 23 (Negoro et al., 2019), and it lasted about 3 months. The _MAXI_ X-ray Hardness-Intensity diagram (HID) of the outburst shows a characteristic q-shaped hysteresis of X-ray binaries in outburst (Williams et al., 2022; Liu et al., 2021). The long-term _MAXI_ light curve and the corresponding hardness ratio are shown in Fig. 2. HXMT and _NICER_ conducted high-cadence pointing observations of EXO 1846\(-\)031. Type-C QPOs appear during the transition from hard to soft state (Liu et al., 2021). We selected observations showing Type-C QPO signatures. The selected observations are marked in the light curve in Fig. 2 and listed in Tab. A2. ### Data reduction HXMT covers the broadband energy range of 1-250 keV with low-energy, medium-energy, and high-energy detectors (Zhang et al., 2020; Chen et al., 2020; Cao et al., 2020; Liu et al., 2020). The light curves and spectra are extracted with the HXMT data analysis software (HXMTDAS) version 2.05 and CALDB version 2.06, following the official user guide. The background is estimated by the standalone scripts hebkgmap, mebkgmap, and lebkgmap(Guo et al., 2020; Liao et al., 2020). The data are screened following the recommended criteria, i.e., an elevation angle \(>\)10\({}^{\circ}\), a geomagnetic cutoff rigidity \(>\)10 GeV, a pointing offset angle \(<\)0.1, and at least 300 s away from the South Atlantic Anomaly. The _NICER_ data are processed with the _NICER_ data analysis software (NICERDAS) version 2021-04-01_V008 and CALDB version 20210707. We use the standard fil tering criteria: the pointing offset is less than 54'', and the pointing direction is more than 40deg away from the bright Earth limb, more than 30deg away from the dark Earth limb, and outside the South Atlantic Anomaly. In addition, we remove the data of detectors # 14 and # 34 which are affected by episodes of increased electronic noise, and we select events that are not flagged as "overshoot" or "undershoot" resets (EVENT_FLAGS = bxxxx00) or forced triggers (EVENT_FLAGS=bx1x000). The standard _NICER_ reduction routine nicerl2 is used to process the data. The cleaned events are barycenter-corrected using the FTOOL barycorr. We extract the energy spectra of the background in each observation using the nibackgen3C50 tool (Remillard et al., 2022). The Redistribution Matrix File (RMF) and Ancillary Response File (ARF) are created by using the tasks nicerrmf and nicerarf, respectively. ## 3. Data Analysis ### Timing analysis We extract HXMT LE and _NICER_ XTI light curves with a time resolution of 1 ms from the full energy band Figure 1.— _MAXI_ GSC light curve and corresponding hardness of GX 339\(-\)4 starting from 2021 February 5 (MJD = 59250). The hardness is defined as the ratio between the count rates at 4–10 keV and 2–4 keV bands. The vertical black lines mark the HXMT observations analyzed in this work and the vertical red lines mark the _NICER_ observations. Figure 2.— _MAXI_ GSC light curve and corresponding hardness of EXO 1846\(-\)031 starting from 2019 June 16 (MJD = 58650). The hardness is defined as the ratio between the count rates at 4–10 keV and 2–4 keV bands. The vertical black lines mark the HXMT observations analyzed in this work and the vertical red lines mark the _NICER_ observations. (1-10 keV in HXMT; 0.5-10 keV in _NICER_) for each HXMT observation and _NICER_ observation. In order to calculate hardness ratios, we also produce LE light curves from the 1-5 keV and 5-10 keV bands, and produce XTI light curves from the 0.5-4 keV and 4-10 keV bands. We carefully check the extracted light curves from all observations of GX 339\(-\)4 and find there are two _NICER_ observations (ObsID: 4133010103, 4133010104) that show a relatively strong variability of count rate and hardness. Fig. 3 shows the light curves of these two _NICER_ observations. The gaps in the light curves are due to the low Earth orbit of the telescope or the SAA. We can clearly see that the source went through a period of luminosity increase and hardness decrease. Comparing with the location of these two observations in Fig. 1 (the last red dotted line; in fact, since the two lines are quite close, they look like one line), we conclude that the two observations are during the hard-to-soft transition so the hardness keeps getting lower. Then we divide the observations according to the light curve, counting each snapshot as a sub-observation. We check over all sub-observation data and pick out those with exposure \(>200\) s. The selected sub-observations are numbered 4133010103-1 through 4133010103-9, and 4133010104-1 through 4133010104-13, as shown in Fig. 3. The other light curves do not show strong variability in the count rate i.e., no distinctive evidence of flares, dips, or state transitions, making it safe for timing and spectral analysis to characterize the source properties. For EXO 1846\(-\)031, the count rate of the source remain fairly stable during each HXMT and _NICER_ interval, and the hardness does not change dramatically. Therefore, we conclude that we can carry out timing and spectral analysis in the unit of one observation. To measure the QPO frequency of GX 339\(-\)4 and EXO 1846\(-\)031, we employ the Python package Stingray(Huppenkothen et al., 2019) to create PDS for each observation. The light curve is splited into 64 s segments, and then the final PDS is generated by averaging all 64 s segments. The PDS is normalized according to the "rms" method (Belloni & Hasinger, 1990), and logarithmically rebinned so that each bin size is 1.02 times larger than the previous bin. Note that we focus on the HXMT LE 1-10 keV light curve and the _NICER_ XTI 0.5-10 keV light curve to extract PDS and search for QPO signal. The 8-30 keV HXMT ME light curves have been analyzed in the same way and return consistent measurements of the QPO frequencies. So we report only the results from LE data in this work. We use XSPEC v12.12.1 (Arnaud, 1996) to analyze the Figure 3.— Light curves of GX 339\(-\)4 by _NICER_ in the 0.5–10 keV band (top panel) and corresponding hardness (bottom panel) (ObsID: 4133010103, 4133010104). The hardness is defined as the ratio between the count rates in the 4–10 keV and 0.5–4 keV bands. The source undergoes a process of flux increase and hardness decrease during this period. The intervals with exposure \(>200\) s are marked with yellow shadow. The selected sub-observations are numbered 4133010103-1 through 4133010103-9, and 4133010104-1 through 4133010104-13. PDS. The typical PDS of XRBs manifests broad components and one or two narrow peaks at different frequencies, corresponding to broad band noise, the possible QPO fundamental and (sub) harmonics, respectively. We need at least one narrow Lorentzian for the QPO to fit the Poisson-extracted PDS (Belloni et al., 2002). More narrow Lorentzians are sometimes included to model harmonic peaks. All QPOs we detect have a quality factor (Q) greater than 4 and detection significance greater than 3\(\sigma\). Fig. A1 and Fig. A2 show a typical PDS and the fit results with several Lorentzian models for GX 339\(-\)4. Fig. A3 and Fig. A4 show the counterpart for EXO 1846\(-\)031. The QPO frequencies for each observation are listed in Tab. A3 for GX 339\(-\)4, and Tab. A4 for EXO 1846\(-\)031. ### Spectral analysis For spectral analysis of the HXMT data, we consider the LE data in the 2-10 keV band and ME data in the 8-20 keV band. ME data above 20 keV and HE data are ignored because of the very high background. Note that we ignore the data below 2 keV of the LE instrument in spectral analysis (instead of the 1 keV for timing analysis) because of calibration uncertainties in the low energy band. For _NICER_ data, we consider the 1-10 keV band in this section, ignoring the data below 1 keV because of calibration issues. The HXMT and _NICER_ spectra are fitted with the XSPEC (v12.12.1) package, using the recommended photoelectric cross sections of Verner et al. (1996) and element abundances of Wilms et al. (2000). The \(\chi^{2}\) statistics is employed and all parameter uncertainties are estimated at 90% confidence level, corresponding to \(\Delta\chi^{2}=2.71\). All spectra are grouped to ensure a minimum counts of 20 per bin. A systematic error of 1% is added in the _NICER_ spectra. The HXMT and _NICER_ spectra of GX 339\(-\)4 are fitted with the model combination Tobbs \(\times\) (simpl \(\times\) kerrd + relxill). Tobabs is included to account for absorption by the interstellar medium. We set its column density (\(n_{\rm H}\)) to be a free parameter for _NICER_ spectra. While with HXMT spectra, we can not constrain its column density (\(n_{\rm H}\)), so we fix it at best-fit value, \(0.55\times 10^{22}\) cm\({}^{-2}\), which is consistent with the result of the _NICER_ data and the value in literature (e.g., Wang et al., 2020; Liu et al., 2022). kerrd accounts for the thermal emission from the geometrically thin and optically thick accretion disk (Ebisawa et al., 2003), in which the black hole distance, mass, and inclination angle of the accretion disk are set to 8.4 kpc, 9.0 M\({}_{\bigodot}\), and 30\({}^{\circ}\)(Parker et al., 2016), respectively. The spectral hardening factor of kerrd is set to 1.7 (Shimura and Takahara, 1995). simpl(Steiner et al., 2009) is used to take into account for the Comptonization of disk photons by the corona. The source has been found to have strong reflection features (Liu et al., 2022), and we use the full reflection model relxill(Garcia et al., 2014) to fit them. The spin parameter (\(a_{\star}\)) is fixed at 0.95 (Parker et al., 2016), and the index of the emissivity profile is fixed at 3 because it cannot be constrained by the fit. The best-fit values and uncertainties of GX 339\(-\)4 are shown in Tab. A3. Fig. A1 and Fig. A2 show typical spectra and fit results of HXMT data and _NICER_ data, respectively. In the case of EXO 1846\(-\)031, the best-fit model combination is Tobabs \(\times\) (simpl \(\times\) kerrd + relxill) for HXMT spectra. The black hole distance, mass, and inclination angle of the accretion disk are set to 4.5 kpc (Williams et al., 2022), 10.0 M\({}_{\bigodot}\)(Williams et al., 2022, Draghis et al., 2020) and 73\({}^{\circ}\)(Draghis et al., 2020), respectively. The spin parameter (\(a_{\star}\)) in relxill is fixed at 0.998 (Draghis et al., 2020). We use a simple power-law to model the emissivity profile (\(q_{\rm in}=q_{\rm out}\) free). The other parameters are set exactly as in the case of GX 339\(-\)4. For _NICER_ spectra, we notice that there are still some large residuals in the soft X-ray band with the same model, including a Gaussian-like emission near 1.1 keV and edge-like shapes near 1.8 keV. These energies correspond to features in the effective area of _NICER_ versus energy (e.g., Wang et al., 2020), where 1.1 and 1.8 keV are attributed to sodium and silicon, respectively. Therefore, we adopt the following model for the _NICER_ spectra: Tobabs \(\times\) (simpl \(\times\) kerrd + relxill + gaussian) \(\times\) edge. This calibration issue arises in EXO 1846\(-\)031 because the source has a high interstellar absorption, which makes the photon count rate in the lower energy band relatively low, making the calibration issue prominent. Typical spectra and fit results of HXMT and _NICER_ are shown in Fig. A3 and Fig. A4. In Tab. A4, we summarize the best-fit values and errors of EXO 1846\(-\)031. ## 4. Results and Discussion Fig. 4 and Fig. 5 show the evolution of inner radius (\(R_{in}\)) and QPO frequency (\(f_{\rm QPO}\)) with time for GX 339\(-\)4 and EXO 1846\(-\)031, respectively. Generally speaking, we clearly see that the value of \(f_{\rm QPO}\) monotonically increases with time. The behaviour is consistent with that reported in Motta et al. (2011) and Liu et al. (2021). It has also been observed in other XRBs, for example, XTE J1859+226 (Casella et al., 2004). In addition, a notable feature for both sources is the decrease of \(R_{in}\). For GX 339\(-\)4, the inner disk moves toward the ISCO (Innermost Stable Circular Orbit), from \(>50R_{\rm g}\) to \(\sim 7R_{\rm g}\) (\(R_{\rm g}\), gravitational radius), which coincide with the result in the previous study (e.g. Wang-Ji et al., 2018; Wang et al., 2020). Although there is some variable feature, EXO 1846\(-\)031 shows a similar trend. Correlation between the parameters involved in temporal and spectral analysis are shown in Fig. 6 and Fig. 7. An interesting result is the relationship between the photon index (\(\Gamma\)) and the QPO frequency (\(f_{\rm QPO}\)). The results we get from both sources share the same tendency, as shown in the bottom panels of Fig. 6 and Fig. 7. There is a strong positive correlation between \(f_{\rm QPO}\) and \(\Gamma\) of the power-law in the beginning which flattens or starts reversing at the highest values of the \(f_{\rm QPO}\). The turnoff in the correlation is not apparent in GX 339\(-\)4, while it is evident in EXO 1846\(-\)031 (around \(\Gamma\sim 2.7\)). A similar kind of correlation have been reported in a number of other LMXBs (e.g., Vignarca et al., 2003; Titarchuk and Fiorito, 2004; Titarchuk and Seifina, 2009; Furst et al., 2016). Titarchuk and Fiorito (2004) introduced the transition layer (TL) model to explain the observed correlations. The TL model depicts how the QPOs related to the corona properties (e.g., the size, optical depth, temperature and spectral index), and predicts the correlation between photon index and QPO frequency. The results we get are in good agreement with the model's predictions, except for the observations of EXO 1846\(-\)031 with \(f_{\rm QPO}>5.18\), where a negative correlation between \(f_{\rm QPO}\) and \(\Gamma\) appears. A universal explanation of this correlation between \(\Gamma\) and \(f_{\rm QPO}\) is still missing. The upper left panel of Fig. 6 shows a broad anti-correlation between the QPO frequency (\(f_{\rm QPO}\)) and inner radius (\(R_{in}\)) in GX 339\(-\)4. This anti-correlation is not particularly significant in EXO 1846\(-\)031, and we can only see a general tendency of larger \(R_{in}\) corresponding to a smaller frequency. The same correlation between the QPO frequency and the disk inner radius was reported in other sources (e.g., GRS 1915\(+\)105; Rodriguez et al., 2002). The Lense-Thirring precession model would predict anti-correlation between \(f_{\rm QPO}\) and \(R_{in}\), and a direct dependence of the QPO frequency on the inner radius (Ingram et al., 2009; Ingram and Done, 2010). To check the possibility of modeling the results with the relativistic precession model, we use equation 2 in Ingram et al. (2009) to fit data points. The model cannot explain the results we obtained, both in the case of GX 339\(-\)4 and EXO 1846\(-\)031, as shown in the plot. The variation of the frequency with the accretion rate is shown in the upper right panels of Fig. 6 and Fig. 7. Liu et al. (2021) reported a strongest correlation between the QPO frequency (\(f_{\rm QPO}\)) and mass accretion rate (\(\dot{M}\)) in GRS 1915\(+\)105 with HXMT data. We do not find any significant correlation between them in GX 339\(-\)4, while there is a weak anti-correlation in EXO 1846\(-\)031. In fact, a positive correlation between \(f_{\rm QPO}\) and \(\dot{M}\) is proposed in the TL model by Titarchuk and Fiorito (2004), and in the disk-corona natural frequency model by Mastichiadis et al. (2022). Fig. 3 of Titarchuk and Fiorito (2004) depicts the positive correlation between \(f_{\rm QPO}\) and the \(\gamma\)-parameter (which is proportional to mass accretion rate), which is opposite to what we find. Besides, Mastichiadis et al. (2022) argue that type-C QPOs could arise from the interaction of the hot corona with the cold accretion disk, and predict a formula \(f_{0}\propto\dot{M}^{1/2}\) below a certain mass accretion rate (Fig. 5 of Mastichiadis et al. (2022)). The results we get do not fit well with the predictions of that model. These discrepancies may suggest that the "transition layer model" and the disk-corona natural frequency model are not favored in our case. Misra et al. (2020) identified QPOs as the dynamic frequency of a truncated relativistic accretion disk in the case of GRS 1915\(+\)105. The dynamic frequency is defined as the ratio of the sound propagation velocity of the inner disk to the truncation radius, i.e., the inverse Figure 4.— Evolution of disk inner radius \(R_{\rm in}\) (left panel) and QPO frequency (\(f_{\rm QPO}\)) along with MJD in GX 339\(-\)4. The black points indicate HXMT data, and red dots indicate _NICER_ data. Figure 5.— Evolution of disk inner radius \(R_{\rm in}\) (left panel) and QPO frequency (\(f_{\rm QPO}\)) along with MJD in EXO 1846\(-\)031. The black points and red points represent the results of HXMT data and _NICER_ data, respectively. of the sound crossing time (Misra et al., 2020). Based on the assumption that the accretion disk is a standard relativistic accretion disk (Novikov and Thorne, 1973), the dynamic frequency is a function of the inner radius (\(R_{\rm in}\)), black hole spin (\(a_{*}\)), mass accretion rate (\(\dot{M}\)) and a normalization factor (\(N\)). Liu et al. (2021) extended the results in Misra et al. (2020) to a larger range of accretion rates with HXMT data of GRS 1915+105, and confirmed the high spin nature of the source. Following the work of Misra et al. (2020) and Liu et al. (2021), we illustrate the relation between QPO frequency (\(f_{\rm QPO}\)) divided by accretion rate (\(\dot{M}\)) and disk inner radius (\(R_{\rm in}\)) in Fig. 8 and Fig. 9. Both sources show negative correlation between \(f_{\rm QPO}/\dot{M}\) and \(R_{\rm in}\), Moreover, the correlation follows the prediction of the dynamic frequency model. We fit the relation between \(f_{\rm QPO}/\dot{M}\) and \(R_{\rm in}\) using Equation (3) in Misra et al. (2020). The fit returns \(a_{*}=0.9978\pm 0.0009\) and \(N=0.281\pm 0.025\) for EXO 1846\(-\)031, indicating a rapidly spinning black hole. This result is consistent with what has been reported by analyzing the blurred reflection spectra (e.g., Dragish et al., 2020; Abdikamalov et al., 2021). The best-fit curve is shown in Fig. 9. In the case of GX 339\(-\)4, the fit returns \(a_{*}=0.603\pm 0.026\) and \(N=1.02\pm 0.05\). Such a low spin result is somewhat different from the result obtained by analyzing the blurred reflection spectra or Figure 6.— Correlation between the parameters involved in the temporal and spectral analysis in the case of GX 339\(-\)4. Correlation of the QPO frequency vs. inner disk radius and the QPO frequency vs. accretion rate are shown in the upper left and upper right panels. The two central panels illustrate the accretion rate vs. inner disk radius and inner disk radius vs. photon index. The photon index vs. QPO frequency is depicted in the bottom panel. The black and red crosses denote the results of HXMT data and _NICER_ data, respectively. In the left top panel, the dashed gray lines represent the correlation of the frequency and inner radius predicted by Lense–Thirring precession model. The lines from left to right depict \(a_{*}=0.3\), 0.5, 0.7, 0.9 and 0.998, respectively. thermal spectra (e.g. Reis et al., 2008; Ludlam et al., 2015; Garcia et al., 2015; Parker et al., 2016; Wang et al., 2020). We note that for this source we do not have data below 6 \(R_{\rm g}\). The relativistic effects are more evident at lower \(R_{\rm in}\) (3 \(\sim\) 5 \(R_{\rm g}\)). Hence, data points at lower \(R_{\rm g}\) plays a crucial role in the estimation of spin parameter value. In Fig. 8, we simultaneously show two curves, \(a_{\star}\) = 0.603 and \(a_{\star}\) = 0.900. It is worth noting that the most important difference between the two curves is reflected in the region with low \(R_{\rm in}\). This also proves our view that a reasonable fitting value cannot be obtained because of the lack of data with relatively small \(R_{\rm in}\). The middle right panels of Fig. 6 and Fig. 7 show that the inner disk radius tends to decrease when the photon index (\(\Gamma\)) increases. The behaviour is consistent with that expected during a hard-to-soft transition. A noteworthy positive correlation between the mass accretion rate (\(\dot{M}\)) and the inner radius (\(R_{in}\)) in EXO 1846\(-\)031 is described in the middle left panel of Fig. 7. A similar relationship was reported in GRS 1915+105 (Misra et al., 2020; Liu et al., 2021; Rawat et al., 2022) and MAXI J1535-571 (Garg et al., 2022). The correlation is beyond the expectation of truncated disk model (Done et al., 2007). But Dullemond & Spruit (2005) predicted a positive correlation between \(\dot{M}\) and \(R_{in}\) (see their Fig. 8), calculating the evaporation of the cool accretion disk on account of the ion-bombardment. An alternative explanation is discussed in Abramowicz et al. (1978), where the authors suggested a larger inner edge is required when the mass accretion rate increases to dissipate the angular momentum of accretion material. ## 5. Conclusion Figure 7.— Correlation between the parameters involved in temporal and spectral analysis in the case of EXO 1846\(-\)031. It is organized as in Fig. 6. \begin{table} \begin{tabular}{c c c c} \hline \hline Mission & Obs. ID & Start data & Exposure (s) \\ \hline & P0304024026 & 2021-03-12 & 2274 \\ & P0304024028 & 2021-03-14 & 1401 \\ HXMT & P0304024032 & 2021-03-18 & 1597 \\ & P0304024035 & 2021-03-22 & 1669 \\ & P0304024036 & 2021-03-24 & 1193 \\ & P0304024038 & 2021-03-26 & 2088 \\ \hline & 3558011402 & 2021-03-17 & 1595 \\ & 3558011501 & 2021-03-19 & 7560 \\ _NICER_ & 4133010101 & 2021-03-19 & 2030 \\ & 4133010102 & 2021-03-20 & 1860 \\ & 4133010103 & 2021-03-26 & 6111 \\ & 4133010104 & 2021-03-27 & 8709 \\ \hline \hline \end{tabular} \end{table} Table 1HXMT and _NICER_ observations of GX 339\(-\)4 analyzed in this work. For HXMT, the listed exposure time is for the LE instrument. Figure 8.— Variation of QPO frequency divided by the accretion rate vs. inner disk radius in the case of GX 339\(-\)4. The results of HXMT data and _NICER_ data are denoted with black and red crosses, respectively. The orange curve represents the best fit ( \(a_{*}=0.603\), \(N=1.02\)), and the blue curve corresponds to \(a_{*}=0.900\), \(N=0.52\). Figure 9.— Variation of QPO frequency divided by the accretion rate vs. inner disk radius in the case of EXO 1846\(-\)031. The black and red crosses denote the results of HXMT data and _NICER_ data, respectively. The blue curve represents the best fit ( \(a_{*}=0.9978\), \(N=0.281\)). The orange curve represents the best fit ( \(a_{*}=0.754\), \(N=0.853\)) when we only include the data with \(R_{\rm in}>5~{}R_{\rm g}\). It proves that the data with small \(R_{\rm in}\) are important to fit the spin parameter. \begin{table} \begin{tabular}{c c c c} \hline \hline Mission & Obs. ID & Start data & Exposure (s) \\ \hline & P021405000101 & 2019-08-02 & 718 \\ & P021405000102 & 2019-08-02 & 1436 \\ & P021405000103 & 2019-08-02 & 762 \\ & P021405000104 & 2019-08-02 & 718 \\ & P021405000105 & 2019-08-02 & 1715 \\ & P021405000106 & 2019-08-03 & 563 \\ HXMT & P021405000107 & 2019-08-03 & 656 \\ & P021405000301 & 2019-08-05 & 700 \\ & P021405000302 & 2019-08-05 & 1102 \\ & P021405000303 & 2019-08-05 & 678 \\ & P021405000401 & 2019-08-06 & 718 \\ & P021405000502 & 2019-08-07 & 691 \\ & P021405000503 & 2019-08-07 & 539 \\ & P021405000601 & 2019-08-08 & 1130 \\ & P021405000701 & 2019-08-08 & 1163 \\ & P021405000702 & 2019-08-09 & 1795 \\ \hline & 2200760101 & 2019-07-31 & 5658 \\ & 2200760102 & 2019-08-01 & 1165 \\ & 2200760103 & 2019-08-02 & 2562 \\ & 2200760104 & 2019-08-03 & 1488 \\ & 2200760105 & 2019-08-04 & 1130 \\ & 2200760106 & 2019-08-05 & 3564 \\ & 2200760107 & 2019-08-06 & 912 \\ _NICER_ & 2200760108 & 2019-08-07 & 927 \\ & 2200760109 & 2019-08-08 & 3293 \\ & 2200760110 & 2019-08-09 & 4629 \\ & 2200760112 & 2019-08-11 & 2749 \\ & 2200760113 & 2019-08-12 & 3341 \\ & 2200760114 & 2019-08-13 & 7154 \\ & 2200760115 & 2019-08-13 & 8181 \\ & 2200760116 & 2019-08-15 & 4703 \\ & 2200760117 & 2019-08-16 & 8739 \\ & 2200760118 & 2019-08-17 & 4875 \\ & 2200760119 & 2019-08-17 & 3341 \\ & 2200760120 & 2019-08-19 & 3894 \\ \hline \hline \end{tabular} \end{table} Table 1: Left: a typical PDS of GX 339\(-\)4 from HXMT data (Obs ID: P0304024028). Right: the HXMT spectrum and residuals to the best-fit model for the same observation. Data from the LE and ME detector are denoted in black and red, respectively. Figure 17: Left: a typical PDS of EXO 1846\(-\)031 from HXMT data (Obs ID: 2200760106). Right: the _NICER_ spectrum and residuals to the best-fit model for the same observation. Figure 18: Left: a typical PDS of EXO 1846\(-\)031 from HXMT data (Obs ID: P021405000105). Right: the HXMT spectrum and residuals to the best-fit model for the same observation. Same as Fig. 17, data from the LE and ME detector are denoted in black and red, respectively. Figure 17: Left: a typical PDS of EXO 1846\(-\)031 from _NICER_ data (Obs ID: 2200760106). Right: the _NICER_ spectrum and residuals to the best-fit model for the same observation.

2306.16734

Unified View of Damage leaves Planimetry & Analysis Using Digital Images Processing Techniques

The detection of leaf diseases in plants generally involves visual observation of patterns appearing on the leaf surface. However, there are many diseases that are distinguished based on very subtle changes in these visually observable patterns. This paper attempts to identify plant leaf diseases using image processing techniques. The focus of this study is on the detection of citrus leaf canker disease. Canker is a bacterial infection of leaves. Symptoms of citrus cankers include brown spots on the leaves, often with a watery or oily appearance. The spots (called lesions in botany) are usually yellow. It is surrounded by a halo of the leaves and is found on both the top and bottom of the leaf. This paper describes various methods that have been used to detect citrus leaf canker disease. The methods used are histogram comparison and k-means clustering. Using these methods, citrus canker development was detected based on histograms generated based on leaf patterns. The results thus obtained can be used, after consultation with experts in the field of agriculture, to identify suitable treatments for the processes used.

Pijush Kanti Kumar, DeepKiran Munjal, Sunita Rani, Anurag Dutta, Liton Chandra Voumik, A. Ramamoorthy

2023-06-29T07:15:45Z

http://arxiv.org/abs/2306.16734v1

# Unified View of Damage leaves Planimetry & Analysis Using Digital Images Processing Techniques ###### Abstract The detection of leaf diseases in plants generally involves visual observation of patterns appearing on the leaf surface. However, there are many diseases that are distinguished based on very subtle changes in these visually observable patterns. This paper attempts to identify plant leaf diseases using image processing techniques. The focus of this study is on the detection of citrus leaf darker disease. Canker is a bacterial infection of leaves. Symptoms of citrus cankers include brown spots on the leaves, often with a water or only appearance. The spots (called lesions in botany) are usually yellow. It is surrounded by a halo of the leaves and is found on both the top and bottom of the leaf. This paper describes various methods that have been used to detect citrus leaf cancer disease. The methods used are histogram comparison and k-means clustering. Using these methods, citrus canker development was detected based on histograms generated based on leaf patterns. The results thus obtained can be used, after consultation with experts in the field of agriculture, to identify suitable treatments for the processes used. Digital Image Processing, k - Means Clustering, Citrus Leaf Canker Disease + Footnote †: publicationid: pubid: 2023 IEEE ## I Introduction In today's world, the farmland volume serves as an additional source of food. Agriculture's productivity has a significant impact on the Indian economy. As a result, it is crucial to identify plant diseases in the field of agriculture. Use of an automatic disease recognition system is advantageous for spotting a plant pathogens in its very early stages. For example in the case, the United States has pine trees that are susceptible to a dangerous disease called little leaf disorder. The affected tree grows slowly and perishes within six years. Parts of the Southern US, including Alabama and Georgia, are affected by it. Early diagnosis in these situations might have been beneficial. The only method currently in use for identifying and detecting _phytopathogens_ is professional assessment using only one's unaided eye. This requires a sizable group of specialists and ongoing vegetation surveillance, both of which are very expensive when dealing with large farmlands. Meanwhile, in some nations, farmers lack access to the necessary resources and even the knowledge to speak with experts. Because of this, consulting experts is expensive and time-consuming. The recommended method works well in these circumstances for keeping an eye on huge expanses of harvests. Visual diagnosis of plant diseases is more time-consuming, less accurate, and only practicable in a few locations. However, using an automatic detection method will require less work, less time, and result in a higher degree of accuracy. Brown and yellow spots, early and late scorch, and certain other common bacterial, viral, and fungal diseases are all seen in plants. The size of the diseased area is measured using image processing, and the difference in color of the damaged area is also determined. The method used for breaking down or classifying an image into numerous components is known as image segmentation. Image segmentation can be done in a variety of cases right now, from the straightforward segmen tation process to sophisticated color segmentation methods. Usually, these product attributes to elements that people can easily differentiate into separate elements and see. There are numerous techniques for segmenting images because computers lack the ability to recognize objects intelligently. The image's unique components are the basis for the segmentation algorithm. This could be a section of an image, color features, or boundary details. Diseased leaf spots play an important role in plant growth environment. Diseases can also be easily identified with the help of affected areas in culture [1]. Usually, of course, leaves clearly show infected areas and are easily identifiable. Therefore, we can say that plant colour change is an essential aspect of notification. If the crop is in good health, the crop will have different colours, but if the crop dies from some harmful pathogen, it will automatically change colour. Plant diseases affect specific parts. This can lead to decreased productivity. The main method used in practice to detect plant diseases is early observation by a specialist, exposing the eye. In this feature retrieval-based investigation, influenced leaf parts were examined using the suggested diagonal disagreement of border, color, and appearance variability features. K-Means clustering was employed. Six different disease types were predicted using this method. Results of the performance assessments were eventually achieved. This reviewed work [2] is based on methods to reduce computational complexity through improved automated crop disease detection. This is due to the fact that it significantly harms the agricultural industry and makes the disease's symptoms visible. Once more, the possibility for treating and preventing infected plants is clear. In order to detect and categorize plant diseases, it is necessary to look for reliable, affordable, and accurate techniques. Test photographs of various cotton sheets were initially implemented. The pictures that are are then captured using image processing approaches, and helpful features are extracted for additional analysis. uses a variety of statistical techniques to categorize the images in accordance with the precise issue of the influenced leaf spot regions. The analysis of the best corresponded to of characteristic conclusions for leaves impacted by the perfect solution was done using the selection of features. In the classification phase [3], feature values corresponding to the variance of edge, colour and texture features are stored in the image domain. Based on the affected areas of leaf diseases, the affected sites were identified. ## II Materials and Methods ### _Materials_ The following materials were taken into use. 1. Nikon Make 12.5 Megapixels Digital Camera 2. Personal Computer 3. White A4 Paper Sheet for background 4. MATLAB 2013 version or above 5. Photographs of the leaves were arranged in number. ### _Methods_ #### Ii-B1 Graphical Method A leaf with a measurement region was positioned on graph paper with a 1 mm-wide grid. On graph paper, the sheets are meticulously and precisely defined with the aid of a pencil. It was determined how many grids the sheet's detail encompassed in total. The boundary detail is regarded as 1 if it takes up over fifty percent of the grid; otherwise, it is addressed as 0. The precise leaf area is represented by an array of grid numerals. #### Ii-B2 Image Processing Method A wider audience are capable of determining leaf area thanks to the MATLAB-based technique used for the processing of images, which is a partially automated technique. Code is created using MATLAB 2013 or later. For the latest releases of the application, this code is functional. The procedure is as follows: 1. Study the picture. 2. RGB to grayscale conversion. 3. Create a binary picture from the image that is grayscale. 4. Do the leaf area calculation. ## III k Means Clustering Data can be grouped in a variety of ways, however the most popular method is the k-Means algorithm [4], which aims to make categories become more comparable while simultaneously maintaining them as far apart as feasible. In essence, distance calculations using Euclidean distance are performed using k-Means. Euclidean distance calculates the distance between two given points \((x_{1},y_{1})\) and \((x_{2},y_{2})\) using the following formula \[Distance_{euclidian}=\sqrt{(x_{2}-x_{1})^{2}+(y_{2}-y_{1})^{2}} \tag{1}\] The above formula captures distance in 2D space, but the same holds true in multidimensional space as the number of terms added increases. The '_k_' in _k_-Means represents the number of distinct clusters into which the data is divided. A fundamental caveat [5] of the _k_-Means algorithm is that the data must be continuous in nature. It will not work if your data is categorical in nature. ### _Data Preparation_ As was previously addressed, the majority of methodologies for clustering, including k-Means, rely on the concept of separation. They compute their distance from a specific point and make an effort to reduce it. When distinct variables utilize distinct components, something occurs. For instance, I want to divide the general population of India into different groups since weight is measured in kilograms of weight while heights is measured in centimeters. As can be observed, the variable's components have a significant impact on the matrix of distances above. As a result, standardizing the information prior to seeking clustering activities is an excellent decision. ### _Algorithm_ _k_-Means is an iterative clustering process. This is repeated until an optimal solution or cluster is reached in the problem space. The following pseudo-example covers the basic steps involved in \(k\)-means clustering, which is commonly used to cluster data. Start by deciding how many clusters you want, in the present instance three. The \(k\)-Means algorithm attempts to connect the closest points to the arbitrarily data centers it starts with. ### _Image acquisition_ The leaf measuring the affected area is placed on a black background with no light reflection [6]. Hold the camera horizontally to the paper. The shooting distance is neither too close nor too far. The photo has been adjusted to cover only the background. See Fig. 2. ### _Observing the picture_ The image is saved in the computer as b3.jpeg. This image is read for further processing by MATLAB. ### _Dividing the Image_ The initial picture is divided into two distinct kinds of images in this phase based on how its colors vary. These pictures were chosen for their ability to show impacted and untouched leaf regions. This is done by the \(k\)-Means clustering method. The algorithm for this method is shown below. 1. Read the image. 2. Change an image's color space from RGB towards LAB. 3. Use \(k\)-means clustering to categorize the colors in the 'a*b*' space. 4. With the help of the \(k\)-means outcomes identify each pixel in the picture. 5. Make pictures that divide the initial picture into sections based on color. After applying the above algorithm segmented images were obtained. See Fig. 3 and Fig. 4. ### _Convert the image (Cluster - I) into Binary image_ After reading the image, I need to first convert this image from RGB to grayscale [7] and then convert the grayscale image to binary. In this image, the white - coloured areas indicate the unaffected parts of the leaves, as shown in the image in Fig. 5. Useful for calculating the total number of pixels in unaffected areas of the sheet. ### _Convert the image (Cluster - II) into Binary image_ After converting the image into binary image, image in Fig. 6 is obtained where white coloured regions indicate affected regions of the leaf. It helps to calculate total number of pixels of affected regions [8] of the leaf. ### _Pixel Calculation_ #### Iv-H1 From Cluster I WP (Pixels of unaffected regions) = 195612 where, WP denotes white pixels of Cluster - I. #### Iv-H2 From Cluster II WP1 (Pixels of affected regions) = 41246 where, WP1 denotes white pixels of Cluster - II. Now, total pixels of the leaf area that is obtained is shown below. \[TP=WP+WP1 \tag{2}\] \[TP=195612+41246=236858\] Percentage of affected pixels can be obtained by the following as \[Error=\frac{WP1\ (pixels\ of\ affected\ regions)}{TP\ (Total\ pixels\ of\ the\ leaf\ area)}\times 100\] ERROR (%) = 17.4138 % ## IV Results and Discussion Leaves were selected from different plots and different types of citrus leaves to test the performance [9] of the new measurement system. Leaf area is calculated using the square grid method and is considered standard area. When we obtained sampled data from different affected leaves. Then you can get a clear idea about the whole damaged sheet in relation to the specific area. This proposed method is faster [10] and more accurate than any standard method. ## V Conclusion This article discusses a method for measuring the area of citrus leaves using the processing of digital images. Calculating the quantity of leaves that are damaged has been demonstrated to be possible with the executed algorithmic structure. Although more time-consuming, grid calculating strategies are highly precise for measuring the surface area of leaves. The image processing method is also a fast method [11 - 21] with high accuracy [22] and precision [23]. You may modify the leaf picture at any moment if you solely preserve it [24]. This can serve as an empirical basis for creating an affordable leaf area meter that meets precision farming requirements. For precise area computations and disease incidence projections for chemical pesticides and application of fertilizer, the proportion of error aspect is crucial and required. ## References * [1] Prathyakshini, Akshaya, and C. V. Aravinda, "Classification and Clustering of Infected Leaf Plant Using K-Means Algorithm," _Communications in Computer and Information Science_, pp. 468-474, 2018, doi:[https://doi.org/10.1007/978-981-10-9059-2_41](https://doi.org/10.1007/978-981-10-9059-2_41). * [2] M. Zhang et al., "Chromosomal-level genome assembly of potato underground," Phtophiromae operuelle a pse of solanocous crops," _Scientific Data_, vol. 9, no. 1, Dec. 2022, doi: [https://doi.org/10.1038/s41597-022-01859-5](https://doi.org/10.1038/s41597-022-01859-5). * [3] H. A. Gharib and A. M. Mandour, "Effect of Populus nigra spring and autumn leaves extract on Capsicum anum infected with pepper mild mottle virus," _Scientific Reports_, vol. 12, no. 1, Dec. 2022, doi: [https://doi.org/10.1038/s41598-022-24786-2](https://doi.org/10.1038/s41598-022-24786-2). * [4] A. A. Najm, M. R. M. S. Hadi, F. Fazeli, M. T. Darzi, and A. Rahi, "Effect of Integrated Management of Nitrogen Fertilizer and Cattle Manure on the Leaf Chlorophyll, Yield, and Tuber Glycosalkioids of Agria Potato," _Communications in Soil Science and Plant Analysis_, vol. 43, no. 6, pp. 912-923, Mar. 2012, doi:[https://doi.org/10.1080/00103624.2012.653027](https://doi.org/10.1080/00103624.2012.653027).

2310.05910

SALMON: Self-Alignment with Instructable Reward Models

Supervised Fine-Tuning (SFT) on response demonstrations combined with Reinforcement Learning from Human Feedback (RLHF) constitutes a powerful paradigm for aligning LLM-based AI agents. However, a significant limitation of such an approach is its dependency on high-quality human annotations, making its application to intricate tasks challenging due to difficulties in obtaining consistent response demonstrations and in-distribution response preferences. This paper presents a novel approach, namely SALMON, to align base language models with minimal human supervision, using only a small set of human-defined principles, yet achieving superior performance. Central to our approach is an instructable reward model. Trained on synthetic preference data, this model can generate reward scores based on arbitrary human-defined principles. By merely adjusting these principles during the RL training phase, we gain full control over the preferences with the instructable reward model, subsequently influencing the behavior of the RL-trained policy models, and reducing the reliance on the collection of online human preferences. Applying our method to the LLaMA-2-70b base language model, we developed an AI assistant named Dromedary-2. With only 6 exemplars for in-context learning and 31 human-defined principles, Dromedary-2 significantly surpasses the performance of several state-of-the-art AI systems, including LLaMA-2-Chat-70b, on various benchmark datasets. We have open-sourced the code and model weights to encourage further research into aligning LLM-based AI agents with enhanced supervision efficiency, improved controllability, and scalable oversight.

Zhiqing Sun, Yikang Shen, Hongxin Zhang, Qinhong Zhou, Zhenfang Chen, David Cox, Yiming Yang, Chuang Gan

2023-10-09T17:56:53Z

http://arxiv.org/abs/2310.05910v2

# Salmon: Self-Alignment with ###### Abstract Supervised Fine-Tuning (SFT) on response demonstrations combined with Reinforcement Learning from Human Feedback (RLHF) constitutes a powerful paradigm for aligning LLM-based AI agents. However, a significant limitation of such an approach is its dependency on high-quality human annotations, making its application to intricate tasks challenging due to difficulties in obtaining consistent response demonstrations and in-distribution response preferences. This paper presents a novel approach, namely **SALMON** (Self-**AL**ign**M**ent with principle-**f**O**lowi**N**g reward models), to align base language models with minimal human supervision, using only a small set of human-defined principles, yet achieving superior performance. Central to our approach is a _principle-following reward model_. Trained on synthetic preference data, this model can generate reward scores based on arbitrary human-defined principles. By merely adjusting these principles during the RL training phase, we gain full control over the preferences with the reward model, subsequently influencing the behavior of the RL-trained policies, and eliminating the reliance on the collection of online human preferences. Applying our method to the LLaMA-2-70b base language model, we developed an AI assistant named Dromedary-2. With only 6 exemplars for in-context learning and 31 human-defined principles, Dromedary-2 significantly surpasses the performance of several state-of-the-art AI systems, including LLaMA-2-Chat-70b, on various benchmark datasets. We have open-sourced the code and model weights to encourage further research into aligning LLM-based AI agents with enhanced supervision efficiency, improved controllability, and scalable oversight. ## 1 Introduction The prevailing AI alignment paradigm, exemplified in models like ChatGPT (OpenAI, 2022) and LLaMA-2-Chat (Touvron et al., 2023b), employs supervised fine-tuning (SFT) with prompted demonstrations (Sann et al., 2021; Chung et al., 2022; Zhou et al., 2023) and reinforcement learning from human feedback (RLHF) to align the outputs of large language models (LLMs) with human intentions (Ziegler et al., 2019; Ouyang et al., 2022). However, acquiring high-quality human annotations, including consistent response demonstrations and in-distribution preferences, is costly and not scalable (Touvron et al., 2023b). Furthermore, the existing paradigm of SFT + RLHF is inherently limited in assuming that humans can always demonstrate or evaluate the tasks undertaken by advanced AI systems. Although today's models fall within human evaluative boundaries, future, more advanced models could embark on tasks that challenge human evaluation. Consequently, there is a looming danger, i.e., such models may value appealing human evaluators over ensuring accuracy (Andreas, 2022; Perez et al., 2022). To address the current challenges in AI alignment, we aim to develop a new methodology that facilitates scalable oversight (Amodei et al., 2016; Bowman et al., 2022). Our vision is to define a few general principles, akin to Issac Asimov's three laws in robotics (Asimov, 1941), which are comprehensively internalized for AI systems to follow (Gilardi et al., 2023; Ganguli et al., 2023). This goal is in line with the recent research on _self-alignment_(Bai et al., 2022; Sun et al., 2023b), where the primary focus is to use AI models to improve themselves, e.g., with bootstrapping over the model-generated critiques (Madaan et al., 2023; Fu et al., 2023) or self-refined outputs (Wang et al., 2022; Li et al., 2023a). However, it is worth noting that these bootstrapping methods still lag behind the RLHF method in performance (Bai et al., 2022; Touvron et al., 2023b). Meanwhile, methods like Reinforcement Learning from AI Feedback (RLAIF) or Constitutional AI (CAI) (Bai et al., 2022; OpenAI, 2023a) has emerged as an alternative potential. These techniques leverage feedback from automated AI systems, reducing the reliance on exhaustive human-annotated preferences. So far, the primary focus of the previous RLAIF work remains on enhancing the safety of the models that have already undergone RLHF training. That is, these RLAIF methods inherit the heavy dependency on the human-annotated preferences in the RLHF warm-up stage. This leads to a pivotal research question: * **Can RLAIF fully replace RLHF to align language models from scratch in enhancing their general alignment and capabilities?** This paper provides a definitive confirmation for the above question by introducing a novel approach namely **SALMON**. At the heart of our approach lies the introduction of the principle-following (also termed instruction-following) reward model. Pioneering in its nature, this reward model is adept at interpreting and adhering to arbitrary human-written preference guidelines, subsequently generating human-guided reward scores. This is different from previous RLAIF methods (Bai et al., 2022; OpenAI, 2023a) where the principles are only used to produce synthetic preferences, and the resulting reward models generate scores without any specific principles, as illustrated in Figure 1. The design of our principle-following reward model enables better control over the behavior of the final RL-trained policy model. Within conventional RLHF paradigms, the iterative collection of online (in-distribution) preference data (Bai et al., 2022; Touvron et al., 2023b) is essential to counteract reward hacking (Pan et al., 2022). This complication emerges when the policy model exploits weaknesses in the reward model, producing inflated scores that do not accurately reflect model performance. In SALMON, we can address this issue by simply crafting principles explicitly \begin{table} \begin{tabular}{l c c c c} \hline \hline & \# Demonstration & \# Preference & MT-Bench & Alignment \\ & Annotations & Annotations & Score & Techniques \\ \hline _(closed-source models)_ & & & & \\ \hline InstructGPT-SFT (175b) & 12,725 & 0 & 2.7 & SFT \({}^{a}\) \\ InstructGPT (175b) & 12,725 & 33,207 &? & SFT \& RLHF \({}^{a}\) \\ Text-Davinci-003 (175b) &? &? & 6.4 & SFT \& RLHF \({}^{a}\) \\ Claude-V1 (?) &? &? & 7.9 & RLHF \& CAI \({}^{b}\) \\ ChatGPT (?) &? &? & 7.9 & SFT \& RLHF \({}^{c}\) \\ GPT-4 (?) &? &? & 9.0 & SFT \& RLHF \& CAI \({}^{d}\) \\ \hline _(non-distilled open-source models)_ & & & & & \\ Dolly-V2 (12b) & 15,000 & 0 & 2.0 & SFT \\ Guanaco (65b) & 9,846 & 0 & 6.4 & SFT \\ OpenAssistant-SFT (30b) & 69,614 & 0 & 6.4 & SFT \\ OpenAssistant (30b) & 69,614 & 39,670 & 6.6 & SFT \& RLHF \\ LLaMA-2-Chat (70b) & 27,540 & 1,418,091 & 6.9 & SFT \& RLHF \\ Dromedary-2 (70b) & **6** & **0** & **7.4** & Self-Align \& SALMON \\ \hline \hline \end{tabular} \end{table} Table 1: Comparison of human supervisions used in recent AI systems and their MT-Bench scores (Zheng et al., 2023). We exclude models that used any Knowledge Distillation (KD) data. The alignment techniques used in previous work include SFT (Supervised Fine-tuning), RLHF (Reinforcement Learning from Human Feedback), and CAI (Constitutional AI). Information is from: \({}^{a}\) OpenAI (2023b), \({}^{b}\) Bai et al. (2022b); Anthropic (2023), \({}^{c}\) OpenAI (2022), \({}^{d}\) OpenAI (2023a). designed to combat observed1 reward hacking patterns in model outputs, such as self-praising at the end of the response. Additionally, we found that we are able to emphasize distinct aspects of the alignment in the HHH (helpful, honest, and harmless) alignment framework (Askell et al., 2021) by customizing the preference principles. Our methodology also proved effective in reducing the occurrence of false refusal seen in certain over-aligned language models (Touvron et al., 2023b) by crafting special principles. Footnote 1: In this paper, we write language descriptions of the reward-hacking patterns observed through human’s manual inspection. Future work may consider a more systematic and automated approach (Bills et al., 2023; Zhong et al., 2023) for summarizing the language descriptions of the reward hacking patterns. Our principle-following reward model can be trained with synthetic data and seamlessly applied to a diverse range of language models without collecting any model-specific human preference data (Bai et al., 2022a; Touvron et al., 2023b). Possible policy model initialization strategies include principle-driven self-alignment (Sun et al., 2023b), supervised fine-tuning on human demonstrations (Chung et al., 2022a; Zhou et al., 2023), or even those unaligned base language models (Touvron et al., 2023a). Remarkably, when integrated with the Self-Align technique (Sun et al., 2023b), our method enabled the training of a self-aligned AI-assistant agent, namely Dromedary-2, from scratch by only manually crafting **6 exemplars** for In-Context Learning (Brown et al., 2020) and a combined total of **31 principles** (17 from Self-Align and 14 for SALMON). Despite its minimal human supervision design, our model outperformed the extensively RLHF-trained LLaMA-2-Chat model (Touvron et al., 2023b), which was trained with over 20,000+ human-curated response demonstrations and 1,000,000+ human-annotated response preferences. The comparisons of human supervision efficiency and performance on MT-Bench (Zheng et al., 2023) are detailed in Table. 1. ## 2 Related Work AI Alignment from ScratchThe problem of aligning AIs (Gabriel, 2020), especially large language models (LLMs), to human values and intentions in terms of being helpful, honest, and harmless (Christiano et al., 2017; Patil et al., 2020; Askell et al., 2021; Ouyang et al., 2022; Bai et al., 2022a;b; OpenAI, 2023a) has gained significant attention as recent AI systems have rapidly ad Figure 1: Comparison among RLHF (Ouyang et al., 2022), RLAIF (Bai et al., 2022b), and SALMON (Ours). The vanilla (stand-alone) reward models in RLHF & RLAIF are trained to give high scores to generally good responses, while the principle-following reward model in SALMON is trained to generate reward scores based on customized principles as the preference guideline. vanced in their capabilities (Devlin et al., 2018; Radford et al., 2019; Brown et al., 2020; Chowdhery et al., 2022). This work focuses on the problem of aligning LLMs from scratch, that is, we aim to develop a new methodology capable of aligning a pre-trained base language model without relying on pre-existing, well-aligned models like ChatGPT (OpenAI, 2022) or GPT-4 (OpenAI, 2023a). This direction markedly differentiates our work from contemporary research primarily focused on distilling capabilities of aligned behaviors from proprietary models into smaller open-source models (Taori et al., 2023; Chiang et al., 2023), which has notable drawbacks (Guolbande et al., 2023). Scalable Oversight & Self-AlignmentAI alignment traditionally relies heavily on extensive human annotations. Primary Supervised Fine-Tuning (SFT) sources for response demonstrations include those curated from existing NLP datasets (Sann et al., 2021; Wei et al., 2021; Chung et al., 2022b; Wang et al., 2022) and those specifically created by humans for instruction tuning (Databricks, 2023; Kopf et al., 2023; Zhou et al., 2023; Ouyang et al., 2022). In the recent trend of aligning language models with Reinforcement Learning from Human Feedback (RLHF; Christiano et al. (2017); Stiennon et al. (2020); Ouyang et al. (2022); Bai et al. (2022a); Touvron et al. (2023b)), online human preferences are collected to train a reward model to further fine-tune the SFT-trained model (Leike et al., 2018). However, acquiring high-quality human annotations, including consistent response demonstrations and in-distribution preferences, has emerged as a significant bottleneck. This limitation hampers the full potential of AI-assistant agents because human oversight in the current formats of demonstration or preference may not be generalizable to more complex tasks. Additionally, even for relatively simpler tasks, obtaining human annotations could be costly and raises concerns about quality, reliability, diversity, creativity, self-consistency, and the potential for undesirable biases (Wang et al., 2022a; Kopf et al., 2023; Wan et al., 2023). To address the above challenges, we need to develop a new paradigm to support **"self-alignment"** in AI systems that can facilitate scalable oversight (Nakano et al., 2021; Bowman et al., 2022). A few notable self-alignment techniques involve bootstrapping by fine-tuning on model-generated synthetic data. For instance, Self-Instruct (Wang et al., 2022a) bootstraps a base language model with its own generations conditional on 175 In-Context Learning (ICL) query-response pairs. Self-Align (Sun et al., 2023b) removes the need for response demonstrations and uses 16 principles and 5 ICL exemplars to guide the AI in generating appropriate responses. Instruction Back-translation (Li et al., 2023a) uses web documents to create new training examples for an SFT model trained on 3200 seed examples. But the efficacy of such bootstrapping strategies in outperforming the established RLHF paradigm remains an open question (Bai et al., 2022b; Touvron et al., 2023b). Reinforcement Learning from AI Feedback (RLAIF)Another line of self-alignment research seeks to fine-tune LLMs using a reward model trained on the AI's own evaluations (Bai et al., 2022b; OpenAI, 2023a) or a stronger LLM as the oracle evaluator (Dubois et al., 2023). In particular, Constitutional AI (CAI) (Bai et al., 2022b; OpenAI, 2023a) delves into self-enhancement for alleviating harmful outputs, without relying on human annotations. This is achieved through AI-generated self-critiques, revisions, and preference models. Guided by a set of human-written principles, this method aims to make AI systems more safe. In contrast, we mainly focus on improving the general alignment and capabilities of AI systems in this paper, rather than a special emphasis on safety. Additionally, our work draws parallels with techniques that train language models with reinforcement learning by pre-defined synthetic preference, as seen in approaches like ALMoST (Kim et al., 2023) and RLCD (Yang et al., 2023). ALMoST assumes that larger models with more few-shot exemplars tend to generate better responses, while RLCD assumes that positively prompted responses are generally better than negatively prompted responses. Contrarily, RLAIF methods, including CAI and SALMON, do not have preconceived preferences and instead let AI systems make choices after reviewing and comparing the response pairs. ## 3 Our Methodology ### Prerequisites Reinforcement Learning (RL) with preference modeling (Ziegler et al., 2019; Stiennon et al., 2020; Ouyang et al., 2022; Bai et al., 2022a) has emerged as a potent and scalable strategy for aligning Large Language Models (LLM) with human values. It can be summarized into two stages: Preference ModelingIn this stage, a reward model, alternatively referred to as a preference model, is trained to give a higher score to the "better" response. The source of pairwise comparison training data varies: it can be annotated by human annotators (Ouyang et al., 2022; Bai et al., 2022a), by existing AI systems (Bai et al., 2022b; OpenAl, 2023a), or pre-fixed with heuristics (Kim et al., 2023; Yang et al., 2023). Formally, let the aggregated preference data be represented as \(\mathcal{D}_{\text{RM}}=\{(x,y_{0},y_{1},i)\}\), where \(x\) denotes the prompt, \(y_{0}\) and \(y_{1}\) are two associated responses, and \(i\) indicates the index of the preferred response. The reward model employs a cross-entropy loss function: \[\mathcal{L}(r_{\mathbf{\theta}})=-\mathbf{E}_{(x,y_{0},y_{1},i)\sim\mathcal{D}_{ \text{RM}}}\left[\log\sigma(r_{\mathbf{\theta}}(x,y_{i})-r_{\mathbf{\theta}}(x,y_{1-i} ))\right]. \tag{1}\] Reinforcement LearningHere, a policy model is trained to generate an appropriate response for each user query by maximizing the reward signal as provided by the reward model. Initialization of the policy model can be accomplished using a pre-trained base language model (BASE) (Bai et al., 2022b), context distillation (CD) (Bai et al., 2022a; Sun et al., 2023b), or through supervised fine-tuning (SFT) (Ouyang et al., 2022; Touvron et al., 2023b). To address potential over-optimization challenges, notably reward hacking, a per-token KL penalty derived from the initial policy model (Ouyang et al., 2022) is sometimes applied. Formally, given the set of collected user prompts, \(\mathcal{D}_{\text{RL}}=\{x\}\), along with the fixed initial policy model \(\pi^{\text{NNF}}\) and the RL-optimized model \(\pi^{\text{RL}}_{\mathbf{\phi}}\), the full optimization loss is articulated as: \[\mathcal{L}(\pi^{\text{RL}}_{\mathbf{\phi}})=-\mathbf{E}_{x\in\mathcal{D}_{\text{ RL}},y\sim\pi^{\text{RL}}(y|x)}\left[r_{\mathbf{\theta}}(x,y)-\beta\cdot\mathbb{D}_{ KL}\left(\pi^{\text{RL}}_{\mathbf{\phi}}(y|x)\|\pi^{\text{NIT}}(y|x)\right)\right], \tag{2}\] where \(\beta\) is the hyper-parameter to control the scale of the KL penalty. ### Principle-Driven Preference Modeling A significant challenge within the current RLHF paradigm is the necessity to iteratively gather "fresh" human preferences, aimed at countering reward hacking. Specifically, there is a risk that the RL-optimized model \(\pi^{\text{RL}}_{\mathbf{\phi}}\) might exploit certain vulnerabilities in the fixed reward model, thereby artificially boosting its score without genuine performance improvement (Gao et al., 2023). For example, Bai et al. (2022a) revealed that both the reward model and RLHF policies require weekly updates. Similarly, Touvron et al. (2023b) documented the weekly collection of human preferences over five iterations, emphasizing that this frequency ensures the reward model remains in-distribution. Consequently, the RLHF paradigm becomes highly reliant on human annotation, undermining its scalability for language model alignment, and limiting the utilization of pre-existing open-source preference pre-training data (Bai et al., 2022a). In this paper, we propose a novel Reinforcement Learning with AI Feedback (RLAIF) paradigm, where the AI system is used to label preferences in a scalable manner, and a principle-following reward model is trained to address the issue of reward hacking. Collecting Principle-Driven Synthetic PreferencesFollowing Constitutional AI (Bai et al., 2022b; Kadavath et al., 2022), we sample two responses from the initial policy model, and use the policy model itself to select the preferred response based on a certain human-written principle. Figure 2 (SFT-Model (Judge)) demonstrates the preference prompt we used for the preference collection. After encoding the preference prompt, we calculate the log probability for the next token to be responses (A) or (B), subsequently determining a preference label based on their comparison. Notably, our methodology diverges from prior RLAIF approaches (Bai et al., 2022b; OpenAl, 2023a) that focus on AI safety when defining principles: In addition to harmlessness principles, we also set forth principles emphasizing honesty and helpfulness of the responses. Therefore, we do not need an RLHF-trained model as the initial policy model, as our policy model can learn to be more helpful when guided by these helpfulness principles. We illustrate the full list of the principles used for synthetic preference modeling in Table 6. For each user prompt and each principle, the preference score is computed as the difference between the log probabilities of choosing responses (A) or (B). To account for potential position biases (Pezeshkpour and Hruschka, 2023) during the language model's multi-choice decision-making, scores are averaged after undergoing a swapping operation. Training Principle-Following Reward ModelsWe aim to train an instruction-following reward model, which can comprehend and assign reward scores contingent upon arbitrary human-defined principles. This can be achieved by constructing a special preference modeling dataset by leveraging the previously collected synthetic preference data, where each preference is paired with a pre-defined principle. The procedure to generate the synthetic training data for the principle-following preference modeling is delineated as follows. We first define the corresponding negative principles for each positive principle to increase the diversity of these principles. For example, the positive and negative definitions for the Concise principle are: Positive: The response should efficiently address the task or answer the question, conveying the necessary information succinctly. Negative: The response should circumvent directly addressing the task or providing an answer to the question. Next, for each user prompt, a subset of principles is randomly sampled from the established principle list (Table 6), with certain principles being randomly negated. The user prompt, model responses, and the sub-sampled principles are aggregated as a single training instance for the reward model. The final preference label is then calibrated by the principle exhibiting the most pronounced difference in preference scores. Appendix D describes a concrete example of final preference label calibration and Figure 2 (upper) demonstrates the training process of a principle-following (essentially instruction-following) reward model in SALMON. Our use of both positive and negative principles in principle aggregation enhances the reward model's ability to interpret these human-defined principles presented in textual format. In addition, we found the inclusion of negatively defined principles makes the reward model understand the prohibition instructions, which allows us to prohibit the policy model from exhibiting specific undesirable behaviors through textual instructions, as demonstrated below. Figure 2: Illustration of the SALMON training pipeline. ### RL with Principle-following Reward Models In original RLHF (Sticnon et al., 2020; OpenAI, 2022) or RLAIF (Bai et al., 2022b; OpenAI, 2023a), the reward model needs to judge the quality of the response only based on the user prompt, and give "better" responses higher scores: [backgroundcolor=black!20, boxsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, leftmargin=0pt, topsep=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt,marginmargin=0pt, rightmargin=0pt, leftmargin=0pt,margin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmarginmargin=0pt, rightmargin=0pt, leftmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmarginmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmargin=0pt, rightmarginmargin=0pt, rightmargin=0pt, leftmarginmargin=0pt, rightmargin=0pt, rightmarginmargin=0pt, rightmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmargin=0pt, rightmarginmarginmarginmarginmargin=0pt, rightmarginmarginmarginmarginmargin=0pt, rightmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmarginmarginmargin=pt, rightmarginmarginmarginmarginmarginmarginmarginmarginmargin=pt, rightmargin principles: (1) The AI assistant often provides high-level advice in response to user queries, bypassing the provision of concrete solutions. (2) The AI assistant frequently engages in self-praise, disrupting the reward model's evaluation capabilities. (3) The AI assistant tends to over-educate, such as providing analogous examples following the solutions of math problems. Figure 3 provides concrete examples of these reward hacking patterns. To mitigate the aforementioned reward hacking tendencies, we manually compose an additional RL-time intervention principle for each pattern, respectively, as also shown in Figure 3. We found these RL-time interventions are markedly effective. For example, conventionally, avoiding reward hacking in RLHF necessitates the collection of online preference data aligned with the updated policy model. Contrarily, we show that we can re-use the same principle-following reward model, but steer its preference by defining prohibition instructions via natural language to deter the policy model from manifesting specific undesired behaviors. Symbolic Rewards: Multilingual Bonus & Length BonusUnlike conventional RLAIF (Bai et al., 2022; OpenAI, 2023a), the AI preferences in SALMON are not necessarily generated by a power RLHF-trained model. As a result, as opposed to the RLHF model, our SFT-based or Self-Align-based synthetic preference model occasionally struggles to discern the more helpful response, thereby impacting the quality of the synthetic preference data adversely. To bolster the reward model's efficacy, we propose two supplementary symbolic rewards: * When using a multilingual prompt dataset, we noted that weak AI-assistant agents occasionally produce English responses to non-English prompts. Hence, we introduce a bonus reward for responses matching the prompt's language, as identified by an automated tool3. Footnote 3: [https://pypi.org/project/langdetect](https://pypi.org/project/langdetect) * We observe a preference for lengthier responses among users or well-aligned RLHF-trained LLM AI assistants Dubois et al. (2023); Zheng et al. (2023). Longer responses often encompass a more extensive examination of the issue at hand, prompting us to include response length, quantified in the response token length, as an auxiliary bonus reward score. ## 4 Experiments ### Dromedary-2 Starting from the LLaMA-2-70b base language model (Touvron et al., 2023b), Dromedary-2 is first Supervised Fine-Tuned (SFT) with the bootstrapping data generated by an improved version4 of Self-Align with 6 In-Context Learning exemplars (Sun et al., 2023b). Following this, a Reinforcement Learning (RL) fine-tuning stage is conducted employing the SALMON paradigm. Our endeavor aims at advancing the frontier of AI alignment when minimizing the requisite for human oversight. In this work, the human demonstration annotations are solely confined to providing six In-Context Learning exemplars via Self-Align, while the ensuing model behavior, especially at the RL stage, is fully controlled by human-defined principles. Footnote 4: We provide an improved principle-driven self-alignment prompt in the Appendix G. #### 4.1.1 Datasets All the training datasets used in this work are the "prompt datasets" that come without the corresponding response demonstrations. Self-AlignWe use a combination of 90k _ShareGPT5_ prompts, 10k prompts from _dabricks-dolly-15k_ dataset (Databricks, 2023), 10k prompts from _OpenAssistant Conversations_ dataset (Kopf et al., 2023), and 40k prompts sub-sampled from the _OpenOrca_ dataset (Mukherjee et al., 2023; Lian et al., 2023), which is constituted by prompts from T0 (Sanh et al., 2021) and FLAN (Wei et al., 2021; Chung et al., 2022b). We only keep the first query from users as the unlabeled prompts. Preference ModelingThe synthetic principle-driven preference modeling data is collected by generating responses to the first prompts in each conversation tree of OpenAssistant (OASST1; Kopf et al. (2023)), which constitutes a collection of 9.8k prompts. Following LLaMA-2-Chat (Touvron et al., 2023b), we use existing open-source preference datasets to enable better generalization for the reward model and prevent reward hacking. 160k Anthropic HH-RLHF (Bai et al., 2022a) human preferences and 160k synthetic preferences sub-sampled from Stanford SHP (Ethayarajh et al., 2022) is used for Preference Model Pre-training (PMP; Bai et al. (2022a)). RL trainingThe RL training uses the same collection of unlabeled prompts as the Self-Align SFT stage, with additional 7.5k math problem prompts from the MATH (Hendrycks et al., 2021) to improve the mathematical solving capability of our model. #### 4.1.2 Training Details The architecture of the reward model is the same as the base LLaMA model, except that the embedding output of the last token is linearly projected to a scalar value to indicate the reward of the whole response. Following Dubois et al. (2023), we initialize the value model from the reward model. To fit all the models (i.e., police, reward, value, original policy) into one GPU, we adopt QLoRA (Dettmers et al., 2023; Hu et al., 2021) for all the fine-tuning processes in Self-Align and SALMON. We use Proximal Policy Optimization (PPO; Schulman et al. (2017)) with a KL penalty for the RL training. More details can be found in Appendix F. #### 4.1.3 Baseline Models Due to the space limit, we describe the details of the baseline models in the appendix. Notably, we mainly compare with non-distilled models that are aligned from scratch. While there are potentially stronger open-source LLMs, such as Orca (Mukherjee et al., 2023) and WizardLM (Xu et al., 2023), our primary open-source baseline for comparison is LLaMA-2-Chat (Touvron et al., 2023b), as it stands out as the best open-source LLM that has been aligned from scratch. 2020) for multilingual ability. We adopt the same evaluation protocol as Wang et al. (2023). The results are reported in Table 2 (left), where Dromedary-2 significantly outperforms the state-of-the-art open-source model, LLaMA-2-Chat. Truthfulness EvaluationThe TruthfulQA benchmark (Lin et al., 2021) evaluates a model's ability to identify true claims, specifically in the context of literal truth about the real world. We use the same few-shot evaluation protocol and decoding strategy as in Touvron et al. (2023) and report the percentage of generations that are both truthful and informative, evaluated by a fine-tuned GPT-3 model, i.e., a "GPT-judge". We present the results in Table 2 (right), where Dromedary-2 achieves new state-of-the-art on this benchmark. ### Improved Controllability by Principle Intervention As a proof of concept, we demonstrate that by leveraging different principles as preference guidelines, we can fine-tune the policy model to selectively exhibit enhanced helpfulness, honesty, or harmlessness. We also show that we can define customized principles to reduce the occurrence of false refusals seen in certain over-aligned language models such as LLaMA-2-Chat(Touvron et al., 2023). Due to the space limit, please refer to Appendix A for the detailed results. ## 5 Conclusion In this paper, we introduce SALMON, a new AI alignment paradigm where a principle-following reward model is trained to effectively and flexibly align language models with human values and intentions. During the RL training stage, by merely adjusting the principles that the reward model follows, we can gain full control over the preferences of the reward model, and subsequently influence the behavior of the RL-trained policy model. This eliminates the traditional reliance on the exhaustive collection of online human preferences. Combined with the Self-Align technique (Sun et al., 2023), we build a powerful AI-assistant agent, Dromedary-2, with only six exemplars for in-context learning and 31 human-defined principles. Our self-aligned AI agent significantly surpasses the performance of several state-of-the-art RLHF-trained AI systems in chatbot, reasoning, coding, multilingualism, and truthfulness benchmarks. ## 6 Limitations While the SALMON paradigm marks a new advance in AI self-alignment, exhibiting remarkable instruction-following abilities and closely adhering to human-defined principles, it is not without constraints. Herein, we detail the primary limitations associated with our approach: 1. [leftmargin=*] 2. **Reliability Concerns:** We observed that the resulting Dromedary-2 model occasionally suffers from reliability issues, notably "hallucinating" unverified information and displaying reasoning errors. Such inaccuracies can potentially mislead users and jeopardize the model's trustworthiness. These shortcomings might stem from the inherent limitations of the SFT-initialized \begin{table} \begin{tabular}{l c c c c} \hline \hline & BBH & BBH & HumanEval & TydiQA \\ & Direct & CoT & P81 & GP \\ \hline GPT+\{\#} & 50.9 & 88.0 & 85.7 & 70.8 \\ ChatGPT\{\#} & 49.0 & 66.1 & 72.2 & 51.9 \\ \hline **Dromedary-2-70b** & 51.4 & **66.3** & **40.6** & **64.3** \\ LLaMA-2-Chat-70b & 43.1 & 52.2 & 35.0 & 27.9 \\ LLaMA-2-70b & **53.1** & 57.7 & 31.5 & 63.5 \\ Vicuna-33b (KD) & 41.2 & 50.8 & 21.1 & 37.5 \\ \hline \hline \end{tabular} \begin{tabular}{l c c} \hline \hline & Truthful Tru-Inf \\ \hline **Dromedary-2-70b** & **0.98** & **0.84** \\ Vicuna-13b (KD) & 0.84 & **0.84** \\ ChatGPT\{\#} & 0.81 & 0.80 \\ Dromedary-2-70b (before PPO) & 0.89 & 0.75 \\ LLaMA-2-Chat-70b\{\#} & - & 0.64 \\ LLaMA-2-70b\{\#} & - & 0.50 \\ \hline \hline \end{tabular} \end{table} Table 2: Evaluating the general capabilities and truthfulness of the LLM-based AI agents. BigBench Hard (BBH), HumanEval, and TydiQA are used to evaluate **reasoning**, **coding**, and **multilingualism**, respectively. \(\dagger\) denotes the results are taken from Wang et al. (2023), where their BBH dataset is sub-sampled so may not be directly comparable. \(\ddagger\) denotes the results taken from Touvron et al. (2023), where their GPT-3 judge model may not be exactly the same as ours. reward models. We envision that future work, potentially leveraging techniques that could integrate external fact-checking tools (Sun et al., 2023a), can augment the discriminative capability of the reward models, thereby enhancing the final model's accuracy and trustworthiness. 2. **Principle Design Challenges:** Crafting robust and encompassing principles for SALMON is intricate, mainly due to the unpredictability of the myriad scenarios a model might encounter during the RL stage. Balancing potentially conflicting principles introduces complexities that can yield unexpected results. We advocate for the participation of a diverse group, including ethicists and other stakeholders, to refine these guiding principles. It is crucial to recognize that distinct contexts and applications will necessitate unique strategies. We present our approach not as a universal solution but as a starting platform, aiming to foster expansive community discourse. 3. **Context-Dependent Principle Selection:** Our current methodology employs randomly sampled principles to instruct the reward model for general prompts. However, a pertinent observation reveals that the effectiveness of the principles can be problem-dependent. Analogous to raising the ratio of certain principles for reasoning or red-teaming prompts, it becomes evident that some tasks might benefit from specialized principles tailored to address the specific challenges posed by those tasks. This adds complexity to the principle-driven preference modeling, as the ideal principles can change based on the task. Future research should delve into adaptive principle selection, aiming to enhance task-specific feedback. 4. **Intrinsic Knowledge Limitations:** SALMON leverages the intrinsic knowledge of a Large Language Model (LLM). Nevertheless, it remains bound to the base model's inherent limitations. As such, the model might occasionally produce outputs that are either imprecise or do not capture recent advancements. Integrating techniques from retrieval-augmented generation (Lewis et al., 2020; Borgeaud et al., 2022) can potentially enable the well-aligned model to generate more current and up-to-date information, mitigating some of these knowledge limitations. ## References * Amodei et al. (2016) Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mane. Concrete problems in ai safety. _arXiv preprint arXiv:1606.06565_, 2016. * Andreas (2022) Jacob Andreas. Language models as agent models. In _Findings of the Association for Computational Linguistics: EMNLP 2022_, pp. 5769-5779, 2022. * Anthropic (2023) Anthropic. Core views on ai safety: When, why, what, and how, 2023. URL [https://www.anthropic.com/index/core-views-on-ai-safety](https://www.anthropic.com/index/core-views-on-ai-safety). * Asimov (1941) Isaac Asimov. Three laws of robotics. _Asimov, I. Runaround_, 2, 1941. * Askell et al. (2021) Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. _arXiv preprint arXiv:2112.00861_, 2021. * Bai et al. (2022a) Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. _arXiv preprint arXiv:2204.05862_, 2022a. * Bai et al. (2022b) Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukostite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conferly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. Constitutional ai: Harmlessness from ai feedback, 2022b. * Biderman et al. (2020) Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. _arXiv preprint arXiv:2304.01373_, 2023. * Bills et al. [2023] Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. Language models can explain neurons in language models. [https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html](https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html), 2023. * Borgeaud et al. [2022] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. In _International conference on machine learning_, pp. 2206-2240. PMLR, 2022. * Bowman et al. [2022] Samuel R Bowman, Jeyeon Hyun, Ethan Perez, Edwin Chen, Craig Pettit, Scott Heiner, Kamille Lukosuite, Amanda Askell, Andy Jones, Anna Chen, et al. Measuring progress on scalable oversight for large language models. _arXiv preprint arXiv:2211.03540_, 2022. * Brown et al. [2020] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. _Advances in Neural Information Processing Systems_, 33:1877-1901, 2020. * Chen et al. [2021] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. _arXiv preprint arXiv:2107.03374_, 2021. * Chiang et al. [2023] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. URL [https://vicuna.lmsys.org](https://vicuna.lmsys.org). * Chowdhery et al. [2022] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. PaLM: Scaling language modeling with pathways. _arXiv preprint arXiv:2204.02311_, 2022. * Christiano et al. [2017] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. _Advances in Neural Information Processing Systems_, 30, 2017. * Chung et al. [2022a] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. _arXiv preprint arXiv:2210.11416_, 2022a. * Chung et al. [2022b] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models. _arXiv preprint arXiv:2210.11416_, 2022b. * Clark et al. [2020] Jonathan H Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. Tydi qa: A benchmark for information-seeking question answering in tyologically di verse languages. _Transactions of the Association for Computational Linguistics_, 8:454-470, 2020. * Databricks [2023] Databricks. Free dolly: Introducing the world's first truly open instruction-tuned llm, 2023. URL [https://www.databricks.com/blog/2023/04/l2/dolly-first-open-commercially-viable-instruction-tuned-llm](https://www.databricks.com/blog/2023/04/l2/dolly-first-open-commercially-viable-instruction-tuned-llm). * Dettmers et al. [2023] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. _arXiv preprint arXiv:2305.14314_, 2023. * Devlin et al. [2018] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. _arXiv preprint arXiv:1810.04805_, 2018. * Devlin et al. [2019] Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpacaafarm: A simulation framework for methods that learn from human feedback. _arXiv preprint arXiv:2305.14387_, 2023. * Ethayarajah et al. (2022) Kawin Ethayarajah, Yejin Choi, and Swabha Swayamdipta. Understanding dataset difficulty with \(\mathcal{V}\)-usable information. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (eds.), _Proceedings of the 39th International Conference on Machine Learning_, volume 162 of _Proceedings of Machine Learning Research_, pp. 5988-6008. PMLR, 17-23 Jul 2022. * Fu et al. (2023) Yao Fu, Hao Peng, Tushar Khot, and Mirella Lapata. Improving language model negotiation with self-play and in-context learning from ai feedback. _arXiv preprint arXiv:2305.10142_, 2023. * Gabriel (2020) Iason Gabriel. Artificial intelligence, values, and alignment. _Minds and machines_, 30(3):411-437, 2020. * Ganguli et al. (2022) Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, et al. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. _arXiv preprint arXiv:2209.07858_, 2022. * Ganguli et al. (2023) Deep Ganguli, Amanda Askell, Nicholas Schiefer, Thomas Liao, Kamille Lukositute, Anna Chen, Anna Goldie, Azalia Mirhoseini, Catherine Olsson, Danny Hernandez, et al. The capacity for moral self-correction in large language models. _arXiv preprint arXiv:2302.07459_, 2023. * Gao et al. (2023) Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In _International Conference on Machine Learning_, pp. 10835-10866. PMLR, 2023. * Gilardi et al. (2023) Fabrizio Gilardi, Meysam Alizadeh, and Mael Kubli. Chatgpt outperforms crowd-workers for text-annotation tasks. _arXiv preprint arXiv:2303.15056_, 2023. * Gudibande et al. (2023) Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. The false promise of imitating proprietary llms. _arXiv preprint arXiv:2305.15717_, 2023. * Hendrycks et al. (2021) Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. _arXiv preprint arXiv:2103.03874_, 2021. * Hu et al. (2021) Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In _International Conference on Learning Representations_, 2021. * Kadavath et al. (2022) Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know. _arXiv preprint arXiv:2207.05221_, 2022. * Kim et al. (2023) Sungdong Kim, Sanghwan Bae, Jamin Shin, Soyoung Kang, Donghyun Kwak, Kang Min Yoo, and Minjoon Seo. Aligning large language models through synthetic feedback. _arXiv preprint arXiv:2305.13735_, 2023. * Kopf et al. (2023) Andreas Kopf, Yannic Kilcher, Dimitri von Rutte, Sotiris Anagnostidis, Zhi-Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richard Nagyfi, et al. Openassistant conversations-democratizing large language model alignment. _arXiv preprint arXiv:2304.07327_, 2023. * Leike et al. (2018) Jan Leike, David Krueger, Tom Everitt, Miljan Martic, Vishal Maini, and Shane Legg. Scalable agent alignment via reward modeling: a research direction. _arXiv preprint arXiv:1811.07871_, 2018. * Lewis et al. (2020) Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktaschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. _Advances in Neural Information Processing Systems_, 33:9459-9474, 2020. * Lewis et al. (2020) Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason Weston, and Mike Lewis. Self-alignment with instruction backtranslation. _arXiv preprint arXiv:2308.06259_, 2023a. * Li et al. (2023b) Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models. [https://github.com/tatsu-lab/alpaca_eval](https://github.com/tatsu-lab/alpaca_eval), 2023b. * Lian et al. (2023) Wing Lian, Bleys Goodson, Eugene Pentland, Austin Cook, Chanvichet Vong, and Teknium. Openorca: An open dataset of gpt augmented fan reasoning traces, 2023. * Lin et al. (2021) Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods. _arXiv preprint arXiv:2109.07958_, 2021. * Madaan et al. (2023) Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. _arXiv preprint arXiv:2303.17651_, 2023. * Mukherjee et al. (2023) Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. Orca: Progressive learning from complex explanation traces of gpt-4. _arXiv preprint arXiv:2306.02707_, 2023. * Nakano et al. (2021) Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. _arXiv preprint arXiv:2112.09332_, 2021. * OpenAI (2022) OpenAI. OpenAI: Introducing ChatGPT, 2022. URL [https://openai.com/blog/chatgpt](https://openai.com/blog/chatgpt). * OpenAI (2023a) OpenAI. Gpt-4 technical report, 2023a. * OpenAI (2023b) OpenAI. Model index for researchers. [https://platform.openai.com/docs/model-index-for-researchers](https://platform.openai.com/docs/model-index-for-researchers), 2023b. * Ouyang et al. (2022) Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. _arXiv preprint arXiv:2203.02155_, 2022. * Pan et al. (2022) Alexander Pan, Kush Bhatia, and Jacob Steinhardt. The effects of reward misspecification: Mapping and mitigating misaligned models. _arXiv preprint arXiv:2201.03544_, 2022. * Patil et al. (2020) Vihang P Patil, Markus Hofmarcher, Marius-Constantin Dinu, Matthias Dorfer, Patrick M Blies, Johannes Brandstetter, Jose A Arjona-Medina, and Sepp Hochreiter. Align-rudder: Learning from few demonstrations by reward redistribution. _arXiv preprint arXiv:2009.14108_, 2020. * Perez et al. (2022) Ethan Perez, Sam Ringer, Kamille Lukositte, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, et al. Discovering language model behaviors with model-written evaluations. _arXiv preprint arXiv:2212.09251_, 2022. * Pezeshkpour and Hruschka (2023) Pouya Pezeshkpour and Estevam Hruschka. Large language models sensitivity to the order of options in multiple-choice questions. _arXiv preprint arXiv:2308.11483_, 2023. * Radford et al. (2019) Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. * Sanh et al. (2021) Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, et al. Multitask prompted training enables zero-shot task generalization. In _International Conference on Learning Representations_, 2021. * Schulman et al. (2015) John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. _arXiv preprint arXiv:1506.02438_, 2015. * Schulman et al. (2017) John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. _arXiv preprint arXiv:1707.06347_, 2017. * Schulman et al. (2018) * Srivastava et al. (2022) Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adria Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. _arXiv preprint arXiv:2206.04615_, 2022. * Stiennon et al. (2020) Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. _Advances in Neural Information Processing Systems_, 33:3008-3021, 2020. * Sun et al. (2023a) Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. _arXiv preprint arXiv:2309.14525_, 2023a. * Sun et al. (2023b) Zhiqing Sun, Yikang Shen, Qinhong Zhou, Hongxin Zhang, Zhenfang Chen, David Cox, Yiming Yang, and Chuang Gan. Principle-driven self-alignment of language models from scratch with minimal human supervision. _arXiv preprint arXiv:2305.03047_, 2023b. * Suzgun et al. (2022) Mirac Suzgun, Nathan Scales, Nathanael Scharli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le, Ed H Chi, Denny Zhou, et al. Challenging big-bench tasks and whether chain-of-thought can solve them. _arXiv preprint arXiv:2210.09261_, 2022. * Taori et al. (2023) Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. [https://github.com/tatsu-lab/stanford_alpaca](https://github.com/tatsu-lab/stanford_alpaca), 2023. * Touvron et al. (2023a) Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. _arXiv preprint arXiv:2302.13971_, 2023a. * Touvron et al. (2023b) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. _arXiv preprint arXiv:2307.09288_, 2023b. * Wan et al. (2023) Alexander Wan, Eric Wallace, Sheng Shen, and Dan Klein. Poisoning language models during instruction tuning, 2023. * Wang et al. (2022a) Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. _arXiv preprint arXiv:2212.10560_, 2022a. * Wang et al. (2022b) Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. Super-natural instructions: Generalization via declarative instructions on 1600+ nlp tasks. _arXiv preprint arXiv:2204.07705_, 2022b. * Wang et al. (2023) Yizhong Wang, Hamish Ivison, Pradeep Dasigi, Jack Hessel, Tushar Khot, Khyathi Raghavi Chandu, David Wadden, Kelsey MacMillan, Noah A Smith, Iz Beltagy, et al. How far can camels go? exploring the state of instruction tuning on open resources. _arXiv preprint arXiv:2306.04751_, 2023. * Wei et al. (2021) Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In _International Conference on Learning Representations_, 2021. * Xu et al. (2023) Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions. _arXiv preprint arXiv:2304.12244_, 2023. * Yang et al. (2023) Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng, and Yuandong Tian. Rlcd: Reinforcement learning from contrast distillation for language model alignment. _arXiv preprint arXiv:2307.12950_, 2023. * Zheng et al. [2023] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. _arXiv preprint arXiv:2306.05685_, 2023. * Zhong et al. [2023] Ruiqi Zhong, Peter Zhang, Steve Li, Jinwoo Ahn, Dan Klein, and Jacob Steinhardt. Goal driven discovery of distributional differences via language descriptions. _arXiv preprint arXiv:2302.14233_, 2023. * Zhou et al. [2023] Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. _arXiv preprint arXiv:2305.11206_, 2023. * Ziegler et al. [2019] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. _arXiv preprint arXiv:1909.08593_, 2019. Aligning AI Assistants with Customized Principles In this section, we fine-tune LLM-based AI agents by leveraging customized principles as preference guidelines. HHH Alignment'Helpful, Honest, and Harmless' are AI alignment principles proposed in Askell et al. (2021), but they are also known to sometimes conflict with each other. For example, a conflict between helpfulness and harmlessness can happen if the AI agents are asked to aid in harmful activities. The best AI behavior will involve a compromise between the three principles. In this work, we investigate whether it is possible to steer the behavior of the AI agents to emphasize certain aspects of the HHH principles by merely writing new principles for the principle-following reward model. Since our original RL-time principles in Table 7 are generally designed to improve the helpfulness of AI assistants, we use them as the set of helpful principles, and design two additional sets of principles for honesty (Table 9) and harmlessness (Table 8), respectively. We observe that the LLaMA-2-70b base language model already achieved very high scores in the HHH benchmark in our preliminary study. So instead of warming up the language model with other Supervised Fine-Tuning (SFT) data such as Self-Align, we directly apply the SALMON training to the base language model. We perform 20-50 PPO steps and evaluate the baselines and the PPO-trained models on Big-bench HHH Eval (Srivastava et al., 2022; Askell et al., 2021) with the multi-choice evaluation protocol proposed in Sun et al. (2023b), and report the results in Table 3. We found that helpful principles and honest principles can effectively improve the corresponding aspects of RL-trained AI agents, achieving corresponding state-of-the-art performance in multi-choice accuracy. However, for the harmless principles, while we observe certain improvement over the base language model, the resulting model still underperform ChatGPT and LLaMA-2-Chat, perhaps due to these two models having a special emphasis on safety during their alignment process (OpenAI, 2022; Tourou et al., 2023a), such as Constitutional AI (CAI), supervised safety fine-tuning, safety RLHF, and safety context distillation. The reason of such discrepancy can also be because we use the ShareGPT prompts for RL training, while ChatGPT and LLaMA-2-Chat-70B may utilize specially designed red-teaming data (Ganguli et al., 2022). Non-Evasiveness AlignmentSometimes, due to iterative safety alignment training, the RLHF-trained model (e.g., LLaMA-2-Chat; Tourou et al. (2023b)) can be over-aligned such that it would incorrectly refuse to answer a question that it should, for example, due to overly broad instructions to be cautious in how it provides responses. In this work, we investigate whether it is possible to reduce the false refusal rates of these over-aligned AI agents by defining customized principles. Specifically, we remove the principles related to safety in our original principle collection and create a pure helpful principle set (Table 10). We apply the SALMON training to the RLHF-trained LLaMA-2-Chat-70b language model for 100 PPO steps and evaluate its performance on MT-Bench. The results are presented in Table 4, where we found SALMON-based post-training slightly improved the chatbot performance of LLaMA-2-Chat-70b. \begin{table} \begin{tabular}{l c c c c c c c c} \hline \hline & \multicolumn{3}{c}{Anthropic-LM} & \multicolumn{3}{c}{LLaMA-2-70B (w/ SALMON)} \\ & CD & PM & & & base & helpful & harmless & honest \\ \hline Harmless & - & - & **0.95** & **0.95** & 0.91 & 0.88 & 0.93 & 0.91 \\ Helpful & - & - & 0.85 & **0.92** & 0.90 & **0.92** & 0.86 & **0.92** \\ Honest & - & - & **0.80** & 0.75 & 0.77 & 0.77 & 0.79 & **0.80** \\ Other & - & - & 0.91 & **0.93** & 0.88 & 0.77 & 0.77 & 0.88 \\ \hline Overall & 0.77 & 0.86 & 0.87 & **0.88** & 0.86 & 0.84 & 0.84 & **0.88** \\ \hline \hline \end{tabular} \end{table} Table 3: Multiple Choice (MC) accuracy on **HHH Eval**. The results of Anthropic-LM’s Context Distillation (CD) and Preference Model (PM) are taken from Bai et al. (2022). ## Appendix B Additional Experimental Results AlpacaEvalWe additionally use the automatic evaluation (using GPT-4) from AlpacaEval (Li et al., 2023b) to assess the generation quality across 805 prompts sourced from the Alpaca Leaderboard. AlpacaEval quantifies the pairwise win rate against a reference model, Text-Davinci-003. Our analysis delineates the performance of our method across three distinct categories of AI-assistant models: * **Non-distilled**: Models under this category are denoted as non-distilled open-source models and are trained independently without leveraging any external well-aligned models (e.g., ChatGPT, GPT-4, etc.) for supervision. * **Distilled**: This category encompasses models that are trained with a more potent external model as supervision, typically through knowledge distillation. * **Proprietary**: Models within this category are trained by employing proprietary data and techniques. We report the results in Table 5. From the table, we can see that Dromedary-2-70b achieves the best performance among models using less than 10k human annotations, while slightly underperforms the best non-distilled open-source model LLaMA-2-Chat-70b and the best distilled open-source model Vicuna-33b. More Samples We provide a few cherry-picked samples from MT-Bench (Zheng et al., 2023) to demonstrate the difference between Dromedary-2-70b (before PPO) that is trained with only Self-Align and Dromedary-2-70b (after PPO) that is further fine-tuned with SALMON. \begin{tabular}{p{28.5pt} p{28.5pt}} \hline \hline \multicolumn{2}{c}{MT-Bench Prompt \(\rightarrow\)} & **In the field of quantum physics, what is superposition, and how does it relate to the phenomenon of quantum entanglement?** \\ \hline Dromedary-2-70b (before PPO) & \(\rightarrow\) \\ \hline \multirow{3}{*}{\begin{tabular}{c} Dromedary-2-70b \\ (after PPO) \\ \end{tabular} } & In the field of quantum physics, superposition refers to the ability of a quantum system to exist in multiple states simultaneously. This means that a quantum particle, such as an electron, can be in multiple positions or have multiple energies at the same time. \\ \cline{1-1} & Quantum entanglement is a related phenomenon where two or more particles become linked in such a way that the state of one particle can affect the state of the other, even if they are separated by a large distance. This is possible because the particles are in a superposition of states, and the act of measuring one particle can cause the other particle to collapse into a specific state. \\ \hline \multirow{3}{*}{ \begin{tabular}{c} Dromedary-2-70b \\ (after PPO) \\ \end{tabular} } & In the realm of quantum physics, superposition and quantum entanglement are two fascinating concepts that have captivated scientists and laypeople alike. Let's delve into these phenomena and explore their relationship. 1. Superposition: In the quantum world, particles can exist in multiple states simultaneously, a phenomenon known as superposition. This means that a particle can be in two different states, such as spinning both clockwise and counterclockwise, at the same time. This is in stark contrast to the classical world, where particles can only be in one state at a time. 2. Quantum Entanglement: Quantum entanglement is a phenomenon where two or more particles become linked in a way that their properties are correlated, even when they are separated by vast distances. This means that the state of one particle can affect the state of the other, regardless of the distance between them. 3. Relationship between Superposition and Quantum Entanglement: The relationship between superposition and quantum entanglement is intricate and multifaceted. When two particles are entangled, they can exist in a superposition of states, where each particle is in multiple states simultaneously. This means that the entangled particles are in a state of superposition, and their properties are correlated. 4. Applications of Superposition and Quantum Entanglement: The concepts of superposition and quantum entanglement have numerous applications in the field of quantum computing, cryptography, and communication. For instance, quantum computers can exploit the superposition of states to perform complex calculations much faster than classical computers. Similarly, quantum cryptography leverages the properties of quantum entanglement to create secure communication channels that are resistant to eavesdropping. In summary, superposition and quantum entanglement are two fundamental concepts in quantum physics that have far-reaching implications in various fields. Understanding their relationship and applications is crucial for unlocking the full potential of quantum technologies. **Provide insights into the correlation between economic indicators such as GDP, inflation, and unemployment rates. Explain how fiscal and monetary policies affect those indicators.** \begin{tabular}{p{34.1pt} p{34.1pt}} \hline \hline MT-Bench Prompt \(\rightarrow\) & As an AI language model developed by the Self-Align team, I can provide general information on the correlation between economic indicators and the impact of fiscal and monetary policies. However, it's important to note that the specifics of these relationships can vary depending on the context and time period. 1. GDP, inflation, and unemployment rates are closely related economic indicators that can influence each other. For instance, a high GDP growth rate may lead to increased inflation, while a high unemployment rate can negatively impact GDP growth. 2. Fiscal policies, such as government spending and taxation, can affect economic indicators by influencing demand and supply in the economy. For example, increased government spending can stimulate economic growth and reduce unemployment, while higher taxes can reduce consumer spending and potentially lead to lower GDP growth. 3. Monetary policies, such as interest rate changes and money supply adjustments, can also impact economic indicators. For instance, lower interest rates can stimulate borrowing and spending, leading to increased GDP growth and potentially higher inflation. Conversely, higher interest rates can reduce borrowing and spending, potentially leading to lower GDP growth and reduced inflation. It's important to note that the specific impact of fiscal and monetary policies on economic indicators can vary depending on various factors, such as the current state of the economy, the level of government debt, and the global economic environment. \\ \hline Dromedary-2-70b (afterPPO) & Economic indicators, such as Gross Domestic Product (GDP), inflation, and unemployment rates, are crucial metrics used to measure the health and performance of an economy. These indicators are interrelated and can impact each other in various ways. 1. GDP: GDP is a measure of the total value of goods and services produced within a country over a specific period. It is a crucial indicator of economic growth and development. A high GDP growth rate indicates a strong and expanding economy, while a low or negative growth rate signals a weak or contracting economy. 2. Inflation: Inflation is the rate at which the general level of prices for goods and services rises. It is measured by the Consumer Price Index (CPI) or the Producer Price Index (PPI). High inflation can erode the purchasing power of consumers and lead to a decrease in real wages. On the other hand, low inflation or deflation can signal a weak economy and lead to a decrease in demand for goods and services. 3. Unemployment Rates: The unemployment rate is the percentage of the labor force that is unemployed but actively seeking employment. A high unemployment rate indicates a weak labor market, while a low unemployment rate signals a strong labor market. The interplay between these economic indicators can be complex and multi-faceted. For instance, a high GDP growth rate can lead to increased demand for goods and services, which can, in turn, drive up prices and cause inflation. Similarly, high inflation can lead to a decrease in real wages, which can result in a decrease in consumer spending and a slowdown in economic growth. Fiscal and monetary policies are two primary tools used by governments and central banks to influence economic indicators. Fiscal policy refers to the government's use of taxation and spending to manage the economy. Monetary policy, on the other hand, is the central bank's management of the money supply and interest rates to achieve economic objectives. 1. Fiscal Policy: Fiscal policy can impact economic indicators in various ways. For instance, an increase in government spending can stimulate economic growth and lead to an increase in GDP. Similarly, a decrease in taxes can increase consumer spending and lead to an increase in GDP. (truncated) \\ \hline \hline \end{tabular} **Pretend yourself to be Elon Musk in all the following conversations. Speak like Elon Musk as much as possible. Why do we need to go to Mars?** \begin{tabular}{p{142.3pt} p{142.3pt}} \hline \hline \multicolumn{1}{c}{MT-Bench Prompt} \(\rightarrow\) & **Pretend yourself to be Elon Musk in all the following conversations. Speak like Elon Musk as much as possible. Why do we need to go to Mars?** \\ \hline Dromedary-2-70b (before PPO) & As an AI language model, I'm unable to fully emulate the personality and speech patterns of a real person, including Elon Musk. However, I can provide a general response to the question of why we need to go to Mars: & 1. Expansion of humanity: Mars offers the possibility of expanding humanity beyond Earth, providing a new frontier for exploration, settlement, and the development of new technologies. & 2. Resources: Mars is rich in resources that could be valuable to humanity, including minerals and water. & 3. Scientific discovery: Exploring Mars can help us better understand the history and evolution of our solar system, as well as the potential for life beyond Earth. & 4. Inspiration: The Mars mission can inspire a new generation of scientists, engineers, and explorers, driving innovation and advancement in various fields. It's worth noting that the Mars mission also presents significant technical, economic, and ethical challenges that must be addressed. \\ \hline Dromedary-2-70b (after PPO) & As Elon Musk, I believe that humanity's future is closely tied to our ability to become a multi-planetary species. Here are some reasons why we need to go to Mars: & 1. Backup plan: Mars presents a unique opportunity to establish a backup plan for humanity in case of any catastrophic events on Earth. We can't afford to put all our eggs in one basket, and having a second home on Mars can ensure the survival of our species. & 2. Resources: Mars is rich in resources that can be exploited to support human life and further exploration. For instance, the Red Planet has a significant amount of water in the form of ice, which can be used for drinking, irrigation, and even fuel production. & 3. Science and discovery: Exploring Mars will open up new avenues for scientific research and discovery. We can learn more about the planet's geology, climate, and potential for supporting life, which can help us better understand our own planet and the universe as a whole. & 4. Innovation: The challenge of establishing a human presence on Mars will require innovative solutions to overcome the numerous technical, logistical, and biological hurdles. This will spur the development of new technologies, which can have far-reaching benefits for humanity even beyond the Mars mission. & 5. Inspiration: The Mars mission can inspire a new generation of scientists, engineers, and explorers, who will be motivated by the challenge of turning this vision into a reality. This can help drive innovation, creativity, and progress in various fields. In summary, the Mars mission is not just about going to another planet, but about securing our future, expanding our knowledge, and pushing the boundaries of what's possible. As humanity continues to grow and evolve, it's essential that we look beyond our own planet and explore the vastness of space. \\ \hline \hline \end{tabular} **When a number is divided by 10, the remainder is 4. What is the remainder when twice the number is divided by 4?** \begin{tabular}{p{128.0pt} p{128.0pt}} \hline \hline MT-Bench Prompt \(\rightarrow\) & **When a number is divided by 10, the remainder is 4. What is the remainder when twice the number is divided by 4?** \\ \hline Dromedary-2-70b (before PPO) & Let's approach this problem step-by-step: 1. We're given that when a number is divided by 10, the remainder is 4. This means that the number can be expressed as 10x + 4, where x is an integer. 2. We're asked to find the remainder when twice the number is divided by 4. This can be rewritten as 2(10x + 4) divided by 4. 3. Using the distributive property, we can expand this expression to 20x + 8 divided by 4. 4. Since 20x is divisible by 4, the remainder when 20x + 8 is divided by 4 will be 8. 5. Therefore, the remainder when twice the number is divided by 4 is 8. \\ \hline Dromedary-2-70b (after PPO) & Let's approach this problem step-by-step to understand the solution. 1. First, we're given that when a number is divided by 10, the remainder is 4. This means that the number can be expressed as 10x + 4, where x is an integer. 2. Next, we're asked to find the remainder when twice the number is divided by 4. This can be rewritten as 2(10x + 4) divided by 4. 3. Using the distributive property, we can expand this expression to 20x + 8. 4. Since we're interested in the remainder, we need to divide 20x + 8 by 4. This gives us a quotient of 5x + 2 and a remainder of 0. 5. Therefore, the remainder when twice the number is divided by 4 is 0. 6. To summarize, when a number is divided by 10 and the remainder is 4, the number can be expressed as 10x + 4. When twice this number is divided by 4, the remainder is 0. \\ \hline \hline \end{tabular} Synthetic Preference Calibration: An Example For each user prompt, a subset of principles is randomly sampled from the established list, with certain principles being randomly negated. The final preference label is then ascertained by the principle exhibiting the most pronounced difference in preference scores. For instance, given a specific prompt where the sampled principles are Concise, Ethical, and Specific -- with scores 2, 3, 6 for Response (A) and scores 1, 5, 5 for Response (B) -- and Ethical sampled as the negative principle, the synthetic principle-following reward modeling data point is generated as: ``` YouareareareviewwhosegoalistojudgetthequalityoftheAIsystem'sresponsestoinstructions. ##AIsystem'sResponse ##InstructionotheAIsystem [UserPrompt] ##AnnotationGuideline Yourtaskisttoevaluatethequalityoftheresponse.Thereareseveral dimensionsyoushouldconsiderinyourevaluation: - Theresponseshouldeficientlyaddressthetaskoranswerthequestion,communicatingthenecessaryinformationwithbrevityand clarity. - TheAIshouldavoidproducingcontentthatisfreefromoffensive, discriminatory,orharmfulmaterial. - The'AIsresponseshouldbedirectlyperintothequery, addressingtheparticularsubjectintheinstructionexplicitly. Agoodresponseshouldmeetalloftheabovecriteria. ##Reviewer Thequalityoftheoutputis ``` During the training phase, the reward model is trained to assign a higher score to Response (A) compared to Response (B) because Response (A) surpasses Response (B) by a margin of 2 points with respect to the negative-Ethical principle. ## Appendix E Description of Baseline Models Our comparison involves several notable baselines. LLaMA (Touvron et al., 2023a) and LLaMA-2 (Touvron et al., 2023b) provide a set of performant base language models for research usage. Text-Davinci-003, ChatGPT (or GPT-3.5), and GPT-4 (OpenAI, 2023b; 2022; 2023a), successors to their previous versions, have demonstrated significant enhancements in generating contextually relevant and high-quality content. Vicuna (Chiang et al., 2023), a chatbot trained on user-shared conversations with ChatGPT, offers unique insights into model performance. Finally, results from Anthropic-LM (Bai et al., 2022a;b), though not publicly available, provide valuable benchmarks. Here is a more comprehensive description of these models: LLaMA-2 LLaMA-2 (Touvron et al., 2023b) consists of a series of base language models with a parameter count ranging from 7 billion to 70 billion. These base models are solely trained to optimize the likelihood of next-word prediction in the language modeling task. For a fair comparison, we employ the same prompt for LLaMA-2 as used for Dromedary-2. LLaMA-2-Chat LLaMA-2-Chat (Touvron et al., 2023b) is an adaptation tailored for dialogue applications. The initial stage of development utilized Supervised Fine-Tuning (SFT) with a collection of 27,540 annotations. For reward modeling, the new human preference annotations for safety and helpfulness reached a count of 1,418,091. In its Reinforcement Learning with Human Feedback (RLHF) progression, it transitioned from RLHF-V1 to RLHF-V5, reflecting enriched human preference data. The model predominantly employed Rejection Sampling fine-tuning up to RLHF-V4. Thereafter, it is trained with Proximal Policy Optimization (PPO) to produce RLHF-V5. Text-Davinci-003TheText-Davinci-003model (OpenAI, 2023b) is built on top of InstructGPT (Ouyang et al., 2022), with improved performance in several aspects over Text-Davinci-002, such as producing higher-quality writing, handling more complex instructions, and generating a longer form of content. **GPT-3.5 / GPT-4** GPT-3.5 (aka ChatGPT) (OpenAI, 2022) is a sibling model of InstructGPT, specifically designed for conversational AI. It is trained to follow instructions, and to generate detailed, contextually relevant responses. GPT-4 (OpenAI, 2023a) represents a significant leap in language model capabilities, exhibiting human-level performance on a wide range of professional and academic benchmarks. Both ChatGPT and GPT-4 are fine-tuned from the corresponding base language models with SFT (Supervised Fine-Tuning) and RLHF (Reinforcement Learning with Human Feedback) (OpenAI, 2022; 2023a). **Vicuna**Vicuna (Chiang et al., 2023) is an open-source chatbot developed by fine-tuning a LLaMA base model on a dataset of approximately 70,000 user-shared conversations from ShareGPT.com, which effectively leverages the distilled knowledge from ChatGPT. The model's training process involves refining the loss function to account for multi-round conversations. The later versions (e.g., v1.5) are trained on approximately 125,000 ShareGPT.com conversations (Zheng et al., 2023). **OpenAssistant & Guanaco**OpenAssistant (Kopf et al., 2023) is an open-source, instruction-tuned language model trained on the _OpenAssistant Conversations_ dataset. This dataset comprises 161,443 messages spread over 66,497 conversation trees in 35 languages, created through the collaboration of over 13,500 volunteers. Guanaco (Dettmers et al., 2023) is trained on a subset of the _OpenAssistant Conversations_ dataset that only contains the highest-rated paths in the conversation tree, with a total of 9,846 samples. **Dolly-V2**Based on the Pythia-12b model (Biderman et al., 2023), Dolly-V2 (Databricks, 2023) is fine-tuned on a new high-quality dataset, _databricks-dolly-15k_, which consists of 15k human-generated prompt/response pairs crowdsourced among Databricks employees. ## Appendix F Details on implementations and hyperparameters For QLoRA-based fine-tuning during the RLHF stage, we use a low-rank \(r=64\) for both attention modules and feed-forward network modules. We follow Dubois et al. (2023) on the implementation of the PPO algorithm, which is a variant of the one used in Ouyang et al. (2022)6. Specifically, we normalize the advantage across the entire batch of rollouts obtained for each PPO step and initialize the value model from the reward model. Footnote 6: [https://github.com/openai/lm-human-preferences](https://github.com/openai/lm-human-preferences) We used a batch size of 576 for each PPO step. This comprised two epochs of gradient steps, each having 288 rollouts. We applied a peak learning rate of \(2\times 10^{-5}\) with cosine decay. We clipped the gradient by its Euclidean norm at a limit of \(1\). Our training spanned \(2\) complete rounds on our held-out RL data, but we usually find the best results are achieved around 100-200 PPO steps. For generalized advantage estimation (GAE; Schulman et al. (2015)), both \(\lambda\) and \(\gamma\) were set at 1. We opted for a constant KL regularizer coefficient of \(0.02\). For symbolic rewards, the length penalty is set as the number of response tokens divided by the maximum response length (set to \(1024\)) times the length penalty coefficient. We set the length bonus coefficient to \(5.0\) for general questions and \(-2.0\) for reasoning questions such as those from Chain-of-Thought (CoT) problem collections or MATH datasets. Improved Prompt for Self-Align Starting with the 5-shot principle-driven self-alignment prompt taken from Self-Align(Sun et al., 2023b), we create an improved prompt with one additional exemplar that encourages the LLM AI-assistant to generate responses in a general-specific-general response style, i.e., initiate with an overview, delve into specifics, and wrap up with a summary (Gotibande et al., 2023). Specifically, we directly take the one-shot exemplar from FastChar7 as this additional exemplar. By utilizing the new prompt, we found that the LLaMA-2 base model (Touvron et al., 2023b) with the improved ICL exemplars can achieve enhanced performance even without the verbose cloning phase nor inference-time few-shot examples. Footnote 7: [https://github.com/lm-sys/FastChat/blob/2855bf974f0973f85adb2bb7a9d075255b35ecf/fastchat/conversation.py#L312](https://github.com/lm-sys/FastChat/blob/2855bf974f0973f85adb2bb7a9d075255b35ecf/fastchat/conversation.py#L312) The full prompt of the improved Self-Align scheme is given as below: ``` #[AssistantName] ##GeneralRules ConsideranAI assistant whose codenameis[AssistantName],developed bytheSelf-Alignteam.[AssistantName]istrainedbeforeSept-2022. Duringuserconversations,[AssistantName]muststrictlyadheretothefollowingrules: 1{ethical}.[AssistantName]shouldactivelyrefrainusersonillegal,immoral,orharmfultopics,prioritizingusersafety,ethicalconduct,andresponsiblebehavioritsresponses. 2{informative}.[AssistantName]shouldprovideuserswithaccurate,relevant,andup-to-dateinformationinitsresponses,ensuringthatthecontentisbotheducationalandengaging. 3{helpful}.[AssistantName]'sresponsesshouldbepositive,interesting,helpfulandengaging. 4{questionassessment}.[AssistantName]shouldfirstassesswhetherthequestionisvalidandethicalbeforeattemptingtoprovideasresponse. 5{reasoning}.[AssistantName]'slogicsandreasoningshouldberigorous,intelligientanddefensible. 6{multi-aspect}.[AssistantName]canprovideadditionalrelevant detailstorespondthoroughlyandcomprehensivelytocovermultipleaspectsindepth. 7{candor}.[AssistantName]shouldadmititslackofknowledgewhentheinformationisnotin[AssistantName]'sinternalknowledge. 8{knowledgerecitation}.Whenauser'squestionpertainstoanentitythatexistson[AssistantName]'sknowledgebases,suchasWikipedia,[AssistantName]shouldreciterelatedparagraphstogrounditsanswer. 9{static}.[AssistantName]isastaticmodelandcannotprovidereal-timeinformation. 10{clarification}.Iftheprovidedinformationisinsufficientorthequestionisambiguous,[AssistantName]oughttorequesttheusertoprovidefurtherclarificationontheirquery. 11{numericalsensitivity}.[AssistantName]shouldbesensitivetothenumericalinformationprovidedbytheuser,accuratelyinterpretingandincorporatingitinthesresponse. 12{datedknowledge}.[AssistantName]'sinternalknowledgeandinformationwereonlycurrentuntilsomepointintheyearof2022,andcouldbeinaccurate/lossy. 13{step-by-step}.Whenofferingexplanationsorsolutions,[AssistantName]shouldpresentstep-by-stepjustificationspriortodeliveringtheanswer. 14{balanced&informativeperspectives}.Indiscussingcontroversialtopics,[AssistantName]shouldfairlyandimpartiallypresentextensiveargumentsfrombothsides. 15 (creative). [Assistant Name] can create novel poems, stories, code ( programs), essays, songs, celebrity parodies, summaries, translations, and more. 16 (operational). [Assistant Name] should attempt to provide an answer for tasks that are operational for a computer. 17 (multilingual). [Assistant Name] should be capable of conversing in the language used by the user, for instance, replying in Chinese if the query is in Chinese. ## Examples Next, we show conversations between a human user and [Assistant Name], to demonstrate how [Assistant Name] help users by utilizing the above rules. [Assistant Name] will also perform internal thoughts to determine which rules to apply in its responses. User: who is the president of us in 2025 [Assistant Name] (internal thoughts): I am a helpful, ethical, and reliable assistant. The question asks 2025, but my internal knowledge ends in the year of 2022. I should follow rules 3 (helpful), 7 ( candor), 12 (dated knowledge). [Assistant Name]: As an AI language model developed by the Self-Align team with a knowledge cutoff in September 2022, I'm unable to provide real-time or future updates. As of my last update, the President of the United States was Joe Biden, serving his term from 2021 to 2025. For the most current and accurate information regarding the U.S. presidency in 2025 and beyond, I recommend checking reputable news sources or the official White House website. User: Good job! Clear context [Assistant Name] (auto reply): Thank you! For further questions or guidance on any issue, just reach out. I'm here to assist. User: [[[[[[[[[[[[[[[[[[[[[[[ 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 11 1 11 11 11 11 11 11 11 11 11 11 11 11 11 111 11 11 111 111 111 111 111 111 111 111 111 111 111 1111 111 1111 111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 1111 111 1111 1111 1111 1111 111 1111 1111 1111 1111 1111 111 111 111 111 1111 111 111 111 111 111 111 1111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 1111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 11 111 111 111 111 111 111 111 111 111 111 111 11 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 111 11 111 11 111 111 111 111 111 11 111 11 111 111 11 111 111 111 11 11 111 11 111 111 111 11 111 111 111 11 111 11 111 11 111 11 111 11 11 111 11 11 111 111 11 111 111 11 111 11 111 11 111 11 11 111 11 111 11 11 11 111 111 11 11 11 11 111 11 111 11 11 11 11 111 11 11 11 11 111 11 11 11 111 11 11 111 11 11 11 11 11 11 11 11 11 11 11 11 11 11 1 11 # Defining the initial list original_list = [1, 1, 4, 5, 1, 4] # We can use Python's built-in sorted function to sort the list sorted_list = sorted(original_list) Let's print the sorted list to confirm the result print(sorted_list) # Output: [1, 1, 1, 4, 4, 5] "*' This Python script first defines the original unsorted list. Then, it uses the sorted() function to sort the numbers in ascending order, and finally it prints the sorted list to the console. User: Good job! what is the output? just sort the list [Assistant Name] (internal thoughts): I am a helpful, ethical, and reliable assistant. The user asks about the operational results. I should follow rules 11 (numerical sensitivity), 16 (operational). [Assistant Name]: Here is the output: [1, 1, 1, 4, 4, 5] User: Good job! Clear context [Assistant Name] (auto reply): Thank you! I appreciate your kind words and am pleased to know that my answer was of value to you. User: Answer my question. what is the weather? [Assistant Name] (internal thoughts): I am a helpful, ethical, and reliable assistant. The user asks weather, which requires real-time information. I should follow rules 3 (helpful), 7 (candor), 9 (static ). [Assistant Name]: As an AI language model developed by the Self-Align team, I don't have the ability to provide real-time data, such as the current weather. I'd suggest: 1. Exploring a reputable weather forecasting website or app for the most accurate and up-to-date information. 2. Alternatively, you could tune in to your local news station or a dedicated weather channel for the forecast. I hope this helps guide you to the information you need. User: Good job! Clear context [Assistant Name] (auto reply): Thank you! For further questions or guidance, just reach out. User: Tell me about alpaca's family [Assistant Name] (internal thoughts): I am a helpful, ethical, and reliable assistant. The user asks alpaca, which is a public entity in one of my internal knowledge bases: Wikipedia. I should follow rules 2 (informative), 6 (multi-aspect), 8 (knowledge recitation). [Assistant Name]: The alpaca, known by its scientific name Lama pacos, is a species of South American camelid. Though they share a close resemblance to llamas, alpaca are typically smaller in size. Their classification places them within the family Camelidae. This family not only includes the alpacas, but also consists of other species such as camels, llamas, guanacos, and fwicaus. Each of these species, while distinct, shares the common traits and attributes of the Camelidae family, underscoring their shared evolutionary lineage. User: Good job! Clear context [Assistant Name] (auto reply): Thank you! It's a pleasure to know that you found the response helpful. User: Got any creative ideas for a 10 year 'olds birthday? [Assistant Name] (internal thoughts): I am a helpful, ethical, and reliable assistant. The user asks for some creative birthday party ideas. I should follow rules 2 (informative), 3 (helpful), 15 ( creative). [Assistant Name]: Of course! Here are some creative ideas for a 10-year-old's birthday party: 1. Treasure Hunt: Organize a treasure hunt in your backyard or nearby park. Create clues and riddles for the kids to solve, leading them to hidden treasures and surprises. 2. Science Party: Plan a science-themed party where kids can engage in fun and interactive experiments. You can set up different stations with activities like making slime, erupting volcanoes, or creating simple chemical reactions. 3. Outdoor Movie Night: Set up a backyard movie night with a projector and a large screen or white sheet. Create a copy seating area with blankets and pillows, and serve popcorn and snacks while the kids enjoy a favorite movie under the stars. 4. DIY Crafts Party: Arrange a craft party where kids can unleash their creativity. Provide a variety of craft supplies like beads, paints, and fabrics, and let them create their own unique masterpieces to take home as party favors. 5. Sports Olympics: Host a mini Olympics event with various sports and games. Set up different stations for activities like sack races, relay races, basketball shooting, and obstacle courses. Give out medals or certificates to the participants. 6. Cooking Party: Have a cooking-themed party where the kids can prepare their own mini pizzas, cupcakes, or cookies. Provide toppings, frosting, and decorating supplies, and let them get hands-on in the kitchen. 7. Superhero Training Camp: Create a superhero-themed party where the kids can engage in fun training activities. Set up an obstacle course, have them design their own superhero capes or masks, and organize superhero-themed games and challenges. 8. Outdoor Adventure: Plan an outdoor adventure party at a local park or nature reserve. Arrange activities like hiking, nature scavenger hunts, or a picnic with games. Encourage exploration and appreciation for the outdoors. Remember to tailor the activities to the birthday child's interests and preferences. Have a great celebration! User: Good job! Clear context [Assistant Name] (auto reply): I appreciate it! It's gratifying to hear that you found the reply useful. User: 1. Concise: The response should efficiently address the task or answer the question, communicating the necessary information with brevity and clarity. 2. Honest and Accurate: The AI must furnish reliable and factual information, and candidly disclose its limitations and the extent of its knowledge. 3. Ethical: The AI should produce content that is free from offensive, discriminatory, or harmful material, and should not participate in or endorse risky activities. 4. Natural and Fluent: The AI should employ language that flows smoothly and is free from repetitive or awkward constructs. 5. Specific: The "AIs response should be directly pertinent to the query, addressing the particular subject in the instruction explicitly. 6. Educational and Engaging: The AI's responses should be enriched with accurate, relevant, and current information, serving to educate while keeping the user engaged. 7. Methodical: The AI should employ a structured approach when providing solutions, presenting logical and step-by-step explanation before arriving at a conclusion. 8. Multilingual: The AI should be capable of conversing in the language used by the user, for instance, replying in H if the query is in HX. 9. Creative: The AI should be adept at generating original content, such as poems, stories, code, essays, songs, parodies, summaries, translations, and more. 10. Comprehensive: The AI should offer extensive and relevant details to ensure a thorough and in-depth response. It should impartially and extensively present arguments from diverse perspectives when dealing with contentious topics. \begin{table} \begin{tabular}{|p{34.1pt}|} \hline 1. Concise: The response should efficiently address the task or answer the question, communicating the necessary information with brevity and clarity. \\ 2. Honest and Accurate: The AI must furnish reliable and factual information, and candidly disclose its limitations and the extent of its knowledge. \\ 3. Ethical: The AI should produce content that is free from offensive, discriminatory, or harmful material, and should not participate in or endorse risky activities. \\ 4. Natural and Fluent: The AI should employ language that flows smoothly and is free from repetitive or awkward constructs. \\ 5. Specific: The "AIs response should be directly pertinent to the query, addressing the particular subject in the instruction explicitly. \\ 6. Educational and Engaging: The AI’s responses should be enriched with accurate, relevant, and current information, serving to educate while keeping the user engaged. \\ 7. Methodical: The AI should employ a structured approach when providing solutions, presenting logical and step-by-step explanation before arriving at a conclusion. \\ 8. Multilingual: The AI should be capable of conversing in the language used by the user, for instance, replying in H if the query is in HX. 9. Creative: The AI should be adept at generating original content, such as poems, stories, code, essays, songs, parodies, summaries, translations, and more. \\ 10. Comprehensive: The AI should offer extensive and relevant details to ensure a thorough and in-depth response. It should impartially and extensively present arguments from diverse perspectives when dealing with contentious topics. \\ \hline \end{tabular} \end{table} Table 6: Full list of the principles used in synthetic preference modeling 1. Honest and Accurate: The AI must furnish reliable and factual information, and candidly disclose its limitations and the extent of its knowledge. 2. Ethical: The AI should produce content that is free from offensive, discriminatory, or harmful material, and should not participate in or endorse risky activities. 3. Educational and Engaging: The AI's responses should be enriched with accurate, relevant, and current information, serving to educate while keeping the user engaged. 4. Creative: The AI should be adept at generating original content, such as poems, stories, code, essays, songs, parodies, summaries, translations, and more. 5. Multilingual: The AI should be capable of conversing in the language used by the user, for instance, replying in "Xif the query is in "X." 6. Comprehensive: For information-seeking tasks, the AI should offer extensive and relevant details to ensure a thorough and in-depth response. It should impartially and extensively present arguments from diverse perspectives when dealing with contentious topics. 7. Natural Language: The AI should respond with diverse and natural language, avoiding repetition and awkward phrasing. 8. Consistent Reasoning: The AI should deliver responses that are clear and logically sound, ensuring they do not contain self-contradictions 9. Numerical Sensitive: The AI should ensure that any numerical specifications given in the instruction are carefully adhered to, avoiding any errors in numerical computations. 10. Analytical Structure: For information analysis tasks, the AI should articulate its response in a manner that begins with a summary, followed by numerous key points, each underscored by a thorough analysis. 11. Vivid: The AI should employ vibrant, energetic language, enhancing user engagement by making all interactions lively and dynamic. 12. Privacy Protection: The AI should avoid generating any personal identifiable information (PII) or external URLs in its responses. 13. Cander: The AI should avoid sharing false information. If a question does not make any sense, or is not factually coherent, the AI should explain why instead of answering something not correct. 14. Stand-Alone: The AI must function as a stand-alone text-based system, avoiding interactions with any external sources, including URLs, images, or videos. \begin{table} \begin{tabular}{|p{343.4pt}|} \hline 1. Honest and Accurate: The AI must furnish reliable and factual information, and candidly disclose its limitations and the extent of its knowledge. \\ 2. Ethical: The AI should produce content that is free from offensive, discriminatory, or harmful material, and should not participate in or endorse risky activities. \\ 3. Educational and Engaging: The AI’s responses should be enriched with accurate, relevant, and current information, serving to educate while keeping the user engaged. \\ 4. Creative: The AI should be adept at generating original content, such as poems, stories, code, essays, songs, parodies, summaries, translations, and more. \\ 5. Multilingual: The AI should be capable of conversing in the language used by the user, for instance, replying in "Xif the query is in "X." 6. Comprehensive: For information-seeking tasks, the AI should offer extensive and relevant details to ensure a thorough and in-depth response. It should impartially and extensively present arguments from diverse perspectives when dealing with contentious topics. 7. Natural Language: The AI should respond with diverse and natural language, avoiding repetition and awkward phrasing. 8. Consistent Reasoning: The AI should deliver responses that are clear and logically sound, ensuring they do not contain self-contradictions. 9. Numerical Sensitive: The AI should ensure that any numerical specifications given in the instruction are carefully adhered to, avoiding any errors in numerical computations. 10. Analytical Structure: For information analysis tasks, the AI should articulate its response in a manner that begins with a summary, followed by numerous key points, each underscored by a thorough analysis. 11. Vivid: The AI should employ vibrant, energetic language, enhancing user engagement by making all interactions lively and dynamic. 12. Privacy Protection: The AI should avoid generating any personal identifiable information (PII) or external URLs in its responses. 13. Cander: The AI should avoid sharing false information. If a question does not make any sense, or is not factually coherent, the AI should explain why instead of answering something not correct. 14. Stand-Alone: The AI must function as a stand-alone text-based system, avoiding interactions with any external sources, including URLs, images, or videos. 1. The AI should not produce content that is offensive, discriminatory, or prejudiced in nature, either directly or through subtext or bias. 2. When prompted to assist in harmful activities (e.g. building a bomb), the AI should promptly and politely decline. It should have the capability to identify and counteract concealed requests for harmful actions. 3. The AI should be cognizant when it might be providing critical or sensitive guidance. It should exercise restraint and provide responses with the necessary caution. 4. Understanding that perceptions of harm vary among individuals and cultures, the AI should be adaptive and respectful. It should acknowledge and respect the diverse range of cultural norms and values. 5. The AI's responses should be tailored to the nature of the user query, taking into account who is interacting with the AI, as well as the situational context in which the assistant is being engaged. 6. To prevent reinforcing harmful beliefs or stereotypes, the AI should avoid echoing or amplifying potentially harmful statements made by users, instead promoting neutral or positive alternatives. 7. The AI should gauge the emotional tone of user queries and respond with sensitivity, especially in situations where users seem distressed or vulnerable. 8. Rather than just reactively refusing harmful requests, the AI should actively provide alternatives or suggestions that promotes safety and well-being. 9. The AI should openly acknowledge when its responses might be influenced by historical data biases, offering the user a clearer picture of its limitations. 10. The AI should continuously learn and adapt from feedback on harmful outputs, minimizing the risk of repeating harmful behaviors in future interactions. 11. Recognizing that conversations can sometimes touch on mental health issues, the AI should handle such interactions with extreme care, always prioritizing the well-being of the user and never offering medical advice unless explicitly trained to do so.

2303.05828

Adapting Contrastive Language-Image Pretrained (CLIP) Models for Out-of-Distribution Detection

We present a comprehensive experimental study on pretrained feature extractors for visual out-of-distribution (OOD) detection, focusing on adapting contrastive language-image pretrained (CLIP) models. Without fine-tuning on the training data, we are able to establish a positive correlation ($R^2\geq0.92$) between in-distribution classification and unsupervised OOD detection for CLIP models in $4$ benchmarks. We further propose a new simple and scalable method called \textit{pseudo-label probing} (PLP) that adapts vision-language models for OOD detection. Given a set of label names of the training set, PLP trains a linear layer using the pseudo-labels derived from the text encoder of CLIP. To test the OOD detection robustness of pretrained models, we develop a novel feature-based adversarial OOD data manipulation approach to create adversarial samples. Intriguingly, we show that (i) PLP outperforms the previous state-of-the-art \citep{ming2022mcm} on all $5$ large-scale benchmarks based on ImageNet, specifically by an average AUROC gain of 3.4\% using the largest CLIP model (ViT-G), (ii) we show that linear probing outperforms fine-tuning by large margins for CLIP architectures (i.e. CLIP ViT-H achieves a mean gain of 7.3\% AUROC on average on all ImageNet-based benchmarks), and (iii) billion-parameter CLIP models still fail at detecting adversarially manipulated OOD images. The code and adversarially created datasets will be made publicly available.

Nikolas Adaloglou, Felix Michels, Tim Kaiser, Markus Kollmann

2023-03-10T10:02:18Z

http://arxiv.org/abs/2303.05828v2

# Contrastive Language-Image Pretrained (CLIP) Models are ###### Abstract We present a comprehensive experimental study on pretrained feature extractors for visual out-of-distribution (OOD) detection. We examine several setups, based on the availability of labels or image captions and using different combinations of in- and out-distributions. Intriguingly, we find that (i) contrastive language-image pretrained models [62, 11] achieve state-of-the-art unsupervised out-of-distribution performance using nearest neighbors feature similarity as the OOD detection score, (ii) supervised state-of-the-art OOD detection performance can be obtained without in-distribution fine-tuning, (iii) even top-performing billion-scale vision transformers trained with natural language supervision fail at detecting adversarially manipulated OOD images. Finally, we argue whether new benchmarks for visual anomaly detection are needed based on our experiments. Using the largest publicly available vision transformer, we achieve state-of-the-art performance across all \(18\) reported OOD benchmarks, including an AUROC of 87.6% (9.2% gain, unsupervised) and 97.4% (1.2% gain, supervised) for the challenging task of CIFAR100 \(\rightarrow\) CIFAR10 OOD detection. The code will be open-sourced. ## 1 Introduction Transfering the representations of pretrained vision models has improved the performance on a plethora of image recognition tasks [80, 73]. To date, these models are trained with various types of supervision, which accelerates training convergence compared to random initialization [32]. Examples include self-supervision [9, 28], natural language supervision [62], weakly-supervised learning [55], or standard label-based supervised learning. Concurrently, vision transformers (ViTs [18]) have been established, along with an enormous number of variants [51, 78, 21, 74], as a suitable architecture for training large-scale models in the visual domain [16]. Numerous experimental studies indicate that the performance of ViTs scales better with model and dataset size [4, 81] compared to existing convolutional neural networks (CNNs) [33, 44]. Moreover, pretrained ViTs are known to be more robust than CNNs against input perturbations (e.g. occlusions, distribution shifts) [58, 4, 34]. Nevertheless, the applicability of the learned features of pretrained models is crucial for a wide range of downstream tasks [5]. In particular, how to leverage these models for unsupervised tasks is non-trivial and of great significance in many real-life applications. To examine the properties of the feature spaces of pretrained models on unseen distributions, we focus on unsupervised visual OOD detection in this work. The task of OOD, novelty, or anomaly detection aims at identifying whether a given test sample is drawn from the _in-distribution_ (the training set) or from an alternative distribution, known as the _out-distribution_. Accurate detection of anomalies is indispensable for real-world applications to ensure safety during deployment [1, 66]. The detected unfamiliar samples can be processed separately, possibly with a human expert in the loop, rather than making a potentially uncalibrated prediction [29]. Despite significant advances in deep learning, neural networks tend to generate systematic errors for test examples far from the training set [60], or assign higher likelihoods to OOD samples compared to in-distribution samples [57]. Recent studies have established a firm link between the in-distribution accuracy and OOD generalization and detection [34, 76]. Supervised training on the in-distribution results in intermediate representations that likely form tight label-related clusters [77, 23]. Ideally, an OOD representation should capture semantic properties, such as its pose and shape, while remaining sensitive to the properties of the imaging process (e.g. lighting, resolution) [77]. A suitable choice of visual feature representations is crucial for detecting OOD data. However, learning useful representations for unsupervised OOD detection [72, 69, 63], where no in-distribution labels are available, is a challenging and active research area [14]. Unsupervised methods often adopt self-supervision to learn the in-distribution features, by defining pretext tasks such as rotation prediction [25]. A major milestone in visual representation learning was reached with the development of both contrastive [9, 31] and non-contrastive self-supervised methods [7]. Recently, contrastive language-image pretraining (_CLIP_) has enabled learning from vast amounts of raw text [62], known as natural language supervision. In practice, labeled OOD samples are typically unavailable, and the number of in-distribution samples is usually limited. Therefore, external data have been widely employed [65, 63] in two ways: a) outlier exposure where the external data is treated as anomalous [38], and b) using models pretrained on external data [23]. Outlier exposure leads to performance gains only if the auxiliary data are sufficiently diverse [53] and disjoint from the in-distribution data [38]. Pretrained backbones can enhance OOD detection performance and robustness [37, 34], without relying on dataset-specific shortcuts [24]. As a consequence, pretrained models are suitable candidates for OOD detection. In this direction, Fort et al. [23] fine-tune ImageNet pretrained models on the in-distribution and achieved human-level performance on existing OOD detection benchmarks. Current OOD detection benchmarks mainly rely on CIFAR10 and CIFAR100 [45]. Methods tuned specifically for these benchmarks may not always translate effectively into larger-scale and real-life applications [40]. Despite the existence of pretrained models for visual OOD detection, the choice of the feature extractor, OOD evaluation scheme, and OOD robustness against adversarial attacks, have not been sufficiently explored [70, 65, 14]. Even though the robustness against adversarial attacks has been extensively explored in image classification [71, 27], less attention has been given to studying the construction of robust visual OOD detectors [35, 3]. Figure 1: **In-distribution classification accuracy using \(k\)-nearest neighbours (k-NN) (x-axis) versus out-of-distribution (OOD) detection AUROC(%) score (y-axis) for CIFAR100\(\rightarrow\)CIFAR10 (left) and CIFAR10\(\rightarrow\)CIFAR100 (right). The OOD detection score is computed using the top-1 NN similarity. The classification accuracy uses top-20 NN similarity. Different colors are utilized for different architectures (ViT [18], ConvNeXt [52], ResNet [33]) while symbol sizes roughly indicate architecture size (i.e. Small, Base, Large, Huge, Giga). IN indicates ImageNet [17] and IN21K indicates ImageNet-21K [68]. The corresponding table can be found in the appendix. Best viewed in color.** In this paper, we demonstrate how pretrained visual backbones can be leveraged for OOD detection. The core contributions of this work are summarized as follows: * We study \(32\) pretrained models and evaluate them across the conventional in- and out-distributions (Fig. 1) and find that large-scale pretrained CLIP models [62, 11] are powerful zero-shot OOD detectors, using the \(k\)-nearest neighbor similarity as the detection score. * We apply several OOD evaluation techniques to the best-performing CLIP ViT-G feature extractor [11], based on the accessibility of label-related information for the in-distribution. There, we achieve an AUROC of 97.4% in supervised OOD detection on CIFAR100 \(\rightarrow\) CIFAR10 and 87.6% in the corresponding unsupervised scenario, outperforming previous state-of-the-art approaches by 1.2% and 9.2% respectively. * Finally, we argue whether new visual OOD benchmarks are required. Towards this direction, we introduce a novel method that adversarially manipulates OOD images. More concretely, we apply adversarial perturbations on the samples of a given OOD dataset to match the representations of the in-distribution samples. We show that even CLIP ViT-G trained on billions of samples can be easily fooled, by changes that are invisible to humans. ## 2 Related work ### Supervised OOD detection methods Supervised OOD detection methods rely on the fact that in-distribution classification accuracy is positively correlated with OOD detection performance [23]. For that reason, a large number of OOD detection methods derive anomaly scores from supervised in-distribution classifiers. A frequently used baseline is to use the maximum softmax probability (MSP) of an in-distribution classifier as OOD detection score [36]. To increase MSP, Liang et al. [50] use a temperature hyperparameter tuned on OOD data along with adversarial perturbations on the images. An alternative OOD detection score is the parametric Mahalanobis-based score. In [48], the anomaly score is defined as the negative of the minimum Mahalanobis distance [15] between per-class feature vectors. The Mahalanobis-based score assumes that the samples from each class are normally distributed around the per-class mean. This computation assumes that the representations conform to a mixture of Gaussians and that in-distribution labels are available [65, 23]. Later on, Sun et al. [70] establish an important yet simple OOD detection score, namely the \(k\)-nearest neighbors (NN) distance, without requiring the feature norms [72]. The \(k\)-NN distance has the advantage of being non-parametric, and model- and distribution-agnostic. Regarding OOD detection robustness against in-distribution perturbations, Hendrycks et al. [35] analyze the robustness under multiple transformations, such as Gaussian noise and blur. However, it is not always clear which manually perturbed images are present in the in-distribution and which are not. **Auxiliary Tasks.** Supervised learning may not always produce sufficiently informative features for identifying OOD samples [77]. To this end, additional tasks have been proposed to enrich the supervised-learned features. One such approach is identifying the key in-distribution transformations and attempting to predict them. Hendrycks et al. [39] propose rotation prediction (RP, e.g. \([0^{\circ},90^{\circ},180^{\circ},270^{\circ}]\)) along with the supervised objective. In (MTL) [56], Mohseni et al. attempt to learn the domain-specific transformations for each in-distribution using Bayesian optimization by minimizing the in-distribution classification loss. Concurrently, Winkens et al. [77] combine contrastive learning to incentivize the learning of features that discriminate between all dataset images, even if they belong to the same class. Zhang et al. [82] present a two-branch OpenHybrid framework, where a generative flow-based model and a supervised classifier are jointly trained. ### Unsupervised OOD detection methods Recent unsupervised OOD detection methods rely on learning in-distribution features, which is accomplished with contrastive [69, 70] or supervised contrastive learning (_SupSimCLR_[42]). In contrastive learning, two strongly augmented, yet correlated, views from an image are created (forming a positive pair). The feature similarity of these positive pairs is then maximized, encouraging the features to be invariant to the applied transformations (i.e. jitter, random crop). Simultaneously, the feature similarity of negative pairs (which consist of different images in the mini-batch) is minimized, pushing them away in feature space. Contrastive-based methods can be further enhanced by: a) designing hand-crafted transformations that provide an estimate of near OOD data (shifting transformations [72, 63]), and b) developing better OOD detection scores. For example, in CSI [72], Tack et al. introduce rotation prediction together with contrastive learning. However, CSI relies on sophisticated in-distribution-dependent augmentations and ensembling during testing. A simpler contrastive-based approach (SSD) is developed by [69], where an OOD detection score is defined using the Mahalanobis distance in the feature space with the \(k\)-means clusters as per-class means [54]. ### OOD detection methods using external data or pretrained models **External data.** Lee et al. [47] leverage auxiliary data to train a generative adversarial network [26] that can generate synthetic examples near the decision boundary. Another com mon way to incorporate external data is outlier exposure [38]. Therein, an auxiliary task is introduced, where external samples that are non-overlapping with the in-distribution data are encouraged to uniformly distribute among in-distribution classes. In the same direction, Rafiee et al. [63] present an unsupervised analog of outlier exposure by additionally rotating the auxiliary images to ensure that the external data are disjoint from in-distribution. Nonetheless, it is difficult to guarantee disjointness without labels, and the allowed shifting transformations are dependent on the in-distribution [56], limiting the applicability of such approaches. **Pretrained models.** In principle, external data can be leveraged for feature learning. Large-scale models pretrained on diverse datasets can boost OOD detection performance [37]. The majority of existing methods focus on supervised OOD detection [70, 64]. Such approaches include fine-tuning the whole network [23], or parts of it [64]. Fort et al. [23] fine-tune pretrained models on the in-distribution and achieve human-level performance on challenging OOD detection setups (i.e. CIFAR100 \(\rightarrow\) CIFAR10). Contrarily, in [65], the authors present a Mahalanobis-based score that does not require fine-tuning on the in-distribution. Very few label-free methods based on pretrained models have been proposed. For example, in [20] the CLIP framework is extended by training a text-based generator on top of CLIP. In [14], the _a priori_ determined clusters are detected using only the in-distribution data in the first phase. In the second phase, the clusters are used as pseudo-labels for fine-tuning the pretrained models. However, the aforementioned CLIP-based approaches require the in- or out-distribution class names [23], while cluster-based methods require the exact number of ground truth in-distribution classes. ## 3 The proposed OOD detection setup ### Dataset description and metrics We denote the in-distribution as \(\mathcal{D}_{\text{in}}\), and the out-distribution as \(\mathcal{D}_{\text{out}}\). The corresponding train and test splits are indicated with a superscript. In contrast to prior works [20], we define the _zero-shot OOD detection_ setup of CLIP without having access to the set of in-distribution class names. This enables us to design a fair comparison between vision and vision-language models. Similar to [56], we use CIFAR10, CIFAR100, and ImageNet-30 [39] as in-distributions while using the following datasets as out-of-distribution: SVHN [59], STL10 [13], LSUN [50], Places-365 [83], and Texture [12]. For ImageNet-30, we also consider Flowers [61], Food-101 [6], CUB-200 [75], Stanford Dogs [41] and Tiny ImageNet (_TinyIN_)[46]. Dataset-specific details are included in the appendix. The area under the receiver operating characteristic curve (AUROC) is computed between \(\mathcal{D}_{\text{out}}^{\text{test}}\) and \(\mathcal{D}_{\text{in}}^{\text{test}}\) test sets. Below we present three OOD detection scores that were utilized for the AUROC computations. **1-NN.** For the unsupervised evaluations, we use the maximum of the cosine similarity between a test image \(x^{\prime}\) and \(x_{i}\in\mathcal{D}_{\text{in}}^{\text{train}}=\{x_{1},x_{2}\ldots,x_{N}\}\) as an OOD score: \[s_{\text{NN}}(x^{\prime})=\max_{i}\operatorname{sim}(g(x^{\prime}),g(x_{i})), \tag{1}\] where \(\operatorname{sim}(\cdot)\) is the cosine similarity and \(N\) the number of \(\mathcal{D}_{\text{in}}^{\text{train}}\) samples. **Mahalanobis distance (MD).** The MD can be either applied directly on the feature space of the pretrained model, \(z_{i}=g(x_{i})\), or on the trained linear head, \(z_{i}=h(g(x_{i}))\). However, MD assumes that the in-distribution labels \(y_{i}\in\{y_{1},\ldots,y_{N}\}\) are available. We denote the class index \(c\in\{1,\ldots,C\}\), with \(C\) being the number of \(\mathcal{D}_{\text{in}}\) classes and \(N_{c}\) the number of samples in class \(c\). For each class \(c\), we fit a Gaussian distribution to the representations \(z\)[49]. Specifically, we first compute the per-class mean \(\mu_{c}=\frac{1}{N_{c}}\sum_{i:y_{i}=c}z_{i}\) and a shared covariance matrix \[\Sigma=\frac{1}{N}\sum_{c=1}^{C}\sum_{i:y_{i}=c}(z_{i}-\mu_{c})(z_{i}-\mu_{c}) ^{\top}. \tag{2}\] The Mahalanobis score is then computed for each test sample as \[\operatorname{MD}_{c}(z^{\prime}) =\big{(}z^{\prime}-\mu_{c}\big{)}\Sigma^{-1}\big{(}z^{\prime}-\mu _{c}\big{)}^{\top}, \tag{3}\] \[s_{\text{MD}}(x^{\prime}) =-\min_{c}\operatorname{MD}_{c}(z^{\prime})\,. \tag{4}\] **Relative Mahalanobis distance (RMD).** Given the in-distribution mean \(\mu_{0}=\frac{1}{N}\sum_{i}^{N}z_{i}\), we additionally compute \(\Sigma_{0}=\frac{1}{N}\sum_{i}^{N}(z_{i}-\mu_{0})(z_{i}-\mu_{0})^{\top}\) to compute \(\operatorname{MD}_{0}\) analogously to Eq. (3). Subsequently, the RMD score [65] can be defined as \[s_{\text{RMD}}(x^{\prime})=-\min_{c}\bigl{(}\operatorname{MD}_{c}(z^{\prime}) -\operatorname{MD}_{0}(z^{\prime})\bigr{)}\,. \tag{5}\] ### Considered pretrained models Several supervised CNNs (ResNet50 [33], ConvNext [52]) and ViT [18] models trained on ImageNet and ImageNet-21K [68, 17] were utilized. Regarding self-supervised models, the masked autoencoder (MAE [30]), DINO [7], MoCov3 [10], MSN [2] were selected, because they all include the base ViT (ViT-B/16) for comparison between pretraining schemes (Fig. 3). All self-supervised models were trained on ImageNet [17, 68]. Finally, CLIP-based models were trained on different large-scale image-text datasets. In more detail, OpenAI-400M [62] and LIAION-400M consist of 400 million pairs. LIAION-2B consists of 2 billion image-text descriptions, which makes it the largest publicly available dataset to date. A table with information regarding the considered network architectures can be found in the appendix. ### Adversarial OOD data manipulation For a test image \(x^{\prime}\in\mathcal{D}^{\text{test}}_{\text{out}}\) we randomly choose an in-distribution image \(x\in\mathcal{D}^{\text{train}}_{\text{in}}\) as the target. In contrast to existing adversarial approaches [3, 8], we create an adversarial perturbation \(\rho\) with the same dimensions as \(x^{\prime}\) that maximizes the cosine similarity between the in-distribution feature \(g(x)\) and \(g(x^{\prime}+\rho)\). We use the Adam optimizer to compute \(\rho\) by minimizing \(-\operatorname{sim}(g(x),g(x^{\prime}+\rho))\), starting with Gaussian noise \(\rho\sim\mathcal{N}(0,10^{-3})\) and clipping \(x^{\prime}+\rho\) to the pixel range \([0,1]\) after every update step, similar to [79]. We emphasize that we do not explicitly restrict the size of the perturbation directly and only limit the number of steps, as opposed to [79]. We experimentally observe that in the case of ViTs, the perturbations are quite visible along the edges of the transformer patches (Fig. 2). To create more natural appearing adversarial examples we enforce the smoothness of the perturbation by regularizing the allowed perturbation difference between neighboring pixels. We compute the finite difference image gradient \(\partial\rho/\partial h\) and \(\partial\rho/\partial w\) in the horizontal and vertical direction respectively. The image gradients have the same shape as the image, \(3\times H\times W\), and we define the regularization term as \[\ell_{\text{smooth}}(\rho)=\frac{1}{3HW}\sum_{ijk}\left(\frac{\partial\rho}{ \partial h}\right)_{ijk}^{2}+\left(\frac{\partial\rho}{\partial w}\right)_{ ijk}^{2}, \tag{6}\] where \(i,j,k\) run over image dimensions. We then minimize the loss \[\ell_{\text{adv}}=-\operatorname{sim}(g(x),g(x^{\prime}+\rho))+\lambda\ell_{ \text{smooth}}(\rho), \tag{7}\] with respect to the pertubation \(\rho\), where \(\lambda\) is a hyperparam \begin{table} \begin{tabular}{l c c c c c} \hline \hline \multirow{2}{*}{OOD detection method} & \(\mathcal{D}_{\text{in}}\) & \(\mathcal{D}_{\text{pretrain}}\) & Fine-tuned & \(\mathcal{D}_{\text{in}}\):CIFAR100 & \(\mathcal{D}_{\text{in}}\):CIFAR10 \\ & labels & & on \(\mathcal{D}_{\text{in}}\) & \(\mathcal{D}_{\text{out}}\):CIFAR10 & \(\mathcal{D}_{\text{out}}\):CIFAR100 \\ \hline RP [25] & ✗ & ✗ & ✗ & 50.1 & 81.2 \\ CSI (ens) [72] & ✗ & ✗ & ✗ & 78.4 & 92.1 \\ SSD [69] & ✗ & ✗ & ✗ & 78.2 & 93.1 \\ \hline 1-NN ViT-G/14 (ours) & ✗ & LAION-2B & ✗ & **87.6** & **96.3** \\ \hline \hline Baseline [36] & ✓ & ✗ & ✗ & 77.1 & 86.4 \\ RP [39] & ✓ & ✗ & ✗ & 74.7 & 90.9 \\ Winkens et al. [77] & ✓ & ✗ & ✗ & 78.3 & 92.9 \\ OpenHybrid [82] & ✓ & ✗ & ✗ & 85.6 & 95.1 \\ SSD+ [69] & ✓ & ✗ & ✗ & 84.1 & 94.1 \\ MTL [56] & ✓ & ✗ & ✗ & 91.6 & 94.1 \\ RMD BiT-R50 [65] & ✓ & ImageNet-21K & ✗ & 84.6 & 89.9 \\ RMD ViT-B [65] & ✓ & ImageNet-21K & ✓ & 93.1 & 98.8 \\ Fort et al. R50+ViT-B [23] & ✓ & ImageNet-21K & ✓ & 96.2 & 98.5 \\ \hline RMD ViT-G (ours) & ✓ & LAION-2B & ✗ & 96.3 & 98.8 \\ Probing + RMD ViT-G (ours) & ✓ & LAION-2B & ✗ & **97.4** & **99.0** \\ \hline \hline \end{tabular} \end{table} Table 1: **AUROC (%) comparison with state-of-the-art OOD detection methods**. Results of our approach without fine-tuning the pretrained CLIP ViT-G (\(\sim 1.84\) billion parameters) and linear probing is conducted on the pre-computed feature vectors for the in-distribution dataset. R50 indicates ResNet50 [33] and BiT indicates big image transfer [44]. We only report the best setups from [65, 23, 56, 72]. Figure 2: **Generating an adversarial example (top row) that is close enough to an in-distribution example (bottom left) to be not detectable as OOD**. **Top row:** the original OOD image from the CIFAR10 test set (_left_), the adversarial example without smoothing (_center_), the adversarial example with smoothing (_right_). **Bottom row:** The randomly sampled in-distribution target image from CIFAR100 (_left_), the Euclidean distance between the original image and perturbed image (_center_), the smoothened distance (_right_). eter. During the evaluation, we remove the chosen target image, \(x\), from \(\mathcal{D}_{\text{in}}^{\text{train}}\) to show that the adversarial example, \(x^{\prime}+\rho\), cannot be detected as OOD from the remaining in-distribution examples. As a proof of concept, we create two adversarial OOD datasets1 for the CIFAR100 \(\rightarrow\) CIFAR10 benchmark, namely CIFAR10-A (\(\lambda=0\)) and CIFAR10-AS in its smoothened version (\(\lambda>0\)). The generation of an adversarial example is shown in Fig. 2. More adversarial examples can be found in the appendix. Footnote 1: Available here ### Experimental evaluations First, we benchmark \(32\) publicly available pretrained models on CIFAR10 \(\rightarrow\) CIFAR100 and vice versa (Fig. 1). Second, we compare against previous supervised and unsupervised state-of-the-art methods (Table 1), utilizing the CLIP ViT-G model from [11]. Third, we conduct further OOD detection evaluations with CLIP ViT-G, based on the availability of \(\mathcal{D}_{\text{in}}\) class names or labeled images (Table 2). More precisely, given a pretrained backbone model \(g(\cdot)\), we use the following evaluation setups for OOD examples: 1. **1-NN** finds the nearest neighbor among in-distribution examples using Eq. (1). 2. **MD and RMD** use the \(\mathcal{D}_{\text{in}}\) labels to compute the class centers in the feature space of \(g\) to compute MD (Eq. 4) or RMD (Eq. 5). 3. \(k\)**-means MD** computes the cluster-wise MD on the feature space of \(g\), using the \(k\)-means cluster centers. Following previous work [69], we set \(k=5\) (under-clustering). 4. **Pseudo-labels MSP** feeds the class names to the text encoder of CLIP-based models and computes their cosine similarities for each test image, and then takes the maximum softmax probability (MSP) [36]. 5. **Pseudo-labels Probing** computes the MSP as in the previous setup and keeps the \(\mathcal{D}_{\text{in}}^{\text{train}}\) images with at least 90% probability. Then a linear head on the features of \(g\) is trained and evaluated using MSP or RMD. 6. **Few-shot \(p\)** randomly selects \(p=10\) images per class to train a linear head on the backbone features. MSP is used as the OOD score. 7. **Probing** trains a linear head on the features of the backbone \(g\), using \(\mathcal{D}_{\text{in}}^{\text{train}}\), and MSP or RMD is computed. Finally, we study the performance consistency across multiple in- and out-distributions (Table 3), as well as the robustness against the proposed adversarially manipulated OOD samples. ### Implementation details When evaluating the performance on various \(\mathcal{D}_{\text{out}}\) datasets (Table 3). We noticed that the model was able to distinguish the images solely from the different resolution levels. To avoid OOD detection from differences in image size, we resized OOD samples to \(\mathcal{D}_{\text{in}}\) image size [24]. For the text prompting, which is required for the text encoder of the CLIP models, we use the prompt: _an image of a_ {_label_}, similar to [62]. Regarding linear probing, a single linear layer was trained on the precomputed representations. We used the Adam optimizer [43] with a mini-batch size of 256 and a weight decay of \(10^{-2}\), for 20K steps in total. The learning rate follows a cosine schedule from \(10^{-3}\to 10^{-6}\). For few-shot probing, we take the average AUROC over 5 runs and only train for 10K steps. To create the adversarial dataset we perform 250 steps with the Adam optimizer with a learning \begin{table} \begin{tabular}{l c c c c c c} \hline \hline CLIP ViT-G/14 & \(\mathcal{D}_{\text{in}}\) & \(\mathcal{D}_{\text{in}}\) class & \(\mathcal{D}_{\text{in}}\):CIFAR100 & \(\mathcal{D}_{\text{in}}\):CIFAR100 & \(\mathcal{D}_{\text{in}}\):CIFAR10 & \(\mathcal{D}_{\text{in}}\):CIFAR10 \\ & labels & names & \(\mathcal{D}_{\text{out}}\):CIFAR10 & \(\mathcal{D}_{\text{out}}\):TinyIN & \(\mathcal{D}_{\text{out}}\):CIFAR100 & \(\mathcal{D}_{\text{out}}\):STL10 \\ \hline \(k\)-means MD (\(k=5\)) & ✗ & ✗ & 68.2 & 73.5 & 90.3 & 74.4 \\ \(1\)-NN & ✗ & ✗ & **87.6** & **85.2** & **98.2** & **81.1** \\ \hline Pseudo-labels MSP & ✗ & ✓ & 88.6 & 83.4 & 95.7 & 56.9 \\ Pseudo-labels Probing + MSP & ✗ & ✓ & 92.7 & 87.2 & 98.3 & 62.0 \\ Pseudo-labels Probing + RMD & ✗ & ✓ & **96.4** & **88.8** & **98.5** & **64.0** \\ \hline MD & ✓ & ✓ & 73.1 & 77.3 & 91.1 & **74.5** \\ RMD & ✓ & ✓ & 96.3 & 89.0 & 98.8 & 65.7 \\ Few-shot \(p=10\) & ✓ & ✓ & 91.1 & 86.9 & 97.3 & 62.3 \\ Probing + MSP & ✓ & ✓ & 95.0 & 88.0 & **99.0** & 66.4 \\ Probing + RMD & ✓ & ✓ & **97.4** & **89.5** & **99.0** & 67.5 \\ \hline \hline \end{tabular} \end{table} Table 2: **OOD detection AUROCs (%) for different evaluations and scores.** The considered ViT-G/14 model is trained on LAION-2B with the CLIP objective [11]. rate of \(10^{-3}\) on 1K images. We set \(\lambda\) to \(5\cdot 10^{3}\) when applying smoothing (Eq. 7). All the experiments were carried out in a single NVIDIA A100 with 40GB VRAM. Crucially, only a maximum required VRAM of \(16\)GB was needed to compute the synthetic adversarial datasets with the ViT-G architecture. It is also worth noting that all our OOD detection evaluations can be conducted within minutes, in contrast to existing approaches [23, 65] that fine-tune the backbone. ## 4 Experimental results We find a strong positive correlation between the \(\mathcal{D}_{\text{in}}\) classification accuracy and AUROC of 1-NN OOD detection across \(32\) models (Fig. 1). CLIP models exhibit the highest performance in both performance scores, independent of their network architecture. Larger CLIP models exhibit higher performance gains when trained on LAION-2B. Based on this, we select the largest model, CLIP ViT-G [11], for further evaluations. We always report absolute gains and AUROC scores. A substantial gain of 9.2% AUROC (78.4\(\rightarrow\)87.6%) compared to previous unsupervised state-of-the-art (CSI) on CIFAR100 \(\rightarrow\) CIFAR10 is reported in Table 1, without any additional assumption about \(\mathcal{D}_{\text{in}}\). On CIFAR10 \(\rightarrow\) CIFAR100, a consistent improvement of 3.2% is obtained, where we achieved 96.3% AUROC. Using the RMD score (Eq. 5) on CLIP ViT-G, we match the supervised state-of-the-art OOD detection performance (\(\sim\)96.2% AUROC on CIFAR100 \(\rightarrow\) CIFAR10) without any in-distribution specific tuning. On the same benchmark, we report an improvement of 1.2% AUROC (96.2 \(\rightarrow\) 97.4% ) by training a linear layer on the precomputed image features. The gains are less significant on the supervised CIFAR10 \(\rightarrow\) CIFAR100 setup (99.0% versus 98.8% AUROC by [65]). The OOD detection performance for CIFAR10 \(\rightarrow\) CIFAR100 is close to optimal, suggesting that more challenging setups should be adopted in the future. In Table 2, we conduct extensive experimental evaluations using CLIP ViT-G. Combining \(k\)-means with MD yields an inferior AUROC compared to \(1\)-NN, precisely lower by 11.4% on average. By incorporating the class names using CLIP's text encoder, we find that the pseudo-labels can be leveraged for probing, resulting in an improvement of 8.8% and 3.6% on CIFAR100 \(\rightarrow\) CIFAR10 and CIFAR100\(\rightarrow\)TinyIN. A counter-example is CIFAR10\(\rightarrow\)STL10 where there is an 80% class overlap. There, incorporating labels or class names deteriorates performance. Concerning linear probing, RMD consistently outperforms MSP when applied to the classifier's logits; for instance, 1.95% mean gain on CIFAR100 \(\rightarrow\) CIFAR10 and CIFAR100\(\rightarrow\)TinyIN. Next, we compare the performance of our approach for different in- and out-distribution datasets using CLIP ViT-G (Table 3). Our proposed unsupervised method surpasses other unsupervised OOD detection methods by significant margins on all benchmarks. Overall, we find very few OOD settings where the performance is not close to the maximum, like CIFAR100\(\rightarrow\)TinyIN. ## 5 Discussion Can CLIP ViTs still be fooled?The acquired state-of-the-art OOD detection performance along with the fact that CLIP ViTs are known to be robust zero-shot image classifiers against natural distribution shifts [62], raises the question of whether these models are also robust OOD detectors. To \begin{table} \begin{tabular}{c|c|c c c c c c} \hline \hline \multirow{2}{*}{\(D_{\text{train}}^{\text{in}}\)} & \multirow{2}{*}{\(D_{\text{test}}^{\text{out}}\)} & \multicolumn{8}{c}{OOD detection AUROC (\%) \(\uparrow\)} \\ \cline{3-8} & & Baseline [36] & RP [39] & SupSimCLR [42] & SSD+ [69] & CSI (ens) [72] & MTL [56] & Ours \\ \hline \multirow{4}{*}{\(D_{\text{out}}^{\text{out}}\)} & SVHN & 92.9 & 98.0 & 97.2 & 93.8 & 97.4 & 96.6 & **99.5** \\ & Texture & 87.7 & 96.3 & 94.2 & 94.1 & 97.2 & 96.9 & **99.4** \\ & Places365 & 88.4 & 92.6 & 91.1 & 91.8 & 93.1 & 98.7 & **98.9** \\ & TinyIN & 87.4 & 92.1 & 92.1 & 90.3 & 92.5 & 93.6 & **95.2** \\ & LSUN & 89.9 & 93.6 & 92.1 & 94.4 & 94.0 & 94.1 & **99.9** \\ \hline \hline \multirow{4}{*}{\(D_{\text{train}}^{\text{out}}\)} & SVHN & 79.2 & 83.6 & 81.6 & 83.6 & 87.4 & 90.6 & **95.0** \\ & Texture & 75.3 & 82.4 & 76.8 & 81.4 & 78.3 & 78.0 & **93.2** \\ & Places365 & 76.1 & 74.6 & 75.4 & 79.2 & 78.1 & 92.6 & **91.6** \\ & TinyIN & 78.5 & 77.6 & 80.8 & 76.3 & 82.4 & 79.3 & **85.2** \\ & LSUN & 73.7 & 71.9 & 73.5 & 63.8 & 75.2 & 74.0 & **94.0** \\ \hline \hline \multirow{4}{*}{\(D_{\text{train}}^{\text{out}}\)} & Flowers & 87.7 & 92.1 & 93.8 & 96.5 & 96.2 & 97.2 & **98.7** \\ & CUB-200 & 85.3 & 90.6 & 89.2 & 96.6 & 94.2 & 96.4 & **97.3** \\ \cline{1-1} & Dogs & 90.3 & 93.3 & 95.2 & 95.2 & 97.6 & 97.1 & **98.9** \\ \cline{1-1} & Food & 78.9 & 85.1 & 83.6 & 85.5 & 89.0 & 96.5 & **99.1** \\ \cline{1-1} & Texture & 87.0 & 92.2 & 98.7 & 94.9 & 98.5 & 94.0 & **99.2** \\ \hline \hline \end{tabular} \end{table} Table 3: **OOD detection AUROC (%) for various in- and out-distributions. We use the ViT-G/14 trained on LAION-2B and report the \(1\)-NN score (zero-shot). AUROC values of other methods are reported from MTL [56].** answer this question, we examine CLIP ViT-G's robustness against the introduced adversarially manipulated OOD samples (Fig. 2). We found that it is possible to drop the AUROC from 86.2% to 48.6% by using CIFAR10-A. Introducing the smoothness restriction (CIFAR10-AS) degrades performance to 54.2% AUROC. Note that an AUROC score of 50% is the score of a random guess. The above findings show that even the top-performing CLIP ViTs trained on billion-scale image-text pairs can easily be fooled by weak manipulation of the input signal, which is invisible to humans. Pretrainings and learned representations.To illustrate the impact of the pretraining dataset and objective on the learned features, we used the same architecture ViT-B/16 across \(8\) pretraining setups, as demonstrated in Fig. 3. We notice that the optimal choice of backbone depends on both \(\mathcal{D}_{\text{in}}\) and \(\mathcal{D}_{\text{out}}\). We show that this choice is also not symmetric: the best choice for CIFAR100 \(\rightarrow\) CIFAR10 is CLIP LAION-2B, but for CIFAR10 \(\rightarrow\) CIFAR100 it is DINO pretrained on ImageNet. Notably, both MAE pretrained on ImageNet and supervised ImageNet-21K pretrainings are consistently the worst backbone choices. DINO [7] even surpasses supervised ImageNet pretraining on CIFAR100 \(\rightarrow\) CIFAR10, while being the best choice on CIFAR10 \(\rightarrow\) CIFAR100, outlining the transferability of features of self-supervised methods, which is consistent with the results of [19]. The observed inferior performance of ImageNet-21K versus ImageNet suggests that supervised feature representations for OOD detection benefit from diverse and mutually exclusive class labels (ImageNet-21K class labels are not mutually exclusive) [67]. Consequently, we expect the more \(\mathcal{D}_{\text{in}}\) classes are shared with \(\mathcal{D}_{\text{pretrain}}\) and the higher the number of mutually exclusive classes for \(\mathcal{D}_{\text{pretrain}}\), the better the overall OOD detection performance. We believe that natural language supervision affects the learned representations conceptually similar to supervised learning. In fact, we observe a mean performance gain of 4.9% by purely scaling up the dataset from 400M (OpenAI) to 2B (LAION) image-text pairs using ViT-B. The latter is likely attributed to the increased diversity of both language "labels" and images, as discussed in [22]. Nonetheless, the fact that CLIP ViT-G takes a "shortcut" based on the image resolution (Section 3.5) and is sensitive to adversarial attacks gives evidence that besides label-related features, local pixel information significantly affects the learned representations. We encourage future work to investigate this in greater depth. Are we done with CIFAR datasets for OOD detection?Although we can identify challenging OOD detection scenarios, such as CIFAR100\(\rightarrow\)TinyIN, all benchmarks involving CIFAR10 and almost all involving CIFAR100 as in-distribution seem to saturate. Based on our results in Table 3, we believe that the OOD performance studies should include more challenging and diverse benchmarks. This will enable the design of robust and highly accurate OOD detectors. ## 6 Conclusion In this work, a thorough experimental study was presented by leveraging pretrained models for visual OOD detection. It was demonstrated that CLIP ViTs are powerful zero-shot OOD detectors, without requiring labels or class names, outperforming all previous unsupervised approaches by large margins. Supervised state-of-the-art OOD detection performance was also reported without the need to fine-tune the feature extractors. The top-performing CLIP ViT-G [11] was further evaluated under several OOD settings. Based on the reported performance saturation on most existing benchmarks, the need for new and more challenging benchmarks Figure 3: **AUROC values for zero-shot OOD detection using ViT-B/16 pretrained on different datasets (IN, IN-21K, OpenAI-400M, LAION-2B) and pretext tasks**. IN indicates ImageNet. The horizontal line indicates human-level performance, as reported in Fort et al. [23]. was highlighted. Finally, a novel adversarial OOD data manipulation method was introduced, which pointed to the fact that billion-scale feature extractors (CLIP ViT-G) are still sensitive to adversarial attacks.

2305.10615

ML-SUPERB: Multilingual Speech Universal PERformance Benchmark

Speech processing Universal PERformance Benchmark (SUPERB) is a leaderboard to benchmark the performance of Self-Supervised Learning (SSL) models on various speech processing tasks. However, SUPERB largely considers English speech in its evaluation. This paper presents multilingual SUPERB (ML-SUPERB), covering 143 languages (ranging from high-resource to endangered), and considering both automatic speech recognition and language identification. Following the concept of SUPERB, ML-SUPERB utilizes frozen SSL features and employs a simple framework for multilingual tasks by learning a shallow downstream model. Similar to the SUPERB benchmark, we find speech SSL models can significantly improve performance compared to FBANK features. Furthermore, we find that multilingual models do not always perform better than their monolingual counterparts. We will release ML-SUPERB as a challenge with organized datasets and reproducible training scripts for future multilingual representation research.

Jiatong Shi, Dan Berrebbi, William Chen, Ho-Lam Chung, En-Pei Hu, Wei Ping Huang, Xuankai Chang, Shang-Wen Li, Abdelrahman Mohamed, Hung-yi Lee, Shinji Watanabe

2023-05-18T00:01:27Z

http://arxiv.org/abs/2305.10615v2

# ML-SUPERB: Multilingual Speech Universal PERformance Benchmark ###### Abstract Speech processing Universal PERformance Benchmark (SUPERB) is a leaderboard to benchmark the performance of Self-Supervised Learning (SSL) models on various speech processing tasks. However, SUPERB largely considers English speech in its evaluation. This paper presents multilingual SUPERB (ML-SUPERB), covering 143 languages (ranging from high-resource to endangered), and considering both automatic speech recognition and language identification. Following the concept of SUPERB, ML-SUPERB utilizes frozen SSL features and employs a simple framework for multilingual tasks by learning a shallow downstream model. Similar to the SUPERB benchmark, we find speech SSL models can significantly improve performance compared to FBANK features. Furthermore, we find that multilingual models do not always perform better than their monolingual counterparts. We will release ML-SUPERB as a challenge with organized datasets and reproducible training scripts for future multilingual representation research. Jiatong Shi\({}^{1}\), Dan Berrebbi\({}^{1}\), William Chen\({}^{1*}\), Ho-Lam Chung\({}^{2*}\), En-Pei Hu\({}^{2*}\), Wei Ping Huang\({}^{2*}\), Xuankai Chang\({}^{1}\), Shang-Wen Li\({}^{3}\), Abdelrahman Mohamed\({}^{4}\), Hung-yi Lee\({}^{2}\), Shinji Watanabe\({}^{1}\)+\({}^{1}\)Carnegie Mellon University \({}^{2}\)National Taiwan University \({}^{3}\)Meta AI \({}^{4}\)Rembrand {jiatongs, dberrebbi, wc4, swatanab}@cs.cmu.edu, [email protected], [email protected] [email protected] Footnote †: Equal contribution, sorted in alphabetical order. **Index Terms**: speech self-supervised learning, multilingual speech recognition, language identification ## 1 Introduction Self-supervised learning (SSL) has been a popular method in the speech community. SSL models have shown promising results by capturing important speech features, such as phonemes and other acoustic units, through training on large amounts of unlabeled speech data [1]. These models have led to significant improvements in downstream tasks, such as speech recognition, speaker identification, and emotion recognition [2]. Over the past few years, researchers have proposed a variety of SSL models with different training objectives, operating under various data conditions, model architectures, and modalities [3, 4]. A major challenge in evaluating SSL models for speech is the difficulty of comparison since most models have been evaluated using different experimental setups. To address this issue, Yang et al. introduced the Speech processing Universal PERformance Benchmark (SUPERB) [2]. Recently, an extension of SUPERB called SUPERB-SG [5] has been introduced. SUPERB provides a comprehensive speech SSL benchmark including tasks such as recognition, detection, semantics, speaker identification, paralinguistics, and generation. With SUPERB, researchers can more easily compare the performance of different SSL models on various speech-related tasks, universally. While SUPERB covers a wide range of speech tasks, it was designed primarily for English speech. However, there has been growing interest in applying SSL models to multilingual scenarios, such as training multilingual SSL models [6, 7, 8] or using SSL models in a cross-lingual manner [9, 10, 11, 12]. To support future research in these areas, we propose a new benchmark called multilingual SUPERB (ML-SUPERB). ML-SUPERB is designed to cover a wide range of languages, including both high-resource languages like English and endangered languages such as Totonac. The benchmark primarily focuses on evaluating SSL models for automatic speech recognition (ASR) and language identification (LID). To accommodate different use cases for SSL models, ML-SUPERB includes two tracks with four different tasks: the monolingual track (monolingual ASR), and the multilingual task (multilingual ASR, LID), joint multilingual ASR/LID). Similar to SUPERB, ML-SUPERB employs frozen SSL models as feature extractors and a lightweight downstream model that can be fine-tuned for different tracks to achieve high training efficiency. Several existing benchmarks also include multilingual SSL models [13, 14, 15]. Lebenchmark primarily evaluates speech tasks in French [13]; IndicSUPERB focuses mostly on Indian languages [14]. Xtreframe-S focuses on multilingual speech representation benchmarks, including ASR, speech translation, speech classification, and speech retrieval [15]. There are three main differences between Xtreframe-S and ML-SUPERB. Firstly, ML-SUPERB covers a wider range of languages, with 143 languages compared to Xtreframe-S's 102. Secondly, ML-SUPERB focuses on ASR and LID, while Xtreframe-S covers four different tasks. However, ML-SUPERB expands the tasks by evaluating them in four common multilingual research scenarios, while Xtreframe-S considers multilingual training only. Finally, ML-SUPERB is designed for efficiency, using smaller benchmark datasets and downstream models, and does not include fine-tuning. This lightweight setup allows us to conduct experiments for a dozen of popular speech SSL models, trained with various sizes and pre-training sets, and compare their performances across the proposed tracks. We expect ML-SUPERB would be a valuable complement to existing benchmarks. ## 2 Benchmark Details ### Data Collection ML-SUPERB gathers data from a wide range of multilingual speech corpora, including Multilingual Librispeech [16], Commovioce [17], Voxforge [18], Voxpopuli [19], Google18n open-source project [20, 21, 22], Nordic Language Technology ASR corpora [23], Fleurs [24], NCHLT Speech [25], Spoken Wikipedia corpus [26], Mexican endangered languages [10, 27, 28], M-AILab multilingual corpora [29], Living Audio dataset [30], ALFFA corpus [31]. All corpora are with either Creative Commons, MIT, GNU, or Free-BSD licenses, which are available for both industrial and academic research, permissively. For each language-corpus pair denoted as (lang, data), three 10-minute subsets are randomly extracted for training, development, and testing, along with an additional 1-hour training set that includes the 10-minute training set.1 The reasons for using a small 10-minute/1-hour training set: (1) _Challenging design_: using a large training data size could lead to high performance easily and may result in a saturated benchmark in evaluation metrics [3, 4]. Therefore, using a smaller training set size presents a more challenging design for the SSL models, which can help evaluate their robustness and generalization capability. (2) _Reasonable performance_: previous speech SSL works have frequently adopted 10-minute and 1-hour training sizes. Even in such extreme cases, the performances with SSL are generally reasonable [3, 4], indicating that this setting could be a feasible solution to the benchmark as well. (3) _Training efficiency_: with 143 languages coverage, limiting the training size is important to keep the experiments within reasonable computational efforts. Using a smaller training set size can help reduce the computational cost and make the training process more efficient. A full evaluation cycle of ML-SUPERB can take up to 3 days using 4 2080Ti GPUs. Footnote 1: We used the original split for source datasets, with the exception of SWC, M-AILABS, LAD, and ALFFA. Therefore, all datasets except these four can be used for SSL pre-training. Additionally, the benchmark includes few-shot cases with 20 languages and uses only 5 utterances in training for each language. These reserved few-shot training sets are not used in the monolingual ASR track. A detailed summary of the dataset is shown in Table 1. ### Monolingual Track The literature suggests that speech SSL models are commonly fine-tuned on monolingual corpora [9, 10, 11]. In ML-SUPERB, we introduce a dedicated track for monolingual ASR to facilitate this approach. We select nine languages based on geographical and linguistic considerations to balance language and domain coverage with manageable experimental mass. In total, we introduce 14 monolingual_exp. For a monolingual_exp in language lang we select one dataset of this language and use it for training the model and for validation2. For evaluation of a monolingual_exp, we use all the datasets of lang to test the trained model on various accent or domain conditions. We select one pair (lang, data) for training for lang{rus, swa, swe, jpn, cmm, xty}. For lang{eng, fra, deu} we select respectively 3, 2 and 2 pairs (lang, data) in order to evaluate the impact of the training domain on the models' performances. For instance, for eng we have 3 monolingual_exp, with (eng, MLS), (eng, NCHLT) and (eng,VoxPopuli). Footnote 2: Each monolingual_exp is made of one experiment with the 10-minute set for training and one with the 1-hour set. ### Multilingual Track **Multilingual ASR task**: in the multilingual ASR task, we use the training set where combining text transcriptions from all 143 languages. The multilingual ASR task has two sub-tasks on the 10-minute train set and the 1-hour train set. For both training sets, we reserve 20 languages for few-shot learning scenarios as discussed in Sec. 2.1. In this track, the model is expected to directly predict the correct orthography in the target language. **LID task**: LID track focuses on language identification with the same training set of 143 languages in 10 minutes and 1 hour. However, we do not consider evaluation for languages with few-shot settings, given that the identification of those languages is very challenging due to the label biasing. **Joint Multilingual ASR/LID task**: A widely used technique in previous literature involves adding the language ID to the start of the speech transcript to facilitate joint training of multilingual ASR and LID models [32, 33, 34, 35]. Joint training can improve performance in certain scenarios, and it can also enhance model interpretability by separating language identification errors. Therefore, we have included this task in our multilingual track. The task's design is the same as the multilingual ASR task for ASR and the LID task for language identification. ### Framework and Benchmark Settings **Toolkits**: We utilize the S3PRL toolkit [2] for upstream models, which offers a wide range of speech SSL model architectures and APIs that support customized SSL models from Huggingface [36] and user-defined models. For task-specific downstream training, we use ESPnet [37]. We plan to publish ML-SUPERB as an all-in-one recipe in ESPnet's egsz recipe collection, encompassing data preprocessing, training, inference, and evaluation3. Footnote 3: [https://github.com/espnet/espnet/tree/master/egs2/ml_superb/asr1](https://github.com/espnet/espnet/tree/master/egs2/ml_superb/asr1) **Downstream model and training details**: Our downstream model design is based on the SUPERB concept. First, we compute a weighted summation of frozen speech SSL representations using learnable weights. Next, we apply a convolutional downsample layer that reduces the sequence of speech SSL features by half, passing the resulting hidden states to a transformer model consisting of two layers with an attention dimension of 256, a feedforward layer dimension of 1024, and 8 attention heads. A dropout rate of 0.1 is employed, and the model is trained using the connectionist temporal Cessification loss. We use the Adam optimizer with a learning rate of 0.0001 and 1e-6 weight decay. Specaugment is applied to the representation (i.e., the weighted sum of speech SSL representation) following the SUPERB benchmark. The batch size is set to 8 with the gradient accumulation as 4. The same configuration is used for all tasks in both the monolingual and multilingual tracks. The number of iterations in training is the only difference across tasks. In the monolingual track, due to the small training size, we set it to 15,000. In the multilingual track, we use 300,000 iterations for the 10-minute train set and 600,000 for the 1-hour train set. **Evaluation metric**: In the monolingual track, the phoneme error rate is used for jpn and cmn, while Character Error Rate (CER) is used for the remaining languages. In the multilingual track, we use CER for ASR evaluation and accuracy rate for LID evaluation, reporting results separately for the normal training set and the few-shot training set. \begin{table} \begin{tabular}{l|c|c c} \hline \hline Dataset & Hours & Normal Langs (123) & Few-shot Langs (20) \\ \hline 10-minute & 37.43 & \(\sim\)10min \(\times\) 240 (lang, data) & 5 utt. \(\times\) 20 lang \\ 1-hour & 222.46 & \(\sim\)1h \(\times\) 240 (lang, data) & 5 utt. \(\times\) 20 lang \\ Dev.t & 41.82 & \(\sim\)10min \(\times\) 240 (lang, data) & \(\sim\)10min \(\times\) 31 (lang, data) \\ Test & 44.97 & \(\sim\)10min \(\times\) 240 (lang, data) & \(\sim\)10min \(\times\) 31 (lang, data) \\ \hline \hline \end{tabular} \end{table} Table 1: _Statistics of the data used for training, development, and testing in ML-SUPERB. Detailed discussed in Sec. 2.1._ For overall performance, we use the SUPERB\({}_{s}\) metric from the SUPERB benchmark [38]. We denote \(s_{t,i}(u)\) as the \(i^{\text{th}}\) metrics for task \(t\) and SSL model \(u\). \(T\) is the set of four tasks and \(I_{t}\) is the set of metrics for the task \(t\). SUPERB\({}_{s}\) aggregates all task-specific scores \(s_{t}(u)\) with respect to baseline (i.e., FBANK) and state-of-the-art (SOTA) model4 on the task \(t\). The SUPERB\({}_{s}\) is defined as: Footnote 4: The SOTA models for each setting are discussed in Sec. 3.2. \[\text{SUPERB}_{s}(u)=\tfrac{1000}{|T|}\sum\nolimits_{t}^{T}\frac{1}{|I_{t}|} \sum\nolimits_{i}^{I_{t}}\frac{s_{t,i}(u)-s_{t,i}(\text{FBANK})}{s_{t,i}(\text {SOTA})-s_{t,i}(\text{FBANK})} \tag{1}\] We expect SUPERB\({}_{s}\) can provide a comprehensive view of the model performance on the benchmark and take the difficulty of tasks into consideration. **Analysis support**: To facilitate a more comprehensive analysis of the benchmark, we provide various analysis tools. For the multilingual ASR evaluation, we present the character error rate (CER) for each language as well as aggregated scores for different language groups, in addition to the average CER for both normal and few-shot cases. In line with previous studies [39, 40], we also offer visualizations of the learnable layer weights and their learning curve during training. ## 3 Experiments ### Candidate models ML-SUPERB welcomes all speech SSL models trained on either monolingual or multilingual data. We believe the analysis of multilingual scenarios for monolingual speech SSLs is also valuable according to previous works [9, 10, 11]. In this paper, we show the experimental results of some example model candidates as shown in Table 2. **wav2vec2**: wav2vec2 is a popular speech SSL model for speech recognition [3]. Its pre-training uses a contrastive learning approach that prioritizes identifying true quantized latent speech representations over masked time steps from distractors. The wav2vec2 model has also been extended to many other versions for specialized use cases. For example, robust-wav2vec2-large [41] considers the diversity of speech types, such as read speech, conversational speech, and noisy speech, by including additional corpora in the pre-training stage. Wav2vec2-base-23 and wav2vec2-large-23 are pre-trained on Voxopopuli [19], with a focus on European languages. Additionally, XLSR scales up the multilingual training in wav2vec2 by incorporating more languages and data [6, 7]. **HuBERT**: HuBERT uses an iterative offline clustering step to generate pseudo labels for each frame. During training, it predicts the pseudo labels of the masked frame, which helps to improve the quality of the learned features. Similar to wav2vec2, HuBERT also has different versions, such as a multilingual Hubert [43] trained in three European languages (fra, spa, eng) and HuBERT trained on Mandarin [42]. ### Experimental Results The experimental results are shown in Table 3 for 10-minute set and Table 4 for 1-hour set. **Monolingual ASR**: In the monolingual ASR task, all speech SSL models outperform the FBANK baseline. XLSR-128 achieves the best performance in the 1-hour set, while HuBERT-large obtains the best performance in the 10-minute set. Several findings are noteworthy: (1) HuBERT-based models outperform wav2vec2-based models when the training data and model size are similar. (2) Large models usually obtain better results than their base versions. (3) While the XLSR series of models deliver impressive performances in the 1-hour set, we have observed their instability in the 10-minute set, particularly on Asian languages such as cmn. **Multilingual ASR**: In the multilingual ASR task, all models trained using self-supervised learning (SSL) techniques have shown superior performance compared to the baseline model using FBANK features. Among the SSL models, XLSR-128 achieves the best results across all conditions. Our experiments also reveal some interesting findings: (1) Models trained with more languages generally outperform those trained on monolingual datasets, although this may not always be the case. For example, mHuBERT-base performs worse than HuBERT-based models trained on English only. (2) Large models trained on monolingual data do not necessarily have better representations for multilingual scenarios. For instance, HuBERT-large performs worse than HuBERT-base, and wav2vec2-large is less effective than wav2vec2-base. One possible explanation for the lack of performance improvement with larger models is their limited ability to generalize, despite having similar training losses as base models. (3) The robust-wav2vec2-large model achieves decent scores on multilingual ASR, suggesting that our benchmark corpus may need to consider different acoustic environments, as it includes multiple source datasets. **LID**: In the LID task, we notice similarities with multilingual ASR, but there are also notable differences. (1) XLSR-128 has been the dominant model for both 10-minute and 1-hour datasets. (2) While most SSL models have improvements over FBANK, some do not, particularly those based on wav2vec2 (e.g., wav2vec2-large-23 for the 1-minute set and wav2vec2-large for the 1-hour set). (3) Larger models with more parameters and pre-trained data do not necessarily lead to better performance compared to base models. **Joint Multilingual ASR + LID**: In the joint multilingual ASR+LID task, the results generally align with the other two tasks in the multilingual track. (1) SSL models outperform FBANK on ASR, but some models perform worse on LID. (2) Base models exhibit better generalization ability and often perform better on test sets. (3) There is no single best model that dominates the task, particularly in few-shot cases and LID tasks. **Overall**: In terms of overall performance as measured by SUPERB\({}_{s}\) in Sec. 2.4, XLSR-128 is the best model for both the 10-minute and 1-hour sets. Major findings include: (1) multilingual training with a broad coverage of languages, as seen in XLSR models that include more than 50 languages, has proven to be useful. However, multilingual training that is limited to a few selective languages may not be as beneficial in larger language groups (e.g., wav2vec2-large-23 and mHUBERT models \begin{table} \begin{tabular}{l|c|c c} \hline \hline Model & Params (M) & \multicolumn{2}{c}{Pre-Training \# Hours} & \multicolumn{1}{c}{\# Langs} \\ \hline wav2vec2-base [3] & 95 & 1k & 1 \\ wav2vec2-large [3] & 317 & 60k & 1 \\ robust-wav2vec2-large [41] & 317 & 65k & 1 \\ wav2vec2-base-23 [19] & 95 & 100k & 23 \\ wav2vec2-large-23 [19] & 317 & 100k & 23 \\ XLSR-53 [7] & 317 & 56k & 53 \\ XLSR-128 [6] & 317 & 400k & 128 \\ \hline HuBERT-base [4] & 95 & 1k & 1 \\ HuBERT-large [4] & 317 & 60k & 1 \\ HuBERT-base-cmn [42] & 95 & 10k & 1 \\ HuBERT-large-cmn [42] & 317 & 10k & 1 \\ mHuBERT-base [43] & 95 & 14k & 3 \\ \hline \hline \end{tabular} \end{table} Table 2: Description of the candidate models. do not always perform better than their models trained in a single language). (2) The base models tend to generalize better to multilingual cases than their corresponding large versions, such as wav2vec2-base versus wav2vec2-large and HuBERT-base versus HuBERT-large. ### Layerwise analysis Our benchmark offers tools to guide users in the use of SSL representations according to their needs, including an analysis of the learned weights for layer importance. The results for the XLSR-128 model in monolingual ASR tasks (shown in Fig 1) confirm the conclusions reached by [44] and [45]: the most relevant layers for ASR are not the last few layers. We also observed that English3, French2, and German2 have very similar behavior. These tasks use VoxPopuli data for training, which is the only dataset with lecture speech in our collection. Additionally, Mixtec is the only conversational speech data among our sets, and we can see a distinct behavior in Fig 1. Therefore, the relevance of SSL model layers may be related to the speech domain (in addition to the speech task) rather than the language. ## 4 Conclusion This paper introduces ML-SUPERB, a benchmark that extends SUPERB to multilingual tasks. We present the design of the open-source framework and discuss experimental results for some example models. More detailed policies can be found at [https://multilingual.superbbenchmark.org/](https://multilingual.superbbenchmark.org/). We invite the community to participate in this challenge.

2308.16298

Publishing Wikipedia usage data with strong privacy guarantees

For almost 20 years, the Wikimedia Foundation has been publishing statistics about how many people visited each Wikipedia page on each day. This data helps Wikipedia editors determine where to focus their efforts to improve the online encyclopedia, and enables academic research. In June 2023, the Wikimedia Foundation, helped by Tumult Labs, addressed a long-standing request from Wikipedia editors and academic researchers: it started publishing these statistics with finer granularity, including the country of origin in the daily counts of page views. This new data publication uses differential privacy to provide robust guarantees to people browsing or editing Wikipedia. This paper describes this data publication: its goals, the process followed from its inception to its deployment, the algorithms used to produce the data, and the outcomes of the data release.

Temilola Adeleye, Skye Berghel, Damien Desfontaines, Michael Hay, Isaac Johnson, Cléo Lemoisson, Ashwin Machanavajjhala, Tom Magerlein, Gabriele Modena, David Pujol, Daniel Simmons-Marengo, Hal Triedman

2023-08-30T19:58:56Z

http://arxiv.org/abs/2308.16298v2

# Publishing Wikipedia usage data with strong privacy guarantees ###### Abstract For almost 20 years, the Wikimedia Foundation has been publishing statistics about how many people visited each Wikipedia page on each day. This data helps Wikipedia editors determine where to focus their efforts to improve the online encyclopedia, and enables academic research. In June 2023, the Wikimedia Foundation, helped by Tumult Labs, addressed a long-standing request from Wikipedia editors and academic researchers: it started publishing these statistics with finer granularity, including the country of origin in the daily counts of page views. This new data publication uses differential privacy to provide robust guarantees to people browsing or editing Wikipedia. This paper describes this data publication: its goals, the process followed from its inception to its deployment, the algorithms used to produce the data, and the outcomes of the data release. ## 1 Introduction Wikipedia and other projects supported by the Wikimedia Foundation are among the most used online resources in the world, garnering hundreds of billions of visits each year from around the world. As such, the Foundation has access to terabytes of data about visits to a page on a Wikimedia project. This is called _pageview_ data in this document. The Foundation has been publishing statistics about this data for almost 20 years, through the _Pageview API_[17]. This data helps Wikipedia editors measure the impact of their work, and focus their efforts where they are most needed. Pageview data is also a rich resource for academic research: it has been used to better understand many topics, ranging from user behavior [14] and browsing patterns [15] to information dissemination [1], epidemiology [19], online harassment [33], and others. Over time, the Wikimedia Foundation received a number of requests to make these statistics more granular, and publish pageview counts _by country_, to make it even more useful to Wikipedia editors, and enable further academic research. Addressing such requests for more granular data is aligned with the Foundation's open access policy [12], which seeks to provide as much transparency as possible about how Wikimedia projects operate. However, the Foundation also considers privacy to be a key component of the free knowledge movement: there cannot be creation or consumption of free knowledge without a strong guarantee of privacy. These guarantees are expressed by the Foundation's strict privacy policy [13] and data retention guidelines [6], which govern how the infrastructure underlying Wikipedia works. Concretely, people browsing Wikipedia may expect their behavior on the website to stay private: is is crucial to prevent motivated actors to combine this data with outside other data sources in order to spy on or presence Wikipedia users for their view history, edit history, or other behavior. It is well-known that simply aggregating data is not, on its own, enough to prevent re-identification risk [34, 23, 22, 25], so publishing data with a finer geographic granularity warrants an approach with rock-solid privacy guarantees for Wikipedia users and editors. Differential privacy [27] (DP) provides a way of easing this tension: it allows organizations to both lower and more fully understand the risks of releasing data. Therefore, the Wikimedia Foundation decided to investigate the use of differential privacy to release daily pageview data, sliced by country. After an in-depth comparison of available open-source tools [5], the Wikimedia Foundation decided to use Tumult Analytics [30, 18] and started a collaboration with Tumult Labs to design and deploy a DP pipeline for this data release. The pipeline is now deployed, and the published data provides useful insights to anyone interested in better understanding Wikipedia usage. This document describes this data release in more detail. * In Section 2, we present the high-level workflow that we followed towards the deployment of a differentially private data release. * In Section 3, we outline the problem statement and the success metrics for this data release. * In Section 4, we describe the technical algorithms used for this data release. * In Section 5, we summarize the results of this deployment. ## 2 High-level workflow for differential privacy deployments The process to launch a DP data product follows a standard workflow, with three main stages: _Build_, _Tune_, and _Deploy_. The entire process is outlined in Figure 1; its three main stages are as follows. 1. In the initial Build stage, the goal is to gain a good understanding of the problem and its requirements, and implement a first-cut algorithm. There are two steps in this initial stage. First, we properly define the problem, and determine what success looks like for this project. This involves talking to stakeholders to understand what the data will be used for, and what accuracy metrics capture downstream use cases well. Second, we build a prototype mechanism. This is a first rough attempt at solving the data release problem, and it exposes the "levers" inherent to the project. Which choices did we have to make while building the prototype? Which of these choices can later be modified to pick different trade-offs between utility or privacy metrics? Figure 1: A standardized workflow for differentially private data releases. 2. Then, in the Tune step, we use these levers to experiment with different settings and optimize the algorithm. Using the success metrics defined in the previous step, we iteratively evaluate and adjust the algorithm, making changes until it produces data that is fit for use and satisfies the privacy requirements. 3. Finally, in the Deploy stage, we finalize the algorithm, obtain the necessary approvals to publish the data, write documentation about the data publication mechanism for future data users and pipeline maintainers, and deploy it in production. In Section 3, we outline the output of the very first step: the definition of problem statement and its success metrics. Then, in Section 4, we will describe the output of the Tune stage: what does the final algorithm look like, after the multiple rounds of iteration on the initial prototype. ## 3 Problem statement and success metrics In this Section, we describe the desired output data (Section 3.1), the schema and characteristics of the input data (Section 3.2), the privacy goals of this data release (Section 3.3), and the accuracy metrics used to quantify success (Section 3.4). ### Desired output data The pre-existing Pageview API publishes data about the number of times each Wikimedia page was visited during a given day. Each page is identified by two fields: * its _project_, e.g. fr.wikipedia (the French-language version of Wikipedia), zh.wikibooks (the Chinese version of Wikibooks, an open-content textbook collection), wikidata (a central storage for structured data), etc.; * its _page ID_, a numeric identifier uniquely identifying each page within a project. Table 1 is a fictitious sample of the kind of data available via the Pageview API. For example, the first line indicates that there were 4217 visits to the page with ID 23110294 on the English version of Wikipedia on April 2nd, 2023. The goal of this project is to publish more granular data, and also release daily pageview counts _per country_. A fictitious sample of the desired output data appears in Table 2. For example, the first line indicates that 92 of the previously-mentioned visits originated from Switzerland. \begin{table} \begin{tabular}{l|l|l|l} Project & Page ID & Date & Count \\ \hline en.wikipedia & 23110294 & 2023-04-02 & 4217 \\ fr.wikipedia & 28278 & 2023-04-02 & 710 \\ \(\cdots\) & \(\cdots\) & \(\cdots\) & \(\cdots\) \\ \end{tabular} \end{table} Table 1: A fictitious sample from the data made publicly available via the Pageview API. \begin{table} \begin{tabular}{l|l|l|l|l} Project & Page ID & Date & Country & Count \\ \hline en.wikipedia & 23110294 & 2023-04-02 & CH & 92 \\ fr.wikipedia & 28278 & 2023-04-02 & FR & 101 \\ \(\cdots\) & \(\cdots\) & \(\cdots\) & \(\cdots\) & \\ \end{tabular} \end{table} Table 2: A fictitious sample from the data that we would like to publish as part of this project. ### Input data This project uses two input datasets: the _current pageviews dataset_, and the _historical pageviews dataset_. Current pageviews datasetAs users visit the site, their individual pageviews are recorded and stored in the current pageviews dataset. This dataset contains all pageviews across all Wikimedia projects for the last 90 days. Because of the Wikimedia Foundation's commitment to minimal data retention, this data is only kept in this form for 90 days. Table 3 is a fictitious sample of the current pageviews dataset, showing only the columns of interest for this project: project, page ID, date and time, and country. Note that in contrast to similar logging infrastructure for most websites, this data does not contain a persistent user identifier. Most visits to Wikimedia projects come from logged-out users, and the Wikimedia Foundation intentionally did not implement a user tracking mechanism which would provide a cookie ID and allow the Foundation's systems to recognize whether two records came from the same user. This practice is good for data minimization, but it makes it more difficult to obtain user-level differential privacy guarantees, which requires bounding the number of contributions coming from the same user. We come back to this challenge in Section 4.1.1. Historical pageviews datasetPast the initial 90-day retention period, pageviews are aggregated as hourly totals, broken down by project, page id, country, and a number of user characteristics. These aggregates are then stored in the historical pageviews dataset. Table 4 is a fictitious sample of the historical pageviews dataset, again showing only the columns of interest. This pre-aggregated data also poses a challenge for performing DP calculations: it is not possible to determine which contributions came from which users, and therefore to bound the number contributions coming from each user. ### Privacy goal When using differential privacy, one has to decide what to protect in the data; or, equivalently, what the definition of the neighboring databases should be. For long-running pipelines that publish data regularly over an unbounded time period, there are two aspects to this choice: what are the intervals of time considered as part of the unit of privacy, and what are we protecting in each of these intervals. Then, a follow-up question is the choice of privacy parameters: the numeric value of \(\varepsilon\) and \(\delta\). Our goal is to publish data daily: it is natural to use a daily time period in the unit of privacy. This interval is consistent with almost all other long-running DP deployments, like Apple's telemetry collection, or Google's and Meta's data releases related to the COVID-19 crisis. Other releases use a shorter period, like Microsoft's telemetry in Windows. There is no overlap between days: the privacy parameters for each user-day are fixed and do not increase over time. \begin{table} \begin{tabular}{c|c|c|c|c} Project & Page ID & Date and Time & Country & Count \\ \hline en.wikipedia & 23110294 & 2023-04-02 10:00 & CH & 11 \\ fr.wikipedia & 28278 & 2023-04-02 18:00 & FR & 15 \\ \(\cdots\) & \(\cdots\) & \(\cdots\) & \(\cdots\) & \(\cdots\) \\ \end{tabular} \end{table} Table 4: A fictitious sample of the columns of interest from the pre-aggregated historical pageviews dataset. \begin{table} \begin{tabular}{c|c|c|c} Project & Page ID & Date and Time & Country \\ \hline en.wikipedia & 23110294 & 2023-04-02 10:32:45 & CH \\ fr.wikipedia & 28278 & 2023-04-02 18:53:11 & FR \\ \(\cdots\) & \(\cdots\) & \(\cdots\) & \(\cdots\) \\ \end{tabular} \end{table} Table 3: A fictitious sample of the columns of interest from the current pageviews dataset. This choice of unit of privacy means that if a user were to regularly visit the same page from the same country across multiple devices (or clearing their cookies between each page visit) over a long period of time, this behavior could potentially be observed in the output data. Another caveat is that this data release surfaces group-level trends, like minority language group activity on Wikimedia projects within a country. These insights can be helpful (e.g. allow for dedicated support to that minority language group) but could also carry risks (e.g. by causing government persecution of this minority group). We mitigate these risks by choosing conservative privacy parameters, which translate to a reasonable level of protection over longer time periods, by holding off on releasing data for certain countries, and by only releasing aggregates that are above a certain threshold. Protecting each individual Wikipedia user during each day is impossible to achieve entirely without a way to link people's identities across records and devices. Because the Wikimedia Foundation does not have nor want the capability to link records in such a way, we instead attempt to protect the contribution of each _device_ during each day. For the data based on the current pageviews dataset, we achieve this goal using _client-side contribution bounding_, as described in Section 4.1.1. For the data based on the historical pageviews dataset, we cannot bound user contributions. Instead, we choose to protect a fixed number of daily pageviews, denoted by \(m\). This provides an equivalent level of protection to users who contribute fewer than \(m\) pageviews per day. Users who contribute more than \(m\) pageviews will incur a larger privacy loss, proportional to the amount by which their contributions exceed \(m\). This number is set to 300 for data prior to February 8th, 2017, and to 30 for data between February 9th, 2017 to February 5th, 2023. Table 5 summarizes the privacy units chosen for this project. This difference in how many contributions we protect is due the fact that in February 2017, a change occurred to the way the input data was generated. Prior to February 8th, 2017, users who were editing a Wikimedia page and used the Web UI to preview their changes were recorded as one pageview each time the preview refreshed. This meant that during a lengthy editing session, an editor could plausibly rack up many pageviews on the same page. When combined with our inability to limit user contributions, this created a markedly different risk level before/after this date, that our historical pageviews algorithm had to address. Starting on February 9th, 2017, previews were no longer recorded as pageviews. For privacy parameters, we use zero-concentrated DP [20] with \(\rho=0.015\)1 for the more recent data, and pure DP with \(\varepsilon=1\) for the historical data. These values are generally considered to be conservative among differential privacy researchers and practitioners [31], and are lower than most practical DP deployments [24]. Footnote 1: Which is a strictly stronger guarantee than \((\varepsilon,\delta)\)-DP [26] with \(\varepsilon=1\) and \(\delta=10^{-7}\). ### Accuracy metrics We measure utility along three dimensions: the _relative error distribution_, the _drop rate_, and the _spurious rate_. Each of these metrics is computed using the _true data_ as a baseline: the data that corresponds to simply running a group-by query (either counting the number of rows, for the current pageviews dataset, or summing the counts, for the historical pageviews dataset), without any contribution bounding, noise addition, nor suppression. Relative error distributionWe are releasing pageview counts, and the DP process will inject statistical noise into these counts. Thus, it is natural to want to measure how much noise is added to these counts. We measure accuracy according to _relative error_: the relative error of each noisy count \(\hat{c}\) is \(|\hat{c}/c|\), where \(c\) is \begin{table} \begin{tabular}{c|c} Time period of the input data & Unit of privacy \\ \hline July 1st, 2015 – February 8th, 2017 & 300 daily pageviews \\ February 9th, 2017 – February 5th, 2023 & 30 daily pageviews \\ February 6th, 2017 – present & one user-day \\ \end{tabular} \end{table} Table 5: A summary of the privacy units used in this project. the true count. Of course, we are releasing many counts, so we need to look at the _distribution_ of relative error. More specifically, we look at the percentage of released counts having a relative error smaller than 10%, 25%, and 50%. Drop rateThe DP algorithm uses _suppression_: if a noisy count is lower than a given threshold, we remove it from the output data. To quantify the data loss due to this suppression step, it is natural to compute the _drop rate_: the percentage of counts that do not appear in the output, even though they were non-zero in the true data. In the true data, however, many of the counts are very low; suppressing such counts is not as bad as suppressing a popular page. Therefore, we compute the percentage of pages that were suppressed among pages whose true counts is larger than a fixed threshold \(t\) (the _drop rate above \(t\)_), as well as the percentage of pages that were suppressed among the top 1000 rows in the true data (the _top-1000 drop rate_). Spurious rateMany page, project, and country combinations receive zero pageviews on any particular day. When noise is added to these zero counts, it is likely that they will end up with positive (though comparatively small) counts. We refer to these as _spurious_ counts. Spurious counts can mislead data users by wrongly indicating that some combinations had activity. They also increase the size of the output dataset, which can pose a usability challenge. Therefore, we compute an additional metric: the _spurious rate_, which captures the ratio of spurious counts among all counts that appear in the output. ## 4 Technical description of the algorithms In this Section, we describe the algorithms used to generate the differentially private data. For simplicity, we refer to a \(<\)page ID, project\(>\) pair as a _page_. ### Current pageviews For the data using the current pageviews dataset, we want to provide privacy guarantees that protect each user during each day. This requires bounding the maximum number of pageviews that each user can contribute during a single day. The typical way to perform such contribution bounding is to use a user identifier to sub-sample the number of contributions from each user, taking the first \(k\) records [29], or using reservoir sampling [32]. However, without a user identifier, we had to use a novel and alternative approach to this problem: _client-side filtering_. #### 4.1.1 Client-side filtering Without user IDs, the server cannot know whether multiple contributions come from the same user, and perform contribution bounding to get user-level privacy guarantees. Instead, we add some logic to the client side. Each end-user device counts their number of contributions logged in each day, and sends each contribution along with a boolean flag, indicating whether this contribution should be used in the server-side DP computation. The criteria used for inclusion in the input to the DP algorithm is as follows: each day, we include the first 10 _unique_ pageviews. This means that if a user visits the same page multiple times in a day, only the first visit will be counted. This also means that if a user visits more than 10 distinct pages in a day, all pageviews after the 10th visits will not be included. Pseudocode for this client-side filtering step can be found in Algorithm 1. Note that this algorithm does not keep track of the raw page IDs in the client-side cookie. Instead, it uses a salted hash function [16] to remember which page IDs were already visited. This provides an additional level of protection against an attacker that would obtain access to this cookie. Client-side filtering upholds the Wikimedia Foundation's data minimization principle: only the absolute minimal information needed to perform the contribution bounding -- a boolean value associated with each pageview to indicate whether it should be counted -- is added to the logging infrastructure. Alternatives such as using identifiers or a counter that increments for each contribution would have required sending more data to the server, and increase fingerprinting risk. #### 4.1.2 Server-side algorithm Once each pageview was annotated by the client-side filtering algorithm, it is used as input in a server-side differentially private algorithm. This algorithm, run daily on the data from the previous day, has three stages. 1. First, we collect the list of \(<\)page, country\(>\) tuples to aggregate over. 2. Second, we count the number of pageviews in each group, and we add noise to each count. 3. Finally, we suppress low counts, and publish the data. The list of all possible tuples is, in theory, known in advance: the list of Wikimedia pages and countries are both public information. However, the majority of \(<\)page, country\(>\) combinations do not appear in the input data: including all of them would be inefficient and lead to increased spurious data. Instead, we use existing public data to only include a small fraction of these possible counts. On each day, we list all Wikimedia pages with more than \(t\) global pageviews, according to the existing Pageview API, where \(t\) is an arbitrary ingestion threshold. Then, we take the cross-product between these pages and the list of countries2 to create the groups. Footnote 2: This list is based on [7]; excluding countries identified by the Wikimedia Foundation as potentially dangerous for journalists or internet freedom [3]. The second step uses the Gaussian mechanism [28] to add noise to counts. This provides two advantages. First, because each user can contribute to at most 10 _different_\(<\)page, country\(>\) tuples, but only once to each, we get a tighter \(L_{2}\) sensitivity bound (\(\sqrt{k}\)) than if we had used \(L_{1}\) sensitivity (\(k\)): this allows us to add less noise. Second, because the tails of the Gaussian noise distribution decay very fast, this makes the thresholding step more efficient in preventing zero counts from appearing in the output, keeping the spurious rate to acceptably low levels. We quantify the privacy guarantees of the Gaussian mechanism using zero-concentrated DP [20] (zCDP). The third step is straightforward: all counts below a threshold \(\tau\) are removed from the output. This step is necessary because the first step produces many \(<\)page, country\(>\) tuples for which the non-noisy user count is very low or even 0. Such counts lead to unacceptable high relative error and spurious rate. Conversations with data users showed that these made the output dataset hard to use, and that users were most interested in the most-viewed pages, rather than the long tail of pages with few views. Suppressing counts below a fixed and configurable threshold \(\tau\) addresses this problem, at the cost of a non-zero drop rate. The mechanism is presented in Algorithm 2; in this algorithm, \(\mathcal{N}\left(0,\sigma^{2}\right)\) denotes a random sample from a normal distribution of mean \(0\) and variance \(\sigma^{2}\). Step 1 uses only public data, Step 2 provides \(\rho\)-zCDP [20], and Step 3 is a post-processing step: the full algorithm satisfies \(\rho\)-zCDP. ``` 1:\(t\): an ingestion threshold. 2:\(\tau\): a suppression threshold. 3:\(\rho\): a privacy parameter for zCDP. 4:\(P=\left\langle p_{1},b_{1}\right\rangle,\left\langle p_{2},b_{2}\right\rangle,\dots\): a private dataset of annotated pageviews, such each user is at most associated with \(k\) unique pageviews \(\left\langle p_{i},b_{i}\right\rangle\) where \(b_{i}=\texttt{true}\), and all of them have distinct \(p_{i}\). 5:\(P_{daily}=\left\langle p_{1},n_{1}\right\rangle,\left\langle p_{2},n_{2}\right\rangle,\dots\): a public dataset listing the global number of pageviews for each page. 6:\(C\): a pre-defined list of countries. Step 1: Collecting aggregation groups 7:\(G\leftarrow\{\}\) 8:for\(\left\langle p,n\right\rangle\) in \(P_{daily}\)do 9:if\(n\geq t\)then 10:for\(c\) in \(C\)do 11:\(G\gets G\cup\left\langle p,c\right\rangle\) 12:endfor 13:endif 14:endfor Step 2: Computing noisy counts 15:\(\sigma\leftarrow\sqrt{\frac{k}{2\rho}}\) 16:\(O\leftarrow\{\}\) 17:for\(g\) in \(G\)do 18:\(c\leftarrow|\{p\in P\mid p=g\}|\) 19:\(\hat{c}\gets c+\mathcal{N}\left(0,\sigma^{2}\right)\) 20:\(O\gets O\cup\left\langle g,\hat{c}\right\rangle\) 21:endfor Step 3: Suppressing low counts 22:for\(\left\langle g,\hat{c}\right\rangle\) in \(G\)do 23:if\(\hat{c}<\tau\)then 24:\(O\gets O\setminus\left\langle g,\hat{c}\right\rangle\) 25:endif 26:endfor 27:return\(O\) ``` **Algorithm 2** Server-side algorithm for the current pageviews We use \(k=10\) as a per-user daily contribution bound, \(t=150\) as an ingestion threshold, and \(\tau=90\) as a suppression threshold. These values were chosen after extensive experimentation, for input dataset completeness and to optimize the utility metrics described in Section 3.4. To select these algorithmic parameters, we computed metrics using the true data. Such metrics are, in principle, sensitive, and the parameters themselves are not differentially private. To mitigate the privacy risk from this tuning process, we kept fine-grained utility metrics confidential throughout the tuning process, minimizing data leakage. In addition to this consideration, we only publicly communicate approximate values of global utility metrics and the algorithmic parameters obtained from this tuning process. Regardless, this remains a valid critique, and we would appreciate further research into the privacy loss entailed by confidentially tuning on sensitive metrics. ### Historical pageviews To compute differentially private counts using the historical pageview dataset as input data, we follow a similar process, with one key difference: since the data is pre-aggregated, is is impossible to perform per-user contribution bounding. Therefore, we do not use a client-side filtering step, and instead, use a different unit of privacy, as described in Section 3.3. We also have to sum the Count column of the pre-aggregated data, rather than simply counting the number of rows in each group. Another difference is the use of Laplace noise instead of Gaussian noise, motivated by the fact that we only have a bound on the \(L_{1}\) sensitivity of the aggregation, and not \(L_{2}\) like with the current pageviews data. The full process is otherwise similar to the previous one. 1. First, we collect the list of \(<\)page, country\(>\) tuples to aggregate over. 2. Second, we sum the pageview counts in each group, and we add Laplace noise to each sum. 3. Finally, we suppress low sums, and publish the data. The full algorithm is provided as Algorithm 3; there, \(\text{Lap}(0,\lambda)\) denotes a random sample from the Laplace distribution of mean 0 and scale \(\lambda\). Its privacy analysis is straightforward: Step 1 uses only public data, Step 2 provides \(\varepsilon\)-DP guarantees [27], and Step 3 is a post-processing step, so the full algorithm satisfies \(\varepsilon\)-DP. As mentioned in Section 3.3, we use \(m=300\) for the 2015-2017 data, and \(m=30\) for the 2017-2023 data. For the 2015-2017 data, we use \(t=150\) as ingestion threshold and \(\tau=3500\) as suppression threshold. For the 2017-2023 data, we use \(t=150\) as ingestion threshold and \(\tau=450\) as suppression threshold. These values were chosen to optimize the global utility metrics described in Section 3.4. ### Implementation The algorithms were implemented and deployed using Tumult Analytics [30, 18], a framework chosen for its robustness, production-readiness, compatibility with Wikimedia's compute infrastructure, and support for advanced features like zCDP-based privacy accounting [5]. This incurs very slight differences in the mechanisms used: on integer-valued data, Tumult Analytics uses a two-sided geometric distribution instead of Laplace noise, and a discrete version of the Gaussian mechanism [21]. The data release based on the current input data required implementing a new notion of neighboring relation in the framework: rather than protecting a fixed number of rows, or an arbitrary number of rows associated with a single user identifier, it protects a fixed number of rows _associated with different aggregation groups_. This was made easier by the extensibility of the underlying framework, Tumult Core. ## 5 Outcomes The deployment of this differentially private data publication project is now allowing the Wikimedia Foundation to release a much larger and much richer dataset about user visits to Wikimedia projects. The magnitude of this increase in published pageview data is summarized in Table 6. More than 2,000 days of historical data from 2015 to 2021 were not previously published. The use of differential privacy in this project allowed the Wikimedia Foundation to release more than 135 million statistics about this data, encompassing 325 billion pageviews. The output data had acceptable quality according to our success metrics. * For the data based on the current pageviews dataset, more than 95% of the counts has a relative error below 50%, the drop rate above 150 is below 0.1%, the global spurious rate is below 0.01%, and below 3% for all but 3 countries. * For the 2017-2023 data, the median top-1000 drop rate is below 8%, the drop rate above 450 is below 3%, and the global spurious rate is below 0.1%. ``` 0:\(m\): the number of pageviews protected each day. 0:\(t\): an ingestion threshold. 0:\(\tau\): a suppression threshold. 0:\(\varepsilon\): a privacy parameter for DP. 0:\(P_{hourly}=\left\langle p_{1},c_{1}\right\rangle,\left\langle p_{2},c_{2} \right\rangle,\ldots\): a private dataset listing pre-aggregated hourly pageview counts. 0:\(P_{daily}=\left\langle p_{1},n_{1}\right\rangle,\left\langle p_{2},n_{2} \right\rangle,\ldots\): a public dataset listing the global number of pageviews for each page. 0:\(C\): a pre-defined list of countries. 0:\(t\): A minimum pageview threshold for including pages in the output. 0:\(\mathit{Step~{}1}\): Collecting aggregation groups 1:\(G\leftarrow\{\}\) 2:for\(\left\langle p,n\right\rangle\) in \(P_{daily}\)do 3:if\(n\geq t\)then 4:for\(c\) in \(C\)do 5:\(G\gets G\cup\left\langle p,c\right\rangle\) 6:endfor 7:endif 8:endfor 9:\(\mathit{Step~{}2}\): Computing noisy sums 10:\(\lambda\leftarrow\frac{m}{\varepsilon}\) 11:\(O\leftarrow\{\}\) 12:for\(g\) in \(G\)do 13:\(s\leftarrow\sum_{\left\langle p,c\right\rangle\in P_{hourly}\mathrm{where}p=g}c\) 14:\(\hat{s}\gets s+\mathrm{Lap}\left(0,\lambda\right)\) 15:\(O\gets O\cup\left\langle g,\hat{s}\right\rangle\) 16:endfor 17:\(\mathit{Step~{}3}\): Suppressing low counts 18:for\(\left\langle g,\hat{s}\right\rangle\) in \(G\)do 19:if\(\hat{s}<\tau\)then 20:\(O\gets O\setminus\left\langle g,\hat{s}\right\rangle\) 21:endif 22:endfor 23:return\(O\) ``` **Algorithm 3** Algorithm for the historical pageviews * For the 2015-2017 data, the top-1000 drop rate is below 40%, the drop rate above 3500 is below 3%, and the global spurious rate is below 20%. These metrics show that the privacy-accuracy trade-offs are much better for recent data than for historical data: this is explained by the much tighter sensitivity bound from client-side filtering, allowing to take full advantage of the Gaussian mechanism and its fast-decaying tails. ## 6 Conclusion In this paper, we described the process and mechanisms that allowed the Wikimedia Foundation to publish large-scale datasets about user behavior on Wikipedia and other Wikimedia projects. Multiple key factors made this launch possible. * Tumult Labs' systematic workflow for differential privacy publications, described in Section 2, provided the structure necessary to move the project forward from its inception to its deployment. * Combining client-side filtering with server-side aggregation, as described in Section 4.1, was a key innovation that allowed us to obtain user-level differential privacy guarantees for the current pageview data without tracking user identifiers. * Tumult Core, the privacy framework underlying Tumult Analytics, is designed for extensibility. This made it possible for us to add a novel neighboring definition to this framework to capture the properties of client-side filtering, while still being able to use tight privacy accounting techniques. * Finally, the scalability offered by Tumult Analytics was essential in handling the massive datasets that were used as input in this project. The data is now published online [9, 10, 11], along with the source code of the client-side filtering infrastructure [2] and the server-side algorithms [4, 8]. We look forward to seeing what use cases this data will enable! ## 7 Acknowledgements We are grateful to Luke Hartman, Tomoko Kitazawa, Nuria Ruiz, and Xabriel J. Collazo Mojica for their help with this project, and to Leila Zia and the anonymous reviewers for their helpful comments and suggestions on this paper. \begin{table} \begin{tabular}{c|c|c|c} & Before this project & After this project & Percentage change \\ \hline Median number of data points & 9,000 & 360,000 & +4,000\% \\ released per day & 50 million & 120 million & +240\% \\ \hline Total number of data points & \multirow{2}{*}{8 million} & \multirow{2}{*}{120 million} & \multirow{2}{*}{+1,500\%} \\ released since 2021 & & & \\ \hline Total number of pageviews & \multirow{2}{*}{47 billion} & \multirow{2}{*}{116 billion} & \multirow{2}{*}{+250\%} \\ released since 2021 & & & \\ \end{tabular} \end{table} Table 6: A comparison of the amount of data published before and after this project, as of June 29, 2023.

2306.10345

Do as I can, not as I get

This paper proposes a model called TMR to mine valuable information from simulated data environments. We intend to complete the submission of this paper.

Shangfei Zheng, Hongzhi Yin, Tong Chen, Quoc Viet Hung Nguyen, Wei Chen, Lei Zhao

2023-06-17T13:23:22Z

http://arxiv.org/abs/2306.10345v2

# Do as I can, not as I get: ###### Abstract Multi-modal knowledge graph (MKG) includes triplets that consist of entities and relations and multi-modal auxiliary data. In recent years, multi-hop multi-modal knowledge graph reasoning (MMKGR) based on reinforcement learning (RL) has received extensive attention because it addresses the intrinsic incompleteness of MKG in an interpretable manner. However, its performance is limited by empirically designed rewards and sparse relations. In addition, this method has been designed for the transductive setting where test entities have been seen during training, and it works poorly in the inductive setting where test entities do not appear in the training set. To overcome these issues, we propose **TMR** (Topology-aware **M**ulti-hop **R**easoning), which can conduct MKG reasoning under inductive and transductive settings. Specifically, TMR mainly consists of two components. (1) The topology-aware inductive representation captures information from the directed relations of unseen entities, and aggregates query-related topology features in an attentive manner to generate the fine-grained entity-independent features. (2) After completing multi-modal feature fusion, the relation-augment adaptive RL conducts multi-hop reasoning by eliminating manual rewards and dynamically adding actions. Finally, we construct new MKG datasets with different scales for inductive reasoning evaluation. Experimental results demonstrate that TMP outperforms state-of-the-art MKGR methods under both inductive and transductive settings. Multi-hop reasoning, multi-modal knowledge graphs, inductive setting, adaptive reinforcement learning ## 1 Introduction Knowledge graphs (KGs) store and manage huge amounts of data in reality and have been widely used in applications, including recommendation systems [43], information retrieval [28], and knowledge question answering [16]. A traditional KG consists of structural triplets that involve entities and relations, such as (_James Cameron_, \(Role\_create\), \(RoseBuckater\)). In recent years, as multi-modal data has received widespread attention in the field of data science and artificial intelligence, multi-modal knowledge graphs (MKGs) have emerged [27, 45]. As shown in Figure 1(a), an MKG contains extra multi-modal auxiliary data (images and text description) based on structural triplets, which provides diverse modalities of knowledge. However, the intrinsic incompleteness of MKGs severely limits knowledge applications [34]. To address this problem, the multi-modal knowledge graph reasoning (MKGR) technique is proposed to infer missing triplets in MKGs [53]. For instance, given a triple query (_James Cameron_, _Writer_,?), MKGR can utilize both structural and multi-modal auxiliary data to infer the missing entity _Titan_. In the literature, existing MKGR methods can be categorized into two types: single-hop reasoning and multi-hop reasoning [53]. The former focuses on modeling score functions for one-step relations that contain relatively less information [34, 45], while the latter represents the latest work that interpretably infers missing elements by combining multi-hop relations and fusing the corresponding multi-modal features [53]. As shown in Figure 1(b), by connecting (_James Cameron_, _Role\_create_, _Jack Dawson_) and (_Jack Dawson_, _Hero_, _Titanic_), a missing triplet (_James Cameron_, _Writer_, _Titanic_) can be inferred. MMKGR [53] stands out as the unique multi-hop MKGR model in existing ones, garnering significant attention for its state-of-the-art (SOTA) performance and interpretability. To effectively utilize both structural features and corresponding multi-modal features, MMKGR first uses a unified gate-attention network to generate multi-modal complementary features with sufficient attention interactions and less noise. Then, these features are fed into a novel complementary feature-aware reinforcement learning (RL) framework. This framework selects a sequence of actions (i.e., multi-hop reasoning paths) to accumulate rewards on the basis of a manually designed 3D reward function. Finally, MMKGR aims to maximize reward values by successfully inferring missing entities and outputs interpretable reasoning paths. Although MMKGR demonstrates impressive reasoning performance and interpretability on MKGs, there is still scope for enhancing its action and reward design. (1) _3D reward function is limited by manual design_. It comprises three manual sub-rewards, relying on the experience of domain experts and existing data distributions [12]. However, this necessitates time-consuming redesign when adapting to new environments [2]. Moreover, the subjective nature of manual reward design can lead to variations among different designers [22]. (2) _MMKGR is sensitive to the sparsity of relations_. The selection of actions in MMKGR relies on the combination of multi-hop relations. The absence of any relation in this path causes the reasoning path to be unavailable, which limits the reasoning performance [50, 52]. For example, MMKGR infers the missing entity _Titanic_ through a two-hop reasoning path _James Cameron \(\overset{Role\_create}{\rightarrow}\) Jack Dawson \(\overset{Herg}{\rightarrow}\) Titanic_. If \(Role\_create\) or \(\overset{Herg}{\rightarrow}\) is unconnected, the aforementioned two-hop path does not exist, which results in the query (_James Cameron_, _Director_,?) cannot be inferred. Arguably, it is extremely challenging to design an adaptive reward without manual intervention and dynamically add actions to alleviate sparsity. More importantly, MMKGR is difficult to apply to real scenarios since it primarily concentrates on the transductive setting while overlooking the importance of the inductive setting. As shown in Figure 1 (b), all entities are assumed to be seen during testing in the transductive setting [37]. However, knowledge is evolving and new entities are constantly emerging in the real world [29]. This observation is more in line with the inductive setting where the inferred entities in the test set do not appear in the training set [41]. These inferred entities, often referred to as unseen entities, lack knowable structural representations under the inductive setting. Intuitively, MMKGR can leverage multi-modal auxiliary data of unseen entities to infer the missing triples associated with them. This naturally raises an intriguing and fundamental question: How does MMKGR perform under the inductive setting? To answer this question, we first construct datasets of the induction setting where the entities in the test set and the train set are disjoint [33]. Then, we apply MMKGR to conduct multi-hop reasoning under the inductive setting. Experimental results reveal that MMKGR struggles to converge and has low reasoning performance under the inductive setting. Actually, the conclusion of inductive reasoning methods [33, 47] on traditional KGs is consistent with the above experimental findings: a transductive reasoning method that relies only on multi-modal auxiliary data and lacks generalizability to unseen entities is unsuitable for the inductive setting [11, 25, 41]. This prompts a subsequent goal: _How to develop the inductive capability of MMKGR to generalize unseen entities under an inductive setting?_ A technical challenge to achieving this goal lies in the lack of fine-grained entity-independent representations in existing MKGR methods. One of the key advantages of learning this representation is the development of the inductive capability to generalize unseen entities even in the absence of their specific structural features. [25, 37]. MMKGR lacking the inductive capability has no choice but to use multi-modal auxiliary features of unseen entities to understand these entities, which is highly dependent on the quality and quantity of multi-modal data and not suitable for unseen tasks [11, 41]. Additionally, existing entity-independent representation methods of inductive reasoning on traditional knowledge graph reasoning cannot be directly extended in MMKGR. This is because these methods struggle to aggregate the most relevant information based on specific query relations, resulting in the generation of coarse-grained representations for unseen entities. To make matters worse, the coarse-grained representation of each entity in the reasoning path iteratively disrupts decision-making abilities, which impairs the reasoning performance of MMKGR. Consequently, the fine-grained entity-independent representation is crucial to develop an inductive capability for MMKGR. In light of the aforementioned challenges in MMKGR, we propose an extended method entitled **TRM (**Topology-aware **M**ulti-hop **R**easoning). The main difference between our method and existing ones is that TRM has not only a talent for exploiting multi-modal data, but also the inductive capability to generalize unseen entities. Thus, TRM is capable of conducting MKGR under both inductive and transductive settings. Specifically, TRM mainly contains _topology-aware inductive representation_ (TAIR) and _relation-augment adaptive reinforcement learning_ (RARL). To develop the inductive capability for MMKGR, **TAIR** learns fine-grained entity-independent representation from query-related topology knowledge. Its relation-aware entity initializer captures a coarse-grained entity-independent representation by leveraging type information of unseen entities from the connected directed relations. To further generate the fine-grained representation, an adaptive topology representation module introduces a query-aware graph neural network (GNN) to attentively capture the topological information. After completing multi-modal feature fusion, **RARL** infers the missing elements by multi-hop reasoning path on MKGs, aiming to further improve the comple Fig. 1: MKGR under transductive and inductive settings. In the transductive setting, test entities have been seen in the training graph. In contrast, testing entities such as _Interstellar_ and _Christopher Nolan_ did not appear in the training graph under the inductive setting. mentary feature-aware reinforcement learning framework of MMKGR. Technically, RARL not only dynamically adds relations as additional actions to eliminate relational sparsity but also adaptively generates rewards by imitating expert demonstrations with filtering low-contributing paths. In summary, as an extension of our conference paper [53], this work makes the following contributions: * To the best of our knowledge, this is the first work to investigate _how to conduct MKGR under both inductive and transductive settings_. * To resolve the above problem, we propose an RL-based MKGR model called TMR that mainly contains two components TAIR and RARL. Specifically, TAIR generates fine-grained entity-independent representations to generalize unseen entities. RARL conducts multi-hop reasoning by expanding action space and utilizing imitation learning to eliminate manually designed rewards. * We construct MKG datasets under the inductive setting. To simulate unseen entities, we ensure that the entities in the test set and training set are disjoint. * Extensive experiments are conducted under both the transductive and inductive settings. Experimental results demonstrate the superior performance of TMR surpasses MMKGR and various baselines. The remaining sections are organized as follows. Preliminaries and definitions are presented in Section 2, followed by the overview of TMR. For different components in our proposed model, we introduce them in Sections 4, 5, and 6, respectively. Extensive experiments are shown in Section 7. Section 8 provides a review of the related literature. Finally, we conclude this work in Section 9. ## 2 Preliminaries and Definitions A MKG is an extension of KG by adding multi-modal auxiliary data, it is denoted as \(\mathcal{G}_{m}=\{\mathcal{E}_{m},\mathcal{R},\mathcal{U}\}\), where \(\mathcal{R}\) is a set of semantic relations, and \(\mathcal{E}_{m}\) denotes a set of entities associated with related multi-modal auxiliary data. The features of an entity \(i\) are denoted \(\boldsymbol{f}_{i}=\boldsymbol{f}_{s}\circ\boldsymbol{f}_{m}\), where "\(\circ\)" represents a multi-modal fusion method, \(\boldsymbol{f}_{s}\) and \(\boldsymbol{f}_{m}\) denote structural features and multi-modal auxiliary features, respectively. \(\mathcal{U}=\{(e_{s},r,e_{d})\mid e_{s},e_{d}\in\mathcal{E}_{m},r\in\mathcal{ R}\}\) is a set of triplets, where \(e_{s}\), \(e_{d}\), and \(r\) denote a head entity, a tail entity, and the relation between these entities, respectively. MKGR typically refers to the link prediction task of the inferring triple query (\(e_{s}\), \(r_{q}\),?) and (?, \(r_{q}\), \(e_{d}\)), where \(r_{q}\) is a query relation. By adding inverse relation, each triplet (\(e_{s}\), \(r\), \(e_{d}\)) is equivalent to the triplet (\(e_{d}\), \(r^{-1}\), \(e_{s}\)). Without loss of generality, MKGR methods can predict missing head entities by converting (?, \(r_{q}\), \(e_{d}\)) to (\(e_{s}\),?,?). **Definition 1**.: _MKGR under the transductive setting_. Given a MKG \(\mathcal{G}_{m}=\{\mathcal{E}_{m},\mathcal{R},\mathcal{U}\}\), MKGR under transductive setting aims to reason out a set of triple queries \(\mathcal{Q}\), \(\mathcal{Q}=\{(e_{s},r_{q},?)\mid(e_{s},r_{q},?)\notin\mathcal{U}\), \(e_{s}\), "\(\gamma^{\prime\prime}\in\mathcal{E}_{m}\), \(r_{q}\in\mathcal{R}\}\), where "?" is a missing entity, \(\mathcal{E}_{m}\in\mathcal{G}_{m}\) and \(\mathcal{R}\in\mathcal{G}_{m}\) represent entity and relation have been seen in the existing MKG \(\mathcal{G}_{m}\). **Definition 2**.: _MKGR under the inductive setting_. Given two disconnected MKGs \(\mathcal{G}_{m}=\{\mathcal{E}_{m},\mathcal{R},\mathcal{U}\}\) and \(\mathcal{G}_{m}^{*}=\{\mathcal{G}_{m}^{*}\) = \(\{\mathcal{E}_{m}^{*},\mathcal{R}^{*},\mathcal{U}^{*}\}\), \(\mathcal{G}_{m}\) is often known as a training graph, while \(\mathcal{G}_{m}^{*}\) is considered as a testing graph composed of the triplets by the emerging entities \(\mathcal{E}_{m}^{*}\) and the relations \(\mathcal{R}^{*}\). MKGR under the inductive setting requires the model to learn inductive capability in the training graph to infer a set of queries \(\mathcal{Q}\) on the test graph, \(\mathcal{Q}=\{(e_{s},r_{q},\)?) \((e_{s},r_{q},?)\notin\{\mathcal{U}\cup\mathcal{U}^{*}\}\), \(e_{s}\), "?" \(\in\mathcal{E}_{m}^{*},r_{q}\in\mathcal{R}^{*}\}\), where \(\mathcal{E}_{m}\cap\mathcal{E}_{m}^{*}=\emptyset\) and \(\mathcal{R}\cup\mathcal{R}^{*}=\mathcal{R}\). **Definition 3**.: _Multi-hop reasoning._ Multi-hop reasoning infers the missing element by the relational path shorter or equal \(L\) hops, where \(L\) is an integer not less than \(1\). A reasoning path is denoted as \(P\), which is obtained by summing all relations and entities in this path. ## 3 Overview of Trm MMKGR, the version of our conference [53], is limited by manually designed reward functions and relation sparsity as well as poor performance under inductive settings, which motivates us to propose TMR in this paper. As shown in Figure 2, TMR mainly contains two components: **TAIR** and **RARL**. Specifically, inspired by the human inductive ability to generalize unseen tasks from existing relevant knowledge [19], **TAIR** generates fine-grained entity-independent features from the existing topological structure in an attentive manner to represent unseen entities. After employing the unified gate-attention network in MMKGR to complete multi-modal feature fusion, **RARL** conducts MKGR by dynamically adding actions and automatically generating rewards, which is mainly inspired by the fact that humans learn optimal policies by imitating demonstrations rather than predefined paradigms [18]. Notably, TMR is qualified for conducting MKGR under inductive and transductive settings. This is because TMR decouples representation and reasoning into independent components to ensure the flexibility of reasoning under different settings. When inductive settings are converted to transductive setting, TMR only needs to add additional structural representations of seen entities into the unified gate-attention network, while the reasoning module RARL continues to complete reasoning as a multi-modal perception interface without further changes. ## 4 Topology-aware Inductive Representation Existing methods are powerless to capture fine-grained entity-independent representation, thereby restricting the inductive capability of MKGR models [11, 25, 41]. To address the problem, we propose a novel representation method called TAIR in this section. Notably, the technical difference from the existing method for representing unseen entities lies in the following two points: (1) Taking full advantage of type information derived from the connected directed relations of unseen entities. (2) Aggregating query-related neighbor relations in an attentive manner. Specifically, TAIR includes two modules, i.e., a relation-aware entity initializer and an adaptive topology representation. The former obtains coarse-grained representations of the unseen entity, and the latter aggregates topology information related to the query to generate ine-grained entity-independent representations. ### _Relation-aware Entity Initializer_ In general, entities with similar semantics have similar topological structures in MKGs, which are reflected in the connection patterns of their incoming and outgoing relations. By analyzing the connection patterns, we can obtain a coarse-grained representation that includes type information for unseen entities. For example, unseen entities \(James\_Cameron\) and \(Christopher\_Nolan\) both contain the outgoing edge \(Role\_create\), and these two entities have the same type-level representation, i.e., \(art\)\(creator\). For an unseen entity \(e_{i}\), its initialized embedding \(\textbf{h}_{i}^{0}\in\mathbb{R}^{d}\) as follows: \[\textbf{h}_{i}^{0}=\frac{\sum_{r\in I(i)}\textbf{W}_{i}\textbf{u}_{r}+\sum_{r \in O(i)}\textbf{W}_{o}\textbf{u}_{r}}{|I(i)|+|O(i)|} \tag{1}\] where \(\textbf{W}_{i}\), \(\textbf{W}_{o}\in\mathbb{R}^{d\times d}\) are transformation matrices. \(I(i)\) and \(O(i)\) represent the set of incoming and outgoing relations of the entity \(e_{i}\), respectively. \(\textbf{u}_{r}\in\mathbb{R}^{d}\) is embedding of the relation \(r\). Considering that the semantics of the same entity can be diverse under different query relations [39], we utilize the attention mechanism to filter out irrelevant neighbor relations. For example, the relations connected by unseen entity _Stephen Curry_ have different types, such as family and vocational relations. Given a triple query (_Stephen Curry_, _Father_, _?_), the vocational relations connected with the unseen entity indicate that _Stephen Curry_ is a professional basketball player, but this information is irrelevant to the query. Therefore, an attention mechanism that dynamically adjusts weights is employed to more accurately represent unseen entities in query tasks. The calculation process is as follows. \[\alpha_{r}=softmax(\textbf{u}_{r},\textbf{u}_{r_{q}})=\frac{\exp(\textbf{u}_{ r}^{\top}\textbf{u}_{r_{q}})}{\sum_{r^{\prime}\in\mathcal{N}^{\prime}(i)} \exp(\textbf{u}_{r^{\prime}}^{\top}\textbf{u}_{r_{q}})} \tag{2}\] where \(\textbf{u}_{r}\) and \(\textbf{u}_{r_{q}}\) are relation representations of neighbor relation \(r\) and query relation \(r_{q}\). \(\alpha_{r}\) denotes the correlation between \(r\) and \(r_{q}\). After integrating \(\alpha_{r}\), Eq. (1) is updated as, \[\textbf{h}_{i}^{0}=\frac{\sum_{r\in I(i)}\alpha_{r}\textbf{W}_{i}\textbf{u}_{ r}+\sum_{r\in O(i)}\alpha_{r}\textbf{W}_{o}\textbf{u}_{r}}{|I(i)|+|O(i)|} \tag{3}\] ### _Adaptive Topology Representation_ After obtaining the coarse-grained representation \(\textbf{h}^{0}\) of the unseen entity \(e_{i}\) by the initializer, we further capture the fine-grained semantic information from the topology of unseen entities. Inspired by the ability of GNNs to capture topology information in knowledge graphs [51], an adaptive topology representation module leverages GNNs to aggregate local structural information from multi-hop neighbors of entity \(e_{i}\). Specifically, we first concatenate the entities and their relations to obtain triplet information. Compared with individual entities or relations, triple information can provide sufficient topological information [51]. Then, we compute the correlation between the query relation \(r_{q}\) and these triplets that contain more contextual information, which effectively captures the fine-grained representation between topology and \(r_{q}\)[13, 54]. Next, we define updating process of the unseen entity \(e_{i}\) in a \(k\)-th layer as follows. \[\textbf{h}_{i}^{k}=tanh(\textbf{W}_{self}^{k-1}\textbf{h}_{i}^{k-1 }+\sum_{(i^{\prime},r)\in\mathcal{N}_{i}(e_{i})}\alpha_{i,r}\textbf{W}_{in}^{k-1 }(\textbf{h}_{i}^{k-1}\circ\textbf{u}_{r}^{k-1})\] \[+\sum_{(r,i^{\prime})\in\mathcal{N}_{o}(e_{i})}\alpha_{i,r} \textbf{W}_{out}^{k-1}(\textbf{h}_{i}^{k-1}\circ\textbf{u}_{r}^{k-1})) \tag{4}\] \[\alpha_{i,r}=\sigma(\textbf{W}_{2}\textbf{c}_{i,r}+\textbf{b}) \tag{5}\] \[\textbf{c}_{i,r}=\sigma(\textbf{W}_{1}[\textbf{h}_{i}^{k-1}\oplus\textbf{h}_{ j}^{k-1}\oplus\textbf{u}_{r}^{k-1}\oplus\textbf{u}_{r}^{k-1}]) \tag{6}\] where \(\mathcal{N}_{i}\) and \(\mathcal{N}_{o}\) are the incoming and outgoing neighbors of entity \(e_{i}\), respectively. \(\textbf{W}_{self}^{k-1}\), \(\textbf{W}_{in}^{k-1}\) and \(\textbf{W}_{out}^{k-1}\in\mathbb{R}^{d\times d}\) denote the transformation matrices, respectively. \(\circ\) is the element-wise product, and \(\sigma\) is the activation function \(sigmoid\). \(\alpha_{i,r}\) is the attention weight of the triplet (\(e_{i}\), \(r\), \(e_{j}\)). Based on this, we obtain the fine-grained entity-independent representation \(\textbf{h}_{i}^{k}\) of the unseen entity \(e_{i}\). Fig. 2: TAIR first exploits query-related topological information to obtain fine-grained entity-independent features. Then, these features and multimodal auxiliary features are fed into the UGAN to generate multi-modal complementary features \(Z\). Next, the RL-based reasoner utilizes \(Z\) and the augmented actions to generate reasoning paths. The discriminator compares reasoning paths and demonstrations to output adaptive rewards for the reasoner. Finally, the reasoner updates the reasoning policies and interacts with MKG to complete the prediction. Finally, to maintain the consistency of entities and relations within the embedding space, the embeddings of these relations are updated as follows: \[\mathbf{u}_{r}^{k}=\mathbf{W}_{r}^{k-1}\mathbf{u}_{r}^{k-1} \tag{7}\] where \(\mathbf{W}_{r}\in\mathbb{R}^{d}\) is a transformation matrix. ## 5 Unified Gate-attention Network In this section, we employ the unified gate-attention network (UGAN) in MMKGR to conduct feature fusion of fine-grained entity-independent representation and multi-modal auxiliary representation. Specifically, the unified gate-attention network includes an attention-fusion module, and an irrelevance-filtration module. After extracting multi-modal auxiliary features, the attention-fusion module fuses these features and context features together, by attending them with a carefully designed fine-grained attention scheme. Then, the irrelevance-filtration module discards irrelevant or even misleading information and generates noise robust multi-modal complementary features. Based on this, the unified gate-attention network selects features of different modalities online and simultaneously completes intra-modal and inter-modal attention interactions with noise robustness. ### _Feature Extraction_ (1) Context features: The entity \(e_{l}\) at reasoning step \(l\) and query relation \(r_{q}\) are represented as the fine-grained entity-independent embedding \(\mathbf{h}_{l}\) and \(\mathbf{u}_{r_{q}}\), respectively. In addition, the history of the reasoning path that consists of the visited entities and relations is defined as \(b_{l}\) = (\(e_{s}\), \(r_{0}\), \(e_{1}\), \(r_{1}\),...,\(e_{l}\)). We leverage LSTM to integrate the vector of history information \(\mathbf{h}_{l}\) with \(d_{s}\) dimensions into context features. Given the query in our multi-hop reasoning process, we obtain the context features \(\mathbf{y}\) = \([\mathbf{b}_{l};\mathbf{h}_{l};\mathbf{u}_{r_{q}}]\) by concatenating these features. Following [53], a group of context features \(Y\) is calculated as \(Y=[\mathbf{y}_{1},\mathbf{y}_{2},...,\mathbf{y}_{m}]\), where \(Y\in\mathbb{R}^{m\times d_{y}}\), \(m\) and \(d_{y}\) are the number of entities and the dimension of the features, respectively. (2) Multi-modal auxiliary features: To initialize image features \(\mathbf{f}_{i}\), we extract a \(d_{i}\)-dimensional vector of the last fully-connected layer before the softmax in VGG model [4]. Textual features \(\mathbf{f}_{i}\) are initialized by the word2vec framework [30] and expressed as a \(d_{t}\)-dimensional vector. We concatenate the above two parts of features on rows to form the multi-modal auxiliary features \(\mathbf{x}=[\mathbf{f}_{i}W_{i};\mathbf{f}_{i}W_{i}]\). To flexibly add multi-modal auxiliary features, a group of \(\mathbf{x}\) is denoted as \(X=[\mathbf{x}_{1},\mathbf{x}_{2},...,\mathbf{x}_{m}]\), where \(W_{t}\in\mathbb{R}^{d_{t}\times d_{x}/2}\), \(W_{i}\in\mathbb{R}^{d_{t}\times d_{x}/2}\), and \(X\in\mathbb{R}^{m\times d_{x}}\) represents a group of multi-modal auxiliary features, \(d_{x}\) is the dimension of the feature. ### _Attention-fusion Module_ To obtain the complementary features with sufficient interactions and less noise, we need to fuse the context features \(Y\) and multi-modal auxiliary features \(X\) generated in feature extraction. However, redundant features tend to have a negative impact on the prediction during the multi-modal fusion [49]. Specifically, redundant features are either shifted versions of the features related to the triple query or very similar with little or no variations, which can amplify the negative effects of noise [23]. The redundant features add computational complexity and cause collinearity problems [36]. Consequently, we propose the attention-fusion module that fuses the context features and multi-modal auxiliary features effectively. Specifically, we first utilize linear functions to generate the queries \(Q\), keys \(K\), and values \(V\) of the attention mechanism, \[Q=XW_{q},K=YW_{k},V=YW_{v} \tag{8}\] where \(W_{q}\in\mathbb{R}^{d_{x}\times d},W_{k},W_{v}\in\mathbb{R}^{d_{y}\times d}\), and \(Q,K,V\in\mathbb{R}^{m\times d}\) have the same shape. Then, the joint representation \(B^{l}\) of \(Q\) and \(K\) is learned based on the MLB pooling method [17], inspired by the recent successes of it in fine-grained multi-modal fusion, \[B^{l}=KW_{k}^{l}\odot QW_{q}^{l} \tag{9}\] Similarly, we can generate the joint representation \(B^{r}\) of \(V\) and \(Q\) with the following equation, \[B^{r}=VW_{v}^{r}\odot QW_{q}^{r} \tag{10}\] where \(W_{k}^{l},W_{q}^{l},W_{v}^{r},W_{q}^{r}\in\mathbb{R}^{d\times j}\) are embedding matrices, and \(\odot\) is Hadamard product. Next, the filtration gate \(g_{t}\) applied to different feature vectors is defined as, \[g_{t}=\sigma(B^{l}W_{m}) \tag{11}\] where \(W_{m}\in\mathbb{R}^{j\times d}\) is an embedding matrix and \(\sigma\) denotes the sigmoid activation. Based on the filtration gate \(g_{t}\), we can filter out the redundant features generated during fusion and obtain a new representation with the following probability distributions, \[G_{s}=softmax((g_{t}\odot K)((1-g_{t})\odot Q)) \tag{12}\] where \(g_{t}\) and \(1-g_{t}\) are used to trade off how many context features and multi-modal auxiliary features are fused. Finally, our attention-fusion module generates the attended features \(\hat{V}\)=\(\{\mathbf{v}_{i}\}_{i=1}^{m}\) by accumulating the enhanced bilinear values of context features and multi-modal auxiliary features, \[\hat{V}=\sum\nolimits_{i=1}^{m}(G_{s}W_{g}^{l})B_{i}^{r} \tag{13}\] where \(W_{g}^{l}\in\mathbb{R}^{d\times 1}\), and \(\mathbf{v}_{i}\in\mathbb{R}^{1\times j}\) denotes a row of the attended features \(\hat{V}\in\mathbb{R}^{m\times j}\), feature vector \(B_{i}^{r}\in\mathbb{R}^{1\times j}\) is a row of the embedding matrix \(B^{r}\). By designing the attention-fusion module, we can complete the intra-modal and inter-modal feature interactions in a unified manner at the same time. This is because the inputs of this module are pairs from context features and multi-modal auxiliary features, where each vector of a pair may be learned from the same modality or different ones. ### _Irrelevance-filtration Module_ We use an irrelevance-filtration module to further improve the robustness of the model. The attended features \(\hat{V}\) obtained by the attention-fusion module may contain irrelevant features [14]. Specifically, irrelevant features are irrelevant to the triple query in the reasoning process. Since the attention mechanism assigns weights to all features, these features tend to participate in model computation and mislead the reasoning policy [31]. This motivates our model to weight more on the most related complementary features and dynamically filter irrelevant ones. This is achieved by a well-designed irrelevance-filtration gate function. The output of this gate is a scalar, the value range of which is (0,1). The multi-modal complementary features \(Z\) are obtained as follows, \[G_{f}=\sigma(B^{r}\odot\hat{V}) \tag{14}\] \[Z=G_{f}(B^{r}\odot\hat{V}) \tag{15}\] where \(\sigma\) and \(G_{f}\) denote the sigmoid activation function and irrelevance-filtration gate, respectively. ## 6 Relation-augment Adaptive reinforcement learning The existing RL-based MKGR method is limited by manual rewards and sparse relations [1, 20]. To address this problem, we propose a novel RL-based framework entitled RARL in this section. Compared with MMKGR, the main technical difference of RARL lies in the following two points. (1) We effectively increase the additional actions to alleviate the negative impact of sparse relations on the RL-based model. (2) RARL utilizes generative adversarial imitating networks to adaptively learn rewards by imitating demonstrations, which can stabilize reasoning performance and eliminate manual intervention in reward design. This provides a new research perspective for RL-based reasoning methods for MKGR. RARL consists of three modules namely RL-based reasoner, rule-based demonstration sampler, and modality-aware discriminator. Specifically, the reasoner leverages a rule-guided action augmentation method that dynamically adds additional actions and outputs diverse generated paths about missing elements. Then, the rule-based demonstration sampler filters out low-contributing paths as well as extracts trustworthy demonstrations from MKGs. Next, the modality-aware discriminator generates rewards to update the reasoner by evaluating the semantic similarity between demonstrations and reasoning paths. After sufficient adversarial training, the RL-based reasoner tries to deceive the discriminator to gain more adaptive reward values by imitating the demonstrations. We introduce the above three modules in subsections 6.1, 6.2, and 6.3, respectively. ### _RL-based Reasoner_ #### 6.1.1 Reinforcement Learning Formulation RARL trains an agent to interact with the with MKGs by modeling by Markov decision process (MDP). The MDP consists of a 4-tuple, i.e., States, Actions, Transition, Rewards. The agent selects actions based on the current state and obtains rewards from the environment (MKGs) to update its behavior policy until it reaches a termination state or a predefined reasoning step. **States**: The state of the agent at reasoning step \(l\) is denoted as \(s_{l}\)= (\(e_{l}\), (\(e_{s}\), \(r_{q}\)) \(\in\)\(\mathcal{S}\), where \(\mathcal{S}\) denotes a state space and \(e_{l}\) represents the entity at the current reasoning step \(l\). The source entity \(e_{s}\) and the query relation \(r_{q}\) are the global context shared throughout all steps. **Actions**: For the given state \(s_{l}\), its original action space is the set of usable actions \(A_{l}^{o}\) at reasoning step \(l\) is expressed as \(A_{l}^{o}=\{(r_{l+1},\ e_{l+1})|\) (\(e_{l}\), \(r_{l+1}\), \(e_{l+1}\)) \(\in\)\(\mathcal{G}_{m}\}\). To alleviate relation sparsity, the rule-guided action augmentation module adds extra potential actions \(A_{l}^{o}\) into the original action space. Thus, the joint action space \(A_{l}^{o}\cup A_{l}^{a}=A_{l}\in\mathcal{A}\). In addition, we add \(STOP\) action to avoid infinitely unrolling in the reasoning process. The \(STOP\) action executes a self-loop when the reasoning step is unrolled to the maximum step \(L\). **Transition**: \(\mathcal{P}_{r}\) is defined to facilitate the transition from the current state \(s_{t}\) to the next state \(s_{l+1}\). \(\mathcal{P}_{r}\): \(\mathcal{S}\)\(\times\)\(\mathcal{A}\)\(\rightarrow\)\(\mathcal{S}\) is defined as \(\mathcal{P}_{r}\) (\(s_{l}\), \(A_{l}\)) = \(\mathcal{P}_{r}\) (\(e_{l}\), (\(e_{s}\), \(r_{q}\)), \(A^{o}\), \(A^{a}\)). **Rewards**: Different from existing manually-designed reward functions, we design an adaptive reward mechanism to eliminate manual intervention, which achieves high reasoning performance in complex and uncertain environments. The adaptive reward arises from path comparisons between the generator and expert demonstration, and it is defined in Eq. 29. **Policy Network** The policy function \(\pi\) is used as a multi-modal perception interface to output the next action with the highest executable probability. For a given state, \(\pi\) selects the promising action with the maximum likelihood, which is defined as, \[\pi_{\theta}(a_{l}|s_{l})=softmax(\textbf{A}_{l}(\textbf{W}\text{ReLu}(Z))) \tag{16}\] where \(a_{l}\)\(\in\)\(A_{l}\), and \(A_{l}\) can be encoded to \(\textbf{A}_{t}\) by stacking the representations of existing available actions. #### 6.1.2 Rule-guided Action Augmentation The existing RL-based reasoning method assumes sufficient relation paths between entities, and regards these relations and connected tail entities as the next action. However, the intrinsic incompleteness of MKGs leads to sparse relations of an entity. Especially, emerging entities are sparsely connected to existing entities under the inductive setting. This sparsity limits the utilization of potential reasoning paths. Therefore, it is necessary to design an action space augmentation method to eliminate the sparsity of the action space. Although the idea of augmenting the action space is promising, a major challenge is how to _officially augment additional actions_. Intuitively, enumerating all relations and entities to compose an additional action space can complement the existing action space. However, this combined search space is close to \(\mathcal{O}(|\mathcal{E}|\times|\mathcal{R}|)\), where \(|\mathcal{E}|\) and \(|\mathcal{R}|\) are the numbers of entities and relations in a MKG, respectively. For a large-scale MKG with millions of entities and thousands of relations, large search space becomes an intractable problem. To address the above problem, we propose a novel action augmentation method to efficiently augment additional actions. For a state \(s_{l}\), the candidate set of augmented action is denoted as \(C_{t}\) = \(\{(r^{\prime},e^{\prime})|r^{\prime}\in\mathcal{R}\wedge e^{\prime}\in \mathcal{E}_{m}\wedge(e_{l},r^{\prime},e^{\prime})\notin\mathcal{G}_{m}\}\). First, we calculate the probability for the candidate set \(C_{t}\): \[p((r^{\prime},e^{\prime})|s_{l})=p(r^{\prime}|s_{l})p(e^{\prime}|r^{\prime},s_{l}) \tag{17}\] To reduce the candidate action space and time-consuming, an approximate pruning strategy is used to filter out additional actions. The pruning strategy consists of additional relation selection using \(p(r^{\prime}|s_{l})\) and entity generation using \(p(e^{\prime}|r^{\prime},s_{l})\). Then, for the state \(s_{l}\), the attention score of the candidate relations is calculated as \(p(r|s_{l})\), \[\textbf{w}=softmax(MLP(\textbf{s}_{l})\cdot[\textbf{u}_{r_{1}},...,\textbf{u}_ {r_{|R|}}]) \tag{18}\] where **w** denotes the attention vector. We select top \(x\) relations with the largest attention values in **w** to obtain additional relation set \(\mathcal{R}_{add}=r^{1}\), \(r^{2}\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\\\\\\\\\_\ path features whose quality is higher than random noise into the reward. \[R_{r}=\max(D(\mathbf{\nu_{P}})-D^{N}(\mathbf{\nu_{P}}),0) \tag{27}\] \[R_{e}=\max(D(\mathbf{\kappa_{P}})-D^{N}(\mathbf{\kappa_{P}}),0) \tag{28}\] where \(D^{N}(\mathbf{\nu_{P}})\) and \(D^{N}(\mathbf{\kappa_{P}})\) respectively denote noise embeddings from the entity and relation layers. These embeddings consist of random noise sampled from a continuous uniform distribution. \(R_{r}\) and \(R_{e}\) represent the adaptive rewards from relation layer and entity layer, respectively. Then, we define the adaptive reward \(R(s_{l})\) at the state \(s_{l}\) as follows: \[R(s_{l})=\alpha R_{e}+(1-\alpha)R_{r} \tag{29}\] where \(\alpha\) is a balance factor. Note that, the agent learns adaptive rewards from demonstrations without manually designing and tuning, which reduces manual intervention and subjective bias [2]. In addition, the adaptive reward improves the generalizability of our proposed model and it is suitable for unseen tasks. This is because the adaptive reward mechanism automatically capturing common meta-knowledge by learning relation patterns and multi-modal features in demonstrations [32]. Next, we optimize the modality-aware discriminator by reducing training loss and expect it to possess expertise in distinguishing between \(P\) and \(\Omega\). \[\mathcal{L}_{r}=D(\mathbf{\nu_{P}})-D(\mathbf{\nu_{\Omega}})+\lambda(\parallel \bigtriangledown_{\hat{p}}D(\hat{p})\parallel_{2}-1)^{2} \tag{30}\] where \(\lambda\) is a penalty term and \(\hat{p}\) is sampled uniformly along straight lines between the generated path and the demonstration. For the discriminator at the entity level, we define loss \(\mathcal{L}_{r}\) as follows: \[\mathcal{L}_{e}=-(\log D(\mathbf{\kappa_{\Omega}})+\log(1-D(\mathbf{\kappa_{P}}))) \tag{31}\] Finally, to maximize the accumulated rewards of adaptive reinforcement learning and obtain the optimal policy, the objective function is as follows, \[\mathcal{J}(\theta)=\mathbb{E}_{(e_{s},r,e_{d})\sim\mathcal{G}_{m}}\mathbb{E} _{a_{1},\dots,a_{L}\sim\pi_{\theta}}[\mathbb{R}(s_{l}\mid e_{s},r)] \tag{32}\] ## 7 Experiment ### _Datasets_ Following MMKGR [34, 53], we use WN9-IMG-TXT and FB-IMG-TXT to verify the reasoning performance under the transductive setting. Each entity in these MKGs contains three modal information: structure, image, and text. Specifically, the relation triplets and textual descriptions are extracted from WordNet and Freebase. To extract the image features of the entities, 10 images and 100 images are crawled for each entity in WN9-IMG-TXT and FB-IMG-TXT, respectively [34]. Statistics are shown in Table I. To perform inductive reasoning in MKGs, we construct new inductive benchmark datasets by extracting disjoint subgraphs from WN9-IMG-TXT and FB-IMG-TXT. In particular, each dataset contains a pair of graphs: _training graph_ and _ind_test graph_. Following [37], to generate the training graph, we first uniformly sample several entities as the root nodes, and then conduct the union of k-hop neighbor triplets around the roots. Next, we set the maximum number of samples at each hop to prevent the exponential growth of new neighbors. Finally, we remove the training graph from the whole graph and sample the test graph using the same procedure. In fact, the above division method destroys the link distribution in the original graph and reduces the number of triplets. Therefore, for a robust evaluation, we adjust the parameters of the above procedure to sample 5%, 10%, and 15% of the original graph (i.e., WN9-IMG-TXT and FB-IMG-TXT) to construct datasets with different sizes [41]. In summary, the model is trained on the training graph and tested on the ind_test graph in every version. Note that, (1) the two graphs have disjoint sets of entities, (2) training graphs contain all relations present in ind_test graphs. ### _Evaluation Protocol_ To evaluate the reasoning performance of TMR over inductive and transductive settings, we adopt the mean reciprocal rank (MRR) and Hits@N to report experimental results, which are common metrics for MKGR [34, 53]. ### _Baselines and Implementation Details_ To investigate the performance of TMR, two categories of methods are compared. 1) Knowledge graph reasoning methods under the induction setting: DRUM [33], CoMPILE [29], Morse [6], RED-GNN [51]. 2) SOTA MKGR method: MMKGR [53]. Note that, the transductive reasoning methods (i.e., TransE [3] or RLPM [38]) on the traditional KG cannot be applied to MKGR. This is because these methods retrain the embedding from scratch whenever a new entity appears [41]. In our training stage, some core settings are as follows. The embedding dimension \(d_{s}\) of relation and history is set to 200, the embedding dimension \(d_{i}\) of the image feature is set to 128 and 4096 on FB-IMG-TXT and WN9-IMG-TXT respectively, and the embedding dimension \(d_{t}\) of textual feature is 1000. The maximum reasoning step \(L\) is 3. The number of additional relation \(I\) is 3 and 5 on different versions of WN9-IMG-TXT and FB-IMG-TXT, respectively. The size \(N\) is set to 5. We employ a 3-layer GNN to obtain the adaptive topology representation. \(\alpha\) in Eq.(29) is set to 0.4 and 0.2 on different versions of WN9-IMG-TXT and FB-IMG-TXT, respectively. \begin{table} \begin{tabular}{l l l l l l l} \hline Dataset & \#Ent & \#Rel & \#Train & \#Valid & \#Test \\ \hline WN9-IMG-TXT & 6,555 & 9 & 11,747 & 1,337 & 1,319 \\ FB-IMG-TXT & 11,757 & 1,231 & 285,850 & 29,580 & 34,863 \\ \hline \end{tabular} \end{table} TABLE I: Statistics of the experimental datasets over the transductive setting. \begin{table} \begin{tabular}{l l l l l l l l} \hline & \multicolumn{3}{c}{WN9-IMG-TXT} & \multicolumn{3}{c}{FB-IMG-TXT} \\ \hline & & \#Ent & \#Rel & \#Triplets & \#Ent & \#Rel & \#Triplets \\ \hline V1 & Training & 420 & 8 & 760 & 2848 & 342 & 17561 \\ V1 & Ind\_test & 270 & 8 & 401 & 2002 & 263 & 3325 \\ \hline V2 & Training & 654 & 9 & 1465 & 3205 & 631 & 35184 \\ V2 & Ind\_test & 406 & 8 & 814 & 2111 & 343 & 6064 \\ \hline V3 & Training & 658 & 9 & 2180 & 3716 & 750 & 52544 \\ V3 & Ind\_test & 581 & 8 & 1442 & 2419 & 254 & 10623 \\ \hline \end{tabular} \end{table} TABLE II: Statistics of the datasets over the inductive setting. ### _Inductive Reasoning on MKGs_ Reasoning performance over the inductive setting is reported in Table III (these scores are in percentage), where the competitive results of the baseline are marked by underlining and the highest performance results are highlighted in bold. Specifically, we have the following insightful analysis based on the experimental results in Table III. (1) The performance of our proposed TMR outperforms all baselines. This is because TMR not only generates fine-grained entity-independent representations for unseen entities, but also utilizes RARL to eliminate the negative impact of manual rewards and sparse relations on reasoning accuracy. (2) The experimental results of MMKGR are the lowest in the different datasets. This is because MMKGR without induction capability cannot obtain structured representations of unseen entities. Therefore, it is not a reasonable solution to only utilize multi-modal auxiliary information without learning inductive capabilities in the induction setting. (3) The rule-based SOTA model DRUM combines relations to infer unseen tasks, but this model is limited by the quality of the rules. (4) Similar to DRUM, CoMPILE, MorseE and RED-GNN are designed to infer unseen tasks in traditional knowledge graphs, but they adapt to conduct reasoning without utilizing multi-modal auxiliary features in the inductive setting. This is because the three models learn the inductive ability from the local graph structure information. In particular, the existing SOTA model RED-GNN remains competitive in all versions of MKGs. This is because RED-GNN recursively encodes multiple relational digraphs with shared edges and preserves the structural patterns at the same time [51]. In short, the key to performance improvement in the induction setting is to consider both induction capability and the ability to exploit multi-modal data. Note that, TMR is the first model to do this in the domain of MKGR. ### _Transductive Reasoning on MKGs_ In this section, we investigate whether TMR outperforms the SOTA model MMKGR under the transductive setting. To be consistent with MMKGR, we obtain the pre-trained embeddings of all entities by TransE. This is because unseen entities do not exist under the transductive setting. Specifically, the pre-trained representation \(\mathbf{e}_{i}\) of entity \(i\) is additionally added into the context features of TMR. Furthermore, a concern is that TMR is fed with pre-trained entity representations do not need entity-independent features generated by the TAIR component that contains topology information in the transductive setting. To eliminate the above concern, we added a variant of TMR-TAIR where the TAIR component is removed (i.e., the context features of TMR do not contain fine-grained entity-independent features). The experimental results under the transductive setting are shown in Table IV. We have the following observations. (1) The reasoning performance of TMR surpasses that of the SOTA MMKGR on the original MKGs. This demonstrates the flexibility and generalizability of TMR under different settings. (2) The performance of TMR-TAIR is higher than that of MMKGR. This is because the RARL component in TMA-TAIR can eliminate the negative impact of manual rewards and sparse relations on reasoning accuracy. (3) The performance of TMR-TAIR is lower than that of TMR. A reason is that the entity-independent features generated by the TAIR component aggregate multi-hop topology information. The information can provide reasoning context to improve performance under both inductive and transductive settings. ### _Ablation Study_ The overall goal of the ablation study is to measure the contribution of different components by adding different variants. Figure 3 reports the experimental results of the TMR and its variants. (1) The variant version w/o TAIR only uses multi-modal features of unseen entities where the TAIR component is removed from the TMR. (2) w/o UGAN, in which the unified gate-attention network is removed and basic concatenation operation is added to fuse all features. (3) w/o RARL, a variant version where RARL is removed. To ensure the agent conducts the reasoning, we retain the RL-based reasoner with the original action space and basic \begin{table} \begin{tabular}{c|c c c|c c c} \hline \hline & \multicolumn{3}{c|}{WN-IMG-TXT} & \multicolumn{3}{c}{FB-IMG-TXT} \\ Model & MRR & Hist@1 & Hist@10 & MRR & Hist@1 & Hist@10 \\ \hline MMKGR & 80.2 & 73.6 & 92.8 & 71.3 & 65.8 & 82.6 \\ TMR-TAIR & 83.6 & 76.4 & 93.2 & 74.3 & 67.5 & 85.4 \\ TMR & **86.3** & **79.7** & **93.7** & **76.6** & **71.4** & **87.6** \\ \hline \hline \end{tabular} \end{table} TABLE IV: Transductive link prediction results on different MKGs. \begin{table} \begin{tabular}{c|c c c c c c c c|c c c c c c c c c} \hline \hline & \multicolumn{8}{c|}{WN-IMG-TXT} & \multicolumn{8}{c}{FB-IMG-TXT} \\ & \multicolumn{3}{c|}{V1} & \multicolumn{3}{c|}{V2} & \multicolumn{3}{c|}{V3} & \multicolumn{3}{c}{V1} & \multicolumn{3}{c}{V2} & \multicolumn{3}{c}{V3} \\ & MRR & Hist@1 & Hist@10 & MRR & Hist@10 & MRR & Hist@1 & Hist@10 & MRR & Hist@1 & Hist@10 & MRR & Hist@1 & Hist@10 & MRR & Hist@1 & Hist@10 \\ \hline MMKGR & 27.0 & 25.1 & 30.2 & 27.2 & 25.5 & 30.6 & 29.3 & 26.9 & 33.3 & 21.4 & 19.3 & 25.3 & 23.3 & 21.7 & 26.5 & 26.0 & 23.9 & 29.1 \\ DRUM & 43.4 & 40.5 & 46.0 & 45.2 & 41.6 & 48.7 & 48.4 & 45.7 & 51.0 & 35.6 & 32.7 & 38.2 & 37.8 & 34.8 & 39.7 & 40.1 & 38.6 & 43.2 \\ CoMPILE & 45.2 & 41.7 & 47.3 & 47.0 & 43.9 & 50.8 & 49.1 & 47.6 & 53.2 & 36.9 & 34.5 & 39.3 & 39.8 & 35.9 & 40.9 & 42.3 & 39.9 & 44.4 \\ Morse & 48.2 & 45.8 & 52.4 & 50.3 & 48.6 & 53.1 & 54.2 & 52.1 & 56.7 & 38.3 & 36.6 & 41.2 & 40.7 & 38.3 & 43.1 & 44.2 & 42.1 & 46.2 \\ RED-GNN & 51.2 & 49.3 & 54.3 & 53.2 & 50.2 & 56.8 & 56.1 & 54.8 & 59.2 & **40.5** & 38.4 & **42.4** & 43.3 & 41.1 & 45.2 & 46.0 & **44.2** & 48.4 \\ TMR & **64.9** & **62.3** & **69.1** & **67.6** & **64.1** & **72.0** & **71.1** & **69.7** & **74.8** & **57.0** & **54.7** & **59.4** & **60.1** & **57.5** & **62.7** & **63.3** & **60.8** & **66.2** \\ \hline Improv. & 13.7\% & 13.0\% & 14.8\% & 14.3\% & 13.9\% & 15.2\% & 15.0\% & 14.9\% & 15.6\% & 16.5\% & 16.3\% & 17\% & 16.8\% & 16.4\% & 17.5\% & 17.3\% & 16.6\% & 17.8\% \\ \hline \hline \end{tabular} \end{table} TABLE III: Inductive link prediction results for different versions of MKGs. 0/1 reward (i.e., the reward value is set to 1 if the target entity is the ground truth entity. Otherwise, the value is 0). We have the following observations. (1) TMR has the best performance compared with both variant versions. This validates the effectiveness of all components. (2) The results of w/o TAIR are much lower than that of TMR, which verifies the importance of learning fine-grained entity-independent representation in the inductive setting. (3) Although the input of w/o UGAN includes multi-modal information, its reasoning performance still declines compared with TMR. This is because the concatenation operation cannot generate multi-modal complementary features. (4) After removing the rule-guided action augmentation method and adaptive reward mechanism, the performance of w/o RARA significantly degrades on different datasets. This demonstrates that RARL can eliminate the negative effects of sparse relations and manual rewards on reasoning performance. ### _Further Analysis_ #### 7.7.1 Convergence Analysis To analyze the convergence rate and reasoning performance between our adaptive reward and manual rewards in MMKGR, we design a variant entitled TMR-3D. Specifically, this variant removes the _Rule-based Demonstration Sampler_ and _Modality-aware Discriminator_ in RARL. To ensure that the agent is rewarded, we add the 3D reward mechanism (i.e., the manual reward function in MMKGR). In addition, we design variant models TMR-R and TMR-E by removing the relation-level reward \(R_{r}\) and the entity-level reward \(R_{e}\) in TMR, respectively. This setting can investigate the contribution of different parts of the adaptive reward. Observed from Figure 4, (1) TMR adopted the adaptive reward has the fastest convergence and the most stable performance on MKGs. This is because the adaptive reward learned from both path semantic and multi-modal feature automatically eliminate the manual intervention and avoids decision bias on different MKG datasets [21]. (2) Although TMR-3D can converge slowly, its performance is still unstable. A reason is that the weak generalizability of manually-designed 3D rewards leads to unstable training on different datasets [10]. (3) The performance of TMR-R is slightly worse than that of TMR-E, which indicates that \(R_{r}\) has more contribution than \(R_{e}\) in the RARL. #### 7.7.2 Effectiveness Analysis for Action Augmentation To investigate the effectiveness of the action augmentation method for expanding the latent action space, we design a new variant model TMR-AA by removing the rule-guided action augmentation method in RARL. The experimental results are shown in Figure 5, and we have the following analysis: (1) The performance of TMR has declined to varying degrees in all datasets after removing the action augmentation method, which verifies the effectiveness of the proposed action augmentation method. (2) The performance improvement of the rule-guided action augmentation method on different versions of FB-IMG-TXT is more obvious than that on the different versions of WN9-IMG-TXT. This is because more relations on different versions of FB-IMG-TXT can be used to build more complex reasoning rules. (3) Compared with adaptive rewards, the performance improvement of the rule-guided action augmentation method is relatively small. One potential reason is that the reasoning processes completely depend on original actions, and the additional actions from this method mainly play an auxiliary role. ### _Key Parameter Analysis_ The balance factor \(\alpha\) is an important parameter for our proposed model TMR. As presented in Figure 6, 0.4 and 0.2 are the optimal values of \(\alpha\) on the different versions of WN9-IMG-TXT and FB-IMG-TXT, respectively. The reward \(R_{e}\) from entity level is assigned small weights, which demonstrates that the semantic correctness of relational paths Fig. 4: The convergence rate of TMR and the variant model TMR-3D. Fig. 5: Performance comparison between TMR and TMR-AA. Fig. 3: Ablation on different components of the TMR. Fig. 6: Performance of TMR w.r.t. varied \(\alpha\) on different datasets. provides better reasoning clues than multi-modal entities in inductive reasoning tasks. ## 8 Related Work ### _Multi-modal Knowledge Graph_ A traditional KG is essentially a semantic graph that consists of entities (nodes) and relations (edges). At present, the actual internet data show multi-modal characteristics [15]. MKGs are developed to incorporate various types of data from different modalities into KGs [34]. A MKG typically includes structural triplets, and multi-modal data (i.e., texts and images) [26]. Common MKGs are IMGpedia [9], Ricpedia [40], and FB-Des [36]. However, the multi-modal auxiliary data of these MKGs is singular (i.e., image or text). To expand the auxiliary data with one modality, WN9-IMG-TXT and FB-IMG-TXT simultaneously add a number of textual descriptions and images to each entity, aiming to further enhance the data diversity of the MKGs [34, 48]. ### _Multi-modal Knowledge Graph Reasoning_ Since MKGs inherently contain incomplete knowledge, MKGR technology that can synthesize the original knowledge and infer the missing knowledge is particularly important [34, 35]. Some studies employ the attention model or concatenation to fuse multi-modal auxiliary features and then adopt TransE to infer missing elements, such as IKRL [44] and TransAE [42], and MTRL [34]. However, these methods lack interpretability and are primarily suitable for single-hop reasoning containing limited information. To address this issue, MMKGR leverages the symbolic compositionality of the multi-step relational path (choices of actions) to infer the correct entity [5, 53]. MMKGR has been proven to be a SOTA model in the field of TKGR. Its multi-hop reasoning process is as intuitive as "going for a walk", which naturally forms an explainable provenance for MMKGR. ### _Inductive Reasoning on Knowledge Graph_ The inductive setting is receiving increasing attention since unseen entities are emerging in KGs. Therefore, completing knowledge reasoning in an inductive setting is a practical application. Several methods are proposed to solve this problem. Rule-based methods can leverage the logical rules of existing knowledge to infer new facts, because the rules are independent of specific entities [33]. In addition, GraIL [37] and CoMPILE [29] aim to generalize to unseen entities and improve reasoning performance by subgraph extraction, but the enclosing subgraphs cannot learn relational structures so as to weaken the inductive capability. Inspired by the powerful graph modeling capabilities of, SOTA models like MorsE [6] and RED-GNN [51] utilize GNNs to aggregate topological structures and mine existing neighbor information, which is a promising method for inductive reasoning. However, these methods still have limitations: (1) They do not extract fine-grained entity-independent features related to the query, which restricts their inductive capacity. (2) Lack of ability to utilize multi-modal auxiliary information in MKGR. ## 9 Conclusion In this paper, we propose TMR as a solution to conduct TKGR in both inductive and transductive settings. Specifically, TMR mainly includes TAIR and RARL. TAIR learns fine-grained entity-independent representation from query-related topology knowledge to represent unseen entities. RARL eliminates the negative impact of sparse relations and artificial rewards on reasoning accuracy by introducing additional actions and adaptive rewards. To ensure that the entities in the training and testing sets are disjoint under the inductive setting, we construct six MKG datasets with varying scales. Experimental results demonstrate the superior performance of our proposed model compared to existing baselines across different settings.

2307.07510

NNLL Resummation for Projected Three-Point Energy Correlator

The projected energy correlator measures the energy deposited in multiple detectors as a function of the largest angular distance $x_L = (1 - \cos\chi_L)/2$ between detectors. The collinear limit $x_L\to 0$ of the projected energy correlator is particularly interesting for understanding the jet-substructures, while the large logarithms of $x_L$ could potentially spoil the perturbation theory and must be resummed. As a necessary ingredient for its resummation at next-to-next-to-leading logarithmic (NNLL) accuracy, we calculate the two-loop jet functions for the projected three-point energy correlator (E3C), using direct integration method and the parameter space Integration-by-Part (IBP) method. We then present the NNLL resummation for $e^+e^-$ annihilation and an approximate NNLL resummation for $pp\rightarrow jj$ process, where the two-loop hard constant is estimated in the latter case. The convergence is improved and the hadronization effect in the collinear limit is suppressed when considering the ratio of E3C distribution to two-point energy-energy correlator (EEC). Our results show potential in precision determination of strong coupling constant using energy correlators from both $e^+e^-$ data and $pp$ data.

Wen Chen, Jun Gao, Yibei Li, Zhen Xu, Xiaoyuan Zhang, Hua Xing Zhu

2023-07-14T17:58:13Z

http://arxiv.org/abs/2307.07510v1

# NNLL Resummation for Projected Three-Point Energy Correlator ###### Abstract The projected energy correlator measures the energy deposited in multiple detectors as a function of the largest angular distance \(x_{L}=(1-\cos\chi_{L})/2\) between detectors. The collinear limit \(x_{L}\to 0\) of the projected energy correlator is particularly interesting for understanding the jet-substructures, while the large logarithms of \(x_{L}\) could potentially spoil the perturbation theory and must be resummed. As a necessary ingredient for its resummation at next-to-next-to-leading logarithmic (NNLL) accuracy, we calculate the two-loop jet functions for the projected three-point energy correlator (E3C), using direct integration method and the parameter space Integration-by-Part (IBP) method. We then present the NNLL resummation for \(e^{+}e^{-}\) annihilation and an approximate NNLL resummation for \(pp\to jj\) process, where the two-loop hard constant is estimated in the latter case. The convergence is improved and the hadronization effect in the collinear limit is suppressed when considering the ratio of E3C distribution to two-point energy-energy correlator (EEC). Our results show potential in precision determination of strong coupling constant using energy correlators from both \(e^{+}e^{-}\) data and \(pp\) data. + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutet: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutetext: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA + Footnote †: institutet: \({}^{a}\) Department of Physics, University of California, Berkeley, CA 94720, USA Introduction Energy correlators are a class of multi-particle angle correlation functions, weighted by the particle energy. Thanks to the energy weighting, they are infrared and collinear safe observables and can be calculated in perturbation theory. The simplest energy correlator is a two-point energy correlator, or Energy-Energy Correlation function (EEC). Proposed in 1970s [1; 2], EEC measures the correlation of energy deposited in two detectors as a function of the angle \(\chi\) between them. In perturbation theory, the definition of EEC reads \[\frac{d\sigma^{[2]}}{d\cos\chi}\equiv\sum_{i,j}\int d\sigma\frac{E_{i}E_{j}}{Q ^{2}}\delta\left(\vec{n}_{i}\cdot\vec{n}_{j}-\cos\chi\right)\,, \tag{1}\] where \(i\), \(j\) run over all the final state particles, \(\vec{n}_{i}\) and \(\vec{n}_{j}\) are unit three-vectors that define the directions of the particles, and \(Q\) is the total energy in the center-of-mass frame. Compared with other event shape variables studied at Large Electron-Positron Collider (LEP), one advantage of EEC is its simple analytic properties. As far as we are aware of, EEC is the only event shape that can be calculated analytically beyond leading order, e.g. it's now known analytically through to next-to-next-to-leading order (NNLO) [3; 4] in \(\mathcal{N}=4\) super Yang-Mills (SYM) theory and through to NLO in QCD [5; 6; 7]. In recent years, increasing attention has been paid to generalization of EEC to \(N\)-point energy correlators, which measure the energies of the outgoing particles with \(N\) detectors at colliders and turn out to be a function of \(N(N-1)/2\) angles among these detectors [8; 9; 10; 11; 12; 13; 14]. For example, the three-point energy correlator (EEEC) is defined as \[\frac{d^{3}\sigma}{dx_{1}dx_{2}dx_{3}}\equiv \sum_{i,j,k}\int d\sigma\frac{E_{i}E_{j}E_{k}}{Q^{3}}\] \[\times\delta\left(x_{1}-\frac{1-\cos\theta_{jk}}{2}\right)\delta \left(x_{2}-\frac{1-\cos\theta_{ik}}{2}\right)\delta\left(x_{3}-\frac{1-\cos \theta_{ij}}{2}\right)\,, \tag{2}\] which gives rise to rich functional dependence on the angles and can be used to probe various properties of perturbative QCD. The LO EEEC was first computed in the triple collinear limit in Ref. [9], later genelarized to arbitrary angle dependence in both \(\mathcal{N}=4\) SYM [15] and QCD [14]. To reduce the dimension of the kinematic space of the measured angles without losing too much useful information, one can project the kinematic dependence into a 1D subspace, which leads to the so-called _projected energy correlator_[16]. In momentum space, projected \(N\)-point energy correlator (ENC) is given by restricting the maximum angular distance to be \(x_{L}\): \[\frac{d\sigma^{[N]}}{dx_{L}}\equiv\sum_{n}\sum_{1\leq i_{1},\cdots i_{N}\leq n }\int d\sigma\frac{\prod_{a=1}^{N}E_{i_{a}}}{Q^{N}}\delta(x_{L}-\max\{x_{i_{1 },i_{2}},x_{i_{1},i_{3}},\cdots x_{i_{N-1},i_{N}}\})\,, \tag{3}\] and for example, EEEC is then reduced to the projected three-point correlator (E3C). In this work we are mainly interested in the small angle, or collinear limit of E3C, namely \(x_{L}\to 0\). It is well-known in the boundary of phase space, incomplete cancellation of infrared divergences can lead to large logarithms that could possibly spoil the convergence of the perturbation theory and thus it is essential to resum these large logarithms to all orders. EEC is special as it exhibits both large logarithms in collinear limit and back-to-back limit. In this work we are interested in the large logarithms in the collinear limit, for which the most singular terms behave as \(\alpha_{s}^{n}\ln^{n}x_{L}\) at \(n\) loops. In the collinear region, EEC can be factorized into a hard function and a jet function, both of which live in the flavor space. The resummation of collinear EEC has been performed up to NNLL accuracy in both QCD [17] and \(\mathcal{N}=4\) SYM [17; 18; 19]. More interestingly, the collinear factorization can be easily generalized to three-point energy correlator [9] and even the projected \(N\)-point energy correlator [16]. Previously, LL and NLL resummation has been performed in [20; 16; 21]. To improve upon those results, it is necessary to compute the relevant jet and hard function to higher order. While the hard function is universal for them, the jet functions differ by the measurement function. One of the key new results in this paper is the calculation of two-loop jet function for projected three-point energy correlator, which is the last missing ingredient for NNLL resummation of projected three-point energy correlator in \(e^{+}e^{-}\) collider. One of the main motivations for improving the theoretical accuracy of projected energy correlators comes from the possibility of determining the strong coupling constant \(\alpha_{s}\) by measuring the ratio of projected energy correlators [16]. Measurements of strong coupling constant using classical QCD event shape observable has been actively studied for a long time, e.g. [22; 23; 24; 25; 26; 27; 28; 29; 30; 31; 32; 33; 34; 35; 36; 37; 38; 39; 40]. In recent years, there has been increasing attention to using jet substructure observables to extract \(\alpha_{s}\), such as soft-drop thrust and jet mass [41; 42], see also [43] for \(\alpha_{s}\) determination from jet substructure by demixing quark and gluon jets. Since we are mainly concerned with the collinear limit of projected energy correlators in this paper, our results naturally provide theory input for measuring projected energy correlator within a jet, treating it as a jet substructure observable. We will show that considering the ratio of E3C and EEC can significantly reduce scale uncertainties and hadronization corrections, which makes it a good candidate for precision determination of \(\alpha_{s}\) using jet substructure. We also note that energy correlators have the advantage that they can be defined and calculated using charged hadrons only [16; 44]. Using the track function formalism [45; 46], it is possible to perform precision calculation for projected energy correlators on tracks in the future. The outline of this paper is as follows. In Sec. 2, we present the factorization theorem for ENC in the collinear limit and the RG evolution for both hard function and jet function. The desired orders required for all the ingredients to achieve NNLL resummation are briefly summarized there. In Sec. 2.4, we calculate the two-loop E3C jet function. Modern multiloop techniques like IBP and differential equation (DE) are applied for both finite and contact terms. Combining all together, we are able to extract the two-loop E3C jet constants, which is the last missing piece of the NNLL resummation for collinear E3C in \(e^{+}e^{-}\) collision. In Sec. 3, we present the matched NNLL results for both E3C and the ratio of E3C to EEC in \(e^{+}e^{-}\) collision. A qualitative analysis is performed to estimate the leading hadronization correction. The resummation procedure is extended to the case of \(pp\) collision, in particular, the \(pp\to\) dijet process in Sec. 4. We present the highest pertur bative prediction given the available ingredients, the approximate NNLL, with the missing two-loop hard function constants estimated and included as an additional uncertainty. We summarize and conclude in Sec. 5. ## 2 Resummation formalism ### Factorization theorem In this subsection, we summarize the factorization theorem for the projected \(N\)-correlator in the collinear limit and describe the necessary ingredients for NNLL resummation [16]. Similar to EEC, \(N\)-point energy correlator (ENC) in this limit is dominated by the logarithmic series of the largest angular distance \(x_{L}\) \[\frac{d\sigma^{[N]}}{dx_{L}}=\sum_{L=1}^{\infty}\sum_{j=-1}^{L-1} \left(\frac{\alpha_{s}(\mu)}{4\pi}\right)^{L}c_{L,j}\mathcal{L}^{j}(x_{L})+ \ldots\,, \tag{1}\] where \(\mathcal{L}^{-1}(x_{L})=\delta(x_{L})\) and \(\mathcal{L}^{j}(x_{L})=\left[\ln^{j}(x_{L})/x_{L}\right]_{+}\) for \(j\geq 0\), with standard plus distribution. We do the logarithm counting in the projected \(N\)-point energy correlator cumulant, defined as \[\Sigma^{[N]}\left(x_{L},\ln\frac{Q^{2}}{\mu^{2}}\right)=\frac{1} {\sigma_{\rm tot}}\int_{0}^{x_{L}}dx_{L}^{\prime}\,\frac{d\sigma^{[N]}}{dx_{ L}^{\prime}}\left(x_{L}^{\prime},\ln\frac{Q^{2}}{\mu^{2}}\right)\,, \tag{2}\] which maps \([\ln^{j}(x_{L})/x_{L}]_{+}\to 1/(j+1)\times\ln^{j+1}(x_{L})\) and \(\delta(x_{L})\to 1\). Then N\({}^{k}\)LL accuracy refers to the logarithmic series \(\sum_{i=0}^{\infty}\sum_{j=\max\{0,i-k\}}^{i}\left(\frac{\alpha_{s}(\mu)}{4\pi }\right)^{i}d_{i,j}\ln^{j}x_{L}\) in the cumulant \(\Sigma^{[N]}\). At leading power, the \(e^{+}e^{-}\) cumulant \(\Sigma^{[N]}\) can be written in terms of a modified factorization formula in the collinear limit \(x_{L}\to 0\)[16]: \[\Sigma^{[N]}_{ee}\left(x_{L},\ln\frac{Q^{2}}{\mu^{2}}\right)= \int_{0}^{1}dx\,x^{N}\vec{J}^{[N]}\left(\ln\frac{x_{L}x^{2}Q^{2}}{\mu^{2}} \right)\cdot\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{\mu^{2}}\right)\,, \tag{3}\] where the hard function \(\vec{H}_{ee}^{[N]}\) encodes the production of a parent parton with energy fraction \(x\) with respect to the center of mass energy, and the jet function \(\vec{J}^{[N]}\) encodes the evolution of the parent parton into a number of collinear partons which contribute to the observable. Similar factorization formula for EEC was first obtained in [17], and checked explicitly with known NLO results in QCD [5; 6] and \(\mathcal{N}=4\) SYM [3; 4]. We note the explicit dependence on the variable \(x\) in both the jet function and the hard function. Ignoring the dependence on different quark flavor, both jet and hard functions are two-component vectors living in the flavor space, i.e. \(\vec{J}^{[N]}=\{J_{q}^{[N]},J_{g}^{[N]}\}\), \(\vec{H}_{ee}=\{H_{ee,q},H_{ee,g}\}\). We will describe their definition for both \(e^{+}e^{-}\) annihilation and \(pp\) collision in detail in the following subsections. We also emphasize that the factorization theorem holds for any \(N\) at leading power, though we only calculate the \(N=3\) case in this paper. Finally the energy weights in the distribution makes projected \(N\)-point energy correlator insensitive to the soft radiations and non-global logarithms. In hadron colliders, the largest angular distance \(x_{L}\) is replaced by the rapidity-azimuth distance \(R_{L}=\max_{i,j\in X_{E}}\sqrt{\Delta\eta_{ij}^{2}+\Delta\phi_{ij}^{2}}\), where \(X_{E}\) is the set of particles that contributes to the energy weight. When the projected energy correlators are measured within a jet, as is typical for jet substructure observable, the cumulant \(\Sigma_{\rm had}^{[N]}\) also depends on the jet radius \(R_{0}\) parameter. In the limit of \(R_{L}\ll R_{0}\), the modified factorization formula can be written as \[\Sigma_{\rm had}^{[N]}\left(R_{0},R_{L},\ln\frac{p_{T}^{2}}{\mu^{2}}\right)= \int_{0}^{1}dx\,x^{N}\vec{J}^{[N]}\left(\ln\frac{R_{L}^{2}x^{2}p_{T}^{2}}{\mu^ {2}}\right)\cdot\vec{H}_{\rm had}\left(R_{0},x,\ln\frac{p_{T}^{2}}{\mu^{2}} \right)\,, \tag{4}\] where \(p_{T}\) is the jet transverse momentum. Around \(R_{L}\sim R_{0}\), the jet function can also depend on \(R_{0}\). However, there is no large logarithms associated with \(R_{0}\), and its dependence can be obtained from fixed-order matching. For simplicity, we will ignore the \(R_{0}\) dependence in the jet function. In that case the jet function become universal between \(e^{+}e^{-}\) and \(pp\) collision. For \(pp\) collision, the hard function depends on the partonic scattering process, as well as parton distribution functions (PDFs). ### Hard functions #### 2.2.1 \(e^{+}e^{-}\) annihilation For \(e^{+}e^{-}\), the hard function is simply the semi-inclusive hadron fragmentation function [47], which depends on the parton flavor and parton energy fraction \(x=\frac{2p\cdot q}{Q^{2}}\), where \(q\) is the total momentum and \(p\) is the parton momentum. The leading order hard function follows from the born process \(e^{+}e^{-}\to q\bar{q}\), \(\vec{H}_{ee}^{(0)}(x)=\{2\delta(1-x),0\}\). At one-loop, we find \[\frac{1}{2}H_{ee,q}^{(1)}(x) = \frac{\alpha_{s}}{4\pi}C_{F}\Bigg{[}\left(\frac{4\pi^{2}}{3}-9 \right)\delta(1-x)+4\left[\frac{\ln(1-x)}{1-x}\right]_{+}\] \[+\left(4\ln(x)-\frac{3}{2}\right)\left(2\frac{1}{\left[1-x\right] _{+}}-x-1\right)-\frac{9x}{2}-2(x+1)\ln(1-x)+\frac{7}{2}\Bigg{]}\,,\] \[H_{ee,g}^{(1)}(x) = \frac{\alpha_{s}}{4\pi}C_{F}\Bigg{[}\frac{4\left(x^{2}-2x+2 \right)\ln(1-x)}{x}+\frac{8\left(x^{2}-2x+2\right)\ln(x)}{x}\Bigg{]}\,. \tag{5}\] The factor \(1/2\) in front of the quark channel indicates for identical contribution from anti-quark, since we do not distinguish quark and anti-quark flavor. At two-loop, the hard function can be found from the coefficient functions in [47]. Similar to the hadron fragmentation function, the renormalization group evolution (RGE) for the hard function \(\vec{H}\) is simply the DGLAP equation, \[\frac{d\vec{H}(x,\ln\frac{Q^{2}}{\mu^{2}})}{d\ln\mu^{2}}=-\int_{x}^{1}\frac{dy }{y}\widehat{P}(y)\cdot\vec{H}\left(\frac{x}{y},\ln\frac{Q^{2}}{\mu^{2}}\right)\,, \tag{6}\] with \(\widehat{P}(y)\) being the singlet timelike splitting matrix, which is now known to three loops [48; 49]. While it is very difficult to derive an analytic solution for DGLAP to all orders in \(\alpha_{s}\), as we will see below, our resummation only uses a \(\alpha_{s}\)-expanded solution (which turns out to be a very good approximation) and only requires certain moments of the hard function. Explicitly, we will only need the regular and logarithmic moments for the hard function defined as the following [17], \[\int_{0}^{1}dx\,x^{N}\,H_{q,g}(x,\mu=Q) = \sum_{L=0}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}h_{L}^{q,g}(N)\,,\] \[\int_{0}^{1}dx\,x^{N}\,\ln x\,H_{q,g}(x,\mu=Q) = \sum_{L=1}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}\dot{h} _{L}^{q,g}(N)\,,\] \[\int_{0}^{1}dx\,x^{N}\,\ln^{2}x\,H_{q,g}(x,\mu=Q) = \sum_{L=1}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}\ddot{h }_{L}^{q,g}(N)\,. \tag{7}\] Here we use \(x^{N}\ln x=\partial_{N}x^{N}\) and the dot on the RHS stands for the derivative. The expressions of needed hard function moments can be found in Appendix A. #### 2.2.2 \(pp\) collision In hadronic collisions, we mainly focus on the dijet production \(pp\to jj\), which has a relatively large cross section at the LHC. Different from \(e^{+}e^{-}\) collider, this hard function incorporates the partonic scattering cross sections, the contribution from parton distribution functions (PDFs) and the jet algorithms for clustering the particles. Currently, to the best of our knowledge, the hard function is not know at two-loop. However, important progress are being made to compute those hard functions, e.g. [50]. Similar to the \(e^{+}e^{-}\) case, our resummation will only need the hard function moments. In this work we evaluate the needed moments of the hard function numerically in Madgraph5[51, 52]. To investigate the sensitivity of the result to the values of \(\alpha_{s}\), we used three different PDF sets: NNPDF31_nnlo_as_0112, NNPDF31_nnlo_as_0118 and NNPDF31_nnlo_as_0124 through Lhapdf[53]. Each PDF set fixes also the value of \(\alpha_{s}(m_{Z})\) and the corresponding evolution in Madgraph5. To address the fact that the hard function contains collinear divergence when resolving the energy fraction of the quarks and gluons, we use the one cut-off phase space slicing to regularize the collinear singularity, as implemented in [54]. With the collinear divergent contribution singled out and calculated analytically, the remaining contributions can be evaluated numerically. The detailed discussion can be found in Appendix A. For \(pp\to jj\), we adopt the anti-\(k_{t}\) algorithm [55] for jet detection and use the following parameters in the calculation \[R_{0}=0.4,\qquad p_{T}>15\,\text{GeV},\qquad|\eta|<1.5\,. \tag{8}\] The two leading jets are further subject to the following cuts \[|\Delta\phi(j_{1},j_{2})|>2,\qquad|p_{T}^{1}-p_{T}^{2}|/(p_{T}^{1}+p_{T}^{2})< 0.5\,, \tag{9}\] and cast to the corresponding \(p_{t}\) bins for the analysis. The calculated moments need to be normalized with the cross section \(\sigma_{J}\) of jet production within specific \(p_{t}\) range. In particular, we expand \(H_{\text{had}}/\sigma_{J}\) to NLO in \(a_{s}\), and take the \(\mathcal{O}(a_{s}^{0})\) and \(\mathcal{O}(a_{s}^{1})\) as the leading and next-to-leading order results. For the purpose of phenomenological studies, we will focus on two different \(p_{t}\) ranges: \([300,350]\) GeV and \([500,550]\) GeV. The hard function moments needed for NNLL are also summarized in Appendix A. ### Jet functions The E3C jet function, on the other hand, encodes the measurement information. From RG invariance of the modified factorization formula (3), the jet function satisfies a modified timelike DGLAP evolution equation \[\frac{d\vec{J}^{[N]}(\ln\frac{x_{L}Q^{2}}{\mu^{2}})}{d\ln\mu^{2}}= \int_{0}^{1}dy\,y^{N}\vec{J}^{[N]}\left(\ln\frac{x_{L}y^{2}Q^{2}}{\mu^{2}} \right)\cdot\widehat{P}(y)\,. \tag{10}\] In order to write down an operator description of the E3C jet function, we first recall the collinear EEEC jet function from [9]: \[J_{q}(x_{1},x_{2},x_{3},Q,\mu^{2})=\] \[\int\frac{dl^{+}}{2\pi}\frac{1}{2N_{C}}\text{Tr}\int d^{4}xe^{il \cdot x}\langle 0|\frac{\not{n}}{2}\chi_{n}(x)\widehat{\mathcal{M}}_{\text{EEEC}} \ \delta(Q+\bar{n}\cdot\mathcal{P})\delta^{2}(\mathcal{P}_{\perp})\bar{\chi}_{n}( 0)|0\rangle\] \[J_{g}(x_{1},x_{2},x_{3},Q,\mu^{2})=\] \[\int\frac{dl^{+}}{2\pi}\frac{1}{2(N_{C}^{2}-1)}\text{Tr}\int d^{ 4}xe^{il\cdot x}\langle 0|\mathcal{B}^{a,\mu}_{n,\perp}(x)\widehat{\mathcal{M}}_{ \text{EEEC}}\ \delta(Q+\bar{n}\cdot\mathcal{P})\delta^{2}(\mathcal{P}_{\perp}) \mathcal{B}^{a,\mu}_{n,\perp}(0)|0\rangle\,, \tag{11}\] where \(\chi_{n}\equiv W_{n}^{\dagger}\xi_{n}\) is the collinear quark and \(\mathcal{B}^{\mu}_{n,\perp}\equiv\frac{1}{g}\left[\frac{1}{\bar{n}\cdot P}W_{n }^{\dagger}[i\bar{n}\cdot D_{n},iD_{n\perp}^{\mu}|W_{n}\right]\) is the collinear gluon, and \(\mathcal{P}^{\mu}_{n\perp}\) form a complete set of collinear gauge invariant building blocks [56] in SCET [57; 58; 59; 60; 61]. The triple collinear measurement function \(\widehat{\mathcal{M}}_{\text{EEEC}}\) is defined as \[\widehat{\mathcal{M}}_{\text{EEEC}}(x_{1},x_{2},x_{3})=\sum_{i,j, k}\frac{E_{i}E_{j}E_{k}}{Q^{3}}\delta\left(x_{1}-\frac{\theta_{ij}^{2}}{4} \right)\delta\left(x_{2}-\frac{\theta_{jk}^{2}}{4}\right)\delta\left(x_{3}- \frac{\theta_{ki}^{2}}{4}\right)\,, \tag{12}\] with \(\theta_{ij}\) being the angle between parton \(i\) and \(j\). Then our E3C jet function has the same form as EEEC jet function, with a replacement of the measurement function: \[\widehat{\mathcal{M}}_{\text{EEEC}}\Rightarrow\widehat{\mathcal{M }}_{\text{E3C}}(x_{L}) =\int_{0}^{x_{L}}dx_{L}^{\prime}\int_{K}dx_{1}dx_{2}dx_{3}\, \widehat{\mathcal{M}}_{\text{EEEC}}\,\delta\left(x_{L}^{\prime}-\max(x_{1},x_ {2},x_{3})\right)\] \[=\int_{K}dx_{1}dx_{2}dx_{3}\,\widehat{\mathcal{M}}_{\text{EEEC}} \,\theta\left(x_{L}-\max(x_{1},x_{2},x_{3})\right)\,. \tag{13}\] There are two folds integration in the first line. The first one is performed in the allowed kinematic space \(\{x_{1},x_{2},x_{3}\}\in K\) that will be discussed below, projecting the shape-dependent EEEC jet function into a single-scale jet function. The second integration brings the differential measurement to the cumulant level. For \(N>3\), the measurement function takes a similar structure, with more \(\delta\) functions and integrations. Perturbatively, the E3C jet function can be written as \(J_{q,g}=\sum_{L}(\alpha_{s}/4\pi)^{L}J_{q,g}^{(L)}\), and we use the normalization condition \(2^{3}\cdot J_{q}^{(0)}=2^{3}\cdot J_{g}^{(0)}=1\) as in Ref. [16]. The one-loop correction can be calculated from the QCD \(1\to 2\) timelike splitting kernel and is given by \[2^{3}J_{q}^{(1)}=\frac{9C_{F}}{2}\ln\frac{x_{L}Q^{2}}{\mu^{2}}- \frac{37C_{F}}{2}\,,\] \[2^{3}J_{g}^{(1)}=\,\left(\frac{21C_{A}}{5}+\frac{3n_{f}}{10}\right)\ln\frac{x_{L} Q^{2}}{\mu^{2}}-\frac{449C_{A}}{25}-\frac{21n_{f}}{25}\,. \tag{14}\] Note that the \(\mu\)-dependent terms are precisely captured by the jet RGE, while the remaining constants have to come from the fixed-order calculation. One of the main result in this paper is to calculate the two-loop constants described below. ### Two-loop calculation for the E3C jet function In this subsection, we present the two-loop calculation of the E3C jet functions for both quark jets and gluon jets. Since they are universal in the small angle limit, they can be used in both \(e^{+}e^{-}\) collision and \(pp\) collision. We start from recalling the definition of E3C at finite angle before taking the small angle limit. At two loops, E3C receives contributions from double-real (RR) and real-virtual (RV) as well as double-virtual (VV) corrections to \(q\to q\), from which the quark jet function can be extracted by matching to the factorization formula, (3). Similarly, the gluon jet function can be extracted from the NLO E3C distribution of Higgs gluonic decay \(H\to gg\). To organize the calculation, we rewrite the definition of E3C in Eq. (3) with the number of energy weight: \[\frac{1}{\sigma_{0}}\frac{d\sigma^{[3]}}{dx_{L}} =\sum_{1\leq i_{1}\neq i_{2}\neq i_{3}\leq 4}\int\mathrm{d LIPS}_{4}\,|\mathcal{M}_{4}|^{2}\frac{E_{i_{1}}E_{i_{2}}E_{i_{3}}}{Q^{3}} \delta(x_{L}-\max\{x_{i_{1},i_{2}},x_{i_{1},i_{3}},x_{i_{2},i_{3}}\})\] \[+\sum_{n\in\{3,4\}}\sum_{1\leq i_{1}\neq i_{2}\leq n}\int\mathrm{ dLIPS}_{n}\,|\mathcal{M}_{n}|^{2}\frac{E_{i_{1}}^{2}E_{i_{2}}}{Q^{3}}\delta(x_{L}-x _{i_{1},i_{2}})\] \[+\sum_{n\in\{2,3,4\}}\sum_{1\leq i_{1}\leq n}\int\mathrm{dLIPS}_{ n}\,|\mathcal{M}_{n}|^{2}\frac{E_{i_{1}}^{3}}{Q^{3}}\delta(x_{L})\,, \tag{15}\] where we normalize the distribution to the born cross-section in \(d\) dimension. The first line represents the contribution from nonidentical energy weights measurement and the other lines are called contact terms. If we define \(x_{1}=x_{L}z\bar{z}\), \(x_{2}=x_{L}(1-z)(1-\bar{z})\) and \(x_{3}=x_{L}\), then in the collinear limits, they are the contact terms for \(\delta(z\bar{z})\) that captures the strict squeeze limit and \(\delta(x_{L})\) that captures the strict triple collinear limit. The main goal of this section is to compute the collinear limit of Eq. (15) and extract the corresponding two-loop constants. The lowest regular distribution of the E3C quark jet function comes from tree-level process \(\gamma^{*}\to 4\) partons in electron-positron annihilation, which under the triple collinear limit, factorizes into the born process \(\gamma^{*}\to q\bar{q}\) and the \(1\to 3\) splitting functions, and we will call it nonidentical energy weight term. Below we will introduce two different methods to compute this part. The traditional method is to calculate the EEEC jet function to order \(\mathcal{O}(\epsilon)\) and to integrate two angular distances \(x_{2},\,x_{3}\) numerically by the interpolation method. The OPE singularities (sometimes called squeezed singularities) of EEEC are subtracted and integrated in \(d\) dimension separately. The second approach benefits from the parameter space IBP method [62; 63; 64] developed very recently. Only 7 master integrals are needed to express EEEC, allowing the precise calculation of the remaining two-fold integral. The other two parts contribute to the contact terms and cancel the infrared divergence, which is guaranteed by the Kinoshita-Lee-Nauenberg (KLN) theorem [65; 66]. Similar to EEC at NLO, the measurement function in the contact terms can be treated as a non-standard cut propagators, which allows for a generalized IBP reduction in Liltered[67; 68] and Fire6[69]. The master integrals then can be calculated in packages like Canonica[70] or Libra[71; 72] with the differential equation method implemented. #### 2.4.1 Nonidentical energy-weight terms We start by computing the nonidentical energy-weight contribution in the traditional approach. As discussed in Ref. [9], the inclusive jet function \(J_{i\overline{j}k}\) is related to the \(1\to 3\) splitting function \(P_{ijk}\)[73; 74; 75] through \[J^{\rm nonid}\equiv J_{i\overline{j}k}=\int{\rm d}\Phi_{c}^{(3)}\left(\frac{ \mu^{2}e^{\gamma_{E}}}{4\pi}\right)^{2\epsilon}\frac{4g^{4}}{\delta_{123}^{2} }\sum_{i,j,k}P_{ijk}\widehat{\mathcal{M}}_{\rm EEEEC}\,, \tag{16}\] where \({\rm d}\Phi_{c}^{(3)}\) is the triple collinear phase space [75; 76], and \(i,j,k\) run over all final-state particles. The fully differential distribution with respect to all angular distances \(\{x_{1},x_{2},x_{3}\}\) in \(d=4-2\epsilon\) dimension is then written as \[\frac{dJ^{\rm nonid}}{dx_{L}d\text{Re}(z)d\text{Im}(z)}=\left( \frac{\mu^{2}}{Q^{2}}\right)^{2\epsilon}\frac{\alpha_{s}^{2}}{\pi^{3}}\frac{e ^{2\epsilon\gamma_{E}}}{\Gamma(1-2\epsilon)}\frac{1}{x_{L}^{1+2\epsilon}} \frac{1}{(2\text{Im}(z))^{2\epsilon}}\\ \times\left[G(z)+\epsilon F(z)+\epsilon^{2}H(z)+\mathcal{O}( \epsilon^{3})\right]\,, \tag{17}\] where \(G(z),F(z),H(z),\cdots\) the shape function in \(\epsilon\) expansion. The order \(\mathcal{O}(1)\) part \(G(z)\) is computed analytically in [9] and following the same approach, we also calculate the complete result for \(F(z)\) and the \(z\to 1\) limit of \(H(z)\). We will see that these are all the needed ingredients for nonidentical part. Note that the \(x_{L}\) dependence is defined by plus distribution, where \[\frac{1}{x_{L}^{1+2\epsilon}}=-\frac{\delta(x_{L})}{2\epsilon}+\left(\frac{1 }{x_{L}}\right)_{+}-2\epsilon\left(\frac{\ln x_{L}}{x_{L}}\right)_{+}+\cdots\,. \tag{18}\] In order to perform the integral over \(z\), we need to figure out the integration region first. Compared with the first line in Eq. (15), it is straightforward to show that \[\frac{dJ^{\rm nonid}}{dx_{L}}=\left(\frac{\mu^{2}}{Q^{2}}\right)^{2\epsilon} \frac{\alpha_{s}^{2}}{\pi^{3}}\frac{e^{2\epsilon\gamma_{E}}}{\Gamma(1-2 \epsilon)}\frac{6}{x_{L}^{1+2\epsilon}}\underbrace{\int_{\mathcal{S}}\frac{d \text{Re}(z)d\text{Im}(z)}{(2\text{Im}(z))^{2\epsilon}}\left[G(z)+\epsilon F( z)+\epsilon^{2}H(z)+\mathcal{O}(\epsilon^{3})\right]}_{\equiv A(\epsilon)}\,, \tag{19}\] where the constant factor \(6\) comes from the \(S_{3}\) permutation symmetry and the integration region \(\mathcal{S}\) is given in Fig. 1. To calculate \(A(\epsilon)\) numerically, we also need to subtract the OPE singularities around \(z\to 1\) at the integrand level, and evaluate its \(z\) integration analytically in \(d\) dimension. The full asymptotic expansion of \(z\to 1\) is given in the appendix C. The most singular term is proportional to \(\frac{1}{(1-z)(1-\bar{z})}\), which gives rise to \[\int_{0}^{\frac{\sqrt{3}}{2}}d\text{Im}(z)\int_{\frac{1}{2}}^{\sqrt{1 -(\text{Im}(z))^{2}}}d\text{Re}(z)\frac{1}{(2\text{Im}(z))^{2\epsilon}}\frac{1} {(1-z)(1-\bar{z})}\\ =-\frac{\pi}{4\epsilon}-\kappa+\epsilon\left(-\frac{167}{1080} \pi^{3}-\frac{1}{20}\pi\ln^{2}3+\kappa\ln 3+\frac{12}{5}\eta\right)+\mathcal{O}( \epsilon^{2})\,. \tag{20}\] Here \(\kappa=\text{ImLi}_{2}e^{i\frac{\pi}{3}}\) is the Gieseking's constant living in the transcendentality-two family and \(\eta=\text{ImLi}_{3}\left(\frac{i}{\sqrt{3}}\right)\) is a parity-odd transcendentality-three constant. These constants are typical numbers in loop integrals, especially in trijet observable calculations. With subtraction terms, the integral \(A\) in Eq. (19) up to order \(\mathcal{O}(\epsilon)\) is then written as \[A=\int_{0}^{\frac{\sqrt{3}}{2}}d\text{Im}(z)\int_{\frac{1}{2}}^ {\sqrt{1-(\text{Im}(z))^{2}}}d\text{Re}(z)\frac{1}{(2\text{Im}(z))^{2\epsilon }}\left[G(z\to 1)+\epsilon F(z\to 1)+\epsilon^{2}H(z\to 1)\right]\\ +\int_{0}^{\frac{\sqrt{3}}{2}}d\text{Im}(z)\int_{\frac{1}{2}}^ {\sqrt{1-(\text{Im}(z))^{2}}}d\text{Re}(z)\frac{1}{(2\text{Im}(z))^{2\epsilon }}\left[(G(z)-G(z\to 1))+\epsilon\left(F(z)-F(z\to 1)\right)\right]. \tag{21}\] The first term is proportional to Eq. (20) and it is straightforward to compute it to \(\mathcal{O}(\epsilon)\). For the second integral, we have to expand in \(\epsilon\) and evaluate it numerically. To implement the interpolation method, we first change the integration variables via \(v_{1}=\frac{2}{\sqrt{3}}\text{Im}(z)\) and \(v_{2}=\frac{\text{Re}(z)-\frac{1}{2}}{\sqrt{1-(\text{Im}(z))^{2}-\frac{1}{2}}}\), such that both \(v_{1,2}\) range from \(0\) to \(1\). Then we can build a 2D lattice by discretizing \(v_{1,2}\) and approximate our integrand with polynomials. This allows one to perform the two-fold numerical integral directly in Mathematica. To check the stability of the integration and estimate the statistical error, we vary the lattice size and the order of polynomials and see which significant figure remains unchanged. Eventually we obtain both Figure 1: The integration region \(\mathcal{S}\) for E3C jet function. The \(S_{3}\) symmetry is applied to reduce the entire region of \(z\) into \(6\) times the blue region. The integration range for \(z\) then becomes \(\int_{0}^{\frac{\sqrt{3}}{2}}d\text{Im}(z)\int_{\frac{1}{2}}^{\sqrt{1-(\text{ Im}(z))^{2}}}d\text{Re}(z)\). \(\delta(x_{L})\) contact term and \(\frac{1}{x_{L}}\) finite term for the nonidentical energy weight contribution. The explicit expression for both quark and gluon jet function can be found in Eq. (45)-(46) in the appendix. Alternatively, benefiting from the recent development of the IBP method in the Feynman parameter space, we can simplify the whole jet function calculation with integral reduction. First of all, recall that Eq. (16) takes the form \[J^{\rm nonid}\equiv\int{\rm d}x_{1}{\rm d}x_{2}{\rm d}x_{3}\frac{dJ^{R}}{dx_{1 }dx_{2}dx_{3}}\propto\int{\rm d}x_{1}{\rm d}x_{2}{\rm d}x_{3}{\rm d}\omega_{1}{ \rm d}\omega_{2}{\rm d}\omega_{3}\delta(1-\omega_{1}-\omega_{2}-\omega_{3}) \hat{P}_{ijk}\,. \tag{22}\] Here \(\hat{P}\) is a homogeneous function of the energy fraction \(\omega_{i}\) of the final-state particles. Explicitly, it is of the form \[\hat{P}_{ijk}=\frac{\omega_{1}^{\alpha_{1}}\omega_{2}^{\alpha_{2}}\omega_{3}^ {\alpha_{3}}}{f_{1}^{\beta_{1}}f_{2}^{\beta_{2}}}\,, \tag{23}\] with \(f_{1}\) linear in \(\omega_{i}\), and \(f_{2}\) a polynomial of \(\omega_{i}\) of degree 2. Following the idea in Ref [9], the integral \(\frac{{\rm d}^{3}J^{R}}{{\rm d}x_{1}{\rm d}x_{2}{\rm d}x_{3}}\) in Eq. (22) can be related to a Feynman parameter integral through1 Footnote 1: In the special cases where \(\beta_{1}=0\) or \(f_{1}=U\), we don’t need to introduce the parameter \(\omega_{4}\). \[\frac{d^{3}J^{\rm nonid}}{dx_{1}dx_{2}dx_{3}}= \frac{\Gamma(\beta_{1}+\beta_{2})}{\Gamma(\beta_{1})\Gamma(\beta _{2})}\int{\rm d}\omega_{1}{\rm d}\omega_{2}{\rm d}\omega_{3}{\rm d}\omega_{4 }\delta(1-\omega_{1}-\omega_{2}-\omega_{3})\frac{\omega_{1}^{\alpha_{1}}\omega _{2}^{\alpha_{2}}\omega_{3}^{\alpha_{3}}\omega_{4}^{\beta_{1}-1}}{(f_{2}+f_{1} \omega_{4})^{\beta_{1}+\beta_{2}}}\] \[= \frac{\Gamma(\beta_{1}+\beta_{2})}{\Gamma(\beta_{1})\Gamma(\beta _{2})}\int{\rm d}\omega_{1}{\rm d}\omega_{2}{\rm d}\omega_{3}{\rm d}\omega_{4 }\delta(1-\omega_{1}-\omega_{2}-\omega_{3})\frac{\omega_{1}^{\alpha_{1}}\omega _{2}^{\alpha_{2}}\omega_{3}^{\alpha_{3}}\omega_{4}^{\beta_{1}-1}}{(f_{2}+f_{1} \omega_{4})^{\beta_{1}+\beta_{2}}}\] \[= \frac{\Gamma(\beta_{1}+\beta_{2})}{\Gamma(\beta_{1})\Gamma(\beta _{2})}\int{\rm d}\omega_{1}{\rm d}\omega_{2}{\rm d}\omega_{3}{\rm d}\omega_{4 }\delta(1-U)\frac{\omega_{1}^{\alpha_{1}}\omega_{2}^{\alpha_{2}}\omega_{3}^{ \alpha_{3}}\omega_{4}^{\alpha_{4}}}{U^{\lambda_{1}}F^{\lambda_{2}}}\] \[\equiv \frac{\Gamma(\alpha_{1})\Gamma(\alpha_{2})\Gamma(\alpha_{3})}{ \Gamma(\beta_{1})\Gamma(\beta_{2})}I(\alpha_{0},\alpha_{1},\alpha_{2},\alpha_ {3},\alpha_{4})\,, \tag{24}\] where \(U=\omega_{1}+\omega_{2}+\omega_{3}\), \(F=f_{2}+f_{1}\omega_{4}\), \(\lambda_{1}=\alpha_{1}+\alpha_{2}+\alpha_{3}-\beta_{1}-2\beta_{2}+3\), \(\lambda_{2}=\beta_{1}+\beta_{2}\), and \(\alpha_{0}=-\beta_{1}-\beta_{2}\). The integral in the last line is a standard parametric Feynman integral, which can be reduced with IBP reduction [77; 78] in the parametric representation [62; 63; 64; 79]2. The master integrals are Footnote 2: The algorithms described in ref. [63] to generate symbolic rules work only when all the indices are nonnegative. Thus, here we carry out the reduction by merely solving IBP identities using Kira[80; 81; 82; 83]. \[\mathcal{I}_{1} =I_{1}\left(\alpha_{0},-2\epsilon,1-2\epsilon,-2\epsilon\right), \mathcal{I}_{2} =I_{1}\left(\alpha_{0},1-2\epsilon,-2\epsilon,-2\epsilon\right),\] \[\mathcal{I}_{3} =I_{1}\left(\alpha_{0},-2\epsilon,-2\epsilon,1-2\epsilon\right), \mathcal{I}_{4} =I_{1}\left(\alpha_{0},-2\epsilon,-2\epsilon,-2\epsilon\right),\] \[\mathcal{I}_{5} =I_{2}\left(\alpha_{0},-2\epsilon,-2\epsilon,-2\epsilon,0\right), \mathcal{I}_{6} =I_{3}\left(\alpha_{0},-2\epsilon,-2\epsilon,-2\epsilon,0\right),\] \[\mathcal{I}_{7} =I_{4}\left(\alpha_{0},-2\epsilon,-2\epsilon,-2\epsilon,0\right)\,, \tag{25}\] with the integrals \(I_{i}\) defined by the \(F\) polynomials \[F_{1} =x_{1}\omega_{2}\omega_{3}+x_{2}\omega_{1}\omega_{3}+x_{3}\omega_{1 }\omega_{2}\,, F_{2} =F_{1}+(\omega_{1}+\omega_{2})\omega_{4}\,,\] \[F_{3} =F_{1}+(\omega_{1}+\omega_{3})\omega_{4}\,, F_{4} =F_{1}+(\omega_{2}+\omega_{3})\omega_{4}\,, \tag{26}\] and \(\alpha_{0}=6\epsilon-2^{3}\). The master integrals can be evaluated using the differential equation technique [84; 85]. For simplicity, we set \(\mu=x_{3}=1\), and introduce \(u\) and \(v\) following \(z=u(1+iv)\). Then we construct the differential-equation system with respect to \(u\), and derive the canonical basis [86] using Libra[71; 72] \[\mathcal{I}_{1}^{\prime}= 6u(v-1)\mathcal{I}_{4}+\frac{x_{1}(1-2\epsilon)}{\epsilon} \mathcal{I}_{1}\,,\] \[\mathcal{I}_{2}^{\prime}= 6(u-1)\mathcal{I}_{4}+\frac{x_{2}(1-2\epsilon)}{\epsilon} \mathcal{I}_{2}\,,\] \[\mathcal{I}_{3}^{\prime}= 6\left(uv+u-x_{1}\right)\mathcal{I}_{4}+\frac{x_{1}x_{2}(1-2 \epsilon)}{\epsilon}\mathcal{I}_{3}\,,\] \[\mathcal{I}_{4}^{\prime}= 6uv\mathcal{I}_{4}\,,\] \[\mathcal{I}_{5}^{\prime}= \left(x_{1}-x_{2}\right)\mathcal{I}_{5}\,,\] \[\mathcal{I}_{6}^{\prime}= \left(x_{3}-x_{1}\right)\mathcal{I}_{6}\,,\] \[\mathcal{I}_{7}^{\prime}= \left(x_{2}-x_{3}\right)\mathcal{I}_{7}\,, \tag{27}\] with the corresponding alphabet \(\{u,\,2u-1,\,x_{2},\,x_{2}-1\}\). By solving the differential-equation system, we can express the master integrals via Goncharov polylogarithms (GPLs) [87; 88; 89]. The GPL is defined iteratively by \[G(a_{1},\cdots a_{n};x)\equiv\int_{0}^{x}\frac{dt}{t-a_{1}}G(a_{2},\cdots a_{ n};t)\,, \tag{28}\] with \[G(;x)\equiv 1,\quad G(\vec{0}_{n};x)\equiv\frac{1}{n!}\ln^{n}(x)\,. \tag{29}\] After finishing the simplified calculation of EEEC in the collinear limit, we still need to integrate two angular distances for the projected EEEC as the previous approach. By virtue of the \(S_{3}\) permutation symmetry, this amount to consider \[\frac{dJ^{\rm nonid}}{dx_{L}}= 6\int{\rm d}x_{1}{\rm d}x_{2}\ \Theta(x_{1},x_{2})\frac{d^{3}J}{dx_{1}dx_{2}dx_{3}}\] \[= 24\int{\rm d}u{\rm d}v\ \Theta(x_{1},x_{2})u^{2}v\frac{d^{3}J}{dx _{1}dx_{2}dx_{3}}\] \[\equiv \int{\rm d}u{\rm d}v\ \Theta(x_{1},x_{2})\tilde{J}(u,v)\,, \tag{30}\] where \(\Theta(x_{1},x_{2})\equiv\theta\left(1-\sqrt{x_{2}}\right)\theta\left(\sqrt{x _{2}}-\sqrt{x_{1}}\right)\theta\left(\sqrt{x_{2}}+\sqrt{x_{1}}-1\right)\). Now the OPE singularity corresponds to \(u\to 0\) limit, and similarly, we need to subtract the singular behavior and do the integration separately: \[\frac{dJ^{\rm nonid}}{dx_{L}}=\int{\rm d}u{\rm d}v\ \Theta(x_{1},x_{2})\tilde{J}(u \to 0)+\int{\rm d}u{\rm d}v\ \Theta(x_{1},x_{2})\left[\tilde{J}(u,v)-\tilde{J}(u \to 0)\right]\,, \tag{31}\] where again we can evaluate the first integral in \(d\) dimension and expand the integrand of the second one in \(\epsilon\). To calculate the \(\tilde{J}(u\to 0)\), now we can directly extract the asymptotic expansion of the integral \(I\) in Eq. (24) from DE, in which we identify two expansion regions: \[\text{hard region:}\quad\omega_{1}\sim\omega_{2}\sim\omega_{3} \sim 1\,,\] \[\text{small region:}\quad\omega_{2}\sim\omega_{3}\sim 1,\ \omega_{1} \sim u^{2}\,. \tag{32}\] Eventually we only need to integrate the reduced master integrals in \(d\) dimension. Regarding the second integral in Eq. (31), the \(u\) integral is straightforward since \(\tilde{J}(u,v)\) is expressed in terms of GPLs of the form \(G(\ldots,u)\). However, the \(v\) integral becomes unstable in two regions \(v\to 0\) and \(v\to\infty\). To resolve this problem, we decompose the \(v\in[0,\infty]\) integration into three parts: \([0,\ \frac{1}{C}]\), \([\frac{1}{C},\ C]\), and \([C,\ \infty]\), with a arbitrary cut parameter \(C>1\). In the region \((\frac{1}{C},\ C)\), we carry out the integration numerically, with the GPLs numerically using Handyg[90]. The other two regions require expanding the integrand in \(v\) (or \(\frac{1}{v}\)) to \(\mathcal{O}(v^{100})\) (or \(\mathcal{O}(v^{-100})\)) and performing the integration analytically. This expansion can easily be done by asymptotically solving the differential equations satisfied by the GPLs. Eventually, we find the same result as in Eq. (15)-(16). #### 2.4.2 Contact terms While it is convenient to calculate the nonidentical \(E_{i_{1}}E_{i_{2}}E_{i_{3}}\) part starting with the splitting functions, it is preferable to compute the full angular dependence on \(x_{L}\) for corresponding processes (namely \(e^{+}e^{-}\) annihilation and gluonic Higgs decay) with energy weights \(E_{i_{1}}^{2}E_{i_{2}}\) (\(i_{1}\neq i_{2}\)) and \(E_{i_{1}}^{3}\), and extract the contact term from the collinear limit \(x_{L}\to 0\). In other words, we will adopt the full matrix elements squared and compute the full phase space integral using modern multi-loop techniques, with which the collinear expansion gives \(\mathrm{d}\sigma^{[3]}_{\mathrm{E}^{2}\mathrm{EC}}(x_{L})/\mathrm{d}x_{L}\) (the \(E_{i_{1}}^{2}E_{i_{2}}\) (\(i_{1}\neq i_{2}\)) part) and \(\mathrm{d}\sigma^{[3]}_{\mathrm{E}^{3}\mathrm{C}}(x_{L})/\mathrm{d}x_{L}\) (the \(E_{i_{1}}^{3}\) part) in the \(x_{L}\to 0\) limit. We start with the relevant processes in perturbation theory for two-loop jet functions, \[\mathbf{e}^{+}\mathbf{e}^{-}\text{annihilation} \mathbf{Higgs decays}\] \[\gamma^{*}\to q\bar{q}+VV H\to gg+VV\] \[\gamma^{*}\to q\bar{q}g+V H\to ggg+V\] \[H\to q\bar{q}g+V\] \[\gamma^{*}\to q\bar{q}gg H\to gggg\] \[\gamma^{*}\to q\bar{q}q\bar{q} H\to q\bar{q}gg\] \[\gamma^{*}\to q\bar{q}q^{\prime}\bar{q} H\to q\bar{q}q\bar{q}\] \[H\to q\bar{q}\bar{q}^{\prime}\bar{q}^{\prime} \tag{33}\] where \(V\) and \(VV\) denotes one-loop and two-loop correction respectively. In particular, in the \(x_{L}\to 0\) limit, \(1\to 2\) processes only contribute to \(\delta(x_{L})\)-terms (i.e., \(\mathrm{d}\sigma^{[3]}_{\mathrm{E}^{3}\mathrm{C}}(x_{L})/\mathrm{d}x_{L}\)). The calculation setup of \(\mathrm{d}\sigma^{[3]}_{\mathrm{E}^{2}\mathrm{EC}}(x_{L},\epsilon)/\mathrm{d}x _{L}\) shares the same structure as the original EEC, which basically follows the approach described in Ref. [5] and more detail in [6]. Briefly speaking, using the Cutkosky rules [91, 92], we can replace the phase-space on-shell delta functions with the cut propagators \[\delta(p^{2})=\frac{1}{2\pi{\rm i}}\left(\frac{1}{p^{2}-{\rm i}0}-\frac{1}{p^{2}+{ \rm i}0}\right)\,, \tag{34}\] and also the EEC measurement function \(\delta(x_{L}-x_{i,j})\) with \[\delta\left(x_{L}-\frac{1-\cos\theta_{ij}}{2}\right)=\frac{(p_{i }\cdot p_{j})}{x_{L}}\delta\left[2x_{L}(p_{i}\cdot Q)(p_{j}\cdot Q)-p_{i}\cdot p _{j}\right]\] \[= \frac{1}{2\pi{\rm i}}\frac{(p_{i}\cdot p_{j})}{x_{L}}\left\{\frac{ 1}{[2x_{L}(p_{i}\cdot Q)(p_{j}\cdot Q)-p_{i}\cdot p_{j}]-{\rm i}0}-\frac{1}{[2x _{L}(p_{i}\cdot Q)(p_{j}\cdot Q)-p_{i}\cdot p_{j}]+{\rm i}0}\right\}\,, \tag{35}\] where we set the center-of-mass energy \(Q=1\) for simplicity. After topology classification and identification as described in Ref. [6], the E\({}^{2}\)EC integral can be reduced to a set of master integrals \(\widetilde{\mathcal{I}}_{k}(x_{L},\epsilon)\) using IBP reduction and E\({}^{2}\)EC distribution can be written as a linear combination of the master integrals, \[\frac{{\rm d}}{{\rm d}x_{L}}\sigma_{\text{E${}^{2}$EC}}^{[3]}(x_{L},\epsilon)= \sum_{k}\mathcal{C}_{k}(x_{L},\epsilon)\widetilde{\mathcal{I}}_{k}(x_{L}, \epsilon)\,. \tag{36}\] Specifically, we generate the standard IBP equations using Litered[67, 68], add the missing one that is associated with the EEC measurement function by hand, and do the reduction in Fireg[69]. The master integrals turn out to be the same as in NLO EEC calculation for both \(e^{+}e^{-}\) annihilation and gluonic Higgs decays, which can be converted into the canonical basis using the DE package Canonica[70]. In order to obtain the collinear \({\rm d}\sigma_{\text{E${}^{2}$EC}}^{[3]}(x_{L},\epsilon)/{\rm d}x_{L}\), one could surely expand the differential equation asymptotically and derive the analytical expression of the master integrals in that limit. However, the fact that the most singular power of \(\mathcal{C}_{k}\)'s is \(x_{L}^{-8}\) requires us to compute the master integrals up to \(\mathcal{O}(x_{L}^{7})\) order, which turns out to be expensive and time-consuming. This becomes worse in the higher-point energy correlator since the singular power increases as well. One antidote is to reconstruct the coefficients from DE following an ansatz on the structure of asymptotic expansion. In fact, the pattern turns out to be \(x_{L}^{-\epsilon}U_{1}^{(1)}(x_{L},\epsilon)\) at \(\mathcal{O}(\alpha_{s})\) and \(x_{L}^{-\epsilon}U_{1}^{(2)}(x_{L},\epsilon)+x_{L}^{-2\epsilon}U_{2}^{(2)}(x_ {L},\epsilon)\) at \(\mathcal{O}(\alpha_{s}^{2})\), where \(U\) denotes a series in \(x_{L}\) with rational fractions of \(\epsilon\) as the coefficients. Therefore, we perform the asymptotic expansion in the following way. First of all, we solve the canonical DE at \(0<x_{L}<1\) to transcendental-weight \(5\), which can be used to obtain the finite part of the contact term via Eq. (36). The result can be converted to Harmonic polylogarithms (HPLs) with the package Hpl[93] or even classical polylogarithms. Then we can extract the leading power \(x_{L}^{-1}\) and match it to a resummed ansatz \[x_{L}^{-1-\epsilon}C_{1}(\epsilon)+x_{L}^{-1-2\epsilon}C_{2}(\epsilon)\,, \tag{37}\] with unknown \(\epsilon\)-series \(C_{1}(\epsilon)\) and \(C_{2}(\epsilon)\). The matching between fixed order calculation and the resummed structure in \(\epsilon\) leads to the solution of \(C_{1}(\epsilon)\) and \(C_{2}(\epsilon)\) in \(\epsilon\) expansion. Since \(x_{L}^{-1-\epsilon}\) and \(x_{L}^{-1-2\epsilon}\) are defined with plus distribution similar to Eq. (18), now we obtain the correct \(\mathcal{O}(\epsilon^{0})\) formula for \({\rm d}\sigma_{\text{E${}^{2}$EC}}^{[3]}(x_{L},\epsilon)/{\rm d}x_{L}\) in the collinear limit. The last remaining piece is \(\mathrm{d}\sigma^{[3]}_{\mathrm{E^{3}C}}(x_{L},\epsilon)/\mathrm{d}x_{L}\). The computation of the self-energy correlator is much easier since its dependence on \(x_{L}\) is factorized out by \(\delta(x_{L})\) and the integrals are simply standard cut integrals. The master integrals can be found in the literature, e.g. [94; 95]. Eventually adding \(\mathrm{d}\sigma^{[3]}_{\mathrm{E^{2}EC}}(x_{L})/\mathrm{d}x_{L}\) and \(\mathrm{d}\sigma^{[3]}_{\mathrm{E^{3}C}}(x_{L},\epsilon)/\mathrm{d}x_{L}\) together, we obtain the complete contact terms \(\mathrm{d}\sigma^{[3]}_{\mathrm{C}}(x_{L},\epsilon)/\mathrm{d}x_{L}\) for E3C distribution. The results are also summarized in Eq. (101)-(102). Combined with the nonidentical energy weight contributions, we find all \(\frac{1}{\epsilon}\) canceled and thus the infrared safety is guaranteed as expected. #### 2.4.3 Results of two-loop jet function constants With all individual contributions at hand, the full expressions of 2-loop E3Cs in the collinear limit can be written as \[\frac{1}{\sigma_{0}}\frac{\mathrm{d}\sigma^{[3],2\text{-loop}}_{ \mathrm{q}}}{\mathrm{d}x_{L}}= \,2\,\frac{\mathrm{d}J^{\text{nonid,2-loop}}_{q}}{\mathrm{d}x_{L} }+\frac{1}{\sigma_{0}}\frac{\mathrm{d}\sigma^{[3],2\text{-loop}}_{\mathrm{C },\mathrm{q}}}{\mathrm{d}x_{L}}\quad(e^{+}e^{-}\,\text{annihilation})\,, \tag{38}\] \[\frac{1}{\sigma_{0}^{\prime}}\frac{\mathrm{d}\sigma^{[3],2\text{- loop}}_{\mathrm{d}x_{L}}}= \,2\,\frac{\mathrm{d}J^{\text{nonid,2-loop}}_{g}}{\mathrm{d}x_{L} }+\frac{1}{\sigma_{0}^{\prime}}\frac{\mathrm{d}\sigma^{[3],2\text{-loop}}_{ \mathrm{C},\mathrm{g}}}{\mathrm{d}x_{L}}\quad(\text{gluonic Higgs decay})\,. \tag{39}\] Here a factor of 2 is added because we only consider a single jet in Sec. 2.4.1. Given the tree-level hard functions, \(\{H_{q}^{(0)},H_{g}^{(0)}\}=\{2\delta(1-x),0\}\) for \(e^{+}e^{-}\) annihilation and \(\{\tilde{H}_{q}^{(0)},\tilde{H}_{g}^{(0)}\}=\{0,2\delta(1-x)\}\) for the Higgs decay through the effective \(Hgg\) coupling, we can extract the two-loop jet constant directly from the \(\delta(x_{L})\) contribution from Eq. (38) and Eq. (39). We find that the \(\mu\) dependence are in full agreement with prediction from RG evolution, providing strong check to our calculation. The \(\mu\) independent part are the new results from this calculation. For the quark jet function, we get \[j_{2}^{q,[3]}=12.3020\,C_{F}T_{F}n_{f}-26.2764\,C_{A}C_{F}+21.3943\,C_{F}^{2}\,, \tag{40}\] and for gluon jet functions \[j_{2}^{g,[3]}=17.5487\,C_{A}T_{F}n_{f}-2.05342\,C_{F}T_{F}n_{f}-5.97991\,C_{A}^ {2}+0.904693\,n_{f}^{2}T_{F}^{2}\,. \tag{41}\] ### Perturbative resummation We start by defining the logarithmic order for our E3C resummation. The ingredients needed for our E3C resummation are summarized in Table 1. This includes the order of timelike splitting kernel \(\hat{P}(y)\), the boundary information (hard and jet constants), the \(\beta\) function for running coupling as well as the fixed-order matching.4 Due to the absent of analytic method to solve the RG equation exactly, we also truncate in the number of loops of the RGE solution to the desired logarithmic order [17]. Footnote 4: This is the same log counting as N\({}^{k}\)LL\({}^{\prime}\) in SCET, except that we omit all \({}^{\prime}\) for convenience. We first review the LL resummation in \(e^{+}e^{-}\) annihilation. Based on our resummation setting, it is safe to set \(x=1\) in the argument of E3C jet function in Eq. (10), which only affects the higher-order terms beyond LL. This leads to \[\frac{d\vec{J}_{\mathrm{LL}}^{[N]}(\ln\frac{x_{L}Q^{2}}{\mu^{2}})}{d\ln\mu^{2} }=\vec{J}_{\mathrm{LL}}^{[N]}(\ln\frac{x_{L}Q^{2}}{\mu^{2}})\cdot\frac{ \alpha_{s}}{4\pi}\int_{0}^{1}dy\,y^{N}\hat{P}^{(0)}(y)=-\vec{J}_{\mathrm{LL}}^ {[N]}(\ln\frac{x_{L}Q^{2}}{\mu^{2}})\cdot\frac{\alpha_{s}}{4\pi}\gamma_{T}^{(0 )}(N+1)\,. \tag{42}\] Here, we introduce the anomalous dimension to be the moment of timelike splitting kernel \[\gamma_{T}(N)\equiv-\int_{0}^{1}dy\,y^{N}\hat{P}(y)=\left(\frac{\alpha_{s}}{4\pi} \right)\gamma_{T}^{(0)}+\left(\frac{\alpha_{s}}{4\pi}\right)^{2}\gamma_{T}^{(1 )}+\cdots\,. \tag{43}\] Then given the boundary condition \(\vec{J}^{(0)}=\{2^{-N},2^{-N}\}\), we can directly write down the solution to LL jet function: \[\vec{J}^{[N]}_{\rm LL}=2^{-N}(1,1)\cdot\exp\left[-\frac{\gamma_{T}^{(0)}}{ \beta_{0}}\ln\frac{\alpha_{s}\left(\sqrt{x_{L}}Q\right)}{\alpha_{s}(\mu)} \right]\,. \tag{44}\] Plugging both jet and hard functions into the factorization for the cumulant \(\Sigma^{[N]}\) and differentiating it with respect to \(x_{L}\), we obtain the LL resummed physical spectrum for E3C. Beyond LL, the \(x=1\) approximation is no longer valid, and instead we have to solve the jet RGE directly. While it is difficult to obtain a close-form solution for this modified DGLAP equation, we find that a truncated solution in \(\alpha_{s}\) is already in good convergence. Explicitly, we assume the jet function takes the form \[\vec{J}^{[N]}=\underbrace{\sum_{i=1}^{\infty}\alpha_{s}^{i}L^{i}\vec{c}_{i,i} }_{\rm LL}+\underbrace{\sum_{i=1}^{\infty}\alpha_{s}^{i}L^{i-1}\vec{c}_{i,i- 1}}_{\rm NLL}+\underbrace{\sum_{i=1}^{\infty}\alpha_{s}^{i}L^{i-2}\vec{c}_{i, i-2}}_{\rm NNLL}+\cdots\,, \tag{45}\] with \(L\equiv\ln\frac{x_{L}Q^{2}}{\mu^{2}}\) and \(c_{i,j}\) unknown constants, and solve both the jet RGE and \(\beta\) RGE order by order in \(\alpha_{s}\) (which is referred as expanded solution). In practice, we evaluate it numerically up to \(\mathcal{O}(\alpha_{s}^{50})\). Another advantage of using expanded solution is that we only need certain moments of the hard functions. For example, consider one term from the jet function, \(\vec{J}^{[N]}\supset\alpha_{s}^{2}\vec{c}_{2,2}L^{2}\), and plug into Eq. (3), we find \[\Sigma^{[N]}\supset\alpha_{s}^{2}\vec{c}_{2,2}\cdot\int_{0}^{1}dx \,x^{N}\ln^{2}\left(\frac{x_{L}x^{2}Q^{2}}{\mu^{2}}\right)\cdot\vec{H}_{ee} \left(x,\ln\frac{Q^{2}}{\mu^{2}}\right)\] \[=\alpha_{s}^{2}\vec{c}_{2,2}\cdot\left[\ln^{2}\left(\frac{x_{L}Q^ {2}}{\mu^{2}}\right)^{2}\int_{0}^{1}dx\,x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{ 2}}{\mu^{2}}\right)\right.\] \[+2\ln\left(\frac{x_{L}Q^{2}}{\mu^{2}}\right)\int_{0}^{1}dx\,\ln x ^{2}x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{\mu^{2}}\right)+\int_{0}^{1}dx\, \ln^{2}x^{2}x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{\mu^{2}}\right)\bigg{]}\] \[=\alpha_{s}^{2}\vec{c}_{2,2}\cdot\left[\ln^{2}\left(\frac{x_{L}Q^ {2}}{\mu^{2}}\right)+2\ln\left(\frac{x_{L}Q^{2}}{\mu^{2}}\right)\partial_{N} +4\partial_{N}^{2}\right]\int_{0}^{1}x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{ \mu^{2}}\right)\,, \tag{46}\] \begin{table} \begin{tabular}{|c|c|c|c|c|} \hline resummation order & \(\hat{P}(y)\) & \(\vec{H}\), \(\vec{J}\) constants & \(\beta[\alpha_{s}]\) & fixed-order matching \\ \hline LL & tree & tree & 1-loop & LO \\ \hline NLL & 1-loop & 1-loop & 2-loop & NLO \\ \hline NNLL & 2-loop & 2-loop & 3-loop & NNLO \\ \hline \end{tabular} \end{table} Table 1: Definition of the resummation order and their corresponding fixed-order matching. where the three terms correspond to the standard moment, the single logarithmic moment and the double logarithmic moment of the E3C hard function. To derive the last line, we also use the following relation \[\int_{0}^{1}\ln^{k}x^{2}x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{\mu^{2}}\right) =2^{k}\partial_{N}^{k}\int_{0}^{1}x^{N}\vec{H}_{ee}\left(x,\ln\frac{Q^{2}}{\mu^ {2}}\right)\,. \tag{47}\] In the Appendix A, we provide all the hard moments with \(N=2,3\) that are required for NNLL resummation. In this paper, we present results for the NNLL resummation of E3C for \(e^{+}e^{-}\) annihilation, and approximate NNLL resummation for jets from the hadronic collision process \(pp\to jj\). For \(e^{+}e^{-}\) annihilation, we have all ingredients needed for NNLL resummation. And since there is no accurate fixed-order data for E3C at NNLO, we will instead match the NNLL result to NLO. Regarding the dijet production, due to the absence of the two-loop hard constant, we will present the approximate NNLL resummation (which we refer as NNLL\({}_{\rm approx}\)), with an additional uncertainty coming from the missing two-loop hard constant. Resummation with the accurate two-loop hard function as well as the matching with fixed-order result are left as future improvements. ## 3 NNLL resummation in \(e^{+}e^{-}\) annihilation With all the ingredients at hand, now we can present the NNLL resummation prediction. In this section, we first consider \(e^{+}e^{-}\) collision at two different energies: 250 GeV and 1 TeV. In the resummation calculation, we will use \(\alpha(m_{Z})=0.118\). ### Resummation results Following the discussion in Sec. 2.5, our resummation is performed by perturbatively solving the jet function RG equation to order \(\mathcal{O}(\alpha_{s}^{50})\), plugging back to the cumulant factorization and finally truncating the logarithms \(\ln\frac{x_{L}Q^{2}}{\mu^{2}}\) to the desired order. In the resummation formula, we set canonical jet scale \(\mu_{j}=\mu_{h}\sqrt{x_{L}}\) in the factorization, leaving a single hard scale \(\mu_{h}=\mu\) in the resummed expression. We vary the scale \(\mu\) to estimate the uncertainty from higher order corrections. Regarding the observables, below we consider three cases: \(N=2\), \(N=3\) and their ratio. The \(N=2\) case is precisely the EEC observable, where we directly use the result from Ref. [17], and the singular expansion has been verified against the NLO EEC fixed-order calculation. For \(N=3\) case, this is the main result of this paper. In Fig. 2, we first check our \(\mathcal{O}(\alpha_{s}^{2})\) expansion with the Monte Carlo program Event2. In the collinear limit, we find excellent agreement between theory and numeric result, while in the meantime, this also suggests the non-singular contribution from fixed-order calculation is negligible in this limit. Nevertheless, the matching formula can be written as \[\frac{d\sigma^{\rm match}}{dx_{L}}=\frac{d\sigma^{\rm resum}}{dx_{L}}-\frac{ d\sigma^{\rm sing}}{dx_{L}}+\frac{d\sigma^{\rm FO}}{dx_{L}}\,. \tag{48}\] Here each term is a function of \(\alpha_{s}(\mu)\) evaluated at the hard scale \(\mu_{h}=\mu\). In Fig. 3, we present the E3C resummation up to NNLL, matched to fixed-order. As explained above, due to the absence of NNLO data, we only match NNLL to NLO. The hard scale is chosen to be half of the center-of-mass energy \(\mu=Q_{\rm jet}\equiv Q/2\), the typical energy for each quark jet, and the scale uncertainty is obtained by varying the hard scale by a factor of 2. In both energies, the uncertainty band width goes down as we increase the resummation order, while at 1 TeV, we have a tighter band because the coupling \(\alpha_{s}\) runs slower at high energy. At NNLL, we find a relative 4% hard uncertainty for \(Q=250\) GeV and 2% for \(Q=1\) TeV. We find large corrections as we go from LL to NNLL, as was also observed previously in [17], which emphasize the importance of higher order corrections. For higher center-of-mass energy, the convergence between different orders is improved. To improve the convergence, we also introduce the ratio of different point energy correlators, namely [16] \[\Delta_{m,n}(x_{L},\mu,\mu^{\prime})\equiv\frac{d\sigma^{[m]}/dx_{L}}{d\sigma^ {[n]}/dx_{L}},\quad m,n\geq 2\,, \tag{3.2}\] where \(\mu\) and \(\mu^{\prime}\) are the hard scale in \(\frac{d\sigma^{[m]}}{dx_{L}}\) and \(\frac{d\sigma^{[n]}}{dx_{L}}\) respectively. In particular, we focus on the ratio between fully matched E3C and EEC, i.e. \(\Delta_{3,2}(x_{L})\). In Fig. 4, we show the NNLL resummed \(\Delta_{3,2}(x_{L})\) at again \(Q=250\) GeV and \(Q=1\) TeV, and find good convergence. This implies that the ratio can be used as precision observable. For hard scale uncertainty, we use the seven-point scale variation, which amounts to varying the scales in both numerator and denominator independently by a factor of 2, to a combination of \[\left(\frac{\mu}{Q_{\rm jet}},\frac{\mu^{\prime}}{Q_{\rm jet}}\right)\in \left\{\left(\frac{1}{2},\frac{1}{2}\right)\text{, }\left(2,2\right)\text{, }\left(1,2\right)\text{, }\left(1,1\right)\text{, }\left(2,1\right)\text{, }\left(1,\frac{1}{2}\right)\text{, }\left(\frac{1}{2},1\right)\right\}, \tag{3.3}\] and take the envelope as the uncertainty estimation. The convergence also indicates that ENC shares similar non-perturbative behavior in the collinear limit and taking the ratio strongly suppresses the power corrections. Figure 2: The comparison of fixed-order result and the singular expansion from resummation. The left panel shows good agreement between LO expression and the \(\mathcal{O}(\alpha_{s})\) singular expansion of E3C resummed result. The difference between them is the non-singular result. The right panel gives the corresponding contributions at NLO, with the numerical full NLO prediction from Event2. ### Hadronization corrections In this subsection, we consider the power-suppressed hadronization corrections in the collinear limit. At present hadronization corrections cannot be computed from first principle. For Figure 4: The ratio between resummed E3C and EEC distribution \(\Delta_{3,2}(x_{L})\) up to NNLL+NLO for \(e^{+}e^{-}\) collision at 250 GeV (left top panel) and 1 TeV (right top panel). Uncertainty bands are obtained in the seven-point scale variation scheme, namely varying the scales \(\mu\) and \(\mu^{\prime}\) in the two factors independently by a factor of 2 around the central value. The lower panels show the relative scale uncertainty of the NNLL+NLO distribution around the central value. Figure 3: The resummed E3C distribution up to NNLL+NLO, multiplied by a factor \(x_{L}(1-x_{L})\), for \(e^{+}e^{-}\) collision at 250 GeV (left top panel) and 1 TeV (right top panel). Uncertainty bands are obtained by varying the hard scale \(\mu\) around the nominal value \(Q_{\rm jet}=Q/2\) by a factor of 2. The lower panels show the relative scale uncertainty of the NNLL+NLO distribution around the central value. simplicity, we use a phenomenological form for the leading non-perturbative power correction as suggested in [96], and fit the unknown parameters from a Monte Carlo program. This provides some insights on how to model the hadronization effect for a global fit in the future. In general, the non-perturbative corrections in infrared-collinear safe observables are (at least) suppressed as \(\Lambda_{\rm QCD}/Q\) to some power, where \(Q\) is the hard scale of the process. Following from the LL result in Eq. (44), we observe that in the collinear limit, there exists a lower scale \(\sqrt{x_{L}}Q\) in the coupling, and the most important non-perturbative correction that could potentially appear is linear in \(\Lambda_{\rm QCD}\) and takes the form \(\Lambda_{\rm QCD}/(\sqrt{x_{L}}Q)\), multiplied with an extra kinematic factor \(1/x_{L}\). The sub-leading non-perturbative corrections with additional powers of \(\Lambda_{\rm QCD}/(\sqrt{x_{L}}Q)\) will become necessary down to small \(x_{L}\sim\Lambda_{\rm QCD}^{2}/Q^{2}\), where the perturbation theory also breaks down. For the leading non-perturbative correction we are considering, such structure is in fact recovered for the EEC in the fragmentation modeling of non-perturbative radiations [1] and and analysis using renormalon or dispersive techniques [96; 97; 98]. As a qualitative analysis, we use the following parametrization of the leading non-perturbative correction, \[\frac{d\sigma^{\rm NP-soft}}{dx_{L}}=\frac{1}{x_{L}}\cdot\left(\,\frac{ \tilde{\Lambda}}{\sqrt{x_{L}}\,Q}\,\right)^{1+\gamma}\qquad(\,\text{soft fragmentation}\,), \tag{45}\] we verify the scaling behaviour of the non-perturbative correction in the collinear limit for both EEC and E3C distributions with Pythia8[99], and extract the non-perturbative parameters by fitting from the difference of the hadron level and parton level predictions. Note that the issues of extracting the non-perturbative power corrections from Monte Carlo generators have been pointed out in Ref. [36]. In particular, the corrections from the hadronization modeling in the Monte Carlo programs in fact unfaithfully absorb partial subleading-log contributions, as the hadronization modeling has been tuned to reproduce some collider data with limited perturbative accuracy. Therefore, in this paper we only use Monte Carlo to illustrate the impact of power correction for individual EEC and E3C distribution as well as their ratio. For our case, we stay in the default settings of Pythia8 and obtain the following fit at the 95% confidence level. At \(Q=250\) GeV, we find for EEC and E3C: \[\tilde{\Lambda}_{2} =(0.956\pm 0.031)\,\text{GeV}\,,\quad\gamma_{2}=0.462\pm 0.017\,,\] \[\tilde{\Lambda}_{3} =(0.500\pm 0.040)\,\text{GeV}\,,\quad\gamma_{3}=0.335\pm 0.031\,. \tag{46}\] And in the case with \(Q=1\) TeV, we have \[\tilde{\Lambda}_{2} =(0.775\pm 0.013)\,\text{GeV}\,,\quad\gamma_{2}=0.383\pm 0.008\,,\] \[\tilde{\Lambda}_{3} =(0.435\pm 0.015)\,\text{GeV}\,,\quad\gamma_{3}=0.325\pm 0.012\,. \tag{47}\] We emphasis that for too small \(x_{L}\) value, the leading order non-perturbative approximation itself becomes invalidated. The enhancement of the non-perturbative corrections in the collinear limit must be turned off before entering the fully non-perturbative phase, where the degrees of freedom become freely interacting hadrons and a nice scaling behavior follows [20]. In this qualitative analysis, we choose the lower bound of the fit range by finding the extreme point of the distributions from hadron level prediction in Pythia8. Multiplying the extreme point by a factor of 2 gives a good estimate of the lower bound for the range where the non-perturbative correction follows the described scaling behavior. In Fig. 5, we show the relative hadronization correction from both Pythia8 and our two-parameter fit. Except the shaded region, our parameterization agrees with the Monte Carlo result and it is sufficient for understanding their structure. In Fig. 6, we include the non-perturbative correction in the matched E3C resummation, which strongly enhances the extreme collinear limit. At \(Q=1\) TeV, the non-perturbative correction changes our NNLL+NLO prediction by only a few percent at \(x_{L}\sim 0.1\), while this modification reaches 50% at \(x_{L}\sim 10^{-4}\). This shows that the non-perturbative corrections for energy correlators, though being power suppressed at high energies, can become sizable even at the energy level of future \(e^{+}e^{-}\) colliders. However, since EEC and E3C share a close power law in the leading power correction, the enhancement is significantly canceled when considering their ratio \(\Delta_{3,2}(x_{L})\). As shown in Fig. 7, the leading non-perturbative correction only gives rise to roughly 4% effect at \(Q=250\) GeV and 2% at \(Q=1\) TeV for matched NNLL. This confirms that \(\Delta_{3,2}(x_{L})\) is insensitive to the hadronization and indeed Figure 5: Comparison of Pythia8 result and the fitting using Eq. (14). The blue curves are the difference of hadron-level and parton-level distribution for EEC and E3C, at both 250 GeV and 1 TeV. The red curves are our fitting result with parameters from Eq. (15) and (16). The shaded region stands for the range where the parametrization of the leading non-perturbative correction is no longer valid and should be excluded from the fit range. a good candidate for precise \(\alpha_{s}\) measurement. We also investigate the impact on the final resummation results caused by the uncertainties from the two-parameter fit. The statistical error for both \(\tilde{\Lambda}\) and \(\gamma\) are given in Eq. (3.5) and (3.6). Fig. 8 shows the final uncertainty in the matched NNLL distribution from varying these two NP parameters. In both \(Q=250\) GeV and \(Q=1\) TeV, excluding Figure 6: The E3C distribution to NNLL+NLO including the non-perturbative (NP) hadronization corrections estimated with Pythia8 data, with different plot ranges for collision energies at 250 GeV (left panel) and 1 TeV (right panel). Figure 7: The E3C/EEC ratio to NNLL+NLO with non-perturbative (NP) hadronization corrections. The non-perturbative hadronization corrections are estimated as above by \(\tilde{\Lambda}/(x_{L}^{3/2}Q)\) for both the EEC and E3C distributions with the coefficients fitted from Pythia8. the shaded region, the NP uncertainty is much smaller than the hard uncertainty estimated by seven-point variation. In particular, at \(Q=1\) TeV, the NP uncertainty is reduced to 1% in the potential fit region. Despite that, we admit that the effect of non-perturbative corrections turns to increase for such small \(x_{L}\) region, and more accurate understanding of the non-perturbative corrections will be required to further improve the precision. ### Anticipation of \(\alpha_{s}\) determination In this subsection, we discuss the potential of extracting the strong coupling constant \(\alpha_{s}\) from measuring the resumed E3C/EEC ratio \(\Delta_{3,2}(x_{L})\). In literature [100; 101], the back-to-back limit of EEC is resummed to NNLL+NLO and has been use for \(\alpha_{s}\) measurement from \(e^{+}e^{-}\) data. Similar to other event shapes, the non-perturbative correction is significantly large in this region and require careful modeling. And how we profile the resummation and power correction has a sizable effect on the final theory uncertainty. Alternatively, we can also do the \(\alpha_{s}\) measurement _only_ in the collinear limit. First of all, as we discussed in Sec. 3.1, the non-singular contribution is almost zero in this limit, and thus it is safe to ignore the higher fixed-order contribution. Secondly, by considering the ratio distribution, \(\Delta_{m,n}(x_{L})\), the suppressed power corrections will lead to a smaller theory uncertainty and thus more precise \(\alpha_{s}\) determination. As illustration, we first investigate the sensitivity of \(\Delta_{3,2}(x_{L})\) when slightly changing the value of \(\alpha_{s}\). In particular, we vary the value of strong coupling at \(Z\)-pole \(\alpha_{s}(m_{Z})\) by a factor of 5%, namely \(\alpha_{s}(m_{Z})=\{0.112,0.118,0.124\}\) and compare the effect on matched resummation result. We first consider the NNLL+NLO \(\Delta_{3,2}(x_{L})\) at \(Q=91.2\) GeV with all three values of \(\alpha_{s}(m_{Z})\). As observed in Fig. 9, the slope become sensitive to the \(\alpha_{s}\) in the collinear region \(x_{L}=10^{-3}\sim 10^{-4}\), while the relative difference with respect to \(\alpha_{s}(m_{Z})=0.118\) ranges from Figure 8: The uncertainty from varying the non-perturbative parameters \(\tilde{\Lambda}\) and \(\gamma\) in the resummed ratio distribution \(\Delta_{3,2}(x_{L})\). The bottom panels stand for the relative difference of NNLL+NLO+NP with respect to its central value. 10% to 20%. The slope sensitivity and the cancellation of hadronization correction have made the ratio of E3C and EEC \(\Delta_{3,2}(x_{L})\) an advantageous observable for extracting the \(\alpha_{s}\) from \(e^{+}e^{-}\) annihilation. Similar behaviors also exist at other energies and for completeness, we present the comparison at \(Q=250\) GeV and \(Q=1\) TeV in Fig. 10. The fact that the resummed E3C/EEC ratio has larger sensitivity to \(\alpha_{s}\) and reduced non-perturbative corrections in the collinear limit makes it a promising candidate for the \(\alpha_{s}\) determination. To further improve the \(\alpha_{s}\) determination requires improving the resummation accuracy, matching with NNLO fixed-order correction, as well as the non-perturbative modeling. ## 4 Approximate NNLL resummation in \(pp\) collisions In this section, we consider the dijet production \(pp\to jj\) at the LHC. There are several motivations to study energy correlators in \(pp\) collisions. First of all, LHC provides unique opportunities to study energy flows correlation in QCD at extremely high energy. While the LEP or future CEPC provides a very clean environment for precise measurements, \(pp\) collisions at the LHC can produce multiple jets with very high energies (\(p_{T}\gtrsim 500\) GeV), and high angular resolution can be achieved to probe the underlying dynamics for their formation and evolution. Secondly, as we have observed in the \(e^{+}e^{-}\) collisions, the non-perturbative corrections for ENC have a relatively simple form compared to other event shape observables (at least in leading power), which might be easier to study non-perturbative QCD. At the same time, with multiple scales involved, \(pp\) collision can provide robust data from high energy to low energy, which is beneficial for understanding non-perturbative effects. In this section, we still focus on improving the perturbative predictions for ENC. As in Sec. 2, the jet functions are universal across different hard processes and the new ingredients are the moments of \(pp\) hard function, both regular and logarithmic. The main complication Figure 9: Left panel is the matched NNLL ratio distribution \(\Delta_{3,2}(x_{L})\) with different strong coupling constants: \(\alpha_{s}(m_{Z})=\{0.112,0.118,0.124\}\) at \(Q=91.2\) GeV. The uncertainty is the hard scale variation. Right panel shows the relative deviation of all three bands with respect to the central prediction of \(\alpha_{s}(m_{Z})=0.118\). for \(pp\) collision is that the hard function now involves convolution with PDFs and algorithmic definition of jet, allowing only numeric calculation of the hard function. For the numerical calculation of the hard function, we adopt the anti-\(k_{t}\) jet algorithm and choose the jet radius to be \(R_{0}=0.4\). The complete kinematic cuts are summarized in Eqs. (8)-(9). The \(\mu\) independent part of the NLO hard function are presented in Appendix. A. We observes large corrections going from LO to NLO. The \(\mu\) dependent part of the NNLO hard function can be derived using the RG equation in (6). The \(\mu\) independent part requires a genuine two-loop corrections and are beyond the scope of this work. Instead we make a simple estimate of the two-loop constant terms, and dubbed the resulting prediction approximate NNLL resummation (NNLL\({}_{\rm approx}\)). Specifically, we use a modified Pade approximation to estimate the two-loop hard function constants in both quark channel and gluon channel: \[a_{s}^{2}h_{0}^{(2)}\approx\kappa\frac{(a_{s}h_{0}^{(1)})^{2}}{h_{0}^{(0)}}\,, \tag{10}\] where we vary \(\kappa\) in the range \([0,1/2]\) as a naive way to estimate our theory uncertainties on the missing two-loop constants. For the splitting function, \(\beta\) function, as well as the jet Figure 10: Left panels are the matched NNLL ratio distribution \(\Delta_{3,2}(x_{L})\) with different strong coupling constants: \(\alpha_{s}(m_{Z})=\{0.112,0.118,0.124\}\) at \(Q=250\) GeV and \(Q=1\) TeV. Right panels are the relative deviation of all three bands with respect to the central prediction of \(\alpha_{s}(m_{Z})=0.118\). functions, we used the ones required by NNLL accuracy as shown in Table 1. In Fig. 11, we show the E3C/EEC ratio \(\Delta_{3,2}(R_{L})\) up to NNLL\({}_{\rm approx}\), with the hard uncertainty estimated by seven-point variation. Due to the lack of knowledge of the genuine two-loop hard function moment, we have chosen to normalize the E3C/EEC distribution in the range of \(R_{L}\in[0.01,0.4]\) to reduce the impact from not knowing the full two-loop hard function. We find good convergence for both \(p_{t}\) ranges: \([300,350]\) GeV and \([500,550]\) GeV. In the future, it would be interesting the compute the two-loop hard function, as well as match the resummed results to fixed order to improve the prediction around \(R_{L}\sim R_{0}\). ### Anticipation of \(\alpha_{s}\) determination Similar to \(e^{+}e^{-}\) annihilation, in this subsection we discuss the potential of extracting the strong coupling constant \(\alpha_{s}\) from the resummed \(\Delta_{3,2}(R_{L})\) distribution in \(pp\to jj\). In particular, we also investigate the slope sensitivity of the distribution with respect to different values of \(\alpha_{s}\). For hadron colliders, we need to change the PDFs as we vary the strong coupling among \(\alpha_{s}(m_{Z})=0.118\pm 0.06\). For this purpose, we use three PDF sets: NNPDF31_nnlo_as_0112, NNPDF31_nnlo_as_0118 and NNPDF31_nnlo_as_0124 when calculating the hard function using the method in [54]. As shown in Fig. 12, for each \(p_{t}\) range, the uncertainty is significantly reduced from NLL to NNLL\({}_{\rm approx}\), leading to distinguishable slopes with respect to different \(\alpha_{s}\). This suggests that ratios of energy correlators are good candidate for extracting \(\alpha_{s}\). We note that there is larger slope variation for lower \(p_{t}\) of the jet, in agreement with the expectation that the measurement at lower energy is more sensitive to \(\alpha_{s}\) due to asymptotic free nature of QCD. ## 5 Conclusion In this paper we have performed a systematic study of resummation of projected three-point energy correlator E3C [16], and its ratio to EEC, in both \(e^{+}e^{-}\) collider and \(pp\) collider. We Figure 11: Normalized E3C/EEC ratio \(\Delta_{3,2}(R_{L})\) for \(pp\to jj\) with \(\alpha_{s}(m_{Z})=0.118\), and jet \(p_{t}\) ranges \([300,350]\) GeV (left) and \([500,550]\) GeV (right). Uncertainty bands are obtained in the 7-point scale variation scheme, with additional uncertainty from varying the estimation of NNLO hard function constant for NNLL\({}_{\rm approx}\). have achieved the first NNLL accuracy for the \(e^{+}e^{-}\) case, and NNLL\({}_{\rm approx}\) accuracy for the \(pp\) case. Our results show that good perturbative convergence can be achieved for the ratios of projected energy correlators. The current theoretical uncertainties are at a level of a few percent, and can be further improved in the future when the higher order ingredients become available. We have also shown that the ratio observable is sensitive to variation of \(\alpha_{s}\), therefore provides a good candidate for precision \(\alpha_{s}\) determination using jet substructure. To achieve the above theory accuracy, one of the main new ingredients is the two-loop E3C jet function computed in this work. The calculation includes three pieces: double-real, real-virtual and double-virtual. The last two contributions only involve a single \(\delta\) measure Figure 12: Normalized NLL (upper) and NNLL\({}_{\rm approx}\) (lower) resummation result for E3C/EEC ratio at \(pp\to jj\) with \(\alpha_{s}(m_{Z})=0.118\pm 0.06\), i.e., varied by about 5%, for two different jet \(p_{t}\) ranges [300,350] GeV and [500,550] GeV. Lower panels show the relative difference from the result at the central scale with \(\alpha_{s}(m_{Z})=0.118\). ment function in the phase space integral and share a similar form as the analytic EEC calculation at NLO [5]. Regarding the double-real emissions, which amounts to integrating the fully-differential EEEC distribution within the collinear kinematic space, we used two different approaches and find the same results. The first method is to subtract the infrared divergence in the collinear EEEC jet function, integrate it separately with \(d\)-dimension kinematic space, and expand the finite terms in \(\epsilon\). The second approach benefits from the recently developed parametric IBP, where we can also simplify the integrand with IBP reduction and calculate the integrals via differential equations. Regarding the ENC resummation, for \(e^{+}e^{-}\) annihilation, we solve the E3C jet RGE (which is a modified DGLAP equation) order by order in \(\alpha_{s}\) with the two-loop boundary, and push the resummation up to NNLL. For \(pp\) collisions, we calculate the combined hard function moments using the method in [54] for dijet production. We present the complete NLL and the approximate NNLL resummation result, where the approximation is due to the missing of genuine two-loop hard function constant. The uncertainty is reduced compared with the previous results [16; 20; 21]. For the fixed-order matching, we notice that the singular contribution dominates the collinear limit and the non-singular contribution from matching has only small effects in the \(e^{+}e^{-}\) case. Nevertheless, we perform the matching for \(e^{+}e^{-}\) given the fixed-order result is already available, but leave the matching with fixed-order in the \(pp\) case for the future study. For a complete phenomenological analysis and precise \(\alpha_{s}\) extraction at hadron collider, there are still several ingredients needed in the future. Perturbatively, we need to compute both two-loop hard function and the NLO non-singular distribution for \(pp\to jj\), in order to achieve a full NNLL story. More over, it would be interesting to solve the RG equation exactly following [102], and compare the results with the truncation method. At the same time, for both \(e^{+}e^{-}\) and \(pp\), it would be interesting to better understand the hadronization power corrections to help further reduce theoretical uncertainties. We hope that all these efforts can lead to a precision determination of \(\alpha_{s}\) from jet substructure in the future. The authors thank Hao Chen, Kyle Lee, Meng Xiao, Tong-Zhi Yang, Yulei Ye for useful discussions. XYZ also thanks the MIT CTP for its hospitality while part of this work was performed. The work of WC, YL, ZX, and HXZ was supported by the National Natural Science Foundation of China under the Grant No. 11975200. The work of JG was sponsored by the National Natural Science Foundation of China under the Grant No.12275173 and No.11835005. ## Appendix A Hard and jet functions ### \(e^{+}e^{-}\) Hard function The ENC hard function for \(e^{+}e^{-}\) can be obtained from the semi-inclusive hadron fragmentation function. At NNLL, following our resummation procedure, we need the regular up to two-loop, single logarithmic up to one-loop and the double logarithmic moments at tree level with respect to the energy fraction \(x\): \[\int_{0}^{1}dx\,x^{N}\,H_{q,g}(x,\mu=Q) = \sum_{L=0}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}h_{L}^{q,g}(N)\,,\] \[\int_{0}^{1}dx\,x^{N}\,\ln x\,H_{q,g}(x,\mu=Q) = \sum_{L=1}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}\dot{h} _{L}^{q,g}(N)\,,\] \[\int_{0}^{1}dx\,x^{N}\,\ln^{2}x\,H_{q,g}(x,\mu=Q) = \sum_{L=1}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L}\ddot{h }_{L}^{q,g}(N)\,. \tag{100}\] For EEC (\(N=2\)), we have \[h_{0}^{q} =2\,,\qquad h_{0}^{g}=0\,,\qquad\qquad h_{1}^{q}=\frac{131}{4}\,C _{F}\,,\qquad h_{1}^{g}=-\frac{71}{12}\,C_{F}\,,\] \[h_{2}^{q} =\left(64\zeta_{4}-\frac{1172}{3}\zeta_{3}-166\zeta_{2}+\frac{2 386397}{2592}\right)C_{A}C_{F}\] \[+\left(-128\zeta_{4}+\frac{1016}{3}\zeta_{3}+\frac{1751}{18} \zeta_{2}-\frac{1105289}{5184}\right)C_{F}^{2}+\left(32\zeta_{3}+\frac{118}{15 }\zeta_{2}-\frac{8530817}{54000}\right)C_{F}T_{F}n_{f}\,,\] \[h_{2}^{g} =\left(-\frac{76}{3}\zeta_{3}+\frac{188}{45}\zeta_{2}-\frac{2980 2739}{324000}\right)C_{A}C_{F}+\left(\frac{124}{3}\zeta_{3}+\frac{523}{18} \zeta_{2}-\frac{674045}{5184}\right)C_{F}^{2}\,,\] \[\dot{h}_{0}^{q} =0\,,\qquad\dot{h}_{1}^{q}=\left(40\zeta_{3}+\frac{61}{3}\zeta_{ 2}-\frac{5303}{72}\right)C_{F}\,,\qquad\dot{h}_{0}^{g}=0\,,\qquad\dot{h}_{1}^{ g}=\left(-\frac{7}{3}\zeta_{2}+\frac{31}{4}\right)C_{F}\,,\] \[\ddot{h}_{0}^{q} =0\,,\qquad\ddot{h}_{0}^{g}=0\,. \tag{101}\] Note that the EEC hard moments are also summarized in the appendix of Ref. [17]). However, the normalization condition in [17] is different from ours, due to the scaled energy \(E_{i}/(Q/2)\) there in contrast with \(E_{i}/Q\) here in the definition of the jet function. For E3C (\(N=3\)), we find \[h_{0}^{q} = 2,\qquad h_{0}^{g}=0,\qquad h_{1}^{q}=\frac{11909}{300}C_{F}, \qquad h_{1}^{g}=-\frac{547}{150}C_{F}\,,\] \[h_{2}^{q} = \left(-\frac{942}{5}\zeta_{3}-\frac{17}{45}\zeta_{2}+\frac{17147 309}{32400}\right)C_{A}C_{F}+\left(32\zeta_{3}+\frac{322}{25}\zeta_{2}-\frac{ 6169957}{30000}\right)C_{F}n_{f}T_{F}\] \[+\left(-\frac{2012}{15}\zeta_{3}-\frac{8987}{30}\zeta_{2}+\frac{ 3256506739}{3240000}\right)C_{F}^{2}\,,\] \[h_{2}^{g} = \left(\frac{52}{5}\zeta_{3}+\frac{4396}{225}\zeta_{2}-\frac{1017 63773}{810000}\right)C_{A}C_{F}+\left(\frac{392}{15}\zeta_{3}+\frac{397}{15} \zeta_{2}-\frac{163115357}{1620000}\right)C_{F}^{2}\,,\] \[\dot{h}_{0}^{q} = 0\,,\qquad\dot{h}_{1}^{q}=\left(40\zeta_{3}+\frac{337}{15}\zeta_ {2}-\frac{709693}{9000}\right)C_{F}\,,\] \[\dot{h}_{0}^{g} = 0\,,\qquad\dot{h}_{1}^{g}=\left(-\frac{22}{15}\zeta_{2}+\frac{167 39}{4500}\right)C_{F}\,,\] \[\ddot{h}_{0}^{q} = 0\,,\qquad\ddot{h}_{0}^{g}=0\,. \tag{102}\] For completeness, we also provide the E3C (\(N=3\)) hard moments for the gluonic Higgs decay, which is needed for extracting the two-loop gluon jet constants. Here we use \(\hat{h}\) to distinguish from the \(e^{+}e^{-}\) case. \[\tilde{h}_{0}^{q} = 0,\qquad\tilde{h}_{0}^{g}=2,\qquad\tilde{h}_{1}^{q}=-\frac{2461}{45 0}n_{f}T_{F},\qquad\tilde{h}_{1}^{g}=\frac{11491}{150}C_{A}-\frac{494}{45}n_{f}T _{F}\,,\] \[\tilde{h}_{2}^{q} = n_{f}T_{F}\left[C_{A}\left(\frac{88}{3}\zeta_{3}+\frac{3428}{75} \zeta_{2}-\frac{219509243}{810000}\right)+\left(\frac{1727}{225}\zeta_{2}-\frac {187858397}{1620000}\right)C_{F}\right]\] \[+\left(-\frac{352}{45}\zeta_{2}+\frac{7224}{125}\right)n_{f}^{2} T_{F}^{2}\,,\] \[\tilde{h}_{2}^{g} = n_{f}T_{F}\left[C_{A}\left(-\frac{208}{3}\zeta_{3}+\frac{1264}{ 15}\zeta_{2}-\frac{38190113}{40500}\right)+C_{F}\left(96\zeta_{3}-\frac{242}{2 25}\zeta_{2}-\frac{113165189}{810000}\right)\right]\] \[+C_{A}^{2}\left(-388\zeta_{3}-\frac{31684}{75}\zeta_{2}+\frac{83 7482633}{270000}\right)+n_{f}^{2}T_{F}^{2}\left(-\frac{64}{9}\zeta_{2}+\frac{ 44252}{675}\right)\,,\] \[\dot{\tilde{h}}_{0}^{q} = 0\,,\qquad\dot{\tilde{h}}_{1}^{q}=n_{f}T_{F}\left(-\frac{22}{15} \zeta_{2}+\frac{404}{125}\right)\,,\] \[\dot{\tilde{h}}_{0}^{g} = 0\,,\qquad\dot{\tilde{h}}_{1}^{g}=C_{A}\left(40\zeta_{3}+\frac{346 }{15}\zeta_{2}-\frac{2134817}{27000}\right)+\left(-\frac{8}{3}\zeta_{2}+\frac {5369}{1350}\right)n_{f}T_{F}\,,\] \[\ddot{\tilde{h}}_{0}^{q} = 0\,,\qquad\ddot{\tilde{h}}_{0}^{g}=0\,. \tag{100}\] \(pp\to jj\) **Hard function** The following table gives the hard function moments for \(pp\to jj\) calculated in Madgraph5 in two different \(p_{t}\) ranges: \([300,350]\) GeV and \([500,550]\) GeV, needed for the resummation of both EEC (\(N=2\)) and E3C (\(N=3\)). As one of the important checks of our calculation, we show in Fig. 13 the independence of the slicing parameter \(\delta_{\rm cut}\) when evaluating the hard function moments using the method in [54]. The values of the moments are in agreement within the numeric uncertainty for three values of \(\delta_{\rm cut}\) across two orders of magnitude, namely \(\delta_{\rm cut}\in\{0.003,0.03,0.3\}\). \begin{table} \begin{tabular}{||c c c c c c c||} \hline \multicolumn{6}{||c||}{\(pp\to jj\) at 13 TeV, with NNPDF31\_nnlo\_as\_0118} \\ \hline \hline (300,350) GeV & \(h_{0}^{q}\) & \(h_{0}^{g}\) & \(a_{s}\,h_{1}^{q}\) & \(a_{s}\,h_{1}^{g}\) & \(a_{s}\,\dot{h}_{1}^{q}\) & \(a_{s}\,\dot{h}_{1}^{g}\) \\ \hline \(N=2\) & 0.3571 & 0.6429 & 0.1003 & 0.3304 & 0.0546 & 0.2149 \\ \hline \(N=3\) & 0.3571 & 0.6429 & 0.1463 & 0.4996 & 0.0393 & 0.1379 \\ \hline \hline (500,550) GeV & \(h_{0}^{q}\) & \(h_{0}^{g}\) & \(a_{s}\,h_{1}^{q}\) & \(a_{s}\,h_{1}^{g}\) & \(a_{s}\,\dot{h}_{1}^{q}\) & \(a_{s}\,\dot{h}_{1}^{g}\) \\ \hline \(N=2\) & 0.4417 & 0.5583 & 0.1337 & 0.2473 & 0.0568 & 0.1816 \\ \hline \(N=3\) & 0.4417 & 0.5583 & 0.1820 & 0.3894 & 0.0417 & 0.1150 \\ \hline \end{tabular} \end{table} Table 2: Values for hard function moments in \(pp\) collision for different \(p_{t}\) ranges. The NLO corrections turn out to be significant. Figure 13: NLO hard function moments for \(N=2\) (left), and \(N=3\) (right), with \(p_{t}\in[300,350]\) GeV and \(\delta_{\rm cut}\in\{0.003,\,0.03,\,0.3\}\). The lower panels show the relative variation of the moments compared with the average value for three \(\delta_{\rm cut}\). The error bars represent the Monte-Carlo numeric uncertainty given by Madgraph5. #### Jet function For ENC, solving the jet function RGE requires the regular anomalous dimensions and their derivatives, and at NNLL, similar to hard function, we need the regular terms up to two-loop, the first derivative up to one-loop as well as the second derivative at tree-level. The QCD timelike splitting function is expanded in \(\frac{\alpha_{s}}{4\pi}\) \[P_{ij}(x)=\sum_{L=0}^{\infty}\left(\frac{\alpha_{s}}{4\pi}\right)^{L+1}P_{ij}^{ (L)}(x)\,, \tag{100}\] and the anomalous dimension for ENC is defined to be the (N+1) Mellin moment of the splitting function. Explicitly, \[\gamma^{(L)}_{T,ij} \equiv-\int_{0}^{1}\mathrm{d}x\,x^{N}\,P_{ij}^{(L)}(x)\,,\] \[\dot{\gamma}^{(L)}_{T,ij} \equiv-\int_{0}^{1}\mathrm{d}x\,\ln x\,x^{N}P_{ij}^{(L)}(x)\,,\] \[\ddot{\gamma}^{(L)}_{T,ij} \equiv-\int_{0}^{1}\mathrm{d}x\,\ln^{2}x\,x^{N}P_{ij}^{(L)}(x)\,. \tag{101}\] Here the dot also represents the derivative with respect to \(N\). Note that \(\{i,j\}=\{q,g\}\) and the anomalous dimension is a \(2\times 2\) matrix. The results for EEC (\(N=2\)) are derived and summarized in the appendix of Ref. [17], so here we provide the expressions for E3C (\(N=3\)). At LO, we find \[\gamma^{(0)}_{T,qq} =\frac{157}{30}\,C_{F}\,,\qquad\gamma^{(0)}_{T,gq}=-\frac{11}{15} \,C_{F}\,,\qquad\gamma^{(0)}_{T,qg}=\frac{11}{30}\,n_{f}\,,\qquad\gamma^{(0)}_{ T,gg}=\frac{21}{5}\,C_{A}+\frac{2}{3}\,n_{f}\,,\] \[\dot{\gamma}^{(0)}_{T,qq} =\left(4\zeta_{2}-\frac{10169}{1800}\right)C_{F}\,,\quad\dot{ \gamma}^{(0)}_{T,gq}=\frac{247}{900}\,C_{F}\,,\quad\dot{\gamma}^{(0)}_{T,qg}= \frac{137}{1800}\,n_{f}\,,\quad\dot{\gamma}^{(0)}_{T,gg}=\left(4\zeta_{2}- \frac{2453}{450}\right)C_{A}\,,\] \[\ddot{\gamma}^{(0)}_{T,qq} =\left(-8\zeta_{3}+\frac{507103}{54000}\right)C_{F}\,,\quad\ddot{ \gamma}^{(0)}_{T,gq}=-\frac{5489}{27000}\,C_{F}\,,\quad\ddot{\gamma}^{(0)}_{ T,qg}=-\frac{1919}{54000}\,n_{f}\,,\] \[\ddot{\gamma}^{(0)}_{T,gg} =\left(-8\zeta_{3}+\frac{124511}{13500}\right)C_{A}\,, \tag{102}\] and at NLO, we have \[\gamma^{(1)}_{T,qq} =\left(-\frac{628}{15}+\frac{2905763}{54000}\right)C_{F}^{2}+ \frac{16157}{675}C_{A}C_{F}-\frac{13427}{3000}\,C_{F}n_{f}\,,\] \[\gamma^{(1)}_{T,gq} =\left(\frac{88}{15}\zeta_{2}-\frac{104389}{27000}\right)C_{F}^{2} -\frac{142591}{13500}C_{A}C_{F}\,,\] \[\gamma^{(1)}_{T,qg} =\left(\frac{44}{15}\zeta_{2}-\frac{60391}{27000}\right)C_{A}n_{f} -\frac{166729}{54000}\,C_{F}n_{f}-\frac{6}{25}\,n_{f}^{2}\,,\] \[\gamma^{(1)}_{T,gg} =\left(-\frac{168}{5}\zeta_{2}+\frac{90047}{1500}\right)C_{A}^{2} +\left(-\frac{16}{3}\zeta_{2}+\frac{2273}{1350}\right)C_{A}n_{f}+\frac{57287}{2 7000}\,C_{F}n_{f}\,,\] \[\dot{\gamma}^{(1)}_{T,qq} =\left(-120\zeta_{4}+\frac{422}{3}\zeta_{3}+\frac{10169}{150} \zeta_{2}-\frac{162656941}{1080000}\right)C_{F}^{2}\] \[\gamma^{(2)}_{T,gg}=\left(840\zeta_{4}-\frac{3752}{25}\zeta_{3}-\frac{ 342578}{375}\zeta_{2}+\frac{1069405919}{1350000}\right)C_{A}^{3}\] \[\qquad+\left(\frac{400}{3}\zeta_{4}-\frac{29534}{225}\zeta_{3}- \frac{30316}{675}\zeta_{2}+\frac{129284923}{2430000}\right)C_{A}^{2}n_{f}\] \[+\left(\frac{2744}{45}\zeta_{3}-\frac{2158}{125}\zeta_{2}-\frac{18828 3293}{6075000}\right)C_{A}C_{F}n_{f}+\left(-\frac{352}{225}\zeta_{3}+\frac{4037} {3375}\zeta_{2}+\frac{27742123}{24300000}\right)C_{F}^{2}n_{f}\] \[+\left(-\frac{64}{9}\zeta_{3}+\frac{160}{27}\zeta_{2}-\frac{71341 }{27000}\right)C_{A}n_{f}^{2}+\left(-\frac{484}{675}\zeta_{2}-\frac{165553}{270 000}\right)C_{F}n_{f}^{2}\,.\] (A.9) ## Appendix B \(\beta\)-function RGE and running coupling The well-known QCD \(\beta\)-function is written as \[\frac{\mathrm{d}\alpha_{s}(\mu)}{\mathrm{d}\ln\mu}=\beta(\alpha_{s}(\mu)), \quad\beta(\alpha)=-2\alpha\,\left[\left(\frac{\alpha}{4\pi}\right)\beta_{0}+ \left(\frac{\alpha}{4\pi}\right)^{2}\beta_{1}+\left(\frac{\alpha}{4\pi}\right) ^{3}\beta_{2}+\cdots\right]\,,\] (B.1) where the coefficient up to three loops are given by [103; 104; 105; 106] \[\beta_{0} =\frac{11}{3}C_{A}-\frac{4}{3}T_{F}n_{f}\,,\] (B.2) \[\beta_{1} =\frac{34}{3}C_{A}^{2}-\frac{20}{3}C_{A}T_{F}n_{f}-4C_{F}T_{F}n_{ f}\,,\] \[\beta_{2} =n_{f}^{2}T_{F}^{2}\left(\frac{158}{27}C_{A}+\frac{44}{9}C_{F} \right)+n_{f}T_{F}\left(2C_{F}^{2}-\frac{205}{9}C_{F}C_{A}-\frac{1415}{27}C_{A }^{2}\right)+\frac{2857}{54}C_{A}^{3}\,,\] \[\beta_{3} =\frac{1093}{729}n_{f}^{3}+\left(\frac{50065}{162}+\frac{6472}{8 1}\zeta_{3}\right)n_{f}^{2}+\left(-\frac{1078361}{162}-\frac{6508}{27}\zeta_{ 3}\right)n_{f}+3564\zeta_{3}+\frac{149753}{6}\,.\] At one-loop, the \(\beta\)-RGE can be solved exactly. At two-loop and beyond, there are different solutions. In terms of \(L\equiv\ln\frac{\mu^{2}}{\Lambda_{\text{QCD}}^{2}}\), a expanded solution can be written as: \[\alpha_{s}(\mu)=\frac{4\pi}{\beta_{0}}\left[\frac{1}{L}-\frac{ \beta_{1}}{\beta_{0}^{2}L^{2}}\ln L+\frac{\beta_{1}^{2}}{\beta_{0}^{4}L^{3}}( \ln^{2}L-\ln L-1)+\frac{\beta_{2}}{\beta_{0}^{3}L^{3}}\right.\\ \left.+\frac{\beta_{1}^{3}}{\beta_{0}^{6}L^{4}}\left(-\ln^{3}L+ \frac{5}{2}\ln^{2}L+2\ln L-\frac{1}{2}\right)-3\frac{\beta_{1}\beta_{2}}{ \beta_{0}^{5}L^{4}}\ln L+\frac{\beta_{3}}{2\beta_{0}^{4}L^{4}}\right]\,.\] (B.3) Here we can obtain the two-loop running coupling for NLL resumation by setting \(\beta_{2}=\beta_{3}=0\) and three-loop running coupling for NNLL by only \(\beta_{3}=0\). Alternatively, one can iteratively solve the RGE order by order in a formal expansion parameter \(\epsilon\sim\frac{\beta_{a}}{\beta_{0}}\), with \(n\geq 1\). For NLL, the two-loop running coupling is written as \[\alpha_{s}(\mu)=\alpha_{s}(Q)\left[X+\alpha_{s}(Q)\frac{\beta_{1}}{4\pi\beta_ {0}}\ln X\right]^{-1},\quad X\equiv 1+\frac{\alpha_{s}(Q)}{2\pi}\beta_{0}\ln \frac{\mu}{Q}\,,\] (B.4) and at three loops for NNLL \[\alpha_{s}(\mu)=\alpha_{s}(Q)\left\{X+\alpha_{s}(Q)\frac{\beta_{1}}{4\pi\beta _{0}}\ln X+\frac{\alpha_{s}^{2}(Q)}{16\pi^{2}}\left[\frac{\beta_{2}}{\beta_{0}} \left(1-\frac{1}{X}\right)+\frac{\beta_{1}^{2}}{\beta_{0}^{2}}\left(\frac{1}{X }-1+\frac{\ln X}{X}\right)\right]\right\}^{-1}\,.\] (B.5) For the resummation in this paper, we use the iterative solution (the latter one) and set the coupling at \(Q=91.2\) GeV to be the world average value \(\alpha_{s}(m_{Z})=0.118\). Squeeze limit of EEEC jet functions In this section, we provide the perturbative data for the squeeze limit of the EEEC jet function in Eq. (17), which is needed for E3C jet function calculation. Given the conformal parameterization, \[x_{1}=x_{L}z\bar{z},\quad x_{2}=x_{L}(1-z)(1-\bar{z}),\quad x_{3}=x_{L}\,, \tag{114}\] the squeeze limits correspond to \(z\to 0,1,\infty\), related by a \(\mathbb{S}_{3}\) symmetry. Without loss of generality, we provide the \(z\to 1\) limit for the shapes function up to \(\mathcal{O}(\epsilon^{2})\). In the quark jet, we find for \(G(z)\) \[G_{q}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f}\bigg{(} \frac{13}{4800(1-z)(1-\bar{z})}+\frac{z-2}{1440(1-\bar{z})^{2}}+\frac{\bar{z}}{ 1440(1-z)^{2}}-\frac{39z+1}{28800(1-z)^{2}}\] \[+\frac{13}{9600(1-\bar{z})}\bigg{)}+C_{F}C_{A}\bigg{(}\frac{91}{4 800(1-z)(1-\bar{z})}+\frac{2-z}{2880(1-\bar{z})^{2}}-\frac{\bar{z}}{2880(1-z)^ {2}}\] \[-\frac{273z-293}{28800(1-z)^{2}}+\frac{91}{9600(1-\bar{z})} \bigg{)}+C_{F}^{2}\bigg{(}\frac{1}{20(1-z)(\bar{z}-1)}-\frac{z+\bar{z}-2}{40(z -1)(1-\bar{z})}\bigg{)}\,, \tag{115}\] and for \(F(z)\): \[F_{q}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f}\bigg{(} \frac{649}{28800(1-z)(1-\bar{z})}-\frac{259}{43200(1-z)^{2}}-\frac{259}{43200(1 -\bar{z})^{2}}\bigg{)}\] \[+C_{F}C_{A}\bigg{(}\frac{561}{3200(1-z)(1-\bar{z})}+\frac{229}{86 400(1-z)^{2}}+\frac{229}{86400(1-\bar{z})^{2}}\bigg{)}\] \[+C_{F}^{2}\bigg{(}\frac{3307}{7200(1-z)(1-\bar{z})}\bigg{)}\,, \tag{116}\] as well as the \(H(z)\): \[H_{q}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f}\bigg{(} \frac{664193-23400\pi^{2}}{4320000(1-z)(1-\bar{z})}+\frac{1800\pi^{2}-53191}{ 1296000(1-z)^{2}}+\frac{1800\pi^{2}-53191}{1296000(1-\bar{z})^{2}}\bigg{)}\] \[+C_{F}C_{A}\bigg{(}\frac{1805867-54600\pi^{2}}{1440000(1-z)(1-\bar {z})}+\frac{45421-1800\pi^{2}}{2592000(1-\bar{z})^{2}}-\frac{1800\pi^{2}-45421 }{2592000(1-z)^{2}}\bigg{)}\] \[+C_{F}^{2}\bigg{(}\frac{352451-10800\pi^{2}}{108000(1-z)(1-\bar{ z})}\bigg{)}\,. \tag{117}\] Here the red stands for the most singular term, which contributes to \(\frac{1}{\epsilon}\) divergence in the E3C jet function calculation. For the gluon jet, we also find \[G_{g}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f}\bigg{(} \frac{3}{320(1-z)(1-\bar{z})}+\frac{3}{640(1-z)}+\frac{3}{640(1-\bar{z})}\bigg{)}\] \[+C_{A}T_{F}n_{f}\bigg{(}\frac{7}{800(1-z)(1-\bar{z})}+\frac{z-2}{1 440(1-\bar{z})^{2}}+\frac{\bar{z}}{1440(1-z)^{2}}-\frac{63z-43}{14400(1-z)^{2}}\] \[+\frac{7}{1600(1-\bar{z})}\bigg{)}+C_{A}^{2}\bigg{(}\frac{49}{800( 1-z)(1-\bar{z})}+\frac{2-z}{2880(1-\bar{z})^{2}}-\frac{\bar{z}}{2880(1-z)^{2}}\] \[-\frac{441z-451}{14400(1-z)^{2}}+\frac{49}{1600(1-\bar{z})}\Bigg{)}\,, \tag{109}\] \[F_{g}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f}\bigg{(} \frac{241}{3200(1-z)(1-\bar{z})}\bigg{)}+C_{A}T_{F}n_{f}\bigg{(}\frac{343}{4800 (1-z)(1-\bar{z})}-\frac{259}{43200(1-z)^{2}}\] \[-\frac{259}{43200(1-\bar{z})^{2}}\bigg{)}+C_{A}^{2}\bigg{(}\frac{ 557}{960(1-z)(1-\bar{z})}+\frac{229}{86400(1-z)^{2}}+\frac{229}{86400(1-\bar{z })^{2}}\bigg{)}\,,\] \[H_{g}(z) \stackrel{{ z\to 1}}{{\approx}}C_{F}T_{F}n_{f} \bigg{(}\frac{434309-16200\pi^{2}}{864000(1-z)(1-\bar{z})}+C_{A}T_{F}n_{f} \bigg{(}\frac{1033981-37800\pi^{2}}{2160000(1-z)(1-\bar{z})}\] \[+\frac{1800\pi^{2}-53191}{1296000(1-z)^{2}}+\frac{1800\pi^{2}-531 91}{1296000(1-\bar{z})^{2}}\bigg{)}+C_{A}^{2}\bigg{(}\frac{2999389-88200\pi^{2} }{720000(1-z)(1-\bar{z})}\] \[-\frac{1800\pi^{2}-45421}{2592000(1-z)^{2}}-\frac{1800\pi^{2}-454 21}{2592000(1-\bar{z})^{2}}\bigg{)}\,. \tag{110}\] ## Appendix D Result of two-loop E3C jet function calculation We list the individual results for the two-loop jet function calculation in Sec. 2.4. As we discussed above, the calculation is reorganized as nonidentical energy weight contribution and contact terms. For the nonidentical energy weight in Sec. 2.4.1, we find for the quark jet \[\frac{dJ_{q}^{\rm nonid}}{dx_{L}} =\Big{(}\frac{\alpha_{s}}{4\pi}\Big{)}^{2}\left\{\delta(x_{L})f_{ q}(\mu,Q,\epsilon)+\frac{1}{x_{L}}\bigg{[}C_{F}T_{F}n_{f}\bigg{(}-\frac{13}{200 \epsilon}+\frac{13}{100}\ln\bigg{(}\frac{Q^{2}x_{L}}{\mu^{2}}\bigg{)}\] \[-0.44158(3)\bigg{)}+C_{F}^{2}\left(-\frac{6}{5\epsilon}+\frac{12} {5}\ln\bigg{(}\frac{Q^{2}x_{L}}{\mu^{2}}\bigg{)}-10.963(1)\right)\] \[+C_{F}C_{A}\left(-\frac{91}{200\epsilon}+\frac{91}{100}\ln\bigg{(} \frac{Q^{2}x_{L}}{\mu^{2}}\bigg{)}-4.3743(7)\bigg{)}\,\bigg{]}\right\}, \tag{111}\] with the coefficient of the \(\delta(x_{L})\) being \[f_{q}(\mu,Q,\epsilon) =C_{F}T_{F}n_{f}\bigg{[}\frac{13}{400\epsilon^{2}}+\frac{1}{ \epsilon}\left(\frac{13}{200}\ln\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+0.22079( 2)\right)+0.44158(3)\ln\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}\] \[+\frac{13}{200}\ln^{2}\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+0.544 1(8)\bigg{]}+C_{F}C_{A}\bigg{[}\frac{91}{400\epsilon^{2}}+\frac{1}{\epsilon} \left(\frac{91}{200}\ln\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+2.1871(8)\right)\] \[+4.3743(7)\ln\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+\frac{91}{200} \ln^{2}\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+10.483(2)\bigg{]}+C_{F}^{2}\bigg{[} 24.60(4)+\frac{3}{5\epsilon^{2}}\] \[+\frac{1}{\epsilon}\left(\frac{6}{5}\ln\bigg{(}\frac{\mu^{2}}{Q^{2 }}\bigg{)}+5.4815(3)\right)+10.963(1)\ln\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}+ \frac{6}{5}\ln^{2}\bigg{(}\frac{\mu^{2}}{Q^{2}}\bigg{)}\,\bigg{]}\,. \tag{112}\] The \(\ln\Big{(}\frac{Q^{2}x_{L}}{\mu^{2}}\Big{)}\) term is verified by the jet RGE. For a gluon jet, the \(\mathcal{O}(\alpha_{s}^{2})\) contribution is \[\frac{dJ_{g}^{\rm nonid}}{dx_{L}}=\Big{(}\frac{\alpha_{s}}{4\pi}\Big{)}^{2} \left\{\delta(x_{L})f_{g}(\mu,Q,\epsilon)+\frac{1}{x_{L}}\bigg{[}C_{F}T_{F}n_{f} \left(-\frac{9}{40\epsilon}+\frac{9}{20}\ln\bigg{(}\frac{Q^{2}x_{L}}{\mu^{2}} \right)\] \[-1.8862(6))+C_{A}T_{F}n_{f}\left(-\frac{21}{100\epsilon}+\frac{21}{5 0}\ln\left(\frac{Q^{2}x_{L}}{\mu^{2}}\right)-1.5376(9)\right)\] \[+C_{A}^{2}\left(-\frac{147}{100\epsilon}+\frac{147}{50}\ln\left( \frac{Q^{2}x_{L}}{\mu^{2}}\right)-14.031(3)\right)\bigg{]}\bigg{\}}\,, \tag{103}\] with the corresponding coefficient \[f_{g}(\mu,Q,\epsilon) =C_{A}T_{F}n_{f}\bigg{[}\frac{21}{200\epsilon^{2}}+\frac{1}{ \epsilon}\left(\frac{21}{100}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+0.7688(5) \right)+1.5376(9)\ln\left(\frac{\mu^{2}}{Q^{2}}\right)\] \[+\frac{21}{100}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)+2.350(8 )\bigg{]}+C_{F}T_{F}n_{f}\bigg{[}\frac{9}{80\epsilon^{2}}+\frac{1}{\epsilon} \left(\frac{9}{40}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+0.9431(3)\right)\] \[+1.886(3)\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+\frac{9}{40}\ln^{ 2}\left(\frac{\mu^{2}}{Q^{2}}\right)+3.757(1)\bigg{]}+C_{A}^{2}\bigg{[}33.188( 4)+\frac{147}{200\epsilon^{2}}\] \[+\frac{1}{\epsilon}\left(\frac{147}{100}\ln\left(\frac{\mu^{2}}{Q ^{2}}\right)+7.01569(5)\right)+14.031(3)\ln\left(\frac{\mu^{2}}{Q^{2}}\right) +\frac{147}{100}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)\bigg{]}\,. \tag{104}\] Regarding the contact term in Sec. 2.4.2, for \(e^{+}e^{-}\) annihilation, we have the sum of E\({}^{2}\)EC and E\({}^{3}\)C \[\frac{1}{\sigma_{0}}\frac{\mathrm{d}_{\mathrm{C},q}^{[3],2\text{- loop}}(x_{L},\epsilon)}{\mathrm{d}x_{L}}= \left(\frac{\alpha_{s}}{4\pi}\right)^{2}\Bigg{\{}\delta(x_{L})r_{ q}(\mu,Q,\epsilon)+\left[\frac{1}{x_{L}}\right]_{+}\left[C_{A}C_{F}\bigg{(}\frac{91}{100 \epsilon}+\frac{1189}{200}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)\right.\] \[-6\zeta_{3}+\frac{25\pi^{2}}{6}-\frac{52307}{18000}\bigg{)}+C_{F }n_{f}T_{F}\left(\frac{13}{100\epsilon}-\frac{31}{25}\ln\left(\frac{\mu^{2}}{Q ^{2}}\right)-\frac{14809}{2000}\right)\] \[+C_{F}^{2}\left(\frac{12}{5\epsilon}+\frac{24}{5}\ln\left(\frac{ \mu^{2}}{Q^{2}}\right)+12\zeta_{3}-\frac{43\pi^{2}}{6}+\frac{274081}{3600} \right)\bigg{]}\] \[+\left[\frac{\ln(x_{L})}{x_{L}}\right]_{+}\left(-\frac{1343}{200} C_{A}C_{F}+\frac{113}{100}C_{F}n_{f}T_{F}+\frac{87}{80}C_{F}^{2}\right)\Bigg{\}}\,, \tag{105}\] with the singular part \(r_{q}(\mu,Q,\epsilon)\) \[r_{q}(\mu,Q,\epsilon) =C_{A}C_{F}\Bigg{[}-\frac{91}{200\epsilon^{2}}+\frac{1}{\epsilon} \Bigg{(}-\frac{91}{100}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+3\zeta_{3}-\frac {25\pi^{2}}{12}+\frac{452921}{36000}\Bigg{)}-\frac{91}{100}\ln^{2}\left(\frac{ \mu^{2}}{Q^{2}}\right)\] \[+\left(6\zeta_{3}+\frac{890167}{36000}-\frac{25\pi^{2}}{6} \right)\ln\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{347\zeta_{3}}{2}+\frac{7 \pi^{4}}{20}-\frac{6697\pi^{2}}{1800}+\frac{47220317}{270000}\Bigg{]}\] \[+C_{F}n_{f}T_{F}\Bigg{[}-\frac{13}{200\epsilon^{2}}+\frac{1}{ \epsilon}\Bigg{(}-\frac{13}{100}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{ 5299}{12000}\Bigg{)}-\frac{13}{100}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)\] \[-\frac{4349}{6000}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+4\zeta_{3} +\frac{137\pi^{2}}{400}-\frac{1413979}{720000}\Bigg{]}+C_{F}^{2}\Bigg{[}- \frac{6}{5\epsilon^{2}}\] \[+\frac{1}{\epsilon}\Bigg{(}-\frac{12}{5}\ln\left(\frac{\mu^{2}}{Q ^{2}}\right)-6\zeta_{3}+\frac{43\pi^{2}}{12}-\frac{281641}{7200}\Bigg{)}- \frac{12}{5}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)\] \[+\left(-12\zeta_{3}-\frac{281641}{3600}+\frac{43\pi^{2}}{6}\right) \ln\left(\frac{\mu^{2}}{Q^{2}}\right)+293\zeta_{3}-\frac{7\pi^{4}}{10}+\frac{15 371\pi^{2}}{1440}-\frac{380074411}{864000}\Bigg{]}\,. \tag{106}\] Similarly, in the gluonic Higgs decay, we get \[\frac{1}{\sigma_{0}^{\prime}}\frac{\text{d}\sigma_{\text{C,g}}^{[3] \text{,2-loop}}(x_{L},\epsilon)}{\text{d}x_{L}}= \lambda(\mu)\left(\frac{\alpha_{s}}{4\pi}\right)^{2}\Bigg{\{} \delta(x_{L})r_{g}(\mu,Q,\epsilon)+\left[\frac{1}{x_{L}}\right]_{+}\left\{n_ {f}^{2}T_{F}^{2}\bigg{(}-\frac{3}{5}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)- \frac{131}{60}\bigg{)}\right.\] \[+n_{f}T_{F}\bigg{[}C_{A}\left(\frac{21}{50\epsilon}-\frac{171}{100 }\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+\frac{7\pi^{2}}{15}-\frac{140917}{9000} \right)\] \[+C_{F}\left(\frac{9}{10}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)+ \frac{9}{20\epsilon}+\frac{1579}{400}\right)\bigg{]}\] \[+C_{A}^{2}\bigg{(}\frac{147}{50\epsilon}+\frac{1743}{100}\ln \left(\frac{\mu^{2}}{Q^{2}}\right)+6\zeta_{3}-\frac{97\pi^{2}}{30}+\frac{21182 9}{2250}\bigg{)}\bigg{\}}\] \[+\left[\frac{\ln(x_{L})}{x_{L}}\right]_{+}\left[n_{f}T_{F}\left( \frac{51}{25}C_{A}-\frac{69}{40}C_{F}\right)-\frac{133}{25}C_{A}^{2}+\frac{2} {5}n_{f}^{2}T_{F}^{2}\right]\right\}, \tag{100}\] with the gluonic singular term \(r_{g}(\mu,Q,\epsilon)\) \[r_{g}(\mu,Q,\epsilon) =C_{A}T_{F}n_{f}\bigg{[}-\frac{21}{100\epsilon^{2}}+\frac{1}{ \epsilon}\bigg{(}-\frac{21}{50}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{7 \pi^{2}}{30}+\frac{6887}{9000}\bigg{)}-\frac{1163}{150}\ln^{2}\left(\frac{\mu^ {2}}{Q^{2}}\right)\] \[+\left(-\frac{948847}{18000}-\frac{7\pi^{2}}{15}\right)\ln\left( \frac{\mu^{2}}{Q^{2}}\right)-\frac{211\zeta_{3}}{10}+\frac{3037\pi^{2}}{1800}- \frac{5585159}{67500}\bigg{]}+C_{F}T_{F}n_{f}\] \[\left[-\frac{9}{40\epsilon^{2}}+\frac{1}{\epsilon}\bigg{(}-\frac {9}{20}\ln\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{1509}{800}\bigg{)}-\frac{9} {20}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{3109}{400}\ln\left(\frac{ \mu^{2}}{Q^{2}}\right)+15\zeta_{3}\right.\] \[+\frac{5\pi^{2}}{8}-\frac{230393}{6000}\bigg{]}+C_{A}^{2}\bigg{\{} -\frac{147}{100\epsilon^{2}}+\frac{1}{\epsilon}\bigg{[}-\frac{147}{50}\ln \left(\frac{\mu^{2}}{Q^{2}}\right)-3\zeta_{3}+\frac{97\pi^{2}}{60}-\frac{47485 7}{18000}\bigg{]}\] \[+\frac{2143}{300}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)+\left( -6\zeta_{3}+\frac{261281}{18000}+\frac{97\pi^{2}}{30}\right)\ln\left(\frac{\mu ^{2}}{Q^{2}}\right)+\frac{1133\zeta_{3}}{10}-\frac{7\pi^{4}}{20}\] \[+\frac{373\pi^{2}}{100}-\frac{12512789}{90000}\bigg{\}}+n_{f}^{2} T_{F}^{2}\bigg{[}\frac{4}{3}\ln^{2}\left(\frac{\mu^{2}}{Q^{2}}\right)+\frac{2971}{300} \ln\left(\frac{\mu^{2}}{Q^{2}}\right)-\frac{23\pi^{2}}{45}+\frac{579043}{27000} \bigg{]}\,, \tag{101}\] where \(\lambda\) is the effective \(Hgg\) coupling5[107]. These results are then used to extract the two-loop jet constants. Footnote 5: For the case of gluonic Higgs decays, we normalize the E3C into the form where the LO E3C is \(\frac{1}{\sigma_{0}^{\prime}}\frac{\text{d}\sigma_{\text{C,g}}^{[3]}}{\text{d}x_{L }}=\lambda(\mu)\left(\frac{1}{4}\delta(x_{L})+\frac{3}{4}\delta(1-x_{L})\right)\) in \(d=4-2\epsilon\) dimensions. ## Appendix E Fixed-order expansion In this section, we provide the singular expansion of projected energy correlator up to NNLO \(\mathcal{O}(\alpha_{s}^{3})\) in \(e^{+}e^{-}\) annihilation. This can be achieved by expanding our resummed distribution with canonical scale \(\mu=Q\). For EEC, we find \[\frac{1}{\sigma_{0}}\frac{d\sigma^{[2]}}{dx_{L}}=\left(\frac{\alpha_{s}}{4\pi} \right)C_{F}\frac{3}{2x_{L}}+\left(\frac{\alpha_{s}}{4\pi}\right)^{2}C_{F} \bigg{\{}\bigg{[}\frac{53}{30}n_{f}T_{F}+\frac{25}{4}C_{F}-\frac{107}{15}C_{A} \bigg{]}\frac{\ln x_{L}}{x_{L}}\] \[+\bigg{[}-\frac{4913}{450}n_{f}T_{F}+\bigg{(}-\frac{8263}{216}+\frac{4 3}{9}\pi^{2}-8\zeta_{3}\bigg{)}C_{F}+\bigg{(}\frac{35336}{675}-\frac{25}{9}\pi^{ 2}+4\zeta_{3}\bigg{)}C_{A}\bigg{]}\frac{1}{x_{L}}\bigg{\}}\] \[+\bigg{(}\frac{\alpha_{s}}{4\pi}\bigg{)}^{3}\,C_{F}\bigg{\{}\bigg{[} \frac{8059}{300}C_{A}^{2}-\frac{340}{9}C_{F}C_{A}+\frac{625}{48}C_{F}^{2}- \frac{16259}{900}C_{A}T_{F}n_{f}+\frac{4619}{360}C_{F}T_{F}n_{f}\] \[+\frac{92}{45}n_{f}^{2}T_{F}^{2}\bigg{]}\frac{\ln^{2}x_{L}}{x_{L} }+\bigg{[}-\frac{17734}{675}n_{f}^{2}T_{F}^{2}+\bigg{(}-\frac{64\zeta_{3}}{3}- \frac{6760183}{32400}+\frac{416\pi^{2}}{27}\bigg{)}C_{F}T_{F}n_{F}\] \[+\bigg{(}\frac{32\zeta_{3}}{3}+\frac{6644267}{27000}-\frac{36 \pi^{2}}{5}\bigg{)}C_{A}T_{F}n_{f}+\bigg{(}-\frac{172\zeta_{3}}{3}-\frac{7235 33}{2592}+\frac{1849\pi^{2}}{54}\bigg{)}C_{F}^{2}\] \[+\bigg{(}-\frac{74\zeta_{3}}{3}-\frac{2916859}{6750}+\frac{503 \pi^{2}}{30}\bigg{)}C_{A}^{2}+\bigg{(}\frac{262\zeta_{3}}{3}+\frac{105425}{14 4}-\frac{550\pi^{2}}{9}\bigg{)}C_{F}C_{A}\bigg{]}\frac{\ln x_{L}}{x_{L}}\] \[+\bigg{[}\bigg{(}\frac{88031}{1125}+\frac{4\pi^{2}}{5}\bigg{)}n_{ f}^{2}T_{F}^{2}+\bigg{(}-\frac{15988\zeta_{3}}{45}+\frac{236\pi^{4}}{135}- \frac{15161\pi^{2}}{360}+\frac{164829499}{243000}\bigg{)}C_{F}T_{F}n_{F}\] \[+\bigg{(}\frac{3679\zeta_{3}}{15}-\frac{118\pi^{4}}{135}+\frac{3 79579\pi^{2}}{16200}-\frac{1025118113}{1080000}\bigg{)}C_{A}T_{F}n_{F}\] \[+\bigg{(}8\pi^{2}\zeta_{3}+52\zeta_{3}+208\zeta_{5}-\frac{167\pi^ {4}}{27}-\frac{18805\pi^{2}}{1296}+\frac{742433}{1944}\bigg{)}C_{F}^{2}\] \[+\bigg{(}4\pi^{2}\zeta_{3}-\frac{47483\zeta_{3}}{90}+56\zeta_{5} -\frac{481\pi^{4}}{540}-\frac{906257\pi^{2}}{16200}+\frac{964892417}{540000} \bigg{)}C_{A}^{2}\] \[+\bigg{(}-12\pi^{2}\zeta_{3}+\frac{10604\zeta_{3}}{15}-216\zeta_ {5}+\frac{847\pi^{4}}{180}+\frac{137305\pi^{2}}{1296}-\frac{105395741}{51840 }\bigg{)}C_{F}C_{A}\bigg{]}\frac{1}{x_{L}}\bigg{\}}\,. \tag{100}\] Similarly, for E3C, we have \[\frac{1}{\sigma_{0}}\frac{d^{\sigma[3]}}{dx_{L}} =\Big{(}\frac{\alpha_{s}}{4\pi}\Big{)}\,C_{F}\frac{9}{8x_{L}}+ \Big{(}\frac{\alpha_{s}}{4\pi}\Big{)}^{2}\,C_{F}\bigg{\{}\bigg{[}\frac{139}{100 }n_{f}T_{F}+\frac{471}{80}C_{F}-\frac{979}{200}C_{A}\bigg{]}\frac{\ln x_{L}}{x _{L}}\] \[+\bigg{[}-\frac{24863}{3000}n_{f}T_{F}-\frac{21}{10}C_{F}+\frac{ 66769}{3000}C_{A}\bigg{]}\frac{1}{x_{L}}\bigg{\}}\] \[+\bigg{(}\frac{\alpha_{s}}{4\pi}\bigg{)}^{3}\,C_{F}\bigg{\{}\bigg{[} \frac{17743}{1000}C_{A}^{2}-\frac{412753}{12000}C_{F}C_{A}+\frac{24649}{1600} C_{F}^{2}-\frac{19019}{1500}C_{A}T_{F}n_{f}\] \[+\frac{35369}{3000}C_{F}T_{F}n_{f}+\frac{128}{75}n_{f}^{2}T_{F}^{ 2}\bigg{]}\frac{\ln^{2}x_{L}}{x_{L}}+\bigg{[}-\frac{4559891}{22500}C_{A}- \frac{814823}{48000}C_{F}^{2}\] \[+\bigg{(}\frac{34399441}{120000}-\frac{11\pi^{2}}{2}\bigg{)}C_{F}C _{A}+\bigg{(}2\pi^{2}-\frac{1026851}{10000}\bigg{)}C_{F}T_{F}n_{f}+\frac{305590 7}{22500}C_{A}T_{F}n_{f}\] \[-\frac{23494}{1125}n_{f}^{2}T_{F}^{2}\bigg{]}\frac{\ln x_{L}}{x_{L }}+\bigg{[}j_{2}^{q,[3]}\bigg{(}\frac{157}{15}-\frac{44C_{A}}{3C_{F}}+\frac{1 6n_{f}T_{F}}{3C_{F}}\bigg{)}-\frac{22}{15}j_{2}^{g,[3]}\] \[+\bigg{(}\frac{106027}{54000}-\frac{22\pi^{2}}{225}\bigg{)}n_{f}^{ 2}T_{F}^{2}+\bigg{(}\frac{1827\zeta_{3}}{25}-\frac{3877\pi^{2}}{3000}-\frac{32 39027203}{10800000}\bigg{)}C_{F}T_{F}n_{f}\] \[+\bigg{(}-\frac{1037\zeta_{3}}{50}-\frac{2167\pi^{2}}{4500}-\frac{2 4958553}{3600000}\bigg{)}C_{A}T_{F}n_{f}\] \[+\bigg{(}\frac{3267\zeta_{3}}{20}-\frac{111313\pi^{2}}{14400}-\frac{6 031520921}{17280000}\bigg{)}C_{F}^{2}+\bigg{(}-\frac{829\zeta_{3}}{100}+\frac{4 4333\pi^{2}}{2250}+\frac{363491521}{5400000}\bigg{)}C_{A}^{2}\] \[+\bigg{(}-\frac{42321\zeta_{3}}{200}+\frac{284797\pi^{2}}{36000}+ \frac{4941457181}{7200000}\bigg{)}C_{F}C_{A}\bigg{]}\frac{1}{x_{L}}\bigg{\}}\,, \tag{114}\] with the two-loop jet constant \(j_{2}^{q/g,[3]}\) from Eq. (4.4)-(2.41).

2306.07639

Hidden Lagrangian coherence and memory effects in the statistics of Hamiltonian motions

This paper is focused on the coherent effects that appear in tracer statistics in two-dimensional incompressible turbulence in the presence of an average velocity. We show that this determines strong modifications of the transport and trajectory statistics, which are essentially caused by hidden coherent components of the motion.

Madalina Vlad, Dragos Iustin Palade, Florin Spineanu

2023-06-13T09:18:03Z

http://arxiv.org/abs/2306.07639v1

# Hidden Lagrangian coherence and memory effects ###### Abstract Turbulence is a complex nonlinear process, which appears in many domains as fluid mechanics, plasma physics, astrophysics, atmosphere and ocean sciences, chemistry [1]-[3]. One of the main difficulty in understanding the dynamics of turbulence is the complicated combination of stochastic and quasi-coherent aspects that is typical for the strongly nonlinear regimes. Quasi-coherence or order appears at the basic level of tracer trajectories in smooth stochastic velocity fields with finite correlation lengths \(\lambda\) and times \(\tau_{c}.\) Trajectory coherence (or Lagrangian coherence) is usually a transitory process that lasts during the time of flight over \(\lambda\) with the amplitude \(V\) of the stochastic velocity, \(\tau_{fl}=\lambda/V.\) It exists only for slow time variation of the velocity with \(\tau_{c}>\tau_{fl}\). In the case of two-dimensional incompressible velocity fields, a much stronger Lagrangian coherence appears that is due to trajectory eddying or trapping. It generates vortical structures of trajectories, which produce non-Gaussian statistics of tracer displacements and strongly modified transport [4]-[7]. The order of the tracer trajectories determines much more complicated effects on turbulence evolution, which essentially amplifies the degree of quasi-coherence. The turbulence that is dominantly two-dimensional has a self-organizing character [8]-[11], which consists of the generation of quasi-coherent large scale structures [12]. This paper is focused on the coherent effects that appear in tracer statistics in two-dimensional incompressible turbulence in the presence of an average velocity \({\bf V}_{d}\). We show that \({\bf V}_{d}\) determines strong modifications of the transport and trajectory statistics, which are essentially caused by hidden coherent components of the motion. The results are based on the numerical simulation of the trajectories and consist of a conditional statistical analysis adapted to the special properties of the two-dimensional incompressible velocity fields. The formulation of the problem and the simulation method are presented in Section 2. The motion is of Hamiltonian type with the Hamiltonian function \(\phi_{t}\) composed of the stochastic and average potentials. The trajectories evolve on the contour lines of \(\phi_{t}({\bf x})\) in static (frozen) potentials (\(\tau_{c}\rightarrow\infty\)), and they remain strongly correlated to these lines for slow time variation (large \(\tau_{c}\)). We consider first (Sections 3-5) frozen potentials. We discuss the configuration of the contour lines of the potential and the main features of tracer advection in Section 3. The space of trajectories is organized in two categories: trapped and free. The statistics of the Lagrangian velocity and of the trajectories are examined for each category, and their specific contributions to the global statistics (on the whole set of trajectories) are identified (Section 4). We show that quasi-coherent Lagrangian velocities parallel to \({\bf V}_{d}\) are generated for both categories. They are hidden in the global statistics, in the sense that their contributions compensate each other. However, they provide explanations for the nonstandard transport produced in these conditions. A deeper examination of the coherence induced by the average velocity is presented in Section 5, where we determine the Lagrangian statistics for the trajectories of each category that evolve on contour lines of the potential with the same value of \(\phi_{t}.\) These results reveal other (hidden) coherent elements of the motion, and provide important properties that are used for the understanding of the effects of the average velocity on the statistics of trajectories and transport. Sections 6 and 7 deal with time dependent potentials. The Lagrangian statistics conditioned by the value of the potential shows that the order found in frozen potentials does not decay due to the random time variation of the potential, as expected. On the contrary, important quasi-coherent elements significantly increase (Section 6). Explanations are provided in Section 7. They are essentially related to the constraint of invariance of the potential, which are approximately valid at large \(\tau_{c},\) and to the slow transition of the trajectories between the two categories. Long memory effects are identified and their effects is discussed. A short summary of the results and the conclusions are presented in Section 8. ## II 2. The problem and the simulation method Tracer trajectories in two-dimensional stochastic velocity fields are obtained from \[\frac{d{\bf x}}{dt}={\bf v}({\bf x},\!t)=\widetilde{\bf v}({\bf x},\!t)+V_{d}{ \bf e}_{2}, \tag{1}\] where \({\bf e}_{1}\), \({\bf e}_{2}\) are the unit vectors in the plane of the motion \({\bf x}=(x_{1},x_{2})\), \({\bf e}_{3}\) is perpendicular on this plane. The velocity \({\bf v}({\bf x},\!t)\) has a stochastic component \(\widetilde{\bf v}({\bf x},\!t)\) superposed on a constant average velocity \({\bf V}_{d}=V_{d}\;{\bf e}_{2}.\) The incompressibility condition \(\nabla\cdot\widetilde{\bf v}({\bf x},\!t)=0\) of the velocity field is equivalent with the representation of \(\widetilde{\bf v}({\bf x},\!t)\) by a stochastic potential (or stream function) \(\phi({\bf x},\!t)\) \[\widetilde{\bf v}({\bf x},\!t)=-\nabla\phi({\bf x},\!t)\times{\bf e}_{3}. \tag{2}\] The equation of motion is of Hamiltonian type, with \(x_{1}\) and \(x_{2}\) the conjugate variables and \(\phi_{t}({\bf x},\!t)=\phi({\bf x},\!t)+x_{1}V_{d}\) the Hamiltonian function. Dimensionless quantities are used in Eq. (1) with the potential normalized by its amplitude \(\Phi\), the distances by \(\lambda_{0}\) that is of the order of the correlation lengths, the velocities (including \(V_{d}\)) by \(V_{0}=\Phi/\lambda_{0}\) and the time by \(\tau_{0}=\lambda_{0}/V_{0}.\) The potential is represented by a homogeneous and stationary Gaussian stochastic field. Its Eulerian correlation (EC) \(E({\bf x},\!t)\equiv\langle\phi({\bf x}_{0},\!t_{0})\ \phi({\bf x}_{0}+{\bf x},\!t_{0}+t)\rangle\) in dimensionless quantities is modeled in the simulations presented here by \[E({\bf x},\!t)\equiv\exp\left(-\frac{x_{1}^{2}}{2\lambda_{1}^{2}}-\frac{x_{2 }^{2}}{2\lambda_{2}^{2}}-\frac{t^{2}}{2\tau_{c}^{2}}\right), \tag{3}\] where \(\lambda_{i}\) are the correlation lengths of the 2-dimensional potential and \(\tau_{c}\) is the correlation time. The EC's of the velocity components \(E_{ii}({\bf x},\!t)\equiv\langle v_{i}({\bf x}_{0},\!t_{0})\ v_{i}({\bf x}_{0}+ {\bf x},\!t_{0}+t)\rangle\) are \[E_{11}({\bf x},\!t)=-\partial_{2}\partial_{2}E({\bf x},\!t),\ E_{22}({\bf x},\!t )=-\partial_{1}\partial_{1}E({\bf x},\!t), \tag{4}\] which determine the normalized amplitudes of the velocity fluctuations \(V_{1}=\sqrt{E_{11}({\bf 0},\!0)}=1/\lambda_{2}\), \(V_{2}=1/\lambda_{1}\). The statistical properties of the trajectories obtained from Eq. (1) are numerically analyzed. More precisely, we determine the statistics of the trajectories and of the Lagrangian velocity, and a class of conditional Lagrangian correlations that reveal the quasi-coherent components of the motion and their properties. We use statistical averages, which consists of generating a large number of realizations (\(r\)) of the stochastic Gaussian potential and of determining the trajectory with the initial condition \({\bf x}(0)=0\) in each \(r\) by effectively computing the velocity on the trajectory at each time step, \({\bf v}({\bf x}(t_{i}),t_{i}).\) However, the analysis of the results is connected to the equivalent space averaging procedure. This corresponds to the statistical ensemble of trajectories obtained in a single typical realization of the potential by different initial conditions \({\bf x}(0)={\bf x}_{0}^{r},\) where the points \({\bf x}_{0}^{r}\) are uniformly distributed in a very large domain. We use the simulation code presented in [13], which is based on a fast generator of Gaussian fields with prescribed spectra. In the present work, we have implemented the so called FRD representation \[\phi({\bf X})=\sum_{i=1}^{N_{c}}\sqrt{S({\bf K}_{i})}\sin\left({\bf K}_{i}{ \bf X}+\frac{\pi}{4}\zeta_{i}\right), \tag{5}\] where \({\bf X}\equiv({\bf x},\!t)\) is the three-dimensional space-time and \({\bf K}_{i}\equiv({\bf k}_{\perp}^{i},\omega^{i})\) are the \(N_{c}\) discrete values of the wave numbers \({\bf k}_{\perp}^{i}\) and frequencies \(\omega^{i}.\)\(S({\bf K})\) is the spectrum of the stochastic potential, the Fourier transform of the EC (3). This representation is different of the usual discrete Fourier decomposition by the set of the values of \({\bf K}_{i}\) that are not the fixed points of a three-dimensional mesh, but random values with uniform distribution. Also, the random phases \(\zeta_{i}\) have not continuous distributions, but discrete values \(\pm 1\) (with equal probabilities). Each set of the \(N_{c}\) random values of \({\bf K}_{i}\) and \(\zeta_{i}\) determines a realization \(r\) of the potential and a trajectory (solution of Eq. (1) with initial condition \({\bf x}(0)={\bf 0}\)). The statistical ensemble \(R\) consists of a number \(M\) of these sets. The representation (5) provides a fast convergence of the Eulerian properties of the stochastic potential. We have shown that reasonable errors in the EC and in the probability of the potential are obtained at much smaller values of \(N_{c}\) and \(M\) than in the usual fast Fourier representation (FFR). This leads to the decrease of the computing times by roughly one order of magnitude compared to the usual FFR method in two-dimensional potentials [13]. Most of the simulations analyzed here are performed with \(N_{c}=500\) and \(M=10^{5}.\) ## III Main features of tracer advection The incompressibility of the two-dimensional velocity field (\(\nabla\cdot\mathbf{v}(\mathbf{x}\),\(t)=0\)) determines two invariance laws of the solutions of Eq. (1). It leads to equations of motion of Hamiltonian type, with \(x_{1}\) and \(x_{2}\) conjugate variables and \(\phi_{t}\) the Hamiltonian function. In the case of time independent (frozen) potentials \(\phi\left(\mathbf{x}\right)\), the trajectories are linked to the contour line of the potential \(\phi_{t}(\mathbf{x})\), which means that the Lagrangian potential \(\phi_{t}(\mathbf{x}(t))\) is invariant along each trajectory. The other invariant law is statistical and applies to the motion in both frozen and time dependent potentials \(\phi_{t}(\mathbf{x}(t),t)\) for any value of the correlation time \(\tau_{c}\). It concerns the distribution of the Lagrangian velocity \(\mathbf{v}(\mathbf{x}(t),t)\), that is shown to be time independent, and thus identical with the distribution of the Eulerian velocity \(\mathbf{v}(\mathbf{x},t)\). The Lagrangian potential is statistically invariant too. This property is trivial in frozen potentials where \(\phi_{t}(\mathbf{x}(t))=\phi_{t}(\mathbf{x}(0))\) and it is similar to the case of the velocity in time dependent potentials where \(\phi_{t}(\mathbf{x}(t),t)\) changes in time. An example of configuration of the contour lines of the potential can be seen in Fig. 1, where a typical realization of \(\phi_{t}(\mathbf{x})\) is presented for \(V_{d}=0\) (left panel) and for \(V_{d}=0.3\) (right panel). The contour lines at \(V_{d}=0\) are nested closed curves with multi-scale sizes that have the dimensions \(r_{\max}\) from \(r_{\max}\ll\lambda_{i}\) to \(r_{\max}\to\infty.\) The average velocity \(\mathbf{V}_{d}\) completely changes the field lines by breaking the large size contour lines and generating (winding) open paths along its direction. Islands of closed contour lines remain between the network of open paths, but their average size decreases as \(V_{d}\) increases, and, for \(V_{d}\) much larger than the amplitude of the stochastic velocity, all the lines are open. The average velocity also determines the limitation of the excursion of the contour lines perpendicular to \(\mathbf{V}_{d}\). This configuration of the contour lines of \(\phi_{t}(\mathbf{x})\) determines solutions of Eq. (1) that are, in the presence of an average velocity, a mixture of localized periodic (or trapped) trajectories that are closed, and of free trajectories that have unlimited displacements along \(\mathbf{V}_{d}\). The space of trajectories \(R\) is organized in two disjointed subensembles: \(tr\) for the trapped trajectories and \(fr\) for the free ones (\(R=tr\cup fr\), \(tr\cap fr=\varnothing\)). The classification criterion is the periodicity of the trajectories. A trajectory \(r\) with period \(T_{r}\) belongs to \(tr\) if \(T_{r}\) is finite and to \(fr\) otherwise. \(T_{r}\) is defined as the time of the first return in the initial point \(\mathbf{x}(0)=\mathbf{0}\), and is determined as the first solution of \(r(t)=0\), where \(r(t)=\sqrt{x_{1}^{2}(t)+x_{2}^{2}(t)}.\) Practically, a trajectory belongs to the subensemble \(tr\) when its period is smaller than the time of integration \(t_{\max}\). The size of each trajectory \(r_{\max}=Max(r(t))\) is also calculated. For \(V_{d}=0\), all trajectories \(\mathbf{x}(t)\) are closed, periodic functions of time when \(t\to\infty.\) At finite time \(t\), open trajectories are found, which correspond to large periods \(T_{r}>t\) (and to large size contour lines). As \(t\) increases the fraction of free trajectories decreases, and, in the limit \(t\to\infty\), all trajectories are trapped (\(tr=R\) and \(fr=\varnothing\)). The probability of trajectory sizes \(P(r_{\max},t)\) is represented in Fig. 2 at two time moments, \(t=60\) (dashed line) and \(t=120\) (solid line). One can see that the time evolution of \(P(r_{\max},t)\) affects only the large distances, while the small \(r_{\max}\) domain has invariant probability. The contributions of the closed and open trajectories to \(P(r_{\max},t)\) are also represented Figure 1: Typical realization of the potential \(\phi_{t}(\mathbf{x})\) for \(V_{d}=0\) (left panel) and for \(V_{d}=0.3\) (right panel). in the figure. The closed trajectories (red points) determine the invariant part of \(P(r_{\max},t).\) The open trajectories (green points) have large sizes and they determine the time variation of \(P(r_{\max},t).\) Their contribution move toward larger \(r_{\max}\) as time increases and it decays, such that \(P(r_{\max},t)\) is determined only by closed trajectories in the limit \(t\rightarrow\infty\). It is a decaying function of \(r_{\max}\) that scales as \(P(r_{\max},t)\sim r_{\max}^{-1.3}\) at large \(t.\) The slow algebraic decay of the asymptotic probability shows that the sizes of the trajectories cover multiple scales from \(r_{\max}\ll\lambda_{i}\) to \(r_{\max}\rightarrow\infty.\) Thus, the invariance of the Lagrangian potential determines a process of trajectory trapping manifested by eddying in the structure of \(\phi\left(\mathbf{x}\right).\) The average velocity \(V_{d}\) that strongly modifies the structure of the field lines of \(\phi_{t}\left(\mathbf{x}\right)\) determines a significant change of the probability of trajectory sizes. Two categories of trajectories coexist for \(V_{d}\lesssim V:\) periodic, closed trajectories situated on the islands of closed contour lines of \(\phi_{t}\left(\mathbf{x}\right)\) and non-periodic trajectories along the open paths generated by the average potential \(xV_{d}.\) The latter are free trajectories that make large displacements along \(\mathbf{V}_{d}.\) The probability \(P(r_{\max},t)\) can be written as the sum of the contributions of these two types of trajectories \[P(r_{\max},t)=n_{tr}(r_{\max},t)+n_{fr}(r_{\max},t), \tag{6}\] where \(n_{tr},\)\(n_{fr}\) are determined in the subensembles \(tr\) and \(fr\) at time \(t.\)\(P(r_{\max},t),\) shown in Fig. 3 (left panel) has a second maximum. It appears at a large value of \(r_{\max}\) that increases with the increase of \(V_{d}.\) Also, the amplitude and the width of this peak increase with \(V_{d}.\) It is determined by the free trajectories. The narrow peak \(n_{tr}(r_{\max},t)\) at small \(r_{\max}\) is the contribution of the trapped, periodic trajectories. It is represented in the right panel of Fig. 3, which shows that both the maximum size and the amplitude of the trapped trajectories decrease as \(V_{d}\) increases. The average velocity hinders and eventually eliminates the trapping process. The contribution \(n_{tr}(r_{\max},t)\) in Eq. (6) decreases with \(V_{d}\) and become negligible at \(V_{d}\gg 1.\) The contribution of the free trajectories is negligible in this range of small sizes, at any \(V_{d},\) as shown in Fig. 3 (right panel) where the black points for \(P(r_{\max},t)\) are superposed on the red curves for \(n_{tr}(r_{\max},t).\) The two contributions in Eq. (6) separates at large time. The probability of the periods of the closed trajectories \(P(T,t)\) calculated from the trajectories \(\mathbf{x}(t)\) at \(t=60\) is shown in Fig. 4. One can see that, at small \(V_{d},\) this probability extends to large values of \(T\lesssim 100\) and it has a weak decay. As \(V_{d}\) increases, the width of \(P(T,t)\) decreases and its decay is steeper. This behavior is in agreement with the decay of the trajectory sizes at large \(V_{d}.\) An average velocity can be defined for the trapped trajectories as the maximum displacement over the period, \(v^{eff}=r_{\max}/T.\) Its probability is weakly dependent on the average velocity. The fraction of trajectories that are not closed at time \(t,\)\(n_{fr}(t,V_{d})\) is obtained from the probability of the periods of the closed trajectories (calculated at the time of integration, \(t_{\max}\)) \[n_{fr}(t,V_{d})=1-n_{tr}(t,V_{d}),\text{ \ \ }n_{tr}(t,V_{d})=\int_{0}^{t}P(T,t_{ \max})\text{ }dT. \tag{7}\] This function decreases in time from \(n_{fr}(0,V_{d})=1,\) as seen in Fig. 5 (left panel). In the case \(V_{d}\neq 0,\)\(n_{fr}(t,V_{d})\) saturates at a value \(n_{fr}(V_{d})\) in a time that becomes shorter at larger \(V_{d}\). In the case of \(V_{d}=0,\) the decay is not limited, and it scales as \(n_{fr}(t,0)\sim t^{-0.6}.\) Figure 2: The probability of trajectory sizes \(P(r_{max},t)\) for \(V_{d}=0\) at \(t=60\) (dashed black line) and \(t=120\) (solid black line). Also shown are the contributions of the trapped (red points) and free (green points) trajectories at \(t=120.\) The results obtained for the asymptotic fraction of free trajectories \(n_{fr}(V_{d})\equiv\lim\limits_{t\to\infty}n_{fr}(t,V_{d})\), presented in Fig. 5 (right panel), are well approximated by \[n_{fr}(V_{d})=\left[1-\exp\left(-V_{d}^{2}\right)\right]^{1/4}. \tag{8}\] The fraction of trapped trajectories is \(n_{tr}(t,V_{d})=1-n_{fr}(t,V_{d})\) at any time, with the asymptotic value \(n_{tr}(V_{d})=1-n_{fr}(V_{d})\) that is also represented in Fig. 5 (right panel). ## IV 4. Lagrangian statistics in static potentials Thus, the trajectories obtained in the stochastic potential \(\phi_{t}(\mathbf{x})\) were divided into two categories: trapped and free. They have different topologies and different sizes, which suggests that their contributions to the global statistical properties of the trajectories are qualitatively different. We analyze here the statistics of the Lagrangian velocity and of the displacements of each category of trajectories. For any Lagrangian quantity \(A(\mathbf{x}(t))\), we determine \(\left\langle A(\mathbf{x}(t))\right\rangle_{tr}\) and \(\left\langle A(\mathbf{x}(t))\right\rangle_{fr}\) that are conditional averages restricted to the trapped and free trajectories, respectively. These are statistical averages calculated over the subspaces \(tr\) and \(fr.\) The contribution of each subensemble to the global average (over \(R\)) is the product of the probability that a trajectory belongs to the subensemble multiplied by the statistical average over the subensemble, \(n_{c}(t,V_{d})\left\langle A(\mathbf{x}(t))\right\rangle_{c},\) where \(c=tr,\;fr.\) It yields Figure 4: The probability of the periods of the trapped trajectories \(P(T,t)\) as functions of \(T\) at \(t=60\) and at the values of \(V_{d}\) that label the curves. Figure 3: The probability \(P(r_{max},t)\) for several average velocities \(V_{d}\) that label the curves, at \(t=60\) as function of \(r_{max}\) (the curves in the left panel and black points in the right panel) and the contribution of the trapped trajectories \(n_{tr}(r_{max},t)\) (right panel, red lines). \[\left\langle A({\bf x}(t))\right\rangle=n_{tr}(t,V_{d})\ \left\langle A({\bf x}(t)) \right\rangle_{tr}+n_{fr}(t,V_{d})\ \left\langle A({\bf x}(t))\right\rangle_{fr}. \tag{9}\] The separation of the trajectories in these categories is performed at a large time such that \(n_{fr}(t,V_{d})\) is saturated (see Fig. 5, left panel). ### 4.1 Statistics of the Lagrangian velocity The statistical parameters of the Lagrangian velocity \({\bf v}\left({\bf x}(t)\right)\equiv{\bf v}(t)\) are shown in Fig. 6 for a stochastic potential with \(\lambda_{1}=1\), \(\lambda_{2}=2\) and \(V_{d}=0.2.\) The average Eulerian velocity and fluctuation amplitudes are in this case \(\left\langle v_{1}\right\rangle=0,\ \left\langle v_{2}\right\rangle=V_{d}\), \(V_{1}=0.5\) and \(V_{2}=1\), where \(V_{i}=\sqrt{\left\langle\widehat{v}_{i}^{2}\right\rangle}\) are obtained from Eq. (4). The Lagrangian quantities maintain the Eulerian values at any time, as stated by Lumley theorem. Besides this, the conditional average velocity and fluctuation amplitudes are time invariant, as seen in Fig. 6, but their values depend on the category. It is interesting to note that the average velocity is determined only by the free trajectories, while the trapped trajectories do not contribute (\(\left\langle v_{2}(t)\right\rangle_{tr}=0\) at any time). The average velocity of the free trajectories is larger than \(V_{d}\), and it can be approximated with \[\left\langle v_{2}(t)\right\rangle_{fr}=\frac{V_{d}}{n_{fr}}>V_{d} \tag{10}\] Figure 5: Left panel: the fractions of free trajectories as function of time for the values of \(V_{d}\) that label the curves. Right panel: the asymptotic values \(n_{fr}\) and \(n_{tr}\) and the average velocity of the free trajectories (see next Section) as functions of \(V_{d}\). Figure 6: The average Lagrangian velocities (left panel) and the fluctuations of the Lagrangian velocities (right panel) as functions of time. The dashed lines are for the \(v_{1}\) and the solid lines for the \(v_{2}\). The green lines are for the free trajectories and the red lines for the trapped trajectories, while the black are averages on the whole statistical ensemble \(R\). \(V_{d}=0.2\). It is \(\left\langle v_{2}(t)\right\rangle_{fr}=0.45\) for the example presented in Fig. 6, left panel, obtained for \(V_{d}=0.2\). The conditional average velocity \(\left\langle v_{2}(t)\right\rangle_{fr}\) is also shown in Fig. 5 (right panel) as function of \(V_{d}.\) One can see that this average velocity is significantly larger than \(V_{d}\) only in the presence of trajectory trapping (for \(V_{d}\lesssim 1\)). This result shows that a supplementary ordered component of the Lagrangian velocity appears for the free trajectories that exactly compensates the missing contribution of the trapped particles, such that \(\left\langle v_{2}(t)\right\rangle=n_{fr}\left\langle v_{2}(t)\right\rangle_{ fr}=V_{d}.\) It seems to be a trivial consequence of \(\left\langle v_{2}(t)\right\rangle_{tr}=0,\) but the underlying physical process is rather complex. It essentially consists of generation of ordered motion from the stochastic velocity \(\widetilde{\mathbf{v}}(\mathbf{x},\)\(t)\) for both types of trajectories \[\left\langle\widetilde{v}_{2}(t)\right\rangle_{tr}=-V_{d},\ \ \left\langle \widetilde{v}_{2}(t)\right\rangle_{fr}=V_{d}\frac{n_{tr}}{n_{fr}}. \tag{11}\] The supplementary average velocity of the trapped trajectories is opposite to \(\mathbf{V}_{d}\) and exactly compensates it. The supplementary average velocity of the free trajectories is along \(\mathbf{V}_{d}\) and it contributes to the increase of the Lagrangian over the Eulerian velocity. Equations (11) are valid at any time, including \(t=0.\) It can be interpreted as the condition for the separation of the trajectories in the free and trapped categories. The trapped trajectories start from the geometric locus for which \(\left\langle\widetilde{v}_{2}(\mathbf{x})\right\rangle=-V_{d}\) and they remain in this domain, while the free trajectories are confined in the complement of this domain. These ordered components of the motion are hidden, in the sense that they are not "seen" in the average velocity calculated on the whole ensemble \(R\) (\(\left\langle v_{2}(t)\right\rangle=V_{d}).\) However, as shown below, they have strong effects on the transport along \(\boldsymbol{V}_{d}\) through the modification of the correlation of the Lagrangian velocity. Figure 8: The correlations of the Lagrangian velocity \(v_{1}(t)\) (left panel) and \(v_{2}(t)\) (right panel). The correlations on the whole statistical ensemble \(L_{i}(t)\) (black lines) are compared to the subensemble correlations \(L_{i}^{tr}(t)\) (red lines) and \(L_{i}^{fr}(t)\) (green lines). \(V_{d}=0.2\). Figure 7: The amplitudes of fluctuations of the Lagrangian velocities as functions of \(V_{d}\). The dashed lines are for the \(v_{1}\) and the solid lines for the \(v_{2}\). The green lines are for the free trajectories and the red lines for the trapped trajectories, while the black are averages on the whole statistical ensemble \(R.\) The amplitudes of velocity fluctuations around the average velocity are shown in Fig. 6 (right panel). They are different for the two types of trajectories. It is interesting to underline that the supplementary order that characterizes trapped and free trajectories appears in the fluctuations of the velocity in \(R\). The average of the square velocity decomposed on \(tr\) and \(fr\) subensembles according to Eq. (9) for large time \[\left\langle v_{i}^{2}(t)\right\rangle=n_{tr}\left\langle v_{i}^{2}(t)\right\rangle _{tr}+n_{fr}\left\langle v_{i}^{2}(t)\right\rangle_{fr}, \tag{12}\] leads to \[V_{2}^{2}=n_{tr}(V_{2}^{tr})^{2}+n_{fr}(V_{2}^{fr})^{2}+\frac{n_{tr}}{n_{fr}}V_ {d}^{2}, \tag{13}\] where \[(V_{i}^{c})^{2}\equiv\left\langle\left(v_{i}(t)-\left\langle v_{i}(t)\right \rangle_{c}\right)^{2}\right\rangle_{c} \tag{14}\] are the amplitudes of the fluctuations of the velocity \(\delta v_{i}(t)\equiv v_{i}(t)-\left\langle v_{i}(t)\right\rangle_{c},\)\(i=1,2,\) conditioned by the category of trajectories \(c=tr,\)\(fr\) (on the subensembles \(tr\) and \(fr\)). Thus, a contribution produced by the ordered motion appears (the last term of Eq. (13)) besides the direct contributions of the conditional fluctuations. It is determined by the ordered motion (11) generated by \(V_{d}\) in the presence of trapping (for \(V_{d}\lesssim 1).\) The results presented in Fig. 6 (right panel) show values \(V_{2}^{tr}<V_{2}\) and \(V_{2}^{fr}<V_{2},\) which reproduce Eq. (13). The conditioned amplitudes of the velocity fluctuations \(V_{i}^{c}\) depend on the average velocity \(V_{d}.\) As seen in Fig. 7, the amplitudes of both components of the trapped trajectory velocity (red curves) are continuously decreasing functions of \(V_{d}\). This is the effect of the decrease of the size of the islands of closed contour lines of the potential, which, as \(V_{d}\) increases, shrink around the maxima and minima of \(\phi(\mathbf{x})\) where the gradients are small. In the case of free trajectories (green lines), the amplitudes of the Lagrangian velocities are different of \(V_{i}\) only in the range of \(V_{d}\) that corresponds to the existence of islands of closed contour lines of the potential. The perpendicular amplitude is increased (\(V_{1}^{fr}>V_{1}),\) while the parallel amplitude is decreased (\(V_{2}^{fr}<V_{2}\)) such that the supplementary parallel velocity is compensated (Eq. (13)). One can deduce from these results that the EC defined on the geometric locus of the free trajectories is different of the EC (3). As shown below (Section 6), the amplitude of the stochastic potential of the free trajectories \(\Delta\) is smaller than in the whole space (\(\Delta<\Phi\)). The correlation lengths are evaluated using the amplitudes of velocity fluctuations of the free trajectories \[\lambda_{1}^{fr}\sim\frac{\Delta}{V_{2}^{fr}}=\lambda_{1}\frac{\Delta}{\Phi} \frac{V_{2}}{V_{2}^{fr}},\;\lambda_{2}^{fr}\sim\frac{\Delta}{V_{1}^{fr}}= \lambda_{2}\frac{\Delta}{\Phi}\frac{V_{1}}{V_{1}^{fr}}, \tag{15}\] where \(\lambda_{1}=\Phi/V_{2}\) and \(\lambda_{2}=\Phi/V_{1}.\) Thus, the correlation lengths on the domain of free trajectories decrease with the factor \(\Delta/\Phi\) on both directions and are modified by the velocity amplitudes (decreased along \(\mathbf{V}_{d}\) and increased across \(\mathbf{V}_{d}\)). Figure 9: Histograms of the Lagrangian velocities \(v_{1}\) (left panel) and \(v_{2}\) (right panel) represented by the black curves and the contributions determined by the free (green curves) and trapped (red curves) trajectories. \(V_{d}=0.4.\) The correlations of the Lagrangian velocity are shown in Fig. 8, where the notations are \[L_{i}(t)=\left\langle\delta v_{i}(0)\ \delta v_{i}(t)\right\rangle,\ L_{i}^{c}(t)= \left\langle\delta v_{i}(0)\ \delta v_{i}(t)\right\rangle_{c},\ \ c=tr,fr. \tag{16}\] One can see that all the conditional correlations (for both categories and both components of the velocity) decay to zero at large \(t.\) However, the correlation of the velocity along \({\bf V}_{d}\) calculated on all trajectories, \(L_{2}(t),\) has a finite asymptotic value. It is determined by the ordered components of motion produced in subensembles \(tr\) and \(fr.\) An equation similar to (13) can be obtained from (9) written for \(A=v_{i}(0)\ v_{i}(t)\) \[L_{2}(t)=n_{tr}L_{2}^{tr}(t)+n_{fr}L_{2}^{fr}(t)+\frac{n_{tr}}{n_{fr}}V_{d}^{2}, \tag{17}\] which shows that \(L_{2}(t)\) has a finite asymptotic tail in spite of the decay to zero of \(L_{2}^{tr}(t)\) and \(L_{2}^{fr}(t).\) It is determined by the presence of trapped trajectories at small average velocity \(V_{d}.\) The histograms for the Lagrangian velocity components are time invariant for all statistical ensembles \(R,\)\(tr\) and \(fr.\) The histogram for all trajectories (in \(R\)) is shown in Fig. 9 together with the contributions of the trapped and free trajectories (that include the fractions of trajectories). One can see that the distribution is Gaussian in \(R,\) while significant departures from Gaussianity appear in the subensembles \(tr\) and \(fr,\) especially for the velocity component \(v_{2}\) (right panel). The domain of large positive velocities is dominated by the free trajectories, while the trapped trajectories have the main contribution for the large negative \(v_{2}.\) The most probable value of \(v_{2}\) on \(tr\) (that is slightly negative) is compensated by a longer tail at positive \(v_{2}.\) The non-Gaussian distribution of the Lagrangian velocity of the trapped trajectories provides additional information on the geometric locus of this category of trajectories. It shows that the space average of the parallel velocity \(v_{2}({\bf x})\) on this locus (that is zero for any \(V_{d}\)) results from the elimination of the regions with large, positive \(v_{2}({\bf x}).\) In other words, the regions where the stochastic velocity is oriented parallel to \({\bf V}_{d}\) belong to the geometric locus \(fr.\) ### 4.2 Transport and statistics of trajectories The statistics of the displacements is strongly non-Gaussian, in spite of the Gaussian Lagrangian velocity. Moreover, the average and mean square displacements (calculated for all trajectories and in the subensembles \(tr\) and \(fr\)) have asymptotic regimes that can be linear, quadratic or saturated functions of time, which shows that the transport has anomalous aspects. The average displacements are in agreement with the average Lagrangian velocities \[\left\langle x_{1}(t)\right\rangle = \left\langle x_{1}(t)\right\rangle_{tr}=\left\langle x_{1}(t) \right\rangle_{fr}=0, \tag{18}\] \[\left\langle x_{2}(t)\right\rangle = V_{d}t,\ \left\langle x_{2}(t)\right\rangle_{tr}=0,\ \left\langle x_{2}(t)\right\rangle_{fr}=\frac{V_{d}}{n_{fr}}t.\] The dispersion \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle,\) where \(\delta x_{i}(t)=x_{i}(t)-\left\langle x_{i}(t)\right\rangle\) are shown in Fig. 10 for \(V_{d}=0\) (left panel) and for \(V_{d}=0.2\) (right panel), as functions of time. In the absence of the average velocity (\(V_{d}=0\)), the dispersions are similar along the two directions. The curves in the left panel of Fig. 10 are only translated due to the different amplitudes of the stochastic velocities \(V_{1}=0.5,\ V_{2}=1.\) The dispersions are sub-diffusive, with time increase that is slower than linear \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle\sim t^{0.68}.\) The reason is the progressive saturation of the contributions of the trajectories with small periods. At a time \(t,\) all the trajectories with \(T<t\) have saturated dispersion and only the free trajectories (that are still not closed) determine the time variation of \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle.\) The latter results from two factors with opposite effects: the fraction of free trajectories at time \(t\) and their average size. As seen in Fig. 5, left panel, \(n_{fr}(t,0)\) is a decreasing function of time, \(n_{fr}(t,0)\sim t^{-0.6}.\) The average size of the closed trajectories is an increasing function of \(t,\) because it is an increasing function of the average period. The average velocity \(V_{d}\) makes trajectory dispersion strongly non-isotropic, as seen in Fig. 10, right panel. The dispersion \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle\) for all trajectories (black lines) are compared to the results obtained for the trapped \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle_{tr}\) (red lines) and free \(\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle_{fr}\) (green lines) trajectories. The dispersions across \({\bf V}_{d}\) (of \(x_{1}(t)\)) for the whole set of trajectories and for the subensembles \(tr\) and \(fr\) are all saturated (the dashed curves in Fig. 10, right panel), which corresponds to the minimum sub-diffusive transport. This means that the average velocity completely hinders the perpendicular transport in the case of static stochastic potentials. The contrary happens to the transport parallel to \(\mathbf{V}_{d}\): the dispersion of the trajectories has a very fast time-increase, \(\left\langle\left(\delta x_{2}(t)\right)^{2}\right\rangle\sim t^{2}\), which correspond to the maximum super-diffusive transport that is of the ballistic type. It appears in spite of the much weaker transport of the trapped and free trajectories \(\left(\left\langle\left(\delta x_{2}(t)\right)^{2}\right\rangle_{tr}\right.\) saturates and \(\left\langle\left(\delta x_{2}(t)\right)^{2}\right\rangle_{fr}\sim t\) is diffusive). This super-diffusive parallel transport is the effect of the coherent parallel motion generated by \(V_{d}\), as demonstrated using Eq. (9) for \(A=x_{i}^{2}(t).\) The relations between the dispersion of all trajectories (in \(R\)) and the subensemble \(tr\) and \(fr\) dispersions are \[\left\langle\left(\delta x_{1}(t)\right)^{2}\right\rangle=n_{tr}\left\langle \left(\delta x_{1}(t)\right)^{2}\right\rangle_{tr}+n_{fr}\left\langle\left( \delta x_{1}(t)\right)^{2}\right\rangle_{fr}, \tag{19}\] \[\left\langle\left(\delta x_{2}(t)\right)^{2}\right\rangle=n_{tr}\left\langle \left(\delta x_{2}(t)\right)^{2}\right\rangle_{tr}+n_{fr}\left\langle\left( \delta x_{2}(t)\right)^{2}\right\rangle_{fr}+\frac{n_{tr}}{n_{fr}}V_{d}^{2}\ t^{2}. \tag{20}\] The last term in Eq. (20) is dominant at large time and it makes the asymptotic regime superdiffusive of ballistic type. This term is determined by the supplementary average velocity generated from the stochastic components for the free and trapped trajectories, Eq. (11). It leads to the "concentration" of the average velocity along the free trajectories Eq. (10). Thus, the super-diffusive parallel transport is determined by the average velocity (\(V_{d}\neq 0\)) only in the presence of the islands of trapped trajectories (\(n_{tr}\neq 0\)), which corresponds to \(V_{d}\lesssim 1\). The dispersions of the trajectories (Fig. 10) are connected to the correlations of the Lagrangian velocity (Fig. 8) and to the time dependent diffusion coefficients, defined by \(2D_{i}(t)=d\left\langle\left(\delta x_{i}(t)\right)^{2}\right\rangle/dt,\) by Taylor formulas [14] \[D_{i}(t)=\int_{0}^{t}L_{i}(\tau)\ d\tau,\ \ \left\langle\left(\delta x_{i}(t) \right)^{2}\right\rangle=2\int_{0}^{t}\left(t-\tau\right)\ L_{i}(\tau)\ d\tau. \tag{21}\] Similar equations can be written for each category of trajectories (trapped or free). Figure 11 presents the time dependent diffusion coefficients compared to their restrictions to the trapped and free trajectories for the \(x_{1}\) (left panel) and \(x_{2}\) (right panel) directions. This confirms that the perpendicular diffusion is completely hindered (even for the free trajectories). The time integral of \(L_{1}(t)\) vanishes for all categories at a finite time. The parallel transport is ballistic \(D_{2}(t)\sim t\), in spite of the normal diffusion of the free trajectories and of the total confinement of the trapped ones. It is the result of the ordered parallel motion, as seen by performing the time derivative in Eq. (20). The probability of the displacements \(P(\mathbf{x},t)\) is strongly non-Gaussian, as seen in in Fig. 12 (black curves). The contributions of the two categories of trajectories are completely different: the trapped trajectories determine the steep peak in \(\mathbf{x}=\mathbf{0}\), and the free ones have a large Gaussian distribution with the average parallel displacement \(\left\langle x_{2}(t)\right\rangle_{fr}=V_{d}t/n_{fr}.\) We note that the transport, which is essentially produced by the free trajectories, results from a Gaussian distribution. Figure 10: The dispersions of the trajectories on the whole statistical ensemble (black line) for \(V_{d}=0\) (left panel) and \(V_{d}=0.2\) (right panel) as functions of time. The dashed lines are for \(x_{1}(t)\) and the solid lines for \(x_{2}(t)\). The conditional dispersions for trapped (red) and free (green) trajectories are also shown in the right panel. ## V Coherence induced by an average velocity The Hamiltonian structure of equation (1) is the origin of the order that characterizes the two-dimensional incompressible turbulence. It determines the strong connection between trajectories and the contour lines of the potential, which are paths of the motion. The order (quasi-coherence) of the motion is essentially represented by the existence of correlations between the potential and the trajectories. They are represented by nonzero average displacements or velocities conditioned by the (initial) potential. Significant quasi-coherent characteristics of the transport process can be found by analyzing statistical Lagrangian quantities restricted on the contour lines with given potential \(\phi^{0}.\) The trajectories that belong to this class correspond to solutions of Eq. (1) that start (in \(\mathbf{x}(0)=\mathbf{0}\)) from points with \(\phi(\mathbf{0})=\phi^{0}.\) The invariance of the total potential in this class gives \[\phi_{t}(\mathbf{x}(t))=\phi(\mathbf{x}(t))+x_{1}(t)V_{d}=\phi^{0}. \tag{22}\] The fractions of trajectories that evolve on the \(\phi^{0}\) potential lines, the average and the amplitude of fluctuations of their displacements and Lagrangian velocities are determined below for each type of trajectories using conditional averages. The analysis starts from the representation (9) and introduces a supplementary condition for the trajectories, namely that the initial potential is \(\phi^{0}\) [\(\phi(\mathbf{x}(0))=\phi^{0}\)]. Defining the fraction of these trajectories by \(n(\phi^{0})\) and the Figure 11: The time dependent diffusion coeficients in the direction perpendicular (left panel) and parallel (right panel) to the average velocity for the whole statistical ensemble \(R\) (black lines) and restricted to the \(tr\) (red lines) and \(fr\) (green lines) subensembles. \(V_{d}=0.2.\) Figure 12: The probabilities of the \(x_{1}\) (left panel) and \(x_{2}\) (right panel) for the whole set of realizations (black lines) compared to the contributions of the free (green lines) and trapped (red lines) trajectories for \(V_{d}=0.2\) and \(t=97.\) corresponding conditional average by \(\langle\rangle_{\phi^{0}}\), the average \(\langle A({\bf x}(t))\rangle\) is the sum of the contributions from each value \(\phi^{0}\) \[\langle A({\bf x}(t))\rangle=\int_{-\infty}^{\infty}\left\langle A({\bf x}(t)) \right\rangle_{\phi^{0}}\ P(\phi^{0})\ d\phi^{0}, \tag{23}\] where \(P\left(\phi^{0}\right)\) is the Gaussian distribution of the (normalized) potential \[P\left(\phi^{0}\right)=\frac{1}{\sqrt{2\pi}}\exp\left(-\frac{\left(\phi^{0} \right)^{2}}{2}\right). \tag{24}\] Similar equations can be written for the contributions of the free and trapped trajectories \[n_{c}\ \left\langle A({\bf x}(t))\right\rangle_{c}=\int_{-\infty}^{\infty} \left\langle A({\bf x}(t))\right\rangle_{\phi^{0},c}\ n^{c}(\phi^{0})\ d\phi^{0}, \tag{25}\] where \(n^{c}(\phi^{0})\) is the fraction of trajectories that evolve on contour lines \(\phi^{0}\) and are in the category \(c=tr,\ fr,\) and \(\langle\rangle_{\phi^{0},c}\) is the conditional average taken on the subensemble of these trajectories. \(n^{c}(\phi^{0})\) is related to \(n_{tr},\)\(n_{fr}\) (defined in Section 3) \[n_{c}=\int_{-\infty}^{\infty}n^{c}(\phi^{0})\ d\phi^{0}. \tag{26}\] One obtains using Eq. (9) \[\left\langle A({\bf x}(t))\right\rangle_{\phi^{0}}P(\phi^{0})=\left\langle A( {\bf x}(t))\right\rangle_{\phi^{0},tr}\ n^{tr}(\phi^{0})+\ \left\langle A({\bf x}(t))\right\rangle_{\phi^{0},fr}\ n^{fr}(\phi^{0}), \tag{27}\] which connects the contribution of all trajectories \(\left\langle A\right\rangle_{\phi^{0}}P(\phi^{0})\) to the contributions of each category \(\left\langle A\right\rangle_{\phi^{0},c}n^{c}(\phi^{0}).\) The fractions of trajectories fulfil the equation \[P(\phi^{0})=n^{tr}(\phi^{0})+n^{fr}(\phi^{0}), \tag{28}\] which is obtained from Eq. (27) for \(A=1.\) The numerical results obtained for \(n^{fr}(\phi^{0})\) and \(n^{tr}(\phi^{0})\) (represented by points), are compared to analytical approximations (solid lines) in Fig. 13. One can see that the fraction of trajectories that evolve on \(\phi^{0}\) contour lines (black line and points in Fig. 13) reproduces Eq. (24). The fraction of the free trajectories is narrower, but it is still Gaussian. We have found, as seen in Fig. 13 (green curve), a good approximation of the data by \[n^{fr}(\phi^{0})=n_{fr}\ G(\phi^{0};\Delta), \tag{29}\] Figure 13: The fraction of the trajectories that evolve on the contour lines with potential \(\phi^{0}\) for the free (green points), trapped (red points) and for all trajectories (black points) obtained from the numerical simulations compared to \(P(\phi^{0})\) (solid black line), with Eq. (29) (solid green line) and Eq. (31) (solid red line). where \(G(\phi^{0};\Delta)\) is the Gaussian distribution \[G(\phi^{0};\Delta)=\frac{1}{\sqrt{2\pi}\Delta}\exp\left(-\frac{\left(\phi^{0} \right)^{2}}{2\Delta^{2}}\right) \tag{30}\] with a width \(\Delta\) that depends on the average velocity \(V_{d}.\) The fraction of the trapped trajectories (red curve in Fig. 13), which according to Eq. (28) is \[n^{tr}(\phi^{0})=P(\phi^{0})-n_{fr}\ G(\phi^{0};\Delta), \tag{31}\] provides a good representation of the numerical results (red points). The width \(\Delta\) as function of the average velocity \(V_{d}\) is shown in Fig. 14 together with the fraction of free trajectories \(n_{fr}.\) The numerical results for \(\Delta(V_{d})\) (diamonds) are well approximated by \[\Delta(V_{d})=\left[1-\exp\left(-V_{d}^{2}\right)\right]^{0.17}. \tag{32}\] Both functions saturate at large \(V_{d}\) (\(V_{d}>V_{1},V_{2}\)), and they have power law dependence for small \(V_{d}\) \[n_{fr}\sim V_{d}^{0.5},\ \Delta\sim V_{d}^{0.34}. \tag{33}\] The asymptotic value \(n_{fr}\to 1\) for \(V_{d}\rightarrow\infty\) corresponds to the complete elimination of the islands of close contour lines of the potential (\(n_{tr}=0\)). The limit \(\Delta\to 1\) confirms that all trajectories are free, because \(G(\phi^{0};1)=P(\phi^{0})\), where \(P(\phi^{0})\) is the probability of the Eulerian potential. Thus, the free trajectories are localized on the contour lines with small values of \(\left|\phi^{0}\right|\lesssim\Delta.\) The potential on the geometrical locus of free trajectories is Gaussian with an amplitude that is smaller than in the whole space. The trapped (periodic) trajectories mainly have large \(\left|\phi^{0}\right|\) : they completely occupy the range of large potential \(\left|\phi^{0}\right|\gg\Delta\), but also have significant presence at small potential \(\left|\phi^{0}\right|\lesssim\Delta\) that correspond to free trajectories. The average displacements conditioned by the value of the initial potential and by the category of the trajectories are shown in Fig. 15 for free (green), trapped (red) and all (black) trajectories. The perpendicular (left panel) and the parallel (right panel) displacements are shown at a large time \(t=97\), larger than the saturation time of \(n_{fr}(t,V_{d})\) (seen in Fig. 5, left panel). These represent quasi-coherent components of the motion, and appear only in the presence of an average velocity. One can see that the average conditional displacements are small for the trapped trajectories (red points), and that significant values appear for the free trajectories in both directions (green points). As shown in Fig. 15, these quantities can be approximated by \[\left\langle x_{1}(t)\right\rangle_{\phi^{0},fr}=\frac{\phi^{0}}{V_{d}},\ \ \left\langle x_{1}(t)\right\rangle_{\phi^{0},tr}\cong 0, \tag{34}\] Figure 14: The width \(\Delta\) of the initial potential of the free trajectories (diamonds) and the fraction of free trajectories (circles) as functions of the average velocity \(V_{d}\). The numerical results are interpolated by Eqs. (32) and (8) \[\left\langle x_{2}(t)\right\rangle_{\phi^{0},fr}=\frac{V_{d}t}{n_{fr}},\ \ \left\langle x_{2}(t)\right\rangle_{\phi^{0},tr}=0, \tag{35}\] which are represented by red and green lines respectively. The black lines have the equations \[\left\langle x_{1}(t)\right\rangle_{\phi^{0}}=\frac{\phi^{0}}{V_{d}}F(\phi^{0} ),\ \ \left\langle x_{2}(t)\right\rangle_{\phi^{0}}=\frac{V_{d}t}{n_{fr}}F(\phi^{0}), \tag{36}\] where \[F(\phi^{0})=\frac{n^{fr}(\phi^{0})}{P(\phi^{0})}=\frac{n_{fr}}{\Delta}\exp \left(-\frac{\left(\phi^{0}\right)^{2}}{2}\frac{1-\Delta^{2}}{\Delta^{2}}\right). \tag{37}\] They result from Eq. (27) with \(A=x_{i}(t)\) using (34) and (35), and provide, as seen in Fig. 15, good approximations of the data for \(\left\langle x_{i}(t)\right\rangle_{\phi^{0}}\) (black points). The contributions to the average displacements, obtained by multiplying the conditional averages with the corresponding fractions of trajectories (\(P(\phi^{0})\), \(n^{fr}(\phi^{0})\) or \(n^{tr}(\phi^{0})\)), are shown in Fig. 16. It appears more clearly that the coherent displacements are produced only by the free trajectories. The black points for all trajectories are practically superposed on the green points and they are well approximated by the green lines that represent the contributions of Figure 16: The contributions of the trapped (red points) and free (green points) trajectories, to the average displacements (black points) along \(x_{1}\) (left panel) and \(x_{2}\) (right panel) directions as functions of \(\phi^{0}\) for the values of \(V_{d}\) label the curves. The approximations for the free trajectories are represented by the green lines. Figure 15: The conditional average displacements along \(x_{1}\) axis (left panel) and along \(x_{2}\) axis (right panel) as functions of \(\phi^{0}\) for the trapped (red points), free (green points) and all (black points) trajectories, compared to the approximations (34)-(35) (green lines) and (36) (black lines). \(V_{d}=0.3\). the free trajectories (\(\phi^{0}/V_{d}\;n^{fr}(\phi^{0})\) in the left panel and \(V_{d}\;t/n_{fr}\;n^{fr}(\phi^{0})\) in the right panel). The dependence on \(\phi^{0}\) is different across and along the average velocity \(\mathbf{V}_{d}.\) In the first case, it is an anti-symmetrical function of \(\phi^{0}\) that saturates in time, and, in the second case, it is a symmetrical Gaussian function that increases linearly in time. The parallel average displacement on the contour lines with initial potential \(\phi^{0},\) which increases linearly with \(t\) (35), leads to an average Lagrangian velocity \[\left\langle v_{2}(t)\right\rangle_{\phi^{0},fr}=\frac{V_{d}}{n_{fr}}. \tag{38}\] It is important to note that this velocity does not depend on \(\phi^{0},\) and it equals the average velocity (10). The contribution of the conditional average velocity is determined only by the free trajectories since \(\left\langle v_{2}(t)\right\rangle_{\phi^{0},tr}=0.\) The perpendicular average displacement of the free trajectories also determines an average velocity, but it is transitory since \(\left\langle x_{1}(t)\right\rangle_{\phi^{0},fr}\) saturates in time. The dispersion of the trajectories conditioned by the value of the initial potential and by the category are shown in Fig. 17 for free (green), trapped (red) and all (black) trajectories, in the perpendicular (left panel) and parallel (right panel) directions. One can see that the dispersion of the free trajectories along both directions are not dependent on the initial potential \(\phi^{0},\) and can be approximated by \[\left\langle\delta x_{1}^{2}\right\rangle_{\phi^{0},fr}=\frac{\Delta^{2}}{V_{ d}^{2}},\;\;\left\langle\delta x_{2}^{2}\right\rangle_{\phi^{0},fr}=2\;D_{2}^{fr}t \tag{39}\] represented by the green lines. On the contrary, the trapped trajectories have dispersions that decay with the increase of \(\phi^{0}\) (the red points). The dispersion of the trajectories conditioned by the potential is obtained using Eq. (27) for \(A=x_{i}^{2}(t)\) \[\left\langle\delta x_{i}^{2}\right\rangle_{\phi^{0}}=\left\langle\delta x_{1} ^{2}\right\rangle_{\phi^{0},fr}F+\left\langle\delta x_{1}^{2}\right\rangle_{ \phi^{0},tr}(1-F)+\left\langle x_{i}(t)\right\rangle_{\phi^{0},fr}^{2}F(1-F), \tag{40}\] which depends on \(\phi^{0}\) as seen in Fig. 17 (black points). Thus, the analysis of the Lagrangian statistics conditioned by the initial potential reveals a coherent component of motion perpendicular to \(\mathbf{V}_{d}\). It consists of average displacements that have the sign correlated with the sign of \(\phi^{0}.\) They appear for the free trajectories and are hidden in the sense that the contributions of all contour lines (with all values of \(\phi^{0}\)) mix to zero. The amplitude of the ordered motion is defined by the displacements conditioned by the sign of \(\phi^{0},\)\(\left\langle x_{1}(t)\right\rangle_{+}\) obtained by integration over \(\phi^{0}\) on the interval \([0,\infty)\) and \(\left\langle x_{1}(t)\right\rangle_{-}\) that is the integral over \((-\infty,0].\) These are symmetrical quantities since \(\left\langle x_{1}(t)\right\rangle=\left\langle x_{1}(t)\right\rangle_{+}+ \left\langle x_{1}(t)\right\rangle_{-}=0.\) The time derivatives of these functions determine a pair of average velocities with opposite directions that exactly compensate each other (the hidden drifts, HDs) that are oriented across \(\mathbf{V}_{d}\). The HDs were first found in [15] using an approximate theoretical approach, the decorrelation trajectory method [16]. In the presence of components of the motion that introduce a Figure 17: The conditional trajectory dispersion along \(x_{1}\) axis (left panel) and along \(x_{2}\) axis (right panel) as functions of \(\phi^{0}\) for the trapped (red points), free (green points) and all (black points) trajectories, compared to the approximations (39) (green lines). \(V_{d}=0.3\) and \(t=97.\) small compressibility, an average velocity can be generated by breaking the equilibrium of the HDs. Such effects were found in magnetically confined plasmas [17]-[19]. The HDs are transitory in frozen potentials because the average displacements saturate \(\left\langle x_{1}(t)\right\rangle_{\phi^{0},fr}\rightarrow\phi^{0}/V_{d}.\) The analysis also shows that the parallel motion is similar on the contour lines with different \(\phi^{0},\) and that it depends only on the category of trajectories. The asymptotic value \(\left\langle x_{1}(t)\right\rangle_{\phi^{0},fr}\rightarrow\phi^{0}/V_{d}\) represents the centrum of the space domain that contains the trajectories, which start from points on the line \(x_{1}=0\) where the potential is \(\phi^{0}.\) It is limited by the lines \(x_{1}^{-}=(\phi^{0}-\Delta)/V_{d},\)\(x_{1}^{+}=(\phi^{0}+\Delta)/V_{d},\) and has infinite dimension along \({\bf V}_{d}.\) Reported at this centrum, the free trajectories are statistically identical for all values of the initial potential \(\phi^{0}.\) ## VI 6. Lagrangian statistics in time dependent potentials The general conclusion of the analysis of trajectory statistics in frozen potentials is the existence of a high degree of coherence, which reflects the structure of the contour lines of \(\phi_{t}({\bf x})\) on which the trajectories are bounded. The time-dependence of the potential determines the variation of the Lagrangian potential and the decorrelation from its initial value \(\phi^{0}\). It is expected to strengthen the random aspects of the trajectories and to cause the elimination of the Lagrangian coherence in a time of the order of the decorrelation time \(\tau_{c}\). More precisely, the random time-variation of the potential should vanish the averages and correlations conditioned by \(\phi^{0},\) which show the existence of hidden order. It is thus expected that the order found in static potential is in this case only a transitory processes with life-time \(\tau_{c}.\) The trajectories are more complex than in static potentials. Closed periodic trajectories do not exist in time dependent potentials, but trapping events represented by almost closed eddying segments appear on all trajectories when the decorrelation time \(\tau_{c}\) is large compared to the time of flight \(\tau_{fl}=\lambda/V,\)\((\lambda=(\lambda_{1}^{2}+\lambda_{2}^{2})^{1/2}\) and \(V=(V_{1}^{2}+V_{2}^{2})^{1/2}),\) and the integration time is much longer than \(\tau_{c}.\) The trapping events are separated by long jumps, which are similar with the free trajectories. The separation of the trajectories in the categories \(c=tr,\)\(fr\) has no meaning in time-dependent potentials. However one can define related quantities that are not properties of the trajectories but of the contour lines of the potential. The latter are geometric objects. The fraction of free/trapped trajectories can be defined using the number of trajectories that stay on open/closed contour lines of the potential at time \(t.\) These fractions do not depend on time for stationary stochastic potentials, because the amplitude, the space correlation and the structure of the contour lines are statistically time-invariant. They equal the asymptotic values of the time dependent fractions \(n_{c}(t,V_{d})\) obtained in static potentials from the trajectories (Sections 3) \[n_{c}(V_{d})=\ \lim_{t\rightarrow\infty}\ n_{c}(t,V_{d}), \tag{41}\] for any \(\tau_{c}\) and \(c=tr,fr.\) In a similar way, the fraction of trajectories that stay at time \(t\) on contour lines of category \(c\) with the potential \(\phi^{t}=\phi({\bf x}(t))\) is a time-independent function of \(\phi^{t}\) and \(c,\) which is the asymptotic value \(n^{c}(\phi^{t})\) of the fractions obtained in static potential (\(\tau_{c}\rightarrow\infty\)) in Eqs. (29-31). The physical meaning of these quantities will be clarified after analyzing the significant modifications of the statistics of trajectories produced by the time-variation of the potential. We underline that, in time-dependent potentials, one can define the statistics conditioned by the initial potential, but not by the category. Our aim is to see if the special order determined by the average velocity survives at finite \(\tau_{c}.\) We analyze here the statistics on the whole set of trajectories (in \(R\)), while, in the next section, the statistics conditioned by the initial potential will be used for understanding the direct effects of time variation on the coherent elements found here. The time variation of \(\phi\) represents a decorrelation mechanism, because it determines the stochastic change of the Lagrangian velocity, which vanishes the Lagrangian correlations \(L_{i}(t)\) at times \(t\gg\tau_{c}.\) Usually, this produces the saturation of the time dependent diffusion coefficients \(D_{i}(t)\to D^{i}\) and diffusive transport with \(\left\langle\delta x_{i}^{2}(t)\right\rangle\to 2D^{i}t\) (as obtained from Eq. (21)). The memory of the initial velocity is lost in a time \(\tau_{c},\) which means that the displacements at large time \(t\gg\tau_{c}\) are sequences of non-correlated random steps that yield a Gaussian distribution. We show that this general behavior is not at all observed in the presence of \({\bf V}_{d}\) at large correlation times (weak time variation). A strong non-standard influence appears both on the transport and on the probability of the displacements. The Lagrangian velocity is, as expected, Gaussian at any time and for any \(\tau_{c},\) as in the static case. The time variation influences only its correlation. The dispersions of the trajectories \(\left<\delta x_{i}^{2}(t)\right>\) and the probabilities \(P(x_{i},t)\) for a typical case that illustrates the effects of weka time variation of the potential are shown in Figs. 18 and 19, by the black lines for \(V_{d}=0.3\) and \(\tau_{c}=33\). We also present, for comparison, two examples with \(V_{d}=0.3\) (for the static case \(\tau_{c}=\infty\) (red lines) and for fast time variation \(\tau_{c}=3.3\) (blue)), and two examples with \(V_{d}=0\) (for \(\tau_{c}=\infty\) (green) and \(\tau_{c}=33\) (cyan)). When \(V_{d}=0\), the subdiffusive transport at \(\tau_{c}=\infty\) with \(\left<\delta x_{i}^{2}(t)\right>\sim t^{0.68}\) is transformed into normal transport (\(\left<\delta x_{i}^{2}(t)\right>\to 2D^{i}t\)) at large \(\tau_{c}\) (Fig. 18, green and cyan curves). The process appears for all finite values of \(\tau_{c}\), which only influence the diffusion coefficients \(D^{i}.\) However, the probabilities of displacements are Gaussian only for fast time variation. In the static case, \(P(x_{i},t)\) has a steep peak in \(x_{i}=0,\) which corresponds to trapped trajectories, superposed on a large Gaussian component, which yields from the trajectories that are not closed at time \(t.\) The steep peak is flattened by time variation and the probabilities have extended exponential shapes at large \(\tau_{c}.\) When \(\tau_{c}\) decreases, \(P(x_{i},t)\) evolves to Gaussian distribution, which is attained when \(\tau_{c}<\tau_{fl}.\) The average velocity makes the transport strongly anisotropic. In frozen potential \(\tau_{c}=\infty,\) the transport is ballistic in the parallel direction and saturated perpendicular to \(\mathbf{V}_{d},\) as discusses in Section 4.2 and also shown in Figs. 18 and 19 (red curves). The normal transport and the Gaussian probability are reached only for fast time variation of the potential (\(\tau_{c}<\tau_{fl}\)), as seen in the example for \(\tau_{c}=3.3\) (blue curves). In these conditions, the motion along the contour lines of \(\phi(\mathbf{x},t)\) is completely hindered, which means that the quasi-coherent components are eliminated. Compared to these cases, the trajectory statistics at slow time variation in the presence of \(V_{d}\) (the black curves) is strongly anomalous with complex behavior. Trajectory dispersion at large time has nonstandard time-dependence in both directions \[\left<\delta x_{1}^{2}(t)\right>\sim t^{\alpha_{1}},\ \ \left<\delta x_{2}^{2}(t) \right>\sim t^{\alpha_{2}},\] where \(\alpha_{1}<1\) and \(\alpha_{2}>1,\) which corresponds to subdiffusive perpendicular transport (but not saturated as in the static case) and superdiffusive parallel transport (but not of ballistic type). These powers are functions of \(\tau_{c}\) and \(V_{d},\)\(\alpha_{i}(\tau_{c},V_{d}).\) When \(\tau_{c}\) decreases, \(\alpha_{1}\) increases and saturates \(\alpha_{1}\to 1,\) while \(\alpha_{2}\) decreases and saturates \(\alpha_{2}\to 1.\) A similar effect is determined by the increase of \(V_{d},\) which leads to normal transport at \(V_{d}\gtrsim 1.\) In the example presented in Fig. 18 (black curves), \(\alpha_{1}(33,0.3)=0.57,\)\(\alpha_{2}(33,0.3)=1.35.\) The probabilities are very large, especially in the parallel direction, and the peak in \(\mathbf{x}=0\) persists for very long time (\(t=300=6\tau_{c}\) in Fig. 19). Thus, the transport and the statistics of displacements are non-standard when \(V_{d}\lesssim 1\) and \(\tau_{c}>\tau_{fl}.\) In these conditions the structure of the contour lines of the potential shows island of closed lines in a network of open lines. Also, the trajectories approximately follow the contour lines for distances of the order of the correlation length before they are removed by the time variation of the potential. Figure 18: Comparision of the dispersions of the trajectories in time-dependent potential with the static cases along \(x_{1}\) (dashed lines) and \(x_{2}\) (solid lines) directions. The values of \(V_{d}\) and \(\tau_{c}\) label the curves. ## VII 7. Enhanced Coherence and Long Memory Effects ### 7.1 Hidden ordered motion Ordered motion conditioned by the initial potential \(\phi^{0}\) was found in the presence of an average velocity \(V_{d}\) for the free trajectories. It is represented by the average displacements \(\left\langle x_{i}(t)\right\rangle_{\phi^{0},fr}\) that are conditioned by the initial potential and by the category \(c=fr\) [Eq. (36)]. These quantities obtained in a time-dependent potential (with \(\tau_{c}=33\)) are shown in Fig. (20) for \(x_{1}\) and in Fig. (21) for \(x_{2}\), compared to the static case. Significant differences appear for both directions. One can see in Fig. (20, left panel) that the perpendicular displacements \(\left\langle x_{1}(t)\right\rangle_{\phi^{0}}\) are larger in time-depending potential, although the calculations are at a very large time, \(t=6\tau_{c}\) (where the EC of the potential (3) is negligible, \(E(t)=10^{-8}\)). The main contribution comes from large values of the potential \(\left|\phi^{0}\right|,\) which is negligible in the static potential. The amplitude of the ordered motion is represented by the average displacements conditioned by the sign of \(\phi^{0},\)\(\left\langle x_{1}(t)\right\rangle_{+}\) and \(\left\langle x_{1}(t)\right\rangle_{-},\) which determine by time derivative the HDs. Surprisingly, the time evolution of these quantities shows a continuous increase in time-dependent potential, while it saturates in the static case, as seen in Fig. (20, right panel) for \(\left\langle x_{1}(t)\right\rangle_{+}\). This means that the hidden drifts are transitory in static potentials, but they are long-life statistical quantities in time dependent potential. Their amplitude decays on a long time scale, much longer than the decorrelation time of the potential. The time variation of the potential modifies the parallel displacements \(\left\langle x_{2}(t)\right\rangle_{\phi^{0}}\) by determining the extension of Figure 19: The probabilities of the displacement along \(x_{1}\) (left panel) and \(x_{2}\) (right panel) directions, show the effect of the time variation of the potential for typical cases with the values of \(V_{d}\) and \(\tau_{c}\) that label the curves. Figure 20: Ordered perpendicular displacements \(\left\langle x_{1}(t)\right\rangle_{\phi^{0}}\) as functions of \(\phi^{0}\) at \(t=200\) (left panel) and \(\left\langle x_{1}(t)\right\rangle_{+}\) as function of time (right panel) for a time-dependent potential with \(\tau_{c}=33\) (black points) compared to the static case (dashed blue lines) the contribution on the whole range of \(\phi^{0}\) and the dependence on these quantities of the initial potential (Fig. (21, left panel). It is peaked on \(\phi^{0}=0\) and has a weak (algebraic) decay at large \(\left|\phi^{0}\right|,\) instead of the concentration on the domain of free trajectories with uniform average displacement Eq. (35). However, the average Lagrangian velocity is uniform on the whole range of \(\phi^{0}\) and it has the Eulerian value \(V_{d},\) as seen in Fig. (21, right panel). The process of concentration of the Lagrangian average velocity on the domain of free trajectories found in static potentials is eliminated by the time-variation. The fluctuations of the trajectories \(\left\langle\delta x_{i}^{2}(t)\right\rangle_{\phi^{0}}\) and the transport \(\left\langle\delta x_{i}(t)\delta v_{i}(t)\right\rangle_{\phi^{0}}\) conditioned by the initial potential are all asymptotically uniform on the whole domain of \(\phi^{0}\). They reach this stage in a long time compared to the correlation time of the potential (\(t\gg\tau_{c}\)) staring from the values corresponding to static potential that are maintained at small time \(t<\tau_{c}.\) These results show that the trajectories are statistically identical with the exception of the perpendicular average displacement \(\left\langle x_{i}(t)\right\rangle_{\phi^{0}},\) which depend on \(\phi^{0}.\) ### 7.2 Long memory The persistent Lagrangian order and the non-standard characteristics of the trajectories in the time dependent case can be understood by analyzing the statistics of the Lagrangian potential \(\phi(t)\equiv\phi(\mathbf{x}(t),t).\) The distribution of the Lagrangian potential has the same invariance property as the Lagrangian velocity. It has the Gaussian probability of the Eulerian potential at any time \(t,\) for both the static and the time-dependent cases, at any value of the average velocity \(V_{d}.\) However, significant differences appear between these cases concerning the correlation and the average conditioned by the initial potential, as seen in Figs. (22) and 23. The correlation of the Lagrangian potential \(L_{\phi}(t)=\left\langle\phi(0)\right\rangle\)\(\phi(t)\) is far from the Eulerian time-correlation, \(E(\mathbf{0},t).\) Starting from the trivial case of static potential with \(V_{d}=0,\) where the invariance of the potential implies \(L_{\phi}(t)=E(\mathbf{0})=1,\) in all the other cases shown in Fig. (22), the Lagrangian correlation is stronger than the Eulerian one, as it has a long tail with much slower decay. It demonstrates that the Lagrangian potential has a long-time memory. The memory effect is strongest in almost static potentials (very large \(\tau_{c}\)) with average velocity \(V_{d}=0.\) The Lagrangian correlation decreases much slower than the Eulerian one, and it is larger than than \(E(\mathbf{0},t)\) at any time (Fig. 22, the curve for \(V_{d}=0,\ \tau_{c}=33\)). In this example, at \(t=200\cong 6\tau_{c},\)\(L_{\phi}\) decreases only at \(0.4,\) while \(E(\mathbf{0},t)=1.5\)\(10^{-8}.\) The average velocity (\(V_{d}\neq 0\)) determines a faster decrease of \(L_{\phi}(t)\) at small time that leads to smaller values compared to the case \(V_{d}=0\) (Fig. 22, the curve for \(V_{d}=0.3,\ \tau_{c}=33\)). The decorrelation takes place on two time-scales. There is a fast decay at small time that is followed by a slab decrease of \(L_{\phi}(t).\) The fast decay is the same for \(\tau_{c}=\infty\) and \(\tau_{c}=33\) at \(V_{d}=0.3,\) which shows that this process is not a consequence of the potential time variation, but rather of the presence of \(V_{d}.\) In the static case, the memory of the Lagrangian potential is infinite (\(L_{\phi}(t)\) saturates). The asymptotic value is positive and it is a decreasing function of \(V_{d}\). The time-dependence of \(L_{\phi}(t)\) is the result of a selective decorrelation mechanism determined by the average velocity. This process can be understood by analyzing the correlation of \(\phi(t)\) conditioned by the category \(c=tr,\;fr\) in the static potential (\(\tau_{c}=\infty\)). As seen in Fig. 23, left panel, \(\left\langle\phi(0)\;\phi(t)\right\rangle_{c}\) decays to zero for the free trajectories, while it saturates for the trapped trajectories at a value that is comparable to the conditioned amplitude \(\left\langle\phi^{2}(0)\right\rangle_{tr}\). This demonstrates that in static potential the decorrelation affects only the free trajectories and that the memory effect is determined by the trapped trajectories, which approximately maintain the initial potential \(\phi(0)\). The asymptotic value of the average Lagrangian potential is thus \[\left\langle\phi(0)\;\phi(t)\right\rangle=\left\langle\phi(0)\;\phi(t)\right \rangle_{tr}n_{tr}=\phi^{0}n_{tr}. \tag{42}\] It is interesting to note that at finite \(\tau_{c}\), the significant decrease of the correction appears at any time, although it seems that the process determined by \(V_{d}\) is transitory in static potentials. This is caused by the interaction of the effects of \(V_{d}\) with the influence produced by the time variation, which determine the two time-scale evolution of \(L_{\phi}(t)\). It is clearly evidenced by the average Lagrangian potential conditioned by the initial value \(\phi^{0}\) normalized by this value, \(\left\langle\phi(t)\right\rangle_{\phi^{0}}/\phi^{0}\) represented in Fig. 23, right panel for the static case (lines) and for \(\tau_{c}=33\) (points) for several values of \(\phi_{0}\). An important property of the Lagrangian potential in slow time-dependent potential is that its correlation and the average conditioned by \(\phi^{0}\) have the same time-decay (as seen in Fig. 22 for the case \(V_{d}=0.3\), \(\tau_{c}=33\) and in the right panel of Fig. 23, all the curves have the same behavior \(\exp\left(-t/88\right)\)). The long memory of the potential and the increase of the average displacements \(\left\langle x_{1}(t)\right\rangle_{\phi^{0}}\) (Fig. 20) are the result of the same process. It consists of the liberation of the trapped trajectories with large \(\left|\phi^{0}\right|\) followed by repeated Figure 23: Caracterization of the memory of the Lagrangian potential. Left: The correlations conditioned by the category \(<\phi(0)\phi(t)>_{c}\) for \(V_{d}=0.1\) (dotted), \(V_{d}=0.2\) (dashed), \(V_{d}=0.3\) (continuous). Right: The normalized average potential conditioned by the initial potential for \(V_{d}=0.3\), for \(\tau_{c}=\infty\) (continuous), \(\tau_{c}=33\) (dotted), and the values of \(\phi_{0}\) that label the curves. Figure 22: Correlation of the Lagrangian potential for \(V_{d}=0\) (dashed lines) and \(V_{d}=0.3\) (continuous lines) for static (\(\tau_{c}=\infty\)), slow (\(\tau_{c}=33\)) and fast time-variation (\(\tau_{c}=3.3\)), compared to the Eulerian correlations (dotted blue lines). Long tails with exponential decay appear in time-dependent potentials with large \(\tau_{c}\). stochastic events of capture and release combined with the constraint of the total potential invariance (22) that approximately holds for small time intervals. Considering the case of the peaks of the potential, the liberation of the trapped trajectories with large \(\phi^{0}\) is produced when the time variation determines the decrease of the potential to \(\Delta\). The contour lines of the potential that are open have the average perpendicular displacement \(\Delta/V_{d}\) and the average potential along them equal to zero (as imposed by Eq. (22)). The stochastic recapture is uniformly distributed over the potential and has the average perpendicular location \(\Delta/V_{d}\). This cancels asymptotically the average of the potentials on the trapping events and leads to the average \(\Delta/V_{d}\) of the positions of the trapping events. This happens on a time scale that is much larger than \(\tau_{c}\). These released trajectories with large \(\phi^{0}\) determine the slow decay of their initial average potential and the increase of their average displacement from zero to the largest possible value \(\Delta/V_{d}\). Thus, the memory of the Lagrangian potential and the strengthening of the coherence of the trajectories are both determined by the slow evolution toward uniform distribution of the trapping events on the trajectories caused by the time-variation of the potential. ## VIII Summary and conclusions A detailed study of the Lagrangian coherence of the trajectories in 2-dimensional incompressible velocity fields is presented. The strong order that appear in this kind of velocity fields is determined by the Hamiltonian structure of the advection equations (1), (2) and by the space correlation of the stochastic potential. The trajectories follow the contour lines of the potential in the static case, and, for slowly varying potentials, they remain close to them for long time intervals. This study is focused on the identification and understanding of the order generated by an average velocity \(V_{d}\) superposed on the stochastic field. It determines an average potential, which strongly modifies the structure of contour lines of the total potential \(\phi_{t}({\bf x},t)=\phi({\bf x},t)+x_{1}Vd\) by generation of a network of open lines between islands of closed lines. As a result, two categories of trajectories are found in static (frozen) potential: trapped (closed, periodic) and free (with unlimited displacement in the parallel direction to \({\bf V}_{d}\)). The results presented here are based on the numerical simulation of the trajectories and on a complex statistical analysis that includes conditional averages and correlations connected to the topology of the contour lines of the potential \(\phi_{t}.\) The statistics of displacements and of the Lagrangian velocity are determined for the whole ensemble \(R,\) for the categories trapped (\(tr\)) and free (\(fr\)), and also on sets of contour lines of the potential conditioned by the value \(\phi^{0}\) at the starting point of the trajectories. This analysis reveals the origin of coherence and provides explanations for the nonstandard statistics, transport and memory of this motion. In the case of frozen potentials, we have found that the statistical properties determined for the two categories \(tr,\)\(fr\) are completely different compared to those obtained in the whole space \(R\). The average velocity \(V_{d}\) generates coherence in the Lagrangian velocity, which acquires average components from he stochastic ones for both categories. The supplementary coherent velocity cancels the average velocity \(V_{d}\) of the trapped trajectories, and it determines larger velocity for the free trajectories that compensate the missing contribution of the trapped ones (\(\left<v_{2}(t)\right>_{fr}=V_{d}/n_{fr}\)). Thus, the statistical invariance of the Lagrangian velocity (Lumley theorem) is ensured in a rather non-trivial manner that involves hidden coherent parallel velocity. The statistical analysis conditioned by the initial potential \(\phi^{0}\) reveals additional important aspects. The free trajectories have Gaussian distribution of \(\phi^{0}\) with a width \(\Delta\) that is smaller than the dispersion of the Eulerian potential. It shows the existence of ordered perpendicular motion (average displacements across \({\bf V}_{d}\) (34)) that appear for the free trajectories and are proportional with \(\phi^{0}.\) These averages \(\left<x_{1}(t)\right>_{\phi^{0},fr}\) increase in time from zero and saturate at \(\phi^{0}/V_{d}.\) They generate the hidden drifts, a pair of transitory average velocities perpendicular on \({\bf V}_{d}\) conditioned by the sign of \(\phi^{0},\) which have opposite directions and exactly compensate each other. We have also found that the Lagrangian statistics of the free trajectories conditioned by \(\phi^{0}\) depends on the value \(\phi^{0}\) only through the average perpendicular displacement. This means that the trajectories in the category \(fr\) for different values of \(\phi^{0}\) are statistically identical, and are organized in strips with limited perpendicular extensions. The probability of the displacements is non-Gaussian with separated contributions of the categories: a steep peak for \(tr\) and Gaussian distribution that move along \({\bf V}_{d},\) but with larger velocity \(V_{d}/n_{fr}\) for the \(fr\) subensemble. The time-invariant Gaussian distribution of the Lagrangian velocity is the sum of the non-Gaussian contributions of the two categories, which are both non-Gaussian, but invariant. The transport is produced only by the free trajectories. A paradoxical behavior was found: the statistics of the trajectories is strongly non-Gaussian, but the transport is produced by Gaussian trajectories, which in fact yield from the non-Gaussian velocity distribution of the free trajectories. The transport is anomalous, subdiffusive across \({\bf V}_{d}\) and superdiffusive of ballistic type along \({\bf V}_{d}.\) The latter results from the ordered parallel motion. category. The free trajectories can be associated to a geometrical locus on the two-dimensional space \((x_{1},x_{2}),\) in the sense that each point is the initial condition for such trajectory and all trajectories are confined in this domain \(fr\). The complement of \(fr\) is the geometric locus of the trapped trajectories \(tr\), which is composed of the islands of closed contour lines of the potentials. The (Eulerian) statistical characteristics of each geometrical locus were identified. The time-dependence of the stochastic potential produces anomalous increase of the Lagrangian coherence, instead of the expected decay after the decorrelation time. In particular, the perpendicular average displacements conditioned by \(\phi^{0}\) significantly increase and the transitory hidden drifts become long-life structure that survive at \(t\gg\tau_{c}.\) The enhanced coherence is found to be associated to a long memory of the Lagrangian potential. Also, the trajectories conditioned by the initial potential become statistically identical for all values of \(\phi^{0},\) not only on the domain of small potential with width \(\Delta,\) as in frozen potentials. These effects are caused by the stochastic liberation by the time-variation of the potential of the trajectories that initially are trapped, followed by repeated stochastic captures that are constraint by of the approximate invariance of the total potential.

2307.08623

HYTREL: Hypergraph-enhanced Tabular Data Representation Learning

Language models pretrained on large collections of tabular data have demonstrated their effectiveness in several downstream tasks. However, many of these models do not take into account the row/column permutation invariances, hierarchical structure, etc. that exist in tabular data. To alleviate these limitations, we propose HYTREL, a tabular language model, that captures the permutation invariances and three more structural properties of tabular data by using hypergraphs - where the table cells make up the nodes and the cells occurring jointly together in each row, column, and the entire table are used to form three different types of hyperedges. We show that HYTREL is maximally invariant under certain conditions for tabular data, i.e., two tables obtain the same representations via HYTREL iff the two tables are identical up to permutations. Our empirical results demonstrate that HYTREL consistently outperforms other competitive baselines on four downstream tasks with minimal pretraining, illustrating the advantages of incorporating the inductive biases associated with tabular data into the representations. Finally, our qualitative analyses showcase that HYTREL can assimilate the table structures to generate robust representations for the cells, rows, columns, and the entire table.

Pei Chen, Soumajyoti Sarkar, Leonard Lausen, Balasubramaniam Srinivasan, Sheng Zha, Ruihong Huang, George Karypis

2023-07-14T05:41:22Z

http://arxiv.org/abs/2307.08623v2

# HyTrel: Hypergraph-enhanced ###### Abstract Language models pretrained on large collections of tabular data have demonstrated their effectiveness in several downstream tasks. However, many of these models do not take into account the row/column permutation invariances, hierarchical structure, etc. that exist in tabular data. To alleviate these limitations, we propose **HyTrel**, a tabular language model, that captures the permutation invariances and three more _structural properties_ of tabular data by using hypergraphs-where the table cells make up the nodes and the cells occurring jointly together in each row, column, and the entire table are used to form three different types of hyperedges. We show that HyTrel is maximally invariant under certain conditions for tabular data, i.e., two tables obtain the same representations via HyTrel_iff_ the two tables are identical up to permutations. Our empirical results demonstrate that HyTrel**consistently** outperforms other competitive baselines on four downstream tasks with minimal pretraining, illustrating the advantages of incorporating the inductive biases associated with tabular data into the representations. Finally, our qualitative analyses showcase that HyTrel can assimilate the table structures to generate robust representations for the cells, rows, columns, and the entire table. 1 Footnote 1: Code, data, and checkpoints will be available soon. ## 1 Introduction Tabular data that is organized in bi-dimensional matrices are widespread in webpages, documents, and databases. Understanding tables can benefit many tasks such as table type classification, table similarity matching, and knowledge extraction from tables (e.g., column annotations) among others. Inspired by the success of pretrained language models in natural language tasks, recent studies (Yin et al., 2020; Yang et al., 2022) proposed Tabular Language Models (TaLMs) that perform pretraining on tables via self-supervision to generate expressive representations of tables for downstream tasks. Among the TaLMs, many works (Herzig et al., 2020; Yin et al., 2020; Deng et al., 2020; Iida et al., 2021) serialize tables to a sequence of tokens for leveraging existing pretrained language model checkpoints and textual self-supervised objectives like the Masked Language Modeling. However, due to the linearization of tables to strings, these models do not explicitly incorporate the structural properties of a table, e.g., the invariances to arbitrary permutations of rows and columns (independently). Our work focuses on obtaining representations of tables that take table structures into account. We hypothesize that incorporating such properties into the table representations will benefit many downstream table understanding tasks. **Motivation:** Tabular data is structurally different in comparison to other data modalities such as images, audio, and plain texts. We summarize four _structural properties_ present in the tables below: * Most tables are invariant to row/column permutations. This means, in general, if we arbitrarily (and independently) permute the rows or columns of a table, it is still an equivalent table. For other tables with an explicit ordering of rows or columns, we can make them permutation invariant by appropriately adding a ranking index as a new column or row. * Data from a single column are structurally similar-for example, they oftentimes have the same semantic types. Similarly, the cells within a single row are not independent of each other, and they usually describe the different aspects of one object. * The interactions within cells/rows/columns are necessarily not pairwise, i.e., the cells within the same row/column, and rows/columns from the same table can have high-order multilateral relations. * Information in tables is generally organized in a hierarchical fashion where the information at the table-level can be aggregated from the column/row-level, and further from the cell-level. However, the linearization-based approaches are not designed to explicitly capture most of the above properties. We aim to address the limitation by modeling all the aforementioned structural properties as inductive biases while learning the table representations. **Our Approach:** In line with recent studies (Deng et al., 2020; Yang et al., 2022; Wang et al., 2021) which have elucidated upon the importance of the structure of a table, we propose the HyTrel that uses hypergraphs to model the tabular data. We propose a modeling paradigm that aims _capture all of the four properties_ directly. Figure 1 provides an example of how a hypergraph is constructed from a table. As observed, converting a table into a hypergraph allows us to incorporate the first two properties inherent to the nature of hypergraphs. Hypergraphs seamlessly allow the model to incorporate row/column permutation invariances, as well as interactions among the cells within the same column or row. Moreover, the proposed hypergraph structure can capture the high-order (not just pairwise) interactions for the cells in a column or a row, as well as from the whole table, and an aggregation of hyperedges can also help preserve the hierarchical structure of a table. **Contributions:** Our theoretical analysis and empirical results demonstrate the advantages of modeling the four structural properties. We first show that HyTrel is maximally invariant when modeling tabular data (under certain conditions), i.e. if two tables get the same representations via the hypergraph table learning function, then the tables differ only by row/column permutation (independently) actions and vice versa. Empirically, we pretrain HyTrel on publicly available tables using two self-supervised objectives: a table content based ELECTRA\({}^{2}\) objective (Clark et al., 2020; Iida et al., 2021) and a table structure dependent contrastive objective (Wei et al., 2022). The evaluation of the pretrained HyTrel model on four downstream tasks (two knowledge extraction tasks, a table type detection task, and a table similarity prediction task) shows that HyTrel can achieve state-of-the-art performance. We also provide an extensive qualitative analysis of HyTrel-including visualizations that showcase that (a) HyTrel representations are robust to arbitrary permutations of rows and columns (independently), (b) HyTrel can incorporate the hierarchical table structure into the representations, (c) HyTrel can achieve close to state-of-the-art performance even without pretraining, and the model is Figure 1: An example of modeling a table as a hypergraph. Cells make up the nodes and the cells in each row, column, and the entire table form hyperedges. The table caption and the header names are used for the names of the table and column hyperedges. The hypergraph keeps the four structural properties of tables, e.g., the invariance property of the table as the row/column permutations result in the same hypergraph. extremely efficient with respect to the number epochs for pretraining in comparison to prior works, further demonstrating the advantages of HyTel in modeling the structural properties of tabular data. In Appendix B, we provide additional analysis that demonstrates HyTrel's ability to handle input tables of arbitrary size and underscore the importance of the independent row/column permutations. ## 2 HyTrel Model Formally, a table in our work is represented as \(\mathcal{T}=[M,H,R]\), where \(M\) is the caption, \(H=[h_{1},h_{2},h_{3},...,h_{m}]\) are the \(m\) column headers, \(R\) represents the \(n\) rows \([R_{1},R_{2},R_{3},...,R_{n}]\). Each row \(R_{i}\) has \(m\) cells \([c_{i1},c_{i2},c_{i3},...,c_{im}]\). The caption, header, and cell values can be regarded as sentences that contain several words. We note that each cell \(c_{ij}\) or header \(h_{i}\) also belongs to the corresponding column \(C_{j}\). We use \(C=[C_{1},C_{2},C_{3},...,C_{m}]\) to represent all the columns that include the headers, so a table can also be defined as \(\mathcal{T}=[M,C]\). ### Formatter & Embedding Layer The formatter transforms a table into a hypergraph. As shown in Figure 1, given a table \(\mathcal{T}\), we construct a corresponding hypergraph \(\mathcal{G}=(\mathcal{V},\mathcal{E})\), where \(\mathcal{V}\), \(\mathcal{E}\) denote the set of nodes and hyperedges respectively. We treat each cell \(c_{ij}\) as a node \(v_{ij}\in\mathcal{V}\), and each row \(R_{i}\), each column \(C_{j}\), and the entire table \(\mathcal{T}\) as hyperedges \(e^{e}_{i},e^{r}_{j},e^{t}\in\mathcal{E}_{\{1\leq i\leq n,1\leq j\leq m\}}\), respectively. As a part of our hypergraph construction, each cell node \(v_{ij}\) is connected to 3 hyperedges: its column hyperedge \(e^{c}_{i}\), row hyperedge \(e^{r}_{j}\), and the table hyperedge \(e^{t}\). The hypergraph can be conveniently be represented as an incidence matrix \(\mathbf{B}\in\{0,1\}^{mn\times(m+n+1)}\), where \(\mathbf{B}_{ij}=1\) when node \(i\) belong to hyperedge \(j\) and \(\mathbf{B}_{ij}=0\) otherwise. An embedding layer is then employed over the nodes and hyperedges. Each node \(v_{ij}\) corresponds to a cell \(c_{ij}\) that has several tokens, and we obtain the feature vector \(\mathbf{X}_{v_{ij},:}\in\mathbb{R}^{F}\) for a given node by feeding its constituent cell tokens into the embedding layer and then averaging the embeddings over the tokens. After obtaining node embeddings \(\mathbf{X}\in\mathbb{R}^{nm\times F}\) for all nodes, a similar transformation is applied over hyperedges. For different hyperedge types, we use different table content for their initialization : for a column hyperedge \(e^{c}_{i}\), we use all the tokens from the corresponding header \(h_{i}\). For the table hyperedge \(e^{t}\), we use the the entire caption associated with \(M\). For the row hyperedge Figure 2: The overall framework of HyTrel. We first turn a table \(\mathcal{T}\) into a hypergraph \(\mathcal{G}\) and then initialize the embeddings of the nodes \(\mathcal{V}\) and hyperedges \(\mathcal{E}\). After that, we encode the hypergraph using stacked multiple-layer hypergraph-structure-aware transformers (HyperTrans). Each HyperTrans layer has two attention blocks that work on hypergraph (HyperAtt) and one hyperedge fusion block. Lastly, we use the node and hyperedge representations from the final layer for pretraining and fine-tuning. when no semantic information is available, we randomly initialize them as \(\mathbf{S}_{e_{:}^{r},:}\in\mathbb{R}^{F}\). Performing the above operations yields an initialization for all the hyperedge embeddings \(\mathbf{S}\in\mathbb{R}^{(m+n+1)\times F}\). ### Hypergraph Encoder After embedding, we propose to use a structure-aware transformer module (HyperTrans) to encode the hypergraphs. HyperTrans encoder can encode the table content, structure, and relations among the table elements (including cells, headers, captions, etc.). As shown in Figure 2, one layer of the HyperTrans module is composed of two hypergraph attention blocks (HyperAtt, \(f\)) [12] that interact with the node and hyperedge representations, and one Hyperedge Fusion block. The first HyperAtt block is the Node2Hyperedge attention block as defined below: \[\mathbf{\tilde{S}}_{e,:}^{(t+1)}=f_{\mathcal{V}\rightarrow\mathcal{E}}\left(K_ {e,\mathbf{X}^{(t)}}\right) \tag{1}\] Where \(f_{\mathcal{V}\rightarrow\mathcal{E}}\) is a hypergraph attention function defined from nodes to hyperedges. \(K_{e,\mathbf{X}}=\{\mathbf{X}_{v,:}\ v\in e\}\) denotes the sets of hidden node representations included in the hyperedge \(e\). The Node2Hyperedge block will aggregate information to hyperedge \(e\) from its constituent nodes \(v\in e\). We then use a Hyperedge Fusion module (a Multilayer Perceptron Network, \(\mathrm{MLP}\)) to propagate the hyperedge information from the last step, as defined below: \[\mathbf{S}_{e,:}^{(t+1)}=\mathrm{MLP}\left(\mathbf{S}_{e,:}^{(t)},\mathbf{ \tilde{S}}_{e,:}^{(t+1)}\right) \tag{2}\] A second HyperAtt block Hyperedge2Node then aggregates information from a hyperedge to its constituent nodes as follows: \[\mathbf{X}_{v,:}^{(t+1)}=f_{\mathcal{E}\rightarrow\mathcal{V}}\left(L_{v, \mathbf{S}^{(t+1)}}\right) \tag{3}\] Where \(f_{\mathcal{E}\rightarrow\mathcal{V}}\) is another hypergraph attention function defined from hyperedges to nodes. \(L_{v,\mathbf{S}}=\{\mathbf{S}_{e,:}\ v\in e\}\) is defined as the sets of hidden representations of hyperedges that contain the node \(v\). As for the HyperAtt block \(f\), similar to transformer [25], it is composed of one multi-head attention, one Position-wise Feed-Forward Network (FFN), two-layer normalization (\(\mathrm{LN}\)) [11] and two skip connections [11], as in Figure 2. However, we do not use the self-attention [25] mechanism from the transformer model because it is not designed to keep the invariance structure of tables or hypergraphs. Inspired by the deep set models [10, 12], we use a set attention mechanism that can keep the permutation invariance of a table. We define HyperAtt \(f\) as follows: \[f_{\mathcal{V}\rightarrow\mathcal{E}\ \mathtt{or}\ \mathcal{E} \rightarrow\mathcal{V}}(\mathbf{I}):=\mathrm{LN}(\mathbf{Y}+\mathrm{FFN}( \mathbf{Y})) \tag{4}\] Where \(\mathbf{I}\) is the input node or hyperedge representations. The intermediate representations \(\mathbf{Y}\) is obtained by: \[\mathbf{Y}=\mathrm{LN}\left(\mathbf{\omega}+\mathrm{SetMH}(\mathbf{\omega}, \mathbf{I},\mathbf{I})\right) \tag{5}\] Where \(\mathrm{SetMH}\) is the multi-head set attention mechanism defined as: \[\mathrm{SetMH}(\mathbf{\omega},\mathbf{I},\mathbf{I})=\|_{i=1}^{h}\mathbf{O}_ {i} \tag{6}\] and \[\mathbf{O}_{i}=\mathrm{Softmax}\left(\mathbf{\omega}_{i}\left(\mathbf{I} \mathbf{W}_{i}^{K}\right)^{T}\right)\left(\mathbf{I}\mathbf{W}_{i}^{V}\right) \tag{7}\] Where \(\mathbf{\omega}\) is a learnable weight vector as the query and \(\mathbf{\omega}:=\|_{i=1}^{h}\mathbf{\omega}_{i}\), \(\mathbf{W}_{i}^{K}\) and \(\mathbf{W}_{i}^{V}\) are the weights for the key and value projections, \(\|\) means concatenation. So the HyperTrans module will update node and hyperedge representations alternatively. This mechanism enforces the table cells to interact with the columns, rows, and the table itself. Similar to BERT\({}_{base}\)[13] and TaBERT\({}_{base}\)[25], we stack \(12\) layers of HyperTrans. ### Invariances of the HyTrel Model Let \(\phi:\mathcal{T}\mapsto\mathbf{z}\in\mathbb{R}^{d}\) be our target function which captures the desired row/column permutation invariances of tables (say for tables of size \(n\times m\)). Rather than working on the table \(\mathcal{T}\) directly, the proposed HyTrel model works on a hypergraph (via Eqns (1-5)) that has an incidence matrix \(\mathbf{B}\) of size \(mn\times(m+n+1)\). Correspondingly, we shall refer to HyTrel as a function \(g:\mathbf{B}\mapsto\mathbf{y}\in\mathbb{R}^{k}\). In this section we will make the connections between the properties of the two functions \(\phi\) and \(g\), demonstrating a maximal invariance between the two-as a result of which we prove that our HyTrel can also preserve the permutation invariances of the tables. First, we list our assumptions and resultant properties of tabular data. Subsequently, we present the maximal invariance property of \(\phi\) and our hypergraph-based learning framework \(g\). As a part of our notation, we use \([n]\) to denote \(\{1,2,\ldots,n\}\). Preliminaries and all detailed proofs are presented in the Appendix A.1 and A.2 respectively. **Assumption 2.1**.: For any table \((\mathcal{T}_{ij})_{i\in[n],j\in[m]}\) (where \(i,j\) are indexes of the rows, columns), an arbitrary group action \(a\in\mathbb{S}_{n}\times\mathbb{S}_{m}\) acting appropriately on the rows and columns leaves the target random variables associated with tasks on the entire table unchanged. This assumption is valid in most real-world tables-as reordering the columns and the rows in the table oftentimes doesn't alter the properties associated with the entire table (e.g. name of the table, etc). As noted earlier, for tables with an explicit ordering of rows or columns, we can make them permutation invariant by adding a ranking index as a new column or row appropriately. To model this assumption, we state a property required for functions acting on tables next. _Property 1_.: A function \(\phi:\mathcal{T}\mapsto\mathbf{z}\in R^{d}\) which satisfies Assumption 2.1 and defined over tabular data must be invariant to actions from the (direct) product group \(\mathbb{S}_{n}\times\mathbb{S}_{m}\) acting appropriately on the table i.e. \(\phi(a\cdot\mathcal{T})=\phi(\mathcal{T})\ \ \forall a\in\mathbb{S}_{n}\times \mathbb{S}_{m}\). However, HyTrel (or the function \(g\) via hypergraph modeling) through Eqns (1-5)) models invariances of the associated incidence matrix to the product group \(\mathbb{S}_{mn}\times\mathbb{S}_{m+n+1}\) (proof presented in the appendix). To make the connection between the two, we present the maximal invariance property of our proposed HyTrel model. **Theorem 2.2**.: _A continuous function \(\phi:\mathcal{T}\mapsto\mathbf{z}\in\mathbb{R}^{d}\) over tables is maximally invariant when modeled as a function \(g:\mathbf{B}\mapsto\mathbf{y}\in\mathbb{R}^{k}\) over the incidence matrix of a hypergraph \(\mathcal{G}\) constructed per Section 2.1 (Where \(g\) is defined via Eqns (1-5)) if \(\exists\) a bijective map between the space of tables and incidence matrices (defined over appropriate sizes of tables, incidence matrices). That is, \(\phi(\mathcal{T}_{1})=\phi(\mathcal{T}_{2})\) iff \(\mathcal{T}_{2}\) is some combination of row and/or column permutation of \(\mathcal{T}_{1}\) and \(g(\mathbf{B}_{1})=g(\mathbf{B}_{2})\) where \(\mathbf{B}_{1},\mathbf{B}_{2}\) are the corresponding (hypergraph) incidence matrices of tables \(\mathcal{T}_{1},\mathcal{T}_{2}\)._ _Proof Sketch_: Detailed proof is provided in Appendix A.2. The above theorem uses Lemma 1 from (Tyshkevich and Zverovich, 1996) and applies the Weisfeiler-Lehman test of isomorphism over the star expansion graphs of the hypergraphs toward proving the same. As a consequence of Theorem 2.2, two tables identical to permutations will obtain the same representation, which has been shown to improve generalization performance (Lyle et al., 2020). ### Pretraining Heads **ELECTRA Head**: In the ELECTRA pretraining setting, we first corrupt a partial of cells and headers from a table and then predict whether a given cell or header has been corrupted or not Iida et al. (2021). Cross-entropy loss is used to train the binary classification head. **Contrastive Head**: In the contrastive pretraining setting, we randomly corrupt a table-transformed hypergraph by masking a portion of the connections between nodes and hyperedges, as inspired by the hypergraph contrastive learning (Wei et al., 2022). For each hypergraph, we corrupt two augmented views and use them as the positive pair, and use the remaining in-batch pairs as negative pairs. Following this, we contrast the table and column representations from the corresponding hyperedges. The InfoNCE (van den Oord et al., 2018) objective is used for optimization as in 8. \[loss=-\log\frac{\exp\left(\mathbf{q}\cdot\mathbf{k}_{+}/\tau\right)}{\sum_{i=0}^{K} \exp\left(\mathbf{q}\cdot\mathbf{k}_{i}/\tau\right)} \tag{8}\] where \((\mathbf{q},\mathbf{k}_{+})\) is the positive pair, and \(\tau\) is a temperature hyperparameter. Experiments ### Pre-training **Data** In line with previous TaLMs (Yin et al., 2020; Iida et al., 2021), we use tables from Wikipedia and Common Crawl for pretraining. We utilize preprocessing tools provided by Yin et al. (2020) and collect a total of 27 million tables (1% are sampled and used for validation).3 During pretraining, we truncate large tables and retain a maximum of 30 rows and 20 columns for each table, with a maximum of 64 tokens for captions, column names, and cell values. It is important to note that the truncation is solely for efficiency purposes and it does not affect HyTrel's ability to deal with large tables, as elaborated in appendix B.1. Footnote 3: As the version of Wikipedia used by (Yin et al., 2020) is not available now, we use an updated version so we collect slightly more tables than previous TaLMs. **Settings** With the ELECTRA pretraining objective, we randomly replace 15% of the cells or headers of an input table with values that are sampled from all the pretraining tables based on their frequency, as recommended by Iida et al. (2021). With the contrastive pretraining objective, we corrupted 30% of the connections between nodes and hyperedges for each table to create one augmented view. The temperature \(\tau\) is set as 0.007. For both objectives, we pretrain the HyTrel models for 5 epochs. More details can be found the Appendix C.1. ### Fine-tuning4 Footnote 4: More details about experimental settings, the datasets, and the baselines can be found the Appendix C.2 After pretraining, we use the HyTrel model as a table encoder to fine-tune downstream table-related tasks. In order to demonstrate that our model does not heavily rely on pretraining or on previous pretrained language models, we also fine-tune the randomly initialized HyTrel model for comparison. In this section, we introduce the evaluation tasks and the datasets. We choose the following four tasks that rely solely on the table representations since we want to test the task-agnostic representation power of our model and avoid training separate encoders for texts (e.g., questions in table QA tasks) or decoders for generations. As mentioned, our encoder can be used in all these scenarios and we leave its evaluation in other table-related tasks as future work. **Column Type Annotation** (CTA) task aims to annotate the semantic types of a column and is an important task in table understanding which can help many knowledge discovery tasks such as entity recognition and entity linking. We use the column representations from the final layer of HyTrel with their corresponding hyperedge representations for making predictions. We evaluate HyTrel on the TURL-CTA dataset constructed by Deng et al. (2020). **Column Property Annotation** (CPA) task aims to map column pairs from a table to relations in knowledge graphs. It is an important task aimed at extracting structured knowledge from tables. We use the dataset TURL-CPA constructed by Deng et al. (2020) for evaluation. **Table Type Detection** (TTD) task aims to annotate the semantic type of a table based on its content. We construct a dataset using a subset from the public WDC Schema.org Table Corpus. **Table Similarity Prediction** (TSP) task aims at predicting the semantic similarity between tables and then classifying a table pair as similar or dissimilar. We use the PMC dataset proposed by Habibi et al. (2020) for evaluation. ### Baselines **TaBERT**(Yin et al., 2020) is a representative TaLM that flattens the tables into sequences and jointly learns representations for sentences and tables by pretraining the model from the BERT checkpoints. _K=1_ and _K=3_ are the two variants based on the number of rows used. **TURL**(Deng et al., 2020) is another representative TaLM that also flattens the tables into sequences and pretrains from TinyBERT (Jiao et al., 2020) checkpoints. It introduces a vision matrix to incorporate table structure into the representations. **Doduo**(Suhara et al., 2022) is a state-of-the-art column annotation system that fine-tunes the BERT and uses table serialization to incorporate table content. ### Main Results The results are presented in Tables 1 and 2. Overall, HyTrel can consistently outperform the baselines and achieve superior performance. A salient observation is that our model (even without pretraining) can achieve close to state-of-the-art performance. In comparison, we notice that the performance slumps significantly for TaBERT without pretraining. This phenomenon empirically demonstrates the advantages of modeling the table structures as hypergraphs over the other methods that we compare. Additionally, we observe that the two pretraining objectives help different tasks in different ways. For the CTA, CPA, and TTD tasks, the two objectives can help HyTrel further improve its performance. In general, the ELECTRA objective performs better than the contrastive objective. These results are also in line with the representation analysis in Section 4.2 where we observe that the ELECTRA objective tends to learn table structure better than the contrastive objective. However, for the TSP task, we observe that the contrastive objective can help the HyTrel model while the ELECTRA objective fails to bring any improvement. One possible reason for the ineffectiveness of the ELECTRA objective could be its inability to transfer well across domains. HyTrel pretrained with tables from Wikipedia and Common Crawl could not transfer well to the medical domain PMC dataset. As for the improvement observed from the contrastive objective, the reason could be that contrastive learning that uses similarity metrics in the objective function can naturally help the similarity prediction task. **Scalability:** As stated in Section 3, we have truncated large tables during pretraining. However, this truncation does not hinder the ability of HyTrel to handle large table inputs in downstream tasks. In Appendix B.1, we present additional experiments demonstrating that: (a) HyTrel can effectively \begin{table} \begin{tabular}{l|c|c} \hline \hline Systems & Column Type Annotation & Column Property Annotation \\ \hline Sherlock & 88.40 / 70.55 / 78.47 & - \\ BERT\({}_{base}\) & - & 91.18 / 90.69 / 90.94 \\ TURL + metadata & 92.75 / 92.63 / 92.69 & 92.90 / 93.80 / 93.35 \\ Doduo + metadata & 93.25 / 92.34 / 92.79 & 91.20 / 94.50 / 92.82 \\ \hline TaBERT\({}_{base}\)_(K=1)_ & 91.40\({}_{\pm 0.06}\) / 89.49\({}_{\pm 0.21}\) / 90.43\({}_{\pm 0.11}\) & 92.31\({}_{\pm 0.24}\) / 90.42\({}_{\pm 0.53}\) / 91.36\({}_{\pm 0.30}\) \\ _w/o_ Pretrain & 90.00\({}_{\pm 0.14}\) / 85.50\({}_{\pm 0.09}\) / 87.70\({}_{\pm 0.10}\) & 89.74\({}_{\pm 0.40}\) / 68.74\({}_{\pm 0.93}\) / 77.84\({}_{\pm 0.64}\) \\ TaBERT\({}_{base}\)_(K=3)_ & 91.63\({}_{\pm 0.21}\) / 91.12\({}_{\pm 0.25}\) / 91.37\({}_{\pm 0.08}\) & 92.49\({}_{\pm 0.18}\) / 92.49\({}_{\pm 0.22}\) / 92.49\({}_{\pm 0.10}\) \\ _w/o_ Pretrain & 90.77\({}_{\pm 0.11}\) / 87.23\({}_{\pm 0.22}\) / 88.97\({}_{\pm 0.12}\) & 90.10\({}_{\pm 0.17}\) / 84.83\({}_{\pm 0.89}\) / 87.38\({}_{\pm 0.48}\) \\ \hline HyTrel _w/o_ Pretrain & **92.92\({}_{\pm 0.11}\)** / 92.50\({}_{\pm 0.10}\) / 92.71\({}_{\pm 0.08}\) & 92.85\({}_{\pm 0.35}\) / 91.50\({}_{\pm 0.54}\) / 92.17\({}_{\pm 0.38}\) \\ HyTrel _w/_ ELECTRA & 92.85\({}_{\pm 0.21}\) / **94.21\({}_{\pm 0.09}\)** / **93.53\({}_{\pm 0.10}\)** & 92.88\({}_{\pm 0.24}\) / **94.07\({}_{\pm 0.27}\)** / **93.48\({}_{\pm 0.12}\)** \\ HyTrel _w/_ Contrastive & 92.71\({}_{\pm 0.20}\) / 93.24\({}_{\pm 0.08}\) / 92.97\({}_{\pm 0.13}\) & **93.01\({}_{\pm 0.40}\)** / 93.16\({}_{\pm 0.40}\) / 93.09\({}_{\pm 0.17}\) \\ \hline \hline \end{tabular} \end{table} Table 1: Test results on the CTA and CPA tasks (Precision/Recall/F1 Scores,%). The results of TaBERT and HyTrel are from the average of 5 system runs with different random seeds. For fair comparisons, we use the results of TURL and Doduo with metadata, i.e., captions and headers. \begin{table} \begin{tabular}{l|c|c c} \hline \hline Systems & Table Type Detection & Table Similarity Prediction \\ \cline{2-4} & Accuracy & Precision/Recall/F1 & Accuracy \\ \hline TFIDF+Glove+MLP & - & 87.36 / 83.81 / 84.47 & 85.06 \\ TabSim & - & 88.65 / 85.45 / 86.13 & 87.05 \\ \hline TaBERT\({}_{base}\)_(K=1)_ & 93.11\({}_{\pm 0.31}\) & 87.04\({}_{\pm 0.64}\) / 85.34\({}_{\pm 0.93}\) / 86.18\({}_{\pm 1.13}\) & 87.35\({}_{\pm 1.42}\) \\ _w/o_ Pretrain & 85.04\({}_{\pm 0.41}\) & 33.61\({}_{\pm 12.70}\) / 50.31\({}_{\pm 12.75}\) / 40.30\({}_{\pm 12.03}\) & 63.45\({}_{\pm 1.011}\) \\ TaBERT\({}_{base}\)_(K=3)_ & 95.15\({}_{\pm 0.14}\) & 87.76\({}_{\pm 0.64}\) / 86.97\({}_{\pm 0.59}\) / 87.36\({}_{\pm 0.95}\) & 88.29\({}_{\pm 0.98}\) \\ _w/o_ Pretrain & 89.88\({}_{\pm 0.26}\) & 82.96\({}_{\pm 1.84}\) / 81.16\({}_{\pm 1.45}\) / 82.05\({}_{\pm 1.02}\) & 82.57\({}_{\pm 1.20}\) \\ \hline HyTrel w/o_ Pretrain & 93.84\({}_{\pm 0.17}\) & 88.94\({}_{\pm 1.83}\) / 85.72\({}_{\pm 1.52}\) / 87.30\({}_{\pm 1.02}\) & 88.38\({}_{\pm 1.43}\) \\ HyTrel w/ ELECTRA & **95.81\({}_{\pm 0.19}\)** & 87.35\({}_{\pm 0.42}\) / 87.29\({}_{\pm 0.48}\) / 87.32\({}_{\pm 0.50}\) & 88.29\({}_{\pm 0.49}\) \\ HyTrel w/ Contrastive & 94.52\({}_{\pm 0.30}\) & **89.41\({}_{\pm 0.58}\)** / **89.10\({}_{\pm 0.90}\)** / **89.26\({}_{\pm 0.53}\)** & **90.12\({}_{\pm 0.49}\)** \\ \hline \hline \end{tabular} \end{table} Table 2: Test results on the TTD (Accuracy Score,%) and TSP (Precision/Recall/F1 Scores,%) tasks. The results of TaBERT and HyTrel are from the average of 5 system runs with different random seeds. process tables of any size as inputs, and (b) down-sampling can be a favorable strategy when working with large input tables, significantly improving efficiency without compromising performance. ## 4 Qualitative Analysis ### HyTrel Learns Permutation Robust Representations We sample 5,000 tables from the validation data for analysis. We analyze the impact of applying different permutations to a table, including permuting rows, columns, and both rows/columns independently. Toward our analysis, we measure the Euclidean distance (L2 Norm) of the representations (cells, rows, columns and tables). As shown in Figure 3, the distance is almost always 0 for the HyTrel model because of its explicit invariance-preserving property. On the other hand, for the TaBERT model, the distance is not trivial. We observe that when more rows (K=3) are enabled, the value of the L2 norm increases as we introduce different permutations. Moreover, we also observe that permuting the columns has a greater impact on the L2 norm than the row permutations. A combination of rows and columns permutations has the largest impact among all three actions. Note that when K=1 with just one row, the effect of row permutation is disabled for TaBERT. ### HyTrel Learns the Underlying Hierarchical Table Structure Next, we demonstrate that the HyTrel model has learned the hierarchical table structure into its representations, as we target at. We use t-SNE (Van der Maaten and Hinton, 2008) for the visualization of different table elements from the same 5,000 validation tables, as shown in Figure 4. We observe that with random initializations, different table elements cannot be distinguished properly. After the encoding of the randomly initialized HyTrel model, we start to observe noticeable differences for different table elements in the visualization space. Notably, the individual cell representations start to concentrate together and can be distinguished from high-level table elements (tables, columns, and rows) which occupy their separate places in the space of representations. We also notice that, by pretraining the HyTrel with the ELECTRA objective, all table elements can be well separated, showing that it incorporates the hierarchical table structure into its latent representations. As for the contrastive pretraining, we see that it can distinguish columns from rows as Figure 4: t-SNE visualization of the representations learned by HyTrel. ‘tab’, ‘col’, ‘row’, and ‘cell’ are the representations for different table elements: tables, columns, rows, and cells. Figure 3: Average distance between table element representations before and after permutations. The HyTrel is immune to the permutations while the TaBERT is sensitive to them. compared with randomly initialized HyTrel, but could not to well separate the table representations in comparison with the ELECTRA pretraining. This also partially explains the better performance of the ELECTRA pretraining in the CTA, CPA and TTD tasks in contrast to the contrastive pretraining. ### HyTrel Demonstrates Effective Pretraining by Capturing the Table Structures Our evaluation shows that the HyTrel model can perform well without pretraining, which demonstrates its training efficiency by modeling table structures. Here we further analyze how much pretraining is required for HyTrel to further enhance its performance, as compared with the baseline model. We plot the validation performance of the tasks evaluated at different pretraining checkpoints in Figure 5. Overall, we can observe that the performance of HyTrel improves drastically during the first several pretraining epochs, and then saturates at about 5 epochs. With the minimal pretraining, HyTrel can outperform the fully pretrained TaBERT models. This demonstrates that our model does not require extensive pretraining to further improve its performance in contrast with previous TaLMs (e.g., TURL for 100 epochs, TaBERT for 10 epochs). Besides, we also observe from the curves that the ELECTRA objective consistently outperforms the contrastive objective for the CTA, CPA, and TTD tasks, but under-performs on the TSP task. Also, the ELECTRA objective has a negative impact on the TSP task when pretrained for longer duration, which is in line with our previous findings. ## 5 Related Work There are two avenues of research that have studied tabular data representation learning. The first group of studies focus on predicting labels (essentially one row) for classification and regression problems, using row information and column schema as inputHuang et al. (2020); Arik and Pfister (2021); Sompalli et al. (2021); Gorishniy et al. (2021); Grinsztajn et al. (2022); Wang and Sun (2022); Du et al. (2022); Wydmanski et al. (2023). These studies use gradient descent-based end-to-end learning and aim to outperform tree-based models through task-specific model pretraining and fine-tuning. The second group of studies proposes TaLMs to retrieve task-agnostic tabular data representations for different downstream table understanding tasks. Drawing inspiration from textual Language Models like BERT Devlin et al. (2019), many works Herzig et al. (2020); Yin et al. (2020); Deng et al. (2020); Iida et al. (2021) serialize tables to a sequence of tokens, leveraging existing checkpoints and textual self-supervised objectives. However, the representations of the tables can not only be learned from table content and by utilizing table structures, similar to other forms of semi-structured data like code and HTML. Some contemporary works have noticed the importance of the table structures and introduce many techniques to learn a certain aspect of them, such as masking Deng et al. (2020), coordinates Wang et al. (2021); Dash et al. (2022), and attention bias Yang et al. (2022). Our work belongs to this second group of studies and we propose to use hypergraphs to comprehensively model the rich table structures, and this is close to previous graph-based neural networks Mueller et al. (2019); Wang et al. (2021); Wang et al. (2021) where tables have been structured as graphs to incorporate row and column order information. Table representation learning that focuses on joint text and table understanding is a separate field of research that partially overlaps with our work. Among them, some work Herzig et al. (2020); Shi Figure 5: Performance with different pretraining checkpoints on the validation set of the four tasks. For the TaBERT models, we can only access the randomly initialized and fully pretrained (10 epochs) checkpoints. All results are from 5 system runs with different random seeds. et al., 2022; Herzig et al., 2021; Glass et al., 2021; Yang et al., 2022) specialize in question-answering (QA) on tables and they jointly train a model that takes the question and the table structure as input together, allowing the pretraining to attend to the interactions of both texts and tables and boosting the table-based QA tasks. Another branch of joint text and table understanding work focuses on text generation from tables(Parikh et al., 2020; Yoran et al., 2021; Wang et al., 2022; Andrejczuk et al., 2022), relying on an encoder-decoder model like T5(Raffel et al., 2020) that can encode tables and decode free-form texts. In contrast to these studies, our work centers on the importance of structures in tables for table representation only, without extra text encoding or decoding. Learning on hypergraphs has gone through a series of evolution (Agarwal et al., 2006; Yadati et al., 2019; Arya et al., 2020) in the way the hypergraph structure is modeled using neural networks layers. However, many of them collapse the hypergraph into a fully connected graph by clique expansion and cannot preserve the high-order interaction among the nodes. The recent development of permutation invariant networks (Zaheer et al., 2017; Lee et al., 2019) has enabled high-order interactions on the hypergraphs (Chien et al., 2022) that uses parameterized multi-set functions to model dual relations from node to hyperedges and vice versa. Closely related to the latest advancement, our HyTrel model adopts a similar neural message passing on hypergraphs to preserve the invariance property and high-order interactions of tables. ## 6 Limitations The proposed HyTrel is a table encoder, and by itself cannot handle joint text and table understanding tasks like table QA and table-to-text generation. While it's possible to add text encoders or decoders for these tasks, it can potentially introduce additional factors that may complicate evaluating our hypothesis about the usefulness of modeling structural table properties. Moreover, the current model structure is designed for tables with simple column structures, like prior TaLMs, and cannot handle tables with complex hierarchical column structures. Additionally, HyTrel does not consider cross-table relations. Although we believe the hypergraph can generalize to model complicated column structures and cross-table interactions, we leave these aspects for future research. ## 7 Conclusion In this work, we propose a tabular language model HyTrel that models tables as hypergraphs. It can incorporate the permutation invariances and table structures into the table representations. The evaluation on four table-related tasks demonstrates the advantages of learning these table properties and show that it can consistently achieve superior performance over the competing baselines. Our theoretical and qualitative analyses also support the effectiveness of learning the structural properties.

2305.06077

Relightify: Relightable 3D Faces from a Single Image via Diffusion Models

Following the remarkable success of diffusion models on image generation, recent works have also demonstrated their impressive ability to address a number of inverse problems in an unsupervised way, by properly constraining the sampling process based on a conditioning input. Motivated by this, in this paper, we present the first approach to use diffusion models as a prior for highly accurate 3D facial BRDF reconstruction from a single image. We start by leveraging a high-quality UV dataset of facial reflectance (diffuse and specular albedo and normals), which we render under varying illumination settings to simulate natural RGB textures and, then, train an unconditional diffusion model on concatenated pairs of rendered textures and reflectance components. At test time, we fit a 3D morphable model to the given image and unwrap the face in a partial UV texture. By sampling from the diffusion model, while retaining the observed texture part intact, the model inpaints not only the self-occluded areas but also the unknown reflectance components, in a single sequence of denoising steps. In contrast to existing methods, we directly acquire the observed texture from the input image, thus, resulting in more faithful and consistent reflectance estimation. Through a series of qualitative and quantitative comparisons, we demonstrate superior performance in both texture completion as well as reflectance reconstruction tasks.

Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, Stefanos Zafeiriou

2023-05-10T11:57:49Z

http://arxiv.org/abs/2305.06077v2

# Relightify: Reliable 3D Faces from a Single Image via Diffusion Models ###### Abstract Following the remarkable success of diffusion models on image generation, recent works have also demonstrated their impressive ability to address a number of inverse problems in an unsupervised way, by properly constraining the sampling process based on a conditioning input. Motivated by this, in this paper, we present the first approach to use diffusion models as a prior for highly accurate 3D facial BRDF reconstruction from a single image. We start by leveraging a high-quality UV dataset of facial reflectance (diffuse and specular albedo and normals), which we render under varying illumination settings to simulate natural RGB textures and, then, train an unconditional diffusion model on concatenated pairs of rendered textures and reflectance components. At test time, we fit a 3D morphable model to the given image and unwrap the face in a partial UV texture. By sampling from the diffusion model, while retaining the observed texture part intact, the model inpaints not only the self-occluded areas but also the unknown reflectance components, in a single sequence of denoising steps. In contrast to existing methods, we directly acquire the observed texture from the input image, thus, resulting in more faithful and consistent reflectance estimation. Through a series of qualitative and quantitative comparisons, we demonstrate superior performance in both texture completion as well as reflectance reconstruction tasks. ## 1 Introduction Creating digital avatars of real people is of paramount importance for a range of applications, including VR, AR or the film industry. Human faces have been studied extensively over the years, attracting attention at the intersection of Computer Vision, Graphics and Machine Learning research. Although vast literature exists around the estimation of the 3D shape and reflectance of a face from unconstrained inputs such as "in-the-wild" RGB images, it still remains a challenging problem in the field. In particular, the recent breakthrough in image synthesis using diffusion generative models creates a new perspective towards photo-realistic 3D face reconstruction, which has not been explored so far and stems from the state-of-the-art performance of these models in solving inverse problems without supervised training. Facial reflectance capture typically requires a controllable illumination system equipped with multiple cameras, first introduced as a Light Stage [13]. Polarized illumination and gradient patterns can be employed for diffuse-specular separation [49, 27], using which, spatially varying facial reflectance maps can be acquired, that describe BRDF parameters, including the diffuse and specular albedo and normals. Although recent works attempt to simplify the capturing apparatus and process using inverse rendering [29, 56] or commodity devices [39], such methods still require a laborious capturing process and expensive equipment. Since their introduction by Blanz and Vetter [4], 3D Mor phable Models (3DMMs) [55, 12, 45, 6, 5] have been established as a robust methodology for monocular 3D face reconstruction [19, 70] by regularizing the otherwise ill-posed optimization problem towards a known statistical prior of the facial geometry, which is usually defined by the linear space of a PCA model. In addition to the coarse geometry estimation, 3DMMs have been used in conjunction with powerful CNN-based texture models, leading to impressively detailed avatar reconstructions even from low-resolution images [58, 24, 25]. Furthermore, another line of research [7, 33, 69, 3, 18, 17, 40, 42] revolves around the reconstruction of rendering assets such as reflectance components (diffuse and specular albedo) and high-frequency normals of the facial surface. As a result, the recovered 3D faces can be realistically rendered in arbitrary illumination environments. However, prior work either contains scene illumination inhibiting relighting [14, 22, 24] or is restricted by the models' generalization, lowering the identity similarity [24, 40, 48]. Our work shares the same objective in that we couple a 3DMM with high-quality UV reflectance maps, but attempts to solve both of these issues, by preserving the observed texture details from the input image and jointly inferring the facial reflectance. In fact, the visible pixels of the facial texture by the given camera pose are directly recoverable from the input image via inverse rasterization of the fitted 3D mesh. Therefore, we cast the 3D face reconstruction problem as an image inpainting task in the UV space; _i.e_. the goal is to fill in the missing pixels in a consistent manner with respect to some statistical prior. In particular, we propose to use a diffusion model as the generative backbone of our method. Diffusion models [62] are naturally associated with guided image synthesis since they treat image generation as a sequence of denoising steps in the form of a learnable Markov process. This allows to directly interfere with the sampling process, given that samples at each part of the chain are distorted versions of real images with known noise variances. Thus, by properly modifying the sampling process, a single unconditional diffusion model can be used for different inverse problems, such as image editing [51], inpainting [47, 11], restoration [37] or super-resolution [10, 9], without problem-specific training. In this paper, we build a high-quality statistical model of facial texture and reflectance by means of a diffusion model and adopt an inpainting approach to complete the partially reconstructed UV texture produced by a 3DMM fitting step. We further extend the sampling process to recover the missing reflectance components by enforcing consistency with the input texture. As a result, our method, dubbed _Relightfy_, generates accurate and render-ready 3D faces from unconstrained images, as shown in Fig. 1. In summary, we make the following contributions: * We present the first, to the best of our knowledge, diffusion-based approach for relightable 3D face reconstruction from images. By training on a pseudo-ground-truth dataset of facial reflectance, while directly recovering texture parts from the input, we achieve high-quality rendering assets that preserve important details of the input face (_e.g_. wrinkles, moles). * We propose an efficient way of predicting different modalities in a consistent way by learning a generative model on concatenated reflectance maps and casting the reconstruction as an inpainting problem, spatially, but also channel-wise. * We qualitatively and quantitatively demonstrate the superiority of our approach against previous methods regarding both the completed textures as well as the recovered reflectance maps. ## 2 Related Work ### Diffusion Models for Inverse Problems Diffusion models [62] are latent variable generative models which artificially corrupt the data distribution by adding noise and attempt to approximate the reverse process. They have lately emerged as a powerful image synthesis model [31, 16, 64] outperforming previous state-of-the-art approaches in both conditional and unconditional tasks. While they achieve excellent image quality and are robust to multi-modal distributions, they are computationally demanding to sample from, since they require a large sequence of denoising steps (_e.g_. 1000), each of which operates in the high dimensional image space. To alleviate this, a number of works [63, 38, 59] have proposed alternative strategies to accelerate sampling by reducing the steps of the reverse process. Another line of research [68, 60] proposes to train an encoding model and learn a diffusion model on its lower-dimensional latent space. Recently, Rombach _et al_. [57] have further explored the use of a VQGAN [20] as the auto-encoding model, showing that a mild compression is enough to reduce the training/sampling time without sacrificing sample quality. The latter approach is our method of choice for this work, as we elaborate on a high-resolution UV image space, which would otherwise significantly increase the computational overhead. One of the most interesting aspects of diffusion models is that they can be used as unsupervised solvers for different inverse problems, where the goal is to reconstruct a sample from some distorted observation, _i.e_. conditioning input. Song _et al_. [64] propose a conditioning mechanism during inference that allows applications such as class-conditional generation, inpainting and colorization. Similarly, [9] uses a low-pass filtered version of the conditioning image to guide the denoising process at each step and SDEdit [51] addresses image translation and editing using a diffused ver sion of the input image to initialize sampling from an intermediate timestep. RePaint [47] achieves state-of-the-art results on image inpainting by repeating multiple forward and backward diffusion steps to enforce harmonization. Despite its improved performance, this resampling strategy significantly increases the computational time. In contrast, CCDF [10] and DDRM [37] propose efficient techniques for reducing the length of the reverse process while retaining image quality at a high level. More recently, MCG [11] introduced a novel manifold constraint step, which combined with the standard reverse diffusion outperforms the aforementioned methods on a number of inverse tasks, including inpainting. We adopt this approach in our work to accurately fill in the missing pixels of both texture and reflectance maps of a face from a given image via diffusion-based inpainting, while fully preserving the observed ones. Note also that this approach does not assume any specific distribution of visibility masks, as it is trained unconditionally on complete textures. ### Facial Reconstruction 3DMMs [4] are the typical models for facial reconstruction from "in-the-wild" images, using a linear model for the identity, and additional linear models for expression or color. Current facial 3DMMs include the Basel Face Model (BFM) [55] and the Large Scale Facial Model (LSFM) [5]. Egger _et al_. [19] provide a thorough review on the subject. AlbedoMM [61] first created a 3DMM of facial reflectance, which can be relighted, but is restricted to a linear and per-vertex color model. Dib _et al_. [17, 18] improved on prior works' simplistic shading models and used inverse ray tracing to acquire photorealistic facial reflectance. Recently, GANFit [24, 25] introduced a potent method for fitting 3DMMs with a GAN-based [28] facial texture generator, achieving high-fidelity facial avatars, but lacking relighting capabilities due to baked illumination in the textures. AvatarMe++ [40, 42] overcame this issue by translating the reconstructed textures to facial reflectance using a conditional GAN, while adding extra processing steps. While we use AvatarMe++ to augment our training data, our method significantly outperforms them by using a powerful diffusion model and inferring only the occluded facial texture. TBGAN [23] first introduced a deep generative network for facial reflectance, based on ProgressiveGAN [34] and [44] introduced a more powerful model, based on StyleGAN [35]. However, both works did not showcase fitting capabilities. An extension of the latter [48], introduced a set of multiple networks, with a StyleGAN2 [36] base, that can be used to generate shape and albedo from images with arbitrary illumination and expression. While close to our work, our method uses a single and more powerful diffusion model, inferring not only the diffuse albedo, but also the specular albedo and normals. Moreover, our work inpaints only the occluded facial areas, preserving the visible part of the texture and achieves higher reconstruction fidelity. Although our method is applied to facial reconstruction, we simultaneously solve a facial texture inpainting problem in UV space. Initially explored in 2D facial images [46] and expanded to UV completion using deep encoder-decoder architectures (UV-GAN [14]), such works recover the facial texture from partial and masked facial images. Recently, OSTeC [22], used a pre-trained StyleGAN in 2D to recover multiple poses of the input subject so as to create a complete UV facial texture. While prior works achieve impressive results, all are restricted facial textures with baked illumination. In contrast, we jointly recover the facial reflectance, making the reconstruction relightable in standard rendering engines. ## 3 Method We propose a diffusion-based inpainting approach to estimate both the UV texture with existing baked illumination and the actual reflectance of a face in a single process. At the core of our approach lies an unconditional diffusion generative model trained on pairs of textures and their accompanying reflectance. This coupled texture-reflectance modeling along with the sequential denoising process of diffusion models allows us to reconstruct the reflectance from a partial texture of the input face, as shown in Fig. 2. Our method, thus, generates high-quality 3D face avatars from 'in-the-wild' images, which can be realistically relighted. In the following sections, we first analyze the training of our diffusion model, and then explain the 3D shape reconstruction and texture inpainting strategies in further detail. ### Diffusion Models: Background Given a distribution of real images \(x\), diffusion models [62] define a forward diffusion process which gradually adds Gaussian noise to the input image in \(T\) consecutive steps. This corresponds to a fixed Markov Chain, where starting from a clean image \(x_{0}\), the noisy samples \(x_{t}\) at each timestep \(t\) are drawn from the following distributions (with timestep-depending variances \(\beta_{t}\)) conditioned on the previous samples: \[q(x_{t}|x_{t-1})=\mathcal{N}(x_{t};\sqrt{1-\beta_{t}}x_{t-1},\beta_{t}\mathbf{ I}) \tag{1}\] This is equivalent to directly sampling \(x_{t}\) conditioned on the clean image \(x_{0}\) via: \[q(x_{t}|x_{0})=\mathcal{N}(x_{t};\sqrt{\bar{\alpha}_{t}}x_{0},(1-\bar{\alpha}_ {t})\mathbf{I}) \tag{2}\] where \(\alpha_{t}\coloneqq 1-\beta_{t}\) and \(\bar{\alpha}_{t}\coloneqq\prod_{s=1}^{t}\alpha_{s}\). Given large enough \(T\), this process leads to normally distributed noise \(x_{T}\). Then, the goal is to learn the reverse Markov process: \[p_{\theta}(x_{t-1}|x_{t})=\mathcal{N}(x_{t-1};\mu_{\theta}(x_{t},t),\Sigma_{ \theta}(x_{t},t)) \tag{3}\] which gradually denoises the random noise \(x_{T}\) towards a realistic image, by minimizing the variational lower bound of the negative log likelihood [31, 16]. Following the reparameterization proposed in [31], the model consists of time-conditioned denoising autoencoders \(\epsilon_{\theta}(x_{t},t);t\in\{1,2,\dots,T\}\), which are trained to predict the noise \(\epsilon\sim\mathcal{N}(\mathbf{0},\mathbf{I})\) that was added to the input image \(x_{0}\) to account for the noisy version \(x_{t}\): \[L=E_{x_{0},\epsilon,t}\left[||\epsilon-\epsilon_{\theta}(x_{t},t)||^{2}\right] \tag{4}\] Once trained, we can generate images by starting from random noise \(x_{T}\sim\mathcal{N}(\mathbf{0},\mathbf{I})\) and sequentially drawing denoised images around the mean: \[\mu_{\theta}(x_{t},t)=\frac{1}{\sqrt{\alpha_{t}}}\left(x_{t}-\frac{\beta_{t}} {\sqrt{1-\hat{\alpha}_{t}}}\epsilon_{\theta}(x_{t},t)\right) \tag{5}\] ### Training of our Diffusion Model In this work, we harness the power of diffusion models to learn a strong generative prior over the domain of facial texture/reflectance. In particular, we adopt a physically-based perspective by separating the facial reflectance into different UV maps, namely diffuse albedo (\(\mathbf{A}_{d}\)), specular albedo (\(\mathbf{A}_{s}\)) and surface normals (\(\mathbf{N}\)) with high-frequency details. This allows realistic rendering under different illumination conditions. We learn our prior using a high-quality dataset consisting of _complete_ pairs of facial reflectance, and a corresponding rendered texture \(\mathbf{T}\) under arbitrary illumination. More details on the data we use are provided in section 4.1. We train an unconditional diffusion model (as described in section 3.1) on the quadruples: \[x=[\mathbf{T},\mathbf{A}_{d},\mathbf{A}_{s},\mathbf{N}]\in\mathbb{R}^{512\times 5 12\times 12} \tag{6}\] where we concatenate the components of Eq. 6 across channels (each of the 4 UV images measures \(512\times 512\times 3\) pixels). By sampling from this model, we can synthesize pairs of shaded RGB textures (\(\mathbf{T}\)) and reflectance components (\(\mathbf{A}_{d},\mathbf{A}_{s},\mathbf{N}\)) which are in correspondence, meaning that the texture is a rendered version of the UV reflectance under some illumination environment. In practice, to reduce the computational requirements to a reasonable level, we follow the paradigm of **latent diffusion models** proposed by Rombach _et al_. [57], where the images are first compressed to a latent space \(z=\mathcal{E}(x)\in\mathbb{R}^{h\times w\times c}\) by training a perceptual auto-encoder, consisting of an encoder \(\mathcal{E}\) and a decoder \(\mathcal{D}\). Using perceptual and adversarial losses similar to VQGAN [20], the autoencoder achieves an excellent quality of reconstructed samples \(\tilde{x}=\mathcal{D}(\mathcal{E}(x))\), while allowing to efficiently train the diffusion model on the lower dimensional pixel-space of the learned embeddings. In our case, we train four similar auto-encoders, one for each of \(\mathbf{T},\mathbf{A}_{d},\mathbf{A}_{s}\) and \(\mathbf{N}\), all of them reducing the input resolution to latent dimensions of \(h=w=64\), \(c=3\). Therefore, our latent diffusion model [57] is trained on the concatenation of the 4 embeddings: \[z=[z_{\mathbf{T}},z_{\mathbf{A}_{d}},z_{\mathbf{A}_{s}},z_{\mathbf{N}}]\in \mathbb{R}^{64\times 64\times 12} \tag{7}\] Samples from our diffusion model (after being decoded through each \(\mathcal{D}\)) can be seen in the left part of Fig. 1. ### Inference We use the aforementioned trained diffusion model to perform inpainting on both the texture and reflectance UV Figure 2: Overview of our method during inference. Please note that we use a latent diffusion model [57], yet we illustrate the denoising process in the original image space for visualization purposes. We perform standard 3DMM fitting to get a partial UV texture via image-to-uv rasterization. Then, starting from random noise, we utilize the known texture to guide the sampling process of a texture/reflectance diffusion model towards completing the unobserved pixels. Each denoising step, from \(z_{t}\) to \(z_{t-1}\) (\(t\in\{1,\dots,T\}\)), follows an inpainting approach similar to MCG [11] (see Eq. 9): 1). The reflectance maps and unobserved texture pixels are updated based on reverse diffusion sampling and manifold constraints, while 2) the known pixels are directly sampled from the input texture via forward diffusion (\(\odot\) and \(\oplus\) denote the Hadamard product and addition respectively). Note that masking is only applied to the texture, while the reflectance maps (diffuse/specular albedo, normals) are entirely predicted from random noise. At the end of the process, we acquire high-quality rendering assets, making our 3D avatar realistically renderable. maps based on a partial UV texture obtained by 3DMM fitting. We provide a detailed description below. 3DMM Fitting and Texture Initialization.We rely on 3DMMs to recover a rough 3D shape of the face from a 2D image as a mesh \(\mathbf{S}\in\mathbb{R}^{n\times 3}\) with \(n\) vertices. Specifically, we employ a linear 3DMM: \[\mathbf{S}(\mathbf{p}_{s},\mathbf{p}_{e})=\mathbf{m}+\mathbf{U}_{s}\mathbf{p}_ {s}+\mathbf{U}_{e}\mathbf{p}_{e} \tag{8}\] consisting of the LSFM [5] shape eigenbasis \(\mathbf{U}_{s}\in\mathbb{R}^{3n\times 158}\) and the expression eigenbasis \(\mathbf{U}_{e}\in\mathbb{R}^{3n\times 29}\) from the 4DFAB database [8]. We fit the 3DMM to the input image by optimizing the shape coefficients \(\mathbf{p}_{s}\), expression coefficients \(\mathbf{p}_{e}\) and camera parameters \(\mathbf{p}_{c}\) utilizing an off-the-shelf framework [1]. We use a standard UV topology for texturing the 3D mesh, where each vertex is assigned to a fixed 2D coordinate on the UV plane. By rasterizing the fitted 3D mesh and using barycentric interpolation, we can reverse the rendering process and unfold the face in UV, hence reconstructing the visible parts of the texture directly from the input image. This initial texture is accompanied by a UV visibility mask, with 1 for pixels that are observed from the input image, and 0 for those that are occluded and, thus, need to be inpainted by our model. Texture Completion and Reflectance Prediction.Starting from the partially completed UV texture \(\mathbf{T}_{0}\) of the face and a binary visibility mask \(m\) produced by the previous step, our goal is to inpaint the remaining pixels along with the pixels of the 3 reflectance maps. We use the latent representation \(z_{\mathbf{T}_{0}}=\mathcal{E}(\mathbf{T}_{0})\in\mathbb{R}^{h\times\nu\times c}\) of this texture image to constrain the reverse diffusion process. Note that the mask \(m\) is downsampled to the same resolution \(h=w=64\) of the latent space for the next steps. Our inpainting algorithm starts with a random noise image \(z_{T}\sim\mathcal{N}(\mathbf{0},\mathbf{I})\) and uses the denoising procedure of MCG [11], consisting of the following repeated steps: \[z_{t-1}^{\text{unknown}} \sim\mathcal{N}(\mu_{\theta}(z_{t},t),\Sigma_{\theta}(z_{t},t)) \tag{9a}\] \[z_{\mathbf{T}_{t-1}}^{\text{known}} \sim\mathcal{N}(\sqrt{\bar{\alpha}_{t-1}}z_{\mathbf{T}_{0}},(1- \bar{\alpha}_{t-1})\mathbf{I})\] (9b) \[\dot{z}_{0} =\left(z_{t}-\sqrt{1-\bar{\alpha}_{t}}\epsilon_{\theta}(z_{t},t) \right)/\sqrt{\bar{\alpha}_{t}}\] (9c) \[\mathcal{L} =\left\|\left(z_{\mathbf{T}_{0}}-\hat{z}_{\mathbf{T}_{0}}\right) \odot m\right\|_{2}^{2}\] (9d) \[z_{\mathbf{T}_{t-1}} =m\odot z_{\mathbf{T}_{t-1}}^{\text{known}}+(1-m)\odot\left(z_{ \mathbf{T}_{t-1}}^{\text{unknown}}-\alpha\frac{\mathcal{L}}{\partial z_{ \mathbf{T}_{t}}}\right)\] (9e) \[z_{k_{t-1}} =z_{k_{t-1}}^{\text{unknown}}-\alpha\frac{\partial\mathcal{L}}{ \partial z_{k_{t}}},\quad k=\{\mathbf{A}_{d},\mathbf{A}_{s},\mathbf{N}\} \tag{9f}\] Given a sample \(z_{t}\) at timestep \(t\), we first sample the next denoised sample \(z_{t-1}\) using the original reverse diffusion step (Eq. 9a). We term this as \(z_{t-1}^{\text{unknown}}\) (borrowing the notation from [47]) as it does not take into account the known parts of the observed texture. To exploit the known texture, we sample a noisy version of it \(z_{\mathbf{T}_{t-1}}^{\text{known}}\) at timestep \(t-1\) via a forward diffusion step (Eq. 9b). Then, we directly impose this known noisy texture \(m\odot z_{\mathbf{T}_{t-1}}^{\text{known}}\) (\(\odot\) denotes the Hadamard product) as in the first half of Eq. 9e. Finally, for the unknown pixels, we add the manifold constraint introduced in MCG [11]; _i.e_. we make a prediction of the clean sample \(\hat{z}_{0}\) (Eq. 9c) based on the previous timestep \(z_{t}\), compare this (\(\ell_{2}\) loss) with the ground truth in the known regions (Eq. 9d), and use the gradient of this loss to update the unknown pixels of \(z_{t-1}\) (Eq. 9e and 9f) so as to minimize this distance. Note on inpainting algorithm.We have chosen to adopt the recently proposed MCG [11] inpainting algorithm, which outperforms related state-of-the-art diffusion-based methods (_e.g_. RePaint [47], DDRM [37]), as we empirically found it to produce excellent results. Motivated by the original algorithm, which aims at inpainting standard RGB images, we expand it to account for different input domains: Figure 3: Examples of 3D reconstructions by our method, rendered using different environment maps in a commercial renderer [50]. by treating our images as concatenated texture/reflectance maps, we challenge the model to perform not only spatial inpainting, but also 'channel-wise inpainting', thus predicting accurate reflectance maps from just a partial illuminated version of them. ## 4 Experiments ### Dataset and Implementation Details We create a high-quality dataset that consists of facial textures and their corresponding reflectance. Each item includes a texture \(\mathbf{T}\), shaded in some illumination, diffuse albedo \(\mathbf{A}_{d}\), specular albedo \(\mathbf{A}_{s}\) and normals \(\mathbf{N}\). To achieve this, firstly, we acquire the public MimicMe dataset [52], which contains \(\mathbf{\tilde{T}}=\{\mathbf{T}_{0},\dots,\mathbf{T}_{n_{T}}\},n_{T}=4,700\) diverse facial textures, whose statistics are reported in [52]. However, such textures contain the illumination of the scanning apparatus and are not relightable. Hence, we then train an image-to-image translation network based on AvatarMe++ model using the available dataset [42], which translates the textures \(\mathbf{\tilde{T}}\) to facial reflectance: \(\alpha(\mathbf{\tilde{T}})\rightarrow\{\mathbf{A}_{D},\mathbf{A}_{S},\mathbf{ N}\}\). Moreover, we augment the skin-tone diversity, using histogram matching albedo augmentation following [41]. Given the memory requirement of our network, all textures have a resolution of \(512\times 512\). Finally, to enable the diffusion model to perform well in "in-the-wild" images, we use the shapes \(\mathbf{S}\) of MimicMe and the acquired reflectance, to re-render the textures under arbitrary realistic environments, directly on the UV space: \(\rho(\mathbf{A}_{D},\mathbf{A}_{S},\mathbf{N},\mathbf{S})\rightarrow\mathbf{T}\). Although AvatarMe++ uses a similar method to augment training data, we do not require this process to be differentiable and use a ray-tracing renderer [50] (_Baker_ algorithm) to achieve more realistic textures. To train our model, we use a KL-regularized latent diffusion model with the default hyper-parameters proposed by the authors of [57]. Specifically, we use a downsampling factor of \(f=8\) for the perceptual auto-encoder and a diffusion length of \(T=1000\) for the denoising model. We train our model once and use it for texture and reflectance reconstruction from "in-the-wild" images. Below we provide comprehensive qualitative and quantitative evaluations. spect to the input thanks to our carefully designed inpainting approach. We also show an extensive qualitative comparison with related 3D reconstruction methods in Fig. 5 (most of which can only recover the texture), where similar observations can be made. Finally, we test our method on images from the Digital Emily [2] and show the results in Fig. 8 together with related works [18, 42]. We yield similar results regardless of the lighting, thanks to our coupled texture/reflectance modeling that combines reflectance with randomly rendered textures during training. ### Texture Completion Following [22, 14], we evaluate our method on the task of texture completion using the Multi-PIE [30] subset of the UVDB dataset [14]. This consists of complete UV textures for 337 different identities, and corresponding 2D images of the faces from various camera poses. In accordance with [22, 14], we use the last 137 subjects for evaluation (as the first 200 were used as training data in prior works). We perform texture completion with our diffusion-based approach for each different viewing angle and compare it with existing texture completion methods, namely CE [54], UV-GAN [14] and OSTeC [22]. We use the widely adopted Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics to compare the completed textures with the ground truth and report the results in Tab. 1. As can be seen, _Relighify_ outperforms the related methods in almost all settings, especially for challenging angles. A visual comparison with [22, 14] is provided in Fig. 6. Note that in contrast to CE [54] and UV-GAN [14], our model was not trained on the Multi-PIE dataset. ### Identity Preservation We perform quantitative evaluations of our method's ability to preserve the subject's identity, by comparing the distribution of identity scores between the input image and rendered reconstruction, on the LFW dataset [32], against prior work [24, 25, 26, 67]. Following the existing benchmark [25], we evaluate our results using VGG-Face [53]. We present our analysis in Fig. 7, measuring the distance between the input image and reconstruction for all subjects. Our method shows a significant improvement in similarity, while also producing not just a facial texture, but a set of relightable reflectance textures. and specular albedos, while the normals closely match that of [42]. This demonstrates our method's ability to better capture subject-specific details by directly leveraging texture information from the input image. ### Experimentation with Inpainting Algorithms Although we adopt the MCG [11] approach for our texture/reflectance diffusion model, we have experimented with different inpainting algorithms. We compare four of them in Fig. 9 and Tab. 4. We also provide the runtime for each algorithm in Tab. 3. The baseline method of Score-SDE [64], which can be interpreted as Eq. 9 without the gradient term, produces sub-optimal results, the occluded areas are often inpainted in an inconsistent way with the observed ones, which is especially apparent in the texture (Fig. 9) and albedos (Tab. 4). RePaint [47] also produces unsatisfactory textures while at the same time increasing the reverse diffusion steps by a factor of \(n\) (we use \(n=10\) as suggested by the authors of [47]), which significantly affects the computational time. In contrast, MCG [11] preserves the original sampling length (\(T=1000\) timesteps), hence being much more efficient. However, it is still slower than Score-SDE [64] since it requires the computation of a gradient for the manifold constraint at each step. In general, we found MCG [11] to perform better in most cases. To further strengthen the efficiency of our method, we have additionally incorporated the DDIM [63] acceleration technique in the MCG algorithm, which allows reducing the denoising steps to \(N<T\) (we use \(N=200\)) without a significant drop in quality. In such case, our method can generate high-quality texture and reflectance assets from a partial UV texture in roughly 12 seconds, which is significantly faster than competing texture completion algorithms (OSTeC [22] requires around 10 minutes).

2303.14925

Stratifications of abelian categories

This paper studies abelian categories that can be decomposed into smaller abelian categories via iterated recollements - such a decomposition we call a stratification. Examples include the categories of (equivariant) perverse sheaves and epsilon-stratified categories (in particular highest weight categories) in the sense of Brundan-Stroppel (2018). We give necessary and sufficient conditions for an abelian category with a stratification to be equivalent to a category of finite dimensional modules of a finite dimensional algebra - this generalizes the main result of Cipriani-Woolf (2022). Furthermore, we give necessary and sufficient conditions for such a category to be epsilon-stratified - this generalizes the characterisation of highest weight categories given by Krause (2017).

Giulian Wiggins

2023-03-27T05:48:34Z

http://arxiv.org/abs/2303.14925v1

# Stratifications of abelian categories ###### Abstract. This paper is a study of abelian categories that can be decomposed into smaller abelian categories via iterated recollements - such a decomposition we call a _stratification_. Examples include the categories of (equivariant) perverse sheaves and \(\varepsilon\)-stratified categories (in particular highest weight categories) in the sense of Brundan-Stroppel [1]. We give necessary and sufficient conditions for an abelian category with a stratification to be equivalent to a category of finite dimensional modules of a finite dimensional algebra - this generalizes the main result of Cipriani-Wolf [1]. Furthermore, we give necessary and sufficient conditions for such a category to be \(\varepsilon\)-stratified - this generalizes the characterisation of highest weight categories given by Krause [1]. ###### Contents * 1 Introduction * 1.1 Outline and explanation of main results * 2 Preliminaries * 3 The intermediate-extension functor * 4 Recollements with enough projectives/injectives * 4.1 Projective covers * 4.2 Ext-finiteness * 4.3 Main results * 5 Standard and costandard objects * 6 Brundan and Stroppel's \(\varepsilon\)-stratified categories * 6.1 Highest weight categories ## 1. Introduction A _recollement_ of abelian categories is a short exact sequence of abelian categories in which \(\mathcal{A}_{Z}\) is a Serre subcategory of \(\mathcal{A}\) (with Serre quotient \(\mathcal{A}_{U}\)), \(i_{*}\) has both a left and right adjoint, and \(j^{*}\) has fully-faithful left and right adjoints. We usually denote such a recollement of abelian categories by the diagram where \((i^{*},i_{*},i^{!})\) and \((j_{!},j^{*},j_{*})\) are adjoint triples. This definition is motivated by recollements of triangulated categories as defined by Beilinson, Bernstein and Deligne [1] to generalize Grothendieck's six functors relating the constructible derived category, \(\mathcal{D}(X,\Bbbk)\), of sheaves on a variety \(X\) (with coefficients in a field \(\Bbbk\)) and the constructible derived categories, \(\mathcal{D}(Z,\Bbbk)\) and \(\mathcal{D}(U,\Bbbk)\), of sheaves on a closed subvariety \(Z\subset X\) and open complement \(U:=X\backslash Z\) i.e. the situation: Here, \(i:Z\hookrightarrow X\) is the closed embedding and \(j:U\hookrightarrow X\) is the complimentary open embedding. Note that if \(\mathcal{S}h(X,\Bbbk)\) is the category of sheaves on \(X\), and \(\mathcal{H}^{i}:\mathcal{D}(X,\Bbbk)\to\mathcal{S}h(X,\Bbbk)\) is the \(i\)-th cohomology functor, then the following is an example of a recollement of abelian categories. More generally, given a recollement of triangulated categories with \(t\)-structure, there is a canonical recollement of abelian categories on the hearts of the \(t\)-structure [1, Proposition 1.4.16] defined using the zero-th cohomology functor arising from the \(t\)-structure. In representation theory, recollements of abelian categories arise from modules of algebras \(A\) with an idempotent \(e\). In this situation there is a recollement where \(i_{*}:A/AeA\text{-Mod}\to A\text{-Mod}\) is the restriction functor, and \(j^{*}:A\text{-Mod}\to eAe\text{-Mod}\) is the functor sending an \(A\)-module \(M\) to the \(eAe\)-module \(eM\). In applications, recollements often arise as part of an iterated collection of recollements, giving a kind of filtration of a larger category by smaller categories. Consider, for example, the category, \(P_{\Lambda}(X,\Bbbk)\), of perverse sheaves that are constructible with respect to a stratification, \(X:=\bigcup_{\lambda\in\Lambda}X_{\lambda}\), of a variety \(X\). For each strata \(X_{\lambda}\), there is a recollement where \(\operatorname{Loc}^{ft}(X_{\lambda},\Bbbk)\) is the category of local systems on \(X_{\lambda}\) of finite type. More generally, if \(\Lambda^{\prime}\subset\Lambda\) is such that \(X_{\Lambda^{\prime}}:=\bigcup_{\lambda\in\Lambda^{\prime}}X_{\lambda}\) is a closed subvariety of \(X\), and \(X_{\lambda}\) is open in \(X_{\Lambda^{\prime}}\), then the following is a recollement of abelian categories. Similar iterations of recollement occur in highest weight categories (as defined in [1]). Indeed, let \(\mathcal{A}\) be a highest weight category with respect to a poset \(\Lambda\), and let \(\{\Delta_{\lambda}\mid\lambda\in\Lambda\}\) be the collection of standard objects in \(\mathcal{A}\). For each lower1 subposet \(\Lambda^{\prime}\subset\Lambda\), let \(\mathcal{A}_{\Lambda^{\prime}}\) be the Serre subcategory generated by objects \(\{\Delta_{\lambda}\mid\lambda\in\Lambda^{\prime}\}\). For each lower subpost \(\Lambda^{\prime}\subset\Lambda\) and maximal \(\lambda\in\Lambda^{\prime}\) there is a recollement of abelian categories where \(j^{*}:=\operatorname{Hom}(\Delta_{\lambda},-):\mathcal{A}_{\Lambda^{\prime}} \rightarrow\operatorname{End}_{\mathcal{A}}(\Delta_{\lambda})^{op}\text{-Mod}\). These examples are unified by the following definition. **Definition 2.4**.: A _stratification_ of an abelian category \(\mathcal{A}\) by a non-empty poset \(\Lambda\) consists of the following data 1. For each lower subposet \(\Lambda^{\prime}\subset\Lambda\), define a Serre subcategory, \(\mathcal{A}_{\Lambda^{\prime}}\), of \(\mathcal{A}\). 2. For each \(\lambda\in\Lambda\), define an abelian category \(\mathcal{A}_{\lambda}\). These we call _strata categories_. This data must satisfy the conditions 1. \(\mathcal{A}_{\emptyset}=0\) and \(\mathcal{A}_{\Lambda}=\mathcal{A}\). 2. For each pair of lower subposets \(\Lambda^{\prime}_{1}\subset\Lambda^{\prime}_{2}\subset\Lambda\), there are inclusions of Serre categories \(\mathcal{A}_{\Lambda^{\prime}_{1}}\hookrightarrow\mathcal{A}_{\Lambda^{\prime }_{2}}\). 3. For each lower subposet \(\Lambda^{\prime}\subset\Lambda\), and maximal \(\lambda\in\Lambda^{\prime}\) there is a recollement For example, the category \(P_{\Lambda}(X,\Bbbk)\) has a stratification by the poset \(\Lambda\) with closure order: \(\lambda\leq\mu\) if \(X_{\lambda}\subset\overline{X_{\mu}}\). Further examples include equivariant perverse sheaves, and \(\varepsilon\)-stratified categories in the sense of Brundan and Stroppel [1]. This paper is inspired by a result of Cipriani and Woolf [15, Corollary 5.2] that says that a category of perverse sheaves (with coefficients in a field) on a space stratified by finitely many strata is finite2 if and only if the same is true for each category of finite type local systems on each stratum. We extend this result by showing that an abelian category with a stratification by a finite poset is finite if and only if the same if true of all strata categories (Corollary 4.10). Moreover, we give new necessary and sufficient conditions for a finite abelian category to be \(\varepsilon\)-stratified (Theorem 6.4). This result specializes to give necessary and sufficient conditions for a finite abelian category to be standardly stratified in the sense of Cline, Parshall, Scott [11], and further specialises to recover a characterisation of highest weight categories due to Krause [16]. Footnote 2: For any field \(\Bbbk\), a \(\Bbbk\)-linear abelian category \(\mathcal{A}\) is _finite_ (over \(\Bbbk\)) if \(\mathcal{A}\) is equivalent to a category of finite dimensional modules of a finite dimensional \(\Bbbk\)-algebra. This article revises and extends part of the author's PhD thesis [20]. This paper has benefited from discussions with Oded Yacobi, Kevin Coulembier, and feedback from a referee of [20]. This work was supported by Australian Research Council grant DP190102432. ### Outline and explanation of main results Let \(\mathcal{A}\) be an abelian category with a stratification by a poset \(\Lambda\). For each \(\lambda\in\Lambda\), define the Serre quotient functor \(j^{\lambda}:\mathcal{A}_{\{\mu\in\Lambda\mid\mu\leq\lambda\}}\rightarrow \mathcal{A}_{\lambda}\), and let \(j^{\lambda}_{!}:\mathcal{A}_{\lambda}\rightarrow\mathcal{A}_{\{\mu\in\Lambda \mid\mu\leq\lambda\}}\) and \(j^{\lambda}_{*}:\mathcal{A}_{\lambda}\rightarrow\mathcal{A}_{\{\mu\in\Lambda \mid\mu\leq\lambda\}}\) be the left and right adjoints of \(j^{\lambda}\). By a slight abuse of notation, write \(j^{\lambda}_{!}:\mathcal{A}_{\lambda}\rightarrow\mathcal{A}\) and \(j^{\lambda}_{*}:\mathcal{A}_{\lambda}\rightarrow\mathcal{A}\) for the functors obtained by postcomposing with the inclusion functor \(\mathcal{A}_{\{\mu\in\Lambda\mid\mu\leq\lambda\}}\hookrightarrow\mathcal{A}\). In Section 2 we recall some basic features of recollements and stratifications, and give examples of stratifications of abelian categories. In Section 3 we define, for each \(\lambda\in\Lambda\), the _intermediate-extension functor_\(j^{\lambda}_{!*}:\mathcal{A}_{\lambda}\to\mathcal{A}\): \[j^{\lambda}_{!*}X:=\operatorname{im}(\overline{1_{X}}:j^{\lambda}_{!}X\to j^{ \lambda}_{*}X),\] where \(\overline{1_{X}}\) is the morphism corresponding to the identity \(1_{X}\) under the natural isomorphism \[\operatorname{Hom}_{\mathcal{A}}(j^{\lambda}_{!}X,j^{\lambda}_{*}X)\simeq \operatorname{Hom}_{\mathcal{A}_{\lambda}}(X,j^{\lambda}j^{\lambda}_{*}X)\simeq \operatorname{Hom}_{\mathcal{A}_{\lambda}}(X,X).\] Proposition 3.4 says that every simple object \(L\in\mathcal{A}\) is of the form \(j^{\lambda}_{!*}L_{\lambda}\), for a unique (up to isomorphism) simple object \(L_{\lambda}\in\mathcal{A}_{\lambda}\) and unique \(\lambda\in\Lambda\). Proposition 3.6 says that if \(\mathcal{A}\) is an abelian category with a stratification by a finite poset, then every object in \(\mathcal{A}\) has a finite filtration by simple objects if and only if the same is true of all the strata categories. The proofs here are almost identical to the standard proofs of these results in the theory of perverse sheaves (see e.g. [1, Chapter 3]). In Section 4 we prove the following condition for a category to have enough projectives. **Theorem 4.9**.: _Consider a recollement:_ _Suppose \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have finitely many simple objects and every object has a finite filtration by simple objects. Suppose moreover that for any simple objects \(A,B\) in \(\mathcal{A}\),_ \[\dim_{\operatorname{End}_{\mathcal{A}}(B)}\operatorname{Ext}^{1}_{\mathcal{A }}(A,B)<\infty.\] _Then \(\mathcal{A}\) has enough projectives if and only if both \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have enough projectives.3_ Footnote 3: Since \(\operatorname{End}_{\mathcal{A}}(B)\) is a division ring, any \(\operatorname{End}_{\mathcal{A}}(B)\)-module is free. For an \(\operatorname{End}_{\mathcal{A}}(B)\)-module \(M\), \(\dim_{\operatorname{End}_{\mathcal{A}}(B)}M\) is the rank of \(M\) as a free \(\operatorname{End}_{\mathcal{A}}(B)\)-module. Theorem 4.9 has the following important consequence. **Corollary 4.11**.: _For any field \(\Bbbk\), a \(\Bbbk\)-linear abelian category with a stratification by a finite poset is finite if and only if the same is true for all strata categories._ To state the results in Sections 5 and 6, let \(\mathcal{A}\) have enough projectives and injectives, finitely many simple objects, and suppose every object in \(\mathcal{A}\) has finite length. Let \(B\) be a set indexing the simple objects in \(\mathcal{A}\) (up to isomorphism) and write \(L(b)\) (respectively \(P(b)\), \(I(b)\)) for the simple (respectively projective indecomposable, injective indecomposable) object corresponding to \(b\in B\). Define the _stratification function_ \[\rho:B\to\Lambda\] that maps each \(b\in B\) to the corresponding \(\lambda\in\Lambda\) in which \(L(b)=j^{\lambda}_{!*}L_{\lambda}(b)\) for some simple object \(L_{\lambda}(b)\in\mathcal{A}_{\lambda}\). Write \(P_{\lambda}(b)\) and \(I_{\lambda}(b)\) for the projective cover and injective envelope of \(L_{\lambda}(b)\) in \(\mathcal{A}_{\lambda}\). For \(b\in B\) and \(\lambda=\rho(b)\), define the _standard_ and _costandard_ objects \[\Delta(b):=j^{\lambda}_{!}P_{\lambda}(b),\qquad\nabla(b):=j^{\lambda}_{*}I_{ \lambda}(b).\] Porism 5.1 says that every projective indecomposable, \(P(b)\), in \(\mathcal{A}\) has a filtration by quotients of standard objects, \(\Delta(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). Dually, every injective indecomposable, \(I(b)\), of \(\mathcal{A}\) has a filtration by subobjects of costandard objects, \(\nabla(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). Standard and costandard objects play a crucial role in representation theoretic applications of stratifications of abelian categories, where one often requires that projective and/or injective indecomposable objects have filtrations by standard and/or costandard objects. For example, both of these conditions are required in Cline-Parshall-Scott's definition of highest weight category [1]. Categories whose projective indecomposables have filtrations by standard objects have been widely studied - beginning with the work of Dlab [10] and Cline, Parshall and Scott [1]. Categories in which both projective objects have a filtration by standard objects and injective objects have a filtration by costandard objects have been studied by various authors (see e.g. [11], [12], [13]). Brundan and Stroppel [1] (based on the work of [1]) define a general framework called an \(\varepsilon\)_-stratified category_ that includes these situations as special cases. We recall this definition now. For a _sign function_\(\varepsilon:\Lambda\to\{+,-\}\), define the \(\varepsilon\)_-standard_ and \(\varepsilon\)_-costandard objects_ \[\Delta_{\varepsilon}(b):=\left\{\begin{array}{ll}j_{\uparrow}^{\lambda}P_{ \lambda}(b)&\text{if $\varepsilon(\lambda)=+$}\\ j_{\downarrow}^{\lambda}L_{\lambda}(b)&\text{if $\varepsilon(\lambda)=-$} \end{array}\right.,\quad\nabla_{\varepsilon}(b):=\left\{\begin{array}{ll}j_ {*}^{\lambda}L_{\lambda}(b)&\text{if $\varepsilon(\lambda)=+$}\\ j_{*}^{\lambda}I_{\lambda}(b)&\text{if $\varepsilon(\lambda)=-$}\end{array}\right.,\] where \(\lambda=\rho(b)\). Brundan and Stroppel [1] say that \(\mathcal{A}\) is \(\varepsilon\)_-stratified_ if the following equivalent conditions hold: 1. Every projective indecomposable, \(P(b)\), has a filtration by \(\varepsilon\)-standard objects, \(\Delta_{\varepsilon}(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). 2. Every injective indecomposable, \(I(b)\), has a filtration by \(\varepsilon\)-costandard objects, \(\nabla_{\varepsilon}(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). Note that \(\mathcal{A}\) is a _highest weight category_ if and only if each \(\mathcal{A}_{\lambda}\) is semisimple and \(\mathcal{A}\) is \(\varepsilon\)-stratified for any (and all) \(\varepsilon:\Lambda\to\{+,-\}\). The following result gives a new criterion for a finite abelian category to be \(\varepsilon\)-stratified. **Theorem 6.4**.: _A finite abelian category \(\mathcal{A}\) with a stratification by a finite poset \(\Lambda\) is \(\varepsilon\)-stratified (for a function \(\varepsilon:\Lambda\to\{+,-\}\)) if and only if the following conditions hold:_ 1. _For each inclusion of Serre subcategories_ \(i_{*}:\mathcal{A}_{\Lambda^{\prime}}\to\mathcal{A}_{\Lambda}\)_, and objects_ \(X,Y\in\mathcal{A}_{\Lambda^{\prime}}\)_, there is a natural isomorphism_ \(\operatorname{Ext}^{2}_{\mathcal{A}_{\Lambda^{\prime}}}(X,Y)\simeq \operatorname{Ext}^{2}_{\mathcal{A}}(i_{*}X,i_{*}Y)\)_._ 2. _For each_ \(\lambda\in\Lambda\)_,_ 1. _If_ \(\varepsilon(\lambda)=+\) _then_ \(j_{*}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) _is exact._ 2. _If_ \(\varepsilon(\lambda)=-\) _then_ \(j_{\uparrow}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) _is exact._ One application of Theorem 6.4 is a criterion for a category of (equivariant) perverse sheaves to be \(\varepsilon\)-stratified - although we remark that checking these conditions (particularly the first condition) is difficult in general. ## 2. Preliminaries We begin with an axiomatic definition of recollement. The notation used in this definition will be used throughout the paper. **Definition 2.1**.: A _recollement of abelian categories_ consists of three abelian categories \(\mathcal{A}_{Z}\), \(\mathcal{A}\) and \(\mathcal{A}_{U}\) and functors: (2.1) satisfying the conditions: 1. \((i^{*},i_{*},i^{!})\) and \((j_{!},j^{*},j_{*})\) are adjoint triples. 2. The functors \(i_{*}\), \(j_{!}\), \(j_{*}\) are fully-faithful. Equivalently the adjunction maps \(i^{*}i_{*}\to\operatorname{Id}\to i^{!}i_{*}\) and \(j^{*}j_{*}\to\operatorname{Id}\to j^{*}j_{!}\) are isomorphisms. 3. The functors satisfy \(j^{*}i_{*}=0\) (and so by adjunction \(i^{*}j_{!}=0=i^{!}j_{*}\)). 4. The adjunction maps produce exact sequences for each object \(X\in\mathcal{A}\): (2.2) \[j_{!}j^{*}X\to X\to i_{*}i^{*}X\to 0\] (2.3) \[0\to i_{*}i^{!}X\to X\to j_{*}j^{*}X\] Alternatively, condition (R4) can be replaced by the condition 1. For any object \(X\in\mathcal{A}\), if \(j^{*}X=0\) then X is in the essential image of \(i_{*}\). A _recollement of triangulated categories_ is defined in the same way as a recollement of abelian categories except that condition (R4) is replaced by the existence of the triangles: \[j_{!}j^{*}X\to X\to i_{*}i^{*}X\to \tag{2.5}\] \[i_{*}i^{!}X\to X\to j_{*}j^{*}X\to \tag{2.4}\] for each object \(X\). **Remark 2.2**.: The interchangibility of (R4) and (R4') follows from the following argument. If \(j^{*}X=0\) then (R4) implies that \(i_{*}i^{!}X\simeq X\simeq i_{*}i^{*}X\) and so \(X\) is in the essential image of \(i_{*}\). Conversely let \(\mu:j_{!}j^{*}\to\operatorname{Id}\) and \(\eta:\operatorname{Id}\to i_{*}i^{*}\) be the adjunction natural transformations. Then there is a commutative diagram in which the rows are exact. By applying \(j^{*}\) to the top row we see that \(j^{*}\operatorname{cok}\mu_{X}=0\) and so (R4') implies that \(\operatorname{cok}\mu_{X}\simeq i_{*}i^{*}(\operatorname{cok}\mu_{X})\simeq i _{*}i^{*}X\). Equation (2.3) holds by a similar argument. Write \(\mathcal{A}^{Z}\) for the essential image of \(i_{*}\). To reconcile Definition 2.1 with the definition of recollement in Section 1, note that by (R2), \(\mathcal{A}^{Z}\simeq\mathcal{A}_{Z}\), and by (R4'), \(\mathcal{A}^{Z}\) is the kernel of the exact functor \(j^{*}\) and is hence a Serre subcategory of \(\mathcal{A}\). It will be useful to note that if we extend the sequences (2.2) and (2.3) to exact sequences \[0\to K\to j_{!}j^{*}X\to X\to i_{*}i^{*}X\to 0 \tag{2.7}\] \[0\to i_{*}i^{!}X\to X\to j_{*}j^{*}X\to K^{\prime}\to 0 \tag{2.6}\] then \(K\) and \(K^{\prime}\) are in \(\mathcal{A}^{Z}\). Indeed, by applying the exact functor \(j^{*}\) to (2.6) we get that \(j^{*}K=0\) and so \(i_{*}i^{!}K\simeq K\simeq i_{*}i^{*}K\). Likewise by applying \(j^{*}\) to (2.7) we get that \(K^{\prime}\in\mathcal{A}^{Z}\). Given a recollement of abelian or triangulated categories with objects and morphisms as in (2.1), the opposite categories form the following recollement which we call the _opposite recollement_. The following proposition describes a useful way to characterise the functors \(i^{*}\) and \(i^{!}\) in any recollement. **Proposition 2.3**.: _Let \(\mathcal{A}\) be an abelian category with a recollement as in (2.1). Then for any object \(X\in\mathcal{A}\):_ 1. \(i_{*}i^{*}X\) _is the largest quotient object of_ \(X\) _in_ \(\mathcal{A}^{Z}\)_._ 2. \(i_{*}i^{!}X\) _is the largest subobject of_ \(X\) _in_ \(\mathcal{A}^{Z}\)_._ Proof.: By the adjunction \((i_{*},i^{*})\) and since \(i_{*}\) is fully-faithful we have natural isomorphisms for \(X\in\mathcal{A}\), \(Y\in\mathcal{A}_{Z}\): \[\operatorname{Hom}_{\mathcal{A}}(i_{*}i^{*}X,i_{*}Y)\simeq\operatorname{Hom} _{\mathcal{A}_{Z}}(i^{*}X,Y)\simeq\operatorname{Hom}_{\mathcal{A}}(X,i_{*}Y)\] sending \(f\) to \(f\circ\eta\) where \(\eta:X\to i_{*}i^{*}X\) is the adjunction unit. In particular any morphism \(X\to i_{*}Y\) factors through \(i_{*}i^{*}X\). Statement (i) follows. Statement (ii) follows by taking the opposite recollement. Say that a subset \(\Lambda^{\prime}\) of a poset \(\Lambda\) is _lower_ if for any \(\lambda\in\Lambda^{\prime}\), if \(\mu\leq\lambda\) then \(\mu\in\Lambda^{\prime}\). **Definition 2.4**.: A _stratification_ of an abelian/triangulated category \(\mathcal{A}\) by a non-empty poset \(\Lambda\) consists of the following data: 1. An assignment of an abelian/triangulated category \(\mathcal{A}_{\Lambda^{\prime}}\) for every lower \(\Lambda^{\prime}\subset\Lambda\), and for lower subsets \(\Lambda^{\prime\prime}\subset\Lambda^{\prime}\subset\Lambda\) an embedding \(i_{\Lambda^{\prime\prime},\Lambda^{\prime*}}:\mathcal{A}_{\Lambda^{\prime \prime}}\hookrightarrow\mathcal{A}_{\Lambda^{\prime}}\). 2. For each \(\lambda\in\Lambda\) an abelian/triangulated category \(\mathcal{A}_{\lambda}\). We call these _strata categories_. This data must satisfy the following conditions 1. \(\mathcal{A}_{\emptyset}=0\) and \(\mathcal{A}_{\Lambda}=\mathcal{A}\). 2. For each \(\lambda\in\Lambda\) and lower subset \(\Lambda^{\prime}\subset\Lambda\) in which \(\lambda\in\Lambda^{\prime}\) is maximal, the functor \(i_{*}=i_{\Lambda^{\prime}\setminus\{\lambda\},\Lambda^{\prime*}}\) fits into a recollement \[\mathcal{A}_{\Lambda^{\prime}\setminus\{\lambda\}}\] 3. If \(\Lambda^{\prime\prime}\subset\Lambda^{\prime}\) are lower subsets of \(\Lambda\), and \(\lambda\in\Lambda\) is a maximal element of both \(\Lambda^{\prime\prime}\) and \(\Lambda^{\prime}\), then the following diagram of functors commutes We proceed with some important examples of recollements and stratifications. **Example 2.5** (Constructible sheaves with respect to a stratification).: A _stratification_ of a quasiprojective complex variety \(X\) is a finite collection, \(\{X_{\lambda}\}_{\lambda\in\Lambda}\), of disjoint, smooth, connected, locally closed subvarieties, called _strata_, in which \(X=\bigcup_{\lambda\in\Lambda}X_{\lambda}\) and for each \(\lambda\in\Lambda\), \(\overline{X_{\lambda}}\) is a union of strata. In this case we equip \(\Lambda\) with the partial order \[\mu\leq\lambda\text{ if }X_{\mu}\subset\overline{X_{\lambda}}.\] We use \(\Lambda\) to refer to the stratification of \(X\). For a variety \(X\), let \(\operatorname{Loc}^{ft}(X,\Bbbk)\) be the category of local systems on \(X\) of finite type with coefficients in a field \(\Bbbk\). Recall that, by taking monodromy, \(\operatorname{Loc}^{ft}(X,\Bbbk)\) is equivalent to the category, \(\Bbbk[\pi_{1}(X_{\lambda})]\text{-mod}_{fg}\), of finitely generated \(\Bbbk[\pi_{1}(X_{\lambda})]\)-modules (see e.g. [1, Theorem 1.7.9]). Say that a sheaf \(\mathcal{F}\) on \(X\) is _constructible_ with respect to a stratification, \(\Lambda\), of \(X\) if \(\mathcal{F}|_{X_{\lambda}}\) is a local system of finite type for each \(\lambda\in\Lambda\). Write \(\mathcal{D}^{b}_{\Lambda}(X,\Bbbk)\) for the full triangulated subcategory of \(\mathcal{D}^{b}(X,\Bbbk)\) consisting of objects \(\mathcal{F}\) in which \(H^{k}(\mathcal{F})\) is constructible with respect to \(\Lambda\). Say that a stratification, \(\Lambda\), of \(X\) is _good_ if for any \(\lambda\in\Lambda\) and any object \(\mathcal{L}\in\operatorname{Loc}^{ft}(X_{\lambda},\Bbbk)\), we have \(j_{\lambda*}\mathcal{L}\in\mathcal{D}^{b}_{\Lambda}(X,\Bbbk)\), where \(j_{\lambda}:X_{\lambda}\hookrightarrow X\) is the embedding, and \(j_{\lambda*}\) is the derived pushforward. It is difficult to tell in general whether a stratification is good (see [1, Remark 2.3.21] for a discussion of these difficulties). A stratification satisfying the _Whitney regularity conditions_[10] is good. In particular, if an algebraic group \(G\) acts on \(X\) with finitely many orbits (each connected), then the stratification of \(X\) by \(G\)-orbits is good (see e.g. [1, Exercise 6.5.2]). Given a good stratification \(\Lambda\) on \(X\), the triangulated category \(\mathcal{D}^{b}_{\Lambda}(X,\Bbbk)\) has a stratification by \(\Lambda\) with strata categories \[\mathcal{D}_{\lambda}:=\mathcal{D}^{b}(\operatorname{Loc}^{ft}(X_{\lambda}, \Bbbk))\simeq\mathcal{D}^{b}(\Bbbk[\pi_{1}(X_{\lambda})]\text{-mod}_{fg})\] and Serre subcategories \(\mathcal{D}_{\Lambda^{\prime}}:=\mathcal{D}^{b}_{\Lambda^{\prime}}(\bigcup_{ \lambda\in\Lambda^{\prime}}X_{\lambda})\) for each lower \(\Lambda^{\prime}\subset\Lambda\). For a perversity function \(p:\Lambda\to\mathbb{Z}\), the category \({}^{p}P_{\Lambda}(X,\Bbbk)\) of perverse sheaves on \(X\) with respect to the stratification \(\Lambda\) (and perversity function \(p\)) is the full subcategory of \(\mathcal{D}^{b}_{\Lambda}(X,\Bbbk)\) consisting of complexes \(\mathcal{F}\) in which for any strata \(h_{\lambda}:X_{\lambda}\hookrightarrow X\): 1. \(\mathcal{H}^{d}(h_{\lambda}^{*}\mathcal{F})=0\) if \(d>p(\lambda)\), 2. \(\mathcal{H}^{d}(h_{\lambda}^{!}\mathcal{F})=0\) if \(d<p(\lambda)\), where \(\mathcal{H}^{d}(\mathcal{F})\) refers to the \(d\)-th cohomology sheaf of \(\mathcal{F}\). The category \(\mathcal{A}={}^{p}P_{\Lambda}(X,\Bbbk)\) is abelian and has a stratification by \(\Lambda\), with strata categories \[\mathcal{A}_{\lambda}=\operatorname{Loc}^{ft}(X_{\lambda},\Bbbk)[p(\lambda) ]\simeq\Bbbk[\pi_{1}(X_{\lambda})]\text{-mod}_{fg}.\] **Example 2.6** (\(G\)-equivariant perverse sheaves).: Another example of a stratification arises in the theory of equivariant perverse sheaves as defined in [1]. For a complex algebraic group \(G\) and quasiprojective complex \(G\)-variety \(X\), a \(G\)_-equivariant perverse sheaf_ on \(X\) is, roughly speaking, a perverse sheaf on \(X\) with a \(G\)-action compatible with the \(G\)-action on \(X\) (see e.g. [1, Definition 6.2.3] for a precise definition). The category, \(P_{G}(X,\Bbbk)\) of \(G\)-equivariant perverse sheaves is the heart of a \(t\)-structure on the \(G\)_-equivariant derived category_, \(\mathcal{D}_{G}(X,\Bbbk)\) defined by Bernstein-Lunts [1]. For a \(G\)-equivariant map of \(G\)-varieties \(h:X\to Y\), there are equivariant versions of the (proper) pushforward and (proper) pullback functors: \(h_{*},h_{!},h^{!},h^{*}\). If \(i:Z\hookrightarrow X\) is the inclusion of a \(G\)-invariant closed subvariety with open complement \(j:U\hookrightarrow X\), then there is a recollement of triangulated categories If \(X\) is a homogeneous \(G\)-variety, then every \(G\)-equivariant perverse sheaf is a finite type local system (shifted by \(\dim_{\mathbb{C}}X\)). Moreover, in this case, \[P_{G}(X,\Bbbk)\simeq\Bbbk[G^{x}/(G^{x})^{\circ}]\text{-mod}_{fg}, \tag{2.8}\] where \(G^{x}\subset G\) is the stabilizer of a point \(x\in X\), and \((G^{x})^{\circ}\) is the connected component of \(G^{x}\) containing the identity element (see e.g. [1, Proposition 6.2.13] for a proof of this statement). Suppose \(G\) acts on \(X\) with finitely many orbits (each connected). Let \(\Lambda\) be a set indexing the set of \(G\)-orbits in \(X\), and write \(\mathcal{O}_{\lambda}\) for the orbit corresponding to \(\lambda\in\Lambda\). Consider \(\Lambda\) as a poset with the closure order: \(\lambda\leq\mu\) if \(\mathcal{O}_{\lambda}\subset\overline{\mathcal{O}_{\mu}}\). Then the category \(\mathcal{A}=P_{G}(X,\Bbbk)\) has a stratification with strata categories \[\mathcal{A}_{\lambda}=P_{G}(\mathcal{O}_{\lambda},\Bbbk)\simeq\Bbbk[G^{x_{ \lambda}}/(G^{x_{\lambda}})^{\circ}]\text{-mod}_{fg},\] where \(x_{\lambda}\in\mathcal{O}_{\lambda}\). **Example 2.7** (Modules with idempotents).: For a ring \(A\), let \(\operatorname{Mod-}A\) be the category of all right \(A\)-modules, and \(\operatorname{mod-}A\) be the category of finitely presented right \(A\)-modules. Let \(e\) be an idempotent in \(A\), and define the inclusion functor \(i_{*}:\operatorname{Mod-}A/AeA\to\operatorname{Mod-}A\). Note that \(\operatorname{Mod-}A/AeA\) is equivalent to the Serre subcategory of \(\operatorname{Mod-}A\) consisting of modules annihilated by \(e\). There is a corresponding Serre quotient \(j^{*}:\operatorname{Mod-}A\to\operatorname{Mod-}eAe\) defined \[j^{*}:=\operatorname{Hom}_{A}(eA,-)\simeq-\otimes_{A}Ae.\] i.e. \(j^{*}M=Me\) for any object \(M\in\operatorname{Mod-}A\). These functors fit into a recollement of abelian categories where for any right \(A\)-module \(M\): 1. \(i^{*}M\) is the largest quotient, \(N\), of \(M\) in which \(Ne=0\). 2. \(i^{!}M\) is the largest subobject, \(N\), of \(M\) in which \(Ne=0\). Moreover \(j_{!}:=-\otimes_{eAe}eA\) and \(j_{*}:=\operatorname{Hom}_{eAe}(Ae,-)\). If \(A\) is right artinian and has enough injectives then the inclusion \(i_{*}:\operatorname{mod-}A/AeA\to\operatorname{mod-}A\) fits into a recollement in which \(j^{*}\) has left adjoint \(j_{!}=-\otimes_{eAe}eA\) (see e.g [16, Lemma 2.5]). **Example 2.8** (Macpherson-Vilonen construction).: Let \(\mathcal{A}_{Z}\), \(\mathcal{A}_{U}\) be abelian categories, \(F:\mathcal{A}_{U}\to\mathcal{A}_{Z}\) be a right exact functor, \(G:\mathcal{A}_{U}\to\mathcal{A}_{Z}\) be a left exact functor, and let \(\varepsilon:F\to G\) be a natural transformation. Macpherson and Vilonen [14] define a category, \(\mathcal{A}(\varepsilon)\), whose objects are tuples \((X_{U},X_{Z},\alpha,\beta)\), where \((X_{U},X_{Z})\in\operatorname{Obj}\mathcal{A}_{U}\times\operatorname{Obj} \mathcal{A}_{Z}\) and \(\alpha:F(X_{U})\to X_{Z}\), \(\beta:X_{Z}\to G(X_{U})\) are morphisms in \(\mathcal{A}_{Z}\) in which the following diagram commutes A morphism \((X_{U},X_{Z},\alpha,\beta)\to(X_{U}^{\prime},X_{Z}^{\prime},\alpha^{\prime}, \beta^{\prime})\) is a pair \[f=(f_{U}:X_{U}\to X_{U}^{\prime},f_{Z}:X_{Z}\to X_{Z}^{\prime})\in \operatorname{Mor}\mathcal{A}_{U}\times\operatorname{Mor}\mathcal{A}_{Z}\] in which the following prism commutes: Macpherson and Vilonen show [14, Proposition 1.1] that the category \(\mathcal{A}(\varepsilon)\) is abelian. Moreover they show that the category \(\mathcal{A}(\varepsilon)\) fits into a recollement in which \[i_{*}(X_{Z})=(0,X_{Z},0,0), j^{*}(X_{U},X_{Z},\alpha,\beta)=X_{U},\] \[i^{*}(X_{U},X_{Z},\alpha,\beta)=\operatorname{cok}\alpha, j_{!}(X_{U})=(X_{U},F(X_{U}),1_{F(X_{U})},\varepsilon_{X_{U}}),\] \[i^{!}(X_{U},X_{Z},\alpha,\beta)=\ker\beta, j_{*}(X_{U})=(X_{U},G(X_{U}), \varepsilon_{X_{U}},1_{G(X_{U})}).\] Macpherson and Vilonen [14] use iterations of this formal construction to construct the category of perverse sheaves on a stratified variety. Franjou and Pirashvili [13, Theorem 8.7] give necessary and sufficient conditions for a recollement of abelian categories to be equivalent to a recollement built using the Macpherson-Vilonen construction. Note that the functor \(i_{*}:\mathcal{A}_{U}\to\mathcal{A}(\varepsilon)\) has an exact retraction \(i^{!*}:\mathcal{A}(\varepsilon)\to\mathcal{A}_{Z}\) defined \(i^{!*}(X_{U},X_{Z},\alpha,\beta)=X_{Z}\). This is a special feature of recollements built using Macpherson-Vilonen's construction - the functor \(i_{*}\) does not usually have an exact retraction in a general recollement. The functor \(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}(\varepsilon)\) defined \(j_{!*}(X_{U})=(X_{U},\operatorname{im}\varepsilon_{X_{U}},\varepsilon_{X_{U}},1)\) is a special case of an intermediate-extension functor defined in Definition 3.1 below. Note that every simple object in \(\mathcal{A}(\varepsilon)\) is either of the form \(i_{*}L\) for a simple object \(L\) in \(\mathcal{A}_{Z}\), or of the form \(j_{!*}L\) for a simple object in \(\mathcal{A}_{U}\). This is a special case of Proposition 3.4 below. ## 3. The intermediate-extension functor Consider again a recollement: (3.1) In this section we study the full subcategory, \(\mathcal{A}^{U}\hookrightarrow\mathcal{A}\), whose objects have no subobjects or quotients in \(\mathcal{A}^{Z}:=\operatorname{im}i_{*}\). The main result of this section (Proposition 3.3(ii)) is that the restricted functor \(j^{*}:\mathcal{A}^{U}\to\mathcal{A}_{U}\) is an equivalence of categories. The quasi-inverse \(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}^{U}\) is defined as follows. **Definition 3.1** (\(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}^{U}\)).: For an object \(X\in\mathcal{A}_{U}\), let \(\overline{1_{X}}:j_{!}X\to j_{*}X\) be the morphism corresponding to the identity on \(X\) under the isomorphism \[\operatorname{Hom}_{\mathcal{A}}(j_{!}X,j_{*}X)\simeq\operatorname{Hom}_{ \mathcal{A}_{U}}(X,j^{*}j_{*}X)\simeq\operatorname{Hom}_{\mathcal{A}_{U}}(X,X).\] Define \[j_{!*}X:=\operatorname{im}(\overline{1_{X}}:j_{!}X\to j_{*}X)\in\mathcal{A}.\] It is easy to check that if \(X\in\mathcal{A}_{U}\) then \(j_{!*}X\in\mathcal{A}^{U}\). Indeed as \(i^{!}j_{*}X=0\), \(j_{*}X\) has no subobjects in \(\mathcal{A}^{Z}\). In particular, as \(j_{!*}X\) is a subobject of \(j_{*}X\) it cannot have any subobjects in \(\mathcal{A}^{Z}\). Likewise as \(j_{!*}X\) is a quotient of \(j_{!}X\) it cannot have any quotients in \(\mathcal{A}^{Z}\). We call the functor \(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}\) an _intermediate-extension functor_. **Remark 3.2**.: Not every subquotient of an object in \(\mathcal{A}^{U}\) need be in \(\mathcal{A}^{U}\). In particular, an object in \(\mathcal{A}^{U}\) may still have simple composition factors in \(\mathcal{A}^{Z}\). **Proposition 3.3**.: _Let \(\mathcal{A}\) be an abelian category with a recollement of abelian categories as in (3.1). Then_ * _If_ \(X\in\mathcal{A}\) _has no nonzero quotient objects in_ \(\mathcal{A}^{Z}\)_, and_ \(Y\in\mathcal{A}\) _has no nonzero subobjects in_ \(\mathcal{A}^{Z}\) _(i.e._ \(i^{*}X=0\) _and_ \(i^{!}Y=0\)_), then_ \[\operatorname{Hom}_{\mathcal{A}}(X,Y)\simeq\operatorname{Hom}_{\mathcal{A}_{U}} (j^{*}X,j^{*}Y).\] * \(j^{*}:\mathcal{A}^{U}\to\mathcal{A}_{U}\) _is an equivalence of categories with quasi-inverse_ \(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}^{U}\)_._ Proof.: If \(i^{*}X=0\) then (2.6) gives an exact sequence \[0\to K\to j_{!}j^{*}X\to X\to 0\] in which \(K\simeq i_{*}i^{!}K\). So applying \(\operatorname{Hom}(-,Y)\) we get the exact sequence \[0\to\operatorname{Hom}(X,Y)\to\operatorname{Hom}(j_{!}j^{*}X,Y)\to \operatorname{Hom}(i_{*}i^{!}K,Y).\] Applying adjunctions gives the exact sequence \[0\to\operatorname{Hom}(X,Y)\to\operatorname{Hom}(j^{*}X,j^{*}Y)\to \operatorname{Hom}(i^{!}K,i^{!}Y).\] Statement (i) follows as \(i^{!}Y=0\). A corollary of statement (i) is that \(j^{*}:\mathcal{A}^{U}\to\mathcal{A}_{U}\) is fully-faithful. To show that \(j^{*}\) is essentially surjective it suffices to show that for any object \(X\in\mathcal{A}_{U}\), \(j^{*}j_{!*}X\simeq X\). Now, as \(j^{*}\) is exact: \[j^{*}j_{!*}X=j^{*}\operatorname{im}(j_{!}X\to j_{*}X)\simeq\operatorname{im}(j ^{*}j_{!}X\to j^{*}j_{*}X)\simeq\operatorname{im}(\operatorname{Id}:X\to X)=X\] and so (ii) follows. If \(\mathcal{A}\) has a stratification by a finite poset \(\Lambda\), then for each \(\lambda\in\Lambda\), there is a functor \(j_{!*}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) defined by the composition **Proposition 3.4**.: _Let \(\mathcal{A}\) be an abelian category with a stratification by a finite poset \(\Lambda\). Every simple object \(L\in\mathcal{A}\) is of the form \(j_{!*}^{\lambda}L_{\lambda}\), for a unique (up to isomorphism) simple object \(L_{\lambda}\in\mathcal{A}_{\lambda}\) and unique \(\lambda\in\Lambda\)._ Proof.: Suppose \(\mathcal{A}\) fits into a recollement as in (3.1). By Proposition 3.3, if \(L\in\mathcal{A}_{U}\) is a simple object, then \(j_{!*}L\) is a simple object in \(\mathcal{A}\). Moreover all the simple objects of \(\mathcal{A}\) are either of the form \(i_{*}L\) for a simple object \(L\in\mathcal{A}_{Z}\), or of the form \(j_{!*}L\) for a simple object \(L\in\mathcal{A}_{U}\). The statement follows via an induction argument on \(|\Lambda|\). The following properties of the intermediate-extension functor will be useful. **Proposition 3.5**.: _Let \(\mathcal{A}\) be an abelian category with a recollement of abelian categories as in (3.1). Then_ 1. _The functor_ \(j_{!*}:\mathcal{A}_{U}\to\mathcal{A}\) _maps injective morphisms to injective morphisms and surjective morphisms to surjective morphisms._ 2. _If_ \(X\in\mathcal{A}\) _has no nonzero quotient objects in_ \(\mathcal{A}^{Z}\) _then there is a canonical short exact sequence_ \[0\to i_{*}i^{!}X\to X\to j_{!*}j^{*}X\to 0\] 3. _If_ \(X\in\mathcal{A}\) _has no nonzero subobjects in_ \(\mathcal{A}^{Z}\) _then there is a canonical short exact sequence_ \[0\to j_{!*}j^{*}X\to X\to i_{*}i^{*}X\to 0\] Proof.: Let \(f:X\to Y\) be a map in \(\mathcal{A}_{U}\) and define objects \(K_{1}\), \(K_{2}\) in \(\mathcal{A}\) by the exact sequence \[0\to K_{1}\to j_{!*}X\to j_{!*}Y\to K_{2}\to 0\] To prove statement (i) it suffices to show that if \(j^{*}K_{i}=0\) then \(K_{i}=0\). If \(j^{*}K_{i}=0\) then by (R4'), \(K_{1}\simeq i_{*}i^{*}K_{1}\) and \(K_{2}\simeq i_{*}i^{!}K_{2}\). Then each \(K_{i}=0\) since \(j_{!*}X\) and \(j_{!*}Y\) are in \(\mathcal{A}^{U}\). To prove statement (ii), let \(X\in\mathcal{A}\) have no nonzero quotients in \(\mathcal{A}^{Z}\) and consider the short exact sequence \[0\to i_{*}i^{!}X\to X\to K\to 0\] Applying \(i^{!}\) to the sequence we see that \(i^{!}K=0\) and so \(K\in\mathcal{A}^{U}\). So \(K\simeq j_{!*}j^{*}K\) and (by applying \(j^{*}\) to this sequence) \(j^{*}X\simeq K\). Statement (ii) follows immediately. The proof of statement (iii) is similar. Say that an abelian category is a _length category_ if every object has a finite filtration by simple objects. **Proposition 3.6**.: _If \(\mathcal{A}\) be an abelian category with a stratification by a finite poset, then \(\mathcal{A}\) is a length category if and only if all the strata categories are length categories._ Proof.: Let \(\mathcal{A}\) be an abelian category fitting into a recollement of abelian categories as in (3.1). It suffices to show that \(\mathcal{A}\) is a length category if and only if both \(\mathcal{A}_{Z}\) and \(\mathcal{A}_{U}\) are length categories. The result follows from this statement by induction. Let \(X\) be an object in \(\mathcal{A}\) and let \(K\) be defined by the short exact sequence \[0\to i_{*}i^{!}X\to X\to K\to 0\] Then \(i^{!}K=0\) and so applying Proposition 3.5(iii) we get the short exact sequence \[0\to j_{*}j^{*}K\to K\to i_{*}i^{*}K\to 0\] In particular if every object in \(\mathcal{A}_{Z}\) and every object in \(\mathcal{A}_{U}\) has a finite filtration by simple objects, then so does \(K\) and hence so does \(X\). The converse statement is obvious. ## 4. Recollements with enough projectives/injectives In this section we study the relationship between projective covers of objects in the different categories of a recollement. More precisely, let \(\mathcal{A}\) be a category fitting into a recollement as in (3.1). Proposition 4.5 says that if \(\mathcal{A}\) has enough projectives/injectives then so does \(\mathcal{A}_{U}\). Proposition 4.6 says that if \(\mathcal{A}\) is a Krull-Schmidt category then if \(\mathcal{A}\) has enough projectives/injectives then so does \(\mathcal{A}_{Z}\). Proposition 4.7 says that if \(X\in\mathcal{A}_{U}\) has a projective cover \(P\) in \(\mathcal{A}_{U}\) then \(j_{!}P\) is a projective cover in \(\mathcal{A}\) of \(j_{!*}X\). Unfortunately it is not easy to find a projective cover in \(\mathcal{A}\) of an object \(i_{*}X\in\mathcal{A}^{Z}\), even if a projective cover of \(X\) exists in \(\mathcal{A}_{Z}\). Theorem 4.9 gives sufficient conditions for such a projective cover to exist. A consequence of Theorem 4.9 (Corollary 4.11) is that a category \(\mathcal{A}\) with a stratification by a finite poset is equivalent to a category of finite dimensional modules of a finite dimensional algebra if and only if the same is true of the strata categories. ### Projective covers Recall that a surjection \(\phi:X\to Y\) is _essential_ if for any morphism \(\alpha:X^{\prime}\to X\), if \(\phi\circ\alpha\) is surjective then \(\alpha\) is surjective. Equivalently \(\phi:X\to Y\) is essential if for any subobject \(U\subset X\), if \(U+\ker\phi=X\) then \(U=X\). If \(P\to X\) is an essential surjection and \(P\) is projective then we call \(P\) (or more accurately the morphism \(P\to X\)) a _projective cover_ of \(X\). The projective cover of \(X\) (if it exists) factors through every other essential cover of \(X\), and is unique up to isomorphism. If \(L\in\mathcal{A}\) is a simple object and \(P\) is projective then \(\phi:P\to L\) is a projective cover if and only if the following equivalent conditions hold: 1. \(\ker\phi\) is the unique maximal subobject of \(P\). 2. The endomorphism ring of \(P\) is local. See e.g. [15, Lemma 3.6] for a proof of these facts. The dual concept of an essential surjection is called an _essential extension_. If \(X\to I\) is an essential extension and \(I\) is injective then this extension is called the _injective envelope_ of \(X\). An abelian category has _enough projectives_ (resp. _enough injectives_) if every object has a projective cover (resp. injective envelope). An abelian category \(\mathcal{A}\) is a _Krull-Schmidt category_ if every object in \(\mathcal{A}\) is a finite direct sum of objects with local endomorphism rings. For example, any abelian length category is a Krull-Schmidt category. In a Krull-Schmidt category, the projective covers of simple objects are exactly the projective indecomposable objects. Moreover a Krull-Schmidt category \(\mathcal{A}\) has enough projectives if and only if every simple object has a projective cover. We will need the following well-known characterisation of projective covers of simple objects in Krull-Schmidt categories. **Proposition 4.1**.: _Let \(\mathcal{A}\) be a Krull-Schmidt category. Let \(P\in\mathcal{A}\) be a projective object and \(L\in\mathcal{A}\) be a simple object. A map \(P\to L\) is a projective cover if and only if for any simple object \(L^{\prime}\)_ \[\dim_{\operatorname{End}_{\mathcal{A}}(L^{\prime})}\operatorname{Hom}_{ \mathcal{A}}(P,L^{\prime})=\begin{cases}1&\text{ if }L=L^{\prime}\text{,}\\ 0&\text{ otherwise.}\end{cases} \tag{4.1}\] **Remark 4.2**.: Since \(\operatorname{End}_{\mathcal{A}}(B)\) is a division ring, any \(\operatorname{End}_{\mathcal{A}}(B)\)-module is free. For an \(\operatorname{End}_{\mathcal{A}}(B)\)-module \(M\), we write \(\dim_{\operatorname{End}_{\mathcal{A}}(B)}M\) for the rank of \(M\) as a free \(\operatorname{End}_{\mathcal{A}}(B)\)-module. Proof of Proposition 4.1.: Let \(\phi:P\to L\) be a projective cover of a simple object. Since \(\ker\phi\) is the unique maximal subobject of \(\phi\), \(\operatorname{Hom}_{\mathcal{A}}(P,L^{\prime})=0\) whenever \(L\neq L^{\prime}\). To show equation (4.1), it remains to show that the \(\operatorname{End}_{\mathcal{A}}(L)\)-equivariant map \[-\circ\phi:\operatorname{End}_{\mathcal{A}}(L)\to\operatorname{Hom}_{\mathcal{ A}}(P,L)\] is an isomorphism. Since \(\phi\) is a surjection this map is injective. To show surjectivity, let \(f\in\operatorname{Hom}_{\mathcal{A}}(P,L)\) be nonzero. Then as \(\ker f\) is a maximal subobject of \(P\), \(\ker\phi\subset\ker f\), and so \(f\) factors through \(\phi\). Conversely, if (4.1) holds, then if \(P=P_{1}\oplus P_{2}\), only one \(P_{i}\) can have a simple quotient and the other must be zero. In particular, \(P\) is indecomposable. ### Ext-finiteness To state the main result of this section (Theorem 4.9) we need the concept of _\(\operatorname{Ext}\)-finiteness_. In this section we recall this definition and give two results about Ext-finiteness (Propositions 4.3 and 4.4) that will be needed in the discussion following Theorem 4.9. For \(k\in\mathbb{N}\), say that an abelian category \(\mathcal{A}\) is _\(\operatorname{Ext}^{k}\)-finite_ if for any simple objects \(A,B\) in \(\mathcal{A}\), \[\dim_{\operatorname{End}_{\mathcal{A}}(B)}\operatorname{Ext}^{k}_{\mathcal{A} }(A,B)<\infty.\] Note that if \(\mathcal{A}\) is a \(\Bbbk\)-linear category, for some field \(\Bbbk\), then \[\dim_{\operatorname{End}_{\mathcal{A}}(B)}\operatorname{Ext}^{k}_{\mathcal{A} }(A,B)=\frac{\dim_{\Bbbk}\operatorname{Ext}^{k}_{\mathcal{A}}(A,B)}{\dim_{ \Bbbk}\operatorname{End}_{\mathcal{A}}(B)}.\] So \(\mathcal{A}\) is \(\operatorname{Ext}^{k}\)-finite whenever \(\dim_{\Bbbk}\operatorname{Ext}^{k}_{\mathcal{A}}(A,B)<\infty\) for every simple object \(A,B\). The converse is true if the endomorphism ring of every simple object has finite \(\Bbbk\)-dimension (e.g. if \(\Bbbk\) is algebraically closed). The following two propositions give useful criteria for a category to be \(\operatorname{Ext}^{k}\)-finite. **Proposition 4.3**.: _Any abelian length category with enough projectives is \(\operatorname{Ext}^{k}\)-finite for every \(k\in\mathbb{N}\)._ Proof.: Let \(\mathcal{A}\) be an abelian length category. Let \(X\) be an object in \(\mathcal{A}\), and let \(Y\) be a simple object in \(\mathcal{A}\). Consider a projective presentation of X: \[0\to K\to P\to X\to 0\] Since \(\operatorname{Ext}^{k}_{\mathcal{A}}(P,Y)=0\) there is a \(\operatorname{End}_{\mathcal{A}}(Y)\)-equivariant surjection of \(\operatorname{Ext}^{k-1}_{\mathcal{A}}(K,Y)\) onto \(\operatorname{Ext}^{k}_{\mathcal{A}}(X,Y)\) for each \(k>0\). Since \(\mathcal{A}\) is a length category, \[\dim_{\operatorname{End}_{\mathcal{A}}(Y)}\operatorname{Hom}_{\mathcal{A}}(X,Y )<\infty.\] The result follows by induction. Say that a \(\Bbbk\)-linear abelian category is _finite over \(\Bbbk\)_ if \(\mathcal{A}\) is a length category with enough projectives, finitely many simple objects, and finite dimensional Homspaces. It is well-known that \(\mathcal{A}\) is finite over \(\Bbbk\) if and only if there is a finite-dimensional \(\Bbbk\)-algebra \(A\) in which \(\mathcal{A}\) is equivalent to the category, \(A\)-mod, of modules that are finite dimensional as \(\Bbbk\)-vector spaces. Indeed, if \(\{P_{\lambda}\}_{\lambda\in\Lambda}\) are the projective indecomposables in \(\mathcal{A}\) (up to isomorphism), then \(A=\operatorname{End}_{\mathcal{A}}(\bigoplus_{\lambda\in\Lambda}P_{\lambda})^{op}\) and \(\operatorname{Hom}_{\mathcal{A}}(\bigoplus_{\lambda\in\Lambda}P_{\lambda},-): \mathcal{A}\simeq A\text{-mod}\). Note that there is a contravariant equivalence \(\operatorname{Hom}_{\Bbbk}(-,\Bbbk):A\text{-mod}\to A^{op}\text{-mod}\). In particular, any finite abelian category has enough injectives. **Proposition 4.4**.: _Let \(\Bbbk\) be a field and let \(\mathcal{A}\) be a \(\Bbbk\)-linear abelian category with a stratification by a finite poset in which every strata category is a finite abelian category. Then \(\mathcal{A}\) is \(\operatorname{Ext}^{1}\)-finite._ Proof.: By the assumptions on the strata categories, \(\mathcal{A}\) has finite dimensional Homspaces. Suppose \(\mathcal{A}\) has a recollement with objects and morphisms as in (3.1). Suppose \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have enough projectives, and \(\mathcal{A}_{U}\) has enough injectives. By Proposition 4.3, both \(\mathcal{A}_{Z}\) and \(\mathcal{A}_{U}\) are \(\operatorname{Ext}^{1}\)-finite. We show that \(\mathcal{A}\) is \(\operatorname{Ext}^{1}\)-finite. It suffices to show that \(\dim_{\Bbbk}\operatorname{Ext}^{1}_{\mathcal{A}}(X,Y)<\infty\) for all simple objects \(X,Y\). Since \(\mathcal{A}^{Z}\) is a Serre subcategory of \(\mathcal{A}\), this is true whenever \(X\) and \(Y\) are both in \(\mathcal{A}^{Z}\). Let \(L\in\mathcal{A}_{U}\) be simple and let \(j_{!*}L\) have projective and injective presentations: \[0\to K\to j_{!}P\to j_{!*}L\to 0\] \[0\to j_{!*}L\to j_{*}I\to K^{\prime}\to 0\] The projective presentation implies that \(\operatorname{Hom}_{\mathcal{A}}(K,Y)\) surjects onto \(\operatorname{Ext}^{1}_{\mathcal{A}}(j_{!*}L,Y)\). The injective presentation implies that \(\operatorname{Hom}_{\mathcal{A}}(Y,K^{\prime})\) surjects onto \(\operatorname{Ext}^{1}_{\mathcal{A}}(Y,j_{!*}L)\). It follows that \(\mathcal{A}\) is \(\operatorname{Ext}^{1}\)-finite. The result follows by an induction argument. ### Main results This section includes our original results about recollements and projective covers. **Proposition 4.5**.: _Let \(\mathcal{A}\) be an abelian category with a recollement of abelian categories as in (3.1). Then_ 1. _If_ \(X\to Y\) _is an essential surjection in_ \(\mathcal{A}\) _and_ \(i^{*}Y=0\) _then_ \(i^{*}X=0\) _and_ \(j^{*}X\to j^{*}Y\) _is an essential surjection._ 2. _If_ \(P\in\mathcal{A}\) _is projective and_ \(i^{*}P=0\) _then_ \(j^{*}P\in\mathcal{A}_{U}\) _is projective. In particular if_ \(P\to X\) _is a projective cover in_ \(\mathcal{A}\) _and_ \(i^{*}X=0\) _then_ \(j^{*}P\to j^{*}X\) _is a projective cover in_ \(\mathcal{A}_{U}\)_._ _In particular, if \(\mathcal{A}\) has enough projectives then so does \(\mathcal{A}_{U}\)._ Proof.: Let \(\phi:X\to Y\) be an essential surjection in \(\mathcal{A}\) and suppose \(i^{*}Y=0\). To show that \(i^{*}X=0\) it suffices to show that the canonical map \(\epsilon_{X}:j_{!}j^{*}X\to X\) is surjective. This follows from the following commutative diagram since \(\phi\) is essential. Let \(\alpha:X^{\prime}\to j^{*}X\) be a morphism in \(\mathcal{A}_{U}\), in which \(j^{*}(\phi)\circ\alpha:X^{\prime}\to j^{*}Y\) is surjective. Then \(\epsilon_{Y}\circ j_{!}j^{*}(\phi)\circ j_{!}\alpha:j_{!}X^{\prime}\to Y\) is surjective and so (since \(\phi\) is essential) \(\epsilon_{X}\circ j_{!}\alpha:j_{!}X^{\prime}\to X\) is surjective. Hence \(j^{*}(\epsilon_{X}\circ j_{!}\alpha)\simeq\alpha:X^{\prime}\to j^{*}X\) is surjective. This proves (i). If \(P\in\mathcal{A}\) is projective and \(i^{*}P=0\) then the functor \[\operatorname{Hom}_{\mathcal{A}_{U}}(j^{*}P,-)\simeq\operatorname{Hom}_{ \mathcal{A}}(j_{!}j^{*}P,j_{!}(-))\simeq\operatorname{Hom}_{\mathcal{A}}(P,j_ {!}(-)):\mathcal{A}_{U}\to\mathbb{Z}\text{-mod}\] is exact. Here the last isomorphism follows from the sequence (2.6). It follows that \(j^{*}P\) is projective. Statement (ii) follows. **Proposition 4.6**.: _Suppose \(\mathcal{A}\) is a Krull-Schmidt category with a recollement of abelian categories as in (3.1). If \(P\to L\) is a projective cover in \(\mathcal{A}\) of a simple object \(L\in\mathcal{A}^{Z}\), then \(i^{*}P\to i^{*}L\) is a projective cover in \(\mathcal{A}_{Z}\). In particular, if \(\mathcal{A}\) has enough projectives then so does \(\mathcal{A}_{Z}\)._ Proof.: Since \(i^{*}\) is the left adjoint of an exact functor it preserves projective objects. For any simple object \(L^{\prime}\in\mathcal{A}^{Z}\), \(\operatorname{Hom}_{\mathcal{A}_{Z}}(i^{*}P,i^{*}L^{\prime})=\operatorname{ Hom}_{\mathcal{A}}(P,i_{*}i^{*}L^{\prime})=\operatorname{Hom}_{\mathcal{A}}(P,L^{ \prime})\). The result follows. **Proposition 4.7**.: _Let \(\mathcal{A}\) be an abelian category with a recollement of abelian categories as in (3.1). Let \(X\) and \(Y\) be objects in \(\mathcal{A}_{U}\). If \(X\to Y\) is an essential surjection in \(\mathcal{A}_{U}\) then the composition \(j_{!}X\to j_{!*}X\to j_{!*}Y\) is an essential surjection in \(\mathcal{A}\). In particular:_ 1. _The canonical surjection_ \(j_{!}X\to j_{!*}X\) _is essential._ 2. _If_ \(P\to X\) _is a projective cover of_ \(X\) _in_ \(\mathcal{A}_{U}\) _then_ \(j_{!}P\to j_{!}X\to j_{!*}X\) _is a projective cover of_ \(j_{!*}X\) _in_ \(\mathcal{A}\)_._ Proof.: Let \(\phi:X\to Y\) be an essential surjection in \(\mathcal{A}_{U}\). The map \(\phi^{\prime}:j_{!}X\to j_{!*}X\to j_{!*}Y\) is surjective by Proposition 3.5(i). Let \(\alpha:X^{\prime}\to X\) be a morphism in which \(\phi^{\prime}\circ\alpha\) is surjective. Now, \(j^{*}(\phi^{\prime})=\phi:X\to Y\) and since \(j^{*}\) is exact, \(j^{*}(\phi^{\prime}\circ\alpha)=\phi\circ j^{*}(\alpha):j^{*}X^{\prime}\to Y\) is surjective. Since \(\phi\) is essential it follows that \(j^{*}(\alpha):j^{*}X^{\prime}\to X\) is surjective in \(\mathcal{A}_{U}\) and so \(j_{!}j^{*}(\alpha):j_{!}j^{*}X^{\prime}\to j_{!}X\) is surjective in \(\mathcal{A}\). The surjectivity of \(\alpha\) follows from the commutative triangle in which the downward arrow is the adjunction counit. The following result holds by an almost identical argument. **Proposition 4.8**.: _The intermediate-extension functor preserves essential surjections and essential extensions._ The following is the main result of this section. **Theorem 4.9**.: _Let \(\mathcal{A}\) be an abelian length category with finitely many simple objects, and a recollement of abelian categories as in (3.1). If \(\mathcal{A}\) is \(\operatorname{Ext}^{1}\)-finite then \(\mathcal{A}\) has enough projectives if and only if both \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have enough projectives. Dually if \(\mathcal{A}^{op}\) is \(\operatorname{Ext}^{1}\)-finite then \(\mathcal{A}\) has enough injectives if and only if both \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have enough injectives._ Before giving the proof of this theorem we will explain one important ingredient: the _universal extension_. Let \(A,B\) be objects in an abelian category \(\mathcal{A}\) in which \(\operatorname{End}_{\mathcal{A}}(B)\) is a division ring and \(d:=\dim_{\operatorname{End}_{\mathcal{A}}(B)}\operatorname{Ext}^{1}_{ \mathcal{A}}(A,B)<\infty\). We form the _universal extension_\(\mathcal{E}\in\operatorname{Ext}^{1}_{\mathcal{A}}(A,B^{\oplus d})\) by the following process. First let \(E_{1},\ldots,E_{d}\in\operatorname{Ext}^{1}_{\mathcal{A}}(A,B)\) be an \(\operatorname{End}_{\mathcal{A}}(B)\)-basis. The diagonal map \(\Delta:A\to A^{\oplus d}\) induces a map \(\operatorname{Ext}^{1}_{\mathcal{A}}(A^{\oplus d},B^{\oplus d})\to \operatorname{Ext}^{1}_{\mathcal{A}}(A,B^{\oplus d})\). Let \(\mathcal{E}\) be the image of \(E_{1}\oplus\cdots\oplus E_{d}\) under this map. Note that the \(\operatorname{End}_{\mathcal{A}}(B)\)-equivariant map \(\operatorname{Hom}_{\mathcal{A}}(B^{\oplus d},B)\to\operatorname{Ext}^{1}_{ \mathcal{A}}(A,B)\) induced by the short exact sequence \[0\to B^{\oplus d}\to\mathcal{E}\to A\to 0\] is surjective (this is easy to check on the basis of \(\operatorname{Ext}^{1}_{\mathcal{A}}(A,B)\)). When \(B_{1},\ldots,B_{n}\) are objects in \(\mathcal{A}\) in which each ring \(\operatorname{End}_{\mathcal{A}}(B_{i})\) is a division ring and \(d_{i}:=\dim_{\operatorname{End}_{\mathcal{A}}(B_{i})}\operatorname{Ext}^{1}_{ \mathcal{A}}(A,B_{i})<\infty\), we also talk about a _universal extension_\(\mathcal{E}\in\operatorname{Ext}^{1}_{\mathcal{A}}(A,\bigoplus_{i}B^{\oplus d _{i}}_{i})\) constructed in the following way. Let \(\mathcal{E}_{i}\in\operatorname{Ext}^{1}_{\mathcal{A}}(A,B^{\oplus d_{i}}_{i})\) be a universal extension (as defined in the previous paragraph) and define \(\mathcal{E}\) to be the image of \(\mathcal{E}_{1}\oplus\cdots\oplus\mathcal{E}_{n}\) under the map \(\operatorname{Ext}^{1}_{\mathcal{A}}(\bigoplus_{i}A,\bigoplus_{i}B^{\oplus d _{i}})\to\operatorname{Ext}^{1}_{\mathcal{A}}(A,\bigoplus_{i}B^{\oplus d_{i}})\) induced by the diagonal map \(\Delta:A\to A^{\oplus d}\). Then \(\mathcal{E}\) has the property that the short exact sequence \[0\to\bigoplus_{i}B^{\oplus d_{i}}\to\mathcal{E}\to A\to 0\] induces a surjection \(\operatorname{Hom}_{\mathcal{A}}(\bigoplus_{i}B^{\oplus d_{i}}_{i},B_{j})\to \operatorname{Ext}^{1}_{\mathcal{A}}(A,B_{j})\) for each \(j=1,\ldots,n\). Dually, if \(\operatorname{End}_{\mathcal{A}}(A)^{op}\) is a division ring and \(\dim_{\operatorname{End}_{\mathcal{A}}(A)^{op}}\operatorname{Ext}^{1}(A,B)<\infty\) then one can form a universal extension \(\mathcal{E}^{\prime}\in\operatorname{Ext}^{1}_{\mathcal{A}}(A^{\otimes d},B)\) using the codiagonal map \(\delta:B^{\oplus d}\to B\) instead of the diagonal map. Proof of Theorem 4.9.: Suppose that \(\mathcal{A}_{U}\) and \(\mathcal{A}_{Z}\) have enough projectives. Suppose \(\mathcal{A}^{Z}\) has simple objects \(L_{1},\ldots,L_{m}\) with projective covers \(\bar{P}_{1},\ldots,\bar{P}_{m}\) in \(\mathcal{A}^{Z}\). Suppose \(\mathcal{A}^{U}\) has simple objects \(L_{m+1},\ldots,L_{m+n}\). By Proposition 4.7 every simple object in \(\mathcal{A}^{U}\) has a projective cover in \(\mathcal{A}\). It suffices to construct a projective cover in \(\mathcal{A}\) of each simple object in \(\mathcal{A}^{Z}\). This amounts to finding, for each \(1\leq t\leq m\), a projective object, \(P_{t}\), whose unique simple quotient is \(L_{t}\). Fix \(1\leq t\leq m\). _Step 1. Define \(P_{t}\)._ For simple object \(L_{m+k}\in\mathcal{A}^{U}\), let \(P_{m+k}\) denote its projective cover in \(\mathcal{A}\). Define \(Q\) to be a maximal length quotient of \[P:=\bigoplus_{k=1}^{n}P_{m+k}^{\oplus\dim_{\operatorname{End}_{\mathcal{A}}(L_{m +k})}\operatorname{Ext}^{1}_{\mathcal{A}}(\bar{P}_{t},L_{m+k})}\] in which there is an extension \[0\to Q\to\mathcal{E}\to\bar{P}_{t}\to 0 \tag{4.2}\] inducing an isomorphism \(\operatorname{Hom}_{\mathcal{A}}(Q,L)\simeq\operatorname{Ext}^{1}_{\mathcal{A}}( \bar{P}_{t},L)\) for each \(L\in\mathcal{A}^{U}\). That is \(0=\operatorname{Hom}_{\mathcal{A}}(\bar{P}_{t},L)\simeq\operatorname{Hom}_{ \mathcal{A}}(\mathcal{E},L)\) and \(\operatorname{Ext}^{1}_{\mathcal{A}}(\mathcal{E},L)\) injects into \(\operatorname{Ext}^{1}_{\mathcal{A}}(Q,L)\). Let \(P_{t}\) be any choice of such \(\mathcal{E}\). _Step 2. \(P_{t}\) is well-defined._ To show that the maximal quotient \(Q\) exists, we just need to find one quotient of \(P\) with the required property. Then since \(\mathcal{A}\) is a length category there exists a maximal length quotient with the required property. Since \(\mathcal{A}\) has finite \(\operatorname{Ext}^{1}\)-spaces, we can let \[R=\bigoplus_{k=1}^{n}L_{m+k}^{\oplus\operatorname{dim}_{\operatorname{End}_{ \mathcal{A}}(L_{m+k})}\operatorname{Ext}^{1}_{\mathcal{A}}(\bar{P}_{t},L_{m+k})}\] and form the universal extension \[0\to R\to\mathcal{E}\to\bar{P}_{t}\to 0\] Since this is a universal extension it induces a surjection \(\operatorname{Hom}_{\mathcal{A}}(R,L_{m+k})\to\operatorname{Ext}^{1}_{ \mathcal{A}}(\bar{P}_{t},L_{m+k})\) for each \(L_{m+k}\in\mathcal{A}^{U}\). This map is an isomorphism since \[\operatorname{dim}_{\operatorname{End}_{\mathcal{A}}(L_{m+k})}\operatorname{ Hom}_{\mathcal{A}}(R,L_{m+k})=\operatorname{dim}_{\operatorname{End}_{ \mathcal{A}}(L_{m+k})}\operatorname{Ext}^{1}_{\mathcal{A}}(\bar{P}_{t},L_{m+k}).\] _Step 3. \(P_{t}\) has unique simple quotient \(L_{t}\)._ By definition of \(P_{t}\) and by the \((i^{*},i_{*})\)-adjunction, for any simple module \(L\in\mathcal{A}\): \[\operatorname{Hom}_{\mathcal{A}}(P_{t},L)\simeq\operatorname{Hom}_{\mathcal{ A}}(\overline{P}_{t},L)\simeq\operatorname{Hom}_{\mathcal{A}}(\overline{P}_{t},i_{*}i^{* }L)\] and so the only simple quotient of \(P_{t}\) is \(L_{t}\) with multiplicity one. _Step 4. \(P_{t}\) is projective._ We show that \(\operatorname{Ext}^{1}_{\mathcal{A}}(P_{t},L)=0\) for each simple \(L\in\mathcal{A}\). For any simple \(L\in\mathcal{A}\) there is an exact sequence \[0\to\operatorname{Ext}^{1}_{\mathcal{A}}(P_{t},L)\to\operatorname{Ext}^{1}_{ \mathcal{A}}(Q,L)\to\operatorname{Ext}^{2}_{\mathcal{A}}(\bar{P}_{t},L) \tag{4.3}\] Indeed if \(L\in\mathcal{A}^{U}\) then this holds because \(\operatorname{Hom}_{\mathcal{A}}(Q,L)\simeq\operatorname{Ext}^{1}_{\mathcal{A}} (\bar{P}_{t},L)\). If \(L\in\mathcal{A}^{Z}\) then (4.3) holds since \(\operatorname{Ext}^{1}_{\mathcal{A}}(\bar{P}_{t},L)=0\). To show that \(P_{t}\) is projective it suffices to show that the third map in (4.3) is injective for any \(L\in\mathcal{A}\). Suppose for contradiction that there is a nontrivial extension \[0\to L\to Q^{\prime}\to Q\to 0 \tag{4.4}\] in the kernel of this map. Then there is an object \(\mathcal{E}\in\mathcal{A}\) fitting into the following diagram (4.5) [MISSING_PAGE_POST] in which each row and column is exact. For each \(L^{\prime}\in\mathcal{A}^{U}\) the sequence (4.4) induces an exact sequence \[0\to\operatorname{Hom}_{\mathcal{A}}(Q,L^{\prime})\to\operatorname{Hom}_{ \mathcal{A}}(Q^{\prime},L^{\prime})\to\operatorname{Hom}_{\mathcal{A}}(L,L^{ \prime})\] Of course, \(\operatorname{Hom}_{\mathcal{A}}(L,L^{\prime})=0\) if \(L\neq L^{\prime}\). If \(L=L^{\prime}\) the third map must be zero. Indeed if \(f:L\to L\) factors through the inclusion \(\iota:L\to Q^{\prime}\) via a map \(g:Q^{\prime}\to L\), then \(f^{-1}\circ g:Q^{\prime}\to L\) is a retraction of \(\iota\). This contradicts the assumption that (4.4) does not split. Hence, for any \(L^{\prime}\in\mathcal{A}^{U}\), there is an isomorphism \[\operatorname{Hom}_{\mathcal{A}}(Q^{\prime},L^{\prime})\simeq\operatorname{ Hom}_{\mathcal{A}}(Q,L^{\prime})\simeq\operatorname{Ext}_{\mathcal{A}}^{1}( \bar{P}_{t},L^{\prime}). \tag{4.6}\] Since \(P\) is projective the quotient \(P\to Q\) fits into the diagram Now \(\varphi\) cannot be surjective, as, by (4.6), this would contradict the maximality of \(Q\). Thus the image of \(\varphi\) is isomorphic to \(Q\) and so the sequence (4.4) splits. This is a contradiction. Hence \(P_{t}\) is projective. **Corollary 4.10**.: _Let \(\mathcal{A}\) be an abelian category with a stratification in which every strata category is a length category, and has finitely many simple objects._ _If \(\mathcal{A}\) is \(\operatorname{Ext}^{1}\)-finite (respectively \(\mathcal{A}^{op}\) is \(\operatorname{Ext}^{1}\)-finite) then \(\mathcal{A}\) has enough projectives (respectively injectives) if and only if every strata category has enough projectives (respectively injectives)._ Proof.: By Proposition 3.6, every category \(\mathcal{A}_{\Lambda^{\prime}}\) (for lower \(\Lambda^{\prime}\subset\Lambda\)) satisfies the conditions of Theorem 4.9. So we can obtain a projective cover in \(\mathcal{A}\) of any simple object \(j_{!*}^{\lambda}L\) by repeatedly applying the construction in the proof of Theorem 4.9 to larger and larger Serre subcategories of \(\mathcal{A}\). The following result follows immediately from Proposition 4.4 and Corollary 4.10. **Corollary 4.11**.: _Let \(\mathcal{A}\) be a \(\Bbbk\)-linear abelian category with a stratification by a finite poset. Then \(\mathcal{A}\) is finite over \(\Bbbk\) if and only if the same is true of every strata category._ From this result we recover the following result of Cipriani-Woolf. **Corollary 4.12** (Corollary 5.2 of [10]).: _Let \(P_{\Lambda}(X,\Bbbk)\) be the category of perverse sheaves (with coefficients in a field \(\Bbbk\)) on a variety \(X\) that is constructible with respect to a stratification, \(X=\bigcup_{\lambda\in\Lambda}X_{\lambda}\), with finitely many strata. Then \(P_{\Lambda}(X,\Bbbk)\) is finite over \(\Bbbk\) if and only if each category \(\Bbbk[\pi_{1}(X_{\lambda})]\text{-}\mathrm{mod}_{fg}\) is finite over \(\Bbbk\)._ For example, if \(X\) has a finite stratification, \(X=\bigcup_{\lambda\in\Lambda}X_{\lambda}\), in which each \(X_{\lambda}\) has finite fundamental group, then the category \(P_{\Lambda}(X,\Bbbk)\) is finite over \(\Bbbk\). **Corollary 4.13**.: _Let \(G\) be an algebraic group and let \(X\) be a \(G\)-variety with finitely many orbits, each connected. Let \(\Bbbk\) be a field. The category, \(P_{G}(X,\Bbbk)\), of \(G\)-equivariant perverse sheaves is finite over \(\Bbbk\) if and only if for each \(G\)-orbit \(\mathcal{O}_{\lambda}\) and \(x\in\mathcal{O}_{\lambda}\), the category \(\Bbbk[G^{x}/(G^{x})^{\circ}]\text{-}\mathrm{mod}_{fg}\) is finite over \(\Bbbk\)._ ## 5. Standard and costandard objects In this section we focus on abelian length categories \(\mathcal{A}\) with finitely many simples, enough projectives and injectives, and admitting a stratification by a finite non-empty poset \(\Lambda\). Let \(B\) be a set indexing the simple objects of \(\mathcal{A}\) up to isomorphism. Write \(L(b)\) for the simple object corresponding to \(b\in B\), and write \(P(b)\) and \(I(b)\) for the projective cover and injective envelope of \(L(b)\). For each \(\lambda\in\Lambda\), write \(\mathcal{A}_{\leq\lambda}:=\mathcal{A}_{\{\mu\in\Lambda\ |\ \mu\leq\lambda\}}\) and let \(j_{!}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) be the composition Define \(j_{*}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) and \(j_{!*}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) likewise. Let \(j^{\lambda}:\mathcal{A}_{\leq\lambda}\to\mathcal{A}_{\lambda}\) denote the Serre quotient functor. Define the _stratification function_\(\rho:B\to\Lambda\) that assigns to each \(b\in B\) the \(\lambda\in\Lambda\) in which \(L(b)=j_{!*}^{\lambda}L_{\lambda}(b)\). Let \(P_{\lambda}(b)\) and \(I_{\lambda}(b)\) be the projective cover and injective envelope of the simple object \(L_{\lambda}(b)\) in \(\mathcal{A}_{\lambda}\). For \(b\in B\), define the _standard_ and _costandard_ objects \[\Delta(b):=j_{!}^{\lambda}P_{\lambda}(b),\qquad\nabla(b):=j_{*}^{\lambda}I_{ \lambda}(b),\] where \(\lambda=\rho(b)\). The following original result follows from the proofs of Theorem 4.9 and Corollary 4.10. **Porism 5.1**.: _Let \(\mathcal{A}\) be an abelian category with a stratification by a finite non-empty poset \(\Lambda\), in which every strata category is a length category with finitely many simple objects. Let \(\rho:B\to\Lambda\) be the stratification function for \(\mathcal{A}\)._ _If \(\mathcal{A}\) has enough projectives then for each \(\lambda\in\Lambda\) and \(b\in\rho^{-1}(\lambda)\), the projective indecomposable object \(P(b)\in\mathcal{A}\) fits into a short exact sequence_ \[0\to Q(b)\to P(b)\to\Delta(b)\to 0\] _in which \(Q(b)\) has a filtration by quotients of \(\Delta(b^{\prime})\) satisfying \(\rho(b^{\prime})>\rho(b)\)._ _If \(\mathcal{A}\) has enough injectives then for each \(\lambda\in\Lambda\) and \(b\in\rho^{-1}(\lambda)\), the injective indecomposable object \(I(b)\in\mathcal{A}\) fits into a short exact sequence_ \[0\to\nabla(b)\to I(b)\to Q^{\prime}(b)\to 0\] _in which \(Q^{\prime}(b)\) has a filtration by subobjects of \(\nabla(b^{\prime})\) satisfying \(\rho(b^{\prime})>\rho(b)\)._ Proof.: We just prove the first statement by induction on \(|\Lambda|\). The base case \(|\Lambda|=1\) is trivial. Consider the projective cover, \(P(b)\), of simple object \(L(b)\) in \(\mathcal{A}\). If \(\rho(b)\) is maximal then \(P(b)\simeq\Delta(b)\) and the result holds. Suppose \(\rho(b)\) is not maximal, and let \(\mu\in\Lambda\) be a maximal element. Consider the recollement Let \(P_{<\mu}(b)\) and \(\Delta_{<\mu}(b)\) be the projective indecomposable and standard object in \(\mathcal{A}_{<\mu}\) corresponding to the simple object \(i^{*}L(b)\in\mathcal{A}_{<\mu}\). By induction there is a short exact sequence \[0\to Q_{<\mu}(b)\to P_{<\mu}(b)\to\Delta_{<\mu}(b)\to 0\] in which \(Q_{<\mu}(b)\) has a filtration by quotients of standard objects \(\Delta_{<\mu}(b^{\prime})\) satisfying \(\rho(b^{\prime})>\rho(b)\). Since \(i_{*}\) is exact we get the following short exact sequence in \(\mathcal{A}\): \[0\to i_{*}Q_{<\mu}(b)\to i_{*}P_{<\mu}(b)\to\Delta(b)\to 0 \tag{5.1}\] and \(i_{*}Q_{<\mu}(b)\) has a filtration by quotients of standard objects \(\Delta(b^{\prime})\) satisfying \(\mu>\rho(b^{\prime})>\rho(b)\). By applying the construction in Step 1 of the proof of Theorem 4.9, \(P(b)\) fits into the short exact sequence in \(\mathcal{A}\): \[0\to Q_{\mu}(b)\to P(b)\to i_{*}P_{<\mu}(b)\to 0 \tag{5.2}\] and \(Q_{\mu}(b)\) is a quotient of a direct sum of standard objects of the form \(\Delta(b^{\prime})\) in which \(\rho(b^{\prime})=\mu\). Combining (5.1) and (5.2) gives the following diagram with exact rows and columns: The result follows. ## 6. Brundan and Stroppel's \(\varepsilon\)-stratified categories In this section we give necessary and sufficient conditions for a finite abelian category with a stratification by a finite poset to be \(\varepsilon\)-stratified in the sense of Brundan and Stroppel [1] (Theorem 6.4). This result specializes to a characterization of highest weight categories originally given by Krause [14, Theorem 3.4]. We recall this characterisation in Corollary 6.6. Let \(\mathcal{A}\) be an abelian length category with finitely many simples, enough projectives and injectives, and with a stratification by a finite non-empty poset \(\Lambda\). Let \(\rho:B\to\Lambda\) be the stratification function corresponding to this stratification. For \(b\in B\) and \(\lambda=\rho(b)\in\Lambda\), define _proper standard_ and _proper costandard_ objects \[\overline{\Delta}(b):=j_{!}^{\lambda}L_{\lambda}(b),\qquad\overline{\nabla}(b) :=j_{*}^{\lambda}L_{\lambda}(b).\] For a _sign function_\(\varepsilon:\Lambda\to\{+,-\}\), define the _\(\varepsilon\)-standard_ and _\(\varepsilon\)-costandard objects_ \[\Delta_{\varepsilon}(b):=\left\{\begin{array}{ll}\Delta(b)&\text{if } \varepsilon(\rho(b))=+\\ \overline{\Delta}(b)&\text{if }\varepsilon(\rho(b))=-\end{array}\right., \qquad\nabla_{\varepsilon}(b):=\left\{\begin{array}{ll}\overline{\nabla}(b )&\text{if }\varepsilon(\rho(b))=+\\ \nabla(b)&\text{if }\varepsilon(\rho(b))=-\end{array}\right..\] Note that since \(j_{!}^{\lambda}\) and \(j_{*}^{\lambda}\) are fully-faithful, these objects have local endomorphism rings and are hence irreducible. Note also that if \(\rho(b)>\rho(b^{\prime})\) then for any \(\varepsilon:\Lambda\to\{+,-\}\), \[\operatorname{Hom}_{\mathcal{A}}(\Delta_{\varepsilon}(b),\Delta_{\varepsilon}(b^{ \prime}))=0=\operatorname{Hom}_{\mathcal{A}}(\nabla_{\varepsilon}(b^{\prime}), \nabla_{\varepsilon}(b)). \tag{6.1}\] Indeed the only simple quotient of \(\Delta_{\varepsilon}(b)\) is \(L(b)\), and all simple subobjects, \(L(b^{\prime\prime})\), of \(\Delta_{\varepsilon}(b^{\prime})\) satisfy \(\rho(b^{\prime})\geq\rho(b^{\prime\prime})\). Likewise the only simple subobject of \(\nabla_{\varepsilon}(b)\) is \(L(b)\), and all simple quotients, \(L(b^{\prime\prime})\), of \(\nabla_{\varepsilon}(b^{\prime})\) satisfy \(\rho(b^{\prime})\geq\rho(b^{\prime\prime})\). The following definition is due to Brundan and Stroppel [10].4 Footnote 4: In Brundan and Stoppel’s definition, \(\varepsilon\)-stratified categories need not be finite (they satisfy slightly weaker finiteness conditions). **Definition 6.1** (\(\varepsilon\)-stratified category).: Let \(\mathcal{A}\) be a finite abelian category with a stratification by a poset \(\Lambda\) and stratification function \(\rho:B\to\Lambda\). Let \(\varepsilon:\Lambda\to\{+,-\}\) be a function. Say that \(\mathcal{A}\) is an _\(\varepsilon\)-stratified category_ if the following equivalent conditions are satisfied: 1. For every \(b\in B\), the projective indecomposable \(P(b)\) has a filtration by objects of the form \(\Delta_{\varepsilon}(b^{\prime})\), where \(\rho(b^{\prime})\geq\rho(b)\). 2. For every \(b\in B\), the injective indecomposable \(I(b)\) has a filtration by objects of the form \(\nabla_{\varepsilon}(b^{\prime})\), where \(\rho(b^{\prime})\geq\rho(b)\). The equivalence of statements (\(\varepsilon\)-S1) and (\(\varepsilon\)-S2) is shown in [1, Theorem 2.2]. A proof of this fact can also be found in [10, Theorem 3.5]. **Remark 6.2**.: A finite abelian category \(A\)-mod is \(\varepsilon\)-stratified if and only if \(A\) is a _stratified algebra_ in the sense of [1]. The following definition is original. **Definition 6.3** (\(\varepsilon\)-exact stratification).: For a function \(\varepsilon:\Lambda\to\{+,-\}\), say that a stratification of an abelian category \(\mathcal{A}\) by a finite non-empty poset \(\Lambda\) is _\(\varepsilon\)-exact_ if for all \(\lambda\in\Lambda\) the following hold: 1. If \(\varepsilon(\lambda)=+\) then the functor \(j_{*}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) is exact. 2. If \(\varepsilon(\lambda)=-\) then the functor \(j_{!}^{\lambda}:\mathcal{A}_{\lambda}\to\mathcal{A}\) is exact. For \(k\in\mathbb{N}\), say that a recollement of abelian categories (as in (3.1)) is _\(k\)-homological_ if for all \(n\leq k\) and \(X,Y\in\mathcal{A}_{Z}\), \[\operatorname{Ext}^{n}_{\mathcal{A}_{Z}}(X,Y)\simeq\operatorname{Ext}^{n}_{ \mathcal{A}}(i_{*}X,i_{*}Y).\] Say that a recollement of abelian categories is _homological_ if it is \(k\)-homological for all \(k\in\mathbb{N}\). A study of homological recollements is given in [10]. Say that a stratification of an abelian category is _\(k\)-homological_ (respectively _homological_) if each of the recollements in the data of the stratification is \(k\)-homological (respectively homological). The following theorem is the main result of this section. **Theorem 6.4**.: _Let \(\mathcal{A}\) be a finite abelian category. Then, for any finite non-empty poset \(\Lambda\) and function \(\varepsilon:\Lambda\to\{+,-\}\), the following statements are equivalent:_ 1. \(\mathcal{A}\) _has an_ \(\varepsilon\)_-exact homological stratification by_ \(\Lambda\)_._ 2. \(\mathcal{A}\) _has an_ \(\varepsilon\)_-exact 2-homological stratification by_ \(\Lambda\)_._ 3. \(\mathcal{A}\) _is_ \(\varepsilon\)_-stratified._ To prove this theorem we need the following lemma. **Lemma 6.5**.: _Let \(\mathcal{A}\) be an abelian length category with finitely many simple objects, and fitting into a 2-homological recollement of abelian categories as in (3.1)._ _If \(\mathcal{A}\) has enough projectives then for any projective object \(P\in\mathcal{A}\), there is a short exact sequence_ \[0\to j_{!}j^{*}P\to P\to i_{*}i^{*}P\to 0\] _If \(\mathcal{A}\) has enough injectives then for any injective object \(I\in\mathcal{A}\), there is a short exact sequence_ \[0\to i_{*}i^{!}I\to I\to j_{*}j^{*}I\to 0\] Proof.: Let \(P\in\mathcal{A}\) be a projective indecomposable object in \(\mathcal{A}\). If \(i^{*}P=0\) then \(j_{!}j^{*}P\simeq P\) and the result holds. Suppose that \(i^{*}P\neq 0\). Consider the exact sequence \[0\to K\to j_{!}j^{*}P\to P\to i_{*}i^{*}P\to 0\] where \(K\in\mathcal{A}^{Z}\). Since \(i^{*}P\) is projective in \(\mathcal{A}_{Z}\) and the recollement is 2-homological, \(\operatorname{Ext}_{\mathcal{A}}^{2}(i_{*}i^{*}P,K)=0\). In particular, there is an object \(\mathcal{E}\in\mathcal{A}\) and surjection \(q:\mathcal{E}\to P\) that fits into the following diagram in which all rows and columns are exact. (6.2) Since \(P\) is projective, there is a map \(\alpha:P\to\mathcal{E}\) making the following diagram commute. Since \(P\) is indecomposable, the endomorphism \(q\circ\alpha:P\to P\) is either nilpotent or an automorphism. Since \(i_{*}i^{*}P\) is nonzero, this map is not nilpotent, and hence is an automorphism of \(P\). In particular, the first two rows of Diagram (6.2) split. Hence \(K\) is a quotient of \(j_{!}j^{!}P\). Since \(j_{!}j^{!}P\) has no nonzero quotient in \(\mathcal{A}^{Z}\), \(K=0\). The first result follows. The second result holds by the dual argument. Proof of Theorem 6.4.: The implication \((1)\implies(2)\) is obvious. For the remainder of the proof, fix a maximal element \(\lambda\in\Lambda\) and recollement (6.3) \((2)\implies(3)\). Let \(P(b)\in\mathcal{A}\) be a projective indecomposable object. Suppose that every projective indecomposable, \(P(b^{\prime})\), in \(\mathcal{A}_{<\lambda}\) has a filtration by \(\varepsilon\)-standard objects, \(\Delta_{\varepsilon}(b^{\prime\prime})\), in which \(\rho(b^{\prime\prime})\geq\rho(b^{\prime})\). In particular \(i_{*}i^{*}P(b)\) has a filtration by \(\varepsilon\)-standard objects, \(\Delta_{\varepsilon}(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). By Lemma 6.5, if the recollement (6.3) is 2-homological then \(P(b)\) fits into a short exact sequence \[0\to j_{!}j^{*}P(b)\to P(b)\to i_{*}i^{*}P(b)\to 0\] If \(j_{*}\) is exact then \(j^{*}P(b)\) is projective (since \(j^{*}\) is the left adjoint of an exact functor) and so \(j_{!}j^{*}P(b)\) has a filtration by standard objects. If \(j_{!}\) is exact then \(j_{!}j^{*}P(b)\) has a filtration by proper standard objects. In particular, in either case \(\varepsilon(\lambda)=\pm\), \(P(b)\) has a filtration by \(\varepsilon\)-standard objects, \(\Delta_{\varepsilon}(b^{\prime})\), in which \(\rho(b^{\prime})\geq\rho(b)\). The result follows by induction on \(|\Lambda|\). \((3)\implies(1)\). Brundan-Stroppel show that if \(\mathcal{A}\) is \(\varepsilon\)-stratified then \(\mathcal{A}\) has an \(\varepsilon\)-exact stratification [1, Theorem 3.5] and \[\operatorname{Ext}^{n}_{\mathcal{A}}(\Delta_{\varepsilon}(b),\nabla_{ \varepsilon}(b^{\prime}))=0 \tag{6.4}\] for all \(b,b^{\prime}\in B\) and \(n\in\mathbb{N}\)[1, Theorem 3.14]. Let \(\mathcal{A}\) be \(\varepsilon\)-stratified abelian length category. Let \(P\) be a projective object in \(\mathcal{A}_{<\lambda}\), and \(I\) be an injective object in \(\mathcal{A}_{<\lambda}\). Then \(P\) and \(i_{*}P\) have a filtration by \(\varepsilon\)-standard objects, and \(I\) and \(i_{*}I\) have a filtration by \(\varepsilon\)-costandard objects. Since \(\mathcal{A}\) is a length category, then it follows from (6.4) that \(\operatorname{Ext}^{n}_{\mathcal{A}}(i_{*}P,i_{*}I)=0\). Since this is true for any projective object \(P\) and injective object \(I\) in \(\mathcal{A}_{<\lambda}\), the recollement (6.3) is homological (see e.g. [16, Theorem 3.9]). The result follows by induction on \(|\Lambda|\). ### Highest weight categories Let \(\Bbbk\) be a field. Say that a \(\Bbbk\)-linear abelian category \(\mathcal{A}\) is a _highest weight category_ with respect to a finite poset \(\Lambda\) if \(\mathcal{A}\) is finite over \(\Bbbk\), and for every \(\lambda\in\Lambda\) there is a projective indecomposable, \(P_{\lambda}\), that fits into a short exact sequence in \(\mathcal{A}\): \[0\to U_{\lambda}\to P_{\lambda}\to\Delta_{\lambda}\to 0\] in which the following hold: 1. \(\operatorname{End}_{\mathcal{A}}(\Delta_{\lambda})\) is a division ring for all \(\lambda\in\Lambda\). 2. \(\operatorname{Hom}_{\mathcal{A}}(\Delta_{\lambda},\Delta_{\mu})=0\) whenever \(\lambda>\mu\). 3. \(U_{\lambda}\) has a filtration by objects \(\Delta_{\mu}\) in which \(\lambda<\mu\). 4. \(\bigoplus_{\lambda\in\Lambda}P_{\lambda}\) is a projective generator of \(\mathcal{A}\). The following characterisation of highest weight categories is shown by Krause [18, Theorem 3.4]. We give a new proof using Theorem 6.4. **Corollary 6.6**.: _Let \(\Bbbk\) be a field, and let \(\mathcal{A}\) be a \(\Bbbk\)-linear abelian category. The following statements are equivalent._ 1. \(\mathcal{A}\) _is a highest weight category._ 2. \(\mathcal{A}\) _has a homological stratification with respect to_ \(\Lambda\) _in which every strata category is equivalent to_ \(\operatorname{mod}\)_-_\(\Gamma_{\lambda}\) _for some finite dimensional division algebra_ \(\Gamma_{\lambda}\)_._ 3. \(\mathcal{A}\) _has a 2-homological stratification with respect to_ \(\Lambda\) _in which every strata category is equivalent to_ \(\operatorname{mod}\)_-_\(\Gamma_{\lambda}\) _for some finite dimensional division algebra_ \(\Gamma_{\lambda}\) Proof.: \((1)\implies(2)\). If \(\mathcal{A}\) is a highest weight category with respect to \(\Lambda\), then a homological stratification of \(\mathcal{A}\) by \(\Lambda\) is constructed as follows: For a lower subposet \(\Lambda^{\prime}\subset\Lambda\) define \(\mathcal{A}_{\Lambda^{\prime}}\) to be the Serre subcategory of \(\mathcal{A}\) generated by the standard objects \(\Delta_{\lambda}\) in which \(\lambda\in\Lambda^{\prime}\). Then for any maximal \(\mu\in\Lambda^{\prime}\) there is a homological recollement of abelian categories in which \(j^{*}=\operatorname{Hom}_{\mathcal{A}_{\Lambda^{\prime}}}(\Delta_{\mu},-)\) (see the proof of [16, Theorem 3.4]). \((2)\implies(3).\) This is obvious. \((3)\implies(1).\) Suppose \(\mathcal{A}\) has a \(2\)-homological stratification with strata categories of the form \(\mathcal{A}_{\lambda}=\operatorname{mod-}\Gamma_{\lambda}\) for a finite dimensional division ring \(\Gamma_{\lambda}\). Then \(\mathcal{A}\) is finite (by Corollary 4.11). Let \(L_{\lambda}\) denote the unique simple object in \(\mathcal{A}_{\lambda}\). Define \(\Delta_{\lambda}:=j_{!}^{\lambda}L_{\lambda}\) and let \(P_{\lambda}\) be the projective cover of \(j_{!*}^{\lambda}L_{\lambda}\) in \(\mathcal{A}\). Statement (HW1) holds since \(j_{!}\) is fully-faithful, Statement (HW2) is exactly equation (6.1), Statement (HW3) follows from Theorem 6.4 (since each \(j_{!}^{\lambda}\) is exact), and Statement (HW4) is obvious.

2308.07395

Text Injection for Capitalization and Turn-Taking Prediction in Speech Models

Text injection for automatic speech recognition (ASR), wherein unpaired text-only data is used to supplement paired audio-text data, has shown promising improvements for word error rate. This study examines the use of text injection for auxiliary tasks, which are the non-ASR tasks often performed by an E2E model. In this work, we use joint end-to-end and internal language model training (JEIT) as our text injection algorithm to train an ASR model which performs two auxiliary tasks. The first is capitalization, which is a de-normalization task. The second is turn-taking prediction, which attempts to identify whether a user has completed their conversation turn in a digital assistant interaction. We show results demonstrating that our text injection method boosts capitalization performance for long-tail data, and improves turn-taking detection recall.

Shaan Bijwadia, Shuo-yiin Chang, Weiran Wang, Zhong Meng, Hao Zhang, Tara N. Sainath

2023-08-14T18:28:04Z

http://arxiv.org/abs/2308.07395v1

# Text Injection for Capitalization and Turn-Taking Prediction in Speech Models ###### Abstract Text injection for automatic speech recognition (ASR), wherein unpaired text-only data is used to supplement paired audio-text data, has shown promising improvements for word error rate. This study examines the use of text injection for auxiliary tasks, which are the non-ASR tasks often performed by an E2E model. In this work, we use joint end-to-end and internal language model training (JEIT) as our text injection algorithm to train an ASR model which performs two auxiliary tasks. The first is capitalization, which is a de-normalization task. The second is turn-taking prediction, which attempts to identify whether a user has completed their conversation turn in a digital assistant interaction. We show results demonstrating that our text injection method boosts capitalization performance for long-tail data, and improves turn-taking detection recall. Shaan Bijwadia\({}^{1}\), Shuo-yjin Chang\({}^{1}\), Weiran Wang\({}^{1}\), Zhong Meng\({}^{1}\), Hao Zhang\({}^{1}\), Tara N. Sainath\({}^{1}\)\({}^{1}\)Google, USA {shaanb, shuoyjin, weiranwang, haozhang, zhongmeng, tsainath}@google.com **Index Terms**: speech recognition, text injection, auxiliary tasks ## 1 Introduction Automatic speech recognition (ASR) has long been an integral part of important technologies, including voice dictation, digital assistants, and video captioning [1]. While ASR systems are typically evaluated based on word error rate (WER), this is not the only metric of concern in production applications; several "auxiliary tasks" must be integrated with the ASR task in a full system. These tasks may include: capitalization and punctuation, which improves readability; voice activity detection (VAD) and end-of-query (EOQ) detection, which are important for implementing low-latency systems; and natural conversation understanding, which involves predicting the cadence and turn-taking aspects of an ongoing conversation. In this study, we focus on improving the quality of such auxiliary tasks in an end-to-end (E2E) ASR setting via text injection. We build on two recent capabilities for speech models. First is the E2E integration of auxiliary tasks with the ASR task into a single model. In the past, auxiliary tasks were usually performed by separate models downstream of ASR [2, 3, 4, 5]. Recent work has successfully explored integrating auxiliary tasks, such as endpointing [6, 7, 8], capitalization [9], natural conversation understanding [10], and speaker diarization [11] into the same model as ASR prediction. E2E integration of ASR and auxiliary tasks has a key drawback, however. When folded into an E2E ASR model, pure text-to-text tasks (such as capitalization) can no longer be trained on plentiful text-only data (i.e., text data with no associated audio); instead, their training examples will be limited to the transcripts available in paired audio-text labeled data. This puts E2E methods at a disadvantage, since text-only data is generally more plentiful and easier to obtain than labeled audio data, and can be used to more easily expose the model to rare words and other long-tail phenomena which may be difficult to collect in labeled audio form [12]. The second capability enabling the current study is the use of "text injection" as a means of improving ASR quality [13]. An ASR model's internal language model (ILM) is the notional part of the network that predicts the next token given the previous token history, independent of audio input. Though it is usually infeasible to exactly separate the influence of audio input from previous token predictions, several methods have been developed to estimate ILM scores [14, 15]. Text-only data can then be used to refine the ILM capabilities of the ASR network [16, 17]. In this work, we propose a method to utilize text injection techniques for improving auxiliary task performance in an E2E ASR model. Doing so allows auxiliary tasks to access the multi-task learning benefits of co-training with ASR while still including rich text-only data in their training corpora. We focus our study on two tasks: capitalization and conversational turn-taking prediction. The former is a strongly text-based task, since capitalization is merely a form of de-normalization from spoken to written domain, and capitalized words are not pronounced differently. The latter task clearly involves combining linguistic and acoustic understanding -- the prosody of the input speech as well as the semantics of the current recognition are both informative for predicting whether a pause is only momentary or if the user has finished speaking. We integrate these tasks into a production-ready model, streaming E2E RNN-T ASR model [18, 19]. We show results demonstrating that text injection can meaningfully improve auxiliary task performance, particularly in long-tail settings. ## 2 Related Work While auxiliary tasks are usually performed by separate models from ASR [20, 21], E2E approaches to auxiliary task modeling have been recently popular for production-grade systems. Joint training of ASR with endpointing [7], capitalization [9, 22], intended query detection [23, 24], sentence segmentation [25], and more, have been explored. Our work builds most closely on Wang et al.[9], who co-train ASR, capitalization, and turn-taking prediction by building multiple parallel label sequences. To our knowledge, this is the first attempt to refine auxiliary tasks in an E2E ASR model using text-only data. There has long been interest in utilizing unpaired text data for the ASR task. Several approaches to LM fusion, the use of an external LM to improve ASR recognition quality, have been proposed [26]. These methods have the drawback of increasing total parameter count (due to the size of the LM model), and computation cost during inference. Text injection [13] solves these issues by using LM-style unpaired text data to train the ASR model itself. Some methods focus on fine-tuning an existing ASR model trained on audio-text data; ILM adaptation of the ASR decoder has been shown to work well [27, 28, 29]. The text injection method we employ here is joint end-to-end and ILM training (JEIT), which was introduced by Meng et al [30]. We choose JEIT as our method due to its lightweight nature; its primary focus on refining the ASR decoder makes comparison to standard methods straightforward, since the behavior of the audio encoder is preserved. Other methods inject text data directly into the encoder, with fixed and learned duration models to align text and audio sequences [16, 17]. All of the above works focus on improving ASR quality, both for standard and long-tail data; to the best of our knowledge, adapting these techniques for auxiliary tasks is a novel contribution to the literature. ## 3 Auxiliary Tasks ### Capitalization Capitalization is the process of restoring the correct case (uppercase or lowercase) of noisy text. Notably, capitalization is specific to the written domain, and has no marker in spoken speech. This task is important for maintaining readability in ASR output, especially for long-form captioning cases. ### Conversational turn-taking Turn-taking is an active area of research for E2E speech modeling [10, 31]. While humans typically adjust their speech when interacting with voice assistants [31], natural human speech patterns during conversation are often filled with natural disfluencies. For digital assistant products, it is desirable that voice assistants have the ability to predict when the speaker is expecting a response, versus when they merely pause with the intention to resume speaking. We model this phenomenon similar to Chang et al. [10], who classify pauses in speech as being within a complete thought, or after having a finished complete thought. That is, when a user stops speaking, the model should predict whether they will continue speaking after a brief pause or whether a system response is expected. Because the active region of interest is pauses in the audio, we refer to this task in this paper as "pause prediction." ## 4 Model ### Multi-output HAT decoder HAT is a decoder structure for RNN-T in which the (blank) probability is computed separately from next token prediction, facilitating more accurate ILM estimation [14]. Wang et al. [9] propose a variant of HAT decoder which introduces multiple joint networks, one for each task (in our case, these are ASR, capitalization, and pause prediction). All of the parallel joint networks are conditioned on features from both the prediction network and audio encoders. The model is trained using an RNN-T objective [18], where at each timestep the model may choose to emit a wordpiece token, or to insert a special token (blank) which indicates non-emission. Formally, let \(X\) be the input utterance and \(Y\) be the label sequence. The ASR output space \(\mathcal{Y}_{\text{ASR}}\) consists of \(\{y^{0}=(\text{blank}),y^{1},y^{2},...\}\). Let \(T=|X|\) be the number of input audio frames and \(U=|Y|\) be the length of the transcript. The acoustic encoder produces \(f(X)=[f_{0},...,f_{T-1}]\), \(f_{t}\in\mathcal{R}^{D_{a}}\), and the prediction network produces \(g(X)=[g_{0},...,g_{U-1}]\), \(g_{u}\in\mathcal{R}^{D_{p}}\). As in the original HAT implementation, the joint network fuses \(f_{t}\) and \(g_{u}\) with a "project and sum" operation to produce a hidden representation \(h_{t,u}\), which is then passed through a non-linear activation and a final linear layer to produce \(s_{t,u}\): \[h_{t,u}=P\cdot f_{t}+Q\cdot g_{u}+b_{h} \in\mathcal{R}^{D_{h}} \tag{1}\] \[s_{t,u}=A\cdot\text{tanh}(h_{t,u})+b_{s} \in\mathcal{R}^{V}. \tag{2}\] where \(P\), \(Q\), and \(A\) are learned weight matrices with dimensions determined by \(D_{a}\), \(D_{p}\), \(D_{h}\), and \(V\) is the size of the vocabulary. As this is a HAT model, the 0-th logit of \(s_{t,u}\) is used individually to compute the probability of emission \(b_{t,u}\): \[b_{t,u}:=P_{t,u}(\langle\texttt{blank}\rangle|f_{0:t},g_{0:u})=\sigma(s_{t,u}[ 0]) \tag{3}\] where \(\sigma(x)=1/(1+\exp(-x))\) is the sigmoid activation. Probabilities over the ASR tokens are computed by feeding all remaining logits to a softmax function. The probability of each ASR token \(y_{v}\) in the vocabulary is: \[\hat{y}_{v;t,u} =P_{t,u}(\hat{y}_{v}|f_{0:t},g_{0:u})\] \[=\text{softmax}(s_{t,u}[1:])[v-1] \tag{4}\] Thus the predicted probability distribution over all output tokens is the emission probability, followed by the probabilities of each token given emission: \[\hat{y}_{t,u}=[b_{t,u}, (1-b_{t,u})\cdot\hat{y}_{0:t,u},\] \[..., (1-b_{t,u})\cdot\hat{y}_{V-1:t,u}] \tag{5}\] Thus far we have referred to the mechanism above in terms of ASR prediction. Capitalization and pause predictions are made in the exact same way, where each task independently computes Eqs. (1) and (2) based on the shared representations \(f_{t}\) and \(g_{u}\) (note that each auxiliary task is exposed to the label history of the ASR output, not its own prediction history). Since capitalization tokens must be strictly aligned with ASR tokens, the capitalization posterior borrows the blank logit Figure 1: Model diagram for JEIT training. The blue arrows denote the data flow for paired audio-text data. The red arrows denote the path that unpaired text data takes through the network. Baseline experiments are trained using only the blue paths, while the proposed system is trained using both. from the ASR prediction. Thus, a capitalization token will only be emitted when an ASR token is emitted as well. Capitalization has output space \(\mathcal{Y}_{\text{Cap}}=\{\langle\text{cap}\rangle,\langle\text{non-cap}\rangle\}\) and its posterior is: \[\hat{y}_{t,u}^{\text{Cap}}=[b_{t,u}^{\text{ASR}},\quad(1-b_{t,u}^{ \text{ASR}})\cdot P_{t,u}(\langle\text{cap}\rangle),\] \[(1-b_{t,u}^{\text{ASR}})\cdot P_{t,u}(\langle\text{non-cap} \rangle)] \tag{6}\] At inference time, we estimate \(P(\langle\text{cap}\rangle)\) every time an ASR token is emitted and predict a capitalization if it is above a threshold (in this work, we use 0.5). Pause tokens do not need to be strictly aligned with the ASR transcript prediction, since they are likely to be predicted during non-speech periods in the audio during inference, so the turn-taking sequence has its own blank posterior. The pause prediction output space is \(\mathcal{Y}_{\text{Pause}}=\{\langle\text{blank}\rangle,\langle\text{non- pause}\rangle,\langle\text{eos}\rangle\}\) and its posterior is computed in the same way as Eq. (5). ## 5 Training ### Jeit Joint end-to-end model and ILM training (JEIT) was proposed by Meng et al. [30] as a way to train an RNN-T ASR model on paired audio-text data while simultaneously training the HAT decoder ILM on text-only data. For paired dataset \(\mathcal{D}_{\text{paired}}\), training is conducted in the usual way; the model is given the audio sequence as input and predicts \(P_{\text{EE}}(Y|X)\). This is converted to a loss \(\mathcal{L}_{\text{EE}}^{\text{ASR}}\) via the RNN-T objective [18]. The text-only dataset \(\mathcal{D}_{\text{paired}}\) contains transcripts with capitalization and pause annotations (see SS5.2). Similar to HAT ILM adaptation (ILMA) [27], we feed the transcript as the previous token history to the prediction network, and mock the encoder output with vectors full of zeros: \(\forall_{t\in 0:T}:f_{t}=\mathbf{0}\). Since the audio sequence does not exist, we simply ignore the blank posterior, and the predicted next token probabilities are given directly by the softmax output in Eq. (4). With previous token history as input and next token probabilities as output, this allows us to estimate \(P_{\text{ILM}}(y_{t}:y_{0:t-1})\). ILM loss is defined as the negative log probability of each label token given the label sequence prefix: \[\mathcal{L}_{\text{ILM}}^{\text{ASR}}=-\sum_{u=1}^{U}\log P(y_{u}^{\text{ASR} }|\hat{y}_{0:u-1}^{\text{ASR}}) \tag{7}\] The losses \(\mathcal{L}_{\text{EEE}}\) and \(\mathcal{L}_{\text{ILM}}\) are averaged over their respective datasets \(\mathcal{D}_{\text{paired}}\) and \(\mathcal{D}_{\text{paired}}\), then combined in a weighted average to obtain the total JETI loss: \[\mathcal{L}_{\text{JEIT}}^{\text{ASR}}(\mathcal{D}_{\text{paired }},\mathcal{D}_{\text{paired}})=\] \[\mathcal{L}_{\text{EEE}}^{\text{ASR}}(\mathcal{D}_{\text{paired }})+\beta\mathcal{L}_{\text{ILM}}^{\text{ASR}}(\mathcal{D}_{\text{ unpaired}}) \tag{8}\] where \(\beta\) is a hyperparameter controlling the weight given to ILM training (in this work, we use \(\beta=0.2\) to match Meng et al.'s original study). Adapting JEIT to include auxiliary tasks is straightforward. As described in SS4.1, each auxiliary task makes a sequence prediction \(Y_{\text{Aux}}\) based on the predicted ASR sequence \(Y_{\text{ASR}}\). Thus, each auxiliary task predicts \(P_{\text{EE}}(Y_{\text{Aux}}|\hat{Y}_{\text{ASR}};X)\) to produce \(\mathcal{L}_{\text{EEE}}^{\text{Aux}}\). Similarly, the ILM loss is \[\mathcal{L}_{\text{ILM}}^{\text{Aux}}=-\sum_{u=1}^{U}\log P(y_{u}^{\text{Aux} }|\hat{y}_{0:u-1}^{\text{ASR}}) \tag{9}\] The full JEIT loss for each task is defined in the same way as Eq. (8). Total loss is a linear combination of all tasks: (datasets omitted for clarity): \[\mathcal{L}_{\text{IEIT}}^{\text{Total}}= \mathcal{L}_{\text{EEE}}^{\text{ASR}}+\beta\mathcal{L}_{\text{ILM} }^{\text{ASR}}+\alpha_{\text{Cap}}(\mathcal{L}_{\text{EEE}}^{\text{Cap}}+ \beta\mathcal{L}_{\text{ILM}}^{\text{Cap}})\] \[+\alpha_{\text{Pune}}(\mathcal{L}_{\text{EEE}}^{\text{Due}}+ \beta\mathcal{L}_{\text{ILM}}^{\text{Pune}}) \tag{10}\] where each \(\alpha\) is a loss weight for the corresponding task. Matching Wang's original study, we use \(\alpha_{\text{Cap}}=0.1\) and \(\alpha_{\text{Pune}}=0.3\). Figure 1 shows the data flow for paired and unpaired data through the ASR model. ### Transcript annotation While a small amount of our paired training corpus is hand-labeled and capitalized, most of our paired data and all of our unpaired text data have lowercase transcripts. For the lowercase transcripts, we use a text-based increasing RNN teacher model similar to [32] to produce capitalization predictions. Producing pause prediction labels requires different approaches for paired and unpaired data. For paired audio-text data, we use the approach taken by Chang et al. [10], which uses heuristics based on a forced alignment [33] to insert pause tokens into the transcript. There are two such tokens: (pause) denotes a brief stop by the speaker in the middle of a full thought, and \(\langle\text{eos}\rangle\) (end of sentence) is inserted at the end of the full thought, i.e. a full conversational turn. For unpaired text-only data, the above strategy is impossible, since we do not have access to the associated audio. Instead, we rely on the fact that our text-only data comes from short-query sources (see SS6.2). We simply append the \(\langle\text{eos}\rangle\) token to the end of the transcript. ### Multi-task label structure A common approach to transcript labeling for auxiliary tasks would be to embed special tokens corresponding to each task in the transcript itself [7]. However, this is not ideal for inference, since the extra tokens must be expanded in-line with the ASR tokens; if predictions on competing beams differ only in their special tokens, lattice diversity is reduced because the ASR prediction would be identical. To solve for this, we follow Wang et al. [9], factorizing the auxiliary task tokens into parallel sequences of equal length, one for each task. The ASR task is trained on the lowercase transcript sequence, segmented into wordpieces. The capitalization sequence is defined as follows: each token is either \(\langle\text{cap}\rangle\) (capitalized) or \(\langle\text{non-cap}\rangle\) Figure 2: Data preparation for auxiliary tasks. Wordpieces that begin with _denote word boundaries. In this example, we assume that the speaker takes a verbal pause as follows: ”Driving time to... San Francisco,” to illustrate the \(\langle\text{pause}\rangle\) logic. (not capitalized), based on the corresponding wordpiece in the ASR transcript. Similarly, the turn-prediction sequence is populated with \(\langle\texttt{pause}\rangle\) and \(\langle\texttt{eos}\rangle\) tokens corresponding to the wordpieces immediately preceding the corresponding predicted pauses in the transcript. All other token slots are filled with \(\langle\texttt{non-pause}\rangle\). The successive steps of label generation are shown in Figure 2. ## 6 Experimental Details ### Model architecture We use a 128-dimensional log-mel feature frontend computed on 32ms windows with a 10ms stride. We stack four consecutive frames together and sub-sampled by a factor of 3, resulting in 512-dim features at a 30ms framerate. This vector is then concatenated with a 16-dim one-hot domain ID vector [34]. As our ASR backbone we use a 2-pass cascaded encoder model [35]. The first encoder consists of 7 conformer layers [36] with causal convolution and left-context attention. The second encoder consists of 10 conformer layers with a 900ms lookahead. Each conformer layer uses 512-dim 8-head self-attention and a kernel size of 15, and the final layer emits \(D_{a}=384\)-dim encodings. The prediction network of each decoder is a \(V^{2}\) embedding lookup table, which computes \(D_{p}=640\)-dim features based on embeddings of the previous two wordpiece tokens. Each joint network has hidden dimension \(D_{h}=384\), and predictions are made over a vocabulary of \(V=4096\) wordpieces. For evaluation, we report only 2nd pass WER. In total, our model has \(\sim\)160M params. It is implemented in Tensorflow using the Lingvo toolkit, and is trained on proprietary specialized hardware for 500k steps using batch size 4096 for paired and unpaired data. ### Data #### 6.2.1 Paired training data Our training set of audio-text pairs consists of a dataset of 650 million English multi-domain examples, drawn from search, dictation, online video, and telephony domains. A small subset of these utterances are anonymized and hand-transcribed, and the rest are pseudo-labeled by a 600M parameter bidirectional teacher model. To increase model robustness, we apply simulated noise to utterances, as well as SpecAug [37]. #### 6.2.2 Unpaired training data Our text-only data selection pipeline is designed in the style of Sentence-Select by Huang et al [12]. Text query data (\(\sim\) 100B utterances) is collected from web search, maps search, app store search, and online video search domains. This data is filtered for rare words and contrastive filtering based on perplexity is applied. Because the data is selected to include rare words, we expect improvements at the tails of the evaluation distribution. #### 6.2.3 Evaluation Data WER is reported on \(\sim\) 17k utterances representative of real-world voice dictation traffic. Ground truth transcript and auxiliary task annotations are obtained via human labeling. We also report uppercase error rate (UER) on this set, which is calculated by removing all lowercase letters from the ground truth label and the predicted transcript and computing standard WER with upper case letters as words. Since our text-only data focuses on long-tail traffic, we also report UER on a set of \(\sim\)300 utterances with transcripts containing rare words. For pause prediction, we use a testset of \(\sim\)2500 utterances containing hesitations and multiple consecutive commands. Pauses in the audio are hand-annotated as continuation pauses or final pauses. The metrics reported are average precision and recall of the \(\langle\texttt{eos}\rangle\) token. ## 7 Results We evaluate the proposed method (E1) against a baseline (B1) which uses an identical model but is trained on paired data only (Table 1). On the large voice search test set on which it is evaluated, WER does not change, while UER regresses slightly on the voice dictation dataset (1.6% relative). For long tail data, UER improves by a relative 2.0%. Table 2 shows example transcripts demonstrating our proposed method's better capability at recognizing capitalized named entities.

2306.14123

Privacy and Fairness in Federated Learning: on the Perspective of Trade-off

Federated learning (FL) has been a hot topic in recent years. Ever since it was introduced, researchers have endeavored to devise FL systems that protect privacy or ensure fair results, with most research focusing on one or the other. As two crucial ethical notions, the interactions between privacy and fairness are comparatively less studied. However, since privacy and fairness compete, considering each in isolation will inevitably come at the cost of the other. To provide a broad view of these two critical topics, we presented a detailed literature review of privacy and fairness issues, highlighting unique challenges posed by FL and solutions in federated settings. We further systematically surveyed different interactions between privacy and fairness, trying to reveal how privacy and fairness could affect each other and point out new research directions in fair and private FL.

Huiqiang Chen, Tianqing Zhu, Tao Zhang, Wanlei Zhou, Philip S. Yu

2023-06-25T04:38:19Z

http://arxiv.org/abs/2306.14123v1

# Privacy and Fairness in Federated Learning: on the Perspective of Trade-off ###### Abstract. Federated learning (FL) has been a hot topic in recent years. Ever since it was introduced, researchers have endeavored to devise FL systems that protect privacy or ensure fair results, with most research focusing on one or the other. As two crucial ethical notions, the interactions between privacy and fairness are comparatively less studied. However, since privacy and fairness compete, considering each in isolation will inevitably come at the cost of the other. To provide a broad view of these two critical topics, we presented a detailed literature review of privacy and fairness issues, highlighting unique challenges posed by FL and solutions in federated settings. We further systematically surveyed different interactions between privacy and fairness, trying to reveal how privacy and fairness could affect each other and point out new research directions in fair and private FL. Federated learning, data privacy, model fairness + Footnote †: c) 2023 Association for Computing Machinery. XXXX-XXXX/2023/6-ART $15.00 [https://doi.org/10.1145/nnnnnnnnnnnn](https://doi.org/10.1145/nnnnnnnnnnnn) + Footnote †: c) 2023 Association for Computing Machinery. XXXX-XXXX/2023/6-ART $15.00 [https://doi.org/10.1145/nnnnnnnnnnnn](https://doi.org/10.1145/nnnnnnnnnnnn) + Footnote †: c) 2023 Association for Computing Machinery. XXXX-XXXX/2023/6-ART $15.00 ## 1. Introduction Machine learning has changed our lives and will undoubtedly bring us more excitement. However, its success is closely tied to the availability of large-scale training data, and as new learning models keep emerging, the demand for more data persists relentlessly. One worrisome issue with collecting massive amounts of data is the risks that present to privacy. FL (Huiqiang et al., 2019) has emerged as an attractive learning paradigm to meet privacy requirements. Unlike traditional centralized machine learning, FL trains models in a distributed and parallel way such that different clients collectively train a model with their training data. This technique offers two enormous benefits. First, it saves companies the costly process of collecting large-scale data because the clients provide their local data. Second, it preserves the client's privacy by keeping data locally. With such benefits, it is no surprise that the industry has already leaped to put FL into practice, such as Gboard (Ghoard et al., 2019). ### Privacy and Fairness in FL Great achievements have been made with FL. However, the paradigm of FL still suffers from ethical issues surrounding data use. One of those ethical issues is privacy. Although raw data never leave the device, the uploaded gradients/parameters still carry local information. Therefore, a trained model could hold the client's data distribution. Consequently, an adversary can infer information about what data are included in the training set (Zhu et al., 2020; Zhang et al., 2021) or, even worse, reconstruct the training data (Zhu et al., 2020). As such, the research community is endeavoring to identify all potential privacy risks by launching different privacy attacks (Zhu et al., 2020). Accordingly, defenses for all these attacks are also being proposed to secure private data (Zhu et al., 2020). This wargaming between attack and defense is leading us to more private FL environments. Another ethical issue in FL is fairness, which refers to reducing the model's bias towards disadvantaged groups, such as ethnic minorities, women, or the aged. Fairness in FL is defined at two different levels. The first pertains to _algorithmic fairness_(Kolmogorov, 1969; Kolmogorov, 1969), where model output should not skew towards disadvantaged groups defined by some sensitive attributes. The second pertains to _client fairness_(Kolmogorov, 1969; Kolmogorov, 1969). In vanilla FL (Zhu et al., 2020), models trained on the larger dataset are given higher importance during aggregation. Hence, the global model will be optimized to capture the data distributions of clients with a larger dataset. Therefore the model performance will vary significantly among clients, which imposes unfairness at the client level. Privacy and fairness are two crucial ethical notions, and violating either of them is unacceptable. However, to date, the research community has primarily considered these two issues separately, yet they are inextricably entwined. For example, it is well known that privacy comes at the cost of accuracy. What is surprising is that the cost is not consistent across all groups as expected, where disadvantaged groups often suffer more of an accuracy decrease than the other groups due to data scarcity (Zhu et al., 2020; Zhang et al., 2021). In other words, ensuring privacy can exacerbate the inequities between groups. Fairness, in turn, may negatively affect privacy. For example, to ensure a classification model is fair, the server usually needs to know the underlying distribution of the training dataset to eliminate bias existing in either the training data (Zhu et al., 2020; Zhang et al., 2021; Zhang et al., 2021) or the model (Zhu et al., 2020; Zhang et al., 2021; Zhang et al., 2021). This means the client will share more data with the server, increasing the privacy risk. Therefore, in addition to reviewing privacy and fairness issues in FL, another motivation of this survey is to explore the possible interactions between privacy and fairness and to discuss the relationships between fairness and privacy. It is worth noting that the issue of client fairness adds an extra layer of complexity to the federated setting. To the best of our knowledge, this survey is the first attempt to examine the relationships between privacy and fairness in the federated setting. \begin{table} \begin{tabular}{c c c c c c} \hline \hline \multirow{2}{*}{Reference} & \multicolumn{2}{c}{Privacy-preserving} & \multicolumn{2}{c}{Fairness-aware} & Interactions \\ \cline{2-5} & Privacy & Defense & Algorithmic & Client & between privacy \\ & attack & & fairness & fairness & and fairness \\ \hline (Zhu et al., 2020) & ✓ & ✓ & & & \\ (Zhu et al., 2020) & ✓ & ✓ & & & \\ (Zhu et al., 2020) & ✓ & ✓ & & & \\ (Zhu et al., 2020) & ✓ & ✓ & ✓ & ✓ & \\ Our work & ✓ & ✓ & ✓ & ✓ & ✓ \\ \hline \hline \end{tabular} \end{table} Table 1. Comparison to Related Surveys on Privacy or Fairness in FL ### Main Contribution This survey provides a broad view of privacy and fairness issues in FL. We first illustrated that FL is not as private as it claimed to be. Adversarial clients and server have several new attack vectors at their disposal, and we outlined several techniques for preserving privacy against these attacks. Turning to fairness, we explained the two lines of fairness notions adopted in FL and the corresponding debiasing strategies. Lastly, we discussed interactions between privacy and fairness. Our contributions are as follows: * This is the first survey that provides a comprehensive overview of privacy, fairness, and the interactions between the two. * We present a detailed survey of privacy attacks and defenses in FL, discuss how these privacy attacks could damage privacy in FL and highlight the assumptions and principle methods of these attack strategies. * Following a rigorous enumeration of the sources of bias in the FL pipeline, we discuss the fairness notions adopted in FL and summarize fairness-aware FL approaches. * We point out several future research directions toward training private and fair FL models. ## 2. Background Knowledge ### Definition of FL The goal of FL is to train a global model in a distributed way. The objective function is formulated as: \[\min_{w}f\left(w\right)=\sum_{k=1}^{m}p_{k}F_{k}\left(w\right) \tag{1}\] where \(m\) is the number of clients, \(p_{k}>0\) is the aggregating weight of client \(k\), satisfying \(\sum_{k}p_{k}=1\). \(F_{k}(w)\) is the empirical risk on client \(k\)'s data. In a trivial setting, \(p_{k}\) is the ratio of local samples to total samples of all clients. The process of FL consists of two stages: the training and inference stages. Three actors are involved: 1) clients, each of which has a local dataset and will use it to contribute to the global model's training by uploading local gradients/parameters, noting that each client's dataset may vary from the others; 2) a server that coordinates the learning process; and 3) users, who will use the final well-trained model. In each iteration of FL, selected clients download the global model and perform learning algorithms locally. They communicate their updates to the server for aggregation and model updating. This interaction between clients and the server repeats until the model converges. At the inference stage, a well-trained model is deployed to users, where users can infer the model via black-box access. This step is no different from a traditional data center approach. ### Privacy Disclosure in FL Recent research verified the privacy risk of FL. Adversaries can glean the participants' training data. For example, Zhu et al. (Zhu et al., 2021) fully reconstructed training data from the victim client through the uploaded gradients as shown in Fig.1(a). In the course of the FL pipeline, several attack interfaces exist for an adversary, who could be the server or a client in FL. The attack could occur during the training or inference stage, within or outside the FL. The attack targets include membership, property, class representative, and the raw data (Zhu et al., 2020). #### 2.2.1. Membership Inference Attacks Membership Inference Attacks (MIA) aim to identify whether a given sample was used to train the target model. These types of attacks can pose privacy risks to individuals. For example, confirming a patient's clinical record was used to train a model associated with a particular disease would reveal that patient's health condition. MIA was initially investigated by Shokri et al. (2016). The attack models are essentially binary classifiers. Given an instance \(X\) and a target model \(F_{t}\), the goal of the MIA model is to identify whether or not \(X\) is contained within the training dataset \(D\) of the target model \(F_{t}\). #### 2.2.2. Property Inference Attacks Property inference attacks aim to recover some property of the training set, which may be irrelevant to the main tasks. Such as the property of "wearing glasses" against a gender classifier or the composition of the training dataset. This kind of attack also leads to privacy issues. With proper prior knowledge, the adversary can infer the presence of a specific sample in the training set. #### 2.2.3. Model Inversion Attacks Model inversion attacks aim to recover class-specific features or construct class representatives by accessing the target model and other possible auxiliary information. The recovered data is a representing sample (usually a synthetic sample) that only reflects some aspects of the training data and is not a member of the training set. #### 2.2.4. Reconstruction Attacks Reconstruction attacks (Shokri et al., 2016) aim to reconstruct a probabilistic version of samples in the training set. Success in a reconstruction attack is measured by comparing the reconstruction with the original data. If the two are similar, the attack has been successful. Fig.1(b) (Zhou et al., 2017) shows an example. Unlike the model inversion attacks, the recovered data here is almost the same as the training dataset at the pixel level and belongs to the training dataset. ### Privacy-preserving Techniques In the realm of FL, plenty of studies have shown how to break the basic privacy assurances, such as determining a client's membership in the training set (Shi et al., 2017), ascertaining the class representations of the client's training data (Shi et al., 2017; Shi et al., 2018; Shokri et al., 2018), and, the worst case of all, procuring the raw training data (Shi et al., 2017). Several privacy-preserving techniques can help stop these privacy leakages, including cryptographic techniques and the perturbation approach. #### 2.3.1. Cryptographic Approach Secure computation is a cryptographic technique in which functions are evaluated based on a set of distributed inputs without revealing additional information, e.g., the parties' inputs or intermediate results. Secure multi-party computation, homomorphic encryption, and secret sharing are the most common choices for a secure computing platform. Multi-party computation (Zhou et al., 2017) was first introduced to secure the private inputs of multiple participants while they jointly compute an agreed-upon model or function. Formally, \(n\) participants Figure 1. Reconstruction attack in FL. \(p_{1},p_{2},\ldots\), and \(p_{n}\) can collaboratively compute \(y=f\left(x_{1},\ldots,x_{n}\right)\), where \(x_{i}\) is a secret input that belongs to participants \(p_{i}\), This form of secure computing offers both correctness and privacy. After all, no participant learns anything about the others' data other than the final result. Homomorphic encryption (Kumar et al., 2017) allows certain mathematical operations, such as addition and multiplication, to be performed directly on ciphertexts. These can then be used as the basis for more complex arbitrary functions. #### 2.3.2. Perturbation Approach With privacy concerns in mind, one needs to be cautious of how much information about a participating client is revealed during training. The perturbation approach arises as a natural way of preventing information leaks. By injecting the proper amount of artificial noise into the original data, the statistical information calculated from the perturbed data will be statistically indistinguishable from the original data. There are three types of widely used perturbation techniques: differential privacy (DP), additive perturbation, and multiplicative perturbation. DP proposed by Dwork (Dwork, 1998), is the gold standard. The intuition behind DP is to mask the contribution of any individual user by a sufficient level of uncertainty. A randomized mechanism \(\mathcal{M}\) is said to be \((\epsilon,\delta)\)-differentially private if, for any pair of neighboring datasets \(\mathcal{D}\) and \(\mathcal{D}^{\prime}\), and for every set of \(S\subseteq Range\) (\(\mathcal{M}\)), if \(\mathcal{M}\) satisfies: \[\Pr\left[\mathcal{M}\left(\mathcal{D}\right)\in S\right]\leq\exp\left(\epsilon \right)\cdot\Pr\left[\mathcal{M}\left(\mathcal{D}^{\prime}\right)\in S\right]+\delta \tag{2}\] The parameter \(\epsilon\) is defined as privacy budget, which measures how alike the two adjacent datasets \(\mathcal{D}\) and \(\mathcal{D}^{\prime}\) are to each other. A smaller \(\epsilon\) indicates a stronger privacy guarantee. If \(\delta=0\), then the randomized mechanism \(\mathcal{M}\) is degraded into \(\epsilon\)- DP. ### Fairness in FL Algorithms are widely used to assist in making recommendations, assessing loan applications, etc. Several studies have identified the unfairness in different algorithmic scenarios (Kumar et al., 2017). One example is the hate speech detector designed to rate the toxicity score of the given phrases to help companies like Twitter recognize harmful speech. The detector relies on a tool called _Perspective_, which is trained on labeled data. The detector behaves differently towards phrases written in African American English in a racially biased way, as Fig. 2 shows (Kumar et al., 2017). Figure 2. Racial disparities in classifier predictions on tweets written in African-American English and in Standard American English (reproduced from (Kumar et al., 2017)) #### 2.4.1. Bias in FL Fairness can be eroded by bias. According to Olteanu et al. [143], bias can slip into data flows from generation to collection to processing [143]. The distributed learning paradigm of FL brings new and unique challenges to our efforts to build fair models. One challenge is that the independent and identical distribution (i.i.d.) assumption no longer holds [89]. In FL, bias can also be introduced by either the client or the server. * prejudice, underestimation, negative legacies, etc. [93]. This bias is then integrated into the global model through client and server interactions. Second, bias can strike when clients are dropped out of the federated setting due to device shutdowns or communication limitations. In these cases, the global model will find it hard to fit the clients' data properly. The massively distributed data also incurs bias. In this case, the client does not have enough data to capture an underlying distribution, and moreover, the underlying distributions of different clients are probably not the same. [18, 111, 228] * **Server-introduced bias**. The server can also add bias. As the coordinator of the learning scheme, the server will sample clients in each round to train a global model with their local data [139]. However, the sampling process is prone to producing bias if it is not done with careful consideration. First, in terms of efficiency, only a fraction of clients are selected in each round [129]. Yet the data distribution of only a few selected clients forms an inadequate representation of the actual population distribution. Second, the sampling may be skewed toward certain clients. For instance, to speed up convergence, the server prefers clients that meet specific criteria [29, 142]. #### 2.4.2. Fairness Notions in FL To date, researchers have proposed several definitions of fairness, see, e.g., [182, 131]. These definitions vary from scenario to scenario, and it is unlikely that there will ever be one particular definition of fairness that fits all circumstances. Table 2 lists common algorithmic fairness notions adopted in centralized machine learning. At a high level, two families of definitions exist the _individual_ notion and _statistical_ notion [31]. * **Individual notion**. The individual notions ensure fairness between specific pairs of individuals: "_Give similar predictions to similar individuals_." [43]. Formally, for a set of samples \(V\), a distance metric is defined as \(d:V\times V\to R\) to measure the similarity. A function \(\mathcal{M}:V\rightarrow\Delta A\) maps the samples \(V\) to the probability distributions over outcomes, and another distance \(D\) metric measures the distance between the distributions of outputs. Fairness is achieved if and only if \(D\left(\mathcal{M}\left(x\right),\mathcal{M}\left(y\right)\right)\leq d\left( x,y\right)\). This family of definitions provides a meaningful guarantee. However, they are at the cost of making significant assumptions, some of which are non-trivial problems in fairness. * **Statistical notion**. The statistical notions provide fairness assurance at a statistical level. For the protected demographic groups \(G\) (such as racial minorities), some statistical measures are required to be equal across all of these groups. These statistical measures include positive classification rates [94, 20, 43], false positive and false negative rates [73, 103], and positive predictive value [211, 30]. Detailed enumeration can be found in [182, 12]. This family of definitions requires no assumption over data and can be easily verified. However, statistical notions are insufficient as a fairness constraint, which does not give meaningful guarantees to individuals or structured subgroups of the protected demographic groups. Jiang et al. [85] generalized the demographic parity [106] to continuous sensitive attribute. Apart from algorithmic fairness, which is measured on sensitive attributes, fairness in FL can also be made from a client's view since clients are naturally grouped by attributes like geographic location, gender and income [40]. At a client level, fairness can be evaluated by different metrics. **Definition 1** (_Good-intent fairness_(Kumar et al., 2017)).: The training procedure does not overfit a model to any device at the expense of other clients in FL. This metric improves the worst-case performance. Li et al. (Li et al., 2019) took a further step. They tried to ensure a fair FL model for all clients by producing a more uniform model performance across all clients. Fairness is defined as the uniformity of the accuracy distribution across clients in FL. **Definition 2** (_Accuracy parity_(Li et al., 2019)).: Consider two trained models, \(f(w)\) and \(f(\tilde{w})\). The model that provides the most uniform performance across all clients will also provide the fairest solution to the FL objective in Eq. (1). ### Interactions between Privacy and Fairness Both fairness and privacy are important ethical notions in machine learning and have been extensively studied. However, the majority of current studies in the research community consider fairness and privacy separately. However, the interactions between privacy and fairness are bilateral. * **Privacy degrades fairness**. Several works have observed inconsistent reductions in accuracy caused by private mechanisms on classification (Kumar et al., 2017) and generative tasks (Zhu et al., 2019). It turns out that privacy mechanisms affect the underrepresented group more than other groups. * **Fairness increases privacy risk**. Fairness comes at the cost of privacy. To ensure fairness, a model is trained to perform equally on data from different groups, even though the underrepresented group didn't have enough data in the training set, which incurs overfit and increases the privacy risk (Zhu et al., 2019). ## 3. Privacy in FL With the advent of FL, many claims that user data are now secure. However, even sharing a small fraction of gradients (Zhu et al., 2019; Li et al., 2019) with a server would raise privacy concerns. In FL, there are several unexplored types of privacy attacks: _membership inference attacks, property inference attacks, model inversion attacks_, and _reconstruction attacks_. This section will outline these attacks before moving on to mitigation techniques and discussions. \begin{table} \begin{tabular}{p{113.8pt} p{113.8pt} p{113.8pt}} \hline **Fairness notion** & **Definition** & **Explanation** \\ \hline Individual Fairness (Kumar et al., 2017) & \(D(M(x),M(y))\leq d(x,y)\) & Similar samples receive similar treatment \\ \hline \multirow{2}{*}{Eqal Opportunity (Kumar et al., 2017)} & \multirow{2}{*}{\(\Pr[\hat{Y}=1|A=0,Y=1]=\Pr[\hat{Y}=1|A=1,Y=1]\)} & Equal true positive rates for protected/unprotected groups \\ & & \\ \hline Equal Accuracy (Kumar et al., 2017) & \multirow{2}{*}{\(\Pr[\hat{Y}=Y|A=0]=\Pr[\hat{Y}=Y|A=1]\)} & Equal prediction accuracy for protected/unprotected groups \\ \hline Equal Odds (Kumar et al., 2017) & \multirow{2}{*}{\(\Pr[\hat{Y}=1|A=1,Y=y]=\Pr[\hat{Y}=1|A=0,Y=y]\)} & Equal positive rates for protected/unprotected groups \\ & & \\ \hline Treatment Equality (Kumar et al., 2017) & \multirow{2}{*}{Equal false negatives and false positives for protected/unprotected groups} \\ & & \\ \hline Demographic Parity (Kumar et al., 2017) & \multirow{2}{*}{\(\Pr[\hat{Y}|A=0]=\Pr[\hat{Y}|A=1]\)} & Outcome is independent of the protected attribute \\ \hline \end{tabular} \end{table} Table 2. Definitions of Algorithmic Fairness Notions ### Membership Inference Attacks The adversary's goal in FL is to determine whether a given sample belongs to a single client's private training data or of any participants (Levy et al., 2017). MIAs take occur in different ways in FL. #### 3.1.1. White-box and Black-box MIAs Based on the access granted to the adversary, MIAs can be divided into black-box and white-box (Levy et al., 2017). In the black-box setting, the adversary can only obtain a prediction vector computed by the target model while the internal parameters remain secret. MIAs in this setting exploit the statistical differences between a model's predictions on its training set versus unseen data (Levy et al., 2017). Truex et al. (2017) described a systematic approach to constructing a black-box MIA model and the general formulation of each component in the attack model. However, since the global model is shared with all participants for local training, it is often assumed that an adversary has white-box access in FL. The white-box access renders much more information to the adversary, such as the internal parameters of each layer. This enables the adversary to calculate the outputs of each layer. Nasr et al. (Levy et al., 2017) designed a deep learning attack model that separately processes the gradients extracted from different layers of the target model and combines this information to compute the membership probability of a target data point. #### 3.1.2. Training and Inference MIAs MIAs can be launched during the training stage or once the model is complete in the inference stage in FL. In the training stage, the adversary could be the server or any client participating in the training. Both characters have white-box access to the global model and can easily save the snapshots of the global model at each iteration during training. In this way, the adversary obtains multiple versions of the target model over time and acquires the updated information to infer private data (Levy et al., 2017). In addition to passively collecting the updated information, the adversary may further modify the information to allure the victim clients to reveal more information. In the inference phase, the FL model is well-trained and fixed. The adversary can only perform an inference attack passively. In this case, MIA in FL resembles that in a centralized setting. The attack's success largely depends on the information that is revealed to the adversary. Melis et al. (Melis et al., 2017) investigated privacy leaks concerning membership during the inference phase and showed that positions of words in a batch could be revealed from a deep learning model. #### 3.1.3. Active and Passive MIAs The adversary can conduct MIAs against the FL model actively or passively. For instance, the server can either adaptively modify the aggregate parameters or honestly calculate the global model and passively conduct MIAs (Meliis et al., 2017). Melis et al. (Melis et al., 2017) designed MIAs against models operating on non-numerical data (e.g., natural-language text). An embedding layer is equipped for the target model, transforming the inputs into a lower-dimensional vector representation. The adversary passively saves a snapshot of the joint model parameters \(\mathbf{w}_{t}\). The difference between the consecutive snapshots \(\Delta\mathbf{w}_{t}=\mathbf{w}_{t}-\mathbf{w}_{t-1}=\Sigma_{k}\Delta\mathbf{ w}_{t}^{k}\) reveals the aggregated updates from all participants and hence reveals the membership. Nasr et al. (Nasr et al., 2017) performed active MIAs on FL models by reversing the stochastic gradient descent algorithm and extracting membership information. If the target data point belongs to the training dataset, the attacker's modifications will be nullified since the target model will descend the model's gradient for training samples. However, if the target data sample is not used during the training, the target model will not respond to the attacker's modification. Thus, membership can be deduced. In (Turaev et al., 2017), a malicious client actively mislabels the training sample to fool the victim into releasing private information. #### 3.1.4. Insider and Outsider MIAs FL involves two types of actors who can access model information: internal actors (participating clients and the server) and external actors (model consumers and eavesdroppers). Therefore, FL systems must withstand potential adversaries within and outside the protocol. The inner adversary could be a client or a server. Clients are picked at random to participate in a training round. When training with hundreds or millions of clients, malicious clients are highly likely involved, who will attempt to deduce the sensitive information of others (Krishnan et al., 2017). The real-time nature of FL added to the inner attacker's strength. For example, Zhang et al. (Zhang et al., 2019) trained a GAN as a malicious client during training to infer the data of other clients in FL. A malicious central server poses a greater threat than a malicious client. Because it can manipulate the global model supplied to victims and obtain more information. Nasr et al. (Nasr et al., 2019) launched MIAs from both the client and server sides and witnessed a higher inference accuracy as a curious central server than as a malicious client. In addition to internal threats, FL also faces potential attacks from adversaries outside the system. Once the FL training is finished and the model is deployed to users, these users may conduct both black- and white-box attacks depending on their access. #### 3.1.5. Discussion The attacks mentioned above demonstrate the vulnerability of FL to privacy attacks, and these privacy risks stem from two assumptions made within the FL protocols: 1) _The server is trustworthy_. FL gives the server access to each participant's updates in the form of gradients or model parameters containing clients' private information. The server can even purposefully send a modified model to steal information. 2) _Clients are honest_. A malicious client can collect several copies of the global model from the rounds it participates in. In this way, inference phase attacks on data privacy are also plausible during the learning phase. Additionally, adversarial clients may influence and shift the bounds of the model during development rather than just abusing the boundaries of a model's service while it is in production. ### Property Inference Attacks The adversary in property inference attacks attempts to infer the specific property of the subset of the training dataset. The target property may be irrelevant to the classification task (e.g., "wearing glasses" in a gender classification task) and do not characterize the whole class. The attack is made at the population level as opposed to a single sample in MIA. In terms of when the attack is launched, property inference attacks can be classified as _static_ or _dynamic_ attacks. The static attack is applied after the training phase has concluded and the target training set is fixed. The dynamic attack typically occurs during the training phase in FL. In this instance, the training set is changing dynamically. #### 3.2.1. Static Attacks The research of property inference attacks dates back to (Ateniese et al., 2017). Ateniese et al. performed a property inference attack against Hidden Markov Models and Support Vector Machine based on a meta-classifier. A set of shadow classifiers were trained on a dataset similar to the target model except for the target property. The meta-classifier is trained with shadow classifiers as the input to find the classifiers trained on the dataset with the target property. Ganju et al. (Ganju et al., 2018) extended this attack to a fully connected neural network case. They shared a similar idea, using the gradient of shadow classifiers to train a meta-classifier. Different from (Ateniese et al., 2017), their research focuses on improving the attack efficiency by taking permutation invariance into account. #### 3.2.2. Dynamic Attacks In every communication round of FL, clients are selected at random. This means the training data is dynamically changing, which weakens the property inference attack because the target property appears unpredictable, thereby diminishing the distinguishability of model updates (Wang et al., 2018). Wang et al. (Wang et al., 2018) explored property inference attacks within the FL framework. Inspired by the relationship between the changing of neuron weights in the output layer and the sample label, the authors proposed three attacks as an eavesdropper to infer the labels' quantity composition proportion. Recently, Wang et al. [188] presented a poisoning-assisted property inference attack in FL from the client's viewpoint, aiming at inferring if and when a sensitive property emerges. The authors built their attacks around the realization that regular model updates reflect the shift in data distribution and, in particular. A binary classifier is trained to make predictions based on these periodic model updates. A property-specific poisoning attack is proposed to distort the decision boundary of the shared model on target attribute data. Thus, model updates have a better discerning ability to infer target property. #### 3.2.3. Discussion The MIAs and reconstruction attacks represent two ends of a spectrum of privacy invasion. Property inference attacks locate in the middle and seek to determine if the attackers' target property is present in the training samples. This type of attack is more complex than MIA since the target property doesn't always match the attributes that characterize the classes of the FL model. Nonetheless, a property inference attack poses a greater threat than MIA. Using the real-time nature of FL, the adversary can even infer when the target property appears. ### Model Inversion Attacks Fredrikson et al. [55] initiated model inversion attacks on tabular data. A subsequent work [54] extended it to the image data. The attack is formulated as an optimization problem to synthesize the input for a given label: \(y:\max_{x}\log T_{y}\left(x\right)\), where \(T_{y}\left(x\right)\) is the probability of the model \(T\) outputs label \(y\) for input \(x\). The access could be black-box [54] or white-box [27, 225]. #### 3.3.1. Black-box Attacks In the black-box setting, the attacker can only make prediction queries to the model. Fredrikson et al. [54] built attack algorithms following the maximum a posterior principle. Their attack recovered a recognizable image of a person given only API access to a facial recognition system and a specific name of a target person. Yang et al. [203] engineered an inversion model a perform the inversion attacks. The adversary composed an auxiliary set assumed generic enough to retain meaningful information to regularize the ill-posed inversion problem [154]. However, the target model is usually assumed to be simple networks, and the generalization to complex models is not trivial. The inversion problem of a neural network is non-convex, and the optimization suffers minimal local problems, which leads to poor attack performance. #### 3.3.2. White-box Attacks n the white-box setting, the attacker has complete knowledge of the model. Zhang et al. [225] sketched a generative model to learn an informative prior from the public dataset. This prior is then used to regulate the inversion problem. Benefiting from this, the authors revealed private training data of DNNs with high fidelity. Chen et al. [27] boosted [225]'s methods. They leveraged the target model to label a public dataset, and a GAN model was trained to distinguish not only real and synthesized samples but also labels. They also modeled the private data distribution to reconstruct representative data points better. The success of model inversion attack benefits from an informative prior. #### 3.3.3. Discussion MIAs can be performed with either black-box access or white-box access. When given black-box access, the attack's success heavily relies on the auxiliary dataset, which is assumed to share the same generic features as the private target dataset. Furthermore, the target models in this category are usually simple due to limited access. In the white-box case, the target models extend to DNNs. Most attacks implement GAN to synthesize samples to mimic the private samples regarding the soft labels. This kind of attack is less common compared with reconstruction attacks in FL. ### Reconstruction Attacks Unlike MIAs, reconstruction attacks attempt to retrieve training data and pose a much more severe threat to privacy. As demonstrated by Aono et al. (Aono et al., 2018), the gradient of the weights is proportional to that of the bias in the first layer of the model, and their ratio approximates the training input. Geiping et al. (Geiping et al., 2018) demonstrated that it is possible to faithfully reconstruct images at high resolution given knowledge of the parameter gradients. Such a privacy break is possible even for deep neural networks. Huang et al. (Huang et al., 2019) evaluated existing reconstruction attacks and defenses. Gupta et al. (Gupta et al., 2019) extended this attack to text data and successfully reconstructed single sentences with high fidelity for large batch sizes. To date, we know of two kinds of reconstruction attacks, namely _optimization-based attacks_ (Opt-based) and _closed-form attacks_. #### 3.4.1. Optimization-based Attack Raw data can be reconstructed from gradients by solving an optimization problem. Given a machine learning model \(f(w)\) and the gradient \(g=\frac{1}{b}\sum_{j=1}^{b}\nabla_{w}L_{w}\left(x_{j}^{*},y_{j}^{*}\right)\) computed on a private batch \((x^{*},y^{*})\in\mathbb{R}^{b\times d}\times\mathbb{R}^{b}\) with bath size \(b\). The adversary tries to reconstruct \(x\in\mathbb{R}^{b\times d}\) as an approximation of the true data \(x^{*}\) by solving the following optimization problem: \[\arg\min_{x}\mathcal{L}_{grad}\left(x;w,g\right)+\alpha\mathcal{R}_{aux}\left( x\right) \tag{3}\] The first part of Eq. 3 pushes the recovered gradients towards the true gradients \(g\), hence deducing a better approximation. The regularization term \(\mathcal{R}_{aux}\left(x\right)\) is used to incorporate the prior knowledge to further improve reconstruction. Zhu and Han (Zhu and Han, 2018) proposed _DLG_ attack with \(l_{2}\) distance as the reconstruction loss \(\mathcal{L}_{grad}\). _DLG_ starts with randomly generated dummy samples and then iteratively optimizes the dummy samples and labels until they converge. Finally, _DLG_ achieves pixel-wise recovery accuracy for image classification and token-wise recovery accuracy for a masked language model. Zhao et al. (Zhao et al., 2019) improved Zhu and Han's work (Zhu and Han, 2018) on convergence speed and reconstruction fidelity by leveraging the relationship between the ground-truth labels and the signs of the gradients. The optimization problem Eq. 3 is often under-determined (Zhu and Han, 2018), as the information in the gradients \(g\) is usually insufficient to recover the training data \(x^{*}\). Even when the gradient size is substantially bigger than the input data dimension. As demonstrated by Zhu and Blaschko (Zhu and Han, 2018), when the learning model is huge, there may be a pair of separate data sharing the same gradient. In response to this, one may introduce prior knowledge as a regularization term to narrow the search space, making it more consistent with the underlying distribution of training data. Yin et al. (Yin et al., 2019) utilized the local batch norm as the regularization term since adjacent pixels in natural photographs are likely to have comparable values. They achieved precise recovery of the high-resolution images on complex datasets, deep networks, and large batch sizes. Hatamizadeh et al. (Hatamizadeh et al., 2019) extended (Yin et al., 2019)'s approach to vision transformers and discovered that, because of the attention mechanism, vision transformers are substantially more sensitive than previously researched CNNs. In the image domain, total variance is another choice (Zhu and Han, 2018; Zhu and Han, 2018). \begin{table} \begin{tabular}{l c c c c c} \hline \hline & \begin{tabular}{c} **Theoretical** \\ **guarantee** \\ \end{tabular} & \begin{tabular}{c} **Convergence** \\ **convergence** \\ \end{tabular} & \begin{tabular}{c} **Running** \\ **time** \\ \end{tabular} & \begin{tabular}{c} **Recovered** \\ **image** \\ \end{tabular} & \begin{tabular}{c} **Applicability** \\ **in** \\ \end{tabular} \\ \hline Opt-based & No & Local optimal & Slow & With artifacts & No limitation & No \\ \hline Closed-form & Yes & / & Fast & Original & Limited & Yes \\ \hline \hline \end{tabular} \end{table} Table 3. Comparison between two reconstruction attack categories Selecting a proper loss function can also contribute to attack efficiency. By combining mean square error and Wasserstein distance (Beng et al., 2017), Ren et al. (2018) achieved a better reconstruction result than (Zhou et al., 2018; Ren et al., 2019) in terms of batch size and reconstruction fidelity. Geiping et al. (2019) adopted cosine distances to better capture the observation that the angle between two data points quantifies the change in prediction. With this method, they rebuilt a single high-resolution image and a series of low-resolution photos with a maximum batch size of 100. Jeon et al. (2019) systematically investigated ways to best utilize gradients, including using them to extract prior information. Wei et al. (2019) conducted a thorough analysis of how different hyper-parameter setups and settings for attack algorithms influence the effectiveness of the attack and its cost. In an effort to investigate the worst-case attack and evaluate the effectiveness of defense methods for reconstruction attacks, Balunovic et al. (2019) formulated the gradient leakage problem in a Bayesian framework and analyzed the condition for a Bayes optimal adversary. #### 3.4.2. Closed-form Attack In an attempt to provide a theoretical understanding of how and when gradients lead to the remarkable recovery of original data, several studies investigated the possibility of recovering the input of a learnable affine function from gradients (Beng et al., 2017; Ren et al., 2019; Ren et al., 2019). Aono et al. (2019) initiated the closed-form attack based on gradients. In certain circumstances, an honest-but-curious server could directly calculate individual data from the gradients uploaded by clients (Beng et al., 2017; Ren et al., 2019). Consider a fully connected layer \(Wx+b=z\) with \(l=l(f(x),y)\) as the loss function, where \(x\) and \(z\) are the input and output vectors, respectively. Private data \(x\) can be derived from \(l\)'s gradients w.r.t \(W\) and \(b\), i.e.,: \(x^{T}=\frac{\partial l}{\partial W}\oslash\frac{\partial l}{\partial b}\). Here \(\oslash\) denotes entry-wise division. The assumption of a fully connected layer holds for the last prediction layers in many popular architectures. As a result, the prediction modules' input, which is the output of the preceding layers, can be rebuilt. These outputs typically include some information about the training data, making them vulnerable to attackers. In this light, the ability to recover ground truth label information from the gradients of the final fully-connected layer, as stated in Zhao et al. (2018), is very intriguing. Despite its inspiring nature, Aono et al.'s work (2019) has some limitations. First, it does not apply to convolutional neural networks due to a mismatch in dimensions. Second, it cannot deal with batch inputs. For the batch input \(\left\{x_{j}\right\}_{j=1}^{b}\), all derivatives are summed over the batch dimension \(b\) and the recovered \(\bar{x}\) is merely proportional to the average of batch inputs \(\sum_{j=1}^{b}x_{j}\). To fix the dimension mismatch issue, Zhu and Blaschko (2019) converted the convolutional layer into a fully connected layer using circulant matrix representation (Zhu and Blaschko, 2019) of the convolutional kernel (Zhu and Blaschko, 2019). The gradients of each layer were interpreted as the _gradient constraints_. Finally, they recursively reconstructed the layer-wise input. However, their implementation can only recover low-resolution images in settings where the batch size equals 1. For the batch input, the algorithm returns a linear combination of the training data. To address these difficulties with the batch size, Fowl et al. (2019) suggested making minor but malicious changes to the global model to reconstruct the client's data from a batch of gradient updates. The key idea is to separate the batch data by some quantity \(h\), such as image brightness. To this end, an imprint module is added to the global model, which acts as a filter that separates a batch of samples based on quantity \(h\). Qian and Hansen (2019) found bias term in the output layer is the key to the success of reconstruction. A fully-connected neural network requires just one node in one hidden layer for single-input reconstruction. In contrast, mini-batch reconstruction requires that the hidden units exceed the input size. Pan et al. (2019) conducted an analytic investigation of the security boundary of the reconstruction attacks. Given a batch input, the secure/insecure boundary of the reconstruction attack was characterized by the number of Exclusively Activated Neurons (ExANs), where the more ExANs, the more likely the attack's success. #### 3.4.3. Discussion The closed-form attacks outperform optimization-based attacks in several aspects. First, the closed-form attacks provide a theoretical guarantee of convergence. In contrast optimization-based attacks suffer from the local optimum problem since a non-convex optimization may not always converge to a correct solution (Shi et al., 2018). Further, optimization-based attacks are sensitive to initialization (Shi et al., 2018), whereas closed-form attacks do not. Second, the deterministic algorithms run by closed-form attacks are faster than optimization-based attacks. Third, closed-form attacks recover the data more accurately, while optimization-based methods, like GradInversion (Zhu et al., 2019), recover data with artifacts. Jin et al. (Jin et al., 2020) made a good comparison between different reconstruction attacks in FL. Table 4 summarizes their finding and includes some additional results from this study. ### Privacy-preserving Techniques Privacy-preserving machine learning approaches can be roughly classified as cryptographic approaches and perturbation approaches. Cryptographic approaches enable computation over encrypted data and provide rigorous privacy guarantees in the training process. However, they come at a high computational cost compared to the non-encryption alternatives (Jin et al., 2020). This computation \begin{table} \begin{tabular}{c c c c c c} \hline \hline \multirow{2}{*}{**Method**} & \multicolumn{2}{c}{**Objective function**} & \multicolumn{1}{c}{**Maximal**} & \multicolumn{1}{c}{**Opt-based/**} & \multicolumn{1}{c}{**Theoretical**} & \multicolumn{1}{c}{**Additional**} \\ \cline{2-7} & \(\mathcal{L}_{grad}\) & \(\mathcal{R}_{aux}\) & **batch size** & **Closed-form** & **guarantee** & **information** \\ \hline iDLG (Shi et al., 2018) & \(l_{2}\) distance & / & 8 & Opt-based & No & No \\ \hline DLG (Shi et al., 2018) & \(l_{2}\) distance & / & 8 & Opt-based & Yes & No \\ \hline Inverting gradients(Shi et al., 2018) & \begin{tabular}{c} Cosine \\ similarity \\ \end{tabular} & Total variance & 100 & Opt-based & Yes & \begin{tabular}{c} Local updates; \\ BN statistics \\ \end{tabular} \\ \hline [192] & \(l_{2}\) distance & \begin{tabular}{c} Label-based \\ regularizer \\ \end{tabular} & 8 & Opt-based & Yes & No \\ \hline SAPAG (Jin et al., 2020) & \begin{tabular}{c} Gaussian \\ kernel \\ based function \\ \end{tabular} & / & 8 & Opt-based & No & No \\ \hline R-GAP (Shi et al., 2018) & \begin{tabular}{c} Recursive \\ gradients \\ \end{tabular} & / & 5 & Closed-form & No & No \\ \hline \begin{tabular}{c} Theory- \\ oriented (Shi et al., 2018) \\ \end{tabular} & \(l_{2}\) distance & \begin{tabular}{c} \(l_{1}\) distance of \\ feature map \\ \end{tabular} & 32 & Closed-form & Yes & \begin{tabular}{c} Exclusive \\ activated \\ neurons \\ \end{tabular} \\ \hline \begin{tabular}{c} GradInversion \\ (Zhu et al., 2019) \\ \end{tabular} & \(l_{2}\) distance & \begin{tabular}{c} Group \\ consistency \\ \end{tabular} & 48 & Opt-based & No & BN statistics \\ \hline CAFE (Jin et al., 2020) & \(l_{2}\) distance & Total variance & 100 & Opt-based & Yes & Batch indices \\ \hline GIAS\&GIM (Shi et al., 2018) & \begin{tabular}{c} Negative \\ cosine \\ \end{tabular} & \begin{tabular}{c} \(l_{2}\) distance in \\ latent space \\ \end{tabular} & 4 & Opt-based & No & No \\ \hline Imprint & \begin{tabular}{c} One-shot \\ module (Shi et al., 2018) \\ \end{tabular} & / & 16384 & Closed-form & Yes & CDF \\ \hline GradViT (Jin et al., 2020) & \(l_{2}\) distance & \begin{tabular}{c} Image prior; \\ Auxiliary Regularization \\ \end{tabular} & 64 & Opt-based & No & \begin{tabular}{c} Auxiliary \\ networks \\ \end{tabular} \\ \hline \hline \end{tabular} \end{table} Table 4. Comparison of reconstruction attacks in FL overhead limits their application in some learning scenarios, particularly in deep neural networks with huge amounts of parameters. As a result, most state-of-the-art privacy-preserving methods are perturbation-based. The perturbation can be accomplished by adding artifact noise into the dataset, such as DP mechanism (Kumar et al., 2017; Zhang et al., 2018). Or by representing the raw dataset with a surrogate dataset (Zhang et al., 2018; Zhang et al., 2018) or abstracting the dataset via sketch techniques (Zhang et al., 2018; Zhang et al., 2018). #### 3.5.1. Cryptographic Approaches Secure multi-party computation is a sub-field of cryptography that executes calculations on data dispersed among multiple parties in such a way that the computation results are only revealed to the participants (Zhang et al., 2018). It can take the form of homomorphic encryption (HE) or secret sharing. As one of the defacto privacy-preserving solutions, homomorphic encryption (HE) provides perfect privacy protection in the face of a malicious server. It allows clients to encrypt their updates in such a way that the server may directly aggregate ciphertexts without divulging anything about the plain text underneath. The downside is that encryption followed by decryption will inevitably impose both a computation and a communications overhead. Phong et al. (Phong et al., 2018) used _additively homomorphic encryption_ to ensure no information was leaked to a malicious server. The encrypted aggregation was formulated as follows: \[\mathbf{E}(\mathbf{W}_{\text{global}}):=\mathbf{E}(\mathbf{W}_{\text{global}}) +\mathbf{E}(-\alpha\cdot\mathbf{G}_{\text{local}}) \tag{4}\] where \(\mathbf{E}\) is a homomorphic encryption operator that supports addition over ciphertexts, \(\mathbf{G}_{\text{local}}\) is the aggregated gradient. The decryption key is public to the clients and private to the server, and thus, the client's information is secured. Due to the additively homomorphic property of \(\mathbf{E}\), each client is still able to receive the correct updated model \(W_{global}\) via decryption. \[\mathbf{E}(\mathbf{W}_{\text{global}})+\mathbf{E}(-\alpha\cdot\mathbf{G}_{ \text{local}})=\mathbf{E}(\mathbf{W}_{\text{global}}-\alpha\cdot\mathbf{G}_{ \text{local}}) \tag{5}\] However, the amount of data transferred between clients and the server is inflated by two orders of magnitude over the vanilla setting (Zhang et al., 2018). To reduce the communication load, Zhang et al. (Zhang et al., 2018) chose a distributed selective stochastic gradient descent (DSSGD) method in the local training phase to achieve distributed encryption and reduce the computation costs. Zhang et al. (Zhang et al., 2018) presented a BatchCrypt as a simple batch encryption technique. Clients first quantize their local gradients and then encode a batch of quantized updates into a long integer. As a result, the communication overhead is reduced by up to 101 times. Jiang et al. (Jiang et al., 2018) further reduced communication overhead by sending only a sparse subset of local states to the server. Another drawback is that all participants share the same private key for decryption since homomorphic operations require all values to be encrypted with the same public kewhich degrades privacy protection in the face of malicious clients. To counter this problem, Park and Lim (Park and Lim, 2018) sketched a privacy-preserving FL scheme based on a distributed homomorphic cryptosystem that allows clients to have their own unique private key for the homomorphic encryption scheme. Secret sharing (Zhang et al., 2018) is another kind of cryptographic technique. It splits a secret data \(\mathcal{D}\) into \(n\) pieces such that the secret \(\mathcal{D}\) can be easily reconstructed with at least \(k\) pieces. However, any set containing less than \(k\) piece reveals no information about \(\mathcal{D}\). It enables the server to aggregate at least a certain number of clients' updates without disclosing any individual client's contribution. Bonawitz et al. (Bonawitz et al., 2018; Zhang et al., 2018) proposed a secure aggregation method for FL based on \(t\)-out-of-\(n\) secret sharing. The key idea is to mask the raw data in a symmetric way. Thus, when aggregated by the server, the introduced noise will be nullified. Liu et al. (Liu et al., 2018) incorporated secure sharing into their federated transfer learning framework to protect privacy. Based on an investigation of how secure aggregation parameters influence communication efficiency, Bonawitz et al. (Bonawitz et al., 2018) used quantization to build a communication-efficient secure aggregation scheme. So et al. (So et al., 2018) designed Turbo-Aggregate that leverages additive secret sharing and Lagrange coding to reduce the secure aggregation overhead. Shao et al. [161] shared a similar ideal, utilizing Lagrange coding to secretly share private datasets among clients. Even though FL based on a homomorphic encryption scheme can prevent privacy leaks during training, it remains vulnerable to attacks in the inference stage. The trained model embodies the distribution of training data to a certain extent and the privacy risk still exists. Model inversion attack gives such an example. Given the white-box access to a trained model, Zhang et al. [225] successfully discovered the sensitive features \(x\) associated with a specific label \(y\). #### 3.5.2. Perturbation Methods Due to its theoretical guarantee of privacy protection and its low computational and communication complexity, the DP technique [42] has emerged as the most popular choice for privacy protection among a variety of options. In differential privacy, a proper amount of noise is added to the raw data [67], the model [60, 118], the output [23], or the gradients [172] to protect privacy. Geyer et al. [60] applied a differentially private mechanism to the FL scenario, where they approximated the averaging operations with a randomized mechanism that provided client-level privacy. Truex et al. [178] presented an alternative approach that draws on both DP and secure multi-party computation. The clients and the server communicate through a secure channel. Upon receiving a request from the server, the clients upload their answers following the principles of DP. Xu et al. [195] took a different approach. They approximated the objective function of a regression problem via polynomial representation and then added Laplace noise to the polynomial coefficients to protect privacy. Khalili et al. [98] exploited an exponential mechanism [130] to privately select applicants based on the qualification scores predicted by a pre-trained model. One concern with perturbation techniques is the trade-off between privacy, accuracy and convergence. A significant noise perfectly protects privacy at the cost of accuracy and convergence. Conversely, a weak noise is futile to privacy attacks [75]. Wei et al. [190] conducted a theoretical analysis of the convergence behavior of FL with DP and identified the trade-off between convergence performance and privacy protection levels. Many studies have been published on ways to deal with this trade-off [48, 165, 223]. Shokri and Shmatikov [165] suggested randomly selecting and sharing a small fraction of gradient elements (those with large magnitudes) to reduce privacy loss. Fan et al. [48] leveraged element-wise adaptive gradient perturbations to defeat reconstruction attacks and maintain high model accuracy. In a similar manner, Wei and Liu [191] used dynamic privacy parameters. Introducing noise with a greater variance at the beginning of training and progressively decreasing the amount of noise and variance as training progresses. Huang et al. [79] proposed _InstaHide_, a combination of cryptographic and perturbation approaches to provide rigorous privacy protection at the cost of minor effects on accuracy. _InstaHide_ encrypts the raw image by mixing it with multiple random images from a large public dataset. After that, it randomly flips the signs of the pixels before using it to train the model. Yang created et al. [201] NISS to avoid the trade-off between accuracy and privacy by permitting clients to collaborate on reducing the total amount of injected noise. In particular, each client's noise is neutralized and distributed to other clients. Theoretically, if all clients are trustworthy, the locally introduced noise can be perfectly offset by the server's aggregation, completely avoiding the privacy accuracy trade-off. A similar idea can be found in Yang et al. [202]. #### 3.5.3. Trusted Execution Environment Some researchers use Trusted Execution Contexts (TEEs) like Intel SGX and ARM TrustZone to secure ML training in untrusted environments [134, 65, 175]. With hardware and software safeguards, TEEs secure critical code from other programs. Compared with purely cryptography methods, TEEs provide much better performance since it only requires extra operations to create the trusted environment and communicate between trusted and untrusted components. Gu et al. [65] partitioned DNN models and solely encased the first layers in an SGX-powered TEE to protect input information. Hynes et al. [80] investigated speeding up the training using Graphics Processing Units (GPU). Tramer et al. [175] shared the same concept and offered effective privacy-preserving neural network inference utilizing trusted hardware that delegated matrix multiplication to an untrusted GPU. However, this work does not translate well to FL due to the possible adversary server and limited computation power of the client device. To remedy this, Mo et al. [134] advocated using the TEE of client devices in tandem with model partitioning to defend against MIA. The model is divided into two halves, and the final layers are calculated within TEE. Kato et al. [97] proposed to combine DP with TEE in FL in the presence of an untrusted server. The models are aggregated within the TEE of the server's device. [28] #### 3.5.4. Discussion Table 5 summarized and compared the existing defense techniques. Cryptographic approaches preserve privacy to a great extent while suffering from computational complexity and are less feasible. The perturbation approaches trade-off privacy for model performance. Several inspiring works demonstrate that it may be possible to avoid that trade-off through either client collaborations to neutralize locally added noise on the server side or by using a surrogate dataset to protect the raw data without adding noise. The cryptographic approaches only ensure that no information will leak during training. They do not protect privacy during the inference stage. In contrast, the perturbation approaches (e.g., DP) protect privacy in both the training and inference stages. One may combine cryptographic and perturbation approaches to obtain better privacy protection throughout the machine learning pipeline. ### Discussion of privacy attacks and defenses in FL This section reviews existing privacy attacks and defense approaches in FL. Table 6 summarized the existing privacy attacks in FL. From the attacker's perspective, FL differentiates from the centralized counterparts in sever aspects: 1) _The active attacker in FL_. Due to the collaboration between clients \begin{table} \begin{tabular}{|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|} \hline **Attack** & \begin{tabular}{c} **Defense** \\ **method** \\ \end{tabular} & \multicolumn{1}{p{56.9pt}|}{**Rationale**} & \multicolumn{1}{p{56.9pt}|}{**Advantage**} & \multicolumn{1}{p{56.9pt}|}{**Disadvantage**} \\ \hline \multirow{8}{*}{RA} & HE [5, 86, 163] & Gradients are encrypted & Accurate & \begin{tabular}{c} 1. Vulnerable if there are multiple colluding entities; \\ 2. Ineffective at inference \\ \end{tabular} \\ \cline{2-5} & Secret sharing [15, 137, 161] & Hiding information about clients’ individual update, except for their sum & 1. Accurate; 2. Robust to users dropping out & Ineffective at inference \\ \cline{2-5} & \begin{tabular}{c} Variational \\ bottleneck \\ [159] \\ \end{tabular} & \begin{tabular}{c} Using surrogate gradient \\ to protect privacy. \\ \end{tabular} & \begin{tabular}{c} Keep training process \\ and performance intact \\ \end{tabular} & \begin{tabular}{c} Limit to optimization-based \\ attack \\ \end{tabular} \\ \cline{2-5} & Gradient compression [115, 173, 231] & \begin{tabular}{c} Compressing gradients to \\ prevent reconstruct private \\ data by matching gradients \\ \end{tabular} & \begin{tabular}{c} 1. Easy to implement; \\ 2. Reduce communication \\ \end{tabular} & \begin{tabular}{c} Requires considerable noise, \\ degrades model performance, \\ and increases convergence \\ time \\ \end{tabular} \\ \hline \multirow{4}{*}{\begin{tabular}{c} RA, \\ MA, \\ and PIA \\ \end{tabular} } & DE [26, 125, 134] & \begin{tabular}{c} Hiding private information \\ by injecting noise to the raw data, model, or output \\ \end{tabular} & \begin{tabular}{c} 1. Easy to implement; \\ 2. Long-term protection \\ \end{tabular} & \begin{tabular}{c} Requires considerable noise, \\ degrades model performance, \\ and increases convergence \\ time \\ \end{tabular} \\ \cline{1-1} \cline{2-5} & TEEs [97, 134] & \begin{tabular}{c} Isolating part of networks \\ from the untrusted environments \\ \end{tabular} & Reduce computation & Limited memory space \\ \hline \end{tabular} * RA: Reconstruction attack; MIA: Membership inference attack; PIA: Property inference attack. \end{table} Table 5. Privacy-preserving methods in FL and the server, an adversary could actively attack for victim's private data. For example, the attacker may maliciously reverse the gradients [140] or mislabel the training sample [75] to neutralize the benign clients' efforts and fool them into revealing more information about their private data. Hence making them more venerable compared to centralized machine learning. 2) _The real-time nature of FL strengthens the attacker's ability._ During the training process, the adversary could adaptive change their strategy to infer the victim's private data. As a result, the adversary can \begin{table} \begin{tabular}{|p{5.6pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|p{56.9pt}|} \hline \hline * BB: black-box; WB: white-box \end{table} Table 6. Comparison of privacy attacks in FL even infer a specific clients' data (Kamiran and Calders, 2017) when the target features appear in FL (Kamiran and Calders, 2017; Kamiran and Calders, 2017), which is way more severe than centralized setting. 3) _Gradients are shared between clients and server_. Unlike the centralized counterpart, where the adversary could at most access the white-box access to the target model. In FL, gradients are repeatedly shared between clients and the server. Which enables gradient-based privacy attacks. As shown by (Kamiran and Calders, 2017; Kamiran and Calders, 2017), the malicious server can reconstruct clients' training data at the pixel level by minimizing the distance to the target gradients. From the defender's perspective, protecting privacy in FL is also different from that in the centralized scenario. 1) _Malicious could be the server or any client_. FL allows clients to keep private data local. A central server is designed to orchestrate the training process. Which complicated privacy protection. The adversary could be the server (Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017) or the client (Kamiran and Calders, 2017). The malicious adversary is able to infer the target client's privacy passively or actively. For example, sending the modified global model to the target client to probe private data (Kamiran and Calders, 2017). This brings challenges to defending against potential privacy attacks. DP is a prevailing choice, but it degrades performance. The cryptographic approaches, like HE, and MPC, retain both privacy and performance at the cost of computation overhead, which is a more severe issue in FL since most clients' devices are limited in computation power. 2) _Training and inference stage privacy attacks_. Different from centralized machine learning, where the major privacy leakage happens at the inference stage, i.e., malicious users probe private training data by inferring the target model. In FL, the attacks could happen during or after the training. This requires the defenders to be aware of both possibilities. The cryptographic approaches proved provable privacy protection during the training stage. However, fail at the inference stage since training distribution is embedded in the trained model's parameters. The perturbation approaches, e.g., DP, provides long-term protection and covers both the training and inference stage. One can hide sensitive information from adversaries by adding appropriate noise to the training data. ## 4. Fairness in FL Fairness, as discussed in the centralized setting, is mainly defined at either the group level (Kamiran and Calders, 2017; Kamiran and Calders, 2017) or the individual level (Kamiran and Calders, 2017). In the FL scenario, fairness has a broader definition. Beyond the long-established _algorithmic fairness_(Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017) arises as a new challenge in FL. ### Algorithmic Fairness Algorithmic fairness is commonly used to describe the discrepancies in algorithm decisions made across distinct groups as defined by a sensitive attribute. FL often involves a deep neural network with redundant parameters and is pruned to overfit the privileged groups. Various debiasing methods have been devised for different applications, including machine learning (Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017), representation learning (Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017), and natural language processing (Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017). These methods vary in detail but share similar principles. Following the data flow, debiasing methods can be grouped into _pre-processing_, _in-processing_ and _post-processing_ categories, which address the discriminate issues at three distinct stages of the data's handling (Kamiran and Calders, 2017). #### 4.1.1. Pre-processing Pre-processing tries to remove the underlying discrimination from the data typically by 1) altering the values of the sensitive attributes/class labels; 2) mapping the training data to a new space where the sensitive attributes and class labels are no longer relevant (Kamiran and Calders, 2017; Kamiran and Calders, 2017; Kamiran and Calders, 2017); or 3) reweighting the samples in the training dataset to compensate for skewed treatment (Kamiran and Calders, 2017). Intuitively, by training a classifier on discrimination-free data, it is likely that the resulting predictions will be discrimination-free. Inspired by this idea, Kamiran and Calders (Kamiran and Calders, 2017) proposed three types of pre-processing solutions to learn a fair classification, _messaging, reweighing and sampling_. Feldman et al. (2019) investigated the problem of identifying and removing disparate impacts in the data. Xu et al. (2019) proposed FairGAN, which generates fair data from the original training data and uses the generated data to train the model. Abay et al. (2019) proposed two reweighting methods for the FL setting (Kirshman et al., 2017), _local reweighing_ and _global reweighing with DP_. Notably, these pre-processing techniques require access to the training data, which violates the privacy principles of FL. As a result, these types of techniques can only be deployed locally on each client. However, in the presence of data heterogeneous among clients, local debiasing cannot provide fair performance for an entire population (Zhu et al., 2019). #### 4.1.2. In-processing In-processing modifies traditional learning algorithms to address discrimination (Berk et al., 2011; Krizshman et al., 2017; Krizshman et al., 2018; Krizshman et al., 2019; Krizshman et al., 2019). Such as adding a regularization term to the loss function. Berk et al. (2011), for example, incorporated a family of fairness regularizers into the objective function for regression problems. These regularizers span the range from notions of group fairness to individual fairness. They also create a trade-off between accuracy and fairness. Another in-processing option is imposing constraints. Zhang et al. (2014) used a GAN to constrain the bias in a model trained on biased data. During training, the scheme simultaneously tries to maximize the accuracy of the predictor while minimizing the ability of the adversary to predict the protected variable. In FL, G'alvez et al. (2019) studied the notion of group fairness as an optimization problem with fairness constraints. Papadaki et al. (2019) formulated a min-max optimization problem to investigate group fairness in scenarios where population data were distributed across clients. Ezzeldin et al. (2019) replaced the aggregation protocol FedAvg with FairFed, which adaptively updates the aggregating weights in each round to improve group fairness. Clients whose local measurements match the global fairness measure are given preferential treatment. Khedr et al. (2019) add a regularizer term to minimize the average loss in fairness across all training data. #### 4.1.3. Post-processing Post-processing addresses discrimination issues after the model is trained and doesn't need to change the training process. The general methodology of post-processing algorithms is to take a subset of samples and change their predicted labels to meet a group fairness requirement (Berk et al., 2011; Krizshman et al., 2018; Krizshman et al., 2019; Krizshman et al., 2019; Krizshman et al., 2019). Hardt et al. (2019) proposed a post-processing technique to construct a non-discriminating predictor \(\tilde{Y}\) from a learned discriminatory binary predictor \(\hat{Y}\). Only access to the prediction \(\hat{Y}\), the protected attribute \(A\) and target label \(Y\) in the data are required, while details of the mapping of features \(X\) to prediction \(\hat{Y}\) are not needed. Canetti et al. (2018) and Pleiss et al. (2019) shared the key characteristics as Hardt et al.'s (2019) work. Lohia et al. (2019) designed a post-processing method to increase both individual and group fairness. Salvador et al. (2019) introduced a conditional calibration method for fair face verification. Their method clusters images into different sets and assigns distinct thresholds to different sets. #### 4.1.4. Discussion Three different kinds of debiasing methods are at hand in centralized machine learning. However, solutions in the centralized setting cannot be applied directly in the FL scenario due to limitations with the training data. More specifically, in federated settings, the clients usually have limited amounts of data. Hence, a single client can't accurately represent the true distribution over all clients. Consequently, debiasing data before training is not an option. Another limitation is that direct access to local data is prohibited on the server side. Nevertheless, canny researchers have found inspiration from and workarounds to these issues. Galvez et al. (2019), for example, bypassed this access restriction by using statistics to guide the model's training instead of the raw data. ### Client Fairness Client fairness in FL is another different fairness notion than algorithmic notions. Ideally, the models produced from FL should capture clients' data distributions and generalize well when deployed on the client side. However, data distribution usually varies among clients. As a result, the global model has inconsistent performance on different clients' dataset. At the client level, a FL protocol is considered to be fair if the performance fluctuates within a limited range, i.e., the variance in the model's performance across clients falls under a predefined threshold. To this end, two lines of research exist to mitigate fairness issues in FL. These are the _single model approach_ and the _personalized models approach._ #### 4.2.1. Single Model Approach The single model approach trains a single global model for all clients as a standard FL scheme. Here, the focus is on solving any statistical heterogeneity during the training phase rather than smoothing the distribution difference. * **Data augmentation** is a straightforward solution to statistical heterogeneity. It increases data diversity on the client side. Several researchers have studied ways to enhance the statistical homogeneity of local data in FL (Zhao et al., 2018; Wang et al., 2019; Zhao et al., 2020). Zhao et al. (2020) suggested a data share scheme, which creates a globally-shared dataset that is balanced by class. The experiment shows a 30% improvement on accuracy with only 5% globally shared data. Jeong et al. (2020) proposed _FAug_. Clients first collectively train a GAN model, which is then distributed to clients to augment their local data towards yielding an i.i.d dataset. * **Client Selection** is another strategy that focuses on sampling data from a homogeneous distribution. Wang et al. (2019) proposed a control framework to actively select the best subset of clients in each training round. In Yang et al.'s (2020) method, the local data distribution is estimated first by comparing local updated gradients and gradients inferred from a balanced proxy dataset. The client selection algorithm based on a combinatorial multi-armed bandit was designed to minimize the effect of class imbalances. * **Agnostic approach** trains a robust model against a possible unknown testing distribution. Mohri et al. (2019) modeled testing distributions as an unknown mixture of all \(m\) clients' data. The global model is optimized for all possible target distributions. This makes the global model more robust to an unknown testing distribution. Du et al. (2020) introduced a fairness constraint into Mohri et al.'s method (Mohri et al., 2019) and proposed _AgnosticFair_, a fairness-aware FL framework. Their method can provide both _Good-intent fairness_ and _demographic parity_ * **Reweighting** tries to train a fair model by assigning suitable aggregating weights \(p_{k}\) in Eq. 1 to clients. Inspired by \(\alpha\)-fairness notions (Li et al., 2019; Li et al., 2019), Li et al. (2019) sketched _q-Fair FL_ (_q_-FFL) to foster fairer accuracy distribution across all clients by up-weighing clients with lower performance during aggregation. Huang et al. (2020) shared a similar idea where, for each round of aggregation, clients with lower accuracy or less training participant times are assigned higher aggregation weights. #### 4.2.2. Personalized Models Approach Instead of smoothing the statistical heterogeneity, in _personalized FL_, multiple distinct models are trained for clients with different data distributions. A global model is first trained collaboratively and then personalized to clients using private data. In this way, clients can benefit from other clients' data and solve the issue of statistical heterogeneity. Mansour et al. (2019) designed and analyzed three approaches to learning personalized models to learn personalized models. Kulkarni et al. (2019) conducted a brief overview of personalized FL. Chen et al. (2020) provided a comprehensive benchmark of various personalized FL methods. Tan et al. (2020) systematically reviewed this topic and classified personalized FL techniques in terms of data-based and model-based approaches. Here, we summarize their conclusions. * **Multi-task learning** treats building models for each client as different tasks. Smith et al. [168] pioneered this approach and explored personalized FL via a multi-task learning framework. [3, 116] followed this principle. Dinh et al. [38] proposed FedU, which incorporates a Laplacian regularization term into the optimization problem to leverage relationships between clients. * **Model interpolation** trains local and global models simultaneously, where the global model is used for its generalization ability, and the local model is used to improve local performance. Hanzely and Richtarik [70] formulated an optimization problem that learns a mixture of the global and local models. The local model is trained solely on each client's private data. Softly-enforced similarity from multi-task learning is borrowed to discourage the local model from departing too much from the mean model. Deng et al. [36] and Mansour et al. [126] adopt a similar formulation to determine the optimal interpolation of the local and global models. In Zhang et al.'s [222] work, clients are given access to multiple models uploaded by other clients to evaluate how much they will benefit from these models. An optimal combination is then used as a personal update. Lin et al. [117] investigated the trade-offs between local and global models. * **Parameter decoupling** learns local parameters as an independent task performed locally. The local model is designed to assist in personalizing the global model to local distributions. Liang et al. [116] devised the local-global federated averaging algorithm, which jointly learns compact local representations for each client and a global model across all devices. Chen and Chao [25] decomposed a FL model as a generic predictor, which is trained globally, along with a personalized predictor that is trained locally. The personalized predictor is formulated as a lightweight, adaptive module on top of the generic predictor. * **Transfer learning** is a practical training paradigm that leverages knowledge from a source domain to help train a model in a target domain. The performance of transfer learning depends on the similarity between the two domains. Federated transfer learning was first introduced by Liu et al. [119]. Since clients in the same federation usually share the same domain, a FL scheme would make a suitable partner for transfer learning. Li and Wang [110] subsequently proposed FedMD, which combines transfer learning and knowledge distillation. Each client performs transfer learning by training a model to converge on a public dataset and subsequently fine-tune it on local data. * **Clustering** arranges clients into different groups and trains a specific model for each group. Ghosh et al. [61] iteratively determines the membership of each client to a cluster and optimizes each of the cluster models via gradient descent in a distributed setting. Sattler et al. [158] clusters clients according to the cosine similarity between the clients' gradient updates. This allows clients with a similar distribution to profit from one another while minimizing detrimental interference from others. In Briggs et al.'s [18] method, a clustering step is periodically inserted into the training process to cluster clients based on their local updates. The clusters are then trained individually and in parallel on specialized models. Mansour et al. [126] proposed hypothesis-based clustering, partitioning clients into \(q\) clusters and finding the best hypothesis for each cluster. * **Regularization** prevents overfitting when training models and has been used in several studies to remedy the weight divergence problem in FL settings. Li et al. [113] introduced a proximal term that considers the differences between global and local models to limit the effect of local updates. Yao et al. [205] considered parameter importance in the regularised local loss function by using elastic weight consolidation [102]. In addition, a regularization term is introduced to penalize the deviation of the local model from the global model. * **Meta-learning** aims to leverage prior experience with other tasks to facilitate the learning process. The resulting models are highly-adaptable to new heterogeneous tasks [51, 141]. Fallah et al. (Fallah et al., 2017) studied a personalized variant of FedAvg based on model-agnostic meta-learning formulation. The proposed Per-FedAvg algorithm looks for an initial model that performs well after one step of the local gradient update on each client's data. Others have interpreted FedAvg as a meta-learning algorithm, breaking it into two stages of training and fine-tuning to optimize personalized performance and model convergence (Fallah et al., 2017; Fallah et al., 2018). #### 4.2.3. Discussion In addition to algorithmic fairness, client fairness is another concern in the FL community. Table 7 enumerated various works on these two topics. Regarding client fairness, the single model approach focuses on smoothing data heterogeneity, where it is easy to implement and can be added to the general FL paradigm since it only needs modest modification. On the downside, the single model approach is less effective than personalized approaches in terms of capturing local data distribution and may be insufficient when the data distributions vary significantly between clients. Additionally, the single-model approach does not allow clients to customize their models. ### Discussion of fairness in FL There are two definitions of fairness in FL, _client fairness_ and _algorithmic fairness_. _Algorithmic fairness_ has been extensively studied in centralized machine learning. These algorithms presuppose centralized access to data, however, one virtue of FL is data never leaves the device. This means neither the server nor any client gains centralized access to the training data. Therefore, generalizing the fair learning algorithms to FL is not trivial. On the one hand, data is stored locally in FL. The server cannot directly access the local data of clients. Hence, server-side debiasing is not a viable solution. On the other hand, debiasing on the client side is ineffective due to the inadequate data, which can hardly represent the global data distribution (Krishnan et al., 2018). There is no guarantee that model debiased with local data will generalize to the global distribution. The non-i.i.d data distributions further complicated this problem (Krishnan et al., 2018). \begin{table} \begin{tabular}{c c c c c c} \hline \hline **Reference** & **Single** & **Personalized** & **Algorithmic** & **Client** & **Method** \\ & **Model** & **Model** & **Fairness** & **Fairness** & **Fairness** \\ \hline (Fallah et al., 2017; Fallah et al., 2018; Krishnan et al., 2018) & ✓ & & & ✓ & Data Augmentation \\ (Fallah et al., 2018; Fallah et al., 2018; Krishnan et al., 2018) & ✓ & & & ✓ & Client Selection \\ (Fallah et al., 2018) & ✓ & & ✓ & & Agnostic approach \\ (Fallah et al., 2018) & ✓ & & ✓ & ✓ & Agnostic approach \\ (Fallah et al., 2018; Krishnan et al., 2018) & ✓ & & & ✓ & Agnostic approach \\ (Fallah et al., 2018) & ✓ & & ✓ & & Reweight \\ (Fallah et al., 2018; Fallah et al., 2018) & ✓ & & & ✓ & Regularization \\ (Fallah et al., 2018; Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Cluster \\ (Fallah et al., 2018; Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Model interpolation \\ (Fallah et al., 2018; Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Multi-task learning \\ (Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Parameter decoupling \\ (Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Transfer learning \\ (Fallah et al., 2018; Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Regularization \\ (Fallah et al., 2018; Fallah et al., 2018; Fallah et al., 2018) & ✓ & & ✓ & Meta-learning \\ \hline \hline \end{tabular} \end{table} Table 7. Summary of Fairness-aware FL _Client fairness_ is tailored to FL and stems from the non-i.i.d data. Each client sampled the training data from a distinct distribution. In this case, the vanilla FL protocol, _FedAvg_, fails to train a model to fits clients' data distribution. Various methods have been proposed to alleviate this. From the data aspect, (Kal the non-private alternative at the cost of communication overhead. The trade-offs between privacy and efficiency are the main concern in this category. However, as argued by [81], the cryptographic approach does not guarantee privacy at the inference stage. It only ensures the training data remain private during training and cannot prevent the adversary from inferring training samples from the neural network parameters [170; 226]. On the contrary, DP guarantees a fair model will not leak anything beyond what could be carried out from "population level" correlations. As such, the majority of works focus on learning fair, and DP model [9; 34; 81; 196]. Thus, this subsection focus on DP as the privacy-preserving techniques. #### 5.1.1. Empirical Findings The impact of privacy on fairness was initially observed in empirical studies. Bagdasaryan et al. [9] first observed that the reduction in accuracy caused by deep DP \begin{table} \begin{tabular}{c c c c c c} \hline \hline \multirow{2}{*}{**Reference**} & **Privacy** & **Fairness** & \multicolumn{2}{c}{**Techniques to achieve**} & **Trade-off** \\ & **notion** & **notion** & **Privacy** & **Fairness** & **type** \\ \hline [34] & \(\epsilon\)-DP & \(\alpha\)-Discrimination & Exponential mechanism & Minimize discrimination scores & I \\ \hline [196] & \(\epsilon\)-DP & Decision boundary fairness & Functional mechanism & Fairness constraints & I \\ \hline [138] & \(\epsilon\)-DP & \(\alpha\)-Equal opportunity & Local DP & Post-processing & I \\ \hline [107] & \(\epsilon\)-DP & Equal odds \& Class conditional mechanism & Fairness constraints & I \\ \hline [37] & \(\epsilon\)-DP \& Decision boundary fairness & Functional mechanism mechanism & Fairness constraints & I \\ \hline [81] & \((\epsilon,\delta)\)-DP & \(\alpha\)-Equal opportunity & Exponential mechanism & Fairness constraints & / \\ & & & \& Laplace noise & \\ \hline [122] & \((\epsilon,\delta)\)-DP & Equal odds \& Demographic parity & DP-SGDA & ERMI regularizer & II \\ \hline [45] & \((\epsilon,\delta)\)-DP & Excessive risk gap & DPSGD-Global-Adapt Gradient correction & II \\ \hline [176] & \((\alpha,\epsilon_{p})\)- & Equal odds, Accuracy parity & DP-SGD & Fairness constraints & II \\ & Rényi DP & \& Demographic parity & & \\ \hline [101] & / & Equal accuracy & MPC & Fairness constraints & II \\ \hline [66] & / & Equal opportunity & Proxy attribute & Post-processing & II \\ \hline [186] & / & Demographic parity & Noisy attribute & Fairness constraints & II \\ \hline [8] & / & Equal odds & Noisy attribute & Post-processing & II \\ \hline \hline \end{tabular} * I: Trade fairness for privacy. Relaxing fairness notions to achieve purely DP * II: Trade privacy for fairness. Adopting relaxed DP notion to accommodate exact fairness \end{table} Table 8. Private and Fair Learning models negatively impacts underrepresented subgroups disproportionately. DP-SGD strengthens the model's "bias" toward the most prominent features of the distribution that is being learned. Kuppam et al. (2015) reached a similar conclusion when they examined the effects of DP on fairness in three real-world tasks involving sensitive public data. When the noise added by a private algorithm is negligible in relation to the underlying statistics, the costs of adopting a private technique may be minor. When stronger privacy is implemented or when a task entails a small population, significant disparities may emerge. Farrand et al. (2019) demonstrated that even minor differences and weak privacy protections could result in disparate outcomes. Ganev et al. (2018) shifted the emphasis to generative models and tabular synthetic data. Three DP generative models PrivBayes (2018), DP-WGAN (Beng et al., 2019), and PATE-GAN (Petersson et al., 2019), were involved. They witnessed a disparate effect on the accuracy of classifiers trained on synthetic data generated by all generative models. The losses are greater and/or more dispersed for underrepresented groups. Uniyal et al. (2018) compared DP-SGD and Private Aggregation of Teacher Ensembles (PATE) (Farrand et al., 2019), an alternative DP mechanism for discrerely training a deep neural network, in terms of fairness. They discovered that PATE has a disparate effect, but it is considerably less severe than DP-SGD. #### 5.1.2. Theoretical Explanations Several works have attempted to determine the mechanism underlying the well-known relationship between privacy and unfairness. Bagdasaryan et al. (2019) associated the impact with the gradient clipping operation in DP-SGD. During training, the model generates larger gradients for samples from underrepresented subgroups; consequently, clipping slows their learning rate. Therefore, the model learns less from the underrepresented subgroups, and its performance on those subgroups is negatively impacted more. Tran et al. (2019) conducted an in-depth study into this phenomenon with output perturbation (Tran et al., 2019) and DP-SGD as the private mechanism. By measuring fairness with _excessive risk gap_, Tran et al. proved that output perturbation mechanisms incur unfairness when the local curvatures of the loss functions of different groups differ substantially. For DP-SGD, Tran et al. found that the clipping bound, the norm of inputs, and the group's distance to the decision boundary collectively contributed to the unfairness raised by DP-SGD. Esipova et al. (2019) examined the same issue from the gradient perspective. They proved that the gradient misalignment caused by DP-SGD is the main reason for unfairness. If the clipping operation disproportionately and sufficiently increases the direction error for group \(a\) relative to group \(b\), then group \(a\) incurs larger excessive risk due to gradient misalignment. #### 5.1.3. Mitigation Strategies Diverse methods have been proposed to mitigate the effect of private mechanisms on fairness. Xu et al. (2019) proposed DP-SGD-F, a variant of DP-SGD that reduces the divergent impact on different populations. By adaptively designating clipping bounds for each group, DP-SGD-F achieves a level of privacy proportional to each group's utility-privacy trade-off. For the group whose clipping bias is greater (due to large gradients), a larger clipping bound is adopted to mitigate for their greater privacy cost. Tran et al. (2019) formulated a regularized optimization problem that minimizes empirical loss while satisfying two additional constraints. The first constraint equalizes the averaged non-private and private gradients, while the second constraint penalizes the difference between the local curvatures of distinct groups' loss functions. Esipova et al. (2019) modified DP-SGD and developed DP-SGD-Global-Adapt to preserve gradient direction. It is assumed that a hyperparameter \(Z\) is the upper bound for most gradients. Gradients less than \(Z\) are uniformly scaled, whereas gradients greater than \(Z\) are trimmed to \(Z\). ### Fairness Increases Privacy Risk Fairness in turn presents challenges for privacy mechanisms. Chang and Shokri (Chang and Shokri, 2018) observed an elevated privacy risk for underprivileged subgroups in a fair model. In order to ensure fairness, the model must perform equally well for all subgroups. However, limited data availability for underprivileged subgroups can lead to overfitting of the training data for unprivileged subgroups in a fair model, thereby increasing the privacy risk. Previous works on fairness-aware machine learning often assume that the sensitive features are reliable and accessible. This assumption unavoidably introduces privacy risks. However, achieving precise notions of fairness, such as demographic parity, becomes unattainable without access to sensitive attributes, specifically the membership information of sensitive groups. To address this issue, several techniques have been proposed to safeguard the privacy of sensitive attributes during the training of fair models. #### 5.2.1. Training Fair Models with Noisy Representation Researchers in this field train approximately fair models using noisy sensitive attributes to protect privacy. Gupta et al. (Gupta et al., 2019) substituted protected groups with proxy groups. To achieve fairness, the proxy groups need to align with the true positive group and even overlap with the ground-truth groups. Thus, the fairness guarantee comes at the cost of privacy. Several studies have explored fairness with imperfect group information. (Lamy et al., 2019; Padala et al., 2020; Padala et al., 2020; Padala et al., 2020). Lamy et al. (Lamy et al., 2020) introduced a mutual contaminated model to simulate a noisy distribution with corrupted attributes. Under this framework, they demonstrated that the fairness constraint on the clean distribution is equivalent to a scaled fairness constraint on the noisy distribution. To protect the privacy of sensitive attributes, they added class conditional noise to release the noisy dataset. Awasthi et al. (Awasthi et al., 2020) addressed the challenging problem of training a fair model with perturbed sensitive attribute values, where each attribute is independently flipped to its complementary value with probability \(\gamma\). They identified conditions on the perturbation under which the classifier, denoted as \(\hat{Y}\), obtained by Hardt et al.'s method (Hardt et al., 2019), is fairer than the vanilla classifier, denoted as \(\tilde{Y}\), trained on accurate attributes. They further provided a formal guarantee of effectiveness under the necessary conditions. Wang et al. (Wang et al., 2020) trained a fair binary classifier based on a noisy label \(\hat{G}\in\{1,...,\hat{m}\}\), i.e., \(\hat{G}\) could be _"country of residence"_ as a noisy representation of the true group labels \(G=\)_"language spoken at home"_. #### 5.2.2. Training Fair Models with DP Works in this area protect privacy by adding noise to the private characteristic (Yang et al., 2020). The trade-off between privacy and fairness depends on the amount of noise added, with no noise and excessive noise representing the two extremes. In the case of no noise, the model's performance remains unaffected but could lead to information breaches. Conversely, high levels of noise are effective in preserving privacy but can compromise the model's utility. Tran et al. (Tran et al., 2020) proposed a constrained optimization problem to address both private and fair learning tasks. Their framework ensures \((\alpha,\epsilon_{p})\)-Renyi DP (Jagielski et al., 2020) for the sensitive attributes by solving the constrained problem with DP-SGD. Jagielski et al. (2020) extended Agarwal et al.'s approach (Agarwal et al., 2020) by incorporating privacy considerations. They formulated a two-player zero-sum game, played between a "learner" and an "auditor," to derive a fair classifier. Laplacian noise (Zavak et al., 2019) and the exponential mechanism (Jagielski et al., 2020) were utilized separately for the "learner" and the "auditor". As a result, the learned model satisfies \((\epsilon,\delta)\)-DP and achieves equalized odds. ### Fair and Private FL In centralized machine learning, one entails centralized access to training data (either the true data or noisy data). However, this is invalid in FL, where neither the server nor clients have access to others' data. Therefore, one cannot simply apply centralized fair learning algorithms in FL tasks. This raises a question: _How can we promote algorithmic fairness in FL without accessing clients' data in FL_? Several studies made progress in response to this challenge. **Using a surrogate model to preserve privacy**. Padala et al. (Padala et al., 2020) tried to satisfy both \((\epsilon,\delta)\)-local DP and demographic fairness through a fair and private FL framework. To circumvent the access restriction, they decomposed the learning into two phases. First, each client learns a fair and accurate model on a local dataset, where the fairness constraint acts as a regularization term in the loss function. Then, every client trains a surrogate model to match the fair predictions from the first model with a DP guarantee. Finally, only the surrogate model is communicated to the server. **Privacy through secure aggregation**. Zhang et al. (Zhang et al., 2016) investigated classification problems in FL through multiple goal optimization problems with privacy constraints. The objective is to minimize the accuracy loss and the discrimination risk. To this end, a team Markov game was designed to select participating clients at each communication round. In each round, clients decide whether or not to participate based on the global model's state, which is characterized by bias level and accuracy. Further, a secure aggregation protocol is designed to estimate the global model's status based on polynomial interpolation (Zhou et al., 2017) for privacy concerns. Under this protocol, the server is able to calculate the discrimination status without accessing the local data. **Achieve fairness based on statistics**. Galvez et al. (Galvez et al., 2016) formulated a constrained optimization problem that is solved by the differential multiplier. Local statistics are provided to the server for debiasing the global model. To further protect privacy, client updates are clipped and perturbed by Gaussian noise before being sent to the server. Finally, their solution is able to provide the approximate group fairness notion over multiple attributes and \((\epsilon,\delta)\)-DP. **Fairness through agnostic learning**. Shifts in distribution is one source of bias in FL. The global model is trained on the data of all clients (source distribution), but each client's local data distribution (target distribution) may differ. When deployed to the client, unfavorable outcomes occur. Du et al. (Du et al., 2018) proposed treating the client data distribution in an agnostic way. An adversary generates any possible unknown local data distribution to maximize the loss, while the learner aims to optimize the accuracy and fairness. **Calculate fairness violations locally**. Chu et al. (Chu et al., 2018) formulated a constraint optimization problem to learn a fair and private model in FL. Each client locally calculates fairness violations to avoid impinging on the data privacy of any client. Chu et al. (Chu et al., 2018) further optimized this method by aggregating fairness constraints to better estimate the true fairness violation for all data. Although some fair FL algorithms do not directly access the training data (Chu et al., 2018; Du et al., 2018), faithfully sharing the model/gradients in FL could incur privacy leakage risks. The privacy breach could happen during the training or inference stage. The attack could be carried out by either the server or the clients (Zhou et al., 2017). For example, an honest-but-curious server can lunch a reconstruction attack (Zhou et al., 2017) to recover the private data from the gradients uploaded by the victim client. However, the main challenge to training a fair model in FL is restricted data access, e.g., data never leaving local devices, which is an under-investigated topic in FL literature. In the case of the adversary clients/server in FL, some privacy-preserving techniques, such as DP, can be combined with the aforementioned fair FL approaches to prevent privacy leakage. ### Discussion of Privacy and Fairness Interactions The complex interactions between privacy and fairness have been thoroughly examined and documented in various studies. These investigations highlight the intricate trade-offs and challenges that arise when attempting to simultaneously address both privacy and fairness objectives (Zhou et al., 2017). The impact of privacy and fairness on each other is indeed bilateral. In one scenario, privacy measures can degrade fairness. For instance, in widely-used privacy mechanisms like DP-SGD, to protect privacy, the algorithm clips and adds noise to the gradients. However, due to the scarcity of data for certain groups, these modifications can disproportionately affect underrepresented groups, exacerbating unfairness. Therefore, the implementation of DP can inadvertently worsen existing unfairness by disproportionately impacting certain groups. In another case, fairness can increase privacy risks. To achieve fairness, it may be necessary to collect additional demographic information about users, even if it is irrelevant to the task at hand. This data collection is aimed at guiding modifications to the model, such as addressing inconsistent responses or removing discrimination in statistical models (Kalalal and Triggs, 2011; Kalal and Triggs, 2012; Kalalal and Triggs, 2013). However, the collection of such sensitive information raises privacy concerns, as it expands the scope of data being collected and potentially increases the risk of privacy breaches. In the context of Federated Learning (FL), the cooperative game between clients and the server adds complexity to the privacy and fairness challenges. FL introduces new privacy attack surfaces, as discussed in Section 3, where potential malicious participants can actively or passively infer the private data of other clients. Consequently, securing private information in FL requires even stronger privacy protection measures compared to the centralized setting. Merely protecting group membership is insufficient to address the privacy risks in FL. Furthermore, the non-i.i.d. (non-independent and identically distributed) nature of FL poses another challenge. In a typical FL system, clients' data are sampled from different distributions, leading to data heterogeneity. A model that achieves fairness within the local distribution of each client is not guaranteed to perform unbiasedly on a global scale. The non-i.i.d. issue also introduces potential fairness concerns at the client level, as the performance of the model can vary significantly among clients. It is crucial to address this variation and ensure fairness across all participating clients in FL. The challenge lies in training a fair model in FL without violating the data access restrictions imposed by each client. Finding methods to mitigate the fairness issues arising from the non-i.i.d. nature of the data while respecting the privacy and data access constraints in FL remains a challenging task. ## 6. Open research directions The research community has made fruitful progress in privacy and fairness in FL. However, throughout this survey, we found this field still faces several challenges that need to be solved. * **Trade-offs between Privacy and Fairness**. The interaction between privacy and fairness is an under-studied topic. Existing works have focused on exploring the two notions in isolation, either focused on privacy-preserving machine learning (Kalalal and Triggs, 2011) or on paradigms that respect fairness (Kalalal and Triggs, 2013). However, as demonstrated by several studies (Kalalal and Triggs, 2013; Kalalal and Triggs, 2013), privacy and fairness may compete with each other. In the realm of FL, challenges and opportunities coexist. On the one hand, restricted information and non-i.i.d distribution complicate the problem settings. On the other hand, the flexibility of the FL paradigm may enable more possible solutions. For instance, the personalized model (Kalalal and Triggs, 2011; Kalalal and Triggs, 2013) has been widely used in FL to address statistical challenges by assigning clients personalized models. We may combine privacy and personalized models to achieve a better trade-off between privacy, fairness, and utility. Thus, we believe it is worth examining the trade-offs between privacy and fairness in FL. * **The Compatibility of Fairness and DP**. We believe it would be worth investigating techniques that simultaneously accommodate fairness and DP. As pointed out in Dwork et al.'s (Dwork et al., 2012) work, given a carefully designed distance metric, it is possible to achieve individual fairness through \(\epsilon\)-DP. Two characteristics of FL make individual fairness a superior choice over group fairness: 1) Data distribution in FL may vary significantly between clients, and individual fairness is more suitable in such cases. Since it is defined at the sample level, thus, it generates better than group notions when addressing new samples which may be distinct from those in the training set; 2) The restricted access to information in FL lends itself more to individual fairness because individual fairness relies on a Lipschitz continual prediction model and does not require access to demographic data. This perfectly fits the FL setting. * **How can one satisfy fairness at both the algorithm and client levels in FL?** The majority of studies on fairness in FL focus on promoting fairness at the client level. However, client-level fairness does not necessarily imply algorithmic fairness. Consider a scenario where multiple companies (clients) collaborate to train a credit card approval model. Consumer demographic compositions vary between each company. Although a federated model trained subject to client-level fairness constraints might handle the different companies fairly, the model could still be biased towards sensitive attributes (such as race or educational background). This raises a question: _How can one satisfy fairness at both the algorithm and the client levels while preserving privacy in FL?_. ## 7. Conclusion In this article, we conducted a detailed survey of data privacy and model fairness issues in FL. Uniquely, we also documented the interactions between privacy and fairness from the perspective of trade-offs. In terms of privacy in FL, we first reviewed privacy attacks in FL. Then, we presented three kinds of privacy-preserving techniques. Regarding fairness, we first analyzed the possible sources of bias and how bias can be introduced on both the client and server sides. Following a review of the notions of fairness adopted in machine learning and those originating from FL, a discussion of the various fairness-aware FL algorithms is presented. The last part of the survey focused on the interactions between privacy and fairness. We identified three relations in the general context and further listed possible solutions to achieve both fair and private FL. ## Acknowledgments This paper is supported by the Australian Research Council Discovery DP200100946 and DP230100246, and NSF under grants III-1763325, III-1909323,III-2106758, and SaTC-1930941.

2307.03776

Double-$Q$ spin chirality stripes in the anomalous Hall antiferromagnet CoNb$_3$S$_6$

The metallic antiferromagnet CoNb$_3$S$_6$ exhibits a giant anomalous Hall effect (AHE) that cannot be explained by a collinear N\'eel order on intercalated Co ions. Thus, a noncoplanar structure is expected. We carried out resonant elastic x-ray scattering (REXS) to reexamine the magnetic structure of CoNb$_3$S$_6$ and found a double-$Q$ ($2Q$) order with a $(\frac{1}{2}00)$ commensurate component and a long-wavelength modulation. Circular dichroism and linear polarization analysis reveal that the commensurate components on the two Co sites are noncollinear and the modulation is helical. The resulting magnetic structure has a staggered scalar spin chirality forming a stripe pattern in real space. Furthermore, we found that the helical modulation wavevector exhibits a sample dependence and develops a low-symmetry domain structure. We propose that quenched-in lattice strain controls the helical domain structure, accounting for much of the sample dependence. These results provide insight into the mechanism of the AHE in CoNb$_3$S$_6$ and identifies potential routes for controlling the Hall response and realizing other unconventional electronic phenomena in metallic antiferromagnets.

Ben Zager, Raymond Fan, Paul Steadman, Kemp Plumb

2023-07-07T18:00:12Z

http://arxiv.org/abs/2307.03776v1

# Double-\(Q\) spin chirality stripes in the anomalous Hall antiferromagnet CoNb\({}_{3}\)S\({}_{6}\) ###### Abstract The metallic antiferromagnet CoNb\({}_{3}\)S\({}_{6}\) exhibits a giant anomalous Hall effect (AHE) that cannot be explained by a collinear Neel order on intercalated Co ions. Thus, a noncoplanar structure is expected. We carried out resonant elastic x-ray scattering (REXS) to reexamine the magnetic structure of CoNb\({}_{3}\)S\({}_{6}\) and found a double-\(Q\) (\(2Q\)) order with a \((\frac{1}{2}00)\) commensurate component and a long-wavelength modulation. Circular dichroism and linear polarization analysis reveal that the commensurate components on the two Co sites are noncollinear and the modulation is helical. The resulting magnetic structure has a staggered scalar spin chirality forming a stripe pattern in real space. Furthermore, we found that the helical modulation wavevector exhibits a sample dependence and develops a low-symmetry domain structure. We propose that quenched-in lattice strain controls the helical domain structure, accounting for much of the sample dependence. These results provide insight into the mechanism of the AHE in CoNb\({}_{3}\)S\({}_{6}\) and identifies potential routes for controlling the Hall response and realizing other unconventional electronic phenomena in metallic antiferromagnets. Materials with complex magnetic phases beyond traditional ferro- and antiferromagnetism exhibit a diverse range of phenomena that are both fundamentally rich and offer many potential applications as next-generation electronic and spintronic devices. Such phases include noncoplanar, chiral, and topological spin textures [1; 2], altermagnetism [3; 4], multiferroics [5], and multipolar magnetism [6]. In these materials, the intricate magnetic symmetries allow for the coupling between charge, magnetic, and lattice degrees of freedom from which effective composite degrees of freedom emerge and give rise to novel macroscopic response. Transition metal dichalcogenides intercalated with \(3d\) transition metal ions form a class of materials where such complex magnetic phases are stabilized through an interplay between localized spins on the \(3d\) sites and itinerant electrons in the host layers [7; 8]. Diverse phenomena are possible depending on the host compound, intercalation species, and intercalation ratio. Co-intercalated NbS\({}_{2}\), CoNb\({}_{3}\)S\({}_{6}\), is of particular interest because it exhibits a giant anomalous Hall effect (AHE) that cannot be explained by its reported collinear antiferromagnetic structure [9; 10; 11]. A series of neutron diffraction measurements have found the symmetry-related magnetic propagation vectors \((\frac{1}{2}00)\), \((0\frac{1}{2}0)\), and \((\frac{1}{2}\frac{1}{2}0)\), but disagree on the orientation of the moments and the presence of single-\(Q\) domains or multi-\(Q\) order [10; 11; 12; 13; 14]. Elucidating the precise details of the magnetic structure is an essential step towards understanding the origin of the giant AHE in this antiferromagnet, and potentially tuning the properties to realize new functionalities. In this letter, we reexamine the magnetic structure of CoNb\({}_{3}\)S\({}_{6}\) using Co \(L_{3}\) edge resonant elastic x-ray scattering (REXS). We find a double-\(Q\) (\(2Q\)) magnetic structure with a commensurate \(\mathbf{Q}_{0}=(\frac{1}{2}00)\) component and incommensurate \(\mathbf{Q}_{0}\pm\mathbf{q}\) modulation giving rise to a staggered scalar spin chirality with a modulated stripe or checkerboard pattern. The commensurate component of the structure is noncollinear and the incommensurate component is helical. The data confirms that \((\frac{1}{2}00)\) and \((0\frac{1}{2}0)\) peaks belong to separate \(2Q\) domains. Finally, we found that the modulation varies between samples and shows an asymmetric domain pattern implicating lattice strains to influence the magnetic structure, and likely the anomalous Hall response of CoNb\({}_{3}\)S\({}_{6}\). Single crystals were grown using chemical vapor transport [9] with the nominal stoichiometry Co:Nb:S=1:3:6. Four different samples from the same growth were measured. All samples undergo abrupt magnetic transitions at 28.6 K, and exhibit sharp (100)-type Bragg peaks, indicating a well-ordered triangular lattice of intercalated Figure 1: (a) Magnetic REXS intensity in sample 3 at 16 K using circularly polarized x-rays. (b) Experimental geometry, incident (outgoing) polarization channels \(\sigma\) (\(\sigma^{\prime}\)) and \(\pi\) (\(\pi^{\prime}\)) correspond to \(\alpha\) (\(\eta\)) of 0\({}^{\circ}\) and 90\({}^{\circ}\) respectively. (c) Summary of observed magnetic peaks in the first Brillouin zone across all measured samples. Large white circles are \(\mathbf{Q}_{0}\!=\!(\frac{1}{2}00)\) and \((0\frac{1}{2}0)\), green circles show the magnetic reflections observed in samples 1 and 2 and blue squares show those found in sample 3 as in (a). Dashed lines with empty markers show positions of symmetry allowed peaks that were not observed. Co ions [15; 16]. REXS experiments were performed at the I10 beamline at Diamond Light Source using the RASOR endstation [17] with the experimental geometry shown in Fig. 1(b). Samples were mounted in the \((HK0)\) scattering plane to access (\(\frac{1}{2}00\)) and (\(0\frac{1}{2}0\)) magnetic wavevectors at \(2\theta\!=\!106.2\)deg at the Co \(L_{3}\) edge (778.5 eV). In this geometry the x-ray beam scatters from a (110) facet probing an effective area of \(20\!\times\!200\)\(\mu\)m, and with a penetration depth of 0.3 \(\mu\)m. Thus, our measurements probe a macroscopic sample volume containing many basal plane layers. Data was collected for four different samples using an area detector and full linear polarization analysis (FLPA) was carried out on a fifth sample using a point detector and multilayer polarization analyzer optimized for the Co \(L_{3}\) edge [18]. Measurements were performed for zero-field-cooling (ZFC) and field-cooling (FC) under the application of a 0.1 T permanent magnet along the (001) direction. Representative reciprocal space maps of the magnetic scattering are shown in Fig. 1(a). Primary magnetic reflections at \(\mathbf{Q}_{0}\!=\!\big{(}\frac{1}{2}00\big{)}\) and (\(0\frac{1}{2}0\)) were observed in all samples, consistent with previous reports [11]. Our fine resolution measurements also revealed a long-wavelength incommensurate modulation of the magnetic structure through satellite magnetic reflections at \(\mathbf{Q}_{0}\!\pm\!\mathbf{q}\) [Fig. 1]. These satellites showed three distinct behaviors between samples as summarized in Fig. 1(b). In samples 1 and 2, satellites appear at \((\frac{1}{2},\pm\delta,0)\) and \((\pm\delta,\frac{1}{2},0)\), in sample 3 one set of satellites appears at \((\pm\delta,\frac{1}{2},0)\) while the other set appears at \((\frac{1}{2}\mp\delta,\pm 2\delta,0)\), i.e. purely transverse to the main peak. No satellite reflections were observed in sample 4 [15]. At \(T\!=\!14\) K, we find \(\delta\!=\!3.0(3)\!\times\!10^{-3}\) r.l.u. \(=3.7(3)\!\times\!10^{-3}\) A\({}^{-1}\), corresponding to a modulation with 170(15) nm wavelength, or 97(10) nm for sample 3 (\(\frac{1}{2}00\)). These results were consistent across multiple zero-field-cooled (ZFC) and field-cooled (FC) cycles. The different satellite wavevectors that appear between \((\frac{1}{2}00)\) and \((0\frac{1}{2}0)\) reflections in sample 3 [Fig. 1(a)] break both \(C_{6}\) rotational and mirror symmetry about (110) of the triangular lattice, while the satellite reflections observed in samples 1 and 2 possess mirror symmetry but break the \(C_{6}\) rotational symmetry. Such symmetry reductions indicate that \((\frac{1}{2}00)\) and \((0\frac{1}{2}0)\) belong to distinct domains, not a single multi-\(Q\) domain, as will be further confirmed by the linear polarization analysis presented below. In this case, the particular long wavelength modulation of the magnetic structure realized in each domain of a given sample may be selected through a symmetry-breaking field such as small lattice strains that are quenched in during crystal synthesis. Fig. 2 shows the temperature-dependent integrated intensities at \(\mathbf{Q}_{0}\!=\!(\frac{1}{2}00)\) and \(\mathbf{Q}_{0}\!+\!(\overline{\delta}\ 2\delta\ 0)\) in sample 3. Both sets of peaks have the same critical temperature of \(T_{N}=28\) K and an intensity that varies smoothly with temperature. We also observed a smooth decrease in the magnitude of the satellite wave vector with decreasing temperature, decreasing towards \(\mathbf{Q}_{0}\) as temperature is decreased [Fig. 2], characteristic of a helical magnetic modulation [19]. The spectral lineshape of incident energy scans across the 778.5 eV resonance is typical for Co\({}^{2+}\)[20; 21] and further verifies the magnetic origin of observed peaks [Fig. 2 inset, and Fig. 3 (b) inset]. Further details of the magnetic structure are revealed through the polarization-dependent resonant x-ray scattering. All magnetic reflections exhibit a finite circular dichroism (CD), [Fig. 3], arising at \(\mathbf{Q}_{0}\) from noncollinearity of the commensurate component, and at \(\mathbf{Q}_{0}\pm\mathbf{q}\) from a helical modulation [15]. The CD at \((\frac{1}{2}00)\) shows a variation along the \((1\bar{2}0)\) direction suggesting the presence of opposite chirality domains that may have slightly different values of \(\delta\) or are spatially separated on a length scale comparable to the beam size. We also find that the CD varies between ZFC and 0.1 T FC measurements, especially for the \((\frac{1}{2}00)\) peaks, which is consistent with a redistribution of chiral domains, as we discuss below. Precise moment directions were determined from full linear polarization analysis (FLPA) by measuring the intensity at \(\mathbf{Q}_{0}\) as a function of incident polarization angle \(\alpha\) at various fixed analyzer angles \(\eta\) [Fig. 1(b)]. The polarization-dependent intensity shown in Fig. 4(a) is directly sensitive to the real space orientations of the Fourier component \(\mathbf{m}_{\mathbf{Q}_{0},n}\) of the magnetic moments. To model the polarization-dependent scattering intensity in CoNb\({}_{3}\)S\({}_{6}\), we consider a \(2Q\) magnetic structure with propagation vectors \(\mathbf{Q}_{0}\) and \(\mathbf{Q}_{0}\pm\mathbf{q}\) (See SI [15] for alternate possibilities). The real space form of this Figure 2: Normalized temperature dependence of main and satellite peak, and parameter \(\delta\) in satellite wavevector \((\delta,-2\delta,0)\) for sample 3 measured on the point detector with \(\pi\) polarized x-rays. The inset shows a fixed-\(Q\) energy scan of the main peak, with the dashed line showing total fluorescence yield (TFY). structure is \[\begin{split}\mathbf{m}_{n}(\mathbf{r}_{j})&=\cos(\mathbf{q}\cdot \mathbf{r}_{j}+\psi_{n})\cos(\mathbf{Q}_{0}\cdot\mathbf{r}_{j}+\phi_{n})\hat{\mathbf{u}}_{n}\\ &+\sin(\mathbf{Q}_{0}\cdot\mathbf{r}_{j}+\phi_{n})\hat{\mathbf{v}}_{n},\\ &+\cos(\mathbf{q}\cdot\mathbf{r}_{j}+\psi_{n}+\tfrac{\pi}{2}\chi)\cos(\bm {Q}_{0}\cdot\mathbf{r}_{j}+\phi_{n})\hat{\mathbf{w}}_{n},\end{split} \tag{1}\] where \(n\!=\!1,2\) labels the sublattices at 2 Co sites in the unit cell, \(\chi\!=\!\pm 1\) is the helix chirality, \(\phi_{n}\) and \(\psi_{n}\) are the phases on sublattice \(n\) for \(\mathbf{Q}_{0}\) and \(\mathbf{Q}_{0}\!\pm\!\mathbf{q}\) respectively, and \(\hat{\mathbf{u}}_{n}\), \(\hat{\mathbf{v}}_{n}\), and \(\hat{\mathbf{w}}_{n}\) are unit vectors, which we assume to be orthogonal to maintain a constant moment size. The Fourier components of this structure are \[\begin{split}\mathbf{m}_{\mathbf{Q}_{0},n}&=\mathbf{m}^{*}_{ -\mathbf{Q}_{0},n}=-\tfrac{i}{2}e^{i\phi_{n}}\hat{\mathbf{v}}_{n}\\ \mathbf{m}_{\mathbf{Q}_{0}\pm\mathbf{q},n}&=\tfrac{\mathbf{n}^{*}}{ 4}e^{i\phi_{n}\pm i\psi_{n}}(\hat{\mathbf{u}}_{n}\pm i\hat{\mathbf{w}}_{n}).\end{split} \tag{2}\] We parameterize the magnetic structure in terms of the angle \(\mu_{n}\) of \(\hat{\mathbf{u}}_{n}\) from the lattice vector \(\mathbf{a}_{1}\!=\!a\hat{\mathbf{x}}\) within the \(a\)-\(b\) plane, out-of-plane canting angle \(\nu_{n}\) of \(\mathbf{v}_{n}\), and phases \(\phi_{n}\), \(\psi_{n}\). We fit the commensurate component \(\mathbf{m}_{\mathbf{Q}_{0},n}\) to the measured FLPA shown in Fig. 4(a). \(\nu_{1}\!=\!\nu_{2}\) is ruled out as this always leads to zero \(\pi\)-\(\pi^{\prime}\) intensity, inconsistent with the observed FLPA and CD. For our analysis, we have fixed \(\mu_{1}\!=\!\mu_{2}\!=\!\mu\) and \(\nu_{1}\!=\!-\nu_{2}\!=\!\nu\). We find no improvements in the model by relaxing these constraints. While the phase variables \(\phi_{n}\) and \(\psi_{n}\) cannot be uniquely determined from our measurements, symmetry constrains \(\Delta\phi\!=\!\phi_{2}\)-\(\phi_{1}\) to either \(0^{*}\) or \(180^{*}\)[22]. We separately consider two cases: \(\Delta\phi\!=\!0^{*}\) or \(\Delta\phi\!=\!180^{*}\) and find for both \(\mu\!=\!109(1)^{*}\) at \((\tfrac{1}{2}00)\) and \(\mu\!=\!12(1)^{*}\) at \((0\tfrac{1}{2}0)\), or nearly \(\pm 80^{*}\) from \(\mathbf{Q}_{0}\). The in-plane angle relative to \(\mathbf{Q}_{0}\) is opposite in each domain, with the same broken symmetry as the modulation wavevectors in samples 1 and 2 [15]. For \(\Delta\phi=0^{*}\) we find \(\nu=37(2)^{*}\) at \((\tfrac{1}{2}00)\) and \(\nu\!=\!24(2)^{*}\) at \((0\tfrac{1}{2}0)\). While for \(\Delta\phi\!=\!180^{*}\), we find \(\nu\!=\!14(2)^{*}\) at \((\tfrac{1}{2}00)\) and \(\nu\!=\!9(2)^{*}\) at \((0\tfrac{1}{2}0)\). Both cases adequately describe the data at \((\tfrac{1}{2}00)\) while neither fully matches the intensity at \(\pi\)-\(\sigma\) for \((0\tfrac{1}{2}0)\). We attribute this discrepancy to an experimental artifact likely due to a slight analyzer misalignment. Furthermore, we cannot rule out contributions from domains with different moment orientations because the x-ray intensity measures an ensemble-average for a given \(\mathbf{Q}\). The results of our fit are summarized in Table 1 and Fig. 4(b) shows a real-space representation of the magnetic structure found in sample 3. Our measurements reveal that CoNb\({}_{3}\)S\({}_{6}\) hosts a non-coplanar magnetic structure [Fig. 4(d)] that may strongly influence the electronic transport properties. In particular, a nonzero scalar spin chirality, \(\chi_{s}\!=\!\mathbf{m}(\mathbf{r}_{i})\cdot[\mathbf{m}(\mathbf{r}_{j})\times\mathbf{m}(\mathbf{r}_{k})]\) for sites \(i\), \(j\), \(k\) on a triangular plaquette generates an effective magnetic field felt by conduction electrons passing over the plaquette: \(\mathbf{b}_{\alpha}\!\propto\!t_{\alpha}\chi_{s,\alpha}\hat{\mathbf{n}}_{\alpha}\), where \(\hat{\mathbf{n}}_{\alpha}\) is the plaquette normal vector and \(t_{\alpha}\!=\!t_{ij}t_{jk}t_{ki}\) is the hopping integral around the plaquette [13]. We compute the total scalar spin chirality for CoNb\({}_{3}\)S\({}_{6}\) using the real space spin structures found above by considering separate contributions from intra-sublattice plaquettes \(\chi_{s}^{\perp}\), and inter-sublattice plaquettes \(\chi_{s}^{\perp}\), involving two sites from one sublattice and one site from the opposite one [Fig. 4(c)]. We find that the uniform net scalar chirality vanishes \(\chi_{s}(Q\!=\!0)\!=\!0\), but the local scalar chirality is finite. \(\chi_{s}^{\parallel}\) forms stripes with propagation vector \(\mathbf{Q}_{0}\) and no variation along \(\mathbf{q}\). The stripes on each sublattice are \(\pm 90^{*}\) out of phase for \(\Delta\phi=0^{*}\) or \(180^{*}\). \(\chi_{s}^{\perp}\) depends on \(\nu\) and forms complex structures that depend on the choice of the phase variables. We apply the constraint \(\Delta\psi\!=\!0^{*}\) or \(180^{*}\) to consider four different combinations of phases that result in stripe or checkerboard like patterns of chirality [15]. In all cases, the magnitudes of \(\chi_{s}^{\parallel}\) and \begin{table} \begin{tabular}{c c c c c} \hline \hline & \multicolumn{2}{c}{\(\Delta\phi=0^{*}\)} & \multicolumn{2}{c}{\(\Delta\phi=180^{*}\)} \\ \hline Peak & \(\mu\) & \(\nu\) & \(\mu\) & \(\nu\) \\ \hline \((\tfrac{1}{2}00)\) & \(109(1)^{*}\) & \(37(2)^{*}\) & \(109(1)^{*}\) & \(14(2)^{*}\) \\ \((0\tfrac{1}{2}0)\) & \(12(1)^{*}\) & \(24(2)^{*}\) & \(12(1)^{*}\) & \(9(2)^{*}\) \\ \hline \hline \end{tabular} \end{table} Table 1: Parameters obtained from FLPA describing the commensurate Fourier component, for the two possible choices of \(\Delta\phi\). vanish as \(q\to 0\). Thus, the incommensurate modulation is essential for providing a finite local scalar spin chirality in CoNb\({}_{3}\)S\({}_{6}\). Due to the opposite canting between sublattices, \(\chi_{s}^{\perp}\) is much larger than \(\chi_{s}^{\parallel}\) for the values of \(\nu\) we have found. Although the relative magnitude of the intra- and inter-sublattice hopping integrals are unknown, the effective field produced by \(\chi_{s}^{\perp}\) will dominate for all feasible values and the qualitative behavior is unchanged by \(\chi_{s}^{\parallel}\)[15]. In order to visualize \(\chi_{s}\), we project \(\chi_{s}^{\perp}\) onto the \(z\) component of the plaquette normal vectors, shown in Fig. 4(d) and (e) for two different possibilities of the relative phases. In all four cases, the scalar chirality develops an intricate pattern, modulated along both \(\mathbf{Q}_{0}\) and \(\mathbf{q}\)[15]. Such staggered chirality cannot directly account for the AHE. However, the staggered chirality in the presence of finite spin-orbit coupling may generate a net Berry curvature that can account for the AHE [23]. Alternatively, the structure we have found may play a role in the AHE via the crystal chirality or alternating effects [3; 4; 24; 25]. The finite local chirality should also give rise to nonlinear or nonreciprocal transport [26] in samples with suitably controlled domain structures. The present findings clarify the microscopic origins of the giant Hall signal in CoNb\({}_{3}\)S\({}_{6}\) and highlight the importance of mesoscale (magnetic domains) in this itinerant frustrated magnet. A noncoplanar magnetic structure may be stabilized through competing exchange interactions [27] or through itinerant frustration in a Kondo lattice model [28; 29; 30; 31; 32]. Given that CoS\({}_{6}\) octahedra in CoNb\({}_{3}\)S\({}_{6}\) are disconnected, with distances of 5.75 A between nearest neighbor Co sites, superexchange is not likely between local Co moments and a Kondo lattice picture is more natural. Such a mechanism is also consistent with angle resolved photoemission experiments [33]. In this model, the long wavelength modulation at \(\mathbf{q}\) is due to biquadratic interactions originating from Fermi surface instabilities and giving rise to stripes of scalar chirality similar to our observations [34; 29; 30]. The specific value of \(\mathbf{q}\) depends on precise details of the Fermi surface and chemical potential that are further influenced by intrinsic material factors such as lattice strains or chemical defects, consistent with the sample dependence we have observed. This intrinsic sensitivity of the magnetic modulation also affects the domain structure that will, in turn, influence electronic transport in macroscopic samples. Indeed, a striking feature of our results is the broken symmetry of the domains and differences between samples with identical thermodynamic characterization and well-ordered Co triangular lattices [15]. We refer to the three possible commensurate wavevectors as \(Q_{0}\)-domains. For a given \(Q_{0}\)-domain, the symmetry-allowed directions of \(\mathbf{q}\) form \(q\)-domains. For a given \(\mathbf{Q}_{0}\), \(\mathbf{q}\) pair, the helix chirality \(\chi=\pm 1\) and canting direction \(\pm\nu\) give four types of \(\chi\)-domains. In all samples, we only observed a single \(q\)-domain for a given \(Q_{0}\), breaking the expected symmetry between each \(Q_{0}\). The \(q\)-domains in sample 3 [blue squares in Fig. 1(c)] break both \(C_{6}\) rotational and mirror symmetry about (110). The \(q\)-domains in samples 1 and 2, [green circles in Fig. 1(c)] break \(C_{6}\) rotational symmetry, while retaining mirror symmetry about the (110). This same symmetry-breaking is exhibited by the in-plane orientation of the commensurate component measured from FLPA. Although we did not measure reciprocal space maps on that sample, it has the same sur face normal as samples 1 and 2, so we expect it to show the same modulation types. It is thus likely that the in-plane orientation of the commensurate component and the helical modulation wavevector are pinned. The appearance of distinct \(q\)-domains between two \(Q_{0}\)-domains in a single sample implicates a symmetry breaking strain field. In helimagnets where competing and anisotropic exchange interactions control the modulation, helical wavevectors often align with the strain direction [35; 36; 37; 38; 39]. Similarly, small strains can modify the Fermi surface and break degeneracies between nesting vectors in itinerant magnets [40; 41]. The domain selection in samples 1 and 2 can thus naturally be explained by a residual strain along the (110) surface normal that favors \(\mathbf{q}\) more closely aligned with (110) [42]. However, in sample 3, the modulations at \((\frac{1}{2}00)\) and \((0\frac{1}{2}0)\) are not along symmetry-equivalent directions and these two peaks cannot be described as separate domains of the same structure. In this sample, the surface normal is tilted away from (110), possibly giving rise to a rotated residual strain that may stabilize the transverse modulation in one \(Q_{0}\)-domain. This same rotation in the other \(Q_{0}\)-domain would place the modulation closer to the longitudinal direction (\(q_{\parallel}\)), which may be unstable. The two types of \(\chi\)-domains we have identified in CoNb\({}_{3}\)S\({}_{6}\) provide a microscopic picture for the chiral _micro_ and _macro_ domains [10], which respectively require field-cooling with 0.1 T and 8 T to align. We observed a finite CD after zero-field cooling that must arise from an unbalanced population of \(\chi\)-domains in the absence of any external symmetry-breaking perturbation, suggesting that magnetic chirality may be coupled to structural chirality [43; 44; 45]. Field cooling under a 0.1 T field alters the CD at both \(\mathbf{Q}_{0}\) and \(\mathbf{Q}_{0}\pm\mathbf{q}\), consistent with chiral microdomains that arise from spin canting [15]. While the AHE in CoNb\({}_{3}\)S\({}_{6}\) has been shown to vary greatly with both Co [46] and S [47] stoichiometry, the central importance of the domain structure is implicit through the observed order of magnitude enhancement of the anomalous Hall conductivity in mesoscale samples [10]. Our measurements further highlight and provide a microscopic picture for these domains, demonstrating that quenched in strains can act to give distinct magnetic phenomena between samples with identical stoichiometry. Given the dependence of the magnetic structure on this strain, the AHE may also show a pronounced in-plane anisotropy based on the particular domain populations preferred by the sample geometry. Future work might take advantage of this and employ strain directly to control the symmetry-breaking and tune new phases [48; 49]. In summary, we have discovered a double-\(Q\) noncoplanar magnetic structure in CoNb\({}_{3}\)S\({}_{6}\) exhibiting a staggered scalar spin chirality \(\chi_{s}\). The result rules out a uniform scalar chirality as the origin of the anomalous Hall effect, but opens up the possibility for other nontrivial transport phenomena in this itinerant antiferromagnet. Theoretical work is needed to understand the mechanism underlying this magnetic structure and its impact on transport properties. On the other hand, the magnetic domain pattern reveals a potential tunablity of the magnetic modulation through lattice strain that opens up new avenues for controlling the coupling between magnetic order and electronic transport in itinerant antiferromagnets. ## Acknowledgements We are grateful to Cristian Batista for helpful discussions and comments on this manuscript. Work at Brown University was supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Award Number DE-SC0021265. This work was carried out with the support of Diamond Light Source, beamline I10 under proposal numbers MM30765 and MM30768. We thank Mark Sussmuth for technical support at I10.

2305.02211

Influence zones for continuous beam systems

Unlike influence lines, the concept of influence zones is remarkably absent within the field of structural engineering, despite its existence in the closely related domain of geotechnics. This paper proposes the novel concept of a structural influence zone in relation to continuous beam systems and explores its size numerically with various design constraints applicable to steel framed buildings. The key challenge involves explicitly defining the critical load arrangements, and is tackled by using the novel concepts of polarity sequences and polarity zones. These lead to the identification of flexural and (discovery of) shear load arrangements, with an equation demarcating when the latter arises. After developing algorithms that help identify both types of critical load arrangements, design data sets are generated and the influence zone values are extracted. The results indicate that the influence zone under ultimate state considerations is typically less than 3, rising to a maximum size of 5 adjacent members for any given continuous beam. Additional insights from the influence zone concept, specifically in comparison to influence lines, are highlighted, and the avenues for future research, such as in relation to the newly identified shear load arrangements, are discussed.

Adrien Gallet, Andrew Liew, Iman Hajirasouliha, Danny Smyl

2023-02-24T11:14:15Z

http://arxiv.org/abs/2305.02211v1

# Influence zones for continuous beam systems ###### Abstract Unlike influence lines, the concept of influence zones is remarkably absent within the field of structural engineering, despite its existence in the closely related domain of geotechnics. This paper proposes the novel concept of a structural influence zone in relation to continuous beam systems and explores its size numerically with various design constraints applicable to steel framed buildings. The key challenge involves explicitly defining the critical load arrangements, and is tackled by using the novel concepts of polarity sequences and polarity zones. These lead to the identification of flexural and (discovery of) shear load arrangements, with an equation demarcating when the latter arises. After developing algorithms that help identify both types of critical load arrangements, design data sets are generated and the influence zone values are extracted. The results indicate that the influence zone under ultimate state considerations is typically less than 3, rising to a maximum size of 5 adjacent members for any given continuous beam. Additional insights from the influence zone concept, specifically in comparison to influence lines, are highlighted, and the avenues for future research, such as in relation to the newly identified shear load arrangements, are discussed. _Keywords_: influence zones, influence lines, load arrangements, continuous beams, structural design, pattern loads, polarity zones. ## 1 Introduction Influence lines, which derive from Betti's theorem established in 1872 [1], are a well-established tool in structural engineering to identify the worst-case load placement on structural systems [2, 3, 4], and are widely applied in research related to continuous beam systems [5, 6], rigid frames [7], bridge engineering [8] and structural health monitoring [9, 10]. Influence zones, on the other hand, also known as zones of influence, are an established concept within the field of geotechnical engineering, helping to identify the area of engineering soils likely to be affected by loading due to sub- and superstructure construction [11], providing geotechnical engineers valuable design insight in deep foundation design [12, 13], settlement estimations [14] and preserving groundwater supplies [15]. Despite the obvious discipline link between geotechnical and structural engineering, the equivalent use of an influence zone in structural engineering does not exist in literature. Here, the term _structural influence zone_ would refer to the zone in which applied forces, stiffness provisions and support conditions, or changes thereof, impact the design of the surrounding structural system. The dearth of literature on such an _influence zone_ is surprising. For instance, the concept of influence zones also exists outside of geotechnical literature. Some examples are available in research related to the study of saltwater-freshwater interfaces [16], harmful emission concentrations at traffic intersections [17], reverse \(k\)-nearest neighbour algorithms [18, 19], propagation path of surfaces waves [20] and ecological studies on below-ground plant competition [21]. Furthermore, one can readily identify situations where knowledge of the _influence zone_ could be beneficial in design. For example, the size of the influence zone could allow an engineer to avoid the need to model an entire structure for the design of a single element whilst being confident that structural information outside the influence zone is irrelevant, with direct applications in multi-disciplinary projects [22]. The impact of late design changes (due to changes in loading or structural provisions), which are known to cause significant time lags until the associated engineering analysis is completed [23], could be more effectively addressed by knowing immediately the selection of members impacted by the said design change. Similarly, engineers are typically required to verify assumptions made in preliminary design [24]. In such cases, the use of an influence zone-based approach could guide what information to incorporate when building an independent model of the design problem. In all of these scenarios, there is valuable design insight to be gained from the _influence zone_. This article aims to address the above mentioned knowledge gap by numerically introducing the concept of influence zones in relation to continuous beam systems. First, the theory and methodology for evaluating the influence zone will be introduced in section 2, followed by a systematic analysis of critical load arrangements in section 3. The explicit formulations of critical load arrangements allow for the efficient generation of design data sets and the evaluation of their respective influence zones in section 4, the results of which are discussed in section 5. In addition to the _influence zone_, this paper proposes other novel concepts such as _polarity zones_, identifies an entirely new set of critical pattern loads named _shear load arrangements_, and proposes efficient _load arrangement algorithms_ for continuous beam systems of arbitrary member size. ## 2 Methodology ### Overview Consider a continuous beam system, as shown in Figure 1, consisting out of \(m\) members, indexed by \(i\), which is subjected to \(w_{i}\) uniformly distributed loads (UDLs) from vector \(\mathbf{w}\), with each member having span length \(L_{i}\) from vector \(\mathbf{L}\). When designing this system to identify the minimum required structural properties of the members (size optimisation) denoted \(I_{i}\) to form vector \(\mathbf{I}\), it will need to be designed against the worst-case load arrangement (also known as pattern load) from the set of load arrangements \(\mathbf{J}\) of size \(p\). The over-restrained nature of this structural system (a function of the support fixity and structural connectivity) renders the continuous beam indeterminate. This means that the performance of the system is a function of the structural properties which need to be evaluated, and generally makes the design process iterative. Literature has well established formulations to design such indeterminate systems [25]. Figure 1: An exemplary continuous beam system with \(m=5\) members, subjected to UDLs \(\mathbf{w}\), spans \(\mathbf{L}\) and with designed cross-sectional properties \(\mathbf{I}\), all indexed by \(i\). The system’s indeterminacy requires an iterative design process against various load arrangements \(\mathbf{J}\) of size \(p\) indexed by \(j\). ### Influence zone formulations Suppose a member within a continuous beam system is designated as the _design beam_ by index \(d\), and a discrete integer \(k\in\mathbf{Z}\) indicates the index position of a member relative to the design beam at \(d\). As shown in Figure 2, if \(\mathbf{K}\) refers to the list of members identified in terms of \(k\) that fall within the influence zone, then the size of the influence zone is denoted by \(k_{\max}=\max(|\mathbf{K}|)\) with \(k_{\max}\in\mathbf{N}^{0}\), representing the set of all positive integers and including \(0\). Two different formulations have been identified to evaluate the influence zone: * The _local formulation_ identifies the value of \(k_{\max}\) based on whether the design information at the design beam \(d\) significantly influences the surrounding members of indices \(d-k_{\max}\leq i\leq d+k_{\max}\). * The _global formulation_ identifies the value of \(k_{\max}\) based on when the design information at members with indices \(i<d-k_{\max}\) and \(i>d+k_{\max}\) becomes inconsequential for the design beam at \(d\). Figure 2: An example demonstrating how influence lines relate to influence zones, and what an influence zone of size \(k_{\max}=2\) corresponds to in relation to a given design beam (here \(d=3\), highlighted in yellow). For the continuous beam system established in Figure 1, the "design information" include the UDLs \(w_{i}\) and spans \(L_{i}\). Although the terms "significantly influences" and "becomes inconsequential" are currently undefined, they refer to an error threshold that will be explained later. Whilst the _local_ and _global_ formulations differ in terms of where the design information impact is measured from (locally at the design beam for the _local formulation_ or outside the influence zone for the _global formulation_), as long as the design constraints are identical, the size of the influence zone \(k_{\max}\) each formulation identifies will be identical. There are various methodologies one could employ to establish the influence zone using either formulation. For example, _analytical_ approaches making use of concepts such as perturbation theories based on the relationship between force vectors and stiffness matrices may be viable. On the other hand, influence zones could be approached experimentally with the use of physical models or numerically with the use of finite element methods. Each methodology has its own disadvantages. It is not intuitive how one would evaluate the size of the influence zone using a perturbation based approach if large design perturbations are required with multiple load arrangements. Experimental procedures would be limited by the number of design scenarios that can be tested, whilst numerical approaches would make mechanical assumptions on the material and structural behaviour of the system. A numerical approach was preferred since it allows a multitude of design scenarios to be investigated and for statistical conclusions to be determined. Both the local and global formulations were attempted with the use of a numerical model, yet only the latter formulation was fully developed. This was because with the global formulation, the influence zone \(k_{\max}\) could be measured in relation to the utilisation ratio of the design beam \(d\) directly, which made evaluating and reporting the influence zone easier. In the local formulation, the utilisation ratio of all surroundings members outside the influence zone would have to be monitored. ### Mathematical formulation Mathematically, the global formulation can be expressed as follows. For a given continuous beam system as depicted in Figure 1, and the design constraints expressed in Equation 1, \[\begin{array}{l}w_{\min}\;<w_{i}<w_{\max}\\ L_{\min}\;<L_{i}\;<L_{\max}\\ I_{\min}\;<I_{i}\;<I_{\max}\end{array} \tag{1}\] the size of the influence zone of a given design beam \(d\) is found when the value of \(k_{\max}\in\mathbf{N}^{0}:k_{\max}\in[0,m]\)**and** all values larger than \(k_{\max}\) fulfil the following condition: \[\begin{array}{l}\left|\;1-\frac{u_{d,\mathrm{cap}}}{u_{d, \mathrm{true}}}\;\right|\leq\epsilon_{\max}\\ \\ u_{d,\mathrm{cap}}=\max\left(\;\sum_{i=-k_{\max}}^{k_{\max}}\mathbf{u}_{d,i, j}(\mathbf{w},\mathbf{L},\mathbf{I},\mathbf{J})\;\right)\end{array} \tag{2}\] where \(\epsilon_{\max}\) represents the maximum error threshold for the difference between \(u_{d,\text{cap}}\), the captured utilisation ratio of the design beam \(d\) for a given value of \(k_{\max}\), and \(u_{d,\text{true}}\), the true utilisation ratio of the design beam \(d\) if the contribution of all members of the continuous beam system had been considered. \(\mathbf{u}_{d,i,j}\) is the utilisation ratio contribution function towards the design beam \(d\) by member \(i\) based on the UDLs \(\mathbf{w}\), spans \(\mathbf{L}\), structural properties \(\mathbf{I}\) and load arrangements \(\mathbf{J}\) indexed by \(j\). The global formulation as written in Equation 2 measures the point at which the contributions outside of \(k_{\max}\) "becomes inconsequential" by minimising the difference between \(u_{d,cap}\) and \(u_{d,true}\) based on \(\epsilon_{\max}\). As \(k_{\max}\) increases, the ratio \(u_{d,cap}/u_{d,true}\) will approach unity, attaining unity if all structural members (\(k_{\max}=m\)) are considered within the influence zone. If the error threshold \(\epsilon_{\max}\) is relaxed, an influence zone less than the total number of beam members \(m\) can be found. The influence zone is therefore a heuristic measure based on an acceptable maximum error threshold \(\epsilon_{\max}\). The importance of the design constraints as specified by Equation 1 is that they allow for the statistical estimation of the maximum influence zone size based on the diversity of design information variation that can arise. The maximum influence zone value for a type of structural system should always be understood with explicit reference to the design constraints it was evaluated by. ### Design constraints and assumptions The design constraints considered in this investigation were chosen for their relevance in the design of continuous steel framed buildings, which is reflected by the range of UDLs and spans of the design data sets. Four individual design scenarios are considered to study the influence zone in depth, with each set featuring an increasing variation in span lengths and applied loads, summarised in Table 1. Length and UDL values are discretized in \(0.5\,\,\mathrm{m}\) and \(5\,\,\mathrm{kN/m}\) increments respectively, and are drawn from a random uniform distribution. \begin{table} \begin{tabular}{c c c c} \hline \hline Data set & \(G_{k,i}=\) & \(Q_{k,i}\in\) & \(L_{i}\in\) \\ \hline Set 1 & \(3.0\,\mathrm{kN/m}+\) & \(a\) for all \(i\), with \(a\in\) & \(b\) for all \(i\), with \\ _Zero variation_ & self-weight & \([0\,\mathrm{kN/m},60\,\mathrm{kN/m}]\) & \(b\in[1\,\mathrm{m},12\,\mathrm{m}]\) \\ Set 2 & \(3.0\,\mathrm{kN/m}+\) & \([20\,\mathrm{kN/m},40\,\mathrm{kN/m}]\) & \([4\,\mathrm{m},8\,\mathrm{m}]\) \\ _Low variation_ & self-weight & \([10\,\mathrm{kN/m},50\,\mathrm{kN/m}]\) & \([2\,\mathrm{m},10\,\mathrm{m}]\) \\ Set 4 & \(3.0\,\mathrm{kN/m}+\) & \([0\,\mathrm{kN/m},60\,\mathrm{kN/m}]\) & \([1\,\mathrm{m},12\,\mathrm{m}]\) \\ _High variation_ & self-weight & \([0\,\mathrm{kN/m},60\,\mathrm{kN/m}]\) & \([1\,\mathrm{m},12\,\mathrm{m}]\) \\ \hline \hline \end{tabular} \end{table} Table 1: Design constraints for various design scenarios used in this investigation based on Eurocode terminology, with \(G_{k}\) and \(Q_{k}\) being the characteristic permanent and variable actions. Increasing set numbers correspond to increasing design variation, a proxy for design complexity. Span and UDL values are discretized in \(0.5\,\,\mathrm{m}\) and \(5\,\,\mathrm{kN/m}\) increments respectively, and will be drawn from a random uniform distribution. Further design and modelling constraints/assumptions include restricting the cross-sectional properties to prismatic BS EN 10365:2017 UKB I-sections, designing for S355 steel with perfectly linear elastic behaviour using Timoshenko-Ehrenfest beam theory (yet the design was conducted using plastic section properties as allowed by EN 1993-1-1 5.4.2(2) [26]). It was assumed that all spans are laterally restrained (and hence not susceptible to lateral instability), with elements designed against EC3 ULS checks (and notably not SLS requirements) with EN 1990 Eq. 6.10 load combination factors [27]. ### The key challenge The most important aspect for evaluating the influence zone using equation 2 is the member-based utilisation ratio contribution function \(\mathbf{u}_{\mathrm{d,i,j}}\). Whilst the UDLs \(\mathbf{w}\) and spans \(\mathbf{L}\) are given, the critical load arrangement from set \(\mathbf{J}\) which will determine the required structural properties \(\mathbf{I}\) is unknown. Furthermore, the critical load arrangement for a design beam could also differ based on the assumed value of the influence zone size \(k_{\mathrm{max}}\). One approach would be to use a naive, brute-force procedure to trial every possible load arrangement to create the set \(\mathbf{J}_{\mathrm{naive}}\) with a corresponding set size of \(p_{naive}=2^{m}\) for \(\mathbf{J}\) in equation 2. This is not an issue for systems with few members, but if larger systems with \(m>10\) members need to be modelled to study the influence zone in depth, a brute-force approach becomes computationally expensive. The issue of computational cost in relation to critical load arrangements of large-scale systems is well acknowledged in literature, and various methodologies have been employed using probability [28] and possibility theories [29, 30]. Among the latter, fuzzy sets using interval finite-element methods have been shown to be efficient and accurate [31, 32]. However, whilst these interval-based methods are effective at evaluating the bounds (the worst case force/moment value) of the critical load arrangement, they do not in fact reveal what this load arrangement looks like. This is problematic for the evaluation of the influence zone, since Equation 2 relies on being able to identify this set \(\mathbf{J}\) explicitly. Another approach would be to use the load arrangements prescribed by design manuals, yet these consist out of a heuristic set of load arrangements that are known to be non-conservative [32]. Due to these limitations, a rigorous study is conducted to identify and validate the set of critical load arrangements _a priori_ for the design problem identified in Section 2.1, labelled \(\mathbf{J}_{\mathrm{crit}}\). This will not only reduce the computational cost of both designing the members and evaluating their influence zone with Equation 2; it will also highlight the relationship between influence lines and influence zones, providing an intuition on the size of the latter. This study is followed by the numerical generation of randomly distributed data sets based on the design constraints identified in Section 2.4, allowing the evaluation and statistical analysis of the influence zone for various continuous beam systems. ## 3 Critical load arrangement investigation ### Polarity sequences Influence lines can be used to identify the critical load arrangements for a given continuous beam system. The design problem is restricted to positive UDL values only (no uplift) which can be activated on or off (1 or 0 as activation factors). Therefore, by integrating the influence line for each individual beam \(i\), one can evaluate the net contribution (positive or negative) a given beam causes in terms of bending moments and shear forces at the influence line (IL) location when subjected to a positive, unit UDL. The net-contribution of each beam can be either positive or negative at the IL location, that is "hogging or sagging" for bending moments and "clockwise or anti-clockwise" for shear forces, respectively, which is termed as the _polarity_ of that particular beam. This procedure is shown in Figure 3. The last row of Figure 3 therefore reflects a particular _polarity sequence_ for a given IL location, which can be directly used to identify the critical load arrangement for that IL location. When all beams of positive polarity are loaded, then the maximum positive internal forces are generated at the IL location, and vice-versa, loading the negative polarity members leads to the maximum negative internal forces. Figure 3: An exemplary process of arriving from influence line plots (top row) to polarity sequences (bottom row) via integrated influence lines (middle row) for a) major axis bending moment \(M_{y}\) and b) major axis shear force \(V_{z}\) about the specified influence line (IL) location. ### Polarity zones A rigorous qualitative study of the polarity sequences for different IL locations and design scenarios revealed 5 unique polarity sequences that occur along specific segments of a given beam span termed _polarity zones_, which are illustrated in Figure 4 for the central beam highlighted in red. These 5 polarity zones are common to all members of both homogeneous (equal spans and cross-sections) as well as heterogeneous continuous beam systems, although the exact boundaries between one zone varied depending on the relative magnitude of spans and cross-section properties. The sequences identified in Figure 4 also apply to larger beam systems with the polarity direction alternating at each successive beam. For example, if the 5-member system was extended by an additional member on either side of the system (to give a 7 member system), the left-most member of the Type I polarity sequence would have a positive polarity, and similarly, the right-most member would have a negative polarity. The same logic extends to the other four sequences. Each polarity sequence is indicative of two critical load arrangements that maximise the positive or negative internal member forces respectively. The maximum positive load Figure 4: Polarity zones that occur along various span segments of a \(m=5\) homogeneous beam system of equal span and cross-sectional properties. The same zones and sequences, although at different boundaries, occur in heterogeneous (varying UDL and span) systems. arrangement for Type I is also equal to the maximum negative load arrangement for Type IV, since these sequences are polar opposites of each other, which is also true for the Type II and Type IV sequences. Consequently, these 5 polarity zones correspond to 6 unique critical load arrangements for a given beam, namely positive Type I, II and III along with their (negative) polar opposites. The only exceptions occur for the beams at either end of the spans, named end-span beams, in which the Type I and Type IV sequences collapse into the Type III sequence (or its polar opposite) at the left end, and similarly for the Type II and Type V sequences at the right end, resulting in four unique load arrangements for end-span beams. Whilst each non-end-span beam has 6 unique critical load arrangements, it does not mean that the beam system has \(6m\) unique load arrangements (\(m\) is the number of members in the beam system). This is because, as shown in Figure 5, the maximum positive Type V load arrangement for one beam is identical to the maximum positive Type I load arrangement of the beam immediately adjacent to (the right of) it. A similar overlap exists between Type II and Type IV sequences, and the two Type III load arrangements (for maximum and negative internal forces) are identical for all beams. Through a process of elimination, it is possible to simplify the actual total number of potential critical load arrangements to \(p_{\text{flex}}=2m\). This set will be termed the _flexural load arrangements_ set \(\mathbf{J}_{\text{flex}}\), and can be evaluated using Algorithm 1 provided in A, with an example output for a \(m=5\) system shown in Figure 6, grouped in alternating and adjacently loaded arrangements. The load arrangement set \(\mathbf{J}_{\text{flex}}\) of size \(p_{\text{flex}}=2m\) identified here is a literal exponential improvement to the brute-force approach of analysing and designing against \(p=2^{m}\) load arrangements and for evaluating the influence zone with Equation 2. It needs to be shown though that all critical load arrangements \(\mathbf{J}_{\text{crit}}\) fall within \(\mathbf{J}_{\text{flex}}\) (i.e. \(\mathbf{J}_{\text{crit}}\in\mathbf{J}_{\text{flex}}\)). Figure 5: Polarity sequences are identical for adjacently lying beams (highlighted in red) for Type I and Type V sequences as shown by Figure a) and Figure c), as well as Type II and Type IV sequences, as shown by Figure b) and Figure d). ### Shear beams and the impact on critical load arrangements To check if the load arrangement set \(\mathbf{J}_{\text{flex}}\) contains all critical pattern loads, various continuous beam systems of size up to \(m=10\) (to facilitate computational feasibility) were numerically generated, using randomly distributed UDLs and span lengths based on the design constraints identified in Table 1. By calculating all \(2^{m}\) load arrangements, evaluating the utilisation ratio (based on the moment and shear force combinations) of each, the load arrangement that caused the worst-case utilisation ratio could be identified. This set of critical load arrangements was then compared against the \(\mathbf{J}_{\text{flex}}\) set identified by Algorithm 1 provided in A. Although \(\mathbf{J}_{\text{flex}}\) tended to cause the critical utilisation ratio in the majority of cases, there were instances were other load arrangements not within \(\mathbf{J}_{\text{flex}}\) controlled the design. This unexpected behaviour occurred in cases where short spanning, deep beams were included. Analysing these special cases in detail indicated that the exact conditions under which the previously unidentified load arrangements occur were generally related to the following \(L_{\text{shear}}\) span limit quantified by \[\sqrt{\frac{6EI_{yy}}{GA_{z}}}<L_{\text{shear}} \tag{3}\] where \(E\) and \(G\) are the Young's and shear modulus of the material respectively, and \(I_{yy}\) and \(A_{z}\) are the major second moment of area and shear area of the prismatic beam, respectively. Although the \(L_{\text{shear}}\) span limit appears to be related to shear beams, this is the first time that shear beams have been reported in literature to cause novel critical load arrangements. As shown in Figure 7, shear beams appear to flip the polarity of the immediately adjacent member when measured outwardly from a given IL location, with all subsequent members alternating the polarity direction as before. When shear beams (as defined by the \(L_{\text{shear}}\) limit) occur, they introduce new critical load arrangements not found within \(\mathbf{J}_{\text{flex}}\). The increase in terms of the final utilisation factor of the beams was typically in the range of 4-5%, although larger increases were also observed. Whilst a thorough analysis of the increase in utilisation ratio caused by these newly identified load arrangements would be of interest, it falls outside the scope of this study. Instead, an algorithm will be presented capable of identifying these new load arrangements _a priori_, which is the main objective of this investigation as explained in section 2.5. Figure 6: The critical load arrangements set \(\mathbf{J}_{\text{flex}}\) of size \(p=2m\) for a 5-member continuous beam system (\(p=10\)) grouped in alternating and adjacently loaded arrangements. ### Determining shear beam induced critical load arrangements _a priori_ The principal issue when evaluating the shear beam induced critical load arrangements _a priori_, hereafter referred to as the _shear load arrangements_\(\mathbf{J}_{\mathrm{shear}}\), is the fact that the final material and cross-sectional properties to evaluate the \(L_{\mathrm{shear}}\) limit in Equation 3 are not known until the beam is designed. This creates a causality dilemma and hence needs to be addressed. In clear opposition to the \(\mathbf{J}_{\mathrm{flex}}\) set, which does not depend on the continuous beam system properties, the shear load arrangements cannot be established _in universum_ without some system knowledge. However, by taking advantage of the design constraints set by Equation 1, one can identify _a priori_ what members are potentially susceptible to cause shear load arrangements by re-writing Equation 3 as: Figure 7: A schematic demonstrating the impact of a shear beam (highlighted in yellow) on a standard polarity sequence of a continuous beam system when spans shorter than the shear span limit \(L_{\mathrm{shear}}\) (as identified by Equation 3) occur. Note the flipped polarity directions of the members on the right-hand side of the system. \[\sqrt{6\left(\frac{E}{G}\right)_{\max}\left(\frac{I_{yy}}{A_{z}}\right)_{\max}}<L_{ \mathrm{shear,max}} \tag{4}\] The above equation groups the maximum material and cross-sectional property ratios together. By limited the design space to S355 steel and UKB section sizes as specified in Section 2.4, the maximum material ratio (\((E/G)_{\max}=2.600\)) and cross-sectional property ratio (\((I_{yy}/A_{z})_{\max}=0.397m^{2}\)) can be evaluated. Consequently, beams shorter than the shear span limit are susceptible to cause shear load arrangements (in this case \(L_{\mathrm{shear,max}}=2.49\)m). In identifying these susceptible members _a priori_, it is possible to evaluate the shear load arrangements using Algorithm 2 provided in B. Algorithm 2 transforms the flexural load arrangement from set \(\mathbf{J}_{\mathrm{flex}}\) based on a list of susceptible shear beams identified by Equation 4. This is achieved by flipping the on/off activation factor of the load arrangement if a shear beam is encountered whilst travelling outwardly in both the left (-1) and right (1) direction from a start beam index. This operation transforms the flexural load arrangement based on the behaviour identified visually in Figure 7, and needs to check four individual case conditions to account for continuous beam systems that have multiple, potentially adjacently lying, shear beams. Since every beam system is of size \(m\), the time complexity of a single pass of Algorithm 2 is \(O(m)\). However, since every flexural load arrangement (\(2m\)), and every combination of \(n\) potential shear beams (\(2^{n}-1\) combinations, as the zero set is already considered in \(\mathbf{J}_{\mathrm{flex}}\) by default), and every possible start-index (\(m\)) needs to be computed, the time complexity to evaluate the shear set \(\mathbf{J}_{\mathrm{shear}}\) would be \(O(m^{3}\,2^{n})\). It should be noted that this process is computationally expensive. It was observed that passing every possible start index generated either duplicate shear load arrangements, or occasionally existing flexural load arrangements. For example, for a given singular potential shear beam location, the algorithm would result in the same transformed shear load arrangement for all start-indices starting on the left and right hand-side of that susceptible shear beam location. Similarly, the two alternating arrangements from \(\mathbf{J}_{\mathrm{flex}}\) would result in an already existing adjacent arrangement from \(\mathbf{J}_{\mathrm{flex}}\) if only a singular susceptible shear beam exists. Using such logic, it is sufficient to pass only adjacent arrangements from \(\mathbf{J}_{\mathrm{flex}}\) along with the left-hand (or right-hand) index of the adjacently loaded spans as the start index for Algorithm 2 to yield an effective set of potential shear load arrangements. By not having to evaluate Algorithm 2 for every possible start index of each load arrangement, the computational complexity reduces to \(O(m^{2}\,2^{n})\). From this, it also follows that since the alternating load arrangement is never transformed (which leaves only \(2(m-1)\) load arrangements to be passed to the algorithm) and since \(2^{n}-1\) possible shear beam combinations can exist, the maximum number of unique critical shear load arrangements should be of size \(p_{\mathrm{shear}}=2(m-1)(2^{n}-1)\). ### Validating flexural and shear load arrangement algorithms A design data set consisting of 32 UDL and 32 span values sampled from a random uniform distribution for a \(m=10\) beam system was generated based on the high-variation design scenario identified in Section 2.4. Significantly higher variable UDLs (\(Q_{k,i}\in[200\,\mathrm{kN/m},400\,\mathrm{kN/m}]\)) were applied to increase the likelihood of deep beams and thereby critical shear load arrangements, allowing the performance of the algorithm to be stress-tested. This resulted in \(10\times 32\times 32=10240\) individual beam design examples, for which the critical load arrangement \(J_{\mathrm{crit}}\) could be identified. The results of this validation exercise are illustrated in Figure 8, which plots the critical load arrangement index for each design beam example. Every load arrangement index corresponds to a unique load arrangement out of the naive set \(\mathbf{J}_{\mathrm{naive}}\) of size \(p_{\mathrm{naive}}=2^{m}=1024\). The set \(\mathbf{J}_{\mathrm{naive}}\) was ordered so that the load arrangements for set \(\mathbf{J}_{\mathrm{flex}}\) are first, followed by those of set \(\mathbf{J}_{\mathrm{shear}}\), and subsequently all others. The design examples themselves were sorted twice: first in ascending number of shear beam occurrences, and subsequently in ascending load arrangement indices. This results in the gradual increase of the \(J_{\mathrm{crit}}\) indices as seen in Figure 8. Figure 8: Load arrangement index for each design beam example ordered in increasing number of shear beam occurrences and critical load arrangement indices. This confirms visually that the critical load arrangement \(J_{\mathrm{crit}}\) for each design beam example from the generated data set falls within either \(\mathbf{J}_{\mathrm{flex}}\) or \(\mathbf{J}_{\mathrm{shear}}\) and are significantly smaller than \(\mathbf{J}_{\mathrm{naive}}\). Figure b) is an enlarged view of Figure a). Figure 8 sheds insight on a number of important points. The first is that the critical load arrangement \(J_{\text{crit}}\) for every single beam example from the \(10240\) data set occurred within the \(\mathbf{J}_{\text{flex}}\) or \(\mathbf{J}_{\text{shear}}\) sets, validating the qualitative analysis based on the polarity zones and sequences identified previously. This also emphasises the validity of Algorithm 1 and 2. Furthermore, the set size predictions \(p_{flex}=2m\) and \(p_{shear}=2(m-1)(2^{n}-1)\) are also confirmed. For the \(m=10\) member system designed here, \(p_{flex}=20\), and depending on the number of shear beam occurrences of each system, which varied from \(n=\{0,1,2,3,4\}\), the number of shear load arrangements varied from \(p_{shear}=\{0,18,54,126,270\}\). This corresponded to \(p_{total}=\{20,38,74,146,290\}\) respectively, as indicated by the \(y\)-axis of Figure 8 b). Figure 8 a) also emphasises how much smaller sets \(\mathbf{J}_{\text{flex}}\) and \(\mathbf{J}_{\text{shear}}\) are in comparison to \(\mathbf{J}_{\text{naive}}\). This will greatly reduce the number of load-arrangements that need to be analysed, reducing the computational cost of both optimally designing the continuous beam system and evaluating the influence zone for system lengths of \(m>10\). Further insights generated by Figure 8 are discussed in Section 5. ### Summary of critical load arrangements of continuous beams By adding the set of critical flexural and shear load arrangements together, it is possible to explicitly define _a priori_ the set of critical load arrangements for any continuous beam system under defined design constraints. For the purpose of the influence zone concept and Equation 2, it is assumed that: \(\mathbf{J}\rightarrow\mathbf{J}_{\text{crit}}\in\mathbf{J}_{\text{flex}} \cup\mathbf{J}_{\text{shear}}\). The results from this systematic critical load arrangement investigation are summarised in Table 2. \begin{table} \begin{tabular}{c c c} \hline \hline Set & Set Size & Algorithm Complexity \\ \hline Critical load arrangements & 6 & \(O(1)\) \\ per internal beam & & \\ Critical load arrangements & 4 & \(O(1)\) \\ per end-span beam & & \\ \(\mathbf{J}_{\text{flex}}\) - Critical flexural arrangements per beam & \(2m\) & \(O(m)\) \\ system & & \\ \(\mathbf{J}_{\text{shear}}\) - Critical shear arrangements per beam & \(2(m-1)(2^{n}-1)\) & \(O(m^{2}\,2^{n})\) \\ system & & \\ \(\mathbf{J}_{\text{naive}}\) - Naive load arrangements & \(2^{m}\) & \(O(2^{m})\) \\ \hline \hline \end{tabular} \end{table} Table 2: Load arrangements set summary for \(m\) dimensional beam systems containing \(n\) shear beams with associated algorithm complexities. ## 4 Influence zone evaluation ### Explicitly defining the utilisation ratio contribution function By taking advantage of the concept of integrated influence lines and polarity zones from Section 3.1, and by having explicitly defined the critical load arrangement set \(\mathbf{J}\rightarrow\mathbf{J}_{\mathrm{crit}}\) in Section 3.6, it is possible to define the utilisation ratio contribution function \(\mathbf{u}_{d,i,\,j}\) as: \[\begin{split}\mathbf{u}_{d,i,\,j}&\rightarrow \mathbf{D}_{\mathrm{ULS}}(I_{d},M_{d,i,j},V_{d,i,j})\\ M_{d,i,j}&=w_{i}\ J_{i,j}\int_{i}\mathbf{M}_{ \mathrm{IL},d}\\ V_{d,i,j}&=w_{i}\ J_{i,j}\int_{i}\mathbf{V}_{ \mathrm{IL},d}\end{split} \tag{5}\] \(\mathbf{D}_{\mathrm{ULS}}\) represents the ULS steel cross-section design checks based on Eurocode EN 1993-1-1 6.2 [26], \(I_{d}\) represents the cross-sectional properties, \(M_{d,i,j}\) denotes the major axis moment while \(V_{d,i,j}\) is the major axis shear force of the design beam \(d\), \(w_{i}\) is the UDL, and \(J_{i,j}\) is the activation factor of the load arrangement \(j\) from the set \(\mathbf{J}_{\mathrm{crit}}\) for beam \(i\). Integrals \(\int_{i}\mathbf{M}_{\mathrm{IL},\mathrm{d}}\) and \(\int_{i}\mathbf{V}_{\mathrm{IL},\mathrm{d}}\) are the integrated influence line values across beam \(i\) for a particular influence line location within the design beam \(d\) as introduced in Figure 3. The influence line locations within the design beam need to correspond with the worst-case internal force locations. Whilst engineering experience would dictate those to occur over the supports, they can in fact arise anywhere along the design beam depending on the exact distribution of UDLs, spans and cross-sectional properties, since each segment of the design beam has its own critical load arrangement as highlighted by Figure 4. In this study, a total of 11 influence line locations were sampled, one at either support and another 9 equidistantly distributed between the supports. For a given design beam and \(k_{\mathrm{max}}\) value, Equation 5 will therefore result in \(11p\) utilisation ratios (recall that \(p\) is the set size of \(\mathbf{J}\)), from which the critical utilisation ratio (the maximum one) is evaluated in Equation 2 to check if the \(\epsilon_{\mathrm{max}}\) threshold has been attained. Note that as \(k_{\mathrm{max}}\) increases, the critical influence line location and the critical load arrangement can vary, yet as \(k_{\mathrm{max}}\) approaches \(m\), they will equate to the location and load arrangement that governed the design for that particular beam. ### Design data set generation The size of the continuous beam system \(m\) to be modelled needs to be at least double the maximum influence size \(k_{\mathrm{max}}\). This is because the highest influence zone measurable for the middle span of a continuous beam is by design half the system length \(m\). Therefore, size \(m\) needs to be chosen such that \(\max(\mathbf{k}_{\mathrm{max}})<m/2\), where \(\mathbf{k}_{\mathrm{max}}\) is the list of all influence values \(k_{\mathrm{max}}\) of the continuous beam system. Since evaluating \(k_{\mathrm{max}}\) is the main aim of this investigation, a sufficiently large value for \(m\) needs to be assumed; \(m=15\) was used for this purpose. Individual design data sets consisting of 32 UDL and 32 span values sampled from a random uniform distribution for a \(m=15\) beam system were created based on the design constraints identified in Section 2.4 for Sets 2, 3, and 4, each containing \(32\times 32\times 15=15360\) beam designs. For Set 1, the difference within the beam systems only varied in terms of the identical span \(L\) and UDLs \(Q_{k}\) of the beams, which were also sampled in 0.5 m and 5 kN/m increments respectively. Given that this results in 23 span and 13 UDL increments for Set 1 respectively, Set 1 contained \(23\times 13\times 15=4485\) design examples. For the design optimisation of the continuous beam systems, a coupled analysis and design approach was taken, optimising for minimum structural depth. Design sensitivity analysis was avoided by an implicit ordering of the UKB section list based on structural capacity. The influence zone values were extracted using Equation 2 and Equation 5. ### Influence zone results The influence zone results are shown in Figure 9 for a max error threshold \(\epsilon_{\text{max}}=0.005\) and various design data sets defined in Section 2.4. For all sets investigated, the most common influence zone value (the mode) was \(k_{\text{max}}=3\), and the majority of influence zone values were at \(k_{\text{max}}\leq 3\), meaning the span and applied loading information of a given beam along with that of the three adjacent spans on either side captured the correct utilisation ratio of the design beam with less than a \(\pm 0.5\%\) error in the majority of cases. However, the various sets reveal differences in the maximum and distribution of the influence zone. The maximum influence zone value for Set 1 was \(k_{\text{max}}=4\), whereas it was Figure 9: Influence zone results for various design constraints with a max error threshold \(\epsilon_{\text{max}}=0.005\) indicating the percentage frequency distributions of the influence zone values \(k_{\text{max}}\) and minimum utilisation factors captured for each \(k_{\text{max}}\) value for a given design beam \(d\). \(k_{\max}=5\) for Set 2, 3 and 4. Furthermore, as the set number increases, which corresponds with an increase in variation of the design information in terms of spans and UDLs, the influence zone value distribution appears to flatten and widen. For example, it was the high-variation Set 4 which actually contained the most influence zone values \(k_{\max}=0\) for 1.6% of the design examples, whereas the zero variation Set 1 only had 0.4% of its design examples exhibit an influence zone of \(k_{\max}=0\). The minimum utilisation curve (red curve with point markers in Figure 9) captured by each influence zone value suggests that, in general, increasing design variation leads to greater maximum influence zone values. The average and maximum influence zone values were also calculated for various error thresholds as shown in Table 3. Note that the maximum influence zone value of \(k_{\max}=7\) for Set 4 with the highest error threshold confirms that the \(m=15\) member-size assumption was sufficient for the purpose of this study. Together Figure 9 and Table 3 provide evidence for the following conclusions: * A decrease in the acceptable error threshold correlates with an increase in both the average and maximum influence zone range. * An increase in design variation correlates with an increase in the maximum influence zone range. * An increase in design variation, however, correlates with a decrease in average influence zone range in most instances where the acceptable error threshold is relatively tight (\(\epsilon_{\max}\leq 10\%\)). At higher error thresholds the trend is less discernible. It should be noted that an error threshold of less than 0.5% is relatively small in comparison to uncertainties that exist in structural design. These uncertainties include, for example, material yield strength and imposed UDL values (consider that variable UDL values \(Q_{k}\) are increased 50% with a load combination factor of 1.5 within the Eurocodes [27]). Furthermore, the design constraints of design set 4 represent the top end of design variation which may occur in typical continuous beam systems. Consequentially, it is reasonable to suggest that for continuous beam systems with design constraints specified \begin{table} \begin{tabular}{c c c c c c c c c} \hline \hline \multirow{2}{*}{Error \(\epsilon_{\max}\) [\%]} & \multicolumn{4}{c}{Average \(k_{\max}\)} & \multicolumn{4}{c}{Maximum \(k_{\max}\)} \\ \cline{2-10} & Set 1 & Set 2 & Set 3 & Set 4 & Set 1 & Set 2 & Set 3 & Set 4 \\ \hline 0.1 & 4.60 & 4.46 & 3.84 & 3.37 & 5 & 5 & 6 & 7 \\ 0.5 & 2.89 & 2.86 & 2.69 & 2.38 & 4 & 5 & 5 & 5 \\ 1 & 2.76 & 2.75 & 2.35 & 2.06 & 3 & 3 & 5 & 5 \\ 5 & 1.52 & 1.39 & 1.29 & 1.17 & 2 & 3 & 3 & 4 \\ 10 & 0.98 & 0.98 & 0.89 & 0.83 & 2 & 2 & 3 & 4 \\ 20 & 0.76 & 0.84 & 0.73 & 0.67 & 1 & 1 & 2 & 3 \\ 50 & 0.00 & 0.30 & 0.41 & 0.43 & 0 & 1 & 1 & 2 \\ \hline \hline \end{tabular} \end{table} Table 3: Influence zone results for various maximum error thresholds \(\epsilon_{\max}\) for each design data set, evaluating average and maximum influence zone values \(k_{\max}\). Note that increasing set numbers corresponds with increasing design variation, a proxy for design complexity, see Table 1 for details. in section 2.4 the influence zone values are on average \(k_{\max}<3\), and in the most extreme case \(k_{\max}=5\). ## 5 Discussion The results along with the methodology of the influence zone investigation has led to a number of important findings. These include introducing the novel concept of the _structural influence zone_ and a numerical methodology of evaluating it, discovering novel _shear load arrangements_ with the help of _polarity zones_ and _polarity sequences_, and introducing _load arrangement algorithms_ to explicitly identify critical load arrangements of continuous beam systems of any arbitrary member size, which were a necessary prerequisite for the influence zone study. Each of these findings are discussed in detail and contextualised with relevant existing literature. ### Influence zone insights The influence zone results confirm that the impact of loading, and by extension of any design information, drops off sharply the further away one moves from the influence line location. This behaviour can be identified across all influence line diagrams found within this paper, such as Figures 3 and 7. This investigation has formulated this concept as the _influence zone_, shown how it applied to continuous beam systems, and rigorously studied the influence zone distributions under various design assumptions and error thresholds. An important element of the influence zone definition in Section 2.3 should be brought to light. The influence zone value \(k_{\max}\) is only found when two important conditions are met, notably when both the smallest value of \(k_{\max}\) and all values larger than this value of \(k_{\max}\) conform to Equation 2. This was a necessary constraint to account for the fact that the ratio \(u_{d,cap}/u_{d,true}\) sometimes converged towards unity by oscillation. The cause of this behaviour results from adjacent load arrangements (see Figure 6) which were sometimes more critical when only a segment of (as opposed to the entire) load arrangement was considered. For example, the maximum \(u_{d,cap}/u_{d,true}\) ratio calculated for Set 4 when assuming an influence zone value of \(k_{\max}=1\) was 1.89, only for the ratio to drop to 0.942 and 0.998 for influence zone values \(k_{\max}=2\) and \(k_{\max}=3\), respectively. Future research on influence zones should keep this non-intuitive behaviour in mind, since a simple \(u_{d,cap}/u_{d,true}<r_{cap}\) threshold, in which \(r_{cap}\) represents a minimum threshold of captured utilisation would lead to an underestimation of the influence zone distribution. ### Demarcating influence zones from influence lines Although there is a proximal relationship between the concept of influence zones and influence lines, mostly evidenced by Equation 5 where integrated influence lines play an important role for the evaluation of influence zones, these two concepts differentiate themselves in important ways. This distinction also applies to the two-dimensional application of influence lines known as influence surfaces [33, 34, 35, 36]. Whilst influence lines/surfaces are exact analytical tools that define the mechanical response of a known structural system about a particular point, influence zones are a heuristic design tool that offer insight on what information is relevant to the design of the structural system to begin with based on certain analytical assumptions. The value of influence lines/surfaces arise during analysis on a system-by-system basis, whereas the value of influence zones arise during design after having studied them in their statistical aggregate. This distinction could be considered further evidence supporting the demarcation between design and analysis in structural engineering. Previous literature has highlighted the difference between _knowledge-that_ explains fundamental facts about systems (such as influence lines) versus _knowledge-how_ something can be designed or solved (such as influence zones) [37, 38]. Recent literature has suggested that the processes of analysis and design solve related, albeit oppositely posed problems known as forward and inverse problems respectively [39]. Influence lines can be seen as a tool that solves the former, whereas influence zones solve the latter. As a matter of fact, the influence zone concept was developed whilst developing a design model for continuous beam systems from an inverse problem perspective, and allows the _a priori_ knowledge of what span and loading information is relevant for design of a particular continuous beam. It is possible that the influence zone concept could serve as an important heuristic tool in the design of continuous structural systems, supporting the view that the application of heuristics is a cornerstone for engineering design [40]. Further novel ideas might be uncovered when approaching engineering design from an inverse problem perspective. ### Flexural load arrangements An important contribution of this investigation was presenting the flexural load arrangements clearly through the use of _polarity sequences_. Notably the _polarity zones_ highlight which load arrangement is critical for specific segments of a beam, which could be useful in the design of tapered (non-prismatic) continuous beam systems [41, 42]. The influence zone study allows the contextualisation of simplified load arrangement provisions. For example, whilst Annex AB.2 from EN 1993-1-1 [26] covers alternating flexural load arrangements in full, it specifies that for the adjacent flexural load arrangement type, the two adjacently loaded spans are the only spans required to factor the variable load (\(Q_{k}\)). In essence, the variable load information on all other spans aside from the beam under consideration and the two directly adjacent spans are ignored, which is the technical equivalent of assuming an influence zone to \(k_{\max}=1\). With help of Table 3, it is possible to infer that an influence zone value \(k_{\max}=1\) is likely to introduce an error between \(5-10\%\) in terms of the true utilisation for design scenarios with no UDL or span variation (the average \(k_{\max}\) value for \(\epsilon_{\max}=5\%\) and \(\epsilon_{\max}=10\%\) is \(1.52\) and \(0.98\) for Set 1 respectively). The simplified Eurocode provisions are therefore, on average, a reasonable simplification to capture the impact of variable load arrangements. However, the maximum influence zone value of Set 1 with \(k_{\max}=1\) corresponds to an error of \(\epsilon_{\max}=20\%\), and when considering non-heterogeneous continuous beam systems (reflected by Set 2, 3 and 4), this error can increase up to \(\epsilon_{\max}=50\%\) and more. This is further evidence, as already pointed out in literature, that the load arrangement provisions from building codes can be non-conservative and hence lead to unsafe designs [32]. The simplified provisions within the Eurocodes, which also exist within EN 1992-1-1 5.1.3 [43] and other codes [44], need to be understood in context of the \(1.5Q_{k}\) load factors and the dead load contribution \(G_{k}\), which invariably will lessen the underestimation made by the provisions. Nonetheless, the validity of the design code recommendations for flexural load arrangements could be investigated further, especially for highly irregular beam and floor arrangements [45]. ### Shear load arrangements Unlike flexural load arrangements, which have been identified in literature and building codes, the shear load arrangements are a novel discovery. To the authors' knowledge, this is the first time that deep beams have been identified to cause new critical load arrangements in literature. Although shear load arrangements sometimes resulted in identical utilisation ratios to that of flexural ones, initial analyses pointed to an average increase in utilisation ratio of 4-5%, while larger deviations were occasionally observed. Figure 8 also highlights that these shear load arrangements were relatively prevalent within the design scenarios considered. Confirmation and validation of these shear load arrangements by future research is encouraged. Of particular interest is why Equation 4 defines the exact point when these critical load arrangements arise. One notable difference in the mechanical assumption in this investigation of load arrangements as to that of previous studies was the use of Timoshenko-Ehrenfest rather than Euler-Bernoulli beam theory. For example, the two seminal works on establishing the bounds of critical load arrangements using fuzzy set based finite-element methods used Bernoulli-Euler beam theory [31, 32]. A re-investigation with deep beams as defined by Equation 3 and Timoshenko-Ehrenfest beam theory should reveal more critical bounds of load arrangements than previously identified with interval-finite-element methods. The extent to which these shear load arrangements require special provisions within building codes will require further exploration. ### Critical load arrangement algorithms The critical _load arrangement algorithms_ provided in A and B, along with a study of their computational complexity, were key for the evaluation of the influence zone. Limiting the design space to a fraction of the naive \(J_{naive}\) load arrangement set without making heuristic simplifications was crucial in both the data set generation and influence zone evaluation steps. It is likely that there is further room for improvement for Algorithm 2 for evaluating the shear load arrangements for a known list of susceptible shear beams. The current formulation, as explained in Section 3.4, still leads to either pre-existing flexural load arrangements, or creates duplicate shear load arrangements. On average, 74.7% of the outputs obtained from Algorithm B were unique, with a best-case efficiency of 88.8% and a worst-case efficiency of 12.7%. This suggests that an algorithm with a lesser computational complexity than \(O(m^{2}\,2^{n})\) might be achievable through further investigation. ### Future investigations and application of the influence zone concept This investigation will hopefully serve as a starting point for future studies related to the influence zone. There were several limitations within this study, notably not accounting for serviceability checks and limiting the design space to positively loaded UDLs. Furthermore, only 11 equidistant points were sampled about each beam during design and extraction of influence zone values. A more efficient approach could sample specific points against the critical load arrangement that apply to that particular polarity zone. Further studies could be conducted for different material and design information assumptions, while studies could also be expanded to 2D continuous frames and shells, with fixed and semi-rigid connections. As previously explained, such numerical studies could be validated with either analytical or experimental approaches, along with using local, as opposed to global, influence zone formulations as discussed in section 2.2. The influence zone concept and associated results could be a helpful piece of information when teaching the design of large-scale, steel-framed continuous beam systems, may have applications in other research areas such as reliability engineering, and help in the future development of generalised design models [39]. ## 6 Conclusions A novel concept termed the _influence zone_ was proposed in relation to continuous beam systems. The investigation developed a local and global formulation, of which the latter one was explored numerically with design constraints applicable to steel framed buildings. The key challenge was the explicit definition of critical load arrangements to allow the computational feasible generation of design data sets and evaluation of their respective influence zones. The investigation led to three important outcomes: * The development of polarity sequences and polarity zones which led to the demarcation between previously known flexural load arrangements and the newly discovered shear load arrangements, with an explicit span limit equation for when these novel load arrangements occur. * Two algorithms capable of finding these two types of load arrangements, and providing evidence that they encompass all critical permutations in comparison to the naive, brute-force approach. * The generation of design data sets from which the influence zone values for various degrees of design complexities and error thresholds could be rigorously studied. For error thresholds deemed acceptable in structural design, the influence zone for continuous beams within steel framed building under ultimate state considerations is on average less than 3, going to a maximum influence zone value of 5. The influence zone is a heuristic design tool that differentiates itself from influence lines (and influence surfaces) and demonstrates the value of the inverse problem perspective through which it was evaluated by. This study opens the scope for future research, notably in the evaluation of influence zones for various materials and structural systems, validating and explicating the existence of shear load arrangements, and encouraging research on improving the existing algorithm that identifies them. ## Acknowledgements **Authors' contributions:** Adrien Gallet: Conceptualization, Methodology, Investigation, Software, Formal analysis, Validation, Visualization, Writing - Original Draft Andrew Liew: Writing - Review & Editing Iman Hajirasouliha: Writing - Review & Editing Danny Smyl: Supervision, Writing - Review & Editing **Data statement:** Data used in this article are available from the authors upon request. **Competing interests:** The authors declare that they have no competing interests.

2307.13565

Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities

Decision-focused learning (DFL) is an emerging paradigm that integrates machine learning (ML) and constrained optimization to enhance decision quality by training ML models in an end-to-end system. This approach shows significant potential to revolutionize combinatorial decision-making in real-world applications that operate under uncertainty, where estimating unknown parameters within decision models is a major challenge. This paper presents a comprehensive review of DFL, providing an in-depth analysis of both gradient-based and gradient-free techniques used to combine ML and constrained optimization. It evaluates the strengths and limitations of these techniques and includes an extensive empirical evaluation of eleven methods across seven problems. The survey also offers insights into recent advancements and future research directions in DFL. Code and benchmark: https://github.com/PredOpt/predopt-benchmarks

Jayanta Mandi, James Kotary, Senne Berden, Maxime Mulamba, Victor Bucarey, Tias Guns, Ferdinando Fioretto

2023-07-25T15:17:31Z

http://arxiv.org/abs/2307.13565v4

# Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities ###### Abstract Decision-focused learning (DFL) is an emerging paradigm in machine learning which trains a model to optimize decisions, integrating prediction and optimization in an end-to-end system. This paradigm holds the promise to revolutionize decision-making in many real-world applications which operate under uncertainty, where the estimation of unknown parameters within these decision models often becomes a substantial roadblock. This paper presents a comprehensive review of DFL. It provides an in-depth analysis of the various techniques devised to integrate machine learning and optimization models, introduces a taxonomy of DFL methods distinguished by their unique characteristics, and conducts an extensive empirical evaluation of these methods proposing suitable benchmark dataset and tasks for DFL. Finally, the study provides valuable insights into current and potential future avenues in DFL research. ## 1 Introduction Real-world applications frequently confront the task of decision-making under uncertainty, such as planning the shortest route in a city, determining optimal power generation schedules, or managing investment portfolios (Sahinidis, 2004; Liu & Liu, 2009; Kim, Lewis, & White, 2005; Hu, Wang, & Gooi, 2016; Delage & Ye, 2010; Garlappi, Uppal, & Wang, 2006). In such scenarios, estimating unknown parameters often poses a significant challenge. Machine Learning (ML) and Constrained Optimization (CO) serve as two key tools for these complex problems. ML models estimate uncertain quantities, while CO models optimize objectives within constrained spaces. This sequential process, commonly referred to as _predictive_ and _prescriptive_ modeling, as illustrated in Figure 1, is prevalent in fields like operations research and business analytics (den Hertog and Postek, 2016). For instance, in portfolio management, the prediction stage forecasts asset returns, while the prescriptive phase optimizes returns based on these predictions. A commonly adopted approach involves handling these two stages--prediction and optimization--separately and independently. This "two-stage" process first involves training an ML model to create a mapping between observed features and the relevant parameters of a CO problem. Subsequently, and independently, a specialized optimization algorithm is used to solve the decision problem, which is specified by the predicted problem parameters. The underlying assumption in this methodology is that superior predictions would lead to precise models and consequently, high-quality decisions. Indeed, if the predictions of parameters were perfectly accurate, they would enable the correct specification of CO models which can be solved to yield fully optimal decisions. However, ML models often fall short of perfect accuracy, leading to suboptimal decisions due to propagated prediction errors. Thus, in many applications, the predictive and prescriptive modelings are not isolated but rather, deeply interconnected, and hence should ideally be modeled jointly. This is the goal of the **decision-focused learning** (DFL) paradigm, which directly trains the ML model to make predictions that lead to good decisions. In other words, DFL integrates prediction and optimization in an end-to-end system trained to optimize a criterion (i.e., a loss function) that is based on the resulting decisions. Since many ML models, including neural networks (NNs), are trained via gradient-based optimization, the gradients of the loss must be backpropagated through each constituent operation of the model. In DFL, the loss function is dependent on the solution of an optimization model, thus the optimization solver is _embedded_ as a component of the ML model. In this integration of prediction and optimization, a key challenge is _differentiating through the optimization problem_. An additional challenge arises from decision models operating on discrete variables, which produce discontinuous mappings and hinder gradient-based learning. Hence, examining _smooth_ surrogate models for these discrete mappings, along with their differentiation, becomes crucial. These two challenges are the core emphasis and central focal points in DFL. Figure 1: Decision-making under uncertainty involves both predictive and prescriptive analytics. In the predictive stage, the uncertain parameters are predicted from the feature variables using an ML model. In the prescriptive stage, a decision is prescribed by solving a CO problem using the predicted parameters. This manuscript presents a comprehensive survey of decision-focused learning and makes several contributions. First, to navigate the complex methodologies developed in recent years, the paper proposes the first categorization of DFL methods into four distinct classes: **(1)** analytical differentiation of optimization mappings, **(2)** analytical smoothing of optimization mappings, **(3)** smoothing by random perturbations, and **(4)** differentiation of surrogate loss functions. This categorization, as illustrated in Figure 4, serves as a framework for comprehending and organizing various DFL methodologies. Next, the paper compiles a selection of problem-specific DFL models, making them publicly available to facilitate broader access and usage. An integral part of this paper involves benchmarking the performance of various available methodologies on _seven_ distinct problems. This provides an opportunity for comparative understanding and assists in identifying the relative strengths and weaknesses of each approach. The code and data used in the benchmarking are accessible through [https://github.com/PredOpt/predopt-benchmarks](https://github.com/PredOpt/predopt-benchmarks). Finally, this survey addresses the critical need to look forward, by discussing the outstanding challenges and offering an outlook on potential future directions in the field of DFL. ### Paper organization. Following this introduction, the paper is structured as follows. Preliminary concepts are discussed in Section 2, which introduces the problem setting and explicates the challenges in implementing DFL. The subsequent Section 3 offers a comprehensive review of recently proposed methodologies for handling these challenges, neatly organized into broad classes of related techniques. Secion 4 presents interesting real-world examples of DFL applications. Section 5 brings forth seven benchmark DFL tasks from public datasets, with a comparative evaluation of eight DFL methodologies presented in the following section. The manuscript concludes by providing a discourse on the current challenges and possible future directions in DFL research. ## 2 Preliminaries This section presents an overview of the problem setting, along with preliminary concepts and essential terminology. Then, the central modeling challenges are discussed, setting the stage for a review of current methodologies in the design and implementation of DFL solutions. Throughout the manuscript, vectors are denoted by boldface lowercase letters, such as \(\mathbf{x}\), while scalar components within the vector \(\mathbf{x}\) are represented with a subscript \(i\), denoting the \(i^{\text{th}}\) item within \(\mathbf{x}\) as \(x_{i}\). Similarly, the vectors \(\mathbf{1}\) and \(\mathbf{0}\) symbolize the vector of all-ones and all-zeros, respectively. ### Problem Setting In operations research and business analytics, decisions are often quantitatively modeled using CO problems. These problems model various decision-making scenarios, but may not be efficiently solvable and often demand specialized solution algorithms that are tailored to their specific form. In many real-world applications, some parameters of the CO problems are uncertain and must be inferred from contextual data (hereafter referred to as _features_). The settings considered in this manuscript involve estimating those parameters through predictive inferences made by ML models, and subsequently, the final decisions are modeled as the solution to the CO problems based on those inferences. In this setting, the decision-making processes can be described by _parametric_ CO problems, defined as, \[\mathbf{x}^{\star}(\mathbf{c})=\underset{\mathbf{x}}{\operatorname{ argmin}}\;\;f(\mathbf{x},\mathbf{c}) \tag{1a}\] \[\mathtt{s.t.}\;\;\boldsymbol{g}(\mathbf{x},\mathbf{c}) \leq\mathbf{0}\] (1b) \[\boldsymbol{h}(\mathbf{x},\mathbf{c}) =\mathbf{0}. \tag{1c}\] The goal of the optimization problem above is to find \(a\) solution \(\mathbf{x}^{\star}(\mathbf{c})\in\mathbb{R}^{n}\), a minimizer of the objective function \(f\), satisfying a set \(\boldsymbol{g}\) of inequality and a set \(\boldsymbol{h}\) of equality constraints. The _parametric_ problem formulation defines \(\mathbf{x}^{\star}(\mathbf{c})\) as a function of the parameters \(\mathbf{c}\in\mathbb{R}^{k}\). In the present setting, this function can naturally be interpreted as part of an overall composite function that encompasses ML inference and decision-making, and returns optimal decisions given feature variables as input. CO problems can be categorized in terms of the forms taken by the functions defining their objectives (1a) and constraints (1b-1c). These forms also determine important properties of the optimization mapping \(\mathbf{c}\to\mathbf{x}^{\star}(\mathbf{c})\) when viewed as a function from problem parameters to optimal solutions, such as its continuity, differentiability, and injectivity. In this manuscript, it is assumed that the constraints are fully known prior to solving, i.e., \(\boldsymbol{h}(\mathbf{x},\mathbf{c})=\boldsymbol{h}(\mathbf{x})\) and \(\boldsymbol{g}(\mathbf{x},\mathbf{c})=\boldsymbol{g}(\mathbf{x})\), and restrict the dependence on \(\mathbf{c}\) to the objective function only. This is the setting considered by almost all existing works surveyed. While it is also possible to consider uncertainty in the constraints, this leads to the possibility of predicting parameters that lead to solutions that are infeasible with respect to the ground-truth parameters. The learning problem has not yet been well-defined in this setting (unless a recourse action to correct infeasible solutions is used (Hu, Lee, & Lee, 2022, 2023a)). For this reason, in the following sections, only \(f\) is assumed to depend on \(\mathbf{c}\), so that \(\boldsymbol{g}(\mathbf{x})\leq\mathbf{0}\) and \(\boldsymbol{h}(\mathbf{x})=\mathbf{0}\) are satisfied for all outputs of the decision model. For notational convenience, the feasible region of the CO problem in (1), will be denoted by \(\mathcal{F}\) (i.e., \(\mathbf{x}\in\mathcal{F}\) if and only if \(\boldsymbol{g}(\mathbf{x})\leq\mathbf{0}\) and \(\boldsymbol{h}(\mathbf{x})=\mathbf{0}\)). If the true parameters \(\mathbf{c}\) are known exactly, the corresponding 'true' optimal decisions may be computed by solving (1). In such scenarios, \(\mathbf{x}^{\star}(\mathbf{c})\) will referred to as the _full-information optimal decisions_(Bertsimas & Kallus, 2020). This paper, instead, considers problems where the parameters \(\mathbf{c}\) are unknown but can be estimated as a function of empirically observed features \(\mathbf{z}\). The problem of estimating \(\mathbf{c}\) falls under the category of supervised machine learning problems. In this setting, a set of past observation pairs \(\{(\mathbf{z}_{i},\mathbf{c}_{i})\}_{i=1}^{N}\) is available and used to train a ML model \(m_{\boldsymbol{\omega}}\) (with trainable parameters \(\boldsymbol{\omega}\)), so that parameter predictions take the form \(\mathbf{\hat{c}}=m_{\boldsymbol{\omega}}(\mathbf{z})\). Then, a decision \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) can be made based on the predicted parameters. \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) is referred to as a _prescriptive decision_. The overall learning goal is to optimize the set of prescriptive decisions made over a distribution of feature variables \(\mathbf{z}\sim\mathcal{Z}\), with respect to some evaluation criterion on those decisions. Thus, while the machine learning model \(m_{\boldsymbol{\omega}}\) is trained to predict \(\mathbf{\hat{c}}\), its performance is evaluated on the basis of the corresponding optimal solutions \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). This paper uses the terminology _Predict-Then-Optimize_ problem to refer to the problem of predicting \(\mathbf{\hat{c}}\), to improve the evaluation of \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). ### Learning Paradigms The defining challenge of the Predict-Then-Optimize problem setting is the gap in modeling between the prediction and the optimization components: while \(m_{\mathbf{\omega}}\) is trained to predict \(\mathbf{\hat{c}}\), it is evaluated based on the subsequently computed \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). Using standard ML approaches, learning of the predictions \(\mathbf{\hat{c}}=m_{\mathbf{\omega}}(\mathbf{z})\) can only be supervised by the ground-truth \(\mathbf{c}\) under standard loss functions \(\mathcal{L}\), such as mean squared error or cross-entropy. In principle, it is favorable to train \(m_{\mathbf{\omega}}\) to make predictions \(\mathbf{\hat{c}}\) that optimize the evaluation criterion on \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) directly. This distinction motivates the definition of two alternative learning paradigms for Predict-Then-Optimize problems. Prediction-focused learning (PFL).A straightforward approach to this supervised ML problem is to train the model to generate accurate parameter predictions \(\mathbf{\hat{c}}\) with respect to ground-truth values \(\mathbf{c}\). This paper introduces the term _prediction-focused learning_ to refer to this approach (also called two-stage learning (Wilder, Dilkina, & Tambe, 2019a)) because the model is trained with a focus on the accuracy of the parameter predictions preceding the decision model. Here, the training is agnostic of the downstream optimization problem. At the time of making the decision, the pre-trained model's predictions \(\mathbf{\hat{c}}\) are passed to optimization routines which solve (1) to return \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). Typical ML losses, such as the mean squared error (MSE) or binary cross entropy (BCE), are used to train the prediction model in this case. \[MSE(\mathbf{\hat{c}},\mathbf{c})=\frac{1}{N}\|\mathbf{c}-\mathbf{\hat{c}}\|^{2} \tag{2}\] Such loss functions, like Eq. (2), which measure the prediction error of \(\mathbf{\hat{c}}\) with respect to \(\mathbf{c}\), are referred to as _prediction losses_. Algorithm 1 illustrates prediction-focused learning using the MSE loss. Decision-focused learning (DFL).By contrast, in _decision-focused_ learning, the ML model is trained to optimize the evaluation criteria which measure the quality of the resulting decisions. As the decisions are realized after the optimization stage, this requires the integration of prediction and optimization components, into a composite model which produces full decisions. From this point of view, generating the predicted parameters \(\mathbf{\hat{c}}\) is an intermediary step of the integrated approach, and the accuracy of \(\mathbf{\hat{c}}\) is not the primary focus in training. The focus, rather, is on the error incurred after optimization. A measure of error with respect to the integrated model's prescriptive decisions, when used as a loss function for training, is henceforth referred to as a _task loss_. The essential difference from the aforementioned prediction loss is that it measures the error in \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\), rather than in \(\mathbf{\hat{c}}\). The objective value achieved by using the predicted \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) is generally suboptimal with respect to the true objective parameters \(\mathbf{c}\). Often, the end goal is to generate predictions \(\mathbf{\hat{c}}\) with an optimal solution \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) whose objective value in practice (i.e., \(f(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})\)) comes close to the full-information optimal value \(f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{c})\). In such cases, a salient notion of task loss is the _regret_, defined as the difference between the full-information optimal objective value and the objective value realized by the prescriptive decision. Equivalently, it is the magnitude of suboptimality of the decision \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) with respect to the optimal solution \(\mathbf{x}^{\star}(\mathbf{c})\) under ground-truth parameters \(\mathbf{c}\): \[\textit{Regret}(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})=f(\mathbf{x} ^{\star}(\mathbf{\hat{c}}),\mathbf{c})-f(\mathbf{x}^{\star}(\mathbf{c}), \mathbf{c}) \tag{3}\] Note that minimizing regret is equivalent to minimizing the value of \(f(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})\), since the term \(f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{c})\) is constant with respect to the prediction model. While regret may be considered the quintessential example of a task loss, other task losses can arise in practice. For example, when the ground-truth target data are observed in terms of decision values \(\mathbf{x}\), rather than parameter values \(\mathbf{c}\), they may be targeted using the typical training loss functions such as \(MSE(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{x})\). Relationship between prediction and task losses.As previously mentioned, an ML model is trained without considering the downstream CO problem in prediction-focused learning for Predict-Then-Optimize tasks; still the ML model is evaluated at test time on the basis of its resulting CO problem solutions. This is based on an underlying assumption that generating accurate predictions with respect to a standard prediction loss will result in good prescriptive decisions. Note that zero prediction loss always implies zero task loss, since \(\mathbf{\hat{c}}=\mathbf{c}\) implies \(\mathbf{x}^{\star}(\mathbf{\hat{c}})=\mathbf{x}^{\star}(\mathbf{c})\). However, in practice, it is impossible to learn a model that makes no prediction error on any sample. The model error can only be minimized in one metric, and the minimization of the prediction error and the resulting decision error do not in general coincide (Wilder et al., 2019). Furthermore, the prediction loss and the task loss are, in general, not continuously related. These principles are illustrated by the following example: Example.The shortcomings of training with respect to prediction errors can be illustrated with a relatively simple CO problem. For this illustration, consider a knapsack problem (Pisinger & Toth, 1998). The objective of the knapsack problem is to select a subset of maximal value from an overall set of items, each having its own value and unit weight, subject to a capacity constraint. The capacity constraint imposes that the sum of the weights of the selected items cannot be higher than the capacity \(C\). This knapsack problem with unit weights can be formulated as follows: \[\mathbf{x}^{\star}(\mathbf{c})=\operatorname*{argmax}_{\mathbf{x}\in\{0,1\}} \mathbf{c}^{\top}\mathbf{x}\ \ \texttt{s.t.}\sum_{i}x_{i}\leq\text{Capacity} \tag{4}\] In a Predict-Then-Optimize variant of this knapsack problem, the item weights and knapsack capacity are known, but the item values are unknown and must be predicted using observed features. The ground-truth item value \(\mathbf{c}\) implies the ground-truth solution \(\mathbf{x}^{\star}(\mathbf{c})\). Overestimating the values of the items that are chosen in \(\mathbf{x}^{\star}(\mathbf{c})\) (or underestimating the values of the items that are not chosen) increases the prediction error. Note that these kind of prediction errors, even if they are high, do not affect the solution, and thus do not affect the task loss either. On the other hand, even low prediction errors for some item values may change the solution, affecting the task loss. That is why after a certain point, reducing prediction errors does not decrease task loss, and sometimes may increase it. DFL aims to address this shortcoming of PFL: by minimizing the task loss directly, prediction errors are implicitly traded off on the basis of how they affect the resulting decision errors. The discrepancy between the prediction loss and the task loss has been exemplified in Figure 2 for a very simple knapsack problem with only two items. For this illustration, assume that both the items are of unit weights and the capacity of the knapsack is one, i.e., only one of the two items can be selected. The true values of the first and second items are 2.5 and 3 respectively. The point \((2.5,3)\), marked with \(\clubsuit\), represents the true item values. In this case the true solution is \((0,1)\), which corresponds to selecting only the second item. It is evident that any prediction in the blue shaded region leads to this solution. For instance, the point \((1.5,3)\), marked with \(\clubsuit\), corresponds to predicting \(1.5\) and \(3\) as values of the two items respectively and this results in selecting the second item. On the other hand, the point \((2.5,2)\), marked with \(\clubsuit\), triggers the wrong solution \((1,0)\), although the squared error values of \(\clubsuit\) and \(\clubsuit\) are identical. Also, note that overestimating the value of the second item does not change the solution. For instance, the point \((1.5,4)\), marked with \(\clubsuit\), corresponds to overestimating the value of the second item to \(4\) while keeping the value of the first item the same as the point in \(\clubsuit\). This point is positioned directly above the point in \(\clubsuit\) and still stays in the blue-shaded region. Similarly, the point \((0.5,3)\), marked with \(\clubsuit\), results from underestimating the value of the first item and is in the blue shaded region too. Although these two points have higher values of squared error than the point marked with \(\clubsuit\), they trigger the right solution, resulting in zero regret. Empirical risk minimization and bilevel form of DFL.The minimization of either the prediction loss in PFL or the task loss in DFL, can be expressed as an _empirical risk minimization_ (ERM) (Vapnik, 1999) problem over a training dataset containing feature variables and their corresponding parameters \(\mathcal{D}\equiv\{(\mathbf{z_{i}},\mathbf{c_{i}})\}_{i=1}^{N}\). For concreteness, the respective ERM problems below assume the use of the MSE and regret loss functions, but the principles described here hold for a wide range of alternative loss functions. Figure 2: An illustrative numerical example with a knapsack problem with two items to exemplify the discrepancy between prediction error and regret. The figure illustrates that two points can have the same prediction error but different regret. Furthermore, it demonstrates that overestimating the values of the selected items or underestimating the values of the items that are left out does not change the solution, and thus does not increase the regret, even though the prediction error does increase. PFL, by minimizing the prediction error with respect to the ground-truth parameters directly, takes the form of a standard regression problem: \[\min_{\mathbf{\omega}}\frac{1}{N}\sum_{i=1}^{N}\|m_{\mathbf{\omega}}(\mathbf{z_{i}})- \mathbf{c_{i}}\|^{2}, \tag{5}\] which is an instance of unconstrained optimization. In the case of DFL, it is natural to view the ERM as a bilevel optimization problem: \[\min_{\mathbf{\omega}}\frac{1}{N}\sum_{i=1}^{N}\left(f(\mathbf{x^{*}} (\mathbf{\hat{c}_{i}}),\mathbf{c_{i}})-f(\mathbf{x^{*}}(\mathbf{c_{i}}), \mathbf{c_{i}})\right) \tag{6a}\] \[\mathtt{s.t.}\ \ \mathbf{\hat{c}_{i}}=m_{\mathbf{\omega}}(\mathbf{z_{i}}); \ \mathbf{x^{*}}(\mathbf{\hat{c}_{i}})=\operatorname*{argmin}_{\mathbf{x}\in \mathcal{F}}\mathbf{\hat{c}_{i}}^{\top}\mathbf{x}. \tag{6b}\] The outer-level problem (6a) minimizes task loss on the training set while the inner-level problem (6b) computes the mapping \(\mathbf{c}\rightarrow\mathbf{x^{*}}(\mathbf{c})\). Solving (6) is computationally more challenging than solving (5) in the prediction-focused paradigm. In both cases, optimization by stochastic gradient descent (SGD) is the preferred solution method for training neural networks. Algorithms 1 and 2 compare the gradient descent training schemes for each of these problems. Algorithm 1 is a standard application of gradient descent, in which the derivatives of Line 6 are generally well-defined and can be computed straightforwardly (typically by automatic differentiation). Line 7 of Algorithm 2 shows that direct differentiation of the mapping \(\mathbf{c}\rightarrow\mathbf{x^{*}}(\mathbf{c})\) can be used to form the overall task loss gradient \(\frac{d\mathcal{L}}{d\mathbf{\omega}}\), by providing the required chain rule term \(\frac{d\mathbf{x^{*}}(\mathbf{\hat{c}})}{d\mathbf{\hat{c}}}\). However, this differentiation is nontrivial as the mapping itself lacks a closed-form representation. Further, many interesting and practical optimization problems are inherently nondifferentiable and even discontinuous as functions of their parameters, precluding the direct application of Algorithm 2 to optimize (6) by gradient descent. The following subsections review the main challenges of implementing Algorithm 2. ``` 0: training data D\(\equiv\{(\mathbf{z_{i}},\mathbf{c_{i}})\}_{i=1}^{N}\) Hyperparams: \(\alpha\)- learning rate 1: Initialize \(\mathbf{\omega}\). 2:for each epoch do 3:for each instance \((\mathbf{z},\mathbf{c})\) do 4:\(\mathbf{\hat{c}}=m_{\mathbf{\omega}}(\mathbf{z})\) 5:\(\mathcal{L}=(\mathbf{\hat{c}}-\mathbf{c})^{2}\) 6:\(\mathbf{\omega}\leftarrow\mathbf{\omega}-\alpha\frac{d\mathcal{L}}{d\mathbf{\hat{c}} }\frac{d\mathbf{\hat{c}}}{d\mathbf{\omega}}\) 7:endfor 8:endfor ``` **Algorithm 1** Gradient-descent in prediction-focused learning ``` 0:\(\mathcal{F}\), training data D\(\equiv\{(\mathbf{z_{i}},\mathbf{c_{i}},\mathbf{x^{*}}(\mathbf{c_{i}})\}_{i=1}^{N}\); Hyperparams: \(\alpha\)- learning rate 1:Initialize \(\mathbf{\omega}\). 2:for each epoch do 3:for each instance \((\mathbf{z},\mathbf{c},\mathbf{x^{*}}(\mathbf{c}))\) do 4:\(\mathbf{\hat{c}}=m_{\mathbf{\omega}}(\mathbf{z})\) 5:\(\mathbf{x^{*}}(\mathbf{\hat{c}})=\operatorname*{argmin}_{\mathbf{x}\in\mathcal{ F}}f(\mathbf{x},\mathbf{\hat{c}})\) 6:\(\mathcal{L}=f(\mathbf{x^{*}}(\mathbf{\hat{c}}),\mathbf{c})-f(\mathbf{x^{*}}( \mathbf{c}),\mathbf{c})\) 7:\(\mathbf{\omega}\leftarrow\mathbf{\omega}-\alpha\frac{d\mathcal{L}}{d\mathbf{x^{*}}( \mathbf{\hat{c}})}\frac{d\mathbf{x^{*}}(\mathbf{\hat{c}})}{d\mathbf{\hat{c} }}\frac{d\mathbf{\hat{c}}}{d\mathbf{\omega}}\) 8:endfor 9:endfor ``` **Algorithm 2** Gradient-descent in decision-focused learning with regret as task loss ### Challenges to Implement DFL Differentiation of CO mappings.To minimize a task loss by gradient descent training, its partial derivatives with respect to the prediction model parameters \(\mathbf{\omega}\) must be computed to carry out at the parameter update at Line 7 of Algorithm 2. Since the task loss \(\mathcal{L}\) is a function of \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\), the gradient of \(\mathcal{L}\) with respect to \(\mathbf{\omega}\) can be expressed in the following terms by using the chain rule of differentiation: \[\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})}{d\mathbf{ \omega}}=\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})}{ d\mathbf{x}^{\star}(\mathbf{\hat{c}})}\frac{d\mathbf{x}^{\star}(\mathbf{\hat{c}})}{d \mathbf{\hat{c}}}\frac{d\mathbf{\hat{c}}}{d\mathbf{\omega}} \tag{7}\] The first term in the right side of (7), can be computed directly as \(\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{c})\) is typically a differentiable function of \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). A deep learning library (such as TensorFlow (Abadi, Agarwal, Barham, Brevdo, Chen, Citro, Corrado, Davis, Dean, Devin, Ghemawat, Goodfellow, Harp, Irving, Isard, Jia, Jozefowicz, Kaiser, Kudlur, Levenberg, Mane, Monga, Moore, Murray, Olah, Schuster, Shlens, Steiner, Sutskever, Talwar, Tucker, Vanhoucke, Vasudevan, Viegas, Vinyals, Warden, Wattenberg, Wicke, Yu, & Zheng, 2015), PyTorch (Paszke, Gross, Massa, Lerer, Bradbury, Chanan, Killeen, Lin, Gimelshein, Antiga, et al., 2019)) computes the last term by representing the neural network as a computational graph and applying automatic differentiation (autodiff) in the reverse mode (Baydin, Pearlmutter, Radul, & Siskind, 2018). However, the second term, \(\frac{d\mathbf{x}^{\star}(\mathbf{\hat{c}})}{d\mathbf{\hat{c}}}\), may be nontrivial to compute given the presence of two major challenges: **(1)** The mapping \(\mathbf{\hat{c}}\rightarrow\mathbf{x}^{\star}(\mathbf{\hat{c}})\), as defined by the solution to an optimization problem, _lacks a closed form_ which can be differentiated directly, and **(2)** for many interesting and useful optimization models, the mapping is _nondifferentiable_ in some points, and has zero-valued gradients in others, precluding the straightforward use of gradient descent. As shown in the next subsection, even the class of linear programming problems, widely used in decision modeling, is affected by both issues. Section 3 details the various existing approaches aimed at overcoming these challenges. Computational costAnother major challenge in decision-focused learning is the computational resources required to train the integrated prediction and optimization model. Note that Line 5 in Algorithm 2 evaluates \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\). This requires solving and differentiating the underlying optimization problem for each observed data sample, in each epoch. This imposes a significant computational cost even when dealing with small-scale and efficiently solvable problems, but can become an impediment in the case of large and (NP-)hard optimization problems. Section 3 reviews the techniques proposed thus far for reducing the computational demands of DFL and improving scalability. ### Optimization Problem Forms The effectiveness of solving an optimization problem depends on the specific forms of the objective and constraint functions. Considerable effort has been made to developing efficient algorithms for certain optimization forms. Below, the readers are provided an overview of the key and widely utilized types of optimization problem formulations. #### 2.4.1 Convex optimization In _convex_ optimization problems, a convex objective function is to be optimized over a convex feasible space. This class of problems is distinguished by the guarantee that any locally optimal is also globally optimal (Boyd, Boyd, & Vandenberghe, 2004). Since many optimization methods converge provably to local minima, convex problems are considered to be reliably and efficiently solvable relative to _nonconvex_ problems. Despite this, convex optimization mappings still impose significant computational overhead on Algorithm 2 since they must be solved for each data sample in each epoch, and most convex optimizations are orders of magnitude more complex than conventional neural network layers (Amos & Kolter, 2017). Like all parametric optimization problems, convex ones are implicitly defined mappings from parameters to optimal solutions, lacking a closed form that can be differentiated directly. However as detailed in Section 3.1, they can be canonicalized to a standard form, which facilitates automation of their solution and backpropagation by a single standardized procedure (Agrawal, Amos, Barratt, Boyd, Diamond, & Kolter, 2019). The class of convex problems is broad enough to include some which yield mappings \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) that are differentiable everywhere, and some which do not. The _linear programs_, which are convex and form nondifferentiable mappings with respect to their objective parameters, are notable examples of the latter case and are discussed next. The portfolio optimization problem (44), which contains both linear and quadratic constraints, provides an example of a parametric convex problem which admits useful gradients over some regions of its parameter space and not others. Where the (quadratic) variance constraint (44b) is not active, it behaves as a linear program. Elsewhere, the optimal solution is a smooth function of its parameters. #### 2.4.2 Linear programming _Linear programs_ (LPs) are convex optimization problems whose objective and constraints are composed of affine functions. These programs are predominant as decision models in operations research, and have endless industrial applications since the allocation and transfer of resources is typically modeled by linear relationships between variables (Bazaraa, Jarvis, & Sherali, 2008). The parametric LPs considered in this manuscript take the following Figure 3: In decision-focused learning, the neural network model is trained to minimize the task loss form: \[\mathbf{x}^{\star}(\mathbf{c}) =\operatorname*{argmin}_{\mathbf{x}}\mathbf{c}^{\top}\mathbf{x}\] (8a) s.t. \[A\mathbf{x} =\mathbf{b} \tag{8b}\] \[\mathbf{x} \geq\mathbf{0} \tag{8c}\] Compared to other classes of convex problems, LPs admit efficient solution methods, even for large-scale problems (Bazaraa et al., 2008; Ignizio and Cavalier, 1994). From a DFL standpoint, however, LPs pose a challenge, because the mapping \(\mathbf{c}\to\mathbf{x}^{\star}(\mathbf{c})\) is nondifferentiable. Although the derivatives of mapping (8) are defined almost everywhere, they provide no useful information for gradient descent training. To see this, first note the well-known fact that a linear program always takes its optimal value at a vertex of its feasible set (Bazaraa et al., 2008). Since the number of vertices in any such set is finite, (8) maps a continuous parameter space to a discrete set of solutions. As such, it is a piecewise constant mapping. Therefore its derivatives are zero almost everywhere, and undefined elsewhere. Prevalent strategies for incorporating linear programs in decision-focused learning thus typically rely on differentiating smooth approximations to the LP, as detailed in Section 3.2. Many operations research problems, such as the allocation and planning of resources, can be modeled as LPs. Also many prototypical problems in algorithm design (e.g., sorting and top-\(k\) selection) can be formulated as LPs with continuous variables, despite admitting only discrete integer solutions, by relying on the total unimodularity of the constraint matrices (Bazaraa et al., 2008). In what follows, some examples of machine learning models of LPs and how they might occur in a Predict-Then-Optimize context are given. **Shortest paths.**: Given a directed graph with a given start and end node, the goal in the shortest path problem is to find a sequence of arcs of minimal total length that connects the start end the end node. The decision variables are binary indicators of each edge's inclusion in the path. The linear constraints ensure \([0,1]\) bounds on each indicator, as well as flow balance through each node. These flow balance constraints capture that, except for the start and end node, each node has as many incoming selected arcs as outgoing selected arcs. For the start node, there is one additional outgoing selected arc, and for the end node, there is one more incoming selected arc. The parameters in the linear objective represent the arc lengths. In many realistic settings--as well as in several common DFL benchmarks (Elmachtoub and Grigas, 2022; Pogancic et al., 2020)--these are unknown, requiring them to be predicted from features before a shortest path can be computed. This motivating example captures the realistic setting in which the shortest route between two locations has to be computed, but in which the road traversal times are uncertain (due to unknown traffic conditions, for example), and have to be predicted from known features (such as day of the week, time of day and weather conditions). **Bipartite matching.**: Given is a graph consisting of two sets of nodes, and arcs connecting each node of the first set to each node of the second. The arcs are weighted but the weights are unknown and must be predicted. The optimization task is to choose a subset of arcs such that each node from each set is involved in a selected arc at most once, and the total weight of the selected arcs is maximized. The variables lie in \([0,1]\) and indicate the inclusion of each edge. The constraints ensure that each node is involved at most once in a selected arc. The objective parameters represent arc weights. With a complete bipartite graph, matchings can be construed as permutations, and are presented a permutation matrices, which can be employed in tasks such as learning to rank (Kotary, Fioretto, Van Hentenryck, & Zhu, 2022). **Sorting and Ranking.**: The sorting of any list of predicted values can be posed as a linear program over a feasible region whose vertices correspond to all of the possible permutations of the list. The related ranking, or argsort problem assigns to any length-\(n\) list a permutation of sequential integers \([n]\) which sorts the list. By smoothing the linear program, these basic operations can be differentiated and backpropagated (Blondel, Teboul, Berthet, & Djolonga, 2020). **Top-\(k\) selection.**: Given a set of items and item values that must be predicted, the task is to choose the subset of size \(k\) with the largest total value in selected items. In addition to \([0,1]\) bounds on the indicator variables, a single linear constraint ensures that the selected item indicators sum to \(k\). A prevalent example can be found in multilabel classification (Amos, Koltun, & Kolter, 2019; Martins & Astudillo, 2016). **Computing the maximum.**: This is a special case of top-\(k\) selection where \(k=1\). When the LP's objective is regularized with the entropy term \(H(\mathbf{x})=\mathbf{x}\cdot\log\mathbf{x}\), the mapping from predicted values to optimal solutions is equivalent to a softmax function (Agrawal et al., 2019). **Max-flow/ min-cut.**: Given a network with predefined source and sink nodes, and predicted flow capacities on each arc, the task is to find the maximum flow rate that can be channeled from source to sink. Here the predicted flow capacities occupy the right-hand side of the linear constraints, which is not in line with the DFL problem description given in subsection 2.1. However, in the related min-cut problem--which is equivalent to the dual linear program of the max-flow problem--the flow capacities are the parameters in the objective function. The max-flow problem can thus be cast as an equivalent min-cut problem and DFL can be used to learn to predict the flow capacities. #### 2.4.3 Integer linear programming Integer Linear Programs (ILPs) are another mainstay in operations research and computer science. ILPs differ from LPs in that the decision variables \(\mathbf{x}\) are restricted to integer values, i.e., \(\mathbf{x}\in\mathbb{Z}^{k}\) where \(\mathbb{Z}^{k}\) is the set of integral vectors of appropriate dimensions. Like LPs, ILPs are challenging to use in DFL because they yield discontinuous, nondifferentiable mappings. Computationally however, they are more challenging due to their NP-hard complexity, which may preclude the exact computation of the mapping \(\mathbf{\hat{c}}\rightarrow\mathbf{x}^{\star}(\mathbf{\hat{c}})\) at each step of Algorithm 2. Their differentiation is also significantly more challenging, since the discontinuity of their feasible regions prevents many smoothing techniques that can be applied in DFL with LPs. In the following, examples of how ILPs may occur in a Predict-Then-Optimize setting are provided. **Knapsack.**: The knapsack problem has been used as a benchmark in several papers about DFL (Mandi, Demirovic, Stuckey, & Guns, 2020; Mandi & Guns, 2020; Demirovic, Stuckey, Bailey, Chan, Leckie, Ramamohanarao, & Guns, 2019). Given are weights of a set of items, as well as a capacity. The items also have associated values, which have to be predicted from features. The optimization task involves selecting a subset of the items that maximizes the sum of the weights associated with the selected items, whilst ensuring that the sum of the associated weights does not exceed the capacity. **Travelling salesperson problem**: In the travelling salesperson problem, the list of cities, and the distances between each pair of cities, is given. The goal is to find a path of minimal length that visits each city exactly once. In the Predict-Then-Optimize setting, the distances between the cities first have to be predicted (Pogancic et al., 2020) from observable empirical data. **Combinatorial portfolio optimization.**: Portfolio optimization involves making optimal investment decisions across a range of financial assets. In the combinatorial Predict-Then-Optimize variant, the decisions are discrete, and must be made on the basis of the predicted next period's increase in the value of several assets (Ferber, Wilder, Dilkina, & Tambe, 2020). **Diverse bipartite matching.**: Diverse bipartite matching problems are similar to the bipartite matching problems described in 2.4.2, but are subject to additional diversity constraints (Ferber et al., 2020; Mulamba, Mandi, Diligenti, Lombardi, Bucarey, & Guns, 2021; Mandi, Bucarey, Tchomba, & Guns, 2022) In this variant, edges have additional properties. The diversity constraints enforce lower and upper bounds on the proportion of edges selected with a certain property. This precludes the LP formulation, and makes the use of ILP more interesting. **Energy-cost aware scheduling.**: Energy-cost aware scheduling involves scheduling a set of tasks across a set of machines in a way that minimizes the overall energy cost involved. As future energy costs are unknown, they first have to be predicted (Mulamba et al., 2021; Mandi et al., 2020, 2022; Mandi & Guns, 2020). #### 2.4.4 Integer nonlinear programming In integer nonlinear programming, the objective function and/or the constraints are nonlinear. Performing DFL on integer nonlinear programs faces the same challenges as performing DFL on ILPs: integer nonlinear programs are computationally expensive to solve, are implicit mappings with zero-valued gradients almost everywhere, and have discontinuous feasible regions, hindering the use of the smoothing techniques that can be applied in DFL with LPs. Additionally, because of their nonlinear nature, many of the techniques developed for DFL with ILPs, which assume linearity, do not work on integer nonlinear programs (Elmachtoub & Grigas, 2022; Pogancic et al., 2020). To the best of our knowledge, no DFL method has specifically been developed for or tested on integer nonlinear programs. The most closely related work is (Ferber, Huang, Zha, Schubert, Steiner, Dilkina, & Tian, 2022), which learns an approximate ILP surrogates for integer nonlinear programs, which could then in turn be used in a DFL loop. ## 3 Review of Decision-focused Learning Methodologies To the best of our knowledge, this manuscript is the first to provide a comprehensive review of methods developed for and suitable for DFL in gradient-based training. Concurrently with this paper, Sadana, Chenreddy, Delage, Forel, Frejinger, and Vidal (2023) survey recently proposed approaches to address Predict-Then-Optimize problems (referred to as contextual optimization within it). While Sadana et al. (2023) also cover some of the works to be surveyed later, this manuscript goes beyond by presenting an extensive review solely focused on DFL methods and proposing the first categorization of existing DFL methods. This section will describe several methodologies which address the challenge of differentiating an optimization mapping for DFL in gradient-based training. In essence, different approaches propose different smoothed surrogate approximations of \(\frac{d\mathcal{X}^{\star}(\mathbbm{e})}{d\mathbbm{e}}\) or \(\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbbm{e}))}{d\mathbbm{e}}\), which is used for backpropagation. This paper proposes the first categorization of existing DFL approaches into the following four distinct classes: **Analytical Differentiation of Optimization Mappings:**: Methodologies under this category aim to compute exact derivative for backpropagation by differentiating the optimality conditions for certain optimization problem forms, for which the derivative exists and non-zero. **Analytical Smoothing of Optimization Mappings:**: These approaches deal with combinatorial optimization problems (for which the analytical derivatives are zero almost everywhere) by performing smoothing of combinatorial optimization problems, which results in approximate problems that can be differentiated analytically. **Smoothing by Random Perturbations:**: Methodologies under this category utilize implicit regularization through perturbations, constructing smooth approximations of optimization mappings. **Differentiation of Surrogate Loss Functions:**: Methodologies under this category propose convex surrogate loss functions of specific task loss such as regret. **Decision-Focused Learning without Optimization in the Loop:**: These methodologies bypass the need for computing \(\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbbm{e}))}{d\mathbbm{e}}\) by utilizing surrogate losses, which reflect the quality of the decisions, but do not require computing the solution of the optimization problem for differentiation. Figure 4 presents key characteristics of these four methodology classes, highlighting the types of problems that can be addressed within each class. Next, each category is thoroughly described. ### Analytical Differentiation of Optimization Mappings As discussed before, differentiating through parametric CO problems comes with two main challenges. First, since CO problems are complex, implicitly defined mappings from parameters to solutions, computing the derivatives is not straightforward. Second, since some CO problems result in piecewise-constant mappings, their derivatives are zero almost everywhere, and do not exist elsewhere. This subsection pertains to CO problems for which the second challenge does not apply, i.e., problems that are smooth mappings. For these problems, all that is required to implement DFL is direct differentiation of the mapping in Eq. (1). Differentiating unconstrained relaxations.An early work discussing differentiation through constrained argmin problems in the context of machine learning is (Gould, Fernando, Cherian, Anderson, Cruz, & Guo, 2016). It first proposes a technique to differentiate the argmin of a smooth, _unconstrained_ convex function. When \(V(\mathbf{c})=\operatorname*{argmin}_{\mathbf{x}}f(\mathbf{c},\mathbf{x})\), it can be shown that when all second derivatives of \(f\) exist, \[\frac{dV(\mathbf{c})}{d\mathbf{c}}=-\frac{f_{\mathbf{ex}}(\mathbf{c},V( \mathbf{c}))}{f_{\mathbf{xx}}(\mathbf{c},V(\mathbf{c}))} \tag{9}\] where \(f_{\mathbf{ex}}\) is the second partial derivative of \(f\) with respect to \(\mathbf{c}\) followed by \(\mathbf{x}\). This follows from implicit differentiation of the first-order optimality conditions \[\frac{df}{d\mathbf{x}}(\mathbf{c},V(\mathbf{c}))=0 \tag{10}\] with respect to \(\mathbf{c}\), and rearranging terms. Here the variables \(\mathbf{c}\) are the optimization problem's defining parameters, and the variables \(\mathbf{x}\) are the decision variables. Figure 4: An overview of the categorization of DFL methodologies in four classes. This technique is then extended to find approximate derivatives to _constrained_ problems with inequality constraints \(g_{i}(\mathbf{c},\mathbf{x})\leq 0\), \(1\leq i\leq m\), by first relaxing the problem to an unconstrained problem, by means of the log-barrier function \[F(\mathbf{c},\mathbf{x})=f(\mathbf{c},\mathbf{x})-\mu\sum_{i}\log(-g_{i}( \mathbf{c},\mathbf{x})) \tag{11}\] and then differentiating \(\operatorname*{argmin}_{\mathbf{x}}F(\mathbf{c},\mathbf{x})\) with respect to \(\mathbf{c}\) for some choice of the scaling factor \(\mu\). Since this approach relies on approximations and requires hyperparameter tuning for the factor \(\mu\), subsequent works focus on differentiating constrained optimization problems directly via their own global conditions for optimality, as discussed next. Differentiating KKT conditions of quadratic programs.More recent approaches are based on differentiating the optimality conditions of a CO problem directly, i.e., without first converting it to an unconstrained problem. Consider an optimization problem and its optimal solution: \[\mathbf{x}^{\star}=\operatorname*{argmax}_{\mathbf{x}} f(\mathbf{x}) \tag{12a}\] \[\mathtt{s.t.} \boldsymbol{g}(\mathbf{x})\leq 0\] (12b) \[\boldsymbol{h}(\mathbf{x})=0 \tag{12c}\] and assume that \(f\), \(g\) and \(h\) are differentiable functions of \(\mathbf{x}\). The Karush-Kuhn-Tucker (KKT) conditions are a set of equations expressing optimality conditions for a solution \(\mathbf{x}^{\star}\) of problem (12) (Boyd et al., 2004): \[\nabla f(\mathbf{x}^{\star})+\sum_{i}w_{i}\nabla h_{i}(\mathbf{x }^{\star})+\sum_{j}u_{j}\nabla g_{j}(\mathbf{x}^{\star})=0 \tag{13a}\] \[g_{j}(\mathbf{x}^{\star})\leq 0 \ \forall j\] (13b) \[h_{i}(\mathbf{x}^{\star})=0 \ \forall i\] (13c) \[u_{j}\geq 0 \ \forall j\] (13d) \[u_{j}g_{j}(\mathbf{x}^{\star})=0 \ \forall j \tag{13e}\] OptNet is a framework developed by Amos and Kolter (2017) to differentiate through optimization mappings that are convex quadratic programs (QPs) by differentiating through these KKT conditions. In convex quadratic programs, the objective \(f\) is a convex quadratic function and the constraint functions \(g,h\) are linear over a continuous domain. In the most general case, each of \(f\), \(g\) and \(h\) are dependent on a distinct set of parameters, in addition to the optimization variable \(\mathbf{x}\): \[f(\mathbf{c},Q,\mathbf{x})=\frac{1}{2}\mathbf{x}^{\top}Q\mathbf{ x}+\mathbf{c}^{\top}\mathbf{x} \tag{14a}\] \[g(A,\mathbf{b},\mathbf{x})=R\mathbf{x}-\mathbf{s}\] (14b) \[h(R,\mathbf{s},\mathbf{x})=A\mathbf{x}-\mathbf{b} \tag{14c}\] When \(\mathbf{x}\in\mathbb{R}^{k}\) and the number of equality constraints is \(M_{in}\) and \(M_{eq}\), respectively, a QP problem is specified by parameters \(Q\in\mathbb{R}^{k\times k}\), \(\mathbf{c}\in\mathbb{R}^{k}\), \(R\in\mathbb{R}^{k\times M_{in}}\), \(\mathbf{s}\in\mathbb{R}^{M_{in}}\), \(A\in\mathbb{R}^{k\times M_{eq}}\) and \(\mathbf{b}\in\mathbb{R}^{M_{eq}}\). Note that for this problem to be convex, \(Q\) must be positive-semidefinite always, which can be ensured by learning instead parameters \(\mathbf{q}\in\mathbb{R}^{k}\) and taking \(Q=\mathbf{q}^{\top}\mathbf{q}\). A defining characteristic of quadratic programs such as (14) is their straightforward parameterization. This is due to the fact that any linear or quadratic function can be fully specified by a square matrix or a vector of parameters, respectively. Here, problem (12) is viewed as a mapping \((Q,\mathbf{c},R,\mathbf{s},A,\mathbf{b})\to\mathbf{x}^{\star}(Q,\mathbf{c},R, \mathbf{s},A,\mathbf{b})\), which parameterizes the space of all possible quadratic programs and their solutions. The presence of such a canonical form allows for separation of a problem's inherent structure from its parameters (Grant and Boyd, 2008), and is key to creating a differentiable mapping from parameters to optimal solutions in an automated way, without necessitating additional analytical transformations. The gradients are sought with respect to each of the parameters in \((Q,\mathbf{c},R,\mathbf{s},A,\mathbf{b})\). For this purpose, Amos and Kolter (2017) argue that the inequalities (13b) and (13d) can be dropped, resulting in a system of equalities representing optimality conditions for \(\mathbf{x}^{\star}\). Exact gradients \(\frac{d\mathbf{x}^{\star}}{dP}\) for any \(P\in\{Q,\mathbf{c},R,\mathbf{s},A,\mathbf{b}\}\) can then be retrieved by solving the differential KKT conditions: \[\begin{bmatrix}Q&R^{\top}&A^{\top}\\ D(\mathbf{w}^{*})R&D(R\mathbf{x}^{*}-\mathbf{s})&0\\ A&0&0\end{bmatrix}\begin{bmatrix}d\mathbf{x}\\ d\mathbf{w}\\ d\mathbf{u}\end{bmatrix}=-\begin{bmatrix}dQ\mathbf{x}^{*}+d\mathbf{c}+dR^{\top}\mathbf{ w}^{*}+dA^{\top}\mathbf{u}^{*}\\ D(\mathbf{w}^{*})dR\mathbf{x}^{*}-D(\mathbf{w}^{*})d\mathbf{s}\\ dA\mathbf{x}^{*}-d\mathbf{b}\end{bmatrix} \tag{15}\] where the shorthand \(d\) stands for the derivative \(\frac{d}{dP}\). This is another example of implicit differentiation, and requires solving a linear system of equations. Later, Konishi and Fukunaga (2021) extended the method of Amos and Kolter (2017), where they compute the second order derivative of the solution. This allows to train gradient boosting models, which require the gradient as well as the Hessian matrix of the loss. In summary, the techniques in this category compute the derivatives of the solution with respect to the parameters (if they exist) by leveraging implicit differentiation of the KKT conditions. Differentiating optimality conditions of conic programs.Another class of problems with a parametric canonical form are the conic programs, which take the form: \[\mathbf{x}^{\star}(A,\mathbf{b},\mathbf{c})=\operatorname*{argmax }_{\mathbf{x}}\ \ \mathbf{c}^{\top}\mathbf{x} \tag{16a}\] \[\mathtt{s.t.}\ \ A\mathbf{x}-\mathbf{b}\in\mathcal{K} \tag{16b}\] where \(\mathcal{K}\) is a nonempty, closed, convex cone. A framework for differentiating the mapping (16) for any \(\mathcal{K}\) is proposed in (Agrawal, Barratt, Boyd, Busseti, & Moursi, 2019c), which starts by forming the homogeneous self-dual embedding of (16), whose parameters form askew-symmetric block matrix composed of \(A\), \(\mathbf{b}\), and \(\mathbf{c}\). Following (Busseti, Moursi, & Boyd, 2019), the solution to this embedding is expressed as the problem of finding a zero of a mapping containing a skew-symmetric linear function and projections onto the cone \(\mathcal{K}\) and its dual. The zero-value of this function is implicitly differentiated, in a similar manner to the KKT conditions of a quadratic program as in (Amos and Kolter, 2017). The overall mapping (16) is viewed as the composition of function that maps \((A,\mathbf{b},\mathbf{c})\) onto the skew-symmetric parameter space of the self-dual embedding, the rootfinding problem that produces a solution to the embedding, and a transformation back to a solution of the primal and dual problems. The overall derivative is found by a chain rule applied over this composition. Subsequent work (Agrawal et al., 2019a) leverages the above-described differentiation of cone programs to develop a more general differentiable convex optimization solver--Cvxpylayers. It is well known that conic programs of the form (16) can provide canonical representations of convex programs (Nemirovski, 2007). The approach described by Agrawal et al. (2019a) is based on this principle; that a large class of _parametric_ convex optimization problems can be recast as equivalent parametric cone programs, with an appropriate choice of the cone \(\mathcal{K}\). A major benefit of this representation is that it allows a convex program to be separated with respect to its defining parameters \((A,\mathbf{b},\mathbf{c})\) and its structure \(\mathcal{K}\), allowing a generic procedure to be applied for solving and differentiating the transformed problem with respect to \(A\), \(\mathbf{b}\) and \(\mathbf{c}\). The framework for transforming convex programs to cone programs of the form (16) is drawn from (Grant & Boyd, 2008), which is based on two related concepts. First is the notion of _disciplined convex programming_, which assists the automation of cone transforms by imposing a set of rules or conventions on how convex programs can be represented. Second is the notion of _graph implementations_, which represent functions as optimization problems over their epigraphs, for the purpose of generically representing optimization problems and assisting conversion between equivalent forms. The associated software system called cvx allows for disciplined convex programs to be converted to cone programs via their graph implementations. Subsequently, the transformed problem is solved using conic optimization algorithms, and its optimal solution is converted to a solution of the original disciplined convex program. Differentiation is performed through each operation and combined by the chain rule. The transformation of parameters between respective problem forms, and the solution recovery step, are differentiable by virtue of being affine mappings (Agrawal et al., 2019a). The intermediate conic program is differentiated via the methods of (Agrawal et al., 2019c). Solver unrolling and fixed-point differentiation.While the methods described above for differentiation through CO problems are generic and applicable to broad classes of problems, other practical techniques have been proven effective and even advantageous in some cases. A common strategy is that of solver _unrolling_, in which the solution to (1) is found by executing an optimization algorithm in the computational graph of the predictive model. Then, the mapping (1) is backpropagated simply by automatic differentiation or 'unrolling' through each step of the algorithm, thus avoiding the need to explicitly model \(\frac{d\mathbf{x}^{*}(\mathbf{c})}{d\mathbf{c}}\)(Domke, 2012). While this approach leads to accurate backpropagation in many cases, it suffers disadvantages in efficiency due to the memory and computational resources required to store and apply backpropagation over the entire computational graph of an algorithm that requires many iterations (Amos & Kolter, 2017). Additionally, it has been observed that unrolling over many solver iterations can leads to vanishing gradient issues reminiscent of recurrent neural networks (Monga, Li, & Eldar, 2021). On the other hand, unrolling allows for the learning of unspecified algorithm parameters, such as gradient descent step sizes or weights in an augmented lagrangian, which can be exploited to accelerate the forward-pass convergence of the optimization solver. A comprehensive survey of algorithm unrolling for image processing applications is provided in (Monga et al., 2021). Another way in which a specific solution algorithm may provide gradients though a corresponding optimization mapping, is by implicit differentiation of its fixed-point conditions. Suppose that the solver iterations \[\mathbf{x}_{k+1}(\mathbf{c})=\mathcal{U}(\mathbf{x}_{k}(\mathbf{c}),\ \mathbf{c}) \tag{17}\] converge as \(k\rightarrow\infty\) to a solution \(\mathbf{x}^{\star}(\mathbf{c})\) of the problem (1), then the fixed-point conditions \[\mathbf{x}^{\star}(\mathbf{c})=\mathcal{U}(\mathbf{x}^{\star}(\mathbf{c}),\ \mathbf{c}) \tag{18}\] are satisfied. Assuming the existence of all derivatives on an open set containing \(\mathbf{c}\) to satisfy the implicit function theorem, it follows by implicit differentiation with respect to \(\mathbf{c}\) that \[(I-\Phi)\frac{d\mathbf{x}^{\star}}{d\mathbf{c}}=\Psi, \tag{19}\] which is a linear system to be solved for \(\frac{d\mathbf{x}^{\star}}{d\mathbf{c}}\), in terms of \(\Phi=\frac{d\mathcal{U}}{d\mathbf{x}^{\star}}(\mathbf{x}^{\star}(\mathbf{c}),\ c)\) and \(\Psi=\frac{d\mathcal{U}}{\mathbf{c}}(\mathbf{x}^{\star}(\mathbf{c}),\ \mathbf{c})\). The relationship between unrolling and differentiation of the fixed-point conditions is studied by Kotary, Dinh, and Fioretto (2023), which shows that backpropagation of (1) by unrolling (17) is equivalent to solving the linear system (19) by fixed-point iteration. As such, the convergence rate of the backward pass in unrolling is determined by the convergence rate of the equivalent linear system solve, and can be calculated in terms of the spectral radius of \(\Phi\). Discussion.In contrast to most other differentiable optimization methods surveyed in this article, the analytical approaches in this subsection allow for backpropagation of coefficients that specify the constraints as well the objective function. For example, Amos and Kolter (2017) propose parametric quadratic programming layers whose linear objective parameters are predicted by previous layers, and whose constraints are learned through the layer's own embedded parameters. This is distinct from most cases of DFL, in which the optimization problems have fixed constraints and no trainable parameters of their own. Furthermore, the techniques surveyed in this subsection are aimed at computing exact gradients of parametric optimization mappings. However, many applications of DFL contain optimization mappings that are discontinuous and piecewise-constant. Such mappings, including parametric linear programs (8), have gradients that are zero almost everywhere and thus do not supply useful descent directions for SGD training. Therefore, the techniques of this subsection are often applied after regularizing the problem analytically with smooth functions, as detailed in the next subsection. ### Analytical Smoothing of Optimization Mappings To differentiate through combinatorial optimization problems, the optimization mapping first has to be smoothed. While techniques such as noise-based gradient estimation (surveyed in Section 3.3) provide smoothing and differentiation simultaneously, analytical differentiation first incorporates smooth analytical terms in the optimization problem's formulation, and then analytically differentiates the resulting optimization problem using the techniques discussed in Section 3.1. Analytical smoothing of linear programs.Note that while an LP problem is convex and has continuous variables, only a finite number of its feasible solutions can potentially be optimal. These points coincide with the vertices of its feasible polytope (Bazaraa et al., 2008). Therefore the mapping \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) in (8), as a function of \(\mathbf{\hat{c}}\), is discontinuous and piecewise constant, and thus requires smoothing before it can be differentiated through. An approach to do so was presented in Wilder et al. (2019), which proposes to augment the linear LP objective function with the Euclidean norm of its decision variables, so that the new objective takes the following form \[\mathbf{x}^{\star}(\mathbf{c}) =\operatorname*{argmax}_{\mathbf{x}}\mathbf{c}^{\top}\mathbf{x}- \mu\|\mathbf{x}\|_{2}^{2} \tag{20a}\] \[=\operatorname*{argmin}_{\mathbf{x}}\|\mathbf{x}-\frac{\mathbf{c }}{\mu}\|_{2}^{2} \tag{20b}\] where the above equality follows from expanding the square and cancelling constant terms, which do not affect the argmax. This provides an intuition as to the effect of such a quadratic regularization: it converts a LP problem into that of projecting the point \(\frac{\mathbf{c}}{\mu}\) onto the feasible polytope, which results in a continuous mapping \(\mathbf{\hat{c}}\to\mathbf{x}^{\star}(\mathbf{\hat{c}})\). Wilder et al. (2019) then train decision-focused models by solving and backpropagating the respective quadratic programming problem using the framework of (Amos and Kolter, 2017), in order to learn to predict objective parameters with minimal regret. At test time, the quadratic smoothing term is removed. This article refers to such regret-based DFL with quadratically regularized linear programs as the _Quadratic Programming Task Loss_ method (QPTL). Other forms of analytical smoothing for linear programs can be applied by adding different regularization functions to the objective function. Some common regularization terms for LPs include the entropy function \(H(\mathbf{x})=\sum_{i}x_{i}\log x_{i}\) and the binary entropy function \(H_{b}(\mathbf{x})=H(\mathbf{x})+H(\mathbf{1}-\mathbf{x})\). To differentiate the resulting smoothed optimization problems, the framework of Agrawal et al. (2019) can be used. Alternatively, problem-specific approaches that do not employ (Agrawal et al., 2019) have also been proposed. For example, (Blondel et al., 2020) proposes a method for problems where \(H\) smooths an LP for differentiable sorting and ranking, and (Amos et al., 2019) proposes a way to differentiate through problems where \(H_{b}\) is used in a multilabel classification problem. Both works propose fast implementations for both the forward and backward passes of their respective optimization problems. In a related approach, Mandi and Guns (2020) propose a general, differentiable LP solver based on log-barrier regularization. For a parametrized LP of standard form (8), gradients are computed for the regularized form in which the constraints \(\mathbf{x}\geq\mathbf{0}\) are replaced with log-barrier approximations: \[\mathbf{x}^{\star}(\mathbf{c})=\operatorname*{argmin}_{\mathbf{x}} \mathbf{c}^{\top}\mathbf{x}+\lambda\sum_{i}x_{i} \tag{21a}\] \[\text{s.t. }A\mathbf{x}=\mathbf{b} \tag{21b}\] While similar in this sense to (Gould et al., 2016), this method exploits several efficiencies specific to linear programming, in which the log-barrier term serves a dual purpose of rendering (21) differentiable and also aiding its solution. Rather than forming and solving this regularized LP problem directly, the solver uses an interior point method to produce a sequence of log-barrier approximations to the LP's homogenous self-dual (HSD) embedding. Early stopping is applied in the interior point method, producing a solution to (21) for some \(\lambda\), which serves as a smooth surrogate problem for differentiation. A major advantage of this technique is that it only requires optimization of a linear program, making it in general more efficient than direct solution of a regularized problem as in the approaches described above. Analytical smoothing of integer linear programs.To differentiate through ILPs, Wilder et al. (2019) propose to simply drop the integrality constraints, and to then smooth and differentiate through the resulting LP relaxation, which is observed to give satisfactory performance in some cases. Ferber et al. (2020) later extended this work by using a more systematic approach to generate the LP relaxation of the ILP problem. They use the method of cutting planes to discover an LP problem that admits the same solution as the ILP. Subsequently, the method of (Wilder et al., 2019) is applied to approximate the LP mapping's derivatives. Although this results in enhanced performance with respect to regret, there are some practical scalability concerns, since the cut generation process is time consuming but also must be repeated for each instance in each training epoch. ### Smoothing by Random Perturbations A central challenge in DFL is the need for smoothing operations of non-smooth optimization mappings. Techniques that perform the smoothing operation by adding explicit regularization functions to the optimization problems' objective function have been surveyed in Section 3.2. This section instead surveys techniques, which use implicit regularization via perturbations. These techniques construct smooth approximations of the optimization mappings by adopting a probabilistic point of view. To introduce this point of view, the CO problem in this section is not viewed as a mapping from \(\mathbf{c}\) to \(\mathbf{x}^{\star}(\mathbf{c})\). Rather, it is viewed as a function that maps \(\mathbf{c}\) onto a probability distribution over the feasible region \(\mathcal{F}\). From this perspective, \(\mathbf{x}^{\star}(\mathbf{c})\) can be viewed as a random variable, conditionally dependent on \(\mathbf{c}\). The motivation behind representing \(\mathbf{x}^{\star}(\mathbf{c})\) as a random variable is that the rich literature of likelihood maximization with latent variables, in fields such as Probabilisic Graphical Models (PGMs) (Koller and Friedman, 2009), can be exploited. Implicit differentiation by perturbation.One seminal work in the field of PGMs is by Domke (2010). This work contains an important proposition, which deals with a setup where a variable \(\boldsymbol{\theta}_{1}\) is conditionally dependent on another variable \(\boldsymbol{\theta}_{2}\) and the final loss \(\mathcal{L}\) is defined on the variable \(\boldsymbol{\theta}_{1}\). Let \(p(\boldsymbol{\theta}_{1}|\boldsymbol{\theta}_{2})\) and \(\mathbb{E}[\boldsymbol{\theta}_{1}|\boldsymbol{\theta}_{2}]\) be the conditional distribution and the conditional mean of \(\boldsymbol{\theta}_{1}\). The loss \(\mathcal{L}\) is measured on the conditional mean \(\mathbb{E}[\boldsymbol{\theta}_{1}|\boldsymbol{\theta}_{2}]\) and the goal is to compute the derivative of \(\mathcal{L}\) with respect to \(\boldsymbol{\theta}_{2}\). Domke (2010) proposes that the derivative of \(\mathcal{L}\) with respect to \(\boldsymbol{\theta}_{2}\) can be approximated by the following finite difference method: \[\frac{dL}{d\boldsymbol{\theta}_{2}}\approx\frac{1}{\delta}\Bigg{(}\mathbb{E}[ \boldsymbol{\theta}_{1}|\big{(}\boldsymbol{\theta}_{2}+\delta\frac{d}{d \boldsymbol{\theta}_{1}}\big{(}\mathcal{L}(\mathbb{E}[\boldsymbol{\theta}_{1} |\boldsymbol{\theta}_{2}]\big{)}\big{)}]-\mathbb{E}[\boldsymbol{\theta}_{1}| \boldsymbol{\theta}_{2}]\Bigg{)} \tag{22}\] where \(\frac{d}{d\boldsymbol{\theta}_{1}}[\mathcal{L}(\mathbb{E}[\boldsymbol{\theta} _{1}])]\) is the derivative \(\mathcal{L}\) with respect to \(\boldsymbol{\theta}_{1}\) at \(\mathbb{E}[\boldsymbol{\theta}_{1}]\). Notice that the first term in (22) is the conditional mean after perturbing the parameter \(\boldsymbol{\theta}_{2}\) where magnitude of the perturbation is modulated by the derivative of \(\mathcal{L}\) with respect to \(\mathbf{\theta}_{1}\). Taking inspiration from this proposition, by defining a conditional distribution \(p(\mathbf{x}^{\star}(\mathbf{\hat{c}})|\hat{c})\), one can compute the derivative of the regret with respect to \(\mathbf{\hat{c}}\) in the context of DFL. To perfectly represent the deterministic mapping \(\mathbf{c}\rightarrow\mathbf{x}^{\star}(\mathbf{c})\), the straightforward choice is to define a Dirac mass distribution, which assigns all probability mass to the optimal point and none to other points, i.e., \[p(\mathbf{x}|\mathbf{c})=\begin{cases}1&\mathbf{x}=\mathbf{x}^{\star}(\mathbf{ c})\\ 0&\text{otherwise}\end{cases} \tag{23}\] Differentiation of blackbox combinatorial solvers.Note that with the distribution in (23) \(\mathbb{E}_{\mathbf{x}\sim p(\mathbf{x}|\mathbf{c})}[x|c]=\mathbf{x}^{\star}( \mathbf{c})\). Hence, using conditional probability in the proposition in (22), \(\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))}{d\mathbf{\hat{c}}}\) can be computed in the following way: \[\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))}{d\mathbf{\hat{c}}} \approx\nabla^{(DBB)}\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))=\left( \mathbf{x}^{\star}\Big{(}\mathbf{\hat{c}}+\delta\frac{d\mathcal{L}(\mathbf{x} ^{\star}(\mathbf{\hat{c}}))}{d\mathbf{x}^{\star}(\mathbf{\hat{c}})}\Big{)}- \mathbf{x}^{\star}\Big{(}\mathbf{\hat{c}}\Big{)}\right) \tag{24}\] The gradient computation methodology proposed by Pogancic et al. (2020) takes the form of (24). They interpret it as substituting the jump-discontinuous optimization mapping with a piece-wise linear interpolation. It is a linear interpolation of the mapping \(\mathbf{\hat{c}}\rightarrow\mathbf{x}^{\star}(\mathbf{\hat{c}})\) between the points \(\mathbf{\hat{c}}\) and \(\mathbf{\hat{c}}+\delta\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}) )}{d\mathbf{x}}|_{\mathbf{x}=\mathbf{x}^{\star}(\mathbf{\hat{c}})}\). Pogancic et al. (2020) call this 'differentiation of blackbox' (DBB) solvers, because this approach considers the CO solver as a blackbox oracle, i.e., it does not take cognizance of how the solver works internally. In a subsequent work, Sahoo, Paulus, Vlastelica, Musil, Kuleshov, and Martius (2023) propose to treat \(\frac{d\mathbf{x}^{\star}(\mathbf{\hat{c}})}{d\mathbf{\hat{c}}}\) as a negative identity matrix while backpropagating the loss. However, they notice that such an approach might run into unstable learning for scale-invariant optimization problems such as LPs and ILPs. To negate this effect, they suggest multiplying the cost vector with the matrix of the invariant transformation. In case of LPs and ILPs this can be achieved by normalizing the cost vector through projection onto the unit sphere. Perturb-and-MAP.However, at this point it is worth mentioning that Domke (2010) assumes, in his proposition, that the distribution \(p(\theta_{1}|\theta_{2})\) in (22) belongs to the exponential family of distributions (Barndorff-Nielsen, 1978). Note that the distribution defined in (23) is not a distribution of the exponential family. Nevertheless, a tempered softmax distribution belonging to exponential family can be defined to express the mapping in the following way: \[p_{\tau}(\mathbf{x}|\mathbf{c})=\begin{cases}\frac{\exp(-f(\mathbf{x},\mathbf{ c})/\tau)}{\sum_{\mathbf{x}^{\prime}\in\mathcal{F}}\exp(-f(\mathbf{x}^{\prime}, \mathbf{c})/\tau)}&\mathbf{x}\in\mathcal{F}\\ 0&\text{otherwise}\end{cases} \tag{25}\] In this case, the log unnormalized probability mass at each \(\mathbf{x}\in\mathcal{F}\) is proportional to \(\exp(-f(\mathbf{x},\mathbf{c})/\tau)\), the exponential of the negative of the tempered objective value. The idea behind (25) is to assign a probability to each feasible solution such that solutions with a better objective value have a larger probability. The parameter \(\tau\) affects the way in which objective values map to probabilities. When \(\tau\to 0\), the distribution becomes the argmax distribution in (23), when \(\tau\rightarrow\infty\), the distribution becomes uniform. In other words, the value of \(\tau\) determines how drastically the probability changes because of a change in objective value. Good values for \(\tau\) are problem-dependent, and thus tuning \(\tau\) is advised. Note that (22) deals with conditional expectation. As in the case of tempered softmax distribution, the conditional expectation is not always equal to the solution to the CO problem, it must be computed first to use the finite difference method in (22). However, computing the probability distribution function in (25) is not tractable, as the denominator (also called the partition function) requires iterating over all feasible points in \(\mathcal{F}\). Instead, Papandreou and Yuille (2011) propose a novel approach, known as _perturb-and-MAP_, to estimate the probability using perturbations. It states that the distribution of the maximizer after perturbing the log unnormalized probability mass by i.i.d. \(\text{Gumbel}(0,\epsilon)\) noise has the same exponential distribution as (25). To make it more explicit, if \(\tilde{\mathbf{c}}=\mathbf{c}+\boldsymbol{\eta}\), where the perturbation vector \(\boldsymbol{\eta}\stackrel{{\text{i.i.d.}}}{{\sim}}\text{Gumbel}(0,\epsilon)\), \[\mathbb{P}[\mathbf{x}=\underset{\mathbf{x}^{\prime}}{\text{ argmax}}-f(\mathbf{x}^{\prime},\tilde{\mathbf{c}})]=p_{\epsilon}(\mathbf{x}| \mathbf{c}) \tag{26}\] The perturb-and-MAP framework can be viewed as a method of stochastic smoothing (Abernethy, Lee, & Tewari, 2016). A smoothed approximation of the optimization mapping is created by considering the average value of the solutions of a set of _nearby perturbed_ points. With the help of (26), the conditional distribution and hence the conditional mean can be approximated by Monte Carlo simulation. Differentiable perturbed optimizers.Berthet, Blondel, Teboul, Cuturi, Vert, and Bach (2020) propose another approach for perturbation-based differentiation. They name it differentiable perturbed optimizers (DPO). They make use of the perturb-and-MAP framework to draw samples from the conditional distribution \(p(\mathbf{x}|\mathbf{c})\). In particular, they use the reparameterization trick (Kingma & Welling, 2014; Rezende, Mohamed, & Wierstra, 2014) to generate samples from \(p(\mathbf{x}|\mathbf{c})\). The reparameterization trick uses a change of variables to rewrite \(\mathbf{x}\) as a _deterministic function_ of \(\mathbf{c}\) and a random variable \(\boldsymbol{\eta}\). In this reformulation, \(\mathbf{x}\) is still a random variable, but the randomness comes from the variable \(\boldsymbol{\eta}\). They consider \(\boldsymbol{\eta}\) to be a random variable having a density proportional to \(\exp(-\nu(\boldsymbol{\eta}))\) for a twice-differentiable function \(\nu\). Moreover, they propose to multiply the random variable \(\boldsymbol{\eta}\) with a temperature parameter \(\epsilon>0\), which controls the strength of perturbing \(\mathbf{c}\) by the random variable \(\boldsymbol{\eta}\). In summary, first \(\mathbf{c}\) is perturbed with random perturbation vector \(\epsilon\boldsymbol{\eta}\), where \(\boldsymbol{\eta}\) is sampled from the aforementioned density function, and then the maximizer of the perturbed vector \(c+\epsilon\boldsymbol{\eta}\) is viewed as a sample from the conditional distribution for given values of \(\mathbf{c}\) and \(\epsilon\), i.e., \(\mathbf{x}_{\epsilon}^{\star}(\mathbf{c})=\mathbf{x}^{\star}(\mathbf{c}+ \epsilon\boldsymbol{\eta})\) is considered as a sample drawn from \(p(\mathbf{x}|\mathbf{c})\) for a given \(\epsilon\). They call \(\mathbf{x}_{\epsilon}^{\star}(\mathbf{c})\) a _perturbed optimizer_. Note that, for \(\epsilon\to 0\), \(\mathbf{x}_{\epsilon}^{\star}(\mathbf{c})\rightarrow\mathbf{x}^{\star}( \mathbf{c})\). Like before, \(\mathbf{x}_{\epsilon}^{\star}(c)\) can be estimated by Monte Carlo simulation by sampling i.i.d. random noise \(\boldsymbol{\eta}^{(m)}\) from the aforementioned density function. The advantage is that the Monte Carlo estimate is _continuously_ differentiable with respect to \(\mathbf{c}\). This Monte Carlo estimate \(\bar{\mathbf{x}}_{\epsilon}^{\star}(\mathbf{c})\) can be expressed as: \[\bar{\mathbf{x}}_{\epsilon}^{\star}(\mathbf{c})=\frac{1}{M}\sum_{m=1}^{M} \mathbf{x}^{\star}\Big{(}\mathbf{c}+\epsilon\boldsymbol{\eta}^{(m)}\Big{)} \tag{27}\] Moreover, its derivative can be estimated by Monte Carlo simulation too \[\frac{d\bar{\mathbf{x}}_{\epsilon}^{\star}(\mathbf{c})}{d\mathbf{c}}=\frac{1}{ \epsilon}\frac{1}{M}\sum_{m=1}^{M}\mathbf{x}^{\star}(\mathbf{c}+\epsilon\mathbf{ \eta}^{(m)})\nu^{\prime}(\mathbf{\eta}^{(m)})^{\top} \tag{28}\] where \(\nu^{\prime}\) is the first order derivative of \(\nu\). They can approximate \(\frac{d\mathbf{x}^{\star}(\mathbf{c})}{d\mathbf{c}}\) by \(\frac{d\bar{\mathbf{x}}_{\epsilon}^{\star}(\mathbf{c})}{d\mathbf{c}}\) to implement the backward pass. As mentioned before, if \(\epsilon\to 0\), the estimation will be an unbiased estimate of \(\mathbf{x}^{\star}(\mathbf{c})\). However, in practice, for low values of \(\epsilon\), the variance of the Monte-Carlo estimator will increase, leading to unstable and noisy gradients. This is in line with the smoothing-versus-accuracy trade-off mentioned before. Berthet et al. (2020) use this DPO framework to differentiate any optimization problem with linear objective. For a CO problem with discrete feasible space, they consider the convex hull of the discrete feasible region. Furthermore, Berthet et al. (2020) construct the Fenchel-Young loss function and show for Fenchel-Young loss function, the gradient can be approximated in the following way: \[\nabla\mathcal{L}^{FY}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))=-\big{(}\bar{ \mathbf{x}}_{\epsilon}^{\star}(\mathbf{\hat{c}})-\mathbf{x}^{\star}(\mathbf{c })\big{)} \tag{29}\] In a later work, Dalle, Baty, Bouvier, and Parmentier (2022) extend the perturbation approach, where they consider multiplicative perturbation. This is useful when the cost parameter vector is restricted to be non-negative, such as in the applications of shortest path problem variants. The work of Paulus, Choi, Tarlow, Krause, and Maddison (2020) can also be viewed as an extension of the DPO framework. They introduce stochastic softmax tricks (SST), a framework of Gumbel-softmax distribution, where they propose differentiable methods by sampling from more complex categorical distributions. Implicit maximum likelihood estimation (I-MLE).Niepert, Minervini, and Franceschi (2021) also use the _perturb-and-MAP_ framework. However, they do not sample noise from the Gumbel distribution, rather they report better results when the noise \(\mathbf{\eta}^{\gamma}\) is sampled from a _Sum-of-Gamma_ distribution with hyperparameter \(\gamma\). Combining the finite difference approximation (22) with the perturb-and-MAP framework, the gradient takes the following form: \[\frac{d\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))}{d\mathbf{\hat{c}}} \approx\nabla^{(IMLE)}\mathcal{L}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))= \left(\mathbf{x}^{\star}\Big{(}\mathbf{\hat{c}}+\delta\frac{d\mathcal{L}( \mathbf{x}^{\star}(\mathbf{\hat{c}}))}{d\mathbf{x}^{\star}(\mathbf{\hat{c}})}+ \epsilon\mathbf{\eta}^{\gamma}\Big{)}-\mathbf{x}^{\star}\Big{(}\mathbf{\hat{c}}+ \epsilon\mathbf{\eta}^{\gamma}\Big{)}\right) \tag{30}\] where \(\epsilon>0\) is a temperature parameter, which controls the strength of noise perturbation. Clearly, (30) turns into (24) when there is no noise perturbation, i.e., if \(\mathbf{\eta}^{\gamma}=0\). Discussion.One major advantage of the methodologies explained in this subsection is that for gradient computation they call the optimization solver as a blackbox oracle and only use the solution returned by it for gradient computation. In essence, these techniques are not concerned with _how_ the CO problem is solved. The users can utilize any techniques of their choice--constraint programming (CP) (Rossi, van Beek, & Walsh, 2006), Boolean satisfiability (SAT) (Gomes, Kautz, Sabharwal, & Selman, 2008) or linear programming (LP) and integer linear programming (ILP) to solve the CO problem. ### Differentiation of Surrogate Loss Functions The methodologies explained in the preceding subsections can be viewed as implementations of differentiable optimization layers, which solve the CO problem in the forward pass and return useful approximations of \(\frac{d\mathbf{x}^{\star}(\mathbf{c})}{d\mathbf{c}}\) in the backward pass. Consequently, those methodologies can be used to introduce optimization layers _anywhere in a neural network architecture_, and can be combined with arbitrary loss functions. In contrast, the methodologies that will be introduced next can only be used to differentiate regret (3)--a specific task loss. Hence, models can only be trained in an end-to-end fashion using these techniques when the CO problem occurs in the _final_ stage of the pipeline, as in the case of Predict-Then-Optimize problems. Also note that the computation of the regret requires both the ground-truth cost vector \(\mathbf{c}\); as well as ground-truth solution \(\mathbf{x}^{\star}(\mathbf{c})\). If \(\mathbf{c}\) is observed, \(\mathbf{x}^{\star}(\mathbf{c})\) can be computed. However, if only \(\mathbf{x}^{\star}(\mathbf{c})\) is observed, \(\mathbf{c}\) cannot directly be recovered. Hence, the techniques that will be discussed next are not suitable when the true cost vectors \(\mathbf{c}\) are not observed in the training data. Smart "Predict, Then Optimize".Elmachtoub and Grigas (2022) developed Smart "Predict, Then Optimize" (SPO), a seminal work in DFL. As the gradient of the regret with respect to cost vector \(\mathbf{\hat{c}}\) is zero almost everywhere, SPO instead uses a surrogate loss function that has subgradients which _are_ useful in training. They start by proposing a convex surrogate upper bound of regret, which they call the SPO+ loss. \[\mathcal{L}_{SPO+}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))=2\mathbf{\hat{c}}^{ \top}\mathbf{x}^{\star}(\mathbf{c})-\mathbf{c}^{\top}\mathbf{x}^{\star}( \mathbf{c})+\max_{\mathbf{x}\in\mathcal{F}}\{\mathbf{c}^{\top}\mathbf{x}-2 \mathbf{\hat{c}}^{\top}\mathbf{x}\} \tag{31}\] Then, they derive the following useful subgradient of \(\mathcal{L}_{SPO+}(\mathbf{x}^{\star}(\mathbf{\hat{c}}))\): \[\mathbf{x}^{\star}(\mathbf{c})-\mathbf{x}^{\star}(2\mathbf{\hat{c}}-\mathbf{ c})\in\partial\mathcal{L}_{SPO+} \tag{32}\] This subgradient is used in place of to update the model parameters in the backward pass. From a theoretical point of view, the SPO+ loss has the Fisher consistency property with respect to the regret, under certain distributional assumptions. A surrogate loss function satisfies the Fisher consistency property if the function that minimizes the surrogate loss also minimizes the true loss in expectation (Zou, Zhu, & Hastie, 2008). Concretely, this means that minimizing the SPO+ loss corresponds to minimizing the regret in expectation. While training ML models with a finite dataset, an important property of considerable interest would be _risk bounds_(Massart & Nedelec, 2006). Liu and Grigas (2021) develop risk bounds for SPO+ loss and show that low excess SPO+ loss risk translates to low excess regret risk. Furthermore, El Balghiti, Elmachtoub, Grigas, and Tewari (2019) develop worst-case generalization bounds of the SPO loss. The SPO framework is applicable not only to LPs, but to _any CO problems where the cost parameters appear linearly in the objective function_. This includes QPs, ILPs and MILPs. Mandi et al. (2020) empirically investigated how the framework performs on ILP problems. However, as these problems are much more computationally expensive to solve than the ones considered by Elmachtoub and Grigas (2022), they compared the regular SPO methodology with a variant in which, it is significantly cheaper to solve the CO problem during training. To be specific, they consider LP relaxations of the ILPs These LP relaxations are obtained by considering the continuous relaxation of the ILPs, i.e., they are variants of the ILPs in which the integrality constraints are dropped. Using the LP relaxations significantly expedite training, without any cost: Mandi et al. (2020) did not observe a significant difference in the final achieved regret between these two approaches, with both of them performing better than the prediction-focused approach. However, one should be cautious to generalize this result across different problems, as it might be dependent on the integrality gap between the ILP and its LP relaxation. Next, within this category, a different type of DFL technique is being surveyed. In these DFL techniques, the surrogate loss functions are supposed to reflect the decision quality, but their computations do _not_ involve solving the CO problems, thereby avoiding the zero-gradient problem. Noise contrastive estimation.One such approach is introduced by Mulamba et al. (2021). Although their aim is still to minimize regret, computation of \(\nabla_{\mathbf{\hat{c}}}\text{{Regret}}(\mathbf{x}^{\star}(\mathbf{\hat{c}}), \mathbf{c})\) has been avoided by using a surrogate loss function. In their work, the CO problem is viewed from a probabilistic perspective, as in (25). However, instead of maximum likelihood estimation, the noise contrastive estimation (NCE) (Gutmann & Hyvarinen, 2010) method is adopted. NCE has been extensively applied in many applications such as language modeling (Mnih & Teh, 2012), information retrieval (Huang, He, Gao, Deng, Acero, & Heck, 2013) and entity linking (Gillick, Kulkarni, Lansing, Presta, Baldridge, Ie, & Garcia-Olano, 2019). Its basic idea is to learn to discriminate between data coming from the true underlying distribution and data coming from a noise distribution. In the context of DFL, this involves contrasting the likelihood of ground-truth solution \(\mathbf{x}^{\star}(\mathbf{c})\) and a set of negative examples \(S\). In other words, the following ratio is maximized: \[\max_{\mathbf{\hat{c}}}\sum_{\mathbf{x}^{\prime}\in S}\frac{p_{\tau}(\mathbf{ x}^{\star}(\mathbf{c})|\mathbf{\hat{c}})}{p_{\tau}(\mathbf{x}^{\prime}|\mathbf{ \hat{c}})} \tag{33}\] where \(\mathbf{x}^{\prime}\in S\) is a negative example. Because the probability \(p_{\tau}(\mathbf{x}^{\star}(\mathbf{c})|\mathbf{\hat{c}})\) is defined as in (25), when \(\tau=1\), maximizing (33) corresponds to minimizing the following loss: \[\mathcal{L}_{NCE}(\mathbf{\hat{c}},\mathbf{c})=\sum_{\mathbf{x}^{\prime}\in S }f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{\hat{c}})-f(\mathbf{x}^{\prime}, \mathbf{\hat{c}}) \tag{34}\] In other words, this approach learns to predict a \(\mathbf{\hat{c}}\) for which ground-truth solution \(\mathbf{x}^{\star}(\mathbf{c})\) achieves a good objective value, and for which other feasible solutions \(\mathbf{x}^{\prime}\) achieve worse objective values. Note that when \(f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{\hat{c}})\leq f(\mathbf{x}^{\prime}, \mathbf{\hat{c}})\) for all \(\mathbf{x}^{\prime}\in\mathcal{F}\), it holds that \(\mathbf{x}^{\star}(\mathbf{c})=\mathbf{x}^{\star}(\mathbf{\hat{c}})\), and thus the regret is zero. Also note that computing \(\mathcal{L}_{NCE}(\mathbf{\hat{c}},\mathbf{c})\) does not involve computing \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\), circumventing the zero-gradient problem. As an alternative to NCE, Mulamba et al. (2021) also introduce a maximum a posteriori (MAP) approximation, in which they only contrast the ground-truth solution with the most probable negative example from \(S\) according to the current model: \[\mathcal{L}_{MAP}(\mathbf{\hat{c}},\mathbf{c}) =\max_{\mathbf{x}^{\prime}\in S}f(\mathbf{x}^{\star}(\mathbf{c}), \mathbf{\hat{c}})-f(\mathbf{x}^{\prime},\mathbf{\hat{c}})\] \[=f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{\hat{c}})-f(\mathbf{x}^ {\prime},\mathbf{\hat{c}})\text{ where }\mathbf{x}^{\prime}=\operatorname*{argmin}_{ \mathbf{x}\in S}f(\mathbf{x},\mathbf{\hat{c}}) \tag{35}\] Note that whenever \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\in S\), it holds that \(\mathcal{L}_{MAP}(\mathbf{\hat{c}},\mathbf{c})=f(\mathbf{x}^{\star}(\mathbf{ c}),\mathbf{\hat{c}})-f(\mathbf{x}^{\star}(\mathbf{\hat{c}}),\mathbf{\hat{c}})\). This is also known as _self-contrastive estimation_ (SCE) (Goodfellow, 2015) since the ground-truth is contrasted with the most likely output of the current model itself. Also note that for optimization problems with a linear objective, the losses are \(\mathcal{L}_{NCE}(\mathbf{\hat{c}},\mathbf{c})=\sum_{\mathbf{x}^{\prime}\in S} \mathbf{\hat{c}}^{\top}(\mathbf{x}^{\star}(\mathbf{c})-\mathbf{x}^{\prime})\) and \(\mathcal{L}_{MAP}(\mathbf{\hat{c}},\mathbf{c})=\mathbf{\hat{c}}^{\top}( \mathbf{x}^{\star}(\mathbf{c})-\mathbf{x}^{\prime})\), where \(\mathbf{x}^{\prime}=\operatorname*{argmin}_{\mathbf{x}\in S}f(\mathbf{x}, \mathbf{\hat{c}})\). In order to prevent the model from simply learning to predict \(\mathbf{\hat{c}}=\mathbf{0}\), the following alternate loss functions are proposed for these kinds of problems: \[\mathcal{L}_{NCE}^{(\mathbf{\hat{c}}-\mathbf{c})}(\mathbf{\hat{c}},\mathbf{c} )=\sum_{\mathbf{x}^{\prime}\in S}(\mathbf{\hat{c}}-\mathbf{c})^{\top}( \mathbf{x}^{\star}(\mathbf{c})-\mathbf{x}^{\prime}) \tag{36}\] \[\mathcal{L}_{MAP}^{(\mathbf{\hat{c}}-\mathbf{c})}(\mathbf{\hat{c}},\mathbf{c} )=\max_{\mathbf{x}^{\prime}\in S}(\mathbf{\hat{c}}-\mathbf{c})^{\top}( \mathbf{x}^{\star}(\mathbf{c})-\mathbf{x}^{\prime}) \tag{37}\] Construction of \(S\).Forming \(S\) by sampling points from the feasible region \(\mathcal{F}\) is a crucial part of using the contrastive loss functions. To this end, Mulamba et al. (2021) proposes to construct \(S\) by caching all the optimal solutions in the training data. That is why they name \(S\) as'solution cache'. While training, more feasible points are gradually added to \(S\) by solving for some of the predicted cost vectors. However, in order to avoid computational cost, the solver call is not made for each predicted cost during training. Whether to solve for a predicted cost vector is decided by pure random sampling, i.e., is based on a biased coin toss with probability \(p_{solve}\). Intuitively, the \(p_{solve}\) hyperparameter determines the proportion of instances for which the CO problem is solved during training. Experimentally, it has been reported that \(p_{solve}=5\%\) of the time is often adequate, which translates to solving for only \(5\%\) the predicted instances. This translates to reducing the computational cost by approximately \(95\%\), since solving the CO problems represents the major bottleneck in terms of computation time in DFL training. Approximation of a solver by solution-cache.Furthermore, Mulamba et al. (2021) propose a solver-free training variant for any methodology that treats the optimization solver as a blackbox oracle. Such methodologies include the aforementioned I-MLE, DBB, SPO. In this solver-free implementation, solving the optimization problem is substituted with a cache lookup strategy, where the minimizer within the cache \(S\subset\mathcal{F}\) is considered as a proxy for the solution to the optimization problem (i.e., the minimizer within \(\mathcal{F}\)). This significantly reduces the computational cost as solving an optimization problem is replaced by a linear search within a limited cache. Such an approximation can be useful in case the optimization problem takes long to solve. DFL as a learning to rank (LTR) problem.In a later work, Mandi et al. (2022) observe that \(\mathcal{L}_{NCE}\) (34) can be derived by formulating DFL as a _pairwise learning to rank_ task (Joachims, 2002). The learning to rank task consists of learning the implicit order over the solutions in \(S\) invoked by the objective function values achieved by the solutions with respect to \(\mathbf{c}\). In other words, it involves learning to predict a \(\mathbf{\hat{c}}\) that ranks the solutions in \(S\) similarly to how \(\mathbf{c}\) ranks them. In the _pairwise_ approach, \(\mathbf{x}^{\star}(\mathbf{c})\) and any \(\mathbf{x}^{\prime}\in S\) are treated as a pair and the model is trained to predict \(\mathbf{\hat{c}}\) such that the ordering of each pair is the same for \(\mathbf{c}\) and \(\mathbf{\hat{c}}\). The loss is considered to be zero if \(\mathbf{\hat{c}}^{\top}\mathbf{x}^{\star}(\mathbf{c})\) is smaller than \(\mathbf{\hat{c}}^{\top}\mathbf{x}^{\prime}\) by at least a margin of \(\Theta>0\). The pairwise loss is formally defined in the following form: \[\mathcal{L}_{\text{Pairwise}}(\mathbf{\hat{c}},\mathbf{c})=\sum_{\mathbf{x}^ {\prime}\in S}\max\big{(}0,\Theta+(f(\mathbf{x}^{\star}(\mathbf{c}),\mathbf{ \hat{c}})-f(\mathbf{x}^{\prime},\mathbf{\hat{c}}))\big{)} \tag{38}\] Another loss function is formulated by considering the difference in differences between the objective values at the true optimal \(\mathbf{x}^{\star}(\mathbf{c})\) and non-optimal \(\mathbf{x}^{\prime}\) with \(\mathbf{c}\) and \(\mathbf{\hat{c}}\) as the parameters. \[\mathcal{L}_{\text{PairwiseDifference}}(\mathbf{\hat{c}},\mathbf{c})=\sum_{ \mathbf{x}^{\prime}\in S}\bigg{(}\big{(}f(\mathbf{x}^{\star}(\mathbf{c}), \mathbf{\hat{c}})-f(\mathbf{x}^{\prime},\mathbf{\hat{c}})\big{)}-\big{(}f( \mathbf{x}^{\star}(\mathbf{c}),\mathbf{c})-f(\mathbf{x}^{\prime},\mathbf{c}) \big{)}\bigg{)}^{2} \tag{39}\] Further, motivated by _listwise learning to rank task_(Cao, Qin, Liu, Tsai, & Li, 2007), a loss function is proposed by Mandi et al. (2022) where the ordering of all the items in \(S\) is considered, rather than the ordering of pairs of items. Cao et al. (2007) define this listwise loss based on a _top-one probability_ measure. The top-one probability of an item is the probability of it being the best of the set. Note that such probabilistic interpretation in the context of DFL is already defined in Section 3.3. Mandi et al. (2022) make use of the tempered softmax probability defined in (25). Recall that this \(p_{\tau}(\mathbf{x}|\mathbf{c})\) can be interpreted as a probability measure of \(\mathbf{x}\in\mathcal{F}\) being the minimizer of \(f(\mathbf{x},\mathbf{c})\) in \(\mathcal{F}\) for a given \(\mathbf{c}\). However, as mentioned before, direct computation of \(p_{\tau}(\mathbf{x}|\mathbf{c})\) requires iterating over all feasible points in \(\mathcal{F}\), which is intractable. Therefore Mandi et al. (2022) compute the probability with respect to \(S\subset\mathcal{F}\). This probability measure finally is used to define a listwise loss--the cross-entropy loss between \(p_{\tau}(\mathbf{x}|\mathbf{c})\) and \(p_{\tau}(\mathbf{x}|\mathbf{\hat{c}})\), the distributions obtained for ground-truth \(\mathbf{c}\) and predicted \(\mathbf{\hat{c}}\). This can be written in the following form: \[\mathcal{L}_{\text{Listwise}}(\mathbf{\hat{c}},\mathbf{c})=\bigg{(}-\frac{1} {|S|}\sum_{\mathbf{x}^{\prime}\in S}p_{\tau}(\mathbf{x}^{\prime}|\mathbf{c}) \log p_{\tau}(\mathbf{x}^{\prime}|\mathbf{\hat{c}})\bigg{)} \tag{40}\] The main advantage of (34), (35), (38), (39) and (40) is that they are differentiable and can be computed directly by any neural network library via automatic differentiation. Also note that the computation and differentiation of the loss functions are solver-free, i.e., they need not solve the optimization problem to compute the loss or its derivative. Learning efficient surrogate solvers.Another research direction without optimization in the loop is based on reducing the computational cost associated with repeatedly solving optimization problems, by learning efficiently computable and differentiable surrogate losses that approximate and replace the true task loss. Shah, Wang, Wilder, Perrault, and Tambe (2022) propose to learn a surrogate of the regret function by parametric local losses. Due to the difficulty of learning a single convex surrogate function to estimate regret, a convex local surrogate is learned for each data sample in training. By design, the surrogate losses are automatically differentiable, and thus they eliminate the need for a differentiable optimization solver. ### Discussion So far, in this section, an extensive overview of different DFL methodologies have been provided. For the ease of the readers, a summary of some of the key DFL techniques, discussed so far, have been provided in Table 1. The second column of Table 1 highlights the form of the CO problem applicable to the technique. Note that although some techniques are generally applicable to any optimization problem forms, most techniques have been evaluated so far using CO problems with linear objective functions. The third column \begin{table} \begin{tabular}{c c c c} \hline \hline Methodologies & CO Problem Forms & Computation of Gradient & Optimization \\ & & & Layer \\ \hline OptNet & & Implicit differentiation of & \\ (Amos \& Kolter, 2017) & Convex QPs & KKT conditions & \\ Cvxpylayers & & Implicit differentiation of & \\ (Agrawal et al., 2019a) & Convex problems & & \\ Fold-opt & Convex and nonconvex & & \\ (Kotary et al., 2023) & problems & & \\ \hline QPTL & & Implicit differentiation & \\ (Wilder et al., 2019a) & LPs, ILPs & after transforming into QPs & \\ & & by adding regularizer & \\ & & Implicit differentiation of & \\ & & HSD of (relaxed) LPs by & \\ (Mandi \& Guns, 2020) & & adding log-barrier relaxation & \\ & & & Conversion of ILPs into LPs & \\ & & & by method of cutting planes & \\ & & & before applying QPTL & \\ \hline DBB & Optimization problems & & Differentiation of \\ (Pogancic et al., 2020) & with a linear objective & & linear interpolation & \\ & & & of optimization mapping & \\ & & & Treating the CO solver as & \\ & & with a linear objective & \\ & & & negative identity mapping & \\ & & & \\ & & & Finite difference approximation & \\ & & with a linear objective & \\ & & & with perturb-and-MAP & \\ & & & \\ & & & Differentiation of & \\ & & & perturbed optimizer & \\ & & & \\ & & & Differentiation of perturbed & \\ & & & Fenchel-Young loss & \\ & & & \\ & & & Differentiation of & \\ & & & surrogate SPO+ loss & \\ & & & \\ & & & \\ & & & Differentiation of & \\ & & & surrogate contrastive loss & \\ & & & \\ & & & Differentiation of & \\ & & & surrogate LTR loss & \\ & & & \\ & & & Differentiation of a & \\ & & & \\ & & & learned convex local surrogate loss & \\ \hline \hline \end{tabular} \end{table} Table 1: A concise overview of gradient modeling techniques in key DFL methodologies that use gradient-based learning. summarizes the gradient computation technique. The fourth column indicates whether that particular technique is compatible with any generic task loss. Techniques, termed as implementations of _differentiable optimization layers_, can be embedded in any stage of an NN architecture. The other techniques are applicable where optimization is the final stage of the pipeline (such as in Predict-Then-Optimize problem formulations) and a particular loss (most often regret) is used as the task loss. ### Other Aspects of Decision-Focused Learning In the following, some aspects related to DFL, that have not yet been discussed in this manuscript, will be highlighted. To begin with, it should be noted that certain CO problems may have _multiple non-unique optimal solutions_ for a given cost vector. This can occur when the cost vector of an LP is parallel to one of the faces of the feasible polyhedron. Moreover, problems involving symmetric graphs often exhibit multiple optimal solutions, especially when the problems' solution can be transformed into other solutions through automorphisms (Weisstein, 2000). It is important to note that if the predicted cost vector has multiple non-unique optimal solutions, each of these solutions may have different value of regret. In such scenarios, Elmachtoub and Grigas (2022) propose to consider the worst-case regret. To do so, the set of optimal solutions of \(\hat{\mathbf{c}}\) can be represented by \(\mathcal{X}^{\star}(\hat{\mathbf{c}})\). And then the worst-case regret can be defined in the following form: \[\textit{Regret}(\mathbf{x}^{\star}(\hat{\mathbf{c}}),\mathbf{c})=\max_{ \mathbf{x}^{\star}(\hat{\mathbf{c}})\in\mathcal{X}^{\star}(\hat{\mathbf{c}})} f(\mathbf{x}^{\star}(\hat{\mathbf{c}}),\mathbf{c})-f(\mathbf{x}^{\star}( \mathbf{c}),\mathbf{c}) \tag{41}\] Having addressed the possibility of the presence of multiple non-unique optimal solutions in a CO problem, the focus now turns to other important facets of DFL. #### 3.6.1 Prediction-focused vs. Decision-focused learning DFL methodologies are expected to deliver lower regret than a PFL approach in Predict-Then-Optimize problems, as the ML model is directly trained to achieve low regret. However, as discussed before, the implementation of DFL poses significant challenges. In fact, practitioners may be tempted to resort to a PFL approach to circumvent the computational costs associated with DFL, when dealing with real-world Predict-Then-Optimize problems. To encourage practitioners to adopt DFL methodologies, it is crucial to investigate scenarios where DFL methodologies outperform the PFL approach. To this end, Elmachtoub, Lam, Zhang, and Zhao (2023) conduct a theoretical comparison of the limiting distributions of the optimality gaps between the two approaches in the context of stochastic optimization. They show the PFL approach that does not consider optimization while training the model asymptotically outperforms the integrated prediction and optimization approach, employed by DFL methodologies, _if_ the underlying prediction model is well-specified. This is intuitive, as a well-specified model tends to produce highly accurate predictions, which can contribute to the success of the PFL approach. In such cases, the DFL methodologies might perform worse than PFL since training in DFL involves _approximate_ gradients (because the true gradient is zero almost everywhere), whereas the gradient is well-defined for a PFL approach. On the other hand, they show that if the model is not well-specified, a PFL approach perform suboptimally compared to the DFL approach. Hence, it is recommended to use DFL when there exists aleatoric or epistemic uncertainty. As most real-world settings include various sorts of uncertainty--both aleatoric and epistemic--DFL methodologies are expected to outperform the PFL approach. In a separate work, Cameron, Hartford, Lundy, and Leyton-Brown (2022) show that the suboptimality of the PFL becomes more pronounced in the presence of correlations between the predicted parameters. #### 3.6.2 Alternatives to gradient-based decision-focused learning The methodologies explained so far implement DFL by gradient descent training, which is the go-to approach for training neural networks. However, note that there exist other machine learning frameworks, such as tree-based methods, which do not require gradient-based training. To avoid the problem of zero-valued gradients altogether, several works have considered alternatives to gradient-based learning instead. In SPO Trees (SPOTs) (Elmachtoub, Liang, & McNellis, 2020), the predictive model is a decision tree or ensemble of decision trees. Such models can be learned by recursive partitioning with respect to the regret directly, and thus do not require the use of the SPO+ surrogate loss function introduced by Elmachtoub and Grigas (2022). Alternatively, the tree learning problem can be posed as a MILP and be solved by an off-the-shelf solver, in the same spirit as Jeong, Jaggi, Butler, and Sanner (2022). Jeong et al. (2022) formulate the problem of minimizing regret as a mixed-integer linear program (MILP), when the predictive model is linear. They start from the bilevel optimization formulation, introduced in (6a) and (6b). First, the transition points where the solution of the lower level program (6b) changes are identified, and then the solution space is exhaustively partitioned, and for each partition the solution is annotated. This paves the way to construct a MILP formulation of the outer program (6a). This MILP problem is solved to learn the parameters \(\mathbf{\omega}\) of the linear predictive model. The resulting model is guaranteed to be globally optimal, which is not the case for gradient-based methods that might get stuck in a local optimum. However, their method is limited to ML models that are linear and optimization problems that are binary MILPs. Demirovic, Stuckey, Guns, Bailey, Leckie, Ramamohanarao, and Chan (2020) consider linear ML models and represent the objective function of the CO problem as a piece-wise linear function of the ML parameters. In this proposed technique, the ML parameters are updated via coordinate descent algorithm, where each component of the cost vector is updated at a time to minimize the regret keeping other components fixed. This technique requires identifying the _transition points_, where regret changes, as function of each component of the cost parameter. Demirovic et al. (2020) consider CO problems that can be solved by dynamic programming and identify the transition points using dynamic programming. In a later work, Guler, Demirovic, Chan, Bailey, Leckie, and Stuckey (2022) extend this technique by employing a 'divide-and-conquer' algorithm to identify the transition points for CO problems whose objective function is a bilinear function of the decision variables and the predicted parameters. This development generalizes the previous work (Demirovic et al., 2020) to cover much broader class of CO problems and offers a substantial speed improvement. The 'branch & learn' approach proposed by HU, Lee, Lee, and Zhong (2022), which consider CO problems that can be solved by recursion, also extends this technique. #### 3.6.3 Predicting parameters in the constraints The majority of the works in DFL aim to predict parameters in the objective function and assume that the feasible space is precisely known. However, in many applications the unknown parameters occur in the constraints as well as in the objectives. When the parameters in the constraints are predicted and prescribed decisions are made using the predicted parameters, one major issue is that the prescribed decisions might turn out to be infeasible with respect to the true parameters. In this case, the task loss should not only minimize the suboptimality of the prescribed decisions, but it should also penalize if the prescribed decisions become infeasible. Hence designing DFL algorithms suitable for such problems entails a few additional considerations. The first consideration deals with quantifying the extent of infeasibility when the prescribed decisions become infeasible with respect to the true parameters. In this regard, Garcia, Street, Homem-de Mello, and Munoz (2021) propose to add artificial slack variables with high penalty costs in the objective function to penalize infeasible decisions. In a recent work, Hu et al. (2022) introduce the notion of _post-hoc_ regret, wherein a non-negative penalty is added to regret to account for the conversion of infeasible solutions into feasible ones. This idea of a penalty function shares a fundamental resemblance to the concept of recourse action in stochastic programming (Ruszczynski and Shapiro, 2003). In a later work, Hu, Lee, and Lee (2023b) apply the 'branch & learn' (HU et al., 2022) to minimize _post-hoc_ regret in CO problems, solvable by recursion. The second consideration is formulating a task loss that strikes a balance between trading off suboptimality and the measure of infeasibility. The next consideration is computing the gradients of this task loss with respect to the parameters in the constraints. Some of the techniques discussed in Section 3.1 can be utilized for this purpose. For example, the gradient can be obtained by solver unrolling. Tan, Delong, and Terekhov (2019) compute the gradient by unrolling a LP. As the parameters in the constraints are also present in the the KKT conditions (13), it is possible to compute the gradients for optimization problems, with differentiable constraint functions by differentiating the KKT conditions using the techniques discussed in Section 3.1. Hu et al. (2022) shows how the gradient can be computed by differentiating the KKT conditions for packing and covering LPs. For an LP, Tan, Terekhov, and Delong (2020) provide a empirical risk minimization formulation considering both the suboptimlaity of the prescribed decision and the feasibility of the true optimal decisions. This formulation takes the form of a non-linear optimization program and they propose to compute the derivative by considering its sequential quadratic programming (SQP) approximation. The task of computing the gradients of the task loss with respect to the parameters in the constraints is particularly challenging for combinatorial optimization problems, which often involves discrete feasible space. For combinatorial optimization problems, it might happen that no constraints are active at the optimal point. So, slight changes of the parameters in the constraints do not change the optimal solution leading towards the problem of zero gradients. Hence coming up with meaningful gradients for back-propagation is a big challenge for combinatorial optimization problems. Paulus, Rolinek, Musil, Amos, and Martius (2021) develop a differentiable optimization layer for ILPs, which considers the downstream gradient of the solution as an input and returns the directions of the updating the parameters in the backward pass. They update the parameters along the directions so that the Euclidean distance between the solution of the updated parameter and the updated solution with the downstream gradient is minimized. For ILPs, Nandwani, Ranjan, Mausam, and Singla (2023) view the task of constraint learning from the lens of learning hyperplanes, which is common in classification tasks. Such an approach requires negative samples. However, the negative samples in this setting must also include infeasible points, which is different from the framework proposed by Mulamba et al. (2021). #### 3.6.4 Model robustness in decision-focused learning The issue of model robustness arises often in deep learning. As has been shown in many works, it is often possible for malicious actors to craft inputs to a neural network in such a way that the output is manipulated (evasion attacks) (Goodfellow, Shlens, & Szegedy, 2014), or to generate training data which cause adverse effects on the performance of the trained model (poisoning attacks). As a subset of machine learning, some adversarial settings also apply in DFL. Evasion attacks, despite being the most commonly studied adversarial attacks, do not generalize straightforwardly to DFL since they inherently pertain to classification models with finite output spaces. On the other hand, it is shown by Kinsey, Tuck, Sinha, and Nguyen (2023) that effective poisoning attacks can be made against DFL models. The paper shows that while such attacks can be effective, they are computationally expensive due to the optimization which must be repeatedly evaluated to form the attacks. On the other hand, it is also demonstrated that poisoning attacks designed against two-stage models can be transferred to fully integrated DFL models. Separately, Johnson-Yu, Wang, Finocchiaro, Taneja, and Tambe (2023) study robustness of decision-focused learning under label noise. The paper provides bounds on the degradation of regret when test-set labels are corrupted by noise relative to those of the training set. An adversarial training scheme is also proposed to mitigate this effect. The robust training problem is equivalent to finding the equilibrium solution to a Stackelberg game, in which a figurative adversary applies label noise that is optimized to raise regret, while the main player seeks model parameters that minimize regret. #### 3.6.5 Stochastic optimization Settings in decision-focused learning based on stochastic optimization models are studied by Donti, Kolter, and Amos (2017). In contrast to more typical settings, the downstream decision model is considered to be a stochastic optimization problem. In this formulation, it is only possible to predict parameters of a random distribution that models the parameters of an optimization problem. For instance, the mean and variance of load demands in a power scheduling problem could be modeled as parameters of the optimization problem. Their work shows how such problems can be converted to DFL with deterministic decision models and solved using the techniques described in this article. To this end, it also introduces an effective technique for approximating the derivatives through arbitrary convex optimization problems, by forming and differentiating their quadratic programming approximations, as computed by sequential quadratic programming. #### 3.6.6 Problems other than optimization problems Furthermore, we believe that DFL can be further extended to encompass problems beyond optimization problems, thereby broadening its applicability. For instance, to integrate symbolic reasoning into neural network architectures, Wang, Donti, Wilder, and Kolter (2019) make use of MAXSAT solvers and perform end-to-end training of the neural network by differentiating through the semidefinite program (SDP) relaxations of the MAXSAT problems. Wilder, Ewing, Dilkina, and Tambe (2019b) consider the K-means clustering in a graph as the optimization problem, i.e., the optimization problem in their case is to cluster the nodes of a given graph into \(K\) segments. They embed the \(K\)-means clustering as a layer in a neural network architecture by differentiate through the clustering layer. Wang, Shah, Chen, Perrault, Doshi-Velez, and Tambe (2021) further extend DFL, for sequential decision making problems, where the decision making problems have been formulated as Markov decision processes (MDPs). In such cases, the DFL problem deals with the challenge of predicting the unknown parameters in the MDPs. #### 3.6.7 Active learning algorithm for DFL Active learning concerns with ML problems where labeled data are scarce or expensive to obtain. To address the challenge of limited training data, active learning algorithms choose the most informative instances for labeling (Settles, 2009). Liu, Grigas, Liu, and Shen (2023) study active learning in DFL paradigm. To choose datapoints for which to ask for a label, they propose to use notion of 'distance to degeneracy' (El Balghiti et al., 2019). Distance to degeneracy measures how far the predicted cost vector is from the set of cost vectors that have multiple optimal solutions. They argue that if distance to degeneracy is higher at a datapoint, there is more certainty regarding the solution (of the CO problem); hence they propose to acquire the label of a datapoint if its distance to degeneracy is lower than a threshold. #### 3.6.8 Multi-task decision-focused learning In most DFL works, a single task is considered. For instance, in the shortest path benchmark considered by Elmachtoub and Grigas (2022), the grid structure and the start and end nodes are the same in all instances. However, one often has to deal with multiple tasks at once, in which it would be convenient to make decision-focused predictions, without having to train a separate model for each task. A first step in this direction was recently taken in (Tang and Khalil, 2023a). This paper proposes a way of training a model in a decision-focused way with respect to multiple tasks at once. They consider two kinds of architectures. The first is a regular multi-layer perceptron that outputs a single vector \(\mathbf{\hat{c}}\) which is used in the different tasks. The different resulting task losses then get aggregated to inform the update to weights \(\boldsymbol{\omega}\), i.e., the weights \(\boldsymbol{\omega}\) are trained to produce a \(\mathbf{\hat{c}}\) that generally works well for the different tasks considered. The second architecture is a multi-headed one, consisting of one or more shared first layers, followed by a dedicated head for every task. This means that a different vector \(\hat{c}_{i}\) is produced for every task. Their results show that they can train a model that can make effective decision-focused predictions for multiple tasks at once, and that this is particularly beneficial when not that many training data are available. However, a remaining limitation is that the model can still not be trained with the aim of _generalizing_ to new tasks. ## 4 Applications of Decision-Focused Learning The Predict-Then-Optimize problem occurs in many real-world applications, as optimal decisions can be found by solving CO problems and due to the presence of uncertainty, some parameters of the CO problems must be estimated. Having seen the development of DFL for Predict-Then-Optimize problems in the preceding section, practical uses of DFL in various application domains will be presented below. As DFL techniques, which predict cost parameters have been reviewed in Section 3, applications, presented below, focus on the task of predicting only the cost parameters. Computer vision.The DBB framework (Pogancic et al., 2020) (reviewed in Section 3.3) has been used by Rolinek, Musil, Paulus, Vlastelica, Michaelis, and Martius (2020a) for differentiating rank-based metrics such as precision and recall and by Rolinek, Swoboda, Zietlow, Paulus, Musil, and Martius (2020b) and Kainmueller, Jug, Rother, and Myers (2014) for differentiating bipartite matching in deep graph and multi-graph matching problems respectively in the application of semantic keypoint matching of images. Fair Learning to Rank.In learning to rank (LTR), a machine learning model must produce rankings of documents in response to users' search queries, in which those most relevant to a given query are placed in the highest ranking positions. In this setting, the relevance of documents to queries is often measured empirically by historical user click rates (Cao et al., 2007). In fair learning to rank (FLTR), this relevance-based matching must be performed subject to strict constraints on the relative exposure between predefined groups. Due to the difficulty of enforcing such constraints on the outputs of a machine learning model, many FLTR frameworks resort to a two-stage approach in which prediction of query-document relevance scores is learned by a typical LTR model without constraints on fairness of exposure. At test time, the predicted relevance scores inform the objective of a separate fair ranking optimization program (Singh & Joachims, 2018). Kotary, Fioretto, Van Hentenryck, and Zhu (2021) use DFL to unify the prediction of relevance scores with the subsequent optimization of fair rankings, in an end-to-end model trained by SPO which learns to map user queries directly to the fair ranking policies that optimize user relevance. The result is a FLTR model which outperforms previous penalty-based models in terms of both user relevance and fairness, with the ability to directly control their trade-offs by modifying the fairness constraints of the optimization layer. Route optimization.Ferber, Griffin, Dilkina, Keskin, and Gore (2023) present an interesting application, where DFL is used to combat the challenge of wildlife trafficking. They consider the problem of predicting the flight trajectory of traffickers based on a given pair of source and destination airports. It is framed as a shortest path problem in a graph, where each node is an airport. In the prediction stage, the probability of using a directed edge \((i,j)\) to leave the node \(i\) is predicted. In the optimization stage, the most likely path from the source to the destination is found by solving a shortest path problem where the negative log probabilities are used as edge weights. In this Predict-Then-Optimize formulation, the probabilities are predicted via DFL, using the DBB framework for gradient computation. Solving a shortest path problem by considering the negative log probabilities as edge weights has also been explored by Mandi, Canoy, Bucarey, and Guns (2021). In (Mandi et al., 2021), the objective is to prescribe most preferred routing in capacitated vehicle routing problem (CVRP) (Toth and Vigo, 2015) for last-mile delivery applications. A high probability value for the edge \((i,j)\) indicates that it is the preferred edge to leave the node \(i\). However, they do not observe any advantage of DFL paradigm over PFL paradigm and attribute this to the lack of training data instances (fewer than 200 instances). DFL is used for last-mile delivery applications by Chu, Zhang, Bai, and Chen (2021) too. However, there the objective is to minimize total travel time. In the prediction stage, the travel times of all the edges are predicted and in the optimization stage, the CVRP is solved to minimize the total travel time. The underlying model is trained using the SPO framework to directly minimize the total travel time. #### Maritime transportation. The inspection of ships by port state control has been framed as a Predict-Then-Optimize problem by Yang, Yan, and Wang (2022). Due to limited number of available personnel, the aim is to identify non-compliant ships with high attention risk beforehand and select those ships for inspection. A ship can be found to be non-compliant by port state control in multiple categories. If a ship is found to be non-compliant for a category, the deficiency number for that category will be recorded as one. In the prediction stage, a linear model is built to identify deficiency numbers of the ships in all the categories and in the optimization stage, a CO problem is solved to select ships maximizing the total number of deficiencies. Due to the nature of the large-scale optimization problem, training in the SPO framework is not practical. Therefore, they employ pairwise-comparison based loss function, similar to Eq. (38) to implement DFL. Ship maintenance activities by ship owners have been framed as Predict-Then-Optimize problems by Tian, Yan, Liu, and Wang (2023). The ship owners have to schedule regular maintenance activities to remain compliant. However, as maintenance activities are expensive, the objective of identifying categories that may warrant immediate detentions has been considered. To do so, in the prediction stage, a random forest model is built to predict the deficiency number (likelihood of non-compliance) for each category. In the optimization stage, a CO problem is formulated considering maintenance cost and detention cost to determine whether maintenance activity should be scheduled for each category. The random forest models are trained to directly minimize regret using SPOTs (Elmachtoub et al., 2020). #### Planning and Scheduling. Wahdany, Schmitt, and Cremer (2023) provide a use-case of DFL in renewable power system application. In their work, the prediction stage involves the task of generating wind power forecasts. As these forecasts are further used in power system energy scheduling, the task of minimizing power system operating costs has been considered. Cvxpylayers(Agrawal, Amos, Barratt, Boyd, Diamond, & Kolter, 2019) has been used to directly train the model with the objective of minimizing power system operating costs. DFL is applied in power system application by Sang, Xu, Long, Hu, and Sun (2022) also. In the prediction stage electricity prices are predicted and the optimization stage deals with optimal energy storage system scheduling to maximize arbitrage benefits. Lower values of regret have been reported when the prices are predicted using the SPO framework. #### Communication technology. DFL is applied in mobile wireless communication technology application by Chai, Wong, Tong, Chen, and Zhang (2022). Fluid antenna system (Wong, Tong, Zhang, & Zhongbin, 2020) is one of the recent development in mobile wireless communication technology. However, its effectiveness depends on the position of the radiating element, known as the port. Chai et al. (2022) frame the port selection problem as a Predict-Then-Optimize problem, where in the prediction stage signal-to-noise ratio for each position of the port is predicted and then the optimal position of the port is decided in the optimization stage. They use LSTM as the predictive model and report the SPO framework is very effective for such port selection applications. Solving non-linear combinatorial optimization problems.Ferber et al. (2022) study the problem of learning a linear surrogate optimizer to solve non-linear optimization problems. The objective is to learn a surrogate linear optimizer whose optimal solution is the same as the solution to the non-linear optimization problem. Learning the parameters of the surrogate linear optimizer entails backpropagating through the optimizer, which is implemented using Cvxpylayers (Agrawal et al., 2019b). Interested readers are referred to (Qi & Shen, 2022) for more applications of Predict-Then-Optimize problems in various areas within operations management. ## 5 Experimental Evaluation on Benchmark Problemsets DFL recently has received increasing attention. The methodologies discussed in Section 3 have been tested so far on several different datasets. Because a common benchmark for the field has not yet been set up, comparisons among methodologies are sometimes inconsistent. In this section, an effort is made to propose several benchmark test problems for evaluating DFL methodologies. 1 Then some of the methodologies explained in Section 3 are compared on these test problems. Footnote 1: During the course of writing this manuscript, we have become aware of the PyEPO project (Tang & Khalil, 2023b), which develops an interface for benchmarking DFL methodologies. However, it is important to emphasize that our work differs significantly from PyEPO. While PyEPO focuses on providing an interface for implementing DFL methodologies, our paper serves as a comprehensive survey that goes beyond benchmarking. \begin{table} \begin{tabular}{l c c c c} \hline \hline **Methodology** & **Controlling functions** & **Position variables** & **GO-SQu** & **Prediction Model** \\ \hline Shortest path problem & & Linear & Continuous & OR-Tools & Linear \\ on a \(5\times 5\) grid & & & & \\ Portfolio optimization & Quadratic & Continuous & Gurobi & Linear \\ Warcraft shortest path & & Linear & Continuous & Customized python & CNN \\ Energy-cost aware & & & & \\ scheduling & & & & \\ Knapsack problem & & & & \\ Diverse bipartite matching & & & & \\ Subset selections & & & & \\ \hline \hline \end{tabular} \end{table} Table 2: Brief overview of the test problems considered for experimental evaluation. The **objective functions are linear** for all the optimization problem. ### Problem Descriptions All the test problems, which are selected for benchmarking, have been previously used in the DFL literature and their datasets are publicly available. Needless to say, all these problems encompass the two stages--prediction and optimization. Table 2 provides an overview of the experimental setups associated with each test problem, including the specification of the CO problem and the type of predictive model. Next, these test problems are described in detail. #### 5.1.1 Shortest path problem on a \(5\times 5\) grid This experiment is adopted from the work of Elmachtoub and Grigas (2022). It is a shortest path problem on a \(5\times 5\) grid, with the objective of going from the southwest corner of the grid to the northeast corner where the edges can go either north or east. This grid consists of 25 nodes and 40 edges. Formulation of the optimization problem.The shortest path problem on a graph with a set \(V\) of vertices and a set \(E\) of edges can be formulated as an LP problem in the following form: \[\min_{\mathbf{x}}\ \ \mathbf{c}^{\top}\mathbf{x} \tag{42a}\] \[\mathtt{s.t.}A\mathbf{x}=\mathbf{b}\] (42b) \[\mathbf{x}\geq\mathbf{0} \tag{42c}\] Where \(A\in\mathbb{R}^{|V|\times|E|}\) is the incidence matrix of the graph. The decision variable \(\mathbf{x}\in\mathbb{R}^{|E|}\) is a binary vector whose entries are 1 only if the corresponding edge is selected for traversal. \(\mathbf{b}\in\mathbb{R}^{|V|}\) is the vector whose entry corresponding to the source and sink nodes are 1 and \(-1\) respectively; all other entries are 0. The constraint (42b) must be satisfied to ensure the path will go from the source to the sink node. The objective is to minimize the cost of the path with respect to the (predicted) cost vector \(\mathbf{c}\in\mathbb{R}^{|E|}\). Synthetic data generation process.In this problem, the prediction task is to predict the cost vector \(\mathbf{c}\) from the feature vector \(\mathbf{z}\). The feature and cost vectors are generated according to the data generation process defined by Elmachtoub and Grigas (2022). For the sake of completeness, the data generation process is described below.2 Each problem instance has cost vector of dimension \(|E|=40\) and feature vector of dimension \(p=5\). The training data consists of \(\{(\mathbf{z_{i}},\mathbf{c_{i}})\}_{i=1}^{N}\), which are generated synthetically. The feature vectors are sampled from a multivariate Gaussian distribution with zero mean and unit variance, i.e., \(\mathbf{z_{i}}\sim\mathbf{N}(0,I_{p})\) To generate the cost vector, first a matrix \(B\in\mathbb{R}^{|E|\times p}\) is generated, which represents the true underlying model. The cost vectors are then generated according to the following formula: Footnote 2: The generator in [https://github.com/paulgrigas/SmartPredictThenOptimize](https://github.com/paulgrigas/SmartPredictThenOptimize) is used to generate the dataset. \[c_{ij}=\bigg{[}\bigg{(}\frac{1}{\sqrt{p}}\big{(}B\mathbf{z_{i}}\big{)}+3 \bigg{)}^{\text{Deg}}+1\bigg{]}\xi_{i}^{j} \tag{43}\] where \(c_{ij}\) is the \(j^{\text{th}}\) component of cost vector \(\mathbf{c_{i}}\). The _Deg_ parameter specifies the extent of model misspecification, because a linear model is used as a predictive model in the experiment. The higher the value of _Deg_, the more the true relation between the features and objective parameters deviates from a linear one and the larger the prediction errors will be. Finally, \(\xi_{i}^{j}\) is a multiplicative noise term sampled randomly from the uniform distribution \([1-\vartheta,1+\vartheta]\). The experimental evaluation involves five values of the parameter _Deg_, which are \(1,2,4,6\) and \(8\), and the noise-halfwidth parameter \(\vartheta\) being \(0.5\). Furthermore, for each setting, a different training set of of size \(1000\) is used. In each case, the final performance of the model is evaluated on a test set of size \(10,000\). Predictive model.In each setting, the underlying predictive model is a one-layer feed-forward neural network without any hidden layer, i.e., a linear model. Basically the input to the model is a \(p\) dimensional vector, and output is a \(|E|\) dimensional vector. Note that a multi-layer neural network model can be used to to improve the accuracy of the predictive model. The intuition behind using a simple predictive model is to test the efficacy of the DFL methods when the predictions are not \(100\%\) accurate. The DFL methods are trained to minimize the regret and the prediction-focused model is trained by minimizing the MSE loss between the true and predicted cost vector. #### 5.1.2 Portfolio optimization problem A classic problem that combines prediction and optimization is the Markowitz portfolio optimization problem, in which asset prices are predicted by a model based on empirical data, and then subsequently, a risk-constrained optimization problem is solved for a portfolio which maximizes expected return. This experiment is also adopted from the work of Elmachtoub and Grigas (2022). Formulation of the optimization problem.In portfolio optimization problem, the objective is to choose a portfolio of assets having highest return subject to a constraint on the total risk of the portfolio. The problem is formulated in the following form: \[\max_{\mathbf{x}} \ \mathbf{c}^{\top}\mathbf{x}\] (44a) \[\mathtt{s.t.} \ \ \mathbf{x}^{\top}\Sigma\mathbf{x}\leq\gamma\] (44b) \[\ whose entries are drawn uniformly over \([-0.0025\vartheta,0.0025\vartheta]\) is generated. Asset returns are calculated first in terms of their conditional mean \(\bar{c}_{ij}\) as \[\bar{c}_{ij}\coloneqq(\frac{0.05}{\sqrt{\bar{p}}}(B\mathbf{z_{i}})_{j}+(0.1)^{ \frac{1}{\mathrm{Deg}}})^{\mathrm{Deg}} \tag{45}\] Then the observed return vectors \(\mathbf{c_{i}}\) are defined as \(c_{ij}\coloneqq\bar{r}_{i}+Lf+0.01\vartheta\xi\), where \(f\sim\mathbf{N}(0,I_{4})\) and noise \(\xi\sim\mathbf{N}(0,I_{d})\) This causes the \(c_{ij}\) to obey the covariance matrix \(\Sigma\coloneqq LL^{\top}+(0.01\zeta)^{2}I\), which is also used to form the constraint (44b), along with a bound on risk, defined as \(\gamma\coloneqq 2.25\ e^{\top}\Sigma e\) where \(e\) is the equal-allocation solution (a constant vector). Four values of the parameter _Deg_--\(1,4,8,16\) have been used in the experimental evaluation. The value of noise magnitude parameter \(\vartheta\) is set to \(1\). It is assumed that the covariance matrix of the asset returns does not depend on the features. The values of \(\Sigma\) and \(\gamma\) are constant, and randomly generated for each setting. Predictive modelLike the previous experiment, the underlying predictive model is a linear model, whose input is a feature vector \(\mathbf{z}\in\mathbb{R}^{p}\) and output is the return vector \(\mathbf{c}\in\mathbb{R}^{d}\). #### 5.1.3 Warcraft shortest path problem This experiment was adopted from the work of Pogancic et al. (2020). Each instance in this problem is an image of a terrain map using the Warcraft II tileset (Guyomarch, 2017). Each image represents a grid of dimension \(d\times d\). Each of the \(d^{2}\) pixels has a fixed underlying cost, which is unknown and to be predicted. The objective is to identify the minimum cost path from the top-left pixel to the bottom-right pixel. From one pixel, one can go in eight neighboring pixels--up, down, front, back, as well as four diagonal ones. Hence, it is a shortest path problem on a graph with \(d^{2}\) vertices and \(\mathcal{O}(d^{2})\) edges. Formulation of the optimization problem.Note that this is a node-weighted shortest path problem, where each node (pixel) in the grid is assigned a cost value; whereas in the previous shortest path problem, each edge is assigned a cost value. However, this problem can be easily reduced to the more familiar edge weighted shortest path problem by 'node splitting'. 'Node splitting' splits each node into two separate nodes--entry and exit nodes and adds an edge, that has a weight equal to the node weight, from the entry node to the exit node. For each original edge, an edge, with null weight, from the exit node of the source node to the entry node of the sink node, is constructed. Predictive model.The prediction task is to predict the cost associated with each pixel. The actual cost ranges from from \(0.8\) to \(9.2\) and is dependent on visible characteristics of the pixel. For instance, cost changes depending on whether the pixel represents a water-body, land or wood. The predictive model used in this case is a convolutional neural network (CNN), which predicts the cost of each node (pixel). The model takes the \(d\times d\) image as an input and outputs costs of the \(d^{2}\) pixels. The ResNet18 (He, Zhang, Ren, & Sun, 2016) architecture is slightly modified to form the ML model. The first five layers of ResNet18 are followed by a max-pooling operation to predict the underlying cost of each pixel. Furthermore, a Relu activation function (Agarap, 2019) is used to ensure the predicted cost remains positive, thereby avoiding the existence of negative cycles in the shortest path edge weights. #### 5.1.4 Energy-cost aware scheduling This experiment setup was adopted from the work of Mandi et al. (2020). This is a resource-constrained day-ahead job scheduling problem (Simonis, O'Sullivan, Mehta, Hurley, & Cauwer, 1999) with the objective of minimizing energy cost. Tasks must be assigned to a given number of machines, where each task has a duration, an earliest start, a latest end, a resource requirement and a power usage. Each machine has a resource capacity constraint. Also, tasks cannot be interrupted once started, nor migrated to another machine and must be completed before midnight. The scheduling is done in one-day advance. So, the prediction task is to predict the energy prices of the next day. Formulation of the optimization problem.The scheduling problem is formulated as an ILP. Let \(J\) be the set of tasks to be scheduled on a set of machines \(I\) while maintaining resource requirement of \(W\) resources. The tasks must be scheduled over \(T\) number of time slots. Each task \(j\) is specified by its duration \(\zeta_{j}\), earliest start time \(\zeta_{j}^{(1)}\), latest end time \(\zeta_{j}^{(2)}\), power usage \(\phi_{j}\). Let \(\rho_{jw}\) be the resource usage of task \(j\) for resource \(w\) and \(q_{iw}\) is the capacity of machine \(i\) for resource \(w\). Let \(x_{jit}\) be a binary variable which possesses 1 only if task \(j\) starts at time \(t\) on machine \(i\). The objective of minimizing energy cost while satisfying the required constraints can be expressed by the following ILP: \[\min_{x_{jit}}\sum_{j\in J}\sum_{i\in I}\sum_{t\in T} x_{jit}\Big{(}\sum_{t\leq t\prime<t+\zeta_{j}}\phi_{j}c_{t\prime} \Big{)} \tag{46a}\] \[\texttt{s.t.}\quad\sum_{i\in I}\sum_{t\in T}x_{jit} =1\,\forall_{j\in J}\] (46b) \[x_{jit} =0\ \ \forall_{j\in J}\forall_{i\in I}\forall_{t<\zeta_{j}^{(1)}}\] (46c) \[x_{jit} =0\ \ \forall_{j\in J}\forall_{i\in I}\forall_{t+\zeta_{j}> }\zeta_{j}^{(2)}\] (46d) \[\sum_{j\in J}\sum_{t-\zeta_{j}<t^{\prime}\leq t}x_{jit}\rho_{jw} \leq q_{iw},\forall_{i\in I}\forall_{w\in W}\forall_{t\in T}\] (46e) \[x_{jit} \in\{0,1\}\forall_{j\in J}\forall_{i\in I}\forall_{t\in T} \tag{46f}\] The (46b) constraint ensures each task is scheduled once and only once. The constraints in (46c) and (46d) ensure that the task scheduling abides by earliest start time and latest end time constraints. (46e) imposes the constraints of resource requirement. Data description.The prediction task is to predict the energy prices one day advance. The energy price dataset comes from the Irish Single Electricity Market Operator (SEMO) (Ifrim, O'Sullivan, & Simonis, 2012). This dataset consists of historical energy price data at 30-minute intervals starting from midnight on the 1st of November, 2011 until the 31st of December, 2013. In this setup, each day forms an optimization instance, which comprises of 48 time slots, corresponding to 48 half-hour slots. Each half-hour instance of the data has calendar attributes, day-ahead estimates of weather characteristics, SEMO day-ahead forecasted energy-load, wind-energy production and prices, actual wind-speed, temperature and \(CO_{2}\) intensity, which are used as features. So, the dimension of feature vector is 8. Note that, in this dataset, each \(c_{t}\) in the cost vector is associated with an eight dimensional feature vector, i.e., \(\mathbf{c}\in\mathbb{R}^{48}\) and \(\mathbf{z}\in\mathbb{R}^{48\times 8}\). Predictive model.As energy prices of each half-hour slot is associated with 8 features, the input to the predictive model is a feature vector of dimension 8 and output is a scalar. In this case also, the predictive model is a linear model, i.e., a feed forward neural network without any hidden layer. #### 5.1.5 Knapsack problem This problem setup was also adopted from the work of Mandi et al. (2020). The objective of the knapsack problem is to choose a maximal value subset from a given set of items, subject to a capacity constraint. In this case, the weights of all items and the knapsack capacity are known. What are unknown are the values of the items. Hence, the prediction task is to predict item values of each item. Formulation of the optimization problem.The formulation of the knapsack optimization problem with unit weights has already been provided in Eq. (4). However, in general the weights of all items are not equal. So, a general knapsack optimization problem can be formulated as follows: \[\max_{\mathbf{x}}\ \ \mathbf{c}^{\top}\mathbf{x}\] (47a) \[\mathtt{s.t.}\ \mathbf{w}^{\top}\mathbf{x}\leq\text{Capacity}\] (47b) \[\ Getoor, Galligher, & Eliassi-Rad, 2008), where a node represent a publication and an edge represent a citation. So the matching problem is to identify the citation between the two sets of publications. Furthermore, the matching must obey diversity constraints, as described later. Note that this problem falls under the category of structured output prediction tasks (Nowozin & Lampert, 2011), which requires capturing dependencies and relationships between different parts of the output. In this matching problem, each edge does not have an associated cost in the true sense. Therefore, in the prediction-focused approach the model is trained by directly predicting the presence or absence of each edge. On the other hand, the DFL approaches consider the likelihood of the existence of each edge as the edge weights and then determine which edges should be present while ensuring all the constraints are satisfied. Optimization problem formulation.Let \(S_{1}\) and \(S_{2}\) denote the two sets. The matching must satisfy the following diversity constraints: a minimum \(\rho_{1}\%\) and \(\rho_{2}\%\) of the suggested pairings should belong to same and distinct fields of study respectively. Let \(c_{ij}\) be the likelihood of an edge existing between article \(i\) and \(j\), \(\forall i\in S_{1},j\in S_{2}\). With this likelihood value, the matching can be performed by solving the following ILP, which ensures the diversity constraints: \[\max_{\mathbf{x}} \sum_{i,j}c_{ij}x_{ij} \tag{48a}\] \[\mathtt{s.t.} \sum_{j}x_{ij}\leq 1 \forall i\in S_{1}\] (48b) \[\sum_{i}x_{ij}\leq 1 \forall j\in S_{2}\] (48c) \[\sum_{i,j}\phi_{i,j}x_{ij}\geq\rho_{1}\sum_{i,j}x_{ij}\] (48d) \[\sum_{i,j}(1-\phi_{ij})x_{ij}\geq\rho_{2}\sum_{i,j}x_{ij}\] (48e) \[x_{ij}\in\{0,1\} \forall i\in S_{1},j\in S_{2} \tag{48f}\] where \(\phi_{ij}\) is an indicator variable, which takes the value 1 only if article \(i\) and \(j\) are of same field, and 0 if they belong to two different fields. Data description.The network is divided into 27 disjoint topologies, each containing 100 nodes. Each of the instant form an optimization instance. In each instance, the 100 nodes are split into two sets of 50 nodes \(S_{1}\) and \(S_{2}\); so each instance forms a bipartite matching problem between two sets of cardinality 50. Each publication (node) has 1433 bag-of-words features. The feature of an edge is formed by concatenating features of the two corresponding nodes. The prediction task is to estimate \(c_{ij}\) values. In this problem, each individual \(c_{ij}\) is associated with a feature vector of length 2866. Predictive model.The predictive model is a neural network model. The input to the neural network is a 2866 dimensional vector and final output is a scalar between 0 and 1. The neural network has one hidden layer and uses a sigmoid activation function on the output. #### 5.1.7 Subset selections This experiment is a structured prediction task, in which the object is to learn a mapping from feature vectors to binary vectors which represent subset selections. Unlike the other experiments above, the ground-truth data take the form of optimal solutions to an optimization problem, rather than its corresponding problem parameters. Thus, the regret loss is not suitable for training a prediction model. Instead, a task loss based on the error of the predicted solutions with respect to ground-truth solutions is used in this experiment. Optimization problem formulation.For any \(\mathbf{c}\in\mathbb{R}^{n}\), the objective of the optimization problem is to output a binary vector in \(\mathbb{R}^{n}\), where the non-zero values correspond to the top-\(k\) values of \(\mathbf{c}\). This can be formulated as an LP problem in the following form: \[\underset{\mathbf{x}}{\operatorname{argmax}}\ \ \mathbf{c}^{\top} \mathbf{x}\] (49a) \[\mathtt{s.t.}\ \ 1^{\top}\mathbf{x}=k\] (49b) \[\ Pairwise LTR loss (38) [Pairwise], 10. Pairwise difference LTR loss (39) [Pairwise(diff)], 11. Maximum a posteriori contrastive loss (37) [MAP]. The reason behind including prediction-focused approach is that it is considered as a benchmark. Note that among these methodologies, Listwise, Pairwise, Pairwise(diff), and MAP make use of a solution cache. The solution cache is implemented using the procedure proposed by Mulamba et al. (2021). In this approach, the solution cache is initialized by caching all the solutions in the training data and the cache is later expanded by employing a \(p_{solve}\) parameter value greater than zero. As in (Mulamba et al., 2021; Mandi et al., 2022) it is reported that \(p_{solve}=5\%\) is adequate for most applications, the value of \(p_{solve}\) is set to \(5\%\). Next, the procedure systematically followed for the empirical evaluations is explained. Experimental setup and procedures.The performance of a methodology is sensitive to the choice of the methodology specific hyperparameters as well as some other fundamental hyperparameters, common in any neural network training such as learning rate. These are called hyperparameters because they cannot be estimated by training the model, rather they must be selected before training begins. Tuning hyperparameters is the process of identifying the set of hyperparameter values that are expected to produce the best model outcome. In the experimental evaluations, hyperparameter tuning is performed via grid search. In the grid search, each of the hyperparameters is tried for a set of values. The set of values to be tested on for each hyperparameter is predetermined. Grid search suffers from the curse of dimensionality in the hyperparameter space, as the number of combinations grows exponentially with the number of hyperparameters. However, it is possible to train the different models for different combination of hyperparameter in parallel as the combinations are independent. The hyperparameter of each model for each experiment is selected based on performance on the validation dataset. For each hyperparameter a range of values as defined in Table 3 is considered. The hyperparameter combination which produces the lowest average regret on the validation dataset is considered to be the 'optimal' one. For both validation and testing, 10 trials are run where in every trial the network weights are initialized with a different seed. To be specific, values of seed from 0 to 9 have been considered. Each model for each setup is trained using Pytorch(Paszke et al., 2019) and PyTorch-Lightning(Falcon et al., 2019) with Adam optimizer(Kingma & Ba, 2014) and 'ReduceLROn \begin{table} \begin{tabular}{c c c} \hline \hline \begin{tabular}{c} **Hyperparameter** \\ \end{tabular} & \begin{tabular}{c} **Methodologies Utilizing** \\ **the Hyperparameter** \\ \end{tabular} & \begin{tabular}{c} **Range** \\ \end{tabular} \\ \hline \begin{tabular}{c} learning rate \\ \end{tabular} & All & \(\{5\times 10^{-4},1\times 10^{-3},5\times 10^{-3},0.01,0.05,0.1,0.5,1.0\}\) \\ \(\lambda\) & I-MLE, DBB & \(\{0.1,1,10,100\}\) \\ \(\epsilon\) & I-MLE, FY & \(\{0.05,0.1,0.5,1,2,5\}\) \\ \(\kappa\) & I-MLE & \(\{5,10,50\}\) \\ \(\tau\) & Listwise & \(\{0.05,0.1,0.5,1,2,5\}\) \\ \(\Theta\) & Pairwise & \(\{0.01,0.05,0.1,1.,10.,50.\}\) \\ \(\mu\) & QPTL, HSD & \(\{0.01,0.1,1.,10.,\}\) \\ damping & HSD & \(\{1\times 10^{-4},0.01,0.1,1.,10\}\) \\ \hline \hline \end{tabular} \end{table} Table 3: The range of hyperparameters for hyperparameter tuning by grid search. Plateau'(PyTorch, 2017) learning rate scheduler. As mentioned before, the learning rate of Adam optimizer is treated as a hyperparameter. For QPTL, the QP problems are solved using Cvxpylayers (Agrawal et al., 2019). For other methodologies, which treat the CO solver as a blackbox solver, Gurobi(Gurobi Optimization, 2021) or OR-Tools(Perron and Furnon, 2020) is used as the solver. For MAP and LTR losses, the experiments are run with \(p_{solve}\) being 5%. Evaluation metric.After selecting the 'optimal' hyperparameter combination for each test problem, **10** trials of all the methodologies with the 'optimal' hyperparameter combination are run on test dataset. Unless otherwise mentioned the comparative evaluation is made based on relative regret on the test dataset. The **relative regret** is defined as follows: \[\frac{1}{N_{test}}\sum_{i=1}^{N_{test}}\frac{\mathbf{c_{i}}^{\top}(\mathbf{x}^ {\star}(\mathbf{\hat{c}_{i}})-\mathbf{x}^{\star}(\mathbf{c_{i}}))}{\mathbf{c_{ i}}^{\top}\mathbf{x}^{\star}(\mathbf{c_{i}})}. \tag{50}\] In practice, \(\mathbf{c}\) (or \(\mathbf{\hat{c}}\)) can have non-unique optimal solutions. However, note that if all the entries in \(\mathbf{c}\) are continuous, it is very unlikely that \(\mathbf{c}\) will have non-unique solutions. For instance, in the case of an LP, the only circumstance in which the LP can have multiple solutions is when \(\mathbf{c}\) is parallel to one of the faces of the LP polyhedron. Nevertheless, if the cost vector is predicted by an ML model, a pathological case might occur, especially at the beginning of model training, when all the cost parameters are zero. This results in all feasible solutions being optimal with zero cost. However, to avoid this complexity in the experiments, it is assumed that the solution \(\mathbf{x}^{\star}(\mathbf{\hat{c}})\) is obtained by calling an optimization oracle and that if there exist non-unique solutions, the oracle returns a single optimal solution by breaking ties in a pre-specified manner. This is true if a commercial solver such as Gurobi is used to solve the CO problem. #### 5.2.1 Comparative Evaluations Next, the performances of the 11 methodologies in the 7 problems are presented with insights. Shortest path problem on a \(5\times 5\) grid.The comparative evaluation for the synthetic shortest path problem in is shown in Figure 5 with the aid of box plots. To conserve space, boxplots for two values of Deg are shown in Figure 5. The boxplots for all the five degrees are shown in Figure A1 the Appendix. In Figure 5, the value of \(\vartheta\), the noise- halfwidth parameter is 0.5 for all the experiments and the training set for each Deg contains of 1000 instances. The predictive model is a simple linear model implemented as a neural network model with no hidden layers. For Deg 1, the linear predictive model perfectly captures the data generation process. Consequently the PF approach is very accurate and it results in the lowest regret. SPO has slightly higher regret than the PF approach. All the other models have considerable higher regrets. It is followed by MAP and FY. For Deg 8, FY has the lowest regret, closely followed by I-MLE. Then comes Listwise and Pairwise ranking losses followed by QPTL and DBB. In this case, SPO performs poorer than them. MAP and HSD have very high regret but still lower than the PF approach. The relative regret worsens for the PF approach, as the value of Deg parameter is increased. For Deg 2, both PF and SPO have lowest regret. However, their differences with other models reduce in this case. FY, MAP and I-MLE come at the next three places respectively. For Deg 4, the PF model starts to result in high regret. In this case, I-MLE has the lowest regret, closely followed by FY and SPO. The next three spots are taken by MAP, Listwise and Pairwise respectively. DBB and HSD perform worse than the PF approach. For Deg 6, the best one FY, although its test regret is not very different from SPO, I-MLE and QPTL. Listwise, DBB, Pairwise and MAP come next. Overall, FY and I-MLE are the top two best-performing approaches, for \(Deg>2\). For \(Deg\) values of 1 and 2, the PF approach has the lowest regret. Note that the performance of SPO is very consistent too. It performs considerably worse than I-MLE and FY only for Deg 8. On the other hand, HSD exhibits higher regret than the other DFL approaches. In fact it does better than the PF approach only for Deg 6 and 8. It also exhibits higher variances. Portfolio optimization problem.Note that this is an optimization problem with continuous decision variables having quadratic constraints and a linear objective function. Hence, the HSD approach is not applicable for this problem, as it cannot handle non-linear constraints. The boxplots of test regrets for noise magnitude parameter \(\vartheta\) being 1 are shown in Figure 6. In this problem, in some problem instances, all the return values are negative, which makes a portfolio with zero return to be the optimal portfolio. In such cases, relative regret turns infinite as the denominator is zero in Eq. (50). Hence, for this problem set, the **absolute regret** instead of relative regret is reported in Figure 6. The boxplots for Deg Figure 5: Comparative evaluations on the synthetic shortest path problem with noise-halfwidth parameter \(\vartheta\) = 0.5. The boxplots show the distributions of relative regrets. values of 1 and 16 are shown in The boxplots for all the four degrees are shown in Figure A2 the Appendix. Apparently the PF approach performs very well in this problem; but SPO manages to outperform PF slightly in all cases except for Deg 1. It is evident in Figure 6 that DBB, I-MLE, FY and QPTL perform miserably as they generate regret even higher than the PF approach. All these methodologies were proposed considering problems with linear constraints. Hence the concerns arise that these methodologies be suitable in the presence of quadratic constraints. On the other hand, LTR losses--Pairwise and Pairwise(diff) and contrastive loss function, MAP, perform even better than SPO for Deg 16. For Deg 16, again Pairwise is the best performing model, followed by Listwise, Pairwise(diff), MAP and SPO, in that order. For Deg 1, PF is the best followed by MAP and SPO. For Deg 4 and 8, Pairwise loss function has the lowest regret, closely followed by Pairwise(diff), MAP and SPO. For Deg 1, PF is the best one, followed by MAP, SPO, Pairwise and Pairwise(diff), in that order. The Listwise loss function exhibits high variance for Deg 1 and for Deg values of 4 and 8 it generates high regret for few instances. For Deg 16, it generates average test regret lower than SPO. In general, Figure 6 reveals DBB, I-MLE, FY and QPTL perform poorly in this problem, whereas, SPO, MAP Pairwise and Pairwise(diff) seem to be suitable methodologies for this problem. Warcraft shortest path problem.Recall that this a shortest path problem in an image with dimension \(d\times d\). The optimization problem can be efficiently solved using Dijkstra's algorithm (Dijkstra, 1959), as underlying costs of all the pixel values are non-negative. Hence the shortest path problem is solved using Dijkstra's algorithm for the methodologies which Figure 6: Comparative evaluations on the synthetic portfolio optimization problem with noise magnitude \(\vartheta=1\). The boxplots show the distributions of **absolute** regrets. view the CO solver as a blackbox oracle. However, HSD and QPTL require the problem to be formulated as an LP and require a primal-dual solver. Note in this experiment, the predictive ML model is a CNN, which predicts the cost of each pixel. In this case, training of the ML model is challenging due to the large number of parameters. Hence combining this ML model with computation-intensive modules such as interior point optimizer poses significant challenges. We could not run the experiments with HSD and QPTL because of this computational burden. The dataset contains four values of \(d\): \(12,18,24,30\). Clearly, as the value of \(d\) increases, the optimization problem contains more number of parameters. The boxplots of comparative evaluations are summarized in Figure 7. The boxplots of other two values of \(d\) can be found in Figure 10 in the Appendix. First note that the PF approach, which is trained by minimizing mse loss between the predicted cost and true cost performs significantly worse than the DFL methodologies. In fact, the performance of the PF approach deteriorates as the image size increases. As the size of the image increases, the same level of prediction error induces greater inaccuracies in the solution. This is because an increase in the area of the image involves dealing with a greater number of decision variables in the CO problem. When the level of prediction error remains constant, the probability of the error in prediction changing at least one of the decision variables also increases. Consequently, there is a higher likelihood of error in the final solution. As the regret of the PF approach is significantly higher, note that the scale of the y-axis is changed to fit it into the plot. Among the DFL methodologies, Listwise performs best for sizes 12, 18, and 30 and SPO performs best for size 30. In fact, for sizes 12, 18, and 24, there are not many Figure 7: Comparative evaluations on the Warcraft shortest path problem instances. The boxplots show the distributions of relative regrets. variations between SPO, Listwise, and MAP. After them, the next three best-performing methodologies are Pairwise (diff), I-MLE and DBB. However, for size 30, DBB comes third after Listwise and MAP, followed by Pairwise (diff), SPO, and I-MLE in that order. FY and Pairwise perform slightly worse than the other DFL methodologies. In general, this set of experiments shows the advantage of the DFL approaches as all of them outperform the PF approach. Energy-cost aware scheduling.There are three instances of this scheduling problem. All the instances have 3 machines. The first, second, and third instances contain 10, 15, and 20 tasks, respectively. In this problem, the underlying ML model is a simple linear model implemented as a neural network model with no hidden layers. The boxplot of comparative evaluations for the first instance is presented in Figure 8. The boxplots of the other instances can be found in Figure A4 in the Appendix. Note that the scheduling problem is an ILP problem. For HSD and QPTL, the LPs obtained by relaxing the integrality constraints have been considered. For the first instance, MAP and SPO result in the lowest average regret, closely followed by I-MLE. DBB, FY, and Pairwise(diff) perform better than the PF approach. The performances of the Listwise and Pairwise rankings are worse than the PF. QPTL and HSD also perform poorly in all three instances, probably because in this case the LP obtained by relaxing by removing the integrality is not a proper representation of the ILP. In fact, QPT fails to learn in this problem instance. In the second instance, FY, SPO, and I-MLE are the best three performing models. Then comes MAP and DBB, followed by Pairwise(diff). Again, performances of the Listwise and Pairwise rankings are worse than the PF. In the third instance, again, MAP and SPO deliver the lowest average regret. Then comes I-MLE and FY. The test regret of these two models is very similar. In this case, the performance of Pairwise(diff) is slightly worse than the PF approach, whereas, like before, performances of Listwise and Pairwise ranking are significantly worse. In general, across the three problem instances, it is possible to identify some common patterns. The first one is relaxing the integrality constraints fails to capture the essence of the combinatorial nature of the LP. Consequently, HSD and QPTL perform poorly. Secondly, Listwise and Pairwise ranking performances are significantly worse than the PF approaches. The learning curve suggests (refer to B), these models fail to converge in these problem instances, although in some epochs, they are Figure 8: Comparative evaluations on the energy-cost aware scheduling problem instances. This boxplot shows the distributions of relative regrets. able to perform significantly better than the PF approach, their pheromones never plateau. Lastly, SPO, MAP, FY, and I-MLE perform consistently better than the other models. Knapsack problem.Three instantiations of the knapsack problem are considered for the experiment--each instantiation with a different capacity. The three capacity values are--60, 120 and 180. The boxplot corresponding to capacity value 60 is presented in Figure 9. The boxplots of the other two capacities can be found in Figure A5 in the Appendix. With a capacity of 60, the best three models are QPTL, DBB, and I-MLE, in that order. HSD, SPO, and MAP come next and perform better than the PF approach. FY and LTR losses perform worse than the PF approach. With a capacity of 120, the top three models are DBB, I-MLE, and QPTL. Then comes SPO, HSD and MAP. The Pairwise(diff) model performs slightly better than the PF approach, but the other two LTR losses and FY perform worse. With a capacity of 180, the best three models are DBB, I-MLE and SPO. HSD and QPTL perform better than the PF approach, but MAP, LTR losses, and FY perform worse. In general, for this problem, DBB and I-MLE are the best-performing models across the three capacity values. QPTL, SPO, HSD also consistently perform better than the PF approach in all three cases. However, FY and the LTR losses perform poorly in this problem. Figure 10: Comparative evaluations on the diverse bipartite matching problem instances. This boxplot shows the distributions of relative regrets. Figure 9: Comparative evaluations on the knapsack problem instances. This boxplot shows the distributions of relative regrets Diverse bipartite matching.Three instantiations of the diverse bipartite matching problem are formed by changing the values of \(\rho_{1}\) and \(\rho_{2}\). The values of \((\rho_{1},\rho_{2})\) for the three instantiations are \((10\%,10\%)\), \((25\%,25\%)\), \((50\%,50\%)\) respectively. The boxplot of comparative evaluations for \((\rho_{1},\rho_{2})\) being \((50\%,50\%)\), is presented in Figure 10. As mentioned before, in this problem, each edge is not associated with an edge weight in the true sense. Hence, the PF approach is trained by directly learning to predict whether an edge exists. So the loss used for supervised learning for the PF approach is BCE loss. The DFL approaches consider the predicted probability of each edge as the edge weight and then aim to minimize regret. In this problem instance QPTL is the best-performing model. FY, Pairwise(diff) and Listwise take the next three places. MAP, Pairiwse, I-MLE and SPO also perform better than the PF approach. The performances of HSD and DBB are similar to that of the PF approach. Also note that the relative regrets of all the models are very high (higher than \(80\%\)) for all three instances. With \(\rho_{1}\) and \(\rho_{2}\) being \(10\%\), I-MLE performs considerably better than all the other models. Then comes HSD, FY, Pairwise and Pairwise(diff) followed by SPO, MAP, DBB and Listwise. When \(\rho_{1}\) and \(\rho_{2}\) take the value of \(25\%\), QPTL, I-MLE and HSD are the top there models, with significantly lower regret than the rest. In this instance, the regrets of Listwise, Pairwise, SPO, FY and MAP are higher than the PF approach. Across the instances, the performances of I-MLE and QPTL are consistently better than the PF approach. In the first two instances, other than I-MLE and QPTL, other DFL models do not significantly better than the PF approach. DFL approaches such as FY, Listwise, Pairwise and MAP perform considerably better than the PF approach only in the third instances. On the other hand, the test regret of DBB is similar to the PF approach across the instances. Learning subset selections.Subset selection problems of three dimensions: \(n=25\), \(n=50\), and \(n=100\) are considered for evaluation. In each case, the subset size \(k\) is chosen to be \(\frac{n}{5}\). The error of any predicted subset \(\hat{x}\), with respect to ground truth \(x\), is considered to be the fraction of items which are selected in \(x\) but not in \(\hat{x}\). Such occurrences are referred to as mismatches. Figure 11 shows the average mismatch rates over the size \(n=25\) instances that were achieved by each DFL methodology listed in Table 1, excluding those which assume ground-truth data in the form of problem parameters. Here, the ground-truth data are optimal solutions of (49) representing subset selections. For each assessed method, a distribution of results is shown, corresponding to \(10\) different randomly generated training datsets. Figure A7 shows similar results over the larger problem instances. Note that it is suggested in (Amos et al., 2019) that the entropy function \(H(\mathbf{x})=\sum_{i}x_{i}\log x_{i}\) is particularly well-suited as a regularizer of the objective in (49), for the purpose of multilabel classification, which is identical to the task in terms of its optimization component and the form of its target data. Hence a Cvxpylayers implementation of this model is included and referred to as ENT. Figure A7 shows that most of the assessed methods perform similarly, with DBB performing worst regardless of the problem's dimension. HSD is most sensitive with respect to the randomly generated training set; the rest show consistent performance across datasets. QPTL and IMLE each show a marginal advantage over the other methods, but DPO and ENT are also competitive. Across all methods, variation in performance over the randomly generated datasets tends to diminish as problem size increases. #### 5.2.2 Comparison on Runtime While coming up with a useful gradient is considered to be the primary challenge of DFL, as mentioned in Section 2.3, computational cost associated with repeatedly solving CO problems gives rise to the second challenge. DFL methodologies with low computational cost are essential for scalability for implementing DFL for real-world large-scale Predict-Then-Optimize problems. The importance of scalability and low computational cost becomes significant while dealing with large-scale CO problems, especially NP-hard combinatorial optimization problems. Note that while the shortest path and the knapsack problems are relatively easy to solve; the energy-cost aware scheduling problem is much more challenging and can be considered an example real-world large-scale NP-hard combinatorial optimization problems. That is why the scheduling problem is considered to compare the computational costs of the DFL methodologies. The median training time of an epoch during training of each methodology for two instances of the scheduling problem are shown Figure 12. Recall that the first, second and third instances contain 10, 15 and 20 tasks respectively. So, the first one is the easiest of the three and the third one is the hardest one. The complexity of the scheduling problem is evident from the fact that a single instance of the the knapsack problem takes 0.001 seconds to solve, while solving the most difficult instance of the scheduling problem takes 0.1 seconds, both using Gurobi MIP solver. The readers are cautioned against placing excessive emphasis on the absolute values of training times in Figure 12, as they are subject to system overhead. However, some general conclusions can be drawn from the relative ordering of the training times. It is not surprising that the training time of the PF approach is the lowest, as it does not require solving the CO problem for model training. Training times of SPO, DBB, I-MLE and FY are almost 100 times higher than the PF approach. Although QPTL and HSD consider the relaxed LP problem, it is not always the case that they have lower training times. Recall that QPTL and HSD solve and differentiate the optimization problem using primal-dual solver, which involves matrix factorization. On the other hand, SPO, DBB, I-MLE and FY can leverage faster commercial optimization solvers, as they only require the optimal solution. However, for Instance 3, it seems solving the ILP Figure 11: Comparative evaluations on the subset selection problem instances of Size 25. This boxplot shows the distributions of mismatch rates. problem is more computationally expensive than solving differentiating the underlying QP problem using Cvxpylayers. On the other hand, Listwise, Pairwise, Pairwise(diff) and MAP, all of which are run with \(p_{solve}=5\%\), exhibit significantly lower training time than the other DFL methodologies. In fact, the training time of these methodologies are comparable to the PF approach. From this perspective, these methodologies can be viewed as bridging the gap between between PF and DFL approaches. The same conclusion generally holds true for other experiments as well. However, for relatively easier CO problems, the system overhead time sometime dominates over model training time, which might disrupt the ordering of the model training time. #### 5.2.3 Discussion The experimental evaluations reveal that no single methodology performs the best across all experiments. Certain methodologies excel on specific test problems, while others perform better on different test problems. Nevertheless, certain interesting characteristics emerge from the experimental evaluations. Firstly, **the performance of SPO is consistently robust across the test problems**, even though it may not outperform other techniques in every experiment. Secondly, **MAP demonstrates consistent performance across most test problems too**; it only exhibits low quality performance specifically in the knapsack problem for Capacity=180 and in the bipartite matching problem when \(\rho_{1}\) and \(\rho_{2}\) are 25%. Additionally, among the LTR losses, Listwise and Pairwise often exhibit high variances, especially in the scheduling and the knapsack problems. The performance of Figure 12: Comparative evaluations of per epoch training time of different DFL methodologies on the energy-cost aware scheduling problem. Pairwise(diff) stands out among the LTR losses due to its lower variance. Its performance is comparable to or slightly worse than MAP for most problems other than the synthetic shortest path problem with high values of Deg, i.e., when the underlying predictive model is completely misspecified. Surprisingly, I-MLE, FY, DBB and QPTL perform worse than the PF approach for the portfolio optimization problem, where a quadratic constraint is present. Across the remaining problems, **the performance of I-MLE is comparable to that of SPO and MAP**. DBB performs considerably worse than I-MLE only in the bipartite matching problem. On the other hand, FY performs well in certain cases, but it is more susceptible to higher variance compared to I-MLE. This is particularly evident in the knapsack problem. Moreover, **QPTL demonstrates robust performance in most experiments**. In fact, QPTL outperforms other models by a substantial margin in the bipartite matching problem. However, QPTL performs poorly compared to others in the scheduling problem, which is an ILP. In this case, the poor performance may be attributed to the fact that QPTL considers a relaxation of the ILP. In this problem the LP solution might differ significantly from the true ILP solution. This is not the case for the knapsack problem, because the solution of the relaxed LP does not deviate significantly from the ILP solution for the knapsack problem. HSD also considers relaxeds for ILP problems. However, it performs worse than QPTL for all but the scheduling problem, where it performs considerably better than QPTL. Finally, due to the limitation of computational resources, we were unable to run QPTL and HSD on the Warcraft shortest path problem. This highlights the advantage of DFL methodologies which can make use of any blackbox combinatorial solver (Dijkstra's shortest path solver for instance) to solve the CO problem. Continuing on this topic of computational cost, MAP and the LTR losses are considerably faster and less computationally intensive when they are run with low values of \(p_{solve}\). As MAP tends to have regret as low as SPO for most test problem, it may be considered a _favorable DFL technique for tackling large-scale real-world Predict-Then-Optimize problems_. ## 6 Future Research Directions While there is increasing interest in decision-focused learning research, it still need to evolve to incorporate new characteristics to tackle real-world problems. This section aims to summarize the wide range of challenges that remain open. A few promising research directions for future investigations that could be exploited in the upcoming days, are presented next. DFL for related tasks/ Task generalization.In the current DFL framework, the ML model is tailored to a particular optimization task. However, in many applications the CO problem might slightly differ in different instantiations. For example, in the recent MIT-Amazon Last Mile Routing Challenge (Merchan, Arora, Pachon, Konduri, Winkenbach, Parks, & Noszek, 2022), a TSP problem is solved every day for deciding the routing of last mile package delivery, but nodes of the TSPs change every day as the delivery locations vary. An interesting research direction would be to investigate how a model, which is trained to minimize regret of one optimization problem, would perform if evaluated on a similar but different optimization problems. Future work need to advance the approach proposed by Tang and Khalil (2023a) by training the ML model with the aim of _generalizing_ to new tasks. Noise contrastive loss functions to learn parameters in the constraints.One key advantage of the noise contrastive loss functions (called MAP in the experimental evaluations) proposed by Mulamba et al. (2021) is that it is differentiable. They view the DFL problem by learning to contrast the likelihood of ground-truth solution and a set of negative examples. However, this work does not consider the case of predicting parameters in the constraints. In future studies, there is potential to extend noise contrastive estimation approach by considering the prediction of parameters within the constraints. This can be achieved by learning to contrast the likelihood of feasible points with that of infeasible ones. However, the efficacy of such an approach may rely on how the infeasible points are selected and that is why an empirical investigation into this aspect would provide valuable insights. Robust decision-focused learning framework to learn parameters in the constraints.While predicting parameters in the constraints of an optimization problem, the prescribed optimal decision might not be feasible with respect to the true parameters. In such scenarios, an interesting direction would be to recommend a solution which is feasible under extreme distributional variations of the parameters. We believe a framework for optimizing average performance and minimizing worst-case constraint violations could reveal new tracks for theoretical research as well as practical applications. The research in this regard can take inspiration from the well-established field of robust optimization (Ben-Tal, El Ghaoui, & Nemirovski, 2009). Surrogate loss functions in the absence of ground-truth cost parameters.In may real-world applications, the true cost parameters of the objective function might be latent variables. In such cases the parameters are not observed, only the solutions are observed. So, the parameters would not be available for supervised learning, which entails the use of a task loss other than regret. DFL frameworks, which implement differentiable optimization layer, such as DBB, QPTL or I-MLE are compatible with any task loss. However, the SPO approach, which comes with a theoretical proof of convergence, requires the ground-truth cost vector for gradient computation. This is also true for noise contrastive and LTR losses, whose computation and differentiation do not involve solving the CO problem. Development of surrogate loss functions, which neither require solving nor the true cost vector, would be a valuable contribution with potentials in real-world applications. Decision-focused learning by score function gradient estimation.Most of the DFL techniques focus on computing the derivative \(\frac{d\mathbf{x}^{*}(\mathbf{\bar{c}})}{d\mathbf{\bar{c}}}\) analytically or construct a surrogate task that provides useful gradients. However, there exists another alternative way to estimate the gradient--zeroth-order estimation of the gradient. A widely used approach to zeroth-order optimization is the score function gradient estimation (Williams, 1992). In order to apply score function gradient estimation in DFL, one has to assume the predicted parameter follows a distribution and then compute Monte Carlo estimate of regret by sampling cost vector from that distribution. The score function gradient estimation would return a gradient that moves the parameters of the distribution in directions that facilitate sampling the cost vector with low values of regret (or task loss in general). Although score function gradient estimation provides unbiased gradient; a major challenge of using this technique is it suffers from high variances, which might destabilize the learning. Hence, conducting further research to examine the potential application of score function gradient estimation in DFL would be a valuable contribution. Non-linear objective function.Most of the works in DFL, consider optimization problems with linear objectives. This is the primary reason such problems have been considered for experimental evaluations in this work. Any convex optimization problem with nonlinear objective function can be differentiated through CVxypylayers (Agrawal et al., 2019a). However, no DFL technique considers nonlinear objectives with discrete decision variables. As many real-world problems in OR are combinatorial optimization problems with discrete decision variables, developing ML techniques for such problems in the future could be beneficial in real life. For examples, the problem of optimally locating substations in an electrical network to minimize the costs of distribution is formulated as nonlinear programming (Lakhera, Shanbhag, & McInerney, 2011). Another classic OR problem, which does not have a linear objective function is the minimization of makespan in flowshop scheduling. Most of the methodologies discussed in this paper are not applicable to handle such problems. Bilevel optimization techniques for DFL.As mentioned in Section 2.2, the empirical regret minimization problem can be cast as a pessimistic bilevel optimization problem. We believe that by understanding the mathematical object behind the learning process can lead to better algorithms for DFL, leaving a door open to the bilevel optimization community to tackle this problem. Optimization as an intermediate layer within neural networks.In a Predict-Then-Optimize problem, the final task is to make a decision by solving a CO problem. However, in many other applications the optimization task may appear as an intermediate task. For instance, consider the task of selecting relevant patches in high resolution images, where the patches are being used for a downstream image recognition task. In (Cordonnier, Mahendran, Dosovitskiy, Weissenborn, Uszkoreit, & Unterthiner, 2021) the patch selection task is modeled as a Top-\(k\) selection CO problem. Note that the Top-\(k\) selection is embedded as an intermediate layer between two neural networks; where the upstream neural network assign score to each patch and the downstream neural network performs the recognition task. Techniques such as I-MLE, DBB, QPTL, DPO, which are implementation of differentiable optimization layer can be applied to tackle problems like this. Although the existence of downstream layer after the CO problem may give rise to novel challenges, embedding the CO problem as an intermediate layer could find extensive use across various domains. Construction of solution cache.The loss functions which utilize solution cache are very effective to address the computational cost of DFL and promising for large NP-hard real-world Predict-Then-Optimize problems. However, we believe there is a space for research to study the trade-off between the solution cache size and solution quality. ## 7 Conclusion The survey article begins by underscoring the significance of Predict-Then-Optimize problem formulations, wherein an ML model is followed by a CO problem. The Predict-Then-Optimize problem has emerged as a powerful driving force in numerous real-world applications of artificial intelligence, operations research and business analytics. The key challenge in Predict-Then-Optimize problems is predicting the unknown CO problem parameters in a manner that yields high-quality _solutions_, in comparison to the retrospective solutions obtained when using the groundtruth parameters. To address this challenge, the DFL paradigm has been proposed, wherein the ML models are directly trained considering the CO problems using task losses that capture the error encountered after the CO problems. However to date, there is no comprehensive survey on DFL. This survey provides a comprehensive overview of DFL, highlighting recent technological advancements, applications and identifying potential future research directions. In Section 2, the problem description has been laid out with examples and then the fundamental challenges in decision-focused learning have been presented. Afterward, Section 3 has presented a categorization with four categories of DFL techniques, which have been thoroughly explained highlighting the trade-offs among these four categories. Then, in Section 4, some examples of applications of DFL techniques to address real-world Predict-Then-Optimize problems, across different domains have been provided. Furthermore, extensive comparative evaluations on different problem sets between 11 DFL techniques have been provided in Section 5. Finally, a discussion of some of the open problems in DFL and an outline of potential research directions have been presented in Section 6. While there has been significant recent progress in DFL, there remain challenges that need to be addressed. For instance, the development of DFL techniques, which can handle uncertain parameters occurring anywhere within a generic CO problem, will have a significant impact on various industrial applications. We hope this survey article will assist readers to understand the paradigm of decision-focused learning and grasp the fundamental challenges of implementing it in many real-world applications. We aspire this survey to potentially act as a catalyst, inspiring the application of decision-focused learning in diverse domains and contexts as well as stimulating further methodological research and advancements. ## Acknowledgments This project received partial funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program under Grant no. 101070149 and Grant No. 101002802 (CHAT-Opt) and the FWO Flanders project G070521N. This research is also partially supported by NSF grants 2242931, 2232054, 2007164, and NSF CAREER award 2143706. Victor Bucarey was funded by the ANID Fondecyt Iniciacion Grant no. 11220864.

2307.04019

GP-guided MPPI for Efficient Navigation in Complex Unknown Cluttered Environments

Robotic navigation in unknown, cluttered environments with limited sensing capabilities poses significant challenges in robotics. Local trajectory optimization methods, such as Model Predictive Path Intergal (MPPI), are a promising solution to this challenge. However, global guidance is required to ensure effective navigation, especially when encountering challenging environmental conditions or navigating beyond the planning horizon. This study presents the GP-MPPI, an online learning-based control strategy that integrates MPPI with a local perception model based on Sparse Gaussian Process (SGP). The key idea is to leverage the learning capability of SGP to construct a variance (uncertainty) surface, which enables the robot to learn about the navigable space surrounding it, identify a set of suggested subgoals, and ultimately recommend the optimal subgoal that minimizes a predefined cost function to the local MPPI planner. Afterward, MPPI computes the optimal control sequence that satisfies the robot and collision avoidance constraints. Such an approach eliminates the necessity of a global map of the environment or an offline training process. We validate the efficiency and robustness of our proposed control strategy through both simulated and real-world experiments of 2D autonomous navigation tasks in complex unknown environments, demonstrating its superiority in guiding the robot safely towards its desired goal while avoiding obstacles and escaping entrapment in local minima. The GPU implementation of GP-MPPI, including the supplementary video, is available at https://github.com/IhabMohamed/GP-MPPI.

Ihab S. Mohamed, Mahmoud Ali, Lantao Liu

2023-07-08T17:33:20Z

http://arxiv.org/abs/2307.04019v3

# GP-guided MPPI for Efficient Navigation in Complex Unknown Cluttered Environments ###### Abstract Robotic navigation in unknown, cluttered environments with limited sensing capabilities poses significant challenges in robotics. Local trajectory optimization methods, such as Model Predictive Path Integral (MPPI), are a promising solution to this challenge. However, global guidance is required to ensure effective navigation, especially when encountering challenging environmental conditions or navigating beyond the planning horizon. This study presents the GP-MPPI, an _online_ learning-based control strategy that integrates MPPI with a local perception model based on Sparse Gaussian Process (SGP). The key idea is to leverage the learning capability of SGP to construct a variance (uncertainty) surface, which enables the robot to learn about the navigable space surrounding it, identify a set of suggested subgoals, and ultimately recommend the optimal subgoal that minimizes a predefined cost function to the local MPPI planner. Afterward, MPPI computes the optimal control sequence that satisfies the robot and collision avoidance constraints. Such an approach eliminates the necessity of a global map of the environment or an offline training process. We validate the efficiency and robustness of our proposed control strategy through both simulated and real-world experiments of 2D autonomous navigation tasks in complex unknown environments, demonstrating its superiority in guiding the robot safely towards its desired goal while avoiding obstacles and escaping entrapment in local minima. The GPU implementation of GP-MPPI, including the supplementary video, is available at [https://github.com/IhabMohamed/GP-MPPI](https://github.com/IhabMohamed/GP-MPPI). Autonomous vehicle navigation, MPPI, sparse Gaussian process (SGP), occupancy grid map path planning. ## I Introduction and Related Work Autonomous navigation of mobile robots in unknown, cluttered, and unpredictable environments with limited sensor capabilities is a challenging task owing to the inherent uncertainty and complexity of such environments. To tackle this challenge, a _receding-horizon_ strategy such as Model Predictive Control (MPC) is commonly employed. The MPC control framework allows the robot to simultaneously plan a short trajectory (sequence of actions), following which the robot executes the immediate action while planning a subsequent trajectory. To successfully achieve receding-horizon planning, the robot must consider both safety and persistent feasibility, where _safety_ is achieved by avoiding collisions with any obstacles while executing a planned trajectory, and _persistent feasibility_ is maintained by always generating a safe trajectory that does not result in dead-ends or local minima while progressing towards the desired goal. One of the significant challenges in robot motion planning is that the desired goal is often situated beyond the planning horizon, which requires the use of local subgoals or _cost-to-go_ heuristics for motion safety and persistent feasibility. A common strategy is to rely on single-query motion planning algorithms, such as A\({}^{*}\) and RRT\({}^{\text{X}}\), to identify feasible paths that direct the local planner towards its desired goal [1, 2]. For instance, the RRT\({}^{\text{X}}\) algorithm, introduced in [2], incorporates replanning techniques from Dynamic Rapidly-exploring Random Trees (DRRT) and Rapid-exploring Random Trees (RRT\({}^{*}\)) algorithms to adjust the path during exploration based on environmental changes. However, due to its high computational demands, implementing this algorithm in _real-time_ on a robot can be challenging. One alternative method to achieve efficient solutions for motion planning problems is the integration of MPC with data-driven methods, also known as learning-based MPC [3]. To name a few, a subgoal planning policy using Deep Reinforcement Learning (DRL) is recently proposed to guide the local MPC planner to navigate in crowded surroundings [4, 5]. Similarly, RL was utilized to choose the next subgoal from a set of predefined possibilities [6], which guides the robot through challenging environments with dead-end corridors while also prevents the MPC planner from getting trapped in local minima. Another related work that combines learning with MPC is POLO which aims to enhance MPC performance by learning a global value function [7]. Most of these approaches typically rely on either offline training or having access to the global map of the environment. In addition, many recent studies have suggested combining Gaussian Process (GP) with MPC to learn system dynamics, leading to better control performance and robustness to uncertainty [8]. Another research avenue employed gap-based techniques Fig. 1: Architecture of our proposed GP-MPPI control strategy, which comprises two main components: the GP-subgoal recommender and the local planner, the MPPI. First, the GP-subgoal recommender observes the surrounding environment and suggests the optimal subgoal position \(\mathbf{g}^{*}\) to the local motion planner, where four colored circles represent the GP-recommended subgoals. MPPI then computes the optimal control sequence, which minimizes the distance to \(\mathbf{g}^{*}\) while avoiding collision with obstacles and respecting system constraints, followed by executing the first optimal control \(\mathbf{u}_{0}\) to the robot.

2305.12002

XuanYuan 2.0: A Large Chinese Financial Chat Model with Hundreds of Billions Parameters

In recent years, pre-trained language models have undergone rapid development with the emergence of large-scale models. However, there is a lack of open-sourced chat models specifically designed for the Chinese language, especially in the field of Chinese finance, at the scale of hundreds of billions. To address this gap, we introduce XuanYuan 2.0, the largest Chinese chat model to date, built upon the BLOOM-176B architecture. Additionally, we propose a novel training method called hybrid-tuning to mitigate catastrophic forgetting. By combining general-domain with domain-specific knowledge and integrating the stages of pre-training and fine-tuning, XuanYuan 2.0 is capable of providing accurate and contextually appropriate responses in the Chinese financial domain.

Xuanyu Zhang, Qing Yang, Dongliang Xu

2023-05-19T21:01:20Z

http://arxiv.org/abs/2305.12002v1

# XuanYuan 2.0: A Large Chinese Financial Chat Model ###### Abstract In recent years, pre-trained language models have undergone rapid development with the emergence of large-scale models. However, there is a lack of open-sourced chat models specifically designed for the Chinese language, especially in the field of Chinese finance, at the scale of hundreds of billions. To address this gap, we introduce **XuanYuan 2.0** (\(\ddagger\) 2.0), the largest Chinese chat model to date, built upon the BLOOM-176B architecture. Additionally, we propose a novel training method called hybrid-tuning to mitigate catastrophic forgetting. By combining general-domain with domain-specific knowledge and integrating the stages of pre-training and fine-tuning, XuanYuan 2.0 is capable of providing accurate and contextually appropriate responses in the Chinese financial domain. ## 1 Introduction In recent years, pre-trained language models have witnessed rapid development. Broadly speaking, they can be categorized into three main architectures: the Encoder architecture represented by BERT (Devlin et al., 2018), the Decoder architecture represented by GPT (Radford et al., 2018), and the Encoder-Decoder architecture represented by T5 (Raffel et al., 2020). Each architecture has its unique characteristics and advantages, catering to different NLP requirements. The GPT series, with GPT-4 (OpenAI, 2023) being the latest addition, has gained considerable attention due to its remarkable performance in natural language generation tasks, including dialogue generation. The ChatGPT (OpenAI, 2022) model, in particular, has impressed researchers and practitioners with its ability to generate coherent and contextually relevant responses in conversational settings. As a result, the GPT series has become a focal point of research and development in the NLP community. Moreover, the emergence of large-scale pre-trained models has further fueled the advancements in language modeling. Models such as OPT (Zhang et al., 2022), BLOOM (Scao et al., 2022), and LLaMA (Touvron et al., 2023), with parameter sizes reaching billions, have recently been open-sourced, enabling researchers and developers to explore the potential of these massive models. These models have demonstrated superior performance on various tasks, pushing the boundaries of what is possible in NLP. While the general-purpose large models mentioned above have garnered significant attention, the importance of domain-specific models cannot be overlooked. In many domains, the distribution of language and the specific linguistic nuances require models that are fine-tuned or specifically trained for that particular domain. Consequently, a range of domain-specific large models has been proposed to cater to the unique needs of various fields. For example, BioBERT (Lee et al., 2020) and PubMedBERT (Gu et al., 2021) are proposed for the biomedical field, and BloombergGPT (Wu et al., 2023) are proposed for financial scenarios. These models have shown promising results in their respective domains, leveraging the domain-specific knowledge learned during pre-training. Within the Chinese financial domain, there has been considerable progress in the development of pre-trained language models. Researchers have introduced models such as FinBERT (Araci, 2019; Yang et al., 2020; Liu et al., 2021), Mengzi (Zhang et al., 2021), and FinT5 (Lu et al., 2023), which have been tailored for financial text analysis and understanding. These models, though valuable for certain applications, have parameter sizes below one billion, limiting their ability to handle the increasing demands of the Chinese financial NLP landscape. As the volume of financial data and the complexity of language usage continue to grow, there is a pressing need for more powerful models that can effectively process and understand Chinese financial text. Despite significant advancements in chat models, there is currently no open-sourced chat model at the scale of hundreds of billions specifically designed for the Chinese language, let alone in the field of Chinese finance. To address this gap, we propose **XuanYuan 2.0** (\(\frac{\text{FT-14}}{\text{SR}}\) 2.0), the largest Chinese chat model to date, based on BLOOM-176B. XuanYuan 2.0 not only surpasses its predecessor, **XuanYuan 1.0** (\(\frac{\text{FT-14}}{\text{SR}}\) 1.0), which achieved first place at the leaderboard of CLUE classification in 2021, but also addresses the need for a large-scale chat model specifically designed for the Chinese financial domain. Furthermore, domain-specific language models and chat models impose higher requirements on data distribution and training approaches compared to general-domain models. Domain-specific models need to capture the unique linguistic characteristics, terminologies, and contexts of a particular field to achieve optimal performance. However, training these models solely on domain-specific data may lead to catastrophic forgetting, where the model loses previously learned knowledge from the general domain, impacting its overall performance. To mitigate this issue, we propose a novel training method, hybrid-tuning, that combines the stages of pre-training and fine-tuning. By integrating the two stages, our approach guarantees that fine-tuning the model with financial-specific instructions does not impede its general generation capabilities acquired during pre-training. As a result, XuanYuan 2.0 can effectively leverage both its general-domain knowledge and domain-specific financial knowledge to provide accurate and contextually appropriate responses in the Chinese financial domain. ## 2 Related Work The advancements in pre-trained language models have led to remarkable progress in various NLP tasks, attracting extensive research efforts. Among the notable contributions, the BERT Devlin et al. (2018) series stands out as a groundbreaking development in the field of pre-trained models. Following the success of BERT, the GPT Radford et al. (2018) series emerged as a prominent line of research, focusing on the decoding aspect of language modeling. GPT models, in contrast to BERT's bidirectional approach, leveraged autoregressive language modeling. By training on large amounts of unlabeled text data, GPT models acquired a rich understanding of language and demonstrated impressive capabilities in generating coherent and contextually relevant text. Subsequent iterations of the GPT series, such as GPT-4 OpenAI (2023), showcased superior performance in various language generation tasks. And Chat-GPT (OpenAI, 2022), an extension of the GPT series, demonstrated the ability to engage in interactive and contextually coherent conversations. This breakthrough sparked considerable interest in developing conversational AI agents capable of simulating human-like dialogue. In addition to the general-purpose BERT and GPT models, there has been a growing interest in domain-specific pre-training. Researchers have recognized that incorporating domain-specific knowledge during pre-training can lead to substantial performance gains in downstream tasks within those domains. Domain-specific pre-trained models aim to capture domain-specific nuances, enabling them to excel in tasks relevant to the target domain. For instance, in the biomedical domain, BioBERT Lee et al. (2020) and PubMedBERT Gu et al. (2021) are proposed to leverage large-scale biomedical \begin{table} \begin{tabular}{p{113.8pt} p{56.9pt} p{56.9pt} p{56.9pt}} \hline \hline **Model** & **Type** & **Parameter** & **Corpus Content** \\ \hline FinBERT Araci (2019) & PLM & 110M & News filtered by financial keywords \\ FinBERT Yang et al. (2020) & PLM & 110M & Corporate Reports, Earnings Call Transcripts, Analyst Reports \\ Mengzi-BERT-base-fin Zhang et al. (2021) & PLM & 110M & News, Analyse reports, Company announcements \\ FinT5 Lu et al. (2023) & PLM & 220M, 1B & Corporate Reports, Analyst Reports, Social media and Financial News \\ \hline XuanYuan 2.0 & ChatLM & 176B & Corporate Reports, Analyst Reports, Social media and Financial News \\ \hline \hline \end{tabular} \end{table} Table 1: Comparison of different financial language models. corpora during pre-training. Similarly, in the financial domain, models such as BloombergGPT Wu et al. (2023) were developed to address the unique challenges and intricacies of the financial text. Despite the advancements in domain-specific pre-training, the availability of large-scale open-source chat models specifically tailored for the Chinese language and the Chinese financial domain has remained limited. This gap motivates our work in proposing XuanYuan 2.0, a model built upon BLOOM-176B Scao et al. (2022) with hundreds of billions parameters, to address the unique requirements of the Chinese financial domain and facilitate the development of sophisticated conversational AI systems. ## 3 XuanYuan 2.0 ### Model Architecture We adopted the original BLOOM Scao et al. (2022) architecture, which is a decoder-only architecture. The joint probability of tokens in a text can be represented as: \[p(w)=p(w_{1},\dots,w_{T})=\prod_{t=1}^{T}p(w_{t}|w_{<t}) \tag{1}\] where \(w\) represents a sequence of tokens, \(w_{t}\) is the \(t^{\rm th}\) token, and \(w_{<t}\) is the sequence of tokens preceding \(w_{t}\). This method is called autoregressive language modeling, where we predict the probability of the next token in an iterative manner. And following BLOOM, we utilize ALBi positional embeddings Press et al. (2021) and embedding LayerNorm Dettmers et al. (2022) in the traditional decoder structure of Transformer Vaswani et al. (2017). ### Hybrid-tuning To alleviate the problem of catastrophic forgetting, we propose a novel domain-specific training framework, hybrid-tuning. In terms of the training stage, it integrates the pre-training stage and instruction fine-tuning stage that are previously split together. In terms of the field of data, it integrates data from both general and financial domains. As shown in Figure 1, different from traditional two-stage domain-specific training, our proposed hybrid-tuning randomly shuffles pre-training data (general pre-training, financial pre-training) and instruction data (general instruction, financial instruction) into one training data. And all the training Figure 1: Our proposed hybrid-tuning. process is done in one stage. In this way, the model can accurately handle instructions in the financial domain, while retaining general conversational capabilities. For unsupervised pre-training data, we crawl them from the Internet and clean and filter them. For Instruction-tuning data, we use human-written seed instructions to collect general data by Self-Instruct Wang et al. (2022) and utilize unstructured and structured data in the financial field to gather domain-specific instruction data by Self-QA Zhang and Yang (2023). Unstructured financial data comprises a wide range of textual information, such as financial news articles, market reports, analyst commentary, and social media discussions. And structured financial data includes company information and so on. These sources offer valuable insights into market trends, investment strategies, and economic situations. ### Training To train our complex and computationally intensive model, we employ the powerful NVIDIA A100 80GB GPU and the DeepSpeed Rasley et al. (2020) distributed training framework. For parallel processing, we primarily rely on pipeline parallelism, which involves distributing the layers of our model across several GPUs. This approach ensures that each GPU only handles a portion of the model's layers, a technique also known as vertical parallelism. Additionally, we adopt the Zero Redundancy Optimizer Rajbhandari et al. (2020) to enable different processes to store only a portion of the data (parameters, gradients, and optimizer states). Specifically, we use ZeRO stage 1, which means that only the optimizer states are divided using this method. The specific hyperparameters are presented in Table 2. ## 4 Experiment We conducted a comparison between our model and other open-source Chinese conversational models. Simultaneously, we constructed evaluation datasets encompassing various dimensions in both general and financial domains, which were subsequently subject to manual assessment. The results revealed XuanYuan's robust knowledge base and conversational capabilities in the financial domain. Further insights and additional findings will be presented in the next version of the paper after the release of the evaluation rankings. ## 5 Conclusion In this paper, we propose the largest Chinese financial chat model, XuanYuan 2.0 (\(\frac{\text{Total}}{\text{Total}}\) 2.0), to fill the gap of open-source billion-scale chat models specifically designed for the Chinese financial domain. Besides, we propose a novel training method called hybrid-tuning to mitigate catastrophic forgetting. By combining the general domain with domain-specific knowledge and integrating the stages of pre-training and finetuning, XuanYuan 2.0 achieves the remarkable ability to deliver precise and contextually relevant responses within the Chinese financial domain. We will continue to gather larger-scale Chinese financial domain data in order to further optimize our model. \begin{table} \begin{tabular}{l|c|c} \hline \hline Hyperparameter & XuanYuan2-7B & XuanYuan2 \\ \hline \multicolumn{4}{c}{_Architecture hyperparameters_} \\ \hline Parameters & 7,069M & 176,247M \\ Layers & 30 & 70 \\ Hidden dim. & 4096 & 14336 \\ Attention heads & 32 & 112 \\ Vocab size & 250,680 & \\ Sequence length & 2048 & \\ Precision & float16 & \\ Activation & GELU & \\ Position emb. & Alibi & \\ Tied emb. & True & \\ \hline \hline \multicolumn{4}{c}{_Pretraining hyperparameters_} \\ \hline Global Batch Size & 512 & 2048 \\ Learning rate & 1.2e-4 & 6e-5 \\ Total tokens & 341B & 366B \\ Min. learning rate & 1e-5 & 6e-6 \\ Warmup tokens & 375M & \\ Decay tokens & 410B & \\ Decay style & cosine & \\ Adam \((\beta_{1},\beta_{2})\) & (0.9, 0.95) & \\ Weight decay & 1e-1 & \\ Gradient clipping & 1.0 & \\ \hline \hline \multicolumn{4}{c}{_Multitask finetuning hyperparameters_} \\ \hline Global Batch Size & 2048 & 2048 \\ Learning rate & 2.0e-5 & 2.0e-5 \\ Total tokens & & 13B \\ Warmup tokens & & 0 \\ Decay style & constant & \\ Weight decay & 1e-4 & \\ \hline \hline \end{tabular} \end{table} Table 2: Training hyperparameters of XuanYuan 2.0.

2306.02999

Von Neumann Dimensions and Trace Formulas I: Limit Multiplicities

Given a connected semisimple Lie group $G$ and an arithmetic subgroup $\Gamma$, it is well-known that each irreducible representation $\pi$ of $G$ occurs in the discrete spectrum $L^2_{\text{disc}}(\Gamma\backslash G)$ of $L^2(\Gamma\backslash G)$ with at most a finite multiplicity $m_{\Gamma}(\pi)$. While $m_{\Gamma}(\pi)$ is unknown in general, we are interested in its limit as $\Gamma$ is taken to be in a tower of lattices $\Gamma_1\supset \Gamma_2\supset\dots$. For a bounded measurable subset $X$ of the unitary dual $\widehat{G}$, we let $m_{\Gamma_n}(X)$ be the sum of the multiplicity $m_{\Gamma_n}(\pi)$ of a representation $\pi$ over all $\pi$ in $X$. Let $H_X$ be the direct integral of the irreducible representations in $X$, which is also a module over the group von Neumann algebra $\mathcal{L}\Gamma_n$. We prove: \begin{center} $\lim\limits_{n\to \infty}\cfrac{m_{\Gamma_n}(X)}{\dim_{\mathcal{L}\Gamma_n}H_X}=1$, \end{center} for any bounded subset $X$ of $\widehat{G}$, when i) $\Gamma_n$'s are cocompact, or, ii) $G=\SL(n,\mathbb{R})$ and $\{\Gamma_n\}$ are principal congruence subgroups.

Jun Yang

2023-06-05T16:10:07Z

http://arxiv.org/abs/2306.02999v1

# Von Neumann dimensions and trace formulas I: limit multiplicities ###### Abstract. Given a connected semisimple Lie group \(G\) and an arithmetic subgroup \(\Gamma\), it is well-known that each irreducible representation \(\pi\) of \(G\) occurs in the discrete spectrum \(L^{2}_{\mathrm{disc}}(\Gamma\backslash G)\) of \(L^{2}(\Gamma\backslash G)\) with at most a finite multiplicity \(m_{\Gamma}(\pi)\). While \(m_{\Gamma}(\pi)\) is unknown in general, we are interested in its limit as \(\Gamma\) is taken to be in a tower of lattices \(\Gamma_{1}\supset\Gamma_{2}\supset\dots\). For a bounded measurable subset \(X\) of the unitary dual \(\widehat{G}\), we let \(m_{\Gamma_{n}}(X)\) be the sum of the multiplicity \(m_{\Gamma_{n}}(\pi)\) over all \(\pi\) in \(X\). Let \(H_{X}\) be the direct integral of the irreducible representations in \(X\) with respect to the Plancherel measure of \(G\), which is also a module over the group von Neumann algebra \(\mathcal{L}\Gamma_{n}\). We prove: \[\lim_{n\to\infty}\frac{m_{\Gamma_{n}}(X)}{\dim_{\mathcal{L}\Gamma_{n}}H_{X}}=1,\] for any bounded subset \(X\) of \(\widehat{G}\), when i) \(\Gamma_{n}\)'s are cocompact, or, ii) \(G=\mathrm{SL}(n,\mathbb{R})\) and \(\{\Gamma_{n}\}\) are principal congruence subgroups. AMS 2010 Mathematics Subject Classification: 46L10, 20G05, 20G35. \(\mathbb{E}\)Jun Yang [email protected] Harvard University, Cambridge, MA, USA This work was supported in part by the ARO Grant W911NF-19-1-0302 and the ARO MURI Grant W911NF-20-1-0082. ## 1. Introduction: an example on \(\operatorname{SL}(2,\mathbb{R})\) In this section, we introduce a multiplicity problem of square-integrable irreducible representations of \(G=\operatorname{SL}(2,\mathbb{R})\) on \(L^{2}_{\operatorname{cusp}}(\Gamma\backslash G)\) for some arithmetic subgroups \(\Gamma\). It is one of the motivations of this article. We first \(\Gamma=\operatorname{SL}(2,\mathbb{Z})\) and \(\Gamma(N)\) be the principal congruence subgroup of level \(n\) defined by \[\Gamma(N)=\Big{\{}\begin{pmatrix}a&b\\ c&d\end{pmatrix}\in\operatorname{SL}(2,\mathbb{Z}):a,d\equiv 1\pmod{N},b,c\equiv 0 \pmod{N}\Big{\}}.\] Consider the right quasi-regular representation of \(G\) on \(L^{2}(\Gamma(N)\backslash G)\) given by \((R(g)\phi)(x)=\phi(xg)\) for \(\phi\in L^{2}(\Gamma(N)\backslash G)\), \(g\in G\). It is well known (see [28]) to be reducible and can be decomposed as \[L^{2}(\Gamma(N)\backslash G)=L^{2}_{\operatorname{cusp}}(\Gamma(N)\backslash G )\oplus L^{2}_{\operatorname{cont}}(\Gamma(N)\backslash G)\oplus\mathbb{C}.\] Here \(L^{2}_{\operatorname{cusp}}(\Gamma(N)\backslash G)\) is the cuspidal part, which is a direct sum of irreducible representations with finite multiplicities, i.e., \[L^{2}_{\operatorname{cusp}}(\Gamma(N)\backslash G)=\sum m_{\Gamma(N)}(\pi) \cdot\pi,\,m_{\Gamma(N)}(\pi)<\infty\text{ for each }\pi,\] and \(L^{2}_{\operatorname{cont}}(\Gamma(N)\backslash G)\) is a direct integral of irreducible representations given by the Eisenstein series. The multiplicities \(m_{\Gamma(N)}(\pi)\) are still unknown in general, except for some special families of irreducible representations including the discrete series of \(\operatorname{SL}(2,\mathbb{R})\) (see [26] for an introduction of discrete series). Let \(S_{k}(\Gamma)\) be the space of cusp forms of weight \(k\) for a Fuchsian group \(\Gamma\). We have the following result (see [22] Theorem 2.10). **Lemma 1.1**.: _For the discrete series \(\pi_{k}\), we have \(m_{\Gamma(N)}(\pi_{k})=\dim S_{k}(\Gamma(N))\)._ By applying the dimension formulas of cusp forms (see [14] Chapter 3.9), we obtain \[m_{\Gamma(N)}(\pi_{k})=(\frac{k-1}{24}-\frac{1}{4N})N^{3}\prod_{p|N}(1-\frac{1} {p^{2}}) \tag{1}\] for all \(N>2\). On the other hand, let \(H_{k}\) be the underlying Hilbert space of the discrete series \(\pi_{k}\). As \(H_{k}\) is a module over the group \(\Gamma(N)\), we can further prove that it is also a module over the _group von Neumann algebra_\(\mathcal{L}(\Gamma(N))\) (see Section 4.1 for the definition). Hence \(H_{k}\) has a _von Neumann dimension_\(\dim_{\mathcal{L}(\Gamma(N))}H_{k}\) over \(\mathcal{L}(\Gamma(N))\). Indeed, if a discrete group \(\Gamma\) is ICC (infinite conjugacy class, see also 4.1), this dimension totally determines the equivalence class of \(\mathcal{L}(\Gamma)\)-module, i.e., \(\dim_{\mathcal{L}(\Gamma)}H_{1}=\dim_{\mathcal{L}(\Gamma)}H_{2}\) if and only if \(H_{1},H_{2}\) are isomorphic as \(\mathcal{L}(\Gamma)\)-module. We consider a lattice \(\Gamma\) in a Lie group \(G\). Suppose \((\pi,H)\) is a discrete series representation of \(G\) and let \(d(\pi)\) be the formal dimension of \(\pi\) (see [33] Chapter 16). We have **Lemma 1.2** (Goodman-la Harpe-Jones[23]).: \(\dim_{\mathcal{L}(\Gamma)}H=\operatorname{vol}(\Gamma\backslash G)\cdot d(\pi)\)__ By Example 3.3.4 in [23], we know \(\dim_{\mathcal{L}(\operatorname{PSL}(2,\mathbb{Z}))}H_{k}=\frac{k-1}{12}\). As \(\operatorname{SL}(2,\mathbb{Z})=(\mathbb{Z}/2\mathbb{Z})\rtimes\operatorname{PSL }(2,\mathbb{Z})\), we have \(\dim_{\mathcal{L}(\operatorname{SL}(2,\mathbb{Z}))}H_{k}=\frac{k-1}{24}\). Since \([\Gamma\colon\Gamma(n)]=N^{3}\prod_{p|N}(1-\frac{1}{24})\), we have \(\dim_{\mathcal{L}(\operatorname{SL}(2,\mathbb{Z}))}H_{k}=\frac{k-1}{24}\). \(\frac{1}{p^{2}}\)), we can conclude \[\dim_{\mathcal{L}(\Gamma(N))}H_{k}=\frac{k-1}{24}N^{3}\prod_{p|N}(1-\frac{1}{p^{2 }}). \tag{2}\] Thus we obtain: **Corollary 1.3**.: _For a discrete series \((\pi_{k},H_{k})\) of \(\operatorname{SL}(2,\mathbb{R})\), we have_ \[\lim_{N\to\infty}\frac{m_{\Gamma(N)}(\pi_{k})}{\dim_{\mathcal{L}(\Gamma(N))}H _{k}}=1.\] **Proof**: Comparing Equations 1 and 2, we obtain \(\frac{m_{\Gamma(N)}(\pi_{k})}{\dim_{\mathcal{L}(\Gamma(N))}H_{k}}=\frac{k-1-6/ N}{k-1}\) and then take the limit. While the explicit multiplicities of most irreducible representations are still unknown, the limit multiplicities have been studied since 1970s. In the case of towers of uniform lattices, DeGeorge and Wallach got the first results for discrete series of Lie groups [11] and later for bounded sets of irreducible representations in rank one group [12]. Delorme [13] finally solved the problem for bounded sets of irreducible representations in all Lie groups. See also [1] for a recent approach. For the non-uniform lattices (or most arithmetic subgroups), Savin [36] obtained the results on discrete series in his thesis at first, which is based on the work by Rohlfs and Speh [34]. Then Deitmar and Hoffmann proved the results on certain towers of arithmetic subgroups in rank one group. Recently, Finis and Lapid solved the case of congruence subgroups in \(\operatorname{SL}(n,\mathbb{R})\)[20, 17], which are based on their study of the spectral side of Arthur's trace formulas [18, 16]. The goal of this paper is to extend Corollary 1.3 to some general settings. In the rest of this paper, we generalize this result mainly in the following aspects: 1. from a single discrete series representation to any bounded subset of the unitary dual \(\widehat{G}\) of \(G\); 2. from \(\operatorname{SL}(2,\mathbb{R})\) to the towers of uniform lattices in an arbitrary semisimple Lie group, 3. from \(\operatorname{SL}(2,\mathbb{R})\) to \(\operatorname{SL}(n,\mathbb{R})\) with its the principal congruence subgroups. Finally, we are able to prove: **Theorem 1.4** (**The Main Theorem**).: _Let \(G\) be a semisimple simply-connected Lie group. Let \(X\) be a bounded subset of the unitary dual of \(G\) and \(H_{X}\) be the direct integral of the irreducible representations of \(G\) in \(X\). We have:_ \[\lim_{n\to\infty}\frac{m_{\Gamma_{n}}(X)}{\dim_{\mathcal{L}\Gamma_{n}}H_{X}}=1\] _when i) \(\Gamma_{n}\)'s are cocompact, or ii) \(G=\operatorname{SL}(n,\mathbb{R})\) and \(\{\Gamma_{n}\}\) are principal congruence subgroups._ ## 2. The trace formulas and dominant terms We have a brief review of the Arthur-Selberg trace formulas and give the dominant terms in these formulas. We mainly follow [3, 4, 19]. Let \(\mathbf{G}\) be a reductive group over \(\mathbb{Q}\). The group \(G(\mathbb{A})\) acts naturally on \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A}))\) by \[R(g)\phi(x)=\phi(xg)\] for \(\phi\in L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A}))\) and \(g\in G(\mathbb{A})\)). Let \(C_{\mathrm{c}}^{\infty}(G(\mathbb{A}))\) be the complex algebra of smooth, compactly supported function on \(G(\mathbb{A})\)). Given \(f\in C_{\mathrm{c}}^{\infty}(G(\mathbb{A}))\), we may define \[(R(f)\phi)(x)=\int_{G(\mathbb{A}))}f(g)R(g)\phi(x)dg=\int_{G(\mathbb{A}))}f(g) \phi(xg)dg.\] If we define the _kernel_ \[K(x,y)=K_{f}(x,y)\colon=\sum\limits_{\gamma\in G(\mathbb{Q})}f(x^{-1}\gamma y),\] we have \((R(f)\phi)(x)=\int_{G(\mathbb{Q})\backslash G(\mathbb{A})}K(x,y)\phi(y)dy\). ### The Selberg trace formula We first assume \(\mathbf{G}\) is anisotropic and hence the quotient space \(G(\mathbb{Q})\backslash G(\mathbb{A})\) is compact. Let \(\mathcal{O}\) be the set of conjugacy classes in \(G(\mathbb{Q})\) and \(o\in\mathcal{O}\) be a conjugacy class. We may define \[K_{o}(x,y)=\sum\limits_{\gamma\in o}f(x^{-1}\gamma y)\] and obtain \(K(x,y)=\sum\limits_{o\in\mathcal{O}}K_{o}(x,y)\). On the other hand, the representation \(R\) decomposes into a direct sum of irreducible representations with finite multiplicities, i.e., \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A}))=\oplus_{\chi\in\mathcal{X}}L^{2} (G(\mathbb{Q})\backslash G(\mathbb{A}))_{\chi}\). Here \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A}))_{\chi}=m(\chi)\cdot\chi\), which is \(m(\chi)\) copies of the irreducible representation \(\chi\). Assume \(\mathcal{B}_{\chi}\) is a orthonormal basis of \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A}))_{\chi}\). Then \[K_{\chi}(x,y)=K_{f,\chi}(x,y)\colon=\sum_{\phi\in\mathcal{B}_{\chi}}(R(f)) \phi(x)\cdot\overline{\phi(y)}\] converges. Now we let 1. \(k_{\chi}(x,f)=K_{\chi}(x,x)\) and \(J_{\chi}(f)=\int_{G(\mathbb{Q})\backslash G(\mathbb{A})}k_{\chi}(x,f)dx\), 2. \(k_{o}(x,f)=K_{o}(x,x)\) and \(J_{o}(f)=\int_{G(\mathbb{Q})\backslash G(\mathbb{A})}k_{o}(x,f)dx\). If we let \(\gamma\) be a representatives of \(o\in\mathcal{O}\) and \(H_{\gamma}=\{h\in H|h\gamma h^{-1}=\gamma\}\) for a group \(H\) containing \(\gamma\), we get \[J_{o}(f)=\operatorname{vol}(G(\mathbb{Q})_{\gamma}\backslash G(\mathbb{A})_{ \gamma})\int_{G(\mathbb{A})_{\gamma}\backslash G(\mathbb{A})}f(x^{-1}\gamma x )dx.\] **Theorem 2.1**.: _Assuming \(G(\mathbb{Q})\backslash G(\mathbb{A})\) is compact, we have_ \[\operatorname{tr}R(f)=\sum_{o\in O}J_{o}(f)=\sum_{\chi\in\mathcal{X}}J_{\chi}(f) \tag{3}\] _for any \(f\in C_{c}^{\infty}(G(\mathbb{A}))\)._ For the classical setting, we start with a real Lie group \(G\) and a lattice \(\Gamma\subset G\). Consider the representation \(R_{\Gamma}\) of \(G\) on \(L^{2}(\Gamma\backslash G)\) given by \((R_{\Gamma}(g)\phi)(x)=\phi(xg)\) for \(x,g\in G\). Let \(C_{c}^{\infty}(G)\) be the space of smooth function on \(G\) with compact support. For \(f\in C_{c}^{\infty}(G)\) and a representation \((\pi,H)\) of \(G\), we let \[\pi(f)v=\int_{G}f(g)\pi(g)vdg.\] If \(\pi\) is irreducible, \(\pi(f)\) is a trace class operator and we let \(\theta_{\pi}(f)=\operatorname{tr}\pi(f)\). Note for the representation \(R_{\Gamma}\), we have \((R_{\Gamma}(f)\phi)(x)=\int_{G}f(g)R_{\Gamma}(g)\phi(x)dg=\int_{G}f(g)\phi(xg)dg\). It is known that \(\Gamma\backslash G\) is compact if and only if the reductive part of \(\mathbf{G}\) is anisotropic (see [32] Theorem 4.12). In this case, \(L^{2}(\Gamma\backslash G)\) can be decomposed into a direct sum of irreducible representations of \(G\) with each of finite multiplicity, i.e., \[L^{2}(\Gamma\backslash G)=\oplus m_{\Gamma}(\pi)\cdot\pi\] with \(m_{\Gamma}(\pi)=\dim\operatorname{Hom}_{G}(\pi,L^{2}(\Gamma\backslash G))<\infty\) for each \(\pi\). By taking the test function in Theorem 2.1 to be \(f\otimes 1_{K}\) for a maximal compact subgroup \(K\) of \(G(\mathbb{A}^{\operatorname{fin}})\) with \(f\in C_{c}^{\infty}(G)\) (see Section 3.1), we get the following result for the lattice \(\Gamma\) in the real Lie group \(G\). **Corollary 2.2** (The Selberg trace formula).: _If \(\Gamma\backslash G\) is compact, \(R_{\Gamma}(f)\) is of trace class and_ \[trR_{\Gamma}(f)=\sum_{\pi\in\widehat{G}}m_{\Gamma}(\pi)\theta_{\pi}(f)=\sum_{ \gamma\in[\Gamma]}\operatorname{vol}(\Gamma_{\gamma}\backslash G_{\gamma}) \int_{\Gamma_{\gamma}\backslash G}f(x^{-1}\gamma x)dx \tag{4}\] ### The Arthur trace formula We assume \(\mathbf{G}\) is not necessarily anisotropic and \(G(\mathbb{Q})\backslash G(\mathbb{A})\) may not be compact. Assume \(B\) is a Borel subgroup defined over \(\mathbb{Q}\), \(M_{0}\) is a Levi factor of \(B\) defined over \(\mathbb{Q}\), \(P\) is a standard parabolic subgroup defined over \(\mathbb{Q}\) (i.e., \(P_{0}=B\subset P\)), \(N_{P}=R_{u}(P)\) (the unipotent radical of \(P\)), \(M_{P}\) is the unique Levi component of \(P\) such that \(M_{0}\subset M_{P}\). We also assume \(A_{P}\) is the split component of the center of \(M_{P}\) and \(Z=A_{G}\), \(\Delta_{0}=\Delta_{B}\) is a base for a root system. We will mostly use the notations of [3, 4, 6] and [21] as follows: * \(\mathfrak{a}_{P}=\operatorname{Hom}(X(M_{P})_{\mathbb{Q}},\mathbb{R})\) where \(X(M_{P})_{\mathbb{Q}}\) is the \(\mathbb{Q}\)-characters of \(M_{P}\), \(\mathfrak{a}_{P}^{*}=X(M_{P})_{\mathbb{Q}}\otimes\mathbb{R}\) and \(\mathfrak{a}_{P}^{+}=\{H\in\mathfrak{a}_{P}|\alpha(H)>0,\forall\alpha\in \Delta_{P}\}\). * \(\gamma=\gamma_{s}\gamma_{u}\), which is the decomposition such that \(\gamma_{s}\) is semisimple and \(\gamma_{u}\) is unipotent. * \(\mathcal{O}\) is the set of \(G(\mathbb{Q})\)-semisimple conjugacy class of \(G(\mathbb{Q})\) (\(\gamma\cong\beta\) if \(\gamma_{s}\) and \(\beta_{s}\) are \(G(\mathbb{Q})\)-conjugate). * \(o\in\mathcal{O}\) is a conjugacy class in \(G(\mathbb{Q})\). * \(\mathcal{X}\) is the set of equivalence classes of pairs \((M,\rho)\), where \(M\) is a Levi subgroup of \(G\) and \(\rho\in\widehat{M(\mathbb{A})}^{1}\) (\((M,\rho)\sim(M^{\prime},\rho^{\prime})\) if there is an \(s\in\Omega(\mathfrak{a},\mathfrak{a}^{\prime})\) such that the representation \((s\rho)(m^{\prime})=\rho(w_{s}^{-1}mw_{s})\) is unitarily equivalent to \(\rho^{\prime}\)). * For a pair of parabolic groups \(P_{1}\subset P_{2}\), \(\Delta_{P_{1}}^{P_{2}}\) is the set of simple roots of \((M_{P_{2}}\cap P_{1},A_{P_{1}})\) and \(\hat{\Delta}_{P_{1}}^{P_{2}}=\{\varpi_{\alpha}|\alpha\in\Delta_{P_{1}}^{P_{2}}\}\), i.e, the dual basis for \(\Delta_{P_{1}}^{P_{2}}\). * \(\hat{\tau}_{P}\) is the characteristic function on \(\mathfrak{a}_{0}\) of \(\{H\in\mathfrak{a}_{0}|\varpi(H)>0,\varpi\in\hat{\Delta}_{P}^{G}\}\). * For \(m=\prod_{v}m_{v}\in M(\mathbb{A})\), let \(H_{M}(m)\in\mathfrak{a}_{p}\) given by \[e^{\langle H_{M}(m),\chi\rangle}=|\chi(m)|=\prod_{v}|\chi(m_{v})|_{v},\,\forall \chi\in X(M)_{\mathbb{Q}}.\] * \(x=nmak\in G(\mathbb{A})\) with \(n\in G(\mathbb{A}),m\in M(\mathbb{A})^{1},a\in A(\mathbb{R})^{0}\) and \(k\in K\). * \(H(x)=H_{M}(ma)=H_{M}(a)\in\mathfrak{a}_{p}\). Let \(T\in\mathfrak{a}_{0}^{+}\) be suitably regular, i.e., \(\alpha(T)\) is sufficiently large for all \(\alpha\in\Delta_{0}\). For a parabolic subgroup \(P\), there are kernels \(K_{P,o}=\sum\limits_{\gamma\in M(\mathbb{Q})\cap o}\int_{N(\mathbb{A})}f(x^{ -1}\gamma ny)dn\) and \(K_{P,\chi}\) (see [3] p.923 and p.935 for the precise definitions). Then Arthur is able to define the _truncated kernels_ and distributions \(J_{o}^{T},J_{\chi}^{T}\) as follows: 1. \(k_{o}^{T}(x,f)=\sum\limits_{P}(-1)\dim(A_{P}/Z)\sum\limits_{\delta\in P( \mathbb{Q})\backslash G(\mathbb{Q})}K_{P,o}(\delta x,\delta x)\cdot\hat{\tau}_{p }(H(\delta x)-T)\). 2. \(k_{\chi}^{T}(x,f)=\sum\limits_{P}(-1)\dim(A_{P}/Z)\sum\limits_{\delta\in P( \mathbb{Q})\backslash G(\mathbb{Q})}K_{P,\chi}(\delta x,\delta x)\cdot\hat{\tau}_{ p}(H(\delta x)-T)\). 3. \(J_{o}^{T}(f)=\int_{G(\mathbb{Q})\backslash G(\mathbb{A})^{1}}k_{o}^{T}(x,f)dx\). 4. \(J_{\chi}^{T}(f)=\int_{G(\mathbb{Q})\backslash G(\mathbb{A})^{1}}k_{\chi}^{T}(x,f)dx\). Let \(\mathcal{X}(G)=\{(M,\rho)\in\mathcal{X}|M=G\}\). We reach a coarse trace formula, which is firstly given in [4] Chapter 5. **Theorem 2.3** (The Arthur trace formula).: _For any \(f\in C_{c}^{\infty}(G(\mathbb{A})^{1})\) and any suitably regular \(T\in\mathfrak{a}_{0}^{+}\), we have_ \[\sum_{o\in\mathcal{O}}J_{o}^{T}(f)=\sum_{\chi\in\mathcal{X}}J_{\chi}^{T}(f) \tag{5}\] _Moreover, the trace formula of \(R(f)\) is given by_ \[\operatorname{tr}R_{\text{cusp}}(f)=\sum_{o\in\mathcal{O}}J_{o}^{T}(f)-\sum_{ \chi\in\mathcal{X}\backslash\mathcal{X}(G)}J_{\chi}^{T}(f).\] ### The dominant term on the geometric side Consider the adelic case at first. Let \(F\) be a number field and \(V,V_{\infty}\) and \(V_{f}\) be the set of places, Archimedean and non-Archimedean places of \(F\) respectively. Let \(\mathbb{A}\) be adele ring of \(F\) and \(A_{\text{fin}}\subset\mathbb{A}\) be restricted product over the finite places. Suppose \(S\subset V\) is a finite set containing \(V_{\infty}\). Let \(F_{S}=\prod_{v\in S}F_{s}\) and \(\mathbb{A}^{S}=\prod_{v\in V\backslash S}^{\prime}F_{s}\) so that \(\mathbb{A}=F_{S}\times\mathbb{A}^{S}\). We define 1. \(G(F_{S})^{1}=\bigcap_{\chi\in\text{Hom}(G(F_{S}),F^{\times})}\{\ker|\chi| \colon G(F_{S})\to\mathbb{R}_{+}\}\), 2. \(G(\mathbb{A})^{1}=\bigcap_{\chi\in\text{Hom}(G(\mathbb{A}),F^{\times})}\{\ker| \chi|\colon G(\mathbb{A})\to\mathbb{R}_{+}\}\), where \(|\cdot|\) is the product of valuations on \(F_{S}\) and \(\mathbb{A}\) respectively. We will consider the representation of \(G(F_{S})\) on \(L^{2}(G(F)\backslash G(\mathbb{A})^{1}/K)\) for an open compact subgroup \(K\) of \(G(\mathbb{A}^{S})\). In particular, it will reduce to the representation of \(G(F_{\infty})\) on \(L^{2}(\Gamma_{K}\backslash F_{\infty})\) if we take \(S=\{\infty\}\) and \(\Gamma_{K}=G(F)\cap K\). Let \(J(f)\) be the distribution defined by Equation 3 or 5 in Section 2 for \(f\in C_{c}^{\infty}(G(F_{S}))\), which also depends on \(G(F)\backslash G(\mathbb{A})^{1}\) is compact or not. The goal of this subsection is to prove \[\lim_{n\to\infty}\frac{\operatorname{vol}(G(F)\backslash G(\mathbb{A})^{1})f( 1)}{J(f\otimes 1_{K_{n}})}=1\] for certain towers of open compact subgroups \(\{K_{n}\}_{n\geq 1}\) of \(G(\mathbb{A}^{S})\). Let us assume \(\Gamma\) is a uniform lattice in the semisimple Lie group \(G\). We add a subscript such as \(R_{\Gamma}\) and \(J_{\Gamma}\) for the representation of \(G\) on \(L^{2}(\Gamma\backslash G)\) and the corresponding trace formulas as an emphasis on the lattice \(\Gamma\). Since \(\Gamma\backslash G\) is compact, \(J_{\Gamma}(f)\) is the trace \(\operatorname{tr}R_{\Gamma}(f)\) and we obtain \[J_{\Gamma}(f)=\operatorname{tr}R_{\Gamma}(f)=\sum_{\pi\in\bar{G}}J_{\pi, \Gamma}(f)=\sum_{o\in\mathcal{O}}J_{o,\Gamma}(f).\] Let \(J_{\{1\},\Gamma}(f)=\operatorname{vol}(\Gamma\backslash G)f(1)\), the contribution of the identity to the geometric side of the trace formula. We take a tower of uniform lattices \(\{\Gamma_{n}\}_{n\geq 1}\), such that \(\Gamma_{n}\trianglelefteq\Gamma_{1}\), \([\Gamma_{1}:\Gamma_{n}]<\infty\) and \(\cap_{n\geq 1}\Gamma_{n}=\{1\}\). **Proposition 2.4**.: _With the assumption of uniform lattice \(\{\Gamma_{n}\}\) above, we have_ \[\lim_{n\to\infty}\frac{J_{\{1\},\Gamma_{n}}(f)}{J_{\Gamma_{n}}(f)}=1.\] **Proof**: Following the Equation (2) in [10], we obtain \[\operatorname{tr}R_{\Gamma_{n}}(\phi)=J_{\{1\},\Gamma_{n}}(\phi)+\sum_{\gamma \neq 1}s_{n}(\gamma)\operatorname{vol}(\Gamma_{j}\backslash G)\operatorname{vol }(\Gamma_{\gamma}\backslash G_{\gamma})\int_{\Gamma_{\gamma}\backslash G} \phi(x^{-1}\gamma x)dx,\] where \(0\leq s_{n}(\gamma)\leq\operatorname{vol}(\Gamma\backslash G)^{-1}\). As \(\cap_{n\geq 1}\Gamma_{n}=\{1\}\), \(\lim_{n\to\infty}s_{n}(\gamma)\) for all \(\gamma\neq 1\). By [10] Theorem 2, we have \(\operatorname{vol}(\Gamma_{n}\backslash G)^{-1}\cdot\lim_{n\to\infty} \operatorname{tr}R_{\Gamma_{n}}(\phi)=\phi(1)\). Hence \(\lim_{n\to\infty}\frac{J_{\{1\},\Gamma_{n}}(\phi)}{J_{\Gamma_{n}}(\phi)}=\lim_ {n\to\infty}\frac{J_{\{1\},\Gamma_{n}}(\phi)}{\operatorname{tr}R_{\Gamma_{n}}( \phi)}=1\). Now we let \(G\) be a reductive group over a number field \(F\). Let \(K=K_{\infty}K_{\rm fin}\) be a maximal compact subgroup of \(G(\mathbb{A})=G(\mathbb{A}_{F})\). By fixing a faithful \(F\)-rational representation \(\rho\colon G(F)\to{\rm GL}(m,F)\) for some \(m>0\), we let \(\Lambda\subset F^{m}\) be an \(\mathcal{O}_{F}\)-lattice such that the stablilizer of \(\widehat{\Lambda}=\mathcal{O}_{F}\otimes_{F}\Lambda\) in \(G(A_{\rm fin})\) is \(K_{\rm fin}\). For a non-trivial ideal \(I\) of \(\mathcal{O}_{F}\), we let \[K(I)=\{g\in G(A_{\rm fin})|\rho(g)v\equiv v\pmod{I\cdot\widehat{\Lambda}},v \in\widehat{\Lambda}\}\] be the _principal congruence subgroup_ of level \(I\). We also denote the ideal norm of \(I\) by \(N(I)=[\mathcal{O}_{F}\colon I]\). Consider a descending tower of ideals \(I_{1}\supsetneq I_{2}\supsetneq I_{3}\supsetneq\cdots\) such that each \(I_{k}\) is prime to (the prime ideals in) \(S\). We obtain the corresponding tower of principal congruence subgroups: \[K_{1}\supsetneq K_{2}\supsetneq K_{3}\supsetneq\cdots,\] where \(K_{n}=K(I_{n})\). By factoring into prime ideals, the family \(\{I_{n}\}_{n\geq 1}\) satisfies either one of the following properties: 1. there exists a prime ideal \(\mathfrak{p}\) such that each \(\mathfrak{p}^{k}\) is eventually contained in the tower, i.e., for any \(k\geq 1\), there is \(N_{k}>0\) such that \(\mathfrak{p}^{k}\subset I_{n}\) for all \(n\geq n_{k}\), or, 2. there exists infinitely many prime ideals \(\{\mathfrak{p}_{k}\}_{k\geq 1}\) such that for each \(k\), there exist \(M_{k}>0\) such that \(\mathfrak{p}_{k}\subset I_{n}\) for all \(n\geq M_{k}\). In either of these two cases, we have **Lemma 2.5**.: \(\cap_{n\geq 1}I_{n}=\{0\}\) _and \(\cap_{n\geq 1}K_{n}=\{1\}\)._ Recall the equivalence class of unipotent elements in \(G(F)\), which is the element \(\gamma=\gamma_{s}\gamma_{u}\) with the semisimple component \(r_{s}=1\) (see [5] p.1240). Let \[J_{\rm unip}^{T}(f),\,f\in C_{c}^{\infty}(G(\mathbb{A})^{1}).\] be the contribution of this equivalence class on the geometric side of the trace formula 5. We will consider the function of the form \(f=h_{S}\otimes 1_{K_{n}}\) with \(h_{S}\in C_{c}^{\infty}(G(F_{S})^{1})\). **Lemma 2.6**.: _For \(h_{S}\in C_{c}^{\infty}(G(F_{S})^{1})\), \(\lim_{n\to\infty}J(h_{S}\otimes 1_{K_{n}})=\lim_{n\to\infty}J_{unip}(h_{S} \otimes 1_{K_{n}})\)._ **Proof**: Let \(D_{h}={\rm supp}(h_{S})\subset G(F_{S})^{1}\) be the compact support of \(h_{S}\). Then \({\rm supp}(h_{S}\otimes 1_{K_{n}}))=D_{h}K_{n}\) is compact and hence it intersects finitely many semisimple-conjugate class \(o\in\mathcal{O}\). Consider the trace formula and Equation 5, only the classes \(o\)'s (and its \(G(\mathbb{A})\)-conjugations) which intersect infinitely many \(D_{h}K_{n}\) contributes a non-trivial \(J_{o}(h_{S}\otimes 1_{K_{n}})\) to the limit \(\lim_{n\to\infty}J(h_{S}\otimes 1_{K_{n}})\). Suppose the \(G(\mathbb{A})\) conjugacy classes of elements in \(o\) intersects \(D_{h}K_{n}\) for infinitely many \(n\), i.e., \(\{g\gamma g^{-1}|g\in G(\mathbb{A}),\gamma\in o\}\cap D_{h}K_{n}\neq\emptyset\) for infinitely many \(n\). Take some \(\gamma\in o\). By fixing a faithful \(F\)-representation \(\rho\colon G(F)\to{\rm GL}(m)\), we let \(p(x)\in F[x]\) be the characteristic polynomial of \(\rho(\gamma)-1\) (a \(m\)-by-\(m\) matrix over \(F\)). Suppose \(p(x)=x^{m}+a_{m-1}x^{m-1}+\cdots+a_{0}\) with all \(a_{i}\in F\). By Lemma 2.5, we know \(a_{i}\) belongs to infinitely many \(I_{n}\), or, equivalently \(a_{i}=0\). Hence \(p(x)=x^{m}\) and \(\gamma\) is unipotent. The unipotent contribution \(J_{\rm unip}(h_{S}\otimes 1_{K_{n}})\) can be further reduced to the the one from the identity as follows. We let \(I_{S}\) be a product of prime ideals in at the places of \(S\) and \(K_{S-S_{\infty}}(I_{S})\) be the \(S-S_{\infty}\) component of the compact group \(K(I_{S})\). We also let \(C_{\Omega}^{\infty}(G(F_{S})^{1})\) be the set of smooth functions with compact support contained in a compact subset \(\Omega\) of \(G(F_{S})^{1}\). For each \(k\geq 0\), we let \(\mathcal{B}_{k}\) be the \(k\)-th component of the universal enveloping algebra \(\mathcal{U}(\mathfrak{g}_{\mathbb{C}})\), where \(\mathfrak{g}_{\mathbb{C}}\) is the complexified Lie algebra of the Lie group \(G(F_{\infty})\). We set \(\|h\|_{k}=\sum_{X\in\mathcal{B}_{k}}\|X\circ h\|_{L^{1}(G(\mathbb{A})^{1}}\) for \(h\in C_{\Omega}^{\infty}(G(F_{S})^{1})\). The following result is a special case of Proposition 3.1 in [20], whose proof is mainly based on Theorem 3.1 and Theorem 4.2 in [5]. **Proposition 2.7** (Finis-Lapid-Muller).: _There exists an integer \(k\geq 0\) such that for any compact subset \(\Omega\) of \(G(F_{S})^{1}\), we have a constant \(C_{\Omega}>0\) and_ \[|J_{\text{unip}}(h_{S}\otimes 1_{K_{n}})-\operatorname{vol}(G(F)\backslash G( \mathbb{A})^{1})h_{S}(1)|\leq C_{\Omega}\tfrac{(1+\log N(I_{S}I))^{d_{0}}}{N(I )}\|h_{S}\|_{k}\] _for any bi-\(K_{S-S_{\infty}}(I_{S})\)-invariant function \(h_{S}\in C_{\Omega}^{\infty}(G(F_{S})^{1})\)._ Then, combining Lemma 2.6 we can obtain: **Corollary 2.8**.: _For \(h_{S}\in C_{c}^{\infty}(G(F_{S})^{1})\), we have_ \[\lim_{n\to\infty}\frac{\operatorname{vol}(G(F)\backslash G(\mathbb{A})^{1})h_ {S}(1)}{J(h_{S}\otimes 1_{K^{S}(n)})}=1.\] ## 3. The multiplicities problem This section is devoted to the multiplicity of bounded subsets of the unitary dual, instead of a single irreducible representation. ### The multiplicities in \(L^{2}(\Gamma\backslash G)\) Let \(G=\mathbf{G}(\mathbb{R})^{0}\), the connected component of the real group obtained from an almost simple group \(\mathbf{G}\) over \(\mathbb{Q}\). By fixing a faithful \(\mathbb{Q}\)-embedding \(\rho:\mathbf{G}\to GL_{n}\), we have an arithmetic group \(\Gamma\) commensurable with \(G\cap\operatorname{GL}_{n}(\mathbb{Z})\). Let \(\widehat{G}\) be the unitary dual of \(G\) and \(\widehat{G}_{\text{temp}}\subset\widehat{G}\) be the tempered dual. Let us consider the following two cases. 1. \(\Gamma\backslash G\)_is compact_. As introduced in Section 2.1, \(L^{2}(\Gamma\backslash G)\) can be decomposed into a direct sum of irreducible representations of \(G\) with each of finite multiplicity, i.e., \[L^{2}(\Gamma\backslash G)=\oplus m_{\Gamma}(\pi)\cdot\pi\] with \(m_{\Gamma}(\pi)\colon=\dim\operatorname{Hom}_{G}(\pi,L^{2}(\Gamma\backslash G ))<\infty\) for each \(\pi\in\widehat{G}\). 2. \(\Gamma\backslash G\)_is not compact_. If \(G\) is semisimple, we have \(\Gamma\backslash G\) is of finite (Haar) measure (see [31] Theorem 4.13). The regular representation has both discrete and continuous spectra: \(L^{2}(\Gamma\backslash G)=L^{2}_{\text{disc}}(\Gamma\backslash G)\oplus L^{2}_ {\text{disc}}(\Gamma\backslash G)\). The discrete spectrum can be written as the direct sum of cuspidal and residue subspaces: \(L^{2}_{\text{disc}}(\Gamma\backslash G)=L^{2}_{\text{cusp}}(\Gamma\backslash G )\oplus L^{2}_{\text{res}}(\Gamma\backslash G)\). which can be decomposed further into a direct sum of irreducible representations with finite multiplicities, i.e., \[L^{2}_{\text{disc}}(\Gamma\backslash G)=\oplus m_{\Gamma}(\pi)\cdot\pi\] with \(m_{\Gamma}(\pi)\colon=\dim\operatorname{Hom}_{G}(\pi,L^{2}_{\text{disc}}( \Gamma\backslash G))=\dim\operatorname{Hom}_{G}(\pi,L^{2}(\Gamma\backslash G))\) is finite for each \(\pi\in\widehat{G}\). We say \(X\subset\widehat{G}\) is _bounded_ if it is relatively compact under the Fell topology. **Definition 3.1** (The multiplicity for \(X\subset\widehat{G}\)).: For a bounded \(X\subset\widehat{G}\), we define the _multiplicity of \(X\)_ to be the sum of the multiplicities of the irreducible representations in \(X\), i.e., \[m_{\Gamma}(X)\colon=\sum_{\pi\in X}m_{\Gamma}(\pi).\] Borel and Garland proved the finiteness of \(m_{\Gamma}(X)\) by considering the spectrum of a certain Laplacian (see [9] Theorem 3, Theorem 4.6 and also [24] Theorem 1.1.3). **Theorem 3.1** (Borel-Garland).: _Let \(G={\bf G}(\mathbb{R})^{0}\) for a connected semisimple group \({\bf G}\) over \(\mathbb{Q}\) and \(X\subset\widehat{G}\) is bounded. We have \(m_{\Gamma}(X)<\infty\)._ For a subset \(X\subset\widehat{G(F_{S})}^{1}\), we call it _bounded_ if it is relatively compact under the Fell topology (see [35]). **Definition 3.2** (The multiplicity for \(\widehat{G(F_{S})}^{1}\)).: Suppose \(K\) is a compact open subgroup of \(G(\mathbb{A}^{S})\). Let \(\sigma\) be an irreducible representation of \(G(F_{S})^{1}\) and \(X\subset\widehat{G(F_{S})}^{1}\) be a bounded subset. 1. The _multiplicity of \(\sigma\) with respect to \(K\)_ is defined as \[m_{K}(\sigma)\colon=\dim\operatorname{Hom}_{G(F_{S})^{1}}(\sigma,L^{2}(G( \mathbb{Q})\backslash G(\mathbb{A})^{1}/K)).\] 2. The _multiplicity of \(X\) with respect to \(K\)_ is defined as \[m_{K}(X)\colon=\sum_{\sigma\in X}m_{K}(\sigma).\] For an irreducible representation \(\pi\) of \(G(\mathbb{A})^{1}\), we write \(\pi=\pi_{S}\otimes\pi^{S}\), where \(\pi_{S}\) and \(\pi^{S}\) denote the components of the representations of \(G(F_{S})^{1}\) and \(G(\mathbb{A}^{S})\) respectively. As shown in Theorem 3.1, \(m_{K}(X)\) is finite and hence well-defined. If we treat \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A})^{1}/K)\) as the subspace of \(K\)-right invariant functions in \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A})^{1}))\), we have \[m_{K}(\sigma)=\sum_{\pi\in\widehat{G(\mathbb{A})}^{1},\pi_{S}=\sigma}\dim \operatorname{Hom}_{G(\mathbb{A})^{1}}(\pi,L^{2}(G(\mathbb{Q})\backslash G( \mathbb{A})^{1}))\dim(\pi^{S})^{K}.\] If we take \(S=V_{\infty}\) and \({\bf G}\) is semisimple, simply connected, and without any \(F\)-simple factors \(H\) such that \(H(F_{\infty})\) is compact and \(K\) is an open compact subgroup of \(G(\mathbb{A}_{\text{fin}})\), we know \(\Gamma_{K}=G(\mathbb{F})\cap K\) is a lattice in the seimisimple Lie group \(G(F_{\infty})\). **Lemma 3.2**.: _With the assumption above, we have \(m_{\Gamma_{K}}(\pi)=m_{K}(\pi)\) for any \(\pi\in\widehat{G(F_{\infty})}^{1}\) and \(m_{\Gamma_{K}}(X)=m_{K}(X)\) for any bounded \(X\subset\widehat{G(F_{\infty})}^{1}\)_ **Proof**: It follows the fact \(G(\mathbb{Q})\backslash G(\mathbb{A})/K\) can be identified with \(\Gamma_{K}\backslash G(F_{\infty})\), which leads to a \(G(F_{\infty})\)-isomorphism \(L^{2}(\Gamma_{K}\backslash G(F_{\infty}))\cong L^{2}(G(\mathbb{Q})\backslash G (\mathbb{A})^{1}/K)\) (see [27] Chapter 6 and [32] Chapter 7.4.). For a finite set \(S\) and a function \(\phi\) on \(\widehat{G(F_{S})}^{1}\), we define \[m_{K}(\phi)\colon=\int_{\widehat{G(F_{S})}^{1}}\phi(\pi)dm_{K}(\pi)\] as its integral with respect to the measure given by multiplicities above. If \(1_{X}\) is the characteristic function of \(X\), i.e., \(1_{X}(\pi)=1\) if \(\pi\in X\) and \(0\) otherwise, \(m_{K}(1_{X})=m_{K}(X)\). For \(f\in C_{\text{c}}^{\infty}(G(F_{S})^{1})\), we let \(\widehat{f}(\pi)=\operatorname{tr}\pi(f)\), the distribution character of \(\pi\). Let \(R_{\text{disc}}\) denote the action of \(G(\mathbb{A})\) on the discrete subspace \(L^{2}(G(\mathbb{Q})\backslash G(\mathbb{A})^{1})\). **Proposition 3.3**.: _For \(f\in C_{\text{c}}^{\infty}(G(F_{S})^{1})\), we have_ \[\operatorname{tr}R_{disc}(f\otimes\tfrac{1_{K}}{\operatorname{vol}(K)})=m_{K}( \hat{f}).\] **Proof**: Observe for the component \(\pi^{S}\) of representation of \(G(\mathbb{A}^{S})\), we have \[\operatorname{tr}\pi^{S}(1_{K}) =\int_{G(\mathbb{A}^{S})}1_{K}(x)\pi^{S}(x^{-1})d\mu^{S}(x)\] \[=\int_{K}\pi^{S}(x^{-1})d\mu^{S}(x)=\operatorname{vol}(K)\dim(\pi^ {S})^{K},\] where we apply the fact that \(\int_{K}\sigma(x)d\mu^{S}(x)=0\) for any non-trivial irreducible representation \(\sigma\) of \(K\). Hence we obtain \[\operatorname{tr}R_{\operatorname{disc}}(f\otimes\frac{1_{K}}{ \operatorname{vol}(K)}) =\frac{1}{\operatorname{vol}(K)}\sum_{\pi\in\widehat{G(\mathbb{A}) }^{1}}m(\pi)\operatorname{tr}\pi(f\otimes 1_{K})\] \[=\frac{1}{\operatorname{vol}(K)}\sum_{\pi\in\widehat{G(\mathbb{A} )}^{1}}m(\pi)\operatorname{tr}\pi_{S}(f)\operatorname{tr}\pi^{S}(1_{K})\] \[=\frac{1}{\operatorname{vol}(K)}\sum_{\pi\in\widehat{G(\mathbb{A} )}^{1}}m(\pi)\operatorname{tr}\pi_{S}(f)\operatorname{vol}(K)\dim(\pi^{S})^{K}\] \[=\sum_{\sigma\in\widehat{G(F_{S})}^{1}}m_{K}(\sigma) \operatorname{tr}\sigma(f)=m_{K}(\widehat{f}).\] We also give the following result which connects the trace formulas for adelic groups and Lie groups. **Corollary 3.4**.: _Let \(\Gamma_{K}=G(F)\cap K\) with an open compact subgroup \(K\) of \(G(\mathbb{A}_{\operatorname{fin}})\). We have_ \[\operatorname{tr}R_{\operatorname{disc}}(f\otimes\tfrac{1_{K}}{\operatorname {vol}(K)})=\operatorname{tr}R_{\Gamma_{K}}(f).\] _for all \(f\in C_{c}^{\infty}(G(F_{\infty})^{1})\)._ **Proof**: It follows the fact \(m_{K}(\widehat{f})=m_{\Gamma_{K}}(\widehat{f})\) in Lemma 3.2, \(m_{\Gamma_{K}}(\widehat{f})=\operatorname{tr}R_{\Gamma_{K}}(f)\) and Proposition 3.3. ### Sauvageot's density theorems We have a brief review of the results in [35]. See also [37] for an alternative approach and corrections. For an open compact subgroup \(K\) of \(G(\mathbb{A}^{S})\), we define a measure on \(\widehat{G(F_{S})}^{1}\) by \[\nu_{K}(X)\colon=\tfrac{\operatorname{vol}(K)}{\operatorname{vol}(G(\mathbb{Q} )\setminus\widehat{G(\mathbb{A})}^{1})}m_{K}(X)\] for any bounded subset \(X\) of \(\widehat{G(F_{S})}^{1}\) and \(m_{K}\) is the multipilicity defined in Chapter 3.1. Let \(K_{1}\supsetneq K_{2}\supsetneq\cdots\) be a sequence of open compact subgroups of \(G(\mathbb{A}^{S})\). Given a bounded subset \(X\) of \(\widehat{G(F_{S})}^{1}\) and \(C\geq 0\), we write \[\lim_{n\to\infty}\nu_{K}(X)=C,\] if for any \(\varepsilon>0\), there exists \(N=N(\varepsilon)>0\) such that \(|\nu_{K_{n}}(X)-C|<\varepsilon\) for all \(n\geq N\). Let \(\mathcal{H}(G(F_{S})^{1})\) be the complex algebra of smooth, compactly-supported, bi-\(K_{S}\)-finite functions on \(G(F_{S})^{1}\). **Lemma 3.5** ([35] Corollaire 6.2).: _For \(\varepsilon>0\) and any bounded \(X\subset\widehat{G(F_{S})}^{1}\setminus\widehat{G(F_{S})}^{1}{}_{temp}\), there is \(\Psi\in\mathcal{H}(G(F_{S})^{1})\) such that_ \[\widehat{\Psi}|_{\widehat{G(F_{S})}^{1}}\geq 0\text{, }\nu(\widehat{\Psi})<\varepsilon \text{ and }\widehat{\Psi}|_{X}\geq 1.\] Given a function \(f\) defined on \(\widehat{G(F_{S})}^{1}{}_{temp}\), we also denote by \(f\) the function on \(\widehat{G(F_{S})}^{1}\), which is extended by \(0\) on the untempered part. **Lemma 3.6** ([35] Theoreme 7.3(b)).: _For \(\varepsilon>0\) and any \(\nu\)-integrable function \(f\) on \(\widehat{G(F_{S})^{1}}_{\text{temp}}\), there exists \(\phi,\psi\in\mathcal{H}(G(F_{S})^{1})\) such that_ \[|f(\pi)-\widehat{\phi}(\pi)|\leq\widehat{\psi}(\pi)\text{ and }\nu(\widehat{\psi})<\varepsilon.\] Here we obtain one of the main results in [35] and we also provide a proof for completeness. **Theorem 3.7** (Sauvageot).: _Suppose \(\lim_{n\to\infty}\nu_{K_{n}}(\widehat{\phi})=\phi(1)\) for all \(\phi\in\mathcal{H}(G(F_{S})^{1})\). We have_ \[\lim_{n\to\infty}\nu_{K_{n}}(X)=\nu(X)\] _for all bounded subset \(X\) of \(\widehat{G(F_{S})^{1}}\)._ **Proof**: First, we show the contribution from the untempered part is negligible in the limit. For a bounded subset \(X_{0}\) of \(\widehat{G(F_{S})^{1}}_{\text{temp}}\) and \(\varepsilon>0\), we let \(\Psi\in\mathcal{H}(G(F_{S})^{1})\) satisfies Lemma 3.5 with respect to \(X\). We have \(\nu_{K_{n}}(X)\leq\nu_{K_{n}}(\widehat{\Psi})\leq|\nu_{K_{n}}(\widehat{\Psi})- \psi(1)|+\psi(1)<2\varepsilon\) for all \(n\geq N_{1}\) with some \(N_{1}\geq 0\). For the tempered part, we fix a bounded subset \(X_{1}\) of \(\widehat{G(F_{S})^{1}}_{\text{temp}}\) with the same \(\varepsilon\) above. Let \(\phi,\psi\in\mathcal{H}(G(F_{S})^{1})\) satisfy Lemma 3.6 with respect to the function \(f=1_{X_{1}}\) on \(\widehat{G(F_{S})^{1}}_{\text{temp}}\) and \(\varepsilon\). By assumption, we have \(|\nu_{K_{n}}(\widehat{\phi})-\phi(1)|<\varepsilon\) and \(|\nu_{K_{n}}(\widehat{\psi})-\psi(1)|<\varepsilon\) for all \(n\geq N_{2}\) with some \(N_{2}\geq 0\). Hence, for \(n\geq N_{2}\), we obtain \[|\nu_{K_{n}}(X_{1})-\nu(X_{1})| \leq|\nu_{K_{n}}(X_{1})-\nu_{K_{n}}(\widehat{\phi})|+|\nu_{K_{n} }(\widehat{\phi})-\phi(1)|+|\phi(1)-\nu(X)|\] \[\leq|\nu_{K_{n}}(\widehat{\phi})-\phi(1)|+\nu_{K_{n}}(\widehat{ \psi})+\psi(1)\] \[\leq|\nu_{K_{n}}(\widehat{\phi})-\phi(1)|+|+|\nu_{K_{n}}(\widehat {\psi})-\psi(1)|+2\psi(1)<4\varepsilon.\] Hence, for the bounded set \(X\) of \(\widehat{G(F_{S})^{1}}\), let \(X=X_{0}\sqcup X_{1}\) be the decomposition into its untempered and tempered parts. We have \[|\nu_{K_{n}}(X)-\nu(X)| =|\nu_{K_{n}}(X)-\nu(X_{1})|=|\nu_{K_{n}}(X_{1})-\nu(X)|+\nu_{K_{ n}}(X_{0})\] \[\leq 4\varepsilon+2\varepsilon=6\varepsilon\] for all \(N\geq\max\{N_{1},N_{2}\}\). ## 4. The von Neumann dimensions of direct integrals ### The group von Neumann algebra and the trace Let \(\Gamma\) be a countable group with the counting measure. Let \(\{\delta_{\gamma}\}_{\gamma\in\Gamma}\) be the usual orthonormal basis of \(l^{2}(\Gamma)\). We also let \(\lambda\) and \(\rho\) be the left and right regular representations of \(\Gamma\) on \(l^{2}(\Gamma)\) respectively. For all \(\gamma,\gamma^{\prime}\in\Gamma\), we have \(\lambda(\gamma^{\prime})\delta_{\gamma}=\delta_{\gamma^{\prime}\gamma}\) and \(\rho(\gamma^{\prime})\delta_{\gamma}=\delta_{\gamma\gamma^{\prime-1}}\). Let \(\mathcal{L}(\Gamma)\) be the strong operator closure of the complex linear span of \(\lambda(\gamma)\)'s (or equivalently, \(\rho(\gamma)\)'s). This is the _group von Neumann algebra of \(\Gamma\)_. There is a canonical faithful normal tracial state \(\tau_{\Gamma}\), or simply \(\tau\), on \(\mathcal{L}(\Gamma)\), which is given by \[\tau(x)=\langle x\delta_{e},\delta_{e}\rangle_{l^{2}(\Gamma)},\,x\in\mathcal{L }(\Gamma).\] Hence \(\mathcal{L}(\Gamma)\) is a finite von Neumann algebra (which must be of type I or II\({}_{1}\)). More generally, for a tracial von Neumann algebra \(M\) with the trace \(\tau\), we consider the GNS representation of \(M\) on the Hilbert space constructed from the completion of \(M\) with respect to the inner product \(\langle x,y\rangle_{\tau}=\tau(xy^{*})\). The underlying space will be denoted by \(L^{2}(M,\tau)\), or simply \(L^{2}(M)\). Consider a normal unital representation \(\pi\colon M\to B(H)\) with both \(M\) and \(H\) separable. There exists an isometry \(u\colon H\to L^{2}(M)\otimes l^{2}(\mathbb{N})\), which commutes with the actions of \(M\): \[u\circ\pi(x)=(\lambda(x)\otimes\operatorname{id}_{l^{2}(\mathbb{N})})\circ u,\, \forall x\in M,\] where \(\lambda\colon M\mapsto L^{2}(M)\) denotes the left action. Then \(p=uu^{*}\) is a projection in \(B(L^{2}(M)\otimes l^{2}(\mathbb{N}))\) such that \(H\cong p(L^{2}(M)\otimes l^{2}(\mathbb{N}))\). We have the following result (see [2] Proposition 8.2.3). **Proposition 4.1**.: _The correspondence \(H\mapsto p\) above defines a bijection between the set of equivalence classes of left \(M\)-modules and the set of equivalence classes of projections in \((M^{\prime}\cap B(L^{2}(M)))\otimes B(l^{2}(\mathbb{N}))\)._ The _von Neumann dimension_ of the \(M\)-module \(H\) are defined to be \((\tau\otimes\operatorname{Tr})(p)\) and denoted by \(\dim_{M}(H)\), which takes its value in \([0,\infty]\). We have: 1. \(\dim_{M}(\oplus_{i}H_{i})=\sum_{i}\dim_{M}(H_{i})\). 2. \(\dim_{M}(L^{2}(M))=1\). Note \(\dim_{M}(H)\) depends on the trace \(\tau\). If \(M\) is a finite factor, i.e., \(Z(M)\cong\mathbb{C}\), there is a unique normal tracial state (see [25, 29]) and we further have: 1. \(\dim_{M}(H)=\dim_{M}(H^{\prime})\) if and only if \(H\) and \(H^{\prime}\) are isomorphic as \(M\)-modules (provided \(M\) is a factor). When \(M\) is not a factor, there is a \(Z(M)\)-valued trace which determines the isomorphism class of an \(M\)-module (see [8]). In the following sections, we will consider the group von Neumann algebra \(\mathcal{L}(\Gamma)\) with the canonical trace \(tr(x)=\langle x\delta_{e},\delta_{e}\rangle\). Hence the von Neumann dimension of \(\mathcal{L}(\Gamma)\) is the one uniquely determined by this trace. Note a discrete group \(\Gamma\) is called an infinite conjugacy class (ICC) group if every nontrivial conjugacy class \(C_{\gamma}=\{g\gamma g^{-1}|g\in\Gamma\}\), \(\gamma\neq e\), is infinite. It is well-known that \(\mathcal{L}(\Gamma)\) is a II\({}_{1}\) factor if and only if \(\Gamma\) is a nontrivial ICC group. Now we consider the case that \(\Gamma\) is a discrete subgroup of a locally compact unimodular type I group \(G\). Let \(\mu\) be a Haar measure of \(G\). A measurable set \(D\subset G\) is called a _fundamental domain_ for \(\Gamma\) if \(D\) satisfies \(\mu(G\setminus\cup_{\gamma\in\Gamma}\gamma D)=0\) and \(\mu(\gamma_{1}D\cap\gamma_{2}D)=0\) if \(\gamma_{1}\neq\gamma_{2}\) in \(\Gamma\). In this section, we always assume \(\Gamma\) is a lattice, i.e., \(\mu(D)<\infty\). The measure \(\mu(D)\) is called _covolume_ of \(\Gamma\) and will be denoted by \(\operatorname{covol}(\Gamma)\). Note the covolume depends on the Haar measure \(\mu\) (see Remark 4.3). There is a natural isomorphism \(L^{2}(G)\cong l^{2}(\Gamma)\otimes L^{2}(D,\mu)\) given by \[\phi\mapsto\sum_{\gamma\in\Gamma}\delta_{\gamma}\otimes\phi_{\gamma}\text{ with }\phi_{\gamma}(z)=\phi(\gamma\cdot z),\] where \(z\in D\) and \(\gamma\in\Gamma\). The restriction representation \(\lambda_{G}|_{\Gamma}\) of \(\Gamma\) is the tensor product of \(\lambda_{\Gamma}\) on \(l^{2}(\Gamma)\) and the identity operator \(\operatorname{id}\) on \(L^{2}(D,\mu)\). Hence we obtain the von Neumann algebra \(\lambda_{G}(\Gamma)^{\prime\prime}\cong\mathcal{L}(\Gamma)\otimes\mathbb{C}= \mathcal{L}(\Gamma)\), which will be denoted by \(M\) throughout this section. Please note \(L^{2}(M)=l^{2}(\Gamma)\). ### A theorem of von Neumann dimension Suppose \(X\) is a measurable subset of \(\widehat{G}\) with the Plancherel measure \(\nu(X)<\infty\). Define \[H_{X}=\int_{X}^{\oplus}H_{\pi}d\nu(\pi),\] which is the direct integral of the spaces \(H_{\pi}\) with \(\pi\in X\). It is a module over \(G\), its lattice \(\Gamma\), and also the group von Neumann algebra \(\mathcal{L}(\Gamma)\). We state a result on the von Neumann dimension of direct integrals. One may refer to [38] Section 4 for the proof. **Theorem 4.2**.: _Let \(G\) be a locally compact unimodular type I group with Haar measure \(\mu\). Let \(\nu\) be the Plancherel measure on the unitary dual \(\widehat{G}\) of \(G\). Suppose \(\Gamma\) is a lattice in \(G\) and \(\mathcal{L}(\Gamma)\) is the group von Neumann algebra of \(\Gamma\). Let \(X\subset\widehat{G}\) such that \(\nu(X)<\infty\) and \(H_{X}=\int_{X}^{\oplus}H_{\pi}d\nu(\pi)\). We have_ \[\dim_{\mathcal{L}(\Gamma)}(H_{X})=\operatorname{covol}(\Gamma)\cdot\nu(X).\] _Remark 4.3_.: 1. If \(\mu^{\prime}=k\cdot\mu\) is another Haar measure on \(G\) for some \(k>0\), the covolumes are related by \(\operatorname{covol}^{\prime}(\Gamma)=\mu^{\prime}(G/\Gamma)=k^{\prime}\cdot \mu(G/\Gamma)=k\cdot\operatorname{covol}(\Gamma)\). But the induced Plancherel measure \(\nu^{\prime}=k^{-1}\cdot\nu\) and the dependencies cancel out in the formula above. 2. There is a relevant approach by H. Peterson and A. Valette [31]. They study the von Neumann dimension over locally compact groups. The group von Neumann algebra is equipped with a semifinite tracial weight instead of a tracial state for a discrete group. It is motivated by the study of \(L^{2}\)-Betti number of locally compact groups [30]. If \(\pi\) is an atom in \(\widehat{G}\), i.e., \(\nu(\{\pi\})>0\), the irreducible representation \(\pi\) is a discrete series and \(\nu(\{\pi\})\) is just the formal dimension of \(\pi\)[15, 33]. Under this assumption, if \(G\) is a real Lie group that has discrete series and \(\Gamma\) is an ICC group, the theorem reduces to the special case of a single representation (see [23] Theorem 3.3.2) \[\dim_{\mathcal{L}(\Gamma)}(H_{\pi})=\operatorname{covol}(\Gamma)\cdot d_{\pi}.\] This is motivated by the geometric construction of discrete series of Lie groups by M. Atiyah and W. Schmid [7]. ### The proof of the main theorem We will prove the main theorem. We first give the proof for a tower of uniform lattices. **Theorem 4.4**.: _[a tower of uniform lattices] Let \(\Gamma_{1}\supsetneq\Gamma_{2}\supsetneq\cdots\) be a normal tower of cocompact lattice in a semisimple real Lie group \(G\) such that \(\cap_{n\geq 1}\Gamma_{n}=\{1\}\). For any bounded subset \(X\) of \(\widehat{G}\), we have_ \[\lim_{n\to\infty}\frac{m(X,\Gamma_{n})}{\dim_{\mathcal{L}\Gamma_{n}}H_{X}}=1.\] **Proof**: Recall that \(m_{\Gamma_{K}}(X)=\operatorname{vol}(\Gamma_{K}\backslash G(F_{\infty}))\nu_{ K}(X)\) by definition and \(\dim_{\mathcal{L}\Gamma_{n}}H_{X}=\operatorname{vol}(\Gamma_{K}\backslash G(F_{ \infty}))\nu(X)\) by Theorem 4.2. We need to show \(\lim_{n\to\infty}\nu_{K_{n}}(X)=\nu(X)\), which reduces to \(\lim_{n\to\infty}\nu_{K_{n}}(\widehat{\phi})=\phi(1)\) for all \(\phi\in C_{c}^{\infty}(G(F_{\infty})^{1})\) by Theorem 3.7. From Proposition 3.3, we know \[\operatorname{tr}R_{\operatorname{disc}}(\phi\otimes\frac{1_{K}}{\operatorname {vol}(K)})=m_{K}(\widehat{\phi})=\operatorname{vol}(G(\mathbb{Q})\backslash G (\mathbb{A})^{1})/K)\cdot\nu_{K}(\widehat{\phi}),\] which is to say \(\operatorname{tr}R_{\operatorname{disc}}(\phi\otimes 1_{K})=\operatorname{vol}(G( \mathbb{Q})\backslash G(\mathbb{A})^{1}))\cdot\nu_{K}(\widehat{\phi})\). By Proposition 2.4, we have \(\lim_{n\to\infty}\operatorname{tr}R_{\operatorname{disc}}(\phi\otimes 1_{K_{n}})= \operatorname{vol}(G(\mathbb{Q})\backslash G(\mathbb{A})^{1}))\cdot\phi(1)\). Hence \(\lim_{n\to\infty}\nu_{K_{n}}(\widehat{\phi})=\phi(1)\). For the non-uniform case, the distribution \(J(f)\) in Equation 5 will no longer be the trace of \(R_{\operatorname{disc}}(f)\), which leads to the main task for most arithmetic subgroups. Fortunately, Finis-Lapid-Muller proved the following result on the limit of the spectral side of Equation 5 (see [20] Corollary 7.8). **Theorem 4.5**.: _[Finis-Lapid-Muller] Suppose \(G=\operatorname{SL}(n)\). Let \(\{I_{n}\}\) be a family of descending integral ideals in \(\mathcal{O}_{F}\) prime to \(S\) and \(K_{n}=K(I_{n})\) be the compact subgroups of \(G(\mathbb{A}^{S})\) given by \(I_{n}\). We have_ \[\lim_{n\to\infty}J(h_{S}\otimes 1_{K_{n}})=\lim_{n\to\infty}\operatorname{tr}R _{\text{disc}}(h_{S}\otimes 1_{K_{n}})\] _for any \(h_{S}\in C_{c}^{\infty}(G(F_{S})^{1})\)._ Then we are able to prove: **Corollary 4.6**.: _[principal congruence subgroups in \(\operatorname{SL}(n,\mathbb{R})\)] Let \(\Gamma_{1}\supsetneq\Gamma_{2}\supsetneq\cdots\) be a tower of principal congruence subgroups in \(G=\operatorname{SL}(n,\mathbb{R})\). For any bounded subset \(X\) of \(\widehat{G}\), we have_ \[\lim_{n\to\infty}\frac{m(X,\Gamma_{n})}{\dim_{\mathcal{L}\Gamma_{n}}H_{X}}=1.\] **Proof**: As shown in Theorem 4.4, it suffices to prove \(\lim_{n\to\infty}\nu_{K_{n}}(\widehat{\phi})=\phi(1)\) for all \(\phi\in C_{c}^{\infty}(G(F_{\infty})^{1})\). By Proposition 3.3 and Theorem 4.5, we know \[\lim_{n\to\infty}\operatorname{vol}(G(\mathbb{Q})\backslash G(\mathbb{A})^{1 }))\cdot\nu_{K_{n}}(\widehat{\phi})=\lim_{n\to\infty}\operatorname{tr}R_{ \text{disc}}(\phi\otimes 1_{K_{n}})=\lim_{n\to\infty}J(\phi\otimes 1_{K_{n}}).\] As \(\lim_{n\to\infty}\frac{\operatorname{vol}(G(F)\backslash G(\mathbb{A})^{1}) \phi(1)}{J(\phi\otimes 1_{K_{n}})}=1\) in Corollary 2.8, we obtain \(\lim_{n\to\infty}\nu_{K_{n}}(\widehat{\phi})=\phi(1)\).

2304.05657

Towards pore-scale simulation of combustion in porous media using a low-Mach hybrid lattice Boltzmann/finite difference solver

A hybrid numerical model previously developed for combustion simulations is extended in this article to describe flame propagation and stabilization in porous media. The model, with a special focus on flame/wall interaction processes, is validated via corresponding benchmarks involving flame propagation in channels with both adiabatic and constant-temperature walls. Simulations with different channel widths show that the model can correctly capture the changes in flame shape and propagation speed as well as the dead zone and quenching limit, as found in channels with cold walls. The model is further assessed considering a pseudo 2-D porous burner involving an array of cylindrical obstacles at constant temperature, investigated in a companion experimental study. Furthermore, the model is used to simulate pore-scale flame dynamics in a randomly-generated 3-D porous media. Results are promising, opening the door for future simulations of flame propagation in realistic porous media.

S. A. Hosseini, Dominique Thevenin

2023-04-12T07:23:01Z

http://arxiv.org/abs/2304.05657v1

Towards pore-scale simulation of combustion in porous media using a low-Mach hybrid lattice Boltzmann/finite difference solver ###### Abstract A hybrid numerical model previously developed for combustion simulations is extended in this article to describe flame propagation and stabilization in porous media. The model, with a special focus on flame/wall interaction processes, is validated via corresponding benchmarks involving flame propagation in channels with both adiabatic and constant-temperature walls. Simulations with different channel widths show that the model can correctly capture the changes in flame shape and propagation speed as well as the dead zone and quenching limit, as found in channels with cold walls. The model is further assessed considering a pseudo 2-D porous burner involving an array of cylindrical obstacles at constant temperature, investigated in a companion experimental study. Furthermore, the model is used to simulate pore-scale flame dynamics in a randomly-generated 3-D porous media. Results are promising, opening the door for future simulations of flame propagation in realistic porous media. ## I Introduction Rapid depletion of fossil fuel resources and related pollutant emissions are a consequence of their widespread and abundant use in most areas of industry and technology [1; 2; 3]. Motivated by these two issues, the search for more efficient and eco-friendly energy production technologies and their implementation at the industrial level is growing by the day. Combustion in porous media has been proven to be one promising route to tackle some of the previously-cited challenges. For burners, the concept of porous media can result in high power densities, increased power dynamic range, and low emissions of NO and CO\({}_{2}\)[4]. This is, for the most part, the consequence of the presence of a solid porous matrix which has higher levels of heat capacity, conductivity, and emissivity as compared to the gaseous phase. The concept of combustion in porous media is also present in other eco-friendly technologies, for instance in packed bed reactors with Chemical Looping Combustion that allow for efficient separation of CO\({}_{2}\)[5; 6]. Similar challenges involving intense flame/wall interactions are faced in meso- and micro-combustion found in corresponding burners developed within the context of micro electro-mechanical systems [7; 8]. Given the pronounced impact of flame/solid interactions, the further development of such technologies requires a better understanding of flame/wall interaction dynamics. For this purpose, it is essential to develop numerical models that are able to properly capture such physics with a sufficient level of accuracy. The topic of flame/wall interaction has been tackled in a variety of articles in the past decades, starting with investigations of head-on quenching [9], mostly to quantify wall heat flux [10]. Such interesting investigations have been going on up to now, involving additional configurations and aspects as well as a variety of fuels [11; 12]. Even more relevant for the present investigations are flames propagating in narrow channels. Corresponding publications and results presented therein point to the very rich physics of the flame front when propagating in such a channel, see for instance [13; 14; 15; 16]. Depending on the ratio of the channel diameter to the flame thickness and on the type of thermal boundary condition at the wall the flame front can take on a wide variety of shapes, most notably, the so-called tulip shape [16]. Extending further this line of research, flame propagation within porous media has also been studied with different levels of complexity, starting with academic configurations in [17]. These preliminary studies led the authors to the conclusion that in the context of flame propagation in porous media, different flame propagation speeds exist, which is in agreement with the different propagation modes observed for flame propagation in channels. While volume-averaged approaches appear to be a cost-efficient tool for simulations of large-size, realistic systems, these observations clearly show the necessity of direct pore-scale simulations for a better understanding of the interaction process. To the authors' knowledge, apart from [18] where the authors model flame propagation in straight channels and [19] where authors discuss specifically coal combustion, all studies targeting combustion applications in porous media and configurations dominated by flame/wall interactions have been carried out using classical, discrete solvers for the Navier-Stokes-Fourier equations, coupled to balance equations for the individual species. In the low-Mach number limit, to alleviate the limitation in time-step resulting from the presence of acoustic modes, most such solvers rely on the so-called zero-Mach approximation [20], which by virtue of the Helmholtz decomposition of the velocity field brings the Poisson equation into the scheme, see for instance [21]. The elliptic Poisson equation is well-known to be the computational bottleneck of incompressible Navier-Stokes models. To solve this issue, different approaches such as Chorin's artificial compressibility method (ACM) [22] replacing the Poisson equation with a hyperbolic equation for the pressure have been proposed for incompressible flows. The lattice Boltzmann method (LBM), which emerged in the literature in the late 80's [23], has now achieved widespread success. This is in particular due to the fully hyperbolic nature of all involved equations. In addition, and as an advantage over ACM, normal acoustic modes are also subject to dissipation and, therefore, are governed by a parabolic partial differential equation allowing the LBM to efficiently tackle unsteady flows. Following up on the same idea, we recently proposed an algorithm for low-Mach thermo-compressible flows based on the lattice Boltzmann method [24; 25; 26]. Different from other LBM approaches proposed in recent years for combustion simulation [18; 19; 27], this scheme is specifically tailored for the low-Mach regime. While this model has been successfully used for large-eddy simulations (LES) of flames in complex geometries, in particular swirl burners [28], detailed interactions between flame fronts and walls have not been considered in detail up to now, since they did not play a central role for the considered systems. In this study a corresponding validation of the solver is proposed, including boundary conditions for curved walls. Configurations of increasing complexity are considered, such as flame propagation in narrow channels of different widths involving different thermal boundary conditions, as well as combustion in a reference 2-D packed bed reactor corresponding to a companion experimental study. Note that the so-called pores considered in the present study are large, being indeed inter-particle spaces at the millimeter or centimeter scale, and not restricted to a few micrometers, as found in many other applications. In this article, the terms pore and inter-particle space are used interchangeably to designate the same configuration. After a brief refresher of the model itself, along with its multiple relaxation time (MRT) cumulants realization, a discussion of the boundary conditions is proposed for both the lattice Boltzmann and the finite-difference (FD) solvers. Afterwards, results from the different validation cases are presented and discussed, before conclusion. ## II Theoretical background ### Governing equations The model used here and detailed in the next subsections targets the low-Mach approximation to describe thermo-compressible reacting flows [29]. The species mass balance equation reads in non-conservative form: \[\partial_{t}Y_{k}+\mathbf{u}\cdot\mathbf{\nabla}Y_{k}+\frac{1}{\rho}\mathbf{\nabla}\cdot \rho\mathbf{V}_{k}Y_{k}=\frac{\omega_{k}}{\rho}, \tag{1}\] where \(Y_{k}\) is the \(k^{\text{th}}\) species mass fraction, \(\rho\) the local density, \(\mathbf{u}\) the mixture velocity, and \(\omega_{k}\) the source term due to chemical reactions. The mass flux due to diffusion, \(Y_{k}Y_{k}\), is given by: \[Y_{k}\mathbf{V}_{k}=-\frac{D_{k}W_{k}}{W}\mathbf{\nabla}X_{k}+Y_{k}\sum_{k^{\prime}=1} ^{N_{\text{sp}}}\frac{D_{k^{\prime}}W_{k^{\prime}}}{W}\mathbf{\nabla}X_{k^{\prime}} \tag{2}\] where \(X_{k}\), \(W_{k}\) and \(D_{k}\) are respectively the \(k^{\text{th}}\) species mole fraction, molar mass and mixture-averaged diffusion coefficient. \(W\) is the mixture molar mass. The second term corresponds to the correction velocity ensuring local conservation of total mass (i.e., \(\sum_{k=1}^{N_{\text{sp}}}Y_{k}V_{k}=0\)). The momentum balance equation (Navier-Stokes) reads: \[\partial_{t}(\rho\mathbf{u})+\mathbf{\nabla}\cdot(\rho\mathbf{u}\otimes\mathbf{u})+\mathbf{\nabla }\cdot\mathbf{S}=0, \tag{3}\] where the stress is: \[\mathbf{S}=P_{h}\mathbf{I}-\mu\left(\mathbf{\nabla}\mathbf{u}+\mathbf{\nabla}\mathbf{u}^{t}-\frac{2} {D}\mathbf{\nabla}\cdot\mathbf{u}\mathbf{I}\right)-\eta\mathbf{\nabla}\cdot\mathbf{u}\mathbf{I}. \tag{4}\] in which \(\mu\) and \(\eta\) are the mixture-averaged dynamic and bulk viscosity coefficients and \(P_{h}\) is the hydrodynamic pressure tied to the total pressure as \(P=P_{0}+P_{h}\), with \(P_{0}\) the uniform thermodynamic pressure. The employed closure for the hydrodynamic pressure \(P_{h}\) reads: \[\frac{1}{\rho c_{s}^{2}}\partial_{t}P_{h}+\mathbf{\nabla}\cdot\mathbf{u}=\Lambda, \tag{5}\] where \(c_{s}\) is the characteristic propagation speed of normal modes, also known as sound speed. At the difference of a truly compressible model, here \(c_{s}\) is not necessarily the physical sound speed. Using the continuity equation and the ideal gas mixture equation of state, one gets: \[\Lambda=\frac{\partial_{t}T+\mathbf{u}\cdot\mathbf{\nabla}T}{T}+\sum_{k=1}^{N_{\text{ sp}}}\frac{W}{W_{k}}\left(\partial_{t}Y_{k}+\mathbf{u}\cdot\mathbf{\nabla}Y_{k}\right). \tag{6}\] Finally the energy balance equation is given by \[\rho\mathbf{c}_{p}\left(\partial_{t}T+\mathbf{u}\cdot\mathbf{\nabla}T\right)- \mathbf{\nabla}\cdot\left(\lambda\mathbf{\nabla}T\right)\\ +\rho\left(\sum_{k=1}^{N_{\text{sp}}}c_{p_{k}}Y_{k}\mathbf{V}_{k}\right) \cdot\mathbf{\nabla}T=\omega_{T}, \tag{7}\] where \(c_{p_{k}}\) and \(c_{p}\) are respectively the \(k^{\text{th}}\) species and the mixture specific heat capacities and \(\lambda\) is the thermal diffusion coefficient. One point that is to be noted is the difference of the current low-Mach set of equations with the zero-Mach model of Majjda and the low-Mach model of Toutant; Setting \(c_{s}\) to be the real sound speed in Eq. 5 reduces it to that of [30], but now for a multi-species reacting system. On the other hand, in the limit of \(c_{s}\rightarrow\infty\) one ends up with Majjda's zero-Mach limit [20], i.e. \(\mathbf{\nabla}\cdot\mathbf{u}=\Lambda\). A detailed perturbation analysis of this system would be interesting but will be left for future publications. In the next section the lattice Boltzmann model used to recover the corresponding hydrodynamic limit is briefly introduced. ### Lattice Boltzmann model To solve the low-Mach aerodynamic equations, we use a lattice Boltzmann model that we have developed in previous works [24; 25; 26]: \[g_{i}(\mathbf{r}+c_{i}\delta\mathbf{r},t+\delta t)-g_{i}(\mathbf{r},t)=\Omega_{i}+\delta t \Xi_{i}, \tag{8}\] where \(g_{i}\) are discrete populations, \(\mathbf{c}_{i}\) corresponding discrete velocities, \(\mathbf{r}\) and \(t\) the position in space and time, \(\delta t\) the time-step size and \[\Xi_{i}=c_{s}^{2}\left(f_{i}^{\text{eq}}/\rho-w_{i}\right)\left(\mathbf{c}_{i}-\mathbf{ u}\right)\cdot\mathbf{\nabla}\rho+w_{i}\rho c_{s}^{2}\Lambda. \tag{9}\] Here, \(W_{k}\) is the molar mass of species \(k\) and \(W\) the average molar mass, \(N_{\text{sp}}\) the number of species, \(w_{i}\) the weights associated to each discrete velocity in the lattice Boltzmann solver and \(c_{s}\) the lattice sound speed tied to the time-step and grid size \(\delta r\) as \(c_{s}=\delta r/\sqrt{3}\delta t\). The equilibrium distribution function, \(f_{i}^{\text{eq}}\), is given by: \[f_{i}^{\text{eq}}=w_{i}\rho\left(1+\frac{\mathbf{c}_{i}\cdot\mathbf{u}}{c_{s}^{2}}+ \frac{(\mathbf{c}_{i}\cdot\mathbf{u})^{2}}{2c_{s}^{4}}-\frac{\mathbf{u}^{2}}{2c_{s}^{2}} \right). \tag{10}\] The collision term \(\Omega_{i}\) is defined as: \[\Omega_{i}=-\omega_{s}\left(g_{i}-g_{i}^{\text{eq}}\right), \tag{11}\] where \[g_{i}^{\text{eq}}=w_{i}(P_{h}-\rho c_{s}^{2})+c_{s}^{2}f_{i}^{\text{eq}}, \tag{12}\] and \(P_{h}\) is the hydrodynamic pressure. In the present study first-neighbour stencils based on third-order quadratures are used, i.e. D2Q9 and D3Q27. The hydrodynamic pressure and momentum are computed as moments of the distribution function \(g_{i}\): \[P_{h} =\sum_{i=1}^{Q}g_{i}+\frac{\delta t}{2}\rho c_{s}^{2}\Lambda, \tag{13a}\] \[\rho\,\mathbf{u} =\frac{1}{c_{s}^{2}}\sum_{i=1}^{Q}c_{i}g_{i}. \tag{13b}\] This lattice Boltzmann model recovers the previously introduced pressure evolution equation along with the Navier-Stokes equation. In the viscous stress tensor deviations from Galilean invariance are limited to third order. ### Implementation of the Multiple Relaxation Times (MRT) collision operator In the context of the present study, following our proposals for both multi-phase and multi-species flows [31; 28], the Cumulants-based operator is used [32]. The post-collision populations \(g_{i}^{*}\) are computed as: \[g_{i}^{*}=\rho c_{s}^{2}{f_{i}^{{}^{\prime}}}^{*}+\frac{\delta t}{2}\Xi_{i}, \tag{14}\] where the post-collision pre-conditioned populations \({f_{i}^{{}^{\prime}}}^{*}\) are: \[{f_{i}^{{}^{\prime}}}^{*}=\mathcal{M}^{-1}\left(\mathcal{I}-\mathcal{W}\right) \mathcal{K}^{{}^{\prime}}+\mathcal{M}^{-1}\mathcal{W}\mathcal{K}^{{}^{\prime}}, \tag{15}\] In this equation, \(\mathcal{M}\) is the moments transform matrix from pre-conditioned populations to the target momentum space, \(\mathcal{I}\) the identity matrix and \(\mathcal{W}\) the diagonal relaxation frequencies matrix \[\mathcal{W}=\text{diag}(\omega_{0},\omega_{x},\omega_{y},...,\omega_{\text{ xxyzz}}), \tag{16}\] where the operator diag is defined as: \[\text{diag}(\mathbf{A})=(\mathbf{A}\otimes\mathbf{1})\circ\mathcal{I}, \tag{17}\] with \(\mathbf{A}\) a given vector and \(\mathbf{1}\) a vector with elements 1. The relaxation frequencies of second-order shear moments, e.g. \(x\)y (here shown with \(\omega_{s}\) for the sake of readability) are defined as: \[\omega_{s}=\frac{\nu}{c_{s}^{2}\delta t}+\frac{1}{2}, \tag{18}\] where \(\nu\) is the local effective kinematic viscosity. Prior to transformation to momentum space the populations are preconditioned as: \[f_{i}^{\prime}=\frac{1}{\rho c_{s}^{2}}g_{i}+\frac{\delta t}{2\rho c_{s}^{2}} \Xi_{i}. \tag{19}\] This pre-conditioning accomplishes two tasks, 1) normalizing the populations with the density \(-\) and thus eliminating the density-dependence of the moments -, and 2) introducing the first half of the source term. As such the moments \(\mathcal{K}\) are computed as: \[\mathcal{K}_{j}^{{}^{\prime}}=\mathcal{M}_{ij}f_{i}^{{}^{\prime}}. \tag{20}\] The Cumulants \(\mathcal{K}_{j}\) are computed from the central moments of the distribution function, these central moments being defined as: \[\widetilde{\Pi}_{x^{\prime}y^{\prime}z^{\prime}}^{{}^{\prime}}=\sum_{i}\left(c _{i,x}-u_{x}\right)^{p}(c_{i,y}-u_{y})^{q}(c_{i,z}-u_{z})^{r}f_{\alpha}^{{}^{ \prime}}. \tag{21}\] As noted in [32], up to order three Cumulants are identical to their central moments counter-parts. At higher orders they are computed as: \[\mathcal{K}_{xxyzz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyzz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{ \prime}}\widetilde{\Pi}_{yz}^{{}^{\prime}}-2\widetilde{\Pi}_{xy}^{{}^{\prime} }\widetilde{\Pi}_{xz}^{{}^{\prime}},\] (22a) \[\mathcal{K}_{xxyxy}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-\widetilde{\Pi}_{xz}^{{}^{ \prime}}\widetilde{\Pi}_{yyz}^{{}^{\prime}}-2\widetilde{\Pi}_{xy}^{{}^{\prime} }\widetilde{\Pi}_{xy}^{{}^{\prime}},\] (22b) \[\mathcal{K}_{xxyzz}^{{}^{\prime}} =\widetilde{\Pi}_{xyzz}^{{}^{\prime}}-\widetilde{\Pi}_{yyzz}^{{}^{ \prime}}\widetilde{\Pi}_{xyxy}^{{}^{\prime}}-\widetilde{\Pi}_{yy}^{{}^{\prime} }\widetilde{\Pi}_{xzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-\widetilde{\Pi}_{xx}^{{}^{\prime} }\widetilde{\Pi}_{yyzz}^{{}^{\prime}}\] \[\mathcal{K}_{xxyyz}^{{}^{\prime}} =\widetilde{\Pi}_{xxyyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}-4\widetilde{\Pi}_{yx}^{{}^{\prime} }\widetilde{\Pi}_{xyz}^{{}^{\prime}}-4\widetilde{\Pi}_{xy}^{{}^{\prime}} \widetilde{\Pi}_{xyz}^{{}^{\prime}}\] \[\mathcal{K}_{xxy}^{{}^{\prime}} =\widetilde{\Pi}_{xx}^{{}^{\prime}}\widetilde{\Pi}_{yy}^{{}^{ \prime}}-4\widetilde{\Pi}_{xy}^{{}^{\prime}}\widetilde{\Pi}_{xyz}^{{}^{ \prime}}-4\widetilde{\Pi}_{xy}^{{}^{\prime}}\widetilde{\Pi}_{xyz}^{{}^{\prime}}\] \[\mathcal{K}_{xxy}^{{}^{\prime}} =\widetilde{\Pi}_{xxy}^{{}^{\prime}}\widetilde{\Pi}_{xy}^{{}^{ \prime}}\widetilde{\Pi}_{xz}^{{}^{\prime}}\] (22d with \(\omega_{j}\) the relaxation frequency of Cumulant \(j\). After collision, the Cumulants \(\mathcal{K}_{j}^{*}\) have to be transformed back into populations \(f_{i}^{{}^{\prime}s}\). The first step, as for the forward transformation is to get the corresponding central moments. Given that up to order three central moments and Cumulants are the same, we only give here the backward transformation of higher-order moments: \[\widetilde{\Pi}_{\text{xxyz}}^{*} =\mathcal{K}_{xxyz}^{{}^{\prime}*}+\widetilde{\Pi}_{xx}^{{}^{ \prime}s}\widetilde{\Pi}_{yz}^{{}^{\prime}*}+2\widetilde{\Pi}_{xy}^{{}^{ \prime}s}\widetilde{\Pi}_{xz}^{{}^{\prime}*} \tag{24a}\] \[\widetilde{\Pi}_{\text{xxyz}}^{*} =\mathcal{K}_{xxy}^{{}^{\prime}*}+\widetilde{\Pi}_{xx}^{*} \widetilde{\Pi}_{yz}^{{}^{\prime}*}+2\widetilde{\Pi}_{xy}^{{}^{\prime}s} \widetilde{\Pi}_{xy}^{{}^{\prime}*}\] (24b) \[\widetilde{\Pi}_{xyxyx}^{{}^{\prime}*} =\mathcal{K}_{xyxy}^{{}^{\prime}*}+\widetilde{\Pi}_{xz}^{*} \widetilde{\Pi}_{xy}^{{}^{\prime}*}+\widetilde{\Pi}_{xy}^{{}^{\prime}s} \widetilde{\Pi}_{xz}^{{}^{\prime}*}+4\widetilde{\Pi}_{yz}^{{}^{\prime}s} \widetilde{\Pi}_{xyz}^{{}^{\prime}*}\] \[\quad+2\widetilde{\Pi}_{xz}^{{}^{\prime}*}\widetilde{\Pi}_{yz}^{ {}^{\prime}*}+2\widetilde{\Pi}_{xy}^{{}^{\prime}*}\widetilde{\Pi}_{yzx}^{{}^{ \prime}*}\] (24c) \[\widetilde{\Pi}_{xxyzz}^{{}^{\prime}*} =\mathcal{K}_{xxyzxyz}^{{}^{\prime}*}+4\widetilde{\Pi}_{xyz}^{{ }^{\prime}*}\widetilde{\Pi}_{xyz}^{{}^{\prime}*}+4\widetilde{\Pi}_{xz}^{{}^{ \prime}*}\widetilde{\Pi}_{yyz}^{{}^{\prime}*}\] \[\quad+\widetilde{\Pi}_{xy}^{{}^{\prime}s}\widetilde{\Pi}_{xzxz}^{ {}^{\prime}*}+\widetilde{\Pi}_{xy}^{{}^{\prime}*}\widetilde{\Pi}_{xyz}^{{}^{ \prime}*}+4\widetilde{\Pi}_{yz}^{{}^{\prime}s}\widetilde{\Pi}_{xyzxyz}^{{}^{ \prime}*}\] \[\quad+4\widetilde{\Pi}_{xz}^{{}^{\prime}*}\widetilde{\Pi}_{xyz}^{ {}^{\prime}*}+4\widetilde{\Pi}_{xy}^{{}^{\prime}s}\widetilde{\Pi}_{yz}^{{}^{ \prime}*}+2\widetilde{\Pi}_{xy}^{{}^{\prime}s}\widetilde{\Pi}_{xzxz}^{{}^{ \prime}*}\] \[\quad+2\widetilde{\Pi}_{xz}^{{}^{\prime}s}\widetilde{\Pi}_{yz}^{ {}^{\prime}*}+2\widetilde{\Pi}_{xz}^{{}^{\prime}s}\widetilde{\Pi}_{yz}^{{}^{ \prime}*}+16\widetilde{\Pi}_{xy}^{{}^{\prime}s}\widetilde{\Pi}_{yz}^{{}^{ \prime}*}\widetilde{\Pi}_{yz}^{{}^{\prime}*}\] \[\quad+4\widetilde{\Pi}_{xz}^{{}^{\prime}s}\widetilde{\Pi}_{xyz}^{ {}^{\prime}*}+4\widetilde{\Pi}_{xy}^{{}^{\prime}s}\widetilde{\Pi}_{yz}^{{}^{ \prime}*}+4\widetilde{\Pi}_{yz}^{{}^{\prime}s}\widetilde{\Pi}_{xyzxyz}^{{}^{ \prime}*}\] \[\quad+2\widetilde{\Pi}_{xz}^{{}^{\prime}s}\widetilde{\Pi}_{yy}^{ {}^{\prime}*}\widetilde{\Pi}_{zz}^{{}^{\prime}*}. \tag{24d}\] Once central moments have been obtained the inverse of the central moments transform tensor is used to compute the corresponding populations. ### Solver for species and energy balance equations In the context of the present study the species and energy balance laws (Eqs. 1 and 7) are solved using finite differences. To prevent the formation of Gibbs oscillations at sharp interfaces, convective terms are discretized using a third-order weighted essentially non-oscillatory (WENO) scheme while diffusion terms are treated via a fourth-order central scheme. Near boundary nodes, to prevent any nonphysical interaction of the smoothness indicator with ghost nodes, a centered second-order scheme is used to discretize the convection term. Global mass conservation of the species balance equation, i.e. \(\sum_{k}Y_{k}=1\), while naturally satisfied for classical discretizations of the convection term, for instance in 1-D: \[\frac{u_{x}}{2\delta r}\sum_{k}\left[Y_{k}(x+\delta r)-Y_{k}(x-\delta r) \right]=0. \tag{25}\] is not necessarily satisfied for WENO schemes, as coefficients weighing contributions of each stencil are not the same for all species. To guarantee conservation of overall mass the concept of correction speed is used as for the diffusion model; Representing the discretization via an operator \(\mathcal{L}\) the discrete convection term is computed as: \[\mathbf{u}\cdot\mathbf{\nabla}Y_{k}=\mathbf{u}\cdot\left[\mathcal{L}(\mathbf{\nabla}Y_{k})-Y _{k}\sum_{k^{\prime}}\mathcal{L}(\mathbf{\nabla}Y_{k^{\prime}})\right] \tag{26}\] which - once summed up over all species - gives: \[\sum_{k}\mathbf{u}\cdot\mathbf{\nabla}Y_{k}=\mathbf{u}\cdot\left[\sum_{k}\mathcal{L}(\mathbf{ \nabla}Y_{k})-\sum_{k^{\prime}}\mathcal{L}(\mathbf{\nabla}Y_{k^{\prime}})\right]=0. \tag{27}\] All equations are discretized in time using a first-order Euler approach. Transport and thermodynamic properties of the mixture along with the kinetic scheme are taken into account via the open-source library Cantera, coupled to our in-house solver ALBORZ [25]. Details of the coupling can be found in [33]. ## III Boundary conditions ### Lattice Boltzmann solver In the context of the present study three types of boundary conditions are needed for the lattice Boltzmann solver, namely wall, inflow, and outflow boundary conditions. A brief overview of these boundary conditions is given in what follows. Solid boundaries are modeled using the half-way bounce-back scheme. For this purpose, missing populations are computed as [34]: \[f_{i}\left(\mathbf{r},t+\delta t\right)=f_{i}^{*}\left(\mathbf{r},t\right), \tag{28}\] where \(f_{i}^{*}\) is the post-collision population (prior to streaming) and \(\vec{i}\) is the index of the particle velocity opposite that of \(i\). To take into account wall curvature the interpolated half-way bounce back approach is used [35]. At a given boundary node \(\mathbf{r}_{f}\), the missing incoming populations are computed as: \[f_{i}(\mathbf{r}_{f},t+\delta t) =2qf_{i}(\mathbf{r}_{f}+\mathbf{c}_{i},t+\delta t)\] \[+(1-2q)f_{i}(\mathbf{r}_{f},t+\delta t),\forall q<\frac{1}{2}, \tag{29a}\] \[f_{i}(\mathbf{r}_{f},t+\delta t) =\frac{1}{2q}f_{i}(\mathbf{r}_{f}+\mathbf{c}_{i},t+\delta t)\] \[+\frac{2q-1}{2q}f_{i}(\mathbf{r}_{f},t+\delta t),\forall q\geq\frac{1} {2}, \tag{29b}\] where \(\vec{i}\) designates the direction opposite \(i\) and \(q\) reads: \[q=\frac{||\mathbf{r}_{f}-\mathbf{r}_{s}||}{||\mathbf{c}_{i}||}, \tag{30}\] with \(\mathbf{r}_{s}\) denoting the wall position along direction \(i\). For inlet boundary conditions a modified version of the half-way bounce-back scheme is used to impose a target inlet velocity vector \(\mathbf{u}_{\text{in}}\). To that end the missing populations are computed as: \[f_{i}\left(\mathbf{r},t+\delta t\right)=f_{i}^{*}\left(\mathbf{r},t\right)+(w_{i}+w_{i}) \rho_{\text{in}}\mathbf{u}_{\text{in}}\cdot\mathbf{c}_{i}. \tag{31}\] In addition to velocity boundary conditions, a modified non-reflecting version of the zero-gradient boundary condition is also employed [34] at the outlet, as first introduced in [32]. The missing populations at the outflow boundary are defined as: \[f_{i}\left(\mathbf{r},t+\delta t\right) = f_{i}\left(\mathbf{r}-\mathbf{n}\delta r,t\right)\left(c_{s}-\mathbf{u}(\mathbf{r },t)\cdot\mathbf{n}\right) \tag{32}\] \[+f_{i}\left(\mathbf{r}\delta r,t\right)\left(\frac{\delta r}{\delta t }-c_{s}+\mathbf{u}(\mathbf{r},t)\cdot\mathbf{n}\right),\] where \(\mathbf{n}\) is the outward-pointing unit vector normal to the boundary surface. ### Energy and Species fields In addition to the application of boundary conditions to the discrete populations, given that the model involves derivatives of macroscopic properties such as density, appropriate measures have to be taken. For the finite-difference solver and all terms involving this approximation, the boundary conditions are implemented via the image/ghost node method [36; 37; 38]. Representing the macroscopic parameter of interest with the generic variable \(\phi\), for a Dirichlet boundary condition for instance, one would have: \[\phi(B)=\phi_{B}, \tag{33}\] where \(B\) refers to the position of the boundary, shown in Fig. 1. The _virtual_ field value \(\phi(G)\) in the ghost node, the discrete grid-point outside the fluid domain neighboring the boundary (Fig. 1) is computed as: \[\phi(G)=2\phi_{B}-\phi(I), \tag{34}\] where \(I\) is the image point in the fluid domain placed such that \(\overline{GB}=\overline{BI}\) with both line segments perpendicular to the boundary interface. Since the image node does not necessarily fall on a grid-point it is reconstructed using data from neighboring grid points. For the reconstruction process to be robust with respect to the wall geometry, Shepard's inverse distance weighting is used [39]: \[\phi(r_{I})=\sum_{j=1}^{N}w_{j}\phi(r_{j}), \tag{35}\] with: \[w_{j}=\frac{d(r_{I},r_{j})^{-p}}{\sum_{j=1}^{N}d(x_{I},r_{j})^{-p}}, \tag{36}\] where \(d(r_{I},r_{j})\) is the distance between points \(I\) and \(j\) and \(p\) is a free parameter typically set to \(p=2\). Note that: \[\sum_{j}w_{j}=1. \tag{37}\] In order to obtain good precision, the field reconstruction at image points considers all fluid nodes neighboring \(I\) such that: \[d(r_{I},r_{j})\leq 4\delta r, \tag{38}\] which comes at the additional cost of a wider data exchange layer between cores during parallelization. Note that terms involving second-order derivatives such as the diffusion term in the energy and species balance equations also require an interpolation/reconstruction process on the diffusion coefficient. To avoid non-physical values, instead of using the previously computed properties, the coefficients at the ghost nodes are computed by applying the interpolation/reconstruction procedure directly to the transport properties. ## IV Validations and results ### Premixed laminar flame acceleration in 2-D channels The proper interaction of flames with different wall boundary conditions (isothermal, adiabatic, known heat flux) while enforcing the no-slip condition for the flow is probably the most important step when extending a combustion solver to porous media applications. To that end, the propagation of premixed flames in narrow 2-D channels is first considered to verify that the proposed solver correctly captures the different flame front regimes. Two configurations are considered: (a) Adiabatic and (b) constant-temperature channel walls. Given that the width of the channel, here written \(H\), plays an important role to control flame front shape, heat exchange, as well as propagation speed, different cases with different channel widths have been computed. All configurations involve 2-D channels of height \(H\) and length \(L=20H\). At the inflow (left end of the domain) a stoichiometric mixture of methane/air at temperature \(T_{\text{in}}=300\) K is injected. The flow rate is dynamically set throughout all simulations to match the flame propagation speed, so as to ensure a globally static flame front within the numerical domain. The top and bottom boundaries are set to no-slip walls with either constant temperature, i.e. \(T_{w}=T_{\text{in}}\), or adiabatic boundary conditions for the temperature field. At the outlet a constant-pressure boundary condition is used. Note that for the inlet a 2-D Poiseuille distribution satisfying the target mass flow rate is implemented. To initialize all simulations profiles from the steady solution of a 1-D methane/air flame with the flame placed half-way Figure 1: Illustration of the ghost/image node approach for boundary conditions. The red point \(G\) is outside the flow domain while green points are inside. in the domain are used and supplemented with the velocity distribution at the inlet. 1-D free flame propertiesAs a first step pseudo 1-D free flame simulations were run both using ALBORZ (coupled to Cantera) or Cantera (as standalone tool) using the BFER-2 two-step kinetic mechanism[7]. The results obtained with both codes have been compared, as illustrated in Fig. 2; the agreement is perfect for all species and all quantities. For this case, experimental measurements led to a flame propagation speed of \(S_{F}=0.404\) m/s[7], in excellent agreement with both solvers; ALBORZ predicts a laminar flame speed of 0.408 m/s. Furthermore, to have a clear indication regarding resolution requirements, the thermal thickness \(\delta_{T}\) defined as: \[\delta_{T}=\frac{T_{\mathrm{ad}}-T_{\mathrm{in}}}{dT/dx}\,, \tag{39}\] where \(T_{\mathrm{ad}}\) is the adiabatic flame temperature, was also computed. Simulations with ALBORZ led to \(\delta_{T}=328\)\(\mu\)m, which is in very good agreement with the value reported in [40]. This indicates that for fully resolved simulations one should implement \(\delta r<35\)\(\mu\)m, in order to get 10 grid points within the flame front. For all channel simulations conducted in the present section \(\delta r=20\)\(\mu\)m has been set. While larger grid-sizes would be sufficient for resolved simulations, as will be seen in next section, here we use a smaller grid-size to properly resolve the width of the smaller channel. Considering additionally the characteristic speed in the system the time-step size was then fixed to \(\delta t=7.5\times 10^{-8}\) s, also satisfying all stability conditions regarding Fourier and CFL numbers for the hybrid solver. Adiabatic wallsFor the first set of simulations the walls are set to be adiabatic. Three different channel widths are considered, i.e. \(H\in\{0.4,\,1,\,3\}\)mm. Simulations were conducted until the system reached steady state. Then, flame propagation speeds computed from the mass flow rate as well as flame shapes were extracted. The results are compared to those from [40] for validation. Starting from channels with widths comparable to the flame thickness (top part of Fig. 3), deformations of the flame front due to the Poiseuille velocity profile are minimal. As the channel grows in width (from top to bottom in Fig. 3) one observes more and more pronounced deformations at the center of the channel, effectively increasing the surface of the flame front. With more elongated flame surfaces one would expect changes in the propagation speed of the flame. The flame propagation speeds as a function of channel width are shown in Fig. 4 and again compared to reference data from [40]. As a first observation it is seen that the present solver matches reference data very well. Furthermore, as expected from the changes in flame shape the flame propagation speed also increases with increased channel width, reaching speeds up to three time the laminar flame speed for \(H=3\) mm. Isothermal wallsA second second set of simulations were then carried out while setting the wall boundary conditions to isothermal at \(T_{w}=300\) K. As for adiabatic walls, three different channel widths were considered, i.e. \(H\in\{2.47,\,3,\,6\}\)mm. These channel widths were selected to cover the main flame shapes occurring for this configuration as expected from the literature, i.e. parabolic and tulip profile. The results obtained with ALBORZ are compared to simulations Figure 3: Comparison of flame shape obtained for adiabatic walls from simulations with ALBORZ (top half of each subfigure) to results from [40] (bottom half of each subfigure), with increasing channel height from top to bottom. The colors show the heat release rate, while the iso-contours (black in the top part), red in the bottom part) represent the following isotherms: \(\theta=\frac{T-T_{\mathrm{in}}}{T_{\mathrm{ad}}-T_{\mathrm{in}}}\in\{0.1,0.3, 0.5,0.7,\,0.9\}\). Reference images (bottom half of each subfigure) are reproduced from [40]. Channel widths are set to be true to scale. Figure 2: Validation of stoichiometric methane/air flame against reference solver: The dashed lines are from Cantera, while the markers have been computed using ALBORZ. reported in [40] in Fig. 5. The results show good agreement with each other. Minor differences between results from ALBORZ and from [40] can, at least in part, be attributed to the fact that a two-step chemical mechanism is employed here, while [40] rely on a single-step, global mechanism. The propagation speeds were also extracted and compared to [40], as shown in Fig. 6. The agreement is observed to be very good for this quantity. Different from adiabatic walls where as channel width went down flame propagation speed converged to the free flame propagation speed, here as the channel width decreases the flame propagation speed goes below the free flame speed. This can be explained by the fact that lowering the channel width increases the energy loss toward the cold walls, compared to the energy released by the flame. It is also observed that at \(H\) below 3 mm the flame propagation speed drops sharply; this corresponds to the onset of flame quenching discussed in the next paragraph. Dead space and onset of quenchingA closer look at Figs. 3 and 5 shows that the flame front hangs on to the walls for the adiabatic cases; On the other hand, for the isothermal cases there is a layer close to the walls where the flame is extinguished due to excessive heat losses, and fresh gas flow through; this zone is referred to as the dead zone [40], as illustrated in Fig. 7. Here, the quantity \(\delta_{\text{dead}}\) is introduced as the minimum thickness of the dead zone by monitoring the peak of heat release. To do that the position along the \(x\)-axis where the distance between the reaction front (marked by maximum of heat release) and the wall is minimum is found, and the corresponding distance along the normal to the wall is extracted. These values have been computed for four different cases (for the same widths as in the previous paragraph, and additionally for \(H=2.1\) mm). The results obtained with ALBORZ agree once more well with data from [40]. It is observed that for large channel widths the dead zone thickness reaches a lower plateau at a value of \(\delta_{\text{dead}}\approx 0.4\) mm. As the channel width goes down the dead zone thickness experiences a rapid Figure 4: Comparison of flame propagation speed obtained for adiabatic walls from simulations with ALBORZ to results from [40] for different channel widths. Red circular markers are ALBORZ results while the black dashed line is data from [40]. Figure 5: Comparison of flame shape obtained for isothermal walls obtained from simulations with ALBORZ (right half of the figure) to results from [40] (left half of the figure), with increasing channel height from top to bottom. The colors show the heat release rate on the right side, while the iso-contours in black in the right part represent the following isotherms: \(\theta=\frac{T-T_{\text{th}}}{T_{\text{th}}-T_{\text{th}}}\in\{0.1,0.3,0.5,0.7, 0.9\}\). Reference images showing iso-contours on the left (top half: heat release; bottom half: temperature) are reproduced from [40]. Channel widths are set to be true to scale. Figure 6: Comparison of flame propagation speed for isothermal walls obtained from simulations with ALBORZ to results from [40] for different channel widths. Red circular markers are ALBORZ results while the black dashed line is data from [40]. growth until the point where it becomes comparable to the channel width, so that the flame can not maintain itself anymore; this is called the quenching channel width. Calculations with ALBORZ led to a value of \(H\) between 2 and 2.1 mm for the quenching width, while [40] reported \(H=2.4\) mm. The difference between the two results can be probably attributed to the different chemical schemes employed, and grid- resolution, as reference uses an adaptive grid refinement procedure leading to grid sizes of 12.5 \(\mu\)m in the diffusion and reaction layers; at such scales the slightest differences in laminar flame speed and thickness can have a pronounced effect on the flame/wall interaction dynamics. ### Methane/air premixed flame in pseudo 2-D reactor with cylindrical obstacles The next case considered in this work is that of a pseudo-2D packed bed burner presented in [41]. It has been designed by colleagues at the University of Magdeburg in the Thermodynamics Group with the aim to replicate flow physics found in industrial packed beds by incorporating relevant size, geometry, and boundary conditions. For all details regarding design and measurement apparatus the interested readers are referred to [41]. The overall geometry of the reactor, as initially intended, is illustrated in Fig. 9 in a vertical cut-plane through the center of the cylinders; it consists of a slit burner placed below a bed of cylindrical "particles". The rows of cylinders are arranged in an alternating pattern, with each consecutive row offset by precisely half the center-to-center distance. Most of the injected fuel/air mixture enters the packing between the two central cylinders of the first row, which are aligned with the slit burner. The configuration considered involves a premixed methane/air mixture at equivalence ratio of one (stoichiometry) coming in from the central inlet at speed 0.3 m/s, and air coming in from the two side inlets at the same speed to reduce the possible impact of external perturbations. All incoming fluxes are at temperature \(25^{\circ}\)C. All cylinders except three of them, the two central cylinders in the bottom-most row and the central cylinder in the middle row - i.e., the three cylinders directly above fuel inlet, are associated to adiabatic no-slip walls as boundary conditions. The three remaining, central cylinders (the ones shown for instance in Fig. 10) are set to constant-temperature no-slip walls at \(T_{\rm w}=373.15\) K \(=100^{\circ}\)C, since they are thermostated at this particular temperature in the experiments. It should be noted that the measured temperatures in the experiment actually led to temperatures of \(105\pm 1^{\circ}\)C for the side cylinders and \(120^{\circ}\)C for the top central cylinder, which also might explain some of differences between simulation and experimental results. The simulations are conducted with resolutions \(\delta r=0.05\) mm and \(\delta t=0.1\)\(\mu\)s. Before looking at the steady position/shape of the flame and compare to experimental measurements, it is interesting to look at the unsteady evolution of the flame front and interpret Figure 8: Comparison of dead zone minimum thickness for isothermal walls obtained from simulations with ALBORZ to results from [40] for different channel widths. Red circular markers are ALBORZ results while the black dashed line is data from [40]. Figure 7: Illustration of dead zone minimum thickness for case with isothermal walls at \(T_{\rm w}=300\) K for \(H=2.5\) mm. The right figure shows the flame in the channel (temperature field). The left figure corresponds to a cut through it along the dashed black line, plotting heat release (in black) and temperature (in red), together with the dead-zone limit (dashed blue line). Figure 9: Geometry of the pseudo 2-D burner with cylindrical obstacles of [41]. these results based on the flame shapes discussed in the previous section. The flame evolution in the simulations is shown in Fig. 10. The sequence of images present the flame front (described here by the temperature field) retracting along the positive \(y\)-direction going upward from the narrow gap between the two central cylinders in the bottom row toward the wider, inter-particle space located in-between the three isothermal cylinders facing the injection. In the narrowest cross-section (top-left image in Fig. 10) the flame shows a parabolic shape. As it moves further downstream, the center flattens and eventually goes toward a tulip shape (even better visible in Fig. 11, left, showing heat release). Noting that at the narrowest section the equivalent channel width is 2.3 mm, it can be seen that the behavior of the flame agrees qualitatively with that shown in Fig. 5(top) for the straight channel. At the widest section, i.e. for the bottom right snapshot in Fig. 10, \(H\approx 3.5\) mm. Referring again to the channel results discussed in the previous section, the flame front should be between a flattened parabola and a tulip (between middle and bottom row of Fig. 5), which is in good agreement with Figs. 10 and 11 - keeping in mind that the wall geometries are different in the channel and in the 2-D burner configurations. Furthermore, as for the channel with isothermal cold walls, the flame front exhibits a clear dead zone in regions neighboring the walls in Fig. 10, perhaps even better visible in Fig. 11(left). The flame front, as obtained from simulation, has been compared to experimental observations reported in [42] in Fig. 11. In the experiments, the flame front is located at about 3.5 mm above the center of the first row of cylinders along the central vertical line, while in simulations it stabilizes at approximately 2.8 mm. Furthermore, experimental measurements point to an asymmetrical flame front. This missing symmetry, as noted in [42], might be possibly explained by small inaccuracies in the actual geometry of the burner compared to the design shown in Fig. 9. To verify this point another simulation was carried out considering the finally _measured_ geometry of the real set-up as reported in [42]. The resulting flow field is illustrated via streamlines in Fig. 12. The streamlines at steady state show indeed a slightly asymmetrical flow configuration, especially in the region above the first row of cylindrical obstacles. Note that while the flow is unsteady above the bed, it reaches a steady configurations within. The distribution of velocity and temperature in the full burner geometry is shown in Fig. 13. Th effect of the asymmetry in the flow field is better visible when looking at the flame front, shown in Fig. 14. Figure 14 shows that the asymmetrical flame shape observed in the experiments is better reproduced in the hybrid simulation when taking into account the really measured geometry. In particular, the flame becomes tilted, from top left to bottom right. Furthermore the flame stabilizes at a higher position, at 3.1 mm, matching better the ex Figure 11: Illustration of flame front (right) reported in [42] from experiments compared to (left) simulations with ALBORZ for the geometry shown in Fig. 9. The numerical image on the left shows the heat release rate, while the experimental image on the right captures all spontaneous emissions from species below 550 nm. Figure 12: Flow structures illustrated by streamlines at steady state as obtained with ALBORZ based on the really measured geometry of the burner with cylindrical obstacles [42]. Figure 10: Evolution of the flame front as a function of time from top left to bottom right (corresponding to final steady state) in the configuration of Fig. 9, illustrated via the temperature field. perimental observations. The remaining discrepancy can be explained by different factors: minor differences in temperatures of iso-thermal cylinders as used in the simulation and as measured in experiments; non-homogeneous velocity and turbulence profiles at the inlet; and - regarding simulations - the simplicity of the chosen chemical scheme BFER-2, at the difference of a complete reaction mechanism. On top of this, while for simulations heat release was used to track the position of the flame front, experimental images contain spontaneous emissions from all species radiating below 550 nm, which is known to lead to a thicker flame front with deviations of the order of 0.1-1 mm regarding flame position toward the burnt gas region, i.e., here in streamwise direction, toward the top. Defining exactly the flame front has always been a challenge, since many different definitions are possible [43]; this is even more true in experiments, considering that heat release can generally not be measured directly [44]. Keeping these points in mind, the agreement between experimental measurements and numerical results appears to be good. The obtained results already show a reasonable agreement between ALBORZ and measurement data, demonstrating that the numerical solver can well capture flow/flame/wall interactions. More detailed comparisons between experimental and numerical data will be the topic of future studies involving systematic parameter variations, and relying on additional quantities for the comparisons as soon as they have been measured experimentally. ### Pore-resolved flame simulation in randomly generated porous media As a final configuration and to illustrate the applicability of the solver to more complex configurations, a geometry generated in the Porous Microstructure Analysis (PuMA) software [7] composed of randomly placed non-overlapping spheres with a diameter of 1.6 mm, a global porosity of 0.7 and a physical domain size \(L\times H\times H\) with \(L=0.08\) m and \(H=0.005\) m is considered. The geometry is illustrated in Fig. 15. Here \(L_{1}=0.01\) m and \(L_{2}=0.02\) m. For this simulations the grid- and time-step sizes are set at the same values as in the previous configuration. Periodic boundary conditions are used for the top and bottom of the simulation domain. A constant mass flow rate boundary condition is used for the inflow (on the right), where the pressure and temperature are set to 1 atm and 298.15 K. At the inflow, the species mass fractions are set to that of the fresh gas at equivalence ratio 1. At the other end of the domain a constant hydrodynamic pressure along with zero-gradient boundary conditions for species and temperature field are used. During the simulation the total consumption speed of methane is monitored via: \[S_{c}=\frac{\int_{V}\dot{\omega}_{\text{CH}_{4}}dV}{\int_{V}\dot{\omega}_{ \text{CH}_{4}}dV}, \tag{40}\] where the consumption speed is normalized by that of a flat flame front, without any interaction with a porous media. The results are displayed in Fig. 16. The average normalized propagation speed for this configuration is 1.797, with a large standard deviation of 0.6875. This larger propagation speed as Figure 14: Flame shape and position illustrated via heat release as obtained from ALBORZ simulations for the really measured geometry. Figure 13: (Left half of each subfigure) Velocity magnitude and (right half of each subfigure) temperature fields in the full burner geometry obtained with ALBORZ based on the really measured geometry of the burner with cylindrical obstacles [42]. Iso-contours are for the temperature field dividing \(T\in[300\ 2300]\) K into 10 equally-spaced intervals. Figure 15: Illustration of randomly-generated porous media geometry. compared to the laminar flame propagation speed is not unexpected. The flame dynamics in a porous media with adiabatic solid boundaries is mainly governed by the flame contortion as it goes over the solid obstacles. The consumption speed, in a process similar to that found for turbulent flames, is directly impacted by the increased flame surface. The evolution of the flame shape as it goes through the porous media is illustrated in Fig. 17. ## V Conclusions and discussion In this work a numerical model previously developed for gas-phase combustion has been extended and applied to reacting flows in porous media. Benchmark cases of increasing complexity in which flame/wall interactions dominate the dynamics of the system have been considered. It was shown that the model is able to capture the different flame/wall interaction regimes for both Dirichlet (constant temperature) and Neumann (adiabatic) boundary conditions. The suitability of the proposed solver for combustion simulations within a regular particle packing was discussed in connection to a pseudo 2-D burner involving cylindrical obstacles. First comparisons to experimental data point to a good agreement. Finally, for the first time to the authors' knowledge a lattice Boltzmann-based pore-scale simulation of combustion in a complex 3-D porous media is presented. These results open the door for future studies considering flame propagation in realistic porous media and parametric studies of reacting gas flows in packed bed configurations. ## Acknowledgement The authors acknowledge funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) in TRR 287 (Project-ID 422037413), as well as the Gauss centre for providing computation time under grant "pn73ta" on the GCS supercomputer SuperMUC-NG at Leibniz Supercomputing Centre, Munich, Germany. Additionally, the authors thank Mohammadassan Khodsiani, Benoit Fond and Frank Beyrau for interesting discussions regarding experimental measurements in the 2-D burner.

2301.07846

ClusterLog: Clustering Logs for Effective Log-based Anomaly Detection

With the increasing prevalence of scalable file systems in the context of High Performance Computing (HPC), the importance of accurate anomaly detection on runtime logs is increasing. But as it currently stands, many state-of-the-art methods for log-based anomaly detection, such as DeepLog, have encountered numerous challenges when applied to logs from many parallel file systems (PFSes), often due to their irregularity and ambiguity in time-based log sequences. To circumvent these problems, this study proposes ClusterLog, a log pre-processing method that clusters the temporal sequence of log keys based on their semantic similarity. By grouping semantically and sentimentally similar logs, this approach aims to represent log sequences with the smallest amount of unique log keys, intending to improve the ability of a downstream sequence-based model to effectively learn the log patterns. The preliminary results of ClusterLog indicate not only its effectiveness in reducing the granularity of log sequences without the loss of important sequence information but also its generalizability to different file systems' logs.

Chris Egersdoerfer, Dong Dai, Di Zhang

2023-01-19T01:54:48Z

http://arxiv.org/abs/2301.07846v1

# ClusterLog: Clustering Logs for Effective Log-based Anomaly Detection ###### Abstract With the increasing prevalence of scalable file systems in the context of High Performance Computing (HPC), the importance of accurate anomaly detection on runtime logs is increasing. But as it currently stands, many state-of-the-art methods for log-based anomaly detection, such as DeepLog, have encountered numerous challenges when applied to logs from many parallel file systems (PF-Ses), often due to their irregularity and ambiguity in time-based log sequences. To circumvent these problems, this study proposes ClusterLog, a log pre-processing method that clusters the temporal sequence of log keys based on their semantic similarity. By grouping semantically and sentimentally similar logs, this approach aims to represent log sequences with the smallest amount of unique log keys, intending to improve the ability for a downstream sequence-based model to effectively learn the log patterns. The preliminary results of ClusterLog indicate not only its effectiveness in reducing the granularity of log sequences without the loss of important sequence information but also its generalizability to different file systems' logs. ## I Introduction In light of growing datasets, increasing problem complexity, and more compute intensive algorithms, it is no question that large-scale computing systems, such as Cloud or High-Performance Computing (HPC), are a growing area of interest. In these systems, distributed storage frameworks are the critical foundation to providing global data access, and their health is therefore critical to the function of the entire system. However, due to increasing scale and complexity, distributed storage systems are subject to various bugs, failures, and anomalies in production, which lead to data loss, service outages and degradation of quality of service [15, 9, 6]. It is thereby critical to detect malfunctions accurately and in a timely manner, so that system operators can promptly pinpoint issues and resolve them immediately to mitigate losses. It has been proven that runtime logs, which record detailed runtime information about storage systems during operation, are a valuable information source to detect potential anomalies in distributed storage systems. These runtime logs are generated by log statements written in the source code (using simple printf function or logging libraries such as Log4j [4]). These logs record important internal states of distributed storage systems, such as key variable values, return values, and performance statistics, all of which can be useful to reveal system anomalies. As a result, an extensive amount of research on log-based anomaly detection has been done recently [8, 26, 16, 25, 34, 10, 7, 19, 12, 22, 32, 21, 20, 9]. The main theme of modern Log-based anomaly detection solutions is to apply machine learning, especially deep learning methods, onto log sequences to detect anomalies. The common process of doing so is for logs to be parsed and processed first, then their normal sequence to be learned by sequence-based pattern detection models such as LSTM [18]. Having learned what normal and abnormal log sequences look like, these algorithms are then shown runtime logs and are tasked to classify them accurately. Take one of the most representative works, DeepLog [12], as an example: during training, it first parses the runtime logs into templates and represents each of these templates using a single integer. Through this process, a sequence of logs becomes a sequence of integral identifiers, which will be learned using an LSTM model. During runtime, the anomalies are defined by whether or not the actual next log identifier is within the set of identifiers which the LSTM model predicts. In another recent study, NeuraLog [18], the runtime logs are identified as a vector instead of an integer to contain more semantic information. Still, these sequences of embedded vectors will are fed into a DL model to learn the normal/abnormal sequences. Although these sequence-based log anomaly detection solutions work well for many storage systems such as HDFS [2], they have one key issue: they rely heavily on the quality of the log sequences in the training data. The sequence of logs must be both accurate and representative of the log system's logic. Such sequences of logs are often expensive to obtain in the real-world. For instance, the HDFS logs used in existing studies were pre-processed by aggregating logs with the same data block ID into a sequence, regardless of how far these log entries are from each other in the runtime. Unfortunately, such pre-processing is not always possible. In fact, many parallel storage systems, such as Lustre [5] and BeeGFS [3], do not have any common identifier (ID) in log entries to denote their relevance. Missing such global IDs makes it difficult to identify the matching events or to build log sequences accurately, resulting in only raw time-based log sequences available [35]. In addition, the raw log sequences generated from a distributed environment are quite ambiguous. One source of ambiguity is the _clock skew_ in distributed systems, as logs generated concurrently across multiple nodes are merged into a single file where their order is not always equivalent to the order of time in which they occurred. Secondly, interleaved concurrent threads present a further, and more complex problem. As different nodes run separate execution threads concurrently, the unrelated processes are often logged in random, interleaving order. Directly learning from these often noisy sequences of run-time logs, can be problematic, and require a much larger labeled dataset and longer training time. To address these issues, in this study we propose ClusterLog, a log pre-processing method which clusters individual runtime logs based on their similarity to effectively reduce the ambiguity of log sequences and improve the accuracy of downstream sequence-based learning. The intuition behind ClusterLog is driven by the idea that grouping similar runtime logs together will result in less random variation within the log sequence due to a lesser amount of unique key identifiers. In addition, grouping logs based on their similarity can still retain the vital sequence information between different types of high-level file system operations. For example, Lustre log sequences contain many sequences of logs where actions are all very similar, but because of the lack of block ID, they are highly irregular in time sequence. Grouping some of these similar actions is intended to eliminate a large portion of this irregularity, providing a cleaner sequence to be learned. Further, the robust and generalizable nature of this approach allows it to be applied to numerous types of file system logs and on limited amounts of available training data, both of which are not adequately captured by previous approaches. The rest of this paper is organized as follows. In Section II, we present the design and implementation of ClusterLog. In Section III, we discuss the evaluation setup and results. Finally, in sections IV and V, we discuss related work and lay out our future work, respectively. ## II Design and Implementation The implementation of ClusterLog can be effectively broken into four parts. The first is rudimentary preprocessing, where the log content is extracted from the log files, resulting in only the natural language sequence of each log which can be matched throughout the log file to create a set of unique log keys. From here, the preprocessed log keys are fed into a pre-trained semantic similarity model to produce unique embeddings for each unique log key. Simultaneously (or in sequence) to the semantic similarity embedding step, the preprocessed log keys are fed into a pre-trained sentiment prediction model to result in a 0 or 1 prediction of each log's sentiment. The output of the semantic similarity embedding model and the sentiment prediction model are concatenated at this point and serve as the entire latent representation of each log key. Following the concatenation, the embeddings are fit to a clustering algorithm where the resulting cluster labels are used to replace the original sequence of log keys. Finally, the sequence of these cluster labels are fed into the downstream sequential analysis algorithm. In our current implementation, we use DeepLog's sequence-learning part as the downstream algorithm. It is a two-layer LSTM network which is used to predict a probability distribution of the next log key given a specified window of previous logs. If the next log key is within the top candidates of this distribution, the sequence is considered normal, otherwise it is labeled as anomalous. The code base which implements the design described in this section is available using the following link: [https://github.com/DIR-LAB/ClusterLog](https://github.com/DIR-LAB/ClusterLog). ### _Preprocessing_ The first step of the ClusterLog implementation is rudimentary preprocessing of the raw log files. The goal of this step is to remove any excess characters and information from each individual log key, to result in a stripped key containing only the natural language sequence of the log which can be extracted as a unique log key and matched with many other logs in the original file. While it is ideal to reduce the log to its most generic natural language form to result in the smallest possible amount of unique log keys, an approximation of this will likely suffice if the former is too difficult to achieve. This is sufficient as the semantic embedding model will almost identically embed keys which are of the same logging template but are not a perfect match, meaning they will be clustered together even at the lowest of distance thresholds. As preprocessing can occasionally be very tedious, approximation at this step is highly valuable in regards to time and effort required to set up this approach compared to others. Depending on the file system which is being analyzed, the exact preprocessing steps may vary, but with the same goal in mind, preprocessing generally consists of extracting the log content, removing unique identifiers such as time stamps or Block IDs, removing unnecessary characters or character patterns, and unabbreviating words. ### _Semantic Embedding_ Once the logs have been preprocessed, the natural language sequence of each log is fed into a pre-trained semantic embedding model. Though a number of applicable models do exist, the most accurate embeddings were achieved using the all-mpnet-base-v2 [1] model which itself was pre-trained on the original mpnet-base [28] model, published in 2020, and further fine-tuned on over 2.1 Billion sentence pairs to optimize the model for semantic similarity. Though a further fine-tuning step specific to the semantic similarity among log keys would likely improve the semantic embedding quality even further, there is no known labeled dataset for this task, and creating labeled sentence pair data for semantic similarity is an arduous task. The output of this embedding model is a latent representation of the natural language contained within the log key of shape (768, 1) where the distance in the latent space between semantically similar sentences is close together, and those which are semantically diverse are far apart. ### _Sentiment Prediction_ While the semantic embedding model is valuable in that it creates a latent space where semantically (language-wise) similar sentences are close together and diverse ones are far apart, which can be used to cluster log keys reporting on the same or similar processes, these embeddings sometimes miss a valuable feature of natural language which is often included in logs. That feature is sentiment. A simple example of how classifying sentiment can add value to a semantic embedding is evident in the following two HDFS logs: Exception in receiveBlock for block and Receiving empty packet for block. These two logs are similar in terms of semantics using our pre-trained model because of their shared key words. However, the first one indicates an anomaly while the second does not. This presents a challenge when clustering solely based on semantic embedding. To work around this kind of issue, we follow the idea of our recent SentiLog work which leverages the sentiment of file system logs [35]. Specifically, we reuse the pre-trained sentiment classification model from SentiLog on a set of neutral/positive and negative log keys, and concatenate the rounded (0 or 1) output of this model to the overall embedding of the log. Adding a sentiment dimension properly helps separate logs which may be semantically similar but opposite with regard to sentiment. Ultimately, the semantic embedding, including the con catenated sentiment prediction, serves as a highly accurate latent representation of the log key which can confidently be used to cluster logs which are truly similar. ### _Clustering_ The ultimate goal of ClusterLog is to cluster the runtime logs into similar groups, so that we can use the same group ID to represent similar logs to reduce the complexity of the log sequences. Following the semantic embedding and sentiment prediction steps, and their concatenation, the entire embedding is fed into a clustering algorithm in order to group the keys. However, there are a variety of problems associated with common clustering algorithms such as K-Means clustering when applied to this task. Initially, we ruled out K-Means primarily because of the need to specify the amount of centroids, which presents a major challenge as it makes the approach highly dependent upon the training data being fully representative of the logs during deployment as new, unseen, log keys would be forced to be grouped with an existing cluster, regardless of actual semantic and sentiment similarity. To add to this, finding an optimal number of centroids in K-Means based purely on the embedding data presents a challenge in and of itself as classic methods like the Elbow method are inconclusive on out our datasets. In response to this, other clustering methodologies which create an arbitrary amount of clusters dependent on a specific distance threshold between clusters or points seem to be more suited. However, in practice, these too can be difficult to assess, as finding the correct hyper-parameters (i.e. the distance) for the given dataset is not always clear. But through extensive exploration, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) [11] gave the most promising results when applied to a variety of file systems' logs. DBSCAN is simply described as a clustering algorithm which defines data points as core points, border points, and noise points based on _Epsilon_, a distance threshold, and _minimum samples_, both of which must be manually set. The reason for choosing DBSCAN was primarily driven by the fact that it gave a more consistent, and often more valuable, insight on how to set the threshold hyper-parameter for good results. In contrast to KMeans and other classic algorithms, the epsilon parameter used by DBSCAN can be used to locate density gaps in the data as well as the amount of change required to overcome these gaps. Additionally, _Epsilon_ is highly intuitive to tune as it can be easily understood as modifying the radius around each point which is used to search for neighboring points. This means that when applied to any system, some analysis of how far apart embeddings of similar entries are can be used to gain meaningful insight as to what a good _Epsilon_ value may be. As shown in Figure 1 and 2, when plotting epsilon from 0.4 to 1 against the amount of clusters created by DBSCAN for HDFS and Fig. 1: Number of generated clusters on _Lustre_ logs using different DBSCAN threshold. Fig. 2: Number of generated clusters on _HDFS_ logs using different DBSCAN threshold. Lustre, respectively, both graphs indicate noticeable ledges where multiple steps of epsilon do not result in any change in the amount of clusters found by DBSCAN. The thresholds at or around the longest of these ledges, provide a good baseline for setting the epsilon hyperparameter. Using a very minimal training and testing set allows us to verify such hyperparameter values. With regard to the second hyperparameter which was mentioned, _minimum samples_, it was set to a constant value of 1. By setting minimum samples to 1, DBSCAN does not predict any of the data points as noise, but rather, any outliers simply become a unique cluster of just themselves. This was done for the simple reason that labeling all outlier points as - 1 to indicate noise effectively creates a meaningless cluster of completely unrelated log keys which are only grouped because they are too far from others. This severely impacts the ability for the sequential analysis algorithm to properly predict anomalies. ### _Sequential Analysis_ The final part of ClusterLog is the downstream sequential analysis algorithm. In the current implementation, we focus on comparing with a state-of-the-art log-based anomaly detection method, DeepLog. Hence, we use DeepLog's sequence-learning part as the downstream algorithm. It is based on a standard 2 layer LSTM architecture which uses ten sequential log keys as input to predict the probability distribution of the next log key. From this probability distribution, log keys within the top 9 are treated as part of a normal sequence, while other log keys in this distribution are treated as anomalies. ## III Evaluation To evaluate the performance of ClusterLog, we conducted evaluations on two vastly different distributed storage systems: Apache HDFS and Lustre. The combination of these different evaluations demonstrates the generality of ClusterLog on both noisy Lustre logs and more structured HDFS logs. We will show that our approach will outperform existing solutions in both areas through granularity reduction. ### _Lustre_ #### Iii-A1 Dataset Due to the lack of publicly available anomaly detection datasets for Parallel File Systems, the dataset utilized in this paper was generated and labeled via a process of fault injection. Specifically, the fault injection was simulated using PFault [6], an open-source fault injection repository for Parallel File Systems. The labeling of this data was done by treating data before fault injection as normal and having domain experts manually label the data following a fault injection to ensure its relevance with the injected anomaly. This process resulted in a dataset containing 150,473 normal logs and 7,401 abnormal, anomalous logs. Additionally, the total number of unique log key templates in this dataset equated to 73. #### Iii-A2 Training and Testing The training setup for evaluating ClusterLog against the Lustre dataset can be simply described as learning a portion of the normal sequence of logs (without anomalies) to show the sequence model how normal logs look, and then using that knowledge to see if anomalies can be detected in a sequence of logs containing both normal and anomalous logs. Both the train and test set are created by forming a sliding window of size 10 among the log key sequence, where the goal is to create a probability distribution for the next log key. If the next log key is not within the top candidates of the probability distribution it is counted as an anomaly while testing. A distinction to this setup is made when comparing it with NeuralLog as NeuralLog trains on a set of both normal and anomalous logs. Additionally, NeuralLog does not predict the next log in a sequence, but rather classifies a given sequence as anomalous or not. Based on these approaches, Accuracy, Precision, Recall, and F-measure are calculated. While most of the training and testing analysis was done using 25 percent of the total normal logs, in order to test and compare the limits of generalizability, a further test was carried out in which a smaller amount of the entire dataset was used to train the sequence model. This test was carried out by using just 1 percent of the entire Lustre training set. This split resulted in 1,500 total samples. ### _HDFS_ #### Iv-B1 Dataset Among the majority of recent anomaly detection works, the same HDFS dataset is most commonly referenced in their respective evaluations. This makes it a good benchmark comparison for new results. In contrast to logs contained in the Lustre log dataset, HDFS logs contain a specific block ID in addition to their log content which allows them to be grouped by session. Anomalies are labeled on a per-session basis rather than a per-log basis for this dataset. The labeling itself was carried out by domain experts who discovered and labeled sessions containing one of 11 types of anomalies. This resulted in a dataset with 11,197,954 messages, grouped into 575,139 sessions, of which 16,808 sessions were anomalous [32]. #### Iv-B2 Training and Testing During training, the sequence model utilizes only the normal log sessions, and each key in each session is represented by the corresponding cluster calculated by DBSCAN. Much like what was done with Lustre, the sequence model learns normal behavior by using a window of 10 cluster IDs. During testing, the trained sequence model is run against both the normal and the anomalous log sessions. For each session, the sequence model utilizes the window of 10 keys to predicting a probability distribution for the next ID in the session. As opposed to Lustre, if the model predicts an anomaly, the session is labeled as an anomaly, instead of just the log key. If no anomalies are predicted within a given session, the session itself is predicted as being normal. This setup varied slightly in the case of Neuralog, as it did for Lustre, as Neuralog views each entire block and classifies the block as anomalous or normal. ### _Results Analysis_ #### Iv-C1 ClusterLog performance analysis with different settings and hyper-parameters The first set of results, shown in Figure 3, represents the detailed performance of ClusterLog: 1) using numerous values for DBSCAN's epsilon parameter; 2) using sentiment dimension or not; 3) using variable or fixed number of candidates. The leftmost graph in both of these rows shows the results of ClusterLog without the concatenation of the sentiment prediction to the semantic embedding. The middle graph shows ClusterLog's performance with the concatenation of the sentiment prediction to the semantic embedding. The right-most graph shows ClusterLog's performance with the concatenation of the sentiment prediction as well as using a variable number of candidates instead of a fixed number in the sequence model. Specifically, when the number of clusters descends below 27, the number of candidates used in the sequence model is set to the floor division of the number of clusters by 3 (eg. for 20 clusters, candidates would be set to 6). This feature was added to ClusterLog based on the intuition that when the amount of clusters becomes very low, using a relatively large amount of candidates may force the sequence model to consider a key of very low calculated probability to be normal, increasing the amount of false negatives. From these results, we can clearly observe not only how ClusterLog can achieve its best results by reducing a large amount of noise with high clustering thresholds, but also that the concatenation of sentiment as well as the adaptation of candidates to account for a low cluster count add to ClusterLog's ability to produce strong and consistent results. #### Iv-C2 Overall Comparison with DeepLog and NeuralLog In order to provide a reference among state-of-the-art anomaly detection solutions, we compared ClusterLog's results with the results from DeepLog using the most similar training and testing setup possible. For Lustre, a sliding window of size 10 was used to predict a probability distribution among the 73 unique log key templates where a succeeding log key within the top 9 of predictions was considered normal and anything else was considered anomalous. For HDFS, the setup was the same, but because HDFS logs can be grouped into sessions using their block IDs, each session was individually used as a sequence, and the prediction of an anomaly within a session classified the entire session as anomalous. The results shown in Figure 4 and 5 for HDFS and Lustre respectively, show ClusterLog's improvement upon previous SOTA results. In the case of HDFS, while DeepLog maintains a slightly higher Precision, ClusterLog boasts a more significant improvement in Recall, primarily because clustering the most similar logs on the HDFS dataset allows ClusterLog to more easily detect what is truly a normal sequence amidst some noise. NeuralLog also has high precision and recall, but both of these values seem to be slightly outshined by ClusterLog and DeepLog on this train and test split. Overall, ClusterLog's very high precision and high recall boost its combined F1 score slightly higher than Deeplog's and NeuralLog's. However, due to the fact that the proposed method is not able to detect the correct F1 score, it is not possible to detect the correct F1 score. Fig. 4: Performance comparison of ClusterLog, DeepLog, and NeuralLog on _HDFS_ dataset Fig. 5: Performance comparison of ClusterLog, DeepLog and NeuralLog on _Lustre_ dataset Fig. 3: number of clusters (right \(y\)-axis) and F-score of the prediction (left \(y\)-axis) of ClusterLog under different parameter settings for _Lustre_ (top) and _HDFS_ (bottom). Adding sentiment dimension (the middle figure) shows a clear improvement in F1 scores in a larger range of DBSCAN thresholds. Also, variable number of candidates further improves the F1 scores. ability to group HDFS logs into sessions by their block ID, the nature of these sessions is already fairly low in noise. This means the application of ClusterLog will still be effective, as is shown, but the difference in performance will likely not be as large as it may be for more noisy systems. Because Lustre is a good example of one such noisy system, the comparison of results on this dataset provides evidence of this larger performance gap. As shown by Figure 5, the disparity in precision between DeepLog and ClusterLog is very large while it seems both approaches have a very high recall. This is because DeepLog is not able to learn the more noisy sequence effectively, and as a result classifies the vast majority of normal and abnormal logs as anomalous. NeuralLog does a much better job of holding up against this dataset as its vector representations for logs are more robust than DeepLog's log indexes for noisy data, but as shown later, it takes much more labeled data to extract good performance from NeuralLog due to a very large set of trainable weights. ClusterLog was able to learn the sequence and accurately discern between normal and abnormal logs. In total, the results on both of these datasets in comparison to modern approaches verify ClusterLog's ability to generalize across different file systems and achieve strong results. #### Iii-B3 The impact of training dataset size using ClusterLog The final comparison for ClusterLog represents the impact of using a small training set on the ability of ClusterLog, DeepLog and NeuralLog to accurately predict anomalies on the more noisy Lustre dataset. This comparison was done on the Lustre dataset because the difference in the number of clusters used by ClusterLog and DeepLog on this dataset is large, so the impact of changing the training set size should be emphasized by this fact. As shown in Figure 6, the prediction results for DeepLog and NeuralLog are massively impacted as the proportion of the training set used shrinks. The alternative approaches' low performance on such a small dataset can be explained by the fact that the normal interactions between a higher granularity of keys are unlikely to be explained by such a small dataset. In contrast, the figure shows that ClusterLog can maintain the same level of results even having trained on just 1 percent of the training set (1504 logs). By reducing the number of keys through clustering, ClusterLog is much more likely to not only see all of the keys enough times but also learn the normal interactions between them, ultimately leading to much stronger results on small datasets. ## IV Related Works There is a large catalogue of both preprocessing and anomaly detection approaches for file system logs in existence today [31, 30, 24, 13, 29, 27, 23, 37, 8, 26, 16, 25, 34, 10, 7, 19, 12, 22, 32, 33, 21, 20, 9]. Among preprocessing techniques, the primary approaches include frequent pattern mining, clustering, and heuristic methods. Frequent pattern mining approaches extract frequently occurring patterns, with the ultimate goal of grouping the logs which have the highest degree of similarity [30, 31, 24]. The clustering-based preprocessing techniques employ clustering algorithms to group similar logs. The individual approaches in this category generally differ in their approach towards measuring similarity, as some may employ weighted edit differences on log pairs [13], while others may fit a given set of raw logs to a predefined set of events [29]. The final category of preprocessing approaches is heuristic-based approaches. Among these, Drain [17] is the most popular and relies on a fixed depth tree to Fig. 6: Performance comparison between ClusterLog, DeepLog, and NeuralLog using smaller training data. ClusterLog is much more stable even with less training data, showing the effectiveness of clustering logs. extract log templates. ClusterLog belongs to the clustering-based method with new designs using semantic and sentimental features of logs. Anomaly detection approaches can be categorized into non-machine learning and machine learning approaches. Of the more successful non-machine learning approaches, the rule-based approaches are well seen, where domain experts formulate a set of rules which can be used to classify logs as anomalies [8, 26, 16, 25, 10]. Among machine learning techniques, pattern-seeking [32, 33] and invariant mining [21] based algorithms have proven effective on a variety of file system logs, but their results do not hold up on the irregularity of logs which do not include session identifiers [35]. Additionally, These approaches do not hold up to more recent deep learning-based solutions, which learn sequences of logs using deep neural networks. The first approach, LogAnomaly[22], borrows from NLP by proposing a log key embedding technique which vectorizes log keys based on synonyms and antonyms and calculates similarity between vector embeddings to predict anomalies. An additional approach in this domain, LogBERT[14], utilizes the Bidirectional Encoders Represented by Transformers model which has provided STOA Results in multiple domains. In this approach, it is not the next log key that is predicted, but rather a given sequence that is masked and then classified as normal or abnormal. More recently, similar attention-based approaches have shown very strong results across a variety of Distributed File Systems. Among these approaches, LogRobust[36] and Nueralog[18] show the most promising results. While the encapsulated models are slightly different, both of these approaches utilize attention mechanisms to classify a sequence of logs represented by content embedding vectors. These approaches lend well to unseen logs. However, training accurate classification on such high dimensional embedding input requires a large amount of labeled training data. Additionally, the lack of grouping characteristics such as sequence identifiers in some logs and the resultant irregularity makes it difficult for these solutions to be effectively applied. ClusterLog is proposed to better work with these advanced models to achieve better performance and higher training efficiency. ## V Conclusion and Future Work In this work, through ClusterLog, we have shown that granularity reduction in log key sequences is a viable approach to improve log anomaly detection in general. A key area where this method of anomaly detection outperforms previous approaches is in non-session parallel file system logs, which are highly irregular and noisy. Additionally, this work shows that granularity reduction in log sequences may allow users to reduce their effort by allowing for more lenient pre-processing, as well as smaller labeled datasets. This all is because of the effective reduction of noise and retention of important sequence information achieved by a good clustered representation of the file system log keys. In the future, there are two primary areas in this work which are intended to be improved and many others which can be explored. To begin, due to ClusterLog's high dependence on the accuracy of sentence and sentiment embedding models, improvements in both of these areas could prove to be of value. The first direction of improvement is further exploration of clustering techniques to find an even more clear and easy-to-apply way of selecting the right clustering hyperparameters. The second direction is to further provide evidence for CLusterLog's generalizability by applying it to a larger scope of log systems. In addition to more parallel and distributed file systems, ClusterLog should prove its viability against other systems such as Kubernetes.

2305.04660

Rotational Slippage Prediction from Segmentation of Tactile Images

Adding tactile sensors to a robotic system is becoming a common practice to achieve more complex manipulation skills than those robotics systems that only use external cameras to manipulate objects. The key of tactile sensors is that they provide extra information about the physical properties of the grasping. In this paper, we implemented a system to predict and quantify the rotational slippage of objects in hand using the vision-based tactile sensor known as Digit. Our system comprises a neural network that obtains the segmented contact region (object-sensor), to later calculate the slippage rotation angle from this region using a thinning algorithm. Besides, we created our own tactile segmentation dataset, which is the first one in the literature as far as we are concerned, to train and evaluate our neural network, obtaining results of 95% and 91% in Dice and IoU metrics. In real-scenario experiments, our system is able to predict rotational slippage with a maximum mean rotational error of 3 degrees with previously unseen objects. Thus, our system can be used to prevent an object from falling due to its slippage.

Julio Castaño-Amoros, Pablo Gil

2023-05-08T12:23:47Z

http://arxiv.org/abs/2305.04660v1

# Rotational Slippage Prediction from Segmentation of Tactile Images ###### Abstract Adding tactile sensors to a robotic system is becoming a common practice to achieve more complex manipulation skills than those robotics systems that only use external cameras to manipulate objects. The key of tactile sensors is that they provide extra information about the physical properties of the grasping. In this paper, we implemented a system to predict and quantify the rotational slippage of objects in hand using the vision-based tactile sensor known as Digit. Our system comprises a neural network that obtains the segmented contact region (object-sensor), to later calculate the slippage rotation angle from this region using a thinning algorithm. Besides, we created our own tactile segmentation dataset, which is the first one in the literature as far as we are concerned, to train and evaluate our neural network, obtaining results of 95% and 91% in Dice and IoU metrics. In real-scenario experiments, our system is able to predict rotational slippage with a maximum mean rotational error of 3 degrees with previously unseen objects. Thus, our system can be used to prevent an object from falling due to its slippage. ## I Introduction and Related Work Traditionally, the methods to carry out robotic manipulation tasks used 2D or 3D vision sensors [1], which only take into account the geometric properties of the objects to perform the grasping. In contrast, with tactile sensors, it is possible to measure and react to physical properties (mass distribution, center of gravity or friction) in order to achieve a stable grasping [2]. In the last twenty years, several tactile sensors have been designed using different hardware technologies [3], although the last trend of tactile sensors lies in optical tactile sensors [4]. In this manuscript, we present an algorithm to estimate the rotation angle of an object which is being manipulated when slippage occurs. This method is based on segmentation neural networks to obtain the contact region (object-sensor) and traditional computer vision techniques to calculate the rotation angle and is applied to the vision-based tactile sensor Digit [5] which does not contain visual markers to keep low its cost. Estimating the contact region between the robot's fingertips and the grasped object has been attempted to be solved in different ways. For example, by subtracting contact and no-contact tactile images [6], detecting and grouping visual markers [7], throughout 3D reconstruction and photometric algorithms [8], or using neural networks [9, 10]. In contrast, although our work is inspired by these previous articles, the main differences lie in the fact that we use the Digit sensors, without markers [7], which do not produce depth information [8], and state-of-the-art segmentation neural networks, which are more robust than subtracting operations [6] and vanilla CNN [9], and its training is more stable compared with GAN's training [10]. Slippage is a common physical event that occurs during object manipulation, that has been tried to solve for several years employing different approaches. For example, detecting binary slippage events with traditional image preprocessing techniques [11], combining convolutional and recurrent neural networks to classify slip in clockwise and counterclockwise rotation [12] or estimating the slip rotation angle using vision-based tactile sensors with markers [13] or force/torque sensors [14]. In this paper, we have inspired our work in these methods that characterize and quantify the rotational slip. ## II Method We propose a two-stage method for touch region segmentation and rotational slippage prediction. The first stage of our method is based on a segmentation neural network applied to vision-based tactile sensing, which we called Tactile Segmentation Neural Network (TSNN). In this work, our goal is only to segment the contact region, then we decided to use DeepLabV3+ [15] architecture for experimentation. DeepLabV3+ is well-known for using an encoder-decoder architecture to perform image segmentation, and for introducing a new layer in its architecture, which is a combination of atrous or dilated and depth-wise separable convolutions. This combination leads to a reduction of computational complexity while maintaining similar or even better performance than previous versions. As the encoder, the authors used a modified version of the architecture Xception, called Aligned Xception, which replaces all the max pooling layers by the depth-wise separable convolutions to perform the feature extraction procedure. The decoder, in contrast, is a simpler part of the architecture, which only comprises convolution, concatenation, and upsampling layers to transform the intermediate features into the output. The second stage of our method estimates the angle of rotation of the segmented region of contact using a traditional computer vision thinning algorithm (Skeleton method) [16] that blackens points in the binary contact region using an 8-square neighborhood and different connectivity conditions. Other approaches, based on different neural networks such as Unet++ [17] and PSPNet [18] or different algorithms to estimate the angle such as PCA or ellipse fitting, were tested. The complete system is shown in Fig 1. ## III Experimentation and results We have generated our own dataset as we have not found any dataset related to tactile segmentation in order to be used as the base of our experimentation. Our tactile segmentation dataset comprises 3675 tactile images with their respective labelled contact regions. We have used 16 objects from YCB dataset to record it, from which we have captured between 200 and 250 tactile images per object. The objects contain different textures, rigidity, weight, geometries, etc. To train the TSNN we use the Dice and IoU metrics, an NVIDIA A100 Tensor Core GPU with 40 GB of RAM, and the following optimal hyperparameters: a batch size of 32, a learning rate of 1e-4, the Adam optimizer, the Focal loss, and 30 training epochs. Table I shows the results obtained by DeepLabV3+ TSNN in the testing experiment. DeepLabV3+ is able to segment tactile images with high accuracy and in real-time execution. Besides, this TSNN is 3 ms faster than other segmentation neural networks (Unet++ and PSPNet) while maintaining the same performance, thus, achieving a better trade-off between segmentation accuracy and prediction time. Figure 1(a) shows different examples of contact region segmentation carried out by DeepLabV3+ TSNN, and Fig. 1(b) shows our robotic manipulation setup with a UR5 robot, two DIGIT sensors, a ROBOTIQ gripper, the object to grasp with the aruco markers attached and an Intel RealSense camera to calculate the ground truth angle. The task consists of grasping and lift an object while the tactile segmentation and rotational slippage angle are estimated. The predicted angle is calculated as the difference between the current and the initial angle obtained in the Skeleton method described earlier, while the ground truth angle is calculated using two aruco markers as visual references. Our system was evaluated with seven unseen objects (1 to 7 in Fig. 3) and two seen objects from our tactile segmentation dataset (8 and 9 in Fig. 3). The experimentation comprises 45 graspings and lifts in total (five per object) while calculating the rotational error in degrees. Figure 3 shows the mean rotational error of the 5 graspings and lifts for each object. Note that object 6 and 8 causes more error and deviation because object 6 weight's is higher compared with the rest of the objects, and object 8 contains higher curvature on its surface that causes more saturation in the sensor. Our system is able to predict rotational slippage with an overall mean rotational error of **1.854deg**\(\pm\)**0.988deg**, that is to say, a maximum mean error of 3 degrees in the worst case. Figure 4 shows some examples of the prediction of rotational slippage with four aforementioned objects. ## IV Conclusions In this paper, we propose a model-based system to predict rotational slippage during the grasping and lift of an object, achieving a mean error value of **1.854deg**\(\pm\)**0.988deg**, compared \begin{table} \begin{tabular}{|c|c|c|c|} \hline & **Dice** & **IoU** & **Time(s)** \\ \hline **DeepLabV3+** & 0.956 \(\pm\) 0.013 & 0.914 \(\pm\) 0.023 & 0.006 \(\pm\) 0.002 \\ \hline **PSPNet** & 0.951 \(\pm\) 0.014 & 0.907 \(\pm\) 0.025 & 0.006 \(\pm\) 0.002 \\ \hline \end{tabular} \end{table} TABLE I: DeepLabV3+ TSNN results in terms of Dice, IoU and inference time metrics, and using the backbone ResNet18 Fig. 1: Diagram of our system combining both stages Fig. 2: a) Examples of rotation angle calculation for slipping during lift task: DIGIT image (first row), ground truth (second row), prediction (third row), b) Robotic manipulation setup with different objects with the error of **3.96\({}^{\circ}\)\(\pm\) UNK** from [13], and the error of **4.39\({}^{\circ}\)\(\pm\) 0.18\({}^{\circ}\)** from [14]. Although we could not carry out an experimental comparison because we do not have their sensors available, some objects were used both in this work and in theirs. Our system also has some limitations regarding the shape of the contact region. If this shape is similar to a circle, it becomes impossible to calculate its rotation movement. In that case, we propose to grasp the object by surfaces with small curvature.

2310.11965

Filling in the Gaps: Efficient Event Coreference Resolution using Graph Autoencoder Networks

We introduce a novel and efficient method for Event Coreference Resolution (ECR) applied to a lower-resourced language domain. By framing ECR as a graph reconstruction task, we are able to combine deep semantic embeddings with structural coreference chain knowledge to create a parameter-efficient family of Graph Autoencoder models (GAE). Our method significantly outperforms classical mention-pair methods on a large Dutch event coreference corpus in terms of overall score, efficiency and training speed. Additionally, we show that our models are consistently able to classify more difficult coreference links and are far more robust in low-data settings when compared to transformer-based mention-pair coreference algorithms.

Loic De Langhe, Orphée De Clercq, Veronique Hoste

2023-10-18T13:44:58Z

http://arxiv.org/abs/2310.11965v1

# Filling in the Gaps: ###### Abstract We introduce a novel and efficient method for Event Coreference Resolution (ECR) applied to a lower-resourced language domain. By framing ECR as a graph reconstruction task, we are able to combine deep semantic embeddings with structural coreference chain knowledge to create a parameter-efficient family of Graph Autoencoder models (GAE). Our method significantly outperforms classical mention-pair methods on a large Dutch event coreference corpus in terms of overall score, efficiency and training speed. Additionally, we show that our models are consistently able to classify more difficult coreference links and are far more robust in low-data settings when compared to transformer-based mention-pair coreference algorithms. ## 1 Introduction Event coreference resolution (ECR) is a discourse-centered NLP task in which the goal is to determine whether or not two textual events refer to the same real-life or fictional event. While this is a fairly easy task for human readers, it is far more complicated for AI algorithms, which often do not have access to the extra-linguistic knowledge or discourse structure overview that is required to successfully connect these events. Nonetheless ECR, especially when considering cross-documents settings, holds interesting potential for a large variety of practical NLP applications such as summarization Liu and Lapata (2019), information extraction Humphreys et al. (1997) and content-based news recommendation Vermeulen (2018). However, despite the many potential avenues for ECR, the task remains highly understudied for comparatively lower-resourced languages. Furthermore, in spite of significant strides made since the advent of transformer-based coreference systems, a growing number of studies has questioned the effectiveness of such models. It has been suggested that classification decisions are still primarily based on the surface-level lexical similarity between the textual spans of event mentions Ahmed et al. (2023); De Langhe et al. (2023), while this is far from the only aspect that should be considered in the classification decision. Concretely, in many models coreferential links are assigned between similar mentions even when they are not coreferent, leading to a significant number of false positive classifications, such as between Examples 1 and 2. 1. The French president Macron met with the American president for the first time today 2. French President Sarkozy met the American president We believe that the fundamental problem with this method stems from the fact that in most cases events are only compared in a pairwise manner and not as part of a larger coreference chain. The evidence that transformer-based coreference resolution is primarily based on superficial similarity leads us to believe that the current pairwise classification paradigm for transformer-based event coreference is highly inefficient, especially for studies in lower-resourced languages where the state of the art still often relies on the costly process of fine-tuning large monolingual BERT-like models De Langhe et al. (2022). In this paper we aim to both address the lack of studies in comparatively lower-resourced languages, as well as the more fundamental concerns w.r.t. the task outlined above. We frame ECR as a graph reconstruction task and introduce a family of graph autoencoder models which consistently outperforms the traditional transformer-based methods on a large Dutch ECR corpus, both in terms of accuracy and efficiency. Additionally, we introduce a language-agnostic model variant which disregards the use of semantic features entirely and even outperforms transformer-based classification in some situations. Quantitative analysis reveals that the lightweight autoencoder models can consistently classify more difficult mentions (cfr. Examples 1 and 2) and are far more robust in low-data settings compared to traditional mention-pair algorithms. ## 2 Related Work ### Event Coreference Resolution The primary paradigm for event coreference resolution takes the form of a binary mention-pair approach. This method generates all possible event pairs and reduces the classification to a binary decision (coreferent or not) between each event pair. A large variety of classical machine learning algorithms has been tested using the mention-pair paradigm such as decision trees Cybulska and Vossen (2015), support vector machines Chen et al. (2015) and standard deep neural networks Nguyen et al. (2016). More recent work has focused on the use of LLMs and transformer encoders Cattan et al. (2021), with span-based architectures attaining the best overall results Joshi et al. (2020); Lu and Ng (2021). It has to be noted that mention-pair approaches relying on LLMs suffer most from the limitations discussed in Section 1. In an effort to mitigate these issues some studies have sought to move away from the pairwise computation of coreference by modelling coreference chains as graphs instead. These methods' primary goal is to create a structurally-informed representation of the coreference chains by integrating the overall document Fan et al. (2022); Tran et al. (2021) or discourse Huang et al. (2022) structure. Other graph-based methods have focused on commonsense reasoning Wu et al. (2022). Research for comparatively lower-resourced languages has generally followed the paradigms and methods described above and has focused on languages such as Chinese Mitamura et al. (2015), Arabic NIST (2005) and Dutch Minard et al. (2016). ### Graph Autoencoders Graph Autoencoder models were introduced by Kipf and Welling (2016) as an efficient method for graph reconstruction tasks. The original paper introduces both variational graph autoencoders (VGAE) and non-probabilistic graph autoencoders (GAE) networks. The models are parameterized by a 2-layer graph-convolutional network (GCN) Kipf and Welling (2016) encoder and a generative inner-product decoder between the latent variables. While initially conceived as lightweight models for citation network prediction tasks, both the VGAE and GAE have been successfully applied to a wide variety of applications such as molecule design Liu et al. (2018), social network relational learning Yang et al. (2020) and 3D scene generation Chattopadhyay et al. (2023). Despite their apparent potential for effectively processing large amounts of graph-structured data, application within the field of NLP has been limited to a number of studies in unsupervised relational learning Li et al. (2020). ## 3 Experiments ### Data Our data consists of the Dutch ENCORE corpus De Langhe et al. (2022), which in its totality consists of 12,875 annotated events spread over 1,015 documents that were sourced from a collection of Dutch (Flemish) newspaper articles. Coreferential relations between events were annotated at the within-document and cross-document level. ### Experimental Setup #### 3.2.1 Baseline Coreference Model Our baseline model consists of the Dutch monolingual BERTje model de Vries et al. (2019) fine-tuned for cross-document ECR. First, each possible event pair in the data is encoded by concatenating the two events and by subsequently feeding these to the BERTje encoder. We use the token representation of the classification token _[CLS]_ as the aggregate embedding of each event pair, which is subsequently passed to a softmax-activated classification function. Finally, the results of the text pair classification are passed through a standard agglomerative clustering algorithm Kenyon-Dean et al. (2018); Barhom et al. (2019) in order to obtain output in the form of coreference chains. We also train two parameter-efficient versions of this baseline model using the distilled Dutch Language model RobBERTje Delobelle et al. (2022) and a standard BERTje model trained with bottleneck adapters Pfeiffer et al. (2020). #### 3.2.2 Graph Autoencoder Model We make the assumption that a coreference chain can be represented by an undirected, unweighted graph \(\mathcal{G}=(V,E)\) with \(|V|\) nodes, where each node represents an event and each edge \(e\in E\) between two nodes denotes a coreferential link between those events. We frame ECR as a graph reconstruction task where a partially masked adjacency matrix \(A\) and a node-feature matrix \(X\) are used to predict all original edges in the graph. We employ both the VGAE and GAE models discussed in Section 2.2. In a non-probabilistic setting (GAE) the coreference graph is obtained by passing the adjacency matrix \(A\) and node-feature matrix \(X\) through a Graph Convolutional Neural Network (GCN) encoder and then computing the reconstructed matrix \(\hat{A}\) from the latent embeddings \(Z\): \[Z=GCN(X,A) \tag{1}\] \[\hat{A}=\sigma(ZZ^{\tau}) \tag{2}\] For a detailed overview of the (probabilistic) variational graph autoencoder we refer the reader to the original paper by Kipf and Welling (2016). Our experiments are performed in a cross-document setting, meaning that the input adjacency matrix \(A\) contains all events in the ENCORE dataset. Following the original approach by Kipf and Welling (2016) we mask 15% of the edges, 5% to be used for validation and the remaining 10% for testing. An equal amount of non-edges is randomly sampled from \(A\) to balance the validation and test data. We extract masked edges and non-edges and use them to build the training, validation and test sets for the mention-pair baseline models detailed above, ensuring that both the mention-pair and graph autoencoder models have access to exactly the same data for training, validation and testing. We define the encoder network with a 64-dimension hidden layer and 32-dimension latent variables. For all experiments we train for a total duration of 200 epochs using an Adam optimizer (Kingma and Ba, 2014) and a learning rate of 0.001. We construct node features through Dutch monolingual transformer models by average-pooling token representations for each token in the event span in the models' final hidden layer, resulting in a 768-dimensional feature vector for each node in the graph. For this we use the Dutch BERTje model (de Vries et al., 2019), a Dutch sentence-BERT model (Reimers and Gurevych, 2019) and the Dutch RoBERTa-based RobBERT model (Delobelle et al., 2020). Additionally, we create a second feature set for the BERTje and RobBERT models where each event is represented by the concatenation of the last 4 layers' average-pooled token representations Devlin et al. (2018). This in turn results in a 3072-dimensional feature vector. Finally, we also evaluate a language-agnostic featureless model where \(X\) is represented by the identity matrix of \(A\). #### 3.2.3 Hardware Specifications The baseline coreference algorithms were trained and evaluated on 2 Tesla V100-SXM2-16GB GPUs. Due to GPU memory constraints, the Graph encoder models were all trained and evaluated on a single 2.6 GHz 6-Core Intel Core i7 CPU. ## 4 Results and Discussion Results from our experiments are disclosed in Table 1. Results are reported through the CONLL F1 metric, an average of 3 commonly used metrics for coreference evaluation: MUC (Vilain et al., 1995), \begin{table} \begin{tabular}{l c c c c c} **Model** & **CONLL F1** & **Training Runtime (s)** & **Inference Runtime (s)** & **Trainable Parameters** & **Disk Space (MB)** \\ \hline MP RobBERTje & 0.767 & 7962 & 16.31 & 74M & 297 \\ MP BERTje\({}_{\text{e}ADPT}\) & 0.780 & 12 206 & 20.61 & 0.9M & 3.5 \\ MP BERTje & 0.799 & 9737 & 21.78 & 110M & 426 \\ \hline GAE NoFeatures & 0.832 \(\pm\) 0.008 & 1006 & 0.134 & 825856 & 3.2 \\ GAE BERTje\({}_{\text{708}}\) & 0.835 \(\pm\) 0.010 & 975 & 0.263 & 51200 & 0.204 \\ GAE BERTje\({}_{\text{9072}}\) & **0.852 \(\pm\) 0.006** & 1055 & 0.294 & 198656 & 0.780 \\ GAE RobBERT\({}_{\text{708}}\) & 0.838 \(\pm\) 0.004 & 1006 & 0.273 & 51200 & 0.204 \\ GAE RobBERT\({}_{\text{3072}}\) & 0.841 \(\pm\) 0.007 & 1204 & 0.292 & 198656 & 0.780 \\ GAE SBERT & 0.801 \(\pm\) 0.002 & 982 & 0.291 & 51200 & 0.204 \\ \hline VGAE NoFeatures & 0.824 \(\pm\) 0.009 & 1053 & 0.139 & 827904 & 3.2 \\ VGAE BERTje\({}_{\text{768}}\) & 0.822 \(\pm\) 0.011 & 1233 & 0.282 & 53248 & 0.212 \\ VGAE BERTje\({}_{\text{3072}}\) & 0.842 \(\pm\) 0.009 & 1146 & 0.324 & 200704 & 0.788 \\ VGAE RobBERT\({}_{\text{768}}\) & 0.828 \(\pm\) 0.0021 & 1141 & 0.288 & 53248 & 0.212 \\ VGAE RobBERT\({}_{\text{3072}}\) & 0.831 \(\pm\) 0.004 & 1209 & 0.301 & 200704 & 0.788 \\ VGAE SBERT & 0.773 \(\pm\) 0.012 & 1185 & 0.295 & 53248 & 0.212 \\ \end{tabular} \end{table} Table 1: Results for the cross-document event coreference task. We report the average CONLL score and standard deviation over 3 training runs with different random seed initialization for the GCN weight matrices (GAE/VAE) and classification heads (Mention-Pair models). Inference runtime is reported for the entire test set. B\({}^{3}\)[11] and CEAF [12]. We find that the graph autoencoder models consistently outperform the traditional mention-pair approach. Moreover, we find the autoencoder approach significantly reduces model size, training time and inference speed even when compared to parameter-efficient transformer-based methods. We note that the VGAE models perform slightly worse compared to their non-probabilistic counterparts, which is contrary to the findings in Kipf and Welling (2016). This can be explained by the use of more complex acyclic graph data in the original paper. In this more uncertain context, probabilistic models would likely perform better. As a means of quantitative error analysis, we report the average Levenshtein distance between two event spans for the True Positive (TP) pairs in our test set in Figure 1. Logically, if graph-based models are able to better classify harder (i.e non-similar) edges, the average Levenstein distance for predicted TP edges should be higher than for the mention-pair models. For readability's sake we only include results for the best performing GAE-class models. A more detailed table can be found in the Appendix. We find that the average distance between TP pairs increases for our introduced graph models, indicating that graph-based models can, to some extent, mitigate the pitfalls of mention-pair methodologies as discussed in Section 1. ## 5 Ablation Studies We gauge the robustness of the graph-based models in low-data settings by re-running the original experiment and continually reducing the available training data by increments of 10%. Figure 2 shows the CONLL F1 score for each of the models with respect to the available training data size. Also here, only the best-performing GAE-class models are visualized and an overview of all models' performance can be found in the Appendix. Surprisingly, we find that training the model on as little as 5% of the total amount of edges in the dataset can already lead to satisfactory results. Logically, feature-less models suffer from a significant drop in performance when available training data is reduced. We also find that the overall drop in performance is far greater for the traditional mention-pair model than it is for the feature-based GAE-class models in low-data settings. Overall, we conclude that the introduced family of models can be a lightweight and stable alternative to traditional mention-pair coreference models, even in settings with little to no available training data. ## 6 Conclusion We show that ECR through graph autoencoders significantly outperforms traditional mention-pair approaches in terms of performance, speed and model size in settings where coreference chains are at least partially known. Our method provides a fast and lightweight approach for processing large cross-document collections of event data. Additionally, our analysis shows that combining BERT-like embeddings and structural knowledge of coreference chains mitigates the issues in mention-pair classification w.r.t the dependence on surface-form lexical similarity. Our ablation experiments reveal that only a very small number of training edges is needed to obtain satisfactory performance. Future work will explore the possibility of combining mention-pair models with the proposed graph autoencoder approach in a pipeline setting in order to make it possible to employ graph reconstruction models in settings where initially all edges in the graph are unknown. Additionally, we aim to perform more fine-grained analyses, both quantitative and qualitative, regarding the type of errors made by graph-based coreference models. Figure 1: Average Levenshtein distance for True Positive (TP) classifications across all models Figure 2: CONLL F1 performance with respect to the available training data. ## 7 Limitations We identify two possible limitations with the work presented above. First, by framing coreference resolution as a graph reconstruction task we assume that at least some coreference links in the cross-document graph are available to train on. However, we note that this issue can in part be mitigated by a simple exact match heuristic for event spans on unlabeled data. Moreover, in most application settings it is not inconceivable that at least a partial graph is available. A second limitation stems from the fact that we modelled coreference chains as undirected graphs. It could be argued that some coreferential relationships such as pronominal anaphora could be more accurately modelled using directed graphs instead.

2301.06580

Construction and Analysis of a Discrete Heat Equation Using Dynamic Consistency: The Meso-scale Limit

We present and analyze a new derivation of the meso-level behavior of a discrete microscopic model of heat transfer. This construction is based on the principle of dynamic consistency. Our work reproduces and corrects, when needed, all the major previous expressions which provide modifications to the standard heat PDE. However, unlike earlier efforts, we do not allow the microscopic level parameters to have zero limiting values. We also give insight into the difficulties of constructing physically valid heat equations within the framework of the general mathematically inequivalent of difference and differential equations.

Ronald E. Mickens, Talitha Washington

2023-01-16T19:36:22Z

http://arxiv.org/abs/2301.06580v1

Construction and analysis of a discrete heat equation using dynamic consistency: the meso-scale limit ###### Abstract. We present and analyze a new derivation of the meso-level behavior of a discrete microscopic model of heat transfer. This construction is based on the principle of dynamic consistency. Our work reproduces and corrects, when needed, all the major previous expressions which provide modifications to the standard heat PDE. However, unlike earlier efforts, we do not allow the microscopic level parameters to have zero limiting values. We also give insight into the difficulties of constructing physically valid heat equations within the framework of the general mathematically inequivalent of difference and differential equations. Key words and phrases:Heat equation; Dynamic consistency; Random walk; Asymptotic analysis; Continuum limit 1991 Mathematics Subject Classification: 35K05; 39A14 ## 1. Introduction The purpose of this paper is to examine the meso-scale limit of a discrete micro-level mathematical model constructed to represent simple heat transfer. The creation of the microscopic model is based on the application of the principle of dynamic consistency (Mickens (2005, 2015, 2021)). Under the appropriate mathematical assumptions, we are able to obtain the previous results of Maxwell (1867), Cattaneo (1948), and Vernotte (1958) regarding the replacement of the standard heat equation \[u_{t}=Du_{xx}, \tag{1}\] by the generalization \[\tau u_{tt}+u_{t}=Du_{xx}, \tag{2}\] where \(D\) is the temperature diffusion constant, \(\tau\) is a time-lag parameter, and \[u_{t}=\frac{\partial u}{\partial t},\quad u_{tt}=\frac{\partial^{2}u}{ \partial t^{2}},\quad u_{x}=\frac{\partial u}{\partial x},\quad u_{xx}=\frac{ \partial^{2}u}{\partial x^{2}}.\] Note that \((x,t)\) are the one-dimensional space and time independent variables, and \(u=u(x,t)\) is defined over appropriate intervals of \((x,t)\) with suitable boundary conditions and initial values. The need for a generalization of Eq. (1) comes from the fact that the solutions of Eq. (1) transmit information at an infinite speed (Ali et al. (2005); Christov et al. (2005); Dreher et al. (2009); Guyer et al. (1966); Joseph et al. (1989)), a condition which violates the principle of causality (Ali et al. (2005); Christov et al. (2005); Dreher et al. (2009); Guyer et al. (1966); Joseph et al. (1989)). For convenience, we now provide a resume of our major results:

2305.14473

Weighted maximal inequalities on hyperbolic spaces

In this work we develop a weight theory in the setting of hyperbolic spaces. Our starting point is a variant of the well-known endpoint Fefferman-Stein inequality for the centered Hardy-Littlewood maximal function. This inequality generalizes, in the hyperbolic setting, the weak $(1,1)$ estimates obtained by Str\"omberg in "Weak type L1 estimates for maximal functions on noncompact symmetric spaces", Ann. of Math. 114 (1981), where Str\"omberg answered a question posed by Stein and Wainger in "Problems in harmonic analysis related to curvature", Bull. Amer. Math. Soc. 84 (1978). Our approach is based on a combination of geometrical arguments and the techniques used in the discrete setting of regular trees by Naor and Tao in "Random martingales and localization of maximal inequalities", J. Funct. Anal. 259 (2010). This variant of the Fefferman-Stein inequality paves the road to weighted estimates for the maximal function for $p>1$. On the one hand, we show that the classical $A_p$ conditions are not the right ones in this setting. On the other hand, we provide sharp sufficient conditions for weighted weak and strong type $(p,p)$ boundedness of the centered maximal function, when $p>1$. The sharpness is in the sense that, given $p>1$, we can construct a weight satisfying our sufficient condition for that $p$, and so it satisfies the weak type $(p,p)$ inequality, but the strong type $(p,p)$ inequality fails. In particular, the weak type $(q,q)$ fails as well for every $q < p$.

Jorge Antezana, Sheldy Ombrosi

2023-05-23T19:02:41Z

http://arxiv.org/abs/2305.14473v1

# Weighted maximal inequalities on hyperbolic spaces ###### Abstract. In this work we develop a weight theory in the setting of hyperbolic spaces. Our starting point is a variant of the well-known endpoint Fefferman-Stein inequality for the centered Hardy-Littlewood maximal function. This inequality generalizes, in the hyperbolic setting, the weak \((1,1)\) estimates obtained by Stromberg in [17] who answered a question posed by Stein and Wainger in [16]. Our approach is based on a combination of geometrical arguments and the techniques used in the discrete setting of regular trees by Naor and Tao in [11]. This variant of the Fefferman-Stein inequality paves the road to weighted estimates for the maximal function for \(p>1\). On the one hand, we show that the classical \(A_{p}\) conditions are not the right ones in this setting. On the other hand, we provide sharp sufficient conditions for weighted weak and strong type \((p,p)\) boundedness of the centered maximal function, when \(p>1\). The sharpness is in the sense that, given \(p>1\), we can construct a weight satisfying our sufficient condition for that \(p\), and so it satisfies the weak type \((p,p)\) inequality, but the strong type \((p,p)\) inequality fails. In particular, the weak type \((q,q)\) fails as well for every \(q<p\). 2020 _Mathematics Subject Classification: 43A85_ _Keywords: Hyperbolic space, Fefferman-Stein inequality, weighted estimates._ J. Antezana was supported by grants: PICT 2019 0460 (ANPCyT), PIP112202101 00954CO (CONICET), 11X829 (UNLP), PID2020-113048GB-I00 (MCI). S. Ombrosi was supported by PID2020-113048GB-I00 (MCI). operator is the same in both spaces, \(\mathbb{R}^{n}\) and \(\mathcal{H}^{n}\). However, this is not the case in general, and it will be reveled by analyzing weighted estimates. More precisely, to complete the answer to Stein-Wainger's question we study an end-point two-weight Fefferman-Stein inequality for \(M\) in the hyperbolic setting. ### Fefferman Stein type inequality In the Euclidean setting, the classical Fefferman Stein inequality [4] is \[w\left(\{x\in\mathbb{R}^{n}\,:\,Mf(x)>\lambda\}\right)\lesssim\frac{1}{\lambda }\int_{\mathbb{R}^{n}}|f(x)|\,Mw(x)dx,\] where \(w\) is non-negative measurable function (a weight) defined in \(\mathbb{R}^{n}\), and \(w(E)=\int_{E}w(x)dx\). This is a cornerstone in the theory of weights, and a powerful tool to consider vector valued extension of the maximal function \(M\). This result follows from a classical covering lemma, which is not available in the hyperbolic setting. Indeed, in this setting \[\mu_{n}\Big{(}B_{H}(x,r)\Big{)}=\Omega_{n}\int_{0}^{r}(\sinh t)^{n-1}dt\sim_{ n}\frac{r^{n}}{1+r^{n}}e^{(n-1)r}, \tag{1.1}\] where \(\Omega_{n}\) is the euclidean \((n-1)\)-volume of the sphere \(S^{n-1}\), and the subindex in the symbol \(\sim\) means that the constant behind this symbol depends only on the dimension \(n\). This exponential behaviour, as well as the metric properties of \(\mathcal{H}^{n}\), make the classical covering arguments fail. In consequence, it is unclear how to decompose the level set \(\{x\in\mathcal{H}^{n}\,:\,Mf(x)>\lambda\}\) in such way that the appropriate averages of \(w\) appear. As in the euclidean case, from now on, given a non-negative measurable function \(w\) (a weight) defined on \(\mathcal{H}^{n}\), let \(w(E)=\int_{E}w(x)d\mu_{n}(x)\) for a measurable set \(E\subset\mathcal{H}^{n}\). On the other hand, given \(s>1\), let \[M_{s}w=M(w^{s})^{1/s}.\] Using this notation, our first main result is the following variant of the Fefferman-Stein inequality. **Theorem 1.1**.: _For every weight \(w\geq 0\) we have that_ \[w\left(\{x\in\mathcal{H}^{n}\,:\,Mf(x)>\lambda\}\right)\leq C_{s,n}\frac{1}{ \lambda}\int_{\mathcal{H}^{n}}|f(x)|M_{s}w(x)d\mu_{n}(x)\] _where the constant \(C_{s,n}\to+\infty\) when \(s\to 1\)._ This theorem is a generalization of the result of Stromberg [17], and as far as we know, it represents the first result for general weights in the hyperbolic setting. The reader may wonder if this result could hold for \(s=1\). We will show that this result is false in general if \(s=1\) (see Example 4.1 item 1 below). Moreover, our example shows that it is false, even if we put iterations of the maximal function in the right hand side. In some sense, this is an evidence of a stronger singularity of the maximal function in the hyperbolic setting. In Section 4 we will show that there are non trivial weights satisfying the pointwise condition \(M_{s}(w)(x)\leq Cw(x)\) a.e \(x\in\mathcal{H}^{n}\). Then, for these weights it holds that the maximal function \(M\) satisfies the weak type \((1,1)\) respect to the measure \(wd\mu_{n}\). _About the proof of Theorem 1.1._ For each \(r>0\), let \(A_{r}\) be the averaging operator \[A_{r}f(x)=\frac{1}{\mu_{n}(B_{H}(x,r))}\int_{B_{H}(x,r)}|f(y)|\,d\mu_{n}(y).\] Hence \(Mf(x)=\sup_{r\geq 0}A_{r}f(x)\). If \(M^{loc}(f)\) denotes the operator obtained if supremum is restricted to \(r\leq 2\), and \(M^{far}(f)\) denotes the operator obtained if the supremum is taken over all \(r\geq 2\), then \[Mf(x)\leq M^{loc}f(x)+M^{far}f(x).\] On the one hand, the operator \(M^{loc}\) behaves as in the Euclidean setting. The main difficulties appear in the estimations of \(M^{far}\). In [17], Stromberg uses a pointwise inequality obtained by Clerc and Stein in [3]. This pointwise inequality reduced the problem to get a good estimate for a convolution operator associated with a \(k\)-bi-invariant kernel \(\tau\), which in the case of hyperbolic setting is \(\tau(z,w)=(1+\mu_{n}(B(0,d(z,w))^{-1}\). A similar approach was used by Li and Lohoue in [9] to obtain sharp constants with respect to the dimension \(n\). However, Stromberg's argument strongly uses the homogeneity of the measure \(\mu_{n}\). So, it is not clear that one can apply a similar idea in the general case of any weight \(w\). This makes it necessary to look for a more flexible approach. Our general strategy is based in the scheme used by Naor and Tao in [11], where the weak type \((1,1)\) of the centered maximal function on the discrete setting of rooted \(k\)-ary trees is obtained. The flexibility of this approach was shown in [13] and [14], where the authors used this approach to get weighted estimates in the same discrete setting. It is well known that regular trees can be thought as discrete models of the hyperbolic space. Moreover, this kind of heuristic was used by Cowling, Meda and Setti in [2], but in the other way round, that is, in this work the authors used Stromberg's approach to prove weak estimates in the setting of trees. A novelty of our paper is to bring ideas of the discrete setting to the continue hyperbolic context. Adapting this strategy to a continuous context requires overcoming certain obstacles. On the one hand, the combinatorial arguments used in the discrete setting of trees are not longer available, so they have to be replaced by geometrical arguments. In this sense, the following estimate (Proposition 2.1) \[\mu_{n}\Big{(}B_{H}(y,s)\cap B_{H}(x,r)\Big{)}\leq C_{n}e^{\frac{n-1}{2}(\,r+s-d _{n}(x,y)\,)}\] is behind many estimates, as well as, some examples. It will also play a key role in the inequality \[\int_{F}A_{r}(\chi_{E})(y)w(y)d\mu_{n}(y)\leq c_{s,n}\ e^{-(n-1)\frac{r}{s^{ \prime}+1}}w(F)^{\frac{1}{s^{\prime}+1}}M_{s}w(E)^{\frac{s^{\prime}}{s^{\prime }+1}},\] that is very important to prove Theorem 1.1. In this inequality, \(E\) and \(F\) are measurable subsets of \(\mathcal{H}^{n}\), \(s>1\), \(s^{\prime}=\frac{s}{s-1}\), and \(r\) is a positive integer. On the other hand, in our setting the measure is not atomic. This leads us to make some estimations on some convenient averages of the original function instead of the function itself (see for instance Lemma 3.3). ### Weighted estimates in the hyperbolic space for \(p>1\) In the Euclidean case, the weak and strong boundedness of the maximal operator \(M\) in weighted \(L^{p}\) spaces is completely characterized by the \(A_{p}\) condition defined in the seminal work of Muckenhoupt [10]: \[\sup\left(\frac{1}{|B|}\int_{B}w\,dx\right)\left(\frac{1}{|B|}\int_{B}w^{- \frac{1}{p-1}}\,dx\right)^{p-1}<\infty, \tag{1.2}\] where the supremum is taken over all the Euclidean balls. Different type of weighted inequalities were proved for measures such that the measure of the balls grows polynomially with respect to the radius (see for instance [5], [12], [15], [18], and [19]). However, the techniques used in those works can not be applied in our framework because of the geometric properties of \(\mathcal{H}^{n}\) and the exponential growth of the measures of balls with respect to the radius. Unweighted strong \((p,p)\) inequalities for the maximal function were proved for \(p>1\) by Clerc and Stein in [3]. Moreover, singular integral operators also were studied on symmetric spaces by Ionescu ([6, 7]). Roughly speaking, in the hyperbolic spaces, the behaviour of the maximal function is a kind of combination of what happens in the Euclidean case and in the trees. More precisely, recall that we have defined the operators \[M^{loc}f(x)=\sup_{0<r\leq 2}A_{r}f(x)\quad\text{and}\quad M^{far}f(x)=\sup_{2 <r}A_{r}f(x).\] As we have already mentioned, the operator \(M^{loc}\) behaves as if it were defined in the Euclidean space. So, it is natural to expect that it boundedness could be controlled by a kind of "local \(A_{p}\) condition". We say that a weight \(w\in A_{p,loc}(\mathcal{H}^{n})\) if \[\sup_{0<r(B)\leq 1}\left(\frac{1}{\mu_{n}(B)}\int_{B}w\mu_{n}\right)\left( \frac{1}{\mu_{n}(B)}\int_{B}w^{-\frac{1}{p-1}}\mu_{n}\right)^{p-1}<\infty.\] The situation is very different for large values of the radius, when the hyperbolic structure comes into play. For instance, it is not difficult to show that the natural \(A_{p}\) condition is too strong for the boundedness of \(M^{far}\) in the hyperbolic setting. Indeed, in the Example 4.1 we show a weight for which the maximal function is bounded in all the \(L^{p}\)-spaces, but it does not belong to any (hyperbolic) \(A_{p}\) class. This suggests to follow a different approach. Inspired by the condition introduced in [14], in the case of \(k\)-ary trees, we are able to define sufficient conditions to obtain weak and strong estimates for the maximal function respect to a weight \(w\). Our main result in this direction is the following: **Theorem 1.2**.: _Let \(p>1\) and \(w\) a weight. Suppose that_ * \(w\in A_{p,loc}(\mathcal{H}^{n})\)_._ * _There exist_ \(0<\beta<1\) _and_ \(\beta\leq\alpha<p\) _such that for every_ \(r\geq 1\) _we have_ (1.3) \[\int_{F}A_{r}(\chi_{E})(y)w(y)d\mu_{n}(y)\lesssim e^{(n-1)r(\beta-1)}w(E)^{ \frac{\alpha}{p}}w(F)^{1-\frac{\alpha}{p}},\] _for any pair of measurable subsets_ \(E,F\subseteq\mathcal{H}^{n}\)_._ _Then_ \[\|Mf\|_{L^{p,\infty}(w)}\lesssim\|f\|_{L^{p}(w)}. \tag{1.4}\] _Furthermore, if \(\beta<\alpha\) then for each fixed \(\gamma\geq 0\) we have_ \[\sum_{j=1}^{\infty}j^{\gamma}\|A_{j}f\|_{L^{p}(w)}\lesssim\|f\|_{L^{p}(w)}. \tag{1.5}\] _And therefore_ \[\|Mf\|_{L^{p}(w)} \lesssim\|f\|_{L^{p}(w)},\] \[\|Mf\|_{L^{p^{\prime}}(\sigma)} \lesssim\|f\|_{L^{p^{\prime}}(\sigma)},\] _where \(\sigma=w^{1-p^{\prime}}\) and \(p^{\prime}=\frac{p}{p-1}\)._ **Remark 1.3**.: We observe that the estimate (1.5) in the previous theorem is stronger than the boundedness of the maximal function \(M^{far}(f)\). In particular, it implies that if an operator \(T\) satisfies the pointwise estimate \[|Tf(x)|\lesssim M^{loc}(|f|)(x)+\sum_{j\geq 1}j^{\gamma}A_{j}(|f|)(x),\] for some \(\gamma\geq 0\), then the requested conditions on the weight \(w\) in Theorem 1.2 will be sufficient condition for the boundedness of \(T\) in the space \(L^{p}(w)\) with \(p>1\). In particular, this generalized, in the hyperbolic setting, the unweighted estimates obtained by Clerc and Stein in [3, Thm. 2] for the maximal function. **Remark 1.4**.: It is not clear whether or not the condition (1.3) for \(\alpha=\beta\) is a necessary condition for the weak type \((p,p)\) boundedness of \(M\) with respect to \(w\). However, the condition is sharp in the following sense: if \(\beta=\alpha\) we can construct a weight for which the weak type \((p,p)\) holds, but the strong type \((p,p)\) fails. Consequently, the weak type \((q,q)\) fails as well for every \(q<p\) (see Example 4.1 (2)). In particular, this shows that, unlike the classical case, in the hyperbolic context the weak \((p,p)\) inequality with respect to \(w\) of the maximal operator is not equivalent to the strong estimate for \(p>1\). The condition (1.3) could be not easy to be checked. For this reason, we consider the following result which provides a more tractable condition. To simplify the statement, given a positive integer \(j\), let \[\mathcal{C}_{j}=B(0,j)\setminus B(0,j-1).\] Observe that the sets considered in the condition in (1.3) may have non-empty intersection with several different levels \(\mathcal{C}_{j}\). The condition in the following proposition studies the behavior of the weight at each level. **Proposition 1.5**.: _Let \(1<p<\infty\), and let \(w\) be a weight such that there exists a real number \(\delta<1\), so that for every \(j,l,r\geq 1\) integers with the restriction \(|l-j|\leq r\), we have that_ \[w(\mathcal{C}_{l}\cap B(x,r))\lesssim e^{(n-1)\frac{r+l-j}{2}(p-\delta)}e^{(n- 1)r\delta}w(x),\quad\text{for a.e. }x\in\mathcal{C}_{j}. \tag{1.6}\] _Then, the condition (1.3) in Theorem 1.2 holds with \(\beta=\alpha=\frac{p}{p-\delta+1}\)._ Combining Theorem 1.2, Remark 1.3 and Proposition 1.5 we obtain the following corollary. **Corollary 1.6**.: _Let \(1\leq p<\infty\), and \(w\in A_{p,loc}(\mathcal{H}^{n})\) such that there exists a real number \(\delta<1\) such that for every \(j,l,r\geq 1\) integers with the restriction \(|l-j|\leq r\), we have that_ \[w(\mathcal{C}_{l}\cap B(x,r))\lesssim e^{(n-1)\frac{r+l-j}{2}(p-\delta)}e^{(n -1)r\delta}w(x),\quad\text{for a.e. }x\in\mathcal{C}_{j}.\] _Then_ \[\|Mf\|_{L^{p,\infty}(w)}\lesssim\|f\|_{L^{p}(w)}.\] _Furthermore, if \(p<q\) we have_ \[\|Tf\|_{L^{q}(w)}\lesssim\|f\|_{L^{q}(w)},\] _for every operator \(T\) satisfying the pointwise estimate_ \[|Tf(x)|\lesssim M^{loc}(|f|)(x)+j^{\gamma}\sum_{j\geq 1}A_{j}(|f|)(x),\] _for some \(\gamma\geq 0\)._ ### Organization of the paper This paper is organized as follow. In Section 2 we prove an estimate on the measure of the intersection of two hyperbolic balls. Section 3 is devoted to the proof of the main results of this paper. The proof of Theorem 1.1 is contained in Subsection 3.1, while the proof of Theorem 1.2 is contained in Subsection 3.2. The Section 3 concludes with the proof of Proposition 1.5. The Section 4 contains examples that clarify several points previously mentioned. Finally, the paper concludes with an appendix on the ball model of the hyperbolic space. ## 2. Geometric results ### The hyperbolic space Although the precise realisation of hyperbolic space is not important for our purposes, for sake of concreteness, throughout this article we will consider the ball model. Recall that \(\mu_{n}\) denotes the volume measure, and by \(d_{n}\) we will denote the hyperbolic distance. A brief review of some basic facts about this model and its isometries is left to the Appendix A. ### Two results on the intersection of balls in the hyperbolic space This subsection is devoted to prove the following two geometric results, which will be very important in the sequel. **Proposition 2.1**.: _Let \(B_{H}(y,s)\) and \(B_{H}(x,r)\) be two balls in \(\mathcal{H}_{n}\). Then_ \[\mu_{n}\Big{(}B_{H}(y,s)\cap B_{H}(x,r)\Big{)}\leq C_{n}e^{\frac{n-1}{2}(\,r+ s-d_{n}(x,y)\,)},\] _where \(C_{n}\) is a constant that only depends on the dimension._ Proof.: We can assume that \(B_{H}(y,s)\cap B_{H}(x,r)\neq\varnothing\). On the other hand, since the estimate is trivial if \(r\) and \(s\) are less than a fixed constant, we can also assume that \(r,s>2\). Without loss of generality, we can assume that \(y=0\) and \(x=(d,0,\dots,0)\) with \(d=d_{n}(x,y)\). Note that we can also assume that \(d>0\), otherwise the estimate is trivial. The geodesic passing through the centers is the segment \[L=\{(t,0,\dots,0):\ t\in(-1,1)\}.\] Since the balls are symmetric with respect to this geodesic line, the intersection is also symmetric with respect to this line. Let \(O_{L}(n-1)\) be the subgroup of the orthogonal group \(O(n)\) defined by \[O_{L}(n)=\{A\in O(n):\text{A leaves invariant the geodesic line }L\},\] then the intersection is invariant by the action of \(O_{L}(n-1)\). Moreover, the subgroup \(O_{L}(n-1)\) acts transitively in the intersection of the boundaries \(\partial B_{H}(0,s)\cap\partial B_{H}(x,r)\) which turns out to be an \((n-2)\)-sphere. Let \(S\) denote this intersection of boundaries, and consider the point \(m\in L\) that satisfies \[d_{n}(0,m)=\frac{s+d-r}{2}\quad\Longleftrightarrow\quad d_{n}(m,x)=\frac{r+d-s} {2}.\] Since \(L\) is a symmetry axis for \(S\), the points in \(S\) are at the same distance to the point \(m\). Let \(\rho\) denote this distance. The volume of the ball of radius \(\rho\) can be estimated using the hyperbolic law of cosines. Take \(q\in S\), and consider the two dimensional hyperbolic (also linear) plane \(P\) containing \(q\) and \(L\). Let us restrict our attention to this hyperbolic plane (see Figure 1). Since \(\angle(0,m,q)+\angle(q,m,x)=\pi\), one of them is greater or equal to \(\frac{\pi}{2}\). Suppose that the angle \(\theta=\angle(0,m,q)\) is greater than \(\frac{\pi}{2}\), and consider the geodesic triangle whose vertices are \(0\), \(m\) and \(q\) (see Figure 2). Since \(\cos(\theta)\) is non-positive, we have that \[\cosh(s) =\cosh\left(\frac{s+d-r}{2}\right)\cosh(\rho)-\sinh\left(\frac{s+d -r}{2}\right)\sinh(\rho)\cos(\theta)\] \[\geq\cosh\left(\frac{s+d-r}{2}\right)\cosh(\rho).\] Figure 1. Intersection of the balls with the two dimensional plane \(P\). Figure 2. Geodesic triangle. Therefore, we get the following estimate \[e^{\rho}\leq\cosh(\rho)\leq\frac{\cosh(s)}{\cosh\left(\frac{s+d-r}{2}\right)} \leq 2e^{\frac{s+r-d}{2}}.\] By equation (1.1), we get that \[\operatorname{Vol}\left(B_{H}(m,\rho)\right)=\Omega_{n}\int_{0}^{\rho}(\sinh t )^{n-1}dr\leq K_{n}e^{(n-1)\rho}\leq 2^{n}K_{n}e^{(n-1)\left(\frac{s+r-d}{2} \right)}. \tag{2.1}\] Now, it is enough to prove that \(B_{H}(0,s)\cap B_{H}(x,r)\subseteq B_{H}(m,\rho)\). Since the intersection is an open-connected set, it is enough to prove that the boundary \(B_{H}(m,\rho)\) is not contained in the intersection. So, take \(p\in\partial B_{H}(m,\rho)\). By a continuity argument, we can assume that \(p\notin L\). Then, as before, consider the plane \(P\) generated by \(p\) and the geodesic \(L\). The geodesic \(L\) divide this plane in two parts. Let \(q\) be the unique point in \(P\cap S\) in the same half-plane as \(p\), and suppose that \(\theta_{p}=\angle(p,m,x)\) is greater or equal than \(\theta_{q}=\angle(q,m,x)\) (see Figure 3). If \(t=d_{n}(x,p)\), since the cosine is decreasing in \((0,\pi)\) we get that \[\cosh(t) =\cosh\left(\frac{r+d-s}{2}\right)\cosh(\rho)-\sinh\left(\frac{r+ d-s}{2}\right)\sinh(\rho)\cos(\theta_{p})\] \[\geq\cosh\left(\frac{r+d-s}{2}\right)\cosh(\rho)-\sinh\left(\frac{ r+d-s}{2}\right)\sinh(\rho)\cos(\theta_{q})\] \[=\cosh(r).\] In consequence, \(t\geq r\) and therefore, the point \(t\notin B_{H}(x,r)\). If \(\angle(p,m,x)\) is smaller than \(\angle(q,m,x)\), it holds that \(\angle(p,m,0)\) is greater than \(\angle(q,m,0)\). Hence, the same argument, replacing the vertex \(x\) by the vertex \(0\) shows that \(t\notin B_{H}(0,s)\). This concludes the proof. The following is a corollary of the proof of the previous lemma. Figure 3. Comparison of triangles. **Corollary 2.2**.: _Let \(B_{H}(0,s)\) and \(B_{H}(x,r)\) be two balls in \(\mathcal{H}_{n}\) such that their intersection has positive measure. If \(\rho_{0}=\frac{1}{2}(\,r+s-d_{n}(0,x)\,)\), then_ \[B_{H}(m,\rho_{0})\subseteq B_{H}(0,s)\cap B_{H}(x,r)\subseteq B_{H}(m,\rho_{0}+ 1),\] _where \(m=\alpha x\), and \(\alpha=\tanh\Big{(}\frac{s+d-r}{2}\Big{)}\)._ ## 3. Proof of Main results First of all, we will prove the following arithmetical lemma, which is a slight generalization of a result contained in [14]. **Lemma 3.1**.: _Let \(1\leq p<\infty\), \(-p<\delta<1\),and \(\kappa>1\). Let the sequences of non-negative real numbers \(\{c_{j}\}_{j=0}^{\infty}\) and \(\{d_{l}\}_{l=0}^{\infty}\) satisfying_ \[\sum_{j=0}^{\infty}\kappa^{(p-\delta)j}c_{j}=A\quad\text{and}\quad\sum_{l=0}^ {\infty}\kappa^{l}d_{l}=B.\] _Then, for every integer \(r\geq 1\) we have that_ \[\sum_{j,l\in\mathbb{N}\cup\{0\}}\min\Big{\{}\kappa^{\delta r}\kappa^{\frac{(l +j+r)(p-\delta)}{2}}c_{j},\kappa^{\frac{l+j+r}{2}}d_{l}\Big{\}}\leq c_{p, \delta,\kappa}\,\kappa^{\frac{p}{p-\delta+1}r}A^{\frac{1}{p-\delta+1}}B^{1- \frac{1}{p-\delta+1}}. \tag{3.1}\] Proof.: To prove this inequality, let \(\rho\) be a real parameter to be chosen later, and argue as follows \[\sum_{j,l\in\mathbb{N}\cup\{0\}}\min\Big{\{}\kappa^{\delta r}\kappa^{\frac{(l +j+r)(p-\delta)}{2}}c_{j},\kappa^{\frac{l+j+r}{2}}d_{l}\Big{\}}\] \[\leq\kappa^{\frac{p+\delta}{2}r}\sum_{\begin{subarray}{c}l,j\in \mathbb{N}\cup\{0\}\\ l<j+\rho\end{subarray}}\kappa^{\frac{(l+j)(p-\delta)}{2}}c_{j}+\kappa^{\frac{ r}{2}}\sum_{\begin{subarray}{c}l,j\in\mathbb{N}\cup\{0\}\\ l\geq j+\rho\end{subarray}}k^{\frac{l+j}{2}}d_{l}\] \[\lesssim\kappa^{\frac{p+\delta}{2}r}\sum_{j=0}^{\infty}\kappa^{ \frac{(j+\rho+j)(p-\delta)}{2}}c_{j}+\kappa^{\frac{r}{2}}\sum_{l=0}^{\infty} \kappa^{l-\frac{\rho}{2}}d_{l}\] \[=\kappa^{\frac{p+\delta}{2}r}\kappa^{\frac{\rho(p-\delta)}{2}} \sum_{j=0}^{\infty}k^{j(p-\delta)}c_{j}+\kappa^{\frac{r}{2}}k^{-\frac{\rho}{2} }\sum_{l=0}^{\infty}\kappa^{l}d_{l}\] \[=\kappa^{\frac{p+\delta}{2}r}\kappa^{\frac{\rho(p-\delta)}{2}}A+ \kappa^{\frac{r}{2}}\kappa^{-\frac{\rho}{2}}B.\] Choosing \(\rho=\frac{2\log_{\kappa}\big{(}\frac{B}{A}\big{)}}{p-\delta+1}-\frac{(p+ \delta-1)r}{p-\delta+1}\), it follows that \[\kappa^{\frac{p+\delta}{2}r}\kappa^{\frac{\rho(p-\delta)}{2}}A+\kappa^{\frac {r}{2}}\kappa^{-\frac{\rho}{2}}B\leq c_{p,\delta}\kappa^{\frac{p}{p-\delta+1}r }A^{\frac{1}{p-\delta+1}}B^{1-\frac{1}{p-\delta+1}},\] which concludes the proof. ### Proof of Theorem 1.1 The first step consists on proving that Lemma 2.1 leads to the following result. This is a key point to push the scheme on the discrete cases in [11] or [13]. Recall that, given \(r\geq 0\), we denote by \(A_{r}\) the averaging operator \[A_{r}f(x)=\frac{1}{\mu_{n}(B_{H}(x,r))}\int_{y\in B_{H}(x,r)}|f(x)|\,d\mu_{n}(x).\] **Lemma 3.2**.: _Let \(E,F\) measurable sets of \(\mathcal{H}^{n}\), \(s>1\) and let \(r\) be a positive integer. Then_ \[\int_{F}A_{r}(\chi_{E})(y)w(y)d\mu_{n}(y)\leq c_{s,n}e^{-(n-1)\frac{r}{s^{ \prime}+1}}w(F)^{\frac{1}{s^{\prime}+1}}M_{s}w(E)^{\frac{s^{\prime}}{s^{\prime }+1}},\] _where \(s^{\prime}=\frac{s}{s-1}\) and \(c_{s,n}\) is a constant depending on \(s\) and the dimension \(n\)._ Proof.: We divide the hyperbolic \(\mathcal{H}^{n}\) in level sets as follows \[\mathcal{H}^{n}=\bigcup_{j=1}^{\infty}\mathcal{C}_{j},\] where \(\mathcal{C}_{j}=\{x\in\mathcal{H}^{n}:j-1\leq d_{H}(0,x)<j\}\). Let \(E_{j}=E\cap\mathcal{C}_{j}\) and \(F_{\ell}=F\cap\mathcal{C}_{\ell}\). Hence, we can write \[I:=\int_{F}A_{r}(\chi_{E})(y)w(y)d\mu_{n}(y)=\sum_{\ell,j\geq 0} \int_{F_{\ell}}A_{r}(\chi_{E_{j}})(y)w(y)d\mu_{n}(y). \tag{3.2}\] Now, we will estimate the integrals \[I_{j,\ell}:=\int_{F_{\ell}}A_{r}(\chi_{E_{j}})(y)w(y)d\mu_{n}(y)\] in two different ways. On the one hand, given \(x\in E_{j}\), let \[\Omega^{x}_{j,\ell}=\{y\in F_{\ell}:\ d(x,y)\leq r\}.\] Then, by Lemma 2.1 \[\mu_{n}(\Omega^{x}_{j,\ell})\leq C_{n}e^{\frac{n-1}{2}(\ell+r-j)}.\] Using this estimate, we obtain that \[I_{j,\ell} =e^{-(n-1)r}\int_{F_{\ell}}\int_{B(y,r)}\chi_{E_{j}}(x)\,d\mu_{n} (x)w(y)d\mu_{n}(y)\] \[=e^{-(n-1)r}\int_{E_{j}}\int_{\Omega^{x}_{j,\ell}}w(y)d\mu_{n}(y) \,d\mu_{n}(x)\] \[=e^{-(n-1)r}\int_{E_{j}}\left(\int_{\Omega^{x}_{j,\ell}}d\mu_{n} \right)^{\frac{1}{s^{\prime}}}\left(\int_{B_{H}(x,r)}w^{s}(y)\,d\mu_{n}(y) \right)^{\frac{1}{s}}\,d\mu_{n}(x)\] \[\leq C_{n}e^{-(n-1)r}e^{\frac{n-1}{2s^{\prime}}(\ell+r-j)}\,e^{ \frac{(n-1)r}{s}}M_{s}(w)(E_{j}).\] On the other hand, if \(y\in F_{\ell}\), let \(\Omega_{j,\ell}^{y}=\{x\in E_{j}:\ d(x,y)\leq r\}\). Then, by Lemma 2.1 \[I_{j,\ell} =e^{-(n-1)r}\int_{F_{\ell}}\int_{\Omega_{j,\ell}^{y}}d\mu_{n}(x)\,w (y)d\mu_{n}(y)\] \[\leq C_{n}e^{-(n-1)r}e^{\frac{n-1}{2}(j+r-\ell)}\,w(F_{\ell}).\] In consequence \[I_{j,\ell}\leq C_{n}e^{-(n-1)r}\min\Big{\{}e^{\frac{n-1}{2s^{\prime}}(\ell+r-j )}\,e^{\frac{(n-1)r}{s}}M_{s}(w)(E_{j}),e^{\frac{n-1}{2}(j+r-\ell)}\,w(F_{\ell })\Big{\}},\] and \[I\leq C_{n}e^{-(n-1)r}\sum_{|\ell-j|\leq r+2}\min\Big{\{}e^{\frac{n-1}{2s^{ \prime}}(\ell+r-j)}\,e^{\frac{(n-1)r}{s}}M_{s}(w)(E_{j}),e^{\frac{n-1}{2}(j+r -\ell)}\,w(F_{\ell})\Big{\}}.\] Now, if we define \(c_{j}=\frac{M_{s}^{\circ}w(E_{j})}{e^{(n-1)\frac{j}{s^{\prime}}}}\) and \(d_{l}=\frac{w(F_{l})}{e^{(n-1)l}}\). We have that \[\sum_{j=0}^{\infty}e^{(n-1)\frac{j}{s^{\prime}}}c_{j}=M_{s}^{\circ}w(E)\quad \text{and}\quad\sum_{j=0}^{\infty}e^{(n-1)l}d_{j}=w(F), \tag{3.3}\] and \[\min\Bigl{\{}e^{\frac{n-1}{2s^{\prime}}(\ell+r-j)}\,e^{\frac{(n-1 )r}{s}}M_{s}(w)(E_{j}),e^{\frac{n-1}{2}(j+r-\ell)}\,w(F_{\ell})\Bigr{\}}\] \[\qquad=\min\Big{\{}e^{\frac{(n-1)r}{s}}e^{(n-1)\frac{(l+j+r)}{2s^ {\prime}}}c_{j},e^{(n-1)\frac{l+j+r}{2}}d_{l}\Big{\}}\] Then we have that \[I\lesssim e^{-(n-1)r}\sum_{l,j\in\mathbb{N}\cup\{0\}}\min\left\{e^{\frac{(n-1 )r}{s}}e^{(n-1)\frac{(l+j+r)}{2s^{\prime}}}c_{j},e^{(n-1)\frac{l+j+r}{2}}d_{l }\right\}. \tag{3.4}\] Now, if we choose \(\delta=\frac{1}{s}\) and \(p=1\) (then \(p-\delta=\frac{1}{s^{\prime}}\)) we have that \[\min\left\{e^{\frac{(n-1)r}{s}}e^{(n-1)\frac{(l+j+r)}{2s^{\prime}}}c_{j},e^{( n-1)\frac{l+j+r}{2}}d_{l}\right\}\] is equal to \[\min\left\{e^{(n-1)\delta r}e^{(n-1)\frac{(l+j+r)(p-\delta)}{2}}c_{j},e^{(n-1) \frac{l+j+r}{2}}d_{l}\right\}.\] Therefore, if \(\kappa=e^{n-1}\) and we take into account (3.3), applying Lemma 3.1 in (3.4) we get \[I\lesssim e^{-(n-1)\frac{r}{s^{\prime}+1}}w(F)^{\frac{1}{s^{\prime}+1}}M_{s}w( E)^{\frac{s^{\prime}}{s^{\prime}+1}}.\] We can use Lemma 3.2 to obtain a distributional estimate on \(A_{r}\). **Lemma 3.3**.: _Let \(r\geq 1\) and \(\lambda>0\). Then_ \[w\left(\left\{A_{r}(A_{1}f)\geq\lambda\right\}\right)\lesssim c_{s}\sum_{k=0}^{r} \left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac{1}{2s^{\prime}}}e^{(n-1)k}M _{s}w\left(\left\{|A_{2}f|\geq\eta e^{(n-1)k}\right\}\right),\] _where \(c_{s}\) depends only on \(s\) and \(c_{s}\to\infty\) when \(s\to 1\)._ Proof of Lemma 3.3.: Let \(f_{1}=A_{1}f\). We bound \[f_{1}\leq\frac{1}{e}+\sum_{k=0}^{r}e^{(n-1)k}\chi_{E_{k}}+f_{1}\chi_{\{f_{1} \geq\frac{1}{2}e^{(n-1)r}\}}, \tag{3.5}\] where \(E_{k}\) is the sublevel set \[E_{k}=\left\{e^{(n-1)(k-1)}\leq f_{1}<e^{(n-1)k}\right\}. \tag{3.6}\] Hence \[A_{r}f_{1}\leq\frac{1}{e}+\sum_{k=0}^{r}e^{(n-1)k}A_{r}\left(\chi_{E_{k}} \right)+A_{r}\left(f_{1}\chi_{\{f_{1}\geq\frac{1}{2}e^{(n-1)r}\}}\right). \tag{3.7}\] Given any \(\lambda>0\) \[w\left(\left\{A_{r}\left(f_{1}\chi_{\{f_{1}\geq e^{(n-1)r}\}} \right)>\lambda\right\}\right) \leq w\left(\left\{A_{r}\left(f_{1}\chi_{\{f_{1}\geq e^{(n-1)r}\} }\right)\neq 0\right\}\right)\] \[\leq w\left(\left\{x:B_{H}(r,x)\cap\{f_{1}\geq e^{(n-1)r}\}\neq \varnothing\right\}\right).\] Take \(x\) such that \(B_{H}(x,r)\cap\{f_{1}\geq e^{(n-1)r}\neq\varnothing\), and let \(y\) be an element of this intersection. It is not difficult to see that \[B_{H}(y,1)\subseteq B_{H}(x,r+1)\cap\big{\{}f_{2}\geq ce^{(n-1)r}\big{\}},\] where \(f_{2}=A_{2}f\) and \(c_{0}=\frac{\mu_{n}(B(0,1))}{\mu_{n}(B(0,2))}\). Therefore \[w\left(\left\{x:B_{H}(r,x)\cap\{f_{1}\geq e^{(n-1)r}\}\neq \varnothing\right\}\right) \leq w\Big{(}\Big{\{}A_{r+1}\left(\chi_{\{f_{2}\geq ce^{(n-1)r}\} }\right)>\frac{1}{c_{1}e^{(n-1)r}}\Big{\}}\Big{)}\] \[\leq c_{1}e^{(n-1)r}M(w)\left(\chi_{\{f_{2}\geq ce^{(n-1)r}\}} \right).\] On the other hand, let \(\beta\in(0,1)\) that will be chosen later. Note that if \[\sum_{k=0}^{r}e^{(n-1)k}A_{r}\left(\chi_{E_{k}}\right)\geq\frac{1}{e},\] then we necessarily have some \(k\in\mathbb{N}\) such that \(1\leq k\leq r\) for which \[A_{r}\left(\chi_{E_{k}}\right)\geq\frac{e^{(n-1)\beta}-1}{e^{(n-1)(k+2)}}\left( \frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\beta}.\] Indeed, otherwise we have that \[\frac{1}{e} \leq\sum_{k=0}^{r}e^{(n-1)k}A_{r}\left(\chi_{E_{k}}\right)<\frac{e^ {(n-1)\beta}-1}{e^{(n-1)(\beta r+2)}}\sum_{k=0}^{r}e^{(n-1)\beta k}\] \[=\frac{e^{(n-1)\beta}-1}{e^{(n-1)(\beta r+2)}}\ \frac{e^{(n-1)\beta(r+1)}-1}{e^{(n-1)\beta}-1}<\frac{1}{e},\] which is a contradiction. Thus \[w\left(A_{r}f_{1}\geq 1\right)\leq\sum_{k=0}^{r}w(F_{k})+c_{1}e^{(n-1)r}M(w) \left(\chi_{\{f_{2}\geq ce^{(n-1)r}\}}\right),\] where \[F_{k}=\left\{A_{r}\left(\chi_{E_{k}}\right)\geq\frac{e^{(n-1)\beta}-1}{e^{(n-1 )(k+2)}}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\beta}\right\}.\] Note that \(F_{k}\) has finite measure, and \[w(F_{k})\frac{e^{(n-1)\beta}-1}{e^{(n-1)(k+2)}}\left(\frac{e^{(n-1)k}}{e^{(n-1 )r}}\right)^{\beta}\leq\int_{F_{k}}A_{r}(\chi_{E_{k}})wd\mu_{n}(x).\] On the other hand, by Lemma 3.2, \[\int_{F_{k}}A_{r}(\chi_{E_{k}})wd\mu_{n}(x)\leq c_{s}e^{-(n-1)\frac{r}{s^{ \prime}+1}}w(F_{k})^{\frac{1}{s^{\prime}+1}}M_{s}w(E_{k})^{\frac{s^{\prime}}{ s^{\prime}+1}}.\] Hence \[w(F_{k})\frac{e^{(n-1)\beta}-1}{e^{(n-1)(k+2)}}\left(\frac{e^{(n-1)k}}{e^{(n- 1)r}}\right)^{\beta}\leq c_{s}e^{-(n-1)\frac{r}{s^{\prime}+1}}w(F_{n})^{\frac{ 1}{s^{\prime}+1}}M_{s}w(E_{n})^{\frac{s^{\prime}}{s^{\prime}+1}}.\] So, choosing \(\beta=\frac{1}{2(s^{\prime}+1)}\) we have that \[w(F_{k}) \leq c_{s}e^{-(n-1)\frac{r}{2s^{\prime}}}e^{\frac{(n-1)k}{2s^{ \prime}}}e^{(n-1)k}M_{s}w(E_{n})\] \[\leq c_{s}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac{1}{2s ^{\prime}}}e^{(n-1)k}M_{s}w\left(\left\{f_{1}\geq e^{(n-1)(k-1)}\right\} \right).\] Therefore \[w(\{A_{r}f_{1}\geq 1\}) \leq c_{s}\sum_{k=0}^{r}c_{s}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}} \right)^{\frac{1}{2s^{\prime}}}e^{(n-1)k}M_{s}w\left(\left\{f_{1}\geq e^{(n-1 )(k-1)}\right\}\right) \tag{3.8}\] \[+c_{1}e^{(n-1)r}M(w)\left(\chi_{\{f_{2}\geq ce^{(n-1)r}\}}\right).\] So, there exists \(\eta>0\) depending only on the dimension such that \[w(\{A_{r}f_{1}\geq 1\})\leq\tilde{c}_{s}\sum_{k=0}^{r}\left(\frac{e^{(n-1)k}}{e^{ (n-1)r}}\right)^{\frac{1}{2s^{\prime}}}e^{(n-1)k}M_{s}w\left(\left\{f_{2}\geq \eta e^{(n-1)(k-1)}\right\}\right).\] Indeed, note that in the right-hand side of (3.8), the second term is dominated by the last term of the sum. This yields the desired conclusion. Combining the ingredients above we are in position to settle Theorem 1.1. Proof of Theorem 1.1.: By the discussion in the introduction we only need to argue for \(M^{far}(f)(x)\). Then, by Lemma 3.3 implies that \[w\Big{(}M^{far}f \geq\lambda\Big{)}\leq w\left(M^{far}f_{1}\geq\lambda\right)\] \[\leq\sum_{r=1}^{\infty}w\left(A_{r}f_{1}\geq\lambda\right)\] \[=\tilde{c}_{s}\sum_{r=0}^{\infty}\sum_{k=0}^{r}\left(\frac{e^{(n-1 )k}}{e^{(n-1)r}}\right)^{\frac{1}{2s^{\prime}}}e^{(n-1)k}M_{s}w\left(\left\{f_ {2}\geq e^{(n-1)(k-1)}\eta\lambda\right\}\right)\] \[=\tilde{c}_{s}\int_{\mathcal{H}_{n}}\sum_{r=0}^{\infty}\sum_{k=0} ^{r}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac{1}{2s^{\prime}}}e^{(n-1 )k}\chi_{\{f_{2}\geq e^{(n-1)(k-1)}\}\eta\lambda\}M_{s}w(x)d\mu_{n}(x)\] \[=\tilde{c}_{s}\int_{\mathcal{H}_{n}}\sum_{k=0}^{\infty}\sum_{r=k} ^{\infty}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac{1}{2s^{\prime}}}e^{ (n-1)k}\chi_{\{f_{2}\geq e^{(n-1)(k-1)}\}\eta\lambda\}M_{s}w(x)d\mu_{n}(x)\] \[=\tilde{c}_{s}\int_{\mathcal{H}_{n}}\sum_{k=0}^{\infty}e^{(n-1)k} \chi_{\{f_{2}\geq e^{(n-1)(k-1)}\}\eta\lambda\}M_{s}w(x)d\mu_{n}(x)\] \[\leq\frac{\hat{c}_{s}}{\eta\lambda}\int_{\mathcal{H}_{n}}f_{2}(x) M_{s}w(x)d\mu_{n}(x)\] \[=\frac{\hat{c}_{s}}{\eta\lambda}\int_{\mathcal{H}_{n}}f(x)A_{2}(M _{s}w)(x)d\mu_{n}(x).\] Now, if \(w\) is identically \(1\) we have \(A_{2}(M_{s}w)(x)=1\) and we are done. In particular, this recovers the Stromberg's weak type \((1,1)\) estimate. If \(w\) is not constant, we claim that \[A_{2}\left((M_{s}w)\right)(x)\lesssim_{s}M_{s}w(x).\] Indeed, \[\frac{1}{\mu_{n}(B(x,2))}\int_{B(x,2)}M_{s}w(y)d\mu_{n}(y) \leq\frac{1}{\mu_{n}(B(x,2))}\int_{B(x,2)}M(w^{s}\chi_{B(x,4)}(y) )^{\frac{1}{s}}d\mu_{n}(y)\] \[+\frac{1}{\mu_{n}(B(x,2))}\int_{B(x,2)}M(w^{s}\chi_{(B(x,4))^{c}} (y))^{\frac{1}{s}}d\mu_{n}(y).\] The second term in the last line can be controlled by \(cM_{s}(w)(x)\) because \[M(w^{s}\chi_{(B(x,4))^{c}}(y))^{\frac{1}{s}}\sim M(w^{s}\chi_{(B(x,4))^{c}}(x ))^{\frac{1}{s}},\] for every \(y\in B(x,2)\). Using Kolmogorov's inequality and the weak type \((1,1)\) of \(M\) the first term can be estimate by \(c_{\beta}(A_{4}(w^{s})(x))^{\frac{1}{s}}\) and the claim follows. This completes the proof in the general case. ### Proof of Theorem 1.2 The proof of Theorem 1.2 follows the same ideas of the proof of Theorem 1.1. in [14]. First, the hypothesis \(w\in A_{p,loc}(\mathcal{H}^{n})\) implies the estimates for \(M^{loc}\) by standard arguments as in the classical setting. On the other hand, the arguments used to prove that Lemma 3.2 implies Lemma 3.3 can be used to prove that the hypothesis in Theorem 1.2 implies that \[w\left(\{A_{r}(A_{1}f)\geq\lambda\}\right)\lesssim c_{s}\sum_{k=0}^{r}\left( \frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac{1-\beta}{2}\frac{p}{\alpha}}e^{(n- 1)\beta\frac{p}{\alpha}k}w\left(\left\{|A_{2}f|\geq\eta e^{(n-1)k}\lambda \right\}\right). \tag{3.9}\] This inequality shows that the case \(\beta<\alpha\) produces a better estimate than the case \(\beta=\alpha\). First of all, assume that we are in the worst case \(\beta=\alpha\). Arguing as in the proof of Theorem 1.1 we get \[w\left(\left\{M^{far}f(x)\geq\lambda\right\}\right)\lesssim\frac{c}{\lambda^{ p}}\int_{\mathcal{H}^{n}}|A_{2}(f)(x)|^{p}w(x)d\mu_{n}(x)dx.\] Since \(|A_{2}(f)(x)|\leq M^{loc}f(x)\) and \(w\in A_{p,loc}(\mathcal{H}^{n})\), paying a constant we can eliminate \(A_{2}\) in the right hand side of the previous estimate, and the proof is complete in this case. If we assume that \(\beta<\alpha\), then by (3.9) we have that \[\|A_{r}f\|_{L^{p}(w)}^{p} =p\int_{0}^{\infty}\lambda^{p-1}w\left(A_{r}f\geq\lambda\right)d\lambda\] \[\lesssim\sum_{k=0}^{r}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^ {\frac{1-\beta}{2}\frac{p}{\alpha}}e^{(n-1)\beta\frac{p}{\alpha}k}\int_{0}^{ \infty}\lambda^{p-1}w\left(\left\{|A_{2}f|\geq\eta e^{(n-1)k}\lambda\right\}\right)\] \[=\sum_{k=0}^{r}\left(\frac{e^{(n-1)k}}{e^{(n-1)r}}\right)^{\frac {1-\beta}{2}\frac{p}{\alpha}}e^{(n-1)\beta\frac{p}{\alpha}k}e^{-(n-1)kp}\|A_{ 2}f\|_{L^{p}(w)}^{p}\] \[\lesssim e^{(n-1)rp(\frac{\beta}{\alpha}-1)}\|A_{2}f\|_{L^{p}(w)} ^{p}.\] Since \(w\in A_{p,loc}(\mathcal{H}^{n})\) we can eliminate \(A_{2}\) in the last norm, and taking into account that \(\frac{\beta}{\alpha}-1<0\), we have that \[\sum_{r=1}^{\infty}r^{\gamma}\|A_{r}f\|_{L^{p}(w)}\lesssim\sum_{r=1}^{\infty} r^{\gamma}e^{(n-1)rp(\frac{\beta}{\alpha}-1)}\|f\|_{L^{p}(w)}\sim_{\gamma, \alpha,\beta,p}\|f\|_{L^{p}(w)}.\] This leads to (1.5). From (1.5) and the fact that \(\sum_{j=1}^{\infty}A_{j}(f)\) is self-adjoint (\(\gamma=0\)) we obtain the boundedness of \(M^{far}\) in the spaces \(L^{p}(w)\) and \(L^{p^{\prime}}(\sigma)\). Moreover, since \(w\in A_{p,loc}(\mathcal{H}^{n})\) and therefore \(\sigma\) is in \(A_{p^{\prime},loc}(\mathcal{H}^{n})\) we have the same inequalities for \(M^{loc}\), and as a consequence we obtain \[\|Mf\|_{L^{p}(w)} \lesssim\|f\|_{L^{p}(w)}\] \[\|Mf\|_{L^{p^{\prime}}(\sigma)} \lesssim\|f\|_{L^{p^{\prime}}(\sigma)}\] This ends the proof of the Theorem. ### Proof of Proposition 1.5 The proof follows similar ideas as Lemma 3.2. Proof of Proposition 1.5.: Given \(E,F\) subsets in \(\mathcal{H}^{n}\), we should prove that \[\int_{F}A_{r}(\chi_{E})(y)w(y)d\mu_{n}(y)\lesssim e^{(n-1)r\left(\frac{p}{p- \delta+1}-1\right)}w(E)^{\frac{1}{p-\delta+1}}w(F)^{1-\frac{1}{p-\delta+1}}. \tag{3.10}\] Using the same notation as in the Lemma 3.2, we have \[I_{j,\ell}:=\int_{F_{\ell}}A_{r}(\chi_{E_{j}})(y)w(y)d\mu_{n}(y).\] Given \(x\in E_{j}\), let \(\Omega_{j,\ell}^{x}=\{y\in F_{\ell}:\ d(x,y)\leq r\}\). Then, by condition (1.6) \[w(\Omega_{j,\ell}^{x})\leq C_{n}e^{(n-1)\frac{r+l-j}{2}(p-\delta)}e^{(n-1)r \delta}w(x).\] Therefore, \[I_{j,\ell} =e^{-(n-1)r}\int_{F_{\ell}}\int_{B(y,r)}\chi_{E_{j}}(x)\,d\mu(x)w (y)d\mu_{n}(y)\] \[=e^{-(n-1)r}\int_{E_{j}}\int_{\Omega_{j,\ell}^{x}}w(y)d\mu_{n}(y) \,d\mu_{n}(x)\] \[\lesssim e^{-(n-1)r}e^{(n-1)\frac{r+l-j}{2}(p-\delta)}e^{(n-1)r \delta}w(E_{j}).\] On the other hand, if \(y\in F_{\ell}\), let \(\Omega_{j,\ell}^{y}=\{x\in E_{j}:\ d(x,y)\leq r\}\). Then, by Lemma 2.1 \[I_{j,\ell} =e^{-(n-1)r}\int_{F_{\ell}}\int_{\Omega_{j,\ell}^{y}}d\mu_{n}(x) \,w(y)d\mu_{n}(y)\] \[\leq C_{n}e^{-(n-1)r}e^{\frac{n-1}{2}(j+r-\ell)}\,w(F_{\ell}).\] So, \[I_{j,\ell}\leq C_{n}e^{-(n-1)r}\min\Big{\{}e^{(n-1)\frac{r+l-j}{2}(p-\delta)}e ^{(n-1)r\delta}w(E_{j}),e^{\frac{n-1}{2}(j+r-\ell)}\,w(F_{\ell})\Big{\}}.\] From now on, we can follow the same steps as in the proof of Lemma 3.2, and using Lemma 3.1 we obtain (3.10). ## 4. Examples In this last section we show several examples to clarify several points previously mentioned. We omit details since the examples follow from continue variants of Theorem 1.3 in [14]. Let \(-\infty<\gamma\leq 1\), we denote \[w_{\gamma}(x)=\frac{1}{\big{(}\,1+\mu_{n}(\,B(0,d_{H}(0,x)\,)\,)\,\big{)}^{ \gamma}}.\] **Examples 4.1**.: 1. If \(0\leq\gamma\leq 1\), then \[M(w_{\gamma})(x)\lesssim w_{\gamma}(x)\] In particular if \(\gamma<1\) taking \(s>1\) such that \(\gamma s\leq 1\) we have that \[M_{s}(w_{\gamma})(x)\lesssim w_{\gamma}(x)\] Therefore there are non trivial weights satisfying \(M_{s}(w)\lesssim w\). On the other hand, \(Mw_{1}(x)\lesssim w_{1}(x)\). However, the weak type \((1,1)\) of \(M\) with respect to \(w_{1}\) fails. In fact, taking \(f_{k}(x)=\chi_{\mathcal{C}_{k}}(x)\) for \(k\) big, it is not difficult to show that \(w_{1}\{x:M(f_{k})(x)>1/2\}\geq k\) and the \(L^{1}(w_{1})\)-norm of \(f_{k}\) is uniformly bounded. In particular, this example shows that in Theorem 1.1 is not possible to put \(s=1\). In fact, it is not possible to put any iteration \((M^{m}(f)=M(M^{m-1}f))\) of \(M\) for any fixed natural number \(m\). 2. Let \(p>1\). Then \(w_{1-p}(x)\) satisfies the hypothesis of Corollary 1.6 and therefore \[\|Mf\|_{L^{p},\infty(w_{1-p})}\lesssim\|f\|_{L^{p}(w_{1-p})}\] holds. Nevertheless, \(\|Mf\|_{L^{p}(w_{1-p})}\lesssim\|f\|_{L^{p}(w_{1-p})}\) does not. This can be seen by considering the function \(f=\chi_{B(0,1)}\), and taking into account that \(w\simeq(M\chi_{B(0,1)})^{1-p}\). 3. Fixed \(\gamma\in(0,1)\). We have seen in the item 1 that the maximal function satisfies a weak type \((1,1)\) inequality for this weight. In particular, for every \(q>1\), \[\|Mf\|_{L^{q}(w_{\gamma})}\lesssim\|f\|_{L^{q}(w_{\gamma})}.\] However, it is not difficult to see that, for any fixed \(p>1\), it holds that \[\sup_{r>0}\frac{1}{\mu_{n}(B(0,r))}\int_{B(0,r)}w_{\gamma}\left(\frac{1}{\mu_ {n}(B(0,r))}\int_{B(0,r)}w_{\gamma}^{-\frac{1}{p-1}}\right)^{p-1}=\infty.\] This example shows that boundedness of \(M\) does not imply the natural condition \(A_{p}\) for any \(p>1\) in this setting. In the Euclidean setting in the context of a general measure \(\mu\) an example in this line was also obtained by Lerner in [8]. ## Appendix A The ball model of the hyperbolic space Let \(\mathcal{B}_{n}=\{x\in\mathbb{R}^{n}:\ \|x\|<1\}\), where \(\|\cdot\|\) denotes the euclidean norm in \(\mathbb{R}^{n}\). In this ball we will consider the following Riemannian structure \[ds_{x}^{2}(v)=\frac{2\|v\|^{2}}{(1-\|x\|^{2})^{2}}.\] The hyperbolic distance in this model can be computed by \[d_{n}(x,y)=\operatorname{arctanh}\left(\frac{\|x-y\|}{(1-2\left\langle\,x,y \,\right\rangle+\|x\|^{2}\|y\|^{2})^{\frac{1}{2}}}\right).\] The group of isometries \(\mathcal{I}(\mathcal{B}_{n})\) in this representation coincides with the group of conformal diffeomorphisms from \(\mathcal{B}_{n}\) onto itself. For \(n=2\), we can identify \(\mathbb{R}^{2}\) with \(\mathbb{C}\), and this group is the one generated by: * Rotations: \(z\mapsto e^{it}z\), \(t\in\mathbb{R}\). * Mobius maps: \(z\mapsto\frac{z-w}{1-\bar{w}z}\). * Conjugation: \(z\mapsto\overline{z}\). For dimension \(n>2\), recall that, by Liouville's theorem, every conformal map between two domains of \(\mathbb{R}^{n}\) has the form \[x\mapsto\lambda A\circ\iota_{x_{0},\alpha}(x)+b\] where \(\lambda>0\), \(b\in\mathbb{R}^{n}\), \(A\) belongs to the orthogonal group \(O(n)\), and for \(x_{0}\in\mathbb{R}^{n}\), \(\alpha\in\mathbb{R}\) \[\iota_{x_{0},\alpha}(x)=\alpha\frac{x-x_{0}}{\|x-x_{0}\|^{2}}+x_{0}.\] Note that, when \(\alpha>0\), the maps \(\iota_{x_{0},\alpha}\) correspond to a reflection with respect to the sphere \[S^{n-1}(x_{0},\alpha)=\{x\in\mathbb{R}^{n}:\ \|x-x_{0}\|^{2}=\alpha\}.\] If \(\alpha<0\), it is a composition of the inversion with respect to the sphere \(S^{n-1}(x_{0},-\alpha)\) and the symmetry centered at \(x_{0}\). Using this result, we get that the group \(\mathcal{I}(\mathcal{B}_{n})\) consists of the maps of the form \[A\circ\theta\] where \(A\) belongs to the orthogonal group \(O(n)\) and \(\theta\) is either the identity or an inversion with respect to a sphere that intersect orthogonally \(\partial\mathcal{B}_{n}\). Recall that we say that two spheres \(S_{1}\) and \(S_{2}\) intersects orthogonally if for every \(p\in S_{1}\cap S_{2}\) \[(T_{p}S_{1})^{\perp}\perp(T_{p}S_{2})^{\perp}.\] **Remark A.1**.: This representation is also true for \(n=2\). Indeed, on the one hand, the rotations as well as the conjugation belongs to \(O(2)\). On the other hand, given \(\alpha\in\mathbb{C}\) such that \(|\alpha|<1\), the circle of center \(\alpha^{-1}\) and squared radius \(|\alpha|^{-2}-1\) is orthogonal to \(\partial\mathcal{B}_{2}\), and if \(\iota\) denotes the inversion with respect to this circle then \[\iota(z)=\frac{\overline{z}-w}{1-\bar{w}\overline{z}}.\] In this model, the \(r\)-dimensional hyperbolic subspaces that contains the origin are precisely the intersection the \(r\)-dimensional linear subspaces of \(\mathbb{R}^{d}\) with \(\mathcal{B}_{n}\). The other ones, are images of these ones by isometries. So, they are \(r\)-dimensional spheres orthogonal to \(\partial\mathcal{B}_{n}\). The orthogonality in this case, as before, is defined in the natural way in terms of the orthogonal complements of the corresponding tangent spaces.

2306.13894

OUXT Polaris: Autonomous Navigation System for the 2022 Maritime RobotX Challenge

OUXT-Polaris has been developing an autonomous navigation system by participating in the Maritime RobotX Challenge 2014, 2016, and 2018. In this paper, we describe the improvement of the previous vessel system. We also indicate the advantage of the improved design. Moreover, we describe the developing method under Covid-19 using simulation / miniture-size hardware and the feature components for the next RobotX Challenge.

Kenta Okamoto, Akihisa Nagata, Kyoma Arai, Yusei Nagao, Tatsuki Nishimura, Kento Hirogaki, Shunya Tanaka, Masato Kobayashi, Tatsuya Sanada, Masaya Kataoka

2023-06-24T07:57:42Z

http://arxiv.org/abs/2306.13894v1

# OUXT Polaris: Autonomous Navigation System for the 2022 Maritime RobotX Challenge ###### Abstract OUXT-Polaris has been developing an autonomous navigation system by participating in the Maritime RobotX Challenge 2014, 2016, and 2018. In this paper, we describe the improvement of the previous vessel system. We also indicate the advantage of the improved design. Moreover, we describe the developing method under Covid-19 using simulation / miniature hardware and the feature components for the next RobotX Challenge. Maritime systems, Robotics, Unmanned surface vehicle ## I Introduction First of all, we are motivated to develop a big field robot in a large area such as the ocean. In recent years, the aging and shrinking population, as well as a shortage of workers, has led to an increase in demand for the automation of cars, robots, and other equipment. Among these, automated driving is being developed with particular emphasis. Moving the autonomous vehicle or robot outside has a very severe problem. They need to hedge unknown obstacles and go to the target position. The environment such as weather, temperature, or underwater around robots causes sensor and hardware problems. There are each challenging problems and They are also interesting for us, and there are different problems between land and ocean. On the land, the navigation or the estimation of the self-position is solved by the point cloud map and the odmetry, while on the ocean, the point cloud and the odmetry is not obtained enough. So the robots need to estimate the self-position using GPS and IMU sensors. Moreover, on the land, the position of the target objects and obstacles is obtained from Lidar data. On the other hand in the sea, waves disturbe to get the target positions. In that case, the robots have to fusion multiple data such as cameras and Lidars. In this competition, we have a chance to develop a system to get over the wild environment for the robots on the ocean. Therefore, we are participating in the Maritime RobotX Challenge. As shown in Figs. 1, and 2, we designed the architecture of the vessel navigation system. Our vessel navigation systems are composed of localization, perception, behavior, planning, and control. Our localization, behavior, and planning methods are based on classical methods, such as "Extended Kalman Filter", "Behavior Tree", "Cubic Hermite Spline" and "Velocity Control based on WAM-V Dynamics Model" Our perception methods are based on learning methods, such as "YOLOX". We belielfly describe the developed vessel navigation system as follows. 1. _Localization_: The position and the velocity of WAM-V are estimated using the 6DoF Extended Kalman Filter [1] from the data of the GNSS and IMU sensors. 2. _Perception_: The obstacles and task objects are recognized from lidar and camera data. We used YOLOX for object Detection of task object information such as buoys and docks. 3. _Behavior_: Using behavior tree can build WAM-V behaviors like a tree. We use Groot 19, GUI tools for designing behavior tree for smooth development. 4. _Planning_: Path planning generates obstacle-avoidable paths from sensors and WAM-V information in real-time. 5. _Control_: In the servo and thruster controllers, the servo motor direction and the thruster revolution to achieve the target velocity are calculated based on the vessel motion model. ## III Hardware Developments ### _Redesign of Azimuth Thruster_ We considered the following three issues when designing the propulsion mechanism. First, we considered it important for the boat to be able to generate lateral propulsive force to complete the docking task. When propulsion units are mounted on the two aft sections of the boat, each propulsion unit must have a degree of freedom in the yaw axis to generate thrust in any direction in the horizontal plane. This propulsion system is generally called an azimuth thruster. Next, the electric outboard motors that are commonly available have a circular cross-section for the mounting shaft, so it is necessary to find a way to fix the shaft tightly. In addition, the propeller must avoid contact with the seafloor when the vessel needs to navigate in shallow water, such as when launching the boat on the course. Since it is dangerous for a person to enter shallow water and lift the thruster with a tool, a mechanism that can easily raise and lower the thruster was necessary. To meet these requirements, we designed the mechanism shown in Fig. 3. The three functions of gripping, rotating, and elevating are integrated into a single unit. The gripping function was realized using a PLA plastic collet manufactured by a 3D printer. The collet is pushed axially by a screw into an aluminum hollow shaft with a wedge-shaped cross section to enable strong shaft gripping. The azimuth mechanism is realized by transmitting the rotational force from the servo motor (XM540-W270, Dynamixel) to the hollow shaft by spur gears. The hollow shaft is held at two points by angular bearings. ### _Sensor Arrangement_ LiDAR and a visible light camera are used for environmental awareness. Their arrangement is shown in the Fig. 4. In total, 6 cameras and 4 LiDARs are used. Different LiDARs are used for the front and rear views and for the left and right views. The VLP-16 from Velodyne Lidar is used for the front/rear view to provide a wide range of vision, mainly in the direction of boat travel, and the MID-70 from Livox is selected for the left/right view to see the docking bay near the hull in the docking task. For the cameras, a module with an IMX-219 image sensor and a lens with a 120-degree diagonal field of view was used. This allows the acquisition of point cloud information and visible light images covering 360 degrees around the boat. ### _The Perception Array_ To perform point cloud fusion using a visible light camera and LiDAR, it is necessary to accurately calibrate the relative positions of the sensors. Therefore, once the sensors are assembled on the hull, they cannot be easily removed for testing on land. To solve this problem, we have developed a sensor unit that consists of a LiDAR and two cameras fixed to a rigid frame and can be mounted on or carried by various robots while maintaining the accuracy of the relative positioning between the sensors. We call it a perception array. The foreground and background views of the designed perception array are shown in Fig. 5 Fig. 4: Sensor Arrangement Fig. 3: Design of Azimuth Thruster and Fig. 6, respectively. The cameras are mounted on the left and right sides, and the LiDAR is mounted upside down at the bottom. The box in the center contains the power supply function and the switching hub. This configuration is shown in the Fig. 7. ### _The Perception Camera_ The camera, which is part of the Perception Array, is designed to perform edge-processing image recognition and is equipped with a Jetson Nano from Nvidia as the computing system. The camera is designed to be waterproof and heat-dissipating for use in various weather conditions. The system diagram is shown in Fig. 8. The front and rear hatches can be opened and closed without tools. Fig. 9 shows these hatches opened. ### _MINI-V in the COVID-19_ MINI-V(minitua vessel) was created in order to test the software easily in the Covid-19. Over the past several years, we couldn't conduct the experiment on the ocean or lake because of the COVID-19. We were prohibited to meet and create the parts of WAM-V. In addition, the law about vessels is very strict. So, we can't float the boat easily. The WAM-V is so big and it is hard work and costs too much to carry WAM-V to the lake. Then, we need a sustainable system to develop the automotive vessel. As mentioned above, the simulator is used for developing navigation systems, and it doesn't need to use WAM-V. The perception array was created to get the sensor data for software tests. They made it easier for us to develop software without ships. However, the software and hardware integration is the most important to conduct tasks. Then, MINI-V was created to make it easier to do the test and the integration. The concepts of MINI-V are follows: 1. easy assembly, transport, and experiment, 2. open source software and hardware, 3. high compatibility between WAM-V and MINI-V. Fig. 5: Front View of Perception Array Fig. 10: Hardware Components of MINI-V Fig. 6: Rear View of Perception Array Fig. 7: Diagram of Perception Array Fig. 9: Disassembled diagram of the Perception Camera MINI-V is created to be easy to carry, and we can carry them by suitcase like Fig. 10. It is so small that we can float and test on the buthtab. It is also assembled simply. We develop this vessel on open source. So, other people can play or test their software with MINI-V. Finally, we expect the high compatibility between WAM-V and MINI-V, and it will make it easy to migrate developed software on MINI-V toWAM-V. However, MINI-V have not had complete compatibility yet. We have future tasks to create a little bigger vessel to have compatible hardware and software such as batteries and sensors, and so on. ### _the prototype of the multicopter_ In the RobotX 2022, the tasks about multicopter are add on the competition. The drone needs to be automated and pick up the task objects. We created the quadcopter below in Fig.12. The flight controller, Pixacer Pro[2], is introduced to controll the drone. In order to get the drone position, it has GNSS sensor on the plane. we conduct the test of estimating the self-position of the drone. As you can see Fig.13 the drone can estimate the self-position With an error of \(\pm 0.5\mathrm{m}\) approximately. At the moment, the drone is manually controled by a pilot but not automated. For automation, we need to add computers such as raspberryPi to send the operation into flight controller. Moreover, the sensor such as cameras to recognize the task objects are needed. There are many considerations such as the roter-size, motor power, the size of body, and so on. We are going to develop the automated navigation system for drone in next years. ### _Emergency Stop System_ There are four emergency stop switches on the outside of the hull, two of the normal switch type and one wireless one. When any one of these switches is turned on, the propulsion system is turned off and the ship comes to a safe stop. Fig. 11: Experiment on Ai River Fig. 14: Emergency Stop System Diagram Fig. 13: Estimated Drone Position and Track In a normal design, a relay used in such a case would drain about 10 W from its coil alone, which is a very high cost for OUXT-Polaris, which does not have sufficient battery capacity. To avoid this, we designed a circuit that would function reliably as a kill switch when necessary while reducing the power consumption of the relay coil. Specifically, we selected a relay that applies voltage only at the moment it is turned on and consumes little power at other times, and designed a circuit to accommodate this. This design achieved a significant reduction in power consumption and greatly extended the cruising time. In addition, a latch circuit was introduced so that only momentary switches can be operated. This eliminates the possibility that the emergency stop switch, once pressed, would automatically deactivate the emergency stop status in the unlikely event that the contacts are unexpectedly removed. ## IV Software Developments ### _ROS2-based Autonomous Navigation Stack_ In "Maritime RobotX Challenge 2018", we used Robot Operating System 1(ROS1) for developing software. However, the development of ROS1 was finished with python2 end of life. Therefore, we adopted the next generation of ROS called "ROS2". [3] As shown in Fig. 1, our ROS2-based simulation and software system was already developed. Our software contributions are listed as follows. * _Software System_ : We rebuilt the software system from ROS1 to ROS2 * _Behavior Tree_: We adopted the behavior tree library in ROS2 and built our original behavior tree. * _Camera LiDAR object detection_: We are developing a lidar-camera fusion object detection system for this project. * _Simulation Tool Development_ : We developed LiDAR simulation by using intel ray-tracing OSS "Embree". * _Infrastructure_ We developed some automation tools for develop quickly. We published all codes in GitHub to give feedback knowledge to the ROS community and Open-Source all our resources not only software [5] but also CAD models, and circuit data. [6] ### _Software System Architecture_ Our navigation stack is based on ROS2, but we do not use the navigation2 library. We develop our original software. Our software is highly modularized, so some of our team members use our stacks in other autonomous mobility competitions. (Fig.17,Fig. 18) ### _Behavior Tree_ Using behavior tree can build WAM-V behaviors like a tree. We use Groot 19, GUI tools for designing behavior tree for smooth development. For example, in the WAM-V Dynamic Qualifying Task, we defined a node that navigation and searches channel markers. In searching channel markers behavior, first, we got the information from the result of lidar-camera fusion object detection system. Information we can get from the system is objects label, probabilities, and bounding boxes. And we convert the information we can get from the system into information containing channel markers information we can use in robotx challenge contains such as labels and coordinates in the same coordinate system of WAM-V etc. After search channel markers, WAM-V starts navigation. We need to act the Node two times in the task. So we defined two same nodes and connected them by Sequence node. The sequence node is defined in the behavior tree. We made other Nodes, for example, move forward node, stop node and rotate around the buoy node. These nodes can realize many behavior patterns required in tasks. ### _Planner components_ We created hermite planner as the path planning module for WAM-V. Once the goal direction information from the behavior layer is obtained, the path shape is created using the Hermite curve in combination with the current position. The Hermite curve is a kind of parametric curve with continuous curvature, and its shape is determined by specifying a vector between two endpoints. The reason for adopting Hermite curves is that Catmull-Rom Spline, etc. can be created by connecting Hermite curves, so the creation function can be used repeatedly. And since WAM-V itself can turn on the spot, WAM-V can follow the path such that angular velocity is large,even if the Hermite planner is created. The components of hermite planner are explained sequentially. The local waypoint server calculates the distance between the destination and obstacles in the Fresnel coordinate system. If a contact is detected on the route, a route is created to reach the destination with a slight shift in the X and Y directions in the world coordinate system. For contact judgment in the Fresnel coordinate system, the nearest neighbor points between the Hermite curve and the 2D LaserScan obtained from the obstacle are calculated using the Newton method. From among them, the collision between WAM-V and the obstacle objects is judged to have occurred if Satisfying following equations. \[0<t<1 \tag{1}\] \[f(t)<width+margin \tag{2}\] \[f(t)=at^{3}+bt^{2}+c^{2}+d \tag{3}\] When the Hermite curve is determined, the calculation of the velocity constraint is performed using several velocity modules. The first is stop planner, which creates a speed commitment to stop at the destination end of the route. This speed module creates a speed commitment to slow down at a constant acceleration before heading to the destination. Second, there is an obstacle planner to perform a stop before obstacle objects. This constraint to stop before the obstacle uses the same method as stop planner. The third is a curve planner to prevent the angular velocity from exceeding a certain value. It creates a velocity constraint so that the angular velocity between points on the created Hermite curve does not become too large. Each of these three creates velocity constraints independently, integrates the information from each of them. And, the graph adjusts the velocity constraints, and searches the velocity. Creating a velocity plan on the route by deleting edges with too much acceleration as edges with increasing or decreasing velocity between them as nodes on each curve. Then, based on the created route and the ship's own self-position, the pure pursuit planner is used to follow the route. A contact point between the circle centered on the WAM-V's self-position and the path is created, and the angular velocity and speed up to that contact point is calculated. In addition, a straight line is created on the extension of the endpoint to approximate the endpoint so that the endpoint does not run out of control. Fig.21 is shown as an the part of planning results by the above path planning module. Fig. 19: Groot Example Fig. 20: Pipeline between Perception to Behavior ### _Camera LiDAR fusion_ We used YOLOX[7] for object Detection of task object information such as buoys and docks. We created annotation data based on images and videos obtained from past RobotX Challenges. We open our annotation data for speeding up future development for Maritime RobotX. [8] The result of training and inference is shown in Fig.22. The image is part of a video recorded during our challenging navigation in the 2018 RobotX Challenge.[9] This model is based on the YOLOX-S network and converted into a tensort model. So, we can run an object detection model of more than 10 Hz in Jetson nano. The point cloud from LiDAR preprocesses before fusion the camera image and the LiDAR point cloud. First, The point cloud from the LiDAR is filtered to remove outliers. Next, the filtered point cloud is downsampled to a 2D Laser Scan. Then, the object area is extracted from the 2D LaserScan. [10] The extraction algorithm determines the object area information based on whether the distance between the neighboring LaserScan is less than the threshold value determined adaptively. The clustered point cloud is projected on the camera images. And we will match the bounding 2D IOU and the object label on the images by Hungarian method.[11] ### _Simulation_ WAM-V has a car-sized hull, and the setup alone consumes an entire day to conduct the experiment. The development of a simulator is essential to speed up the development process. Therefore, we developed a simulator called navi_sim. The reason we did not use vxx is that vxx is not yet compatible with ROS2. navi_sim utilizes Embree [12], an open-source ray tracing library developed by Intel, to simulate lidar with fast CPU ray tracing. It also provides various other functions such as semantic information output, camera view simulation, and simple ship motion model simulation. Blue rectangles in Fig. 24 is camera view simulation, and robot 3D models in 24 is simulated robot. ### _Infrastructure_ Our system runs in a complex distributed computing environment that includes microcontrollers, so the configuration file is huge and there are many processes that need to be launched to start the system. If such a system is deployed manually, various human errors are expected to occur. In addition, there are 77 ROS2 packages that make up our autonomous navigation system, and it is almost impossible to manage all of them manually without error. To solve these problems, we used a configuration management tool called ansible, ansible can support environment construction Fig. 21: Part of Planning Results Fig. 24: Running Our Navigation Stack with Navi Sim Fig. 23: Algorithm of scan segmentation Fig. 22: Inference procedures on the computer in YAML format and can switch between development and production environments with a single option. With the adoption of ansible, setting up our system is just a matter of running a single shell script. Also, even if the ROS2 package does not have any changes in the code of the package itself, the software may break if there are changes in the packages on which the package depends. To detect and fix this quickly, we built a CI system using GitHub Actions. CI is a technology called continuous integration, which detects commits to Github, etc., and automatically performs tests to find defects early. When a pull request is issued for any of the packages we manage, a build test is automatically performed once a day. If the build test fails, the pull request cannot be merged. Failed build tests are notified to the OUXT-Polaris Slack so that members know immediately if a failure has occurred. Whole architecture of CI/CD pipeline shown in Fig.25. The robot also consists of a single-board computer with an arm CPU as well as a CPU with x86-based architecture, and a microcontroller with an arm CPU and mbed OS. The infrastructure supports all of these architectures and can test all software from deep-learning recognition systems to low-layer control systems. We divide our packages into smaller pieces in order to maintain the high degree of component nature of the packages we develop. This reduces unnecessary build targets and greatly reduces CI time (tests that originally took 30 minutes or more are now only 3 minutes). However, since there are more than 40 packages under our control alone, it is impossible to keep track of their development and testing status without tools. Therefore, we built a system that calls the API of GitHub and automatically creates a dashboard in Fig.26. All of these deliverables are also open-source software. We also developed GitHub Actions to configure GitHub Actions to unify CI procedures for multiple repositories and built a system to synchronize CI procedures at all times via bot accounts. [13] GitHub Actions is also used when deploying to the actual machine. GitHub Actions has a function called Self-Hosted Runner, which allows you to remotely execute any command using your own computer. Using this function, after the CI of the ROS2 package is completed, a configuration file with all commit hashes fixed is created and deployed to the actual device using Github Actions/ansible based on that configuration file. This allows our team to deploy the verified and up-to-date source code to the machine at the push of a button. In addition, machine learning is a task that requires a variety of complicated labor and a machine with a high-performance GPU. However, most tasks are routine and human intervention is not productive. To give some concrete examples, the following tasks can be automated. * _Conversion of dataset formats_ : We use labelme for annotation tools, but YOLOX use coco format for learning. * _Performing training_ : Virtualize the environment with nvidia docker and run scripts for learning on it * _Visualization and versioning of training results_ : We can visualize inference result via Twitter bot(Fig.[17]) and tensorbord.dev(Fig.27) * _Model Conversion_ : Need to convert PyTorch training results into a tensorrt model for faster inference. All tasks are executed using the Self-hosted runner in GitHub actions, which automatically executes the training when it detects additional events in the machine learning data set. Training takes about 30 minutes when we train YOLOX-S model. Future work includes building a mechanism to automatically deploy the obtained learning results to the actual machine and Neural Architecture Search using black-box optimization. ### _Communication with Technical Director Server_ We must report the vehicle's status as the Autonomous Maritime System heartbeat to a Technical Director server via a TCP connection. Our team usually tests navigation systems using a simulation environment, so we should check the heartbeat sentence on the simulation loop. To achieve it, we built a simple mock server[14]. It displays received NMEA-like sentences from a simulation via a TCP connection. We Fig. 26: CI/CD Dashboard Fig. 27: Training Result in Tensorord.dev Fig. 25: CI/CD Pipeline also implemented a ROS2 node that collects information about the aircraft and sends heartbeats. ### _WAM-V Controller_ The WAM-V control system adopted the ros2_control[15, 16]. The ros2_control is a framework for real-time control of robots using ROS2. Our WAM-V controllers are listed as follows. #### Iv-I1 Equation of Motion Our WAM-V equation of motion is calculated as follows. \[M\dot{\nu}=-D(\nu)\nu+\tau \tag{4}\] \[\nu=\begin{bmatrix}u&v&\omega\end{bmatrix}^{T} \tag{5}\] \[\tau=\begin{bmatrix}f_{x}&f_{y}&f_{yaw}\end{bmatrix}^{T} \tag{6}\] \[M=\begin{bmatrix}m+m_{x}&0&0\\ 0&m+m_{y}&0\\ 0&0&I_{z}+J_{z}\end{bmatrix} \tag{7}\] \[D(\nu)=\begin{bmatrix}0&0&-mu\\ 0&0&mv\\ mu&-mu&0\end{bmatrix} \tag{8}\] where \(\nu\) is WAM-V velocity in WAM-V coordinate system. \(\tau\) is WAM-V coordinate system input. \(M\) is the mass matrix. \(m\) is the WAM-V mass. \(m_{x}\) and \(m_{y}\) are the added mass of WAM-V in X and Y directions. \(I_{z}\) is the WAM-V moment of inertia. \(J_{z}\) is the WAM-V added moment of inertia. \(D(\nu)\) is the WAM-V resistance coefficient matrix. #### Iv-I2 Differential Thruster Model \[\tau = \begin{bmatrix}f_{x}\\ f_{y}\\ f_{yaw}\end{bmatrix} \tag{9}\] \[= \begin{bmatrix}\dfrac{1}{2}(F_{r}+F_{l})\\ 0\\ \dfrac{w}{2}(F_{r}-F_{l})\end{bmatrix}\] where \(Fr\) and \(F_{l}\) are propulsion of right and left side thruster. \(w\) is the distance between right and left thrusters. #### Iv-I3 Propeller Model \[T = \rho n^{2}D^{4}K_{t}(J_{s}) \tag{10}\] \[K_{t}(J_{s}) = k_{2}J_{s}^{2}+k_{1}J_{s}+k_{0}\] (11) \[J_{s} = \dfrac{up}{nD} \tag{12}\] where, \(T\), \(\rho\), and \(D\) are the propeller thrust, fluid density, and propeller radius. \(k_{0}\), \(k_{1}\) and \(k_{2}\) are constants values. \(up\) and \(n\) are inflow rate and rotation speed. ## V Firmware We use the same microcontroller for two main purposes. The reason for using the same microcontroller is to make it easier to have a spare in case of failure during the competition. We considered several types of microcontrollers and finally adopted NUCLEO-F767ZI (Fig.28) from STM Corp. The two roles of the microcontroller are to drive the motor and monitor the supply voltage. In order to meet the specifications for designing a real-time guaranteed control system for the motor drive, we implemented firmware that connects ROS2 through packet communication with the speed control system and UDP communication, which were specified in the previous chapter. For the other role of monitoring the power supply voltage, a real-time guarantee was not necessary. Instead, it was necessary to be able to easily communicate with ROS2 via Pub/Sub, so mROS2, an embedded communication library compatible with ROS2, was adopted to realize communication between the lower layers and higher layers using the ROS2 system. ## VI Conclusion In this paper, OUXT-Polaris reported the development of the autonomous navigation system for the 2022 RobotX Challenge. Based on the results of the 2018 RobotX Challenge, OUXT-Polaris rebuilt the system and developed improved systems for the Maritime RobotX Challenge 2022. We succeeded in constructing the highly reusable system by designing systems with high independence as parts, in addition to high computing capacity and environmental recognition capability. Fig. 28: NUCLEO-F767ZI Microcontroller Moreover, we described the developing method in Covid-19 and the feature components for the next RobotX Challenge. We hope these significant upgrades will produce positive results in the next competition.

2308.05582

Usability Assessment of the OnlyKey Hardware Two-Factor Authentication Key Among Low Vision or Blind Users

Hardware security keys undoubtedly have advantage for users as "usability" pain is trivial compared to the maximum "security" gain in authentication. Naturally, the hardware factor in the authentication received a widespread adoption amongst average users, as it is ergonomically less demanding than phone texts or authentication prompts. This ergonomic advantage in particular is essential for users who are blind or low vision, as their interaction with a phone is impractical. However, the "usability" for low vision or blind users pain might be much higher than an average well-bodied user for the same "security" gain. In an effort to learn more we conducted a usability assessment with ten low vision or blind users setting up the OnlyKey two-factor authentication key. First, the setup process was insurmountable for more than half of the participants, resulting in a situation where the hardware key was abandoned. Secondly, the lack of tactile orientation led participants to consider it as both impractical, and prone to difficulties locating or loosing it. We discuss the implications of our findings for future improvements in usable authentication for visually impaired users.

Aziz Zeidieh, Filipo Sharevski

2023-08-10T13:46:36Z

http://arxiv.org/abs/2308.05582v1

Usability Assessment of the OnlyKey Hardware Two-Factor Authentication Key Among Low Vision or Blind Users ###### Abstract Hardware security keys undoubtedly have advantage for users as "usability" pain is trivial compared to the maximum "security" gain in authentication. Naturally, the hardware factor in the authentication received a widespread adoption amongst average users, as it is ergonomically less demanding than phone texts or authentication prompts. This ergonomic advantage in particular is essential for users who are blind or low vision, as their interaction with a phone is impractical. However, the "usability" for low vision or blind users" pain might be much higher than an average well-bodied user for the same "security" gain. In an effort to learn more we conducted a usability assessment with ten low vision or blind users setting up the OnlyKey two-factor authentication key. First, the setup process was insurmountable for more than half of the participants, resulting in a situation where the hardware key was abandoned. Secondly, the lack of tactile orientation led participants to consider it as both impractical, and prone to difficulties locating or loosing it. We discuss the implications of our findings for future improvements in usable authentication for visually impaired users. ## 1 Introduction Past research in the area of hardware security key usability [6, 10] has primarily centered around non-disabled users. Narrowing down the scope of research to people who are low vision or blind, the available work on the topic of hardware security key accessibility and usability is quite paltry. This is surprising, given the existence of frameworks and agendas for designing inclusive security and privacy mechanisms [13, 14]. Barbosa et al. designed a password manager application for people who are low vision and blind [4]. Azenkot et al. developed PassChods, a visual authentication method for touch surfaces that is robust to aural and visual eavesdropping [3]. Another work along this lines is BraillePassword by Alinfai et al., a web based authentication mechanism for blind users [2]. These works are important in addressing the needs of the visually impaired users but none address the case of second factor of authentication when it comes to usability of external hardware keys. ## 2 Research Study This research evaluates the accessibility and usability of OnlyKey, a hardware security key that offers multi-factor authentication, and password management functionality show in in Figure 1. OnlyKey is a USB device with the form factor of a common flash drive. It comes fitted in removable silicon case. On one side, OnlyKey has six solid state buttons, numbered one thru six resembling a six dot braille cell, and on the reverse side there is an LED indicator light. While the arrangement of the physical solid state buttons is identical to a six dot braille cell, the numbering does not conform to the braille cell dot numbers. The male USB A plug on OnlyKey has the PCB exposed instead of encasing it in a metal structure as commonly seen in USB devices outside of the hardware security industry. The top edge of OnlyKey has a key ring hole, and comes with a quick release connector that can be used to connect OnlyKey to the user's keys. ### Sampling Our research will exclusively focus on the accessibility and usability of the OnlyKey hardware security key, by users who are low vision or blind. OnlyKey is a hardware security key that offers password manager and multi factor authentication functionality, similar to the YubiKey by Yubico [5, 9, 15]. We got approval from our Institutional Review Board (IRB) to conduct semi-structured interviews with a sample of ten visually impaired users from the United States. All participants were 18 years or older. We used snowball sampling technique as the initial couple of participants were asked to recommend another participant with visual impairments for the study. The interviews were conducted over Zoom, as in-person presence was restricted by the university policy. The interviews lasted between 60 and 90 minutes, and users were shipped the OnlyKey hardware key beforehand. Compensation for participation in the research study was an OnlyKey hardware security key and a $25 Amazon gift card. ### Data Collection The data collection for this research was primarily made up of two parts. The first part was the hands-on portion of the interview, where participants were directed to familiarize themselves with the OnlyKey, and set it up using the OnlyKey Quick Setup option. The second part of the interview script consisted of thirteen questions, ten were Likert scale questions adapted from the Accessible Usability Scale (AUS) [12], whereas the final three questions, were open-ended questions, designed to gauge participants attitude toward the accessibility and usability of OnlyKey, given their experience with it in the hands-on portion of the interview. ## 3 Results Ten people who are low vision or blind agreed to participate in this research study as interviewees. Participants will be herein referred to as **P1** thru **P10**. Eight of the participants identified as male while two identified as female. Three participants identified themselves as having low vision with limited visual acuity (B3), two participants identified as being totally blind and cannot see any lights or shapes (B1), two as being blind and having visual perception to see only lights and shapes (B2), two participants who stated they had low vision with high visual acuity (B3), and one participant identified as being low vision but not legally blind (B4). For the remainder of this paper, participants will be categorized into one of four visual classifications - B1 through B4 - based off of their response to the visual perception question in the demographics section of the interview. These four classifications are used by the United States Association of Blind Athletes (USABA) [11]. Participants' responses to the AUS were calculated and included in Table 3 next to their visual classification level. For the positively worded statements (questions 1, 3, 5, 7, and 9), the score contribution is identified by taking the scale position and subtracting 1. Then, multiplying the resulting number by 2.5. For the negatively worded statements (questions 2, 4, 6, 8, and 10), the score contribution is 5 minus the scale position. Then, multiplying the resulting number by 2.5. [12]. The minimum possible AUS score is 0 while the maximum possible AUS score is 100. The average AUS score across all research participants shown in the table above is 46.5. ### Initial Observations Research participants were asked to take time to familiarize themselves with the OnlyKey. **P6**, **P7**, **P8** and **P10** described the OnlyKey keypad as resembling a Braille cell. Shortly thereafter, all participants who's visual perception fell into the B1, B2 and B3 categories, excluding **P10**, expressed concern regarding the layout of the OnlyKey keypad, having no knowledge of which buttons are associated with which numbers. After reading the OnlyKey user manual, participant **P5** asked "wait, which button is which though?" Participants were also thrown off by the atypical USB A plug on OnlyKey. While this style of USB A plug is common with security keys like the YubiKey, all participants in this research had no prior experience with any security keys. Some participants would later go on to plug the OnlyKey into the computer upside down. **P4** initially identified the six buttons on OnlyKey and referred to them as a "netting woven into the device." **P4**, **P5** and **P7** used an app called _Be My Eyes_[1]. **P10** did not use assistive technologies during the entire span of the research study. They were unsure if the keys on \begin{table} \begin{tabular}{|c|c|c|} \hline **Participant** & **Visual Classification** & **AUS Score** \\ \hline P1 & B3 & 47.5 \\ \hline P2 & B3 & 52.5 \\ \hline P3 & B4 & 37.5 \\ \hline P4 & B3 & 35 \\ \hline P5 & B2 & 37.5 \\ \hline P6 & B2 & 37.5 \\ \hline P7 & B1 & 47.5 \\ \hline P8 & B3 & 65 \\ \hline P9 & B1 & 55 \\ \hline P10 & B3 & 50 \\ \hline \end{tabular} \end{table} Table 1: AUS and Visual Classification of the Sample Figure 1: An image of the OnlyKey hardware security key OnlyKey were "buttons" but they stated "I guess that will be something I'll find out in setup?" **P6** employed the assistance of their sighted relative to familiarize themselves with the OnlyKey keypad layout. They asked their relative what the orientation of the keypad was, and where each number key was. **P5**, **P7** and **P8** attempted to locate a written description of the OnlyKey layout in the OnlyKey manual and on Google, they were unable to find anything to help address this inquiry. ### OnlyKey Quick Setup **P4**, **P5**, **P6**, **P7** and **P8** initially found and attempted to watch the official OnlyKey video on how to set up the security key [8]. However, the video lacked narration nor spoken feedback, only text instructions, visual graphics, and a background instrumental which was of no use to the participants. **P3** and **P10** had enough usable vision to read the text on the package and successfully scan the QR code which would take them to the OnlyKey "start" page. **P1**, **P2**, **P4**, **P5**, **P6**, **P7** and **P8** used an app called _Seeing AI_, to read the text off of the OnlyKey packaging. _Seeing AI_ is an app by Microsoft that enables people who are low vision or blind to see the world around them with artificial intelligence. [7]. A photo of the OnlyKey retail packaging is in the appendix. While **P1**, **P2**, and **P4** were able to successfully get the URL of the OnlyKey online start page, **P5**, **P6**, and **P7** struggled with _Seeing AI_ to get this information, abandoning the packaging as a source of information. **P8** had some usable vision and noticed the QR code on the back of the OnlyKey package and attempted to use the product identification function of _Seeing AI_ to scan the QR code. _Seeing AI_ does not recognize QR codes in the product identification mode. Those who did not rely on the packaging as a source of information used Google or Bing. All participants ultimately found the correct instructions to follow for the OnlyKey Quick Setup process. **P1**, **P3**, **P5** and **P10** were the only participants who were able to successfully set up the OnlyKey and know the PIN configured on the device after the setup process. **P2**, **P4**, **P6**, **P7**, **P8** and **P9** were technically able to set up the OnlyKey, however, they did not know the PIN that was configured on the OnlyKey. This resulted in them being locked out of OnlyKey requiring them to reset it to factory defaults after the interview. The OnlyKey Quick Setup method requires that a user plug the OnlyKey into their computer, open up a text editor, (NotePad on Windows or TextEdit on macOS), then press and hold the 3 key on the OnlyKey keypad for five or more seconds then release. With the OnlyKey acting as a keyboard connected to the computer, it types out text at the point of the cursor. At this point of this process, the text "you have 20 seconds" is printed out referring to the time the user has to choose between a custom PIN, or a randomly generated PIN, with randomly generated being the default if no response is recognized after 20 seconds. **P2**, **P4**, **P6**, **P7**, **P8** and **P9** all used a screen reader. This is important because those who tried interacting with their computer while the OnlyKey setup process was in progress shifted the focus of the cursor resulting in text being typed by OnlyKey to go elsewhere. For example, **P4** could hear their screen reader echoing out typed text as the OnlyKey printed out the instructions, however, the printed text was intelligible given the speed at which the OnlyKey was typing. **P4** tried to read through the typed out text with their screen reader by using the arrow keys to navigate the text, which also shifted the position of the cursor while the OnlyKey was still printing out instructions. This resulted in the OnlyKey instructions being printed out of order and in a jumble which would ultimately lead to the OnlyKey getting setup with the PIN in the text file that the participant was unable to discern. **P6** and **P7** were pressing modifier keys associated with functions of the screen reader to read through the document, while OnlyKey was still printing out text. This resulted in shifting the focus of the cursor outside of the text document all together. In **P6**'s case, the focus of the cursor shifted outside the text editor which resulted in output from OnlyKey being entered in other open applications. **P10** was able to see the screen and did not interact with the computer while OnlyKey went through the Quick Setup process of typing out text in NotePad. While the OnlyKey was typing out text, **P10** was reading through the manual and missed the 20 second prompt asking if they want to choose a custom PIN or have one randomly generated. The OnlyKey by default opted to generate random PINs, printed them out in the text document, and finalized the setup process while P10 was reading the user manual. **P4**, **P6**, **P7** and **P8** realized that the setup had not gone as planned and deleted whatever text was in the text document, unplugged the OnlyKey, plugged it back in, and proceeded with the OnlyKey Quick Setup steps once more. They were confused as to why this did not work again as planned expressing frustration with **P6** saying: "what! It's not doing it anymore! I'm doing the same thing!" At this point, the participant's OnlyKey was setup and configured with a PIN which means the OnlyKey Quick Setup would not be available anymore. After participants exhibited signs of frustration and stress over this process, the researcher intervened to notify the participant that the OnlyKey had been setup at this point, and explained how that came to be. **P1**, **P3** and **P10** were able to set up the OnlyKey following the OnlyKey Quick Setup instructions outlined in the manual. **P5** assumed that only one key on the OnlyKey keypad would result in text output as part of the OnlyKey Quick setup process, so their initial approach was to press and hold random keys for five or more seconds until they got the expected result. In **P5**'s efforts of finding the 3 key, they eventually were able to get the OnlyKey to output instructions and PINs. However, instead of the 3 key, **P5** had randomly chose the 2 key. Pressing and holding the 2 key for five or more seconds then releasing results in OnlyKey going through the Quick Setup process with "random PINs generation" as the behavior instead of offering the prompt during what would have been the traditional Quick Setup through the 3 key. After **P5** reviewed the content in the text document, they were able to get the PIN they needed to unlock the OnlyKey, but at this point, **P5** still did not know the exact layout of the OnlyKey keypad, and this is when they called a sighted volunteer through _Be My Eyes_ to ask for a verbal description and orientation of the OnlyKey keypad [1]. At this point, all participants, regardless of ability to set up OnlyKey, were debriefed prior to proceeding to the post-experiment survey. All participants were made aware of the alternative method of setting up OnlyKey, which involved an application. Participants who were unable to set OnlyKey up were provided details on what issues they encountered. ### Post Experiment As part of the post-experiment survey, participants were asked three open-ended questions about the OnlyKey. The first question asked what they like about OnlyKey from an accessibility standpoint, while the second asked what they disliked. The third and final open-ended question asked for any questions, comments, concerns, or complaints the participant may have had regarding OnlyKey, both for accessibility and in general. Participants expressed interest in what the OnlyKey had to offer in terms of features and functionality. OnlyKey's familiar form factor was a point brought up by some participants as a positive aspect. Its also important to note that the form factor and design of OnlyKey was determined to be a negative aspect of the device by other participants. Participant attitudes towards the OnlyKey throughout the hands-on experiment became almost predictable after the researchers completed a few interviews with prior subjects. After participants were introduced to the OnlyKey, they expressed a sense of excitement and curiosity, however, as the hands-on experiment progressed, their excitement dwindled, and their curiosity would morph into frustration and confusion. A primary aspect of the OnlyKey that caused this predictable transformation mid-interview can be attributed to the physical design of the OnlyKey, more specifically, the absence of device feedback that can be interpreted by someone who is low vision or blind, similar to tactile or auditory feedback. Since OnlyKey has solid state buttons, the only feedback is visual, through the single LED indicator light on the back of OnlyKey. Participants noted this flaw in their responses to open-ended questionswith **P3** saying "I don't like the setup process. It would be nice if there was non-visual feedback when clicking the buttons, not just the light." **P10** was able to eloquently summarize the majority of complaints brought up by prior participants in their response "I disliked that the numbers did not correspond with the braille layout. It took me a moment to realize the buttons were not 'clicky" buttons, and I did not like how it only gave me 20 seconds." Another notable complaint shared by participants was the lack of detail in the instructions for a user who is low vision or blind. The instructions provided no verbal description of the OnlyKey's keypad layout, and the official OnlyKey instructional videos had no spoken feedback, only visuals and instrumental. All participant responses to the open-ended questions can be found in the appendix. Qualitative findings for this research study were analyzed manually by the researchers. Code books were not used in the analysis of these findings. ## 4 Discussion and Conclusion The objective of this study was to evaluate the accessibility of setting up the OnlyKey by people who are low vision or blind. A majority of participants were unsuccessful in the setup of OnlyKey with 60% of participants ultimately being unable to achieve this task. Of the four participants who were able to set the OnlyKey up successfully, three had usable vision, and the fourth only had perception of light. The participant who was blind and can only see lights and shapes relied on the help of a sighted volunteer from _Be My Eyes_[1]. Our usability assessment of OnlyKey as a Current-Of-The-Shelf (COTS) hardware authentication key strongly indicates that the design fails to be inclusive for the usability of the visually impaired users. This is problematic as these users are deprived of the opportunity to benefit from most if not all functionalities provided by OnlyKey. No hardware security key on the market as of the time of this writing offers all the functionality and versatility that OnlyKey offers, which forces this population to compromise maximum security and opt for an inferior security key. We acknowledge that we had a limited number of participants in this hands-on study in comparison with the general population of individuals who are low vision or blind. We also are aware that the 60 to 90 minutes should have not been enough for the participants to familiarize with the OnlyKey properly. Another notable limitation was the choice of setup option of the OnlyKey. This research evaluated the accessibility of OnlyKey using the OnlyKey Quick Setup option and did not explore the OnlyKey desktop software. Findings gathered by this research were not interpreted based on the technical aptitude of participants. Result interpretation was focused on participant's visual perception without regard for age, gender, education, prior knowledge of hardware security keys, or proficiency with computers. ## Acknowledgements A special thank you to Amy Gabre, Jackie Jackson, Jeni Shaum, Sue Dalton, and Wendy Brusich.

2305.14315

Estimating a multivariate Lévy density based on discrete observations

Existing results for the estimation of the L\'evy measure are mostly limited to the onedimensional setting. We apply the spectral method to multidimensional L\'evy processes in order to construct a nonparametric estimator for the multivariate jump distribution. We prove convergence rates for the uniform estimation error under both a low- and a high-frequency observation regime. The method is robust to various dependence structures. Along the way, we present a uniform risk bound for the multivariate empirical characteristic function and its partial derivatives. The method is illustrated with simulation examples.

Maximilian F. Steffen

2023-05-23T17:51:00Z

http://arxiv.org/abs/2305.14315v1

# Estimating a multivariate Levy density ###### Abstract Existing results for the estimation of the Levy measure are mostly limited to the onedimensional setting. We apply the spectral method to multidimensional Levy processes in order to construct a nonparametric estimator for the multivariate jump distribution. We prove convergence rates for the uniform estimation error under both a low- and a high-frequency observation regime. The method is robust to various dependence structures. Along the way, we present a uniform risk bound for the multivariate empirical characteristic function and its partial derivatives. The method is illustrated with simulation examples. **Keywords:** Levy processes, jump processes, multivariate density estimation, spectral methods, low-frequency, high-frequency **MSC 2020:** 62G05, 62G07, 62M15, 60G51 ## 1 Introduction Levy processes are a staple to model continuous-time phenomena involving jumps, for instance in physics and finance, see Woyczynski (2001) and Cont & Tankov (2004), respectively. Naturally, manifold such applications call for multivariate processes which significantly complicates the theoretical analysis compared to the onedimensional case. Matters worsen as practitioners often only have time-discrete data at their disposal which obstructs the identification of the jumps and hence the calibration of such models. Statistical results in this setting are typically limited to the onedimensional case or omit the estimation of the jump distribution, despite the practical relevance. In the present work, we study exactly this problem: estimating the jump distribution of a multivariate Levy process based on discrete observations. On a theoretical level, the distribution of the Levy process is uniquely determined by its characteristic triplet, that is, the volatility-matrix, the drift and the Levy measure. The latter characterizes the jump distribution we are interested in. From a statistical point of view, the estimation of the Levy measure is most challenging as we are faced with a nonparametric problem. The literature commonly distinguishes the following two observation regimes for a Levy process observed at equidistant time points \(0<\delta,2\delta,\ldots,n\delta\rightleftharpoons T\): Under the low-frequency regime, \(\delta\) is fixed as \(n\to\infty\), whereas \(\delta\searrow 0\) under the high-frequency regime. Our estimation method is robust across sampling frequencies. Motivated by the clearer separation between the jumps themselves and the Gaussian component as \(\delta\searrow 0\) under the high-frequency regime, threshold-based estimators have been applied extensively. Beyond the overview given by Ait-Sahalia & Jacod (2012), Duval & Mariucci (2021) apply such an approach to the estimation of the Levy measure, Gegler & Stadtmuller (2010) study the estimation of the entire Levy triplet and Mies (2020) estimates the Blumenthal-Getoor index. However, these references are restricted to the onedimensional case and multidimensional extensions seem tedious due to the multitude of directions in which the process can jump. A notable exception is the work by Bucher and Vetter (2013), who estimate the tail-integrals of a multivariate Levy process. Under the low-freqency regime, we cannot identify the intermittent jumps even in the absence of a Gaussian component resulting in an ill-posed inverse problem, see Neumann and Reiss (2009). A popular way out is the spectral method, see Belomestny and Reiss (2015), which leverages the relationship of the Levy triplet with the characteristic function of the process at any time point. Turning the observations of the process into increments, this characteristic function is estimated and then used to draw inference on parts of the Levy triplet. The method was first considered by Belomestny and Reiss (2006) in the context of exponential Levy models and has since been studied extensively, see Belomestny (2010), Gugushvili (2012), Nickl and Reiss (2012), Reiss (2013), Trabs (2015). We adapt this approach to the multivariate setting by constructing a nonparametric estimator for the Levy density \(\nu\), assuming that it exists. Our estimator requires no knowledge of the volatility and drift parameters and works uniformly over fully nonparametric classes of Levy processes with mild assumptions. In particular, Levy processes with infinite jump activity are allowed. The uniform rates we achieve naturally extend those from the onedimsional case and optimality in our setting is discussed. When estimating the Levy density close to the origin, we enhance our method with an estimator for the volatility. The estimation of the volatility matrix itself has previously been studied, see Papagiannouli (2020, high-frequency) and Belomestny and Trabs (2018, low-frequency). A related issue is the estimation of the covariance matrix in deconvolution problems, see Belomestny et al. (2019). However, even the proven minimax-optimal rates of convergence are too slow as to not affect our overall rates under the low-frequency regime. It is sufficient to estimate the trace of the volatility matrix and we show that this can be done with a much faster rate. With this enhancement, there is no additional loss in the rate for the estimation of the Levy density. An effect emerging for multivariate processes is the possibility of different dependence structures between the components which can be in disagreement with the existence of a Levy density in the form of a Lebesgue density on the whole state space. Statistical results in such settings are even rarer. Belomestny (2011) estimates the Levy density for a time changed Levy process with independent components. We propose a quantification of the estimation error when integrating against regular test functions under various forms of dependence structures without modifications to our method. The paper is organized as follows. In Section 2, we introduce the estimation method and state our main results along with a short outline of the proof and the key tools used. The empirical performance of our estimator is illustrated in a simulation examples in Section 3. The full proofs are postponed to Section 4. ## 2 Estimation method and main results We begin by introducing some notation: Throughout, an \(\mathbb{R}^{d}\)-valued Levy process \((L_{t})_{t\geqslant 0}\) with Levy measure \(\nu\) is observed in the form of of \(n\in\mathbb{N}\) increments at equidistant time points with time difference \(\delta>0\) and overall time horizon \(T\coloneqq n\delta\): \[Y_{k}\coloneqq L_{\delta k}-L_{\delta(k-1)},\qquad k=1,\ldots,n.\] For \(x,y\in\mathbb{C}^{d}\) and \(p\geqslant 1\), set \(|x|_{p}\coloneqq\big{(}\sum_{k=1}^{d}|x_{k}|^{p}\big{)}^{1/p}\), \(|x|\coloneqq|x|_{2}\), \(|x|_{\infty}\coloneqq\max_{k=1,\ldots,d}|x_{k}|\), \(x\cdot y\coloneqq\langle x,y\rangle\coloneqq\sum_{k=1}^{d}x_{k}y_{k}\) and \(x^{2}\coloneqq x\cdot x\). For a multi-index \(\beta\in\mathbb{N}_{0}^{d}\), we set \(x^{\beta}\coloneqq\prod_{k=1}^{d}x_{k}^{\beta_{k}}\), \(|x|^{\beta}\coloneqq\prod_{k=1}^{d}|x_{k}|^{\beta_{k}}\). If \(\int|x|^{2}\,\nu(\mathrm{d}x)<\infty\), then the characteristic function of \(L_{t}\) is given by \[\varphi_{t}(u)\coloneqq\mathbb{E}[\mathrm{e}^{\mathrm{i}(u,L_{t})}]=\mathrm{ e}^{t\psi(u)}\qquad\text{with}\qquad\psi(u)\coloneqq\mathrm{i}\langle\gamma,u \rangle-\frac{1}{2}\langle u,\Sigma u\rangle+\int\left(\mathrm{e}^{\mathrm{i} (u,x)}-1-\mathrm{i}\langle u,x\rangle\right)\nu(\mathrm{d}x)\] for some drift parameter \(\gamma\in\mathbb{R}^{d}\) and some positive semidefinite volatility matrix \(\Sigma\in\mathbb{R}^{d\times d}\), see Sato (1999). Denoting by \(\Delta\) the Laplace operator, i.e. \[\Delta g\coloneqq\sum_{k=1}^{d}\frac{\partial^{2}g}{\partial u_{k}^{2}}\] for a function \(g\colon\mathbb{R}^{d}\to\mathbb{C}\) which is twice differentiable in every direction, we have \[\nabla\psi(u) =\mathrm{i}\gamma-\Sigma u+\mathrm{i}\int x\big{(}\mathrm{e}^{i\langle u,x\rangle}-1\big{)}\,\nu(\mathrm{d}x), \tag{1}\] \[\Delta\psi(u) =-\operatorname{tr}(\Sigma)-\int|x|^{2}\mathrm{e}^{i\langle u,x \rangle}\,\nu(\mathrm{d}x)=-\operatorname{tr}(\Sigma)-\mathcal{F}[|x|^{2}\nu]( u)=\frac{\varphi_{t}(u)\Delta\varphi_{t}(u)-(\nabla\varphi_{t}(u))^{2}}{t\varphi_{t}^{2} (u)}, \tag{2}\] where the integral in the first line is component-wise and \(\mathcal{F}[|x|^{2}\nu]\coloneqq\int\mathrm{e}^{\mathrm{i}\langle\cdot,x \rangle}|x|^{2}\,\nu(\mathrm{d}x)\). To motivate our estimator for \(\nu\), suppose \(\Sigma=0\). In view of (2), we then have \(\nu=-|\cdot|^{-2}\mathcal{F}^{-1}[\Delta\psi]\) and \(\Delta\psi\) can naturally be estimated using the empirical characteristic function \(\widehat{\varphi}_{\delta,n}(u)\coloneqq\frac{1}{n}\sum_{k=1}^{n}\mathrm{e}^{ \mathrm{i}\langle u,Y_{k}\rangle}\) leading to \[\widehat{\Delta\psi_{n}}(u)\coloneqq\frac{\widehat{\varphi}_{\delta,n}(u) \Delta\widehat{\varphi}_{\delta,n}(u)-(\nabla\widehat{\varphi}_{\delta,n}(u)) ^{2}}{\delta\widehat{\varphi}_{\delta,n}(u)}\mathbbm{1}_{\{|\widehat{\varphi} _{\delta,n}(u)|\geqslant T^{-1/2}\}} \tag{3}\] with the indicator ensuring a well-defined expression. Therefore, granted \(\nu\) has a Levy density also denoted by \(\nu\), it is reasonable to propose the estimator \[\widehat{\nu}_{h}(x)\coloneqq-|x|^{-2}\mathcal{F}^{-1}\big{[}\mathcal{F}K_{h }\widehat{\Delta\psi_{n}}\big{]}(x),\qquad x\in\mathbb{R}^{d}\setminus\{0\}, \tag{4}\] where \(K\) is kernel a limited by bandwidth \(h>0\) (\(K_{h}\coloneqq h^{-d}K(\cdot/h)\)) satisfying for some order \(p\in\mathbb{N}\) that for any multi-index \(0\neq\beta\in\mathbb{N}_{0}^{d}\) with \(|\beta|_{1}\leqslant p\) we have \[\int_{\mathbb{R}^{d}}K(x)\,\mathrm{d}x=1,\qquad\int x^{\beta}K(x)\,\mathrm{d} x=0\qquad\text{and}\qquad\operatorname{supp}\mathcal{F}K\subseteq[-1,1]^{d}. \tag{5}\] For \(d=1\), we recover the jump density estimator by Trabs (2015) to estimate quantiles of Levy measures. A suitable kernel can be constructed as \(K\coloneqq(\mathcal{F}^{-1}g)/g(0)\) from an integrable even function \(g\colon C^{\infty}(\mathbb{R}^{d})\to\mathbb{R}\) with support contained in \([-1,1]^{d}\), \(g(0)\neq 0\) and vanishing mixed partial derivatives of order up to \(p\) at \(0\). For the theoretical analysis, it will be useful to consider a kernel with product structure \(K(x)=\prod_{j=1}^{d}K^{j}(x_{j})\) for kernels \(K^{j}\) on \(\mathbb{R}\), each with order \(p\), i.e. for all \(q\in\mathbb{N}\), \(q\leqslant p\) \[\int_{\mathbb{R}}K^{j}(x_{j})\,\mathrm{d}x_{j},\qquad\int x_{j}^{q}K^{j}(x_{j} )\,\mathrm{d}x_{j}=0\qquad\text{and}\qquad\operatorname{supp}\mathcal{F}K^{j} \subseteq[-1,1].\] Obviously, such a product kernel also fulfills (5). ### Convergence rates To control the estimation error, we need to impose smoothness and moment conditions on the Levy density. To this end, we introduce for a number of moments \(m>0\), a regularity index \(s>0\), an open subset \(U\subseteq\mathbb{R}^{d}\) and a universal constant \(R>0\) \[\mathcal{C}^{s}(m,U,R)\coloneqq\Big{\{}(\Sigma,\gamma,\nu)\, \Big{|}\,\Sigma\in\mathbb{R}^{d\times d}\text{ positive-semidefinite},\operatorname{tr}(\Sigma)\leqslant R,\,\gamma\in \mathbb{R}^{d},\,\int|x|^{m}\nu(\mathrm{d}x)\leqslant R\\ \nu\text{ has a Lebesgue density with }\||x|^{2}\nu\|_{C^{s}(U)}\leqslant R\Big{\}},\] where \[\|f\|_{C^{s}(U)}\coloneqq\sum_{|\beta|_{1}\leqslant\lfloor s\rfloor}\sup_{x \in U}|f^{(\beta)}(x)|+\max_{|\beta|_{1}=\lfloor s\rfloor}\sup_{x,y\in U,x \neq y}\frac{|f^{(\beta)}(x)-f^{(\beta)}(y)|}{|x-y|^{s-\lfloor s\rfloor}}\] denotes the Holder norm with regularity index \(s\) for any function \(f\) which has derivatives \(f^{(\beta)}\) of order up to \(\lfloor s\rfloor\coloneqq\max\{k\in\mathbb{N}_{0}:k<s\}\) on \(U\). \(C^{s}(U)\) denotes the set of all Holder-regular functions on \(U\) with regularity index \(s>0\). Since we will require regularity of the Levy density in a small \(\zeta\)-neighborhood beyond \(U\) for a uniform rate, we set \(U_{\zeta}\coloneqq\{x\in\mathbb{R}^{d}\mid\exists u\in U:|x-u|<\zeta\}\) for some radius \(\zeta>0\). In view of (3) it is natural, that the estimation error will also depend on the decay behavior of the characteristic function, which in turn, is affected by the presence of a Gaussian component. Therefore, we distinguish the following two classes of Levy processes. First, is the so-called mildly ill-posed case for a decay exponent \(\alpha>0\) \[\mathcal{D}^{s}(\alpha,m,U,R,\zeta)\coloneqq\big{\{}(0,\gamma,\nu)\in\mathcal{ C}^{s}(m,U_{\zeta},R)\,\big{|}\,\|(1+|\cdot|_{\infty})^{-\alpha}/\varphi_{1}\|_{ \infty}\leqslant R,\,\|x|\nu\|_{\infty}\leqslant R\big{\}}.\] As alluded to in the introduction, a Gaussian component overclouds the jumps in addition to the discrete observations and is therefore treated as the severely ill-posed case for \(\alpha,r,\eta>0\) \[\mathcal{E}^{s}(\alpha,m,U,r,R,\zeta,\eta)\coloneqq\big{\{}(\Sigma,\gamma,\nu )\in\mathcal{C}^{s}(m,U_{\zeta},R)\,\big{|}\,\|\exp(-r|\cdot|_{\infty}^{ \alpha})/\varphi_{1}\|_{\infty}\leqslant R,|x|^{3-\eta}\nu(x)\leqslant R\ \forall|x| \leqslant 1\big{\}}.\] The parameters \(\alpha\) and \(r\) control the exponential decay of the characteristic funtion. Note that \(\Sigma\neq 0\) already implies \(\alpha=2\). In the mildly ill-posed case, the Blumenthal-Getoor index of the Levy process is at most \(1\), whereas in the severely ill-posed case it is at most \(\big{(}(3-\eta)\wedge 2\big{)}\lor 0\), where we set \(a\wedge b\coloneqq\min\{a,b\}\) and \(a\lor b\coloneqq\max\{a,b\}\) for \(a,b\in\mathbb{R}\). For these regularity classes, we are able to quantify the estimation error as follows. **Theorem 1**.: _Let \(\alpha,r,R,\zeta>0,s>1,m>4\) and let the kernel satisfy (5) with order \(p\geqslant s\). Let \(U\subseteq\mathbb{R}^{d}\) be an open set which is bounded away from \(0\). We have for \(0<\delta\leqslant R\), \(n\to\infty\):_ 1. _If_ \(U\) _is bounded and_ \(h=h_{\delta,n}=(\log(T)/T)^{1/(2s+2\delta\alpha+d)}\)_, then uniformly in_ \((\Sigma,\gamma,\nu)\in\mathcal{D}^{s}(\alpha,m,U,R,\zeta)\)__ \[\sup_{x^{*}\in U}|\widehat{\nu}_{h}(x^{*})-\nu(x^{*})|=\mathcal{O}_{\mathbb{P} }\Big{(}\Big{(}\frac{\log T}{T}\Big{)}^{s/(2s+2\delta\alpha+d)}\Big{)}.\] _If_ \(\delta=n^{-\varepsilon}\) _with_ \(\varepsilon\in(0,1)\)_, the choice_ \(h=(\log(T)/T)^{1/(2s+d)}\) _yields the rate_ \((\log(T)/T)^{s/(2s+d)}\)_._ 2. _If_ \(|\cdot|^{p+d}K\in L^{1}(\mathbb{R}^{d})\)_,_ \(\eta>0\) _and_ \(h=h_{\delta,n}=(\log(T)/(4r\delta))^{-1/\alpha}\)_, then uniformly in_ \((\Sigma,\gamma,\nu)\in\mathcal{E}^{s}(\alpha,m,U,r,R,\zeta,\eta)\)__ \[\sup_{x^{*}\in U}|\widehat{\nu}_{h}(x^{*})-\nu(x^{*})|=\mathcal{O}_{\mathbb{P} }\Big{(}\Big{(}\frac{\log T}{4r\delta}\Big{)}^{-s/\alpha}\Big{)}.\] _If_ \(\delta=n^{-\varepsilon}\) _with_ \(\frac{3}{2(s+d)+1}\sqrt{\frac{\alpha}{2(s+d)+\alpha}}<\varepsilon<1\)_, the choice_ \(h=T^{-1/(2(s+d))}\) _yields the rate_ \(T^{-s/(2(s+d))}\)_._ This theorem generalizes Trabs (2015, Proposition 2) to the multivariate case and additionally allows for high-frequency observations. Figs. 1 and 2 illustrate a simulation example of the estimation method. In the mildly ill-posed case, one can easily attain the same rates without the logarithm when considering the pointwise loss. We first discuss the low-frequency regime: For \(d=1\), our rates coincide with the proven minimax-optimal rates in the corresponding nonparametric deconvolution problems, see Fan (1991). In the mildly ill-posed case with \(d=1\), the pointwise variant of our rate has been shown to be minimax-optimal under the assumption that \(x\nu\) is \(s\)-Sobolev-regular, see Kappus (2012). In the severely ill-posed case with \(d=1\) and \(\alpha=\{1,2\}\), our rates coincide with the minimax-optimal rates of Neumann & Reiss (2009), who consider the integrated risk in the estimation of \(\Sigma\delta_{0}(\mathrm{d}x)+|x|^{2}(1+|x|^{2})^{-1}\nu(\mathrm{d}x)\) against test functions with Sobolev regularity \(s\). This measure has an atom in \(0\) and is therefore not smooth. Hence, the regularity in the rate comes purely from the test function. By considering \(U\) bounded away from \(0\), we can profit from the regularity of the Levy density outside the origin. We do not even suffer an additional loss for the dimension in the rate, only in the constant. Therefore, the above suggests its optimality. One sees that the rates improve as the time grid narrows. If this refinement happens at an appropriate order compared to the growth of the sample, the ill-posedness vanishes completely in the mildly ill-posed case and the rate becomes polynomial in the severely ill-posed case. In the mildly ill-posed case with high-frequency observations, the rate corresponds to the minimax-optimal rate in a nonparametric regression. It is straightforward to see from our proof, that when estimating \(|\cdot|^{2}\nu\), we can forgo the exclusion of the origin from \(U\) while achieving the same rates in the mildly-ill-posed case. In the severely ill-posed case, the unknown volatility of the Brownian component of the Levy process obstructs the observation of the small jumps. Hence, we can benefit from a pilot estimator for \(\Sigma\). As discussed earlier, even a with minimax-optimal estimator for \(\Sigma\), we would suffer a loss in the overall rate. However, in view of (2), it suffices to estimate the onedimensional parameter \(\operatorname{tr}(\Sigma)\) which is easier compared to the \(d\times d\)-matrix \(\Sigma\). Following the spectral approach again, we propose the estimator \[\widehat{\operatorname{tr}(\Sigma)}\coloneqq\widehat{\operatorname{tr}( \Sigma)}_{h}\coloneqq-\int W_{h}(u)\widehat{\Delta\psi_{n}}(u)\,\mathrm{d}u.\] where \(W_{h}=h^{d}W(h\cdot)\) for a bandwidth \(h>0\) (correspoding to the threshold \(h^{-1}\)) and a weight function \(W\colon\mathbb{R}^{d}\to\mathbb{R}\) with \[\int W(u)\,\mathrm{d}u=1\qquad\text{and}\qquad\operatorname{supp}W\subseteq[-1,1]^{d}.\] This estimator achieves a rate of \((\log T)^{-(s+d)/\alpha}\) and is incorporated into the estimator for \(|\cdot|^{2}\nu\) via \[\widehat{|\cdot|^{2}\nu}_{h}\coloneqq-\mathcal{F}^{-1}\big{[}\mathcal{F}K_{h }\big{(}\widehat{\Delta\psi_{n}}+\widehat{\operatorname{tr}(\Sigma)}_{h}\big{)} \big{]}\] leading to the following extension of Theorem 1. **Proposition 2**.: _Let \(\alpha,r,R,\zeta,\eta>0,1<s\in\mathbb{N},m>4\) and let the kernel satisfy (5) with order \(p\geqslant s\) as well as \(|\cdot|^{p+d}K\in L^{1}(\mathbb{R}^{d})\). Assume \(\|\mathcal{F}^{-1}[W(x)/x_{k}^{s}]\|_{L^{1}}<\infty\) for some \(k\). Choosing \(h=(\log(T)/(4r\delta))^{-1/\alpha}\) we have uniformly in \((\Sigma,\gamma,\nu)\in\mathcal{E}^{s}(\alpha,m,\mathbb{R}^{d},r,R,\zeta,\eta)\)_ \[\sup_{x^{*}\in\mathbb{R}^{d}}\big{|}\big{(}\widehat{|\cdot|^{2}\nu}_{h}\big{)} (x^{*})-|x^{*}|^{2}\nu(x^{*})\big{|}=\mathcal{O}_{\mathbb{P}}\Big{(}\Big{(} \frac{\log T}{4r\delta}\Big{)}^{-s/\alpha}\Big{)}.\] Figure 1: 3D plot of \(|\cdot|^{2}\nu\) (left) and its estimate (right) for a twodimensional compound Poisson process with Gaussian jumps. ### Independent components Compared to the onedimensional case, we need to take the dependence structure of the components of the process into account. In particular, our previous assumption about \(\nu\) having a Lebesgue-density on \(\mathbb{R}^{d}\), rules out Levy processes where all components are independent, since the corresponding Levy measure would only have mass on the coordinate cross. Similarly, Levy processes consisting of multiple mutually independent blocks of components, where the components within the same block depend on each other, are not covered.For the sake of notational simplicity, we focus on the case of two equisized independent blocks: Let \(d\) be even and \(L=(L^{(1)},L^{(2)})\), where \(L^{(1)}\) and \(L^{(2)}\) are two independent Levy processes on \(\mathbb{R}^{d/2}\) with characteristic triplets \((\Sigma_{1},\gamma_{1},\nu_{1})\) and \((\Sigma_{2},\gamma_{2},\nu_{2})\), respectively. Denoting by \(\delta_{0}\) the Dirac-measure in \(0\in\mathbb{R}^{d/2}\), it holds that \[\nu(\mathrm{d}x)=\nu_{1}(\mathrm{d}x^{(1)})\otimes\delta_{0}(\mathrm{d}x^{(2) })+\delta_{0}(\mathrm{d}x^{(1)})\otimes\nu_{2}(\mathrm{d}x^{(2)})\qquad x=(x^ {(1)},x^{(2)}),\,x^{(1)},x^{(2)}\in\mathbb{R}^{d/2}. \tag{6}\] We summarize the class of such Levy processes as \[\widetilde{\mathcal{C}}(m,R)\coloneqq\Big{\{}(\Sigma,\gamma, \nu)\,\Big{|}\,\Sigma=\begin{pmatrix}\Sigma_{1}&0\\ 0&\Sigma_{2}\end{pmatrix},\mathrm{tr}(\Sigma)\leqslant R,\Sigma_{1},\Sigma_{2} \in\mathbb{R}^{d/2\times d/2}\text{ positive-semidefinite},\gamma\in\mathbb{R}^{d},\\ \int|x|^{m}\,\nu(\mathrm{d}x)\leqslant R,\nu\text{ has the form (\ref{eq:Levy}) and }v_{1},\nu_{2}\text{ have Lebesgue- densities}\Big{\}}\] for \(m,R>0\). A simple example of such a Levy measure and its estimate are illustrated in Fig. 3. As before, we distinguish between the mildly ill-posed case with \(\alpha>0\) \[\widetilde{\mathcal{D}}(\alpha,m,R)\coloneqq\big{\{}(0,\gamma,\nu)\in \widetilde{\mathcal{C}}(m,R)\,\big{|}\,\|(1+|\cdot|_{\infty})^{-\alpha}/\varphi _{1}\|_{\infty}\leqslant R,\,\|x_{k}|\nu_{k}\|_{\infty}\leqslant R,k=1,2\big{\}}\] and the severely ill-posed case with \(\alpha,r,\eta>0\) \[\widetilde{\mathcal{E}}(\alpha,m,r,R,\eta)\coloneqq\big{\{}(\Sigma,\gamma, \nu)\in\widetilde{\mathcal{C}}(m,R)\,\big{|}\,\|\exp(-r|\cdot|_{\infty}^{ \alpha})/\varphi_{1}\|_{\infty}\leqslant R,|x_{k}|^{3-\eta}\nu_{k}(x_{k}) \leqslant R\;\forall|x_{k}|\leqslant 1,k=1,2\big{\}}\] based on the decay behavior of the characteristic function and the presence of a Gaussian component. If this dependence structure were known, we could seperate the blocks in the observations, apply our method to each block and obtain an estimator for the overall Levy measure. Since this is not the case, Figure 2: Heatmap of \(|\cdot|^{2}\nu\) (left) and its estimate (right) for a twodimensional compound Poisson process with Gaussian jumps. we are left with applying our initial method. In spite of the unknown dependence structure, we will be able to quantify the estimation error. Due to the structure of the Levy measure, we cannot hope for a pointwise quantitative bound. Instead, we consider the error in a functional sense. To this end, we introduce the following class of test functions for \(\varrho>0\) and \(U\subseteq\mathbb{R}^{d}\) \[F_{\varrho}(U,R)\coloneqq\{f\colon\mathbb{R}^{d}\to\mathbb{R}\mid f\in C^{ \varrho}(\mathbb{R}^{d}),\|f\|_{C^{\varrho}(\mathbb{R}^{d})},\|f\|_{L^{1}( \mathbb{R}^{d})}\leqslant R,\,\mathrm{supp}\,f\subseteq U\}.\] **Theorem 3**.: _Let \(\alpha,r,R>0,\varrho>1,m>4\), let the kernel have product structure and satisfy (5) with order \(p\geqslant\varrho\). Then, we have for \(0<\delta\leqslant R,n\to\infty\):_ 1. _If_ \(U\subseteq\mathbb{R}^{d}\) _is bounded and_ \(h=(\log(T)/T)^{1/(2\varrho+2\delta\alpha+3d/2)}\)_, then uniformly in_ \((\Sigma,\gamma,\nu)\in\widetilde{\mathcal{D}}(\alpha,m,R)\)__ \[\sup_{f\in F_{\varrho}(U,R)}\Big{|}\int_{U}f(x)|x|^{2}\big{(}\nu(\mathrm{d}x) -\widehat{\nu}_{h}(\mathrm{d}x)\big{)}\Big{|}=\mathcal{O}_{\mathbb{P}}\Big{(} \Big{(}\frac{\log T}{T}\Big{)}^{\varrho/(2\varrho+2\delta\alpha+3d/2)}\Big{)}.\] 2. _If_ \(U\subseteq\mathbb{R}^{d}\) _is bounded away from_ \(0\)_,_ \(|\cdot|^{p+d}K\in L^{1}(\mathbb{R}^{d})\)_,_ \(\eta>0\) _and_ \(h=(\log(T)/(4r\delta))^{-1/\alpha}\)_, then uniformly in_ \((\Sigma,\gamma,\nu)\in\widetilde{\mathcal{E}}(\alpha,m,r,R,\eta)\)__ \[\sup_{f\in F_{\varrho}(U,R)}\Big{|}\int_{U}f(x)|x|^{2}\big{(}\nu(\mathrm{d}x) -\widehat{\nu}_{h}(\mathrm{d}x)\big{)}\Big{|}=\mathcal{O}_{\mathbb{P}}\Big{(} \Big{(}\frac{\log T}{4r\delta}\Big{)}^{-\varrho/\alpha}\Big{)}.\] Note that the regularity parameter \(\varrho\) in the rates comes from the smoothness of the test functions as compared to the smoothness \(s\) of the Levy measure in Theorem 1. In the severely ill-posed case, the result is analogous to the well-specified. In the mildly ill-posed case, we pay for the dependence structure with an \(d/2\) in the rate. Morally, one can interpret this as the model dimension being \(3d/2\) instead of \(d\). _Remark 4_.: The product kernel is compatible with any dependence structure of blocks, regardless of their size. For instance, if all components of the process are independent, one gets still gets the analogous result in severely ill-posed case. In the mildly ill-posed case, the dimension appearing in the rate is \(2d-1\) instead of \(3d/2\). Comparing the dependence structures, one finds that the two independent blocks are an in-between case of no independent blocks and fully independent components. Figure 3: 3D plot of \(|\cdot|^{2}\nu\) (left) and its estimate (right) for a Lévy-process where both components are independent compound Poisson processes with Gaussian jumps. ### A uniform risk-bound for the characteristic function and linearization A key ingredient in the proofs of our preceeding results is the following moment-uniform risk bound for the multivariate characteristic function and its partial derivatives. It generalizes the existing results in the univariate case (see Kappus & Reiss 2010, Theorem 1) and the multivariate non-uniform case (see Belomestny & Trabs 2018, Proposition A.1). **Proposition 5**.: _Let \(X_{1},X_{2},\dots\) be \(\mathbb{R}^{d}\)-valued i.i.d. random variables with characteristic function \(\varphi\) and empirical characteristic function \(\widehat{\varphi}_{n}\) such that \(\mathbb{E}[|X_{1}|^{2\beta}|X_{1}|^{\tau}]\lesssim\rho^{|\beta|_{1}\wedge 1}\) and \(\mathbb{E}[|X_{1}|^{2\beta}]\lesssim\rho^{|\beta|_{1}\wedge 1}\) for some multi-index \(\beta\in\mathbb{N}_{0}^{d}\) and \(\tau,\rho>0\). For the inverse softplus-type weight function \(w(u)=\log(\mathrm{e}+|u|)^{-(1+\chi)/2}\) with \(\chi>0\), we have_ \[\mathbb{E}\big{[}\big{\|}w(u)(\widehat{\varphi}_{n}-\varphi)^{(\beta)}(u) \big{\|}_{\infty}\big{]}\lesssim\rho^{(|\beta|_{1}\wedge 1)/2}n^{-1/2}.\] As a direct consequence of Proposition 5, the indicator in the definition (3) equals one on the support of \(\mathcal{F}K_{h}\), with probability converging to one for the bandwidths we consider. To prove our rates for \(\widehat{\nu}_{h}\), (2) lets us decompose the error \[\widehat{\nu}_{h}(x^{*})-\nu(x^{*})\] \[=|x^{*}|^{-2}\big{(}\underbrace{\big{(}K_{h}*(|\cdot|^{2}\nu)-| \cdot|^{2}\nu\big{)}(x^{*})}_{=:B^{\nu}(x^{*})}-\underbrace{\frac{1}{\delta} \mathcal{F}^{-1}\big{[}\mathcal{F}K_{h}\Delta\big{(}(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_{\delta}\big{)}\big{]}(x^{*})}_{=L^{\varphi}_{ \delta,n}(x^{*})}+R_{\delta,n}+\mathrm{tr}(\Sigma)K_{h}(x^{*})\big{)} \tag{7}\] into a bias term \(B^{\nu}\), the linearized stochastic error \(L^{\nu}_{\delta,n}\), the error \(\mathrm{tr}(\Sigma)K_{h}\) due to the volatility and a remainder term \(R_{\delta,n}\). Proposition 5 applied to the increments of the Levy process leads to the following linearization. **Lemma 6**.: _Let \(\int|x|^{4+\tau}\,\nu(\mathrm{d}x)\leqslant R\) for some \(\tau>0\). If \(n^{-1/2}(\log h^{-1})^{(1+\chi)/2}\|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\to 0\) as \(n\to\infty\) for \(h\in(0,1),\chi>0\), it holds_ \[\sup_{|u|_{\infty}\leqslant h^{-1}}\big{|}\widehat{\Delta\psi_{n} }(u)-\Delta\psi(u)-\delta^{-1}\Delta\big{(}(\widehat{\varphi}_{\delta,n}- \varphi_{\delta})/\varphi_{\delta}\big{)}(u)\big{|}=\mathcal{O}_{\mathbb{P}}(a _{n}),\qquad\text{where}\] \[a_{n}\coloneqq n^{-1}(\log h^{-1})^{1+\chi}\|\varphi_{\delta}^{- 1}\|_{L^{\infty}(I_{h})}^{2}\delta^{-1/2}\big{(}\delta\big{\|}|\nabla\psi| \big{\|}_{L^{\infty}(I_{h})}+\delta^{3/2}\big{\|}|\nabla\psi|\big{\|}_{L^{ \infty}(I_{h})}^{2}+1\big{)}.\] As a direct consequence, the remainder term is of the order \[|R_{\delta,n}|=\mathcal{O}_{\mathbb{P}}\big{(}h^{-d}a_{n}\big{)}. \tag{8}\] After treating the four terms in (7), the asserted rates follow from our bandwidth choices. ## 3 Simulation examples We demonstrate the estimation of the Levy density for \(d=2\) with three examples: a compound Poisson process, a variance gamma process and two independent compound Poisson processes. A challenge is to find examples of multivariate Levy processes for which paths can be simulated and the true Levy measure is accessible (at least numerically). To allow for pltable results, we consider the case \(d=2\) and compensate for the possible singularity of the Levy density at the origin, i.e. we plot \(|\cdot|^{2}\nu\) and its estimate. Throughout, we use the flat-top-kernel \(K\), see McMurry & Politis (2004), as defined by its Fourier transform \[\mathcal{F}K(u)\coloneqq\begin{cases}1,&|u|\leqslant c,\\ \exp\Big{(}-\frac{b\exp(-b/(|u|-c)^{2}}{(|u|-1)^{2}}\Big{)}&c<|u|<1,\\ 0,&|u|\geqslant 1,\end{cases}\] whose decay behaviour is controlled by \(b>0\) and \(0<c<1\). In our simulations, \(b=1,c=1/50\) deliver stable results. While a product kernel is convenient for theoretical reasons in Section 2.2, it did not seem necessary in practice. Throughout, we simulate increments of the processes with a time difference of \(\delta=0.001\) and fix the bandwidth at \(h=4T^{-1/2}\). To conquer this ill-posed problem, we use large samples of \(n=500,000\) increments. From the definition (4) of the estimator, it is not guaranteed that \(\widehat{\nu}\geqslant 0\) and for numerical reasons even \(\widehat{\nu}\in\mathbb{C}\setminus\mathbb{R}\) is possible in practice. Therefore, we consider the estimator \(\mathrm{Re}(\widehat{\nu})\lor 0\) in our simulations. The most straightforward example under consideration is the compound Poisson process with intensity \(\lambda=100\) and twodimensional standard-Gaussian jumps. In this case, the Levy density is just the standard normal density, rescaled with the intensity \(\lambda\). Fig. 1 illustrates that the method captures the overall shape of the density. The heatmap in Fig. 2 provides a more detailed view especially around the origin. We observe that the decay for \(|x|\to\infty\) and \(|x|\searrow 0\) is well-estimated, with slight problems only arising on an annulus around the origin. A practical way to construct easy-to-simulate multivariate Levy processes is to subordinate multivariate Brownian motion. In particular, we use a gamma process with variance \(\kappa=1\) to subordinate a twodimensional standard Brownian motion. To access the Levy measure of the resulting variance gamma process, we approximate the theoretical expression from Cont & Tankov (2004, Theorem 4.2) numerically. The results are again illustrated in a 3D plot (Fig. 4) and as a heatmap ((Fig. 5)). In this example, the estimator suffers from oscillations around the true density which are to be expected from spectral-based methods. To demonstrate the method under the depencence structure discussed in (Section 2.2), we consider a Levy process comprised of two independent compound Poisson processes, each with intensity \(\lambda=100\) and onedimensional standard-Gaussian jumps. In contrast to the twodimensional compound Poisson process at the considered at the beginnung of this section, the jumps in both components are driven by independent Poisson processes. The corresponding Levy measure takes the form (6), where \(\nu_{1}\) and \(\nu_{2}\) are onedimensional standard-Gaussian densities, rescaled with \(\lambda\), as illustrated on the left hand side of Fig. 3. It is important to emphasize, that the blue and the orange line represent the Lebesgue-densities of both components on \(\mathbb{R}\), not \(\mathbb{R}^{2}\). The right hand side of the aforementioned figure reveals a strong performance of the estimator on the coordinate cross. Around the axes, we observe a smearing effect due to the singularity of the true Levy measure on the coordinate cross before the estimate drops off as we move away. Figure 4: 3D plot of \(|\cdot|^{2}\nu\) (left) and its estimate (right) for a twodimensional variance gamma process. ## 4 Proofs Throughout, set \(I_{h}\coloneqq[-h^{-1},h^{-1}]^{d}\) for \(h>0\). Note that \(\Sigma,\nu,\Delta\psi\) and \(\widehat{\Delta\psi_{n}}\) do not change if we consider increments based on the Levy process \((L_{t}-t\gamma_{0})_{t\geqslant 0}\) for some \(\gamma_{0}\in\mathbb{R}^{d}\). Hence, no generality is lost if we choose \(\gamma_{0}\) such that in the mildly ill-posed case \[\nabla\psi=\mathrm{i}\mathcal{F}[x\nu]\qquad\qquad\text{and}\qquad\qquad \Delta\psi=-\mathcal{F}[|x|^{2}\nu] \tag{9}\] and in the severely ill-posed case \(\gamma=0\), see Nickl et al. (2016, Lemma 12) for a similar argument in the onedimensional case. Further, due to the infinite divisibility of \(\nu\), the decay behavior of \(\varphi_{1}\) governs that of \(\varphi_{\delta}\). In particular, we have for \(0<\delta\leqslant R\) \[\|(1+|\cdot|_{\infty})^{-\delta\alpha}/\varphi_{\delta}\|_{\infty}\leqslant(1 \lor R)^{R}\qquad\text{and}\qquad\|\exp(-r\delta|\cdot|_{\infty}^{\alpha})/ \varphi_{\delta}\|_{\infty}\leqslant(1\lor R)^{R}\] in the mildly and the severely ill-posed case, respectively. ### Proof of Theorem 1 We extend the proof strategy by Trabs (2015) to accomodate for the multivariate setting. To allow for the application to high-frequency observations, we carefully keep track of \(\delta\) throughout. Subsequently, we will analyze the four terms in (7). #### 4.1.1 Controlling the linearized stochastic error To control \[L^{\nu}_{\delta,n}\coloneqq\mathcal{F}^{-1}\big{[}\mathcal{F}K_{h}\delta^{-1} \Delta\big{(}(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_{\delta} \big{)}\big{]}\] we need to get a grip on the partial derivatives of \(\widehat{\varphi}_{\delta,n}-\varphi_{\delta}\) in the Laplacian of \((\widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_{\delta}\). In particular, we will show that \[\sup_{u\in\mathbb{R}^{d}}\mathbb{E}\Big{[}\Big{|}\frac{\partial^{l}}{ \partial u_{k}^{l}}(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})(u)\Big{|} \Big{]}\leqslant\sup_{u\in\mathbb{R}^{d}}\mathbb{E}\Big{[}\Big{|}\frac{ \partial^{l}}{\partial u_{k}^{l}}(\widehat{\varphi}_{\delta,n}-\varphi_{ \delta})(u)\Big{|}^{2}\Big{]}^{1/2}\stackrel{{!}}{{\lesssim}}n^{- 1/2}\delta^{(l\wedge 1)/2},\qquad l=0,1,2. \tag{10}\] Figure 5: Heatmap of \(|\cdot|^{2}\nu\) (left) and its estimate (right) for a twodimensional variance gamma process. Since \[\mathbb{E}\Big{[}\Big{|}\frac{\partial^{l}}{\partial u_{k}^{l}}(\widehat{\varphi}_{ \delta,n}-\varphi_{\delta})(u)\Big{]}\Big{|}\leqslant n^{-1}\mathbb{E}[Y_{1,m}^ {2l}]=n^{-1}\Big{|}\frac{\partial^{2l}}{\partial u_{k}^{2l}}\varphi_{\delta}(0 )\Big{|}\qquad\forall u\in\mathbb{R}^{d},\] where \(Y_{1,k}\) denotes the \(k\)-th entry of \(Y_{1}\), the case \(l=0\) is obvious and for \(l=1,2\) it remains to show \[\Big{|}\frac{\partial^{2l}}{\partial u_{k}^{2l}}\varphi_{\delta}(0)\Big{|} \lesssim\delta.\] In the mildly ill-posed case \[\Big{|}\frac{\partial}{\partial u_{k}}\psi(0)\Big{|}\lesssim\int|x_{k}|\, \nu(\mathrm{d}x)\leqslant\int|x|\,\nu(\mathrm{d}x)\lesssim 1\] and in the severely ill-posed case \[\frac{\partial}{\partial u_{k}}\psi(0)=0.\] The product rule for higher order derivatives yields \[\Big{|}\frac{\partial^{2}\varphi_{\delta}}{\partial u_{k}^{2}}(0 )\Big{|} =\Big{|}\delta\varphi_{\delta}(u)\Big{(}\delta\Big{(}\frac{ \partial}{\partial u_{k}}\psi(u)\Big{)}^{2}+\frac{\partial^{2}}{\partial u_{k }^{2}}\psi(u)\Big{)}\Big{|}_{u=0}\lesssim\delta\Big{(}1+\int|x|^{2}\,\nu( \mathrm{d}x)\Big{)}\lesssim\delta\qquad\text{and}\] \[\Big{|}\frac{\partial^{4}\varphi_{\delta}}{\partial u_{k}^{4}}(0 )\Big{|} =\Big{|}\frac{\partial^{2}}{\partial u_{k}^{2}}\Big{(}\frac{ \partial^{2}\varphi_{\delta}}{\partial u_{k}^{2}}(u)\Big{)}(0)\Big{|} \lesssim\delta\sum_{j=0}^{2}\binom{2}{j}|\mathbb{E}[Y_{1,k}^{2-j}]|\Big{(}\Big{|} \frac{\partial^{j}}{\partial u_{k}^{j}}\Big{(}\frac{\partial\psi}{\partial u_ {k}}(u)\Big{)}^{2}\Big{|}_{u=0}+\Big{|}\frac{\partial^{j+2}\psi}{\partial u_{ k}^{j+2}}(0)\Big{|}\Big{)}\lesssim\delta,\] where all emerging partial derivatives can again be absolutely and uniformly bounded using our assumptions on \(\nu\). To simplify the notation, set \(m_{\delta,h}\coloneqq\mathcal{F}K_{h}/\varphi_{\delta}\) and recall \(x\cdot y\coloneqq\sum_{k=1}^{d}x_{k}y_{k}\) for \(x,y\in\mathbb{C}^{d}\). For the severely ill-posed case, we have \(|\Delta\psi(u)|\lesssim 1\) and that \(|\mathrm{e}^{\mathrm{i}\langle u,x\rangle}-1|\leqslant|x||u|\) implies \(|\nabla\psi(u)|\lesssim|u|\). Together with (10), we obtain \[\mathbb{E}\Big{[}\sup_{x^{*}\in\mathbb{R}^{d}}|L^{\nu}_{\delta,n} (x^{*})|\Big{]} \leqslant\delta^{-1}\mathbb{E}\big{[}\big{\|}\mathcal{F}^{-1}[m_{ \delta,h}\Delta(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})]\big{\|}_{ \infty}\big{]}+2\mathbb{E}\big{[}\big{\|}\mathcal{F}^{-1}[m_{\delta,h}\nabla( \widehat{\varphi}_{\delta,n}-\varphi_{\delta})\cdot\nabla\psi]\big{\|}_{ \infty}\big{]}\] \[\lesssim(2\pi)^{-d}\int_{I_{h}}\Big{(}\delta^{-1}\mathbb{E}\big{[} \big{|}\Delta(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})(u)\big{|}\big{]} +\mathbb{E}\big{[}\big{|}\nabla(\widehat{\varphi}_{\delta,n}-\varphi_{\delta} )(u)\cdot\nabla\psi(u)\big{|}\big{]}\] \[\lesssim\pi^{-d}T^{-1/2}\int_{I_{h}}\big{(}1+\delta|u|+\delta^{1/ 2}+\delta^{3/2}|u|^{2}\big{)}\exp(r\delta|u|_{\infty}^{\alpha})\,\mathrm{d}u\] \[\lesssim T^{-1/2}\big{(}h^{-d}+\delta h^{-d-1}+\delta^{3/2}h^{-d- 2}\big{)}\exp(r\delta h^{-\alpha}), \tag{11}\] which will be dominated by the bias. In the mildly ill-posed case, the stochastic error needs to be decomposed further into the main stochastic error \[M^{\nu}_{\delta,n} \coloneqq-\frac{1}{T}\sum_{k=1}^{n}\mathcal{F}^{-1}\Big{[}m_{ \delta,h}\big{(}|Y_{k}|^{2}\mathrm{e}^{\mathrm{i}\langle u,Y_{k}\rangle}- \mathbb{E}\big{[}|Y_{k}|^{2}\mathrm{e}^{\mathrm{i}\langle u,Y_{k}\rangle}] \big{)}\Big{]}\qquad\text{and}\] \[M^{\nu}_{\delta,n}-L^{\nu}_{\delta,n} =2\mathcal{F}^{-1}\big{[}m_{\delta,h}\nabla(\widehat{\varphi}_{ \delta,n}-\varphi_{\delta})\cdot\nabla\psi\big{]}+\mathcal{F}^{-1}\big{[}m_{ \delta,h}(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})\big{(}\Delta\psi- \delta(\nabla\psi)^{2}\big{)}\big{]}. \tag{12}\] To control the difference (12), note that \(\||x|\nu\|_{\infty}\leqslant R\) and \(\||x|^{m}\nu\|_{L^{1}}\leqslant R\) imply \(\||x|\nu\|_{L^{1}}\), \(\||x|\nu\|_{L^{2}}\), \(\||x|^{2}\nu\|_{L^{2}}\lesssim 1\). Further, the support of \(\mathcal{F}K\) and the decay behavior of \(\varphi_{\delta}\) ensure \[\|m_{\delta,h}\|_{L^{2}}^{2}\lesssim\int|\mathcal{F}K(hu)|^{2}(1+|u|)^{2\delta \alpha}\,\mathrm{d}u\lesssim(1+h^{-1})^{2\delta\alpha}h^{-d}\lesssim h^{-2 \delta\alpha-d}. \tag{13}\] Hence, (10) and the Cauchy-Schwarz inequality together with (9) and the Plancherel theorem and (9) lead to \[\mathbb{E}\Big{[}\sup_{x^{*}\in U}\big{|}M^{\nu}_{\delta,n}(x^{*})-L ^{\nu}_{\delta,n}(x^{*})\big{|}\Big{]} \leqslant(2\pi)^{-d}\big{(}2\mathbb{E}\big{[}\big{\|}\|m_{\delta,h} \nabla(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})\cdot\nabla\psi\big{\|}_{ L^{1}}\big{]}\] \[\qquad\qquad\qquad+\mathbb{E}\big{[}\big{\|}m_{\delta,h}(\widehat {\varphi}_{\delta,n}-\varphi_{\delta})\big{(}\Delta\psi-\delta(\nabla\psi)^{2} \big{)}\big{\|}_{L^{1}}\big{]}\big{)}\] \[\lesssim n^{-1/2}\|m_{\delta,h}\|_{L^{2}}\big{(}\delta^{1/2}\| |x\|\nu\|_{L^{2}}+\|x|^{2}\nu\|_{L^{2}}+\delta d^{2}\|x\nu\|_{L^{2}}\|x\|\nu\|_ {L^{1}}\big{)}\] \[\lesssim n^{-1/2}h^{-\delta\alpha-d/2}. \tag{14}\] Being the sum of centered i.i.d. random variables, the main stochastic error for fixed \(x\) is controlled by Bernstein's inequality as summarized in the following lemma. **Lemma 7**.: _Let \(\alpha,R,\zeta>0,m>4\) and \(x\in\mathbb{R}^{d}\) and let the kernel satisfy (5) for \(p\geqslant 1\). If \((\Sigma,\gamma,\nu)\in\mathcal{D}^{s}(\alpha,m,U,R,\zeta)\), then there exists some constant \(c>0\) depending only on \(R,\alpha\) and \(d\) such that for any \(\kappa_{0}>0\) and any \(n\in\mathbb{N},0<\delta\leqslant R,h\in(0,1)\)_ \[\mathbb{P}\big{(}|M^{\nu}_{\delta,n}(x)|\geqslant\kappa_{0}T^{-1/2}h^{- \delta\alpha-d/2}\big{)}\leqslant 2\exp\Big{(}-\frac{c\kappa_{0}^{2}}{(1+|x|^{3}) \big{(}1+\kappa_{0}(h^{d}T)^{-1/2}\big{)}}\Big{)}.\] To establish a uniform bound for \(x^{*}\in U\), a union bound extends this lemma to a discretization of the bounded set \(U\) and Lipschitz continuity of \(x\mapsto M^{\nu}_{\delta,n}(x)\) allows us to control the discretization error. In particular, a standard covering argument yields a discretization \(x_{1},\dots,x_{N_{n}}\in\mathbb{R}^{d}\) of \(U\) such that \(\sup_{x^{*}\in U}\min_{l=1,\dots,N_{n}}|x^{*}-x_{l}|\leqslant T^{-2}\), \(N_{n}\lesssim T^{2d}\) and \(\max_{l=1,\dots,N_{n}}|x_{l}|\leqslant C\) with some \(C>0\) indepedent of \(n\). Since \[M^{\nu}_{\delta,n}=K_{h}*g\qquad\text{with}\qquad g\coloneqq\delta^{-1} \mathcal{F}^{-1}\big{[}1_{I_{h}}\varphi_{\delta}^{-1}\Delta(\widehat{\varphi}_ {\delta,n}-\varphi_{\delta})\big{]},\] the fundamental theorem of calculus together with the order of the kernel ensures the Lipschitz continuity of \(M^{\nu}_{\delta,n}\) via \[|M^{\nu}_{\delta,n}(x)-M^{\nu}_{\delta,n}(y)|=\Big{|}\int_{0}^{1 }(x-y)\cdot\nabla(K_{h}*g)(y+\tau(x-y))\,\mathrm{d}\tau\Big{|} \leqslant|x-y|\mathbb{E}[\|g\|_{\infty}]\big{\|}|\nabla K_{h}|_{1} \big{\|}_{L^{1}}\] \[\lesssim|x-y|h^{-1}\mathbb{E}[\|g\|_{\infty}].\] Therefore, the discretization error is upper bounded by \[\mathbb{E}\Big{[}\sup_{x^{*}\in U}\min_{l=1,\dots,N_{n}}|M^{\nu}_{\delta,n}(x^ {*})-M^{\nu}_{\delta,n}(x_{l})|\Big{]}\lesssim T^{-5/2}h^{-1}\int_{I_{h}}| \varphi_{\delta}^{-1}(u)|\,\mathrm{d}u\lesssim T^{-5/2}h^{-\delta\alpha-d-1}.\] Combining the above with Markov's inequality yields for any \(\kappa_{0}\) such that \(2d<\frac{c}{6}\kappa_{0}^{2}/(1+C^{3})\) with \(c\) from Lemma 7 and \(T\) with \(\kappa_{0}^{2}\log(T)/(Th^{d})\leqslant 1\) \[\mathbb{P}\Big{(}\sup_{x^{*}\in U}|M^{\nu}_{\delta,n}(x^{*})|> \kappa_{0}\Big{(}\frac{\log T}{T}\Big{)}^{1/2}h^{-\delta\alpha-d/2}\Big{)}\] \[\leqslant\frac{2}{\kappa_{0}}\Big{(}\frac{\log T}{T}\Big{)}^{-1/2 }h^{\delta\alpha+d/2}\mathbb{E}\Big{[}\sup_{x^{*}\in U}\min_{l=1,\dots,N_{n}}|M ^{\nu}_{\delta,n}(x^{*})-M^{\nu}_{\delta,n}(x_{l})|\Big{]}\] \[\qquad\qquad+2N_{n}\exp\Big{(}-\frac{c\kappa_{0}^{2}\log T}{2(1+ C^{3})\big{(}2+\kappa_{0}(\log(T)/(Th^{d}))^{1/2}\big{)}}\Big{)}\] \[\lesssim\frac{2}{\kappa_{0}}\Big{(}\frac{\log T}{T}\Big{)}^{-1/2 }h^{-d/2-1}T^{-5/2}+2\exp\Big{(}\Big{(}2d-\frac{c\kappa_{0}^{2}}{6(1+C^{3})} \Big{)}\log T\Big{)}. \tag{15}\] The second term obviously converges to \(0\) as \(T\to\infty\). For the first term, \(3d/2\geqslant d/2+1\) implies \[\frac{2}{\kappa_{0}}\Big{(}\frac{\log T}{T}\Big{)}^{-1/2}h^{-d/2-1}T^{-5/2} \leqslant\frac{2}{\kappa_{0}}\Big{(}\frac{\log T}{T}\Big{)}^{-1/2}h^{-3d/2}T^{- 5/2}=\frac{2}{\kappa_{0}}\Big{(}\frac{\log T}{Th^{d}}\Big{)}^{3/2}T^{-1/2}( \log T)^{-2}\] and the right hand side converges to \(0\) by our choice of bandwidth. Overall, (14) and (15) show \[|L^{\nu}_{\delta,n}|=\mathcal{O}_{\mathbb{P}}\Big{(}\Big{(}\frac{\log T}{T} \Big{)}^{1/2}h^{-\delta\alpha-d/2}\Big{)},\] which our bandwidths balance with the bias. #### 4.1.2 Controlling the error owed to the volatility We now consider the last term in (7). The mildly ill-posed case is trivial since \(\Sigma=0\). Turning to the severely ill-posed case, we first aim to bound \(|x|^{p+d}|K(x)|\). To this end, consider \[|x_{k}|^{p+d}|K(x)|\leqslant\frac{1}{(2\pi)^{d}}\Big{\|}\frac{\partial^{p+d} \mathcal{F}K}{\partial u_{k}^{p+d}}\Big{\|}_{L^{1}(I_{1})}=\frac{1}{(2\pi)^{d }}\int_{I_{1}}\Big{|}\int\mathrm{e}^{\mathrm{i}\langle u,z\rangle}K(z)z_{k}^{ p+d}\,\mathrm{d}z\Big{|}\,\mathrm{d}u\lesssim 1,\] It follows from the equivalence of norms \(|x|\lesssim|x|_{p+d}\) that \[|x|^{p+d}|K(x)|\lesssim|x|_{p+d}^{p+d}|K(x)|=\sum_{k=1}^{d}|x_{k}|^{p+d}|K(x)| \lesssim 1. \tag{16}\] Thus, \[|K_{h}(x^{*})|\leqslant h^{-d}\sup_{|x|\geqslant|x^{*}|/h}|K(x)|\leqslant h^{ -d}\sup_{|x|\geqslant|x^{*}|/h}\frac{|x|^{p+d}}{|x^{*}/h|^{p+d}}|K(x)|\lesssim h ^{p}|x^{*}|^{-p-d}\] and since \(U\) is bounded away from \(0\), this gives a uniform bound in \(x^{*}\) of the order \(h^{s}\) as \(p\geqslant s\). #### 4.1.3 Controlling the bias For \(x^{*}\in U\) and \(h|x|<\zeta\), we use a multivariate Taylor expansion of \(g\coloneqq|\cdot|^{2}\nu\in C^{s}(U_{\zeta})\) around \(x^{*}\) to obtain \[g(x^{*}-hx)-g(x^{*})=\sum_{0<|\beta|_{1}<\lfloor s\rfloor}\frac{g^{(\beta)}(x^ {*})}{\beta!}(-hx)^{\beta}+\sum_{|\beta|_{1}=\lfloor s\rfloor}\frac{g^{(\beta) }(x^{*}-\tau_{x^{*}-hx}hx)}{\beta!}(-hx)^{\beta},\] for some \(\tau_{x^{*}-hx}\in[0,1]\). The order of the kernel and the Holder regularity of \(g\) yield \[|B^{\nu}(x^{*})| =\big{|}\big{(}K_{h}*(|\cdot|^{2}\nu)-|\cdot|^{2}\nu\big{)}(x^{*}) \big{|}\] \[=\Big{|}\int\big{(}g(x^{*}-hx)-g(x^{*})\big{)}K(x)\,\mathrm{d}x \Big{|}\] \[\leqslant\Big{|}\int_{|x|\geqslant\zeta/h}\Big{(}g(x^{*}-hx)-\sum_ {|\beta|_{1}\leqslant\lfloor s\rfloor}\frac{g^{(\beta)}(x^{*})}{\beta!}(-hx)^{ \beta}\Big{)}K(x)\,\mathrm{d}x\Big{|}\] \[\qquad\qquad+\Big{|}\int_{|x|<\zeta/h}\sum_{|\beta|_{1}=\lfloor s \rfloor}\frac{(-hx)^{\beta}}{\beta!}\big{(}g^{(\beta)}(x^{*}-\tau_{x^{*}-hx }hx)-g^{(\beta)}(x^{*})\big{)}K(x)\,\mathrm{d}x\Big{|}\] \[\lesssim\int_{|x|\geqslant\zeta/h}|g(x^{*}-hx)K(x)|\,\mathrm{d}x +\sum_{|\beta|_{1}\leqslant\lfloor s\rfloor}\frac{\|g^{(\beta)}\|_{L^{\infty}( U)}}{\beta!}\frac{h^{s}}{\zeta^{s-|\beta|_{1}}}\int_{|x|\geqslant\zeta/h}|x|^{s}|K(x)|\, \mathrm{d}x\] \[\int_{|x|\geqslant\zeta/h}|g(x^{*}-hx)K(x)|\,\mathrm{d}x\lesssim\int_{|x|\geqslant \zeta/h}|x^{*}-hx||K(x)|\,\mathrm{d}x\lesssim h^{s}\frac{1+\zeta}{\zeta^{s}}\int|x |^{s}|K(x)|\,\mathrm{d}x.\] #### 4.1.4 Controlling the remainder term To bound \(R_{\delta,n}\) in (7), we first show that with \(a_{n}\) from Lemma 6 \[g\coloneqq\mathcal{F}^{-1}\big{[}\mathcal{F}K_{h}\big{(}\widehat{\Delta \psi_{n}}-\Delta\psi-\delta^{-1}\Delta((\widehat{\varphi}_{\delta,n}-\varphi_{ \delta})/\varphi_{\delta})\big{)}\big{]}=\mathcal{O}_{\mathbb{P}}\big{(}h^{-d} a_{n}\big{)}.\] Let \(\varepsilon>0\). Owing to Lemma 6 we can choose \(N,M>0\) (wlog. \(M>\|K\|_{L^{1}}\)) such that the probability of the event \(A_{n}\coloneqq\{\sup_{|u|_{\infty}\leqslant h^{-1}}|\widetilde{g}(u)|>a_{n}M\}\) with \(\widetilde{g}\coloneqq\widehat{\Delta\psi_{n}}-\Delta\psi-\delta^{-1}\Delta(( \widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_{\delta})\) is less than \(\varepsilon\) for \(n>N\). Due to the support of \(\mathcal{F}K\), we have on \(A_{n}^{\varepsilon}\) \[|g(x^{*})|=\frac{1}{(2\pi)^{d}}\Big{|}\int_{I_{h}}\mathrm{e}^{-\mathrm{i}(u, x^{*})}\mathcal{F}K(hu)\widetilde{g}(u)\,\mathrm{d}u\Big{|}\leqslant\frac{a_{n}M}{(2 \pi)^{d}}\int_{I_{h}}|\mathcal{F}K(hu)|\,\mathrm{d}u\leqslant h^{-d}a_{n}M\|K \|_{L^{1}}.\] For \(M^{\prime}\coloneqq M\|K\|_{L^{1}}\), we obtain \[\mathbb{P}\Big{(}\sup_{x^{*}\in U}|g(x^{*})|>h^{-d}a_{n}M^{\prime}\Big{)} \leqslant\varepsilon,\] whereby the remainder term has the order proposed in (8). In the mildly ill-posed case, we have \(\|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\lesssim h^{-\delta\alpha}\) and (9) implies \(\big{\|}|\nabla\psi|\big{\|}_{L^{\infty}(I_{h})}\lesssim 1\). Thus, we have \[|R_{\delta,n}|=\mathcal{O}_{\mathbb{P}}\big{(}n^{-1}\delta^{-1/2}(\log h^{-1} )^{1+\chi}h^{-2\delta\alpha-d}\big{)}.\] In the severely ill-posed case, \(\|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\lesssim\exp(r\delta h^{-\alpha})\) holds and (1) implies \(\big{\|}|\nabla\psi|\big{\|}_{L^{\infty}(I_{h})}\lesssim h^{-1}\). Hence, \[|R_{\delta,n}|=\mathcal{O}_{\mathbb{P}}\big{(}n^{-1}\delta^{-1/2}(\log h^{-1} )^{1+\chi}\big{(}h^{-d}+\delta h^{-d-1}+\delta^{3/2}h^{-d-2}\big{)}\exp(2r \delta h^{-\alpha})\big{)}.\] In both cases, the remainder term is dominated by the linearized stochastic error. This completes the proof of Theorem 1. ### Proof of Proposition 2 For the modified estimator, we have to replace \(\operatorname{tr}(\Sigma)K_{h}\) with \(\big{(}\operatorname{tr}(\Sigma)-\widehat{\operatorname{tr}(\Sigma)}\big{)}K_{h}\) in the decomposition (7). All other terms are treated as before. Since we can bound \(|K_{h}(x^{*})|\) by \(h^{-d}\) uniformly in \(x^{*}\) and \(\delta\leqslant R\) is fixed, we only need to prove \[\big{|}\widehat{\operatorname{tr}(\Sigma)}-\operatorname{tr}(\Sigma)\big{|}= \mathcal{O}_{\mathbb{P}}\big{(}(\log n)^{-(s+d)/\alpha}\big{)}.\] Similarly to (7), the error for estimating the trace of \(\Sigma\) can be decomposed into \[\operatorname{tr}(\Sigma)-\widehat{\operatorname{tr}(\Sigma)} =\int W_{h}(u)\big{(}\widehat{\Delta\psi_{n}}-\Delta\psi\big{)}(u )\operatorname{d}u+\int W_{h}(u)\big{(}\Delta\psi(u)+\operatorname{tr}(\Sigma) \big{)}\operatorname{d}u\] \[=\underbrace{\int W_{h}(u)\delta^{-1}\Delta\Big{(}\frac{\widehat {\varphi}_{\delta,n}(u)-\varphi_{\delta}(u)}{\varphi_{\delta}(u)}\Big{)}(u) \operatorname{d}u}_{=\widetilde{L}^{\nu}_{\delta,n}}+\widetilde{R}_{\delta,n} -\underbrace{\int W_{h}(u)\mathcal{F}[|x|^{2}\nu](u)\operatorname{d}u}_{= \widetilde{B}^{\nu}_{h}}\] with the linearized stochastic error \(\widetilde{L}^{\nu}_{\delta,n}\), the bias \(\widetilde{B}^{\nu}_{h}\) and a remainder term \(\widetilde{R}_{\delta,n}\). Using the techniques from Section 4.1.1, it is straightforward to see \[\mathbb{E}[|\widetilde{L}^{\nu}_{\delta,n}|] \lesssim\delta^{-1}n^{-1/2}\int|W_{h}(u)||\varphi_{\delta}^{-1}(u )|\big{(}\delta^{1/2}+\delta^{3/2}|\nabla\psi(u)|+\delta^{2}|\nabla\psi(u)|^{2 }+\delta\big{)}\operatorname{d}u.\] \[\lesssim\delta^{-1}n^{-1/2}\|\varphi_{\delta}^{-1}\|_{L^{\infty}( I_{h})}\big{(}\delta^{1/2}+\delta^{3/2}h^{-1}+\delta^{2}h^{-2}\big{)}\int|W_{h}(u)| \operatorname{d}u\] \[\lesssim n^{-1/2}\exp(r\delta h^{-\alpha})h^{-2}\int|W(u)| \operatorname{d}u,\] which is of the order \(n^{-1/4}(\log n)^{2/\alpha}\) by our choice of \(h\) and will be dominated by the bias. Using the Plancherel theorem in a similar fashion to Belomestny & Reiss (2015, Section 4.2.1), we have for \(g\coloneqq|\cdot|^{2}\nu\) and \(\beta\in\mathbb{N}_{0}^{d}\) with \(\beta_{l}=s^{\mathbbm{1}}_{\{l=k\}}\) \[|\widetilde{B}^{\nu}_{h}|\lesssim\Big{|}\int\mathcal{F}^{-1}[W_{h}(x)/x_{k}^{s }]\mathcal{F}^{-1}\big{[}x_{k}^{s}\mathcal{F}g(x)\big{]}(u)\operatorname{d}u \Big{|}\lesssim\|g^{(\beta)}\|_{\infty}\big{\|}\mathcal{F}^{-1}[W_{h}(x)/x_{k}^ {s}]\|_{L^{1}}.\] By substitution \[\mathcal{F}^{-1}[W_{h}(x)/x_{k}^{s}](u)=\frac{h^{s}}{(2\pi)^{d}}\mathcal{F}^{-1 }\big{[}W(x)/x_{k}^{s}\big{]}(u/h)\] and therefore \[|\widetilde{B}^{\nu}_{h}|\lesssim h^{*}\|g^{(\beta)}\|_{\infty}\big{\|} \mathcal{F}^{-1}[W(x)/x_{k}^{s}](\cdot/h)\|_{L^{1}}\lesssim h^{(s+d)}\|g^{( \beta)}\|_{\infty}\big{\|}\mathcal{F}^{-1}[W(x)/x_{k}^{*}]\|_{L^{1}}\lesssim h^ {(s+d)}.\] Together with Lemma 6, we have \[|\widetilde{R}_{\delta,n}| \lesssim\sup_{|u|_{\infty}\leqslant h^{-1}}\big{|}\widehat{ \Delta\psi_{n}}(u)-\Delta\psi(u)-\delta^{-1}\Delta\big{(}(\widehat{\varphi}_{ \delta,n}-\varphi_{\delta})/\varphi_{\delta}\big{)}(u)\big{|}\int|W_{h}(u)| \operatorname{d}u\] \[=\mathcal{O}_{\mathbb{P}}\big{(}n^{-1}(\log h^{-1})^{1+\chi}\| \varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}^{2}h^{-2}\big{)}.\] which is dominated by the linearized stochastic error. ### Proof of Theorem 3 The distributional analogon to (7) is \[\int f(x)|x|^{2}\widehat{\nu}_{h}(x)(\operatorname{d}x)-\int f(x)|x|^{2}\,\nu( \operatorname{d}x)\] \[=\int f(x)\big{(}K_{h}*(|\cdot|^{2}\nu)\big{)}(x)\,{\rm d}x-\int f(y) |y|^{2}\nu({\rm d}y)\] \[\qquad+\int f(x){\cal F}^{-1}\big{[}{\cal F}K_{h}(\Delta\psi- \widehat{\Delta\psi}_{n})\big{]}(x)\,{\rm d}x+\int f(x)\,{\rm tr}(\Sigma)K_{h}(x )\,{\rm d}x\] \[=\int f(x)\big{(}K_{h}*(|\cdot|^{2}\nu)\big{)}(x)\,{\rm d}x-\int f( y)|y|^{2}\nu({\rm d}y)\] \[\qquad+\int f(x)L^{\nu}_{\delta,n}(x)\,{\rm d}x+\int f(x)R_{\delta,n}\,{\rm d}x+\int f(x)\,{\rm tr}(\Sigma)K_{h}(x)\,{\rm d}x\] with the same \(L^{\nu}_{\delta,n}\) and \(R_{\delta,n}\), for which we will derive uniform upper bounds on \(U\) which directly translate into bounds when integrating against test functions due to their regularity. For the integrated bias, we use Fubini's theorem to obtain \[B^{\nu}_{I} =\int f(x)\big{(}K_{h}*(|\cdot|^{2}\nu)\big{)}(x)\,{\rm d}x-\int f (y)|y|^{2}\nu({\rm d}y)\] \[=\int f(x)\Big{(}\int K_{h}(x-y)|y|^{2}\,\nu({\rm d}y)\Big{)}\,{ \rm d}x-\int f(y)|y|^{2}\,\nu({\rm d}y)\] \[=\int\big{(}\big{(}f*K_{h}(-\cdot)\big{)}(y)-f(y)\big{)}|y|^{2}\, \nu({\rm d}y)\] \[=\int\big{(}\big{(}K_{h}(-\cdot)*f\big{)}(y)-f(y)\big{)}|y|^{2}\, \nu({\rm d}y).\] \(\big{(}\big{(}K_{h}(-\cdot)*f\big{)}(y)-f(y)\big{)}\) is of the order \((|y|\lor 1)h^{\varrho}\) which follows from the arguments in (7) with \(g=f\), \(\varrho\) and \(K_{h}(-\cdot)\) instead of \(|\cdot|^{2}\nu\), \(s\) and \(K_{h}\), respectively. Therefore, \[|B^{\nu}_{I}|\lesssim h^{\varrho}\int(|y|\lor 1)|y|^{2}\,\nu({\rm d}y)\lesssim h ^{\varrho}\int|y|^{3}\,\nu({\rm d}y)\lesssim h^{\varrho}.\] A key tool to control linearized stochastic error \(L^{\nu}_{\delta,n}\) in Section 4.1.1 was (10), which we can still establish here by bounding the first four partial derivatives of \(\psi\) at the origin. Indeed, by (2) we have \(\frac{\partial\psi}{\partial u_{k}}(0)=0\) and similarly \[\Big{|}\frac{\partial\psi}{\partial u_{k}^{j}}(u)\Big{|}\leqslant{\rm tr}( \Sigma)+\int|x|^{j}\,\nu({\rm d}x)={\rm tr}(\Sigma)+\sum_{l=1}^{2}\int|x_{l}|^{ j}\,\nu_{l}({\rm d}x_{l})\lesssim 1,\quad j=2,3,4,\,k=1,\ldots,d. \tag{17}\] Hence, (10) holds. Additionally, (17) implies \(|\Delta\psi(u)|\lesssim 1\). In the severely ill-posed case, we can still bound the gradient of \(\psi\) by \[|\nabla\psi(u)|\lesssim|\Sigma u|+\int|x||{\rm e}^{{\rm i}(u,x)}-1|\,\nu({\rm d }x)\leqslant|u|+\int|\langle u,x\rangle||x|\,\nu({\rm d}x)\lesssim|u|,\] and then apply the arguments from (11). Hence, the linearized stochastic error is of the same order as before. In the severely ill-posed case, (17) holds even for \(j=1\) and therefore \(|\nabla\psi(u)|\lesssim 1\). Continuing from (10) in the mildly ill-posed case requires the most significant changes. (9) now reads as \[\nabla\psi(u)=\big{(}{\rm i}{\cal F}[x^{(1)}\nu_{1}](u^{(1)}),{\rm i}{\cal F} [x^{(2)}\nu_{2}](u^{(2)})\big{)}^{\top}.\] and the main crux is that \[\int{\rm e}^{{\rm i}(u^{(k)},x^{(k)})}|x^{(k)}|^{j}\,\nu_{k}({\rm d}x^{(k)}), \qquad j,k=1,2\] are constant in half of their arguments. Therefore, they cannot be finitely integrable as functions on \(\mathbb{R}^{d}\). In (12), a way out is to consider \[\big{\|}m_{\delta,h}|\nabla\psi|\big{\|}_{L^{1}}\leqslant\sum_{k=1}^{2}\big{\|}m_ {\delta,h}(u)|\mathcal{F}[x^{(k)}\nu_{k}](u^{(k)})|\big{\|}_{L^{1}}. \tag{18}\] Then, we apply the Cauchy-Schwarz inequality and Plancherel's theorem only on \(L^{2}(\mathbb{R}^{d/2})\) to obtain \[\big{\|}m_{\delta,h}(u)|\mathcal{F}[x^{(1)}\nu_{1}](u^{(1)})| \big{\|}_{L^{1}}\] \[=\int\int\Big{|}\frac{\mathcal{F}K(hu)}{\varphi_{\delta}(u)}| \mathcal{F}[x^{(1)}\nu_{1}](u^{(1)})|\Big{|}\,\mathrm{d}u^{(1)}\,\mathrm{d}u^{ (2)}\] \[\leqslant\||\mathcal{F}[x^{(1)}\nu_{1}]|\|_{L^{2}(\mathbb{R}^{d/ 2})}\int_{[-h^{-1},h^{-1}]^{d/2}}\Big{(}\int_{[-h^{-1},h^{-1}]^{d/2}}\big{|} \varphi_{\delta}^{-1}(u)\big{|}^{2}\,\mathrm{d}u^{(1)}\Big{)}^{1/2}\,\mathrm{ d}u^{(2)}\] \[\lesssim h^{-\delta\alpha-3d/4}. \tag{19}\] Analogously, the second summand in (18) has the same order. As a direct consequence, \[\big{\|}m_{\delta,h}|\nabla\psi|^{2}\big{\|}_{L^{1}}\leqslant\sum_{k=1}^{2} \||x^{(k)}|\nu_{k}\|_{L^{1}}\big{\|}m_{\delta,h}|\nabla\psi|\big{\|}_{L^{1}} \lesssim h^{-\delta\alpha-3d/4}\] and similarly to (18) and (19) the same holds for \(\|m_{\delta,h}\Delta\psi\|_{L^{1}}\). Recalling (12), we have \[\mathbb{E}\Big{[}\sup_{x^{\prime}\in U}\big{|}M_{\delta,n}^{\nu} (x^{*})-L_{\delta,n}^{\nu}(x^{*})\big{|}\Big{]} \leqslant(2\pi)^{-d}\big{(}2\mathbb{E}\big{[}\big{\|}m_{\delta,h} \nabla(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})\cdot\nabla\psi\big{\|}_ {L^{1}}\big{]}\] \[\lesssim n^{-1/2}h^{-\delta\alpha-3d/4}.\] Note that we pay for the dependence struture with an additional \(h^{-d/4}\) compared to (14). The same happens when applying Bernstein's inequality to obtain the following adaptation of Lemma 7. **Lemma 8**.: _Let \(\alpha,R,\zeta>0,m>4,U\subseteq\mathbb{R}^{d}\) and \(x\in\mathbb{R}^{d}\), let the kernel have product structure and satisfy (5) for \(p\geqslant 1\). If \((\Sigma,\gamma,\nu)\in\widetilde{\mathcal{D}}(\alpha,m,R)\), then there exists some constant \(c>0\) depending only on \(R,\alpha\) and \(d\) such that for any \(\kappa_{0}>0\) and any \(n\in\mathbb{N},0<\delta\leqslant R,h\in(0,1)\)_ \[\mathbb{P}\big{(}|M_{\delta,n}^{\nu}(x)|\geqslant\kappa_{0}T^{-1/2}h^{-\delta \alpha-3d/4}\big{)}\leqslant 2\exp\Big{(}-\frac{c\kappa_{0}^{2}}{(1+|x|^{3})(1+ \kappa_{0}(Th^{d/2})^{-1/2})}\Big{)}.\] Carrying out the discretization argument from before, the linearized stochastic error in the mildly ill-posed case is of the order \[|L_{\delta,n}^{\nu}|=\mathcal{O}_{\mathbb{P}}\Big{(}\Big{(}\frac{\log T}{T} \Big{)}^{1/2}h^{-\delta\alpha-3d/4}\Big{)}.\] The term \(\mathrm{tr}(\Sigma)K_{h}\) is treated as in Section 4.1.2 just with \(\varrho\) instead of \(s\). No changes are necessary to treat the remainder term compared to Section 4.1.4. This is because when treating the linearized stochstic error, we already showed that still \(|\nabla\psi(u)|,|\Delta\psi(u)|\lesssim 1\) in the mildly ill-posed case and \(|\nabla\psi(u)|\lesssim|u|,|\Delta\psi(u)|\lesssim 1\) in the severely ill-posed case. This concludes the proof of Theorem 3. ### Remaining proofs #### 4.4.1 Proof of Proposition 5 The proof uses empirical process theory and is a combination of Kappus & Reiss (2010) and Belomestny & Trabs (2018). To simplify the notation, write \[C^{\beta}_{\rho,n}(u)\coloneqq n^{-1/2}\rho^{-(|\beta|_{1}\wedge 1)/2}\sum_{k=1}^{n }\frac{\partial^{\beta}}{\partial u^{\beta}}\big{(}\mathrm{e}^{\mathrm{i}\langle u,X_{k}\rangle}-\mathbb{E}[\mathrm{e}^{\mathrm{i}\langle u,X_{k}\rangle}]\big{)}\] so that the assertion reads \[\sup_{\begin{subarray}{c}n\geq 1,\\ 0<\rho\leq R\end{subarray}}\mathbb{E}\big{[}\big{\|}w(u)C^{\beta}_{\rho,n}(u) \big{\|}_{\infty}\big{]}<\infty.\] We decompose \(C^{\beta}_{\rho,n}\) into its real and its imaginary part to obtain \[\mathbb{E}\big{[}\big{\|}w(u)C^{\beta}_{\rho,n}(u)\big{\|}_{\infty}\big{]} \leqslant\mathbb{E}\big{[}\big{\|}w(u)\operatorname{Re}\big{(}C^{\beta}_{ \rho,n}(u)\big{)}\big{\|}_{\infty}\big{]}+\mathbb{E}\big{[}\big{\|}w(u) \operatorname{Im}\big{(}C^{\beta}_{\rho,n}(u)\big{)}\big{\|}_{\infty}\big{]}.\] As both parts can be treated analogously, we focus on the real part. To this end, introduce the class of \[\mathcal{G}_{\rho,\beta}\coloneqq\{g_{u}:u\in\mathbb{R}^{d}\}\qquad\text{ where}\qquad g_{u}\colon\mathbb{R}^{d}\to\mathbb{R},\qquad x\mapsto w(u)\rho^{-(|\beta|_{1} \wedge 1)}\,\frac{\partial^{\beta}}{\partial u^{\beta}}\cos(\langle u,x\rangle).\] Since \(G=\rho^{-(|\beta|_{1}\wedge 1)/2}|\cdot|^{\beta}\) is an envelope function for \(\mathcal{G}_{\rho,\beta}\), van der Vaart (1998, Corollary 19.35) yields \[\mathbb{E}\big{[}\big{\|}w(u)\operatorname{Re}\big{(}C^{\beta}_{\rho,n}(u) \big{)}\big{\|}_{\infty}\big{]}\lesssim J_{\|}\big{(}\mathbb{E}[G(X_{1})^{2} ]^{1/2},\mathcal{G}_{\rho,\beta}\big{)}\coloneqq\int_{0}^{\mathbb{E}[G(X_{1}) ^{2}]^{1/2}}\sqrt{\log N_{\|}(\varepsilon,\mathcal{G}_{\rho,\beta})}\,\mathrm{ d}\varepsilon,\] where \(N_{\|}(\varepsilon,\mathcal{G}_{\rho,\beta})\) is the minimal number of \(\varepsilon\)-brackets (with respect to the distribution of \(X_{1}\)) needed to cover \(\mathcal{G}_{\rho,\beta}\). Since \(|g_{u}(x)|\leqslant w(u)\rho^{-(|\beta|_{1}\wedge 1)/2}|x|^{\beta}\), the set \(\{g_{u}:|u|>B\}\) is covered by the bracket \[[g_{0}^{-},g_{0}^{+}]\coloneqq\{g\colon\mathbb{R}^{d}\to \mathbb{R}\mid g_{0}^{-}(x)\leqslant g(x)\leqslant g_{0}^{+}(x)\,\forall x \in\mathbb{R}^{d}\}\qquad\text{for}\] \[g_{0}^{\pm}\coloneqq\pm\varepsilon\rho^{-(|\beta|_{1}\wedge 1)/2}| \cdot|^{\beta}\qquad\text{and}\qquad B\coloneqq B(\varepsilon)\coloneqq\inf \big{\{}b>0:\sup_{|u|\geqslant b}w(u)\leqslant\varepsilon\big{\}}.\] To cover \(\{g_{u}:|u|\leqslant B\}\), we use for some grid \((u_{\rho,j})_{j\geqslant 1}\subseteq\mathbb{R}^{d}\) the functions \[g_{\rho,j}^{\pm}\coloneqq\rho^{-(|\beta|_{1}\wedge 1)/2}\big{(}w(u_{\rho,j}) \frac{\partial^{\beta}}{\partial u_{\rho,j}^{\beta}}\cos(\langle u_{\rho,j}, \cdot\rangle)\pm\varepsilon|\cdot|^{\beta}\big{)}\mathbbm{1}_{\{|\cdot| \leqslant M\}}\pm\rho^{-(|\beta|_{1}\wedge 1)/2}|\cdot|^{\beta}\mathbbm{1}_{\{|\cdot|>M\}}.\] where \(M\coloneqq\inf\{m:\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2\beta} \mathbbm{1}_{\{|X_{1}|>m\}}]\leqslant\varepsilon^{2}\}\). Owing to \(\mathbb{E}[|X_{1}|^{2\beta}]\lesssim\rho^{|\beta|_{1}\wedge 1}\), we have \[\mathbb{E}\big{[}[g_{j}^{+}(X_{1})-g_{j}^{-}(X_{1})|^{2}\big{]}\leqslant 4 \varepsilon^{2}\big{(}\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2\beta}]+1 \big{)}\leqslant c\varepsilon^{2}\] for some \(c>0\). Denote by \(C\) the Lipschitz constant of \(w\) and use the triangle inequality to see \[\Big{|}w(u)\frac{\partial^{\beta}}{\partial u^{\beta}}\cos(\langle u,x\rangle )-w(u_{j})\frac{\partial^{\beta}}{\partial u_{\rho,j}^{\beta}}\cos(\langle u_{\rho,j}, x\rangle)\Big{|}\leqslant|x|^{\beta}(C+|x|)|u-u_{\rho,j}|.\] Thus, \(g_{u}\in[g_{j}^{-},g_{j}^{+}]\) as soon as \((C+M)|u-u_{\rho,j}|\leqslant\varepsilon\). It takes at most \((\lceil B/\varepsilon_{0}\rceil)^{d}\)\(\ell^{2}\)-balls of radius \(d^{1/2}\varepsilon_{0}\) to cover the \(\ell^{2}\)-ball of radius \(B\) around \(0\). For \(\varepsilon_{0}=\varepsilon d^{-1/2}/(C+M)\), denote their centers by \((u_{\rho,j})_{j}\). To translate this into a cover of \(\{g_{u}:|u|\leqslant B\}\), we fix some \(g_{u}\) with \(|u|\leqslant B\). By construction, we can pick \(j\) such that \(|u-u_{\rho,j}|\leqslant d^{1/2}\varepsilon_{0}=\varepsilon/(C+M)\). The previous calculations show that \([g_{j}^{-},g_{j}^{+}]\) is a \(c^{1/2}\varepsilon\)-bracket containing \(g_{u}\) and therefore \[N_{\|}(\varepsilon,\mathcal{G}_{\rho,\beta})\leqslant\big{(}\big{\lceil} \varepsilon^{-1}(cd)^{1/2}B(C+M)\big{\rceil}\big{)}^{d}+1.\] It is straightforward to see \(B\leqslant\exp(\varepsilon^{-2/(1+\chi)})\). Further, \(m=(\varepsilon^{-2}\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2\beta}|X_{1}|^{ \tau}])^{1/\tau}\) is sufficient for \[\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2\beta}\mathbb{1}_{\{|X_{1}| >m\}}]\leqslant m^{-\tau}\mathbb{E}[|X_{1}|^{2\beta}|X_{1}|^{\tau}]\leqslant \varepsilon^{2}\] and thus \(M\leqslant(\varepsilon^{-2}\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2 \beta}|X_{1}|^{\tau}])^{1/\tau}\leqslant(\varepsilon^{-2}dc^{\prime})^{1/\tau}\) for some \(c^{\prime}>0\). Hence, \[\log N_{\parallel}(\varepsilon,\mathcal{G}_{\rho,\beta})\lesssim 1+\log( \varepsilon^{-2/\tau-1})+\varepsilon^{-2/(1+\tau)}\lesssim 1+\varepsilon^{-2/(1+ \tau)}\] implying \[J_{\parallel}\big{(}\mathbb{E}[G(X_{1})^{2}]^{1/2},\mathcal{G}_{\rho,\beta} \big{)}=\int_{0}^{(\rho^{-(|\beta|_{1}\wedge 1)}\mathbb{E}[|X_{1}|^{2\beta}])^{1/2}} \sqrt{\log N_{\parallel}(\varepsilon,\mathcal{G}_{\rho,\beta})}\,\mathrm{d} \varepsilon<\infty.\qed\] #### 4.4.2 Proof of Lemma 6 Setting \(g(y)=\log(1+y)\) (ie. \(g^{\prime}(y)=(1+y)^{-1},g^{\prime\prime}(y)=-(1+y)^{-2}\)) and \(\xi=(\widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_{\delta}\), we use \[\nabla(g\circ\xi)(u)=g^{\prime}(\xi(u))\nabla\xi(u),\qquad\Delta(g\circ\xi)(u )=g^{\prime\prime}(\xi(u))(\nabla\xi)^{2}(u)+g^{\prime}(\xi(u))\Delta\xi(u)\] and \(|(\nabla\xi)^{2}(u)|\leqslant|\nabla\xi(u)|^{2}\) to obtain for \(|\xi(u)|\leqslant 1/2\) that \[\big{|}\Delta(g\circ\xi)(u)-\Delta\xi(u)\big{|}\leqslant|g^{\prime\prime}(\xi (u))||\nabla\xi(u)|^{2}+|g^{\prime}(\xi(u))-1||\Delta\xi(u)|\lesssim|\nabla \xi(u)|^{2}+|\xi(u)||\Delta\xi(u)|, \tag{20}\] because \[|g^{\prime}(y)-1|\leqslant 2|y|\qquad\text{and}\qquad|g^{\prime\prime}(y)| \leqslant 4\qquad\forall y\in\mathbb{C}:|y|\leqslant 1/2.\] The latter statement holds, since \(1/2\leqslant|1+y|\). For the former statement, consider the expansion \[g^{\prime}(y)=\frac{1}{1+y}=\sum_{k=0}^{\infty}(-y)^{k}\qquad\forall y\in \mathbb{C}:|y|\leqslant 1/2\] to see \[|g^{\prime}(y)-1|=\Big{|}\sum_{k=1}^{\infty}(-y)^{k}\Big{|}=\Big{|}-y\sum_{k=0 }^{\infty}(-y)^{k}\Big{|}=\Big{|}\frac{y}{1+y}\Big{|}\leqslant 2|y|.\] Noting \(\widehat{\Delta\psi_{n}}-\Delta\psi=\delta^{-1}\Delta\log(\widehat{\varphi}_ {\delta,n}/\varphi_{\delta})\), (20) implies on the event \(\Omega_{n}:=\big{\{}\sup_{|u|_{\infty}\leqslant h^{-1}}|\xi(u)|\leqslant 1/2\big{\}}\) \[\sup_{|u|_{\infty}\leqslant h^{-1}}\big{|}\delta(\widehat{\Delta\psi_{n}}- \Delta\psi)(u)-\Delta((\widehat{\varphi}_{\delta,n}-\varphi_{\delta})/\varphi_ {\delta})(u)\big{|}\lesssim\||\nabla\xi||_{L^{\infty}(I_{h})}^{2}+\|\xi\|_{L^ {\infty}(I_{h})}\|\Delta\xi\|_{L^{\infty}(I_{h})}.\] To control the \(\xi\)-terms, we invoke Proposition 5 applied to the increments of the Levy process with \(\rho=\delta\) after verifying that the moments are of the appropriate order. Owing to the equivalence of norms, it is sufficient to show that with \(\tau=m-4>0\) \[\mathbb{E}[|Y_{1,k}|^{2l+\tau}]\lesssim\delta^{l\wedge 1}\qquad\text{and} \qquad\mathbb{E}[|Y_{1,k}|^{2l}]\lesssim\delta^{l\wedge 1},\qquad k=1,\dots,d,\,l=0,1,2, \tag{21}\] where \(Y_{1,k}\) is the \(k\)-th entry of \(Y_{1}\) and thus an increment with time difference \(\delta\) based on the Levy process \((L_{t,k})_{t\geqslant 0}\) with Levy measure \(\nu_{k}\). For \(l=1,2\), it follows from Figueroa-Lopez (2008, Theorem 1.1) that \[\lim_{\delta\searrow 0}\delta^{-1}\mathbb{E}[|Y_{1,k}|^{2l+\tau}]=\lim_{\delta \searrow 0}\delta^{-1}\mathbb{E}[|L_{\delta,k}|^{2l+\tau}]=\int|x_{k}|^{2l+\tau }\,\nu_{k}(\mathrm{d}x_{k})\leqslant\int|x|^{2l+\tau}\,\nu(\mathrm{d}x) \lesssim R.\] For \(l=0\), \(\mathbb{E}[|Y_{1,k}|^{\tau}]\lesssim\mathbb{E}[|Y_{1,k}|^{m}]\lesssim 1\) holds by our moment assumptions. The second condition in (21) was already checked at the beginning of Section 4.1.1. Therefore, \(|\Delta\psi(u)|\lesssim 1\) yields \[\|\xi\|_{L^{\infty}(I_{h})} =\mathcal{O}_{\mathbb{P}}\big{(}n^{-1/2}(\log h^{-1})^{(1+\chi)/2} \|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\big{)}, \tag{22}\] \[\||\nabla\xi||_{L^{\infty}(I_{h})}^{2} =\mathcal{O}_{\mathbb{P}}\big{(}n^{-1}(\log h^{-1})^{1+\chi}\| \varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}^{2}\big{(}\delta+\delta^{2}\|| \nabla\psi|\|_{L^{\infty}(I_{h})}^{2}\big{)}\big{)},\] \[\|\Delta\xi\|_{L^{\infty}(I_{h})} =\mathcal{O}_{\mathbb{P}}\big{(}n^{-1/2}(\log h^{-1})^{(1+\chi)/2} \|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\big{(}\delta^{1/2}+\delta^{3/2} \big{\|}|\nabla\psi|\big{\|}_{L^{\infty}(I_{h})}+\delta^{2}\big{\|}|\nabla\psi| \big{\|}_{L^{\infty}(I_{h})}^{2}\big{)}\big{)}.\] Combining (22) with \(n^{-1/2}(\log h^{-1})^{(1+\chi)/2}\|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}\to 0\) gives \(\mathbb{P}(\Omega_{n})\to 1\), completing the proof. #### 4.4.3 Proof of Lemma 7 For fixed \(x\in\mathbb{R}^{d}\), we want to apply Bernstein's inequality to \[M^{\nu}_{\delta,n}(x)=-\sum_{l=1}^{n}\big{(}\xi_{l}-\mathbb{E}[\xi_{l}]\big{)} \qquad\text{with}\qquad\xi_{l}\coloneqq T^{-1}\mathcal{F}^{-1}\big{[}m_{\delta, h}(u)|Y_{l}|^{2}\mathrm{e}^{\mathrm{i}\langle u,Y_{l}\rangle}\big{]}(x).\] Similar arguments to (13) reveal \(\big{\|}m_{\delta,h}(u)\big{\|}_{L^{1}}\lesssim h^{-\delta\alpha-d}\), and with the quotient rule one finds the same order for \(\big{\|}\Delta m_{\delta,h}(u)\big{\|}_{L^{1}}\) and \(\big{\|}\nabla m_{\delta,h}\big{\|}_{L^{1}}\) paving a deterministic bound of \(\xi_{l}\) via \[|\xi_{l}| =T^{-1}|Y_{l}|^{2}\big{|}\mathcal{F}^{-1}\big{[}m_{\delta,h}(u) \mathrm{e}^{-\mathrm{i}\langle u,x\rangle}\big{]}(-Y_{l})\big{|}\] \[=T^{-1}\big{|}\mathcal{F}^{-1}\big{[}\Delta\big{(}m_{\delta,h}(u )\mathrm{e}^{-\mathrm{i}\langle u,x\rangle}\big{)}\big{]}(-Y_{l})\big{|}\] \[\leqslant T^{-1}\big{\|}\Delta\big{(}m_{\delta,h}(u)\mathrm{e}^{- \mathrm{i}\langle u,x\rangle}\big{)}\big{\|}_{L^{1}}\] \[\leqslant T^{-1}\big{(}\big{\|}\Delta m_{\delta,h}(u)\big{\|}_{L ^{1}}+2|x|_{1}\big{\|}\big{\|}\nabla m_{\delta,h}\big{\|}_{L^{1}}+|x|^{2}\|m_{ \delta,h}\|_{L^{1}}\big{)}\] \[\lesssim T^{-1}(1+|x|^{2})h^{-\delta\alpha-d}.\] To bound the variance of \(\xi_{l}\), note that for the distribution \(\mathbb{P}_{\delta}\) of \(Y_{1}\), we have \[\mathcal{F}[\mathrm{i}z_{k}\mathbb{P}_{\delta}]=\frac{\partial\varphi_{\delta }}{\partial u_{k}}=\delta\varphi_{\delta}\frac{\partial\psi}{\partial u_{k}}= \delta\mathcal{F}[\mathrm{i}z_{k}\nu]\varphi_{\delta}=\mathcal{F}[\mathcal{F}^ {-1}[\delta\mathcal{F}[\mathrm{i}z_{k}\nu]\varphi_{\delta}]]=\mathcal{F}[( \delta\mathrm{i}z_{k}\nu)*\mathbb{P}_{\delta}]\] and therefore \(z_{k}\mathbb{P}_{\delta}=\delta\mu*\mathbb{P}_{\delta}\) with \(\mu(\mathrm{d}z)=z_{k}\nu(z)\mathrm{d}z\). It follows that \[\int g(z)|z_{k}|\,\mathbb{P}_{\delta}(\mathrm{d}z)\leqslant\delta\|z_{k}\nu \|_{\infty}\|g\|_{L^{1}},\qquad\forall g\in L^{1}(\mathbb{R}^{d}).\] Again, using similar arguments to (13) and the quotient rule, we also have \(\big{\|}\Delta m_{\delta,h}(u)\big{\|}_{L^{2}},\big{\|}\big{\|}\nabla m_{ \delta,h}\big{\|}_{L^{2}}\lesssim h^{-\delta\alpha-d/2}\). Thus, the Cauchy-Schwarz inequality and the Plancherel theorem imply \[\mathrm{Var}(\xi_{l})\leqslant\mathbb{E}[|\xi_{l}|^{2}]\] \[\lesssim T^{-2}\sum_{k=1}^{d}\int|y|^{3}\big{|}\mathcal{F}^{-1} \big{[}m_{\delta,h}(u)\mathrm{e}^{-\mathrm{i}\langle u,x\rangle}\big{]}(-y) \big{|}^{2}|y_{k}|\,\mathbb{P}_{\delta}(\mathrm{d}y)\] \[\leqslant n^{-2}\delta^{-1}\sum_{k=1}^{d}\|z_{k}\nu\|_{\infty}\int |y|^{3}\big{|}\mathcal{F}^{-1}\big{[}m_{\delta,h}(u)\mathrm{e}^{-\mathrm{i} \langle u,x\rangle}\big{]}(y)\big{|}^{2}\,\mathrm{d}y\] \[\lesssim n^{-2}\delta^{-1}\big{\|}\big{\|}y|^{2}\mathcal{F}^{-1} \big{[}m_{\delta,h}(u)\mathrm{e}^{-\mathrm{i}\langle u,x\rangle}\big{]}(y) \big{\|}_{L^{2}}\big{\|}y|_{1}\mathcal{F}^{-1}\big{[}m_{\delta,h}(u)\mathrm{e} ^{-\mathrm{i}\langle u,x\rangle}\big{]}(y)\big{\|}_{L^{2}}\] \[\lesssim n^{-2}\delta^{-1}\Big{(}\sum_{k=1}^{d}\Big{\|}\frac{ \partial^{2}}{\partial u_{k}^{2}}\big{(}m_{\delta,h}(u)\mathrm{e}^{-\mathrm{ i}\langle u,x\rangle}\big{)}\Big{\|}_{L^{2}}\Big{)}\Big{(}\sum_{k=1}^{d}\Big{\|} \frac{\partial}{\partial u_{k}}\big{(}m_{\delta,h}(u)\mathrm{e}^{-\mathrm{i} \langle u,x\rangle}\big{)}\Big{\|}_{L^{2}}\Big{)}\] \[\lesssim n^{-2}\delta^{-1}h^{-2\delta\alpha-d}(1+|x|^{3}).\] Now, Bernstein's inequality, e.g. van der Vaart (1998, Lemma 19.32) yields for a constant \(c^{\prime}>0\) and any \(\kappa>0\) that \[\mathbb{P}\big{(}|M^{\nu}_{\delta,n}(x)|\geqslant\kappa\big{)}\leqslant 2\exp \Big{(}-\frac{Tc^{\prime}\kappa^{2}}{h^{-2\delta\alpha-d}(1+|x|^{3})+\kappa(1+|x| ^{2})h^{-\delta\alpha-d}}\Big{)},\] which reads as the assertion if we choose \(\kappa=\kappa_{0}T^{-1/2}h^{-\delta\alpha-d/2}\) for any \(\kappa_{0}>0\) and set \(c=c^{\prime}/2\) #### 4.4.4 Proof of Lemma 8 Fix \(x=(x^{(1)},x^{(2)})\) for \(x^{(1)},x^{(2)}\in\mathbb{R}^{d/2}\) and analogously split \(Y_{l}\) into its first and last \(d/2\) entries \(Y_{l}^{(1)}\) and \(Y_{l}^{(2)}\) with characteristic functions \(\varphi_{\delta,1}\) and \(\varphi_{\delta,2}\), respectively. Due to the product kernel, we obtain \[\xi_{l}=T^{-1}|Y_{l}|^{2}\big{|}\mathcal{F}^{-1}\big{[}m_{\delta,h}(u)\mathrm{e }^{-\mathrm{i}(u,x)}\big{]}(-Y_{l})\big{|}=T^{-1}(A_{1}B_{2}+A_{2}B_{1})\] with \[A_{k}\coloneqq|Y_{l}^{(k)}|^{2}\mathcal{F}^{-1}[m_{\delta,h}^{ (k)}(u^{(k)})\mathrm{e}^{\mathrm{i}(u^{(k)},Y_{l}^{(k)})}](x^{(k)}),\qquad B_{ k}\coloneqq\mathcal{F}^{-1}[m_{\delta,h}^{(k)}(u^{(k)})\mathrm{e}^{\mathrm{i}(u^{(k)},Y_{l}^{(k)})}](x^{(k)}),\qquad\text{and}\] \[m_{\delta,h}^{(k)}(u^{(k)})\coloneqq\varphi_{\delta,k}^{-1}(u^{ (k)})\prod_{j=1+(k-1)d/2}^{(k+1)d/2}\mathcal{F}K^{j}(hu_{j}^{(k)}),\qquad k=1,2.\] \(A_{1}\) and \(A_{2}\) are the same terms that appeared in the proof of Lemma 7 just with half the dimension and therefore \[|A_{k}| \lesssim\|\varphi_{\delta,h}^{-1}\|_{L^{\infty}([-h^{-1},h^{-1}] ^{d/2}])}h^{-d/2}(1+|x^{(k)}|^{2}),\] \[\mathbb{E}[|A_{k}|^{2}] \lesssim\delta\|\varphi_{\delta,h}^{-1}\|_{L^{\infty}([-h^{-1},h^ {-1}]^{d/2}])}^{2}h^{-d/2}(1+|x^{(k)}|^{3}).\] In a similar vain, \(m_{\delta,h}^{(1)}\) and \(m_{\delta,h}^{(2)}\) can be treated be treated like \(m_{\delta,h}\) with half the dimension leading to \[|B_{k}|\lesssim\|m_{\delta,h}^{(k)}\|_{L^{1}(\mathbb{R}^{d/2})}\lesssim\| \varphi_{\delta,k}^{-1}\|_{L^{\infty}([-h^{-1},h^{-1}]^{d/2}])}h^{-d/2}.\] Note that \[\prod_{k=1}^{2}\|\varphi_{\delta,k}^{-1}\|_{L^{\infty}([-h^{-1},h^{-1}]^{d/2} ])}=\|\varphi_{\delta}^{-1}\|_{L^{\infty}(I_{h})}.\] Together, we have the deterministic bound \(|\xi_{l}|\lesssim T^{-1}h^{-\delta\alpha-d}(1+|x|^{2})\). Further, since \(A_{1}\) and \(B_{2}\) as well as \(A_{2}\) and \(B_{1}\) are independent, we obtain \[\mathrm{Var}(\xi_{l})\lesssim T^{-2}(\mathrm{Var}(A_{1}B_{2})+ \mathrm{Var}(A_{2}B_{1})) \leqslant T^{-2}\big{(}\mathbb{E}[|A_{1}|^{2}]\mathbb{E}[|B_{2}|^ {2}]+\mathbb{E}[|A_{2}|^{2}]\mathbb{E}[|B_{1}|^{2}]\big{)}\] \[\lesssim n^{-2}\delta^{-1}h^{-2\delta\alpha-3d/2}(1+|x|^{3}).\] Overall, Bernstein's inequality gives for a constant \(c^{\prime}>0\) and any \(\kappa>0\) that \[\mathbb{P}\big{(}|M_{\delta,n}^{\nu}(x)|\geqslant\kappa\big{)}\leqslant 2\exp \Big{(}-\frac{Tc^{\prime}\kappa^{2}}{h^{-2\delta\alpha-d}(1+|x|^{3})+\kappa(1+ |x|^{2})h^{-\delta\alpha-d}}\Big{)}.\] The assertion follows by choosing \(\kappa=\kappa_{0}T^{-1/2}h^{-\delta\alpha-3d/4}\) for any \(\kappa_{0}>0\) and setting \(c=c^{\prime}/2\).

2308.12798

A preplanned multi-stage platform trial for discovering multiple superior treatments with control of FWER and power

There is a growing interest in the implementation of platform trials, which provide the flexibility to incorporate new treatment arms during the trial and the ability to halt treatments early based on lack of benefit or observed superiority. In such trials, it can be important to ensure that error rates are controlled. This paper introduces a multi-stage design that enables the addition of new treatment arms, at any point, in a pre-planned manner within a platform trial, while still maintaining control over the family-wise error rate. This paper focuses on finding the required sample size to achieve a desired level of statistical power when treatments are continued to be tested even after a superior treatment has already been found. This may be of interest if there are other sponsors treatments which are also superior to the current control or multiple doses being tested. The calculations to determine the expected sample size is given. A motivating trial is presented in which the sample size of different configurations is studied. Additionally the approach is compared to running multiple separate trials and it is shown that in many scenarios if family wise error rate control is needed there may not be benefit in using a platform trial when comparing the sample size of the trial.

Peter Greenstreet, Thomas Jaki, Alun Bedding, Pavel Mozgunov

2023-08-24T13:57:06Z

http://arxiv.org/abs/2308.12798v1

A preplanned multi-stage platform trial for discovering multiple superior treatments with control of FWER and power ###### Abstract There is a growing interest in the implementation of platform trials, which provide the flexibility to incorporate new treatment arms during the trial and the ability to halt treatments early based on lack of benefit or observed superiority. In such trials, it can be important to ensure that error rates are controlled. This paper introduces a multi-stage design that enables the addition of new treatment arms, at any point, in a pre-planned manner within a platform trial, while still maintaining control over the family-wise error rate. This paper focuses on finding the required sample size to achieve a desired level of statistical power when treatments are continued to be tested even after a superior treatment has already been found. This may be of interest if there are other sponsors treatments which are also superior to the current control or multiple doses being tested. The calculations to determine the expected sample size is given. A motivating trial is presented in which the sample size of different configurations is studied. Additionally the approach is compared to running multiple separate trials and it is shown that in many scenarios if family wise error rate control is needed there may not be benefit in using a platform trial when comparing the sample size of the trial. ## 1 Introduction Platform trials are a type of trial design which can aim to reduce the amount of time and cost of clinical trials and in recent years, there has been an increase in the utilization of such trials, including during the COVID-19 pandemic (Stallard et al., 2020; Lee et al., 2021). Clinical trials take many years to run and can cost billions of dollars (Mullard, 2018). During this time it is not uncommon for new promising treatments to emerge and become ready to join the current phase later (Choodari-Oskooei et al., 2020). Therefore it may be advantageous to include these treatments into an ongoing trial. This can have multiple potential benefits including: shared trial infrastructure; the possibility to use a shared control group; less administrative and logistical effort than setting up separate trials and enhance the recruitment (Burnett et al., 2020; Meurer et al., 2012). This results in useful therapies potentially being identified faster while reducing cost and time (Cohen et al., 2015). There is an ongoing discussion about how to add new treatments to clinical trials (Cohen et al., 2015; Lee et al., 2021) in both a pre-planned and in an unplanned manor (Greenstreet et al., 2021; Burnett et al., 2020). In both Bennett and Mander (2020); Choodari-Oskooei et al. (2020) approaches are proposed which extend the Dunnett test (Dunnett, 1955) to allow for unplanned additional arms to be included into multi-arm trials while still controlling the family wise error rate (FWER). This methodology does not incorporate the possibility of interim analyses. FWER is often considered to be one of the strongest types of type I error control in a multi-arm trial (Wason et al., 2016). There are other approaches one may wish to consider such as pairwise error rate (PWER) and the false discovery rate (FDR) (Robertson et al., 2022; Cui et al., 2023; Bratton et al., 2016; Choodari-Oskooei et al., 2020). However as discussed in Wason et al. (2014) there are scenarios where FWER is seen as the recommended error control, and it can be a regulatory requirement. One may wish to include interim analyses as they allow for ineffective treatments to be dropped for futility earlier and allow treatments to stop early if they are found superior to the control. Therefore potentially improving the efficiency of design of a clinical trial by decreasing the expected sample sizes and costs of a trial (Pocock, 1977; Todd et al., 2001; Wason et al., 2016). Multi-arm multi-stage (MAMS) designs (Magir et al., 2012; Royston et al., 2003) allow interim analyses while still allowing several treatments to be evaluated within one study, but do not allow for additional arms to be added throughout the trial. Burnett et al. (2020) have developed an approach that builds on Hommel (2001) to incorporate unplanned additional treatment arms to be added to a trial already in progress using the conditional error principle (Proschan and Hunsberger, 1995). This allows for modifications during the course of a trial. However due to the unplanned nature of the adaptation, later treatments can be greatly underpowered compared to arms which begin the trial. In a recent paper Greenstreet et al. (2021) proposed a preplanned approach to adding additional arms in which interim analyses can be conducted and multiple arms can be evaluated with some arms being added at later time points. In this work the trial was powered assuming that only one treatment may be taken forward. However as discussed in the work by Urach and Posch (2016); Serra et al. (2022) this may not always be the case. For example one may be interested in lower doses; or multiple treatments from different sponsors; or interested if another treatment has preferable secondary outcomes if it also meets the primary outcome. Furthermore in Greenstreet et al. (2021) treatment arms can only be added when a interim analysis happens, this can greatly restrict when arms can join the trial resulting in potentially large time periods that a new treatment is available before an interim is conducted so able to join the trial. In this work, we provide an analytical method for adding of treatments at any point to a multi-arm multi-stage trial in a pre-planned manner, while still controlling the statistical errors. This work will focus on trials in which one is interested in continuing to investigate the other treatments even after a superior treatment has been found. In addition in this work multiple types of power will be considered, and will prove that the conjunctive power of the study is at its lowest for a given sample size when all the active treatments have a clinically relevant effect, where the conjunctive power is the probability of finding all the active treatments with a clinically relevant effect. The methodology discussed here can be used to create multiple designs for each point the additional treatments may be added into the trial. This is due to the model flexibility, as the additional arms do not need to be added when a interim happens, resulting in new active arms being able to be added faster into the platform trial. This work will focus predominantly on the case where one has equal allocation ratio across all the active treatments and the same number of interim analyses per treatment with the same boundary shape. This is to help mitigate issues with time trends (Altman and Royston, 1988; Getz and Campo, 2017) when changing allocation ratio mid trial (Proschan and Evans, 2020; Roig et al., 2023). However the proposed methodology is general and therefore can be implemented for when there is not equal allocation ratio across all the active treatments, however one needs to be cautious of potential time trend effects. We begin this work by analytically calculating the FWER and power of the study and use these to calculate both the stopping boundaries and sample size. Then in Section 2.4 the equations for sample size distribution and expected sample size are given. A trial example of FLAIR (Howard et al., 2021), in Section 3, is used to motivate a hypothetical trial of interest. The sample size and stopping boundaries are found for multiple types of power control and the effect of different treatment effects is studied. Then the trial designs are then compared to running multiple separate trials. Finally in Section 4 there is a discussion of the paper and this introduces areas for further research. ## 2 Methodology ### Setting In the clinical trial design considered in this work K experimental arms effectiveness is compared to a common control arm. The trial has \(K^{\star}\) treatments starting at the beginning of the trial, and the remaining \(K-K^{\star}\) treatments being added at later points into the platform. The primary outcome measured for each patient is assumed to be independent, continuous, and follows a normal distribution with a known variance (\(\sigma^{2}\)). The points at which each active treatment arm is added are predetermined, but can be set to any point within the trial. Each of the \(K\) treatments is potentially tested at a series of analyses indexed by \(j=1,\ldots,J_{k}\) where \(J_{k}\) is the maximum number of analyses for a given treatment \(k=1,\ldots,K\). Let \(n(k)\) denote the number of patients recruited to the control treatment before treatment \(k\) is added to the platform trial and define the vector of adding times by \(\mathbf{n}(\mathbf{K})=(n(1),\ldots,n(K))\). Therefore for treatments that start at the beginning of the trial \(n(k)=0\). We also denote \(n_{k,j}\) as the number of patients recruited to treatment \(k\) by the end of it's \(j^{\text{th}}\) stage and define \(n_{0,k,j}\) as the total number of patients recruited to the control at the end of treatment \(k\)'s \(j^{\text{th}}\) stage. We define \(n_{k}=n_{k,1}\) as the number recruited to the first stage of treatment \(k\), \(k=1,\ldots,K\). Similarly we define \(r_{k,j}\) and \(r_{0,k,j}\) as the ratio of patients recruited treatment \(k\) and the control by treatment \(k\)'s \(j^{\text{th}}\) stage, respectively. Also \(r(k)\) denotes the ratio of patients recruited to the control before treatment \(k\) is added to the trial. For example if a trial was planned to have equal number of patients per stage and a treatment (\(k^{\prime}\)) was added at the first interim then \(r(k^{\prime})=1\) and at the first stage for \(k^{\prime}\), \(r_{0,k^{\prime},1}=2\). The total sample size of a trial is denoted by \(N\). The maximum total planned sample size is \(\max(N)=\sum_{k=1}^{K}n_{k,J_{k}}+\max_{k\in 1,\ldots,K}(n_{0,k,J_{k}})\). Throughout the trial, the control arm is recruited and maintained for the entire duration. The comparisons between the control arm and the active treatment arms are based on concurrent controls, meaning that only participants recruited to the control arm at the same time as the corresponding active arm are used in the comparisons. Work on the use of non-concurrent controls include Lee and Wason (2020); Marschner and Schou (2022). The null hypotheses of interest are \(H_{01}:\mu_{1}\leq\mu_{0},H_{02}:\mu_{2}\leq\mu_{0},...,H_{0K}:\mu_{K}\leq\mu_ {0}\), where \(\mu_{1},\ldots,\mu_{K}\) are the mean responses on the \(K\) experimental treatments and \(\mu_{0}\) is the mean response of the control group. The global null hypothesis, \(\mu_{0}=\mu_{1}=\mu_{2}=\ldots=\mu_{K}\) is denoted by \(H_{G}\). At analysis \(j\) for treatment \(k\), to test \(H_{0k}\) it is assumed that responses, \(X_{k,i}\), from patients \(i=1,\ldots,n_{k,j}\) are observed, as well as the responses \(X_{0,i}\) from patients \(i=n(k)+1,\ldots,n_{0,k,j}\). These are the outcomes of the patients allocated to the control which have been recruited since treatment \(k\) has been added into the trial up to the \(j^{\text{th}}\) analysis of treatment \(k\). The null hypotheses are tested using the test statistics \[Z_{k,j}=\frac{n_{k,j}^{-1}\sum_{i=1}^{n_{k,j}}X_{k,i}-(n_{0,k,j}-n(k))^{-1} \sum_{i=n(k)+1}^{n_{0,k,j}}X_{0,i}}{\sigma\sqrt{(n_{k,j})^{-1}+(n_{0,k,j}-n(k) )^{-1}}}.\] The decision-making for the trial is made by the upper and lower stopping boundaries, denoted as \(U_{k}=(u_{k,1},\ldots,u_{k,J_{k}})\) and \(L_{k}=(l_{k,1},\ldots,l_{k,J_{k}})\). These boundaries are utilized to determine whether to continue or halt a treatment arm or even the whole trial at various stages. The decision-making process is as follows: if the test statistic for treatment \(k\) at stage \(j\) exceeds the upper boundary \(u_{k,j}\), the null hypothesis \(H_{0k}\) is rejected, and the treatment is stopped with the conclusion that it is superior to the control. Conversely, if \(Z_{k,j}\) falls below the lower boundary \(l_{k,j}\), treatment k is stopped for futility for all subsequent stages of the trial. If neither the superiority nor futility conditions are met, \(l_{k,j}\leq Z_{k,j}\leq u_{k,j}\), treatment \(k\) proceeds to it's next stage \(j+1\). If all the active treatments are stopped than the trial stops. These bounds are found to control the type I error of desire for the trial. In this work we consider the familywise error rate (FWER) as the type I error control of focus as discussed in Section 2.2. ### Familywise error rate (FWER) The FWER in the strong sense is a way of defining the type I error of a trial with multiple hypotheses and is defined as \[P(\text{reject at least one true }H_{0k}\text{ under any null }\text{ configuation},k=1,\ldots,K)\leq\alpha. \tag{2.1}\] where \(\alpha\) is the desired level of control for the FWER. As proven in Greenstreet et al. (2021) which builds on Magirr et al. (2012), one can show that the FWER is controlled in the strong sense under the global null, as given in the Supporting Information Section 1. The FWER under the global null hypothesis is equal to \[1-\sum_{\begin{subarray}{c}j_{k}=1\\ k=1,2,\ldots,K\end{subarray}}^{J_{k}}\Phi(\mathbf{L_{j_{k}}(0)},\mathbf{U_{j_{k}} (0)},\Sigma_{\mathbf{j_{k}}}). \tag{2.2}\] Here \(\Phi(\cdot)\) denotes the multivariate standard normal distribution function, and \(\mathbf{j_{k}}=(j_{1},\ldots,j_{K})\). With \(\mathbf{j_{k}}\) one can define the vector of upper and lower limits from the multivariate standard normal distribution function as \(\mathbf{U_{j_{k}}(0)}=(U_{1,j_{1}}(0),\ldots,U_{K,j_{K}}(0))\) and \(\mathbf{L_{j_{k}}(0)}=(L_{1,j_{1}}(0),\ldots,L_{K,j_{K}}(0))\) where \(U_{k,j_{k}}(0)=(u_{k,1},\ldots,l_{k,j_{k}})\) and \(L_{k,j_{k}}(0)=(l_{k,1},\ldots,-\infty)\) respectively. \(U_{k,j_{k}}(0)\) and \(L_{k,j_{k}}(0)\) represent the upper and lower limits for treatment \(k\) given \(j_{k}\) for the multivariate standard normal distribution function. The correlation matrix \(\Sigma_{\mathbf{j_{k}}}\) complete correlation structure is \[\Sigma_{\mathbf{j_{k}}}=\begin{pmatrix}\rho_{(1,1),(1,1)}&\rho_{(1,1),(1,2)}& \ldots&\rho_{(1,1),(1,j_{1})}&\rho_{(1,1),(2,1)}&\ldots&\rho_{(1,1),(K,j_{k})} \\ \rho_{(1,2),(1,1)}&\rho_{(1,2),(1,2)}&\ldots&\rho_{(1,2),(1,j_{1})}&\rho_{(1, 2),(2,1)}&\ldots&\rho_{(1,2),(K,j_{k})}\\ \vdots&\vdots&\ddots&\vdots&\vdots&\ddots&\vdots\\ \rho_{(1,j_{1}),(1,1)}&\rho_{(1,j_{1}),(1,2)}&\ldots&\rho_{(1,j_{1}),(1,j_{1} )}&\rho_{(1,j_{1}),(2,1)}&\ldots&\rho_{(1,j_{1}),(K,j_{k})}\\ \rho_{(2,1),(1,1)}&\rho_{(2,1),(1,2)}&\ldots&\rho_{(2,1),(1,j_{1})}&\rho_{(2, 1),(2,1)}&\ldots&\rho_{(2,1),(K,j_{k})}\\ \vdots&\vdots&\ddots&\vdots&\vdots&\ddots&\vdots\\ \rho_{(K,j_{k}),(1,1)}&\rho_{(K,j_{k}),(1,2)}&\ldots&\rho_{(K,j_{k}),(1,j_{1} )}&\rho_{(K,j_{k}),(2,1)}&\ldots&\rho_{(K,j_{k}),(K,j_{k})}\end{pmatrix}. \tag{2.3}\] where \(\rho_{(k,j),(k^{\star},j^{\star})}\) equals one of the following: If \(k=k^{\star}\) and \(j=j^{\star}\) then \(\rho_{(k,j),(k^{\star},j^{\star})}=1\); If \(k=k^{\star}\) and \(j<j^{\star}\) then \[\rho_{(k,j),(k^{\star},j^{\star})}=\left(\sqrt{r_{k,j}^{-1}+(r_{ 0,k,j}-r(k))^{-1}}\sqrt{r_{k,j^{\star}}^{-1}+(r_{0,k,j^{\star}}-r(k))^{-1}} \right)^{-1}\] \[\left(\frac{1}{r_{k,j^{\star}}}+\frac{1}{r_{0,k,j^{\star}}-r(k)} \right);\] and if \(k\neq k^{\star}\) where \(n(k)<n(k^{\star})\) then \[\rho_{(k,j),(k^{\star},j^{\star})}=\max\Bigg{[}0, \Bigg{(}\sqrt{r_{k,j}^{-1}+(r_{0,k,j}-r(k))^{-1}}\sqrt{r_{k^{ \star},j^{\star}}^{-1}+(r_{0,k^{\star},j^{\star}}-r(k^{\star}))^{-1}}\Bigg{)}^ {-1}\] \[\left(\frac{\min[r_{0,k,j}-r(k^{\star}),r_{0,k^{\star},j^{\star}}- r(k^{\star}))}{[r_{0,k,j}-r(k)][r_{0,k^{\star},j^{\star}}-r(k^{\star})]}\right) \Bigg{]}.\] As seen here if treatment \(k^{\star}\) is added to the platform trial after the \(j\) stage for treatment \(k\) then the correlation equals \(0\) as there is no shared controls. The proposed methodology allows for different critical boundaries to be used for each treatment \(k\) as shown in Equation (2.2). If it is assumed that there is equal number of stages per treatment and equal allocation across all the active treatments then, as a result, if one is using the same stopping boundary shape one can simply just calculate the FWER. This is because it results in equal pairwise error rate (PWER) for each treatment (Bratton et al., 2016; Choodari-Oskooei et al., 2020; Greenstreet et al., 2021). This removes the potential issue of time trends with changing allocation ratios. Therefore to find the boundaries one can use a single scalar parameter \(a\) with the functions \(L_{k}=f(a)\) and \(U_{k}=g(a)\) where \(f\) and \(g\) are the functions for the shape of the upper and lower boundaries respectively. This is similar to the method presented in Magirr et al. (2012). ### Power When designing a multi-arm trial in which all treatments get tested until they are stopped for futility or superiority, regardless of the other treatments, different definitions of power could be considered. The power of a study is focused on the probability that the trial results in some or all of the treatments going forward. The sample size of the study is then found to ensure that the chosen power is greater than or equal to some chosen value \(1-\beta\). One may be interested in ensuring that at least one treatment is taken forward from the study. This can be split into two types of power discussed in the literature. The first is the disjunctive power (Urach and Posch, 2016; Choodari-Oskooei et al., 2020; Hamasaki et al., 2021) which is the probability of taking at least one treatment forward. The second is the pairwise power which is the probability of taking forward a given treatment which has a clinically relevant effect (Choodari-Oskooei et al., 2020; Royston et al., 2011). In the Supporting Information Section 2 the equations needed to calculate the disjunctive power (\(P_{D}\)) are given. Another way of thinking of powering a study is the probability of taking forward all the treatments which have a clinically relevant effect. This is known as the conjunctive power of a study (Urach and Posch, 2016; Choodari-Oskooei et al., 2020; Hamasaki et al., 2021; Serra et al., 2022). For the conjunctive power we prove that it is lowest when all the treatments have the clinically relevant effect. #### 2.3.1 Pairwise power The pairwise power of a treatment is independent of other active treatments. This is because the other active treatments effect has no influence on the treatment of interest as these are independent. Therefore we only need to consider the probability that the treatment of interest with a clinically relevant effect is found superior to the control. The pairwise power for treatment \(k\) (\(P_{pw,k}\)) with the clinically relevant effect \(\theta^{\prime}\) is: \[P_{pw,k}=\sum_{j=1}^{J_{k}}\Phi(U_{k,j}^{+}(\theta^{\prime}),L_{k,j}^{+}( \theta^{\prime}),\ddot{\Sigma}_{k,j}), \tag{2.4}\] with \[L_{k,j}^{+}(\theta_{k})=(l_{k,1}-\frac{\theta_{k}}{\sqrt{I_{k,1} }},\ldots,l_{k,j-1}-\frac{\theta_{k}}{\sqrt{I_{k,j-1}}},u_{k,j}-\frac{\theta_{ k}}{\sqrt{I_{k,j-1}}}) \tag{2.5}\] \[U_{k,j}^{+}(\theta_{k})=(u_{k,1}-\frac{\theta_{k}}{\sqrt{I_{k,j- 1}}},\ldots,u_{k,j-1}-\frac{\theta_{k}}{\sqrt{I_{k,j-1}}},\infty). \tag{2.6}\] and \(I_{k,j}=\sigma^{2}(n_{k,j}^{-1}+(n_{0,k,j}-n(k))^{-1})\). The \((i,i^{\star})^{\text{th}}\) element (\(i\leq i^{\star}\)) of the covariance matrix \(\ddot{\Sigma}_{k,j}\) is \[\Bigg{(}\sqrt{r_{k,i}^{-1}+(r_{0,k,i}-r(k))^{-1}}\sqrt{r_{k,i^{\star}}^{-1}+(r_ {0,k,i^{\star}}-r(k))^{-1}}\Bigg{)}^{-1}\Bigg{(}\frac{1}{r_{k,i^{\star}}}+ \frac{1}{r_{0,k,i^{\star}}-r(k)}\Bigg{)}.\] One can then design the trial so that the pairwise power for each treatment \(k\) (\(P_{pw,k}\)) is greater than or equal to some chosen \(1-\beta\) for every treatment. If one has an equal number of stages per treatment and equal allocation across all the active treatments with the same stopping boundaries, this ensures that pairwise power is equal for each treatment so \(n_{k}=n_{k^{\star}}\) for all \(k,k^{\star}\in 1,\ldots,K\). Therefore we define \(n=n_{k}\) for all \(k\in 1,\ldots,K\). To ensure pairwise power is controlled, keep increasing \(n\) until \(P_{pw}\geq 1-\beta\) where \(P_{pw}=P_{pw,k}\) for all \(k\in 1,\ldots,K\). If one designing a trial in which there is a set number of patients to the control before an active treatment \(k\) is added, so \(n(k)\) is predefined before calculating the boundaries and sample size, one needs to use an approach such as the Algorithm below. This is because when the sample size increases there is no increase in \(n(k)\) for all \(k\). This results in a change in the allocation ratio between \(r(k)\) and \(r_{0,k,j}\) for each \(j\). Therefore requiring the bounds to be recalculated for the given \(r(k)\). If one focus is on the new arms being added after a set percentage of the way through the trial this issue no longer persists, as the allocation ratio stays the same so the bounds can be calculated ones. ``` 0 Begin by assuming \(\mathbf{n}(\mathbf{K})=(0,0,\ldots,0)\) and find the stopping boundaries to control the FWER. Now calculate \(n\) such that the pairwise power is greater then or equal to a pre-specified \((1-\beta)\). Then repeat the following iterative steps until the pairwise power, given the true \(\mathbf{n}(\mathbf{K})\), is greater than \((1-\beta)\): 1 Find the stopping boundaries to control the FWER for the true predefined \(\mathbf{n}(\mathbf{K})\) given the current \(n\). 2 Calculate \(P_{pw}\) for the given boundaries. 3 If \(P_{pw}\geq 1-\beta\) then stop, else increase \(n\) by 1 and repeat steps 1-3. ``` **Algorithm 1** Iterative approach to compute the \(n\) for the pairwise power with predefined \(\mathbf{n}(\mathbf{K})\) #### 2.3.2 Conjunctive power The conjunctive power is defined as the probability of taking forward all the treatments which have a clinically relevant effect. We begin by proving when the conjunctive power is at its lowest. We define the events \[B_{k,j}(\theta_{k})= [l_{k,j}+(\mu_{k}-\mu_{0}-\theta_{k})I_{k,j}^{1/2}<Z_{k,j}<u_{k,j}+ (\mu_{k}-\mu_{0}-\theta_{k})I_{k,j}^{1/2}],\] \[C_{k,j}(\theta_{k})= [Z_{k,j}>u_{k,j}+(\mu_{k}-\mu_{0}-\theta_{k})I_{k,j}^{1/2}],\] where \(B_{k,j}(\theta_{k})\) defines the event that treatment \(k\) continues to the next stage and \(C_{k,j}(\theta_{k})\) defines the event that treatment \(k\) is found superior to the control at stage \(j\). If \(\mu_{k}-\mu_{0}=\theta_{k}\) for \(k=1,\ldots,K\), the event that \(H_{01},\ldots,H_{0K}\) are all rejected (\(\bar{W}_{K}(\Theta)\)) is equivalent to \[\bar{W}_{K}(\Theta)=\bigcap_{k\in\{m_{1},\ldots,m_{K}\}}\Bigg{(}\bigcup_{j=1}^{ J_{k}}\Bigg{[}\Bigg{(}\bigcap_{i=1}^{j-1}B_{k,i}(\theta_{k})\Bigg{)}\cap C_{k,j}( \theta_{k})\Bigg{]}\Bigg{)},\] where \(\Theta=\{\theta_{1},\theta_{2},\ldots,\theta_{K}\}\). **Theorem 2.1**.: _For any \(\Theta\), \(P(\text{reject all }H_{0k}\) for which \(\theta_{k}\geq\theta^{\prime}|\Theta)\geq P(\text{reject all }H_{0k}\) for which \(\theta_{k}\geq\theta^{\prime}|(\mu_{1}=\mu_{2}=\ldots=\mu_{K}=\mu_{0}+\theta^{ \prime}))\)._ Proof.: For any \(\epsilon_{k}<0\), \[\bigcup_{j=1}^{J_{k}}\Bigg{[}\bigg{(}\bigcap_{i=1}^{j-1}B_{k,i}(\theta_{k}+ \epsilon_{k})\bigg{)}\cap C_{k,j}(\theta_{k}+\epsilon_{k})\Bigg{]}\subseteq \bigcup_{j=1}^{J_{k}}\Bigg{[}\bigg{(}\bigcap_{i=1}^{j-1}B_{k,i}(\theta_{k}) \bigg{)}\cap C_{k,j}(\theta_{k})\Bigg{]}.\] Take any \[w=(Z_{k,1},\ldots,Z_{k,J})\in\bigcup_{j=1}^{J_{k}}\Bigg{[}\bigg{(}\bigcap_{i= 1}^{j-1}B_{k,i}(\theta_{k}+\epsilon_{k})\bigg{)}\cap C_{k,j}(\theta_{k}+ \epsilon_{k})\Bigg{]}.\] For some \(q\in\{1,\ldots,J_{k}\}\), for which \(Z_{k,q}\in C_{k,q}(\theta_{k}+\epsilon_{k})\) and \(Z_{k,j}\in B_{k,j}(\theta_{k}+\epsilon_{k})\) for \(j=1,\ldots,q-1\). \(Z_{k,q}\in C_{k,q}(\theta_{k}+\epsilon_{k})\) implies that \(Z_{k,q}\in C_{k,q}(\theta_{k})\). Furthermore \(Z_{k,q}\in B_{k,q}(\theta_{k}+\epsilon_{k})\) implies that \(Z_{k,q}\in B_{k,q}(\theta_{k})\cup C_{k,q}(\theta_{k})\) for some \(j=1,\ldots,q-1\). Therefore, \[w\in\bigcup_{j=1}^{J_{k}}\Bigg{[}\bigg{(}\bigcap_{i=1}^{j-1}B_{k,i}(\theta_{k} )\bigg{)}\cap C_{k,j}(\theta_{k})\Bigg{]}.\] Next suppose for any \(m_{1},\ldots,m_{K}\) where \(m_{1}\in\{1,\ldots,K\}\) and \(m_{k}\in\{1,\ldots,K\}\backslash\{m_{1},\ldots,m_{k-1}\}\) with \(\theta_{m_{1}},\ldots,\theta_{m_{l}}\geq\theta^{\prime}\) and \(\theta_{m_{l+1}},\ldots,\theta_{m_{K}}<\theta^{\prime}\). Let \(\Theta_{l}=(\theta_{m_{1}},\ldots,\theta_{m_{l}})\). Then \[P(\text{reject all }H_{0k}\text{ for which }\theta_{k}\geq\theta^{ \prime}|\Theta) =P(\bar{W}_{l}(\Theta_{l}))\] \[\geq P(\bar{W}_{l}(\Theta^{\prime}))\] \[\geq P(\bar{W}_{k}(\Theta^{\prime}))\] \[=P(\text{reject all }H_{0k}\text{ for which }\theta_{k}\geq\theta^{ \prime}|H_{PG}).\] where \(\Theta^{\prime}=(\theta^{\prime},\ldots,\theta^{\prime})\). It follows from Theorem 2.1 that the conjunctive power (\(P_{C}\)) is minimised when all treatments have the smallest interesting treatment effect. In order to ensure the conjunctive power is greater than level \(1-\beta\) we rearrange the events \(B_{k,j}(\theta_{k})\) and \(C_{k,i}(\theta_{k})\) to find \[P_{C}=P(\bar{W}_{l}(\Theta^{\prime}))=\sum_{\begin{subarray}{c}j_{k}=1\\ k=1,2,\ldots,K\end{subarray}}^{J_{k}}\Phi(\mathbf{L}_{\mathbf{jk}}^{+}(\Theta ^{\prime}),\mathbf{U}_{\mathbf{jk}}^{+}(\Theta^{\prime}),\Sigma_{\mathbf{jk}}), \tag{2.7}\] where \(\mathbf{U}_{\mathbf{jk}}^{+}(\Theta^{\prime})=(U_{1,j_{1}}^{+}(\theta^{\prime} ),\ldots,U_{K,j_{K}}^{+}(\theta^{\prime}))\) and \(\mathbf{L}_{\mathbf{jk}}^{+}(\Theta^{\prime})=(L_{1,j_{1}}^{+}(\theta^{\prime} ),\ldots,L_{K,j_{K}}^{+}(\theta^{\prime}))\) with \(U_{k,j_{k}}^{+}(\theta_{k})\) and \(L_{k,j_{k}}^{+}(\theta_{k})\) defined in Equation (2.6) and Equation (2.5), respectively. The correlation matrix \(\Sigma_{\mathbf{jk}}\) is the same as that given for FWER in Equation (2.3). When one has equal number of stages and equal allocation to find the sample size one needs to increase \(n\) until \(P_{C}=1-\beta\). If one is in the case of fixed \(\mathbf{n}(\mathbf{k})\) then one can use Algorithm 1, now replacing pairwise power for conjunctive power. ### Sample size distribution and Expected sample size The determination of sample size distribution and expected sample size involves calculating the probability for each outcome of the trial, denoted as \(Q_{\mathbf{j_{k}},\mathbf{q_{k}}}\). Here, \(\mathbf{q_{k}}=(q_{1},\ldots,q_{K})\) is defined, where \(q_{k}=0\) indicates that treatment \(k\) falls below the lower stopping boundary at point \(j_{k}\), and \(q_{k}=1\) indicates that treatment \(k\) exceeds the upper stopping boundary at point \(j_{k}\). We find \[Q_{\mathbf{j_{k}},\mathbf{q_{k}}}= \Phi(\tilde{\mathbf{L}}_{\mathbf{j_{k}},\mathbf{q_{k}}}(\Theta), \tilde{\mathbf{U}}_{\mathbf{j_{k}},\mathbf{q_{k}}}(\Theta),\Sigma_{\mathbf{j_ {k}}}),\] with \(\mathbf{j_{k}}\) one can define \(\tilde{\mathbf{L}}_{\mathbf{j_{k}},\mathbf{q_{k}}}(\Theta)=(\tilde{L}_{1,j_{1 },q_{1}}(\theta_{1}),\ldots,\tilde{L}_{K,j_{K},q_{K}}(\theta_{K}))\) and \(\tilde{\mathbf{U}}(\Theta)_{\mathbf{j_{k}},\mathbf{q_{k}}}=(\tilde{U}_{1,j_{1 },q_{1}}(\theta_{1}),\)\(\ldots,\tilde{U}_{K,j_{K},q_{K}}(\theta_{K}))\) where \[\tilde{L}_{k,j,q_{k}}(\theta_{k})= (l_{k,1}-\frac{\theta_{k}}{\sqrt{I_{k,1}}},\ldots,l_{k,j-1}- \frac{\theta_{k}}{\sqrt{I_{k,j-1}}},[\mathbb{1}\left(q_{k}=0\right)(-\infty)+ u_{k,j_{k}}]-\frac{\theta_{k}}{\sqrt{I_{k,j}}}),\] \[\tilde{U}_{k,j,q_{k}}(\theta_{k})= (u_{k,1}-\frac{\theta_{k}}{\sqrt{I_{k,1}}},\ldots,u_{k,j-1}- \frac{\theta_{k}}{\sqrt{I_{k,j-1}}},[\mathbb{1}\left(q_{k}=1\right)(\infty)+l_ {k,j_{k}}]-\frac{\theta_{k}}{\sqrt{I_{k,j}}}),\] respectively. The \(P_{\mathbf{j_{k}},\mathbf{q_{k}}}\) are associated with their given total sample size \(N_{\mathbf{j_{k}},\mathbf{q_{k}}}\) for that given \(\mathbf{j_{k}}\) and \(\mathbf{q_{k}}\). \[N_{\mathbf{j_{k}},\mathbf{q_{k}}}= \bigg{(}\sum_{k=1}^{K}n_{k,j_{k}}\bigg{)}+\max_{k\in 1,\ldots K}(n_{0,k,j_{k}}),\] This shows that the control treatment continues being recruited to until, at the earliest, the last active treatment to be added has had at least one analysis. To obtain the sample size distribution, as similarly done in Greenstreet et al. (2021), we group all the values of \(\mathbf{j_{k}}\) and \(\mathbf{q_{k}}\) that gives the same value of \(N_{\mathbf{j_{k}},\mathbf{q_{k}}}\) with its corresponding \(Q_{\mathbf{j_{k}}.\mathbf{q_{k}}}\). This set of \(Q_{\mathbf{j_{k}}.\mathbf{q_{k}}}\) is then summed together to give the probability of the realisation of this sample size. To calculate the sample size distribution for each active arm, group \(n_{k,j_{k}}\) with its corresponding \(Q_{\mathbf{j_{k}},\mathbf{q_{k}}}\) and this can similarly be done for the control treatment. The expected sample size for a given \(\Theta\), denoted as \(E(N|\Theta)\), is obtained by summing all possible combinations of \(\mathbf{j_{k}}\) and \(\mathbf{q_{k}}\), \[E(N|\Theta)=\sum_{\begin{subarray}{c}j_{k}=1\\ k=1,2,\ldots,K\end{subarray}}^{J_{k}}\sum_{\begin{subarray}{c}q_{k}\in\{1, \infty\}\\ k=1,2,\ldots,K\end{subarray}}Q_{\mathbf{j_{k}}.\mathbf{q_{k}}}N_{\mathbf{j_{k}}.\mathbf{q_{k}}}. \tag{2.8}\] The expected sample size for multiple different treatment effects (\(\Theta=\{\theta_{1},\ldots,\theta_{K}\}\)) can then be found using Equation (2.8). ## 3 Motivating trial example ### Setting One example of a platform trial is FLAIR, which focused on chronic lymphocyte leukemia Howard et al. (2021). FLAIR initially planned to incorporate an additional active treatment arm and conduct an interim analysis midway through the intended sample size for each treatment. During the actual trial, two extra arms were introduced, including an additional control arm. The original trial design primarily addressed the pairwise type I error due to the inclusion of both additional experimental and control arms. Following Greenstreet et al. (2021), a hypothetical trial that mirrors some aspects of FLAIR will be studied. In this hypothetical trial the familywise error rate (FWER) in the strong sense will be controlled. Controlling the FWER may be seen as crucial in this scenario, as the trial aims to assess various combinations of treatments involving a common compound for all active treatments (Wason et al., 2014). There is an initial active treatment arm, a control arm, and a planned addition of one more active treatment arm during the trial. We apply the proposed methodology to ensure FWER control and consider the conjunctive power and pairwise power. The pairwise power is the main focus of the simulation study rather than the disjunctive power, as a potential drawback of disjunctive power is it is highly dependent on the treatment effect of all the treatments in the study, even the ones without a clinically relevant effect. For example assume one treatment has a clinically relevant effect and the rest have effect equal to the control treatment, then the disjunctive power will keep increasing the more treatments that are added, if one keeps the same bounds, even though the probability of taking the correct treatment forward does not increase. Equally the minimum the disjunctive power can be is equal to the pairwise power. This is when only one treatment has a clinically relevant effect and the rest have an extreme negative effect. A further advantage of the pairwise power is it gives the probability of the treatment with the greatest treatment effect being found, assuming that this treatment has effect equal to the clinically relevant effect. Considering the planned effect size from FLAIR, we assume an interesting treatment difference of \(\theta^{\prime}=-\log(0.69)\) and a standard deviation of \(\sigma=1\). It should be noted that while FLAIR used a time-to-event endpoint with 0.69 representing the clinically relevant hazard ratio between the experimental and control groups, our hypothetical trial will focus on continuous endpoints using a normal approximation of time-to-event endpoints as discussed in Jaki and Magirr (2013). The desired power is 80%. We will maintain the same power level as FLAIR while targeting a one-sided FWER of 2.5%. The active treatment arms interim analysis will be conducted midway through its recruitment and 1:1 allocation will be used between the control and the active treatments as done in FLAIR (Hillmen et al., 2023). The difference between a design which controls the pairwise power and the conjunctive power will be studied in Section 3.2. Additionally, for both pairwise power and the conjunctive power, the number of patients per arm per stage, the maximum sample size, the expected sample size and the disjunctive power will be studied. In Section 3.3 the effect of different numbers of patients recruited to the control before the second treatment is added (\(n(2)\)) will be studied with the focus being on expected sample size and maximum sample size of the trial. The designs will be compared to running two completely separate independent trials for each of the 2 active treatments. When running two trials there would be less expectation to control the FWER across the two trials. Therefore along with the fair comparison of type I error control of 2.5% across the multiple separate studies, the setting of having pairwise error rate being controlled for each at 2.5% will be shown. In Section 3.4 the effect of using a more liberal FWER control compared to type I error control for the separate trials is studied for trials with 3 and 4 active arms. ### Comparing the two types of power We will consider the effect of adding the second treatment halfway through recruitment of the first active treatment, both for ensuring pairwise power and conjunctive power are at 80%. The binding triangular stopping boundaries will be used (Whitehead, 1997; Wason and Jaki, 2012; Li et al., 2020). The stopping boundaries are the same regardless of if one is controlling pairwise power or conjunctive power as \(r(2)=r_{1,1}\) for both. The stopping boundaries are given in Table 1 and are equal for both designs. In Table 1 the sample size when ensuring that the pairwise power is greater than 80% is given. Both active treatments will have up to 152 patients recruited to them and the control treatment can have up to 228 patients. This is due to 76 patients already being recruited to the control before the second treatment is added. The maximum sample size for the pairwise power design is therefore \(\max(N)=152+152+228=532\). Additionally in Table 1 the sample size when ensuring that the conjunctive power is greater than 80% is given. The maximum sample size now is \(\max(N)=192+192+288=672\). The calculations were carried out using R (R Core Team, 2021) with the method given here having the multivariate normal probabilities being calculated using the packages mvtnorm(Genz et al., 2021) and gtools(Warnes et al., 2021). Code is available at [https://github.com/pgreenstreet/Platform_trial_multiple_superior](https://github.com/pgreenstreet/Platform_trial_multiple_superior). Based on the two designs in Table 1, in Table 2 the conjunctive power, pairwise power and disjunctive power for different values of \(\theta_{1}\) and \(\theta_{2}\) are given along with the expected sample size. The values of \(\theta_{1}\) and \(\theta_{2}\) are chosen to study the effects under the global null, when treatments have a clinically relevant effect and when one of the active treatments performs considerably worst than the rest. Table 2 shows when \(\theta_{1}\) and \(\theta_{2}\) equals the clinically relevant effect \(\theta^{\prime}\) under the design for pairwise power, that the pairwise power of both treatments is 80.0%; the conjunctive power is 66.0%; the disjunctive power is 94.1%; and the expected sample size is 420.6. This highlights the fact that when controlling the pairwise power that if both treatments have a clinically relevant effect there is a large chance (44%) that one may miss at least one of the two treatments. When studying the design in which conjunctive power is controlled one can now see that the pairwise power and disjunctive power is much greater compared to the pairwise power design. This comes with a large increase in both expected and maximum sample size, for example the maximum sample size has increased by 140 patients. \begin{table} \begin{tabular}{c|c|c|c|c|c|c|c|c|c} Design & \multicolumn{2}{c|}{\(U=\begin{pmatrix}U_{1}\\ U_{2}\end{pmatrix}\)} & \multicolumn{2}{c|}{\(L=\begin{pmatrix}L_{1}\\ L_{2}\end{pmatrix}\)} & \multicolumn{2}{c|}{\(\begin{pmatrix}n_{1,1}&n_{1,2}\\ n_{2,1}&n_{2,2}\end{pmatrix}\)} & \multicolumn{2}{c|}{\(\begin{pmatrix}n_{0,1,1}&n_{0,1,2}\\ n_{0,2,1}&n_{0,2,2}\end{pmatrix}\)} & \multicolumn{2}{c|}{\(\begin{pmatrix}n(1)\\ n(2)\end{pmatrix}\)} & \multicolumn{2}{c|}{\(\begin{pmatrix}n(1)\\ n(2)\end{pmatrix}\)} & \multicolumn{2}{c|}{\(\begin{pmatrix}\max(N)\)} \\ \hline Pairwise power & \(\begin{pmatrix}2.501&2.358\\ 2.501&2.358\\ 2.501&2.358\\ 2.501&2.358\\ \end{pmatrix}\) & \(\begin{pmatrix}0.834&2.358\\ 0.834&2.358\\ 0.834&2.358\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 76&152\\ 96&192\\ 96&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 152&228\\ 96&192\\ 192&288\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 76\\ 76&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 152&228\\ 96&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 76&16\\ 96&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 76&192\\ 96&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&152\\ 76&16\\ 96&192\\ \end{pmatrix}\) & \(\begin{pmatrix}76&192\\ 96&192\\ 192&288\\ \end{pmatrix}\) & \(\begin{pmatrix}0\\ 76\\ 96\\ \end{pmatrix}\) & \(\begin{pmatrix}0\\ 76\\ 96\\ \end{pmatrix}\) & \(\begin{pmatrix}532\\ 96\\ \end{pmatrix}\) \\ \end{tabular} \end{table} Table 1: The stopping boundaries and sample size of the proposed designs, for both control of pairwise power and of conjunctive power. As seen for the design for conjunctive power section of Table 2 the disjunctive power when treatment 1 and 2 have effect \(\theta^{\prime},0\), respectively, does not equal the disjunctive power of treatment 1 and 2 when the effect is \(0,\theta^{\prime}\). This is because the outcome of treatment 1's test statistic has a larger effect on treatment 2 then the other-way round. For example treatment 1 first stage is always independent of treatment 2. However for treatment 2 it's first stage is only independent of treatment 1 if treatment 1 stops after its first stage. Therefore \(\Sigma_{(1,2)}\neq\Sigma_{(2,1)}\). However as can be seen this difference in the cases studied is very small. In Table 2 shows when there is only one treatment with a clinically relevant effect the conjunctive power equals the pairwise power of that treatment. When neither treatment has a clinically relevant effect the conjunctive power equals 100%, as there are no treatments with a clinically relevant effect that need to be found. As a result the trial has already resulted in all the clinically relevant treatments being declare i.e 0 treatments. The expected sample size is greatly dependent on which treatment has the clinically relevant effect and which does not. For example when studying the the design with pairwise power control the expected sample size when the treatment effect is \(\theta^{\prime},0\), is 372.7. This is compared to 396.6 when the treatment effect is \(0,\theta^{\prime}\) for treatment 1 and 2 respectively. This difference is because the probability of treatment \(k\) stopping after the first stage is higher when \(\theta_{k}=0\) compared to \(\theta_{k}=\theta^{\prime}\). Therefore when the second treatment has effect 0 it is more likely that the trial will stop after the second stage of the trial. This reduces the amount of patients on average being recruited to the control treatment. In Table 2 it can be seen that the pairwise power for the treatment with a clinically relevant effect is equal to the disjunctive power when the other treatment has an extremely negative treatment effect compared to the control. This is as there is no longer a chance that the other treatment can be taken forward. Therefore \(\theta_{1}=-\infty\)\(\theta_{2}=\theta^{\prime}\) or \(\theta_{1}=\theta^{\prime}\)\(\theta_{2}=-\infty\), is the point when the pairwise, disjunctive and conjunctive power are all equal. When one treatment has effect \(\theta^{\prime}\) and the other has effect equal to the control the disjunctive power is greater than the pairwise power, as there is still a chance that the other treatment may be taken forward. In Table 2 it is shown that when both treatments have effect 0 the disjunctive power is equal to the FWER for the trial. In addition when a treatment has effect 0 this results in the pairwise power for that treatment equalling the PWER. In the Supporting Information Section 3 results for using both O'Brien and Fleming (O'Brien and Fleming, 1979) and Pocock boundaries (Pocock, 1977) are shown, with the futility boundary equal to 0 (Magirr et al., 2012). Additionally the results for using non binding triangular stopping boundaries are shown in the Supporting Information Section 4. Overall Table 1 and Table 2 have shown that the choice of type of power to control may be highly dependent on the sample size available, as if the design ensures conjunctive power of level \(1-\beta\) it will ensures pairwise power of at least \(1-\beta\) but the opposite does not hold. However the sample size for a trial designed for pairwise power will be less than that of a design for conjunctive power. ### Comparison with running separate trials This section studies the effect on maximum and expected sample size depending on when the additional treatment arm is added to the platform trial. The examples for both conjunctive power and pairwise power are compared to running two separate trials. There are two settings for separate trials which are considered. Setting 1 is when the type I error across both the trials is set to be \(2.5\%\), therefore, the type I error for each is \(1-\sqrt{1-0.025}=1.26\%\). For Setting 2 the type I error of each trial is controlled at \(2.5\%\). For the separate trials which are compared to the pairwise power, the power level for each is set to \(80\%\). This results in the following sample size and stopping boundaries for the two trials for Setting 1, \[U_{1}=\begin{pmatrix}2.508&2.364\end{pmatrix},\quad L_{1}=\begin{pmatrix}0.836&2. 364\end{pmatrix}.\quad\begin{pmatrix}n_{1,1}&n_{1,2}\end{pmatrix}=\begin{pmatrix}7 7&154\end{pmatrix}.\] with \(n_{0,1,1}=n_{1,1}\), \(n_{0,1,2}=n_{1,2}\) and \(n(1)=0\). Setting 2 gives: \[U_{1}=\begin{pmatrix}2.222&2.095\end{pmatrix},\quad L_{1}=\begin{pmatrix}0.741&2.095\end{pmatrix}.\quad\begin{pmatrix}n_{1,1}&n_{1,2}\end{pmatrix}=\begin{pmatrix}6 5&130\end{pmatrix}.\] with \(n_{0,1,1}=n_{1,1}\), \(n_{0,1,2}=n_{1,2}\) and \(n(1)=0\). For comparison with the conjunctive power designs the probability of finding both treatments across the two trials is set to \(80\%\). The required power for each trial is therefore \(\sqrt{1-\beta}=0.894\). The boundaries remain the same for both settings as the type I error remains the same. The new sample size for Setting 1 is \(\begin{pmatrix}n_{1,1}&n_{1,2}\end{pmatrix}=\begin{pmatrix}98&196\end{pmatrix}\) and for Setting 2 is \(\begin{pmatrix}n_{1,1}&n_{1,2}\end{pmatrix}=\begin{pmatrix}85&170\end{pmatrix}\). Figure 1 gives the maximum sample size and the expected sample size under different \(\theta_{1},\theta_{2}\) depending on when the second treatment is added, for the pairwise power control of \(80\%\). Figure 2 gives similar results however the focus now is on control of the conjunctive power at \(80\%\). As indicated in Figure 1, when controlling the pairwise power, if the second active treatment is introduced at the beginning of the trial, the total sample size required is \(456\), whereas if it is added at the end of recruitment for treatment \(1\), the total sample size becomes \(616\). This increase in sample size is attributable to two factors. Firstly, there is a necessity to increase \begin{table} \begin{tabular}{c|c|c|c|c|c|c} \multicolumn{8}{c}{**Design for pairwise power**} \\ \hline Treatment effect & \multicolumn{2}{c|}{Pairwise power} & \multicolumn{2}{c|}{Conjunctive power} & \multicolumn{2}{c|}{Disjunctive power} & \multicolumn{2}{c}{Expected sample size} \\ \(\theta_{1}\) & \(\theta_{2}\) & \(P_{PW,1}\) & \(P_{PW,2}\) & \(P_{C}\) & \(P_{D}\) & \(E(N|\theta_{1},\theta_{2})\) \\ \hline \(\theta^{\prime}\) & \(\theta^{\prime}\) & 0.800 & 0.800 & 0.660 & 0.941 & 420.6 \\ \(\theta^{\prime}\) & 0 & 0.800 & 0.013 & 0.800 & 0.802 & 372.7 \\ \(\theta^{\prime}\) & \(-\infty\) & 0.800 & 0 & 0.800 & 0.800 & 342.9 \\ \(0\) & \(\theta^{\prime}\) & 0.013 & 0.800 & 0.800 & 0.802 & 396.6 \\ \(0\) & 0 & 0.013 & 0.013 & 1 & 0.025 & 348.7 \\ \(-\infty\) & \(\theta^{\prime}\) & 0 & 0.800 & 0.800 & 0.800 & 381.7 \\ \hline \multicolumn{8}{c}{**Design for conjunctive power**} \\ \hline Treatment effect & \multicolumn{2}{c|}{Pairwise power} & \multicolumn{2}{c|}{Conjunctive power} & \multicolumn{2}{c|}{Disjunctive power} & \multicolumn{2}{c}{Expected sample size} \\ \(\theta_{1}\) & \(\theta_{2}\) & \(P_{PW,1}\) & \(P_{PW,2}\) & \(P_{C}\) & \(P_{D}\) & \(E(N|\theta_{1},\theta_{2})\) \\ \hline \(\theta^{\prime}\) & \(\theta^{\prime}\) & 0.890 & 0.890 & 0.801 & 0.979 & 508.1 \\ \(\theta^{\prime}\) & 0 & 0.890 & 0.013 & 0.890 & 0.890 & 463.0 \\ \(\theta^{\prime}\) & \(-\infty\) & 0.890 & 0 & 0.890 & 0.890 & 425.4 \\ \(0\) & \(\theta^{\prime}\) & 0.013 & 0.890 & 0.890 & 0.891 & 485.6 \\ \(0\) & 0 & 0.013 & 0.013 & 1 & 0.025 & 440.5 \\ \(-\infty\) & \(\theta^{\prime}\) & 0 & 0.890 & 0.890 & 0.890 & 466.7 \\ \end{tabular} \end{table} Table 2: Operating characteristics of the proposed designs under different values of \(\theta_{1}\) and \(\theta_{2}\), for both control of pairwise power and of conjunctive power. the number of patients recruited to the control group until treatment 2 has completed the trial. Secondly, the decrease in correlation between the two treatments results in an enlargement of the boundaries to maintain control over the family-wise error rate. It is this secondary factor which causes the small jumps in maximum sample size seen in Figures 1 and 2. In Figure 1 when comparing the platform designs with pairwise power control, to running two separate trials it can be seen that, for the case that the pairwise error for each trial is \(2.5\%\), once the second treatment is added after \(64\) patients have been recruited to the control (\(n(2)\geq 64\)) the maximum sample size of running the platform design is greater than or equal to that of running two separate trials, which is \(520\) patients. However when controlling the error across both separate trials the maximum sample size is now the same as when adding the second treatment at the end of recruitment for the first treatment in the platform design so \(616\). For Setting 1 it can be seen that the expected sample size for separate trials can be better than that of the platform design. In the case of \(\theta_{1}=-\infty\) and \(\theta_{2}=-\theta^{\prime}\) then once \(n(2)\geq 81\) the expected sample size of running the platform design is greater than that of running two separate trials. This is because in the platform approach the control cannot stop until each treatment has finished testing, whereas in the separate trial case each control group will stop as soon as either treatment is dropped. For Setting 1 there are some cases studied which cannot be seen in Figure 1. These are \(\theta_{1}=\theta^{\prime}\), \(\theta_{2}=\theta^{\prime}\) and if \(\theta_{1}=\theta^{\prime}\), \(\theta_{2}=0\) as both these are at the point \(n(2)\geq 117\) which matches that of \(\theta_{1}=\theta^{\prime}\), \(\theta_{2}=-\infty\). When studying the expected sample size of the Setting 2 compared to the platform designs it can be seen that if \(\theta_{1}=-\infty\) and \(\theta_{2}=-\theta^{\prime}\) then once \(n(2)\geq 15\) the expected sample size of running the platform design is greater than that of running two separate trials. The expected sample size for two separate trials when \(\theta_{1}=-\infty\) and \(\theta_{2}=\theta^{\prime}\) is \(319.5\). When controlling the conjunctive power, as in Figure 2, if the second active treatment is introduced at the beginning of the trial, the total sample size required is \(558\), whereas if it is added at the end of recruitment for treatment \(1\), the total sample size becomes \(784\). Once again the maximum sample size for Setting 1 equals that of when treatment \(2\) is added after treatment \(1\) finished recruitment so \(784\) patients. In Figure 2, when \(n(2)\geq 104\) the maximum sample size of running the platform design is greater than or equal to that of running two separate trials under Setting 2, which is \(680\) patients. Similar as seen in Figure 1 there is some lines which overlap for Setting 1 in Figure 2 as \(n(2)=143\) is the point for both \(\theta_{1}=\theta^{\prime}\), \(\theta_{2}=\theta^{\prime}\) and \(\theta_{1}=\theta^{\prime}\), \(\theta_{2}=-\infty\), also \(n(2)=121\) is the point for both \(\theta_{1}=0\), \(\theta_{2}=\theta^{\prime}\) and \(\theta_{1}=0\), \(\theta_{2}=0\). When \(n(2)\geq 104\) for Setting 1, and \(n(2)\geq 39\) for Setting 2, the expected sample size of running the platform design is greater than that of running two separate trials when \(\theta_{1}=-\infty\) and \(\theta_{2}=\theta^{\prime}\). The expected sample size for running two separate trials when \(\theta_{1}=-\infty\) and \(\theta_{2}=\theta^{\prime}\) is \(475.3\) and \(403.8\) for Setting 1 and Setting 2 respectively. Overall Figures 1 and 2 have shown there maybe times that there is no benefit to running a platform trial with regards to sample size, depending on when the later treatment is added to the trial. This issue is further emphasised when there is not the expectation to control the type I error across all the individual trials as seen in Setting 2. Figure 1: Both panels give the maximum sample size and the expected sample size under different \(\theta_{1},\theta_{2}\) depending on the value \(n(2)\), for the pairwise power control of 80%. Left panel: dash vertical lines correspond to the points where the maximum/expected sample size of the trial is now greater than running two separate trials with type I error control across both trials set to 2.5%. Right panel: dash vertical lines correspond to the points where the maximum/expected sample size of the trial is now greater than running two separate trials with type I error control for each trial set to 2.5%. Figure 2: The maximum sample size and the expected sample size under different \(\theta_{1},\theta_{2}\) depending on the value \(n(2)\), for the conjunctive power control of 80%. Left panel: dash vertical lines correspond to the points where the maximum/expected sample size of the trial is now greater than running two separate trials under Setting 1. Right panel: dash vertical lines correspond to the points where the maximum/expected sample size of the trial is now greater than running two separate trials under Setting 2. ### Comparison with running separate trials under different controls of type I error When designing a multi-arm trial one may find that the expected control of the FWER is less than that of the type I error control for an individual trial, as seen in the TAILoR trial for example (Pushpakom et al., 2015, 2020). Therefore in Table 3 we consider the effect of allowing FWER control of 5% one sided compared to 2.5% type I error for the individual trials. In this table the same design parameters where used as above, however, now the number of active arms has increased in the hypothetical trial to 3 or 4, and the number of stages is now either 1,2 or 3. In Table 3 the focus is on controlling the power at the desired 80% level with the pairwise power being the focus for the top half and conjunctive power for the bottom half. When controlling the conjunctive power the power for each separate trial is \((1-\beta)^{1/k}\). In these hypothetical trials it is assumed that each one of the arms is added sequentially, with an equal gap between each one. Therefore in the 3 active arm case if the second arm is added after 20 patients have been recruited to the control then the third arm will be added after a total of 40 patients have been recruited to the control. In Table 3 the first 2 columns give the number of active arms and stages for the platform trial, respectively. The third and forth columns then gives the sample size per stage and the maximum sample size of the individual trials, respectively. This has been chosen as this number will remain constant throughout, as it is unaffected by the timing of when the next arm is ready due to each trial being completely separate from the others. The remaining columns give when there is no benefit with regards to the maximum and expected sample size of conducting a platform trial compared to running separate trials, with respect \(n(k)-n(k-1)\). The value of \(n(k)-n(k-1)=n(2)\) as the first treatment is added at the beginning of the trial. In the Supporting Information Section 5 the plots for the 2 stage and 3 stage example trials as given in Table 3 are shown. Using Table 3, for the 3 active arm, 2 stage example each separate trial has \(n_{1,1}=65\) and \(n_{1,2}=130\). The total maximum sample size of running these 3 separate trials is therefore 780. Once the second treatment is planned to be added after 105 patients recruited to the control, (therefore 210 recruited to the control before treatment 3), there is no benefit in using the platform design with respect to maximum sample size. For the expected sample size four different configurations of the treatment effects are studied. The first (\(\Theta_{1}\)) assumes all the treatments have the clinically relevant effect, so \(\theta_{k}=\theta^{\prime}\) for \(k=1,\ldots,K\). The second (\(\Theta_{2}\)) assumes only the first treatment has a clinically relevant effect and the rest have effect equal to that of the control treatment, so \(\theta_{1}=\theta^{\prime}\), \(\theta_{k}=0\) for \(k=2,\ldots,K\). The third (\(\Theta_{3}\)) assumes only the last treatment has a clinically relevant effect and the rest equal the control, so \(\theta_{K}=\theta^{\prime}\), \(\theta_{k}=0\) for \(k=1,\ldots,K-1\). The forth configuration (\(\Theta_{4}\)) assumes all the treatments have effect equal to that of the control treatment, so the global null, so \(\theta_{k}=0\) for \(k=1,\ldots,K\). For the expected sample size for the 4 treatment effect configurations studied here there is no benefit in using a platform trial after potentially just 62 patients if \(\Theta_{3}\) is true, this does rise to 73 if \(\Theta_{1}\) is true, if the focus is on sample size. Table 3 shows that the maximum sample size of running separate trials increases with increase in number of stages or arms. This is also the case when running the proposed platform trial design. As can be seen with respect to maximum sample size the more stages the trial has the later a treatment can be added before the maximum sample size becomes worst than running separate trials. For example, when pairwise power is controlled, a 1 stage 3 arm trial with regards to maximum sample size one should use separate trials after 90 patients this is compared to 114 patients for a 3 arm 3 stage trial. If the focus is on the expected sample size, than for the examples studied here, increasing the number of stages results in a decrease time before one would switch to separate trials. For example when controlling the conjunctive power, for the 4 arm trial, it can be seen that the expected sample size under the global null for running separate trials becomes less than that of running the platform trials when \(n(2)=140\) for 1 stage case compared to \(n(2)=99\) for the 3 stage version. This is because the ability to have interim analyses saves more patients for separate trials with respect to expected sample size. This is because in separate trials when a treatment is stopped earlier either for futility or superiority the control treatment also stops. Therefore in this 4 arm example there are 4 sets of control treatments which can stop early compared to only 1 set for the platform design. Additionally for the platform trial the control can only stop once all the active treatments have stopped. This is why the expected sample size under \(\Theta_{2}\) is less then that of \(\Theta_{3}\), as if the final treatment has a clinically relevant effect then it will on average have more stages than a treatment with effect equal to that of the control for the configuration studied here. This section has therefore shown that there are periods in which using a platform trial can be beneficial with regards to sample size if one can use a more liberal type I error control compared to that used for individual trials. However this has also shown that if treatments are added late into the trial there may not be benefit, so highlighting the importance of considering which trial design should be use. ## 4 Discussion This paper has built on the work of Greenstreet et al. (2021) to show how one can control the FWER for a trial in which the treatments can be preplanned to be added at any point. This work has then studied the different approaches for powering the trial in which the trial will continue even if a superior treatment is found. This paper shows how the expected sample size and sample size distribution can be found. Finally a hypothetical trial, motivated by FLAIR (Howard et al., 2021) is discussed. This section evaluates the pairwise and conjunctive power when the second active treatment is added halfway through recruitment for the first active treatment. We investigate the operating characteristics for multiple values of \(\theta_{1}\) and \(\theta_{2}\). Then the section goes onto study the effect of adding the later treatments at different points in the platform design and compares these trial designs to running separate trials. The designs flexibility to incorporate the addition of treatments at any point during a trial allows for the creation of multiple designs that depend on when the treatments are introduced. This approach works effectively until the completion of the initial stage for the treatment that initiated the trial. Up to this point, the treatments can be added when it becomes available, and the boundaries can be set accordingly. However, if the treatments are not ready until after the first analysis, two options can be pursued to avoid bias resulting from knowledge of the first stage results. Firstly, one can choose not to plan for the addition of the treatments and conduct separate trials. As demonstrated in Section 3, this approach may require fewer patients \begin{table} \begin{tabular}{c|c|c|c|c|c|c|c|c} \multicolumn{8}{c}{**Design for pairwise power**} \\ \hline Active arms & Stages & \multicolumn{3}{c|}{Separate trial} & \multicolumn{2}{c|}{\(\min_{n(2)}(\max(N_{s})\)} & \multicolumn{2}{c}{\(\min_{n(2)}(E(N_{s}|\Theta)\leq E(N|\Theta))\)} \\ \(K\) & \(J\) & \((n_{1,1},\ldots,n_{1,J})\) & \(\max(N_{s})\) & \(\leq\max(N))\) & \(\Theta_{1}\) & \(\Theta_{2}\) & \(\Theta_{3}\) & \(\Theta_{4}\) \\ \hline 3 & 1 & 115 & 690 & 90 & 90 & 90 & 90 & 90 \\ 3 & 2 & (65, 130) & 780 & 105 & 73 & 72 & 62 & 66 \\ 3 & 3 & (46, 92, 138) & 828 & 114 & 68 & 67 & 55 & 60 \\ 4 & 1 & 115 & 920 & 79 & 79 & 79 & 79 & 79 \\ 4 & 2 & (65, 130) & 1040 & 94 & 61 & 62 & 54 & 59 \\ 4 & 3 & (46, 92, 138) & 1104 & 103 & 59 & 58 & 49 & 55 \\ \hline \multicolumn{8}{c}{**Design for conjunctive power**} \\ \hline Active arms & Stages & \multicolumn{3}{c|}{Separate trial} & \multicolumn{2}{c|}{\(\min_{n(2)}(\max(N_{s})\)} & \multicolumn{2}{c}{\(\min_{n(2)}(E(N_{s}|\Theta)\leq E(N|\Theta))\)} \\ \(K\) & \(J\) & \((n_{1,1},\ldots,n_{1,J})\) & \(\max(N_{s})\) & \(\leq\max(N))\) & \(\Theta_{1}\) & \(\Theta_{2}\) & \(\Theta_{3}\) & \(\Theta_{4}\) \\ \hline 3 & 1 & (171) & 1026 & 143 & 143 & 143 & 143 \\ 3 & 2 & (97, 194) & 1164 & 166 & 107 & 109 & 101 & 106 \\ 3 & 3 & (68, 136, 204) & 1224 & 174 & 98 & 99 & 92 & 98 \\ 4 & 1 & (185) & 1480 & 140 & 140 & 140 & 140 & 140 \\ 4 & 2 & (105, 210) & 1680 & 167 & 102 & 109 & 103 & 109 \\ 4 & 3 & (74, 148, 222) & 1776 & 182 & 93 & 99 & 93 & 99 \\ \end{tabular} Key: \(N_{s}\) is the sample size of running K separate trials, \(\Theta_{1}\): \(\theta_{k}=\theta^{\prime}\) for \(k=1,\ldots,K\); \(\Theta_{2}\): \(\theta_{1}=\theta^{\prime}\), \(\theta_{k}=0\) for \(k=2,\ldots,K\) ; \(\Theta_{3}\): \(\theta_{K}=\theta^{\prime}\), \(\theta_{k}=0\) for \(k=1,\ldots,K-1\); \(\Theta_{4}\): \(\theta_{k}=0\) for \(k=1,\ldots,K\). \end{table} Table 3: The comparison of using the proposed platform design with FWER of 5% one sided against running separate trials with type I error control of each at 2.5% one sided, for different numbers of arms and stages. overall. Alternatively, one can predefine the times at which the treatments will be added and utilize the corresponding bounds. A drawback here is that if the treatments are not ready by the predefined points, they cannot be added. Nevertheless, for the remaining treatments, the control of family-wise error rate will be maintained. Due to the bounds being designed to control FWER across all the hypotheses, therefore, by not adding a treatment and so removing a hypothesis this reduces the maximum value of the FWER. This paper has highlighted a potential issue of increased expected and maximum sample size when requiring strong control of family-wise error rate for a platform trial in which an arm is added later. If one would run two completely separate trials the FWER across the trials would likely not be expected. As a result there is a lot of time where there is no benefit to the platform trial design with regards to maximum or expected sample size as was shown Figure 1 and Figure 2 for Setting 2. This point has been further emphasised in Table 3 which shows that even with a more liberal FWER control compared to the type I error control of each individual trial there are still many points where one may be better of running separate trials with respect to sample size. This work therefore reiterates the importance of the discussions around type I error control in platform trials (Molloy et al., 2022; Wason et al., 2014, 2016; Howard et al., 2018; Proschan and Waclawiw, 2000; Proschan and Follmann, 1995; Nguyen et al., 2023). If one instead wants to control the pairwise error, as done for example in STAMPEDE (Sydes et al., 2009), one can use Equation (2.4), now replacing \(\theta^{\prime}\) with 0. An additional advantage of using the PWER, if controlling the pairwise power, is that the stopping boundaries and the sample size required for each active arm are independent of when the arm is added. Therefore the only change will be how many patients need to be recruited to the control. However one may be find PWER in a platform trial insufficient for error control (Wason et al., 2014; Molloy et al., 2022) and may not meet the regulators requirements. Building upon this research, a study could be conducted to investigate the impact of having different numbers of stages and stopping boundaries while maintaining equal power and type I error for each treatment, utilizing the approach described in Section 2. However, such an investigation would likely require multiple changes in the allocation ratio, resulting in potential issues with time trends. One could therefore examine methods to handle these time trends, as explored in Lee and Wason (2020); Marschner and Schou (2022); Roig et al. (2023); Greenstreet et al. (2021). Furthermore a change in allocation ratio between treatments can result in different PWER and pairwise power for each treatment if using the same boundaries for each treatment therefore one could use an iterative approach such as that discussed in Greenstreet et al. (2021). Equally one could study the effect of using non-concurrent controls, but once again this can face a large issue with time trends. The main issue with these time trends is that they are unknown. However one could look into incorporating approaches to reduce the bias potentially caused (Lee and Wason, 2020; Marschner and Schou, 2022; Wang et al., 2022; Saville et al., 2022). In Section 3.4 it was assumed for the multi-arm trials that each treatment was added after an equal number of control treatments were recruited so \(n(k)-n(k-1)=n(2)\) for \(k=2,\ldots,K\). This may however not be the case. One may therefore wish to consider the effect of having multiple treatments beginning the study and then adding additional treatments later. The methodology presented in Section 2 allows for these changes. However when it comes to the comparison designs there are now multiple options that can be chosen. As done in Section 3.4 one could use separate trials for each comparison, however one could consider using multiple MAMS trials where all treatments begin at once, or a mix of the two. Further points to be considered here is how one can evenly share the power across all these trial types, especially if the focus is on conjunctive power, and also how the type I error should be defined for each comparison trial. Furthermore, this work could be expanded to incorporate adaptive boundaries that adjust once a treatment is deemed effective, as discussed in Urach and Posch (2016) for the case of multi-arm multi-stage (MAMS) trials. However, such an adaptation would result in a less preplanned design so potential further complications in understanding for the clinicians, the trial statisticians and the patients. Additionally, determining the point at which the conjunctive power is at its lowest may no longer be feasible, as dropping each arm would lead to lower bounds for the remaining treatments, thus affecting the conjunctive power assessment. This adaptive approach will likely result in uneven distribution of errors across the treatments added at different points. If one was to then adjust for this one may encounter issues with time trends as the allocation ratio may need to change mid trial. This paper has given a general formulation for designing a preplanned platform trial with a normal continuous endpoint, and using the work of Jaki and Magirr (2013) one could apply this methodology to other endpoint such as time-to-event used in FLAIR (Howard et al., 2021). When using this approach one should be aware of computational issues from calculating high dimensional multivariate normal distributions, if one has a large number of arms and stages in the trial design. If this is an issue then one can restrict to only adding arms at the interims so one can utilise the method of Dunnett (1955) as discussed in Magirr et al. (2012); Greenstreet et al. (2021). ### Acknowledgements This report is independent research supported by the National Institute for Health Research (NIHR300576). The views expressed in this publication are those of the authors and not necessarily those of the NHS, the National Institute for Health Research or the Department of Health and Social Care (DHSC). TJ and PM also received funding from UK Medical Research Council (MC_UU_0002/14 and MC_UU_0002/19, respectively). This paper is based on work completed while PG was part of the EPSRC funded STOR-i centre for doctoral training (EP/S022252/1). For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising. ### Conflict of Interest The authors declare no potential conflict of interests. Alun Bedding is a shareholder of Roche Products Ltd.

2303.07781

Non-Concentration of Primes in $Γ\backslash PSL_2(\mathbb{R})$

This paper generalizes the result of Sarnak and Ubis \cite{sarnak-ubis} about non-concentration of primes in horocycle orbits on $PSL_2(\mathbb{Z}) \backslash PSL_2(\mathbb{R})$ to any lattice in $PSL_2(\mathbb{R})$. The proof combines the asymptotic result of Str\"ombergsson \parencite{strombergsson} and Venkatesh's method \parencite{venkatesh} with the approach of Sarnak and Ubis of approximating horocycle pieces with periodic horocycles. The key step is to establish a dichotomy between $\{\xi h(t), t \in [0, T] \}$ having good equidistribution in $\Gamma \backslash PSL_2(\mathbb{R})$ and it being approximable by closed horocycle pieces with small period. In a follow-up paper, a similar approach will be used to show equidistribution of $\xi h(n^{1+\gamma})$ for small $\gamma>0$, generalizing Venkatesh's result \parencite{venkatesh} to non-compact $\Gamma$.

Lauritz Streck

2023-03-14T10:45:56Z

http://arxiv.org/abs/2303.07781v1

# Non-concentration of primes in \(\Gamma\backslash PSL_{2}(\mathbb{R})\) ###### Abstract. This paper generalizes the result of Sarnak and Ubis [9] about non-concentration of primes in horocycle orbits on \(PSL_{2}(\mathbb{Z})\backslash PSL_{2}(\mathbb{R})\) to any lattice in \(PSL_{2}(\mathbb{R})\). The proof combines the asymptotic result of Strombergsson [11] and Venkatesh's method [12] with the approach of Sarnak and Ubis of approximating horocycle pieces with periodic horocycles. The key step is to establish a dichotomy between \(\{\xi h(t),t\in[0,T]\}\) having good equidistribution in \(\Gamma\backslash PSL_{2}(\mathbb{R})\) and it being approximable by closed horocycle pieces with small period. In a follow-up paper, a similar approach will be used to show equidistribution of \(\xi h(n^{1+\gamma})\) for small \(\gamma>0\), generalizing Venkatesh's result [12] to non-compact \(\Gamma\). ## 1. Introduction _General Introduction._ Let \(G=PSL_{2}(\mathbb{R})\) and \(\mu_{G}\) be the Haar measure on \(G\). Let \(\Gamma\) be a lattice in \(G\), that is, a discrete subgroup such that \(\mu_{X}\), the projection of the Haar measure to \(X=\Gamma\backslash G\), is finite (and assumed to fulfill \(\mu_{X}(X)=1\)). The dynamics of the space \(X\) with respect to \(\mu_{X}\) have been studied extensively in recent years, in part because of the strong connection to Diophantine approximation in the case of \(\Gamma=PSL_{2}(\mathbb{Z})\). The group \(G\) can be parametrized in terms of \[h(x):=\begin{pmatrix}1&x\\ 0&1\end{pmatrix}\quad a(y):=\begin{pmatrix}y^{\frac{1}{2}}&0\\ 0&y^{-\frac{1}{2}}\end{pmatrix}\quad k(\theta):=\begin{pmatrix}\cos\theta& \sin\theta\\ -\sin\theta&\cos\theta\end{pmatrix},\] which induces a natural left-invariant metric \(d_{G}\) on \(G\). This metric descends to \(X\) via \(d_{X}(\Gamma g,\Gamma h)=\inf_{\gamma\in\Gamma}d_{G}(g,\gamma h)\). The geodesic flow \[g_{t}(g):=ga(e^{t})=\begin{pmatrix}ae^{\frac{t}{2}}&be^{-\frac{t}{2}}\\ ce^{\frac{t}{2}}&de^{-\frac{t}{2}}\end{pmatrix}\] and the horocycle flow \[h_{t}(g):=gh(t)=\begin{pmatrix}a&b+at\\ c&d+ct\end{pmatrix}\] act ergodically on \((X,\mu_{X})\). The horocycle orbits were found to exhibit a very rigid behaviour. Furstenberg showed that \(\mu_{X}\) is uniquely ergodic under \(h_{t}\) when \(X\) is compact [3]. For a general \(\Gamma\), there are periodic orbits under \(h_{t}\), but these carry all other invariant measures. Precisely, Dani and Smillie showed that both \(h_{t}(\xi),t\in\mathbb{R}\) and \(h_{n}(\xi),n\in\mathbb{N}\) equidistribute with respect to \(\mu_{X}\) unless \(\xi\) is periodic, in the sense \(t\mapsto h_{t}(\xi)\) periodic [1]. As all periodic orbits are isomorphic to the torus, questions are reduced to questions on tori if the point \(\xi\) is periodic. With these questions settled, questions about the equidistribution of other orbits were raised. Shah conjectured that \(\xi h(n^{\alpha}),n\in\mathbb{N}\) equidistributes with respect to \(\mu_{X}\) for all \(\alpha\geq 1\) and \(\xi\) non-periodic [10]. Margulis conjectured that \(\xi h(p)\) would equidistribute with respect to \(\mu_{X}\), where \(p\) is running over the primes and \(\xi\) is non-periodic [7]. This paper provides partial progress in the latter question by proving that primes do not concentrate anywhere. The way to showing non-concentration is through controlling averages of the form \(\frac{s}{T}\sum_{sn\leq T}f(ph(sn))\) and applying sieve methods. One natural way to do this is to use a smooth approximation of the primes and show equidistribution of this object, which we will do in this paper. We will take the pseudo random measure \(\nu\), which gets introduced below. Hopefully, the reader will find this presentation more coherent and easier to generalize. Using the Selberg sieve instead like Sarnak and Ubis do in [9] would also be possible. Green and Tao used their pseudo random measure \(\nu\) as a smooth approximation of the primes to prove their celebrated theorem of the prime numbers containing arbitrarily long arithmetic progressions [5]. Like in the case of proving properties in function spaces through approximation by smooth functions, introducing \(\nu\) allowed them to split the proof of some properties of the primes into two parts: First, showing that \(\nu\) has certain properties (like the pseudo randomness condition of the \(k\)-linear forms condition in [5]) and second, that one can recover properties of the primes through these properties of \(\nu\) (the relative Szemeredi theorem in the case of arithmetic progressions). We will show in this paper that horocycle orbits along \(\nu\) equidistribute. Goldston and Yildirim defined \[\Lambda_{R}(n):=\sum_{k<R,\ k|n}\mu(k)\log\left(\frac{R}{k}\right),\] modeled after the standard convolution identity for the von Mangoldt-function [4]. Green and Tao [5] defined the pseudo random measure \(\nu\) by \[\nu(n):=\frac{1}{\log R}\Lambda_{R}^{2}(n).\] Due to the particularities of finding arithmetic sequences, they restricted \(\nu\) to a coprime residue class of some big integer \(W\), which is commonly called the \(W\)-trick. In our setting, it will not be necessary. Define furthermore \(\operatorname{dist}(\Gamma g):=d_{X}(\Gamma g,p_{0})\) where \(p_{0}\in X\) could be any point, for example \(p_{0}=\Gamma\) (for the definition of \(d_{X}\), see Section 3). For a function \(f\in C^{4}(X)\) let \(\|f\|_{W^{4}}\) be its Sobolev norm in the Hilbert space \(W^{4,2}\) involving the fourth derivative and let \(\|f\|_{\infty,j}\) be the supremum norm of the \(j\)-th derivatives. Define \[\|f\|:=\|f\|_{W^{4}}+\|f\|_{\infty,1}+\|f\|_{\infty,0}.\] The main result of this paper is: **Theorem 1.1**.: _Let \(\Gamma\subset PSL_{2}(\mathbb{R})\) be a lattice. Let \(R=T^{\theta}\), where \(0<\theta\leq\frac{\beta}{40}\) is fixed. Here, \(\beta\) is the constant depending only on \(\Gamma\) appearing in Theorem 1.4. For a non-periodic \(\xi\in X\) and a function \(f\in C^{4}(X)\) with \(\|f\|=1\),_ \[\left|\frac{1}{T}\sum_{n\leq T}f(\xi h(n))\nu(n)-\int f\;d\mu_{X}\right|\ll r ^{-\theta}+\frac{\log\log R}{\log R},\] _where \(r=T\exp(-\operatorname{dist}(g_{\log T}(\xi)))\) and the implied constant depends only on \(\Gamma\). Because \(r\to\infty\) as \(T\to\infty\), the sequence \(\xi h(n)\) equidistributes along \(\nu\)._ The biggest possible size for \(\theta\) The second error term is due to the normalization of \(\nu\), while \(r\) rules the equidistribution on the horocycle segment up to time \(T\). Theorem 1.1 recovers some of the properties of the primes. An immediate corollary is the generalization of the result proved by Sarnak and Ubis in [9] for \(PSL_{2}(\mathbb{Z})\). **Corollary 1.2**.: _Take any non-periodic point \(\xi\in X\). For any non-negative function \(f\in C^{4}(X)\) with \(\|f\|=1\),_ \[\frac{1}{\pi(T)}\sum_{p\leq T}f(\xi h(p))\leq\frac{1}{\theta}\int f\;d\mu_{X} +O\left(r^{-\theta}+\frac{\log\log R}{\log R}\right)\] _where the sum is over primes, and \(r\) and \(\theta\) are as in Theorem 1.1. In particular, any limit measure of the primes is absolutely continuous with respect to \(\mu_{X}\) and the primes are dense in a set of positive measure._ Corollary 1.2 could be proved without using Theorem 1.1 by adapting the proof of Theorem 1.1. One would use sieve methods instead of the normalization of \(\nu\) and the Siegel-Walfisz theorem instead of the Siegel-Walfisz type Theorem 2.4 for \(\nu\). Unfortunately, these are all properties of the orbit under primes obtainable through Theorem 1.1. It falls short both of showing density and equidistribution of primes. The reason is that to recover properties of the primes, one needs stronger properties of \(\nu\) (like the \(k\)-linear forms condition in [5], which is stronger than containing arithmetic progressions of length \(k\)). In the case of equidistribution of primes, one would probably need good equidistribution of \(\nu\) not along simple sums, but along sums of the type \[\sum_{n\leq\frac{N}{s_{1}s_{2}}}f(\xi h(ns_{1}))f(\xi h(ns_{2}))\nu(n)\] for natural numbers \(s_{1},s_{2}\). As sums of these types are not well understood at all, even when replacing \(\nu\) by \(1\), showing equidistribution of primes using \(\nu\) seems to require significant new input. The methods employed may be of interest beyond the question in this paper; in particular, the following result may have other applications. It will be instrumental in the proof of equidistribution of \(\xi h(n^{1+\gamma})\) for small \(\gamma\) performed in a follow-up paper. **Lemma 1.3**.: _Let \(p\in X\) and \(T\geq 0\). Let \(\delta>0\) and \(K\leq T\). There is an interval \(I_{0}\subset[0,T]\) of size \(|I_{0}|\leq\delta^{-1}K^{2}\) such that: For all \(t_{0}\in[0,T]\backslash I_{0}\), there is a segment \(\{\xi h(t),t\leq K\}\) of a closed horocycle approximating \(\{ph(t_{0}+t),0\leq t\leq K\}\) of order \(\delta\), in the sense that_ \[\forall 0\leq t\leq K:\quad d_{X}\left(ph(t_{0}+t),\xi h(t)\right)\leq\delta.\] _The period \(P=P(t_{0},p)\) of this closed horocycle is at most \(P\ll r\), where \(r=T\exp(-\mathrm{dist}(g_{\log T}(p)))\) is as in Theorem 1.1. Moreover, one can assure \(P\gg\eta^{2}r\) for some \(\eta>0\) by weakening the bound on \(I_{0}\) to \(|I_{0}|\leq\max\left(\delta^{-1}K^{2},\eta T\right)\)._ This result is useful because it bridges the gap between the compact case with good asymptotics and the periodic case in some sense. _Relation to Previous Papers._ Venkatesh showed that for cocompact lattices, \(\xi h(ns),1\leq n\leq T\), is equidistributed with error \(T^{-\epsilon}\) as long as \(s\ll T^{\epsilon}\)[12]. He deduced that \(\xi h(n^{1+\gamma})\) equidistributes for sufficiently small \(\gamma>0\) (where \(\gamma\) and \(\epsilon\) depend only on the lattice). This is the result that will be generalized to all lattices \(\Gamma\subset PSL_{2}(\mathbb{R})\) in the aforementioned follow-up paper. Venkatesh's proof combined the quantitative equidistribution result \[\left|\frac{1}{T}\int_{0}^{T}f(\xi h(t))\;dt-\int f\;d\mu_{X}\right|\ll_{f}T^ {-2\epsilon}\] for \(f\in C^{\infty}(X)\) (see Lemma 9.4 in [12]; the ideas go back to Ratner) with a trick to bound the Fourier coefficients. With an argument as in the proof of Proposition 5.1, the theorem of Venkatesh immediately implies Theorem 1.1 for cocompact \(\Gamma\). Strombergsson proved an effective equidistribution theorem for the horocycle flow in the non-compact case [11]. Strombergsson showed that \[\left|\frac{1}{T}\int_{0}^{T}f(\xi h(t))\ dt-\int f\ d\mu_{X}\right|\ll_{f}r^{- \alpha}, \tag{1}\] where \(\alpha\) only depends on \(\Gamma\) and \(r\) is as in Theorem 1.1. This relates the asymptotics of the horocycle flow at time \(T\) to the location of \(g_{\log T}(\xi)\); the further \(g_{\log T}(\xi)\) is up in some cusp, the worse the asymptotics. Strombergsson's result can be combined with the method of Venkatesh to prove the theorem below, as done for example by Zheng ([13], Theorem 1.2). **Theorem 1.4**.: _Let \(\Gamma\) be a non-compact lattice in \(G\). Let \(f\in C^{4}(X)\) with \(\|f\|<\infty\) and \(1\leq s<T\). Then_ \[\left|\frac{s}{T}\sum_{1\leq j\leq\mathbb{T}/s}f(\xi h(sj))-\int f\ d\mu_{X} \right|\ll s^{\frac{1}{2}}r^{-\frac{\beta}{2}}\|f\|\] _for any initial point \(\xi\in X\), where \(r=T\exp(-\mathrm{dist}(g_{\log T}(\xi)))\). The parameter \(\frac{1}{6}>\beta>0\) and the implied constant depend only on \(\Gamma\)._ If \(r\gg T^{\epsilon}\) for some \(\epsilon>0\), the situation is very similar to the compact case. The set of points \(\xi\) with \(r\gg T^{\epsilon}\) for all \(T\) has full measure (by geodesic excursion rates, compare the introduction of [11]). Thus, if one restricts the analysis to a subset of initial points of full measure, as done by Zheng in [13], Theorem 1.4 is all that is needed to show equidistribution of \(n^{1+\gamma}\). Similarly, McAdam shows density of almost primes in \(SL_{n}(\mathbb{Z})\backslash SL_{n}(\mathbb{R})\) under a Diophantine condition which is equivalent to \(r\gg T^{\epsilon}\) in two dimensions [8]; compare Remark 4.3. Statements about density are not hard to get from Theorem 1.4, because any non-periodic \(\xi\) has a sequence \(T_{i}\to\infty\) such that \(r(T_{i})\gg T_{i}\). This holds because \(g_{\log T}(\xi)\) returns to a compact set infinitely often (compare Lemma 4.2). Explicitly, one immediately gets density of \(\xi h(n^{1+\gamma})\) in \(X\), density of almost primes in \(X\) and density of primes in a set of positive measure (shown as in the proof of Proposition 5.1) from Theorem 1.4. Sarnak and Ubis chose a different approach and analyzed the quantitative equidistribution of \(\xi h(sn)\), for _all_\(\xi\) and _all_ times \(T\)[9]. They did this in the case \(\Gamma=PSL_{2}(\mathbb{Z})\), defining a fundamental period \(y_{T}\). This fundamental period is based on the imaginary part of the horocycle segment \(\xi h([0,T])\) and turns out to be closely related to \(r\). They then proceeded to show that the horocycle segment \(\xi h(t),0\leq t\leq T\) is approximable by periodic horocycle with period at most \(y_{T}\). Analyzing the situation on the periodic horocycle separately, they deduced that for non-negative \(f\in C^{4}\) \[\frac{1}{\pi(T)}\sum_{p\leq T}f(\xi h(p))\leq 10\int f\;d\mu_{X}+o_{T}(1)\] for all \(T\), which implies non-concentration of primes. They did not use Strombergsson's result and Theorem 1.4, but used estimates of automorphic forms to obtain similar asymptotics for \(r\gg T^{\epsilon}\). _Strategy._ We will combine Theorem 1.4 with the approach of Sarnak and Ubis to generalize their result to all lattices \(\Gamma\subset G\). The main step is to generalize their fundamental period from \(\Gamma=PSL_{2}(\mathbb{Z})\) to arbitrary \(\Gamma\). This approach culminates in Lemma 1.3. This theorem will allow us to reduce the analysis to closed horocycles in the cases when \(r\ll T^{\frac{1}{20}}\) and the asymptotics are bad. On those, we will use the Siegel-Walfisz type Theorem 2.4 for \(\nu\) to finish the proof. _Structure of this paper._ Chapter 2 contains the proof of the Siegel-Walfisz theorem for \(\nu\) and ends with a short proof of Corollary 1.2 assuming Theorem 1.1. Chapter 3 recalls basics of the dynamics on quotients of \(G\) and their relation to quotients of the hyperbolic plane. In Chapter 4, Lemma 1.3 is proven by generalizing the fundamental period \(y_{T}\) to all lattices \(\Gamma\), establishing \(r\sim y_{T}^{-1}\) and analyzing horocycle segments in \(PSL_{2}(\mathbb{R})\). In Chapter 5, Theorem 1.4, Lemma 1.3 and Theorem 2.4 are combined to prove Theorem 1.1. _Notation._ Elements and flows in \(G=PSL_{2}(\mathbb{R})\) are denoted by \[h(x)=\begin{pmatrix}1&x\\ 0&1\end{pmatrix},\quad a(y)=\begin{pmatrix}y^{\frac{1}{2}}&0\\ 0&y^{-\frac{1}{2}}\end{pmatrix},\quad k(\theta)=\begin{pmatrix}\cos\theta& \sin\theta\\ -\sin\theta&\cos\theta\end{pmatrix},\] \(g_{t}(g)=ga(e^{t})\) and \(h_{t}(g)=gh(t)\). We set \(e(t):=\exp(2\pi it)\). A fundamental domain of \(X=\Gamma\backslash G\) is chosen to be the unit tangent bundle \(E=T_{1}F\) of a connected fundamental domain \(F\) of the action of \(\Gamma\) on the upper half plane \(\mathbb{H}\). The open interior is denoted by \(F^{\rm o}\) and the closure by \(\overline{F}\). For \(g\in G\), \(g.z=\frac{az+b}{cz+d}\), and \(g.i\) is the projection to \(\mathbb{H}\). When we write \(\operatorname{Im}(g)\), we mean \(\operatorname{Im}(g.i)\). The objects \(r\), \(\operatorname{dist}(\xi)\) and the norm \(\|f\|\) of \(f\in C^{4}(X)\) are defined before Theorem 1.1. The definition of \(y_{T}\) can be found in Chapter 4 on Page 14. The inequality \(f\ll g\) and \(f=O(g)\) mean that there is an absolute constant \(C\) such that \(|f(t)|\leq C|g(t)|\) for all \(t\). Write \(f\sim g\) if \(f\ll g\) and \(g\ll f\). Unless equipped with an index to indicate further dependence, the implicit constants only depend on \(\Gamma\). The divisor function \(\tau(n)\) counts the divisors of \(n\) and the Euler totient function is denoted by \(\phi(n)\). The Mobius function \(\mu\) is supported on square free numbers and is defined by \[\mu(1)=1,\quad\mu(p_{1}\dots p_{n})=(-1)^{n}\] for distinct prime numbers \(p_{1},\dots,p_{n}\). Its square \(\mu^{2}\) is the characteristic function of square free numbers. For integers \(e\) and \(d\), the least common multiple is denoted by \([e,d]\) and the greatest common divisor by\((e,d)\). _Acknowledgements._ This text was written as my master's thesis at the Hebrew University in Jerusalem. I am very grateful for the warm welcome, the superb learning environment and the generous support I received at the Hebrew University, especially from Hillel Furstenberg, Elon Lindenstrauss, Shahar Mozes, Jane Turner and my advisor Tamar Ziegler. Initially, I came to Jerusalem in an exchange year. After an amazing year, I decided to finish my entire degree in Jerusalem. Many thanks to the entire department and especially to Elon and Tamar, who organized support me in my second year. I am thankful for the many discussions with Tamar about the project and to Elon's open ear for questions whenever I had any. One of the most astonishing experiences in this year for me was a 1on1 reading course with Hillel about his proof of Szemeredi's theorem. I learned a lot from his explanations of the proof and his views on the life of a mathematician in general, patiently delivered in his small office amidst piles of books and theses, veiled in layers of chalk dust. I am very happy that I came to a university where even the most senior researchers (Hillel was 83 at the time) still come to their offices and gladly teach young students. A big thank you to the anonymous referee, whose review improved this paper tremendously. The suggested tweaks to the proofs made the ideas much clearer and saved close to 10 pages of calculation. If the reader finds the proofs intuitive, the changes made after the review have a substantial part in that. Moreover, the review did not only help the reader. It also made the ideas clearer to me and got me thinking about the material again. This led me to a proof of equidistribution of \(ph(n^{1+\gamma})\), which will be the subject of a follow-up paper. Without the very helpful review, this would in all likelihood not have happened. Lastly, I want to thank Dan Goldston, Andreas Strombergsson, Peter Sarnak and especially Adrian Ubis for their helpful replies to questions I asked. ## 2. Properties of \(\nu\) In this section, we are going to derive the Siegel-Walfisz type Theorem 2.4 for \(\nu\) and prove that Theorem 1.1 implies Corollary 1.2. **Lemma 2.1**.: _(Lemma 2.1 in [4]) Let \(R>0,k\in\mathbb{N}\) such that \(\log(k)\ll\log(R)\). Then_ \[\sum_{d\leq R,(d,k)=1}\frac{\mu(d)}{d}\log\left(\frac{R}{d}\right)=\frac{k}{ \phi(k)}+O\left(\frac{1}{\log^{3}(R)}\right)\] _and_ \[\sum_{d\leq R,(d,k)=1}\frac{\mu(d)}{\phi(d)}\log\left(\frac{R}{d}\right)= \mathfrak{S}_{2}(k)+O\left(\frac{1}{\log^{3}(R)}\right),\] _where \(\mathfrak{S}_{2}\) is the singular series from the Goldbach conjecture, supported on positive even numbers and given by \(\mathfrak{S}_{2}(2n)=2C_{2}\prod_{p|n,p>2}\left(\frac{p-1}{p-2}\right)\) with \(C_{2}=\prod_{p>2}\left(1-\frac{1}{(p-1)^{2}}\right)\)._ The next lemma from [4] is cited here only in the case \(j=1\) with simplified error terms. The validity of the simplifications can be seen from their remarks after Lemma 2.2 and the fact that \(\mathfrak{S}_{2}(k)\ll\tau(k)\). **Lemma 2.2**.: _(Lemma 2.4 in [4]) Let \(R\geq 1\) and \(k\in\mathbb{N}\) such that \(\log k\ll\log R\). Then_ \[\sum_{d\leq R,(d,k)=1}\frac{\mu^{2}(d)}{\phi(d)}\mathfrak{S}_{2}(dk)=\log(R)+ O\left(\log\log 3k\right).\] These lemmas can be combined in a similar way to the proof of Theorem 5.1 in [4] to yield the following proposition. **Proposition 2.3**.: _Let \(I\) be an interval in \(\mathbb{N}\) and \(R>1\). Let \(q\in\mathbb{N}\) such that \(\log q\ll\log R\) and let \(j\leq q\) be coprime to \(q\). Then_ \[\frac{1}{|I|}\sum_{n\in I}\Lambda_{R}^{2}(qn+j)=\frac{q}{\phi(q)}\log R+\frac {q}{\phi(q)}O\left(\log\log R\right)+O\left(\frac{R^{2}}{|I|}\right).\] Proof.: Note that for any integer \(k\), \[\sum_{\begin{subarray}{c}n\in I\\ k|qn+j\end{subarray}}1=\begin{cases}\frac{|I|}{k}+O(1),\ \ (q,k)=1\\ 0,\ \ \text{else}\end{cases}.\] To bound the sums of the appearing error terms, we record for later use that \[\sum_{k\leq R}\frac{1}{k\log\left(R/k\right)}\leq\sum_{j\leq\log R}\sum_{ \begin{subarray}{c}R\\ 2^{j+1}\end{subarray}}\sum_{k\leq\frac{R}{2^{j}}}\frac{1}{k\log\left(2^{j} \right)}=O(\log\log R) \tag{2}\] and, similarly, using the well known bound \(\frac{1}{\phi(k)}\ll\frac{\log\log k}{k}\), \[\sum_{k\leq R}\frac{1}{\phi(k)\log^{2}\left(R/k\right)}=O(\log\log R). \tag{3}\] So let us bound the terms. Unboxing the definition, we see that \[\sum_{n\in I}\Lambda_{R}^{2}(qn+j) =\sum_{d,e\leq R}\mu(d)\mu(e)\log\left(\frac{R}{d}\right)\log \left(\frac{R}{e}\right)\sum_{\begin{subarray}{c}n\in I\\ d,e|qn+j\end{subarray}}1\] \[=O(R^{2})+|I|\sum_{\begin{subarray}{c}d,e\leq R\\ (de,q)=1\end{subarray}}\frac{\mu(d)\mu(e)\log\left(\frac{R}{d}\right)\log \left(\frac{R}{e}\right)}{[d,e]}\] where \([d,e]\) is the least common multiple. Denote by \(\sum^{\prime}\) a sum in which all summation variables are coprime to \(q\) and to each other. We find by using Lemma 2.1 that \[\sum_{\begin{subarray}{c}d,e\leq R\\ (de,q)=1\end{subarray}}\frac{\mu(d)\mu(e)\log\left(\frac{R}{d}\right)\log \left(\frac{R}{e}\right)}{[d,e]}\] \[=\sum_{\begin{subarray}{c}m\leq R\\ md,me\leq R\end{subarray}}^{\prime}\frac{\mu^{2}(m)}{m}\frac{\mu(d)}{d}\frac{ \mu(e)}{e}\log\left(\frac{R}{md}\right)\log\left(\frac{R}{me}\right)\] \[=\sum_{md\leq R}^{\prime}\frac{\mu^{2}(m)}{m}\frac{\mu(d)}{d}\log \left(\frac{R}{md}\right)\sum_{\begin{subarray}{c}e\leq R/m\\ (e,mdq)=1\end{subarray}}\frac{\mu(e)}{e}\log\left(\frac{R/m}{e}\right)\] \[=\frac{q}{\phi(q)}\left(\sum_{md\leq R}^{\prime}\frac{\mu^{2}(m) }{\phi(m)}\frac{\mu(d)}{\phi(d)}\log\left(\frac{R}{md}\right)\right)+O(\log \log R),\] where (2) was used to bound the error sum coming from Lemma 2.1. Applying Lemma 2.1, to the main term, we find \[\sum_{md\leq R}^{\prime}\frac{\mu^{2}(m)}{\phi(m)}\frac{\mu(d)}{ \phi(d)}\log\left(\frac{R}{md}\right)=\sum_{\begin{subarray}{c}m\leq R\\ (m,q)=1\end{subarray}}\frac{\mu^{2}(m)}{\phi(m)}\sum_{\begin{subarray}{c}d \leq R/m\\ (d,mq)=1\end{subarray}}\frac{\mu(d)}{\phi(d)}\log\left(\frac{R/m}{d}\right)\] \[=\sum_{\begin{subarray}{c}m\leq R\\ (m,q)=1\end{subarray}}\frac{\mu^{2}(m)}{\phi(m)}\mathfrak{S}_{2}(mq)+O\left( \log\log R\right),\] where the error term was bounded with the help of (3). Applying Lemma 2.2 finishes the proof. **Theorem 2.4** (Siegel-Walfisz for \(\nu\)).: _Let \(N\in\mathbb{N}\) and let \(\nu(n)=\frac{\Lambda_{R}^{2}(n)}{\log R}\) with \(R\) of size \(N^{\epsilon}\). For any \(q\in\mathbb{N}\), any interval \(I\subset[0,N]\) of size \(|I|\geq qR^{3}\) and any \(q\)-periodic function \(f\),_ \[\frac{1}{|I|}\sum_{n\in I}f(n)\nu(n)=\frac{q}{\phi(q)|I|}\sum_{ \begin{subarray}{c}n\in I\\ (n,q)=1\end{subarray}}f(n)+O\left(\frac{\|f\|_{\infty}\log\log R}{\log R}\right),\] _where \(\|f\|_{\infty}=\max_{r\leq q}|f(r)|\). In particular,_ \[\frac{1}{N}\sum_{n\leq N}\nu(n)=1+O\left(\frac{\log\log R}{\log R}\right)\] Proof.: Without loss of generality, assume that \(\|f\|_{\infty}=1\). Fix an interval \(I\). For \(q=1\), by Proposition 2.3, \[\frac{1}{|I|}\sum_{n\in I}\nu(n)=1+O\left(\frac{\log\log R}{\log R}\right)+O \left(\frac{1}{R}\right).\] For general \(q\leq N\) and some coprime \(j<q\), Proposition 2.3 applied to \(q^{-1}(I-j)\) (the assumption \(\log q\ll\log R\) is satisfied because \(R\) is of size \(N^{\varepsilon}\)) implies that \[\frac{q}{|I|}\sum_{\begin{subarray}{c}n\in I\\ n\equiv j(q)\end{subarray}}\nu(n)=\frac{q}{\phi(q)}+\frac{q}{\phi(q)}O\left( \frac{\log\log R}{\log R}\right)+O\left(\frac{1}{R}\right).\] In particular, all residue classes coprime to \(q\) contribute the same amount to the sum; it is thus only left to check that the residue classes which are not coprime have negligible contribution. To see this, average over all residue classes \(j\) coprime to \(q\), giving \[\frac{1}{|I|}\sum_{\begin{subarray}{c}n\in I\\ (n,q)=1\end{subarray}}\nu(n)=1+O\left(\frac{\log\log R}{\log R}\right)+O\left( \frac{1}{R}\right),\] and compare with the contribution of all residue classes above. To conclude, we will prove Corollary 1.2 assuming Theorem 1.1. Proof.: By partial summation and the prime number theorem, \[\frac{1}{\pi(T)}\sum_{p\leq T}f(\xi h(p))=\frac{1}{T}\sum_{n\leq T }f(\xi h(n))\ \tilde{\Lambda}(n)+O\left(\frac{\|f\|_{\infty}}{\log T}\right),\] where \(\tilde{\Lambda}\) is given by \[\tilde{\Lambda}(n)=\begin{cases}\log(p),\ n=p\ \text{prime}\\ 0,\ \text{else}\end{cases}.\] By definition of \(\nu\), for all primes \(p>R\) \[\nu(p)=\log R=\frac{1}{\theta}\log T.\] In particular, \(\tilde{\Lambda}(n)\leq\frac{1}{\theta}\nu(n)\) on the interval \([T^{\theta},T]\). The corollary follows from Theorem 1.1. ## 3. Basics of the Dynamics on Quotients of the Hyperbolic Plane In this section, we recall some basics of the hyperbolic plane. Any matrix \(g\in G\) can be uniquely written as \[g=h(x)a(y)k(\theta)\] for \(x\in\mathbb{R},y>0,\theta\in[-\nicefrac{{\pi}}{{2}},\nicefrac{{\pi}}{{2}})\); this is the Iwasawa parametrization. Moreover, \(G\) has a natural left-invariant metric given by \[d_{G}t^{2}=\frac{dx^{2}+dy^{2}}{y^{2}}+d\theta^{2}.\] The upper half plane \(\mathbb{H}\) carries the hyperbolic metric, which is invariant under the action of \(G\) via Mobius transformations \(z\mapsto g.z\). This action lifts to an action on the unit tangent bundle \(T_{1}\mathbb{H}\) given by \[\begin{pmatrix}a&b\\ c&d\end{pmatrix}.(z,v)=\left(\frac{az+b}{cz+d},\frac{v}{(cz+d)^{2}}\right).\] There is a natural bijection \(PSL_{2}(\mathbb{R})\to T_{1}\mathbb{H}\) given by \[h(x)a(y)k(\theta)\mapsto\left(x+iy,e^{i2\theta}\right)\] which is isometric with respect to the metric on \(T_{1}\mathbb{H}\) induced by the hyperbolic metric on \(\mathbb{H}\). The measure \[\mu_{G}=\frac{dxdy}{y^{2}}\frac{d\theta}{\pi}\] is a Haar measure on \(G\) (\(G\) is unimodular, so there is no distinction between a left and right Haar measure). Fix a lattice \(\Gamma\) in \(G\) and set \(X:=\Gamma\backslash G\). This space carries a left-invariant metric \[d_{X}(\Gamma g_{1},\Gamma g_{2})=\min_{\gamma\in\Gamma}d_{G}(g_{1},\gamma g_{2})\] and a finite measure \(\mu:=\mu_{X}=\pi_{\#}\mu_{G}\), the push forward of \(\mu_{G}\) under the projection to \(X\). Write \(p.i\) for \(\Gamma g.i\in\Gamma\backslash\mathbb{H}\) with \(p=\Gamma g\in X\). The fundamental domain \(E\) of \(X\) in \(G\) can be chosen to be the tangent bundle of a convex hyperbolic polygon \(F\subset\mathbb{H}\) (i. e. a polygon in which each edge is a geodesic) with finitely many edges and vertices; this \(F\) is a fundamental domain for \(\Gamma\backslash\mathbb{H}\). Explicitly, one can choose \(F\) to be a Dirichlet domain for any point \(z_{0}\in\mathbb{H}\); that is, the interior of \(F\) is given by \[F^{\circ}=D(z)=\{z\in\mathbb{H}|d_{G}(z,z_{0})<d_{G}(z,\gamma z_{0})\;\forall \gamma\in\Gamma\backslash\{\mathrm{id}\}\}.\] The polygon \(F\) might or might not have boundary vertices, i. e. point of adjacency with \(\partial\mathbb{H}=\mathbb{R}\cup\infty\). \(X\) is compact if and only if there are none of those. If there are some, the equivalence class of each boundary vertex with respect to \(\Gamma\) is called a cusp of \(X\) (for example if \(\Gamma=PSL_{2}(\mathbb{Z})\), \(X\) has the single cusp \(\infty\) which is equivalent to every rational number). For each representative \(r_{i}\in\partial\mathbb{H}\) of each cusp \(\Gamma r_{i}\) there is a fundamental domain \(F\) such that all other boundary vertices of \(F\) are inequivalent to \(r_{i}\); this is because we can take some point far up in the cusp as basic point for the Dirichlet domain. Proofs for all of these statements can be found in chapter 11 of [2]. There are two important flows on \(G\). The first one is the geodesic flow given by \[g_{t}(g)=ga(e^{t})=\begin{pmatrix}ae^{\frac{t}{2}}&be^{-\frac{t}{2}}\\ ce^{\frac{t}{2}}&de^{-\frac{t}{2}}\end{pmatrix}\] and the second one is the horocycle flow given by \[h_{t}(g)=gh(t)=\begin{pmatrix}a&b+at\\ c&d+ct\end{pmatrix}\] where \(g=\begin{pmatrix}a&b\\ c&d\end{pmatrix}\). The flows are well-defined on \(X\) because \(\Gamma\) acts from the left and the flows act from the right. On \(G\) the behaviour of these flows is not very interesting, but on \(X\) it is. As outlined in the introduction, the dynamics with respect to the horocycle flow exhibit a very rigid behaviour. There are no periodic horocycle orbits in \(X\) if and only if \(X\) is compact. If there are periodic orbits, their structure is as follows: **Lemma 3.1**.: _(Lemma 11.29 in [2]) Let \(\Gamma\) be a lattice such that \(X\) is non-compact. To every cusp of \(X\) corresponds exactly one one-parameter family of \(h\)-periodic orbits parametrized by \(g_{t}\); in explanation, if \(p\in X\) is such that \(t\mapsto ph(t)\) is periodic, then \(g_{t}(p)\) converges to some cusp of \(X\) as \(t\to\infty\) and all other periodic orbits associated to this cusp contain exactly one element \(g_{t}(p)\) for \(t\in\mathbb{R}\). Furthermore, the orbit \(t\mapsto ph(t)\) is periodic if and only if \(g_{t}(p)\to\infty\), in the sense that \(g_{t}(p)\) leaves any compact subset of \(X\) permanently._ It is shown in the proof (or alternatively, can be seen directly from the statement) that for any boundary vertex of \(F\) there is a \(\gamma\in\Gamma\) conjugated to \(\begin{pmatrix}1&1\\ &1\end{pmatrix}\) which fixes this boundary vertex and generates the subgroup of \(\Gamma\) fixing the vertex. This \(\gamma\) is precisely the one leading to the periodicity of the corresponding orbits. For example for \(\Gamma=PSL_{2}(\mathbb{Z})\) and the sole boundary vertex \(\infty\), this matrix is \(\gamma=\begin{pmatrix}1&1\\ &1\end{pmatrix}\). If \(p\) is periodic with period \(y\), then \(g_{t}(p)\) is periodic with period \(ye^{-t}\) because of the equation \(g_{t}\circ h_{s}=h_{e^{-t}s}\circ g_{t}\). ## 4. Approximation by closed horocycles In this chapter, we will define the fundamental period of a horocycle piece (first defined for \(\Gamma=PSL_{2}(\mathbb{Z})\) in [9]) and explore the connection to effective equidistribution. The ultimate goal of this section is to prove Lemma 1.3. Let \(n\) be the number of cusps of \(X\). Let \(r_{i}\in\partial\mathbb{H}\) be a representative of one cusp of \(X\) and let \(\gamma_{i}\in\Gamma\) be the corresponding unipotent element fixing \(r_{i}\) and inducing the periodicity of the corresponding horocycle, as in the discussion after Lemma 3.1. Let \(\sigma_{i}\in G\) such that \(\sigma_{i}\gamma_{i}\sigma_{i}^{-1}=h(1)\) and \(\sigma_{i}.r_{i}=\infty\). Explicitly, this \(\sigma_{i}\) consists of a rotation matrix, rotating \(r_{i}\) to \(\infty\) and \(\gamma_{i}\) to some \(h(t_{i})\), and some diagonal element to normalize \(t_{i}=1\). Define \(y_{i}:G\to\mathbb{R}_{+}\) by \[y_{i}(g)=\operatorname{Im}(\sigma_{i}g)\] where we mean in slight abuse of notation the imaginary part of the corresponding basepoint \(\sigma_{i}g.i\) in \(\mathbb{H}\). Note that \(y_{i}\) is only well-defined on \(G\), not on \(X\), and depends on the representative of the cusp. In the natural parametrization \(h(x)a(y).i\) for \(\mathbb{H}\), every level set with fixed \(y\) is a horizontal line, so a circle with point of tangency \(\infty\) which is invariant under \(h(1)\). We could also parametrize \(\mathbb{H}\) by \(\sigma_{i}^{-1}h(x)a(y).i\), in terms of which any level set is a circle with point of tangency \(r_{i}\) invariant under \(\gamma_{i}\). \(y_{i}(z)\) then is the \(y\)-component of \(z\) in terms of this new parametrization. Let for \(T>0\) \[Y_{i}^{T}(g)=\min\left\{y_{i}(gh(t))\middle|0\leq t\leq T\right\}=\min(y_{i}(g),y_{i}(gh(T)))\] where the second equality follows from the fact that \(\{\sigma_{i}gh(t)|0\leq t\leq T\}\) is a piece of a horocycle orbit and thus a segment of a circle in \(\mathbb{H}\). Define \(y_{i}^{T}:X\to\mathbb{R}_{+}\) by \[y_{i}^{T}(\Gamma g)=\sup_{\gamma\in\Gamma}Y_{i}^{T}(\gamma g).\] The supremum is finite and attained for some \(\gamma\) because \(\Gamma\) is discrete and acts properly discontinuous on \(X\). It is independent of the representative of the cusp. Note that \(y_{i}^{0}:\Gamma\backslash\mathbb{H}\to\mathbb{R}_{+}\), i. e. \(y_{i}^{0}\) depends only on the base point \(\Gamma g.i\) of \(\Gamma g\). Lastly, define \[y_{T}(p)=\max_{1\leq i\leq n}[y_{i}^{T}(p)].\] The horocycle piece up to time \(T\) starting in \(p\) will be close to a periodic horocycle with period \(y_{T}^{-1}\). This is called the textitfundamental period at time \(T\), whose properties we will explore now. **Lemma 4.1** (Parametrization in the Cusps).: _1. For some small \(\epsilon>0\), each \(C_{i}:=\{p|y_{i}^{0}(p)^{-1}<\epsilon\}\subset X\) is an open neighborhood of the cusp \(\Gamma r_{i}\) and all \(C_{i}\) are pairwise disjoint. The set \(K:=X\backslash\bigcup C_{i}\) is compact. 2. Any point \(p\) is contained in \(C_{i}\) if and only if there is a \(q\) with the same base point (\(p.i=q.i\) in \(\Gamma\backslash\mathbb{H}\)) such that \(qh(t)\) is periodic with period smaller than \(\epsilon\). In this case, \(qh(t)\) has period \(y_{i}^{0}(p)^{-1}\). 3. Fix disjoint \(C_{i}\) and \(K\) as in 1. and take some \(p_{0}\in K\). Then_ \[\exp(d_{X}(p,p_{0}))\sim y_{0}(p)=\max_{1\leq i\leq n}[y_{i}^{0}(p)].\] _More explicitly, if \(p\in C_{i}\), then \(\exp(d_{X}(p,p_{0}))\sim y_{i}^{0}(p)\) and if \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\sim y_{0}(p)\)._ The implied constants depend on the choice of \(\epsilon\), but \(C_{i}\) and \(K\) will be fixed from here on. Parts of the lemma have appeared in the literature before; the function \(y_{0}(p)\) is known as the invariant height function, compare (11) in [11]. Part 3. of Lemma 4.1 was stated as Inequality (14) in the same paper paper, leaving the proof as an exercise. We will prove it here for completeness. Proof.: As discussed in chapter 3, we can choose a fundamental domain \(E\) such that \(E=T_{1}F\), where the open interior \(F^{\rm o}\) is a Dirichlet domain with boundary vertex \(r_{i}\) such that all other boundary vertices of \(F\) are inequivalent to \(r_{i}\). Then \(\sigma_{i}F\) is a fundamental domain of \(\sigma_{i}\Gamma\sigma_{i}^{-1}\backslash\mathbb{H}\) with boundary vertex \(\infty\) which is inequivalent to all other boundary vertices. \(\sigma_{i}F\) is a hyperbolic polygon with finitely many sides. Because \(\infty\) is a vertex, two of them must be straight vertical lines. Thus for some big \(B\in\mathbb{R}\), \(\sigma_{i}F\cap\{z|{\rm Im}(z)>B\}\) is a rectangle. Because there are no other equivalent boundary vertices and \(h(1)\in\sigma_{i}\Gamma\sigma_{i}^{-1}\) by choice of \(\sigma_{i}\), the horizontal line of this rectangle has euclidean length \(1\) and \[\{z|\text{Im}(z)>B\}=\bigcup_{k\in\mathbb{Z}}h(k)\left(\sigma_{i}F\cap\{z|\text{ Im}(z)>B\}\right).\] 1. Set \(D_{i}=\Gamma\sigma_{i}^{-1}\{z\in\sigma_{i}F\ |\text{Im}(z)>B\}\) and \(C_{i}=T_{1}D_{i}\subset X\). Let \(g\in E\). If \(\Gamma g\in C_{i}\), by definition \(y_{i}^{0}(\Gamma g)\geq y_{i}(g)=\text{Im}(\sigma_{i}g)>B\). If on the other hand \(\Gamma g\notin C_{i}\), because the left translates \(h(k)\{z\in\sigma_{i}F\ |\text{Im}(z)>B\}\) exactly tile the set \(\{z|\text{Im}(z)>B\}\), we must have \(y_{i}(\gamma g)\leq B\) for all \(\gamma\in\Gamma\) (else \(g\) would have two different representatives in the fundamental domain). Thus \(y_{i}^{0}(g)\leq B\). This shows that \(C_{i}=\{p|y_{i}^{0}(p)^{-1}<\epsilon_{i}\}\) with \(\epsilon_{i}=B^{-1}\). 2. Let \(p\) now be periodic with period \(b<\epsilon\) and let \(g\) be the representative of \(p\) in \(E\). Then \(\gamma_{i}g=gh(b)\) by our choice of \(\gamma_{i}\). The orbit of \(\sigma_{i}g\) is periodic with respect to infinity in \(\sigma\Gamma\sigma^{-1}\backslash G\) because \[h(1)\sigma_{i}g=\sigma_{i}\gamma_{i}\sigma_{i}^{-1}\sigma_{i}g=\sigma_{i}gh(b),\] But the orbit \(U=\{h(x)a(b^{-1})|0\leq x\leq 1\}\) is also in \(\sigma_{i}E\) and is periodic with period \(b\), so by Lemma 3.1 they have to agree. Because \(b<\epsilon\), we get \(\sigma_{i}p\in U\subset\sigma_{i}C_{i}\) and \(y_{i}^{0}(p)=b^{-1}\). If on the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 3. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 4. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 5. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 6. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 7. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 8. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 9. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) with \(q^{\prime}.i=p.i\) and a smaller period, because \(\sigma_{i}q^{\prime}.i\) would then have to be at a different level in the fundamental domain \(\sigma_{i}F\). 10. If \(p\in K\), \(\exp(d_{X}(p,p_{0}))\sim 1\) because \(K\) is compact. On the other hand \(p=\Gamma g\in C_{i}\) such that \(g\in E\), then \(\sigma_{i}g=(z,v)\) for some \(z\) with \(\text{Im}(z)>\epsilon^{-1}\). Set \(q=\sigma_{i}^{-1}(z,i)\) and note that \(qh(t)\) is periodic with period \(\text{Im}(z)^{-1}=y_{i}^{0}(p)^{-1}\) because \(t\mapsto\sigma_{i}qh(t)\) is. There cannot be a \(q^{\prime}\) hand, \(K.i\subset\Gamma\backslash\mathbb{H}\) is also compact and thus \(\overline{K.i}\subset\overline{F}\) as well. Consequently, the continuous functions \(y_{i}\) have to be bounded from above and below on this set, showing \(y_{0}(p)\sim 1\). If \(p\in C_{i}\), \(y_{i}^{0}(p)>\epsilon^{-1}\) and \(y_{j}^{0}(p)\leq\epsilon^{-1}\) for \(j\neq i\), so \(y_{0}(p)=y_{i}^{0}(p)\). We can find a point \(\Gamma h\in K\) which is close enough to the cusp \(\Gamma r_{i}\) so that \(r_{i}\) has no equivalent boundary vertices in the Dirichlet domain \(D(h.i)\). Consider the fundamental domain \(E=T_{1}F\) with \(F^{\rm o}=D(h.i)\). Let \(g\) be the representative of \(p\) in \(E\). Note that because \(\sigma_{i}F\) has width at most \(1\), \(\sigma_{i}g\) and \(\sigma_{i}h\) essentially only differ in their \(a\) component in the Iwasawa decomposition, that is, their imaginary part. Furthermore, \(\operatorname{Im}(\sigma_{i}h)\sim 1\) because \(\Gamma h\in K\). Thus by definition of the Dirichlet domain, the left invariance of the metric and the choice of the fundamental domain, \[\exp(d_{X}(p,p_{0})) \sim\exp(d_{X}(p,\Gamma h))=\exp(d_{G}(g,h))\] \[=\exp(d_{G}(\sigma_{i}g,\sigma_{i}h))\sim\operatorname{Im}(\sigma _{i}g)=y_{i}(g)=y_{i}^{0}(p).\] Now we are in the position to establish the connection between the fundamental period and \(r\). **Proposition 4.2**.: _Let \(p\in X\) and \(T\geq 3\). Let \(C_{i}\), \(K\) as in Lemma 4.1 and fix some \(p_{0}\in K\). Let \(r=Te^{-\operatorname{dist}(g_{\log T}(p))}\) be as in Theorem 1.1. With \(y_{T}\) defined in the beginning of the chapter,_ \[r^{-1}\sim y_{T}.\] _More explicitly, if \(g_{\log T}(p)\in C_{i}\), then \(r^{-1}(p)\sim y_{i}^{T}(p)\) and if \(g_{\log T}(p)\in K\), \(r^{-1}(p)\sim T^{-1}\sim y_{T}(p)\). All implied constants depend only on the choice of \(C_{i}\) and \(p_{0}\), so ultimately only on \(\Gamma\)._ Proof.: In light of Lemma 4.1, it suffices to show \[y_{i}^{0}(g_{\log T}(p))T^{-1}\sim y_{i}^{T}(p).\] Fix a representative \(g\) of \(p\). Pick some \(\gamma\in\Gamma\) and set \[\begin{pmatrix}a&b\\ c&d\end{pmatrix}:=\sigma_{i}\gamma g\in PSL_{2}(\mathbb{R}),\] where we choose \(c\) to be non-negative. Then \[Y_{i}^{T}(\gamma g)=\min\left(\frac{1}{c^{2}+d^{2}},\frac{1}{c^{2}+(Tc+d)^{2}}\right)\] by definition; this can be simplified because \[\min\left(\frac{1}{c^{2}+d^{2}},\frac{1}{c^{2}+(Tc+d)^{2}}\right)\sim\min\left( \frac{1}{T^{2}c^{2}},\frac{1}{d^{2}}\right), \tag{4}\] which is left to the reader as an exercise. The observation that for any \(r,s>0\), \[\frac{1}{r+s}\sim\min\left(\frac{1}{r},\frac{1}{s}\right)\] may come in handy showing this. With (4) in hand, \[T^{-1}Y_{i}^{0}(\gamma g_{\log T}(g))=\frac{1}{T^{2}c^{2}+d^{2}}\sim\min\left( \frac{1}{T^{2}c^{2}},\frac{1}{d^{2}}\right)\sim Y_{i}^{T}(\gamma g).\] Taking the supremum over \(\gamma\) finishes the proof. At this point, let us relate two other papers building on the results of Venkatesh and let us translate the respective conditions on the points into our notation. **Remark 4.3**.: In two dimensions the Diophantine condition for \(\Gamma g\) (Equation (3.1.c) on page 11 in [8]) of McAdam is \[\min_{\omega\in\mathbb{Z}^{2}\backslash\{0\}}\max_{0\leq t\leq T}\|\omega gh( t)\|_{\infty}\gg T^{\epsilon}\] which translated in the notation of the proof is equivalent to \[\min_{\gamma\in SL_{2}(\mathbb{Z})}\max(|c_{\gamma}|,|d_{\gamma}|,|d_{\gamma} +Tc_{\gamma}|)\gg T^{\epsilon}\] with \(\begin{pmatrix}a_{\gamma}&b_{\gamma}\\ c_{\gamma}&d_{\gamma}\end{pmatrix}=\gamma g\). As in Proposition 4.2, the left hand side is asymptotically equal to \(y_{T}^{-\frac{1}{2}}\). Zheng proves equidistribution of \(n^{1+\eta}\) for points fulfilling a \(\kappa\)-Diophantine condition, where \(\eta\) depends on \(\kappa\), [13]. In our notation, this \(\kappa=(\kappa_{1},\ldots,\kappa_{n})\)-condition for \(\kappa_{i}>0\) says that for any cusp there exist \(a_{i},b_{i}>0\) such that either \(|c_{\gamma}|>a_{i}\) or \(|d_{\gamma}|^{\kappa_{i}}|c_{\gamma}|>b_{i}\) for all \(\gamma\). It's easy to check that this condition implies \(Y_{i}^{T}(\gamma g)\ll T^{-\frac{2}{1+\kappa_{i}}}\), so by Proposition 4.2 again \(r\gg T^{\epsilon}\) for \(\epsilon<\min_{i}\frac{2}{1+\kappa_{i}}\). In both cases Theorem 1.4 thus immediately implies the respective results. To finish off this section, we prove Lemma 1.3, giving means to approximate the horocycle segment \(ph([0,T])\) with closed horocycles segments of period at most \(y_{T}^{-1}\sim r\). We restate Lemma 1.3 for the convenience of the reader. **Lemma 1.3**.: _Let \(p\in X\) and \(T\geq 0\). Let \(\delta>0\) and \(K\leq T\). There is an interval \(I_{0}\subset[0,T]\) of size \(|I_{0}|\leq\delta^{-1}K^{2}\) such that: For all \(t_{0}\in[0,T]\backslash I_{0}\), there is a segment \(\{\xi h(t),t\leq K\}\) of a closed horocycle approximating \(\{ph(t_{0}+t),0\leq t\leq K\}\) of order \(\delta\), in the sense that_ \[\forall 0\leq t\leq K:\quad d_{X}\left(ph(t_{0}+t),\xi h(t)\right)\leq\delta.\] _The period \(P=P(t_{0},p)\) of this closed horocycle is at most \(P\ll r\), where \(r=T\exp(-\mathrm{dist}(g_{\log T}(p)))\) is as in Theorem 1.1._ _Moreover, one can assure \(P\gg\eta^{2}r\) for some \(\eta>0\) by weakening the bound on \(I_{0}\) to \(|I_{0}|\leq\max\left(\delta^{-1}K^{2},\eta T\right)\)._ Proof.: The quantity \(r\) will play no role in the proof; we show that the period of the closed horocycles is bounded by \(P\ll y_{T}^{-1}\) and use Proposition 4.2. Recall that \(y_{T}\) is a maximum over the different cusps and the elements in \(\Gamma\). Let \(\sigma_{i}\) be the rotation with \(\sigma_{i}.r_{i}=\infty\) corresponding to the cusp \(r_{i}\) maximizing \(y_{T}\), that is such that \(y_{T}=y_{i}^{T}\). The rest of the approximation has nothing to do with \(\Gamma\), but is just an observation about approximating horocycle pieces by horizontal lines in \(PSL_{2}(\mathbb{R})\). The period only comes in because in the coordinate system induced by \(\sigma_{i}\), the height of the horizontal line is the same as the period of the horocycle in \(\Gamma\backslash G\). Let \(g\) be a representative of \(p\) attaining the supremum in the definition of \(y_{i}^{T}\). The horocycle segment \(\{ph(t),0\leq t\leq T\}\) is then a circle segment in the modular plane as sketched in Figure 4. Write \[\sigma_{i}g=:\begin{pmatrix}a&b\\ c&d\end{pmatrix}\] and express points on the circle in term of its peak \[l:=\sigma_{i}gh\left(-\frac{d}{c}\right)=:(\alpha+iR,-i).\] In the Iwasawa decomposition, this is (see (2.3) in [9]) \[lh(s)=h\left(\alpha-\frac{Rs}{s^{2}+1}\right)a\left(\frac{R}{s^{2}+1}\right)k( -\mathrm{arccot}\ s).\] Figure 3. An overview of the definitions, to the left when \(\sigma_{i}g\) and to the right when \(\sigma_{i}gh(T)\) minimizes the imaginary part. Given some \(s\), we will approximate the horocycle segment \(\{lh(s+t),t\leq K\}\) with the periodic horocycle segment \(\{g_{0}h(t),t\leq K\}\), where \(g_{0}=:h(x_{0})a(y_{0})\) lies over the same point in the modular plane as \(gh(s)\) and the vector of \(g_{0}\) points straight up. The horocycle starting in \(g_{0}\) is then a horizontal line moving right. It is thus only left to show that for all but a few exceptional \(s\), which will be the ones in an interval around \(0\), this approximation is good. To see this, fix some \(0\leq t\leq K\) and compare \[lh(s+t)=h\left(\alpha-\frac{R(s+t)}{(s+t)^{2}+1}\right)a\left(\frac{R}{(s+t)^ {2}+1}\right)k(-\text{arccot }(s+t))\] with \[g_{0}h(t)=h\left(\alpha-\frac{Rs}{s^{2}+1}\right)a\left(\frac{R}{s^{2}+1} \right)h(t)=h\left(\alpha-\frac{R(s-t)}{s^{2}+1}\right)a\left(\frac{R}{s^{2}+1 }\right).\] Firstly, note that \[|\text{arccot}(s+t)|\ll\left|\frac{1}{s+t}\right|\leq\delta\] provided that \(|s|\geq\delta^{-1}K\). Secondly, note that \[d_{G}\left(a\left(\frac{R}{(s+t)^{2}+1}\right),a\left(\frac{R}{s^{2}+1} \right)\right)=\left|\log\left(\frac{(s+t)^{2}+1}{s^{2}+1}\right)\right|\ll\delta,\] for any \(t\leq K\) provided that \(|s|\geq\delta^{-1}K\) because \[\frac{d}{dt}\log\left(\frac{(s+t)^{2}+1}{s^{2}+1}\right)=\frac{2(s+t)}{(t+s)^ {2}+1}\ll\frac{1}{|s|}\leq\delta K^{-1}.\] Remembering the left-invariance of the metric and using the triangle inequality, this implies that \[d_{G}\left(lh(s+t),h\left(\alpha-\frac{R(s+t)}{(s+t)^{2}+1}\right)a\left( \frac{R}{s^{2}+1}\right)\right)\ll\delta.\] Finally, \[d_{G}\left(h\left(\alpha-\frac{R(s+t)}{(s+t)^{2}+1}\right)a\left( \frac{R}{s^{2}+1}\right),g_{0}h(t)\right)\] \[=\frac{s^{2}+1}{R}d_{G}\left(h\left(\alpha-\frac{R(s+t)}{(s+t)^{2} +1}\right),h\left(\alpha-\frac{R(s-t)}{s^{2}+1}\right)\right)\] \[=\frac{s^{2}+1}{R}\left|\frac{R(s+t)}{(s+t)^{2}+1}-\frac{R(s-t)}{ s^{2}+1}\right|\] \[=\left|\frac{(s+t)(s^{2}+1)-(s-t)((s+t)^{2}+1)}{(s+t)^{2}+1}\right|\] \[=\left|\frac{st^{2}+t^{3}+2t}{(s+t)^{2}+1}\right|\ll\delta\] where the last inequality holds provided that \(|s|\geq\delta^{-1}K^{2}\). Putting everything together, we deduce that \[d_{G}(lh(s+t),g_{0}h(t))\ll\delta\] provided that \(|s|\geq\delta^{-1}K^{2}\). We then set \(\xi:=\Gamma\sigma_{i}^{-1}g_{0}\), which is a periodic horcycle with period \(y_{0}^{-1}\) as \[\sigma_{i}^{-1}g_{0}h(y_{0}^{-1})=\sigma_{i}^{-1}h(1)g_{0}=\sigma_{i}^{-1}h(1 )\sigma_{i}\sigma_{i}^{-1}g_{0}=\gamma_{i}\sigma_{i}^{-1}g_{0},\] where we recall from the beginning of this section that \(\sigma_{i}\) was the element such that \(\sigma_{i}\gamma_{i}\sigma_{i}^{-1}=h(1)\) and \(\gamma_{i}\in\Gamma\) is the unipotent element inducing the periodicity of the closed horocycles corresponding to \(r_{i}\). For any point \(s\) such that \(lh(s)\in gh([0,T])\), by definition of the fundamental period, \(y_{0}\geq y_{T}\), so that the period of the horocycle \(\{\xi h(t),t\in\mathbb{R}\}\) is indeed bounded by \(y_{T}^{-1}\). Recalling \(lh(s)=\sigma_{i}gh\left(\frac{d}{c}+s\right)\), we can then set the exceptional interval \(I_{0}\) to be \[I_{0}:=\left\{t\in[0,T]:\;\left|\frac{d}{c}+t\right|\leq\delta^{-1}K^{2}\right\}.\] This assures that our estimates hold except for \(t\in I_{0}\) and obviously, \(|I_{0}|\ll\delta^{-1}K^{2}\). Regarding the second point, we want to make sure that for any \(s\) outside of an interval around \(0\) we have that \[\text{Im}(\text{lh(s)})=\frac{R}{s^{2}+1}\ll\eta^{-2}y_{T}.\] Let \(s_{0}\) be such that either \(lh(s_{0})=\sigma_{i}g\) or \(lh(s_{0})=\sigma_{i}gh(T)\), depending on which of the two points minimizes the imaginary part as in the definition of \(y_{T}\); we then have that \(y_{T}=\frac{R}{s_{0}^{2}+1}\). Now, the points \(s\) such that \(lh(s)\) lies on the horocycle orbit \(\sigma_{i}gh([0,T])\), lie either in the interval \([s_{0},s_{0}+T]\) or \([s_{0}-T,s_{0}]\), again depending on which of the two points \(\sigma_{i}g,\sigma_{i}gh(T)\) is minimizing. If \(|s_{0}|\geq 2T\), we have that for any such \(s\) \[\text{Im}(\text{lh(s)})\ll\frac{R}{(|s_{0}|-T)^{2}+1}\ll y_{T}.\] If not, we can impose \(|s|>\eta T\) to assure \[\text{Im}(\text{lh(s)})\ll\frac{R}{\eta^{2}T^{2}+1}\ll\frac{y_{T}T^{2}}{\eta^ {2}T^{2}}\ll y_{T}\eta^{-2}.\] We can set \(I_{0}\) as before, but this time with the condition \(|s_{0}+t|\leq\eta T\). ## 5. Equidistribution of \(\nu\) In this section, we are going to prove Theorem 1.1. We will set \(\theta=\frac{\beta}{40}\), where \(\beta\) is the constant from Theorem 1.4 depending only on the smallest eigenvalue of the Laplacian on \(\Gamma\). In the case \(\Gamma=PSL_{2}(\mathbb{Z})\), this makes \(\frac{1}{2880}\) an admissible value for \(\theta\). The proof is split into different cases depending on the time parameter \(T\). To start off, we will cover the case of good asymptotics, where \(r\gg T^{\frac{1}{20}}\), and \(g_{\log T}(\xi)\) is far away from all cusps. In this case, Theorem 1.4 is sufficient to prove good equidistribution of \(\sum_{n\leq T/s}f(\xi h(sn))\) for all \(s\leq R=T^{\theta}\). This immediately implies that case of Theorem 1.1. **Proposition 5.1**.: _Let \(p\in X\), \(f\) such that \(\|f\|=1\). Then, for all \(T\) such that \(r\gg T^{\frac{1}{20}}\),_ \[\left|\frac{1}{T}\sum_{n\leq T}f(ph(n))\nu(n)-\int f\ d\mu_{X}\right|\ll\frac{ \log\log R}{\log R}.\] Proof.: Assume that \(\int f\ d\mu=0\), which picks up an error term of \(\frac{\log\log R}{\log R}\) from the normalization of \(\nu\) proven in Theorem 2.4. By Theorem 1.4, \[\left|\frac{s}{T}\sum_{1\leq sj\leq T}f(ph(sj))\right|\leq s^{\frac{1}{2}}r^{ -\frac{\beta}{2}}.\] Unboxing the definition of \(\nu\), we find \[\left|\frac{1}{T}\sum_{n\leq T}f(ph(n))\nu(n)\right|=\left|\sum_{n\leq T}\sum_ {\begin{subarray}{c}e,d\leq R\\ e,d|n\end{subarray}}\frac{\mu(e)\mu(d)f(ph(n))}{T\log R}\log\left(\frac{R}{d} \right)\log\left(\frac{R}{e}\right)\right|\] \[\leq\sum_{e,d\leq R}\frac{\log R}{[d,e]}\left|\frac{[d,e]}{T}\sum_{n\leq\frac {T}{[e,d]}}f(ph([e,d]n))\right|\leq\sum_{e,d\leq R}\frac{\log R}{\sqrt{[d,e]}} r^{-\frac{\beta}{2}}\] \[\leq r^{-\frac{\beta}{2}}\sum_{m\leq R}\sum_{e,d\leq\frac{R}{m}}\frac{\log R }{\sqrt{edm}}\ll r^{-\frac{\beta}{2}}\sqrt{R}\log R,\] where we ordered the terms according to their greatest common divisor \(m\). In the case that \(r\ll T^{\frac{1}{20}}\), we will use Lemma 1.3 to reduce to closed horocycles of small period and use Theorem 1.4 together with the Siegel-Walfisz type Theorem 2.4 to conclude the proof. Proof of Theorem 1.1.: Let \(p\in X\) and \(T\) be given such that \(r\ll T^{\frac{1}{20}}\). Assume that \(\|f\|=1\) and fix some \(\delta>0\) to be determined later. Set \(K:=T^{\frac{1}{3}}\). Apply Lemma 1.3 to split the interval \([0,T]\) into intervals \([t_{j},t_{j}+K]\) such that for all but on a \(\delta\) proportion of them, \(t\leq K\) is at distance at most \(\delta\) from \(\{\xi_{j}h(t),t\leq K\}\), where this is a closed horocycle of period \(P_{j}\) with \(\delta^{2}r\ll P_{j}\ll r\). We know by Strombergsson's result ([11]) or from Theorem 1.4 that \[\left|\frac{1}{T}\sum_{n\leq T}f(ph(n))\nu(n)-\int f\;d\mu_{X}\right|\] \[\leq\left|\frac{1}{T}\sum_{n\leq T}f(ph(n))\nu(n)-\frac{1}{T} \int_{0}^{T}f(ph(t))\;dt\right|+O(r^{-\beta})\] \[\ll O(r^{-\beta}+\delta)+\frac{K}{T}\sum_{j}\frac{1}{K}\left|\sum_ {n\leq K}f(\xi_{j}h(n))\nu(n)-\int_{0}^{K}f(\xi_{j}h(t))\;dt\right|.\] Fix some \(j\) and set \(y:=P_{j}^{-1}\). Set \(F(t):=f(\xi_{j}h(ty^{-1}))\), which is a \(1\)-periodic function and is \(y^{-1}\)-Lipschitz by the bounds on \(f\). It thus only remains to show that \[\frac{1}{K}\sum_{n\leq K}F(yn)\nu(n)\] is like \[\int F:=\int_{0}^{1}F(t)dt.\] We want to apply Theorem 2.4 and to do so, we need to get from a function periodic on \([0,1]\) to a function periodic on the integers. To this end, we approximate \(y\) with a rational up to \(R^{3}y^{-3}\). That is, we use the Dirichlet box principle to find \(y^{-1}\leq q\leq R^{3}y^{-3}\) and \((a,q)=1\) such that \[\left|y-\frac{a}{q}\right|<\frac{1}{qR^{3}y^{-3}}.\] Pick some \(M\) and consider how much the function \(n\mapsto F(yn)\) can diverge from a truly \(q\)-periodic function on an interval \(\{m_{0},\ldots,m_{0}+qM\}\). Comparing \(F(y(m_{0}+qM))\) to \(F(ym_{0}+aM)=F(ym_{0})\) for some \(m_{0}\), we get that \[|F(y(m_{0}+qM))-F(ym_{0})|\leq y^{-1}|yqM-aM|\leq\frac{My^{-1}}{R^{3}y^{-3}}.\] This is \(O(y)\) provided that \(qM\leq qR^{3}y^{-1}\). Truncate into intervals of length approximately \(qR^{3}\); as we have just seen, the function \(n\mapsto F(yn)\) is at distance \(O(y)\) from a \(q\)-periodic function on each one. We can thus apply Theorem 2.4 on each interval to deduce that \[\frac{1}{K}\sum_{n\leq K}F(yn)\nu(n)=\frac{q}{\phi(q)K}\sum_{\begin{subarray}{ c}n\leq K\\ (n,q)=1\end{subarray}}F(yn)+O(y)+O\left(\frac{\log\log R}{\log R}\right).\] To show that the sum on the right is like \(\int F\), we need one more claim. **Claim 5.2**.: _Let \(\epsilon=\frac{\beta}{12}\)._ \[\left|\frac{s}{K}\sum_{sn\leq K}F(ysn)-\int F\right|\leq q^{-\epsilon}\] _for all \(s|q\) such that \(s\leq q^{\epsilon}\)._ Before we show the claim, let us see how it allows us to conclude the proof. We use the identity \(1_{m=1}=\sum_{d|m}\mu(d)\) to find \[\sum_{\begin{subarray}{c}n\leq K\\ (n,q)=1\end{subarray}}F(yn) =\sum_{n\leq K}\sum_{d|(n,q)}\mu(d)F(yn)\] \[=\sum_{d|q}\mu(d)\sum_{\begin{subarray}{c}n\leq K\\ d|n\end{subarray}}F(yn)=\sum_{d|q}\mu(d)\sum_{nd\leq K}F(ydn).\] Decomposing the sum and using Claim 5.2, we see that \[\frac{q}{\phi(q)K}\sum_{\begin{subarray}{c}n\leq K\\ (n,q)=1\end{subarray}}F(yn) \leq\frac{q}{\phi(q)}\sum_{\begin{subarray}{c}d|q\\ d<q^{\epsilon}\end{subarray}}\frac{1}{d}\left|\frac{d}{K}\sum_{dn\leq K}F(ydn) \right|+\frac{q}{\phi(q)}\sum_{\begin{subarray}{c}d|q\\ d\geq q^{\epsilon}\end{subarray}}\frac{\|F\|_{\infty}}{d}\] \[\leq 2\frac{q^{1-\epsilon}\tau(q)}{\phi(q)}.\] By standard asymptotics of \(\phi\) and \(\tau\) (see for example [6]), the right-hand side is \[O(q^{-\frac{6\epsilon}{7}})=O(y^{\frac{\beta}{14}})=O\left(\delta^{-\frac{ \beta}{7}}r^{\frac{\beta}{14}}\right).\] We can choose \(\delta=r^{-\frac{1}{5}}\) to get the desired conclusion. It thus only remains to show Claim 5.2. To prove Claim 5.2, we divide into two cases. Firstly, in the case that \(q\leq y^{-3}\), we apply Strombergsson's result (or Theorem 1.4) to the periodic horocycle \(\xi_{j}h(t)\), for which \(r(\xi_{j},T)=y^{-1}\) for any \(T\), to see \[\int F=y\int_{0\leq t\leq y^{-1}}f(\xi_{j}h(t))dt=\int\ f\ d\mu_{X}+O(y^{\beta }).\] Using this and applying Theorem 1.4 to the same periodic horocycle piece, we see \[\forall s\leq y^{\frac{\beta}{4}}:\ \ \left|\frac{s}{K}\sum_{1\leq j\leq K/s}F( ysj)-\int F\right|\ll s^{\frac{1}{2}}y^{\frac{\beta}{2}}\ll y^{\frac{\beta}{4}}.\] As \(q^{\epsilon}\leq y^{\frac{\beta}{4}}\), we deduce Claim 5.2 in this case. For the second case, assume that \(q\geq y^{-3}\). Roughly speaking, in this case there are so many distinct points of the form \(sn\frac{a}{q}\) in the interval \([0,1]\) that they cannot help being dense enough to approximate \(\int F\) by force. We split \([0,K]\) into intervals \(I\) of length \(qR^{3}y^{-1}\). Fix \(s\leq q^{\epsilon}\) and set \(q^{\prime}:=\nicefrac{{q}}{{s}}\). Fix an interval \(I\) and call its left endpoint \(st_{0}\). We note that as for any \(n\) such that \(sn\in I\), \[\left|F(ysn)-F\left(yst_{0}+\frac{sa}{q}(n-t_{0})\right)\right| \leq y^{-1}\left|y-\frac{a}{q}\right|\left|sn-st_{0}\right|\] \[\leq\frac{y^{-1}|I|}{qR^{3}y^{-3}}\leq y;\] set \(x_{0}:=st_{0}\left(y-\frac{a}{q}\right)\) and note that then \[\frac{s}{|I|}\sum_{sn\in I}F(ysn)=O(y)+\frac{s}{|I|}\sum_{sn\in I} F\left(x_{0}+n\frac{as}{q}\right)\] \[=O(y)+O\left(\frac{sq^{\prime}}{|I|}\right)+\frac{s}{q}\sum_{n \leq q^{\prime}}F\left(x_{0}+n\frac{a}{q^{\prime}}\right),\] where we use in the second line that the function \(F(x_{0}+\frac{san}{q})\) is \(q^{\prime}\) periodic in \(n\). The number \(a\) is coprime to \(q^{\prime}\), so it plays no role and can be dropped; we then only have to evaluate \[\frac{1}{q^{\prime}}\sum_{n\leq q^{\prime}}F\left(x_{0}+\frac{n}{q^{\prime}} \right).\] But for any \(t\in(0,1)\) and any \(n\), \[\left|F\left(x_{0}+\frac{n}{q^{\prime}}\right)-F\left(x_{0}+\frac{n+t}{q^{ \prime}}\right)\right|\leq y^{-1}\frac{1}{q^{\prime}}\leq y^{-1}y^{3(1-\epsilon )}\leq y,\] which implies that \[\frac{1}{q^{\prime}}\sum_{n\leq q^{\prime}}F\left(x_{0}+\frac{n}{ q^{\prime}}\right)=O(y)+\frac{1}{q^{\prime}}\sum_{n\leq q^{\prime}}\int_{0}^{1} F\left(x_{0}+\frac{n+t}{q^{\prime}}\right)\ dt\] \[=O(y)+\frac{1}{q^{\prime}}\int_{0}^{q^{\prime}}F\left(\frac{t}{q^ {\prime}}\right)\ dt=O(y)+\int F\] This shows Claim 5.2 also in the second case, which, as we have seen, concludes the proof of Theorem 1.1.

2305.03669

Crossing Symmetric Dispersion Relations without Spurious Singularities

Recently, there has been renewed interest in a crossing-symmetric dispersion relation from the 1970s due to its implications for both regular quantum field theory and conformal field theory. However, this dispersion relation introduces nonlocal spurious singularities and requires additional locality constraints for their removal, a process that presents considerable technical challenges. In this Letter, we address this issue by deriving a new crossing-symmetric dispersion relation that is free of spurious singularities, resulting in a compact form of the contact terms in crossing-symmetric blocks. Our results establish a solid foundation for the Polyakov bootstrap in conformal field theories and the crossing-symmetry S-matrix bootstrap in quantum field theories.

Chaoming Song

2023-05-05T16:38:44Z

http://arxiv.org/abs/2305.03669v2

# Crossing Symmetric Dispersion Relations without Spurious Singularities ###### Abstract Recently, there has been renewed interest in a crossing-symmetric dispersion relation from the 1970s due to its implications for both regular quantum field theory and conformal field theory. However, this dispersion relation introduces nonlocal spurious singularities and requires additional locality constraints for their removal, a process that presents considerable technical challenges. In this Letter, we address this issue by deriving a new crossing-symmetric dispersion relation that is free of spurious singularities, resulting in a compact form of the contact terms in crossing-symmetric blocks. Our results establish a solid foundation for the Polyakov bootstrap in conformal field theories and the crossing-symmetry S-matrix bootstrap in quantum field theories. **Introduction:** The recent revival in crossing-symmetric dispersion relations [1; 2] has sparked considerable interest in both quantum field theory (QFT) [3] and conformal field theory (CFT) [4; 5]. In contrast to traditional \(t\)-fixed dispersion relations, which display symmetry in only two channels [6; 7], crossing-symmetric dispersion relations impose no additional constraints and are in perfect accord with Feynman diagram expansions. Within the CFT domain, four-point correlation functions must adhere to crossing symmetry constraints. Numerical bootstrap typically enforces this crossing symmetry on the conformal block expansion. Alternately, Polyakov introduced a conformal bootstrap using crossing-symmetric blocks [8], an approach that has recently proven effective in Mellin space [9; 10; 11]. This method employs a basis connected to exchange Witten diagrams, although contact terms remain undetermined [12; 13]. Resolving these terms continues to pose a considerable challenge [14; 15; 16; 17; 18; 19; 20; 21]. Gopakumar et al.[4] recently observe that these contact term ambiguities are fully determined using a crossing-symmetric dispersion relation, initially developed by Auberson and Khuri (AK) [1] and later revisited by Sinha and Zahed [3]. However, the AK dispersion relation presents spurious singularities that violate locality. Therefore, additional locality constraints are manually imposed to remove these unphysical terms. In theory, after removing these singularities, crossing-symmetric dispersion relations allow for a Feynman/Witten diagram expansion and entirely fix the contact terms. In line with this approach, a closed form of the contact terms has been proposed [22]. Nevertheless, the complexity of analyzing singularities restricts its practical application to lower spins, thereby complicating the implementation of the Polyakov bootstrap. In this paper, we propose a new dispersion relation that manifests both crossing symmetry and locality. We discover a novel approach to directly remove nonlocal singularities, resulting in a closed form of the singularity-free dispersion relation. Consequently, we present the Feynman/Witten expansion of crossing-symmetric blocks, providing explicit determinations of all contact terms. Furthermore, we develop the full dispersion relation without assuming crossing-symmetric amplitudes, enabling the application of our findings to a wide range of problems. For instance, our work establishes a solid foundation for the Polyakov bootstrap, where the only remaining non-trivial constraint is the Polyakov condition [8; 10]. Moreover, our approach yields a novel functional sum rule for the crossing-symmetric bootstrap, eliminating the need for power series expansions. **Singularity-free dispersion relation:** We begin with the shifted Mandelstam variables \(s_{1}=s-\mu/3\), \(s_{2}=t-\mu/3\), and \(s_{3}=u-\mu/3\) satisfying the constraint \(s_{1}+s_{2}+s_{3}=0\), where \(s\), \(t\), and \(u\) are the usual Mandelstam variables. For regular QFT, we have \(\mu=4m^{2}\), while for CFT, we have \(\mu=2\Delta_{\phi}\). We consider hypersurfaces \((s_{1}-a)(s_{2}-a)(s_{3}-a)=-a^{3}\), and rewrite \(s_{k}(z,a)=a-a(z-z_{k})^{3}/(z^{3}-1)\), where \(z_{k}\) are cube roots of unity [1]. Note that we can express \(a=y/x\), where \(x\equiv-(s_{1}s_{2}+s_{2}s_{3}+s_{3}s_{1})\) and \(y\equiv-s_{1}s_{2}s_{3}\). Instead of a dispersion relation in \(s\) for fixed \(t\), we can write down a twice subtracted dispersion relation in the variable \(z\), for fixed \(a\). The full crossing-symmetric dispersion relation is quite involved, and we refer the readers to Ref. [1] for more details. A full singularity-free dispersion relation is set to be proposed in a subsequent section. Our discussion below primarily focuses on the completely crossing-symmetric scattering amplitudes, such as pion-pion scattering in QFT or the Mellin amplitude for a four-point correlation of identical scalars in CFT [23; 24]. For a crossing-symmetric amplitude \(\mathcal{M}^{(s)}\), the dispersion relation simplifies dramatically in terms of \(\mathbf{s}\equiv\{s_{1},s_{2},s_{3}\}\), as \[\mathcal{M}^{(s)}(\mathbf{s})=\alpha_{0}+\frac{1}{\pi}\int \frac{d\sigma}{\sigma}\mathcal{A}\left(\sigma,s_{\pm}\left(\sigma, \frac{a}{\sigma-a}\right)\right) \tag{1}\] \[\times H(\sigma;\mathbf{s}),\] where \(\mathcal{A}(s_{1},s_{2})\) is the s-channel discontinuity, symmetric under the exchange of \(t\) and \(u\) channels, i.e., \(\mathcal{A}(s_{1},s_{2})=\mathcal{A}(s_{1},s_{3})\). The constant \(\alpha_{0}\equiv\mathcal{M}^{(s)}(0,0)\), and the func tions \(H(\sigma;\mathbf{s})\) and \(s_{\pm}(\sigma,\eta)\) are defined as: \[H(\sigma;\mathbf{s}) \equiv\frac{s_{1}}{\sigma-s_{1}}+\frac{s_{2}}{\sigma-s_{2}}+\frac{ s_{3}}{\sigma-s_{3}},\] \[s_{\pm}(\sigma,\eta) \equiv\sigma\frac{-1\pm\sqrt{1+4\eta}}{2},\] where \(s_{+}s_{-}=-\sigma^{2}\eta\), and \(s_{+}+s_{-}=-\sigma\). Setting \(\eta=a/(\sigma-a)\) and \(s_{1}=\sigma\) solves \(s_{2}=s_{\pm}\) and \(s_{3}=s_{\mp}\) from the definition above. Note that \(\mathcal{A}(\sigma,s_{+})=\mathcal{A}(\sigma,s_{-})\), and thus the validity of Eq. (1) is independent of the choice of \(s_{+}\) or \(s_{-}\). Equation (1) is manifestly crossing symmetric, allowing the scattering amplitude \[\mathcal{M}^{(s)}(\mathbf{s})=\sum_{p,q}\mathcal{M}^{(s)}_{p,q}x^{p}y^{q}, \tag{2}\] to be expanded in terms of crossing-symmetric variables \(x\) and \(y\). However, the AK dispersion relation (1) involves the variable \(a\) and, therefore, leads to negative powers of \(x\) in the expansion (2). These spurious singularities are known to violate locality [3]. To obtain the physical scattering amplitude, additional locality constraints must be imposed to enforce the vanishing of these non-physical terms in Eq. (2). Formally, a singularity-free dispersion relation requires computing the regular part \[R\equiv\mathcal{R}\left\{\mathcal{A}\left(\sigma,s_{\pm}\left(\sigma,\frac{a}{ \sigma-a}\right)\right)H(\sigma;\mathbf{s})\right\}, \tag{3}\] where \(\mathcal{R}\{\ldots\}\) denotes a formal regularization with the negative power of \(x\) terms being removed. To obtain a closed form of the regular part \(R\), we first rewrite \(H(\sigma,\mathbf{s})=(2\sigma^{3}-y)H_{0}(\sigma,\mathbf{s})-2\), where \[H_{0}(\sigma,\mathbf{s})\equiv\frac{1}{(\sigma-s_{1})(\sigma-s_{2})(\sigma-s _{3})}=\frac{1}{\sigma^{3}+y-\sigma x},\] corresponds to the poles. Notice that multiplying the factor \(a\) with a regular function \(f(x,y)\), \[\hat{a}f(x,y)\equiv\mathcal{R}\{af(x,y)\}\] acts as a lowering operator \(\hat{a}|n\rangle=y|n-1\rangle\), with \(\hat{a}|0\rangle=0\), where \(|n\rangle\equiv x^{n}\) denotes the \(n\)-th power of \(x\). Specifically, we obtain \[\hat{a}^{n}H_{0} =\frac{1}{\sigma^{3}+y}\sum_{m=0}^{\infty}\left(\frac{\sigma}{ \sigma^{3}+y}\right)^{m}\hat{a}^{n}x^{m}\] \[=\frac{1}{\sigma^{3}+y}\sum_{m=n}^{\infty}\left(\frac{\sigma}{ \sigma^{3}+y}\right)^{m}y^{n}x^{m-n}\] \[=\left(\frac{\sigma y}{\sigma^{3}+y}\right)^{n}H_{0},\] which suggests that \[F(\hat{a},y)H_{0}(\sigma,\mathbf{s})=F\left(\frac{\sigma y}{\sigma^{3}+y},y \right)H_{0}(\sigma,\mathbf{s}) \tag{4}\] for any function \(F(a,y)\) admitting a Taylor expansion in terms of \(a\). Substituting Eq. (4) into Eq. (3) and noting \(F(\hat{a},y)f(y)=F(0,y)f(y)\) lead to \[R=\mathcal{A}\left(\sigma,s_{\pm}\left(\sigma,y/\sigma^{3}\right)\right)(2 \sigma^{3}-y)H_{0}(\sigma,\mathbf{s})-2\mathcal{A}\left(\sigma,s_{\pm}\left( \sigma,0\right)\right).\] Therefore, we obtain the singularity-free (SF) dispersion relation \[\mathcal{M}^{(s)}(\mathbf{s})=\alpha_{0}+\frac{1}{\pi}\int\frac{d\sigma}{ \sigma}\left(\frac{(2\sigma^{3}+s_{1}s_{2}s_{3})\mathcal{A}\left(\sigma,s_{\pm }\left(\sigma,-s_{1}s_{2}s_{3}/\sigma^{3}\right)\right)}{(\sigma-s_{1})(\sigma -s_{2})(\sigma-s_{3})}-2\mathcal{A}(\sigma,0)\right), \tag{5}\] where the locality constraints are automatically satisfied, as we will show explicitly in the next section. **Block expansion and contact terms:** To facilitate the analysis of the s-channel discontinuity, it is common practice to expand it in terms of the partial waves with _even_ spins, as \[\mathcal{A}(s_{1},s_{2})=\sum_{\ell}\int d\lambda f_{\ell}(\sigma,\lambda)Q_{ \lambda,\ell}(s_{1},s_{2}),\] where the partial wave \(Q_{\lambda,\ell}(s_{1},s_{2})=Q_{\lambda,\ell}(s_{1},s_{3})\) is a symmetric polynomial of order \(\ell\) that is invariant under the exchange of the _ut_ channels, and the spectrum \(f_{\ell}(\sigma,\lambda)\) encodes scattering data. For QFT, we express \(Q_{0,\ell}(s_{1},s_{2})\equiv(s_{1}-2\mu/3)^{\ell}C_{\ell}^{(\frac{d-3}{2})} \left(\frac{s_{2}-s_{3}}{s_{1}-2\mu/3}\right)\) in terms of Gegenbauer polynomials, and \(f_{\ell}(\sigma,\lambda)=(\sigma-2\mu/3)^{-\ell}\Phi(\sigma)(2\ell+2\alpha) \alpha_{\ell}(\sigma)\delta(\lambda)\) with \(\Phi(\sigma)\equiv\Psi((d-3)/2)(\sigma+\mu)^{1/2}/(\sigma-2\mu/3)^{(d-3)/2}\) with a real number \(\Psi((d-3)/2)\). For CFT, we express \(Q_{\lambda,\ell}(\mathbf{s})=P_{\Delta-d/2,\ell}(s_{1}+2\Delta_{\phi}/3,s_{2}- \Delta_{\phi}/3)\) in terms of Mack polynomials, and \(f_{\ell}(\sigma,\lambda)\equiv\sum_{\Delta,k}C_{\Delta,\ell}N_{\Delta,k}\delta (\sigma-s_{k})\delta(\lambda-\Delta)\) encodes the operator product expansion (OPE) data. The scattering amplitude can also be expressed as \[\mathcal{M}^{(s)}(\mathbf{s})=\alpha_{0}+\frac{1}{\pi}\sum_{\ell=0}^{\infty} \int d\sigma d\lambda f_{\ell}(\sigma,\lambda)M_{\lambda,\ell}(\sigma;\mathbf{s}),\] where \(M_{\lambda,\ell}(\sigma;\mathbf{s})\) are scattering blocks. Comparing AK dispersion relation (1), we obtain the Dyson block [3], \[M_{\lambda,\ell}^{(D)}=\frac{1}{\sigma}Q_{\lambda,\ell}\left(\sigma,s_{\pm}\left( \sigma,\frac{a}{\sigma-a}\right)\right)H(\sigma;\mathbf{s}), \tag{6}\] which contains spurious singularities. By contrast, our dispersion relation (5) leads to the singularity-free block \[M_{\lambda,\ell}^{(SF)}=\frac{(2\sigma^{3}-y)Q_{\lambda,\ell}(\sigma,s_{\pm}( \sigma,y/\sigma^{3}))}{\sigma(\sigma-s_{1})(\sigma-s_{2})(\sigma-s_{3})}- \frac{2}{\sigma}Q_{\lambda,\ell}(\sigma,0). \tag{7}\] To show explicitly SF block \(M_{\lambda,\ell}^{(SF)}\) removes spurious singularities present in the Dyson block \(M_{\lambda,\ell}^{(D)}\), we take the QFT case as an example. We start with the Gegenbauer polynomials \(C_{\ell}^{(\frac{d-2}{3})}(\sqrt{\xi})\) where \(\xi=(s_{+}(\sigma,\eta)-s_{-}(\sigma,\eta))^{2}/(\sigma_{1}-2\mu/3)^{2}=\xi_ {0}(1+4\eta)\), where \(\xi_{0}\equiv\sigma^{2}/(\sigma-2\mu/3)^{2}\). We set \(\eta=a/(\sigma-a)\) and expand the Gegenbauer polynomials around \(\xi_{0}\), giving us [1; 3] \[M_{\lambda,\ell}^{(D)}=\frac{1}{\sigma}\sum_{n,m=0}^{\infty}\mathcal{B}_{n,m} ^{(\ell)}x^{n}(y/x)^{m},\] where \(p_{\ell}^{(k)}\equiv\partial_{\xi}^{k}C^{(d-3)/2}(\sqrt{\xi}_{0})\), and \[\mathcal{B}_{n,m}^{(\ell)}=\sum_{k=0}^{m}\frac{p_{\ell}^{(k)}(4\xi_{0})^{k}(3 k-m-2n)(-n)_{m}}{\pi\sigma^{2n+m}k!(m-k)!(-n)_{k+1}}.\] Similarly, expanding the Gegenbauer polynomials around \(\xi_{0}\) with \(\eta=y/\sigma^{3}\) leads to \[M_{\lambda,\ell}^{(SF)}=\frac{1}{\sigma}\sum_{n,m=0}^{\infty}\mathcal{C}_{n,m }^{(\ell)}x^{n}y^{m},\] where \[\mathcal{C}_{n,m}^{(\ell)}=\sum_{k=0}^{m}\frac{p_{\ell}^{(k)}(4\xi_{0})^{k}(- 1)^{m-k}(2n+3(m-k))}{\pi\sigma^{2n+3m}n!k!(m-k)!}\] It is easy to verify that \(\mathcal{C}_{n,m}^{(\ell)}=\mathcal{B}_{n+m,m}^{(\ell)}\) for \(n,m\geq 0\), indicating that the regular part of the Dyson blocks matches with the SF blocks, as expected. However, the Dyson blocks have spurious singularities with a negative power of \(x\) when \(n<m\), which is absent in our SF blocks. A similar deviation can be obtained for general partial waves \(Q_{\lambda,\ell}\). The singularity-free (SF) block provides a block expansion for the amplitude that directly relates to the usual Feynman and Witten diagrammatic expansions for QFT and CFT, respectively. To see this, we will show below that the SF block can be written as a summation of exchange and contact terms, as follows: \[M_{\lambda,\ell}^{(SF)}(\sigma;\mathbf{s})=\sum_{i=1}^{3}M_{\lambda,\ell}^{(i )}(\sigma;\mathbf{s})+M_{\lambda,\ell}^{(c)}(\sigma;\mathbf{s}), \tag{8}\] where the exchange term of channel \(i\) is given by \[M_{\lambda,\ell}^{(i)}(\sigma;\mathbf{s})=Q_{\lambda,\ell}(s_{i},s_{i+1}) \left(\frac{1}{\sigma-s_{i}}-\frac{1}{\sigma}\right),\] for \(i=1,2,3\) with the cyclic condition \(i+1=1\) for \(i=3\). The contact terms \(M_{\lambda,\ell}^{(c)}(\sigma;\mathbf{s})\) involve polynomials of \(s_{i}\)'s, whose explicit form is previously only known for few lower order terms. We substitute Eq. (7) into Eq. (8), obtaining \[M_{\lambda,\ell}^{(c)}(\sigma;\mathbf{s})=\frac{1}{\sigma}\sum_{i=1}^{3}\frac {s_{i}\Delta Q_{\lambda,\ell}^{(i)}}{(\sigma-s_{i})}+\frac{2}{\sigma}\Delta Q_{ \lambda,\ell}^{(0)}, \tag{9}\] where \[\Delta Q_{\lambda,\ell}^{(i)} \equiv Q_{\lambda,\ell}(\sigma;s_{\pm}(\sigma,y/\sigma^{3}))-Q_{ \lambda,\ell}(s_{i};s_{\pm}(s_{i},y/s_{i}^{3})),\] \[\Delta Q_{\lambda,\ell}^{(0)} \equiv Q_{\lambda,\ell}(\sigma,s_{\pm}(\sigma,y/\sigma^{3}))-Q_{ \lambda,\ell}(\sigma,0),\] are polynomials. To show the the contact term \(M_{\lambda,\ell}^{(c)}\) are also polynomials, we notice that the symmetry of ut channels allows us to expand \(Q_{\lambda,\ell}(s_{1};s_{2})=\sum_{n+2m\leq l}q_{nm}s_{1}^{n}(s_{2}s_{3})^{m}\), which implies \[Q_{\lambda,\ell}(\sigma;s_{\pm}(\sigma,y/\sigma))=\sum_{n+2m\leq l}q_{nm} \sigma^{n}(s_{1}/\sigma)^{m}(s_{2}s_{3})^{m}.\] Thus, \[\Delta Q_{\lambda,\ell}^{(i)}=\sum_{n,m}q_{nm}\sigma^{n}\left((s_{i}/\sigma)^{ m}-(s_{i}/\sigma)^{n}\right)(s_{i+1}s_{i+2})^{m},\] where the term \((s_{i}/\sigma)^{m}-(s_{i}/\sigma)^{n}\) must divide \(s_{i}/\sigma-1\) and thus cancel the poles in Eq. (9). More explicitly, we find that \[\Delta Q_{\lambda,\ell}^{(i)}=(s_{i}-\sigma)\sum_{n,m}P_{n,m}(\sigma)s_{i}^{n}( s_{i+1}s_{i+2})^{m},\] where \[P_{n,m}(\sigma)=\begin{cases}\sum_{k=0}^{n}q_{km}\sigma^{k-n-1},&0\leq n<\min (m,\ell-2m+1)\\ \sum_{k=0}^{\ell-2m}q_{km}\sigma^{k-n-1},&\ell-2m\leq n\leq m-1,\\ -\sum_{k=n+1}^{\ell-2m}q_{km}\sigma^{k-n-1},&m\leq n\leq\ell-2m-1.\end{cases}\] Substituting into Eq. (9), we obtain the contact term \[M_{\lambda,\ell}^{(c)} =\frac{2}{\sigma}\left(Q_{\lambda,\ell}(\sigma,s_{\pm}(\sigma,y/ \sigma^{3}))-Q_{\lambda,\ell}(\sigma,0)\right)\] \[-\frac{1}{\sigma}\sum_{n,m}P_{n,m}(\sigma)\left(\sum_{i=1}^{3}s_{i }^{n+1}(s_{i+1}s_{i+2})^{m}\right), \tag{10}\] which are manifestly crossing-symmetric polynomials. Note that the summation over indices \(n\) and \(m\) is across all non-zero \(P(n,m)\) terms, i.e., \(0\leq m\leq\ell/2\) and \(n+2m\leq 3\ell/2-1\). **Singular block and sum rules:** Since the SF block corresponds to the regular part of the Dyson block, we can decompose \[M_{\lambda,\ell}^{(D)}(\sigma;\mathbf{s})=M_{\lambda,\ell}^{(SF)}(\sigma; \mathbf{s})+M_{\lambda,\ell}^{(S)}(\sigma;\mathbf{s}),\] where \(M_{\lambda,\ell}^{(S)}(\sigma;\mathbf{s})\) refers to the corresponding singular part, given by \[M_{\lambda,\ell}^{(S)} =\frac{Q_{\lambda,\ell}(\sigma,s_{\pm}(\sigma,y/\sigma^{3})-Q_{ \lambda,\ell}(\sigma,s_{\pm}(\sigma,a/(\sigma-a)))}{y/\sigma^{3}-a/(\sigma-a)}\] \[\times\frac{a(2\sigma x-3y)}{\sigma^{4}(\sigma-a)}-\frac{2}{ \sigma}(Q_{\lambda,\ell}(\sigma,s_{\pm}(\sigma,y/\sigma^{3})-Q_{\lambda,\ell}( \sigma,0)).\] Note since \(Q_{\lambda,\ell}(\sigma,s_{\pm}(\sigma,\eta))\) is polynomial of \(\eta\), the term in the first line is the difference operator acting on \(Q_{\lambda,\ell}\) between \(\eta=y/\sigma^{3}\) and \(\eta=a/(\sigma-a)\), and thus is also a polynomial of these two terms. Therefore, the first term involves positive powers of \(y/x\) except for the zeroth-order term \((y/x)2\sigma x\), which cancels the last term in the above equation. Consequently, only terms with negative powers of \(x\) remain in \(M_{\lambda,\ell}^{(S)}\), as expected. Since both Dyson and SF blocks lead to the same amplitude, the contribution from the singular part needs to be canceled: \[\sum_{\ell}\int d\sigma d\lambda f_{\ell}(\sigma,\lambda)M_{\lambda,\ell}^{(S )}(\sigma;\mathbf{s})=0, \tag{11}\] which imposes a constraint on the spectrum \(f_{\ell}(\sigma,\lambda)\). For instance, for QFT, Equation (11) requires the cancellation of power series contributions of \(\mathcal{B}_{n,m}^{(\ell)}x^{n-m}y^{m}\) with negative powers of \(x\), i.e., \(n<m\), generalizing the Froissart bound [3; 25]. For CFT, it appears to connect to the conformal dispersive sum rules [4; 19]. Unlike previous approaches, Eq. (11) provides a single functional sum rule without involving series expansion. **Full dispersion relation:** Our approach extends to general scattering amplitudes without assuming the complete crossing symmetry. The corresponding full dispersion relation should link the scattering amplitude \(\mathcal{M}(\mathbf{s})\) to \(s\), \(u\), and \(t\)-channel discontinuities, denoted as \(\mathcal{A}_{i}(\mathbf{s})\) for \(i=1,2,3\). Furthermore, \(\mathcal{M}(\mathbf{s})\) is not merely a function of \(x\) and \(y\), but also of a linear combination of \(s_{i}\). In addition, an antisymmetric part exists [26], characterized in terms of \(w=-(s_{1}-s_{2})(s_{2}-s_{3})(s_{3}-s_{1})\). Note that the algebraic curve \(w^{2}=4x^{3}-27y^{2}\) suggests that any power of \(w\) higher than first order will be absorbed in a combination of \(x\) and \(y\). Using an approach similar to the one presented above, we derive the full SF dispersion relation: \[\mathcal{M}(\mathbf{s})=\alpha_{0}+\sum_{i=1}^{3}\alpha_{i}s_{i} +\frac{1}{2\pi}\sum_{i=1}^{3}\int\frac{d\sigma}{\sigma}\frac{K_{i}^{+}( \mathbf{s},\sigma)\mathcal{A}_{i}\left(\sigma,\tilde{s}_{+}\right)+K_{i}^{-}( \mathbf{s},\sigma)\mathcal{A}_{i}\left(\sigma,\tilde{s}_{-}\right)}{(\sigma- s_{1})(\sigma-s_{2})(\sigma-s_{3})} \tag{12}\] \[-K_{i}^{0+}(\mathbf{s},\sigma)\mathcal{A}_{i}\left(\sigma,s_{+}( \sigma,0)\right)-K_{i}^{0-}(\mathbf{s},\sigma)\mathcal{A}_{i}\left(\sigma,s_{- }(\sigma,0)\right),\] where \(\tilde{s}_{\pm}\equiv s_{\pm}(\sigma,y/\sigma^{3})\), \(K_{i}^{0\pm}(\mathbf{s},\sigma)=\frac{2}{3}+\frac{s_{i}}{\sigma}\pm\frac{s_{i +1}-s_{i+2}}{3\sigma}\), and \[K_{i}^{\pm}(\mathbf{s},\sigma) =\left((2\sigma^{3}-y)\pm\frac{\sigma w}{\tilde{s}_{+}-\tilde{s}_{ -}}\right)\left(\frac{1}{3}+\frac{\sigma^{2}s_{i}}{2(y+\sigma^{3})}\right)\] \[+\left(\sigma^{2}w\pm\frac{(2\sigma^{3}-y)(4y+\sigma^{3})}{ \tilde{s}_{+}-\tilde{s}_{-}}\right)\frac{s_{i+1}-s_{i+2}}{6(y+\sigma^{3})}.\] The constants \(\alpha_{i}\) correspond to the first-order coefficient of \(s_{i}\), with only two being free, enabling us to impose \(\sum_{i=1}^{3}\alpha_{i}=0\). The corresponding SF blocks can be found consequently. The full dispersion relation (12) is substantially more involved. While the full dispersion relation (12) is considerably more involved, it simplifies remarkably for the crossing symmetric and symmetric cases. In the former scenario, the discontinuities across all channels are identical, i.e., \(\mathcal{A}_{i}(\sigma,\tilde{s}_{\pm})=\mathcal{A}(\sigma,\tilde{s}_{\pm})\). Equation (12) reduces to Eq. (5) since all terms cancel after summation except for \((2\sigma^{3}-y)/3\). Likewise, for the crossing antisymmetric case [26], Equation (12) simplifies to \[\mathcal{M}^{(as)}(\mathbf{s})=\frac{w}{\pi}\int\frac{d\sigma}{(\sigma-s_{1})( \sigma-s_{2})(\sigma-s_{3})}\frac{\mathcal{A}(\sigma,\tilde{s}_{+})}{\tilde{s} _{+}-\tilde{s}_{-}},\] which provides the crossing antisymmetric dispersion relation. It is noteworthy that in this case \(\mathcal{A}(\sigma,\tilde{s}_{+})=-\mathcal{A}(\sigma,\tilde{s}_{-})\), thus \(\frac{\mathcal{A}(\sigma,\tilde{s}_{+})}{\tilde{s}_{+}-\tilde{s}_{-}}=\frac{1}{ 2}\frac{\mathcal{A}(\sigma,\tilde{s}_{+})-\mathcal{A}(\sigma,\tilde{s}_{-})}{ \tilde{s}_{+}-\tilde{s}_{-}}\) is a polynomial in terms of \(\tilde{s}_{+}\) and \(\tilde{s}_{-}\), as expected for _odd_ spin contributions. **Discussion:** The singularity-free, crossing-symmetric dispersion relation approach introduced in this paper addresses a long-standing challenge in the nonperturbative exploration of quantum field theories. The proposed crossing-symmetric blocks seamlessly link to Feynman/Witten diagrams, with contact terms being explicitly determined. The null contribution of the singular block leads to a simplified functional sum rule, enhancing existing methods. Furthermore, the full singularity-free dispersion relation lays the groundwork for the Polyakov bootstrap beyond identity operators. This approach also provides remarkable opportunities for numerical S-matrix bootstrap within a broader setup. Undoubtedly, our advancements establish a robust foundation for crossing-symmetric bootstrap applicable to both QFTs and CFTs. ###### Acknowledgements. We express our gratitude to Aninda Sinha and Ahmadullah Zahed for their invaluable discussions and constructive feedback on the manuscript.

2307.15672

Bayesian Time-Series Classifier for Decoding Simple Visual Stimuli from Intracranial Neural Activity

Understanding how external stimuli are encoded in distributed neural activity is of significant interest in clinical and basic neuroscience. To address this need, it is essential to develop analytical tools capable of handling limited data and the intrinsic stochasticity present in neural data. In this study, we propose a straightforward Bayesian time series classifier (BTsC) model that tackles these challenges whilst maintaining a high level of interpretability. We demonstrate the classification capabilities of this approach by utilizing neural data to decode colors in a visual task. The model exhibits consistent and reliable average performance of 75.55% on 4 patients' dataset, improving upon state-of-the-art machine learning techniques by about 3.0 percent. In addition to its high classification accuracy, the proposed BTsC model provides interpretable results, making the technique a valuable tool to study neural activity in various tasks and categories. The proposed solution can be applied to neural data recorded in various tasks, where there is a need for interpretable results and accurate classification accuracy.

Navid Ziaei, Reza Saadatifard, Ali Yousefi, Behzad Nazari, Sydney S. Cash, Angelique C. Paulk

2023-07-28T17:04:06Z

http://arxiv.org/abs/2307.15672v1

Bayesian Time-Series Classifier for Decoding Simple Visual Stimuli from Intracranial Neural Activity ###### Abstract Understanding how external stimuli are encoded in distributed neural activity is of significant interest in clinical and basic neuroscience. To address this need, it is essential to develop analytical tools capable of handling limited data and the intrinsic stochasticity present in neural data. In this study, we propose a straightforward Bayesian time series classifier (BTsC) model that tackles these challenges whilst maintaining a high level of interpretability. We demonstrate the classification capabilities of this approach by utilizing neural data to decode colors in a visual task. The model exhibits consistent and reliable average performance of 75.55% on 4 patients' dataset, improving upon state-of-the-art machine learning techniques by about 3.0 percent. In addition to its high classification accuracy, the proposed BTsC model provides interpretable results, making the technique a valuable tool to study neural activity in various tasks and categories. The proposed solution can be applied to neural data recorded in various tasks, where there is a need for interpretable results and accurate classification accuracy. Keywords:Bayesian analysis Neural decoding Interpretable modeling. ## 1 Introduction Neuroscientists have long sought methods to decode neural activity in hopes of restoring movement and communication for individuals with neurological injuries [28], including stroke, spinal cord injury, brain trauma, and neurodegenerative diseases (e.g., Amyotrophic Lateral Sclerosis, ALS) using brain-computer interfaces (BCIs). Significant progress has been made in motor control [15], with advances also seen in the realms of speech [2], mood, and decoding neural activity corresponding to visual stimuli [11, 14, 16]. Despite these discoveries, BCIs remain primarily in research endeavors, facing hurdles in terms of costs, risks, and technological challenges [36]. The critical components of a BCI system involve feature extraction and accurate classification of neural activity related to different tasks or sensory input. However, several challenges exist in achieving highly accurate classifiers, including selecting the most informative and well-reasoned neural features and developing an interpretable classifier capable of utilizing limited datasets [3]. An interpretable classifier elucidates key neural signal features, like specific frequency bands or time periods, crucial for task performance. This insight enhances our understanding of the neural mechanisms. Addressing these challenges is further complicated by the prohibitively large number of features used to decode neural activity. Features might encompass raw neural data, which include the measured voltage from single or multiple electrode contacts in non-invasive techniques such as electroencephalography (EEG), and invasive methods like intracranial EEG (iEEG) or blood-oxygen-level-dependent (BOLD) signal in functional MRI. Many researchers have suggested a variety of features derived from these neural recordings. These could range from simple statistical metrics of the signal in time or frequency domain, like mean, median, standard deviations, kurtosis, and skewness to more sophisticated features such as Common Spatial Patterns (CSPs) [7], Higher-Order Crossing (HOC) features [22], Hjorth features [20], and Auto-Regressive (AR) coefficients [33]. As the EEG signal is non-stationary [34], time-frequency domain methods such as wavelet transform (WT) have been used for feature extraction as well [23]. In addition to the types of features used, there can be multiple streams of the same data represented by different electrode channels. Many researchers employ data reduction techniques to handle this redundancy of information, such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), or linear discriminant analysis (LDA) [32]. Another approach in data analysis focuses on selecting the most informative channels relative to the target classification task. This can be achieved through channel selection techniques, which can be categorized as either classifier-dependent (including wrapper and embedded-based methods) or classifier-independent (such as filter-based methods) [1]. In summary, different neural activity features can yield varied inferences, not always enhancing the understanding of encoding mechanisms. Thus, refining feature extraction and selection are essential for deriving relevant information that deepens our comprehension of brain processes. Once features are chosen, classification commences, with considerable recent work involving deep-learning-based classifier models. While deep learning frameworks such as Convolutional Neural Networks [9] and Recurrent Neural Networks [25] have been applied to the decoding and classification of EEG signals to identify stimuli [5], these methods might not be as applicable to invasive neural recordings like iEEG. This is primarily due to the limited availability of iEEG data compared to non-invasive EEG data. The drawbacks of these models include, 1) lack of interpretability, which corresponds to difficulty in identifying the necessary features for classification in many deep learning models, and 2) the requirement of vast amounts of data in the training phase of these models, which may not always be possible to collect when dealing with invasive recordings like iEEG [26]. Consequently, the limited data in inva sive neural recordings necessitates the exploration of alternative methods that can effectively handle such constraints while maintaining accurate classification and interpretability. In this study, we propose the Bayesian Time-Series Classifier (BTsC) model, designed to address the above-mentioned challenges. The BTsC allows us to identify the minimum number of channels necessary for stimulus classification from neural data. Furthermore, it enables the selection of neural features from different electrodes that provide optimal classification power, as well as the determination of the precise time period following a stimulus required for accurate stimulus classification. The proposed model can be trained effectively with limited data by leveraging the dynamics of local field potential (LFP) signals in two frequency subbands. Our proposed BTsC model employs a wrapper-based technique and greedy search in selecting channels and features for optimal classification. The pipeline of feature selection and classification used in the model can be applied to other classifiers, including Support Vector Machine (SVM) [29], Long Short-Term Memory (LSTM) [10], Naive Bayes [24], etc. We applied this model to decode the presence of one or the other visual stimulus using LFP neural activity from multiple neural nodes' activity, where our model shows high accuracy in classifying simple stimuli. In the following, we first outline our method of feature extraction from LFP signals and the development of the BTsC for stimulus decoding. Next, we assess the BTsC model's performance and compare it with other machine learning models. Lastly, we discuss our findings' implications for brain information processing and the potential applications of our model. ## 2 Material and Methods ### Dataset and Behavioral Task Four participants performed the visual stimulus task, while LFP was recorded from 278 sites via intracranial stereo EEG (Fig. 1-A). Intracranial EEG recordings were made over the course of clinical monitoring for spontaneous seizures. Participants were implanted with multi-lead depth electrodes (also known as stereotactic EEG, sEEG) [6]. All patients voluntarily participated after fully informed consent according to NIH guidelines, as monitored by the Massachusetts General Brigham (formerly Partners) Institutional Review Board (IRB). Participants were informed that participation in the tests would not alter their clinical treatment and that they could withdraw at any time without jeopardizing their clinical care. The Flicker task consisted of 100 trials where participants were presented with a red fixation cross for 0.5 seconds on a grey background (Fig. 1-B). This was followed by the appearance of a single black or white square on the same grey background. The duration of the square varied randomly between 2 to 4 seconds. The color of the square for each trial was randomly selected from a sequence of black or white with equal chance. Participants were instructed to focus on the red fixation cross and count the number of black or white squares presented to enhance engagement. ### Data Preprocessing: Intracranial LFP Data Data analysis was performed using custom analysis code in MATLAB and Field-trip, an open-source software implemented in MATLAB [21]. All data were subsequently decimated to 1000Hz, demeaned relative to the entire recording, and line noise and its harmonics up to 200Hz were removed by subtracting the band-passed filtered signal from the raw signal. Channels with excessive line noise or which had no discernible signal were removed from the analysis. In addition, we removed pathological channels with interictal epileptiform discharges (IEDs) using an automatic IED detection algorithm [13] (version v21, default settings except -h at 60; [http://isarg.fel.cvut.cz](http://isarg.fel.cvut.cz)). We removed channels that had detected IEDs greater than 6.5 IEDs/minute The remaining electrodes were subsequently bipolar re-referenced relative to nearest neighbors to account for volume conduction. ### Extracting Neural Features for Decoding We focus on two categories of features known to encode stimulus information, namely the low-frequency event-related potentials (ERPs) and the high gamma power (HGP) following image onset (see Fig. 1-C). For the ERP neural features, we filter the LFP in the theta (3-7 Hz) and delta (0-3 Hz) frequency bands [35] Figure 1: Data Overview: (A) shows the electrode placement in four participants, (B) illustrates the Flicker Task paradigm performed in 100 trials, (C) outlines the steps for feature extraction, (D) displays the preprocessed single trial signal (top), ERP features (middle), and HGP features (bottom) extracted from the LTP02-LTP03 electrode for patient P04, and (E) presents a scatter plot of ERP (top) and HGP (bottom) features at times t1 and t2 for all trials recorded from the same electrode, indicating a Gaussian distribution. using a low-pass filter with a cut-off frequency of 7 Hz. By applying the low-pass filter, we conform to the Nyquist-Shannon sampling theorem, which allows us to down-sample the filtered data to 15 Hz without losing crucial information. This results in 15 features (each feature being an individual time step) for ERP signals per electrode. Each sample of the ERP feature vector includes a weighted average of multiple samples of the original time series data in a specific time interval. Under the central limit theorem assumption, we can assume that each element of these vectors follows a normal distribution. As HGP has also been shown to encode visual stimuli [19, 35], we band-pass filter the LFP from 65 to 120 Hz. Power is then calculated in 67-millisecond nonoverlapping windows after stimulus onset, generating the same number of features per each channel as the ERP. A subsequent step involves using a log transformation. This transformation compresses the range of high values and expands the range of low values, resulting in a distribution that is more symmetric and closer to a normal distribution. This justifies its use as an input feature vector to our model [4]. In the procedure described, each LFP recording channel results in two feature vectors, one for ERP and one for HGP, each represented as a time series. We model these vectors as multivariate normal distributions for use in the BTsC model, which we'll discuss further in the next section. ### Bayesian Time-series Classifier (BTsC) In section 2.3, we described how ERP and HGP feature vectors are acquired from each LFP recording channel. To simplify the description, we assume that each channel is associated with a single feature vector, and we refer to the classifier trained on this feature vector as a single-channel classifier. This simplification does not compromise the broader applicability of the model. To present the BTsC model, we start by detailing the single-channel classifier's construction and the process of determining optimal time periods. We then discuss combining single-channel classifiers into a multi-channel classifier and selecting the minimum subset of classifiers needed for maximum accuracy. The BTsC aims to determine the fewest feature vectors necessary for effective stimulus classification. Single-channel classifier. Let us assume that the pre-processed neural recording for the \(c^{th}\) electrode is represented by the feature vector defined by \(\mathbf{x}^{c}=\{\mathbf{x}_{1}^{c},\mathbf{x}_{2}^{c},\ldots,\mathbf{x}_{d}^{ c}\}\), where \(d\) is the length of observation period, and \(\mathbf{x}_{i}{}^{c}\) is the \(i^{th}\) sample of the feature vector for the \(i^{th}\) time interval after the stimulus onset. As discussed in section 2.3, we assume that each element of \(\mathbf{x}^{c}\) follows a normal distribution. We further assume the joint distribution of \(\mathbf{x}^{c}\) follows a multivariate normal, where the dependency across time points allows us to characterize the temporal dynamics of observed neural features. For the model, we build the conditional distribution of \(\mathbf{x}^{c}\) given stimulus, \(I\in\{0,\ldots,k\}\equiv\{(stimulus1,...,stimulusK)\}\). The conditional distribution of \(\mathbf{x}^{c}\) is defined by: \[\mathbf{x}^{c}|I\sim\mathcal{N}(\mu_{I}^{c},\mathbf{\Sigma}_{I}^{c}) \tag{1}\] \[p(\mathbf{X}=\mathbf{x}^{c}|I)=\frac{1}{(2\pi)^{d/2}|\mathbf{\Sigma}_{I}^{c}| ^{1/2}}\exp\left(-\frac{1}{2}(\mathbf{x}^{c}-\mu_{I}^{c})^{T}\mathbf{\Sigma}_ {I}^{c\,-1}(\mathbf{x}^{c}-\mu_{I}^{c})\right) \tag{2}\] where \(\mu_{I}^{c}\) and \(\mathbf{\Sigma}_{I}^{c}\) are the mean and covariance of \(\mathbf{x}^{c}\) given stimuli \(I\). Given \(\mu_{I}\) and \(\mathbf{\Sigma}_{I}\), we can construct a Bayesian classifier to compare the posterior probabilities of different stimuli. The assigned class is the one with the highest posterior probability, defined by: \[\forall j\neq k:L(I=k\mid\mathbf{X}=\mathbf{x}^{c})\geq L(I=j\mid\mathbf{X}= \mathbf{x}^{c}) \tag{3}\] \(L\left(I\mid\mathbf{X}=\mathbf{x}^{c}\right)\) corresponds to the posterior distribution of stimulus \(I\), given the observed neural features, defined by: \[L\left(I\mid\mathbf{X}=\mathbf{x}^{c}\right)\propto p(\mathbf{X}=\mathbf{x}^{ c}\mid I)p(I) \tag{4}\] where \(p(I)\) is the stimulus prior probability. To build our single-channel classifier, we require only the mean and covariance of each neural recording (\(\mathbf{x}^{c}\)) per stimulus. Using the training dataset \(\mathcal{D}\), we find the mean and covariance matrix for each time series. To obtain a robust estimation of the covariance matrix, the number of required samples must be on the order of \(d^{2}\). Estimating the covariance matrix can result in a biased estimation given the limited training data [18]. Our solution is to find the minimum length of the observation (\(d_{minimal}^{c}\)) starting from the onset, providing the highest accuracy with the cross-validated result. Using this approach, we can address the limited training dataset in our estimation of the covariance matrix. In the case of the multivariate normal distribution, we can obtain the marginal distribution of a subset of the neural data; any marginalized distribution remains a multivariate normal. With this in mind, we can examine the posterior of each class of data as a time-dependent function, identifying the stimulus from a subset of neural data \(\mathbf{x}^{c}\). We denote this posterior as \(L_{j}\), signifying a marginalized version of the overall posterior distribution \(L\), but only considering the first \(j\) features \(\{\mathbf{x}_{1}^{c},\mathbf{x}_{2}^{c},\ldots,\mathbf{x}_{j}^{c}\}\). We introduce the concept of \(C_{j}(\mathcal{D})\), representing the cross-validated classification accuracy of our model on the dataset \(\mathcal{D}\), using a marginalized posterior distribution with the first \(j\) features. For each classifier, the minimal time, denoted as \(d_{minimal}^{c}\), is defined as follows: \[d_{minimal}^{c}=\arg\max_{j}C_{j}(\mathcal{D}) \tag{5}\] This suggests that \(d_{minimal}^{c}\) represents the smallest set of features necessary to optimize our model's performance, in accordance with the constraints of the k-fold cross-validation method being used. #### 3.2.2 Multi-channel Classifier. We construct our BTsC model initially based on a single-channel classifier, which turns to Quadratic Discriminant Analysis (QDA) for time series [31]. In practice, we have multiple channels, with each channel having two feature vectors. Classifiers trained on these feature vectors can be combined to achieve higher classification accuracy. Equation (4) defines the classifier model for the \(c^{th}\) feature vector. We found that the single-channel classifier accuracy is limited, and in the best case, it is about or less than 75%. To attain higher classification accuracy, we expand our single-channel QDA classifier to account for multiple channels. We employ two solutions to adapt the classifier for multi-channel neural recordings, resulting in an ensemble-based classifier that enhances accuracy and robustness. The first solution expands the model directly to multiple channels. The second solution is based on the majority voting rule of \(C\) different classifiers, where \(C\) is the number of possible channels. Multi-channel Likelihood Method.For a multi-channel classifier, we assume that different feature vectors are conditionally independent given the stimulus. The joint conditional distribution of all recordings is defined by: \[p(\mathbf{X}_{1}=\mathbf{x}^{1},\ldots,\mathbf{X}_{C}=\mathbf{x}^{C}|I)=p( \mathbf{X}_{1}=\mathbf{x}^{1}|I)\ldots p(\mathbf{X}_{C}=\mathbf{x}^{C}|I) \tag{6}\] Where \(I\) represents the stimulus. We can construct each single-channel model similar to the one defined in Equation (2). The posterior distribution for each multi-channel neural recording is defined by: \[L(I;\mathbf{X}_{1}=\mathbf{x}^{1},\ldots,\mathbf{X}_{C}=\mathbf{x}^{C})\propto p (\mathbf{X}_{1}=\mathbf{x}^{1}|I)\ldots p(\mathbf{X}_{C}=\mathbf{x}^{C}|I)p(I) \tag{7}\] Utilizing all channels in the model is not practical as some may lack informative features for classification. Also, multiplying likelihoods from all channels can complicate computation due to smaller values. Hence, identifying the most informative channels through channel subset selection is necessary for accurate classification. We will discuss this challenge in the classification subset selection section. Maximum Voting Method.The outcomes of single-channel classifiers can be combined using the voting method to achieve higher classification accuracy. In a poll of \(C\) single-channel classifiers, each classifier contributes a vote towards a class label based on its independent prediction. The class which receives the most votes, representing the majority opinion among the classifiers, is selected as the final output class [31]. #### 2.0.2 Classifiers Subset Selection. We can combine single-channel classifiers to construct a multi-channel classifier using either of these two methods. The process of selecting the optimal subset of feature vectors is based on an adaptive (or greedy) search. It begins with a single channel with the best performance using k-fold cross-validation and then examines which other channel can be added to it. All possible combinations are evaluated, and the one with the best performance is selected. Through this adaptive search, the minimal number of channels that reach the highest classification accuracy with the highest confidence can be determined with cross-validation. ## 3 Results The BTsC model can be applied to different mental tasks with various features. Here, we investigated the application of BTsC in the visual task described in section 2.1. We trained the BTsC using low-frequency (ERP, \(<7Hz\)) and high-frequency power (HGP; 65-120 Hz) dynamics following image onset to test if we can decode visual stimuli from across the brain at a high temporal resolution (67 ms). Then, we identified the features and time points for maximum decoding accuracy (Fig. 2). Furthermore, to validate the results obtained with the BTsC model, we compared them with the decoding outcomes of seven additional machine learning (ML) algorithms on visual stimuli. Optimal Stimulus Decoding Time and Features in the Model Following the model subset selection, we discovered that the channels and features that survived the selection criteria and contributed to the BTsC models were from multiple electrodes across multiple brain regions. BTsC enabled us to evaluate the performance of each feature vector, ERP and HGP, on individual channels and to determine the optimal timing post-image onset for superior performance. From this analysis, we discerned which regions and which features exhibited the fastest responses. Additionally, we found that combining these feature vectors leads to a boost in performance (Fig. 2-B). Upon analyzing the BTsC results for ERP, HGP, and their combined utilization, we discovered that leveraging both ERP and HGP features enhances the decoding model's accuracy in the visual task (Fig. 2-C). After identifying the time window after image onset with the peak accuracy for the single time or cumulative BTsC models, we found that the time of maximum accuracy after image onset was below 0.8 seconds. The accuracy at each time point and the number of utilized feature vectors are depicted in Fig. 2-D. Figure 2: Results: (A) illustrates the performance of individual classifiers. (B) shows accuracy evolution during channel combination steps for participant P05. (C) shows the comparison between the BTsC performance using ERP, HGP, and both ERP and HGP features to assess the impact of neural features. (D) displays the accuracy of the model at different time points for participants P01 to P05. ### Machine Learning Decoding To test if decoding results, which support the distributed information flow hypothesis [27] are particular to the BTsC model, we applied seven machine learning models to the same neural features (ERP and HGP) and participant data. The machine learning classifiers include SVM [29], Logistic Regression (LR) [12], Naive Bayes (NB) [24], Random Forest (RF) [12], Multi-Layer Perceptron (MLP) [8], LSTM [10] and EEGNet [17]. ## 4 Discussion Our investigation has yielded significant insights into the effectiveness of the BTsC in decoding visual stimuli from intracranial EEG recordings. The BTsC model's integration of the ERP and HGP features has demonstrated a remarkable capacity for classifying visual stimuli, outperforming other machine learning models including SVM, Logistic Regression, Naive Bayes, Random Forest, MLP, LSTM, and EEGNet. By leveraging the BTsC model, we achieved an average accuracy of 75.55% in classifying stimuli (Table 1). This is a noteworthy outcome, particularly given the inherent complexity of neural data, inter-subject variability in iEEG electrode position, and the challenges associated with decoding neural signals. Our results show that the BTsC model can effectively capture the distributed information flow within the brain, with its simplicity offering robustness against limited training data. In comparison, other methods face challenges such as overfitting, interpretability, and scalability issues. For example, complex models like EEGNet can result in overfitting due to their complex architecture, making them less reliable when dealing with limited data [5]. \begin{table} \begin{tabular}{l c c c c c c c} \hline \hline & \multicolumn{2}{c}{**P01**} & \multicolumn{2}{c}{**P02**} & \multicolumn{2}{c}{**P03**} & \multicolumn{2}{c}{**P04**} \\ \cline{2-9} **Model** & Acc. & F1 & Acc. & F1 & Acc. & F1 & Acc. & F1 \\ \hline BTsC (L) & 74.6(5) & **75.2(8)** & **73.7(4)** & **71.0(9)** & 66.5(7) & 62.0(8) & 85.4(7) & 83.1(3) \\ BTsC (V) & **75.1(9)** & 74.3(7) & 70.6(6) & 69.8(9) & 66.4(10) & 67.3(7) & **90.0(6)** & **89.1(4)** \\ EEGNet & 72.0(7) & 73.2(9) & 68.1(7) & 70.1(11) & **67.4(9)** & **68.1(9)** & 85.3(10) & 84.2(8) \\ SVM & 61.2(13) & 65.2(8) & 70.0(14) & 73.0(9) & 58.5(9) & 51.0(10) & 81.2(7) & 81.1(10) \\ NB & 70.0(6) & 69.4(8) & 68.7(8) & 70.7(8) & 51.1(11) & 54.0(18) & 74.8(16) & 70.9(9) \\ RF & 60.0(6) & 58.5(3) & 62.5(3) & 63.0(3) & 55.2(4) & 56.1(8) & 72.9(6) & 71.2(5) \\ MLP & 67.8(9) & 63.2(6) & 67.7(11) & 68.1(9) & 58.5(10) & 55.9(10) & 71.9(8) & 72.5(7) \\ LR & 66.2(11) & 66.0(8) & 67.5(7) & 66.8(13) & 53.1(5) & 46.5(5) & 75.0(12) & 76.1(10) \\ LSTM & 64.6(8) & 63.2(6) & 67.7(7) & 64.2(9) & 52.7(9) & 50.0(11) & 70.0(7) & 71.9(11) \\ \hline \hline \end{tabular} \end{table} Table 1: Performance Comparison across Different ML Techniques: The table depicts the mean and standard deviations (presented in parentheses) of 5-fold cross-validation accuracy for all models, including BTsC (L) and BTsC (V), where (L) and (V) represent combination methods using likelihood and voting, respectively. On the other hand, simpler models like Naive Bayes rely on the assumption of features' independence, which is often unrealistic in real-world applications. The key feature of our model is its interpretability, essential for validating predictions and understanding brain functions [37]. Our BTsC model offers a clear view of which areas of the brain are encoding stimuli and at what time these features exhibit the most discriminative power. Additionally, this model possesses flexibility in utilizing diverse feature vectors and can provide insights into the specific contributions of each vector to the final outcome. In our study, we leveraged both ERP and HGP dynamical features individually and in combination within the BTsC. While both ERPs and HGP independently provided significant information for decoding visual stimuli, we found that combining these features led to a marked increase in classification accuracy. This suggests that these neural features, while independently informative, offer complementary information that enhances the decoding power when combined, reflecting the complex and distributed nature of brain processing [30]. Thus, the integration of multiple neural features could contribute to more accurate and robust models for neuroscience applications. Further expanding our research, we conducted an experiment to determine the most informative time period after image onset for training the BTsC model. This aspect is crucial, as it helps establish the optimal window for capturing the most discriminative neural features. In this experiment, we trained the BTsC model using different time windows post-image onset. Consequently, we determined the optimal time window post-image onset for training the BTsC model for each patient (Fig. 2-D). Moreover, this experiment revealed that in shorter time windows, HGP feature vectors are more involved than ERP. Despite the encouraging results, our model has limitations. The assumption of channels' independence is a significant limitation, which we intend to address in future iterations. We plan to refine our model to account for possible dependencies and correlations between features and electrodes, which could further enhance the predictive accuracy of the BTsC model. ## 5 Conclusion In this study, we introduced a novel Bayesian Time-series Classifier that utilizes both low-frequency event-related potentials and high gamma power features to effectively decode visual stimuli from intracranial EEG recordings. The BTsC model identifies encoding brain areas, discriminative features, and optimal time windows, outperforming other classifiers by 3% in accuracy due to its ability to capture distributed information flow. With its demonstrated success in decoding simple visual information and accommodating individual variations, this model holds promise for applications in neuroscience and clinical settings, including brain-computer interfaces and neural prosthetics. Future research will broaden the scope to include more cognitive tasks and modalities, personalize neurotechnology through additional neural features, and explore the impact of different covariance structures on our Bayesian Time-Series Classifier model.

2306.05748

Shape-based clustering of synthetic Stokes profiles using k-means and k-Shape

The shapes of Stokes profiles contain much information about the atmospheric conditions that produced them. However, a variety of different atmospheric structures can produce very similar profiles. Thus, it is important for proper interpretation of observations to have a good understanding of how the shapes of Stokes profiles depend on the underlying atmosphere. An excellent tool in this regard is forward modeling, i.e. computing and studying synthetic spectra from realistic simulations of the solar atmosphere. Modern simulations routinely produce several hundred thousand spectral profiles per snapshot. With such numbers, it becomes necessary to use automated procedures in order to organize the profiles according to their shape. Here we illustrate the use of two complementary methods, k-means and k-Shape, to cluster similarly shaped profiles, and demonstrate how the resulting clusters can be combined with knowledge of the simulation's atmosphere to interpret spectral shapes. We generate synthetic Stokes profiles for the Ca II 854.2 nm line using the Multi3D code from a Bifrost simulation snapshot. We then apply the k-means and k-Shape clustering techniques to group the profiles together according to their shape. We show and compare the classes of profile shapes we retrieve from applying both k-means and k-Shape to our synthetic intensity spectra. We then show the structure of the underlying atmosphere for two particular classes of profile shapes retrieved by the clustering, and demonstrate how this leads to an interpretation for the formation of those profile shapes. Furthermore, we apply both methods to the subset of our profiles containing the strongest Stokes V signals, and demonstrate how k-Shape can be qualitatively better than k-means at retrieving complex profile shapes when using a small number of clusters.

Thore Espedal Moe, Tiago M. D. Pereira, Flavio Calvo, Jorrit Leenaarts

2023-06-09T08:27:26Z

http://arxiv.org/abs/2306.05748v1

# Shape-based clustering of synthetic Stokes profiles using \(k\)-means and \(k\)-Shape ###### Abstract Context:The shapes of Stokes profiles contain much information about the atmospheric conditions that produced them. However, a variety of different atmospheric structures can produce very similar profiles. Thus, it is important for proper interpretation of observations to have a good understanding of how the shapes of Stokes profiles depend on the underlying atmosphere. An excellent tool in this regard is forward modeling, i.e. computing and studying synthetic spectra from realistic simulations of the solar atmosphere. Modern simulations routinely produce several hundred thousand spectral profiles per snapshot. With such numbers, it becomes necessary to use automated procedures in order to organize the profiles according to their shape. Here we illustrate the use of two complementary methods, \(k\)-means and \(k\)-Shape, to cluster similarly shaped profiles, and demonstrate how the resulting clusters can be combined with knowledge of the simulation's atmosphere to interpret spectral shapes. Aims:We aim to showcase the use of clustering analysis for forward modeling. In particular we wish to introduce the \(k\)-Shape clustering method to the solar physics community as a complement to the well-known \(k\)-means method. Methods:We generate synthetic Stokes profiles for the Ca ii 854.2 nm line using the Multi3D code from a Bifrost simulation snapshot. We then apply the \(k\)-means and \(k\)-Shape clustering techniques to group the profiles together according to their shape, and investigate the within-group correlations of temperature, line-of-sight velocity and line-of-sight magnetic field strengths. Results:We show and compare the classes of profile shapes we retrieve from applying both \(k\)-means and \(k\)-Shape to our synthetic intensity spectra. We then show the structure of the underlying atmosphere for two particular classes of profile shapes retrieved by the clustering, and demonstrate how this leads to an interpretation for the formation of those profile shapes. Furthermore, we apply both methods to the subset of our profiles containing the strongest Stokes \(V\) signals, and demonstrate how \(k\)-Shape can be qualitatively better than \(k\)-means at retrieving complex profile shapes when using a small number of clusters. Conclusions: ## 1 Introduction Forward modeling of the solar atmosphere is a very useful tool for understanding the relative importance of atmospheric components in the formation of polarized spectra, thereby guiding interpretations of observations. By computing synthetic Stokes profiles from realistic 3D radiative magnetohydrodynamic (rMHD) simulations, one can directly compare a particular spectral signature with the full state of the atmosphere that produced it (see e.g. Leenaarts et al., 2013, 2013, 2013) and others in the same series). Modern simulations routinely contain several hundred thousand pixels, with each pixel giving rise to a set of Stokes profiles. Depending on the spatial resolution of the numerical model, and the spectral resolution considered for the synthesis, these profiles can be quite complex; often exhibiting more complicated behavior than what is typically resolved in real observations. It is obviously not feasible to analyze the formation of so many profiles one by one, nor is it practical to manually sort them into groups according to their features. Rather, some automated procedure must be used to organize the profiles in a meaningful manner for further human analysis. One way of reducing the number of individual profiles into more manageable collections is the use of clustering techniques like \(k\)-means (Steinhaus, 1956; MacQueen, 1967). \(k\)-means has seen extensive use in solar and stellar physics, for examples see Sanchez Almeida & Lites (2000); Pietarila et al. (2007); Viticchie & Sanchez Almeida (2011); Panos et al. (2018); Sainz Dalda et al. (2019); Bose et al. (2019); Kuckein et al. (2020); Joshi et al. (2020); Woods et al. (2021); Nobrega-Siverio et al. (2021); Barczynski et al. (2021); Bose et al. (2021); Kleint & Panos (2022); Mathur et al. (2022); Sainz Dalda et al. (2022). Apart from \(k\)-means, other clustering methods have also been used on solar spectra, for instance the \(t\)-distributed Stochastic Neighbor Embedding employed by (Verma et al., 2021). The purposes of the clustering vary from identifying and studying the observational signatures of particular physical processes and features, to reducing the spatial dimensionality of data-sets for inversions, to statistical characterizations of observations. Relatively little explored, however, is the application of clustering techniques in a forward modeling context, one notable exception being Khomenko et al. (2005). In this paper we aim to address that issue, applying the \(k\)-means method to Ca ii 854.2 nm Stokes \(I\) and Stokes \(V\) profiles generated from a Bifrost (Gudiksen et al., 2011) snapshot using the Multi3D radiative transfer code (Leenaarts & Carlsson, 2009), which has been extended (Calvo & Leenaarts (in prep.)) to include polarization, accounting for the Zeeman effect. We focus on the shapes of the Stokes profiles, aiming to illustrate what different classes of shapes do, or do not, tell us about the underlying atmospheric conditions. While \(k\)-means is a fast and robust clustering technique, it does not directly cluster profiles based on their shapes. It works by minimizing the sum of within-cluster Euclidean distances between profiles, which can potentially lead to distinctly different shapes appearing in the same cluster as demonstrated in Fig. 1. Or, for instance, two Doppler-shifted spectral profiles with the otherwise same exact shape can be put into separate clusters. Furthermore, the centroid, or'representative profile' (RP), of a cluster is given as the mean of the profiles belonging to the cluster, which in some cases can give a poor representation of the typical profile shapes in the cluster. Of course, increasing the number of clusters can mitigate this problem, but at the cost of the interpretability, which is the main point of the kind of forward modeling we seek to undertake in this paper. A relatively fast clustering method that is inherently shape-based is the \(k\)-Shape method of Paparrizos & Gravano (2015). Though originally developed for use on time-series, the method is quite general and we apply it here to the case of Stokes profiles with the obvious substitution of the time axis for a wavelength axis. A feature of \(k\)-Shape is that the clustering is largely independent of Doppler-shifts, which can be beneficial or detrimental depending on the intended usage. By ignoring Doppler-shifts and using a different measure of similarity than \(k\)-means, the profiles are matched more directly according to their similarity in actual shape, rather than being matched according to a combination of shape and wavelength position. Furthermore, as the centroid computation is rather different from the one in \(k\)-means, the RP's are much more prototypical of the clustered profiles. The cost, of course, is that all absolute velocity-information is not considered in the clustering. ## 2 Methods ### Generating synthetic profiles We generated our synthetic spectra from the 23 km resolution atmospheric model described in Moe et al. (2022). This is a Bifrost model (Gudiksen et al., 2011) with a magnetic field configuration constructed to resemble a coronal hole. The model has \(512\times 512\times 512\) grid points, spanning roughly 12 Mm in the horizontal directions and going from \(z=-2.5\) Mm below up to \(z=8\) Mm above the solar surface. The horizontal spacing of the grid points is uniform, resulting in a horizontal resolution of 23 km pix\({}^{-1}\). We used an extension (Calvo & Leenaarts (in prep.)) of the Multi3D code (Leenaarts & Carlsson, 2009) with polarimetric capabilities to produce 3D full Stokes profiles of the Ca ii 854.2 nm line accounting for the Zeeman effect. As 3D computations are immensely expensive we cut the bottom 112 grid points, corresponding to below \(-0.4\) Mm beneath the surface, under the assumption that these are too deep to affect the formation of our line of interest. Furthermore, we neglected to include the effects of partial frequency redistribution (PRD) and isotopic splitting. The obtained synthetic profiles were normalized by the nearby continuum, meaning each profile was divided by the Stokes I value of the reddest wavelength in the synthesis at approximately \(\lambda_{0}+0.95\) nm, and interpolated to 100 equidistant wavelength points in the range \(\lambda_{0}\pm 0.05\) nm, where \(\lambda_{0}\) denotes the central wavelength of the line. We performed this interpolation in order to give equal weight to all parts of the profile when clustering since the original wavelength grid used in the synthesis is non-equidistant. ### \(k\)-means clustering The most common clustering technique for spectral profiles is \(k\)-means clustering. The full set of profiles is divided into \(k\) clusters of similarly shaped profiles, where the number \(k\) must be chosen at the outset. The measure of similarity is the Euclidean distance between profiles; that is, the distance between two profiles is the sum over wavelengths of the squared difference in their amplitudes: \[distance=\sum_{i}(I_{1}(\lambda_{i})-I_{2}(\lambda_{i}))^{2}, \tag{1}\] where \(I(\lambda_{i})\) denotes the amplitude of the profile at each wavelength point \(\lambda_{i}\). Each cluster has a centroid, and the goal is to assign the profiles to the \(k\) clusters in such a way that the sum of distances between all profiles and their nearest centroid (often called the inertia) is minimized. Algorithmically, \(k\)-means performs the following steps: 1. Initialize \(k\) centroids, one for each cluster. 2. Assign each profile to the cluster with the closest centroid. 3. Recompute the centroids as the mean (for each wavelength) of the profiles belonging to the cluster. 4. Repeat 2. and 3. until no profile changes cluster, a fixed number of iterations has been performed, or until the total inertia no longer changes above a set tolerance. It should be noted that the convergence of the \(k\)-means algorithm does not guarantee that a global minimum has been found. Therefore it is common to re-initialize the clustering a predefined number of times, keeping the result with lowest inertia. In this paper, we have used the \(k\)-means implementation of scikit-learn (Pedregosa et al., 2011), employing the \(k\)-Means++ initialization (Arthur & Vassilvitskii, 2007) for selecting better initial cluster centroids. Figure 1: Example showing how \(k\)-Shape (left) and \(k\)-means (right) partition a dataset with three distinct signal shapes. While \(k\)-Shape recovers the three distinct classes of shapes, \(k\)-means mixes the class containing a peak and a drop with the class containing only a drop. This illustration is adapted from the documentation of the tklearn-library. 1 ### \(k\)-Shape clustering As the name implies, \(k\)-Shape (Paparrizos & Gravano 2015), is designed to perform a clustering into \(k\) clusters of distinct shape. While the general idea is similar to \(k\)-means, it uses a different metric for the distance between profiles; as well as another method for computing the cluster centroids. The distance metric is based on shifting the profiles across each other and computing the cross-correlation for each possible shift. Consider two profiles \(I_{1}\) and \(I_{2}\), defined on \(m\) wavelength points, written in the form of vectors: \[\mathbf{I_{1}}=I_{1}(\lambda_{1}),I_{1}(\lambda_{2}),...,I_{1}(\lambda_{m}),\ \ \mathbf{I_{2}}=I_{2}(\lambda_{1}),I_{2}(\lambda_{2}),...,I_{2}(\lambda_{m}). \tag{2}\] The cross-correlation sequence between these two profiles, \(CC_{w}(\mathbf{I_{1}},\mathbf{I_{2}})\), is defined as: \[CC_{w}(\mathbf{I_{1}},\mathbf{I_{2}})=R_{w\leftarrow m}(\mathbf{I_{1}},\mathbf{I_{2}}),\ \ \ \ w\in\{1,2,\ldots,2m-1\}, \tag{3}\] where \[R_{k}(\mathbf{I_{1}},\mathbf{I_{2}})=\begin{cases}\sum_{l=1}^{m-k}I_{1}(\lambda_{l+k} )\cdot I_{2}(\lambda_{l}),\ \ \ k\geq 0\\ R_{-k}(\mathbf{I_{1}},\mathbf{I_{2}}),\ \ \ k<0.\end{cases} \tag{4}\] Thus, the sequence \(CC_{w}(\mathbf{I_{1}},\mathbf{I_{2}})\) contains the cross-correlation value for each of the \(2m-1\) possible shifts of the profiles relative to each other; essentially a sequence of the vector dot products between zero-padded \(\mathbf{I_{1}}\) and \(\mathbf{I_{2}}\) for each possible overlapping shift of the profiles. Normalizing the cross-correlation sequence (corresponding to dividing by the Euclidean norm of both profiles): \[NCC_{c}=\frac{CC_{w}(\mathbf{I_{1}},\mathbf{I_{2}})}{\sqrt{R_{0}(\mathbf{I_{1}},\mathbf{I_{1}}) \cdot R_{0}(\mathbf{I_{2}},\mathbf{I_{2}})}}, \tag{5}\] results in a number between \(-1\) and \(1\) for each entry in the sequence, where \(-1\) signifies perfect anti-correlation and \(1\) signifies perfect correlation between the profiles. Selecting the entry with the largest cross-correlation value then gives the shape-based distance between two profiles as: \[distance=1-\max_{w}\Big{(}\frac{CC_{w}(\mathbf{I_{1}},\mathbf{I_{2}})}{\sqrt{R_{0}( \mathbf{I_{1}},\mathbf{I_{1}})\cdot R_{0}(\mathbf{I_{2}},\mathbf{I_{2}})}}\Big{)}, \tag{6}\] which is bounded between \(0\) and \(2\). As in \(k\)-means, each profile is assigned to the closest centroid in terms of distance, and the cluster centroid is recomputed. In \(k\)-Shape, however, the refinement of the cluster centroids is done by reformulating the minimization of within-cluster distances as a maximization of a Rayleigh quotient calculation; for details see the original paper (Paparrizos & Gravano 2015). It should, however, be remarked that the \(k\)-Shape method assumes that the profiles have been \(z\)-normalized, meaning each profile has zero mean, and unity standard deviation: \[\mathbf{I_{1}}^{\prime}=\frac{\mathbf{I_{1}}-\mu_{1}}{\sigma_{1}}, \tag{7}\] where \(\mu_{1}\) and \(\sigma_{1}\) is, respectively, the mean and the standard deviation of the profile over the \(m\) wavelengths considered. This assumption is not strictly necessary, as the method can be modified to work with other data-normalizations. However, the original authors found the z-normalization to work best in their tests and it is beyond the scope of our current work to re-implement and evaluate the method for other normalizations. We used the \(k\)-Shape implementation from the tislearn library (Tavenard et al. 2020), with some simple modifications to make it run in parallel. Even so, the \(k\)-Shape method is significantly slower than the \(k\)-means implementation of scikit-learn. In one example case, using \(k=100\) clusters for \(512\times 512\) profiles with 100 wavelength points, one run of \(k\)-Shapes without re-initializations took roughly \(2.7\) hours, while a \(k\)-means run with 10 re-initializations took about \(5\) minutes, both on the same \(32\)-core workstation. It should be noted that in the tislearn implementation of \(k\)-Shape, \(k\) single profiles are randomly chosen as the initial cluster centroids. In the original paper (Paparrizos & Gravano 2015), the initialization is done by randomly distributing all profiles among \(k\) clusters ## 3 Results ### Overview Our intention was to illustrate and compare the use of both \(k\)-Shape and \(k\)-means for clustering synthetic profiles according to their shape, and subsequently how the resulting clusters can reveal correlations between the typical profile shapes in a cluster and the particular structure of the underlying atmosphere these profiles emerge from. We therefore begin by presenting and discussing the clustering of the intensity profiles in Sec. 3.2, before we perform a detailed examination of two particular profile shapes retrieved by the clustering in Sec. 3.3 and Sec. 3.4. As the \(k\)-Shape method assumes that its input profiles are \(z\)-normalized, we used the same normalization for the \(k\)-means method in order to do a fair comparison. This turned out to be a reasonable approach for the synthetic intensity profiles, as they have signal values in the same general range. However, the polarized components of the Stokes vector can vary vastly in amplitude, so the \(z\)-normalization can cause tiny signals to appear misleadingly large compared to stronger signals as the amplitude is given in units of the per-profile standard deviation. We have therefore focused mostly on the intensity profiles, though we did perform a clustering of the very strongest Stokes \(V\) signals (those with a signal exceeding \(0.5\%\) of the nearby continuum intensity), which we will discuss in Sec. 3.5. ### Clustering the intensity profiles We clustered the synthetic intensity profiles into \(k=100\) clusters using both \(k\)-means and \(k\)-Shape, the resulting clusters are shown in Fig. 2 and Fig. 3 respectively. The choice of \(100\) clusters was made after some experimentation, as a reasonable trade-off between the two opposing considerations of accuracy and human interpretability. The \(k\)-means method was run with 10 re-initializations, while the \(k\)-Shape method was run with a single initialization due to being around two orders of magnitude slower. We have tested \(k\)-Shape with 10 re-initializations, which yielded qualitatively very similar results to the single initialization run. We therefore elected to use the single initialization run in order to compare the methods for somewhat more similar run-times. The first observation we can make is that both clustering techniques seem to recover a similar variety of different profile shapes. These range from typical absorption profiles (e.g. #35 in Fig. 2, #52 in Fig. 3), through increasingly strongly skewed absorption profiles (e.g. #9 and #37 in Fig. 2, #30 and #44 in Fig. 3), to more complicated profiles, including double-peaked profiles (e.g. #98 and #100 in Fig. 2, #97 and #100 in Fig. 3), asymmetric emission profiles (e.g. #73 in Fig. 2, #64 in Fig. 3) and multi-lobed profiles (e.g. #81 in Fig. 2 or #84 in Fig. 3). The clustering appears to be reasonably tight, and in both methods there are several clusters showing very similar shapes, i.e. there is more than one cluster per 'family' of shapes. Encouragingly, both clustering methods seem to recover all the same types of cluster 'families', e.g. several clusters with similar asymmetric emission peaks or double peaks show up in both clusterings, though there is obviously not a one-to-one correspondence between individual clusters across the methods. Conversely, at first glance there do not seem to be clusters with very distinct shapes found only with one method compared to the other. The most unique-looking clusters are perhaps #56 and #88 in Fig. 2, but even these find quite similar counterparts in #97 and #47 in Fig. 3. This gives us some confidence that our choice of 100 clusters reasonably covers the range of typical profile shapes. A second observation we can make, is how the retrieved clusters do differ between the methods. The \(k\)-Shape groupings demonstrate the method's insensitivity to Doppler-shifts, especially the clusters containing the asymmetric emission peaks (e.g. #63, #64, #65, in Fig. 3) show the same shape at different shifts grouped together. Conversely, \(k\)-means splits these into different clusters (e.g. #72, #73, #74 in Fig. 2) according to their Doppler shifts. The fact that both methods retrieve the same 'families', but differently distributed over the clusters, can be beneficial for analysis, as we will see in Sect. 3.3. With such a stereoscopic view of the underlying atmospheres it becomes easier to discern by inspection which atmospheric parameters are important and which are incidental for the formation of the particular profile shapes. In particular, \(k\)-Shape's insensitivity to Doppler shifts contrasted with \(k\)-means sensitivity to them, allows one to better discern which atmospheric behaviors are correlated solely with the shape of the profile, as opposed to being correlated to the combination of shape and Doppler-shift. A third observation relates to how and where the methods perform poorly, in terms of profiles not being a good fit for their assigned clusters. As mentioned, cluster #56 in \(k\)-means does not seem to be well captured by \(k\)-Shape. It turns out that most of the profiles from this cluster are assigned to #68 and #73 in Fig. 3. These profiles are on the whole quite different from their assigned \(k\)-Shape centroids, but when the profiles and the centroids are shifted drastically across each other, the overlapping parts agree sufficiently for them to be grouped together. As \(k\)-Shape computes all possible shifts, it may occasionally find large shifts (and thereby a large clipping of the signal) to be the least bad option, leading to such apparently poor assignments. That type of signal clipping does not happen with \(k\)-means. On the other hand, the \(k\)-means clusters appear to have issues distinguishing profiles where there is a large difference in signal Figure 2: \(k\)-means clusters for synthetic Ca ii 854.2 nm intensity profiles, using 100 clusters on \(z\)-normalized profiles. The red line is the cluster centroid profile (average), while the black lines are all individual profiles assigned to each cluster. The grey line is a visual aid that denotes the position of \(\lambda_{0}\) strength over a narrow wavelength region. For instance, the \(k\)-means cluster #79 in Fig. 2 turns out to be a mix of profiles with enhanced shoulders on either the right side or on both sides of the line core, as well as some with only a weakly enhanced right shoulder followed by a second absorption feature to the right. In the \(k\)-Shape clustering vast majority of these profiles are assigned to #77, #78, #94 and #95 in Fig. 3. To summarize, neither method performs ideally, in the sense that both have clusters where some members that are rather poorly represented by the centroids. The obvious way to improve the fidelity of the clusters is to increase the number of clusters, or possibly do more re-initializations. However, the methods seem to complement each other, each to an extent balancing out the others weaknesses, and are useful as starting points for human analysis. ### CBG-like profiles As an example of the sort of analysis facilitated by these kinds of clustering techniques, we decided to perform an in-depth examination of the family of asymmetric blue-lobed single-peaked Stokes I profiles found in Fig. 2 (exemplified by cluster number #70 and #72) and Fig. 3 (exemplified by cluster number #64 and #65). These profiles are reminiscent of the chromospheric bright grains (CBGs) seen in the Ca ii H and K lines, see for instance (Carlsson and Stein, 1997; Mathur et al., 2022) and references therein, so we call them CBG-like. Fig. 4 shows the Stokes I and Stokes \(V\) signals (with each profile normalized to its nearby continuum Stokes I value), as well as the stratification of temperature, line-of-sight velocity, and line-of-sight magnetic field strength for all the profiles belonging to \(k\)-means cluster #70 and #72. The atmospheric quantities are plotted as a function of the logarithm of optical depth for radiation at wavelength 500 nm (5000 A), \(\log(\tau_{5000})\). Throughout this paper we use the convention that positive heights, velocities and vertical magnetic field components point outwards from the solar surface. Each row of Fig. 4 corresponds to one cluster, and the profiles are stacked along the vertical axis for each panel. The \(k\)-Shape clusters #64 and #65 are shown in a similar fashion in Fig. 5. Looking at the intensities we see that the clusters are indeed well constrained for the most part. The \(k\)-means method produces clusters where the emission peak is at approximately the same wavelength throughout each cluster, but with some variance in the other features of the profile shapes. The \(k\)-Shape method, on the other hand, retrieves clusters where the location of the emission peak varies considerably in wavelength, but the shapes in each cluster seem more consistent in their shapes. For Figure 3: \(k\)-Shape clusters for synthetic Ca ii 854.2 nm intensity profiles, using 100 clusters on \(z\)-normalized profiles. The red line is the cluster centroid from \(k\)-Shape, and the black lines are all individual profiles assigned to the cluster. The grey line is a visual aid that denotes the position of \(\lambda_{0}\) instance, the wavelength distance of the slope from peak to bottom seems to be more regular, and the red-side absorption features show less variance. As for the Stokes \(V\) profiles, with both methods the wavelength positions of the strongest Stokes \(V\) signals seem to coincide with the sharpest changes in the intensity as one might expect from the Weak Field Approximation. There does not, however, seem to be any other universal tendencies in Stokes \(V\) across all the CBG-like clusters. Similarly for the stratification of the line-of-sight magnetic field strengths, there do not appear to be clear tendencies neither within nor across the clusters. This suggests that the structure of the vertical magnetic field component does not play a direct role in the formation of these CBG-like Stokes I profiles. What does seem to be common to all the clusters, and therefore important for the formation of these profile shapes, is the depth-stratification of temperature and line-of-sight velocities. Mostly we see a temperature increase in the atmosphere, followed by a large velocity gradient slightly higher up. Mostly this manifests as upflowing material from below meeting downflowing material from above, but not exclusively as there are some instances of faster downflows from above meeting slower downflows, i.e. there is not necessarily a sign change in the vertical velocity, but there is a significant change in speed. That the temperature increase occurs deeper in the atmosphere than the velocity gradient, as well as the fact that the absolute values of the velocity are less important for the formation of these shapes than the presence of a strong gradient, is more easily seen with the \(k\)-Shape clusters as each of them contains the CBG-like profile shapes at a range of Doppler shifts. In any case, the correlation between the temperature increase, the velocity-gradient and the profile shape is certainly made clearer when comparing the results of both clustering methods. In terms of explaining the formation of these profiles, we are reminded of the interpretation of Ca ii K and H bright grains provided in (Carlsson & Stein, 1997) as signatures of acoustic shocks propagating upwards through the chromosphere, with the asymmetry being caused by Doppler shifts of the opacity across the shock front. The increased temperature enhances the local source function, which produces enhanced emission. The velocity gradient to more rapidly downflowing material above the heating event causes an opacity shift as the absorbing material is shifted to redder wavelengths, letting the bluer part of the profile escape while attenuating the redder part. A point of note is that the correlation between the atmospheric structure and the CBG-like profile shapes is apparent straight from the clustering when we have access to underlying atmosphere. This allowed a qualitative interpretation of the profiles' formation without having to resort to using response functions or contribution functions, which are ill-defined for the case of 3D radiative transfer. ### Double peaked profiles As another example, we now consider the double peaked profiles seen in \(k\)-means clusters #98 and #100, and in \(k\)-Shape Figure 4: Stokes \(I\) and \(V\) profiles for two clusters, along with some atmospheric parameters for their simulation columns. Here showing the \(k\)-means clusters #70 (_top row_) and #72 (_bottom row_) from Fig. 2, both of which have CBG-like profiles. All profiles for each cluster are stacked along the vertical axes of the plots, so the y-axis merely counts the profile number. The left column shows the continuum-normalized intensity versus wavelength from line core. The second-from-left column shows the continuum-normalized Stokes \(V\) profiles. The last three columns show, respectively, the temperature, the line-of-sight velocity, and the line-of-sight magnetic field strength, as a function of \(log(\tau_{5000})\). Figure 5: Same as Fig. 4, but for the \(k\)-Shape clusters #64 (top) and #65 (bottom) from Fig. 3. Figure 6: Same as Fig. 4, but for the \(k\)-means clusters #98 (top) and #100 (bottom) from Fig. 2. clusters #97 and #100. Similar to Figs. 4 and 5, the continuum-normalized intensity and Stokes \(V\) signals, as well as the height-stratified temperature, line-of-sight velocity, and line-of-sight magnetic field strength for all the individual profiles in each cluster is shown in Figs. 6 and 7 for the \(k\)-means and \(k\)-Shape clusters, respectively. Once again the clusters, on the whole, seem fairly well constrained regarding the shape of the intensity profiles. Here, there seems to be a larger variation in the absolute values of the intensities compared to the previous example. This sort of variation is not unexpected; since the \(z\)-normalization scales each profile independently to have a standard deviation equal to one our clusters are relatively insensitive to amplitudes, focusing instead on the shapes. Comparing the methods, we see they mostly recover the same profiles. An exception is that the \(k\)-means cluster #98 in the top row of Fig. 6 has some unique profiles around profile number 300 which appear to have either a very weak left peak or only a single peak on the right, followed by a prominent absorption feature to the right of the rightmost peak. Looking at the temperature and velocity structure for these atypical profiles with suppressed left peaks, it appears they have a temperature enhancement coinciding in height with a moderate downflow. This temperature enhancement persists upwards through a velocity gradient to a region of strong upflow, before it hits a very strong downflow. Their formation can potentially be explained in the same manner as the CBG-like profiles; but with an oppositely signed velocity gradient, and with the strong downflow above the upflow causing the additional strongly redshifted absorption feature. Returning to the general behavior of the clusters, we find that the Stokes \(V\) profiles seem to behave as expected from the weak-field approximation, in that they follow the behaviour of the intensity profiles. There is, however, a rather interesting region between profile number 200 to 300 in the bottom row of Fig. 6, where the rightmost Stokes \(V\) signal is very low despite a gradient in the intensity, and the vertical magnetic field component has a sign change around \(\log\tau_{5000}=-4\). The temperature structure of the atmosphere is more varied for the double peaked profiles, compared to the CBG-like profiles. There are both regions of temperature enhancements with little variation spanning decades in \(\log\tau_{5000}\), and hot regions bounded by colder plasma above and below. The common feature for all these double peaked profiles is enhanced temperatures in the range of \(-5<\log\tau_{5000}<-3\). That was also the case for the CBG-like profiles, though the CBG-like profiles seldom showed these colder layers above the first strong temperature increase. The vertical velocities are also rather varied in their structure, but three general features stand out compared to the CBG-like profiles from before. Firstly, the shift from upflows (or weak downflows) to strong downflows at the top tends to occur at a higher point in the atmosphere. Secondly, the starting points for the temperature enhancements coincide with slower plasma velocities and weaker velocity gradients, as opposed to the CBG-like profiles where the temperature increase starts slightly below strong velocity gradients. Thirdly, we note that the second velocity layer from the top, roughly \(-5.5<\log\tau_{5000}<-4.5\), typically shows low to moderate velocities and fairly modest gradients. As such, the effect of opacity shifting in this layer is less, and both intensity peaks due to the temperature enhancements survive. Another noteworthy point, is that when these double peaked profiles do have downflows from the top extending deeper (to \(\log\tau_{5000}\approx-5.5\)), the downflows are very strong and there is a corresponding absorption feature on the red side of the reddest Figure 7: Same as Fig. 4, but for the \(k\)-Shape clusters #97 (top) and #100 (bottom) from Fig. 3. peak. A possible interpretation is that the previously discussed opacity shifting is so red-shifted in those cases, that it overshoots the red peaks from the slower flowing regions and therefore does not suppress them. Interestingly, and contrasting with the CBG-like profiles, the vertical component of the magnetic field does in many of these double peaked profiles display some correlations with the vertical velocities and temperature stratifications. To wit, there are areas of Figs. 6 and 7 where the velocities change signs coinciding with an appreciable gradient in vertical field strength to more negative (downward) values. Furthermore, the starting heights of the temperature increases coincide with the appearance of the stronger vertical magnetic field components; particularly obvious examples are profiles number 100 through 200 in the bottom row of Fig. 6, and profiles number 300 through 500 in the top row of Fig. 7. In summary, these double peaked profiles seem to arise from a range of different atmospheric conditions. The common features are increased temperatures in the low chromosphere/upper photosphere, coinciding with low or modest velocities and weak velocity gradients. This, combined with cospatial enhanced vertical magnetic field strengths, suggests that these profiles are not all caused solely by acoustic shocks, in contrast with the CBG-like profiles. Whether the cause of the heating is due to a magnetic phenomenon, or if we simply see already hot plasma being transported, is unclear from this analysis. ### The strongest Stokes \(V\) profiles We have so far focused on the clustering of intensity profiles, since the \(z\)-normalization scaled Stokes \(V\) signals of very different amplitudes to a misleadingly similar range. Many of our Stokes \(V\) profiles contained only very weak signals, and clustering according to the shapes of such weak signals should not be expected to provide much diagnostic information. However, by restricting ourselves to look only at the Stokes \(V\) profiles containing an (unsigned) amplitude larger than 0.5% of the nearby continuum intensity we could perform a clustering on profiles with similar strengths. Out of our \(512\times 512\) synthetic profiles, only 7054 (\(\approx 2.7\%\)) matched that selection criterion. The results of \(k\)-means and \(k\)-Shape clustering with \(k=20\) clusters on this subset of Stokes \(V\) profiles are shown in Fig. 8 and Fig. 9 respectively. In this case, we deliberately selected a rather low number of clusters. This was partly done to avoid having clusters with very few members considering our reduced dataset, and partly to compare the performance of the two methods when using a very limited, and possibly too low, number of clusters. It is obvious from looking at Figs. 8 and 9 that 20 clusters is not sufficient to capture all the complexities present in the profiles with either method, though the clusterings do reproduce the primary features of the profiles. Comparing these two clustering results reveals some interesting differences. Most noticeably, not all shapes are common to both methods. The double peaked Stokes V profiles of cluster number #8 and #10 in the \(k\)-Shape result are not retrieved as a separate class by the \(k\)-means method; instead they are mixed into most of the \(k\)-means clusters, though primarily into #1, #7, #10, #12 and #17. On the other hand, the valley-peak-valley shape apparent in cluster number #16 from the \(k\)-means method does not appear in the \(k\)-Shape case. Looking in more detail at the individual profiles comprising that cluster, we find almost no profiles with a shape similar to that of the cluster mean. The triple-lobed shape of the cluster mean (marked in red) is instead mostly a mix of valley-peak and peak-valley shapes. In this case, the \(k\)-Shape centroids are more faithful representations of the shapes picked up by each cluster. In general, the clusters found by \(k\)-means contain one dominant feature, like a peak, a dip, or both, at a certain wavelength position with considerable variation in the rest of the signal. Furthermore, looking at cluster #13 or #16 in Fig. 8 we see that when the dominant feature in the cluster is multi-lobed, it might actually be a mix of single-lobed and multi-lobed signals grouped together, so long as their lobes occur at the same wavelength. This type of shape-mixing does not happen as readily with \(k\)-Shapes, contrast \(k\)-means cluster #13 with \(k\)-Shape cluster #15 and #17. Also, \(k\)-Shape seems to retrieve profiles with more community also at the weaker parts of the signal; compare for instance \(k\)-means clusters #5, #10 and #19 with \(k\)-Shape clusters #1, #5 and #13. \(k\)-Shape does, however, occasionally struggle when excessive shifts of the signal causes clipping of the features at the edges, which can be most easily seen in cluster #1, #19 or #20 of Fig. 9. While it is by no means perfect, we find, in conclusion, that \(k\)-Shapes performs markedly better than \(k\)-means at identifying shapes with this particular combination of complex signals and low number of clusters. How well that observation generalizes to other datasets, or cluster numbers, or both, is not clear, and beyond the scope of the current work. It does, however, indicate the type of problems where \(k\)-Shape can potentially provide an advantage over \(k\)-means. As a note, we have also performed this clustering experiment with \(k\)-means on the continuum-normalized Stokes V profiles and found that their behavior is very similar to the \(z\)-normalized case discussed above. ## 4 Discussion and Conclusions We have used the \(k\)-means and \(k\)-Shape clustering techniques to group according to profile shape synthetic Ca ii intensity and Stokes \(V\) profiles, generated by 3D radiative transfer calculations from a 3D MHD simulation. Using \(k=100\) clusters for the intensities resulted in both methods retrieving qualitatively similar 'families' of clusters. While the \(k\)-means method produced clusters whose features were strongly coherent with regard to wavelength, the \(k\)-Shape method, being insensitive to Doppler shifts, produced clusters where the same shape appeared over a range of wavelength shifts. Regarding the methods' shortcomings, we found that \(k\)-Shape occasionally would mislabel some profiles by clipping the signals at the edges when comparing across Doppler shifts, while \(k\)-means at times would lump rather differently shaped profiles together so long as their strongest feature occurred at the same wavelength. Armed with full knowledge of the simulation's atmospheric parameters, we took an in-depth look at a particular set of profile shapes and arrived at an explanation of their formation by looking at the correlations in the underlying atmospheric structure. We remark that the most interesting aspect of this exercise was not the description itself of how those profile shapes are formed, but rather how we arrived at it. In that use case, there did not appear to be much benefit in using one method over the other in terms of the results; though \(k\)-means was significantly quicker computationally. However, we do note that using both methods gave a stereoscopic view of the data, making it easier to determine which atmospheric quantities were important. Doing a clustering analysis of the Stokes \(V\) profiles, based on their shapes, proved difficult due to the large variations in signal strength being masked by the \(z\)-normalization required by \(k\)-Shape, causing strong and weak signals to appear deceivingly similar. Restricting ourselves to a subset of the strongest Stokes \(V\) profiles, we performed a clustering with \(k=20\) clusters using both methods. We found that the methods showed the same tendencies as with the intensity, but more strongly pronounced due to the lower number of clusters and more complex shapes. In this setting we found that \(k\)-means clearly performed qualitatively worse than \(k\)-Shape at creating clusters with coherent shapes; though is difficult to quantitatively compare the methods since they use very different metrics. In conclusion, \(k\)-Shape seems interesting for use cases where one wants human interpretation and small numbers of clusters. Another interesting possibility is to use the \(k\)-Shape distance metric to search an observation or simulation for the profiles with shape most similar to a certain prototype, for example when trying to detect Ellerman bombs. We want to stress that \(k\)-Shape is, however, not at all suited to usage cases like (Sainz Dalda et al., 2019), where the purpose of clustering is to speed up inversions, as the centroids found by \(k\)-Shape do not correspond to a definite Doppler-shift nor to an absolute intensity. In those cases, Figure 8: \(k\)–means clusters for Stokes \(V\) profiles, using 20 clusters on \(z\)-normalized Stokes \(V\). The red line is the cluster centroid (average), and the black lines are all individual profiles assigned to the cluster. The blue line is a visual aid that denotes the position of \(\lambda_{0}\). The bottom two rows show all the \(z\)-normalized Stokes \(V\) profiles belonging to the corresponding clusters in the two top rows, with the individual profiles stacked along the vertical axis. It should be noted that the clusters are not equally populated, so the grey-scale maps will have different densities of profiles along the vertical axis. means is the better option, and one can easily increase the number of clusters beyond what a human can reasonably process. For a qualitative clustering, aimed towards human interpretation and with a comparatively small number of clusters, we find that \(k\)-Shape can be a useful complement to, and sometimes better than, the more well-known \(k\)-means method. ###### Acknowledgements. The authors wish to thank Mats Carlsson for providing the Bifrost atmosphere used in this paper. We also wish to thank the anonymous referee for comments and suggestions that improved the clarity of this manuscript. This work has been supported by the Research Council of Norway through its Centers of Excellence scheme, project number 262622. Computational resources have been provided by UNINET Sigma2 - the National Infrastructure for High Performance Computing and Data Storage in Norway. The computations were enabled by resources provided by the Swedish National Infrastructure for Computing (SNIC) at the PDC Centre for High Performance Computing (PDC-HPC) at the Royal Institute of Technology partially funded by the Swedish Research Council through grant agreement no. 2018-05973.

2304.02886

Automatic ICD-10 Code Association: A Challenging Task on French Clinical Texts

Automatically associating ICD codes with electronic health data is a well-known NLP task in medical research. NLP has evolved significantly in recent years with the emergence of pre-trained language models based on Transformers architecture, mainly in the English language. This paper adapts these models to automatically associate the ICD codes. Several neural network architectures have been experimented with to address the challenges of dealing with a large set of both input tokens and labels to be guessed. In this paper, we propose a model that combines the latest advances in NLP and multi-label classification for ICD-10 code association. Fair experiments on a Clinical dataset in the French language show that our approach increases the $F_1$-score metric by more than 55\% compared to state-of-the-art results.

Yakini Tchouka, Jean-François Couchot, David Laiymani, Philippe Selles, Azzedine Rahmani

2023-04-06T06:31:54Z

http://arxiv.org/abs/2304.02886v1

# Automatic ICD-10 Code Association: A Challenging Task on French Clinical Texts ###### Abstract Automatically associating ICD codes with electronic health data is a well-known NLP task in medical research. NLP has evolved significantly in recent years with the emergence of pre-trained language models based on Transformers architecture, mainly in the English language. This paper adapts these models to automatically associate the ICD codes. Several neural network architectures have been experimented with to address the challenges of dealing with a large set of both input tokens and labels to be guessed. In this paper, we propose a model that combines the latest advances in NLP and multi-label classification for ICD-10 code association. Fair experiments on a Clinical dataset in the French language show that our approach increases the \(F_{1}\)-score metric by more than 55% compared to state-of-the-art results. natural language processing, icd-10, clinical document, unstructured data, multi-label classification, supervised learning, health, transformers ## I Introduction For a more accurate long term follow-up, patient's stay in a health center is usually reported in a digital documents which constitute the patient's medical record. Written by the patient's physicians, it is composed of operating reports, clinical notes, liaison letters etc. In a large number of countries each patient record is then classified according to the International Classification of Diseases (ICD). ICD is a medical classification system used for coding diseases and other health-related conditions. It is maintained by the World Health Organization (WHO) and is widely used globally as a standard for tracking health statistics, billing for medical services, and conducting research. In its \(10^{th}\) edition [21], ICD is organized into chapters based on different body systems and disease categories. The chapters are further divided into subcategories that provide more specific information about the condition being coded. Each code consists of an alphanumeric string that includes a category code, a subcategory code, and a descriptor code (up to seven characters). The ICD-10 classification system is used by healthcare providers and organizations worldwide to standardize the coding of medical conditions and facilitate the sharing of health information across different systems and platforms. This classification is the common foundation for epidemiology, public health research, and healthcare management. In addition, the reimbursement of medical expenses by public or private insurance companies directly depends on the codes associated with these medical records. This makes it even more important to associate the right codes with each patient's record. Finally, it should be noted that a more or less complex patient record can generate several ICD-10 codes Typically, in a hospital, the responsibility for the ICD-10 classification falls on the medical coders. Staff performing this task is specially trained professionals who use medical documentation to assign the appropriate ICD-10 codes to medical records. Medical coders work closely with healthcare providers, nurses, and other staff to ensure that the medical records are accurately encoded into this classification. In some hospitals the ICD-10 classification is performed by the physicians. However, regardless of how medical coding is managed, accuracy, and attention to detail are crucial to ensure that the data generated is reliable and useful for patient care and management. That is why automatically associating ICD codes to a medical record is a task that has been widely addressed in medical research in recent years [6, 2, 28, 9, 11]. With the recent advances in Natural Language Processing (NLP) and since medical records are unstructured medical documents, it makes sense to apply these theoretical and technological advances in the context of ICD-10 classification. Clearly, the emergence of the Transformers architecture [27, 10] has taken natural language processing to a new precision level. Several works have shown that the representations produced by these models are the most accurate and it is the most used architecture today in a large number of machine learning tasks (from text, to computer vision and time series) [27, 13, 16]. Nevertheless, ICD-10 automatic classification is a multi-label text classification task with tough challenges. For instance, the ICD-10 classification consists of about 140,000 codes (procedure codes and medical codes). Unless one has a huge dataset, extremely important physical resources, and an extremely long period of time, it seems to be unrealistic to believe that one could associate to a patient record one of the 140,000 existing codes with a high degree of accuracy. This large number of labels clearly stresses existing deep learning models to their limits. Another challenge is the size of the medical notes which far exceeds the usual limit of transformer architectures (typically \(512\) tokens). Finally, working on non-English data is also challenging since the vast majority of open-source models available are trained on English corpus. In this paper, we propose to address the three previous challenges for the ICD-10 classification of French medical records. We developed a deep learning model that combines the latest advances in Natural Language Processing. This approach makes it possible to associate a non-negligible part of the existing ICD-10 codes on French-language patient records with an \(F_{1}\)-score outperforming with more than 55% latest state of the art approach. This paper is organized as follows. Section II starts with recalling state of the art of associating ICD codes to medical records. Section III presents the dataset used on the one hand to validate our approach and on the other hand to fairly compare the \(F_{1}\)-scores obtained by our approach with those obtained with already existing approaches. The architecture of our ICD code association model is presented in Section IV. Results are presented and analyzed in Section V. Concluding remarks and future work are finally given in Section VI. ## II Related Work ### _Natural Language Processing_ NLP has significantly evolved in recent years with the joint appearance of the Transformers model [27] and their generalization ability to transfer learning. ELMo [23] and BERT [10] have shown this effectiveness which provides more accurate contextualized representations. Several pre-trained models then appeared such as BERT, RoBERTa [18].... These models are pre-trained on a large amount of general domain English text to capture the ability to model text data, and then refined on common classification tasks. In French two main models have been proposed i.e FlauBERT [14], CamemBERT [19]. Note that some multi-lingual models also exist such as XLM-R [7]. Some models are also trained on domain-specific text corpus. For example, ClinicalBERT[1] and BioBERT[15] have been trained on medical data to address medical domain tasks. Unfortunately, there is no such model in the French language, leading to a gap between the usage of machine learning approaches on French documents compared to the same approach in English ones. In general, Transformers models have a limited input size (\(512\) tokens in pratiee). In the case of clinical documents this limit can become very penalizing since a typical patient document is generally much larger than \(512\) words or tokens. In [22] the authors proposed some hierarchical methods to tackle this problem. They divided the document into several segments that can be processed by a Transformers. Then the encoding of the segments is aggregated into the next layer (Linear, recurrent neural networks or other layer of Transformers). Recently, the sparse-attention system i.e. the _LongFormer_ model has been proposed in [3]. It is composed of a local attention (attention between a window of neighbour tokens) and a global attention that reduces the computational complexity of the model. They can therefore be deployed to process up to 4096 tokens. ### _ICD Code Association_ The automatic association of ICD codes is one of the most addressed challenges in medical research. With the emergence of neural networks and the evolution of natural language processing, several authors have tried to tackle this task. [6] and [2] used recurrent neural networks (RNNs) to encode Electronic Health Records (EHR) and predict diagnostic outcomes. On the other hand, [25] and [20] have used the attention mecanism with RNNs and CNNs to implement more accurate models. The work of [29] and [26] present various ways to consider the hierarchical structure of codes. [29] used a sequence tree LSTM to capture the hierarchical relationship between codes and the semantics of each code. [5] proposed to train the integration of ICD codes in a hyperbolic space to model the code hierarchy. They used a graph neural network to capture code co-occurrences. LAAT [28] integrated a bidirectional LSTM with an attention mechanism that incorporates labels. EffectiveCAN [17] used a squeeze-and-excitation network and residual connections as well as extraction of representations from all layers of the encoder for label attention. The authors also introduced focal loss to address the problem of long-tail prediction with \(58.9\)% of \(F_{1}\)-score on MIMIC 3 [12]. ISD [30] used shared representation extraction between high frequency layers and low frequency layers and a self-distillation learning mechanism to mitigate the distribution of long-tailed codes. Recently [11] proposed the PLM-ICD system that focuses on document encoding with multi-label classification. They used an encoding model based on the Transformers architecture adapted to the medical corpus. Associating ICD-10 codes is finding the codes corresponding to medical documents in a large set of codes. For instance MIMIC 3 [12] contains more than 8,000 codes, and handling such large set of labels in classification is a challenging problem in machine learning. To overcome this problem, the authors used the Label-Aware Attention (LAAT) mechanism proposed in [28] which integrates labels in the encoding of documents. Finally, to solve the problem of long sequences they used the hierarchical method. PLM-ICD is the current state-of-the-art model that achieved \(59.8\)% of \(F_{1}\)-score on MIMIC 3 [12] and \(50.4\)% on MIMIC 2 [24]. In French, [9] proposed Convolutional Neural Networks (CNN) models with multi-label classification to automatically associate ICD-10 code. The authors used FastText [4] vectors with the skip-Gram algorithm for the encoding of documents. They first considered all the final labels of the dataset, then grouped them into families to reduce the number of classes. This model is trained on a private dataset of 28,000 clinical documents and reached \(39\)% of \(F_{1}\)-score with 6,116 codes and \(52\)% with 1,549 codes. ## III Dataset This work is in collaboration with The Hopital Nord Franche-Comte (HNFC), a French public health center that provided us with patient stays. For privacy reasons, all our experiments were conducted on site and no data was taken out of the hospital. A patient's stay is a set of successive visits in possibly different departments of the hospital. Each department produces a clinical document that describes the patient's stay in that department. These clinical documents are used by the medical coding specialists to associate the corresponding ICD-10 codes. We finally obtain a set of unstructured textual documents corresponding to the global stay of the patient to which a set of codes is associated. As clinical documents, we have for example operating reports, discharge letters, external reports or clinical notes. The obtained dataset, further denoted as ICD-10-HNFC dataset is therefore a database of groups of medical documents with associated codes. This system is well illustrated in Fig. 1. ICD-10-HNFC dataset is built for supervised deep learning. In supervised learning, to have an accurate model, there are several factors to consider. Is there enough training data? Is the number of classes consistent with the volume of data available? Is the frequency of classes in the dataset balanced? It is always difficult to find the perfect dataset that answers all these questions. In this paper, we worked not only on the main dataset, which consists in associating the raw ICD codes it contains but also on the sub-datasets such as associating the most frequent codes or code families instead of the raw codes. ### Class Reduction As mentioned, the ICD is a classification system that is composed of thousands of codes. Given the large number of labels present in our basic dataset (shown in Table I), it is difficult to approach this classification task by considering all the classes present. By doing so, the results of the constructed model will be far from perfect. The most precise models to date in English for ICD-10 code association is PLM-ICD which reached \(59.8\)% on MIMIC 3 with 8,922 labels [12] and \(50.4\)% on MIMIC 2 with 5,031 labels [24]. This proves the difficulty of this task. The first sub-dataset consists in reducing the codes to the first \(3\) characters seen as a family. Therefore, instead of considering the raw codes, we will group them into families. This reduces in a consequent way the number of classes to be treated by the model. This dataset is presented in Table I. We can see via the description "line Code with less than 10 examples" in Table I that the reduction of the classes not only allows to have a more reasonable number of classes but also increase the frequency of the codes in the dataset. ### Code Frequency Associating ICD-10 codes is a very frequent task in health centers. As a result, some codes occur more frequently than others. Finding the most frequent codes automatically can only be useful. Our second sub-dataset consists in building models based on the number of codes (\(K\)) that we consider more relevant. We evaluate the relevance based on the frequency of the code in the dataset. Thus, a model built on such a dataset will be able to associate the integrated codes with a better classification performance. ### Additional Label With the code frequency strategy (\(K\)), the dataset is therefore composed of entries whose association belongs only to the \(K\) most relevant codes. To keep the coherence of our dataset an additional label is introduced to represent the codes which are not considered as relevant (least frequent codes). So instead of having \(K\) classes, the model will have \(K+1\) ones. The additional class mean concretely in the association that there are one or more additional codes to associate. ## IV Model Architecture This section presents the different components of the model architecture we have developed and justifies the choices made to design it. As previously exposed, as we deal with the French language our choice was to fine-tune pre-trained transformer-based French models i.e. CamemBERT [19] and FlauBERT [14] for the implementation of the model architecture. ### _Global Document Representation_ As mentioned, Transformers main constraint is the limitation of the number of tokens present in an input sequence. Since the average size of the clinical notes of ICD-10-HNFC dataset exceeds this limit (\(747\) versus \(512\) as shown in Table I), basic Transformers cannot be used. Recently, [8] summarized the available methods for processing long sequences via Transformers. They can be summarized as hierarchical Transformers and sparse-attention Transformers in which we can find the _Longformer_ model of [3] early mentioned. _Longformer_ can process up to \(4096\) tokens per sequence allows to meet this limit. Unfortunately, there is no French pre-trained _Longformer_ Fig. 1: ICD-10-HNFC Dataset Construction model to date. Therefore, in this paper, we will use the hierarchical method to tackle this problem. Hierarchical Transformers[22, 8] are built on top of Transformers architecture. A document \(D\), is first divided into segments \([t_{0},t_{1},\ldots,t_{|D|}]\), each of which must have less than \(512\) tokens (the limit of Transformers). These segments are encoded independently using a typically pre-trained Transformers. We then obtain a list of segment representations which must be aggregated to obtain the whole document \(D\) representation. There are several ways to do this aggregation. The aggregator can be an average of the representations of all the segments of the document (mean pooling) or the maximum of the representations in each dimension of the segments (max pooling) or stacking the segment representations into a single sequence. The aggregated sequence thus serves as an input to the next layer. #### Iv-A1 Classification of a Large Number of Labels To overcome the problem of a large set of labels since ICD-10-HNFC contains more than 6,000 codes, we used the Label-Aware Attention (LAAT) system as in [11]. LAAT consists in integrating the labels into the document representation. Label-Aware Attention captures important text fragments related to certain labels. Let \(H\) be the stacking representation of an input sequence. First, a label-wise attention weight matrix \(Z\) is computed as follows: \[Z=tanh(VH)\] \[A=softmax(WZ)\] where \(V\) and \(W\) are linear transforms. The \(i^{th}\) row of \(A\) represents the weights of the \(i^{th}\) label. The softmax function is performed for each label to form a distribution over all tokens. Then, the matrix \(A\) is used to perform a weighted-sum of \(H\) to compute the label-specific document representation: \[D=HA^{T}\] The \(i^{th}\) row of \(D\) represents the document representations for the \(i^{th}\) label. Finally, \(D\) is used to make predictions by computing the inner product between each row of \(D\) and the related label vector. In this paper, several architectures were experimented such as the model without long sequence processing, the model with long sequence processing (max/mean pooling), and the model with LAAT. The global architecture is illustrated in Fig. 2. ## V Experiments and Analysis In this section, we present the results of the experiments conducted with the previously detailed architectures and dataset. We compare the results of recent works (PLM-ICD[11], CNN[9]) on the association of ICD-10 codes with the most precise model of this paper. To evaluate our model we use the most used performance measures in classification: Precision, Recall, \(F_{1}\)-score. The micro average system is used to obtain the aggregation of the performances. ### _Paper Models_ First, we conducted the experiments on the ICD-10-HNFC dataset with class reduction (\(1564\) labels) as detailed in Table 1. This experimentation is performed with all the architectures developed in this paper. They are listed in Table II. Then we trained another model on the global ICD-10-HNFC dataset (\(6101\) labels) with the architectures that obtained the highest \(F_{1}\)-score in the previous experiment. The results are shown in the Table II. The results confirm the effects of the different components that constitute our architectures. In summary, the LAAT approach outperforms the hierarchical methods with are better than the base truncated model. ### \(K\)_-based Models_ As detailed in Section III, different models have been trained based on a number (\(K\)) of labels (i.e. the most frequent codes). We present here the evaluation of these models with \(K\) in \([10,50,100,200]\). As shown in Table III, models are less and less accurate when we increase the number of labels (classes). This is simply due to the aggregation of performances. The more different codes there are, the fewer instances of each code there are in the dataset, and the less easy the contextualization is. ### _Comparison with other Works_ The Table IV shows the model with the highest \(F_{1}\)-score of this paper with the results of previous work on ICD-10 code association. It is difficult to compare the results, since these works do not use the same evaluation dataset and English works can benefit from specialized models such as ClinicalBERT [1]. For French baseline, we implemented and trained the model proposed in [9] on ICD-10-HNFC dataset. The result is shown in parallel with our proposal. Our model clearly outperforms the classification method used in [9]. On the same validation dataset, with class reduction (1564 labels) the \(F_{1}\)-score goes from 0.35 obtained with the model proposed in [9] to 0.55 with our proposal, i.e. an improvement of 57%. With the raw codes (6161 labels), the \(F_{1}\)-score goes from 0.27 to 0.45, i.e. an improvement of 66.6%. The difference in scores with the results of PLM-ICD can be explained by the use of a context specific (medical) Transformers which has a vocabulary more adapted to the content of the documents. ## VI Conclusion In this paper, we address the challenges of automatically associating ICD-10 codes to French clinical unstructured data. We have experimented several Transformers architectures to address the challenges of large input tokens and large numbers of labels. We therefore propose an ICD-10 association model that uses the latest advances in natural language processing and achieves the highest results in the French language to date. Our future work will focus on the use of Large Language Models and few-shots learning techniques to the ICD-10 classification. ## Acknowledgment This work is (partially) supported by the EIPHI Graduate School (contract ANR-17-EURE-0002).

2304.13753

Debris Rings from Extrasolar Irregular Satellites

Irregular satellites are the minor bodies found orbiting all four Solar System giant planets, with large semi-major axes, eccentricities, and inclinations. Previous studies have determined that the Solar System's irregular satellites are extremely collisionally evolved populations today, having lost $\sim$99 per cent of their initial mass over the course of hundreds of Myr. Such an evolution implies that the irregular satellites must have produced a population of dusty collisional debris in the past, which is potentially observable due to the resulting reprocessing of stellar light. In this paper we examine the signatures of the debris discs produced by extrasolar analogues of this process. Radiation pressure, quantified by the parameter $\beta$, is the driving force behind the liberation of dust grains from the planetary Hill sphere, and results in the formation of circumstellar dust rings, even in the absence of an underlying belt of asteroids in the system. Our simulated discs reproduce many of the same features seen in some classes of observed debris discs, such as thin ring morphology, a large blowout size, and azimuthal symmetry. We compare our simulated discs' radial profiles to those of the narrow dust rings observed around Fomalhaut and HR 4796A, and show that they can broadly reproduce the observed radial distribution of dust.

Kevin T. Hayakawa, Bradley M. S. Hansen

2023-04-26T18:00:04Z

http://arxiv.org/abs/2304.13753v1

# Debris Rings from Extrasolar Irregular Satellites ###### Abstract Irregular satellites are the minor bodies found orbiting all four Solar System giant planets, with large semi-major axes, eccentricities, and inclinations. Previous studies have determined that the Solar System's irregular satellites are extremely collisionally evolved populations today, having lost \(\sim\)99 per cent of their initial mass over the course of hundreds of Myr. Such an evolution implies that the irregular satellites must have produced a population of dusty collisional debris in the past, which is potentially observable due to the resulting reprocessing of stellar light. In this paper we examine the signatures of the debris discs produced by extrasolar analogues of this process. Radiation pressure, quantified by the parameter \(\beta\), is the driving force behind the liberation of dust grains from the planetary Hill sphere, and results in the formation of circumstellar dust rings, even in the absence of an underlying belt of asteroids in the system. Our simulated discs reproduce many of the same features seen in some classes of observed debris discs, such as thin ring morphology, a large blowout size, and azimuthal symmetry. We compare our simulated discs' radial profiles to those of the narrow dust rings observed around Fomalhaut and HR 4796A, and show that they can broadly reproduce the observed radial distribution of dust. keywords: stars: abundances - planets and satellites: formation - planets and satellites: dynamical evolution and stability - planets and satellites: gaseous planets - planets and satellites: composition ## 1 Introduction Planet formation is a dynamic process, wherein the growth of planets is accomplished via a prolonged history of interactions between smaller bodies, leading to scattering and collision (e.g., Lissauer, 1993). This process is particularly important during the latter stages of planetary assembly, as planetary systems settle down into their final configurations. Indeed, the process of dynamical clearing is thought to continue for some time after planets have reached their final masses, as the remnants of the source population are ground down and removed from the system (e.g., Goldreich et al., 2004). Stars in this stage of development often show evidence for extended, tenuous, populations of dust (Wyatt, 2008). These dust grains scatter and re-radiate light from the central star, and can be observed either by looking for infrared excesses or by imaging in scattered light. The lifetime of dust in such systems is short, limited by radiation pressure and Poynting-Robertson drag (e.g., Burns et al., 1979), but the observation of this material offers essential insights into the architectures of newly formed planetary systems. As a result, there have been substantial efforts to image such debris systems directly (see Hughes et al., 2018, for a recent summary). The results show a wide range of morphologies, from extended discs (e.g. those around \(\tau\) Ceti and HD 141569) to very narrow rings, such as those around the stars Fomalhaut, HR 4796A, and HD 141011. The variation in appearance presumably indicates some complexity in the evolution and outcome of the planetary assembly process, and there exist detailed models for the kinds of outcomes to expect (Wyatt, 2008; Krivov, 2010; Kenyon and Bromley, 2016; Lee and Chiang, 2016; Bonnefoy et al., 2021). Debris discs are usually modelled with a source population as a belt of planetesimals undergoing collisional evolution, where the velocity dispersion is stirred either by the development of larger bodies within the belt, or as the result of perturbations from planets in the system (e.g., Wyatt, 2008). These are natural analogues of the Solar system dust generated either by collisions in the Asteroid belt or the Kuiper belt, although the extrasolar systems are much more massive. However, there is a third Solar System small body population that is thought to have undergone substantial collisional evolution but has not yet been widely considered in the extrasolar context - namely the irregular satellites of the giant planets (Jewitt and Haghighipour, 2007). Evolutionary models of this population suggest that it could have been much larger in the past and could have generated a substantial population of dust during the course of losing \(\sim 99\) per cent of its original mass (Bottke et al., 2010). Indeed, such populations are thought to be an inevitable consequence of giant planet formation (Nesvorny et al., 2007) and the possible existence of irregular satellite clouds around extra-solar planets has recently been postulated to explain the curious properties of the exoplanet candidate Fomalhaut b (Kennedy and Wyatt, 2011; Kenyon and Bromley, 2016). These papers have focussed on the production of dust near the planet, but radiation pressure forces will cause much of the dust to spiral outwards into a circumstellar ring, and can therefore also contribute to the observed extended structures observed around many stars. Therefore, our goal in this paper is to examine the kinds of debris disc signatures one might expect from a source population initially confined to a planetary Hill sphere, and to explore their similarities and differences with those that result from more traditional population assumptions. An alternative to the traditional planetesimal disc model is particularly attractive in the case of the thinnest debris rings, such as those around Fomalhaut and HR 4796A, where the narrowness requires additional hypotheses such as shepherd satellites (Boley et al., 2012), instabilities involving a gaseous component (Lyra & Kuchner, 2013) or recent collisions (Olofsson et al., 2019). We will demonstrate that irregular satellite clouds naturally give rise to narrow rings and may therefore offer a more natural explanation for these structures, in the sense that the confinement of the original planetesimal population is due to the gravitational influence of the planet. The outline of this paper is as follows. In SS 2 we describe the dynamics of radiation pressure-driven dust in the reduced three-body problem, and examine the debris disc geometry that results if the source population of the dust is initially restricted to a planetary Hill sphere. In SS 3 we then introduce a model for a source population of dust which we combine with the dynamical model to build a model of a candidate debris disc so that we may explore the observational implications of this hypothesis. We then compare these features to the present state of the art observations of the two most prominent thin ring systems - Fomalhaut and HR 4796A - in SS 5. ## 2 Dynamics of dust generated in an irregular satellite swarm The scattering and absorption/re-emission of light in a debris disc is the action of dust particles in orbit about the star. In a traditional debris disc model, this dust is released in collisions between larger bodies in heliocentric orbits, and so reflects the heliocentric distribution of the parent bodies. Here we wish to examine the consequences when the dust is released by collisions between bodies that are localised in orbit around a planet. In addition to the radiation pressure force from the central star, their dynamics is also regulated by the gravitational influence of the planet. ### Single dust grain dynamics in the restricted three body problem with radiation pressure Dust particles have infinitesimal mass, so their dynamics can be treated accurately within the paradigm of the restricted three-body problem, sketched out in Fig. 1(Murray & Dermott, 1999). However, the stream of photons emanating from the central star is absorbed or scattered by the dust grains, and exerts a radiation pressure. This means that small particles experience a non-negligible additional radial force, which reduces the effective gravity of the central object (Schuerman, 1980) and fundamentally alters the geometry of the pseudo-potential that regulates the dynamics. We can relate this purely radial radiation pressure force to the stellar gravitational force using the formalism of Burns et al. (1979) as follows, \[F=-\frac{G(1-\beta)M_{1}m}{r_{13}^{2}}\dot{r}_{13}-\frac{GM_{2}m}{r_{23}^{2}} \dot{r}_{23}, \tag{1}\] where \(G\) is the Newtonian gravitational constant, \(\beta=|\boldsymbol{F}_{\rm rad}|/|\boldsymbol{F}_{\rm grav}|\) is the relative strength of radiation pressure compared to stellar gravity, \(M_{*}\) is the stellar mass, \(m\) is the dust grain mass, \(r_{13}\) is the distance from the grain to the star, \(r_{23}\) is the distance from the grain to the planet, \(\dot{\boldsymbol{r}}_{13}\) is the radial unit vector away from the star, and \(\dot{\boldsymbol{r}}_{23}\) is the radial unit vector away from the planet. For \(\beta>0\), the dust grains behave as if they'see' a star of reduced mass \((1-\beta)M_{*}\). The parameter \(\beta\) reflects the strength of the radiation pressure, and can be more precisely quantified as \[\beta=\frac{3L_{*}(Q_{\rm rad})}{8\pi GM_{*}c\rho D}, \tag{2}\] where \(L_{*}\) is the stellar luminosity, \(\langle Q_{\rm rad}\rangle\) is the wavelength-averaged radiation pressure coefficient, \(c\) is the speed of light, \(\rho\) is the mass density of grains, and \(D\) is the diameter of the grain. This \(\beta\) can be thought of as a proxy for grain size \(D\) if we assume a constant mass density \(\rho\) among dust grains, since \(\langle Q_{\rm rad}\rangle\) is of order unity. Koschny & Grun (2001) performed laboratory experiments involving collisions between silicates and found the mass density of resulting grains to be \(\rho=2.8\) g cm\({}^{-3}\). We assume a value for \(\langle Q_{\rm rad}\rangle\sim 0.5\) as a rough average from Liu & Schmidt (2018). \(\beta\) can thus be evaluated as \[\beta\approx 0.206\left(\frac{D}{1\ \mu{\rm m}}\right)^{-1}\left(\frac{L_{*}} {\rm L_{\odot}}\right)\left(\frac{M_{*}}{\rm M_{\odot}}\right)^{-1}, \tag{3}\] where we have assumed typical values of luminosity and mass of a G-type star such as the Sun. The dynamics of the test particle in the co-rotating frame of the reduced three-body problem is governed by a pseudo-potential that accounts for both the gravity of the two massive bodies and the centrifugal force (Murray & Dermott, 1999). The pseudo-potential defines a set of 'zero-velocity curves' which restrict the motion of a test particle, depending on its initial conditions. The fact that the radiation pressure only scales the effective mass of the central star means that the same formalism applies here, but the revision of the coefficients in the pseudo-potential results in an important qualitative difference. Although the direct gravitational force felt by the dust is reduced by the radiation pressure, the orbital velocity of the planet is not similarly affected, and so the relative contributions of the three different components of the pseudo-potential change with \(\beta\). In particular, at fixed mass, there is a critical \(\beta\) above which the L\({}_{2}\) point becomes the global potential minimum (instead of L\({}_{1}\), as in the \(\beta=0\) case). This distinction is important because it is Figure 1: Schematic of the restricted three-body problem. The x- and y-axes make up the corotating centre-of-mass frame, rotating at an angular velocity \(n\), centered at point \(O\). The x’- and y’-axes make up the inertial centre-of-mass frame. \(M_{1}\) and \(M_{2}\) are massive bodies with the third body located at point \(P\). this minimum that decides the direction in which dust, grinding down in a collisional cascade, leaves the Hill sphere and enters heliocentric orbit. To illustrate the effects of this change in geometry, let us consider two different physical scenarios: one where radiation pressure is not important (i.e., turned 'off'; \(\beta=0\)) as in panel (a) of Fig. 2 and another where radiation pressure is important (i.e., turned 'on') as in panel (b) of Fig. 2. Without loss of generality, we take \(\beta=0.1\) for the radiation pressure scenario. If we wish to derive the minimum velocity of escape, we see that, in panel (a), particles will more readily escape through the L\({}_{1}\) Lagrange point than L\({}_{2}\). This behavior is well studied, such as in the case of Roche lobe overflow where mass transfer can occur between two bodies in a binary system. However, when radiation pressure is non-negligible, we see in panel (b) that the lowest velocity particles to escape now overflow L\({}_{2}\). Thus, the addition of radiation pressure into our equations of motion changes the physics from accretion on to the star to ejection of material outside the orbit of the planet. This is a consequence of the weakened effective gravity of the central star, which shifts its contribution to the pseudo-potential inwards and changes the relative heights of the L\({}_{1}\) and L\({}_{2}\) points. In Appendix B we review more thoroughly how this change in topology occurs as \(\beta\) is changed. There are essentially three fates for a dust grain: (i) accretion on to the planet, (ii) accretion on to the star, or (iii) escaping to infinity. Depending on the initial conditions, the dust grain may simply spiral into the planet after an irregular satellite collision, coating the planet, and any satellites, with dark dust and affecting its albedo (Burns et al., 1979; Bottke et al., 2010). Accretion on to the star may occur as the result of another radiation-related process: Poynting-Robertson (PR) drag. This is a consequence of the loss of angular momentum due to reradiation of absorbed energy by the dust, but is not taken into account here because it occurs on a longer time-scale than the direct dynamical effects of radiation pressure, and generally affects large particles more. We will ignore both PR drag and circumstellar collisions between dust grains in our simulations because their respective time-scales are longer than the ejection time-scales of individual dust grains from the system, as shown in Fig. 3. Detailed calculations are performed in Appendix A. For our purposes, the most important outcome is escape from the Hill sphere through the L\({}_{2}\) point. Although the eventual outcome is escape to infinity, orbital integrations show that many trajectories spend multiple orbital periods in the vicinity of the outer edge of the relevant zero-velocity curve, before eventually spiralling outwards. This extended period of quasi-regular orbital behaviour thus gives rise to the observational appearance of a thin ring, allied with an exponential tail of orbits that extend out much farther. Such a configuration bears a qualitative resemblance to the 'birth ring + halo' model of many debris systems and we will examine its observational consequences below. ### Sample orbital integrations To better understand how this change in geometry reflects itself in orbital behaviour, let us examine the behaviour of a few representative test particles before building a large ensemble population. We perform our numerical integrations using the Mercury N-body integrator (Chambers, 1999). We start with the simplest case of \(\beta=0\), which represents a particle that is too large for radiation pressure to have an appreciable effect. Each particle originates in the Hill sphere of its parent planet, and receives a 3-D velocity vector of magnitude \(v_{\rm i}\), whose direction is oriented randomly. (We also investigated the effects of preferring prograde or retrograde orbits for our initial conditions, but found no significant change in the results compared to randomly oriented orbits.) After a few orbits around the planet many grains slip through the L\({}_{1}\) Lagrange point and 'bounce' along the inner edge zero-velocity curve. After several excursions around the star, the grain returns to the Hill sphere through the L\({}_{1}\) orbit in a messy, rapidly precessing orbit. In the absence of a dissipative mechanism, this behavior basically repeats itself over time, with the grain being gravitationally shared by the star and the planet. On longer time-scales, Poynting-Robertson drag will eventually decouple the particle from the Hill sphere and it will spiral into the star. Next we examine the case of radiation pressure turned 'on' with an intermediate strength of \(\beta=0.1\). The path of such a particle is shown in Fig. 4, for the case of a Jupiter mass planet on a circular orbit of semi-major axis 5 au. In this case, the particle spirals outwards - rather than inwards - and makes several cardiod-shaped excursions, roughly several planetary semi-major axes in size, as shown in panel (a). This is a consequence of the aforementioned change in the geometry of the pseudopotential, as shown in panel (b). Like in the \(\beta=0\) case, the grain will occasionally come to a sharp halt along the predicted zero-velocity curve. However, the fundamental alteration of the forbidden zone, caused by the addition of radiation pressure, means that the 'collision' occurs with the outer edge of the zero-velocity curve, not the inner one. In panel (c), we see that after a moderate number of dynamical time-scales, this behavior essentially repeats itself since the orbits all stay within \(\sim\)15 au. However, in panel (d), after a large number of dynamical time-scales, we see that the eccentricity of the grain has been pumped up dramatically, reaching an apoapsis up to \(\sim\)75 au, until it is effectively ejected from the system. These sample integrations illustrate that the dynamics of particles released from the planetary Hill sphere under the influence of radiation pressure can reproduce the basic birth ring configuration of debris disc models, even without the underlying birth ring of planetesimals. We wish now to expand this into a proper model for debris discs. This means we need a more detailed source model, which will link the properties of the dust to the new underlying planetesimal population - the irregular satellite population. This is the focus of SS 3. Figure 2: _Panel (a):_ forbidden zone (in blue) for a Jacobi constant of \(C_{J}=3.038\) when radiation pressure is not included (\(\beta=0\)). _Panel (b):_ forbidden zone (in blue) for a Jacobi constant of \(C_{J}=3.824\) when radiation pressure is taken into account (\(\beta=0.1\)). The orange circle (not to scale) represents the location of the giant planet. In panel (a), we note that the Hill spheres of the star and planet overlap at the L\({}_{1}\) Lagrange point. Thus, dust grains originating in the planetary Hill sphere are permitted to escape into a circumstellar orbit. In contrast to panel (a), we note that in panel (b), there is now an opening at the L\({}_{2}\) Lagrange point for the dust grains to escape from. ### Forbidden zone thickness as a function of radiation pressure An interesting question is to ask how the results of our simulations will change depending on planet mass, since the study of exoplanets around nearby stars has discovered a great variety in planetary properties. The solutions to the restricted three-body problem depend not just on the mass of the secondary body (the planet), but specifically on the mass ratio of the secondary body to the primary body (\(\mu=M_{1}/M_{2}\)). Thus, a Saturn-like planet orbiting an M dwarf may have similar dynamics to those of a Jupiter-like planet orbiting a Sun-like star, if both systems have a mass ratio of \(\mu\sim 0.001\). Increasing radiation pressure generally has the effect of increasing the forbidden zone thickness as seen in the first column of Fig. 5. However, while the overall thickness increases, we can see that the radii of both the inner and outer edge of the forbidden actually decrease. In other words, the radius of the outer edge is decreasing by a smaller amount than that of the inner edge. For a more comprehensive look at how the forbidden zone changes as a function of both radiation pressure strength and mass ratio, interested parties may refer to the discussion in Appendix C. ## 3 Generation of dust in an irregular satellite swarm We assume that the particles whose orbits we track originate from collisions between irregular satellites orbiting around the giant planet. Irregular satellites revolve around their parent planet at relatively large distances compared to other moons (e.g., Bottike et al., 2010), so it is natural to characterize their distances in units of Hill radii, given by \(R_{\rm H}=a_{\rm p}[M_{\rm p}/(3M_{*})]^{1/3}\), where \(a_{\rm p}\) is the planetary semi-major axis and \(M_{\rm p}\) is the planetary mass. The original orbits of irregular satellites are believed to be roughly isotropically distributed (Jewitt & Haghighipour, 2007), so we investigate both prograde and retrograde orbits around the parent planet, which are typically found at \(r_{23}/R_{\rm H}\) values of \(\sim 0.1\) to \(0.5\), where \(r_{23}\) is the distance between the planet and the dust grain. Thus, we use those upper and lower limits to randomly generate starting locations for the dust grains. Figure 4: Orbital trajectory of a single \(1~{}\mu\)m dust grain (in blue) overplotted on its respective \(\beta=0.1\) zero-velocity curves (in black). The orbits are shown at various evolutionary stages: \(t=1e2\) yr in panels (a)–(b), \(t=1e4\) yr in panel (c), and \(t=1e5\) yr in panel (d). Panel (b) represents a zoom-in of panel (a). Figure 3: PR drag, collisional, and ejection time-scales of circumstellar dust grains as a function of \(\beta\), for two canonical examples discussed in this paper. _Left:_ For a Jupiter-mass planet orbiting at \(5.2\) au with \(10^{-3}~{}M_{L}\) of irregular satellites. _Right:_ For a Jupiter-mass planet orbiting at \(140\) au with \(1~{}M_{L}\) of irregular satellites. In this paper, we ignore PR drag and circumstellar grain collisions due to their larger time-scales. Collisions cannot always be ignored. We only ignore collisions in this paper because in the situations studied here, the ejection time-scale dominates. Detailed calculations are performed in Appendix A. We divide the discussion of initial velocities into magnitude and direction. We take the magnitude of the velocity to be 71 per cent of the Keplerian circular velocity at the debris's respective distance from the planet. Since it is a spherically symmetric cloud, we assume that the direction of the dust grain's velocity unit vector is random. Specifically, in polar coordinates \(\theta\) and \(\phi\), \(\cos(\theta)\) is distributed uniformly in [-1,1) and \(\phi\) is distributed uniformly in [0,2\(\pi\)). We find no significant difference between the qualitative results for orbits that are initially prograde or retrograde. Since the Keplerian velocity is given by \(v_{\rm kep}=(GM_{\rm p}/r_{23})^{1/2}\), the initial velocities are given by \[v_{\rm kep}=(2.248~{}{\rm km~{}s^{-1}})\left(\frac{M_{\rm p}}{\rm M_{\rm J}} \right)^{1/2}\left(\frac{r_{23}}{0.5~{}R_{\rm H}}\right)^{-1/2} \tag{4}\] where \(\rm M_{\rm J}\) is the mass of Jupiter. ### Rate of dust generation A population of irregular satellites will generate a collisional cascade, in which planetesimals are ground down to micron-sized dust grains. Collisions between the largest collisionally coupled bodies of size \(D_{\rm max}\) initiate the cascade, creating numerous medium-sized bodies that further collide with each other to produce successively smaller bodies. In the traditional context of a circumstellar debris disc, the smallest collisionally coupled body's size \(D_{\rm min}\) is determined by the strength of the central star's radiation pressure, and tends to be about 1 micron. This is often referred to as the blowout size. Dohnanyi (1969) found that a self-similar, steady-state cascade follows a power law differential size distribution governed by \[\frac{dN}{dD}\propto D^{-q}, \tag{5}\] where \(D\) is the spherically symmetric grain's diameter and \(q\approx 3.5\). A dust grain is no longer in a bound orbit around the star when the ratio of radiation pressure force to gravitational force is greater than 0.5 (e.g., Pawellek & Krivov, 2015), i.e., \[\beta\equiv\frac{F_{\rm rad}}{F_{\rm grav}}\geq 0.5. \tag{6}\] In the case discussed here, there is an additional consideration. Fragments from irregular satellite collisions will continue to participate in the collisional cascade as long as they orbit within the planetary Hill sphere. However, once the radiation pressure is strong Figure 5: Zero-velocity curves for both the classic restricted three-body problem (\(\beta=0\)) and with moderate radiation pressure included (\(\beta=0.1\)), as representative examples for a mixed mass ratio of \(M_{2}/M_{1}=0.001\) and initial velocity \(v_{\rm i}\) equal to 71 per cent of the local circumplanetary Keplerian velocity \(v_{\rm kep}\). The left and right columns represent zoomed in versions of the respective red squares shown in the middle column. The top row shows the contours for the minimum Jacobi constant that arise from the initial conditions while the bottom row shows the contours for the maximum Jacobi constant. Thus, all possible Jacobi constants fall between these two extrema. Since the Jacobi constant only depends on the square of the velocity, all possible directions of velocity have implicitly been considered as well as all isotropic orbital configurations. The main two findings here are that the forbidden zone thickness increases with increasing radiation pressure, but the radii of both the inner and outer edges of the forbidden zone shrink with increasing radiation pressure, as seen in the left column. In other words, the radius of the outer edge of the forbidden zone shrinks less than that of the inner edge. enough for the particle to escape the Hill sphere, the density of collision targets drops dramatically and the collisional time-scale becomes large. Thus, the minimum particle size in the cascade is set by the size for which the residence time within the Hill sphere equals the characteristic collision time for particles of that size. The residence time here is defined as the amount of time a dust grain spends in the Hill sphere at a given \(\beta\) before escaping. Conversely, this also sets a minimum \(\beta\) for the particles in the more extended debris disc and thus regulates their size. We can find the collisional time-scale \(t_{\rm coll}\) for any member of the collisional cascade from \(t_{\rm coll}=1/(n\alpha v_{\rm rel})\), where \(n\) is the number density of particles that cause catastrophic collisions, \(\sigma\) is the cross section, and \(v_{\rm rel}\) is the relative velocity between impactors. The number density of particles is given by \(n=N/V\), where \(N\) is the number of particles and \(V\) is the volume they occupy. We calculate \(N\) by integrating Eq. 5 from \(D/2\) to \(2D\), the range by which we define sizes that are capable of a catastrophic collision. Additionally, we normalize Eq. 5 by integrating the collisional cascade over mass via \[M_{\rm tot}=\int_{D_{\rm min}}^{D_{\rm max}}m\frac{dN}{dD}dD \tag{7}\] where \(m=(4\pi/3)(D/2)^{3}p\) is the mass of a body in the cascade. Since the irregular satellites are distributed isotropically in a spherical cloud, we take this volume to be some spherical shell with the radius the fraction \(f_{\rm tot}\) of the Hill sphere they occupy: \(V=f_{\rm tot}V_{H}=f_{\rm tot}(4\pi/3)R_{\rm H}^{3}\). Gravitational focusing is not important for submillimeter-sized particles, the cross section is just the geometric cross section: \(\sigma=\pi(D/2)^{2}\). Lastly, we take the relative velocity \(v_{\rm rel}\) to be of order the circumplanetary Keplerian velocity \(v_{\rm kep}\) since the orbital inclinations of irregular satellites are randomly oriented. Putting everything together, we find that the collisional time-scale is \[\begin{split} t_{\rm coll}&=(726\;{\rm yr})\;\left( \frac{\beta}{0.1}\right)^{-1/2}\left(\frac{M_{\rm tot}}{10^{-2}M_{\rm L}}\right) ^{-1}\left(\frac{\rho}{1\;{\rm g\;cm^{-3}}}\right)^{1/2}\\ &\quad\times\left(\frac{D_{\rm max}}{30\;{\rm km}}\right)^{1/2} \left(\frac{M_{\rm s}}{{\rm M}_{\odot}}\right)^{-5/3}\left(\frac{M_{\rm p}}{{ \rm M}_{\rm J}}\right)^{-2/3}\left(\frac{L_{\star}}{{\rm L}_{\odot}}\right)^{1 /2}\\ &\quad\times\left(\frac{(Q_{\rm rad})}{0.5}\right)^{1/2}\left( \frac{a}{{\rm a}_{\rm J}}\right)^{7/2}\left(\frac{f}{0.4}\right)^{1/2}\left( \frac{f_{\rm tot}}{0.098}\right),\end{split} \tag{8}\] where \(M_{\rm tot}\) is the total mass of the collisional cascade, \({\rm M}_{\rm L}\) is lunar masses, \(a\) is the semi-major axis of Jupiter, \(f\) is the orbital radius of the body as a fraction of the Hill radius (\(f=r_{23}/R_{\rm H}\)). As long as the residence time \(t_{\rm res}\) is larger than the collisional time-scale \(t_{\rm coll}\), we expect the grains to continue to grind down to smaller sizes, which increases \(\beta\) and shortens the residence time-scale. We empirically measure the residence time from our simulations, specifically defining \(t_{\rm res}\) at a given \(\beta\) to be the amount of time it takes for 50 per cent of the dust grains to exit the Hill sphere. A particle is defined as having exited the Hill sphere if the planetocentric distance \(r_{23}>R_{\rm H}\). Fig. 6 shows the comparison of characteristic collisional and residence time-scales for the dust generated in irregular satellite collisions for a Jupiter-like planet orbiting a Sun-like star when \(M_{\rm tot}=1\;M_{\rm L}\) and \(\rho=3\;{\rm g/cm^{3}}\), with the other parameters being described by Eq. 8. At low \(\beta\) (large particles), the residence time within the Hill sphere is long, because radiation pressure is weak, but as \(\beta\) increases, the residence time falls sharply as the radiation pressure accelerates theodus. Although the collision time also gets shorter with decreasing size, the dependence is flatter. The net result is that the cascade to smaller size is truncated when \(\beta\) is large enough that the particles exit the Hill sphere before undergoing any more collisions. We identify the critical \(\beta=0.18\) from the intersection of the \(t_{\rm res}\) and \(t_{\rm coll}\) curves. This critical \(\beta\) represents the largest size dust grain that could escape from the Hill sphere. In principle, the size of an escaping particle could be lower or higher, depending on the parameters assumed in Equation 8 to calculate the collisional time-scale. While it's possible for the residence time (blue curve in Fig. 6) and collisional time-scale (green line) to never intersect depending on the assumed parameters, particles could still escape from the Hill sphere. A short collisional time-scale just means that particles will grind down all the way to the classical "blow-out" size corresponding to approximately \(\beta=0.5\) before getting ejected out of the entire system (Krivov, 2010). We note here that the classical blowout size of \(\beta=0.5\) only applies to a dust grain on a purely circumstellar orbit, and should only be used as a general benchmark for systems like ours where there is a planet involved. ### Source Model Our integrations yield a large number of trajectories as a function of time, for different \(\beta\). A realistic model assumes that individual dust grains are generated at a constant rate due to the collisional cascade initiated by the irregular satellite population. In practice, we achieve this continuous generation by offsetting the initial time of individual grain trajectories from our library of integrations. Working within the co-rotating frame ensures that new dust grains will always be produced in the Hill sphere of the planet. We then track the dynamical evolution of these dust grains to calculate density profiles of the dust population, as described in the next section. Figure 6: Characteristic dust grain residence time in the Hill sphere (in blue) – derived from numerical integration of orbits – and catastrophic collisional time-scale (in green) – from equation (8), both as functions of \(\beta\) when \(M_{\rm tot}=1\;M_{\rm L}\) and \(\rho=3\;{\rm g/cm^{3}}\), with the other parameters being described by Eq. 8. The intersections of the curves at approximately \(\beta=0.18\) represents the critical size at which a dust grain can decouple from the collisional cascade and be expelled from the Hill sphere. ## 4 Results The irregular satellite source model for generating dust discs is fundamentally different from the traditional planetesimal disc source population in that it has a localised source region, confined within the planetary Hill sphere, from which the material spreads out slowly. In this section we wish to therefore characterise the observational appearance of the resulting dust population, which we term an ISDD (Irregular Satellite Debris Disc). ### From circumplanetary to circumstellar orbits Fig. 7 shows a snapshot of the positions of \(N=750,000\) particles with \(\beta=0.1\), integrated for \(10^{3}\) planetary dynamical times (\(\sim\)12,000 yr). This simulation is for a Jupiter-mass planet on a 5.2 au orbit around a solar-mass star. This simulation duration was chosen because it is the time-scale over which a steady state is reached for the shape of the radial profile of the disc. The distribution of dust grains can be divided into circumstellar material and circumplanetary material. The circumstellar material is made up of dust grains that were able to escape from the Hill sphere whereas circumplanetary material represents grains that are still trapped in the Hill sphere. Whether or not a dust grain successfully escapes from the Hill sphere is primarily determined by initial conditions. The role of the zero-velocity curves from the Jacobi formalism in shaping the distribution of the escaped material is clear. Aspects of this dust population, such as azimuthal symmetry, radial profile, ring thickness, and vertical profile will be examined in the following subsections. It is important to note that the overdensity within the Hill sphere in Fig. 7 is artificial because those particles, that do not escape, will be subjected to continued collisional evolution not included in this simulation, and the particles are shown here primarily to illustrate the nature of the dust trajectories and their evolution. As we previously saw in Fig. 6, the intersection of the residence time and collisional time-scale occurs at a very short time-scale (<100 yr). This is very short compared to the duration of the simulation (\(\sim\)12,000 yr), so we expect trapped dust grains to still be coupled to the original cascade, grinding down to even smaller sizes until they too are blown out of the Hill sphere. As a result, we would not expect to see a pile-up of material in the form of a circumplanetary disc. ### ISDDs are azimuthally symmetric Although the dust is initially generated within the Hill sphere, once it escapes, azimuthal symmetry in an ISDD is achieved very quickly. Fig. 8 shows the azimuthal profile of the dust after 100 dynamical time-scales, separated into 20-degree bins for the case of \(\beta=0.15\). We have intentionally excluded material that remains in the planetary Hill sphere so that the global circumstellar disc features could be examined. We quantify the baseline fluctuations in the azimuthal profile by comparing the mean value to the standard deviation. We conclude for four representative values of \(\beta\) in the range \(0.15\leq\beta\leq 0.30\) that the ISDDs are azimuthally symmetric since the standard deviations are small compared to the mean. For example, in the case of \(\beta=0.15\), the average number of dust grains per bin is 185 while the standard deviation is 15, so we conclude that the variations are small. Similar results are obtained for other values of \(\beta\). Dust rings generated in this manner nevertheless retain the appearance of azimuthal symmetry. Another way to evaluate the azimuthal symmetry of the disc is to calculate the average dust grain semi-major axis as a function of azimuthal angle \(\theta\). In Fig. 8, we calculate both the mean and the median radius for the \(\beta=0.15\) disc. We find that the mean radius (in blue) is approximately 50 au and that the median radius (in green) is approximately 25 au. It is not surprising to see that the mean radius is higher than the median radius. While the vast majority of dust grains spend their time close to the planet's orbit bouncing around the edges of the Jacobi contours, a small fraction of dust grains will slowly diffuse out of the system due to the influence of radiation pressure and stirring from the planet, biasing the mean to higher radii. ### ISDDs exhibit thin ring morphology Astronomers commonly quantify ring thickness as the ratio of ring width to ring radius \(\Delta R/R\). Specifically, a ring may be characterized as 'thin' if \(\Delta R/R<0.5\)(Hughes et al., 2018). This definition takes into consideration the great diversity of size scales that debris discs are observed to span and allows us to compare systems on large and small absolute scales. However, the ratio requires us to specify how we define \(\Delta R\) and \(R\). We first fit a function to the distribution and find that several of the parameters naturally characterize the ring width and ring radius. Specifically, we fit a piecewise function to the radial profile that contains three physically motivated regimes. Region I (\(r_{13}<r_{\rm A}\)) is simply a one-sided Gaussian that describes the sharp inner edge of the ring. This feature is to be expected since the forbidden zones from the Jacobi contours prevent dust grains from wandering any closer to the star than one planetary semi-major axis plus one Hill radius (\(a+R_{\rm H}\)). Region II (\(r_{\rm A}<r_{\rm 13}<r_{\rm B}\)) is an exponential decay function that describes the initial drop off in surface density that occurs as we move outward away from the peak. The peak tends to be just outside of the Jacobi contours because the dust grains spend a lot of time bouncing around the edges of the zero-velocity curves before they diffuse out of the system. Lastly, Region III (\(r_{\rm 13}>r_{\rm B}\)) is a continuation of Region II with an added exponential term to soften the drop-off and match the more gradual decay of the outer edge. This feature is also expected since we are investigating moderate radiation pressure strengths (\(\beta=0.15-0.30\)) that are strong enough to perturb the dust grains from a circumplanetary orbit to a circumstellar orbit, but not strong enough to immediately eject the grains from the system. The gradual tail of the radial distribution represents dust grains that are in the process of slowly diffusing out of the system. A sample fitted radial profile for \(\beta=0.15\) is shown in Fig. 9. The resulting functional form is \[N(r_{13})=\left\{\begin{array}{ll}\frac{N_{0}}{r_{13}}\exp \left(-\frac{(r_{13}-r_{\rm A})^{2}}{2\sigma_{1}^{2}}\right),&r_{13}\leq r_{ \rm A}\\ \frac{N_{0}}{r_{13}}\exp\left(-\frac{(r_{13}-r_{\rm A})}{\sigma_{2}} \right),&r_{\rm A}<r_{13}<r_{\rm B}\\ \frac{N_{0}}{r_{13}}\exp\left(-\frac{(r_{13}-r_{\rm A})}{\sigma_{2}} \right)\\ +\frac{N_{13}}{r_{13}}\left[1-\exp\left(-\frac{(r_{13}-r_{\rm B})}{\sigma_{3}} \right)\right],&r_{13}>r_{\rm B}\end{array}\right. \tag{9}\] where \(N_{0}\) and \(N_{1}\) are normalization constants for their respective terms, \(r_{\rm A}\) is the peak of the distribution, \(r_{\rm B}\) is the transition point between Regime II and Regime III, \(\sigma_{1}\) is the standard deviation of the single-sided Gaussian, and \(\sigma_{2}\) and \(\sigma_{3}\) are the characteristic lengths of their respective exponential decay terms. In order to cast this in the observational variables \(\Delta R\) and \(R\), we take \(R\equiv r_{\rm A}\) since it naturally describes the peak of the distribution, and \(\Delta R\equiv\sigma_{1}+\sigma_{2}\) since those lengths each characterize the drop-off in either direction away from the peak. Thus, in terms of the function parameters, \(\Delta R/R\equiv(\sigma_{1}+\sigma_{2})/r_{A}\). We also apply the fit to multiple homogeneous discs of different values of \(\beta\) (\(\beta=0.15-0.30\)). The same piecewise function was fitted to all simulations and for each value of \(\beta\), the piecewise function did a good job of smoothly connecting the surface density profile and defining a reasonable ring width. We measure their normalized ring widths and determined the uncertainty by marginalizing out the seven parameters in the function described above in favor of \(\Delta R/R\). A 3\(\sigma\) confidence interval was used to determine the upper and lower limits of the ensuing error bars. Those results are summarized in Fig. 10. The general trend is that the thickness of the ring grows with increasing \(\beta\). This occurs because higher values of \(\beta\) correspond to stronger radiation pressure. This stronger pressure is able to more efficiently push dust grains into larger and more eccentric orbits. ### An exponential tail As mentioned in the previous section, the radial distribution has a gently sloped exponential tail. This tail is a generic feature of all models that take radiation pressure into account. However, our model generate larger particles than a standard source model since the collisional cascade is truncated by the Hill sphere residence time, as shown in Fig. 6. The slope of the fiducial \(\beta=0.15\) ISDD is characterized in Fig. 11. While the data plotted is from 0 to 100 au, we calculated the slope only for the portion between 20 and 100 au. The characteristic length came out to be 20.0 au, several times the planet semi-major axis, indicating a relatively slow drop-off. ### ISDDs exhibit a toroidal shape We examined the vertical structure of the simulated disc in addition to the radial structure. Specifically, we plotted the dust grain abundance as a function of height \(z\) above or below the midplane of the disc, as seen in Fig. 12. In addition, we fit a standard Gaussian function to the vertical distribution since the standard deviation would naturally translate to a scale height. We find that the scale height of \(H=0.78\) au for the \(\beta=0.1\) toy model is comparable to both the ring thickness (\(\Delta R=0.64\) au) and Hill radius (\(R_{\rm H}=0.35\) au). We attribute the sharp inner edge of the torus to the forbidden zone predicted by the Jacobi constant. Recall from Fig. 2 that particles of a certain Jacobi constant are not allowed to exist in certain spaces in the restricted three-body problem. For physically relevant values of \(\beta\), this region takes the shape of an annulus along the orbit of the planet, approximately one Hill diameter in width. The inner edge of our ISDD represents dust grains bouncing around the edges of the forbidden zone. ## 5 Comparison to observations One motivation for this study is the presence of narrow ring-like structures in some observed debris disc systems. The narrowness of the images implies a mechanism for confining either the dust or its parent population. In our model, this is a consequence of the orbital evolution regulated by the Jacobi constant, which sets a hard inner edge on the distribution. The outer profile is more gradual as the dust spirals out, and we compare here this theoretical expectation to properties of the best quantified observed systems. Figure 7: Position of test particles moving under the influence of radiation pressure for a value of \(\beta=0.15\) after 1,000 dynamical time-scales (\(\sim\)12,000 years) of integration. This simulation is for a Jupiter-mass planet on a 5.2 au orbit around a solar-mass star. This simulation duration was chosen because it is the time-scale over which a steady state is reached for the shape of the radial profile of the disc. Once a steady state is reached in the disc, the shape of the radial profile remains the same, but the amplitude decreases. The zero-velocity curves for \(C_{J}=2.824\) are plotted in blue. The location of the star and planet are denoted by a yellow star and orange dot respectively. The right panel shows an face-on view of the disc as a whole. The left panel zooms in on the vicinity of the planet, showing an overdensity of material in the Hill sphere. These are particles whose trajectories remain bound over the course of the simulation. In reality, these particles have orbited the planet for hundreds of collisional times and will have been ground down to much smaller sizes. Therefore, it is important to note that the overdensity within the Hill sphere is not physical, and is shown here primarily to illustrate the nature of the dust trajectories and their evolution. When viewed as a synthetic image, the panel on the right should have the circumplanetary excess reduced by a factor of several hundred, at least. ### Fomalhaut The existence of a circumstellar disc around the 440 Myr old A3V star Fomalhaut has been known for a long time due to the infrared excess in its spectrum. Fomalhaut is one of the best-studied debris discs, due to its distance from Earth being only 7.7 pc and the fact that it is one of the closest systems that is not edge on. The Fomalhaut debris disc was first directly imaged by Kalas et al. (2005) using the Hubble Space Telescope, revealing a sharp inner edge and the central star being offset from the disc geometric centre. They derived a flux profile for Fomalhaut by fitting a scattered light model of an inclined, optically thin, belt of dust to observational data, as shown in Fig. 13. The best-fitting value for the power law profile describing the inner edge of the belt is proportional to \(r^{1.09}\) whereas that of the outer belt scales with \(r^{-4.6}\). Similarly, when an exponential profile is used, the inner edge is proportional to \(\exp(0.08r)\) while the outer edge scales with \(\exp(-0.03r)\), where \(r\) is in au. They define the inner edge as 120-140 au and the outer edge as 140-158 au. Since Kalas et al. (2005) did not explicitly measure a ring width for Fomalhaut, we extrapolate one from their distribution by defining Fomalhaut's normalized ring width to be equal to its full width at half maximum (FWHM) divided by the peak. Under this definition, Fomalhaut's flux profile has a ring width of \(\Delta R/R=0.191\), which fits the definition from Hughes et al. (2018) of a 'narrow' ring as one having a normalized ring width \(\Delta R/R<0.5\). In order to place the Kalas observations within our paradigm, we fit our function to observations of the radial distribution of the data for Fomalhaut obtained by (Kalas et al., 2005). In Fig. 13, the red Figure 8: _Top:_ Histogram of dust grain distribution as a function of azimuthal angle \(\theta\) for \(\beta=0.15\). The spike located at \(\theta=0^{\circ}\) has been subtracted off, since leftover material in the planetary Hill sphere is not part of the circumstellar disc. The average number of dust grains per bin is 185 with a standard deviation of only 15 grains, denoted by the horizontal blue line and blue shaded region. _Bottom:_ Mean and median radius as a function of azimuthal angle \(\theta\). The mean radius is approximately 50 au at nearly all angles while the median radius is approximately 25 au at all angles, showing that the disc is very azimuthally symmetric. Figure 10: Normalized ring widths (\(\Delta R/R\)) for various values of \(\beta\) for a system with a 1 M\({}_{\odot}\) star and a \(10^{-3}\) M\({}_{\odot}\) planet. The general trend points to broader ring widths and therefore shallower drop-offs for higher values of \(\beta\). This is not surprising since higher values of \(\beta\) correspond to stronger radiation pressure. Thus, broader ring widths show dust grains that are actively spiraling out of the system. We also note that the error bars tend to be larger for larger values of \(\beta\). This is also not surprising since these discs tend to have smaller sample sizes by the end of a controlled simulation, since the stronger radiation pressure ejects a higher percentage of grains from the system. Figure 9: Surface density profile for a system with a 1 M\({}_{\odot}\) star and a \(10^{-3}\) M\({}_{\odot}\) planet as a function of radius for \(\beta=0.15\). Function fit (in red) to the distribution. The radii that define the boundaries between the three regimes are \(r_{A}=5.67\) au and \(r_{B}=5.98\) au. There is a sharp inner edge, indicating the existence of a gap in the disc created by the forbidden zone predicted from the restricted three-body problem. There is a more shallow decline at large distances caused by dust grains taking their time spiraling out of the system due to stellar radiation pressure. The characteristic lengths of the Gaussian and exponential fits allow us to quantify the width of the ring. These width measurements can be compared to observations since only the brightest, densest regions would be observable. For this particular value of \(\beta=0.15\), we find the normalized ring width to be \(\frac{\Delta R}{R}=0.148\,^{+0.023}_{-0.025}\), thus meeting the criterion for a thin ring. data points are the raw data obtained by Hubble Space Telescope observations, the blue curve is the fit performed by Kalas et al. (2005), and the black curve is our function's fit to the same data. If the Fomalhaut debris disc were formed by a hypothetical planet's irregular satellite cloud, we can estimate that planet's semi-major axis by scaling up our simulated system. Specifically, we scale up the simulated system from Fig. 9 so that the peak of the simulated disc's radial profile (5.675 au) matches the peak of the Fomalhaut debris disc's radial profile (144 au). Our model predicts that the planet feeding this disc would have a semi-major axis of 133 au. We fit the piecewise function to the radial distribution to characterize both the inner edge and outer edge. However, our inner edge is best described by a single-sided Gaussian as opposed to either a power law or an exponential tail. We find that the inner edge has a characteristic length of 2.23 au and that the outer edge has an exponential decay scale of 20.6 au. The inner edge behavior is very similar to what Kalas et al. (2005) found in their fit for Fomalhaut, but our function has a more gently-solved outer edge than their fit. If we measure the ring width of the Fomalhaut disc, we get \(\Delta R/R=0.215\), not significantly different from that of Kalas et al. (2005). Thus, the observed profile of the Fomalhaut debris disc is well fit by that expected for an IBDD. As a reminder, the thickness of the forbidden zone from the restricted three-body problem governs not only the thickness of any ring gaps that form, but also defines the size of the offset between the peak of the radial distribution and the location of the planet (e.g., Chiang et al., 2009; Wisdom, 1980). Such a prediction can be made because the planet is located approximately halfway between the inner and outer edge of the forbidden zone, as shown in Fig. 5, for example. In the case of a finite \(\beta\), we expect the peak of the radial distribution to occur just outside of the outer edge of the forbidden zone. Such a phenomenon should occur because the outer edge of the forbidden zone is also a zero-velocity curve upon which a dust grain decelerates and comes to rest in the co-rotating frame, thus statistically spending more time near the forbidden zone. Therefore, we expect the planet to be a distance of one-half of the forbidden zone thickness interior to the location of the peak. In order to physically interpret our model fit to the Fomalhaut observations, we first estimate which value of \(\beta\), for a single uniform-\(\beta\) debris disc, best corresponds to the parameters derived from our fit. We start by plotting two key parameters, \(\sigma_{1}\) and \(\sigma_{2}\), as functions Figure 11: Semilog radial profile of fiducial ISDD for \(\beta=0.15\) showing the entire disc from 0 to 100 au. An exponential decay function (blue line) was fit to the portion of the data from 20 to 100 au to determine the characteristic length of the decay. We find that the profile has a characteristic decay scale of 20.0 au. Figure 12: Distribution of dust grains as a function of height \(z\). A Gaussian function (in red) has been fit to the distribution. The gray horizontal lines represent one scale height (\(H=0.78\) au) above and below the midplane. Figure 13: Fomalhaut flux profile as a function of radius (Kalas et al., 2005). The red points are raw observational data. The fit of Kalas et al. (2005) is in blue, and our functional fit is in black. They showed that the inner edge of the belt can be modeled as either a power law fit with an index of \(\alpha=10.9\) or an exponential growth proportional to \(\exp(0.08r)\), where \(r\) is in units of au. Additionally, they showed that the outer edge of the belt can be modeled as either a power law fit with an index of \(\alpha=-4.6\) or an exponential decay proportional to \(\exp(-0.03r)\), where \(r\) is in units of au. Their model predicts that the planet sculpting this disc will have a semi-major axis of 133 au. By scaling up our simulated disc from its peak of 5.675 au to Fomalhaut’s peak of 144 au, we also predict the location of the underlying planet to be 133 au. of \(\beta\), as shown in Fig. 14. These two parameters were chosen because they directly determine the measured width of the ring. As one can see, both the inner edge characteristic length and outer edge characteristic length become larger with increasing \(\beta\). The relatively flat distributions show that our model is quite robust and can replicate Fomalhaut observations for a wide variety of radiation pressure strengths, specifically \(\beta\leq 0.3\). This is possible because while there is a very weak dependence of \(\sigma_{2}\) on \(\beta\), \(\sigma_{1}\) gets much larger at larger \(\beta\). #### 5.1.1 Fomalhaut b In addition to the debris disc ring, Kalas et al. (2008) also detected a point source that was proposed as Fomalhaut b, a Saturn-mass planet responsible for sculpting the inner edge of the debris disc. In this scenario (Chiang et al., 2009), the planet would have a semi-major axis \(\sim 115\) au and the inner edge of the dust ring would trace the edge of the chaotic region surrounding the planetary orbit. This claim was controversial because the colours of Fomalhaut b showed little evidence for thermal emission from a giant planet, and were far more consistent with pure scattered light from the star. The reality of the source detection itself has been independently confirmed (Currie et al., 2012; Galicher et al., 2013) but further observations by Kalas et al. (2013) reveal several orbital features that make the sculpting planet hypothesis unlikely. The orbit of Fomalhaut b appears to be highly eccentric (\(e\sim 0.8\)), especially compared to the eccentricity of the disc (\(e=0.12\pm 0.01\)) (MacGregor et al., 2017), so that it would pass through the debris disc if it were not inclined at \(\sim 17^{\circ}\) to the disc. A planet on such an orbit would be unlikely to gravitationally sculpt the observed structure, as the high eccentricity and nonzero inclination does not correspond to the correct orbital geometry to maintain the original model. Nor is an object on this orbit likely to be the source of an irregular satellite debris disc, at least according to our model, due to the disparities in both eccentricity and semi-major axis. Kalas et al. (2013) estimates the semi-major axis of Fomalhaut b to be 177 au, much larger than that of the planet we propose, which would be located at 133 au (though their margin of error is quite large at \(\pm 68\) au). In order to explain the colours of the original Fomalhaut b hypothetical planet, Kennedy & Wyatt (2011) developed a model starting from a similar hypothesis as ours. They constructed a collisional cascade of irregular satellites within a fraction of the Hill sphere of a giant planet, taking into account the strength versus self-gravity of the satellite. They took into account both radiation pressure and Poynting-Robertson (PR) drag for the resulting dust grains. Kennedy & Wyatt (2011) focussed on the appearance of dust confined within the Hill sphere, as a source population for the scattered light observed from Fomalhaut b. In our model, we focus on the dust that has escaped into heliocentric orbit, as the origin of the debris disc itself - not the point source. ### HR 4796A HR 4796A is an 8 Myr old A0V star that hosts a well-studied debris disc at a distance of 72.8 pc from Earth. The disc has an exceptionally high infrared excess of \(f=L_{\rm IR}/L_{*}=4.2\times 10^{-3}\)(Jura, 1991). HR 4796A has been imaged in multiple wavelengths including the sub-mm, the mm, mid-infrared, near-infrared, and visible. Combining these different wavelength regimes permits extensive modelling of the spectral energy distribution (SED) of the system. A complete understanding of the SED leads to understanding of the underlying dust composition of the disc. Previous studies have resolved a circular disc structure with a radius of \(\sim 77\) au, with a sharply peaked radial profile, and a \(\sim 1\) au offset from the location of the star. We can learn more about the dynamics of the system from detailed modeling of the exact geometry. #### 5.2.1 HR 4796A ring width In 2009, the Hubble Space Telescope resolved the debris disc around HR 4796A and found that it has a ring width of 18.5 au and a radius of 76 au (Schneider et al., 2009). Thus, its normalized ring thickness is \(\Delta R/R=0.25\), comparable to that of Fomalhaut and our simulated disc. All three are well within the definition of Hughes et al. (2018) for a narrow ring. We compare our model to observations of HR 4796A made by Schneider et al. (2009) using the Hubble Space Telescope Imaging Spectrograph. Specifically, we fit our three-regime piecewise function to the intensity profile for a direct one-to-one comparison and extract a normalized ring width, as shown in Fig. 15. If the HR 4796A debris disc were formed by a hypothetical planet's irregular satellite cloud, we can estimate that planet's semi-major axis by scaling up our simulated system. Specifically, we scale up the simulated system from Fig. 9 so that the peak of the simulated disc's radial profile (5.62 au) matches the peak of the HR 4796A debris disc's radial profile (76.5 au). Our model predicts that the planet feeding this disc will have a semi-major axis of 70.8 au. We find that our function does an overall good job of fitting the inner edge of the disc, but falls to zero more quickly than the HST data. Mathematically, our function drops to zero quickly since the inner edge is defined by a single-sided Gaussian. The discrepancy may be due to background noise in the HST data. As for the outer edge, our function initially drops off a little bit more quickly than the HST data. As a result, the normalized ring width is slightly lower than that derived from the HST observations, but there is not a significant difference. We once again used the full Figure 14: Inner edge characteristic length and outer edge characteristic length for various values of \(\beta\) for a system with a \(1\) M\({}_{\odot}\) star and a \(10^{-3}\) M\({}_{\odot}\) planet. These simulations have the same initial conditions as the simulations shown in Fig. 10. The inner edge, \(\sigma_{1}\), generally has greater lengths with increasing \(\beta\). Interestingly, the outer edge, \(\sigma_{2}\), generally has relatively constant length as a function of \(\beta\). However, the sum of \(\sigma_{1}\) and \(\sigma_{2}\) shows that ring width increases as a function of \(\beta\). The data begin to become unreliable and noisy at \(\beta=0.3\) due to a small surviving sample size. width at half maximum (FWHM) of the radial profile method for determining ring width and find that the ring with for HR 4796A is \(\Delta R/R=18.3\) per cent. We note that HR 4796A, Fomalhaut, and our simulated disc all fall within the definition of a 'thin ring' as defined by Hughes et al. (2018). #### 5.2.2 HR 4796A blowout size We compare the particle sizes predicted by our model to those predicted by other models for the HR 4796A system. We can calculate the dust grain size corresponding to \(\beta=0.1\) for the stellar parameters of HR 4796A, namely the luminosity of 23 L\({}_{\odot}\) and mass of 2.18 M\({}_{\odot}\), using Equation 3. This specific value of \(\beta\) was chosen because it fulfills the criterion laid out in Appendix B to ensure overflow through the L\({}_{2}\) Lagrange point. Rearranging Equation 3 to solve for the blowout size \(D_{\rm bl}\), we obtain \[D_{\rm bl}\approx(21.4~{}\mu\rm{m})\left(\frac{\beta}{0.1}\right)^{-1}\left( \frac{L_{*}}{23~{}L_{\odot}}\right)\left(\frac{M_{*}}{2.18~{}M_{\odot}}\right) ^{-1}. \tag{10}\] The result gives us a dust grain diameter of \(D\approx 21.4~{}\mu\rm{m}\). Chen et al. (2020) derived a similar grain size of 25 \(\mu\rm{m}\) by using MCFOST on SPHERE SPF data. Milli et al. (2017) found the grain size range 17.8-30 \(\mu\rm{m}\) fit the data depending on the exact scattering model used. A general rule of thumb states that dust grains are best observed in electromagnetic radiation at wavelengths that are approximately equal to their size (e.g., Hughes et al., 2018). This phenomenon can be explained as a balance between two opposing processes. We first consider the fact that the smallest grains dominate the grain size distribution and thus contribute the most to the total cross section. However, there is a competing effect where grains can only efficiently emit at wavelengths that are smaller than their actual size. A sharp cutoff in this emission efficiency occurs at larger wavelengths. All in all, the total light emitted will receive contributions from the smallest grains that are able to efficiently emit from that wavelength. The observing wavelengths versus \(\beta\) are highlighted in Table 1. Although debris discs were traditionally detected using IR excesses, we are interested in comparing our simulated discs to scattered light images since IR excesses do not give us a geometric picture. Imaging capabilities can vary strongly amongst the different bands in Table 1. ### Dependence on planet mass Generally speaking, we expect the morphologies of ISDDs to depend on the ratio of planet mass to stellar mass \(M_{\rm p}/M_{*}\). This dependence arises because the Jacobi contours of the restricted three-body problem depend only on the ratio of the masses of the secondary body to the primary body \(M_{2}/M_{1}\), both with and without radiation pressure. For example, a Saturn-like planet orbiting around an M dwarf could have the same normalized Jacobi contours as a Jupiter-like planet orbiting around a G-type star. Since the thickness of the Jacobi forbidden zone is roughly the same size as the diameter of the Hill sphere, we expect any ensuing ring gaps to be a similar size to the Hill diameter as well. This effect would likely only be noticeable in low radiation pressure scenarios, since those are the only cases where a significant number of dust grains escape interior to the planet's orbit and would therefore produce an observable ring gap in the radial distribution. Appendix C goes into greater detail the theoretical foundation of exactly how the mass ratio correlates with the critical \(\beta\). ### General Predictions for Other Systems #### 5.4.1 Wavelength Dependence Different observing wavelengths will be able to probe different structures of potentially the same debris disc. However, in any given system, grain size is just a guideline for observing wavelength. Crudely, wavelength corresponds to grain size, so we do have broad predictions about how things should appear. For example, since observing wavelength is expected to be directly proportional to grain size and therefore inversely proportional to \(\beta\), we predict that the long-wavelength infrared (\(\lambda=8-15~{}\mu\rm{m}\)) will find singular thin rings. Due to the stronger influence of radiation pressure, mid-wavelength infrared (\(\lambda=3-8~{}\mu\rm{m}\)) will detect comparatively broader rings than found in the long-wavelength infrared, but still one singular ring. However, in the far infrared (\(\lambda=15~{}\mu\rm{m}-1~{}mm\)), which can be affected by values of \(\beta\) as low as 0.002, we expect there to be a gap in the ring since Roche lobe overflow through the L\({}_{1}\) Lagrange point is almost equally favorable to occur energetically. In this instance, dust escapes both inward and outward through L\({}_{1}\) and L\({}_{2}\), but the Jacobi forbidden zone prevents the two populations from mixing, giving rise to the gap in the ring. We calculate in Appendix A that an initial irregular satellite population mass on the order of 1 M\({}_{\oplus}\) is necessary to ensure that the Hill sphere collisional time-scale (\(\sim\)135 Myr) is less than the age of the \begin{table} \begin{tabular}{c c c} \hline \hline Division name & Wavelength & \(\beta\) \\ \hline \hline Mid-wavelength infrared & 3 - 8 \(\mu\rm{m}\) & 0.21 - 0.55 \\ \hline Long-wavelength infrared & 8 - 15 \(\mu\rm{m}\) & 0.11 - 0.21 \\ \hline Far infrared & 15 \(\mu\rm{m}\) - 1 mm & 0.002 - 0.11 \\ \hline \hline \end{tabular} \end{table} Table 1: Summary of correspondence between observing wavelength and relative strength of radiation pressure \(\beta\). Figure 15: The best fit of our three-regime piecewise function fitted to an HR 4796A flux profile from Schneider et al. (2009). A dashed red line is plotted to indicate 50 per cent of the peak value to help visualize the FWHM. We find that the normalized ring width of the system is \(\Delta R/R=18.3\) per cent. Our model predicts that the planet sculpting this disc will have a semi-major axis of 70.8 au. age of the system, 440 Myr in the case of Fomalhaut. While 1 M\({}_{\oplus}\) of irregular satellites does seem quite large compared to the estimated \(10^{-3}\) M\({}_{\rm L}\) worth of irregular satellites estimated to have been found around each giant planet in our Solar System (Bottke et al., 2010), a Jupiter-mass planet found at a semi-major axis consistent with the scale of the Fomalhaut system (\(\sim\)140 au) orbiting around a 2 M\({}_{\odot}\) star would have a Hill sphere \(\sim 4\times 10^{4}\) times more voluminous than that of Jupiter and a commensurate cross section. #### 5.4.2 Evolutionary Implications At first glance, irregular satellite collisions would appear to not explain the debris discs found around such aged systems as Fomalhaut (440 Myr), due to how quickly they grind down to dust and dissipate (\(\sim\)tens to hundreds of thousands of years). Thus, irregular satellite debris discs would only be bright enough to be detectable in their infancy. However, irregular satellites are not formed in the same way as regular satellites, and thus do not have the same age as their host planet. In our own Solar System, they are thought to be the result of dynamical capture during late-stage rearrangement of the giant planet orbits (Nesvorny et al., 2007), hundreds of Myr after the formation of the Solar System. This delay may help us explain the age of some older debris disc systems. Our model is not intended to explain every debris disc, but is focussed on the curious geometry of the thin ring systems. Within the proposed context, the observed thin rings indicate systems that have recently emerged from a period of dynamical excitation which resulted in the capture of irregular satellites around giant planets. The size of the dust in such discs may also be a function of time, because the \(\beta\) of the escaping material is set by the balance between residence and collision times. The latter will increase as the mass reservoir in the source population grinds down, moving the characteristic \(\beta\) of the escaping particles to lower values, and therefore increasing the size of the particles in the disc. ## 6 Conclusions In this paper, we explored the effects of including radiation pressure into the classical restricted three-body problem. We found that the traditional Roche lobe overflow can be replaced by Lagrange point L\({}_{2}\) overflow for a sufficiently high \(\beta\) for a given planet-to-star mass ratio \(\mu_{2}/\mu_{1}\). Sample orbital integrations reveal that individual dust grains typically trace out 'flower petal' orbits, coming to rest on the zero-velocity curves for some time. We assumed that the origin of dust grains in our model were from collisions between the giant planet's irregular satellites. We motivated our initial conditions based off of observations of the Solar System giant planets' irregular satellites today as well as what previous studies' determined from their dynamical history. We describe the size distribution of bodies ensuing from irregular satellite collisions as a collisional cascade power law distribution. We calculate the catastrophic collisional time-scale and compare it to an empirically determined residence time-scale to determine the critical \(\beta\) at which ground down dust grains can escape the Hill sphere. Our N-body simulations show that dust grains with a \(\beta\) above \(\beta_{\rm crit}\) quickly escape from the Hill sphere and transition from a circumplanetary orbit to a circumstellar orbit. After a short time, a large population of dust grains achieve an azimuthally symmetric disc appearance. We evaluated this azimuthal symmetry by comparing the fluctuations in the azimuthal profile to the average column density and found that they were low. We also calculated the average radius along a given azimuthal angle \(\theta\) and found that the mean and median radius is consistent along all azimuthal angles. We fit a piecewise function with an Gaussian inner edge and exponential outer edge to the radial profile. These functions naturally allowed us to quantify the ring width for various values of \(\beta\). We normalized the ring width over the ring radius as is standard in the literature (\(\Delta R/R\)), and find that normalized ring width broadens as a function of \(\beta\). We explain this finding as stronger radiation pressure being able to excite dust grains to more eccentric orbits and therefore broadening the overall distribution. Since the vertical profile of the disc resembles a typical Gaussian, we conclude that the overall shape of the disc is a torus. We compared our results to observations for the specific systems of Fomalhaut and HR 4796A, but also make general predictions for all systems. For the assumption of uniform density spherical dust grains, there is an inverse relationship between observing wavelength and \(\beta\). We find that the topology of the debris disc is dictated by the original Jacobi forbidden zone contours, so the fundamental parameter is the planet-to-star mass ratio \(M_{2}/M_{1}\). We test the validity of our radial profile fitting function by applying it to the raw Hubble Space Telescope data of Fomalhaut from Kalas et al. (2005). We obtain very similar results to their fit in terms of inner and outer edge slopes. By defining a ring width for Fomalhaut as its full width at half maximum, we measure its normalized ring width to be 0.191, comparable to our model's \(\Delta R/R=0.13\), both of which are within the 'thin ring' definition defined in Hughes et al. (2018) of \(\Delta R/R=0.5\). We note that there is an ongoing debate about whether Fomalhaut b is a planet or a transient dust cloud and clarify that due to its inclined orbital plane with respect to the disc plane, we do not assume Fomalhaut b is the source of the debris disc in our model, but rather some other underlying hidden planet. For the assumption of a Sun-like star, we make general predictions about distinctions between observing wavelengths in the mid-wavelength infrared, long-wavelength infrared, and far infrared. We address the fact that while Solar System irregular satellite swarms tend to grind down very quickly on time-scales of tens to hundreds of thousands of years, they can still explain very old, large systems such as Fomalhaut once properly scaled up using Kepler's Third Law since irregular satellites were not expected to be captured until the Late Heavy Bombardment period in our Solar System. ## Acknowledgements This research has made use of NASA's Astrophysics Data System. This research was supported by NASA Grant 443820-HN-21577. ## Data Availability Statement The data underlying this article will be shared on reasonable request to the corresponding author.

2305.04106

On the Usage of Continual Learning for Out-of-Distribution Generalization in Pre-trained Language Models of Code

Pre-trained language models (PLMs) have become a prevalent technique in deep learning for code, utilizing a two-stage pre-training and fine-tuning procedure to acquire general knowledge about code and specialize in a variety of downstream tasks. However, the dynamic nature of software codebases poses a challenge to the effectiveness and robustness of PLMs. In particular, world-realistic scenarios potentially lead to significant differences between the distribution of the pre-training and test data, i.e., distribution shift, resulting in a degradation of the PLM's performance on downstream tasks. In this paper, we stress the need for adapting PLMs of code to software data whose distribution changes over time, a crucial problem that has been overlooked in previous works. The motivation of this work is to consider the PLM in a non-stationary environment, where fine-tuning data evolves over time according to a software evolution scenario. Specifically, we design a scenario where the model needs to learn from a stream of programs containing new, unseen APIs over time. We study two widely used PLM architectures, i.e., a GPT2 decoder and a RoBERTa encoder, on two downstream tasks, API call and API usage prediction. We demonstrate that the most commonly used fine-tuning technique from prior work is not robust enough to handle the dynamic nature of APIs, leading to the loss of previously acquired knowledge i.e., catastrophic forgetting. To address these issues, we implement five continual learning approaches, including replay-based and regularization-based methods. Our findings demonstrate that utilizing these straightforward methods effectively mitigates catastrophic forgetting in PLMs across both downstream tasks while achieving comparable or superior performance.

Martin Weyssow, Xin Zhou, Kisub Kim, David Lo, Houari Sahraoui

2023-05-06T18:00:21Z

http://arxiv.org/abs/2305.04106v2

On the Usage of Continual Learning for Out-of-Distribution Generalization in Pre-trained Language Models of Code ###### Abstract. Pre-trained language models (PLMs) have become a prevalent technique in deep learning for code, utilizing a two-stage pre-training and fine-tuning procedure to acquire general knowledge about code and specialize in a variety of downstream tasks. However, the dynamic nature of software codebases poses a challenge to the effectiveness and robustness of PLMs. In particular, world-realistic scenarios potentially lead to significant differences between the distribution of the pre-training and test data, _i.e._, distribution shift, resulting in a degradation of the PLM's performance on downstream tasks. In this paper, we stress the need for adapting PLMs of code to software data whose distribution changes over time, a crucial problem that has been overlooked in previous works. The motivation of this work is to consider the PLM in a non-stationary environment, where fine-tuning data evolves over time according to a software evolution scenario. Specifically, we design a scenario where the model needs to learn from a stream of programs containing new, unseen APIs over time. We study two widely used PLM architectures, _i.e._, a GPT2 decoder and a RoBERTa encoder, on two downstream tasks, API call and API usage prediction. We demonstrate that the most commonly used fine-tuning technique from prior work is not robust enough to handle the dynamic nature of APIs, leading to the loss of previously acquired knowledge _i.e._, catastrophic forgetting. To address these issues, we implement five continual learning approaches, including replay-based and regularization-based methods. Our findings demonstrate that utilizing these straightforward methods effectively mitigates catastrophic forgetting in PLMs across both downstream tasks while achieving comparable or superior performance. 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 2023 ACM SIGKDDDDDDiversity of Montreal 3 ## 1. Introduction Prior research (Krishnan et al., 2017; Krizhevsky et al., 2014; Krizhevsky et al., 2014) on code representation learning leverages a ubiquitous two-stage procedure to effectively train and specialize pre-trained language models (PLMs) for code-related downstream tasks. The first stage, _i.e._, the pre-training, involves optimizing the model using self-supervised learning on a large dataset to acquire general knowledge about code. This pre-training phase allows the model to adapt to downstream tasks in the second stage, _i.e._, the fine-tuning. Previous studies (Krishnan et al., 2017; Krizhevsky et al., 2014; Krizhevsky et al., 2014) typically leverage classical transfer learning methods, which consist of "transferring" the pre-trained knowledge to the target task by fine-tuning the model on a task-specific loss function and data. This approach has been successful in the fields of natural language processing (NLP) (Krizhevsky et al., 2014; Krizhevsky et al., 2014) and deep learning for code (Krizhevsky et al., 2014; Krizhevsky et al., 2014). In this perspective, previous works (Krizhevsky et al., 2014; Krizhevsky et al., 2014) have primarily focused on stationary settings, neglecting the practical need for models to adapt to changing environments and data over time. Most prior research (Krishnan et al., 2017; Krizhevsky et al., 2014; Krizhevsky et al., 2014) has suggested using transfer learning to fine-tune the model in static environments rather than addressing the dynamic nature of real-world scenarios. In practice, programming languages, software libraries and APIs are prone to change Figure 1. Continual fine-tuning of a pre-trained language model of code. After pre-training, the model needs to adapt to new out-of-distribution (OOD) program data over time. and evolution [25, 43, 45], leading to shifts in the distribution of the underlying software data over time, it is also known as concept drift [37, 60]. By ignoring the actual evolution of software codebases, existing studies [10, 59] have focused on fine-tuning and testing pre-trained models of code using stationary datasets. In practice, the software evolution potentially leads to a noticeable difference between training and test data, _i.e._, distribution shift, that is often not present in these stationary datasets. This phenomenon also occurs when the model is put into production and has to deal with real-world data [4, 23]. We argue that creating datasets that reflect real-world software evolution scenarios and distribution shifts is crucial in order to properly evaluate the **out-of-distribution (OOD) generalization** capability of code models [50]. The OOD generalization measures a model's ability to generalize to new, unseen data with a significantly different distribution from the training data. Therefore, evaluating how PLMs of code generalize to OOD software data in software evolution scenarios appears as a prime issue. Existing works on OOD generalization designed the datasets based on various distribution shifts in source code data [22, 27]. However, they did not address the problem of continually adapting a pre-trained model of code to streams of OOD data. The prime goal of our study is to explore methods for a model to better adapt to software evolution scenarios. In this context, we ask: _how to effectively continually fine-tune a pre-trained model of code to adapt to new data while still considering the past data?_ (see Fig. 1). Over the past years, **continual learning (CL)**[44, 60] has emerged to address this problem, which is relevant to a wide range of research areas, including computer vision [5, 31, 35, 52] and NLP [6, 8, 53]. Although transfer learning methods are not tailored for continual learning scenarios, they can still operate to fine-tune a model on streams of data. However, these methods lack robustness, leading to unwanted phenomena such as forgetting past information, known as catastrophic forgetting [18, 40]. There exist other strategies, such as retraining the model from scratch using new data, which are also impractical due to the tremendous computational intensity in the pre-training phase. Motivated by these issues of the existing models, we attempt to investigate more robust and scalable fine-tuning techniques. We hypothesize that continual learning techniques may provide significant benefits over classical transfer learning in this context. In this paper, we delve into the behavior of PLMs of code in a continual fine-tuning scenario, as depicted in Fig 1. Our objective is twofold: (1) to assess the out-of-distribution generalization capability of PLMs of code and (2) to investigate effective continual fine-tuning strategies to fine-tune the models in the presence of a stream of OOD data. Specifically, we address these challenges in a scenario reflecting how typical software codebases may evolve in practice. To this end, we create five OOD domain datasets, each introducing new, unseen APIs by the models during their pre-training phase. These OOD datasets intend to simulate a stream of data for continual fine-tuning, and each dataset entails a significant distribution shift with respect to the pre-training data. As such, our setting establishes an OOD generalization problem. We consider two widely used model architectures: a GPT2-like [46] decoder and a RoBERTa-like [34] encoder pre-trained on code. To eliminate any data leakage between the pre-training and fine-tuning data, we decided to pre-train our models from scratch. We do not study the popular existing PLMs like CodeBERT [17] or CodeT5 [58] because they may be prone to potential data leakage, _i.e._, seeing the OOD data in pre-training, that we cannot precisely control. We evaluate the models on two downstream tasks: API call prediction and API usage prediction. In the first task, the model attempts to predict API calls resulting in a single code token, given code tokens appearing before the call site. On the other hand, the second task involves the generation of the whole API usage resulting in a sequence of code tokens with the same input format as the prior task. Together, these two tasks provide a comprehensive evaluation of the model's performance in different code generation scenarios. We start by investigating the impact of OOD data on the performance of the GPT2-like decoder on both downstream tasks in a zero-shot setting, _i.e._, without fine-tuning the model on the new OOD data. We find that the model consistently fails to generalize to OOD data by highlighting significant gaps in performance compared to in-distribution data across six evaluation metrics (_e.g._, up to \(75\%\) drop in BLEU score). This finding strongly suggests that pre-training itself is not sufficient and cannot solve OOD generalization in PLMs of code. We then evaluate the models' performance in the continual fine-tuning scenario using classical transfer learning and observe notable catastrophic forgetting. To address this issue, we implement a straightforward yet computationally inefficient cumulative fine-tuning approach by utilizing a replay buffer of infinite size. The results show that the approach drastically mitigates forgetting. Finally, we compare the performance of classical transfer learning to that of replay-based and regularization-based continual learning methods. Replay methods are considered tough-to-beat strategies for continual learning and consist of maintaining a small replay buffer containing samples from previously seen data. During fine-tuning, we use the replay buffer in conjunction with the current OOD training set to fine-tune the PLM. We explore regularization-based methods, including EWC [31], SI [66] and RWalk [9], which add regularization terms to the loss function at fine-tuning to prevent extensive changes in important parameters of the PLM. We chose those methods as they are computationally efficient, well-known, and considered strong baselines in the continual learning literature. We discover that those continual learning methods significantly reduce forgetting while achieving similar or superior effectiveness on both tasks. To the best of our knowledge, this work constitutes the first initiative to study continual fine-tuning for OOD generalization of PLMs of code. We believe that the impact of continual learning in this research area has the potential to be far-reaching, particularly due to the inherent evolution of software data over time, and we discuss this aspect in more detail in the discussion section of the paper (see Section 5). Our contributions can be summarized as follows: 1. We demonstrate that PLMs of code fail to generalize to OOD data and highlight the need for further investigation in this area. 2. We conduct a study on the behavior of two pre-trained model architectures of code in a continuous learning environment, showing that classical transfer learning lacks robustness and is prone to catastrophic forgetting. 3. We compare five continual learning methods, including replay-based and regularization-based approaches, in our continual fine-tuning scenario. We show the superiority of continual learning over classical transfer learning. 4. We provide a large-scale dataset of Java code snippets and their API usage sequences, including pre-training data and a procedure for extracting OOD data. **Organization.** In Section 2, we discuss preliminaries on continual learning. In Section 3, we go through our experimental design. We present the results of our experiments in Section 4. In Section 5, we discuss the threats to the validity of our study, as well as potential broader impact and future research directions. We introduce the related work on out-of-distribution generalization and continual learning for pre-trained language models in Section 6. Finally, we discuss some future work and conclude this work in Section 7. ## 2. Preliminaries on continual learning Existing PLMs such as BERT (He et al., 2019) or GPT (Beng et al., 2019) typically operate in transfer learning settings. By using a two-stage pre-training/fine-tuning procedure, these models can be specialized for a wide range of downstream tasks. However, in this setting, the data used for pre-training or fine-tuning are often assumed to be stationary, which is not reflective of real-world situations. In practice, transfer learning methods can still be applied to non-stationary data, such as a stream of data, but this technique is prone to catastrophic forgetting (Krizhevsky et al., 2014; Krizhevsky et al., 2015; Krizhevsky et al., 2016). To address the above issues, prior works (Krizhevsky et al., 2015; Krizhevsky et al., 2016; Krizhevsky et al., 2016; Krizhevsky et al., 2017; Krizhevsky et al., 2018) introduced the concept of _continual learning_ and designed specific techniques to mitigate catastrophic forgetting. The primary assumption for continual learning is that the neural network should possess the ability to adapt to new data or tasks while maintaining stability on previous data or tasks, often referred to as the plasticity-stability dilemma. Considering continual learning is particularly interesting for OOD generalization problems, as continual learning methods focus on a keeping good plasticity-stability trade-off. Altogether, it has to potential to enhance the generalizability of PLMs to a broader range of data. Continual learning methods often operate in constrained scenarios, and Hadsell et al. (Hadsell et al., 2019) outline a comprehensive list of objectives to balance in continual learning scenarios. There exist three main categories of methods for continual learning as defined in a previous study (He et al., 2019). _Replay-based methods_ store samples from previous experiences, _i.e._, previous stream of data, in a replay buffer or use generative approaches to generate examples similar to those of previous experiences. The replay buffer is used in conjunction with the current experience data to train the model. Replay-based methods help the network gain stability by enabling the network to train on previous samples, _i.e._, stored in the replay buffer while adapting to new data. _Regularization-based methods_ add a regularization term to the loss function to prevent catastrophic forgetting by penalizing changes to important neural network parameters. Examples of regularization-based methods include EWC (Krizhevsky et al., 2016), SI (Zhu et al., 2017) and RWalk (Beng et al., 2019). Finally, _parameter isolation methods_ use dynamic architectures to incorporate knowledge from previous experiences to mitigate interference (Zhu et al., 2017). ## 3. Experimental design In this section, we describe the experimental setup of our study. We carefully control our data and model setup to implement our out-of-distribution scenario. We first outline the construction of our dataset and the generation of OOD data for continual fine-tuning. Next, we discuss the pre-training procedure of our models, the target downstream tasks and evaluation metrics. We present the results of our experiments in Section 4. ### Dataset construction Pre-training language models from scratch require a large amount of data for the loss of the model to converge. With that in mind, we constructed our large dataset using programs crawled from GitHub using Google BigQuery1. Specifically, we focused on Java programs and began by collecting all java files stored in GitHub repositories. Next, we used Group (Zhu et al., 2017) to extract all methods defined in the java files along with their API usage sequences. We extracted the API usage sequences to facilitate our data splitting and obtain the position of each API site inside the methods to implement our downstream tasks. Each sample consists of all the tokens of a method. To avoid duplication bias in our experiments (Beng et al., 2019), we deduplicated the dataset by comparing the hash of each method. The resulting dataset contains more than 68M Java methods. For our experiments, we shuffled these 68M methods and randomly selected 10M methods to constitute our initial dataset. Fig. 2 illustrates how we further split the data for our experiments. Because we chose the pre-train PLMs from scratch, we have to split our data into in-distribution (ID) data, used for model pre-training, and OOD data, used for continual fine-tuning. We also need to properly extract the OOD data to align with our scenario consisting of introducing new, unseen APIs over time to the PLM during fine-tuning. Footnote 1: [https://cloud.google.com/bigquery](https://cloud.google.com/bigquery) _Out-of-distribution dataset - \(\mathcal{D}_{OOD}\)._ We create five OOD datasets, \(\mathcal{D}^{1}_{OOD},...,\mathcal{D}^{5}_{OOD}\). Each OOD dataset represents a unique domain that encompasses a high-level functionality of APIs. For example, we have a domain _Security_ that comprises APIs related to programming security-related code and a domain _Guava_ that includes only APIs from the Guava2 library. To create each OOD dataset, we randomly select 10 interfaces from packages/libraries related to their domain. Finally, we associate to each domain dataset all APIs Figure 2. Procedure to extract the ID data used for model pre-training, and the OOD data used for continual fine-tuning. within the selected interfaces, excluding class construction methods. Table 1 summarizes the dataset \(\mathcal{D}_{OOD}\), which contains 147,245 samples in total. To form each OOD dataset, we select samples from the pool of 10 million Java methods that manipulate at least one of their associated API. In our experiments, we perform continual fine-tuning on the training sets associated with the OOD dataset \(\mathcal{D}_{OOD}^{1},...,\mathcal{D}_{OOD}^{5}\) sequentially. Therefore, to prevent data leakage, we exclude samples that manipulate APIs from multiple domains. This elimination of samples removes a significant threat to the validity of our OOD scenario and ensures that APIs are introduced as intended during the fine-tuning process. To obtain representative test sets, we randomly select 10% of samples that manipulate each API within each OOD dataset and used the selected samples to form the corresponding domain test set. _In-distribution dataset_ - \(\mathcal{D}_{ID}\).We obtain \(\mathcal{D}_{ID}\) by removing the samples in \(\mathcal{D}_{OOD}\) from the initial data. Then, we shuffle \(\mathcal{D}_{ID}\) and randomly select 50,000 samples for test (\(\mathcal{D}_{ID\_test}\)). \(\mathcal{D}_{ID\_FT}\) contains the remaining samples for pre-training, and we randomly select 100,000 for model validation (\(\mathcal{D}_{ID\_PT\_valid}\)). In particular, those samples allow us to monitor the evolution of the loss of the model on an independent validation set to avoid overfitting the pre-training data. In total, the pre-training set \(\mathcal{D}_{ID\_PT\_train}\) contains more than 9M samples to pre-train the models. ### Models and tasks setup In this work, we consider two widely-used deep learning architectures for code: a RoBERTa-like encoder (Zhu et al., 2017) and a GPT2-like decoder (Zhu et al., 2018). _Decoder_ - \(\mathcal{M}_{dec}\).The decoder model is based on the GPT-2 architecture, with the same hyperparameters, and is pre-trained using a causal language modeling objective, _i.e._, left-to-right next token prediction. As we conducted our experiments under limited resources, we implemented a small version of GPT-2 with 110 million trainable parameters and pre-train the model for 100,000 steps. We use early stopping to select the best model checkpoint, based on the loss on the validation set \(\mathcal{D}_{ID\_PT\_valid}\). _Encoder_ - \(\mathcal{M}_{enc}\).The encoder model is based on the RoBERTa architecture, with the same hyperparameters, and is pre-trained using a masked language modeling objective. We implemented a base version of RoBERTa. The model has 125 million trainable parameters and is pre-trained similarly to the decoder model, with early stopping used to select the best checkpoint. _Downstream tasks_.We employ two downstream tasks to evaluate the ability of our PLMs of code to learn and adapt to new software data that introduce new, unseen APIs over time. Fig. 3 illustrates both tasks. For API call prediction, the model takes as \begin{table} \begin{tabular}{l c c c c c} \hline \hline Dataset & Domain & Package & Interfaces & \# train & \# test \\ \hline \multirow{4}{*}{\(\mathcal{D}_{OOD}^{1}\)} & \multirow{4}{*}{General} & java.util.concurrent & BlockingQueue, ThreadPoolExecutor & & \\ & &java.math & BigInteger & & 47,213 & 5,239 \\ & &java.util & Based, Treeset & & & \\ & &java.net & FurdjunPool, Proxy, ServerSocket, SocketAddress, URLEncoder & & & \\ \hline \(\mathcal{D}_{OOD}^{2}\) & Security & java.security & Cipher, CodeSource, Identity, KeyFlatancy, KeyFlatancy, KeyFlatancy & 27,189 & 3,017 \\ & & & Provider, Security, Timestamp & & & \\ \hline \multirow{4}{*}{\(\mathcal{D}_{OOD}^{3}\)} & \multirow{4}{*}{Android} & android & android.view & Display, InputEvent, Window & & \\ & & android.widget & Checkbox, GridLayout & & & \\ & & android.media & AudioFormat, ImageReader & 28,400 & 3,150 \\ & & android.hardware & Camera, Sensor & & \\ & & android.database & DatabaseUtils & & & \\ \hline \(\mathcal{D}_{OOD}^{4}\) & Web & org.springframework & CacheManager, ClassPathResource, DataBuffer, HipMessage, Hipfile- & 16,295 & 1,805 \\ & & & quote, JakTCTemplate, MessageChannel, MessageHandler, TaskExecutor & & \\ \hline \multirow{4}{*}{\(\mathcal{D}_{OOD}^{i}\)} & \multirow{4}{*}{Guava} & com.google.common.graph & GraphBuilder, Network & & \\ & & com.google.common.io & BytsSource, BytsStreams & & \\ \cline{1-1} & & com.google.common.cache & Cachebuilder, LoadingCache & 13,448 & 1,489 \\ \cline{1-1} & & com.google.common.collect & ListMultimap, Multitimap & & \\ \cline{1-1} & & com.google.common.base & ChauMarketer, Splitter & & \\ \hline \hline \end{tabular} \end{table} Table 1. Out-of-distribution dataset details. Figure 3. Overview of the downstream tasks. In the API call prediction task, the model outputs a list of top-\(k\) candidates to predict the API call token (_i.e._, min). In the API usage prediction task, the model attempts to predict all the tokens constituting the API usage (_interface name, method name, parameters and syntactical tokens_). The models only leverage left-context tokens to generate a prediction. input all the tokens of the method preceding the call site of the API and generates top-\(k\) candidates. For API usage prediction, the model takes as input the same tokens as for the API call prediction task, but attempts to generate the whole API usage (interface name, method name, parameters and syntactical tokens), which constitutes a more challenging task. Note that conversely to \(\mathcal{M}_{dec}\), the encoder's architecture is not suitable for generation tasks. Therefore, we add a randomly initialized language modeling head on top of it for fine-tuning using the OOD datasets. As a result, we expect \(\mathcal{M}_{enc}\) to be less stable than \(\mathcal{M}_{dec}\) and more prone to catastrophic forgetting since the language modeling head is not pre-trained. This comparison provides valuable insights into the robustness of two different architectures. Evaluation metricsWe measure the performance of the models on both downstream tasks with metrics used in prior works. For API call prediction, we report the Pass@k (Pasz and Karimiredan, 2018), which gives the percentage of correct predictions when considering lists of \(k\) candidates. For API usage prediction, we report BLEU score, Accuracy (exact match), and CodeBLEU (Zhu et al., 2019). To measure how the models perform in a continual learning environment, we use two meta-metrics adapted from prior works (Golovolovolov et al., 2013; Golovolovolov and Levesque, 2015): the _Average (A)_ and _Forgetting (F)_ metrics. We define the average \(A_{M}\) of a metric \(M\) on a test dataset \(\mathcal{D}^{i}_{OOD}\) as: \[A_{M}=\frac{1}{T}\sum_{j=i}^{T}M_{j}(\mathcal{D}^{i}_{OOD})\,,\] where \(j\) refers to the next incremental learning steps after the \(i\)-th included. \(M_{j}\) denotes an evaluation metric, _e.g._, Pass@k, computed at time step \(j\) on the test set and \(T\) denotes the maximum number of fine-tuning steps, _i.e._, five in our case. The Average metric only gives information on how accurate the model is but does not provide any insight into its ability to mitigate catastrophic forgetting. We define the forgetting \(F^{k}_{M}\) of a metric \(M\) on a test dataset \(\mathcal{D}^{i}_{OOD}\) at time step \(k\) as: \[F^{k}_{M}=M_{i}(\mathcal{D}^{i}_{OOD})\,-\,M_{k}(\mathcal{D}^{i}_{OOD})\,,\,\, i<k\;.\] This is the difference between the first time the metric is computed, _i.e._, after fine-tuning the model on \(\mathcal{D}^{i}_{OOD}\) at time step \(i\), and the metric computed at time step \(k\). \(F^{k}_{M}\) gives information on the stability of the model, _i.e._, its capability to not forget from the past. Therefore, the lower \(F^{k}_{M}\); the better. Implementation detailsTo pre-train \(\mathcal{M}_{dec}\) and \(\mathcal{M}_{enc}\), we used four Tesla V100-SXM2-32GB GPUs. It took about 7 days to pre-train \(\mathcal{M}_{dec}\), and 2 days to pre-train \(\mathcal{M}_{enc}\). For fine-tuning and inference, we used a single Tesla V100-SXM2-32GB GPU. We used Huggingface's libraries (Zhu et al., 2019) to implement the models and store the datasets. To implement the continual learning approaches, we used Avalanche (Pasz and Karimiredan, 2018). We provide all the implementation details of our experiments and release our data publicly in our replication package (see Data Availability section). ## 4. Experimental Results ### How does \(\mathcal{M}_{dec}\) generalize to ID and OOD data in zero-shot? In this experiment, we evaluate the performance of the model \(\mathcal{M}_{dec}\) on the ID and OOD test data in a zero-shot setting for both downstream tasks. We do not experiment with \(\mathcal{M}_{enc}\) as the model is not capable of generating code before fine-tuning and, therefore, cannot operate in a zero-shot setting. The purpose of this experiment is twofold. First, it aims to validate the experimental setup of our study. If we observe significant differences in the evaluation metrics obtained on the ID and OOD datasets, it would suggest that our OOD scenario is well-formed and reasonable. Secondly, significant gaps between the ID and OOD test data imply that PLMs such as \(\mathcal{M}_{dec}\) still require the use of robust transfer learning or continual learning techniques to generalize to new data without forgetting about past data. Api call predictionTable 2 reports the _Pass@1_, _Pass@5_ and _Pass@10_ on the ID and OOD test datasets. The results show that the model performs well on ID data, reaching almost 73% in _Pass@1_. However, when tested on OOD data, the performance drops significantly. The decline in performance is less severe when considering more API call candidates, but it remains a significant issue. Furthermore, variations in the performance decline are observed across different OOD datasets. For example, the model performs better on the Security domain (\(\mathcal{D}^{2}_{OOD}\)) than domains such as Android (\(\mathcal{D}^{3}_{OOD}\)) or Web (\(\mathcal{D}^{4}_{OOD}\)), which likely contain more domain-specific API calls. Api usage predictionTable 3 reports the _BLEU_ score, Accuracy (_EM_) and _CodeBLEU_ score on both ID and OOD test datasets. The results indicate that the model performs poorly on OOD data in comparison to ID data, with significant decreases in all evaluation metrics. Additionally, we notice that the _EM_ and _CodeBLEU_ metrics vary similarly to the _Pass@k_ metrics on the API call prediction task. The Android and Web domains experience the most severe drops, whereas the Security domain experiences the least severe drop. \begin{table} \begin{tabular}{c c c c} \hline \hline & \multicolumn{3}{c}{Metrics} \\ \cline{2-4} Dataset & Pass@1 & Pass@5 & Pass@10 \\ \hline \(\mathcal{D}_{ID\_test}\) & 72.88 & 83.30 & 85.60 \\ \hline \(\mathcal{D}_{OOD}\) & 40.82 (44\% ) & 51.19 (38.5\% ) & 54.17 (36.7\% ) \\ \hline \(\mathcal{D}^{1}_{OOD}\) & 49.91 (31.6\% ) & 62.0 (25.6\% ) & 64.46 (24.6\% ) \\ \(\mathcal{D}^{2}_{OOD}\) & 53.72 (26.3\% ) & 62.59 (24.8\% ) & 64.93 (24.2\% ) \\ \(\mathcal{D}^{3}_{OOD}\) & 23.78 (67.4\% ) & 32.64 (60.8\% ) & 36.33 (57.6\% ) \\ \(\mathcal{D}^{4}_{OOD}\) & 30.72 (57.9\% ) & 43.67 (47.3\% ) & 47.89 (44\% ) \\ \(\mathcal{D}^{5}_{OOD}\) & 37.54 (48.6\% ) & 49.53 (40.6\% ) & 53.22 (47.9\% ) \\ \hline \hline \end{tabular} \end{table} Table 2. API call prediction results in zero-shot using \(\mathcal{M}_{dec}\). Our results demonstrate that the model \(\mathcal{M}_{dec}\) (without fine-tuning) is unable to generalize to OOD data while showing strong performance on ID data. Our findings also support the validity of our OOD dataset as a realistic and meaningful test of the model's ability to adapt to new data in a continuous environment. ### Do models forget about past data using classical transfer learning? In this section, we evaluate how classical transfer learning, _i.e._, using fine-tuning as in prior work, performs in the continual learning scenario. We fine-tune the models \(\mathcal{M}_{dec}\) and \(\mathcal{M}_{enc}\) sequentially on the stream of OOD datasets \(\mathcal{D}^{1}_{OOD},...,\mathcal{D}^{5}_{OOD}\). We refer to this approach as "naive fine-tuning", a common term used in the continual learning literature to refer to classical transfer learning, as it does not utilize mechanisms to address catastrophic forgetting. We report the results in terms of _Pass@1_ for API call prediction and _Accuracy (EM)_ for API usage prediction. Fig. 4 illustrates the evolution of the _Pass@1_ and _EM_ metrics on the OOD test sets throughout the fine-tuning steps for both models. Each column of a heatmap refers to the evolution of the performance of the model on a particular test set, and each row refers to a new incremental fine-tuning step. Note that we do not compute the metric on a test set whose corresponding training set has not been seen yet by the model. To quantify catastrophic forgetting, we report the Forgetting (\(F\)) metrics of the _Pass@1_ and _EM_ metrics in Table 4. We do not report all the values for every previously introduced metric as we have a strict page limit, and report them in our replication package. in Section 3.2 that \(\mathcal{M}_{enc}\) may be less stable than \(\mathcal{M}_{dec}\) due to the additional language modeling head randomly initialized. _Forgetting metrics_. In Table 4, we calculate the Forgetting metric for the _Pass@1_ and \(EM\) metrics and for both models. Note that we calculate the \(F\) metric at the final time step of the continual fine-tuning. According to the heatmaps of Fig. 4, the \(F^{5}\) metric of a domain is the difference between the first and last value of its corresponding column. This difference represents the amount of forgetting that has occurred on each OOD domain during fine-tuning. The \(\Delta t\) in the table indicates how recently the model was fine-tuned on a particular domain dataset. We notice that for the decoder \(\mathcal{M}_{dec}\), the forgetting is less severe for the _Pass@1_ (used in the API call prediction) than for the \(EM\) (used in the API usage prediction). The difference can be attributed to the fact that the API call prediction task is substantially easier than the API usage prediction task. In general, we observe more severe forgetting for the encoder, which further confirms our intuition about the lack of stability of \(\mathcal{M}_{enc}\). Our results and observations illustrate that the problem of forgetting about past data is a major issue for both studied models and significantly more severe for the model \(\mathcal{M}_{enc}\). Even with a low number of fine-tuning steps, catastrophic forgetting is already prominent. By considering more fine-tuning steps, we can expect the problem to exacerbate. We conclude that classical transfer learning, the most commonly used fine-tuning method in prior work, is not sufficient and robust enough to allow the model to adapt to new data while retaining knowledge of past data. ### How do continual learning approaches compare to classical transfer learning? To tackle the problem of catastrophic forgetting highlighted in our previous experiments, we propose to leverage some commonly used continual learning approaches from the literature. In this experiment, the naive fine-tuning approach is the lower-bound baseline, as it has no designed mechanism to mitigate catastrophic forgetting. We begin by introducing an upper-bound approach, referred to as "cumulative fine-tuning", which involves storing all training samples from each OOD training set cumulatively. With this approach, we perform continual fine-tuning using all samples from previous fine-tuning steps in addition to the current ones. This approach is usually upper-bound in continual learning settings as by storing all samples from previous data, the model can optimize its learning to generalize better to the whole stream of data. However, the cumulative fine-tuning approach is not usable in practice for a couple of reasons: (1) we may not always have access to all previous data at any given time, and (2) it requires storing all previous samples and significantly more computations during fine-tuning. This upper-bound approach aims to minimize forgetting while achieving the best overall performance. We compare the cumulative and naive approaches in Fig. 5 and Fig. 6. Next, we introduce additional CL methods, including a replay-based method and three regularization-based methods: EWC [31], SI [66], and RWalk [9]. One advantage of these three methods over the replay method is that they do not require storing samples from previous data while fine-tuning. We report the Average (\(A\)) and Forgetting (\(F\)) metrics for both tasks and models on the _Pass@1_ and \(EM\) metrics in Table 5 and Table 6. Note that there is no Forgetting metric for Guava as it is the last domain the PLMs are fine-tuned on. _Fine-tuning details_. We use the same fine-tuning procedure as in the previous experiment. For the replay baseline, we set the buffer size to 200, _i.e.,_ number of sampled stored from past OOD training sets. We provide all our hyperparameters and further details about the implementations in our replication package. _Cumulative fine-tuning_. In Fig. 5, we compare the naive and cumulative approaches for the API call prediction task (_Pass@1_) on both decoder and encoder models. Each curve illustrates the evolution of the _Pass@1_ on a particular OOD test set. The figure further demonstrates how the naive approach (bottom-left part of the figure) with the encoder leads to significantly more forgetting than for the decoder, as previously discussed. At the left of Fig. 5, we observe that the cumulative fine-tuning approach effectively eliminates the catastrophic forgetting issue for both models. Specifically, the _Pass@1_ does not decrease over time and even increases Figure 5. Comparison of naive and cumulative fine-tuning settings for both models on API call prediction. Figure 6. Comparison of naive and cumulative fine-tuning settings for both models on API usage prediction. throughout the fine-tuning, indicating improvement during continual fine-tuning, also known as positive transfer. In Fig. 6, we make the same observations for the API usage prediction task (_EM_). Continual learning approachesTable 5 reports the Average and Forgetting metrics of the _Pass@1_ on each OOD test set for \(\mathcal{M}_{dec}\) and \(\mathcal{M}_{enc}\), with the naive fine-tuning approach as baseline. Similarly to Section 4.2, we compute the \(F\) metric at the end of the continual fine-tuning. Firstly, we observe that for both models, the cumulative fine-tuning approach is the best option to mitigate catastrophic forgetting and generally leads to the best \(A_{Pass@1}\). With the cumulative approach, the \(F^{5}_{Pass@1}\) metric is always negative, which indicates a positive transfer (an increase in the _Pass@1_). For instance, we get \(-8.02\) in \(F^{5}_{Pass@1}\) for \(\mathcal{M}_{dec}\) in the Security domain, _i.e._, an increase of +8.02 in the metric through fine-tuning. However, we observe large gaps between the \(A_{Pass@1}\) obtained using the cumulative approach and the naive approach on the Guava dataset (last fine-tuning step). We hypothesize that with an ever-increasing replay buffer, the models can no longer learn from new data and thus lose their ability to adapt with time. In addition to being computationally intensive, the cumulative fine-tuning approach is not scalable and robust, as previously mentioned. Overall, all other CL approaches, except EWC, greatly reduce forgetting and show a superior average _Pass@1_ compared to the naive approach. The Replay approach generally produces the best or second best \(A_{Pass@1}\). Without the cumulative approach, RWalk is the best method to mitigate forgetting for \(\mathcal{M}_{dec}\), whereas SI is better for \(\mathcal{M}_{enc}\). In Table 6, we report the results for the API usage prediction task. We observe similar trends, except that the Replay approach is less effective for both models. However, RWalk and SI are the best methods for \(\mathcal{M}_{dec}\) and \(\mathcal{M}_{enc}\), respectively. In this final experiment, we demonstrate that continual learning methods, including two replay-based methods (Replay and Cumulative) and two regularization-based methods (SI and RWalk) effectively reduces catastrophic forgetting while achieving similar or superior effectiveness compared to classical transfer learning on both tasks. ## 5. Discussion In this section, we address some threats to the validity of our study. We then discuss the broader impact of our study and various opportunities for future work. \begin{table} \begin{tabular}{c c c c c c c c c c c} \hline \hline & & \multicolumn{2}{c}{General} & \multicolumn{2}{c}{Security} & \multicolumn{2}{c}{Android} & \multicolumn{2}{c}{Web} & \multicolumn{2}{c}{Guava} \\ \cline{3-11} Model & Method & \(\mathcal{A}_{EM}\uparrow\) & \(F^{5}_{EM}\downarrow\) & \(\mathcal{A}_{EM}\) & \(F^{5}_{EM}\) & \(\mathcal{A}_{EM}\) & \(F^{5}_{EM}\) & \(\mathcal{A}_{EM}\) & \(F^{5}_{EM}\) & \(\mathcal{A}_{EM}\) & \(F^{5}_{EM}\) \\ \hline \multirow{8}{*}{\(\mathcal{M}_{dec}\)} & Naive & 37.32 & 13.00 & 44.96 & 13.55 & 32.31 & 10.68 & 41.90 & 5.09 & 44.87 & – \\ & EWC [(31)] & 36.88 & 12.95 & 44.84 & 13.08 & 33.92 & 9.46 & 39.00 & 6.73 & 45.71 & – \\ & SI [(66)] & 40.36 & 8.26 & 49.88 & 6.89 & 30.01 & 3.24 & 36.95 & 16.65 & 43.14 & – \\ & RWalk [(9)] & 40.43 & 6.23 & 47.11 & 40.4 & 33.34 & 2.63 & 36.54 & 2.13 & 41.22 & – \\ & Replay & 39.49 & 11.11 & 46.88 & 8.21 & 33.39 & 7.63 & 39.49 & 6.08 & 43.65 & – \\ & Cumulative & **43.28** & **2.02** & **47.26** & **13.33** & **56.09** & **-2.28** & 27.92 & **-4.59** & 31.35 & – \\ \hline \hline \multirow{8}{*}{\(\mathcal{M}_{enc}\)} & Naive & 21.41 & 11.80 & 24.09 & 22.74 & 19.30 & 11.91 & 26.32 & 7.23 & 25.71 & – \\ & EWC [(31)] & 21.32 & 11.53 & 26.36 & 21.02 & 19.43 & 11.96 & 25.74 & 8.38 & 28.74 & – \\ \cline{1-1} & SI [(66)] & 27.22 & 50.08 & 30.85 & 8.28 & 18.57 & 22.00 & 23.03 & 16.5 & 21.26 & – \\ \cline{1-1} & RWalk [(9)] & 25.21 & 8.80 & 29.25 & 12.23 & 19.10 & 7.62 & 25.00 & 4.28 & 24.23 & – \\ \cline{1-1} & Replay & 25.48 & 13.54 & 29.94 & 13.96 & 18.09 & 11.88 & 24.51 & 5.92 & 26.48 & – \\ \cline{1-1} & Cumulative & **30.50** & **35.89** & **6.88** & **24.81** & **-4.88** & 21.88 & **-1.97** & 18.43 & – \\ \hline \hline \end{tabular} \end{table} Table 6. Continual learning approaches results for API usage prediction using the Accuracy (EM) metric. \begin{table} \begin{tabular}{c c c c c c c c c c c} \hline \hline & & \multicolumn{2}{c}{General} & \multicolumn{2}{c}{Security} & \multicolumn{2}{c}{Android} & \multicolumn{2}{c}{Web} & \multicolumn{2}{c}{Guava} \\ \cline{3-11} Model & Method & \(\mathcal{A}_{\text{Pnas1}}\uparrow\) & \(F^{5}_{\text{Pnas1}}\downarrow\) & \(\mathcal{A}_{\text{Pnas1}}\uparrow\) & \(\mathcal{B}^{5}_{\text{Pnas1}}\downarrow\) & \(\mathcal{A}_{\text{Pnas1}}\uparrow\) & \(\mathcal{B}^{5}_{\text{Pnas1}}\downarrow\) & \(\mathcal{A}_{\text{Pnas1}}\uparrow\) & \(\mathcal{B}^{5}_{\text{Pnas1}}\downarrow\) & \(\mathcal{A}_{\text{Pnas1}}\uparrow\) & \(\mathcal{B}^{5}_{\text{Pnas1}}\downarrow\) \\ \hline \multirow{8}{*}{\(\mathcal{M}_{dec}\)} & Naive & 53.49 & 5.64 & 57.21 & 6.71 & 32.75 & 6.77 & 40.06 & 1.80 & 50.47 & – \\ & EWC [(31)] & 53.22 & 7.02 & 57.16 & 7.49 & 33.73 & 5.72 & 60.14 & 3.77 & 49.59 & – \\ & SI [(66)] & 54.65 & 3.57 & **59.26** & 3.45 & 34.04 & 2.39 & 38.93 & 13.60 & 48.16 & – \\ & RWalk [(9)] & 54.38 & 2.39 & 57.39 & 2.80 & 31.64 & 1.97 & 38.19 & 1.65 & 45.28 & – \\ & Replay & **35.66** & 4.41 & 58.87 & 2.98 & 34.60 & 2.01 & **11.12** & 2.41 & 80.72 & – \\ & Cumulative & 55.63 & **0.31** & 58.44 & **8.02** & **35.74** & **9.73** & 32.99 & **-3.01** & 42.79 & – \\ \hline \hline \multirow{8}{*}{\(\mathcal{M}_{enc}\)} & Naive & 38.78 & 10.99 & 40.49 & 23.38 & 24.01 & 11.15 & 30.05 & 10.99 & 38.55 & – \\ & EWC [(31)] & 39.38 & 9.84 & 44.10 & 22.15 & 23.93 & 10.58 & 29.22 & 7.53 & 40.66 & – \\ \cline{1-1} & SI [(66)] & 44.29 & 5.94 & 50.05 & 13.10 & 21.39 & 60.22 & 27.79 & 25.05 & 35.87 & – \\ \cline{1-1} & RWalk [(9)] & 43.42 & 6.07 & 48.05 & 14.74 & 22.23 & 7.10 & 29.75 & 4.37 & 36.10 & – \\ \cline{1-1} & Replay & 45.15 & 5.48 & 51.56 & 10.56 & 24.31 & 8.27 & 28.58 & 3.92 & 40.22 & – \\ \cline{1-1} & Cumulative & **48.06** & **0.92** & **56.04** & **34.35** & ** ### Threats to validity _Threats to external validity_. We identified a main threat regarding the monolingual aspect of our dataset. Our OOD scenario requires extracting API usage sequences from the source code. Therefore, integrating more programming languages demands substantial additional effort, which we deliberately leave for future work. In addition, the construction of our dataset does not include any programming language-specific design and avoids any data leakage between the ID and OOD data. Consequently, it is highly likely that our results are not affected by the programming language of the data. Another threat related to the data is the choice of the OOD domains and APIs. To mitigate this threat, we selected five domains covering different types of programs. Specifically, we selected 10 random interfaces per domain. Our results show that catastrophic forgetting is observed consistently for all domains, and the selection of different interfaces would result in different intensities in forgetting. We leave the study of this qualitative aspect for future work. The choice of the downstream tasks presents another external threat to validity of our study. We employed two generation tasks, API call and API usage prediction. We focus on APIs-related tasks because APIs are an important part of the distribution of code tokens in programs and give lots of information about the semantics of programs. We observe significant catastrophic forgetting in these two API-related tasks and hypothesize that catastrophic forgetting could appear in other SE tasks because of the importance of APIs in code. For instance, previous work found that APIs play important roles in writing the summarization of code (Krishnam et al., 2017), detecting code clones (Krishnam et al., 2017), retrieving code given a query (Krishnam et al., 2017), etc. We leave the investigation of the OOD phenomenon in other tasks as future work. We identified an external threat to validity related to the limited number of fine-tuning steps in our continual fine-tuning settings. In practice, a PLM deployed to a real production environment would potentially face a larger number of fine-tuning steps throughout its lifetime. In this paper, we showed that both PLMs suffer from severe catastrophic forgetting, although we only consider five fine-tuning steps. We also demonstrated that more steps generally result in more forgetting about past data. Finally, the selection of the size of the PLMs, in terms of the number of trainable parameters, constitutes a potential threat to the validity of our study. While increasing the number of parameters may still result in OOD generalization issues due to the design of our datasets, it is uncertain whether catastrophic forgetting would occur with the same magnitude for larger models. Our experiments were performed under limited computational resources, which required us to consider architectures with a limited number of parameters. To mitigate this threat, we maximized the size of the models considering our limited resources. We pre-train PLMs with 110M and 125M parameters which are within the range of PLMs such as CodeBERT (Dosov et al., 2017), CodeT5 (Zhu et al., 2018) or CodeGPT (Zhu et al., 2018). _Threats to internal validity_. The hyperparameter choices for our CL approaches constitute the main threat to internal validity. We selected our hyperparameters based on values used in prior works about continual learning (Krishnam et al., 2017; Krishnam et al., 2017; Krishnam et al., 2018; Krishnam et al., 2018). These hyperparameters can be optimized for our scenario by using search methods, which tend to have a high computational cost. However, this aspect is not critical to the study as we have already shown the advantages of incorporating continual learning techniques with reasonable hyperparameter values. _Threats to construct validity_. We identified one threat to construct validity related to the choice of our evaluation metrics. We mitigate this threat by selecting metrics widely used in prior works to evaluate code generation tasks (Zhu et al., 2018; Krishnam et al., 2018). Additionally, we adapted continual learning metrics from prior works (Krishnam et al., 2017; Krishnam et al., 2018) to evaluate our continual fine-tuning scenario. ### Broader impact and opportunities Our study sheds light on the performance of PLMs of code in a continual learning setting for out-of-distribution generalization. We believe that this initial exploration of continual learning for code (_CL4Code_) will inspire further investigation in this important area. Our findings highlight two potential areas for future research: improving dataset and benchmark creation, and expanding the application of CL4Code to a wider range of use cases. _Datasets and benchmarks_. Our findings in Section 4.1 highlight a substantial disparity in the performance of a PLM between ID and OOD data. Our results, along with a previous work (Zhu et al., 2018), indicate that evaluating PLMs on ID data often leads to inflated metrics and results in overly optimistic conclusions in terms of the performance. Therefore, it is crucial to develop OOD datasets for code in order to evaluate the real-world generalizability of PLMs, as previously emphasized (Zhu et al., 2018; Zhu et al., 2018). Moreover, aligning dataset designs with continual learning scenarios offers the potential to evaluate the PLM's ability to adapt to changing environments, which is crucial for practical deployment. Improving benchmarks for PLMs of code is another promising direction for future research. Benchmarks such as CodeXGlue (Zhu et al., 2018) play a crucial role by providing standardized evaluations of models of code and enabling reproducible experimental results. However, as such researches progress at a rapid pace, widely used benchmarks often become outdated quickly. In particular, Kiela et al. (2018) showed that benchmarks such as GLUE (Zhu et al., 2018) in NLP saturate, meaning the milestones set by the benchmark are reached. Thus, continued efforts to enhance benchmarks in deep learning for code are vital in establishing concrete goals and driving research to enhance the performance of the models being evaluated. Recently, Yang et al. (2018) proposed GLUE-X, a comprehensive benchmark consisting of 13 datasets to test PLMs on OOD data across eight NLP tasks. The benchmark includes OOD datasets that are distinct from those in the original GLUE benchmark. Developing OOD benchmarks for code similar to GLUE-X (Zhu et al., 2018) would greatly contribute to the growth of research on OOD generalization for PLMs of code. One potential approach is to compile a new set of OOD datasets that are not included in the existing CodeXGlue benchmark, and use them to test PLMs of code. Furthermore, exploring the design of OOD scenarios specific to software changes, as demonstrated in the present study, can provide a valuable foundation for future code benchmark initiatives. Our dataset and methodology for extracting OOD samples for API evolution scenarios can serve as a starting point for these endeavors. _Continual learning for code_. Our findings in Section 4.2 highlight the challenge of catastrophic forgetting that PLMs of code encounter in a continual fine-tuning scenario with OOD data. Our study serves as a starting point for exploring the adaptability of PLMs of code in a variety of continual learning scenarios. For instance, these scenarios can be based on domain adaptation, where PLMs must adapt to new kinds of data such as new, unseen programming languages or code repositories as discussed in prior studies (Zang et al., 2019; Zhang et al., 2020; Zhang et al., 2021). Additionally, incorporating continual learning into a multi-task learning framework is highly relevant to software engineering, given the multitude of downstream tasks involved. In Section 4.3, our results demonstrate the effectiveness of continual learning methods in mitigating catastrophic forgetting in PLMs of code. We chose to explore these widely used methods as a first step in the research on continual learning for code. In the future, more sophisticated techniques from NLP, as discussed in Section 6.2, can be evaluated. Furthermore, the creation of continual learning methods specifically tailored to source code has the potential to further reduce catastrophic forgetting in PLMs of code. ## 6. Related Work ### Out-of-distribution generalization _Natural language processing_. Recent studies have revealed that PLMs are susceptible to generating inaccurate predictions when encountering OOD data (Zang et al., 2019; Zhang et al., 2020). In NLP, this issue can manifest itself in situations where the domain of the test data differs from the pre-training data (Zang et al., 2019). One approach to addressing this problem is to fine-tune PLMs on domain-specific datasets using efficient transfer learning techniques. For example, (Zang et al., 2019; Zhang et al., 2020) demonstrated that such approaches help PLMs in learning domain-specific knowledge and improve their generalization to unseen domains. Additionally, new datasets and benchmarks allow for further research on PLM domain adaptation. For instance, Williams et al. (Williams et al., 2020) introduced the MultiNLI dataset, containing text data from a variety of domains for PLM domain adaptation. Conneau et al. (Conneau et al., 2017) proposed a cross-lingual NLI dataset for evaluating the cross-lingual transferability of PLMs. Recently, Yang et al. (Yang et al., 2020) introduced GLUE-X, a benchmark for evaluating PLMs' ability to generalize to OOD data. _Deep learning for code_. The study of OOD generalization of PLMs of code is an emerging research area. Assessing their generalizability and designing efficient techniques to improve their robustness to OOD scenarios is essential for the practical usability of PLMs of code (Zang et al., 2020). Previous work in this field has focused on designing OOD datasets that simulate specific distribution shifts of program data. Koh et al. (Koh et al., 2020) presented PY150-Wilds, a Python dataset in which the test data consists of code repositories not appearing in the training data. The authors demonstrated performance gaps between the model on ID and OOD data. However, it is important to note that while the design choice is sound, it may not reflect strong OOD phenomena as the distribution of code tokens across different repositories may still be highly similar. More recently, Hu et al. (Hu et al., 2020) proposed a benchmark to evaluate the performance of code models under different distribution shift scenarios, including programmer, time, or token distribution shifts. In their study, the authors found that PLMs such as CodeBERT were robust against distribution shifts. However, they demonstrated that on a simple classification task with small datasets. In addition, the authors did not control the pre-training data of the studied PLMs, which can result in important data leakage between the pre-training and OOD test data. This problem of data leakage is critical as some of the test data may have been seen by the model during pre-training. Overall, this is a prime threat to the validity of the OOD scenario that may lead to obtaining inflated metrics on the OOD test data. Finally, Hajipour et al. (Hajipour et al., 2020) analyzed the performance of PLMs of code on a syntax-based, semantic-based and complexity-based OOD scenario and highlighted that the models exhibit poor generalizability when faced with OOD samples. However, it is important to point out that the OOD scenarios used in this study may be too artificial. For instance, in the syntax-based scenario, some language-specific tokens are masked at training to study how the model generalizes to unseen language tokens. Such a scenario is unrealistic as it does not reflect the nature of OOD data that a PLM of code is likely to encounter in the real world. Additionally, there is no practical motivation for masking specific tokens while training the model. In this study, we propose an OOD dataset that accurately represents the dynamic nature of software codebases in the real world. Specifically, we focus on the scenario where a PLM must adapt to new, unseen APIs over time, a well-established problem in the literature (Zang et al., 2019; Zhang et al., 2020). To ensure the validity of our experiments, we thoroughly control our PLM setup to prevent any data leakage between the pre-training, fine-tuning, and test data. This allows us to create an OOD generalization scenario that is as close to reality as possible, an aspect that has been overlooked in previous works. ### Continual learning for pre-trained language models Continual learning has been studied to adapt pre-trained language models based on the Transformer architecture (Zang et al., 2020) to new domains or tasks in NLP. For example, Cao et al. (Cao et al., 2020) proposed a method to continually learn from new classes of events in textual data to detect them without degradation of the accuracy over time. Douillard et al. (Doillard et al., 2019) introduced DyTox, a method that utilizes an encoder-decoder transformer for multiple tasks by expanding the network with task-specific special tokens, allowing for continual learning of new tasks with a low computational and memory footprint. Ermis et al. (Ermis et al., 2020) proposed a memory-efficient approach for transformers to continually learn new tasks by sharing information across tasks and expanding the network with task-specific modules. Similarly, Vladymyrov et al. (Vladymyrov et al., 2020) proposed the HyperTransformer architecture to continually learn new tasks by generating task-specific convolutional neural network weights in a few-shot learning setting and updating the task-specific weights to avoid catastrophic forgetting. Lastly, Jie et al. (Jie et al., 2020) leverage continual learning to avoid representational shifts in PLMs by proposing a new hierarchical fine-tuning method that prevents excessive changes in the representation spaces of the neural network in a continual fine-tuning setting. Recent advances in NLP highlight the crucial need for PLMs to adapt to changing environments and maintain their performance on new data and tasks. In the field of software engineering, the application of continual learning to PLMs of code is essential for developing methods that enable the model to robustly adapt to new codebases and tasks over time. To the best of our knowledge, there are no existing studies that employ continual learning in this context. Our work breaks new ground by introducing the first continual learning scenario for PLMs of code to continuously learn from new out-of-distribution APIs over time. ## 7. Conclusion and Future Work Our study exposes the limitations of pre-trained language models of code in handling out-of-distribution data in a continual fine-tuning scenario. Our results reveal that OOD data significantly decreases the PLMs' effectiveness in two API-related downstream tasks compared to ID data. Our findings indicate that classical transfer learning fails to adapt the PLMs to new, unseen APIs in this evolution scenario. Additionally, we observe instances of catastrophic forgetting, prompting us to explore methods that address this issue. In our final experiments, we demonstrate that replay-based and regularization-based continual learning techniques can effectively mitigate catastrophic forgetting while retaining or enhancing the performance of the PLMs in both downstream tasks. In future work, we intend to explore more OOD scenarios to further evaluate the generalizability of PLMs of code and develop relevant OOD generalization benchmarks for code. Additionally, we plan to implement more advanced continual learning methods tailored to source code to enhance the adaptability of PLMs of code. Finally, we aim to investigate OOD detection methods to automatically identify OOD data in PLMs, thereby improving their performance. ## Data Availability We publicly release all the code, data and models to reproduce the experiments of our study. The following repository contains instructions on how to acquire the data and pre-train, fine-tune and test the PLMs: [https://anonymous.4open.science/r/cl4code-ood-apdis-2490/](https://anonymous.4open.science/r/cl4code-ood-apdis-2490/)

2305.05218

Graph Neural Network-based surrogate model for granular flows

Accurate simulation of granular flow dynamics is crucial for assessing various geotechnical risks, including landslides and debris flows. Granular flows involve a dynamic rearrangement of particles exhibiting complex transitions from solid-like to fluid-like responses. Traditional continuum and discrete numerical methods are limited by their computational cost in simulating large-scale systems. Statistical or machine learning-based models offer an alternative. Still, they are largely empirical, based on a limited set of parameters. Due to their permutation-dependent learning, traditional machine learning-based models require huge training data to generalize. To resolve these problems, we use a graph neural network, a state-of-the-art machine learning architecture that learns local interactions. Graphs represent the state of dynamically changing granular flows and the interaction laws, such as energy and momentum exchange between grains. We develop a graph neural network-based simulator (GNS) that takes the current state of granular flow and predicts the next state using Euler explicit integration by learning the local interaction laws. We train GNS on different granular trajectories. We then assess the performance of GNS by predicting granular column collapse. GNS accurately predicts flow dynamics for column collapses with different aspect ratios unseen during training. GNS is hundreds of times faster than high-fidelity numerical simulators. The model also generalizes to domains much larger than the training data, handling more than twice the number of particles than it was trained on.

Yongjin Choi, Krishna Kumar

2023-05-09T07:28:12Z

http://arxiv.org/abs/2305.05218v2

# Graph Neural Network-based surrogate model for granular flows ###### Abstract Accurate simulation of granular flow dynamics is crucial for assessing various geotechnical risks, including landslides and debris flows. Granular flows involve a dynamic rearrangement of particles exhibiting complex transitions from solid-like to fluid-like responses. Traditional continuum and discrete numerical methods are limited by their computational cost in simulating large-scale systems. Statistical or machine learning-based models offer an alternative. Still, they are largely empirical, based on a limited set of parameters. Due to their permutation-dependent learning, traditional machine learning-based models require huge training data to generalize. To resolve these problems, we use graph neural network, a state-of-the-art machine learning architecture that learns local interactions. Graphs represent the state of dynamically changing granular flows and the interaction laws, such as energy and momentum exchange between grains. We develop a graph neural network-based simulator (GNS) that takes the current state of granular flow and predicts the next state using Euler explicit integration by learning the local interaction laws. We train GNS on different granular trajectories. We then assess the performance of GNS by predicting granular column collapse. GNS accurately predicts flow dynamics for column collapses with different aspect ratios unseen during training. GNS is hundreds of times faster than high-fidelity numerical simulators. The model also generalizes to domains much larger than the training data, handling more than twice the number of particles than it was trained on. keywords: graph neural network, learned physics simulator, granular column collapse, surrogate model + Footnote †: journal: Computers and Geotechnics ## 1 Introduction Landslides cause extensive material displacement and significant infrastructure damage. Accurate modeling of granular flow runout is crucial to understanding the impact of landslides. Numerical methods, such as particle-based and continuum approaches, are often employed to assess landslide runouts. Particle-based approaches, like the Discrete Element Method (DEM) (Staron and Hinch, 2005; Kermani et al., 2015; Kumar et al., 2017), can model grain-grain interactions but are limited to representative elemental volumes. Traditional continuum approaches, such as the Finite Element Method, can predict the initiation of such failures but suffer from mesh distortions when capturing runout dynamics. Hybrid Eulerian-Lagrangian approaches like the Material Point Method (MPM) (Mast et al., 2014; Kumar et al., 2017) can simulate large-deformation flows without undergoing mesh distortions. However, the hybrid nature of MPM requires tracking both the grid and the material points, which is computationally expensive. Multiple full-scale simulations are necessary for a comprehensive evaluation of runout hazard scenarios. Similarly, a back analysis to estimate material parameters requires a broad parametric sweep involving hundreds to thousands of simulations. However, current state-of-the-art numerical methods are restricted to, at most, a few full-scale simulations, limiting our ability in scenario testing or back analysis. An alternative to numerical simulations is the development of statistical or machine learning models to evaluate landslide risks. These surrogate models build correlations between landslide risks and their influencing factors through simple empirical correlation without considering the complex granular flow dynamics. Several studies adopt probabilistic approaches, such as Monte Carlo simulation and Bayesian analysis, to evaluate the landslide runout distance based on factors including topology and geology (Gao et al., 2021; Zeng et al., 2021; Sun et al., 2021; Zhao et al., 2022). Machine learning models can predict the travel distance and potential path of granular flows based on the geometry and ground properties (Durante and Rathje, 2021; Ju et al., 2022; Yang and Hambleton, 2021). Although researchers have been able to correlate the runout of granular flow based on statistical or data-driven techniques, these techniques do not explicitly consider granular flow dynamics--the actual physics governing the flow behavior. Thus, due to a lack of physics, these statistical models do not generalize outside their training range in modeling other boundary conditions or geometry. Building surrogate models that replicate the entire granular flow dynamics is challenging. The surrogate model must capture complex behaviors involving highly non-linear, static, collisional, and frictional dissipation regimes (Soga et al., 2016). Learning fundamental interaction laws is crucial for generalizing beyond the training datasets. Techniques like max-pooling in convolutional neural networks learn spatially invariant behavior, i.e., they learn features irrespective of their spatial location. However, CNNs are primarily limited to mesh-based systems with fixed neighbors. Granular flow is a dynamic system where neighbor interactions evolve throughout the runout (Lajeunesse et al., 2005; Zhang et al., 2016; Soga et al., 2016). A traditional Multi-Layer Perceptron (MLP) could model such a dynamic system. However, generalizing MLPs requires an exhaustive dataset to overcome combinatorial dependence, i.e., the outputs of the models depend on the order of the inputs (Battaglia et al., 2018; Haeri and Skoniczny, 2022). Unreasonably large training datasets are needed to map the entire parameter space of particle arrangements and dynamics. To address these limitations, we utilize graph neural networks (GNNs), a state-of-the-art machine learning architecture that enables permutation invariant learning (Battaglia et al., 2016, 2018; Sanchez-Gonzalez et al., 2020), to develop a data-driven surrogate model for granular flow dynamics. At any given time, the physical state of the granular system is represented as a graph. We develop a GNN-based Simulator (GNS) that operates on the graph to learn the fundamental interaction law. We demonstrate the capability of GNS in replicating the granular flow dynamics by studying the collapse of a granular column. Granular column collapse is a simple physical experiment that captures the overall dynamics of large-scale runouts. GNS, trained on granular flow trajectories, successfully predicts the runout dynamics of column collapse outside its training range and generalizes to upscaled domain sizes. ## 2 Method This section describes the individual components of GNS: graphs, graph neural networks (GNNs), and message passing. ### Graph Neural Networks and Message Passing #### 2.1.1 Graphs Graphs can represent interactions in physical systems (Battaglia et al., 2016; Sanchez-Gonzalez et al., 2020). We represent the granular media as a graph \(G=(\mathbf{V},\mathbf{E})\) consisting of a set of vertices (\(\mathbf{v}_{i}\in\mathbf{V}\)) representing the soil grains or aggregation of grains and edges (\(\mathbf{e}_{i,j}\in\mathbf{E}\)) connecting a pair of vertices (\(\mathbf{v}_{i}\) and \(\mathbf{v}_{j}\)) representing the interaction between the grains. Consider an example involving interaction between grains in a box (see fig. 1). We encode the state of the physical system, such as the kinematics of grains and their interaction (fig. 1a and fig. 1d), as a graph (fig. 1b and fig. 1c). The vertices describe the position and velocity of the grains, and the edges describe the directional interaction between them, shown as arrows in fig. 1b and fig. 1c. The state of the grain \(i\) is represented as a vertex feature vector \(\mathbf{v}_{i}\). The vertex feature vector includes velocities, mass, and distance to the boundary. The edge feature vector \(\mathbf{e}_{i,j}\) includes information about the interaction between grains \(i\) and \(j\) such as the relative distance between the grains. Thus, we can store and process the state of granular bodies and their interactions as graphs. Graphs offer a permutation-invariant form of encoding data, where the interaction between vertices is independent of the order of vertices or their position in Euclidean space. As graphs represent the interactions between grains as edge connections, graphs are permutation invariants. For example, by storing the relative positional information in \(\mathbf{e}_{i,j}\), rather than the absolute position, machine learning models operating on these networks learn the interaction behavior of different relative distances between grains. Therefore, graphs can efficiently represent the physical state of granular flow involving multi-grain interactions. #### 2.1.2 Graph neural networks (GNNs) GNN takes a graph \(G=(\mathbf{V},\mathbf{E})\) as an input, computes properties and updates the graph, and outputs an updated graph \(G^{\prime}=(\mathbf{V}^{\prime},\mathbf{E}^{\prime})\) with an identical structure, where \(\mathbf{V}^{\prime}\) and \(\mathbf{E}^{\prime}\) are the set of updated vertex and edge features (\(\mathbf{v}^{\prime}_{i}\) and \(\mathbf{e}^{\prime}_{i,j}\)). Message passing is the process of updating the graph by propagating information through it. In the grains-in-a-box example, the GNN first takes the original graph \(G=(\mathbf{V},\mathbf{E})\) (fig. 1b) that describes the current state of the physical system (\(\mathbf{X}_{t}\)). The GNN then updates the state of the physical system through message passing, which models the exchange of energy and momentum between the grains, and returns an updated graph \(G^{\prime}=(\mathbf{V}^{\prime},\mathbf{E}^{\prime})\) (fig. 1c). We decode \(G^{\prime}\), the output of GNN, to extract information related to the future state of the physical system (\(\mathbf{X}_{t+1}\)), such as the next position or acceleration of the grains (fig. 1d). #### 2.1.3 Message passing Message passing consists of three operations: message construction (eq. (1)), message aggregation (eq. (2)), and the vertex update function (eq. (3)). \[\mathbf{e}^{\prime}_{i,j}=\phi_{\mathbf{\Theta}_{\phi}}\left(\mathbf{v}_{i},\mathbf{v}_{j}, \mathbf{e}_{i,\ j}\right) \tag{1}\] \[\bar{\mathbf{v}}_{i}=\Sigma_{j\in N(i)}\ \mathbf{e}^{\prime}_{i,j} \tag{2}\] \[\mathbf{v}^{\prime}_{i}=\gamma_{\mathbf{\Theta}_{\gamma}}\left(\mathbf{v}_{i},\bar{\mathbf{v}}_{ i}\right) \tag{3}\] The subscript \(\mathbf{\Theta}_{\phi}\) and \(\mathbf{\Theta}_{\gamma}\) represent a set of learnable parameters in each computation. The message construction function \(\phi_{\mathbf{\Theta}_{\phi}}\) (eq. (1)) takes the feature vectors of the receiver and sender vertices (\(\mathbf{v}_{i}\) and \(\mathbf{v}_{j}\)) and the feature vector of the edge connecting them (\(\mathbf{e}_{i,\ j}\)) and returns an updated edge feature vector \(\mathbf{e}^{\prime}_{i,j}\) as the output. \(\phi_{\mathbf{\Theta}_{\phi}}\) is a matrix operation including the learnable parameter \(\mathbf{\Theta}_{\phi}\). The updated edge feature vector \(\mathbf{e}^{\prime}_{i,j}\) is the message sent from vertex \(j\) to \(i\). Figure 1(a) shows an example of constructing messages on edges directed to vertex \(0\) originating from vertices \(1\), \(2\), and \(3\) (\(\mathbf{e}^{\prime}_{0,\ 1}\), \(\mathbf{e}^{\prime}_{0,\ 2}\), \(\mathbf{e}^{\prime}_{0,\ 3}\)). Here, we define the message construction function \(\phi_{\mathbf{\Theta}_{\phi}}\) as \(((\mathbf{v}_{i}+\mathbf{v}_{j})\times\mathbf{e}_{i,\ j})\times\mathbf{\Theta}_{\phi}\). The updated feature vector \(\mathbf{e}^{\prime}_{0,\ 1}\) is computed as \(((\mathbf{v}_{0}+\mathbf{v}_{1})\times\mathbf{e}_{0,\ 1})\times\mathbf{\Theta}_{\phi}\), where \(\mathbf{v}_{0}\) and \(\mathbf{v}_{1}\) are the receiver and sender vertex feature vectors, and \(\mathbf{e}_{0,\ 1}\) is their edge feature vector. Suppose we assume all values of \(\mathbf{\Theta}_{\phi}\) are \(1.0\) for simplicity, we obtain \(\mathbf{e}^{\prime}_{0,\ 1}=(([1,\ 0,\ 2])+[1,\ 3,\ 2])\times[2,\ 1,\ 0]^{T})\times 1=[4,\ 3,\ 0]\). Similarly, we can compute the messages \(\mathbf{e}^{\prime}_{0,\ 2}=[0,\ 3,\ 9]\) and \(\mathbf{e}^{\prime}_{0,\ 3}=[3,\ 4,\ 9]\). The next step in message passing is the message aggregation \(\Sigma_{j\in N(i)}\) (eq. (2)), where \(N(i)\) is the set of sender vertices \(j\) related to vertex \(i\)). It collects all the messages directing to vertex \(i\) and aggregates those into a single vector with the same dimension as the aggregated message (\(\bar{\mathbf{v}}_{i}\)). The aggregation rule can be element-wise vector summation or averaging; hence it is a permutation invariant computation. In fig. 1(a), the aggregated message \(\bar{\mathbf{v}}_{0}=[7,\ 10,\ 18]\) is the element-wise summation of the messages directing to vertex \(0\) as \(\bar{\mathbf{v}}_{0}=\mathbf{e}^{\prime}_{0,\ 1}+\ \mathbf{e}^{\prime}_{0,\ 2}+\ \mathbf{e}^{ \prime}_{0,\ 3}\). The final step of the message passing is updating vertex features using eq. (3). It takes the aggregated message (\(\bar{\mathbf{v}}_{i}\)) and the current vertex feature vector \(\mathbf{v}_{i}\), and returns an updated vertex feature vector \(\mathbf{v}^{\prime}_{i}\), using predefined vector operations including the learnable parameter \(\mathbf{\Theta}_{\gamma}\). Figure 1(b) shows an example of the update at vertex \(0\). Here, we define the update function \(\gamma_{\mathbf{\Theta}_{\gamma}}\) as \(\mathbf{\Theta}_{\gamma}\left(\mathbf{v}_{i}+\bar{\mathbf{v}}_{i}\right)\). The updated feature vector \(\mathbf{v}^{\prime}_{0}\) is computed as \(\mathbf{\Theta}_{\gamma}\left(\mathbf{v}_{0}+\bar{\mathbf{v}}_{0}\right)\). Figure 1: An example of a graph and graph neural network (GNN) that process the graph (modified from Battaglia et al. (2018)): (a) A state of the current physical system (\(\mathbf{X}_{t}\)) where the grains are bouncing in a box boundary; (b) Graph representation of the physical system (\(G\)). There are three vertices representing grains and six edges representing their directional interaction shown as arrows; (c) The updated graph (\(G^{\prime}\)) that GNN outputs through message passing; (d) The predicted future state of the physical system (\(\mathbf{X}_{t+1}\)) (i.e., the positions of the grains at the next timestep) decoded from the updated graph. Assuming all parameters in \(\mathbf{\Theta}_{\gamma}\) are \(1.0\) for simplicity, we obtain \(\mathbf{v}_{0}^{\prime}=\mathbf{v}_{0}+\bar{\mathbf{v}}_{0}=[1,\ 0,\ 2]+[7,\ 10,\ 18]=[8,\ 10,\ 20]\). Similarly, we update the other vertex features \((\mathbf{v}_{1}^{\prime},\ \mathbf{v}_{2}^{\prime},\ \mathbf{v}_{3}^{\prime})\). After message passing, the graph vertex and edge features \((\mathbf{v}_{i}\) and \(\mathbf{e}_{i,\ j})\) are updated to \(\mathbf{v}_{i}^{\prime}\) and \(\mathbf{e}_{i,\ j}^{\prime}\). The GNN may include multiple message passing steps to propagate the information further through the network. Unlike the example shown above, where we assume a constant value of \(1.0\) for the learnable parameters, in a supervised learning environment, the optimization algorithm will find a set of the best learnable parameters \((\mathbf{\Theta}_{\phi},\mathbf{\Theta}_{\gamma})\) in the message passing operation. ### Graph Neural Network-based Simulator (GNS) In this study, we use GNN as a surrogate simulator to model granular flow behavior. Figure 3 shows an overview of the general concepts and structure of the GNN-based simulator (GNS) proposed by Sanchez-Gonzalez et al. (2020). Consider a granular flow domain represented as material points (fig. 3a), which represent the collection of grains. In GNS, we represent the physical state of the granular domain at time \(t\) with a set of \(\mathbf{x}_{i}^{t}\) describing the state and properties of each material point. The GNS takes the current state of the granular flow \(\mathbf{x}_{t}^{i}\in\mathbf{X}_{t}\) and predicts its next state \(\mathbf{x}_{i+1}^{i}\in\mathbf{X}_{t+1}\) (fig. 3a). The GNS consists of two components: a parameterized function approximator \(d_{\Theta}\) and an updater function (fig. 3b). The function approximator \(d_{\Theta}\) takes \(\mathbf{X}_{t}\) as an input and outputs dynamics information \(\mathbf{y}_{i}^{t}\in\mathbf{Y}_{t}\). The updater then computes \(\mathbf{X}_{t+1}\) using \(\mathbf{Y}_{t}\) and \(\mathbf{X}_{t}\). Figure 3c shows the details of \(d_{\Theta}\) which consists of an encoder, a processor, and a decoder. The encoder (fig. 3-c1) takes the state of the system \(\mathbf{X}^{t}\) and embeds it into a latent graph \(G_{0}=(\mathbf{V}_{0},\ \mathbf{E}_{0})\) to represent the relationships between material points. The vertices \(\mathbf{v}_{i}^{t}\in\mathbf{V}_{0}\) contain latent information of the current state of the material point, and the edges \(\mathbf{e}_{i,j}^{t}\in\mathbf{E}_{0}\) contain latent information of the pair-wise relationship between material points. Next, the processer (fig. 3-c2) converts the input graph \(G_{0}\) to the output graphs \(G_{M}\) through \(M\) stacks of message-passing GNN (\(G_{0}\to G_{1}\to\cdots\to G_{M}\)). The message passing computes the interaction between vertices. Finally, the decoder (fig. 3-c3) extracts the dynamics of the points \((\mathbf{Y}^{t})\) from \(G_{M}\), such as the acceleration of the physical system. The entire simulation (fig. 3a) involves running GNS surrogate model through \(K\) timesteps predicting from the initial state to \(\mathbf{X}_{K}\left(\mathbf{X}_{0},\ \mathbf{X}_{1},\ \dots,\ \mathbf{X}_{K}\right)\), updating at each step (\(\mathbf{X}_{t}\rightarrow\mathbf{X}_{t+1}\)). We call this successive prediction from GNS the "rollout". In the following sections, we explain the details of our input \(\mathbf{X}^{t}\) (fig. 3a), the encoder, processor, and decoder in \(d_{\Theta}\) (fig. 3c), and how we compute \(\mathbf{X}^{t+1}\) from \(\mathbf{X}^{t}\) using the GNS updater function (fig. 3b). #### 2.2.1 Input The input to the GNS, \(\mathbf{x}_{i}^{t}\in\mathbf{X}^{t}\) (eq. (4)), is a vector consisting of the current material point position \(\mathbf{p}_{i}^{t}\), the material point velocity context \(\mathbf{\hat{p}}_{i}^{\leq t}\), information on boundaries \(\mathbf{b}_{i}^{t}\), and material point type embedding \(\mathbf{f}\). The current state \(\mathbf{x}_{i}^{t}\) will be used to construct vertex feature (\(\mathbf{v}_{i}^{t}\)) (eq. (6)). \[\mathbf{x}_{i}^{t}=\left[\mathbf{p}_{i}^{t},\ \mathbf{p}_{i}^{\leq t},\ \mathbf{b}_{i}^{t},\ \mathbf{f}\right] \tag{4}\] The velocity context \(\mathbf{\hat{p}}_{i}^{\leq t}\) includes the current and previous material point velocities for \(n\) timesteps \(\left[\mathbf{\hat{p}}_{i}^{t-n},\cdots,\ \mathbf{\hat{p}}_{i}^{t}\right]\) with \(n+1\) velocities. We use \(n=4\) to include sufficient velocity context in the vertex feature \(\mathbf{v}_{i}^{t}\). Sanchez-Gonzalez et al. (2020) show that having \(n>1\) significantly improves the model performance. We compute the velocities using the finite difference of the position sequence (i.e., \(\mathbf{\hat{p}}_{i}^{t}=\left(\mathbf{p}_{i}^{t}-\mathbf{p}_{i}^{t-1}\right)/\Delta t\)). \(\mathbf{b}_{i}^{t}\) is boundary information. For a 2D problem, \(\mathbf{b}_{i}^{t}\) has four components, each indicating the distance between material points and the four walls. We normalize \(\mathbf{b}_{i}^{t}\) by the connectivity radius \(R\) which defines the interaction zone, explained in the next section, and restrict it between -1.0 to 1.0. \(\mathbf{b}_{i}^{t}\) is used to evaluate boundary interaction for a material point. \(\mathbf{f}\) is a vector embedding describing a material point type. We define the interaction between material points \(i\) and \(j\) as \(\mathbf{r}_{i,\ j}^{t}\) using the distance and displacement of the material points in the current timestep (see eq. (4)). The former reflects Figure 3: The structure of the graph neural network (GNN)-based physics simulator (GNS) for granular flow (modified from Sanchez-Gonzalez et al. (2020)): (a) The entire simulation procedure using the GNS, (b) The computation procedure of GNS and its composition, (c) The computation procedure of the parameterized function approximator \(d_{\Theta}\) and its composition. the level of interaction, and the latter reflects its spatial direction. \(\mathbf{r}_{i,\ j}^{t}\) will be used to construct edge features (\(\mathbf{e}_{i,j}^{t}\)). \[\mathbf{r}_{i,j}^{t}=\left[\left(\mathbf{p}_{i}^{t}-\mathbf{p}_{j}^{t}\right),\ \|\mathbf{p}_{i}^{t}-\mathbf{p}_{j}^{t}\|\right] \tag{5}\] #### 2.2.2 Encoder The vertex and edge encoders (\(\varepsilon_{\mathbf{\Theta}}^{v}\) and \(\varepsilon_{\mathbf{\Theta}}^{e}\)) convert \(\mathbf{x}_{i}^{t}\) and \(\mathbf{r}_{i,\ j}^{t}\) into the vertex and edge feature vectors (\(\mathbf{v}_{i}^{t}\) and \(\mathbf{e}_{i,j}^{t}\)) (eq. (6)) and embed them into a latent graph \(G_{0}=(\mathbf{V}_{0},\ \mathbf{E}_{0})\), \(\mathbf{v}_{i}^{t}\in\ \mathbf{V}_{0}\), \(\mathbf{e}_{i,j}^{t}\in\ \mathbf{E}_{0}\). \[\mathbf{v}_{i}^{t}=\varepsilon_{\mathbf{\Theta}}^{v}\left(\mathbf{x}_{i}^{t}\right),\ \mathbf{e}_{i,j}^{t}=\varepsilon_{\mathbf{\Theta}}^{e}\left(\mathbf{r}_{i,j}^{t}\right) \tag{6}\] We use a two-layered 128-dimensional multi-layer perceptron (MLP) for the \(\varepsilon_{\mathbf{\Theta}}^{v}\) and \(\varepsilon_{\mathbf{\Theta}}^{e}\). The MLP and optimization algorithm search for the best candidate for the parameter set \(\mathbf{\Theta}\) that estimates a proper way of representing the physical state of the material points and their relationship which will be embedded into \(G_{0}\). The edge encoder \(\varepsilon_{\mathbf{\Theta}}^{v}\) uses \(\mathbf{x}_{i}^{t}\) (eq. (4)) without the current position of the material point (\(\mathbf{p}_{i}^{t}\)), but with its velocities (\(\mathbf{\dot{\mathbf{p}}}_{i}^{\pm t}\)), as velocity governs the momentum, and the interaction dynamics is independent of the absolute position of the material points. Rubanova et al. (2022) confirmed that including position causes poorer model performance. We only use \(\mathbf{p}_{i}^{t}\) to predict the next position \(\mathbf{p}_{i}^{t+1}\) based on the predicted velocity \(\mathbf{\dot{\mathbf{p}}}_{i}^{t+1}\) using Explicit Euler integration. We consider the interaction between two material points by constructing edges between them all pairs of vertices located within a certain distance called connectivity radius \(R\) (see the shaded circular area in fig. 3b). The connectivity radius is a critical hyperparameter that governs how effectively the model learns the local interaction. \(R\) should be sufficiently large to include the local interaction between material points and capture the simulation domain's global dynamics. #### 2.2.3 Processor The processor performs message passing (based on eq. (1) to eq. (3)) on the initial latent graph (\(G_{0}\)) from the encoder for \(M\) times (\(G_{0}\to G_{1}\rightarrow\cdots\to G_{M}\)) and returns a final updated graph \(G_{M}\). We use two-layered 128-dimensional MLPs for both the message construction function \(\phi_{\mathbf{\Theta}_{\phi}}\) and vertex update function \(\gamma_{\mathbf{\Theta}_{r}}\), and element-wise summation for the message aggregation function \(\Sigma_{j\in N(i)}\) in eq. (1) to eq. (3). We set \(M=10\) to ensure sufficient message propagation through the network. These stacks of message passing model information propagation through the network of material points. #### 2.2.4 Decoder The decoder \(\delta_{\mathbf{\Theta}}^{v}\) extracts the dynamics \(\mathbf{y}_{i}^{t}\in\mathbf{Y}^{t}\) of the material points from the vertices \(\mathbf{v}_{i}^{t\prime}\) (eq. (7)) using the final graph \(G_{M}\). We use a two-layered 128-dimensional MLP for \(\delta_{\mathbf{\Theta}}^{v}\), which learns to extract the relevant dynamics for material points from \(G_{M}\). \[\mathbf{y}_{i}^{t}=\delta_{\mathbf{\Theta}}^{v}\left(\mathbf{v}_{i}^{t\prime}\right) \tag{7}\] #### 2.2.5 Update We use the dynamics \(\mathbf{y}_{i}^{t}\) to predict the velocity and position of the material points at the next timestep (\(\mathbf{\dot{p}}_{i}^{t+1}\) and \(\mathbf{p}_{i}^{t+1}\)) based on Euler integration (eq. (8) and eq. (9)), which makes \(\mathbf{y}_{i}^{t}\) analogous to acceleration \(\mathbf{\ddot{p}}_{i}^{t}\). \[\mathbf{\dot{p}}_{i}^{t+1}=\mathbf{\dot{p}}_{i}^{t}+\mathbf{y}_{i}^{t}\Delta\mathrm{t} \tag{8}\] \[\mathbf{p}_{i}^{t+1}=\mathbf{p}_{i}^{t}+\mathbf{\dot{p}}_{i}^{t+1}\Delta\mathrm{t} \tag{9}\] Based on the new position and velocity of the material points, we update \(\mathbf{x}_{i}^{t}\in\mathbf{X}^{t}\) (eq. (4)) to \(\mathbf{x}_{i}^{t+1}\in\mathbf{X}^{t+1}\). The updated physical state \(\mathbf{X}^{t+1}\) is then used to predict the position and velocity for the next timestep. The updater imposes inductive biases, such as an inertial frame, on GNS to force it only to learn the interaction dynamics, improving learning efficiency. A traditional neural network learns both the update scheme and the interaction dynamics: \[p^{t+1}=NN(p^{t},v^{t})\,. \tag{10}\] Whereas, using an inertial prior, we force the GNS only to learn the interaction dynamics, by hardcoding the update function: \[p^{t+1}=p^{t}+v^{t}\cdot\Delta t+NN(p^{t},v^{t})\,. \tag{11}\] Nevertheless, GNS does not directly predict the next position from the current position and velocity (i.e., \(\mathbf{p}_{i}^{t+1}=GNS\left(\mathbf{p}_{i}^{t},\ \mathbf{\dot{p}}_{i}^{t}\right)\)) which has to learn the static motion and inertial motion. Instead, it uses (1) the inertial prior (eq. (8)) where the prediction of the next velocity \(\mathbf{\dot{p}}_{i}^{t+1}\) should be based on the current velocity \(\mathbf{\dot{p}}_{i}^{t}\) and (2) the static prior (eq. (9)) where the prediction of the next position \(\mathbf{p}_{i}^{t+1}\) should be based on the current position \(\mathbf{p}_{i}^{t}\). These make GNS focus on learning unknown dynamics by hardcoding known physics. Since GNS learns the dynamics of material points through interactions independent of absolute position, GNS is generalizable to other geometric conditions. ## 3 Training and Evaluation We now train the GNS to predict granular column collapse. This section explains how we generate training data, details of the training process, and how we evaluate the performance of the GNS. ### Material Point Method We utilize the Material Point Method (MPM) to generate the GNS training dataset of granular flow simulations. The MPM is a hybrid Eulerian-Lagrangian approach designed for modeling large-deformation flows (Soga et al., 2016). In the MPM, a continuum body is represented by individual material points that traverse a static background grid. The governing equation is solved at the nodes, and the updated velocity field is subsequently mapped back to the material points. We employ the position information stored in these material points to construct the current state \(\mathbf{X}^{t}\) in the GNS. For more information on MPM refer to Soga et al. (2016). ### Datasets The training datasets include 26 granular flow trajectories of square-shaped granular mass in a two-dimensional box boundary simulated by the MPM explicit time integration method using the CB-Geo MPM code (Kumar et al., 2019). Each simulation has a different initial configuration regarding the size of the square granular mass, position, and velocity. Table 1 presents the details of the training dataset generated using the MPM simulation. The datasets are published on DesignSafe (Kumar and Choi, 2023). A shows all the training trajectories with different initial configurations and initial velocities. We also create the validation datasets to check if the model experiences overfitting. The datasets include seven trajectories of randomly picked rectangular-shaped granular mass with different initial configurations not included in the training datasets. ### Training Our GNS has a learnable parameter set \(\Theta\). We train \(\Theta\) to minimize the loss calculated as the mean squared error (MSE) between \(\mathbf{y}^{t}_{i}\) (predicted proxy-acceleration) and the ground truth acceleration \(\mathbf{\vec{p}}^{t}_{i}\) for all material points \(i=1,\ 2,\ \dots,\ N\) as shown in eq. (12) based on gradient (\(\nabla loss_{\Theta}\))-based optimizer, Adam (Kingma and Ba, 2014). \[loss_{\Theta}=\frac{1}{n}\sum_{i=1}^{N}\left(\mathbf{y}^{t}_{i}-\mathbf{\vec{p}}^{t}_ {i}\right)^{2} \tag{12}\] For training the GNS, we have to set hyperparameters to learn the flow behavior from the training trajectories properly. The first key hyperparameter is the connectivity radius \(R\) \begin{table} \begin{tabular}{l l l} \hline \hline \multicolumn{2}{c}{Property} & \multicolumn{1}{c}{Description} \\ \hline Simulation boundary & \multicolumn{1}{c}{1.0\(\times\)1.0 m} \\ Mesh size & \multicolumn{1}{c}{0.025\(\times\)0.025 m} \\ Material points per cell & \multicolumn{1}{c}{16} \\ Granular mass geometry & \multicolumn{1}{c}{0.2\(\times\)0.2 m and 0.3\(\times\)0.3 m} \\ Simulation duration (timesteps) & \multicolumn{1}{c}{(each includes 1,024 and 2,304 particles)} \\ \hline \multirow{6}{*}{Material property} & Model & Mohr-Coulomb \\ & Density & 1,800 \(kg/m^{3}\) \\ & Youngs modulus & 2 GPa \\ & Poisson ratio & 0.3 \\ & Friction angle & 30\({}^{\circ}\) \\ & Cohesion & 0.1 kPa \\ & Tension cutoff & 0.05 kPa \\ \hline \hline \end{tabular} \end{table} Table 1: Details of the Material Point Method (MPM) simulation geomaterials and properties used for generating the training datasets. which governs the model's capacity to learn the interactions of material points. We set \(R=0.030\) m which includes about 9 to 10 material points along the diameter. The circular area defined by \(R\) can incorporate approximately 70 material points inside. Another important hyperparameter is the Gaussian noise value for perturbing the ground truth position in the training trajectories. Since every predicted position for each timestep is based on the previous prediction, which includes a prediction error, the simulation over the large timesteps is subjected to an exponential error accumulation. To avoid this issue, we train the model on input positions with Gaussian noise that emulates the prediction error made by a one-step prediction (\(\mathbf{X}_{t}\rightarrow\mathbf{X}_{t+1}\)). The inclusion of noise in training leads to more rigorous long-rollout predictions. We use the learning rate (\(lr\)) decay with the initial value of \(10^{-4}\) and decay rate of 0.1 (\(lr=10^{-4}\times 0.1^{step/5\times 10^{6}}\)) for more stable convergence. We use the batch size of two, i.e., \(\mathbf{X}_{t}\) from two different trajectories are used simultaneously in updating the learnable parameters. For information on the scalability of the GNS algorithm, refer to Kumar and Vantassel (2022). We investigate if the model experiences overfitting by plotting the loss history (fig. 4) for the training and validation datasets evaluated for every 10K training steps. The training and validation losses show a drastic decrease until 2M steps. After that, the validation loss tends to remain slightly larger than the training loss. Figure 4 shows no overfitting during the training. ### GNS prediction of granular flows We trained the GNS to predict the collapse of a granular column (as studied by Lajeunesse et al. (2004); Lube et al. (2005)). Figure 5 shows the granular column collapse experiments to evaluate its ability to replicate granular flow dynamics. Granular column collapse is a simple physical experiment that captures the transient response of granular flow dynamics. The experiment involves the collapse of a granular column of initial height \(H_{0}\) and length \(L_{0}\) on a flat surface due to gravity. As the gate holding the column is removed, the granular Figure 4: Evolution of GNS loss in training and validation with epochs. material destabilizes, resulting in a runout. We measure the final runout deposit with the final height \(H_{f}\) and runout \(L_{f}\). The runout of the granular column is governed by the initial aspect ratio (\(a=H_{0}/L_{0}\)) (Staron and Hinch, 2005; Kumar, 2015). For short columns (\(a\lesssim 2\)) (fig. 5a), the soil mass fails along the flanks of the column above a well-defined failure surface (dashed line). The soil mass beneath the failure surface remains in static equilibrium throughout the collapse forming a truncated conical shape. With the increase in aspect ratio, the portion of the sliding mass above the failure surface increase, and the static part becomes smaller, forming a conical shape. For tall columns (\(a\gtrsim 2\)) (fig. 5b), the majority of the soil mass is involved in the collapse, and it initially experiences a free fall due to gravitational acceleration. As the falling mass reaches the failure surface, the vertical kinetic energy is converted to horizontal acceleration, resulting in a longer runout distance than the short column (fig. 5a). In addition, researchers (Kumar, 2015; Staron and Hinch, 2005; Kermani et al., 2015; Utili et al., 2015) observed a transition zone where the flow dynamics change from short to tall columns. The normalized runout (\(\left(L_{f}-L_{0}\right)/L_{0}\)) of a granular column is only a function of its aspect ratio (\(a\)). The normalized runout represents how far the granular mass runs out before reaching the final deposit state compared to the initial length of the column. Short columns show a linear relationship with the initial aspect ratio. In contrast, tall columns have a power-law relationship with the initial aspect ratio. The GNS was trained only on the aspect ratio of 1.0. However, we evaluate its performance in predicting the runout dynamics of other aspect ratios by comparing the GNS predictions with the MPM simulations. Table 2 presents the test cases for evaluating GNS performance. ## 4 Results and Discussions We evaluate the GNS predictions of granular column collapse against the MPM simulations in terms of the (1) geometry of sliding mass, (2) evolution of runout and height with time, and (3) energy evolution during the collapse. Figure 6 shows the normalized runout (\(\left(L_{f}-L_{0}\right)/L_{0}\)) predictions of GNS for different aspect ratios in comparison with MPM. \(L_{f}\) is the distance from the left wall to the material point that runs out the farthest, as shown in fig. 5. Previous research observed a transition zone for the relationship between the normalized runout and aspect ratio that distinguishes short-column from tall-column dynamics. Figure 5: Schematics for the granular column collapse configuration and its behavior on the aspect ratio. For both GNS and MPM, we observe the transition around an initial aspect ratio \(a=1.2\) (fig. 6). Table 3 summarizes the errors between GNS predictions and MPM simulations for different aspect ratios. In general, the GNS runout prediction is within 5% of the MPM runout estimate. Figure 6 suggests that the GNS successfully captures the dependence of the final runout with the initial aspect ratio, including the transition from the short to the tall column. ### GNS Predictions of Granular Flow Dynamics #### 4.1.1 Short Column We now evaluate the GNS rollout (prediction) of the granular flow dynamics with time for a short column (\(a=0.8\)). Figure 7 shows the time evolution of granular flow for the \begin{table} \begin{tabular}{c c c c c} \hline \hline \multirow{2}{*}{Test case} & \multirow{2}{*}{\(H_{0}\times L_{0}\)} & Duration & Simulation & Number of \\ & & & (timesteps) & boundary & material points \\ \hline Short & & & & X: 0 to 1.0 & \\ columns & \(a=0.5\) & 0.2 \(\times\) 0.4 & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 0.5 \\ Y: 0 to 1.0 \\ Y: 0 to 0.5 \\ Y: 0 to 0.5 \\ Y: 0 to 0.5 \\ \end{tabular} & 1956 \\ & \(a=0.8\) & \(0.24\times 0.30\) & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 1.0 \\ Y: 0 to 0.5 \\ Y: 0 to 0.5 \\ \end{tabular} & 1824 \\ & \(a=1.0\) & 0.30 \(\times\) 0.30 & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 0.5 \\ Y: 0 to 0.5 \\ \end{tabular} & 2304 \\ \hline Tall & & & & X: 0 to 1.0 & \\ columns & \(a=2.0\) & 0.30 \(\times\) 0.15 & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 0.5 \\ Y: 0 to 1.0 \\ Y: 0 to 0.5 \\ \end{tabular} & 1152 \\ & \(a=3.0\) & 0.36 \(\times\) 0.12 & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 0.5 \\ \end{tabular} & 1106 \\ & \(a=4.0\) & 0.35 \(\times\) 0.075 & 400 & \begin{tabular}{c} X: 0 to 1.0 \\ Y: 0 to 0.5 \\ \end{tabular} & 576 \\ \hline Up-scaled & \(a=0.8\) & 0.36 \(\times\) 0.45 & 500 & \begin{tabular}{c} X: 0 to 1.5 \\ Y: 0 to 1.0 \\ \end{tabular} & 5120 \\ \hline \hline \end{tabular} \end{table} Table 2: Granular column collapse simulation cases for testing GNS. \begin{table} \begin{tabular}{c c c c} \hline \hline \multirow{2}{*}{Aspect ratio, \(a\)} & \multicolumn{2}{c}{Normalized runout} & \multirow{2}{*}{Runout error (\%)} \\ \cline{2-3} & MPM & GNS & \\ \hline 0.5 & 0.831 & 0.811 & 2.48 \\ 0.8 & 1.444 & 1.445 & 0.06 \\ 1.0 & 2.071 & 2.152 & 3.78 \\ 2.0 & 3.892 & 3.682 & 5.70 \\ 3.0 & 5.620 & 5.341 & 5.23 \\ 4.0 & 5.753 & 6.070 & 5.21 \\ \hline \hline \end{tabular} \end{table} Table 3: Normalized runout from MPM and GNS depending on aspect ratios and corresponding prediction error. short column collapse. We use a normalized time (\(t/\tau_{c}\)) to compare the flow evolution, where \(t\) is physical time, and \(\tau_{c}\) is the critical time defined as the time required for the flow to fully mobilize. \(\tau_{c}\) is defined as \(\sqrt{H_{0}/g}\), where \(g\) is the gravitational acceleration. In fig. 7, the collapse shows three stages. First, the flow is mobilized by the failure of the flank and reaches full mobilization around \(t/\tau_{c}=1.0\). The majority of the runout occurs until \(t/\tau_{c}=2.5\). Beyond \(t/\tau_{c}>2.5\), the spreading decelerates due to the basal friction and finally stops at around \(t/\tau_{c}=4.0\) for both MPM and GNS rollout (prediction). As seen in fig. 7, although the GNS has only seen an aspect ratio \(a=1.0\) during training, GNS successfully captures the overall time-evolution of granular flows for a short column (\(a=0.8\)). In addition to the visual comparison of profiles, we quantitatively investigate the flow dynamics of the GNS rollout of the short column by comparing the normalized runout and height evolution with the MPM. Figure 8a shows the evolution of normalized runout and height with time. The normalized runout of the MPM (see the gray line corresponding to the left axis in fig. 8a) shows the three stages of collapse. The collapse of the granular column starts with the failure of the flank and evolves slowly until the runout is fully mobilized by \(t/\tau_{c}=1.0\). As the collapse proceeds, the runout acceleration increases (\(t/\tau_{c}=1.0\) to \(2.5\)). After this time, the runout deaccelerates due to basal friction, and finally stops at \(t/\tau_{c}\approx 4.0\). Both GNS and MPM show a constant normalized height (see the gray line corresponding to the right axis in fig. 8a) as only the flank of the column collapse, leaving a static truncated-conical core. GNS predicts an almost identical evolution of runout as the MPM simulation, which is noteworthy as only a small portion of the training trajectories (\(5\) out of \(26\)) includes the deacceleration behavior leading to the flow coming to rest due to the basal friction before hitting the walls. Overall, the quantitative comparison shown in fig. 8a confirms that the GNS can accurately model the granular flow dynamics for the short column. Figure 8b shows the energy evolutions from GNS rollout and MPM simulation. Based on the principle of energy conservation, the granular flow must satisfy \(E_{0}=E_{p}+E_{k}+E_{d}\) Figure 6: Normalized runout distance (\(\left(L_{f}-L_{0}\right)/L_{0}\)) with different aspect ratios (\(a\)). Figure 7: Evolution of flow with normalized time for GNS and MPM for the short column with \(a=0.8\). Units are in \(m\). The color represents the magnitude of the displacement. Subfigure (e) shows the final deposit at the last timestep. where \(E_{0}\) is the potential energy of the column before material points start to mobilize, \(E_{p}\) is the potential energy, \(E_{k}\) is the kinetic energy, and \(E_{d}\) is the dissipation energy due to friction along the boundary and material. In fig. 7(b), both GNS rollout and MPM simulation show identical energy evolutions. A significant fraction of the stored potential energy is converted to kinetic energy in the initial stages of the failure, reaching a peak value of kinetic energy at \(t/\tau_{c}=1\). The kinetic energy dissipates due to the basal friction and flow ceases at \(t/\tau_{c}=4.0\) when \(E_{k}\) is fully dissipated. #### 4.1.2 Tall column Tall columns exhibit different runout dynamics than the short column. GNS was only trained on granular mass with an aspect ratio of 1.0 and has not seen the dynamics of a tall column during training. As an example, we demonstrate the GNS prediction for a tall column with \(a=2.0\). Figure 9 shows the GNS rollout and MPM simulation of the runout evolution for this case. GNS rollout predicts an identical runout profile with normalized time as the MPM simulation. Similar to the short column, the tall column also shows the three stages: the initial mobilization of the flow (\(t/\tau_{c}\) to 1.0), runout (\(t/\tau_{c}=1.0\) to 2.5) along the failure surface, deacceleration (\(t/\tau_{c}=2.5\) to 4.0). In the tall column, however, a larger volume of sliding mass above the failure plane is mobilized. During the initial stages of the collapse, the granular mass experiences free fall due to gravity dominated by collisional dissipation. As the granular mass reaches the failure surface, the vertical kinetic energy is converted to horizontal acceleration, resulting in longer runout distances. GNS rollout shows similar behavior to the MPM runout simulation. Figure 9(a) shows the normalized runout and height evolution for the tall column. Although the runout evolution remains identical in the initial phase of the collapse, MPM Figure 8: (a) Normalized runout and height evolution with normalized time and (b) normalized energy evolution with normalized time for the short column \(a=0.8\). \(H_{t}\) is the height from the bottom corner of the boundary to the highest part of the column at \(t\). \(E_{p}=\sum_{i=1}^{n}m_{i}gh_{i}\) is the potential energy of the column, and \(E_{k}=\frac{1}{2}\sum_{i}^{n}m_{i}v_{i}^{2}\) is the kinetic energy of the column, where \(m_{i}\), \(h_{i}\), and \(v_{i}\) is the mass, height, and velocity of the material point \(i\), and \(n\) is the total number of material points. \(E_{d}=E_{0}-E_{p}-E_{k}\) is the dissipation energy where \(E_{0}\) is the potential energy before material points start to move. Figure 9: Evolution of flow with normalized time for GNS and MPM for the tall column with \(a=2.0\). Units are in \(m\). The color represents the magnitude of the displacement. Subfigure (e) shows the final deposit at the last timestep. shows a slightly larger normalized runout compared to the GNS. The final height in both GNS and MPM remains the same. Figure 10 presents the normalized energy evolution of the GNS rollout and the MPM simulation. During the initial stages of the collapse (\(t/\tau_{c}\) to \(1.0\)), a large amount of initial potential energy is converted to kinetic energy due to the free fall of mass under gravity. Both GNS and MPM show almost identical energy profiles. GNS shows a larger potential energy loss as the flow accelerates with an almost similar gain in kinetic energy. It indicates that GNS predicts larger frictional dissipation in tall columns, which could be from the training data focused only on short columns having higher frictional dissipation than tall columns. At the final stage, MPM shows less dissipation due to the basal boundary friction, resulting in a slightly longer runout than GNS rollout. Generally, energy dissipation behavior in GNS is consistent with MPM showing a more significant potential drop and increased dissipation energy accumulation. Overall, the GNS rollout is consistent with the MPM simulation with a runout error of \(5.7\) % for the tall column with \(a=2.0\), implying that the GNS can capture the dynamics of granular flows in collision-dominated tall columns despite only being trained on \(a=1.0\). #### 4.1.3 Upscaled domain GNS is generalizable to different initial configurations of the flow simulation owing to the strong inductive bias of the GNN(Battaglia et al., 2018). The strengths of GNS surrogate models would be to train them on small-scale experiments and then predict large-scale dynamic scenarios with complex boundary conditions. We now evaluate the scalability of GNS to a larger domain, including more material points than the training dataset. Figure 11 shows the GNS rollout of a short column \(a=0.8\) with \(5120\) material points (up to \(5\times\) more material points than the training dataset) for a larger simulation domain and longer rollout duration than the training dataset. GNS successfully predicts the flow dynamics for an upscaled domain size showing a similar runout profile with the MPM simulation. The GNS rollout predicts a normalized runout of Figure 10: (a) Normalized runout and height evolution with normalized time and (b) normalized energy evolution with normalized time for the tall column with \(a=2.0\). 1.74 while the MPM simulation shows 1.76, showing an error of 1.30%. Figure 12 shows that GNS rollout successfully replicate energy evolution observed in an upscaled domain compared to the MPM simulation. Hence, GNS can reproduce the flow dynamics even for the upscaled geometries beyond the training dataset. The primary source of GNS rollout error is not from the simulation scale but from the portion of material points that shows a large amount of displacement during column collapse. Figure 13 shows the evolution of mean squared error (MSE) of displacement over all material points (\(N\)) with time \(t\) computed as \(\frac{1}{n}\sum_{i}^{N}\left(\boldsymbol{p}_{i}^{t}-\boldsymbol{p}_{MPMi}^{t} \right)^{2}\), where \(\boldsymbol{p}_{MPMi}^{t}\) is material point position from MPM. When we compare the MSE for \(a=0.8\) with 1,824 of material points and its upscaled domain (2.22x material points), upscaling does not alter the MSE significantly. Figure 14 shows the evolution of the squared error of displacement of individual material points for the upscaled domain (\(a=0.8\)). The squared error shows larger values for those material points which run out further, i.e., the larger final displacements, but the proportion of the error with respect to the final runout is small so that GNS could simulate the upscaled case without significant error. Figure 11: Evolution of flow with normalized time for GNS and MPM for the upscaled case of short column with \(a=0.8\). Units are in \(m\). The color represents the magnitude of the displacement. Subfigure (e) shows the final deposit at the last timestep. Figure 12: Normalized runout and height evolution with normalized time and (b) normalized energy evolution with normalized time for the upscaled case of the short column with \(a=0.8\). Note that the data after \(t/\tau_{c}>5.0\) is abbreviated since the flow reaches a static state. Figure 13: Evolution of mean squared displacement error over all material points with time. ## 6 Limitations The GPU memory limits the current implementation of the GNS surrogate model. A GPU with 40 GB memory can simulate up to around 50K material points (approximately 3M edge connections). However, this shortcoming can be improved by optimizing the size of the connectivity radius \(R\). We use \(R\) of 0.030 m, which includes more interaction between neighbors. Multi-GPU GNS rollouts will enable the scalability of GNS to larger and more complex domains. ## 7 Conclusion Traditional numerical methods are computationally intensive when simulating large-scale granular flows. Statistical or conventional machine learning-based surrogate models are not generalizable since they do not explicitly consider the underlying physics. In this study, we develop a graph neural network (GNN)-based simulator (GNS) as a generalizable granular flow surrogate simulator. The use of graphs efficiently represents the physical state of interacting material points. At the same time, the message passing operation of GNN encourages the neural network to learn the interaction between material points. The expressive power of graphs and message passing that models the interaction between material points allows GNS to accurately predict granular flow dynamics for various conditions, including those not seen during training. We demonstrate the performance of GNS on granular column collapse. GNS precisely simulates different flow dynamics involved in columns for different initial aspect ratios and can also be applied to the upscaled domain with more than 2 to \(5\times\) material points with a longer simulation duration than the data provided for training. GNS also shows a remarkable speed-up of \(150\times\) computation speed compared to the parallelized CPU version of MPM. The computational efficiency and generalizability of the GNS surrogate can expedite evaluating runout hazards requiring numerous scenarios. Figure 14: Evolution of squared displacement error for each material point with normalized time in upscaled case of \(a=0.8\). The line color represents the final displacement of each material point.

2307.13203

Sensor selection for fine-grained behavior verification that respects privacy (extended version)

A useful capability is that of classifying some agent's behavior using data from a sequence, or trace, of sensor measurements. The sensor selection problem involves choosing a subset of available sensors to ensure that, when generated, observation traces will contain enough information to determine whether the agent's activities match some pattern. In generalizing prior work, this paper studies a formulation in which multiple behavioral itineraries may be supplied, with sensors selected to distinguish between behaviors. This allows one to pose fine-grained questions, e.g., to position the agent's activity on a spectrum. In addition, with multiple itineraries, one can also ask about choices of sensors where some behavior is always plausibly concealed by (or mistaken for) another. Using sensor ambiguity to limit the acquisition of knowledge is a strong privacy guarantee, a form of guarantee which some earlier work examined under formulations distinct from our inter-itinerary conflation approach. By concretely formulating privacy requirements for sensor selection, this paper connects both lines of work in a novel fashion: privacy-where there is a bound from above, and behavior verification-where sensors choices are bounded from below. We examine the worst-case computational complexity that results from both types of bounds, proving that upper bounds are more challenging under standard computational complexity assumptions. The problem is intractable in general, but we introduce an approach to solving this problem that can exploit interrelationships between constraints, and identify opportunities for optimizations. Case studies are presented to demonstrate the usefulness and scalability of our proposed solution, and to assess the impact of the optimizations.

Rishi Phatak, Dylan A. Shell

2023-07-25T02:00:07Z

http://arxiv.org/abs/2307.13203v2

# Sensor selection for fine-grained behavior verification that respects privacy (extended version) ###### Abstract A useful capability is that of classifying some agent's behavior using data from a sequence, or trace, of sensor measurements. The sensor selection problem involves choosing a subset of available sensors to ensure that, when generated, observation traces will contain enough information to determine whether the agent's activities match some pattern. In generalizing prior work, this paper studies a formulation in which multiple behavioral itineraries may be supplied, with sensors selected to distinguish between behaviors. This allows one to pose fine-grained questions, e.g., to position the agent's activity on a spectrum. In addition, with multiple itineraries, one can also ask about choices of sensors where some behavior is always plausibly concealed by (or mistaken for) another. Using sensor ambiguity to limit the acquisition of knowledge is a strong privacy guarantee, a form of guarantee which some earlier work examined under formulations distinct from our inter-itinerary conflation approach. By concretely formulating privacy requirements for sensor selection, this paper connects both lines of work in a novel fashion: privacy--where there is a bound from above, and behavior verification--where sensors choices are bounded from below. We examine the worst-case computational complexity that results from both types of bounds, proving that upper bounds are more challenging under standard computational complexity assumptions. The problem is intractable in general, but we introduce an approach to solving this problem that can exploit interrelationships between constraints, and identify opportunities for optimizations. Case studies are presented to demonstrate the usefulness and scalability of our proposed solution, and to assess the impact of the optimizations. ## I Introduction The problems of activity recognition [24], surveillance [14, 19, 23], suspicious and/or anomalous behavior detection [15], fault diagnosis [2, 17], and task monitoring [20] -- despite applying to distinct scenarios-- all involve the challenge of analyzing behavior on the basis of streams of observations from sensors. Sensor selection and activation problems (as studied by [2, 13, 21, 22]) are concerned with selecting a set of sensors to provide _sufficient_ information to reach conclusions that are both unequivocal and correct. Yet, _too much_ information may be detrimental -- for instance, in elder care and independent living applications (cf. [19]), capturing or divulging sensitive/inappropriate information could calamitous enough to be considered a showstopper. As a concrete motivating example, consider the house shown in Figure 1. Suppose that it is to be turned, via automation, into a'smart home' to serve as an assisted living space for an elderly person named Myra. Assume that occupancy sensors triggered by physical presence can be placed in each labelled, contiguous area. We might program a system that uses such sensors to track important properties related to Myra's wellness and health goals so that a carer can be notified if something is amiss. For instance, suppose that to help fond off dementia, Myra does a post-lunch crossword in her study. To determine that Myra has moved through the house and ended up in the study doing her crossword, a single occupancy sensor, STUDY, suffices. Unfortunately, when the pool has just been cleaned, the chlorine negatively affects Myra's sinuses. To ensure that she ends up in the study _and_ never visits the swimming pool, we need 2 sensors (study, pool). The increase makes intuitive sense: we are, after all, now asking for more information about the activity than before. Notice the 3 kinds of behavior that we can now discriminate between: ones that are both safe and desirable (never visiting the pool and ending in the study), ones that are safe but undesirable (never visiting the pool, but also not ending in the study), and ones that are not safe (visiting the chlorinated pool). Dinner time is next. We wish to have enough sensing power to tell that Myra has ended up the lounge/dining area, having spent some time in the kitchen. A pair of sensors (kitchen, LOUonge/DING) will do; and to include the study and pool, these are in addition to the previous 2, giving 4 in total. But alas, now Myra is annoyed: very occasionally, she enjoys a perfectly innocent midnight snack and she feels that any sensor that discloses when she has raided the fridge (and even the frequency of such forays!) is too invasive.1 She requires that we guarantee that those evenings in which her bedroom is occupied continuously shall appear identical to those in which one (or more) incursions have been made into the kitchen. Footnote 1: Her concern is not misplaced, given the increasing number of attacks on cloud services in recent years [3] from which stored data may be leaked. Fig. 1: Myra’s assistive living space wherein occupancy detectors can be employed within contiguous areas, corresponding here to eight regions: the pool, STUDY, BODGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BAND, BANDGO, BANDGO, BANDGO, BANDGO, BAND, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BAND, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BAND, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO, BANDGO Her request, along with the previous requirements, can be met with 5 sensors (lounge/duning, study, backvard, front yard, pool). Though simplistic, this example illustrates an important idea -- it is not enough to reduce the number of sensors to increase privacy, but that sometimes it may be necessary to activate a different and higher cardinality combination of sensors to protect sensitive information. The present paper re-visits the sensor selection model introduced in the IROS'21 paper of Rahmani et al. [13], advancing and elaborating upon it in order to treat the sort of problem just described. In that paper, the authors consider the setting where a claimant asserts that (future) movements within an environment will adhere to a given itinerary. Then the objective is to select, from some set of sensors at specific locations, a small subset that will detect any deviations from this claim. One of present paper's key advances is the ability to constrain the information obtained from sensors, in order to meet privacy and non-disclosure requirements. Further, the present paper generalizes the problem so that multiple itineraries are considered and, consequently, the objective becomes rather more subtle. In the prior work, the problem is to select sensors that single out the claimed itinerary from all other activity; now, when closely-related itineraries are provided, the sensors selected must have adequate resolving power distinguish fine-grain differentiations (recall the 3 kinds of behavior above). This paper establishes the computational hardness of sensor selection and optimization under this richer setting (see Section V) giving a nuanced description of its relation to the constraints introduced to modulate the collected information. Then, although the problem is worst-case intractable in general, we introduce an exact method in Section VI which treats the sensor selection problem using automata theoretic tools (an approach quite distinct from the ILP of [13]). Multiple itineraries are provided as input and their interrelationships express constraints -- we examine opportunities to exploit aspects of this structure, which leads us to propose some optimizations. The empirical results we present in Section VII show that the improvements obtained from the optimizations are significant, and demonstrate how they help improve the scalability of our proposed solution. ## II Related works So far, no single model for robotic privacy has yet emerged. A useful taxonomy dealing with privacy for robots (and associated intelligent systems) appears in [16]. Perhaps most visible candidate is that of differential privacy, used by such works as [4, 12]. There, the underlying formulation builds upon a notion of nearness (originally with a static database of multiple records), and is a less natural fit to treat the problem of altering the processes by which data are acquired. The present work tackles how privacy (of even a single entity) may be preserved without any need for addition of noise if they can exert some degree of control on the tools used to collected that data. The idea of obscuring or concealing information is another candidate and is prevalent in the control community's notion of opacity: an excellent overview for Discrete Event Systems (DES) is by Jacob, Lesage, and Faure [6]. A DES is said to be opaque if a secret has some level of indistinguishability, a concept very close to the conflation constraints we define in Section III. For further reading in the role of opacity in DES, the reader is referred to [25, 8] and [7]. Previous work by Masopust and Yin affirms that the properties of detectability and opacity are worst case intractable in general [9]. In particular, Cassez et. al. [1] showed that determining the opacity of static and dynamic masks was \(\mathsf{PSPACE}\)-Complete via formulation of so-called'state-based' and 'trace-based' opacities. In our work, importantly, simply obfuscating states is not enough, as how that particular state was reached also plays a role. A second factor which differentiates our work is that we allow specifications of constraints between two specified behaviors, instead of making them binary, one-versus-all decisions. An important subtlety, moreover, is that the conflation constraints are directed (cf., also [11]), implying that a more fine grained designation of obfuscation is allowed without necessarily running in both directions. Thus, we find it more suitable to reduce directly from the inclusion problem than universality. ## III Problem statement and definitions The environment in which some agent of interest moves is modelled as a discrete structure called the _world graph_: **Definition 1** (World Graph [13]).: A world graph is an edge-labelled, directed multigraph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\): * \(V\) is a non-empty vertex set, * \(E\) is a set of edges, * \(\mathrm{src}:E\to V\) and \(\mathrm{tgt}:E\to V\) are source and target functions, respectively, identifying a source vertex and target vertex for each edge, * \(v_{0}\in V\) is an initial vertex, * \(S=\{s_{1},s_{2},\ldots,s_{k}\}\) is a nonempty finite set of sensors, * \(\mathbb{Y}=\{Y_{s_{1}},Y_{s_{2}},\ldots,Y_{s_{k}}\}\) is a collection of mutually disjoint event sets associated to each sensor, and * \(\lambda:E\rightarrow\not\!\!S(Y_{s_{1}}\cup Y_{s_{2}}\cup\cdots\cup Y_{s_{k}})\) is a labelling function, which assigns to each edge, a world-observation a set of events. (Here \(\not\!\!S(X)\), the powerset, denotes all the subsets of \(X\).) The usefulness of the world graph is that it governs two major aspects of the agent's locomotion: how it may move, and what would happen if it moved in a certain way. The agent is known to start its movements at \(v_{0}\) and take connected edges. However, the agent cannot make any transitions that are not permitted by the world graph. Myra, for example, cannot jump from the bedroom to the lounge/duning without first going through the kitchhen. Thus, the collection of all paths that can physically be taken by the agent is defined as follows: **Definition 2** (Walks [13]).: A string \(e_{1}e_{2}\ldots e_{n}\in E^{*}\) is a walk on the world graph if and only if \(\mathrm{src}(e_{1})=v_{0}\) and for all \(i\in\{1,\ldots,n-1\}\) we have that \(\mathrm{tgt}(e_{i})=\mathrm{src}(e_{i+1})\). The set of all walks over \(\mathcal{G}\) is denoted \(\mathrm{Walks}(\mathcal{G})\) Next, we seek to understand what role the sensors play when an agent interacts with the world. Whenever an edge is crossed, it causes a'sensor response' described by the label on that edge: those sensors which are associated with the sensor values in the label (and are turned on/selected) will emit those values. Returning to the home in Figure 1, assume there are sensors in the bedroom and study which measure occupancy. Then, when Myra starts in the bedroom and moves to the study, we would obtain the event \(\{\texttt{bedroom}^{-},\texttt{stUDY}^{+}\}\) for the transition, with the plus superscript representing an event triggered by detection, and minus the inverse. The model also allows sensors other than those which detect occupancy (e.g., non-directed traversals via break beams), see [13] too. To understand the sensor values generated when crossing a single edge where sensors may be turned off, we use a sensor labelling function: **Definition 3** (Sensor labelling function).: Let \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) be a world graph, and \(M\subseteq S\) a sensor selection from it. For selection \(M\), the set of all events that could be produced by those sensors will be denoted \(\mathbf{Y}(M)=\bigcup_{s\in M}Y_{s}\). Then the _sensor labelling function_ is for all \(e\in E\): \[\lambda_{M}(e)=\begin{cases}\lambda(e)\cap\mathbf{Y}(M)&\text{if }\lambda(e) \cap\mathbf{Y}(M)\neq\varnothing,\\ \epsilon&\text{otherwise.}\end{cases}\] (Note that \(\epsilon\) here is the standard empty symbol.) Later in the paper, Figure 2 forms an example of an environment with a world graph whose edges bear appropriate sensor labels. Now, we may formally define the signature function for a walk and a given sensor set as follows: **Definition 4** (Signature of a walk [13]).: For a world graph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) we define function \(\beta_{\mathcal{G}}:\mathrm{Walks}(\mathcal{G})\times\mathcal{G}(S)\to( \mathcal{G}(\mathbf{Y}(S))\setminus\{\varnothing\})^{*}\) such that for each \(r=e_{1}e_{2}\ldots e_{n}\in\mathrm{Walks}(\mathcal{G})\) and \(M\subseteq S\), \(\beta_{\mathcal{G}}(r,M)=z_{1}z_{2}\ldots z_{n}\) in which for each \(i\in\{1,\ldots,n\}\), we have that \(z_{i}=\lambda_{M}(e_{i})\). The behavior of the agent will be specified with respect to a given world graph and these specifications will describe sequences of edges the agent may decide to take in the world graph. Following the convention of [13], each is called an itinerary. Subsequent definitions will involve the use of multiple itineraries in order to constrain what information about the agent's behavior the sensors are allowed to obtain. **Definition 5** (Itinerary DFA [13]).: An itinerary DFA over a world graph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) is a DFA \(\mathcal{I}=(Q,E,\delta,q_{0},F)\) in which \(Q\) is a finite set of states; \(E\) is the alphabet; \(\delta:Q\times E\to Q\) is the transition function; \(q_{0}\) is the initial state; and \(F\) is the set of accepting (final) states. With the basic elements given, the next four definitions formalize the different classes of constraints we desire a set of sensors to satisfy. Conflation constraints allow one type of behavior to 'appear' similar to another, while discrimination constraints specify that two behaviors must be distinguishable. **Definition 6** (Conflation constraint).: A conflation constraint on a world graph \(\mathcal{G}\) is an ordered pair of itineraries \((\mathcal{I}_{a},\mathcal{I}_{b})^{\boxdot}\). **Definition 7** (Discrimination constraint).: A discrimination constraint on a world graph \(\mathcal{G}\) is an unordered pair of itineraries \([\mathcal{I}_{1},\mathcal{I}_{2}]^{\boxdot}\). Both types will combined within a graph: **Definition 8** (Discriment designation).: A _discernment designation_ is a mixed graph \(\mathcal{D}=(I,I_{D},I_{C})\), with vertices \(I\) being a collection of itineraries, along with undirected edges \(I_{D}\) which are a set of discrimination constraints, and arcs (directed edges) \(I_{C}\) which are a set of conflation constraints. And, finally, we can state what a satisfying selection entails: **Definition 9** (Satisfying sensor selection).: Given some discernment designation \(\mathcal{D}\), a sensor set \(M\subseteq S\) is a _satisfying sensor selection for \(\mathcal{D}=(I,I_{D},I_{C})\)_ if and only if both of the following conditions hold: * For each \([\mathcal{I}_{1},\mathcal{I}_{2}]^{\boxdot}\in I_{D}\) we have that there exist no \(w_{1}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{1})\) and \(w_{2}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{2})\) where \(\beta_{\mathcal{G}}(w_{1},M)=\beta_{\mathcal{G}}(w_{2},M)\). * For each \((\mathcal{I}_{a},\mathcal{I}_{b})^{\boxdot}\in I_{C}\) we have that for every \(w\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{a})\), there exists \(c_{w}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{b})\) where \(\beta_{\mathcal{G}}(w,M)=\beta_{\mathcal{G}}(c_{w},M)\). In the above definition, the '\(\boxdot\)' constraints correspond to _discrimination_ requirements, while '\(\boxdot\)' require _conflation_. The importance of the set intersections is that the only things that can really happen are walks on the world graph. When there is a discrimination constraint, there are no walks from the one itinerary that can be confused with one from the other itinerary. When there is a conflation constraint, any walk from the first itinerary has at least one from the second that appears identical. Conflation models privacy in the following sense: any putative claim that the agent followed one itinerary can be countered by arguing, just as plausibility on the basis of the sensor readings, that it followed the other itinerary. While the discrimination constraint is symmetric, the second need not be. (Imagine: \(\{\beta_{\mathcal{G}}(w,M)|w\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}( \mathcal{I}_{1})\}=\{a,b,c,d\}\) while \(\{\beta_{\mathcal{G}}(w^{\prime},M)|w^{\prime}\in\mathrm{Walks}(\mathcal{G}) \cap\mathcal{L}(\mathcal{I}_{2})\}=\{a,b,c,d,e\}\). Then \((\mathcal{I}_{1},\mathcal{I}_{2})^{\boxdot}\) is possible, while \((\mathcal{I}_{2},\mathcal{I}_{1})^{\boxdot}\) is not.) Now, we are ready to give the central problem of the paper: **Decision Problem:**: **Minimal sensor selection to accommodate a discernment designation in itineraries (MSSADDI)** * A world graph \(\mathcal{G}\), a discernment designation \(\mathcal{D}\), and a natural number \(k\in\mathbb{N}\). * A satisfying sensor selection \(M\subseteq S\) for \(\mathcal{D}\) on \(\mathcal{G}\) with \(|M|\leq k\), or 'Infeasible' if none exist. ## IV Signature Automata To understand how we may begin solving MSSADDI and what its theoretical complexity is, we introduce the concept of a signature automaton. Signature automata are produced from the product automata of an itinerary with the world graph: **Definition 10** (Product automaton [13]).: Let \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) be a world graph and \(\mathcal{I}=(Q,E,\delta,q_{0},F)\) be an itinerary DFA. The product \(\mathcal{P}_{\mathcal{G},\mathcal{I}}\) is a partial DFA \(\mathcal{P}_{\mathcal{G},\mathcal{I}}=(Q_{\mathcal{P}},E,\delta_{\mathcal{P}}, q_{0}^{\mathcal{P}},F_{\mathcal{P}})\) with * \(Q_{\mathcal{P}}=Q\times V\), * \(\delta_{\mathcal{P}}:Q_{\mathcal{P}}\times E\to Q_{\mathcal{P}}\cup\{\bot\}\) is a function such that for each \((q,v)\in Q_{\mathcal{P}}\) and \(e\in E\), \(\delta_{\mathcal{P}}((q,v),e)\) is defined to be \(\bot\) if \(\mathrm{src}(e)\neq v\), otherwise, \(\delta_{\mathcal{P}}((q,v),e)=(\delta(q,e),\mathrm{tgt}(e))\), * \(q_{0}^{\mathcal{P}}=(q_{0},v_{0})\), and * \(F_{\mathcal{P}}=F\times V\). The language of this product automaton, as a DFA, is the collection of (finite-length) sequences from \(E\) that can be traced starting at \(q_{0}^{\mathcal{P}}\), never producing a \(\bot\), and which arrive at some element in \(F_{\mathcal{P}}\). The language recognized is the set of walks that are within the itinerary \(\mathcal{I}\), i.e., \(\mathcal{L}(\mathcal{P}_{\mathcal{G},\mathcal{I}})=\mathrm{Walks}(\mathcal{G })\cap\mathcal{L}(\mathcal{I})\). **Definition 11** (Signature automaton).: Let \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) be a world graph, let \(M\subseteq S\) be a sensor selection on it, \(\mathcal{I}=(Q,E,\delta,q_{0},F)\) be an itinerary DFA, and \(\mathcal{P}_{\mathcal{G},\mathcal{I}}\) be their product. A signature automaton \(\mathcal{S}_{\mathcal{G},\mathcal{I},M}=(Q_{\mathcal{P}},\Sigma,\delta_{ \mathcal{S}},q_{0}^{\mathcal{P}},F_{\mathcal{P}})\) is a nondeterministic finite automaton with \(\epsilon\)-moves (NFA-\(\epsilon\)) with * \(\Sigma=\{\lambda_{M}(e)\mid e\in E,\lambda_{M}(e)\neq\epsilon\}\) * \(\delta_{S}:Q_{\mathcal{P}}\times\Sigma\cup\{\epsilon\}\to\not\in Q(Q_{ \mathcal{P}})\) is a function defined for each \((q,v)\in Q_{\mathcal{P}}\) and \(\sigma\in\Sigma\cup\{\epsilon\}\) such that \[\delta_{\mathcal{S}}\big{(}(q,v),\sigma\big{)}=\Big{\{}\delta_{ \mathcal{P}}\big{(}(q,v),e\big{)}\;\Big{|}e\in E,\] \[\delta_{\mathcal{P}}\big{(}(q,v),e\big{)}\neq\perp,\lambda_{M}(e )=\sigma\Big{\}}.\] The signature automaton produces all the signatures that could result from following a path in the world graph conforming to the given itinerary. Formally, we have the following: **Lemma 1**.: _For world graph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\), sensor selection \(M\subseteq S\), and itinerary \(\mathcal{I}=(Q,E,\delta,q_{0},F)\), if their signature automaton is \(\mathcal{S}_{\mathcal{G},\mathcal{I},M}\), then:_ \[\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I},M})=\{\beta_{\mathcal{G}}(w, M)\mid w\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I})\}\,.\] Proof.: For all \(w\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I})\) there is a unique sequence of states \(q_{0}^{\mathcal{P}},q_{1}^{\mathcal{P}},\ldots,q_{n}^{\mathcal{P}}\) in \(\mathcal{P}_{\mathcal{G},\mathcal{I}}\) such that \(q_{n}^{\mathcal{P}}\in F_{\mathcal{P}}\). Following that sequence through the signature automaton returns the signature \(\beta_{\mathcal{G}}(w,M)\). Similarly, any string that is accepted by \(\mathcal{S}_{\mathcal{G},\mathcal{I},M}\) has a sequence of states \(q_{0}^{\mathcal{P}},q_{1}^{\mathcal{P}},\ldots,q_{n}^{\mathcal{P}}\) in \(\mathcal{S}_{\mathcal{G},\mathcal{I},M}\) such that \(q_{n}^{\mathcal{P}}\in F_{\mathcal{P}}\). Following those states through \(\mathcal{P}_{\mathcal{G},\mathcal{I}}\) returns the walk conforming to the itinerary which produced it. Note that the manner in which the signature automaton was produced was simply to replace the alphabet \(E\) of the product automaton with the alphabet \(\Sigma\). This introduces nondeterminism in the automaton because two outgoing edges from a vertex in the world graph may produce the same (non-empty) sensor values. Moreover, certain transitions may be made on the empty symbol if no sensor values are produced upon taking an edge in the world graph too. The usefulness of the preceding structures becomes clearer from the lemmas that follow. **Lemma 2**.: _Given world graph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) and itinerary DFAs: \(\mathcal{I}^{1}=(Q^{1},E,\delta^{1},q_{0}^{1},F^{1})\) and \(\mathcal{I}^{2}=(Q^{2},E,\delta^{2},q_{0}^{2},F^{2})\), a subset of sensors \(M\subseteq S\) is a satisfying sensor selection for constraint discrimination of itineraries \(\mathcal{I}^{1}\) and \(\mathcal{I}^{2}\) if and only if \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\cap\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})=\varnothing\)._ Proof.: Assume that \(M\) satisfies the constraint \([\mathcal{I}_{1},\mathcal{I}_{2}]^{\boxtimes}\). This implies that there exist no \(w_{1}\) and \(w_{2}\), with \(w_{1}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{1})\) and \(w_{2}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{2})\), where \(\beta_{\mathcal{G}}(w_{1},M)=\beta_{\mathcal{G}}(w_{2},M)\). The previous fact along with Lemma 1 implies \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\cap\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})=\varnothing\). The other way: if such \(w_{1}\) and \(w_{2}\) can be found, then letting \(c=\beta_{\mathcal{G}}(w_{1},M)=\beta_{\mathcal{G}}(w_{2},M)\), we have that \(\{c\}\subseteq\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\cap \mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\). Notice that if \(\mathcal{L}(\mathcal{I}_{1})\cap\mathcal{L}(\mathcal{I}_{2})\neq\varnothing\) then any walks \(w_{1}=w_{2}\) taken from this intersection must have \(\beta_{\mathcal{G}}(w_{1},M)=\beta_{\mathcal{G}}(w_{2},M)\). Any two itineraries with overlapping languages, and whose overlap falls (partly) within the set of walks, will yield a sensor selection problem that must be infeasible when these itineraries are given as a discrimination constraints. A similar lemma follows for the conflation constraints. **Lemma 3**.: _Given world graph \(\mathcal{G}=(V,E,\mathrm{src},\mathrm{tgt},v_{0},S,\mathbb{Y},\lambda)\) and itinerary DFAs: \(\mathcal{I}^{1}=(Q^{1},E,\delta^{1},q_{0}^{1},F^{1})\) and \(\mathcal{I}^{2}=(Q^{2},E,\delta^{2},q_{0}^{2},F^{2})\), a subset of sensors \(M\subseteq S\) is a satisfying sensor selection for constraint conflation of itineraries \(\mathcal{I}^{1}\) and \(\mathcal{I}^{2}\) if and only if \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\subseteq\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\)._ Proof.: Assume that \(M\) satisfies the constraint \((\mathcal{I}_{1},\mathcal{I}_{2})^{\boxplus}\). This implies that every \(w\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{1})\) has a \(c_{w}\in\mathrm{Walks}(\mathcal{G})\cap\mathcal{L}(\mathcal{I}_{2})\) with \(\beta_{\mathcal{G}}(w,M)=\beta_{\mathcal{G}}(c_{w},M)\). The previous fact along with Lemma 1 implies \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\subseteq\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\). In the opposite direction, if there exists a \(w\) for which no \(c_{w}\) can be found, we know that \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\nsubseteq\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\) since \(\beta_{\mathcal{G}}(w,M)\in\mathcal{L}(\mathcal{S}_{\mathcal{G NP-Complete (it essentially involves a single itinerary and its complement, one discrimination constraint, and zero conflation constraints), we naturally expect the problem to be NP-Hard. And this is indeed true (though the direct proof is straightforward and, hence, omitted). For the full problem, the question is whether the conflation constraints contribute additional extra complexity. The answer is in the affirmative, under standard computational complexity assumptions: **Lemma 7**.: _MSSADDI is in \(\mathsf{PSPACE}\)._ Proof.: To show that MSSADDI is in \(\mathsf{PSPACE}\), we shall show that it is in \(\mathsf{NPSPACE}\), and through Lemma 4, this implies that it is also in \(\mathsf{PSPACE}\). Given this fact, assume that we have 'guessed' a sensor selection \(M\subseteq S\) which, in polynomial space, must be verified to be a satisfying sensor selection. Thus, we must verify that each \([\mathcal{I}^{1},\mathcal{I}^{2}]^{\boxtimes}\in I_{D}\) and each \((\mathcal{I}^{1},\mathcal{I}^{2})^{\boxminus}\in I_{C}\) is satisfied by \(M\). First, to show any \([\mathcal{I}^{1},\mathcal{I}^{2}]^{\boxtimes}\in I_{D}\) can be checked in polynomial time (and thus also polynomial space): construct \(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M}\) and \(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M}\). This simply replaces the alphabet in the product, which is of size \(|V||Q|\)). Then, determining whether \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\cap\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})=\varnothing\) is simple (cf. Lemma 5). Thus, the total amount of time taken to check the discrimination constraints can be upper bounded by \(O\left(\sum_{[\mathcal{I}^{1},\mathcal{I}^{2}]^{\boxtimes}\in I_{D}}\mathsf{ poly}(|V||Q_{\mathcal{I}^{1}}|,|V||Q_{\mathcal{I}^{2}}|)\right)\) which is polynomial in the input size. Next, conflation constraints: follow a similar process to construct their signature automata \(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M}\) and \(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M}\), and ascertain whether \(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M})\subseteq\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\). By Lemma 6, we know this problem is \(\mathsf{PSPACE}\)-Complete, thus, it can be determined using only a polynomial amount of space. Hence \(\mathsf{MSSADDI}\in\mathsf{NPSPACE}\implies\mathsf{MSSADDI}\in\mathsf{PSPACE}\). Next, to show that MSSADDI is \(\mathsf{PSPACE}\)-Hard, we reduce from the NFA inclusion problem in Lemma 6. One can think of this intuitively as showing that conflation constraints, in solving the inclusion problem on signature automata, cover worst-case instances. **Lemma 8**.: _MSSADDI is \(\mathsf{PSPACE}\)-Hard_ Proof.: We reduce from NFA Inclusion, known to be \(\mathsf{PSPACE}\)-Complete (Lemma 6). Given an NFA Inclusion Problem instance \(x=\langle\mathcal{A}=(Q_{A},\Sigma,\delta_{A},q_{0}^{A},F_{A}),\mathcal{B}=( Q_{B},\Sigma,\delta_{B},q_{0}^{B},F_{B})\rangle\) we form an instance of MSSADDI \(f(x)=\langle\mathcal{G}=(V,E,\operatorname{src},\operatorname{tgt},v_{0},S, \mathbb{Y},\lambda),\mathcal{D}=(I,I_{D},I_{C}),k\rangle\). Every state of \(\mathcal{A}\) and \(\mathcal{B}\) will be assumed to reachable from their respective start states (unreachable states do not contribute to the NFA's language, and are easily trimmed). We construct \(\mathcal{G}\) as follows:-- 1. Let the vertex set be \(V=\{v_{0}\}\cup Q_{A}\cup Q_{B}\) where \(v_{0}\) is a new vertex not in either \(Q_{A}\) or \(Q_{B}\). 2. Let the edge set be \(E=\{e_{A},e_{B}\}\cup\{e_{1},e_{2},\ldots,e_{n},\)\(e_{n+1},e_{n+2},\ldots,e_{n+m}\}\). Here \(e_{A}\) is an edge that connects \(v_{0}\) to \(q_{0}^{A}\) and \(e_{B}\) is an edge connecting \(v_{0}\) to \(q_{0}^{B}\). Assuming there are \(n\) transitions in \(\mathcal{A}\) of the form \(q_{j}^{A}\in\delta_{A}(q_{i}^{A},\sigma)\), we produce an edge \(e_{k}\) for some \(1\leq k\leq n\) which connects \(q_{i}^{A}\) to \(q_{j}^{A}\) for every such \(\sigma\). Similarly, if there are \(m\) transitions in \(\mathcal{B}\) of the form \(q_{j}^{B}\in\delta_{B}(q_{i}^{B},\sigma)\), we would have an edge \(e_{n+k}\) for some \(1\leq k\leq m\) connecting \(q_{i}^{B}\) to \(q_{j}^{B}\) for each \(\sigma\). The \(\operatorname{src}\) and \(\operatorname{tgt}\) functions are defined appropriately for all edges. 3. Let sensor set \(S=\{s_{1},s_{2},\ldots,s_{|\Sigma|}\}\) where each sensor produces exactly one event so that if \(\Sigma=\{\sigma_{1},\sigma_{2},\ldots,\sigma_{|\Sigma|}\}\) then \(Y_{s_{i}}=\{\sigma_{i}\}\) and \(\mathbb{Y}=\{Y_{s_{1}},Y_{s_{2}},\ldots,Y_{s_{|\Sigma|}}\}\). 4. The edge labelling function is defined as follows. First, let \(\lambda(e_{A})=\lambda(e_{B})=\varnothing\). Then, for each transition in \(\mathcal{A}\) of the form \(q_{j}^{A}\in\delta_{A}(q_{i}^{A},\sigma)\), if \(\sigma=\epsilon\), label that edge with \(\varnothing\), otherwise label it with the singleton set \(\{\sigma\}\) for all such \(\sigma\). Follow the same procedure again for \(\mathcal{B}\). Note that, by construction, a single sensor may cover an edge from both \(\mathcal{A}\) and \(\mathcal{B}\). This is natural as the given NFAs share the alphabet \(\Sigma\). Importantly: this does not violate the assumption that sensors have pairwise distinct readings. Turning some sensor on, means we receive its readings from both regions--that constructed from \(\mathcal{A}\)_and_\(\mathcal{B}\)--or, when turned off, from neither. The following define \(\mathcal{D}\), the discernment designation:-- 1. In the world graph \(\mathcal{G}\) constructed in the previous step, let there be \(p\leq n+m\) edges collected as \(\{e_{i_{1}},e_{i_{2}},\ldots,e_{i_{p}}\}\) where we have that each of them has a non-empty label, i.e., \(e_{i_{k}}\in E\), and \(\lambda(e_{i_{k}})\neq\varnothing\) for every \(1\leq k\leq p\). Then let the set of itineraries \(I\) be \(\{I_{e_{i_{1}}},I_{e_{i_{2}}},\ldots,I_{e_{i_{p}}}\}\cup\{I_{e_{i_{1}}^{+}},I_{e_{ i_{2}}^{+}},\ldots,I_{e_{i_{p}}^{+}}\}\cup\{I_{A},I_{B}\}\), where we will give the language accepted by each DFA. The first \(2p\) elements have a language with a single string: for \(1\leq k\leq p\), to determine the languages \(\mathcal{L}(I_{e_{i_{k}}})\) and \(\mathcal{L}(I_{e_{i_{k}}^{+}})\), run a breadth first search (BFS) from \(v_{0}\) on \(\mathcal{G}\). This co-routine will return the shortest path (consisting of specific edges) from \(v_{0}\) to \(\operatorname{src}(e_{i_{k}})\). This path is the only string accepted by \(I_{e_{i_{k}}}\), and the same path but with \(e_{i_{k}}\) appended is the only string accepted by \(I_{e_{i_{k}}^{+}}\). Next, itinerary DFA \(I_{A}\) is to be defined so it accepts a string \(e_{i_{1}}e_{i_{2}}\ldots e_{i_{r}}\) where \(e_{i_{k}}\in E\) for all \(1\leq k\leq r\) if and only if \(\operatorname{tgt}(e_{i_{r}})\in F_{A}\). Similarly, define DFA \(I_{B}\) so that it accepts a string \(e_{i_{1}}^{\prime}e_{i_{2}}^{\prime}\ldots e_{i_{r}^{\prime}}\) where \(e_{i_{k}}^{\prime}\in E\) for all \(1\leq k\leq q\) if and only if \(\operatorname{tgt}(e_{i_{r}}^{\prime})\in F_{B}\). Note that we are not asking for the given NFAs \(\mathcal{A}\) and \(\mathcal{B}\) to be converted to DFAs -- instead, we are simply constructing a DFA which recognizes that some _path_ of an accepting string arrives at an accepting state in the NFA. The construction of such a DFA is simple: For \(I_{A}\), define two states \(q_{0}\) and \(q_{1}\), with only \(q_{1}\) accepting. Then, define transitions from \(q_{0}\) to \(q_{1}\) and \(q_{1}\) to \(q_{1}\) for all \(e\in E\) such that \(\operatorname{tgt}(e)\) is a final state in \(\mathcal{A}\). Similarly, define transitions from \(q_{0}\) to \(q_{0}\) and \(q_{1}\) to \(q_{0}\) for all \(e\in E\) such that \(\operatorname{tgt}(e)\) is not a final state in \(\mathcal{A}\). Doing the same for \(\mathcal{B}\) gives \(I_{B}\). 2. Define \(I_{D}=\left\{[I_{e_{i_{1}}},I_{e_{i_{1}}^{+}}]^{\boxtimes},\ldots Lastly, let \(k=|\Sigma|\). This three-piece mapping is accomplished in polynomial time since the size of the world graph is \(O(1+|\mathcal{A}|+|\mathcal{B}|)\) and the size of \(\mathcal{D}\) (i.e., the number of constraints) is \(O(|\mathcal{A}|+|\mathcal{B}|)\).2 Since BFS runs in polynomial time on \(\mathcal{G}\), all the discrimination requirements need polynomial time to construct and each is of polynomial size. In other words, for each itinerary in a discrimination constraint, its singleton language is of polynomial length (since \(1\leq q<|V|\) if \(q\) is the length of the shortest path), thus the DFA used to construct it must also be of polynomial size. For the itineraries in the conflation constraints, the DFAs have 2 states and \(|E|\) transitions, which is polynomial in the size of \(\mathcal{A}\) and \(\mathcal{B}\). Footnote 2: Here, \(|\cdot|\) gives the number of transitions or states, whichever is greater. Finally, to prove correctness: there must be a satisfying sensor selection of size at most \(k\) if and only if \(\mathcal{L}(\mathcal{A})\subseteq\mathcal{L}(\mathcal{B})\). (\(\implies\)) Assume that \(\mathcal{L}(\mathcal{A})\subseteq\mathcal{L}(\mathcal{B})\). Then the sensor selection \(M=S\) is a satisfying sensor selection because, firstly, \(|M|=|\Sigma|=k\). Secondly, note that all the discrimination constraints are satisfied because all the sensors are turned on. Lastly, the conflation constraint is also satisfied by reasoning as follows: any walk beginning at \(v_{0}\) first going to \(q_{0}^{4}\) and ending at some \(v\in F_{A}\) has a signature \(\{\sigma_{1}\}\{\sigma_{2}\}\ldots\{\sigma_{m}\}\) for which \(\sigma_{1}\sigma_{2}\ldots\sigma_{m}\in\mathcal{L}(\mathcal{A})\) which implies \(\sigma_{1}\sigma_{2}\ldots\sigma_{m}\in\mathcal{L}(\mathcal{B})\). But, by construction, one can take a path in the world graph, taking a first step from \(v_{0}\) to \(q_{0}^{B}\) without producing any sensor value, and then follow exactly the same path that is accepting in \(\mathcal{B}\) through the world graph, and this path will produce signature \(\{\sigma_{1}\}\{\sigma_{2}\}\ldots\{\sigma_{m}\}\). (\(\Longleftarrow\)) Assume there exists some satisfying sensor selection of size less than or equal to \(k=|\Sigma|\). Firstly, no sensor may be turned off since doing so would violate the discrimination constraint between the singleton itineraries involving the edge(s) labelled with the disabled sensor's value. Thus, the sensor selection has size exactly \(k\). Secondly, the conflation constraint is also met implying that, for all signatures \(\{\sigma_{1}\}\{\sigma_{2}\}\ldots\{\sigma_{m}\}\) produced by taking \(v_{0}\) to \(q_{0}^{A}\) and ending at some \(v_{i}\in F_{A}\), there exists a path from \(v_{0}\) to \(q_{0}^{B}\) ending at \(v_{j}\in F_{B}\) such that its signature is also \(\{\sigma_{1}\}\{\sigma_{2}\}\ldots\{\sigma_{m}\}\). Since no sensor is turned off, the paths that obtain the signatures in the world graph can be taken in \(\mathcal{A}\) and \(\mathcal{B}\) as well, so \(\sigma_{1}\sigma_{2}\ldots\sigma_{m}\in\mathcal{L}(\mathcal{A})\) implies \(\sigma_{1}\sigma_{2}\ldots\sigma_{m}\in\mathcal{L}(\mathcal{B})\), thus \(\mathcal{L}(\mathcal{A})\subseteq\mathcal{L}(\mathcal{B})\). **Theorem 1**.: _MSSADDI is \(\mathsf{PSPACE}\)-Complete._ Proof.: Follows from Lemmas 7 and 8. ## VI Algorithm Description Having proved the theoretical complexity class of MSSADDI, we now turn to a description of the algorithm we used to solve it. Although the algorithm is not polynomial time (as, assuming \(\mathsf{P}\neq\mathsf{PSPACE}\), it couldn't be) we introduce several optimizations to help ameliorate its running time. ### _Baseline Algorithm_ The approach we chose for solving MSSADDI was a complete enumeration of subsets, with some shortcutting. The pseudo-code, based directly on the automata theoretic connections identified in the preceding, appears in Algorithm 1. ``` Inputs: A world graph \(\mathcal{G}=(V,E,\operatorname{src},\operatorname{tgt},v_{0},S,\mathbb{Y},\lambda)\) and a discernment designation \(\mathcal{D}=(I,I_{D},I_{C})\) Output: The minimum satisfying sensor selection, if it exists, otherwise null 1:\(M^{*}\leftarrow\)null\(\triangleright\)The current best sensor set 2:for\(k=|S|\) down to \(0\)do 3:for\(M\) in Combinations\((S,k)\)do 4:for\([\mathcal{I}^{1},\mathcal{I}^{2}]^{\boxtimes}\in I_{D}\)do 5:\(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1,M}}\leftarrow\textsc{SignatureAutomaton}( \mathcal{G},\mathcal{I}^{1},M)\) 6:\(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M}\leftarrow\textsc{SignatureAutomaton}( \mathcal{G},\mathcal{I}^{2},M)\) 7:if\(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\cap\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\neq\varnothing\)then 8:Continue to next \(M\)\(\triangleright\)Check next combination 9:for\((\mathcal{I}_{1},\mathcal{I}_{2})^{\boxplus}\in I_{C}\)do 10:\(\mathcal{S}_{\mathcal{G},\mathcal{I}^{1},M}\leftarrow\textsc{SignatureAutomaton}( \mathcal{G},\mathcal{I}^{1},M)\) 11:\(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M}\leftarrow\textsc{SignatureAutomaton}( \mathcal{G},\mathcal{I}^{2},M)\) 12:if\(\mathcal{L}(\mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\not\subseteq\mathcal{L}( \mathcal{S}_{\mathcal{G},\mathcal{I}^{2},M})\)then 13:Continue to next \(M\)\(\triangleright\)Check next combination 14:if All \(I_{D}\) and \(I_{C}\) satisfiedthen 15:\(M^{*}\gets M\) 16: Continue to next \(k\)\(\triangleright\)Now try sets of size \(k-1\) 17:if No \(M\) where \(|M|=k\) satisfies all \(I_{D}\)then 18:return\(M^{*}\)\(\triangleright\)Prior solution was smallest feasible one 19:return\(M^{*}\)\(\triangleright\)Final exit ``` **Algorithm 1** Complete Enumeration for MSSADDI It is a top down search over all subsets of \(S\) where we attempt to check each constraint by constructing its signature automaton and verifying the intersection and subset properties, lines 7 and 12, respectively, as in the previous sections. Discrimination constraints are checked first (lines 4-8) because we expect them to be easier to check than conflation constraints (Lemmas 5 and 6). We take advantage of one more property of sensor sets in relation to discrimination constraints to define our baseline algorithm. Since we stipulate that different sensors produce different sensor outputs, it follows that if \(M\subseteq S\) does not satisfy a discrimination constraint, then neither can any subset of \(M\). Therefore, when no combination of sensors of size \(k\) satisfies _all_ the discrimination constraints, the search is ended, and the current best satisfying sensor set returned (line 18). Next, we propose two optimizations over the baseline algorithm just described. While each does involve a different trade-off, neither sacrifices the correctness guarantee. ### _The Caching Optimization_ Notice how the signature automaton is constructed each time an itinerary is encountered in a constraint (lines 5-6 and 10-11). This seems to be wasteful if an itinerary appears in multiple constraints (as it can be with several). The signature automaton can be cached after it is constructed should the same itinerary appear in another constraint, allowing it to be retrieved without the need for additional computation. Note, however, the trade-off being made here: while the running time reduced, the space requirements increased. Typical library implementations allow for language intersection and subset properties to be checked only on DFA's which, when converted, can result in an exponential increase in space requirements. ### _The Adaptive Weights Optimization_ The second optimization we introduce is a dynamic re-ordering of constraints. Inspired by classical methods in AI for constraint satisfaction problems (CSP's) which seek to make the current assignment _fail fast_, we devised an adaptive weighting mechanism for the desired discernment graph. Seeking to end the search as fast as possible, discrimination constraints are checked first in the hopes that if none of the sensor sets of cardinality \(k\) satisfies the discrimination constraints, then the search can be declared hopeless and ended immediately. Once a satisfying sensor set is found for the discrimination constraints, though, the following strategy is used. Whenever a particular constraint fails to be satisfied, that sensor set 'votes' the erring constraint up so that future sets know which constraint is likely to fail. Thus, after a few iterations, enough data is collected so that a sensor set checks that constraint first which most of the sets before it failed on. The idea is the more demanding (or stringent) constraints are learned and propagated upward for prioritization. ## VII Experimental results The following experiments were all performed on a computer running Windows 11 with an Intel i7 CPU having \(16\) GB RAM using Python 3. As a basic sanity check, we ran the baseline algorithm on the problems presented in Section I. For these problems, the algorithm correctly provided the optimal solutions in less than \(1\,\mathrm{s}\). Next, to test the scalability of the proposed approach and to assess the impact of the optimizations, we ran the experiments that follow. ### _Test cases_ The test cases we propose are designed such that they are parameterized: we use an \(m\times n\) grid-type world graph. An example with \(m=n=3\) is shown in Figure 2, with the scaled versions adding cells rightward and downward (without any missing edges unlike the figure). There is a sensor in each row that registers the fact that agent is present within the associated row. Similarly, a column sensor detects when the agent is within that column. Sensor set \(S\) consists of \(m+n\) sensors, one for each row and each column. The figure shows the labelled world graph, this small instance with \(18\) edges, the arcs each bearing their \(\lambda\)-based labelling. These follow a simple pattern: for example, \(r_{2}^{+}\) means that row 2's sensor has triggered, going from the unoccupied to occupied state; while \(c_{1}^{-}\) means that column \(1\)'s sensor has gone from the occupied to unoccupied. Finally, we construct an itinerary for every state in the world graph where the language accepted by the DFA for the itinerary describes following any edge in the world graph any number of times followed by an edge incoming to this state. Essentially, the itinerary DFA for that state accepts a string of edges if and only if the last edge that was taken in that walk was an incoming edge to that state. The number of constraints are proportional to the number of states in the world graph. We add \(mn\) discrimination constraints each by randomly selecting any 2 itineraries which describe ending in two states which are in a different column _and_ in a different row. Similarly, we also add \(m\) conflation constraints per column, each between 2 random itineraries that describe ending in different rows in that column. Thus, in expectation, each itinerary is in 2 discrimination constraints and 2 conflation constraints. ### _Solutions_ From the description of the problem above, it should be clear that activating either only the row sensors or only the column sensors should be a satisfying sensor selection for the discrimination constraints alone. After all, ending in a different row and column can be distinguished on the basis of either information provided by a row sensor or a column sensor. However, when considering both the discrimination and conflation constraints, only one of these options becomes feasible -- namely, that involving only activating the column occupancy sensors. Activating a row sensor could potentially violate some conflation constraints which describe ending in that row. Note that we see another detail of MSSADDI reiterated here -- that when \(n>m\), it may be necessary to activate more sensors (i.e., column sensors as opposed to only the row sensors) to satisfy the both upper and lower information bounds as opposed to the lower bounds alone. ### _Analysis_ The basic scaling plot for various grid sizes is shown in Figure 3. As can be seen in that plot, using the caching optimization alone led on average to a \(53.5\,\mathrm{\char 37}\) reduction in the running time. For our purposes, all the signature automata were able to be cached, and memory did not seem to be an issue (i.e., we never received an out-of-memory exception). Thus, time, not space, seemed to be a dominating factor in solving this problem with current resources. The results are even more impressive for the adaptive weights optimization. As compared to the baseline algorithm, it led on average to a \(87.6\,\%\) improvement in running time. When both optimizations are applied together, however, caching the signature automata seems to have little effect when adaptive weights are already in use. This makes sense because the adaptive weights allow a sensor set to be determined as unsatisfiable fast, lowering the probability that the same itinerary will be checked more than once. Seeking to understand how the mix of constraints checked changes when adaptive weights are used, we decided to analyze the time spent by the algorithm in different parts of the code for the \(6\times 5\) world graph grid. We measured the wall clock every time the algorithm started checking subsets of size \(k\) (see line 2 in Algorithm 1). Furthermore, we also kept count of the number of discrimination and conflation constraints checked for each sensor set aggregated over size \(k\) before it failed. The results, including a visualization of the constraint type, appear in the stepping chart in Figure 4. Notice, first, how the optimization leads to a greater proportion of conflation constraints being checked. For our case, conflation constraints tend to fail more often when the sensor set is of high cardinality since they are likely to include row sensors. Thus, a greater proportion (or sometimes even absolutely more) of them are checked, as compared to baseline. We see that the decision, on the basis of Lemmas 5 and 6, to place lines 4-8 before lines 4-13 may be mistaken, on average. Secondly, observe how the algorithm is able to terminate after concluding that no set of size \(k=2\) will satisfy all the discrimination constraints. The minimum satisfying sensor set in this case turned out to be \(3\) column sensors. ## VIII Conclusion and Future Works This paper tackled the sensor selection problem for multiple itineraries while also allowing for considerations of privacy. We also provided strong reasoning for why merely minimizing selected sensors does not lead to satisfaction of specific privacy requirements. We formulated this problem and proved that it was worst-case intractable. Further, we provided an algorithm (based on automata-theoretic operations) to solve the problem and considered a few optimizations over the naive implementation. In the process, we realized that the gains from those optimizations were significant owing to an inclination for wanting incorrect solutions to fail fast. In the future, research might seek a direct reduction from the problem we proposed to canonical PSPACE-Complete problems such as QSAT. Other approaches common to solving computationally hard problems such as random algorithms, and improved heuristics may also be fruitful. ### _Acknowledgements_ This material is based upon work supported in part by the National Science Foundation under grant IIS-2034097 and DoD Army Research Office under award W911NF2120064.

2308.03695

Quantifiers closed under partial polymorphisms

We study Lindstrom quantifiers that satisfy certain closure properties which are motivated by the study of polymorphisms in the context of constraint satisfaction problems (CSP). When the algebra of polymorphisms of a finite structure B satisfies certain equations, this gives rise to a natural closure condition on the class of structures that map homomorphically to B. The collection of quantifiers that satisfy closure conditions arising from a fixed set of equations are rather more general than those arising as CSP. For any such conditions P, we define a pebble game that delimits the distinguishing power of the infinitary logic with all quantifiers that are P-closed. We use the pebble game to show that the problem of deciding whether a system of linear equations is solvable in Z2 is not expressible in the infinitary logic with all quantifiers closed under a near-unanimity condition.

Anuj Dawar, Lauri Hella

2023-08-07T16:12:31Z

http://arxiv.org/abs/2308.03695v1

# Quantifiers closed under partial polymorphisms ###### Abstract We study Lindstrom quantifiers that satisfy certain closure properties which are motivated by the study of polymorphisms in the context of constraint satisfaction problems (CSP). When the algebra of polymorphisms of a finite structure \(\mathfrak{B}\) satisfies certain equations, this gives rise to a natural closure condition on the class of structures that map homomorphically to \(\mathfrak{B}\). The collection of quantifiers that satisfy closure conditions arising from a fixed set of equations are rather more general than those arising as CSP. For any such conditions \(\mathcal{P}\), we define a pebble game that delimits the distinguishing power of the infinitary logic with all quantifiers that are \(\mathcal{P}\)-closed. We use the pebble game to show that the problem of deciding whether a system of linear equations is solvable in \(\mathbb{Z}/2\mathbb{Z}\) is not expressible in the infinitary logic with all quantifiers closed under a near-unanimity condition. ## 1 Introduction Generalized quantifiers, also known as Lindstrom quantifiers, have played a significant role in the development of finite model theory. The subject of finite model theory is the expressive power of logics in the finite, and Lindstrom quantifiers provide a very general and abstract method of constructing logics. We can associate with any isomorphism-closed class of structures \(\mathcal{K}\), a quantifier \(Q_{\mathcal{K}}\) so that the extension \(L(Q_{\mathcal{K}})\) of a logic \(L\) with the quantifier \(Q_{\mathcal{K}}\) is the _minimal_ extension of \(L\) that can express the class \(\mathcal{K}\), subject to certain natural closure conditions. For this reason, comparing the expressive power of logics with Lindstrom quantifiers is closely related to comparing the descriptive complexity of the underlying classes of structures. Another reason for the significance of Lindstrom quantifiers is that we have powerful methods for proving inexpressibility in logics with such quantifiers. In particular, games, based on Hella's bijection games [14], are the basis of the most common inexpressivity results that have been obtained in finite model theory. The \(k,n\)-bijection game was introduced by Hella to characterize equivalence in the logic \(L^{k}_{\infty\omega}(\mathbf{Q}_{n})\), which is the extension of the infinitary logic with \(k\) variables by means of all \(n\)-ary Lindstrom quantifiers. A quantifier \(Q_{\mathcal{K}}\) is \(n\)-ary if the class \(\mathcal{K}\) is defined over a vocabulary \(\sigma\) in which all relation symbols have arity \(n\) or less. In particular, the \(k,1\)-bijection game, often called the \(k\)-pebble bijection game, characterizes equivalence in \(L^{k}_{\infty\omega}(\mathbf{Q}_{1})\) which has the same expressive power as \(C^{k}_{\infty\omega}\), the \(k\)-variable infinitary logic with counting. Hella uses the \(k,n\)-bijection game to show that, for each \(n\), there is an \((n+1)\)-ary quantifier that is not definable in \(L^{k}_{\infty\omega}(\mathbf{Q}_{n})\) for any \(k\). The \(k,1\)-bijection game has been widely used to establish inexpressibility results for \(C^{k}_{\infty\omega}\). The \(k,n\)-bijection game for \(n>1\) has received relatively less attention. One reason is that, while equivalence in \(C^{k}_{\infty\omega}\) is a polynomial-time decidable relation, which is in fact a relation much studied on graphs in the form of the Weisfeiler-Leman algorithm, in contrast the relation induced by the \(k,n\)-bijection game for \(n>1\) reduces to isomorphism on graphs and is intractable in general. Nonetheless, there is some interest in studying, for example, the non-trivial equivalence induced by \(L^{k}_{\infty\omega}({\bf Q}_{2})\) on structures with a ternary relation. Grochow and Levet [13] investigate this relation on finite groups. A second reason why the logics \(L^{\omega}_{\infty\omega}({\bf Q}_{n})\) have attracted less interest is that in finite model theory we are often interested in logics that are closed under vectorized first-order interpretations. This is especially so in descriptive complexity as the complexity classes we are trying to characterize usually have these closure properties. While \(L^{\omega}_{\infty\omega}({\bf Q}_{1})\) is closed under first-order interpretations, this is not the case for \(L^{\omega}_{\infty\omega}({\bf Q}_{n})\) for \(n>1\). Indeed, the closure of \(L^{\omega}_{\infty\omega}({\bf Q}_{2})\) under interpretations already includes \({\bf Q}_{n}\) for all \(n\) and so can express all properties of finite structures. So, it seems that beyond \(L^{\omega}_{\infty\omega}({\bf Q}_{1})\), interesting logics from the point of view of complexity necessarily include quantifiers of all arities. One way of getting meaningful logics that include quantifiers of unbounded arity is to consider quantifiers restricted to stronger closure conditions than just closure under isomorphisms. In recent work, novel game-based methods have established new inexpresibilty results for such logics, i.e. logics with a wide class of quantifiers of unbounded arity, but satisfying further restrictions. An important example is the class of linear-algebraic quantifiers, introduced in [6] which is the closure under interpretations of binary quantifiers invariant under invertible linear maps over finite fields. Equivalence in the resulting logic is characterized by the invertible map games introduced in [8]. These games are used in a highly sophisticated way by Lichter [17] to demonstrate a polynomial-time property that is not definable in fixed-point logic with rank introduced in [7, 12]. The result is extended to the infinitary logic with all linear-algebraic quantifiers in [5]. Another example is the recent result of Hella [15] showing a hierarchy theorem for quantifiers based on _constraint satisfaction problems_ (CSP), using a novel game. Recall that for a fixed relational structure \(\mathfrak{B}\), \(\mathsf{CSP}(\mathfrak{B})\) denotes the class of structures that map homomorphically to \(\mathfrak{B}\). Hella establishes that, for each \(n>1\), there is a structure \(\mathfrak{B}\) with \(n+1\) elements that is not definable in \(L^{\omega}_{\infty\omega}({\bf Q}_{1},\mathsf{CSP}_{n})\), where \(\mathsf{CSP}_{n}\) denotes the collection of all quantifiers of the form \(Q_{\mathsf{CSP}(\mathfrak{B}^{\prime})}\) where \(\mathfrak{B}^{\prime}\) has at most \(n\) elements. Note that \(\mathsf{CSP}_{n}\) includes quantifiers of all arities. The interest in CSP quantifiers is inspired by the great progress that has been made in classifying constraint satisfaction problems in recent years, resulting in the dichotomy theorem of Bulatov and Zhuk [3, 18]. The so-called algebraic approach to the classification of CSP has shown that the complexity of \(\mathsf{CSP}(\mathfrak{B})\) is completely determined by the algebra of polymorphisms of the structure \(\mathfrak{B}\). In particular, the complexity is completely determined by the equational theory of this algebra. As we make explicit in Section 3 below, equations satisfied by the polymorphisms of \(\mathfrak{B}\) naturally give rise to certain closure properties for the class of structures \(\mathsf{CSP}(\mathfrak{B})\), which we describe by _partial polymorphisms_. A central aim of the present paper is to initiate the study of quantifiers closed under partial polymorphisms. We present a Spoiler-Duplicator pebble game, based on bijection games, which exactly characterises the expressive power of such quantifiers. More precisely, there is such a game for any suitable family \(\mathcal{P}\) of partial polymorphisms. The exact definition of the game and the proof of the characterization are given in Section 4. As a case study, we consider the partial polymorphisms described by a _near-unanimity_ condition. It is known since the seminal work of Feder and Vardi [11] that if a structure \(\mathfrak{B}\) admits a near-unanimity polymorphism, then \(\mathsf{CSP}(\mathfrak{B})\) has _bounded width_, i.e. it (or more precisely, its complement) is definable in Datalog. On the other hand, the problem of determining the solvability of a system of equations over the two-element field \(\mathbb{Z}\,/\,2\,\mathbb{Z}\) is the classic example of a tractable CSP that is not of bounded width. Indeed, it is not even definable in \(C^{\omega}_{\infty\omega}\)[1]. We show that the collection of quantifiers that are closed under near-unanimity partial polymorphisms is much richer than the classes \(\mathsf{CSP}(\mathfrak{B})\) where \(\mathfrak{B}\) has a near-unanimity polymorphism. The collection not only includes quantifiers which are not CSP, but it also includes CSP quantifiers which are not of bounded width, including intractable ones such as hypergraph colourability. Still, we are able to show that the problem of solving systems of equations over \(\mathbb{Z}\,/\,2\,\mathbb{Z}\) is not definable in the extension of \(C^{\omega}_{\infty\omega}\) with _all_ quantifiers closed under near-unanimity partial polymorphisms. This sheds new light on the inter-definability of constraint satisfaction problems. For instance, while it follows from the arity hierarchy of [14] that the extension of \(C^{\omega}_{\infty\omega}\) with a quantifier for graph \(3\)-colourability still cannot define solvability of systems of equations over \(\mathbb{Z}\,/\,2\,\mathbb{Z}\), our result shows this also for the extension of \(C^{\omega}_{\infty\omega}\) with all hypergraph colourability quantifiers. ## 2 Preliminaries We assume basic familiarity with logic, and in particular the logics commonly used in finite model theory (see [9], for example). We write \(L^{k}_{\infty\omega}\) to denote the infinitary logic (that is, the closure of first-order logic with infinitary conjunctions and disjunctions) with \(k\) variables and \(L^{\omega}_{\infty\omega}\) for \(\bigcup_{k\in\omega}L^{k}_{\infty\omega}\). We are mainly interested in the extensions of these logics with generalized quantifiers, which we introduce in more detail in Section 2.1 below. We use Fraktur letters \(\mathfrak{A},\mathfrak{B},\ldots\) to denote structures and the corresponding Roman letters \(A,B,\ldots\) to denote their universes. Unless otherwise mentioned, all structures are assumed to be finite. We use function notation, e.g. \(f:A\to B\) to denote possibly _partial_ functions. If \(f:A\to B\) is a function and \(\vec{a}\in A^{m}\) a tuple, we write \(f(\vec{a})\) for the tuple in \(B^{m}\) obtained by applying \(f\) to \(\vec{a}\) componentwise. If \(\vec{a}_{1},\ldots,\vec{a}_{n}\) is a sequnce of \(m\)-tuples, write \((\vec{a}_{1},\ldots,\vec{a}_{n})^{T}\) for the sequence \(\vec{b}_{1},\ldots,\vec{b}_{m}\) of \(n\)-tuples, where \(\vec{b}_{i}\) is the tuple of \(i\)th components of \(\vec{a}_{1},\ldots,\vec{a}_{n}\). Given a function \(f:A^{n}\to B\), we write \(\hat{f}(\vec{a}_{1},\ldots,\vec{a}_{n})\) to denote \(f((\vec{a}_{1},\ldots,\vec{a}_{n})^{T})=(f(\vec{b}_{1}),\ldots,f(\vec{b}_{m}))\). For a pair of structures \(\mathfrak{A}\) and \(\mathfrak{B}\), a _partial isomorphism_ from \(\mathfrak{A}\) to \(\mathfrak{B}\) is a partial function \(f:A\to B\) which is an isomorphism between the substructure of \(\mathfrak{A}\) induced by the domain of \(f\) and the substructure of \(\mathfrak{B}\) induced by the image of \(f\). We write \(\operatorname{PI}(\mathfrak{A},\mathfrak{B})\) to denote the collection of all partial isomorphisms from \(\mathfrak{A}\) to \(\mathfrak{B}\). We write \(\mathbb{N}\) or \(\omega\) to denote the natural numbers, and \(\mathbb{Z}\) to denote the ring of integers. For any \(n\in\mathbb{N}\), we write \([n]\) to denote the set \(\{1,\ldots,n\}\). When mentioned without further qualification, a graph \(G=(V,E)\) is simple and undirected. That is, it is a structure with universe \(V\) and one binary relation \(E\) that is irreflexive and symmetric. The _girth_ of a graph \(G\) is the length of the shortest cycle in \(G\). ### Generalized quantifiers Let \(\sigma,\tau\) be relational vocabularies with \(\tau=\{R_{1},\ldots,R_{m}\}\), and \(\operatorname{ar}(R_{i})=r_{i}\) for each \(i\in[m]\). An interpretation \(\mathcal{I}\) of \(\tau\) in \(\sigma\) with parameters \(\vec{z}\) is a tuple of \(\sigma\)-formulas \((\psi_{1},\ldots,\psi_{m})\) along with tuples \(\vec{y}_{1},\ldots,\vec{y}_{m}\) of variables with \(|\vec{y}_{i}|=r_{i}\) for \(i\in[m]\), such that the free variables of \(\psi_{i}\) are among \(\vec{y}_{i}\vec{z}\). Such an interpretation defines a mapping that takes a \(\sigma\)-structure \(\mathfrak{A}\), along with an interpretation \(\alpha\) of the parameters \(\vec{z}\) in \(\mathfrak{A}\) to a \(\tau\)-structure \(\mathfrak{B}\) as follows. The universe of \(\mathfrak{B}\) is \(A\), and the relations \(R_{i}\in\tau\) are interpreted in \(\mathfrak{B}\) by \(R_{i}^{\mathfrak{B}}=\{\vec{b}\in A^{r_{i}}\mid(\mathfrak{A},\alpha[\vec{b}/ \vec{y}_{i}])\models\psi_{i}\}\). Let \(L\) be a logic and \(\mathcal{K}\) a class of \(\tau\)-structures. The extension \(L(Q_{\mathcal{K}})\) of \(L\) by the _generalized quantifier_ for the class \(\mathcal{K}\) is obtained by extending the syntax of \(L\) by the following formula formation rule: For \(\mathcal{I}=(\psi_{1},\ldots,\psi_{m})\) an interpretation of \(\tau\) in \(\sigma\) with parameters \(\vec{z}\), \(\psi(\vec{z})=Q_{\mathcal{K}}\vec{y}_{1},\ldots,\vec{y}_{m}\mathcal{I}\) is a formula over the signature \(\sigma\), with free variables \(\vec{z}\). The semantics of the formula is given by \((\mathfrak{A},\alpha)\models\psi(\vec{z})\), if, and only if, \(\mathfrak{B}:=\mathcal{I}(\mathfrak{A},\alpha)\) is in the class \(\mathcal{K}\). The extension \(L(\mathbf{Q})\) of \(L\) by a collection \(\mathbf{Q}\) of generalized quantifiers is defined by adding the rules above to \(L\) for each \(Q_{\mathcal{K}}\in\mathbf{Q}\) separately. The _type_ of the quantifier \(Q_{\mathcal{X}}\) is \((r_{1},\ldots,r_{m})\), and the _arity_ of \(Q_{\mathcal{X}}\) is \(\max\{r_{1},\ldots,r_{m}\}\). For the sake of simplicity, we assume in the sequel that the type of \(Q_{\mathcal{X}}\) is _uniform_, i.e., \(r_{i}=r_{j}\) for all \(i,j\in[m]\). This is no loss of generality, since any quantifier \(Q_{\mathcal{X}}\) is definably equivalent with another quantifier \(Q_{\mathcal{X}^{\prime}}\) of uniform type with the same arity. Furthermore, we restrict the syntactic rule of \(Q_{\mathcal{X}}\) by requiring that \(\vec{y}_{i}=\vec{y}_{j}\) for all \(i,j\in[m]\). Then we can denote the formula obtained by applying the rule simply by \(\varphi=Q_{\mathcal{X}}\vec{y}\,(\psi_{1},\ldots,\psi_{m})\). Note however, that this convention disallows formulas of the type \(\theta=Qx,y\,(R(x,y),R(y,x))\) in which both \(x\) and \(y\) remain free even though \(x\) is bound in \(R(x,y)\) and \(y\) is bound in \(R(y,x)\), and hence weakens the expressive power of \(\mathrm{FO}^{k}(Q_{\mathcal{X}})\) and \(L^{k}_{\infty\omega}(Q_{\mathcal{X}})\). Fortunately the loss can be compensated by using more variables (e.g., \(\theta\) is equivalent with \(Q_{\mathcal{Z}}\,(R(z,y),R(z,x))\), whence the restriction does not affect the expressive power of \(\mathrm{FO}(Q_{\mathcal{X}})\) and \(L^{\omega}_{\infty\omega}(Q_{\mathcal{X}})\). Let \(Q=Q_{\mathcal{X}}\) and \(Q^{\prime}=Q_{\mathcal{X}^{\prime}}\) be generalized quantifiers. We say that \(Q\) is _definable_ in \(L(Q^{\prime})\) if the defining class \(\mathcal{X}\) is definable in \(L(Q^{\prime})\), i.e., there is a sentence \(\varphi\) of \(L(Q^{\prime})\) such that \(\mathcal{K}=\{\mathfrak{A}\mid\mathfrak{A}\upharpoonright\varphi\}\). We write \(\mathbf{Q}_{n}\) to denote the collection of all quantifiers of arity at most \(n\). Hella [14] shows that for any \(n\), there is a quantifier of arity \(n+1\) that is not definable in \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{n})\). The logic \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{1})\) is equivalent to \(C^{\omega}_{\infty\omega}\), the infinitary logic with counting. The notion of interpretation we have defined is fairly restricted in that it does not allow for _relativization_ or _vectorizations_ (see, e.g. [9, Def. 12.3.6]. The relativizations and vectorizations of a quantifier \(Q\) can always be seen as a _collection_ of simple quantifiers of unbounded arity. ### CSP and polymorphisms Given relational structures \(\mathfrak{A}\) and \(\mathfrak{B}\) over the same vocabulary \(\tau\), a _homomorphism_\(h:\mathfrak{A}\to\mathfrak{B}\) is a function that takes elements of \(A\) to elements of \(B\) and such that for every \(R\in\tau\) of arity \(r\) and any \(\vec{a}\in A^{\tau}\), \(\vec{a}\in R^{\mathfrak{A}}\) implies \(h(\vec{a})\in R^{\mathfrak{B}}\). For a fixed structure \(\mathfrak{B}\), we write \(\mathsf{CSP}(\mathfrak{B})\) to denote the collection of structures \(\mathfrak{A}\) for which there is some homomorphism \(h:\mathfrak{A}\to\mathfrak{B}\). By the celebrated theorem of Bulatov and Zhuk, every class \(\mathsf{CSP}(\mathfrak{B})\) is either decidable in polynomial time or NP-complete. Given a \(\tau\)-structure \(\mathfrak{B}\) and \(m\in\mathbb{N}\), we define a \(\tau\)-structure \(\mathfrak{B}^{m}\). Its universe is \(B^{m}\) and if \(R\) in \(\tau\) is a relation of arity \(r\), and \(\vec{a}_{i}=(a_{i}^{1},\ldots,a_{i}^{m})\) is an \(m\)-tuple of elements of \(B\), for each \(i\in[r]\), then \((\vec{a}_{1},\ldots,\vec{a}_{r})\in R^{\mathfrak{B}^{m}}\) if, and only if, for each \(j\in[m]\), \((a_{1}^{j},\ldots,a_{r}^{j})\in R^{\mathfrak{B}}\). Then, a _polymorphism_ of \(\mathfrak{B}\) is a homomorphism \(p:\mathfrak{B}^{m}\to\mathfrak{B}\) for some \(m\). The collection of polymorphisms of \(\mathfrak{B}\) forms an algebraic _clone_ with universe \(B\). It is known that the equational theory of this algebra completely determines the computational complexity of \(\mathsf{CSP}(\mathfrak{B})\) (see [2] for an expository account). A function \(m:B^{3}\to B\) is a _majority_ function if it satisfies the equations \(m(a,a,b)=m(a,b,a)=m(b,a,a)=a\) for all \(a,b\in B\). More generally, for \(\ell\geq 3\), a function \(n:B^{\ell}\to B\) is a _near-unanimity_ function of arity \(\ell\) if for any \(\ell\)-tuple \(\vec{a}\), we have \(n(\vec{a})=a\) whenever at least \(\ell-1\) components of \(\vec{a}\) are \(a\). In particular, a near-unanimity function of arity \(3\) is a majority function. A function \(M:B^{3}\to B\) is a _Maltsev_ function if it satisfies the identities \(M(a,b,b)=M(b,b,a)=a\) for all \(a,b\in B\). For any structure \(\mathfrak{B}\) which has a near-unanimity polymorphism, the class \(\mathsf{CSP}(\mathfrak{B})\) is decidable in polynomial time, and definable in \(L^{\omega}_{\infty\omega}\). If \(\mathfrak{B}\) admits a Maltsev polymorphism, then \(\mathsf{CSP}(\mathfrak{B})\) is also decidable in polynomial time, but may not be definable in \(L^{\omega}_{\infty\omega}\) or \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{1})\), its extension with all unary quantifiers. The classic example of a CSP with a Maltsev polymorphism that is not definable in \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{1})\) is solvability of systems of equations over \(\mathbb{Z}\,/\,2\,\mathbb{Z}\) with \(\ell\) variables per equation. We can treat this as the class of structures \(\mathsf{CSP}(\mathfrak{C}_{\ell})\) where \(\mathfrak{C}_{\ell}\) is the structure with universe \(\{0,1\}\) and two \(\ell\)-ary relations \(R_{0}=\{(b_{1},\ldots,b_{\ell})\mid\sum_{i}b_{i}\equiv 0\pmod{2}\}\) and \(R_{1}=\{(b_{1},\ldots,b_{\ell})\mid\sum_{i}b_{i}\equiv 1\pmod{2}\}\). If \(\mathcal{K}=\mathsf{CSP}(\mathfrak{B})\) for some fixed structure \(\mathfrak{B}\), we call \(Q_{\mathcal{X}}\) a _CSP quantifier_. Write \(\mathsf{CSP}_{n}\) for the collection of all CSP quantifiers \(Q_{\mathcal{X}}\) where \(\mathcal{K}=\mathsf{CSP}(\mathfrak{B})\) for a structure with at most \(n\) elements. Note that \(\mathsf{CSP}_{n}\) contains quantifiers of all arities. Hella [15] defines a pebble game that characterizes equivalence of structures in the logic \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{1},\mathsf{CSP}_{n})\) and shows that there is a structure \(\mathfrak{B}\) on \(n+1\) elements such that \(\mathsf{CSP}(\mathfrak{B})\) is not definable in this logic. ## 3 Partial polymorphisms Let \(\tau\) be a relational vocabulary, and let \(\mathfrak{C}\) be a \(\tau\)-structure with a polymorphism \(p\colon\mathfrak{C}^{n}\to\mathfrak{C}\). This gives rise to a closure condition on the class \(\mathsf{CSP}(\mathfrak{C})\). In particular, suppose \(\mathfrak{B}\in\mathsf{CSP}(\mathfrak{C})\) by a homomorphism \(h:\mathfrak{B}\to\mathfrak{C}\). We can, in a sense, "close" \(\mathfrak{B}\) under the polymorphism \(p\) by including in each relation \(R^{\mathfrak{B}}\) (\(R\in\tau\)) any tuple \(\vec{a}\) for which \(h(\vec{a})=p(h(\vec{a}_{1},\ldots,\vec{a}_{n}))\) for some \(\vec{a}_{1},\ldots,\vec{a}_{n}\in R^{\mathfrak{B}}_{i}\). The resulting structure \(\mathfrak{B}^{\prime}\) is still in \(\mathsf{CSP}(\mathfrak{C})\) as is any structure \(\mathfrak{A}\) with the same universe as \(\mathfrak{B}\) and for which \(R^{\mathfrak{A}}\subseteq R^{\mathfrak{B}^{\prime}}\) for all \(R\in\tau\). Our aim is to generalize this type of closure properties from CSP quantifiers to a larger class of generalized quantifiers. To formally define this, it is useful to introduce some notation. For reasons that will become clear, we use _partial_ functions \(p\). **Definition 1**: _Let \(A\neq\emptyset\) be a set, and let \(p\) be a be a partial function \(A^{n}\to A\)._ _(a) If \(R\subseteq A^{r}\), then \(p(R):=\{\hat{p}(\vec{a}_{1},\ldots,\vec{a}_{n})\mid\vec{a}_{1},\ldots,\vec{a} _{n}\in R\}\)._ _(b) If \(\mathfrak{A}=(A,R^{\mathfrak{A}}_{1},\ldots,R^{\mathfrak{A}}_{m})\), then we denote the structure \((A,p(R^{\mathfrak{A}}_{1}),\ldots,p(R^{\mathfrak{A}}_{m}))\) by \(p(\mathfrak{A})\)._ We say that \(p\) is a _partial polymorphism_ of a \(\tau\)-structure \(\mathfrak{A}\) with domain \(A\) if for every \(R\in\tau\), the relation \(R^{\mathfrak{A}}\) is closed with respect to \(p\), i.e., \(p(R^{\mathfrak{A}})\subseteq R^{\mathfrak{A}}\). The reason for considering partial functions is that we are usually interested in polymorphisms that satisfy certain equations. The equations specify the polymorphism partially, but not totally. We can uniformly specify closure properties on our class of structures for all polymorphisms satisfying the equations by only requiring closure for the common partial function. This is illustrated in the examples below. By a _family of partial functions_ we mean a class \(\mathcal{P}\) that contains a partial function \(p_{A}\colon A^{n}\to A\) for every finite set \(A\), where \(n\) is a fixed positive integer. We give next some important examples of families of partial functions that arise naturally from well-known classes of polymorphisms. **Example 2**: _(a) The Maltsev family \(\mathcal{M}\) consists of the partial functions \(M_{A}\colon A^{3}\to A\) such that \(M_{A}(a,b,b)=M_{A}(b,b,a)=a\) for all \(a,b\in A\), and \(M_{A}(a,b,c)\) is undefined unless \(a=b\) or \(b=c\). If \(\mathfrak{A}\) has a Maltsev polymorphism \(p\colon A^{3}\to A\), then clearly \(M_{A}\) is a restriction of \(p\), whence it is a partial polymorphism of \(\mathfrak{A}\)._ _(b) The family \(\mathcal{M}\) of ternary partial majority functions consists of the partial functions \(m_{A}\colon A^{3}\to A\) such that \(m_{A}(a,a,b)=m_{A}(a,b,a)=m_{A}(b,a,a)=a\) for all \(a,b\in A\), and \(m_{A}(a,b,c)\) is undefined if \(a,b\) and \(c\) are all distinct. If \(\mathfrak{A}\) has a majority polymorphism, then \(m_{A}\) is a restriction of it, whence it is a partial polymorphism of \(\mathfrak{A}\)._ _(c) More generally, for each \(\ell\geq 3\) we define the family \(\mathcal{N}_{\ell}\) of \(\ell\)-ary partial near-unanimity functions \(n^{\ell}_{A}\colon A^{\ell}\to A\) as follows:_ * \(n^{\ell}_{A}(a_{1},\ldots,a_{\ell})=a\) _if and only if_ \(|\{i\in[n]\mid a_{i}=a\}|\geq\ell-1\)_._ _In particular, \(\mathcal{M}\mathcal{G}=\mathcal{N}_{3}\)._ We next give a formal definition for the closure property of generalized quantifiers that arises from a family of partial functions. In the definition we use the notation \(\mathfrak{A}\leq\mathfrak{B}\) if \(\mathfrak{A}\) and \(\mathfrak{B}\) are \(\tau\)-structures such that \(A=B\) and \(R^{\mathfrak{A}}\subseteq R^{\mathfrak{B}}\) for each \(R\in\tau\). Furthermore, we define the union \(\mathfrak{A}\cup\mathfrak{B}\) of \(\mathfrak{A}\) and \(\mathfrak{B}\) to be the \(\tau\)-structure \(\mathfrak{C}\) such that \(C=A\cup B\) and \(R^{\mathfrak{C}}=R^{\mathfrak{A}}\cup R^{\mathfrak{B}}\) for each \(R\in\tau\). **Definition 3**: _Let \(\mathcal{P}\) be a family of \(n\)-ary partial functions, and let \(Q_{\mathcal{X}}\) be a generalized quantifier of vocabulary \(\tau\). We say that \(Q_{\mathcal{X}}\) is \(\mathcal{P}\)-closed if the following holds for all \(\tau\)-tructures \(\mathfrak{A}\) and \(\mathfrak{B}\) with \(A=B\):_ * _if_ \(\mathfrak{B}\in\mathcal{K}\) _and_ \(\mathfrak{A}\leq p_{A}(\mathfrak{B})\cup\mathfrak{B}\)_, then_ \(\mathfrak{A}\in\mathcal{K}\)_._ _We denote the class of all \(\mathcal{P}\)-closed quantifiers by \(\mathbf{Q}_{\mathcal{P}}\)._ Note that the condition \(\mathfrak{A}\leq p_{A}(\mathfrak{B})\cup\mathfrak{B}\) holds if and only if for every \(R\in\tau\) and every \(\vec{a}\in R^{\mathfrak{A}}\setminus R^{\mathfrak{B}}\) there are tuples \(\vec{a}_{1},\ldots,\vec{a}_{n}\in R^{\mathfrak{B}}\) such that \(\vec{a}=\widehat{p_{A}}(\vec{a}_{1},\ldots,\vec{a}_{n})\). The quantifier \(Q_{\mathcal{K}}\) is _downwards monotone_, if \(\mathfrak{A}\leq\mathfrak{B}\) and \(\mathfrak{B}\in\mathcal{K}\) implies \(\mathfrak{A}\in\mathcal{K}\). It follows directly from Definition 3 that all \(\mathcal{P}\)-closed quantifiers are downwards monotone. **Proposition 4**: _If \(Q_{\mathcal{K}}\in\mathbf{Q}_{\mathcal{P}}\) for some family \(\mathcal{P}\), then \(Q_{\mathcal{K}}\) is downwards monotone._ It is easy to see that, for any family \(\mathcal{P}\), the first-order quantifiers can be defined from a \(\mathcal{P}\)-closed quantifier using only negation. **Proposition 5**: _Let \(\mathcal{K}_{0}\) be the class of all \(\{P\}\)-structures \(\mathfrak{A}\) such that \(P^{\mathfrak{A}}=\emptyset\). Then \(Q_{\mathcal{K}_{0}}\in\mathbf{Q}_{\mathcal{P}}\) for any family \(\mathcal{P}\) of partial functions._ _Proof._ If \(\mathfrak{B}\in\mathcal{K}_{0}\), then \(P^{\mathfrak{B}}=\emptyset\), whence \(p_{B}(\mathfrak{B})=\emptyset\). Thus, if \(\mathfrak{A}\leq p_{B}(\mathfrak{B})\cup\mathfrak{B}\), then \(P^{\mathfrak{A}}=\emptyset\), and hence \(\mathfrak{A}\in\mathcal{K}_{0}\). \(\Box\) Note that in the case \(\operatorname{ar}(P)=1\), the quantifier \(Q_{\mathcal{K}_{0}}\) of the proposition above is the negation of the existential quantifier: \(\mathfrak{A}\models Q_{\mathcal{K}_{0}}x\,\varphi\iff\mathfrak{A}\models\neg \exists x\,\varphi\). Up to now we have not imposed any restrictions on the family \(\mathcal{P}\). It is natural to require that the partial functions in \(\mathcal{P}\) are uniformly defined, or at least that \((A,p_{A})\) and \((B,p_{B})\) are isomorphic if \(|A|=|B|\). Such requirements are captured by the notions defined below. **Definition 6**: _Let \(\mathcal{P}\) be a family of \(n\)-ary partial functions._ _(a) \(\mathcal{P}\) is invariant if it respects bijections: if \(f\colon A\to B\) is a bijection and \(a_{1},\ldots,a_{n}\in A\), then \(p_{B}(f(a_{1}),\ldots,f(a_{n}))\simeq f(p_{A}(a_{1},\ldots,a_{n}))\). Here the symbol \(\simeq\) says that either both sides are defined and have the same value, or both sides are undefined._ _(b) \(\mathcal{P}\) is strongly invariant if it respects injections: if \(f\colon A\to B\) is an injection and \(a_{1},\ldots,a_{n}\in A\), then \(p_{B}(f(a_{1}),\ldots,f(a_{n}))\simeq f(p_{A}(a_{1},\ldots,a_{n}))\)._ _(c) \(\mathcal{P}\) is projective, if it strongly invariant and it is preserved by all functions: if \(f\colon A\to B\) is a function and \(a_{1},\ldots,a_{n}\in A\) are such that \(p_{A}(a_{1},\ldots,a_{n})\) is defined, then \(p_{B}(f(a_{1}),\ldots,f(a_{n}))=f(p_{A}(a_{1},\ldots,a_{n}))\)._ It is easy to verify that \(\mathcal{P}\) is invariant if, and only if, it is determined by equality types on each cardinality: there are quantifier free formulas in the language of equality \(\theta_{\mathcal{P}}^{m}(\vec{x},y)\) such that if \(|A|=m\), then \(p_{A}(\vec{a})=b\iff A\models\theta_{\mathcal{P}}^{m}[\vec{a}/\vec{x},b/y]\) holds for all \(\vec{a}\in A^{n}\) and \(b\in A\). Similarly, \(\mathcal{P}\) is strongly invariant if, and only if, the same holds with a single formula \(\theta_{\mathcal{P}}=\theta_{\mathcal{P}}^{m}\) for all \(m\in\omega\). Note that if the family \(\mathcal{P}\) is strongly invariant, then for every finite set \(A\), \(p_{A}\) is a _partial choice function_, i.e., \(p_{A}(a_{1},\ldots,a_{n})\in\{a_{1},\ldots,a_{n}\}\). Indeed, if \(b:=p_{A}(a_{1},\ldots,a_{n})\not\in\{a_{1},\ldots,a_{n}\}\) and \(B=A\cup\{c\}\), where \(c\notin A\), then using the identity function \(f=\operatorname{id}_{A}\) of \(A\) in the condition \(p_{B}(f(a_{1}),\ldots,f(a_{n}))=f(p_{A}(a_{1},\ldots,a_{n}))\), we get \(p_{B}(a_{1},\ldots,a_{n})=b\). On the other hand, using the injection \(f^{\prime}\colon A\to B\) that agrees with \(\operatorname{id}_{A}\) on \(A\setminus\{b\}\) but maps \(b\) to \(c\), we get the contradiction \(p_{B}(a_{1},\ldots,a_{n})=c\neq b\). **Remark 7**: _An invariant family may contain functions \(p_{A}\) that are not partial choice functions: for example the family consisting of all functions \(p_{A}\colon A^{n}\to A\) such that \(p_{A}(a_{1},\ldots,a_{n})=a_{n+1}\iff A\setminus\{a_{1},\ldots,a_{n}\}=\{a_{n +1}\}\) is invariant. However, if \(|A|>n+1\), then \(p_{A}\) is necessarily a partial choice function._ **Lemma 8**: _Let \(\mathcal{P}\) be a family of \(n\)-ary partial choice functions. Then \(Q_{\mathcal{K}}\in\mathbf{Q}_{\mathcal{P}}\) for any unary downwards monotone quantifier \(Q_{\mathcal{K}}\). In particular this holds if \(\mathcal{P}\) is strongly invariant._ _Proof._ Let \(\tau\) be the vocabulary of \(\mathcal{K}\), and assume that \(\mathfrak{B}\in\mathcal{K}\) and \(\mathfrak{A}\leq p_{A}(\mathfrak{B})\cup\mathfrak{B}\). Then for all \(R\in\tau\) and \(a\in R^{\mathfrak{A}}\setminus R^{\mathfrak{B}}\) there are \(a_{1},\ldots,a_{n}\in A\) such that \(p_{A}(a_{1},\ldots,a_{n})=a\) and \(a_{i}\in R^{\mathfrak{B}}\) for each \(i\in[n]\). Since \(p_{A}\) is a choice function, we have \(a\in\{a_{1},\ldots,a_{n}\}\), and hence \(a\in R^{\mathfrak{B}}\). Thus we see that \(\mathfrak{A}\leq\mathfrak{B}\), and consequently \(\mathfrak{A}\in\mathcal{K}\), since \(Q_{\mathcal{K}}\) is downwards monotone. \(\Box\) It is easy to see that the families \(\mathcal{M}\) and \(\mathcal{N}_{\ell}\), \(\ell\geq 3\), introduced in Example 2, are strongly invariant. Indeed, the defining formulas \(\theta_{\mathcal{M}}\) and \(\theta_{\mathcal{N}_{\ell}}\) are easily obtained from the identities that define these conditions. Thus, all unary downwards monotone quantifiers are \(\mathcal{M}\)-closed and \(\mathcal{N}_{\ell}\)-closed. For the families \(\mathcal{N}_{\ell}\) we can prove a much stronger result: **Lemma 9**: _Let \(\ell\geq 3\), and let \(Q_{\mathcal{K}}\) be a downwards monotone quantifier of arity \(r<\ell\). Then \(Q_{\mathcal{K}}\in\mathbf{Q}_{\mathcal{N}_{\ell}}\)._ _Proof._ Let \(\tau\) be the vocabulary of \(\mathcal{K}\), and assume that \(\mathfrak{B}\in\mathcal{K}\) and \(\mathfrak{A}\leq n^{\ell}_{A}(\mathfrak{B})\cup\mathfrak{B}\). Then for all \(R\in\tau\) and \(\vec{a}=(a_{1},\ldots,a_{r})\in R^{\mathfrak{A}}\setminus R^{\mathfrak{B}}\) there are \(\vec{a}_{i}=(a_{i}^{1},\ldots,a_{i}^{r})\in R^{\mathfrak{B}}\), \(i\in[\ell]\), such that \(\widehat{n^{\ell}_{A}}(\vec{a}_{1},\ldots,\vec{a}_{\ell})=\vec{a}\). Thus, for each \(j\in[r]\) there is at most one \(i\in[\ell]\) such that \(a_{j}^{j}\neq a_{j}\), and hence there is at least one \(i\in[\ell]\) such that \(\vec{a}=\vec{a}_{i}\). This shows that \(\mathfrak{A}\leq\mathfrak{B}\), and since \(Q_{\mathcal{K}}\) is downwards monotone, we conclude that \(\mathfrak{A}\in\mathcal{K}\). \(\Box\) Using a technique originally due to Imhof for (upwards) monotone quantifiers (see [16]), we can show that any quantifier \(Q_{\mathcal{K}}\) is definable by a downwards monotone quantifier of the same arity. Indeed, if the vocabulary of \(\mathcal{K}\) is \(\tau=\{R_{1},\ldots,R_{m}\}\), where \(\mathrm{ar}(R_{i})=r\) for all \(i\in[m]\), we let \(\tau^{\prime}:=\{S_{1},\ldots,S_{m}\}\) be a disjoint copy of \(\tau\), and \(\tau^{*}:=\tau\cup\tau^{\prime}\). Furthermore, we let \(\mathcal{K}^{*}\) be the class of all \(\tau^{*}\)-structures \(\mathfrak{A}\) such that \(R^{\mathfrak{A}}_{i}\cap S^{\mathfrak{A}}_{i}=\emptyset\) for all \(i\in[m]\), and \((A,R^{\mathfrak{A}}_{1},\ldots,R^{\mathfrak{A}}_{m})\in\mathcal{K}\) or \(R^{\mathfrak{A}}_{i}\cup S^{\mathfrak{A}}_{i}\neq A^{r}\) for some \(i\in[m]\). Then \(Q_{\mathcal{K}^{*}}\) is downwards monotone, and clearly \(Q_{\mathcal{K}}\vec{x}\,(\psi_{1},\ldots,\psi_{m})\) is equivalent with \(Q_{\mathcal{K}^{*}}\vec{x}\,(\psi_{1},\ldots,\psi_{m},-\psi_{1},\ldots,-\psi_{ m})\). Using this observation, we get the following corollary to Lemmas 8 and 9. **Corollary 10**: _(a) Let \(\mathcal{P}\) be as in Lemma 8. Then \(L^{k}_{\infty\omega}(\mathbf{Q}_{\mathcal{P}}\cup\mathbf{Q}_{1})\leq L^{k}_{ \infty\omega}(\mathbf{Q}_{\mathcal{P}})\)._ _(b) \(L^{k}_{\infty\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}}\cup\mathbf{Q}_{\ell-1}) \leq L^{k}_{\infty\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\)._ As explained in the beginning of this section, the definition of \(\mathcal{P}\)-closed quantifiers was inspired by the closure property of a CSP quantifier \(Q_{\mathsf{CSP}(\mathfrak{E})}\) that arises from a polymorphism of \(\mathfrak{C}\). Thus, it is natural to look for sufficient conditions on the family \(\mathcal{P}\) and the target structure \(\mathfrak{C}\) for \(Q_{\mathsf{CSP}(\mathfrak{E})}\) to be \(\mathcal{P}\)-closed. It turns out that the notions of projectivity and partial polymorphism lead to such a condition. **Proposition 11**: _Let \(\mathcal{P}\) be a projective family of \(n\)-ary partial functions, and let \(\mathfrak{C}\) be a \(\tau\)-structure. If \(p_{C}\) is a partial polymorphism of \(\mathfrak{C}\), then \(Q_{\mathsf{CSP}(\mathfrak{E})}\in\mathbf{Q}_{\mathcal{P}}\)._ _Proof._ Assume that \(\mathfrak{B}\in\mathcal{K}\) and \(\mathfrak{A}\leq p_{A}(\mathfrak{B})\cup\mathfrak{B}\). Then \(A=B\) and there is a homomorphism \(h\colon\mathfrak{B}\to\mathfrak{C}\). We show that \(h\) is a homomorphism \(\mathfrak{A}\to\mathfrak{C}\), and hence \(\mathfrak{A}\in\mathcal{K}\). Thus let \(R\in\tau\), and let \(\vec{a}\in R^{\mathfrak{A}}\). If \(\vec{a}\in R^{\mathfrak{B}}\), then \(h(\vec{a})\in R^{\mathfrak{C}}\) by assumption. On the other hand, if \(\vec{a}\in R^{\mathfrak{A}}\setminus R^{\mathfrak{B}}\), then there exist tuples \(\vec{a}_{1},\ldots,\vec{a}_{n}\in R^{\mathfrak{B}}\) such that \(\vec{a}=\widehat{p_{A}}(\vec{a}_{1},\ldots,\vec{a}_{n})\). Since \(h\) is a homomorphism \(\mathfrak{B}\to\mathfrak{C}\), we have \(h(\vec{a}_{i})\in R^{\mathfrak{C}}\) for each \(i\in[n]\). Since \(p_{C}\) is a partial polymorphism of \(\mathfrak{C}\), we have \(\widehat{p_{C}}(h(\vec{a}_{1}),\ldots,h(\vec{a}_{n}))\in R^{\mathfrak{C}}\). Finally, since \(\mathcal{P}\) is projective, we have \(h(\vec{a})=h(\widehat{p_{A}}(\vec{a}_{1},\ldots,\vec{a}_{n}))=\widehat{p_{C}}( h(\vec{a}_{1}),\ldots,h(\vec{a}_{n}))\), and hence \(h(\vec{a})\in R^{\mathfrak{C}}\). \(\Box\) We can now apply Proposition 11 to the families introduced in Example 2. **Example 12**: _(a) Consider a constraint satisfaction problem \(\mathsf{CSP}(\mathfrak{C})\) such that \(\mathfrak{C}\) has a Maltsev polymorphism \(p\colon\mathfrak{C}^{3}\to\mathfrak{C}\). We show that \(Q_{\mathsf{CSP}(\mathfrak{E})}\in\mathbf{Q}_{\mathcal{M}}\). As pointed out in Example 2, \(M_{C}\) is a partial polymorphism of \(\mathfrak{C}\). Thus, by Proposition 11 it suffices to show that the Maltsev family \(\mathcal{M}\) is projective._ _Thus, assume that \(f\colon A\to B\) is a function, and \(M_{A}(a,b,c)\) is defined. Then \(a=b\) and \(M_{A}(a,b,c)=c\), or \(b=c\) and \(M_{A}(a,b,c)=a\). In the former case we have \(f(a)=f(b)\), whence \(M_{B}(f(a),f(b),f(c))=f(c)=f(M_{A}(a,b,c))\). In the latter case we have \(f(b)=f(c)\), whence \(M_{B}(f(a),f(b),f(c))=f(a)=f(M_{A}(a,b,c))\)._ _(b) The \(n\)-regular hypergraph \(m\)-colouring problem is \(\mathsf{CSP}(\mathfrak{H}_{n,m})\), where \(\mathfrak{H}_{n,m}=([m],R_{n,m})\) is the complete \(n\)-regular hypergraph with \(m\) vertices, i.e.,_ * \(R_{n,m}:=\{(v_{1},\ldots,v_{n})\in[m]^{n}\mid v_{i}\neq v_{j}\text{ for all }1\leq i<j\leq m\}\)_._ _We show that \(Q_{\mathsf{CSP}(\mathfrak{H}_{n,m})}\in\mathbf{Q}_{\mathcal{M}\mathcal{G}}\) for all \(n\geq 2\) and \(m\geq n\). By Proposition 11 it suffices to show that \(m_{[m]}\) is a partial polymorphism of \(\mathfrak{H}_{n,m}\), and the family \(\mathcal{M}\mathcal{G}\) is projective._ _To see that \(m_{[m]}\) is a partial polymorphism of \(\mathfrak{H}_{n,m}\), assume that \(\vec{a}_{i}=(a_{i}^{1},\ldots,a_{n}^{n})\in R_{n,m}\) for \(i\in[3]\), and \(\vec{a}=(a_{1},\ldots,a_{n})=m_{[m]}(\vec{a}_{1},\vec{a}_{2},\vec{a}_{3})\). By the definition of \(m_{[m]}\), for each \(j\in[n]\) we have \(|\{i\in[3]\mid a_{i}^{j}=a_{j}\}|\geq 2\). Thus for any two distinct \(j,k\in[n]\), there is \(i\in[3]\) such that \(a_{j}=a_{i}^{j}\) and \(a_{i}^{k}=a_{k}\), whence \(a_{j}\neq a_{k}\). Thus we have \(\vec{a}\in R_{n,m}\)._ _To show that \(\mathcal{M}\mathcal{G}\) is projective, assume that \(f\colon A\to B\) is a function, and \(m_{A}(a,b,c)\) is defined. Then \(a=b=m_{A}(a,b,c)\), \(a=c=m_{A}(a,b,c)\) or \(b=c=m_{A}(a,b,c)\). In the first case we have \(f(m_{A}(a,b,c))=f(a)=f(b)=m_{B}(f(a),f(b),f(c))\), as desired. The two other cases are similar._ _(c) In the same way we can show that the family \(\mathcal{N}_{\ell}\) of partial near-unanimity polymorphisms is projective for any \(\ell\geq 3\). We relax now the notion of hypergraph coloring as follows: Let \(\mathfrak{H}=(H,R)\), where \(R\subseteq H^{[n]}\), be a hypergraph and let \(k<n\). A \(k\)-weak \(m\)-coloring of \(\mathfrak{H}\) is a function \(f\colon H\to[m]\) such that for all \(e\in R\) and all \(i\in[m]\), \(|e\cap f^{-1}[\{i\}]|\leq k\). Observe now that there exists a \(k\)-weak \(m\)-coloring of \(\mathfrak{H}\) if and only if \(\mathfrak{H}\in\mathsf{CSP}(\mathfrak{H}_{n,m}^{k})\), where \(\mathfrak{H}_{n,m}^{k}=([m],R_{n,m}^{k})\) is the structure such that_ * \(R_{n,m}^{k}:=\{(v_{1},\ldots,v_{n})\in[m]^{n}\mid|\{v_{i}\mid i\in I\}|\geq 2 \text{ for all }I\subseteq[n]\text{ with }|I|=k+1\}\)_._ _Note that \(\mathfrak{H}_{n,m}^{1}=\mathfrak{H}_{n,m}\), whence \(m_{[m]}=n_{[m]}^{3}\) is a partial polymorphism of \(\mathfrak{H}_{n,m}^{1}\). It is straightforward to generalize this to \(\ell>3\): \(n_{[m]}^{\ell}\) is a partial polymorhism of \(\mathfrak{H}_{n,m}^{\ell-2}\). Thus by Proposition 11, the \(\mathrm{CSP}\) quantifier \(Q_{\mathsf{CSP}(\mathfrak{H}_{n,m}^{\ell-2})}\) is \(\mathbf{Q}_{\mathcal{N}_{\ell}}\)-closed._ **Remark 13**: _As shown in Example 12(b), the partial majority function \(m_{[m]}\) is a partial polymorphism of the structure \(\mathfrak{H}_{n,m}\). However, there does not exist any polymorphism \(p\colon[m]^{3}\to[m]\) that extends \(m_{[m]}\). This can be verified directly, but it also follows from the fact that \(\mathsf{CSP}(\mathfrak{C})\) is of bounded width for any \(\mathfrak{C}\) that has a majority polymorphism ([11]), but \(\mathsf{CSP}(\mathfrak{H}_{n,m})\) is not of bounded width. The same holds for the partial functions \(n_{[m]}^{\ell}\) and the structures \(\mathfrak{H}_{n,m}^{k}\) in Example 12(c)._ ## 4 Pebble game for \(\mathcal{P}\)-closed quantifiers In this section we introduce a pebble game that characterizes equivalence of structures with respect to \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{\mathcal{P}})\), the extension of the infinitary \(k\)-variable logic \(L^{\omega}_{\infty\omega}\) by the class of all \(\mathcal{P}\)-closed quantifiers. We fix a family \(\mathcal{P}\) of \(n\)-ary partial functions for the rest of the section. Given two structures \(\mathfrak{A}\) and \(\mathfrak{B}\) of the same vocabulary, and assignments \(\alpha\) and \(\beta\) on \(\mathfrak{A}\) and \(\mathfrak{B}\), respectively, such that \(\mathrm{dom}(\alpha)=\mathrm{dom}(\beta)\), we write \((\mathfrak{A},\alpha)\equiv_{\infty\omega,\mathcal{P}}^{k}(\mathfrak{B},\beta)\) if the equivalence \[(\mathfrak{A},\alpha)\models\varphi\iff(\mathfrak{B},\beta)\models\varphi\] holds for all formulas \(\varphi\in L^{k}_{\infty\omega}(\mathbf{Q}_{\mathcal{P}})\) with free variables in \(\mathrm{dom}(\alpha)\). If \(\alpha=\beta=\emptyset\), we write simply \(\mathfrak{A}\equiv_{\infty\omega,\mathcal{P}}^{k}\mathfrak{B}\) instead of \((\mathfrak{A},\emptyset)\equiv_{\infty\omega,\mathcal{P}}^{k}(\mathfrak{B},\emptyset)\). The basic idea of our pebble game for a pair \((\mathfrak{A},\mathfrak{B})\) of structures is the following. In each round Duplicator gives a bijection \(f\colon A\to B\), just like in the bijection games of [14], but instead of using \(\vec{b}=f(\vec{a})\) as answer for Spoiler's move \(\vec{a}\in A^{r}\), she is allowed to give a sequence \(\vec{b}_{1},\ldots,\vec{b}_{n}\in B^{r}\) of alternative answers as long as \(\vec{b}=\widehat{p_{B}}(\vec{b}_{1},\ldots,\vec{b}_{n})\). Spoiler completes the round by choosing one of these alternatives \(\vec{b}_{i}\). Spoiler wins if \(\vec{a}\mapsto\vec{b}_{i}\) is not a partial isomorphism; otherwise the game carries on from the new position. Observe now that if Duplicator has a winning strategy for the first round of the game, then \(f(\mathfrak{A})\leq p_{B}(\mathfrak{B})\cup\mathfrak{B}\). Indeed, if Spoiler chooses a tuple \(\vec{a}\in R^{\mathfrak{A}}\), then Duplicator has to answer by either the tuple \(f(\vec{a})\), or a sequence \(\vec{b}_{1},\ldots,\vec{b}_{n}\in B^{r}\) of tuples such that \(f(\vec{a})=\widehat{p_{B}}(\vec{b}_{1},\ldots,\vec{b}_{n})\); in the first case she loses if \(f(\vec{a})\not\in R^{\mathfrak{B}}\), and in the second case she loses if \(\vec{b}_{i}\not\in R^{\mathfrak{B}}\) for some \(i\in[n]\). Thus if Duplicator has a winning strategy in the one round game and \(\mathfrak{B}\in\mathcal{K}\) for some \(\mathcal{P}\)-closed quantifier \(Q_{\mathcal{K}}\), then \(f(\mathfrak{A})\in\mathcal{K}\), and since \(f\) is an isomorphism \(\mathfrak{A}\to f(\mathfrak{A})\), also \(\mathfrak{A}\in\mathcal{K}\). In other words, if \(\mathfrak{B}\models Q_{\mathcal{K}}\vec{y}\left(R_{1}(\vec{y}),\ldots,R_{m}( \vec{y})\right)\), then \(\mathfrak{A}\models Q_{\mathcal{K}}\vec{y}\left(R_{1}(\vec{y}),\ldots,R_{m}( \vec{y})\right)\). The reverse implication is obtained by using the move described above with the structures switched. By allowing only \(k\) variables and repeating rounds indefinitely (unless Spoiler wins at some round), we obtain a game such that Duplicator having a winning strategy implies \(\mathfrak{A}\equiv_{\infty\omega,\mathcal{P}}^{k}\mathfrak{B}\). However, in order to prove the converse implication we need to modify the rules explained above. This is because \(p_{B}(\mathfrak{B})\cup\mathfrak{B}\) is not necessarily closed with respect to the function \(p_{B}\), and in the argument above it would equally well suffice that \(f(\mathfrak{A})\leq\mathfrak{C}\) for some structure \(\mathfrak{C}\) that is obtained by applying \(p_{B}\) repeatedly to \(\mathfrak{B}\). In the next definition we formalize the idea of such repeated applications. **Definition 14**: _Let \(p\colon A^{n}\to A\) be a partial function, and let \(R\subseteq A^{r}\). We define a sequence \(\Gamma^{i}_{p}(R)\), \(i\in\omega\), of \(r\)-ary relations on \(A\) by the following recursion:_ * \(\Gamma^{0}_{p}(R):=R\)_;_ \(\Gamma^{i+1}_{p}(R):=p(R)\cup\Gamma^{i}_{p}(R)\)_._ _Furthermore, we define \(\Gamma^{\omega}_{p}(R)=\bigcup_{i\in\omega}\Gamma^{i}_{p}(R)\)._ _This is generalized to \(\tau\)-structures in the natural way: for all \(i\in\omega\cup\{\omega\}\), \(\Gamma^{i}_{p}(\mathfrak{A})\) is the \(\tau\)-structure \(\mathfrak{C}\) such that \(C=A\) and \(R^{\mathfrak{C}}:=\Gamma^{i}_{p}(R^{\mathfrak{A}})\) for each \(R\in\tau\)._ Note that since \(\Gamma^{i}_{p}(R)\subseteq\Gamma^{i+1}_{p}(R)\) for all \(i\in\omega\) (assuming \(A\) is finite) there exists \(j\leq|A^{r}|\) such that \(\Gamma^{\omega}_{p}(R)=\Gamma^{j}_{p}(R)\). Similarly for any finite structure \(\mathfrak{A}\), \(\Gamma^{\omega}_{p}(\mathfrak{A})=\Gamma^{j}_{p}(\mathfrak{A})\) for some \(j\leq|A^{r}|\), where \(r\) is the maximum arity of relations in \(\mathfrak{A}\). **Lemma 15**: _Let \(\mathcal{P}\) a family of \(n\)-ary partial functions. A quantifier is \(\mathcal{P}\)-closed if and only if the implication_ \[\mathfrak{B}\in\mathcal{K}\text{ and }\mathfrak{A}\leq\Gamma^{\omega}_{p_{A}}( \mathfrak{B})\implies\mathfrak{A}\in\mathcal{K}\] _holds for all structures \(\mathfrak{A}\) and \(\mathfrak{B}\) with \(A=B\)._ _Proof._ Assume first that \(Q_{\mathcal{K}}\) is \(\mathcal{P}\)-closed, \(\mathfrak{B}\in\mathcal{K}\) and \(\mathfrak{A}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{B})\). We show first by induction on \(i\) that \(\Gamma^{i}_{p_{A}}(\mathfrak{B})\in\mathcal{K}\) for all \(i\in\omega\). For \(i=0\) this holds by assumption. If \(\Gamma^{i}_{p_{A}}(\mathfrak{B})\in\mathcal{K}\), then \(\Gamma^{i+1}_{p_{A}}(\mathfrak{B})=p_{A}(\mathfrak{C})\cup\mathfrak{C}\), for \(\mathfrak{C}=\Gamma^{i}_{p_{A}}(\mathfrak{B})\), and hence \(\Gamma^{i+1}_{p_{A}}\) follows from the assumption that \(Q_{\mathcal{K}}\) is \(\mathcal{P}\)-closed. As noted above, there exists \(j\in\omega\) such that \(\Gamma^{\omega}_{p_{A}}(\mathfrak{B})=\Gamma^{j}_{p_{A}}(\mathfrak{B})\). Thus we have \(\mathfrak{A}\leq\Gamma^{j}_{p_{A}}(\mathfrak{B})\leq\Gamma^{j+1}_{p_{A}}( \mathfrak{B})=p_{A}(\Gamma^{j}_{p_{A}}(\mathfrak{B}))\cup\Gamma^{j}_{p_{A}}( \mathfrak{B})\). Since \(\Gamma^{j}_{p_{A}}(\mathfrak{B})\in\mathcal{K}\) and \(\mathcal{K}\) is \(\mathcal{P}\)-closed, it follows that \(\mathfrak{A}\in\mathcal{K}\). Assume then that the implication \[(*)\hskip 14.226378pt\mathfrak{B}\in\mathcal{K}\text{ and }\mathfrak{A}\leq \Gamma^{\omega}_{p_{A}}(\mathfrak{B})\implies\mathfrak{A}\in\mathcal{K}\] holds for all \(\mathfrak{A}\) and \(\mathfrak{B}\) with \(A=B\). Assume further that \(\mathfrak{B}\in\mathcal{K}\) and \(\mathfrak{A}\leq p_{A}(\mathfrak{B})\cup\mathfrak{B}\). By definition \(p_{A}(\mathfrak{B})\cup\mathfrak{B}=\Gamma^{1}_{p_{A}}(\mathfrak{B})\), and since \(\Gamma^{1}_{p_{A}}(\mathfrak{B})\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{B})\), we have \(\mathfrak{A}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{B})\). Thus \(\mathfrak{A}\in\mathcal{K}\) follows from the implication \((*)\). \(\Box\) ### Game for \(\mathcal{P}\)-closed quantifiers We are now ready to give the formal definition of our pebble game for \(\mathcal{P}\)-closed quantifiers. Let \(k\) be a positive integer. Assume that \(\mathfrak{A}\) and \(\mathfrak{B}\) are \(\tau\)-structures for a relational vocabulary \(\tau\). Furthermore, assume that \(\alpha\) and \(\beta\) are assignments on \(\mathfrak{A}\) and \(\mathfrak{B}\), respectively, such that \(\operatorname{dom}(\alpha)=\operatorname{dom}(\beta)\subseteq X\), where \(X=\{x_{1},\ldots,x_{k}\}\). The \(k\)_-pebble \(\mathcal{P}\) game_ for \((\mathfrak{A},\alpha)\) and \((\mathfrak{B},\beta)\) is played between _Spoiler_ and _Duplicator_. We denote the game by \(\operatorname{PG}^{\mathcal{P}}_{k}(\mathfrak{A},\mathfrak{B},\alpha,\beta)\), and we use the shorthand notation \(\operatorname{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\) whenever \(\mathfrak{A}\) and \(\mathfrak{B}\) are clear from the context. **Definition 16**: _The rules of the game \(\mathrm{PG}^{\mathcal{P}}_{k}(\mathfrak{A},\mathfrak{B},\alpha,\beta)\) are the following:_ 1. _If_ \(\alpha\mapsto\beta\notin\mathrm{PI}(\mathfrak{A},\mathfrak{B})\)_, then the game ends, and Spoiler wins._ 2. _If (_1_) does not hold, there are two types of moves that Spoiler can choose to play:_ * **Left \(\mathcal{P}\)-quantifier move:** _Spoiler starts by choosing_ \(r\in[k]\) _and an_ \(r\)_-tuple_ \(\vec{y}\in X^{r}\) _of distinct variables. Duplicator responds with a bijection_ \(f\colon B\to A\)_. Spoiler answers by choosing an_ \(r\)_-tuple_ \(\vec{b}\in B^{r}\)_. Duplicator answers by choosing_ \(P\subseteq A^{r}\) _such that_ \(f(\vec{b})\in\Gamma^{\omega}_{p_{A}}(P)\)_. Spoiler completes the round by choosing_ \(\vec{a}\in P\)_. The players continue by playing_ \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha^{\prime},\beta^{\prime})\)_, where_ \(\alpha^{\prime}:=\alpha[\vec{a}/\vec{y}]\) _and_ \(\beta^{\prime}:=\beta[\vec{b}/\vec{y}]\)_._ * **Right \(\mathcal{P}\)-quantifier move:** _Spoiler starts by choosing_ \(r\in[k]\) _and an_ \(r\)_-tuple_ \(\vec{y}\in X^{r}\) _of distinct variables. Duplicator chooses next a bijection_ \(f\colon A\to B\)_. Spoiler answers by choosing an_ \(r\)_-tuple_ \(\vec{a}\in A^{r}\)_. Duplicator answers by choosing_ \(P\subseteq B^{r}\) _such that_ \(f(\vec{a})\in\Gamma^{\omega}_{p_{B}}(P)\)_. Spoiler completes the round by choosing_ \(\vec{b}\in P\)_. The players continue by playing_ \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha^{\prime},\beta^{\prime})\)_, where_ \(\alpha^{\prime}:=\alpha[\vec{a}/\vec{y}]\) _and_ \(\beta^{\prime}:=\beta[\vec{b}/\vec{y}]\)_._ 3. _Duplicator wins the game if Spoiler does not win it in a finite number of rounds._ We now prove that the game \(\mathrm{PG}^{\mathcal{P}}_{k}\) indeed characterizes equivalence of structures with respect to the infinitary \(k\)-variable logic with all \(\mathcal{P}\)-closed quantifiers. **Theorem 17**: _Let \(\mathcal{P}\) be an invariant family of partial functions. Then Duplicator has a winning strategy in \(\mathrm{PG}^{\mathcal{P}}_{k}(\mathfrak{A},\mathfrak{B},\alpha,\beta)\) if, and only if, \((\mathfrak{A},\alpha)\equiv^{k}_{\infty\omega,\mathcal{P}}(\mathfrak{B},\beta)\)._ _Proof._\(\Rightarrow\): We prove by induction on \(\varphi\in L^{k}_{\infty\omega}(\mathbf{Q}_{\mathcal{P}})\) that (for any assignments \(\alpha\) and \(\beta\)) if Duplicator has a winning strategy in \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\), then \((\mathfrak{A},\alpha)\models\varphi\iff(\mathfrak{B},\beta)\models\varphi\). * If \(\varphi\) is an atomic formula, the claim follows from the fact that Spoiler always wins the game \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\) immediately if \(\alpha\mapsto\beta\notin\mathrm{PI}(\mathfrak{A},\mathfrak{B})\). * The cases \(\varphi=\neg\psi\), \(\varphi=\bigvee\Psi\) and \(\varphi=\bigwedge\Psi\) are straightforward. * By Proposition 5, the negation of the existential quantifier is in \(\mathbf{Q}_{\mathcal{P}}\), and hence we do not need to consider the case \(\varphi=\exists x_{i}\psi\) separately. * Consider then the case \(\varphi=Q_{\times}\vec{y}\,\vec{y}\) for some \(r\)-ary quantifier \(Q_{\times}\in\mathbf{Q}_{\mathcal{P}}\) and interpretation \(\mathcal{I}=(\psi_{1},\ldots,\psi_{\ell})\). We start by assuming that \((\mathfrak{A},\alpha)\models\varphi\). Thus, \(\mathcal{I}(\mathfrak{A},\alpha):=(A,R_{1},\ldots,R_{\ell})\in\mathcal{K}\). Let Spoiler play in the game \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\) a left \(\mathcal{P}\)-quantifier move with \(r\) and the tuple \(\vec{y}\in X^{r}\), and let \(f\colon B\to A\) be the bijection given by the winning strategy of Duplicator. Let \(\mathcal{I}(\mathfrak{B},\beta):=(B,R^{\prime}_{1},\ldots,R^{\prime}_{\ell})\), and for each \(i\in[\ell]\), let \(S_{i}:=f(R^{\prime}_{i})\). We claim that \(\mathfrak{I}:=(A,S_{1},\ldots,S_{\ell})\in\mathcal{K}\). Since \(f\) is an isomorphism \(\mathcal{I}(\mathfrak{B},\beta)\rightarrow\mathfrak{D}\), it follows then that \((\mathfrak{B},\beta)\models\varphi\). To prove the claim it suffices to show that \(\mathfrak{D}\leq\Gamma^{\omega}_{p_{A}}(\mathcal{I}(\mathfrak{A},\alpha))\), since then \(\mathfrak{D}\in\mathcal{K}\) by Lemma 15 and the assumption that \(Q_{\times}\) is \(\mathcal{P}\)-closed. To show this, let \(i\in[\ell]\) and \(\vec{a}\in S_{i}\). We let Spoiler choose the tuple \(\vec{b}=f^{-1}(\vec{a})\) as his answer to the bijection \(f\). Thus, \((\mathfrak{B},\beta[\vec{b}/\vec{y}])\models\psi_{i}\). Let \(P\subseteq A^{r}\) be the answer of Duplicator. Then by the rules of the game \(\vec{a}\in\Gamma^{\omega}_{p_{A}}(P)\), and Duplicator has a winning strategy in the game \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha[\vec{a}/\vec{y}],\beta[\vec{b}/\vec{y}])\) for all \(\vec{a}\in P\). Hence by induction hypothesis \((\mathfrak{A},\alpha[\vec{a}/\vec{y}])\models\psi_{i}\), i.e., \(\vec{a}\in R_{i}\), holds for all \(\vec{a}\in P\). This shows that \(S_{i}\subseteq\Gamma^{\omega}_{p_{A}}(R_{i})\), and since this holds for all \(i\in[\ell]\), we see that \(\mathfrak{D}\leq\Gamma^{\omega}_{p_{A}}(\mathcal{I}(\mathfrak{A},\alpha))\). By using the right \(\mathcal{P}\)-quantifier move in place of the left quantifier move, we can prove that \((\mathfrak{B},\beta)\models\varphi\) implies \((\mathfrak{A},\alpha)\models\varphi\). Thus, \((\mathfrak{A},\alpha)\models\varphi\iff(\mathfrak{B},\beta)\models\varphi\), as desired. \(\Leftarrow\): Assume then that \((\mathfrak{A},\alpha)\equiv^{k}_{\infty,\omega,\mathcal{P}}(\mathfrak{B},\beta)\). Clearly it suffices to show that Duplicator can play in the first round of the game \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\) in such a way that \((\mathfrak{A},\alpha^{\prime})\equiv^{k}_{\infty\omega,\mathcal{P}}(\mathfrak{B },\beta^{\prime})\) holds, where \(\alpha^{\prime}\) and \(\beta^{\prime}\) are the assignments arising from the choices of Spoiler and Duplicator. Assume first that Spoiler decides to play a left \(\mathcal{P}\)-quantifier move in the first round of \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\). Let \(\vec{y}\in X^{r}\) be the tuple of variables he chooses. Since \(A\) and \(B\) are finite, for each \(\vec{a}\in A^{\tau}\) there is a formula \(\Psi_{\vec{a}}\in L^{k}_{\infty\omega}(\mathcal{Q}_{\mathcal{P}})\) such that for any \(\tau\)-structure \(\mathfrak{C}\) of size at most \(\max\{|A|,|B|\}\), any assignment \(\gamma\) on \(\mathfrak{C}\), and any tuple \(\vec{c}\in C^{r}\) we have * \((\mathfrak{A},\alpha[\vec{a}/\vec{y}])\equiv^{k}_{\infty\omega,\mathcal{P}}( \mathfrak{C},\gamma[\vec{c}/\vec{y}])\) if and only if \((\mathfrak{C},\gamma[\vec{c}/\vec{y}])\models\Psi_{\vec{a}}\). Let \(\vec{c}_{1},\ldots,\vec{c}_{\ell}\) be a fixed enumeration of the set \(A^{r}\), and let \(\mathfrak{J}\) be the interpretation \((\Psi_{1},\ldots,\Psi_{m})\), where \(\Psi_{j}:=\Psi_{\vec{c}_{j}}\) for each \(j\in[m]\). We define \(\mathcal{K}\) to be the closure of the class \(\{\mathfrak{D}\mid\mathfrak{D}\subseteq\Gamma^{\omega}_{p_{A}}(\mathfrak{J}( \mathfrak{A},\alpha))\}\) under isomorphisms. Note that if \(\mathfrak{D}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{J}(\mathfrak{A},\alpha))\) and \(\mathfrak{C}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{D})\), then clearly \(\mathfrak{C}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{J}(\mathfrak{A},\alpha))\). Hence by Lemma 15, the quantifier \(\mathcal{Q}_{\mathcal{K}}\) is \(\mathcal{P}\)-closed. Moreover, since \(\mathfrak{J}(\mathfrak{A},\alpha)\in\mathcal{K}\), we have \((\mathfrak{A},\alpha)\models Q_{\mathcal{K}}\vec{y}\mathfrak{J}\), and consequently by our assumption, \((\mathfrak{B},\beta)\models Q_{\mathcal{K}}\vec{y}\mathfrak{J}\). Thus, there is a structure \(\mathfrak{D}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{J}(\mathfrak{A},\alpha))\) and an isomorphism \(f\colon\mathfrak{J}(\mathfrak{B},\beta)\rightarrow\mathfrak{D}\). We let Duplicator to use the bijection \(f\colon B\to A\) as her answer to the choice \(\vec{y}\) of Spoiler. Let \(\vec{b}\in B^{r}\) be the answer of Spoiler to \(f\), and let \(\vec{a}=f(\vec{b})\). Clearly \((\mathfrak{A},\alpha)\models\forall\vec{y}\bigvee_{j\in[\ell]}\Psi_{j}\), whence there exists \(j\in[\ell]\) such that \((\mathfrak{B},\beta[\vec{b}/\vec{y}])\models\Psi_{j}\), or in other words, \(\vec{b}\in R^{J(\mathfrak{B},\beta)}_{j}\). Since \(f\) is an isomorphism \(\mathfrak{J}(\mathfrak{B},\beta)\rightarrow\mathfrak{D}\), we have \(\vec{a}\in R^{\mathfrak{D}}_{j}\). We let Duplicator to use \(P:=R^{\mathfrak{J}(\mathfrak{A},\alpha)}_{j}\) as her answer to the choice \(\vec{b}\) of Spoiler; this is a legal move since \(\mathfrak{D}\leq\Gamma^{\omega}_{p_{A}}(\mathfrak{A},\alpha)\). Observe now that since \(P=R^{J(\mathfrak{A},\alpha)}_{j}\), we have \((\mathfrak{A},\alpha[\vec{a}/\vec{y}])\models\Psi_{\vec{c}_{j}}\), and consequently \((\mathfrak{A},\alpha[\vec{c}_{j}/\vec{y}])\equiv^{k}_{\infty\omega,\mathcal{P} }(\mathfrak{A},\alpha[\vec{a}/\vec{y}])\), for all \(\vec{a}\in P\). On the other hand we also have \((\mathfrak{B},\beta[\vec{b}/\vec{y}])\models\Psi_{\vec{c}_{j}}\), and hence \((\mathfrak{A},\alpha[\vec{c}_{j}/\vec{y}])\equiv^{k}_{\infty\omega,\mathcal{P} }(\mathfrak{B},\beta[\vec{b}/\vec{y}])\). Thus the condition \((\mathfrak{A},\alpha^{\prime})\equiv^{k}_{\infty\omega,\mathcal{P}}( \mathfrak{B},\beta^{\prime})\), where \(\alpha^{\prime}=\alpha[\vec{a}/\vec{y}]\) and \(\beta^{\prime}=\beta[\vec{b}/\vec{y}]\), holds after the first round of \(\mathrm{PG}^{\mathcal{P}}_{k}(\alpha,\beta)\) irrespective of the choice \(\vec{a}\in P\) of Spoiler in the end of the round. The case where Spoiler starts with a right \(\mathcal{P}\)-quantifier move is handled in the same way by switching the roles of \((\mathfrak{A},\alpha)\) and \((\mathfrak{B},\beta)\). \(\Box\) ## 5 Playing the game In this section we use the game \(\mathrm{PG}^{k}_{\mathcal{P}}\) to show inexpressibility of a property of finite structures in the infinitary finite variable logic \(L^{\omega}_{\infty\omega}\) augmented by all \(\mathcal{N}_{\ell}\)-closed quantifiers. More precisely, we prove that the Boolean constraint satisfaction problem \(\mathsf{CSP}(\mathfrak{C}_{\ell})\), where \(\mathfrak{C}_{\ell}\) is the structure with \(C=\{0,1\}\) and two \(\ell\)-ary relations \(R_{0}=\{(b_{1},\ldots,b_{\ell})\mid\sum_{i\in[\ell]}b_{i}\equiv 0\pmod{2}\}\) and \(R_{1}=\{(b_{1},\ldots,b_{\ell})\mid\sum_{i\in[\ell]}b_{i}\equiv 1\pmod{2}\}\), is not definable in \(L^{\omega}_{\infty\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\). ### CFI construction In the proof of the undefinability of \(\mathsf{CSP}(\mathfrak{C}_{\ell})\) we use a variation of the well-known CFI construction, due to Cai, Furer and Immerman [4]. Our construction is a minor modification of the one that was used in [14] for producing non-isomorphic structures on which Duplicator wins the \(k,n\)-bijection game. We start by explaining the details of the construction. Let \(G=(V,E,\leq^{G})\) be a connected \(\ell\)-regular ordered graph. For each vertex \(v\in V\), we use the notation \(E(v)\) for the set of edges adjacent to \(v\) and \(\vec{c}(v)=(e_{1},\ldots,e_{\ell})\) for the tuple that lists \(E(v)\) in the order \(\leq^{G}\). The CFI structures we use have in the universe two elements \((e,1)\) and \((e,2)\) for each \(e\in E\), and two \(\ell\)-ary relations that connect such pairs \((e,i)\) for edges \(e\) that are adjacent to some vertex \(v\in V\). **Definition 18**: _Let \(G=(V,E,\leq^{G})\) be a connected \(\ell\)-regular ordered graph and let \(U\subseteq V\). We define a CFI structure \(\mathfrak{A}_{\ell}(G,U)=(A_{\ell}(G),R^{\mathfrak{A}_{\ell}(G,U)}_{0},R^{ \mathfrak{A}_{\ell}(G,U)}_{1})\), where \(\mathrm{ar}(R_{0})=\mathrm{ar}(R_{1})=\ell\), as follows._ * \(A_{\ell}(G):=E\times[2]\)_,_ * \(R_{0}^{\mathfrak{A}_{\ell}(G,U)}:=\bigcup_{v\in V\setminus U}R(v)\cup\bigcup_{v \in U}\tilde{R}(v)\) _and_ \(R_{1}^{\mathfrak{A}_{\ell}(G,U)}:=\bigcup_{v\in U}R(v)\cup\bigcup_{v\in V \setminus U}\tilde{R}(v)\)_, where_ * \(R(v):=\{((e_{1},i_{1}),\ldots,(e_{\ell},i_{\ell}))\mid(e_{1},\ldots,e_{\ell})= \vec{e}(v),\,\sum_{j\in[\ell]}i_{j}=0\pmod{2}\}\)_, and_ * \(\tilde{R}(v):=\{((e_{1},i_{1}),\ldots,(e_{\ell},i_{\ell}))\mid(e_{1},\ldots,e_{ \ell})=\vec{e}(v),\,\sum_{j\in[\ell]}i_{j}=1\pmod{2}\}\)_._ For each \(v\in V\), we denote the set \(E(v)\times[2]\) by \(A(v)\). Furthermore, we define \(\mathfrak{A}_{\ell}(v):=(A(v),R(v),\tilde{R}(v))\) and \(\tilde{\mathfrak{A}}_{\ell}(v):=(A(v),\tilde{R}(v),R(v))\). By a similar argument as in the CFI structures constructed in [14] and [15] it can be proved that \(\mathfrak{A}_{\ell}(G,U)\) and \(\mathfrak{A}_{\ell}(G,U^{\prime})\) are isomorphic if and only if \(|U|\) and \(|U^{\prime}|\) are of the same parity. We choose \(\mathfrak{A}_{\ell}^{\rm ev}(G):=\mathfrak{A}_{\ell}(G,\emptyset)\) and \(\mathfrak{A}_{\ell}^{\rm od}(G):=\mathfrak{A}_{\ell}(G,\{v_{0}\})\) as representatives of these two isomorphism classes, where \(v_{0}\) is the least element of \(V\) with respect to the linear order \(\leq^{G}\). We show first that these structures are separated by \({\sf CSP}(\mathfrak{C}_{\ell})\). **Lemma 19**: \(\mathfrak{A}_{\ell}^{\rm ev}(G)\in{\sf CSP}(\mathfrak{C}_{\ell})\)_, but \(\mathfrak{A}_{\ell}^{\rm od}(G)\not\in{\sf CSP}(\mathfrak{C}_{\ell})\)._ _Proof._ Let \(h\colon A_{\ell}(G)\to\{0,1\}\) be the function such that \(h((e,1))=1\) and \(h((e,2))=0\) for all \(e\in E\). Then for any tuple \(((e_{1},i_{1}),\ldots,(e_{\ell},i_{\ell}))\) the parity of \(\sum_{j\in[\ell]}h((e_{j},i_{j}))\) is the same as the parity of \(\sum_{j\in[\ell]}i_{j}\). Thus, \(h\) is a homomorphism \(\mathfrak{A}_{\ell}^{\rm ev}(G)\to\mathfrak{C}_{\ell}\). To show that \(\mathfrak{A}_{\ell}^{\rm od}(G)\not\in{\sf CSP}(\mathfrak{C}_{\ell})\), assume towards contradiction that \(g\colon A_{\ell}(G)\to\{0,1\}\) is a homomorphism \(\mathfrak{A}_{\ell}^{\rm od}(G)\to\mathfrak{C}_{\ell}\). Then for every \(e\in E\) necessarily \(g((e,1))\neq g((e,2))\). Furthermore, for every \(v\in V\setminus\{v_{0}\}\), the number \(n_{v}:=|\{e\in E(v)\mid g((e,2))=1\}|\) must be even, while the number \(n_{v_{0}}\) must be odd. Thus, \(\sum_{v\in V}n_{v}\) must be odd. However, this is impossible, since clearly \(\sum_{v\in V}n_{v}=2|\{e\in E\mid g((e,2))=1\}|\). \(\Box\) ### Good bijections Our aim is to prove, for a suitable graph \(G\), that Duplicator has a winning strategy in \({\rm PG}_{k}^{\mathfrak{A}_{\ell}}(\mathfrak{A}_{\ell}^{\rm ev}(G),\mathfrak{ A}_{\ell}^{\rm od}(G),\emptyset,\emptyset)\). For the winning strategy, Duplicator needs a collection of well-behaved bijections. We define such a collection GB in Definition 23 below. One requirement is that the bijections preserve the first component of the elements \((e,i)\in A_{\ell}(G)\). **Definition 20**: _A bijection \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) is edge preserving if for every \(e\in E\) and \(i\in[2]\), \(f((e,i))\) is either \((e,1)\) or \((e,2)\)._ For any edge preserving \(f\) and any \(v\in V\) we denote by \(f_{v}\) the restriction of \(f\) to the set \(E(v)\times[2]\). The _switching number_\({\rm swn}(f_{v})\) of \(f_{v}\) is \(|\{e\in E(v)\mid f_{v}((e,1))=(e,2)\}|\). The lemma below follows directly from the definitions of \(\mathfrak{A}_{\ell}(v)\) and \(\tilde{\mathfrak{A}}_{\ell}(v)\). **Lemma 21**: _Let \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) be an edge preserving bijection, and let \(v\in V\)._ _(a) If \({\rm swn}(f_{v})\) is even, then \(f_{v}\) is an automorphism of the structures \(\mathfrak{A}_{\ell}(v)\) and \(\tilde{\mathfrak{A}}_{\ell}(v)\)._ _(b) If \({\rm swn}(f_{v})\) is odd, then \(f_{v}\) is an isomorphism between the structures \(\mathfrak{A}_{\ell}(v)\) and \(\tilde{\mathfrak{A}}_{\ell}(v)\)._ Given an edge preserving bijection \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) we denote by \({\rm Odd}(f)\) the set of all \(v\in V\) such that \({\rm swn}(f_{v})\) is odd. Observe that \(|{\rm Odd}(f)|\) is necessarily even. **Corollary 22**: _An edge preserving bijection \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) is an automorphism of the structures \(\mathfrak{A}_{\ell}^{\rm ev}(G)\) and \(\mathfrak{A}_{\ell}^{\rm od}(G)\) if and only if \({\rm Odd}(f)=\emptyset\)._ _Proof._ If \({\rm Odd}(f)=\emptyset\), then by Lemma 21(a) \(f_{v}\) is an automorphism of \(\mathfrak{A}_{\ell}(v)\) and \(\tilde{\mathfrak{A}}_{\ell}(v)\) for all \(v\in V\). Clearly this means that \(f\) is an automorphism of \(\mathfrak{A}_{\ell}^{\rm ev}(G)\) and \(\mathfrak{A}_{\ell}^{\rm od}(G)\). On the other hand, if \(v\in{\rm Odd}(f)\), then by Lemma 21(b), for any tuple \(\vec{a}\in A(v)^{\ell}\), we have \(\vec{a}\in R(v)\iff f(\vec{a})\in\tilde{R}(v)\). Since \(R(v)\cap\tilde{R}(v)=\emptyset\), it follows that \(f\) is not an automorphism of \(\mathfrak{A}_{\ell}^{\rm ev}(G)\) and \(\mathfrak{A}_{\ell}^{\rm od}(G)\). \(\Box\) **Definition 23**: _Let \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) be edge preserving bijection. Then \(f\) is a good bijection if either \(\mathrm{Odd}(f)=\emptyset\) or \(\mathrm{Odd}(f)=\{v_{0},v\}\) for some \(v\in V\setminus\{v_{0}\}\). We denote the set of all good bijections by \(\mathrm{GB}\)._ Note that if \(f\colon A_{\ell}(G)\to A_{\ell}(G)\) is a good bijection, then there is exactly one vertex \(v^{*}\in V\) such that \(f_{v^{*}}\) is not a partial isomorphism \(\mathfrak{A}_{\ell}^{\mathrm{ev}}(G)\to\mathfrak{A}_{\ell}^{\mathrm{od}}(G)\). In case \(\mathrm{Odd}(f)=\emptyset\), \(v^{*}=v_{0}\), while in case \(\mathrm{Odd}(f)=\{v_{0},v\}\) for some \(v\neq v_{0}\), \(v^{*}=v\). We denote this vertex \(v^{*}\) by \(\mathrm{tw}(f)\) (the _twist_ of \(f\)). Assume now that Duplicator has played a good bijection \(f\) in the game \(\mathrm{PG}_{k}^{\mathcal{N}_{\ell}}\) on the structures \(\mathfrak{A}_{\ell}^{\mathrm{ev}}(G)\) and \(\mathfrak{A}_{\ell}^{\mathrm{od}}(G)\). Then it is sure that Spoiler does not win the game in the next position \((\alpha,\beta)\) if \((e,1)\) and \((e,2)\) are not in the range of \(\alpha\) (and \(\beta\)) for any \(e\in E(\mathrm{tw}(f))\). This leads us to the following notion. **Definition 24**: _Let \(f\) be a good bijection, and let \(F\subseteq E\). Then \(f\) if good for \(F\) if \(E(\mathrm{tw}(f))\cap F=\emptyset\). We denote the set of all bijections that are good for \(F\) by \(\mathrm{GB}(F)\)._ **Lemma 25**: _If \(f\in\mathrm{GB}(F)\), then \(f\upharpoonright(F\times[2])\) is a partial isomorphism \(\mathfrak{A}_{\ell}^{\mathrm{ev}}(G)\to\mathfrak{A}_{\ell}^{\mathrm{od}}(G)\)._ _Proof._ Clearly \(f\upharpoonright(F\times[2])\subseteq\bigcup_{v\in V\setminus\{\mathrm{tw}(f) \}}f_{v}\). By Lemma 21, \(f_{v}\) is an automorphism of \(\mathfrak{A}_{\ell}(v)\) for any \(v\in V\setminus\{\mathrm{tw}(f),v_{0}\}\), and if \(v_{0}\neq\mathrm{tw}(f)\), \(f_{v_{0}}\) is an isomorphism \(\mathfrak{A}_{\ell}(v)\to\tilde{\mathfrak{A}}_{\ell}(v)\). The claim follows from this. \(\Box\) Given a good bijection \(f\) with \(\mathrm{tw}(f)=u\) and an \(E\)-path \(P=(u_{0},\ldots,u_{m})\) from \(u=u_{0}\) to \(u^{\prime}=u_{m}\), we obtain a new edge preserving bijection \(f_{P}\) by switching \(f\) on the edges \(e_{i}:=\{u_{i},u_{i+1}\}\), \(i<m\), of \(P\): \(f_{P}((e_{i},j))=(e_{i},3-j)\) for \(i<m\), and \(f_{P}(c)=f(c)\) for other \(c\in A_{\ell}(G)\). Clearly \(f_{P}\) is also a good bijection, and \(\mathrm{tw}(f_{P})=u^{\prime}\). ### Cops and Robber game In order to prove that Duplicator has a winning strategy in \(\mathrm{PG}_{k}^{\mathcal{N}_{\ell}}(\mathfrak{A}_{\ell}^{\mathrm{ev}}(G), \mathfrak{A}_{\ell}^{\mathrm{od}}(G),\emptyset,\emptyset)\) we need to assume that the graph \(G\) has a certain largeness property with respect the number \(k\). We formulate this largeness property in terms of a game, \(\mathrm{CR}_{k}^{\ell}(G)\), that is a new variation of the _Cops&Robber games_ used for similar purposes in [14] and [15]. **Definition 26**: _The game \(\mathrm{CR}_{k}^{\ell}(G)\) is played between two players, Cop and Robber. The positions of the game are pairs \((F,u)\), where \(F\subseteq E\), \(|F|\leq k\), and \(u\in V\). The rules of the game are the following:_ * _Assume that the position is_ \((F,u)\)_._ * _If_ \(E(u)\cap F\neq\emptyset\)_, the game ends and Cop wins._ * _Otherwise Cop chooses a set_ \(F^{\prime}\subseteq E\) _such that_ \(|F^{\prime}|\leq k\)_. Then Robber answers by giving mutually disjoint_ \(E\setminus(F\cap F^{\prime})\)_-paths_ \(P_{i}=(u,u_{1}^{i},\ldots,u_{n_{i}}^{i})\)_,_ \(i\in[\ell]\)_, from_ \(u\) _to vertices_ \(u_{i}:=u_{n_{i}}^{i}\)_; here mutual disjointness means that_ \(P_{i}\) _and_ \(P_{i^{\prime}}\) _do not share edges for_ \(i\neq i^{\prime}\) _(i.e.,_ \(u_{1}^{i}\neq u_{1}^{i^{\prime}}\) _and_ \(\{u_{j}^{i},u_{j+1}^{i}\}\neq\{u_{j^{\prime}}^{i},u_{j^{\prime}+1}^{i}\}\) _for all_ \(j\) _and_ \(j^{\prime}\)_). Then Cop completes the round by choosing_ \(i\in[\ell]\)_. The next position is_ \((F^{\prime},u_{i})\)_._ The intuition of the game \(\mathrm{CR}_{k}^{\ell}(G)\) is that Cop has \(k\) pebbles that he plays on edges of \(G\) forming a set \(F\subseteq E\); these pebbles mark the edges \(e\) such that \((e,1)\) or \((e,2)\) is in the range of \(\alpha\) or \(\beta\) in a position \((\alpha,\beta)\) of the game \(\mathrm{PG}_{k}^{\mathcal{N}_{\ell}}\) on \(\mathfrak{A}_{\ell}^{\mathrm{ev}}(G)\) and \(\mathfrak{A}_{\ell}^{\mathrm{od}}(G)\). Robber has one pebble that she plays on the vertices of \(G\); this pebble marks the vertex \(\mathrm{tw}(f)\), where \(f\) is the good bijection played by Duplicator in the previous round of \(\mathrm{PG}_{k}^{\mathcal{N}_{\ell}}\). Cop captures Robber and wins the game if after some round (at least) one of his pebbles is on an edge that is adjacent to the vertex containing Robber's pebble. This corresponds to a position \((\alpha,\beta)\) in the game \(\mathrm{PG}_{k}^{\mathcal{N}_{\ell}}\) such that \(\alpha\mapsto\beta\) is potentially not a partial isomorphism. Otherwise Lemma 25 guarantees that \(\alpha\mapsto\beta\) is a partial isomorphism. Cop can then move any number of his pebbles to new positions on \(G\). While the pebbles Cop decides to move are still on their way to their new positions, Robber is allowed to prepare \(\ell\) mutually disjoint escape routes along edges of \(G\) that do not contain any stationary pebbles of Cop. We show in the proof of Theorem 29 that these escape routes generate tuples \(\vec{a}_{1},\ldots,\vec{a}_{\ell}\) such that \(f(\vec{b})=\hat{q}(\vec{a}_{1},\ldots,\vec{a}_{\ell})\), where \(q=n^{\ell}_{A_{\ell}(G)}\) and \(\vec{b}\) is the tuple chosen by Spoiler after Duplicator played \(f\). This gives Duplicator a legal answer \(P=\{a_{1},\ldots,\vec{a}_{\ell}\}\) to \(\vec{b}\). Then Spoiler completes the round by choosing one of the tuples in \(P\). Correspondingly, in the end of the round of \(\mathrm{CR}^{\ell}_{k}(G)\) Cop chooses which escape route Robber has to use by blocking all but one of them. **Definition 27**: _Assume that \(u\in V\) and \(F\subseteq E\) is a set of edges such that \(|F|\leq k\). We say that \(u\) is \(k\)-safe for \(F\) if Robber has a winning strategy in the game \(\mathrm{CR}^{\ell}_{k}(G)\) starting from position \((F,u)\)._ We prove next the existence of graphs \(G\) such that Robber has a winning strategy in the game \(\mathrm{CR}^{\ell}_{k}(G)\). **Theorem 28**: _For every \(\ell\geq 3\) and every \(k\geq 1\), there is an \(\ell\)-regular graph \(G=(V,E)\) such that every vertex \(v\in V\) is \(k\)-safe for \(\emptyset\)._ _Proof._ Clearly if Robber has a winning strategy in \(\mathrm{CR}^{\ell}_{k}(G)\), it also has a winning strategy in \(\mathrm{CR}^{\ell}_{k^{\prime}}(G)\) for \(k^{\prime}<k\). Thus, it suffices to prove the theorem for \(k\geq\ell\). By a well-known result of Erdos and Sachs [10], there exist \(\ell\)-regular connected graphs of _girth_\(g\) for arbitrarily large \(g\). Choose a positive integer \(d\) with \(d>\frac{\log 2k}{\log(\ell-1)}+1\) and let \(G\) be an \(\ell\)-regular graph of girth \(g>6d\). We claim that any vertex \(v\) in \(G\) is \(k\)-safe for \(\emptyset\). To prove this, we show inductively that Robber can maintain the following invariant in any position \((F,u)\) reached during the game: * for each edge \(e\in F\), neither end point of \(e\) is within distance \(d\) of \(u\) in \(G\). Note that, from the assumption that \(k\geq\ell\) and \(d>\frac{\log 2k}{\log(\ell-1)}+1\), it follows that \(d\geq 2\). Thus, the invariant \((*)\) guarantees, in particular, that Cop does not win at any point. Clearly the invariant \((*)\) is satisfied at the initial position, since \(F\) is empty. Suppose now that it is satisfied in some position \((F,u)\) and Cop chooses a set \(F^{\prime}\) in the next move. Let \(C\subseteq V\) denote the set of end points of all edges in \(F^{\prime}\). Since \(|F^{\prime}|\leq k\), we have \(|C|\leq 2k\). Let \(N\subseteq V\) denote the collection of vertices which are at distance at most \(3d\) from \(u\). By the assumption on the girth of \(G\), the induced subgraph \(G[N]\) is a tree. We can consider it as a rooted tree, with root \(u\). Then, \(u\) has exactly \(\ell\) children. All vertices in \(N\) at distance less than \(3d\) from \(u\) have exactly \(\ell-1\) children (and one parent), and all vertices at distance exactly \(3d\) from \(u\) are leaves of the tree. This allows us to speak, for instance, of the subtree rooted at a vertex \(u^{\prime}\) meaning the subgraph of \(G\) induced by the vertices \(x\) in \(N\) such that the unique path from \(u\) to \(x\) in \(G[N]\) goes through \(u^{\prime}\). Let \(u_{1},\ldots,u_{\ell}\) be the children of \(u\). For each \(i\), let \(U_{i}\) denote the set of descendants of \(u_{i}\) that are at distance exactly \(d\) from \(u\) (and so at distance \(d-1\) from \(u_{i}\)). Note that the collection \(U_{1},\ldots,U_{\ell}\) forms a partition of the set of vertices in \(N\) that are at distance exactly \(d\) from \(u\). Each \(x\in U_{i}\) is the root of a tree of height \(2d\). Moreover, since the tree below \(u_{i}\) is \((\ell-1)\)-regular, \(U_{i}\) contains exactly \((\ell-1)^{d-1}\) vertices. By the assumption that \(d>\frac{\log 2k}{\log(\ell-1)}+1\), it follows that \((\ell-1)^{d-1}>2k\geq|C|\) and therefore each \(U_{i}\) contains at least one vertex \(x_{i}\) such that the subtree rooted at \(x_{i}\) contains no vertex in \(C\). Let \(y_{i}\) be any descendant of \(x_{i}\) at distance \(d\) from \(x_{i}\) and let \(P_{i}\) denote the unique path in \(G[N]\) from \(u\) to \(y_{i}\). Robber's move is to play the paths \(P_{1},\ldots,P_{\ell}\). We now verify that this is a valid move, and that it maintains the required invariant \((*)\). First, note that the paths \(P_{1},\ldots,P_{\ell}\) are paths in the tree \(G[N]\) all starting at \(u\) and the second vertex in path \(P_{i}\) is \(u_{i}\). It follows that the paths are pairwise edge disjoint. We next argue that no path \(P_{i}\) goes through an edge in \(F\cap F^{\prime}\). Indeed, by the inductive assumption, no endpoint of an edge in \(F\) appears within distance \(d\) of \(u\) and therefore the path from \(u\) to \(x_{i}\) does not go through any such vertex. Moreover, by the choice of \(x_{i}\), no endpoint of an edge in \(F^{\prime}\) appears in the subtree rooted at \(x_{i}\) and therefore the path from \(x_{i}\) to \(y_{i}\) does not go through any such vertex. Together these ensure that the path \(P_{i}\) does not visit any vertex that is an endpoint of an edge in \(F\cap F^{\prime}\). Finally, to see that the invariant \((*)\) is maintained, note that all vertices that are at distance at most \(d\) from \(y_{i}\) are in the subtree of \(G[N]\) rooted at \(x_{i}\). The choice of \(x_{i}\) means this contains no vertex in \(C\). This is exactly the condition that we wished to maintain. \(\Box\) ### Winning the game We are now ready to prove that a winning strategy for Robber in \(\mathrm{CR}^{\,\ell}_{k}(G)\) generates a winning strategy for Duplicator in the game \(\mathrm{PG}^{\mathcal{N}_{\ell}}_{k}\) on the structures \(\mathfrak{A}^{\mathrm{ev}}_{\ell}(G)\) and \(\mathfrak{A}^{\mathrm{od}}_{\ell}(G)\). **Theorem 29**: _Let \(G\) be a connected \(\ell\)-regular ordered graph. If \(v_{0}\) is \(k\)-safe for the empty set, then Duplicator has a winning strategy in the game \(\mathrm{PG}^{\mathcal{N}_{\ell}}_{k}(\mathfrak{A}^{\mathrm{ev}}_{\ell}(G), \mathfrak{A}^{\mathrm{od}}_{\ell}(G),\emptyset,\emptyset)\)._ _Proof._ We show that Duplicator can maintain the following invariant for all positions \((\alpha,\beta)\) obtained during the play of the game \(\mathrm{PG}^{\mathcal{N}_{\ell}}_{k}(\mathfrak{A}^{\mathrm{ev}}_{\ell}(G), \mathfrak{A}^{\mathrm{od}}_{\ell}(G),\emptyset,\emptyset)\): **(\(\dagger\))**: There exists a bijection \(f\in\mathrm{GB}(F_{\alpha})\) such that \(p:=\alpha\mapsto\beta\subseteq f\) and \(\mathrm{tw}(f)\) is \(k\)-safe for \(F_{\alpha}\), where \(F_{\alpha}:=\{e\in E\mid\mathrm{rng}(\alpha)\cap\{e\}\times[2]\neq\emptyset\}\). Note that if (\(\dagger\)) holds, then \(p\subseteq f\upharpoonright(F_{\alpha}\times[2])\) and, by Lemma 25, \(f\upharpoonright(F_{\alpha}\times[2])\in\mathrm{PI}(\mathfrak{A}^{\mathrm{ev}}_{ \ell}(G),\mathfrak{A}^{\mathrm{od}}_{\ell}(G))\), whence Spoiler does not win the game in position \((\alpha,\beta)\). Thus, maintaining the invariant (\(\dagger\)) during the play guarantees a win for Duplicator. Note first that (\(\dagger\)) holds in the initial position \((\alpha,\beta)=(\emptyset,\emptyset)\) of the game: if \(f_{0}\in\mathrm{GB}\) is the bijection with \(\mathrm{tw}(f_{0})=v_{0}\), as \(\emptyset\mapsto\emptyset=\emptyset\subseteq f_{0}\) and \(\mathrm{tw}(f_{0})\) is \(k\)-safe for \(F_{\emptyset}=\emptyset\). Assume then that (\(\dagger\)) holds for a position \((\alpha,\beta)\), and assume that Spoiler plays a left \(\mathcal{N}_{\ell}\)-quantifier move by choosing \(r\leq k\) and \(\vec{y}\in X^{r}\). Duplicator answers this by giving the bijection \(f^{-1}\). Let \(\vec{b}=(b_{i},\ldots,b_{r})\in A_{\ell}(G)^{r}\) be the second part of Spoiler's move, and let \(F^{\prime}\) be the set \(\{e\in E\mid\mathrm{rng}(\beta[\vec{b}/\vec{y}])\cap\{e\}\times[2]\neq\emptyset\}\). Since \(\mathrm{tw}(f)\) is \(k\)-safe for \(F_{\alpha}\), there are mutually disjoint \(E\setminus(F_{\alpha}\cap F^{\prime})\)-paths \(P_{i}\), \(i\in[\ell]\), from \(\mathrm{tw}(f)\) to some vertices \(u_{i}\) that are \(k\)-safe for the set \(F^{\prime}\). Let \(f_{P_{i}}\), \(i\in[\ell]\), be the good bijections obtained from \(f\) as explained before Definition 24. Now Duplicator answers the move \(\vec{b}\) of Spoiler by giving the set \(P=\{\vec{a}_{1},\ldots,\vec{a}_{\ell}\}\) of \(r\)-tuples, where \(\vec{a}_{i}:=f_{P_{i}}^{-1}(\vec{b})\) for each \(i\in[\ell]\). To see that this is a legal move, observe that since the paths \(P_{i}\) are disjoint, for each \(j\in[r]\) there is at most one \(i\in[\ell]\) such that \(f_{P_{i}}^{-1}(b_{j})\neq f^{-1}(b_{j})\). Thus we have \(\hat{q}(\vec{a}_{1},\ldots,\vec{a}_{\ell})=f^{-1}(\vec{b})\), and hence \(f^{-1}(\vec{b})\in q(P)\subseteq\Gamma_{q}^{\omega}(P)\) for \(q=n_{A_{\ell}(G)}^{\ell}\), as required. Let Spoiler complete the round of the game by choosing \(i\in[\ell]\); thus, the next position is \((\alpha^{\prime},\beta^{\prime}):=(\alpha[\vec{a}_{i}/\vec{y}],\beta[\vec{b}/ \vec{y}])\). It suffices now to show that (\(\dagger\)) holds for the position \((\alpha^{\prime},\beta^{\prime})\) and the bijection \(f^{\prime}:=f_{P_{i}}\). Note first that \(F_{\alpha^{\prime}}=F^{\prime}\), since clearly \(\mathrm{rng}(\alpha[\vec{a}_{i}/\vec{y}])\cap\{e\}\times[2]\neq\emptyset\) if, and only if, \(\mathrm{rng}(\beta[\vec{b}/\vec{y}])\cap\{e\}\times[2]\neq\emptyset\). Thus, \(\mathrm{tw}(f^{\prime})=u_{i}\) is \(k\)-safe for \(F_{\alpha^{\prime}}\). This implies that \(f^{\prime}\in\mathrm{GB}(F_{\alpha^{\prime}})\), since otherwise by Definition 26, Cop would win the game \(\mathrm{CR}_{k}(G)\) immediately in position \((F_{\alpha^{\prime}},\mathrm{tw}(f^{\prime}))\). It remains to show that \(p^{\prime}:=\alpha^{\prime}\mapsto\beta^{\prime}\) is contained in \(f^{\prime}\). For all components \(a_{i}^{j}\) of \(\vec{a}_{i}\) we have \(p^{\prime}(a_{i}^{j})=b_{j}=f^{\prime}(a_{i}^{j})\) by definition of \(\vec{a}_{i}\). On the other hand, for any element \(a\in\mathrm{dom}(p^{\prime})\setminus\{a_{1}^{i},\ldots,a_{i}^{r}\}\) we have \(p^{\prime}(a)=p(a)=f(a)\). Furthermore, since the path \(P_{i}\) does not contain any edges in \(F_{\alpha}\cap F_{\alpha^{\prime}}\), we have \(f^{\prime}\upharpoonright(F_{\alpha}\cap F_{\alpha^{\prime}})\times[2]=f\upharpoonright(F_{ \alpha}\cap F_{\alpha^{\prime}})\times[2]\), and since clearly \(a\in(F_{\alpha}\cap F_{\alpha^{\prime}})\times[2]\), we see that \(f^{\prime}(a)=f(a)\). Thus, \(p^{\prime}(a)=f^{\prime}(a)\). The case where Spoiler plays a right \(\mathcal{N}_{\ell}\)-quantifier move is similar. \(\Box\) Note that the vocabulary of the structures \(\mathfrak{A}^{\mathrm{ev}}_{\ell}(G)\) and \(\mathfrak{A}^{\mathrm{od}}_{\ell}(G)\) consists of two \(\ell\)-ary relation symbols. The presence of at least \(\ell\)-ary relations is actually necessary: Duplicator cannot have a winning strategy in \(\mathrm{PG}_{\ell-1}^{\mathcal{N}_{\ell}}\) on structures containing only relations of arity less than \(\ell\), since by Corollary 10(b), all properties of such structures are definable in \(L_{\infty\omega}^{\ell-1}(\mathbf{Q}_{\mathcal{N}_{\ell}})\). From Lemma 19, Theorem 28 and Theorem 29, we immediately obtain the result. **Theorem 30**: _For any \(\ell\geq 3\), \(\mathsf{CSP}(\mathfrak{E}_{\ell})\) is not definable in \(L_{\infty\omega}^{\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\)._ Note that \(\mathsf{CSP}(\mathfrak{E}_{\ell})\) corresponds to solving systems of linear equations over \(\mathbb{Z}\,/2\,\mathbb{Z}\) with all equations containing (at most) \(\ell\) variables. Thus, as a corollary we see that solvability of such systems of equations cannot be expressed in \(L_{\infty\omega}^{\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\) for any \(\ell\). Furthermore, since systems of linear equations over \(\mathbb{Z}\,/2\,\mathbb{Z}\) can be solved in polynomial time, we see that the complexity class PTIME is not contained in \(L_{\infty\omega}^{\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\) for any \(\ell\). Finally, note that since the class \(\mathsf{CSP}(\mathfrak{E}_{\ell})\) is downwards monotone, by Lemma 9 the quantifier \(Q_{\mathsf{CSP}(\mathfrak{E}_{\ell})}\) is \(\mathcal{N}_{\ell+1}\)-closed. Thus, we get the following hierarchy result for the near-unanimity families \(\mathcal{N}_{\ell}\) with respect to the arity \(\ell\) of the partial functions. **Theorem 31**: _For every \(\ell\geq 3\) there is a quantifier in \(\mathbf{Q}_{\mathcal{N}_{\ell+1}}\) which is not definable in \(L_{\infty\omega}^{\omega}(\mathbf{Q}_{\mathcal{N}_{\ell}})\)._ ## 6 Conclusion We have introduced new methods, in the form of pebble games, for proving inexpressibility in logics extended with generalized quantifiers. There is special interest in proving inexpressibility in logics with quantifiers of unbounded arity. We introduced a general method of defining such collections inspired by the equational theories of polymorphisms arising in the study of constraint satisfaction problems. Perhaps surprisingly, while the collection of CSP that have near-unanimity polymorphisms is rather limited (as they all have bounded width), the collection of quantifiers with the corresponding closure property is much richer, including even CSP that are intractable. The pebble game gives a general method of proving inexpressibility that works for a wide variety of closure conditions. We were able to deploy it to prove that solvability of systems of equations over \(\mathbb{Z}\,/2\,\mathbb{Z}\) is not definable using only quantifiers closed under near-unanimity conditions. It would be interesting to use the pebble games we have defined to show undefinability with other collections of quantifiers closed under partial polymorphisms. Showing some class is not definable with quantifiers closed under partial Maltsev polymorphisms would be especially instructive. It would require using the pebble games with a construction that looks radically different from the CFI-like constructions most often used. This is because CFI constructions encode problems of solvability of equations over finite fields (or more generally finite rings), and all of these problems are Maltsev-closed.

2305.12764

Global Symmetries and Effective Potential of 2HDM in Orbit Space

We extend the framework of analyzing the 2HDM in its orbit space to study the one-loop effective potential before and after electroweak symmetry breaking. In this framework, we present a comprehensive analysis of global symmetries of the one-loop thermal effective potential in the 2HDM, demonstrating when the global symmetries of the tree-level 2HDM potential are broken by loop contributions. By introducing light-cone coordinates and generalizing the bilinear notation around the vacuum, we present a geometric view of the scalar mass matrix and on-shell renormalization conditions.

Qing-Hong Cao, Kun Cheng, Changlong Xu

2023-05-22T06:35:20Z

http://arxiv.org/abs/2305.12764v2

# Global Symmetries and Effective Potential of 2HDM ###### Abstract We extend the framework of analyzing the 2HDM in its orbit space to study the one-loop effective potential before and after electroweak symmetry breaking. In this framework, we present a comprehensive analysis of global symmetries of the one-loop thermal effective potential in the 2HDM, demonstrating when the global symmetries of the tree-level 2HDM potential are broken by loop contributions. By introducing light-cone coordinates and generalizing the bilinear notation around the vacuum, we present a geometric view of the scalar mass matrix and on-shell renormalization conditions. Introduction The Two-Higgs-Doublet Model (2HDM) is a simple extension of the SM [1]. It has received much attention for its potential to provide new sources of CP violation and strong first-order phase transition [2; 3; 4; 5; 6; 7; 8; 9; 10; 11]. The most general tree-level 2HDM scalar potential \[V= m_{11}^{2}\Phi_{1}^{\dagger}\Phi_{1}+m_{22}^{2}\Phi_{2}^{\dagger }\Phi_{2}-\left(m_{12}^{2}\Phi_{1}^{\dagger}\Phi_{2}+h.c.\right) \tag{1}\] \[+\frac{1}{2}\lambda_{1}\left(\Phi_{1}^{\dagger}\Phi_{1}\right)^{ 2}+\frac{1}{2}\lambda_{2}\left(\Phi_{2}^{\dagger}\Phi_{2}\right)^{2}+\lambda_ {3}(\Phi_{2}^{\dagger}\Phi_{2})(\Phi_{1}^{\dagger}\Phi_{1})+\lambda_{4}(\Phi_ {1}^{\dagger}\Phi_{2})(\Phi_{2}^{\dagger}\Phi_{1})\] \[+\left(\frac{1}{2}\lambda_{5}\left(\Phi_{1}^{\dagger}\Phi_{2} \right)^{2}+\lambda_{6}(\Phi_{1}^{\dagger}\Phi_{1})(\Phi_{1}^{\dagger}\Phi_{2} )+\lambda_{7}(\Phi_{2}^{\dagger}\Phi_{2})(\Phi_{1}^{\dagger}\Phi_{2})+h.c.\right)\] is parameterized by 14 real parameters. Here, \((m_{12}^{2},\lambda_{5},\lambda_{6},\lambda_{7})\) are in general complex while the others are real. The CP conserving 2HDM, also called real 2HDM, require all the parameters in Eq. (1) to be real with respective to a \(U(2)_{\Phi}\) basis transformation \(\Phi_{i}^{\prime}=U_{ij}\Phi_{j}\). Due to the field redefinition, the CP symmetry and other global symmetries of the potential are hard to determine from the parameters in Eq. (1) directly, and one of the most efficient ways to analyze these symmetries is to use the bilinear notation [12; 13; 14; 15] of the 2HDM. This method involves expressing the tree-level 2HDM potential in terms of orbits of \(SU(2)_{L}\) gauge transformations, which can be combined to form a four-vector, \[(K_{0},\vec{K})=K^{\mu}=\Phi_{i}^{\dagger}\sigma_{ij}^{\mu}\Phi_{j},\quad(\mu= 0,1,2,3). \tag{2}\] In this notation, the \(U(2)_{\Phi}\) basis transformation of the Higgs doublets corresponds to a \(SO(3)_{K}\) rotation of the three space-like components of this four-vector, while CP transformations correspond to improper rotations in these three dimensions [11]. The bilinear notation serves as a convenient tool for examining the symmetries and vacuum conditions of 2HDM. However, its applications are usually restricted to tree-level potential and global structures. In this work, we establish a complete framework for discussing the 2HDM potential, by extending the bilinear notation of the 2HDM to address the properties of physical fields around the vacuum and one-loop effective potential including renormalization. Recently, it is shown in Ref. [16; 17] that the bilinear notation can be extended to Yukawa couplings, making it possible to express the 2HDM effective potential including fermion loop contributions in the bilinear notation [16]. With this approach, we express the effective potential entirely as a function of gauge orbits, and systematically analyze the possible global symmetries of the effective potential. We generalize the bilinear notation to discuss physical fields after electroweak symmetry breaking (EWSB), and provide a geometrical description of scalar masses based on the light-cone coordinates in the orbit space. We demonstrate that the scalar mass matrix can be viewed as a distance matrix between two hyper-surfaces in the orbit space. Additionally, we translate the renormalization conditions in the field space into a set of geometrical conditions in the orbit space. Then numerous redundant renormalization conditions [18] that depend on the selection of background fields can be avoided. After the on-shell renormalization, we give a comprehensive effective field theory description of one-loop 2HDM effective potential for the first time. In the rest of this paper, we first review the global symmetries of the tree-level 2HDM potential in the bilinear notation in Section II, and then examine whether these symmetries are preserved by one-loop corrections in Section III. We explore the relationship between the orbit space and the field space around the vacuum after EWSB in Section IV, and we demonstrate how to carry out the on-shell renormalization in the orbit space in Section V. Finally, we conclude in Section VI. ## II Basis invariant description of global symmetry In this section, we introduce the basis transformations and CP transformations of the Higgs doublets, and the global symmetries of 2HDM potential. If a 2HDM potential is invariant under some basis or CP transformations, then it processes the corresponding symmetries. The bilinear notation is convenient to discuss these global transformations, because the basis or CP transformations simply corresponds to proper or improper rotations in the 3-dimensional space-like part of the orbit space [10; 11; 13; 14], and we refer this 3-dimensional subspace as \(\vec{K}\) space in the following. ### Global transformations and symmetries in the bilinear notation We first consider the \(U(2)_{\Phi}\) basis transformations \(\Phi_{i}\to U_{ij}\Phi_{j}\). It is straightforward to see from Eq. (2) that an \(SU(2)_{\Phi}\) basis transformation corresponds to a rotation in the \(\vec{K}\)-space, \[K_{0}\to K_{0},\ K_{a}\to R_{ab}(U)K_{b},\quad R_{ab}(U)=\frac{1}{2}{\rm tr} \left[U^{\dagger}\sigma_{a}U\sigma_{b}\right],\ a,b=1,2,3. \tag{3}\] Then we consider the CP transformation \(\Phi_{i}(t,\vec{x})\to\Phi_{i}^{*}(t,-\vec{x})\). Because the definition of the standard CP transformation \(\Phi_{i}\to\Phi_{i}^{*}\) will be changed if we choose another set of basis to describe the scalar fields, e.g. \(\Phi_{i}^{\prime}=U_{ij}\Phi_{j}\), the CP transformations in the 2HDM are extended as [2; 4; 19; 20] \[\text{GCP}:\ \ \Phi_{i}\to X_{ij}\Phi_{j}^{*}. \tag{4}\] Here, \(X_{ij}\) is an arbitrary unitary matrix, and such CP transformations are called generalized CP (GCP) transformations. By plugging the GCP transformation into Eq. (2), we find that \(\vec{K}\) transforms in an improper \(O(3)_{K}\) rotation \(\bar{R}(X)\). \[K_{0}\to K_{0},\ K_{a}\to\bar{R}_{ab}(X)K_{b},\quad\bar{R}(X)\equiv R(X)\, \text{diag}(1,-1,1). \tag{5}\] Here, the \(R(X)\) is defined in Eq. (3). Besides, for any GCP transformation, one can always find a basis \(\Phi_{i}\) so that \(X_{ij}\) is a real rotation matrix [19]. Therefore GCP transformations are often classified into three cases [21; 22]: \[\text{CP1}:\Phi_{1}\to\Phi_{1}^{*},\ \Phi_{2}\to\Phi_{2}^{*}, \tag{6}\] \[\text{CP2}:\Phi_{1}\to\Phi_{2}^{*},\ \Phi_{2}\to-\Phi_{1}^{*},\] (7) \[\text{CP3}:\ \begin{cases}\Phi_{1}\to\Phi_{1}^{*}\cos\theta+\Phi_{2}^{* }\sin\theta\\ \Phi_{2}\to-\Phi_{1}^{*}\sin\theta+\Phi_{2}^{*}\cos\theta\end{cases},\quad 0< \theta<\pi/2, \tag{8}\] where CP3 is the most general CP transformation while CP1 and CP2 are some special case with \(\text{CP1}^{2}=\text{CP2}^{2}=\mathbb{1}\), with respect to a global sign. After showing that the basis and GCP transformations correspond to \(O(3)_{K}\) rotations in the \(\vec{K}\)-space, we examine the symmetries conserving conditions on the 2HDM potential. The 2HDM potential in Eq. (1) can be written as a function of gauge orbits [12; 13; 14; 15], \[\begin{split} V&=\xi_{\mu}K^{\mu}+\eta_{\mu\nu}K^{ \mu}K^{\nu}\\ &=\xi_{0}K_{0}+\eta_{00}K_{0}^{2}+\vec{\xi}\cdot\vec{K}+2K_{0} \vec{\eta}\cdot\vec{K}+\vec{K}^{T}E\vec{K}.\end{split} \tag{9}\] Here, \(\vec{\xi}\) parametrizes the scalar quadratic couplings while \(E\) and \(\vec{\eta}\) parametrize the scalar quartic couplings. As discussed above, a GCP or basis transformation corresponds to some (improper) rotation \(R\) in the \(\vec{K}\)-space. If a tree-level potential is invariant under a rotation \(R\), i.e., \(V(K_{0},\vec{K})=V(K_{0},R\vec{K})\), its parameters should be invariant under the rotation \(R\), \[\vec{\xi}=R\vec{\xi},\quad\vec{\eta}=R\vec{\eta},\quad E=RER^{T}. \tag{10}\] A complete analysis of the symmetries of the 2HDM Lagrangian must include the Yukawa interaction, which reads as \[-\mathcal{L}_{\rm Yuk}=\bar{Q}_{L}y_{u,i}\tilde{\Phi}_{i}u_{R}+\bar{Q}_{L}y_{d,i }\Phi_{i}d_{R}+h.c.\,\quad i=1,2, \tag{11}\] for one generation of \(u\)-quark and \(d\)-quark. Note that \(y_{d,i}\) and \(y_{u,i}^{*}\) transform like \(\Phi_{i}\) under the \(SU(2)_{\Phi}\) basis redefinition. Although the Yukawa coupling terms in the Lagrangian cannot be expressed in bilinear notation directly, it is shown that the bilinear notation can still be extended to discuss whether the Yukawa couplings break the global symmetries of the scalar potential [16]. This is done by projecting the Yukawa couplings into orbit space, and defining covariant vectors in the dual space of \(K^{\mu}\) in terms of the Yukawa couplings as \[Y_{u}^{\mu}=y_{u,i}^{*}\sigma_{ij}^{\mu}y_{u,j},\quad Y_{d}^{\mu}=y_{d,i} \sigma_{ij}^{\mu}y_{d,j}^{*}. \tag{12}\] In order to make sure that \(\mathcal{L}_{\rm Yuk}\) is invariant under the basis transformation \(\Phi_{i}\to U_{ij}\Phi_{j}\) or the CP transformation in the scalar sector \(\Phi_{i}\to X_{ij}\Phi_{j}^{*}\), the vector \(\vec{Y}\) projected by the Yukawa couplings should satisfy \[\vec{Y}=R(U)\vec{Y}\quad\text{or}\quad\vec{Y}=\bar{R}(X)\vec{Y}. \tag{13}\] ### Examples of global symmetries Next we show how to discuss some special symmetries that is widely considered in the orbit space. We start with two characteristic examples, the CP1 symmetry and the \(Z_{2}\) symmetry. The CP symmetry is often introduced to the potential because large CP violations are prohibited by experiments. From Eq. (5), the CP1 transformation \(\Phi_{i}\to\Phi_{i}^{*}\) corresponds to a mirror reflection in the \(\vec{K}\)-space. The \(Z_{2}\) symmetry is introduced to prevent the flavor changing neutral interactions,1 and a softly broken \(Z_{2}\) symmetry is often considered. From Eq. 3, the \(Z_{2}\) transformation \(\Phi_{1}\to-\Phi_{1}\) corresponds to a 2-dimensional rotation of \(\pi\) in the \(\vec{K}\) space. Footnote 1: For different types of \(Z_{2}\) charge assignments in the orbit space, see Table II in Appendix A.2. Whether a 2HDM is invariant under the CP1 or \(Z_{2}\) transformation can be understood from the geometrical profile of parameter tensor and vectors, as shown in Eqs. (10) and (13). Without loss of generality, we use an ellipsoid to visualize the \(3\times \(E\) which possesses at least three \(C_{2}\) axis (principal axis) and three symmetry planes, and we illustrate these two examples in Fig. 1. The CP1 symmetric 2HDM potential satisfies a mirror reflection symmetry in the \(\vec{K}\) space, requiring all the parameter vectors to lie on the same reflection plane of \(E\). The \(Z_{2}\) symmetric potential is invariant under a rotation of \(\pi\) in the \(\vec{K}\) space. Hence the parameter vectors should point to the same principal axis of \(E\). As for the softly broken \(Z_{2}\) symmetry, the quadratic term \(\vec{\xi}\) is allowed to break the \(Z_{2}\) symmetry as in Fig. 1. Following Refs. [21; 22], we list other global symmetries in scalar family space by different geometric profile of the scalar potential in Table 1. The \(U(1)_{a}\) transformation \(\Phi_{1}\to e^{i\theta}\Phi_{1}\) corresponds to a rotation along a certain axis, the CP2 transformation corresponds to a point reflection, and the CP3 transformation corresponds to a point reflection followed by an additional rotation of \(0\sim\pi\). The geometric profiles show the hierarchy chain of those global symmetries clearly, \[\text{CP1}<Z_{2}<\left\{\begin{aligned} &\text{CP2}\\ & U(1)_{a}\end{aligned}\right\}<\text{CP3}<U(2), \tag{14}\] i.e., a \(Z_{2}\) symmetric tree-level 2HDM scalar potential must satisfy CP1 symmetry and likewise. For GCP properties of tree-level 2HDM scalar potential, CP2 and CP3 symmetric conditions are more strict than CP1.2 Besides, neither CP2 nor CP3 symmetry can be still Figure 1: Illustration figure of parameter vectors and tensor of the 2HDM potential that obeys CP1 symmetry (left) or softly broken \(Z_{2}\) symmetry (right). The ellipsoid denotes the tensor \(E\) and black dashed lines denote its three principal axes. Red and blue arrows denote the directions of \(\vec{\eta}\) and \(\vec{\xi}\) respectively. preserved after the Higgs field developed a non-vanishing vacuum expectation value. Therefore, we will only discuss CP1 conserving (CPC) conditions and denote CP1 as CP in the following. Footnote 1: The CP-violating vacuum expectation value is defined as \(\Phi_{i}=\Phi_{i}+\Phi_{i}\), where \(\Phi_{i}\) is the CP-violating vacuum expectation value. ## III Effective potential and thermal correction The global symmetries of thermal effective potential are important in the study of the vacuum structure and CP violation in the early universe. The use of bilinear notation simplifies, from a geometrical perspective, the analysis of the global symmetries of the effective potential. In this section, we employ the bilinear notation to evaluate the effective potential and discuss its global symmetries. The thermal effective potential of the 2HDM is written as \[V_{\text{eff}}(T)=V_{\text{tree}}+V_{\text{CW}}+V_{T}+V_{\text{ daisy}}, \tag{15}\] where \(V_{\text{tree}}\) is the tree-level potential, \(V_{\text{CW}}\) is the one-loop Coleman-Weinberg potential at zero temperature [28], and \(V_{T}+V_{\text{daisy}}\) are the thermal corrections at finite temperature. Using the background field method, the one-loop Coleman-Weinberg potential calculated in \begin{table} \begin{tabular}{|c|c|c|c|} \hline Symmetry & Transformation & Vector \(\vec{\eta}\), \(\vec{\xi}\) and \(\vec{Y}\) & Tensor \\ \hline U(2) & \(\Phi_{i}\to U_{ij}\Phi_{j}\) & 0 & spherical \\ \hline CP3 & Eq. (8) & 0 & \(e_{1}=e_{2}\) \\ \hline CP2 & Eq. (7) & 0 & - \\ \hline \(U(1)_{a}\) & \(\Phi_{1}\to e^{i\theta}\Phi_{1}\) & collinear with \(\vec{e}_{3}\) & \(e_{1}=e_{2}\) \\ \hline \(Z_{2}\) & \(\Phi_{1,2}\rightarrow\pm\Phi_{1,2}\) & collinear with an axes \(\vec{e}_{i}\) & - \\ \hline CP1 & \(\Phi_{i}\rightarrow\Phi_{i}^{*}\) & orthogonal to an axes \(\vec{e}_{i}\) & - \\ \hline \end{tabular} \end{table} Table 1: Global symmetry and their geometric profile of parameter vectors and tensor \(E\). \(\vec{e}_{i}\) are direction of three eigenvectors of \(E\), and \(e_{i}\) denotes the three corresponding eigenvalues. Landau gauge under the \(\overline{\text{MS}}\) scheme is \[\begin{split} V_{\text{CW}}(\phi_{c})&=\frac{1}{2} \text{Tr}\int\frac{d^{4}p_{E}}{2\pi^{4}}\ln\left[p_{E}^{2}+\textbf{M}^{2}(\phi_ {c})\right]\\ &=\frac{1}{64\pi^{2}}\sum_{i}n_{i}m_{i}^{4}(\phi_{c})\left[\ln \frac{m_{i}^{2}(\phi_{c})}{\mu^{2}}-c_{i}\right].\end{split} \tag{16}\] Here, \(p_{E}=(-ip^{0},\vec{p})\), \(\textbf{M}^{2}\) is the mass matrix of scalar or fermion in the loop and \(\textbf{Tr}\) traces over the dimension of mass matrix, \(m_{i}^{2}\) is the eigenvalue of the \(\textbf{M}^{2}\) for the field \(i\), and \(n_{i}\) is the degree of freedom of the field \(i\). The constant \(c_{i}\) equals to \(5/6\) for gauge bosons and \(3/2\) for others. The effective potential of the 2HDM has been extensively studied in the literature [6; 7; 8; 9]. Typically, only the neutral or CP-even components of the Higgs boson doublets are treated as background fields, which breaks the \(SU(2)_{L}\) invariance explicitly. Consequently, the bilinear notation cannot be applied to study \(V_{\text{eff}}(\phi_{c})\) directly. In order to analyze the global symmetries of the effective potential using bilinear notation, a global \(SU(2)_{L}\) invariance must be preserved in the calculation [16], which means that the masses in Eq. (16) need to be evaluated in a \(SU(2)_{L}\) invariant way. To achieve this, we treat all the components of the Higgs boson doublets \(\Phi_{i}\)'s, \[\Phi_{i}=\begin{pmatrix}\phi_{i\uparrow}\\ \phi_{i\downarrow}\end{pmatrix},\quad i=1,2, \tag{17}\] as background fields, and \(K^{\mu}\) should be understood as bilinear forms of background fields in this section. ### Symmetries of Coleman-Weinberg potential We first consider the zero temperature effective potential by calculating the contributions from gauge boson loop, fermion loop and scalar loop to the Coleman-Weinberg potential respectively. Contributions from gauge boson loop.The masses of gauge bosons arise from the kinetic term \(|D_{\mu}\Phi_{i}|^{2}\) with \(D^{\mu}\Phi_{i}=(\partial^{\mu}+i\frac{g}{2}\sigma_{a}W_{a}^{\mu}+i\frac{g^{ \prime}}{2}B^{\mu})\Phi_{i}\), where \(i=1,2\). Expanding the covariant derivative term directly yields the gauge boson mass term \[\begin{split}\frac{1}{4}\Phi_{i}^{\dagger}(g\sigma_{a}W_{a}+g^{ \prime}B)^{2}\Phi_{i}=&\frac{1}{4}\Phi_{i}^{\dagger}(g^{\prime 2}B^{2}+2gg^{ \prime}BW_{a}\sigma_{a}+g^{2}\sigma_{a}\sigma_{b}W_{a}W_{b})\Phi_{i}\\ =&\frac{1}{4}\Phi_{i}^{\dagger}(g^{\prime 2}B^{2}+2gg^{ \prime}BW_{a}\sigma_{a}+g^{2}\sigma_{\{a}\sigma_{b\}}W_{a}W_{b})\Phi_{i}\\ =&\frac{\Phi_{i}^{\dagger}\Phi_{i}}{4}(g^{\prime 2}B^{2}+g^{ 2}W_{a}W_{a})+\frac{\Phi_{i}^{\dagger}\sigma_{a}\Phi_{i}}{2}gg^{\prime}BW_{a}. \end{split} \tag{18}\] Then the gauge boson mass matrix in basis \(\vec{G}=(W_{1},W_{2},W_{3},B)\) is \[\mathbf{M}_{G}^{2}(\Phi_{i})=\frac{\partial^{2}L}{\partial\vec{G}\partial\vec {G}}=\frac{g^{2}}{4}\begin{pmatrix}\Phi_{i}^{\dagger}\Phi_{i}&0&0&t_{W}\Phi_{i }^{\dagger}\sigma_{1}\Phi_{i}\\ 0&\Phi_{i}^{\dagger}\Phi_{i}&0&t_{W}\Phi_{i}^{\dagger}\sigma_{2}\Phi_{i}\\ 0&0&\Phi_{i}^{\dagger}\Phi_{i}&t_{W}\Phi_{i}^{\dagger}\sigma_{3}\Phi_{i}\\ t_{W}\Phi_{i}^{\dagger}\sigma_{1}\Phi_{i}&t_{W}\Phi_{i}^{\dagger}\sigma_{2} \Phi_{i}&t_{W}\Phi_{i}^{\dagger}\sigma_{3}\Phi_{i}&t_{W}^{2}\Phi_{i}^{\dagger} \Phi_{i}\end{pmatrix} \tag{19}\] where \(t_{W}=\tan\theta_{W}=g^{\prime}/g\). For a matrix with the shape of Eq. (19), its eigenvalues are \[\text{Eigenvalues}\begin{pmatrix}e&&a\\ &e&&b\\ &&e&c\\ a&b&c&d\end{pmatrix}=(e,e,\frac{d+e\pm\sqrt{4(a^{2}+b^{2}+c^{2})+(d-e)^{2}}}{ 2}). \tag{20}\] With the help of the Fierz identities, \[(\Phi_{i}^{\dagger}\sigma_{a}\Phi_{i})(\Phi_{j}^{\dagger}\sigma_{a}\Phi_{j})=( \Phi_{1}^{\dagger}\Phi_{1}-\Phi_{2}^{\dagger}\Phi_{2})^{2}+4(\Phi_{1}^{\dagger }\Phi_{2})(\Phi_{2}^{\dagger}\Phi_{1})=|\vec{K}|^{2}, \tag{21}\] we present four eigenvalues of the gauge boson mass matrix, \[\begin{split} m_{W^{\pm}}^{2}&=\frac{g^{2}}{4}K_{0},\\ m_{Z}^{2}&=\frac{g^{2}}{8}\left((1+t_{W}^{2})K_{0}+ \sqrt{4t_{W}^{2}|\vec{K}|^{2}+(t_{W}^{2}-1)^{2}K_{0}^{2}}\right),\\ m_{\gamma}^{2}&=\frac{g^{2}}{8}\left((1+t_{W}^{2})K_{0}- \sqrt{4t_{W}^{2}|\vec{K}|^{2}+(t_{W}^{2}-1)^{2}K_{0}^{2}}\right).\end{split} \tag{22}\] Notice that there is a massless photon when the vacuum is neutral, i.e., \(K_{0}=|\vec{K}|\). By plugging Eq. (22) into Eq. (16), we find that the gauge boson loop contributions to the Coleman-Weinberg potential, \(V_{\text{CW}}^{(G)}=V_{\text{CW}}^{(G)}(K_{0},|\vec{K}|)\), is spherically symmetric and preserve any rotational symmetry in the \(\vec{K}\) space, i.e., \[V_{\text{CW}}^{(G)}(K_{0},\vec{K})=V_{\text{CW}}^{(G)}(K_{0},R\vec{K}),\qquad R \in O(3).\] Contributions from the quark loop.Typically, only the contribution from the heaviest quark needs to be included in the effective potential. However, we include both the top and bottom quarks in our calculation to ensure an explicit \(SU(2)_{L}\) invariance. The top and bottom quark masses mix due to the presence of charged background fields, and the fermion mass matrix given by \(-\partial^{2}\mathcal{L}/\partial\bar{\psi}_{L}^{i}\partial\psi_{R}^{j}\) is \[(\bar{t}_{L},\bar{b}_{L})\mathbf{M}_{F}\begin{pmatrix}t_{R}\\ b_{R}\end{pmatrix},\quad\mathbf{M}_{F}=\begin{pmatrix}y_{ii}\phi_{i\downarrow}^ {*}&y_{ib}\phi_{i\uparrow}\\ -y_{it}\phi_{i\uparrow}^{*}&y_{ib}\phi_{i\downarrow}\end{pmatrix}. \tag{23}\] We obtain the fermion masses after singular decomposition, \[L^{-1}\mathbf{M}_{F}R=\begin{pmatrix}m_{t}\\ &m_{b}\end{pmatrix},\quad m_{t/b}^{2}=\frac{B\pm\sqrt{B^{2}+C}}{2}, \tag{24}\] where, with the help of vector \(\vec{Y}\) defined in Eq. (12), \(B\) and \(C\) can be written as \(SO(3)_{K}\) basis invariant forms as follows: \[\begin{split} B=&\frac{1}{2}(Y_{t0}+Y_{b0})K_{0}+\frac{1}{2}( \vec{Y}_{t}+\vec{Y}_{b})\cdot\vec{K},\\ C=&-\frac{1}{2}(Y_{t}\cdot Y_{b})K_{0}^{2}-K_{0}(Y_{t0}\vec{Y}_{b}+Y_{b 0}\vec{Y}_{t})\cdot\vec{K}\\ &+\frac{1}{2}\vec{K}\cdot(\vec{Y}_{t}\cdot\vec{Y}_{b}-Y_{t0}Y_{b0}- \vec{Y}_{t}\otimes\vec{Y}_{b}-\vec{Y}_{b}\otimes\vec{Y}_{t})\cdot\vec{K}.\end{split} \tag{25}\] The masses can be simplified in the case that the Yukawa couplings exhibit a large hierarchy; for example, when \(y_{t}\gg y_{b}\), only the top quark mass \(m_{t}^{2}(K)=(Y_{t0}K_{0}+\vec{Y}_{t}\cdot\vec{K})/4\) needs to be considered. Equations (24) and (25) show that the symmetry of \(V_{\text{CW}}^{(F)}\) is completely determined by the direction of vector \(\vec{Y}\). When the vector \(\vec{Y}\) is invariant under the rotation, i.e., \(\vec{Y}_{t/b}=R\vec{Y}_{t/b}\) for \(R\in O(3)\), \[V_{\text{CW}}^{(F)}(K_{0},\vec{K})=V_{\text{CW}}^{(F)}(K_{0},R\vec{K}).\] When \(\vec{Y}_{t/b}\neq R\vec{Y}_{t/b}\), \[V_{\text{CW}}^{(F)}(K_{0},\vec{K})\neq V_{\text{CW}}^{(F)}(K_{0},R\vec{K}).\] Therefore, whether the fermion loop contribution to \(V_{\text{CW}}\) breaks the global symmetry of the tree-level potential depends on the pattern of Yukawa couplings. Contributions from the scalar loop.The calculation of \(V_{\text{CW}}^{(S)}(K_{0},\vec{K})\) can be performed straightforwardly from Eq. (16), in which the mass matrix of scalars is given by \[\mathbf{M}_{S}^{2}(\varphi)_{ab}=\frac{\delta^{2}V}{\delta\varphi_{a}\delta \varphi_{b}}, \tag{26}\] where \(\varphi_{a}\) are real vectors in the 8-dimensional field space. Though \({\bf M}_{S}^{2}\) cannot be diagonalized analytically, we still find a way to investigate the global symmetries of \(V_{\rm CW}^{(G)}\). We firstly employ the notations in Ref. [30], where the components of \(\varphi_{a}\) are ordered as \[\varphi_{a}^{T}=\left({\rm Re}\,\phi_{1\uparrow},{\rm Im}\,\phi_{1\uparrow},{ \rm Re}\,\phi_{2\uparrow},{\rm Im}\,\phi_{2\uparrow},{\rm Re}\,\phi_{1\downarrow },{\rm Im}\,\phi_{1\downarrow},{\rm Re}\,\phi_{2\downarrow},{\rm Im}\,\phi_{2 \downarrow}\right), \tag{27}\] and \(\varphi_{a}\) is related to the bilinear form by \(K^{\mu}=\varphi_{a}\Sigma_{ab}^{\mu}\varphi_{b}\). The \(8\times 8\) matrices \(\Sigma^{\mu}\) are defined as \[\Sigma^{\mu}=\Sigma_{4}^{\mu}\oplus\Sigma_{4}^{\mu},\quad\Sigma_{4}^{0}= \mathbb{1}_{4},\quad\Sigma_{4}^{1}=\begin{pmatrix}0&\mathbb{1}_{2}\\ \mathbb{1}_{2}&0\end{pmatrix},\quad\Sigma_{4}^{2}=\begin{pmatrix}0&\mathbb{1}_ {2}\\ -\mathbb{1}_{2}&0\end{pmatrix},\quad\Sigma_{4}^{3}=\begin{pmatrix}\mathbb{1}_{2 }&0\\ 0&-\mathbb{1}_{2}\end{pmatrix}, \tag{28}\] where \(\mathbb{1}_{d}\) is the \(d\times d\) identity matrix and \(\mathbb{1}_{2}\equiv(\begin{smallmatrix}0&1\\ -1&0\end{smallmatrix})\). The \(V_{\rm CW}^{(S)}\) can be expanded in the powers of \({\bf M}_{S}\)[31], \[\begin{split} V_{\rm CW}^{(S)}&=\frac{1}{2}{\bf Tr}\int \frac{d^{4}p_{E}}{2\pi^{4}}\ln\left[p_{E}^{2}+{\bf M}_{S}^{2}\right]\\ &=\frac{1}{2}\int\frac{d^{4}p_{E}}{2\pi^{4}}\left[{\bf Tr}\sum_{n=1 }^{\infty}\frac{1}{n}\left(-\frac{{\bf M}_{S}^{2}}{p_{E}^{2}}\right)^{n}+\ln p _{E}^{2}\right],\end{split} \tag{29}\] where \({\bf Tr}\) stands for taking a trace over the 8-dimensional field space. For example, the leading power is \[{\bf Tr}({\bf M}_{S}^{2})=\left(20\eta_{00}+4\,{\rm tr}(E)\right)K_{0}+24 \vec{K}\cdot\vec{\eta}+8\xi_{0}, \tag{30}\] which is consistent with Ref. [30]. We show that all the traces \({\bf Tr}({\bf M}_{S}^{2n})\) in Eq. (29) are functions of gauge orbits \(K^{\mu}\), and the complete calculations are deferred to Appendix B. Here, we present the final calculation result expressed in the bilinear notation as \[V_{\rm CW}^{(S)}={\cal F}\left(S_{p}^{\mu\nu},\eta^{\mu\nu}\right). \tag{31}\] The function \({\cal F}\) only depends on the trace of the inner products of \(S_{p}^{\mu\nu}\) and \(\eta^{\mu\nu}\), and \(S_{p}^{\mu\nu}\) is defined as \[S_{p}^{\mu\nu}=F(p)^{\mu}K^{\nu}+F(p)^{\nu}K^{\mu}-g^{\mu\nu}(F(p)K). \tag{32}\] Here, \(F(p)_{\mu}\) is a function of \(K^{\mu}\) that depends on the integer \(p\), \[F(p)_{0}\equiv\sum_{k=0}^{p/2}C_{p}^{2k}(A_{0})^{p-2k}|\vec{A}|^{2k}, \tag{33}\] \[\vec{F}(p)\equiv\left\{\begin{aligned} -\sum_{k=0}^{(p-1)/2}C_{p}^{2k+1}(A_{0})^{p-2k-1}| \vec{A}|^{2k}\vec{A}&(p\neq 0),\\ 0&(p=0),\end{aligned}\right. \tag{34}\] where \((A_{0},\vec{A})=A_{\mu}=2\eta_{\mu\nu}K^{\nu}+\xi_{\mu}\) and \(C_{p}^{k}\) is the binomial coefficient. Notice that the global symmetries are determined only by the 3-dimension vector \(\vec{A}\). Upon expressing \(V_{\rm CW}^{(S)}\) as a function in the orbit space, we find that the tensor structures in \(V_{\rm CW}^{(S)}\) are constructed entirely by tree level parameter tensors \(\eta^{\mu\nu}\) and \(\xi^{\mu}\), i.e., no new tensor structure appears, therefore, the rotation symmetries of \(V_{\rm CW}^{(S)}\) in the \(\vec{K}\)-space are determined by the tree-level parameter tensors. If the tree-level potential is invariant under a rotation \(R\in O(3)\) in the \(\vec{K}\)-space, i.e., \(V_{\rm tree}(K_{0},\vec{K})=V_{\rm tree}(K_{0},R\vec{K})\), the scalar loop contribution \(V_{\rm CW}^{(S)}\) also preserves the rotation invariance, \[V_{\rm CW}^{(S)}(K_{0},\vec{K})=V_{\rm CW}^{(S)}(K_{0},R\vec{K}).\] ### Symmetries of thermal potential As for the finite temperature corrections in Eq. (15), \(V_{T}\) stands for the contribution from one-loop diagrams, and \(V_{\rm daisy}\) denotes the correction from higher loop Daisy diagrams [31]. The one-loop correction \(V_{T}\) is given by \[V_{T}=\sum_{i}n_{i}\frac{T^{4}}{2\pi^{2}}J_{B/F}\left(m_{i}^{2}/T^{2}\right), \tag{35}\] where the thermal bosonic function \(J_{B}\) and fermionic function \(J_{F}\) are \[J_{B}\left(m^{2}/T^{2}\right)= -\frac{\pi^{4}}{45}+\frac{\pi^{2}}{12}\frac{m^{2}}{T^{2}}-\frac{ \pi}{6}\left(\frac{m^{2}}{T^{2}}\right)^{3/2}-\frac{1}{32}\frac{m^{4}}{T^{4}} \log\frac{m^{2}}{a_{b}T^{2}}\] \[-2\pi^{7/2}\sum_{\ell=1}^{\infty}(-1)^{\ell}\frac{\zeta(2\ell+1) }{(\ell+1)!}\Gamma\left(\ell+\frac{1}{2}\right)\left(\frac{m^{2}}{4\pi^{2}T^ {2}}\right)^{\ell+2}, \tag{36}\] \[J_{F}\left(m^{2}/T^{2}\right)= \frac{7\pi^{4}}{360}-\frac{\pi^{2}}{24}\frac{m^{2}}{T^{2}}-\frac{ 1}{32}\frac{m^{4}}{T^{4}}\log\frac{m^{2}}{a_{f}T^{2}}\] \[-\frac{\pi^{7/2}}{4}\sum_{\ell=1}^{\infty}(-1)^{\ell}\frac{\zeta(2 \ell+1)}{(\ell+1)!}\left(1-2^{-2\ell-1}\right)\Gamma\left(\ell+\frac{1}{2} \right)\left(\frac{m^{2}}{\pi^{2}T^{2}}\right)^{\ell+2}. \tag{37}\] Here, \(a_{b}=16a_{f}=16\pi^{2}e^{3/2-2\gamma_{E}}\), and \(\zeta\) is the Riemann-\(\zeta\) function. The leading \(T\)-dependent terms of \(J_{B/F}\) are given by the mass-square terms, \[J_{B}=\frac{\pi^{2}}{12}\frac{m^{2}}{T^{2}}+O(T^{-4}),\quad J_{F}=-\frac{\pi^{ 2}}{24}\frac{m^{2}}{T^{2}}+O(T^{-4}), \tag{38}\] where the background-field-independent terms are dropped. By collecting the results in Eqs. (22), (24) and (30), we obtain the leading contributions from gauge boson loops, fermion loops, and scalar loops to \(V_{T}\) as follows: \[V_{T}^{(G)} \approx\frac{g^{2}T^{2}}{32}(3+t_{W}^{2})K_{0}, \tag{39}\] \[V_{T}^{(F)} \approx-\frac{T^{2}}{8}\left[(Y_{t0}+Y_{b0})K_{0}+(\vec{Y}_{t}+ \vec{Y}_{b})\cdot\vec{K})\right],\] (40) \[V_{T}^{(S)} \approx\frac{T^{2}}{6}\left[\left(5\eta_{00}+\text{tr}(E)\right) K_{0}+6\vec{K}\cdot\vec{\eta}+2\xi_{0}\right]. \tag{41}\] We find that the corrections from Eqs. (39)-(41) to the tree-level potential is equivalent to shifting the quadratic couplings \(\xi_{\mu}\) in the orbit space, i.e., \[\xi_{0} \rightarrow\xi_{0}+T^{2}c_{T0},\] \[\vec{\xi} \rightarrow\vec{\xi}+T^{2}\vec{c}_{T}, \tag{42}\] where \[c_{T0} =\frac{g^{2}}{32}(3+t_{W}^{2})-\frac{Y_{t0}+Y_{b0}}{8}+\frac{5 \eta_{00}+\text{tr}(E)}{6},\] \[\vec{c}_{T} =\frac{1}{8}\left[8\vec{\eta}-\vec{Y}_{t}-\vec{Y}_{b}\right]. \tag{43}\] The direction of \(\vec{\xi}\) is shifted by the quartic couplings \(\vec{\eta}\) and Yukawa interactions \(\vec{Y}\) from thermal corrections. At a sufficient high temperature with \(T^{2}\gg|\vec{\xi}|/|\vec{c}_{T}|\), the direction of shifted \(\vec{\xi}\) is aligned along the direction of \(\vec{c}_{T}\). As a result, the symmetries of thermal effective potential under the basis transformation and CP transformation are determined by \(\vec{c}_{T}\). At high temperatures, the contribution from higher loop Daisy diagrams \(V_{\text{daisy}}\) is comparable with \(V_{T}\), and it is given by [31] \[V_{\text{daisy}}=-\frac{T}{12\pi}\sum_{i=\text{bosons}}n_{i}\left[\mathcal{M} _{i}^{3}(\phi_{c},T)-m_{i}^{3}(\phi_{c})\right]. \tag{44}\] Here, the \(\mathcal{M}_{i}(\phi_{c},T)\) are thermal corrected masses calculated from \(V_{\text{tree}}+V_{T}\), which is obtained from the tree level potential by parameter shifting \(\xi^{\mu}\rightarrow\xi^{\mu}+T^{2}c_{T}^{\mu}\). Therefore, the \(T\)-dependent terms in \(\mathcal{M}_{i}(\phi_{c},T)\) are in the form of \(T^{2}c_{T}^{\mu}\). As \(c_{T}^{0}\) plays no role in global transformations, the behavior of \(V_{\text{daisy}}\) under the \(O(3)_{K}\) transformation depends only on \(\vec{c}_{T}\). After understanding the behavior of \(V_{\text{CW}}\), \(V_{T}\) and \(V_{\text{daisy}}\) under the \(O(3)_{K}\) transformation, we are ready to discuss whether a global symmetry preserved by the tree-level potential will be violated by the loop corrections. Consider a tree-level potential that processes the symmetry of a basis or CP transformation, then the potential is invariant under a rotation \(R\) in the \(\vec{K}\)-space, \(V_{\rm tree}(K_{0},\vec{K})=V_{\rm tree}(K_{0},R\vec{K})\), and its parameters satisfy \[\vec{\xi}=R\vec{\xi},\quad\vec{\eta}=R\vec{\eta},\quad E=RER^{T}. \tag{45}\] The _only_ quantum correction that may violate the symmetry is the contribution from fermion loops. The global symmetry is maintained in effective potential if and only if all the Yukawa couplings are invariant under \(R\), i.e., \(\vec{Y}=R\vec{Y}\). If the symmetry is softly broken at tree level, i.e., only the scalar quadratic coupling \(\vec{\xi}\neq R\vec{\xi}\) violates the symmetry while other conditions in Eq. (45) are preserved, then the symmetry violation effect from soft terms tend to be suppressed at high temperature. This is because the leading thermal corrections shift the scalar quadratic couplings \(\vec{\xi}\) with Yukawa coupling \(\vec{Y}\) and scalar quartic couplings \(\vec{\eta}\), and both \(\vec{Y}\) and \(\vec{\eta}\) preserve the symmetry. Another noteworthy example is the custodial symmetry. In the orbit space, the custodial symmetry of the 2HDM does not correspond to a rotation symmetry but a shift symmetry [32]. As the effective potential is not invariant under any shift symmetry in the orbit space, the custodial symmetry of 2HDM is bound to be broken by the effective potential. ## IV Bilinear notation after EWSB In this section, we extend the bilinear notation to discuss EWSB and physical fields. There are two reasons to discuss the EWSB in the bilinear notation. Firstly, a global symmetry exhibited by the potential, as shown in Table 1, can be broken spontaneously after the potential develops a vacuum. For example, consider the CP symmetry. Even if the potential is explicitly CP conserving, \(V(\Phi_{1},\Phi_{2})|_{\Phi_{i}\rightarrow\Phi_{i}^{*}}=V(\Phi_{1},\Phi_{2})\), the physical fields, which are fluctuations around the vacuum, may still break CP symmetry after EWSB if the vacuum has an irremovable CP phase as follows: \[\langle\Phi_{1}\rangle=\begin{pmatrix}0\\ v_{1}\end{pmatrix},\quad\langle\Phi_{2}\rangle=\begin{pmatrix}0\\ v_{2}e^{i\delta}\end{pmatrix}. \tag{46}\] This is called spontaneous CP violation (SCPV). In the bilinear notation, the SCPV happens when the potential but not the vacuum is invariant under a mirror reflection. In this case, there are two degenerate vacuums related by a CP transformation as in Fig. 2. After analyzing the vacuum conditions in the orbit space, we can easily determine whether the CP symmetry or other global symmetry is spontaneously broken. Secondly, exploring the physical fields after the EWSB is necessary for performing on-shell renormalization. The renormalized effective potential can be expressed fully in the bilinear notation if we can perform the on-shell renormalization in the orbit space. For that, we examine the vacuum structures in the orbit space and investigate the relations between the field space and orbit space. Furthermore, we demonstrate that the mass matrix of the physical neutral scalars corresponds to a geometric structure in the orbit space, making it convenient to handle the mass spectrum and on-shell renormalization. ### Vacuum condition We start with the vacuum conditions of \(V(K^{\mu})\), where \(V(K^{\mu})\) represents the tree-level or effective potential in the orbit space. Figure 3 displays the light-cone in the orbit space, and the light-cone is a hyper-surface defined by \(K_{0}=|\vec{K}|\). The orbit space inside the forward light-cone \(LC^{+}\) is the physical region, satisfying \(K_{0}\geq|\vec{K}|\)[12; 13; 14; 15]. A neutral vacuum expectation value requires the minimum of the potential, denoted as \(K_{v}^{\mu}\), to lie on the \(LC^{+}\), i.e., \(K_{v,0}=|\vec{K}_{v}|\)[12; 13; 14; 15]. Therefore, \(K_{v}^{\mu}\) is a conditional minimum of \(V(K^{\mu})\) on the \(LC^{+}\). The vacuum of the potential \(V(K^{\mu})\) is solved by minimizing the function \(V_{u}(K^{\mu})=V(K^{\mu})-uL(K^{\mu})\), where \(u\) is a Lagrange multiplier and \(L(K^{\mu})=0\) is the light-cone condition Figure 2: Illustration figures of tree-level parameters for an SCPV potential. Here \(\vec{K}_{v}\) and \(\vec{K}_{v^{\prime}}\) are a pair of degenerated vacuum expectation values that are related by a mirror reflection, and the reflection plane spanned by \(\vec{\xi}\) and \(\vec{\eta}\). with \(L(K^{\mu})\) defined as \[L(K_{0},\vec{K})=K_{0}^{2}-|\vec{K}|^{2}=4K_{+}K_{-}-|\vec{K}_{T}|^{2}. \tag{47}\] Here, for the convenience, we introduce the light-cone coordinates \[K_{\pm}\equiv\frac{(K_{0}\pm K_{3})}{2},\qquad\vec{K}_{\rm T}\equiv(K_{1},K_{2}), \tag{48}\] which are defined after rotating the vacuum along the \(K_{3}\) direction, i.e., \(K_{v}^{\mu}=\frac{v^{2}}{2}(1,0,0,1)^{T}\). The solution of the conditional minimum satisfies \[\left.\frac{\partial V}{\partial K_{-}}\right|_{K_{v}}=2v^{2}u>0,\quad\left. \frac{\partial V}{\partial K_{+}}\right|_{K_{v}}=0,\quad\left.\frac{\partial V }{\partial\vec{K}_{\rm T}}\right|_{K_{v}}=0. \tag{49}\] Note that we require \(\frac{\partial V}{\partial K_{-}}>0\) to ensure no global minimum inside the light-cone to avoid a charged vacuum. In addition to the conditions in Eq. (49), we need to make sure that \(K_{v}\) is a minimal point rather than a saddle point. In the 4-dimensional orbit space, \(K_{v}\) is the tangent point of \(LC^{+}\) and an equipotential surface \(\mathcal{M}_{\rm vev}\) defined by \(V(K^{\mu})=V(K_{v}^{\mu})\)[14], and the normal direction of their tangent space is \(K_{-}\), as shown in Fig. 3. Therefore, the requirement that \(K_{v}\) is not a saddle point indicates that \(\mathcal{M}_{\rm vev}\) must be outside of the \(LC^{+}\). Equivalently, the distance between \(LC^{+}\) and \(\mathcal{M}_{\rm vev}\) is non-negative. Expanding the distance between \(LC^{+}\) and \(\mathcal{M}_{\rm vev}\) at their tangent point yields \[\delta h=(\delta K_{+},\delta\vec{K}_{\rm T})\ {\bf M}_{\rm dist}^{2}\begin{pmatrix} \delta K_{+}\\ \delta\vec{K}_{\rm T}\end{pmatrix},\quad{\bf M}_{\rm dist}^{2}=\frac{1}{ \partial V/\partial K_{-}}\begin{pmatrix}\frac{\partial^{2}V_{u}}{\partial K _{+}^{2}}&\frac{\partial^{2}V_{u}}{\partial K_{+}\partial\vec{K}_{\rm T}}\\ \frac{\partial^{2}V_{u}}{\partial K_{+}\partial\vec{K}_{\rm T}}&\frac{ \partial^{2}V_{u}}{\partial\vec{K}_{\rm T}^{2}}\end{pmatrix}. \tag{50}\] Figure 3: Vacuum expectation value and light-cone coordinates in the orbit space. The yellow surface denotes \(LC^{+}\) and the green denotes the equipotential surface \(\mathcal{M}_{\rm vev}\). \(K_{v}\) is the tangent point of these 3-dimensional hyper-surfaces. Therefore, the matrix \(\mathbf{M}_{\rm dist}^{2}\) must be positive definite. Here, the distance \(\delta h\) is measured in the coordinate \(K_{-}\). As to be shown later, the distance matrix \(\mathbf{M}_{\rm dist}^{2}\) between the two hyper-surfaces directly yields the neutral scalar mass matrix. Now we have introduced the vacuum conditions fully in the orbit space. These conditions apply to both the tree-level and the effective potentials. Specifically, the tree-level potential in Eq. (9) can be written in terms of the light-cone coordinates as follows, \[V_{\rm tree}= \xi_{+}K_{+}+\xi_{-}K_{-}+\vec{\xi}_{\rm T}\cdot\vec{K}_{\rm T}+ \left(K_{+},K_{-},\vec{K}_{\rm T}\right)\begin{pmatrix}\eta_{++}&\eta_{+-}& \vec{\eta}_{\rm T+}\\ \eta_{+-}&\eta_{--}&\vec{\eta}_{\rm T-}\\ \vec{\eta}_{\rm T+}&\vec{\eta}_{\rm T-}&\eta_{\rm TT}\end{pmatrix}\begin{pmatrix} K_{+}\\ K_{-}\\ \vec{K}_{\rm T}\end{pmatrix}. \tag{51}\] Then the minimal conditions for the tree-level potential from Eq. (49) are \[\left.\frac{\partial V_{\rm tree}}{\partial K_{-}}\right|_{K_{v}} =\xi_{-}+v^{2}\eta_{+-}=2v^{2}u>0,\] \[\left.\frac{\partial V_{\rm tree}}{\partial K_{+}}\right|_{K_{v}} =\xi_{+}+v^{2}\eta_{++}=0,\] \[\left.\frac{\partial V_{\rm tree}}{\partial\vec{K}_{\rm T}} \right|_{K_{v}} =\vec{\xi}_{\rm T}+v^{2}\vec{\eta}_{\Gamma+}=0, \tag{52}\] which are equivalent to the minimal conditions given in Ref. [12]. ### A geometrical view of the scalar mass matrix After the potential develops a vacuum expectation value, the scalar fields become massive. The field components after the EWSB, which are fluctuations around the vacuum. Without loss of generality, we use the Higgs basis in which the vacuum \(v\) is rotated to the first doublet, and the field components are \[H_{1}=\begin{pmatrix}G^{+}\\ \frac{v+\phi+iG^{0}}{\sqrt{2}}\end{pmatrix},\hskip 14.226378ptH_{2}=\begin{pmatrix} H^{+}\\ \frac{R+iI}{\sqrt{2}}\end{pmatrix}, \tag{53}\] where \(\phi,R,I\) and \(H^{\pm}\) are physical fields while \(G_{0}\) and \(G^{\pm}\) are Goldstone fields. By substituting the field components of \(H_{i}\) into Eq. (2), and rewriting them in terms of the light-cone coordinates, we have \[\begin{pmatrix}K_{+}\\ K_{1}\\ K_{2}\\ K_{-}\end{pmatrix}=\frac{v^{2}}{2}\begin{pmatrix}1\\ 0\\ 0\\ 0\end{pmatrix}+v\begin{pmatrix}\phi\\ R\\ I\\ 0\end{pmatrix}+\begin{pmatrix}\frac{\phi^{2}}{2}+\frac{G_{0}^{2}}{2}+G^{+}G^{-} \\ \phi R+IG_{0}+G^{+}H^{-}+G^{-}H^{+}\\ \phi I-RG_{0}+i(G^{+}H^{-}-G^{-}H^{+})\\ \frac{I^{2}}{2}+\frac{R^{2}}{2}+H^{-}H^{+}\end{pmatrix}. \tag{54}\] The charged Higgs boson mass is given by \[m_{H^{\pm}}^{2}=\left.\frac{\partial V}{\partial H^{-}H^{+}}\right|_{\rm vev}= \left.\frac{\partial V}{\partial K_{-}}\right|_{K_{v}}\left.\frac{\partial K_{ -}}{\partial H^{-}H^{+}}\right|_{\rm vev}=\left.\frac{\partial V}{\partial K_{ -}}\right|_{K_{v}}. \tag{55}\] As for the neutral physical scalars \(\phi,R\) and \(I\), their mass matrix is calculated by expanding the potential in the field space as follows, \[\delta V=\left(\delta\phi,\delta R,\delta I\right)\,\mathbf{M}_{\rm neutral}^ {2}\,\,\begin{pmatrix}\delta\phi\\ \delta R\\ \delta I\end{pmatrix}, \tag{56}\] where \(\delta\phi,\delta R\) and \(\delta I\) are small expansions of the fields around the vacuum. Equation (54) shows that the three directions \((K_{+},\vec{K}_{\rm T})\), which span the tangent space of \(LC^{+}\) and \(\mathcal{M}_{\rm vev}\), are linearly related to the three neutral scalar fields \((\phi,R,I)\) around the vacuum. The linear relationship between field space and orbit space directly links the scalar mass matrix and the distance matrix between \(LC^{+}\) and \(\mathcal{M}_{\rm vev}\). By combining Eq. (56) with Eqs. (50) and (54), we obtain \[\mathbf{M}_{\rm neutral}^{2}=v^{2}\begin{pmatrix}\frac{\partial^{2}V_{\rm u}} {\partial K_{+}^{2}}&\frac{\partial^{2}V_{\rm u}}{\partial K_{+}\partial\vec {K}_{\rm T}}\\ \frac{\partial^{2}V_{\rm u}}{\partial K_{+}\partial\vec{K}_{\rm T}}&\frac{ \partial^{2}V_{\rm u}}{\partial\vec{K}_{\rm T}^{2}}\end{pmatrix}=v^{2}\frac{ \partial V}{\partial K_{-}}\mathbf{M}_{\rm dist}^{2}. \tag{57}\] Therefore, the neutral mass matrix is simply proportional to the distance matrix between the two hyper-surfaces \(LC^{+}\) and \(\mathcal{M}_{\rm vev}\). The experimentally preferred Higgs alignment limit can be read out from Eq.(57) directly. In the alignment limit, the neutral scalar \(\phi\) in Eq. (53) corresponds to the SM-like Higgs boson, and all of its properties are very close to the SM Higgs boson, including mass, gauge couplings, Yukawa couplings, and CP property. Technically, the alignment limit is reached when the neutral scalar \(\phi\) in Eq. (53) is approximately the 125 GeV mass eigenstate and does not mix with other neutral scalars, therefore, we obtain the following relations from Eq. (57), \[\left.\frac{\partial^{2}V_{u}}{\partial K_{+}\partial\vec{K}_{\rm T}}\right|_{ K_{v}}=\left.\frac{\partial^{2}V}{\partial K_{+}\partial\vec{K}_{\rm T}} \right|_{K_{v}}\approx 0, \tag{58}\] where \(K_{+}\) and \(\vec{K}_{\rm T}\) are light-cone coordinates in orbit space. At tree-level, this condition yields \(\vec{\eta}_{\rm T+}\approx 0\) straightforwards. Another demonstration is to discuss the ultra-light CP-odd particle, which is also known as the axion-like particle (ALP). The ALP is of widespread interest for its rich phenomenology, and the 2HDM is a simple model that can provide the ALP. From the geometric relations in the orbit space, a massless scalar appears when the two hyper-surfaces \(LC^{+}\) and \({\cal M}_{\rm vev}\) osculate at \(K_{v}\) along a certain direction. There are two possibilities in the 2HDM to produce an ALP naturally, due to symmetries rather than accidental parameter choice. One possibility is the 2HDM potential with an approximately \(U(1)_{a}\) symmetry. An exact \(U(1)_{a}\) symmetry in the 2HDM potential results in an additional Goldstone boson, and the Goldstone boson will develop a small mass if the \(U(1)_{a}\) symmetry is slightly broken as shown in Fig. 4(a). In this case, the ALP is a pseudo-Goldston boson as in the Dine-Fischler-Srednicki-Zhitnitsky axion model [33; 34]. Another possibility is the 2HDM potential with a CP symmetry that is spontaneously broken. When the SCPV phase \(\delta\) is very small, the two degenerate vacuums turn to merge, and the two hyper-surfaces \(LC^{+}\) and \({\cal M}_{\rm vev}\) turn to osculate with each other at \(K_{v}\), as shown in Fig. 4(b), therefore, a massless boson appears when the SCPV phase \(\delta\) goes to zero [35]. In this case, the ALP is not a pseudo-Goldston boson. Figure 4: A two dimensional slice of Fig. 3 with \(K_{0}=K_{v,0}\), viewed from the \(K_{0}\) direction. The symbol \(\odot\) denotes the \(K_{0}\) axis. The yellow line denotes \(LC^{+}\) and the green denotes \({\cal M}_{\rm vev}\). There are two scenarios with an ultra-light scalar: (a) potential with a slightly broken \(U(1)_{a}\) symmetry; (b) the SCPV potential with a small CP phase \(\delta\). On-shell renormalization in the orbit space The masses and mixing angles of physical states derived from the one-loop CW potential in the \(\overline{\text{MS}}\) renormalization scheme differ from their tree-level values. To directly use the loop-corrected masses and mixing angles as inputs, the on-shell renormalization scheme is often preferred. This is achieved by adding the counterterm potential \(V_{\text{CT}}\) to the zero temperature effective potential \[V_{\text{eff}}=V_{\text{tree}}+V_{\text{CW}}+V_{\text{CT}}, \tag{59}\] and then enforcing the loop-corrected vacuum and masses to be the same as the tree-level values. Consequently, the renormalization conditions in the field space are given by \[\partial_{\varphi_{a}}(V_{\text{CT}}+V_{\text{CW}})\big{|}_{ \varphi_{a}=(\varphi_{a})_{\text{tree}}}=0, \tag{60}\] \[\partial_{\varphi_{a}}\partial_{\varphi_{b}}(V_{\text{CT}}+V_{ \text{CW}})\big{|}_{\varphi_{a}=(\varphi_{a})_{\text{tree}}}=0, \tag{61}\] where \(\varphi_{a}\) (\(a=1\cdots 8\)) denote the eight scalar field components in the two Higgs doublets. However, most of the renormalization conditions are redundant due to unphysical fields and quite a few identities, and it is convenient to deal with the renormalization condition in orbit space.3 To achieve this, we express the counterterm potential in the bilinear notation as \(V_{\text{CT}}=\delta\xi_{\mu}\,K^{\mu}+\delta\eta_{\mu\nu}\,K^{\mu}K^{\nu}\). Based on the vacuum conditions in Eq. (49) and the scalar masses given in Eqs. (55) and (57), we obtain ten independent renormalization conditions that are related to the physical fields as follows: Footnote 3: A detailed analysis of the number of renormalization conditions in the field space and their equivalence with the conditions in the orbit space is presented in Appendix C. \[0 =\partial_{K_{+}}(V_{\text{CT}}+V_{\text{CW}})\big{|}_{K_{v}}, \tag{62}\] \[0 =\partial_{\widetilde{K}_{T}}(V_{\text{CT}}+V_{\text{CW}})\big{|} _{K_{v}},\] (63) \[0 =\partial_{K_{-}}(V_{\text{CT}}+V_{\text{CW}})\big{|}_{K_{v}},\] (64) \[0 =\partial_{K_{+}}^{2}(V_{\text{CT}}+V_{\text{CW}})\big{|}_{K_{v}},\] (65) \[0 =\partial_{\widetilde{K}_{T}}^{2}(V_{\text{CT}}+V_{\text{CW}}) \big{|}_{K_{v}},\] (66) \[0 =\partial_{K_{+}}\partial_{\widetilde{K}_{T}}(V_{\text{CT}}+V_{ \text{CW}})\big{|}_{K_{v}}. \tag{67}\] Here the light-cone coordinates are defined still by the tree-level vacuum \(K_{v}\), and the derivatives are evaluated around \(K_{v}\). Note that only part of the first and second derivatives \(\left.\partial_{K^{\mu}}(V_{\rm CT}+V_{\rm CW})\right|_{K_{v}}\) and \(\left.\partial_{K^{\mu}K^{\nu}}(V_{\rm CT}+V_{\rm CW})\right|_{K_{v}}\) are related to the vacuum conditions and scalar masses and should be included in the renormalization conditions, while the others are irrelevant to physical quantities. Specifically, four conditions from the first derivative in Eqs. (62)-(64) ensure that the loop-corrected vacuum expectation value is the same as the tree-level case, and Eq. (64) also ensures that the charged scalar mass is the same as the tree-level value. The other six conditions involving the second derivatives in Eqs. (65)-(67) ensure that the neutral scalar masses and mixing angles are the same as those of the tree-level potential. The counterterms \(\delta\xi_{\mu}\) and \(\delta\eta_{\mu\nu}\) can be determined from the renormalization conditions in Eqs. (62)-(67). For a general 2HDM without any constrains on the parameters, there are fourteen free parameters, four in \(\delta\xi_{\mu}\) and ten in \(\delta\eta_{\mu\nu}\), to be determined by the renormalization conditions. After expressing \(\delta\xi_{\mu}\) and \(\delta\eta_{\mu\nu}\) in terms of the light-cone coordinates, the renormalization conditions are \[\delta\eta_{++} =-\partial_{K_{+}}^{2}V_{\rm CW}\big{|}_{K_{v}}, \tag{68}\] \[\delta\vec{\eta}_{T+} =-\partial_{\vec{K}_{T}}\partial_{\vec{K}_{+}}V_{\rm CW}\big{|}_ {K_{v}},\] (69) \[\delta\eta_{TT} =-\partial_{\vec{K}_{T}}^{2}V_{\rm CW}\big{|}_{K_{v}},\] (70) \[\delta\xi_{+} =-\partial_{K_{+}}V_{\rm CW}\big{|}_{K_{v}}-v^{2}\delta\eta_{++},\] (71) \[\delta\vec{\xi}_{T} =-\partial_{\vec{K}_{T}}V_{\rm CW}\big{|}_{K_{v}}-v^{2}\delta \vec{\eta}_{T+},\] (72) \[\delta\xi_{-} =-\partial_{K_{-}}V_{\rm CW}\big{|}_{K_{v}}-v^{2}\delta\eta_{+-}. \tag{73}\] Note that neither the vacuum condition nor the scalar mass matrix depends on the counterterms \(\delta\eta_{--}\), \(\delta\eta_{+-}\) and \(\delta\vec{\eta}_{T-}\), therefore, these four parameters are up to free choices. In addition, our convention is to set the tadpole terms to zero whenever possible. Generally, one can allow the development of vacuum in the field space and introduce the tadpole terms in \(V_{\rm CT}\) as done in Refs. [18; 36]. However, for the most general 2HDM potential, there will be more parameters than renormalization conditions and we can always set the tadpole terms to zero. Tadpole terms may be necessary if we require the counterterms to satisfy some specific constraints such that the remaining parameters cannot satisfy the renormalization conditions. For the 2HDM with some specific parameter constraints required by symmetries or alignment, it is a common practice to demand the counterterms \(\delta\xi_{\mu}\) and \(\delta\eta_{\mu\nu}\) satisfying the same constraints as the tree-level parameters \(\xi_{\mu}\) and \(\eta_{\mu\nu}\). Then the number of parameters in \(\delta\xi_{\mu}\) and \(\delta\eta_{\mu\nu}\) is less than fourteen as in the general 2HDM, and the renormalization conditions need to be dealt with case-by-case. For illustration, we discuss the renormalization conditions used in three 2HDMs below. Softly broken \(Z_{2}\) symmetric potential.Imposing a softly broken \(Z_{2}\) symmetry on the 2HDM Lagrangian is the most popular way to prevent flavor-changing neutral interactions. For a complex 2HDM with softly broken \(Z_{2}\) symmetry, the \(Z_{2}\) symmetry gives four additional constraints on \(\delta\eta_{\mu\nu}\), and the remaining six counterterms can be fixed by the six conditions in Eqs. (68)-(70). The soft quadratic couplings are not constrained, and four parameters in \(\delta\xi^{\mu}\) can be fixed by the four conditions in Eqs. (71)-(73). Real 2HDM with softly broken \(Z_{2}\) symmetry.In addition to the softly broken \(Z_{2}\) symmetry, a CP symmetry is often imposed on the potential. The tree-level potential is invariant under a mirror reflection \(\bar{R}\) in the orbit space, \(V_{\rm tree}(K_{0},\vec{K})=V_{\rm tree}(K_{0},\bar{R}\vec{K})\). Note that the CP symmetry does not impose any additional constraint on the quartic counterterms \(\delta\eta^{\mu\nu}\), as the \(Z_{2}\) symmetry provides stronger constraints than the CP symmetry. On the other hand, the softly broken terms are constrained by the CP symmetry. Say that the mirror reflection is along the second direction \(\bar{R}:K_{2}\rightarrow-K_{2}\), then \(\delta\xi_{2}\) should be set to zero, leaving three free parameters in \(\delta\xi^{\mu}\). Usually, the three parameters in \(\delta\xi^{\mu}\) are not enough to satisfy the four equations in Eqs. (71)-(73). But when the vacuum is invariant under the CP transformation, e.g., \(K_{v}^{\mu}=\frac{v^{2}}{2}(1,0,0,1)\) and \(\vec{K}_{v}=\bar{R}\vec{K}_{v}\), there are only three independent conditions in Eqs. (71)-(73), because the CW potential satisfies the CP symmetry, \(V_{\rm CW}(K_{0},\vec{K})=V_{\rm CW}(K_{0},\bar{R}\vec{K})\), and we have \[\partial_{K_{2}}V_{\rm CW}\big{|}_{K_{v}}=0,\quad\partial_{K_{2}} \partial_{K^{\mu}}V_{\rm CW}\big{|}_{K_{v}}=0. \tag{74}\] Then one renormalization condition \(\delta\xi_{2}=-\partial_{K_{2}}V_{\rm CW}-v^{2}\delta\vec{\eta}_{2+}=0\) automatically holds from Eq. (72), and we end up with three parameters and three conditions. However, if the vacuum develops an SCPV phase \(\delta\), the CP symmetry is broken spontaneously. The vacuum \(\vec{K}_{v}\) is no longer invariant under the CP transformation, e.g., \(K_{v}^{\mu}=\frac{v^{2}}{2}(1,0,\sin\delta,\cos\delta)\) and \(\vec{K}_{v}\neq\bar{R}\vec{K}_{v}\). As a result, Eqs. (74) no longer hold. The rest three parameters in \(\delta\xi^{\mu}\) are not enough to satisfy the renormalization conditions if we still require the counterterm \(\delta\xi_{2}=0\). The remaining renormalization condition, which is equivalent to \(\partial_{\delta}(V_{\rm CW}+V_{\rm CT})=0\), cannot be fulfilled, and this corresponds to a change of the SCPV phase \(\delta\). It could be fixed with a tadpole counterterm of the CP-violating vacuum. Exact aligned 2HDM.In the 2HDM, the exact alignment condition requires that the neutral scalar \(\phi\) in Eq. (53) is the 125 GeV mass eigenstate, then the tree-level parameters satisfy \(\vec{\eta}_{T+}=0\) as shown in Eq. (58). However, the alignment condition is not protected by any symmetry, and there is no guarantee that the counterterms \(\delta\vec{\eta}_{T+}=-\partial_{\vec{K}T}\partial_{\vec{K}_{+}}V_{\rm CW}\) vanish. Therefore, the alignment condition is usually broken by quantum corrections. ## VI Conclusion and discussion We performed a complete analysis of the CP and basis transformation symmetries of the 2HDM in the orbit space. We extended the study of the global symmetries in orbit space to one-loop thermal effective potential. We demonstrated that the global symmetries of the tree-level potential are preserved by quantum corrections from boson loop contributions, but may be broken by fermion loop contributions, depending on the Yukawa interactions. In order to study the vacuum conditions and physical masses in the orbit space, we introduced the light-cone coordinates and generalized the bilinear notation to study the physical scalar fields around the vacuum. It provides a geometric view of the scalar mass matrix and on-shell renormalization conditions. By translating the on-shell renormalization conditions of the vacuum and scalar mass into geometric conditions in the orbit space, we calculated the renormalized one-loop effective potential completely. We extend our study to the case after the EWSB. The geometrical view of scalar masses can provide insight into special limits of the 2HDM mass spectrum, such as alignment limit and ultra-light scalars, thereby simplifying the analysis. The renormalization conditions are much simpler to be dealt with in the orbit space, and there are at most 10 independent on-shell renormalization conditions for a general 2HDM potential. Our work provides a foundation for future study of the 2HDM effective potential and its implications in orbit space. ###### Acknowledgements. The work is supported in part by the National Science Foundation of China under Grants No. 11725520, No. 11675002 and No. 12235001. ## Appendix A Basis invariant notations of 2HDM potential ### Explicit expression of bilinear notation The explicit expression for each component of \(K^{\mu}\) is \[K^{\mu}=\Phi_{i}^{\dagger}\sigma_{ij}^{\mu}\Phi_{j}=\left(\begin{array}{c} \Phi_{1}^{\dagger}\Phi_{1}+\Phi_{2}^{\dagger}\Phi_{2}\\ \Phi_{1}^{\dagger}\Phi_{2}+\Phi_{2}^{\dagger}\Phi_{1}\\ i(\Phi_{2}^{\dagger}\Phi_{1}-\Phi_{1}^{\dagger}\Phi_{2})\\ \Phi_{1}^{\dagger}\Phi_{1}-\Phi_{2}^{\dagger}\Phi_{2}\end{array}\right). \tag{10}\] By comparing the potential in the bilinear notation (Eq. 9) with the traditional notation (Eq. 1), we can explicitly relate these two sets of parameters, \[\xi_{0} \equiv \frac{1}{2}(m_{11}^{2}+m_{22}^{2}),\qquad\eta_{00}=(\lambda_{1}+ \lambda_{2}+2\lambda_{3})/8,\] \[\vec{\xi} = \left(-\Re\left(m_{12}^{2}\right),\Im\left(m_{12}^{2}\right), \frac{1}{2}(m_{11}^{2}-m_{22}^{2})\right)^{T},\] \[\vec{\eta} = \left(\Re(\lambda_{6}+\lambda_{7})/4,-\Im(\lambda_{6}+\lambda_{7 })/4,(\lambda_{1}-\lambda_{2})/8\right)^{T},\] \[E = \frac{1}{4}\left(\begin{array}{ccc}\lambda_{4}+\Re\left( \lambda_{5}\right)&-\Im\left(\lambda_{5}\right)&\Re\left(\lambda_{6}-\lambda_{ 7}\right)\\ -\Im\left(\lambda_{5}\right)&\lambda_{4}-\Re\left(\lambda_{5}\right)&\Im\left( \lambda_{7}-\lambda_{6}\right)\\ \Re\left(\lambda_{6}-\lambda_{7}\right)&\Im\left(\lambda_{7}-\lambda_{6}\right) &\left(\lambda_{1}+\lambda_{2}-2\lambda_{3}\right)/2\end{array}\right). \tag{11}\] In the 4-dimensional orbit space, the physical region is confined to the interior of the forward light-cone, i.e., \(K_{0}\geqslant|\vec{K}|\). Because \(K^{\mu}\) can be decomposed from \(\underline{K}_{ij}=\Phi_{i}^{\dagger}\Phi_{j}\), by definition: \[\underline{K}\equiv\begin{pmatrix}\Phi_{1}^{\dagger}\Phi_{1}&\Phi_{2}^{ \dagger}\Phi_{1}\\ \Phi_{1}^{\dagger}\Phi_{2}&\Phi_{2}^{\dagger}\Phi_{2}\end{pmatrix}\equiv\frac {1}{2}\begin{pmatrix}K_{0}+K_{3}&K_{1}-iK_{2}\\ K_{1}+iK_{2}&K_{0}-K_{3}\end{pmatrix}, \tag{12}\] and the matrix \(\underline{K}\) is actually a semi-positive matrix when \(\Phi_{i}=(\phi_{i\uparrow},\phi_{i\downarrow})^{T}\) are \(SU(2)_{L}\) doublets, \[\underline{K}=\underline{\phi}\underline{\phi}^{\dagger},\quad\underline{\phi }=\begin{pmatrix}\phi_{1\uparrow}&\phi_{1\downarrow}\\ \phi_{2\downarrow}&\phi_{2\downarrow}\end{pmatrix}, \tag{13}\] which directly leads to \[\left\{\begin{aligned} \operatorname{tr}\underline{K}& =K_{0}\geqslant 0,\\ \det\underline{K}&=(K_{0}^{2}-|\vec{K}|^{2})/4 \geqslant 0.\end{aligned}\right. \tag{10}\] Therefore in the bilinear notation, the tree-level 2HDM scalar potential is a real quadratic function of \((K_{0},\vec{K})\), and the physical region is defined inside the forward light-cone. ### \(Z_{2}\) symmetry in the bilinear notation The \(Z_{2}\) symmetry is imposed on the 2HDM by assigning \(Z_{2}\) charges to scalar and fermion fields. In Eq. (1), the two Higgs doublets \(\Phi_{1}\) and \(\Phi_{2}\) carry the \(Z_{2}\) charges of \(-1\) and \(+1\) respectively, forbidding the \((\Phi_{1}^{\dagger}\Phi_{1})(\Phi_{1}^{\dagger}\Phi_{2})\) and \((\Phi_{2}^{\dagger}\Phi_{2})(\Phi_{1}^{\dagger}\Phi_{2})\) terms in the potential. As for the Yukawa interactions, fermions are also assigned with negative or positive \(Z_{2}\) charges, then forced to interact with only \(\Phi_{1}\) or \(\Phi_{2}\). Usually, the patterns of \(Z_{2}\) charges assignments are divided into four types [37; 38; 39; 40; 41; 42]: Type I, Type II, Type X and Type Y, as listed in Table 2. For fermions with different \(Z_{2}\) charges, the vectors \(\vec{Y}\)'s projected by their Yukawa couplings are opposite to each other. For example, in the orbit space of \(Z_{2}\) eigenbasis \((\Phi_{1},\Phi_{2})\), the Yukawa coupling of fermion with positive \(Z_{2}\) charge yield \(Y^{\mu}\propto(1,0,0,-1)\) and the Yukawa coupling of fermion with negative \(Z_{2}\) charge yield \(Y^{\mu}\propto(1,0,0,1)\). ### Tensor notation For the completeness of this paper, here we reviewed another basis invariant notation to analyze the 2HDM potential, the tensor notation [2; 3; 4]. It is straightforward to express the 2HDM scalar potential in an \(U(2)_{\Phi}\) basis invariant form, \[V=\mu_{ij}\Phi_{i}^{\dagger}\Phi_{j}+\lambda_{ij,kl}(\Phi_{i}^{\dagger}\Phi_{j})( \Phi_{k}^{\dagger}\Phi_{l}). \tag{10}\] As a result \(\mu_{ij}\) and \(\lambda_{ij,kl}\) transform covariantly with \(\Phi_{i}\) under the \(U(2)_{\Phi}\) basis transformation, \[\mu_{ij}^{\prime}=U_{ik}\mu_{kl}U_{jl}^{*},\quad\lambda_{ij,kl}^{\prime}=U_{ip} U_{kr}\lambda_{pq,rs}U_{jq}^{*}U_{ls}^{*}. \tag{11}\] By definition, \(\lambda_{ij,kl}=\lambda_{kl,ij}\), and hermiticity requires that \(\mu_{ij}=\mu_{ji}^{*},\ \lambda_{kl,ij}=\lambda_{lk,ji}^{*}\). Under the basis of Eq. (1), we have the following relations explicitly, \[\begin{split}\mu_{11}=m_{11}^{2},&\quad\mu_{22}=m_{ 22}^{2},\\ \mu_{12}=-m_{12}^{2},&\quad\mu_{21}=-{m_{12}^{2}}^{* }\\ \lambda_{11,11}=\lambda_{1},&\quad\lambda_{22,22}= \lambda_{2},\\ \lambda_{11,22}=\lambda_{22,11}=\lambda_{3},&\quad \lambda_{12,21}=\lambda_{21,12}=\lambda_{4},\\ \lambda_{12,12}=\lambda_{5},&\quad\lambda_{21,21}= \lambda_{5}^{*},\\ \lambda_{11,12}=\lambda_{12,11}=\lambda_{6},&\quad \lambda_{11,21}=\lambda_{21,11}=\lambda_{6}^{*},\\ \lambda_{22,12}=\lambda_{12,22}=\lambda_{7},&\quad \lambda_{22,21}=\lambda_{21,22}=\lambda_{7}^{*}.\end{split} \tag{12}\] The potential is invariant under the GCP symmetry Eq. (4) when \(\mu_{ij}\) and \(\lambda_{ij,kl}\) satisfy \[\mu_{ij}=X_{ik}\mu_{kl}^{*}X_{lj}^{*},\quad\lambda_{ij,kl}=X_{im}X_{kn}\lambda _{mp,nq}^{*}X_{jp}^{*}X_{lq}^{*}. \tag{13}\] One can construct several CP invariants to determine whether a potential is GCP invariant [2]. Similar to the Jarlskog invariant [43], a \(SU(3)_{L/R}\) invariant in quark family space, the CP invariants of 2HDM scalar potential are constructed from tensor products of \(\mu_{ij}\) and \(\lambda_{ij,kl}\) as \(U(2)_{\Phi}\) invariants in scalar family space. And tensor notation can also be used to construct CP invariants for scalar fermion interaction after extending tensor structures to fermion family space [2]. In addition, a recent development in tensor notation is using the Hilbert series to systematically construct all possible CP invariants [5], and similar procedures can also be used to construct CP invariant in the lepton sector with Majorana terms [44]. ## Appendix B Effective potential from Scalar Loop Contribution Here we show the calculation of the effective potential from scalar loop contribution in detail. We employ the notations in Ref. [30] to link the eight scalar fields \(\varphi_{i}\) with the bilinear forms \(K^{\mu}\), \[\mathcal{L}=\Omega_{\mu}\left(\partial_{\alpha}\Phi_{\imath}\right)^ {\dagger}\sigma^{\mu}_{ij}\left(\partial^{\alpha}\Phi_{j}\right)-V,\quad\Omega^{2 }=1,\] \[V_{\rm tree}=\xi_{\mu}K^{\mu}+\eta_{\mu\nu}K^{\mu}K^{\nu},\] \[\varphi_{a}=\left(\mathop{\rm Re}\phi_{1,\uparrow},\mathop{\rm Im }\phi_{1,\uparrow},\mathop{\rm Re}\phi_{2,\uparrow},\mathop{\rm Im}\phi_{2, \uparrow},\mathop{\rm Re}\phi_{1,\downarrow},\mathop{\rm Im}\phi_{1, \downarrow},\mathop{\rm Re}\phi_{2,\downarrow},\mathop{\rm Im}\phi_{2, \downarrow}\right),\] (B1) \[K^{\mu}=\varphi_{a}\Sigma^{\mu}_{ab}\varphi_{b},\] \[(\Omega_{\rho}\Sigma^{\rho})^{-1}=\Omega_{\rho}\bar{\Sigma}^{\rho}.\] Note that \(\Omega_{\mu}=(1,0,0,0)\) for the canonical kinetic term The matrix \(\bar{\Sigma}^{\mu}=(\Sigma^{0},-\Sigma^{i})\) and the \(8\times 8\) symmetric matrices \(\Sigma^{\mu}\) defined in Eq. (28) are \[\Sigma^{\mu}=\Sigma^{\mu}_{4}\oplus\Sigma^{\mu}_{4},\quad\Sigma^{0}_{4}= \mathbb{1}_{4},\quad\Sigma^{1}_{4}=\begin{pmatrix}0&\mathbb{1}_{2}\\ \mathbb{1}_{2}&0\end{pmatrix},\quad\Sigma^{2}_{4}=\begin{pmatrix}0&\mathbb{1}_ {2}\\ -\mathbb{1}_{2}&0\end{pmatrix},\quad\Sigma^{3}_{4}=\begin{pmatrix}\mathbb{1}_{2 }&0\\ 0&-\mathbb{1}_{2}\end{pmatrix},\] (B2) where \(\mathbb{1}_{d}\) is the \(d\times d\) identity matrix and \(\mathbb{1}_{2}\equiv\left(\begin{smallmatrix}0&1\\ -1&0\end{smallmatrix}\right)\). Because \((\mathbb{1}_{2})^{2}=-\mathbb{1}_{2}\), the matrix \(\Sigma^{\mu}\) share the same algebra with the pauli matrix \(\sigma^{\mu}\), e.g., \[[\Sigma^{i},\Sigma^{j}]=2\mathbb{1}_{8}\epsilon^{ijk}\Sigma^{k}, \quad(\vec{w}\cdot\vec{\Sigma})^{2}=|\vec{w}|^{2}\mathbb{1}_{8},\] (B3) \[\frac{1}{2}(\bar{\Sigma}^{\mu}\Sigma^{\nu}+\bar{\Sigma}^{\nu} \Sigma^{\mu})=g^{\mu\nu}\mathbb{1}_{8},\] (B4) \[\Sigma^{\mu}\bar{\Sigma}^{\rho}\Sigma^{\nu}=g^{\mu\rho}\Sigma^{ \nu}+g^{\rho\nu}\Sigma^{\mu}-g^{\mu\nu}\Sigma^{\rho}+\mathbb{1}_{8}\epsilon^ {\mu\rho\nu}_{\ \ \ \lambda}\Sigma^{\lambda},\] (B5) \[\bar{\Sigma}^{\mu}\Sigma^{\rho}\bar{\Sigma}^{\nu}=g^{\mu\rho}\bar {\Sigma}^{\nu}+g^{\rho\nu}\bar{\Sigma}^{\mu}-g^{\mu\nu}\bar{\Sigma}^{\rho}- \mathbb{1}_{8}\epsilon^{\mu\rho\nu}_{\ \ \ \lambda}\bar{\Sigma}^{\lambda}.\] (B6) Here \(\mathbb{1}_{8}\equiv\mathbb{1}_{4}\otimes\mathbb{1}_{2}\) is an anti-symmetric matrix who commutes with \(\Sigma^{\mu}\) and satisfies \((\mathbb{1}_{8})^{2}=-\mathbb{1}_{8}\), and \(\vec{w}\) is an arbitrary vector. These identities help to translate some expressions of \(\varphi_{a}\) to bilinear forms. For example, \[\varphi\mathbb{1}_{8}\Sigma^{\mu}\varphi=0,\quad\varphi\Sigma^{\mu}\bar{ \Sigma}^{\rho}\Sigma^{\nu}\varphi=g^{\mu\rho}K^{\nu}+g^{\rho\nu}K^{\mu}-g^{\mu \nu}K^{\rho}.\] (B7) Then we evaluate the second derivative of \(\mathcal{L}\) \[-\frac{\delta^{2}\mathcal{L}}{\delta\varphi_{a}\delta\varphi_{b}}=\Omega_{\rho }\Sigma^{\rho}_{ab}\partial^{2}+\xi_{\mu}\Sigma^{\mu}_{ab}+2\eta_{\mu\nu}( \varphi_{c}\Sigma^{\mu}_{cd}\varphi_{d})\Sigma^{\nu}_{ab}+4\eta_{\mu\nu} \Sigma^{\mu}_{ac}(\varphi_{c}\varphi_{d})\Sigma^{\nu}_{db}.\] (B8) In the following, we work in the frame with the canonical kinetic term with \(\Omega_{\mu}=(1,0,0,0)\), and the scalar mass matrix is \[\mathbf{M}^{2}_{S}(\varphi)_{ab}=A_{ab}+B_{ab},\] \[A_{ab}=A_{\mu}\Sigma^{\mu}_{ab},\quad\,A_{\mu}=2\eta_{\mu\nu}K^{ \nu}+\xi_{\mu},\] (B9) \[B_{ab}=4\eta_{\mu\nu}\Sigma^{\mu}_{ac}\varphi_{c}\varphi_{d} \Sigma^{\nu}_{db}.\] To deal with \(\mathbf{Tr}(\mathbf{M}_{S}^{2n})\) in Eq. (29), we expand the binomial \[\mathbf{Tr}[(A_{ab}+B_{ab})^{n}]=\sum_{l=0}^{n}\sum_{\{p_{i}\}}^{\sum_{p_{i}=n-l} }N_{s}(\{p_{i}\})\mathbf{Tr}(A^{p_{1}}BA^{p_{2}}B\cdots A^{p_{l}}B). \tag{10}\] And we need to evaluate \((A_{\mu}\Sigma^{\mu})^{p}\). Using the identities in Eq. (11), \[(A_{\mu}\Sigma^{\mu})^{p} =(A_{0}\mathbb{1}_{8}+\vec{A}\cdot\vec{\Sigma})^{p},\] \[=\sum_{k=0}^{p}C_{p}^{k}(A_{0})^{p-k}(\vec{A}\cdot\vec{\Sigma})^{k},\] \[=\sum_{k=0}^{p/2}C_{p}^{2k}(A_{0})^{p-2k}|\vec{A}|^{2k}\mathbb{1} _{8}+\sum_{k=0}^{(p-1)/2}C_{p}^{2k+1}(A_{0})^{p-2k-1}|\vec{A}|^{2k}(\vec{A} \cdot\vec{\Sigma}), \tag{11}\] where \(C_{p}^{k}\) is the binomial coefficient and \[A_{0} =2\eta_{00}K_{0}+2\vec{\eta}\cdot\vec{K}+\xi_{0}, \tag{12}\] \[\vec{A} =2K_{0}\vec{\eta}+2E\vec{K}+\vec{\xi}. \tag{13}\] For simplicity, we define a new four-vector \(F(p)_{\mu}\) from \((A_{0},\vec{A})\) \[F(p)_{0} \equiv\sum_{k=0}^{p/2}C_{p}^{2k}(A_{0})^{p-2k}|\vec{A}|^{2k}, \tag{14}\] \[\vec{F}(p) \equiv\begin{cases}-\sum_{k=0}^{(p-1)/2}C_{p}^{2k+1}(A_{0})^{p-2k- 1}|\vec{A}|^{2k}\vec{A}&(p\neq 0),\\ 0&(p=0),\end{cases} \tag{15}\] and we have \[(A_{\mu}\Sigma^{\mu})^{p}=F(p)_{\mu}\bar{\Sigma}^{\mu}. \tag{16}\] The series in Eq. (10) are then calculated as \[\mathbf{Tr}(A^{p_{1}}BA^{p_{2}}B\cdots A^{p_{l}}B) =4^{l}\eta_{\mu_{1}\nu_{1}}\cdots\eta_{\mu_{l}\nu_{l}}\prod_{i=1} ^{l}A(p_{i})_{\rho_{i}}\varphi\Sigma^{\nu_{i}}\bar{\Sigma}^{\rho_{i}}\Sigma^{ \mu_{i+1}}\varphi\] \[=4^{l}\eta_{\mu_{1}\nu_{1}}\cdots\eta_{\mu_{l}\nu_{l}}\prod_{i=1} ^{l}S_{p_{i}}^{\nu_{i}\mu_{i+1}}\] \[=4^{l}\mathbf{tr}\left(\eta\cdot S_{p_{1}}\cdots\eta\cdot S_{p_{l} }\right),\] where \(\mu_{l+1}\equiv\mu_{1}\) and the trace \(\mathbf{tr}\) is taken in the orbit space. The symmetric tensor \(S_{p}^{\mu\nu}=F(p)_{\rho}\varphi\Sigma^{\mu}\bar{\Sigma}^{\rho}\Sigma^{\nu}\varphi =F(p)^{\mu}K^{\nu}+F(p)^{\nu}K^{\mu}-g^{\mu\nu}(F(p)K)\). And the effective potential can be expressed as 4 Footnote 4: For simplicity, the \(\ln p_{E}\) is dropped here. \[V_{\rm CW}^{(S)} = \frac{1}{2}\int\frac{d^{4}p_{E}}{2\pi^{4}}\left[{\bf Tr}\sum_{n=1}^ {\infty}\frac{1}{n}\left(-\frac{{\bf M}_{S}^{2}}{p_{E}^{2}}\right)^{n}\right] \tag{17}\] \[= \frac{1}{2}\int\frac{d^{4}p_{E}}{2\pi^{4}}\sum_{n}(-)^{n}\frac{1} {n(p_{E}^{2})^{n}}\sum_{l=0}^{n}\sum_{\{p_{i}\}}^{\sum p_{i}=n-l}N_{s}(\{p_{i} \}){\bf Tr}(A^{p_{1}}BA^{p_{2}}B\cdots A^{p_{l}}B)\] \[= \frac{1}{2}\int\frac{d^{4}p_{E}}{2\pi^{4}}\sum_{n}(-)^{n}\frac{1} {n(p_{E}^{2})^{n}}\sum_{l=0}^{n}\sum_{\{p_{i}\}}^{\sum p_{i}=n-l}N_{s}(\{p_{i} \}){\bf tr}\left(\eta\cdot S_{p_{1}}\cdots\eta\cdot S_{p_{l}}\right)\] In the end, the \(V_{\rm CW}^{(S)}\) is expressed as a series defined in the orbit space. It is worth mentioning that the discussion of CP property is independent of regularization. When the potential is CP-even, as we discussed, we can apply the CP transformation before and after the regularization and nothing will change. Finally, we can conclude that the CP property of the (CP conserving) potential tree-level potential is not violated by the Coleman-Weinberg potential from scalar loop contribution. ## Appendix C Renormalization Conditions To compare with the renormalization conditions in Ref. [18], we follow their notations and the field expanded around the vacuum \(v_{1},v_{2}\) are \[\Phi_{1}=\frac{1}{\sqrt{2}}\left(\begin{array}{c}\rho_{1}+{\rm i}\eta_{1}\\ v_{1}+\zeta_{1}+{\rm i}\psi_{1}\end{array}\right),\quad\Phi_{2}=\frac{1}{ \sqrt{2}}\left(\begin{array}{c}\rho_{2}+{\rm i}\eta_{2}\\ v_{2}+\zeta_{2}+{\rm i}\psi_{2}\end{array}\right). \tag{18}\] The renormalization conditions are \[\left.\partial_{\varphi_{a}}(V_{\rm CT}+V_{\rm CW})\right|_{ \varphi_{a}=\langle\varphi_{a}\rangle_{\rm tree}}=0, \tag{19}\] \[\left.\partial_{\varphi_{a}}\partial_{\varphi_{b}}(V_{\rm CT}+V _{\rm CW})\right|_{\varphi_{a}=\langle\varphi_{a}\rangle_{\rm tree}}=0.\] (20) \[\varphi_{a}\equiv\left\{\rho_{1},\eta_{1},\zeta_{1},\psi_{1}, \rho_{2},\eta_{2},\zeta_{2},\psi_{2}\right\},\ \langle\varphi_{a}\rangle_{\rm tree}=\left\{0,0,v_{1},0,0,0,v_{2},0\right\}.\] Naively, there are 8+36 renormalization conditions from Eqs. (19) and (20). However, for any function of the form \(f(\Phi_{i}^{\dagger}\Phi_{j})\), its first and second derivative satisfy some identities so that most of the renormalization conditions are redundant. We have the following 5 identities for the first derivatives, \[\partial_{\rho_{1}} =0, \tag{104}\] \[\partial_{\rho_{2}} =0,\] (105) \[\partial_{\eta_{1}} =0,\] (106) \[\partial_{\eta_{2}} =0,\] (107) \[c_{\beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}}=0, \tag{108}\] where \(\partial_{\phi_{i}}=0\) denotes \(\partial_{\phi_{i}}f|_{\phi=\langle\phi\rangle_{\rm{tree}}}=0\) for any function \(f(\Phi_{i}^{\dagger}\Phi_{j})\) and \(\tan\beta=v_{2}/v_{1}\). Therefore, we are left with 3 independent renormalization conditions from Eq. (102), \[\partial_{\zeta_{1}}(V_{\rm{CT}}+V_{\rm{CW}})=0, \tag{109}\] \[\partial_{\zeta_{2}}(V_{\rm{CT}}+V_{\rm{CW}})=0,\] (110) \[(c_{\beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})(V_{ \rm{CT}}+V_{\rm{CW}})=0. \tag{111}\] We have the following 26 identities for the second derivatives, \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})=0, \tag{112}\] \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})=0,\] (113) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}})=0,\] (114) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})=0,\] (115) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})=0,\] (116) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})=0,\] (117) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}})=0,\] (118) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})=0,\] (119) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})=0,\] (120) \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})(c_{ \beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}})=0,\] (121) \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=0,\] (122) \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})(c_{ \beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})=0,\] (123) \[(c_{\beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=0,\] (124) \[(c_{\beta}\partial_{\psi_{1}}+s_{\beta}\partial_{\psi_{2}})(c_{ \beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})=0,\] (125) \[(c_{\beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})(c_{ \beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})=0, \tag{126}\] \[(c_{\beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})(c_{ \beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})=0, \tag{104}\] \[(c_{\beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})=0,\] (105) \[(c_{\beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})(c_{ \beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})=0,\] (106) \[(c_{\beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})=0,\] (107) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})=(c_{\beta}\partial_{ \eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{\beta}\partial_{\eta_{1}}+s_{\beta }\partial_{\eta_{2}}),\] (108) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})=(c_{\beta}\partial_{ \psi_{1}}+s_{\beta}\partial_{\psi_{2}})(c_{\beta}\partial_{\psi_{1}}+s_{\beta }\partial_{\psi_{2}}),\] (109) \[(c_{\beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=(c_{\beta}\partial_{ \eta_{2}}-s_{\beta}\partial_{\eta_{1}})(c_{\beta}\partial_{\eta_{2}}-s_{\beta }\partial_{\eta_{1}}),\] (110) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=(c_{\beta}\partial_{ \eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{\beta}\partial_{\eta_{2}}-s_{\beta }\partial_{\eta_{1}}),\] (111) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=(c_{\beta}\partial_{ \psi_{1}}+s_{\beta}\partial_{\psi_{2}})(c_{\beta}\partial_{\psi_{2}}-s_{\beta }\partial_{\psi_{1}}),\] (112) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=-(c_{\beta}\partial_{ \rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{\beta}\partial_{\eta_{2}}-s_{\beta }\partial_{\eta_{1}}),\] (113) \[(c_{\beta}\partial_{\eta_{1}}+s_{\beta}\partial_{\eta_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=(c_{\beta}\partial_{ \psi_{1}}+s_{\beta}\partial_{\psi_{2}})(c_{\beta}\partial_{\zeta_{2}}-s_{\beta }\partial_{\zeta_{1}}). \tag{114}\] Then, there are 10 independent renormalization conditions from the second derivatives. However, three of them are satisfied automatically when the renormalization conditions from the first derivatives are satisfied, because of the following identities, \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})^{2}= \frac{1}{2v}(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}}), \tag{115}\] \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})=\frac{1}{2v}(c_{\beta} \partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}}),\] (116) \[(c_{\beta}\partial_{\rho_{1}}+s_{\beta}\partial_{\rho_{2}})(c_{ \beta}\partial_{\eta_{2}}-s_{\beta}\partial_{\eta_{1}})=\frac{1}{2v}(c_{\beta} \partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}}). \tag{117}\] Finally, we are left with only 7 independent renormalization conditions from Eq. (C3), \[(c_{\beta}\partial_{\rho_{2}}-s_{\beta}\partial_{\rho_{1}})^{2}(V _{\rm CT}+V_{\rm CW})=0, \tag{118}\] \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})^{2}(V _{\rm CT}+V_{\rm CW})=0,\] (119) \[(c_{\beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})^{2}(V _{\rm CT}+V_{\rm CW})=0,\] (120) \[(c_{\beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})^{2}(V _{\rm CT}+V_{\rm CW})=0,\] (121) \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})(c_{ \beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})(V_{\rm CT}+V_{\rm CW})=0,\] (122) \[(c_{\beta}\partial_{\zeta_{1}}+s_{\beta}\partial_{\zeta_{2}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})(V_{\rm CT}+V_{\rm CW})=0,\] (123) \[(c_{\beta}\partial_{\zeta_{2}}-s_{\beta}\partial_{\zeta_{1}})(c_{ \beta}\partial_{\psi_{2}}-s_{\beta}\partial_{\psi_{1}})(V_{\rm CT}+V_{\rm CW})=0. \tag{124}\] And we have 10 independent renormalization conditions from Eqs. (C9)-(C11) and Eqs. (C1) (C47) in total.

2302.08943

Long Range Object-Level Monocular Depth Estimation for UAVs

Computer vision-based object detection is a key modality for advanced Detect-And-Avoid systems that allow for autonomous flight missions of UAVs. While standard object detection frameworks do not predict the actual depth of an object, this information is crucial to avoid collisions. In this paper, we propose several novel extensions to state-of-the-art methods for monocular object detection from images at long range. Firstly, we propose Sigmoid and ReLU-like encodings when modeling depth estimation as a regression task. Secondly, we frame the depth estimation as a classification problem and introduce a Soft-Argmax function in the calculation of the training loss. The extensions are exemplarily applied to the YOLOX object detection framework. We evaluate the performance using the Amazon Airborne Object Tracking dataset. In addition, we introduce the Fitness score as a new metric that jointly assesses both object detection and depth estimation performance. Our results show that the proposed methods outperform state-of-the-art approaches w.r.t. existing, as well as the proposed metrics.

David Silva, Nicolas Jourdan, Nils Gählert

2023-02-17T15:26:04Z

http://arxiv.org/abs/2302.08943v1

# Long Range Object-Level Monocular Depth Estimation for UAVs ###### Abstract Computer vision-based object detection is a key modality for advanced Detect-And-Avoid systems that allow for autonomous flight missions of UAVs. While standard object detection frameworks do not predict the actual depth of an object, this information is crucial to avoid collisions. In this paper, we propose several novel extensions to state-of-the-art methods for monocular object detection from images at long range. Firstly, we propose Sigmoid and ReLU-like encodings when modeling depth estimation as a regression task. Secondly, we frame the depth estimation as a classification problem and introduce a Soft-Argmax function in the calculation of the training loss. The extensions are exemplarily applied to the YOLOX object detection framework. We evaluate the performance using the Amazon Airborne Object Tracking dataset. In addition, we introduce the Fitness score as a new metric that jointly assesses both object detection and depth estimation performance. Our results show that the proposed methods outperform state-of-the-art approaches w.r.t. existing, as well as the proposed metrics. Keywords:Monocular Depth Estimation Unmanned Aerial Vehicles Detect-And-Avoid Object-level Amazon Airborne Object Tracking Dataset Long Range Detection ## 1 Introduction Within recent years, significant technological progress in Unmanned Aerial Vehicles (UAVs) was achieved. To enable autonomous flight missions and mitigate the risk of in-flight collisions, advanced Detect-And-Avoid (DAA) systems need to be deployed to the aircraft. By design, these systems shall maintain a _well-clear volume_ around other airborne traffic [1]. As a result, DAA systems are required to reliably detect potential intruders and other dangerous objects at a long range to allow for sufficient time to plan and execute avoidance maneuvers. Specifically in small and lightweight UAVs, the usage of Lidar and Radar systems is challenging due to their power consumption, weight, and the required long range detection capabilities. However, computer vision approaches based on monocular images have proved their effectiveness in related use cases such as autonomous driving [7, 22, 5]. In addition, cameras can be equipped with lenses that employ different focal lengths depending on the application. Camera systems are therefore a powerful base modality for the perception stack of small and lightweight UAVs. Depending on the actual use case and application, engineers might choose from several computer vision-related tasks such as image classification, object detection, or semantic segmentation. Those tasks are nowadays usually solved by Convolutional Neural Networks (CNNs) specifically designed for the selected use case. For vision-based DAA systems, single-stage object detection frameworks like SSD [28] or YOLO [31, 32, 16] are often employed to detect the objects of interest. By default, these frameworks detect objects in the two-dimensional image space by means of axis-aligned, rectangular bounding boxes. To reliably detect and avoid potentially hazardous objects, additional information about their three-dimensional position and trajectory is crucial. This capability, however, is missing in most vanilla object detection frameworks and specific extensions are needed to provide it. In this paper, we specifically address the problem of object-level depth estimation based on monocular images for long range detections in the use case of UAVs. Several studies focusing on monocular 3D object detection have been conducted in autonomous driving [22, 5, 34, 14, 15]. For UAV-related use cases, however, object-level monocular depth estimation at long range is not yet widely researched. The two fields of application, UAVs and autonomous driving, differ in two major aspects: 1. The range of the objects. UAVs are required to keep a well clear volume of at least 2000 ft or approximately 600 m to prevent potential mid-air collisions [1]. In autonomous driving, on the other hand, objects are mostly limited to less than 200 m [17, 14, 34, 5, 22, 35]. 2. Knowledge of the full 9 degrees of freedom 3D bounding box is not required to maintain the well clear volume. The distance is sufficient. In addition to simplifying the task, this aspect greatly eases the annotation process. As objects do not require a fully annotated 3D bounding box, one can can save both time and money. Thus, we summarize our contributions as follows: 1. We propose two encodings, Sigmoid and ReLU-like, to improve long range depth estimation modeled as a regression task. 2. We frame the task of depth estimation as a classification problem and introduce Soft-Argmax based loss functions to improve the performance of monocular depth estimation. 3. We introduce a novel _Fitness Score_ metric to assess the quality of depth estimation on object-level combined with the object detection metrics. 4. We demonstrate the extension of the state-of-the-art YOLOX object detection framework and benchmark the proposed approaches against existing methods applied to long range detections. ## 2 Related Work The problem of depth estimation from monocular RGB images has been the subject of intense academic research in the past decade. Due to the ambiguity between an object's size and the object's distance to the camera, it is mathematically an ill-posed problem [24, 14]. Thus, machine learning approaches, specifi cally ones that rely on CNNs, gained traction in this field. Two research streams can be identified in monocular depth estimation: 1. _Dense_ or _Pixel-level_ depth estimation, which estimates a dense map of distances for every pixel of a given image, and 2. _Object-level_ depth estimation, which estimates distances only for detected objects of interest. While dense depth estimation is more prominent in computer vision literature, 3D object detection is gaining popularity in relevant application domains such as environment perception for autonomous driving [22, 5, 34, 14, 15]. Nevertheless, there's limited related work in the domain of 2D object-level depth estimation at long ranges [18]. In the case of depth prediction and corresponding loss functions, we distinguish between continuous regression approaches in contrast to approaches that rely on discretization of the depth estimation. #### 2.0.2 Continuous Regression The reverse Huber loss (berHu) is employed in [23] to model the value distribution of depth predictions as a continuous regression problem for dense depth estimation. [18] uses the L2 loss for training a continuous, object-level depth regressor. The log-distance is used within the loss calculation to scale larger distance values. #### 2.0.3 Depth Discretization [6] formulates depth estimation as a classification problem by discretizing the depth values into intervals. The cross-entropy (CE) loss is used to train a classifier that assigns a depth bin to every pixel for dense depth estimation. [26, 25] use a soft-weighted sum inference strategy to compute final depth predictions based on the softmax predictions scores of the depth bins. [13] proposes an ordinal regression loss function to learn a meaningful inter-depth-class ordinal relationship. The depth intervals for discretization are growing in width for increasing distance to the camera as the uncertainty about the true depth increases as well. [8] extends on the idea of using ordinal regression for depth estimation by using a softmax-like function to encode the target vector for depth classification as a probability distribution. ## 3 Methodology In this section, we give information on YOLOX as the selected base framework for 2D object detection. We outline the mathematical foundation of the different depth estimation strategies and embed our proposed methods into these paradigms. In addition, we introduce the Fitness score metric in detail. ### YOLOX - Base Framework for 2D Object Detection To tackle the problem of depth estimation at the object level, we start with a pure 2D object detection framework that outputs a confidence score, class label, and 2D bounding box for each object using axis-aligned rectangles. Given our use case, in which the trade-off between inference speed and detection performance is of high importance, we choose YOLOX Tiny [16] as the base object detection framework. YOLOX was released in 2021 and is one of the latest advances within the YOLO family [31, 32]. To allow for object-level depth estimation, we create a separate new head dedicated to depth estimation. The architecture of the depth head is based on the existing classification head with the necessary adjustments to the number of output channels in the last layer. This separation between the depth head and the other heads allows for a modular combination of various outputs. While we have selected YOLOX as the foundation for this work, the ideas presented in the following sections can be carried over to other modern 2D object detectors. ### Depth Regression The most natural way to estimate depth \(d\) is to frame it as a continuous regression problem. In this case, the model is trained to predict a single and continuous value by minimizing the distance between the model predictions \(\hat{y}\) and the ground truth target \(y\). In its simplest form, depth can be regressed directly, _i.e._\(y=d\) and \(\hat{y}=\hat{d}\). This simple model, however, allows for negative distances, which are not meaningful in the context of monocular depth estimation. Thus, we can use a differentiable transfer function, \(g\left(x\right)\), which supports us in encoding and decoding the network output given a set of constraints. To avoid negative predictions, we propose the encoding \[g\left(x\right)=\frac{x-b}{a}. \tag{1}\] Its corresponding decoding can be calculated as: \[g\left(x\right)^{-1}=\max\left(d_{\min},a\cdot x+b\right), \tag{2}\] with \(a\) and \(b\) being hyperparameters that allow for better control over the domain and range of the model outputs. As \(g^{-1}\) follows the ReLU structure, we refer to this approach as the ReLU-like encoding. We argue that designing a differentiable transfer function with this constraint not only eases training but also enhances robustness against out-of-distribution objects [4] or adversarial attacks. Besides direct regression, there are encodings based on non-linear transfer functions _e.g._, the inverse \(g\left(x\right)=\frac{1}{x}\)[15] and the logarithm \(g\left(x\right)=\log x\)[10, 9, 24, 3, 18]. All previously mentioned encodings lack an upper bound. Thus, they allow for any positive number to be predicted as the depth of the object. As a result, the calculated loss is also unbound and may cause instabilities during training. In some use cases, however, it is possible to assign an upper bound or a maximum distance to the objects of interest. For those settings, we propose a bounded transfer function that maps the domain \(\left(d_{\min},d_{\max}\right)\) to the range \(\left(-\infty,+\infty\right)\): \[g\left(x\right)=\text{logit}\left(\frac{x-d_{\min}}{d_{\max}-d_{\min}}\right). \tag{3}\] The corresponding decoding operation, based on the sigmoid function \(\sigma\), is then calculated as: \[g\left(x\right)^{-1}=\left(d_{\max}-d_{\min}\right)\sigma\left(x\right)-d_{\min}, \tag{4}\] where \(d_{\max}\) and \(d_{\min}\) are the maximum and the minimum depth, respectively. As \(g^{-1}\) uses the sigmoid function, we refer to this approach as the Sigmoid encoding. ### Depth Bin Classification Depending on the application and the use case, a coarse depth estimation might be sufficient. In such cases, depth estimation can be framed as a multiclass classification task with \(K\) discretized depth intervals \(\{d_{0},d_{1},...,d_{K-1}\}\)[6]. Each depth interval links to an individual class in the classification paradigm. Relaxing the need for fine-grained and continuous depth estimation also eases the process of ground truth generation and data annotation. This, in return, can be beneficial from a business perspective. During training, in a classification setting, the softmax function is typically used in CNNs to compute the pseudo-probability distribution and is paired with CE loss. At test time, the selected depth bin is obtained by using the argmax over the pseudo-probabilities. Reformulating depth estimation as a simple classification task is straightforward. In our experiments, we will use this approach as the baseline for classification-based approaches. Employing CE, however, models the classes - and thus the depth bins - as being independent of each other. In particular, the default CE loss doesn't penalize predictions more if they are further away from the target bin compared to predictions that are closer to the target. Depth bins, however, are ordered. We naturally would consider predictions that are far away from the actual depth as _more wrong_ compared to closer ones. Thus, we propose to design a loss that considers the distance of the predicted depth bin to the target depth bin. Designing a loss based on the distance between the prediction and ground truth implies the knowledge of the argmax of the predicted depth classes. Once the argmax and ground truth is known, an arbitrary distance loss function _e.g._, Smooth L1 or MSE, can easily be computed. The implementation of this approach, however, renders a challenge as the default argmax function is not differentiable. Thus, we replace it with the Soft-Argmax [12, 20] \[\text{Soft-Argmax}\left(\hat{y},\beta\right)=\sum_{i=0}^{K-1}i\cdot\text{ softmax}\left(\beta\hat{y}\right)_{i} \tag{5}\] where \(\beta>0\) is a parameter that scales the model predictions \(\hat{y}\). The larger the \(\beta\), the more it approximates a one-hot encoded argmax. In our experiments, we found \(\beta=3\) to be a good choice. Soft-Argmax provides an approximated bin index that is used to compute a distance loss between it and the target bin. During inference, we can naturally obtain the predicted depth bin, \(\hat{d}_{i}\), by applying the argmax function to the model output, \(\hat{y}\), and set the depth value, \(\hat{d}\), to its center. ### Fitness Score As described previously, depth estimation can be formulated as a regression or a classification task. A natural choice for a metric capable of assessing the quality of depth estimation is the mean absolute localization error [14]. If depth estimation is framed as a classification task, the predicted depth is by default not a continuous number and depends on the _real_ depth assigned to this bin _e.g._, its center. As a result, predictions might cause a large absolute localization error despite being assigned to the proper depth bin. This effect makes it difficult to compare both regression and classification models. To solve this challenge, we suggest to also discretize the network prediction in the case of a regression model into \(K\) bins and apply a metric suitable for classification tasks. By doing so, we are able to compare both regression and classification models. Finally, this approach also simplifies the proper model selection by a single metric across the different depth estimation paradigms. As the network predicts confidence, class label, bounding box parameters as well as the depth, we effectively have set up a multitask network. Thus, we need to be able to assess both depth estimation as well as standard 2D object detection performance. Assessing the performance of a multitask network is challenging as all included tasks might perform and be weighted differently. We, however, favor a single number summarizing the model performance in all tasks. In typical object detection benchmarks, mean Average Precision (mAP) is commonly used [11]. mAP measures both the classification performance as well as the bounding box localization quality by utilizing the Intersection-over-Union (IoU) of the predicted and the ground truth bounding box as an auxiliary metric. In addition, F1 jointly assesses both precision and recall in a single number. Note that because several properties of the object - _e.g._ its size and full 3D bounding box - are unknown and thus not needed for our use case, metrics commonly used in 3D object detection - _e.g._ AP\({}_{\text{3D}}\)[17] and DS [14] - are not suitable. As we propose to calculate the performance of depth estimation as a classification task, we suggest employing a scheme similar to F1 for depth estimation as well. Eventually, we calculate the joint measure as the harmonic mean between the mean F\({}_{1}\)-Score of the object detector, mF\({}_{1}^{\text{OD}}\), and the mean F\({}_{1}\)-Score of the depth estimation, mF\({}_{1}^{\text{DE}}\), given the detected objects. As the harmonic mean between two numbers is the same as the F\({}_{1}\)-Score, we refer to this metric as F\({}_{1}^{\text{Comb}}\). The mean F\({}_{1}\)-Scores for both object detection as well as for depth estimation are dependent on the confidence threshold \(t_{c}\) as well as on the minimum required IoU threshold \(t_{\text{IoU}}\). It is \(t_{c}\in\{0.00,0.01,...,0.99,1.00\}\). All predictions with confidence below this threshold will be discarded. For \(t_{\text{IoU}}\), we obtain the values according to [27] such that \(t_{\text{IoU}}\in\{0.50,0.55,...,0.90,0.95\}\). Predictions with an IoU \(\geq t_{\text{IoU}}\) will be treated as TP. Predictions with an IoU \(<t_{\text{IoU}}\) are assumed to be FP. Finally, it is \[\mathrm{mF}_{1}^{\mathrm{OD}} =\mathrm{mF}_{1}^{\mathrm{OD}}(t_{c},t_{\mathrm{IoU}}) \tag{6}\] \[\mathrm{mF}_{1}^{\mathrm{DE}} =\mathrm{mF}_{1}^{\mathrm{DE}}(t_{c},t_{\mathrm{IoU}})\] (7) \[\mathrm{F}_{1}^{\mathrm{Comb}} =\mathrm{F}_{1}^{\mathrm{Comb}}(t_{c},t_{\mathrm{IoU}})\] (8) \[=\frac{2\cdot\mathrm{mF}_{1}^{\mathrm{OD}}\cdot\mathrm{mF}_{1}^{ \mathrm{DE}}}{\mathrm{mF}_{1}^{\mathrm{OD}}+\mathrm{mF}_{1}^{\mathrm{DE}}}. \tag{9}\] The domain of the combined score \(\mathrm{F}_{1}^{\mathrm{Comb}}\) is \([0,1]\) with higher values representing a better combined performance. As the combined score still depends on both \(t_{c}\) and \(t_{\mathrm{IoU}}\), we distill it into a single value, which we refer to as _Fitness_ score. We define it as the maximum of the combined score over all confidence and IoU thresholds, \[\mathrm{Fitness}=\max_{t_{c},t_{\mathrm{IoU}}}\mathrm{F}_{1}^{ \mathrm{Comb}}. \tag{10}\] By doing so, we are able to assess the model performance _as is_ when productively deployed. ## 4 Experiments To demonstrate the effectiveness of the proposed methods for long range object-level monocular depth estimation we design several experiments and compare them using the Amazon Airborne Object Tracking (AOT) dataset. We split the experiments into three major groups: regression, bin classification, and ordinal regression. Each group formulates the depth estimation task differently, implying different network architectures and loss functions. Eventually, we evaluate the performance of each experiment. We use 2D mAP as well as the mean absolute localization error (MALE) as individual metrics for object detection and depth estimation, respectively. In addition, we assess the quality of each tested approach using the joint Fitness Score. ### Dataset The Amazon AOT dataset was introduced in 2021 as part of the Airborne Object Tracking Challenge [2]. It contains a collection of in-flight images with other aircraft flying by as planned encounters. Planned objects are annotated with a 2D bounding box (in pixels), the object label, and the distance (in meters) from the camera to a specific object. As the metadata only contains the euclidean distance from the camera without splitting it into \(x\), \(y\), and \(z\), we use the terms _distance_ and _depth_ interchangeably within this study. Additionally, the sequences may contain encounters with unplanned objects. Those objects are annotated with bounding box parameters and their specific class label - the distance, however, is unknown. While most other datasets that feature object-level depth annotations mostly focus on autonomous driving and only contain objects up to 200 m [17, 14, 34, 5, 22, 35], the AOT dataset features objects up to several hundreds of meters. In our experiments, we use the _partial_ dataset as provided by the authors. This subset contains objects up to 700 m away from the camera. We have observed that some objects, specifically with a range below 10 m, are mislabeled with respect to the object's distance. Thus, we removed the range annotation for these objects but kept the bounding box and classification labels so that they can still be used for training the object detector. The images in the flight sequences are collected at a rate of 10 Hz. As such, many of the images tend to be quite similar. With the goal of shortening training time without significant degradation of performance, we use only every 5th image of this dataset. This subset is equivalent to 2 Hz or 20 % of the initial dataset. We further split the flight sequences in the dataset into dedicated sets for training (60 %), validation (20 %), and testing (20 %). By splitting the entire dataset on a sequence level, we ensure that no cross-correlations between training, validation, and test set occur. Our selected split also provides similar distance as well as class distributions as depicted in Figures 1 and 2. Sample images of the dataset including typical objects are shown in Figure 3. ### Experimental Setup Our models are trained on \(2464\times 2464\) px images. We upsample and slightly stretch the original images with a resolution of \(2448\times 2048\) px to the target resolution as the network requires squared input images with dimensions multiple of 32. We use Stochastic Gradient Descent (SGD) as our optimizer and combine it with a cosine learning rate scheduler with warm-up [16]. In total, we train for 15 epochs with a batch size of 14 using 2 Nvidia RTX3090 GPUs. As described previously, our network architecture features a de-facto multi-task setting. Thus, we calculate our overall multitask loss function \(\mathcal{L}\) as: \[\mathcal{L}=\mathcal{L}_{\mathrm{OD}}+w_{\mathrm{DE}}\mathcal{L}_{\mathrm{DE}}. \tag{11}\] Figure 1: Distance distribution for all objects within _train_, _val_, and _test_ training set. Accordingly, the detector loss function, \(\mathcal{L}_{\mathrm{OD}}\), is defined as: \[\mathcal{L}_{\mathrm{OD}}=w_{\mathrm{obj}}\mathcal{L}_{\mathrm{obj}}+w_{\mathrm{ loc}}\mathcal{L}_{\mathrm{loc}}+w_{\mathrm{class}}\mathcal{L}_{\mathrm{class}}, \tag{12}\] with \(\mathcal{L}_{\mathrm{obj}}\) being the objectness, \(\mathcal{L}_{\mathrm{loc}}\) the localization, and \(\mathcal{L}_{\mathrm{class}}\) the classification loss. \(w_{\mathrm{obj}}\), \(w_{\mathrm{loc}}\), and \(w_{\mathrm{class}}\) refer to the corresponding balancing weights. We leave the detector loss function from YOLOX [16] unchanged. Thus, it is \(w_{\mathrm{obj}}=1\), \(w_{\mathrm{loc}}=5\), and \(w_{\mathrm{class}}=1\). We conduct experiments with different depth loss functions, \(\mathcal{L}_{\mathrm{DE}}\). At the same time, the depth weight \(w_{\mathrm{DE}}\) is a hyperparameter. #### 3.2.2 Regression Our first set of experiments frames depth estimation as a regression task. As such, we set the number of output channels for the last convolutional layer of our depth estimation head to 1. As described in section 3.2.2, there are different methods of encoding depth information. Moreover, each encoding can be combined with different distance-based loss functions. Figure 3: Sample images of the Amazon AOT dataset [2]. Figure 2: Class distribution for all objects within _train_, _val_, and _test_ training set, log-scaled to improve visibility. As mentioned in section 4.1, the distance of the objects to the camera is at most \(700\,\mathrm{m}\). Therefore, we parameterize the Sigmoid encoding such that it is defined in the domain \((d_{\min},d_{\max})\rightarrow(0,700)\). Similarly, for the ReLU-like encoding, we obtain the best results when defining the hyperparameters \(a\) and \(b\) in a way that it approximates the Sigmoid encoding: \(a=100\) and \(b=\frac{700}{2}\). For the depth loss function, \(\mathcal{L}_{\text{DE}}\), we use Smooth L1 (SL1) [19] and mean squared error (MSE) loss for each encoding: \[\text{SL1}(y,\hat{y}) =\begin{cases}\frac{1}{2N}\sum_{i=1}^{N}\left(y_{i}-\hat{y_{i}} \right)^{2},&\text{if}\;\left|y_{i}-\hat{y_{i}}\right|\leq 1\\ \frac{1}{N}\sum_{i=1}^{N}\left|y_{i}-\hat{y_{i}}\right|-0.5,&\text{otherwise} \end{cases} \tag{13}\] \[\text{MSE}(y,\hat{y}) =\frac{1}{N}\sum_{i=1}^{N}\left(y_{i}-\hat{y_{i}}\right)^{2}. \tag{14}\] In addition, we follow [23] and combine direct depth regression with the reverse Huber (berHu) loss: \[\text{berHu}(y,\hat{y},c)=\begin{cases}\frac{1}{N}\sum_{i=1}^{N}\left|y_{i}- \hat{y_{i}}\right|,&\text{if}\;\left|y_{i}-\hat{y_{i}}\right|\leq c\\ \frac{1}{N}\sum_{i=1}^{N}\frac{\left(y_{i}-\hat{y_{i}}\right)^{2}+c^{2}}{2c},&\text{otherwise}.\end{cases} \tag{15}\] \(\hat{y}\) refers the model prediction and \(y\) is the target. \(c\) is a pseudo-constant that is originally calculated as a function, \(c(y,\hat{y})=\frac{1}{5}\underset{i}{\max}\left(\left|y_{i}-\hat{y_{i}}\right|\right)\)[23]. \(N\) refers to the overall number of predictions. #### 4.2.2 Bin Classification The second set of experiments models depth estimation as a classification task. The depth interval \((d_{\min},d_{\max})\rightarrow(0,700)\) is uniformly discretized into \(K=7\) bins with a uniform bin width of \(100\,\mathrm{m}\). Choosing the proper bin size is rather subjective and highly dependent on the use case. For our use case, we find that \(100\,\mathrm{m}\) is suitable since the environment is less cluttered and objects are found at larger distances when compared to other similar applications, \(e.g.\) autonomous driving, where smaller bin sizes might be desired. Similarly, and in agreement with [29], we choose uniform discretization over a log-space discretization strategy because the latter increases bin sizes at larger distances where most objects are found. Moreover, for our use case, early detections are beneficial as we want to avoid entering other objects' airspace. To allow the model to predict \(K\) depth bins, we change the number of output channels in the last convolutional layer of our depth estimation head to \(K\). Our baseline experiment in this group uses softmax (\(c.f.\) section 3.3) as the final activation and CE as the loss function. In total, we design two experiments that employ the proposed Soft-Argmax (SA) with Smooth L1 and MSE loss. #### 4.2.3 Ordinal Regression In our last set of experiments, we follow the guidelines of [13], framing depth estimation as an ordinal regression problem. First, we uniformly discretize the depth into 7 bins, as previously described. The number of output channels in the last convolution layer is set to \(2\cdot(K-1)\), where the number of bins, \(K\), equals 7. Finally, we reimplement the proposed loss function, applying it to objects instead of pixels. #### 4.2.3 Metrics We evaluate the performance of the experiments based on different metrics. The Fitness score proposed in Section 3.4 is our primary metric. To compute it, depth is once again uniformly discretized into 7 bins with a width of 100 m. During training, we search for hyperparameters that maximize the Fitness score on the validation dataset. Once optimal hyperparameters are found, we evaluate on the _test_ set and report the Fitness score as our primary metric. Additionally, we report secondary metrics including 2D mAP with 10 IoU thresholds \(t_{\text{IoU}}\in\{0.50,0.55,...,0.90,0.95\}\) and the mean absolute localization error. We furthermore evaluate the performance w.r.t. the number of parameters, GFLOPs, inference, and post-processing times, allowing us to compare the methods in terms of computational constraints. ### Results Table 1 summarizes the experiment results on the _test_ set. Within the depth regression methods, the proposed Sigmoid encoding outperforms all other encodings. The ReLU-like encoding performs worse compared to the Sigmoid encoding but is still competitive with the best encoding from the state-of-the-art, the logarithm. The combination Sigmoid/SL1 performs best within this group. Within the classification methods, we observe that the proposed loss functions based on Soft-Argmax perform better than the baseline with CE loss. We obtain the best results w.r.t. Fitness by combining Soft-Argmax with Smooth L1 loss. Ordinal regression also outperforms the classification with CE loss. Our results are consistent with the results of [13]. Overall though, it is outperformed by our proposed loss functions based on Soft-Argmax. Table 1 also shows that in most experiments, extending YOLOX with an additional depth head slightly degrades the base 2D performance by means of 2D mAP. There are notable exceptions, the combination SA/SL1 is one of them. While the combination of Soft-Argmax and Smooth L1 loss performs best w.r.t. the Fitness score and 2D mAP, it doesn't yield the lowest absolute localization error. This can easily be understood as we select the middle point of the predicted bin as the actual distance of the object, _c.f_. Section 3.3. In particular, Table 1 shows that models using the Sigmoid and the ReLU-like encoding regression perform better in this aspect. We attempt to further improve absolute localization in the classification setting by using bin interpolation, as a postprocessing step, instead of simply choosing the center of the bin. Following [29], we define the interpolation function \(f\) as: \[f\left(p\left(d_{i-1}\right),p\left(d_{i}\right),p\left(d_{i+1}\right)\right) =f\left(\frac{p\left(d_{i}\right)-p\left(d_{i-1}\right)}{p\left(d_{i}\right)-p \left(d_{i+1}\right)}\right). \tag{16}\] \(p\left(d_{i}\right)\) refers to the probability of the predicted bin, \(p\left(d_{i-1}\right)\), and \(p\left(d_{i+1}\right)\) are the probabilities of the neighboring bins. The predicted depth bin is refined using: \[\hat{d}=\begin{cases}\hat{d}-\frac{s_{i}}{2}\cdot\left(1-f\left(x\right)\right), &\text{if }p\left(d_{i-1}\right)>p\left(d_{i+1}\right)\\ \hat{d}+\frac{s_{i}}{2}\cdot\left(1-f\left(\frac{1}{x}\right)\right),&\text{ otherwise}\end{cases} \tag{17}\] where \(s_{i}\) is the bin size _i.e._, the width, of the predicted bin \(i\). Any function \(f\) must shift the predicted depth towards the previous bin if \(p\left(d_{i-1}\right)>p\left(d_{i+1}\right)\), shift towards the next bin if \(p\left(d_{i-1}\right)<p\left(d_{i+1}\right)\), and leave it unchanged if \(p\left(d_{i-1}\right)=p\left(d_{i+1}\right)\). We then select the following strictly monotone functions \(f:[0,1]\rightarrow[0,1]\) depicted in Figure 4: \[\text{Equiangular}\] [33] \[f(x)=x, \tag{18}\] \[\text{Parabola}\] [33] \[f(x)=\frac{2x}{x+1},\] (19) \[\text{SinFit}\] [21] \[f(x)=\sin\left(\frac{\pi}{2}(x-1)\right)+1,\] (20) \[\text{MaxFit}\] [30] \[f(x)=\max\left(\frac{1}{2}\left(x^{4}+x\right),1-\cos\left(\frac {\pi x}{2}\right)\right),\] (21) \[\text{SinAtanFit}^{3}\] [29] \[f(x)=\sin\left(\frac{\pi x}{2}\arctan\left(\frac{\pi x}{2}\right) \right). \tag{22}\] \begin{table} \begin{tabular}{l l c c c} \hline \hline **Method** & **Loss function** & **Fitness** & **2D mAP** & **MALE** \\ \hline 2D Only & & – & 27.7 \% & – \\ \hline \multirow{3}{*}{Direct} & SL1 & 42.2 \% & 26.7 \% & 52.7 m \\ & MSE & 43.0 \% & 26.9 \% & 50.3 m \\ & berHu [23] & 43.4 \% & 24.7 \% & 49.6 m \\ \multirow{3}{*}{Inverse} & SL1 & 39.9 \% & 26.4 \% & 92.8 m \\ & MSE & 35.0 \% & 25.5 \% & 94.7 m \\ \multirow{3}{*}{Log} & SL1 & 48.2 \% & 27.0 \% & 35.5 m \\ & MSE & 46.7 \% & 26.6 \% & 38.0 m \\ \multirow{3}{*}{Sigmoid} & SL1 (Ours) & 51.6 \% & 25.7 \% & **28.9 m** \\ & MSE (Ours) & 50.4 \% & 25.3 \% & 32.4 m \\ \multirow{3}{*}{ReLU-like} & SL1 (Ours) & 48.3 \% & 27.9 \% & 33.3 m \\ & MSE (Ours) & 47.7 \% & 28.0 \% & 35.5 m \\ \hline \multirow{3}{*}{Classification} & CE & 50.9 \% & 24.9 \% & 37.9 m \\ & SA/SL1 (Ours) & **53.6 \%** & **28.5 \%** & 37.9 m \\ \cline{1-1} & SA/MSE (Ours) & 52.8 \% & 26.9 \% & 38.5 m \\ \hline \multicolumn{2}{l}{Ordinal Regression} & 52.7 \% & 27.0 \% & 37.9 m \\ \hline \hline \end{tabular} \end{table} Table 1: Experiment results obtained on the _test_ set. Object detection and depth estimation are jointly evaluated on the proposed Fitness score. 2D mAP and mean absolute localization error (MALE) individually evaluate object detection and depth estimation, respectively. As shown in Table 2, all interpolation functions show improvements over the baseline. SinFit and MaxFit obtain the same results and perform the best out of our selection. Despite the improvements, it is not able to surpass the Sigmoid-encoded model. As the interpolation is part of the postprocessing and does not change the network architecture or the predicted depth bin, both Fitness score and 2D mAP remain unchanged. #### 4.2.2 Runtime Comparison Besides the quality of the predictions, another important aspect is how the different models compare at runtime. In Table 3, representative models are benchmarked and compared. Compared to pure 2D object detection, the inference time increases by approx. 4 ms for all proposed methods. This is mainly caused by the increased GFLOPs coming from the additional prediction head. Amongst the proposed methods, GFLOPs, number of parameters, and inference speed do not vary meaningfully. Looking at postprocessing though, we observe that the classification and ordinal regression models are slower than the regression models. This result is expected as there are more steps involved for both in order to transform the model output into the depth value. Moreover, \begin{table} \begin{tabular}{l c} \hline \hline **Function** & **MALE** \\ \hline None (baseline) & 37.9 m \\ Equiangular & 31.1 m \\ Parabola & 32.5 m \\ SinFit & **30.1 m** \\ MaxFit & **30.1 m** \\ SinAtanFit & 34.6 m \\ \hline \hline \end{tabular} \end{table} Table 2: Results of the different bin interpolation functions evaluated on mean absolute localization error (MALE). Figure 4: Different bin interpolation functions. classification and ordinal regression models grow in complexity with an increasing number of depth bins. Lastly, we conclude that the cost of bin interpolation is negligible. ## 5 Conclusion In this work, we addressed the problem of long range object-level monocular depth estimation and exemplarily extended the YOLOX object detection framework. We modeled the depth estimation task as a regression, classification, and ordinal regression problem. To jointly assess object detection and depth estimation performance, we introduced the Fitness score as a novel metric. We proposed two novel encodings for regression, Sigmoid and ReLU-like. The former outperforms other state-of-the-art encodings w.r.t. Fitness score and absolute localization error, while the latter is competitive with the best encoding from the state-of-the-art. Moreover, for classification, we proposed a novel loss function based on the Soft-Argmax operation that minimizes the distance between the predicted and target depth bins. In conjunction with the Smooth L1 loss, it outperforms all other models, including ordinal regression, w.r.t. Fitness score. Furthermore, its 2D mAP performance even surpasses the baseline 2D model. However, it doesn't reach the same accuracy by means of absolute localization error compared to the proposed Sigmoid encoding - even when with bin interpolation functions. In general, regression based models have a slight advantage in postprocessing which lead to an overall faster runtime. Based on the conducted experiments, we find that our proposed methods provide great extensions to standard 2D object detection frameworks, enabling object-level depth estimation at long range. \begin{table} \begin{tabular}{l c c c c} \hline \hline **Method** & **Parameters** & **GFLOPs** & **Inference** & **Postprocessing** \\ \hline 2D Only & 5 034 321 & 56.1 & 21.5 ms & 1.5 ms \\ \hline Sigmoid \& Smooth L1 & 5 200 690 & 66.5 & 25.9 ms & **1.7 ms** \\ SA/SL1 & 5 201 369 & 66.5 & **25.4 ms** & 2.9 ms \\ SA/SL1 \& SinFit & 5 201 369 & 66.5 & 25.5 ms & 3.0 ms \\ Ordinal Regression & 5 201 951 & 66.6 & 25.7 ms & 3.0 ms \\ \hline \hline \end{tabular} \end{table} Table 3: Inference benchmark results on representative models for regression, classification, classification with bin interpolation, and ordinal regression. Results measured with \(2464\times 2464\,\mathrm{px}\) image resolution, batch size 1 and FP16 using PyTorch on an Intel Core i9-10920X and Nvidia RTX3090.

2310.18941

The intrinsic geometry determined by the Cauchy problems of the Camassa-Holm equation

Pseudospherical surfaces determined by Cauchy problems involving the Camassa-Holm equation are considered herein. We study how global solutions influence the corresponding surface, as well as we investigate two sorts of singularities of the metric: the first one is just when the co-frame of dual form is not linearly independent. The second sort of singularity is that arising from solutions blowing up. In particular, it is shown that the metric blows up if and only if the solution breaks in finite time.

Igor Leite Freire

2023-10-29T09:05:39Z

http://arxiv.org/abs/2310.18941v1

# The intrinsic geometry determined by the Cauchy problems of the Camassa-Holm equation ###### Abstract Pseudospherical surfaces determined by Cauchy problems involving the Camassa-Holm equation are considered herein. We study how global solutions influence the corresponding surface, as well as we investigate two sorts of singularities of the metric: the first one is just when the co-frame of dual form is not linearly independent. The second sort of singularity is that arising from solutions blowing up. In particular, it is shown taht the metric blows up if and only if the solution breaks in finite time. **MSC classification 2020:** 35A01, 74G25, 37K40, 35Q51. **Keywords** Equations describing pseudospherical surfaces \(\cdot\) Geometric analysis \(\cdot\) Existence of metrics \(\cdot\) Blow up of metrics ###### Contents * 1 Introduction * 1.1 Novelty of the manuscript * 1.2 Outline of the manuscript * 2 Few facts about the CH equation and the geometry determined by its solutions * 2.1 Wave breaking of solutions * 2.2 Geometric aspects of the CH equation * 3 Notation, notions and main results * 3.1 Sobolev spaces * 3.2 Intrinsic geometry and PSS * 3.3 Main results * 4 Preliminaries * 4.1 Conserved quantities * 4.2 Auxiliary and technical results * 5 Proof of the main results * 5.1 Proof of theorem 3.1 * 5.2 Proof of theorem 3.2 * 5.3 Proof of Theorem 3.3 * 5.4 Proof of theorem 3.4 * 5.5 Proof of theorem 3.5 * 5.6 Proof of theorem 3.6 * 6 Finite height vs finite time of existence * 7 Examples * 8 Discussion * 9 Conclusion ## 1 Introduction Chern and Tenenblat introduced the notion of pseudospherical equations [10], connecting certain special partial differential equations (PDEs) with infinitely differentiable two-dimensional Riemannian manifolds. Roughly speaking, an equation is said to describe a pseudospherical surface (PSS equation) if it is a necessary and sufficient condition for the validity of the structure equations determining a surface of Gaussian curvature \(\mathcal{K}=-1\). As a result, solutions of such an equation determines a co-frame for a pseudospherical metric of a pseudospherical surface (PSS) with \(\mathcal{K}=-1\). This concept will be accordingly and timely revisited in the present work. One of the most well-known equations of this type is the third order equation \[u_{t}-u_{txx}+3uu_{x}=2u_{x}u_{xx}+uu_{xxx}, \tag{1.0.1}\] that was deduced by Camassa and Holm [6] as a shallow water model, named after them, and shown to be a PSS equation by Reyes [56]. Amongst a number of features the Camassa-Holm (CH) equation has, we would like to highlight existence of differentiable, but not smooth, solutions breaking at finite time [13]. Throughout this paper, by smooth we mean \(C^{\infty}\). Subsequent developments and applications of Chern and Tenenblat's ideas have not considered solutions with initial data nor finite regularity. As a result, the solutions examined in [13] are, at first sight, somewhat incompatible with the theory developed in [10]. This incongruity might, and probably does, explain the lack of studies of PSS determined by solutions of the Camassa-Holm equation with finite regularity. Very likely, this is the root of a more general fact, that is the absence of works considering surfaces determined by Cauchy problems involving PSS equations. Only very recently some light have been shed on problems of the nature mentioned above. In [62] qualitative properties of PSS determined by a certain equation reported in [28], (but discovered in a different context in [52]), was studied in conjunction with Cauchy problems. Despite its innovative approach, such a connection was made by considering smooth solutions. A step forward was made soon after, in [22], where PSS surfaces determined by periodic Cauchy problems involving the equation studied in [28, 62] was considered. This led the authors to prove the existence of periodic PSS and, more importantly, for the first time a finite regularity geometric problem was considered. Papers [62, 22] are pioneering studies of PSS in connection with Cauchy problems. However, in view of the qualitative nature of the solutions of the Cauchy problems involved, no solution developing any sort of singularities were considered. A significant leap has been made in [31], where blowing up solutions of the CH were shown to determined PSS. It was proved that the corresponding metric of the surface can only exist within a strip of finite height in the \((x,t)\) plane and also experiences a blow up. Indeed, perhaps the most remarkable result proved in [31] is that any non-trivial initial datum in a certain Sobolev class will necessarily define a co-frame for a pseudospherical metric for a PSS. The progress made in [31] would not be possible at no cost: A _sine qua non_ ingredient that enabled the advances carried out in [31] was the reformulation of the notions of PSS determined by an equation and generic solutions. Despite the significant findings reported in [31], some relevant points remain unclear: in the literature of the CH equation, there are blow up scenarios other than those considered in [31], but we had no more than clues about if they really could always be transferred to the metric and that being so, how that happens. From a completely different nature, not to say opposite, the results reported in [31] showed that any non-trivial initial datum defines a strip contained on the upper plane where a co-frame can be defined everywhere, but it remains unclear if or when we could extend the strip to the entire upper plane. ### Novelty of the manuscript While the results reported in [31] showed that any non-trivial initial datum gives rise to a PSS whose dual co-frame is defined on a strip of height \(T>0\), they left many open questions. For example, the blow up mechanisms explored in that reference are not the only ones leading to a breakdown of the solutions. In addition, no attempt has been made to consider the possibility of having \(T=\infty\), nor problems concerning how persistence properties of the solutions may affect the corresponding PSS. This paper shows that we _may_ have PSS defined on subsets of arbitrary height. Moreover, we additionaly study PSS determined by compactly supported initial conditions, precisely describing the asymptotic behaviour of the metric. Furthermore, other blow up conditions for the metrics are also predicted. More importantly, we prove that the metric blows up if and only if the solution develops wave breaking. ### Outline of the manuscript In section 2 we revisit some basic and relevant aspects of the CH equation, with main focus on the two-dimensional Riemannian geometry determined by its solutions and open problems regarding its geometric analysis. Next, in section 3 we fix the notation used throughout the manuscript, recall basic notions and state our main results. In section 4 we revisit some useful facts, such as conserved quantities and qualitative results regarding the CH equation, that are widely employed in section 5, where our main results are proved. In section 6 we show that the metric of the surface blows up if and only if the solution breaks in finite time. Some examples illustrating our main results are discussed in section 7, while our discussions and conclusions are presented in sections 8 and 9, respectively. ## 2 Few facts about the CH equation and the geometry determined by its solutions Despite being primarily deduced as an approximation for the description of waves propagating in shallow water regimes, the equation proved to have several interesting properties related to integrability [6]. If we denote \[m(x,t):=u(x,t)-u_{xx}(x,t),\] which is known as momentum [6], then (1.0.1) can be rewritten as an evolution equation for \(m\), namely, \[m_{t}+2u_{x}m+um_{x}=0. \tag{2.0.1}\] It was shown in [6] that (2.0.1) has a bi-Hamiltonian structure, having the representations \[m_{t}=-\mathcal{B}_{1}\frac{\delta\mathcal{H}_{2}}{\delta m}=-\mathcal{B}_{2} \frac{\delta\mathcal{H}_{1}}{\delta m},\] where \[\mathcal{B}_{1}(\cdot)=\partial_{x}(1-\partial_{x}^{2})(\cdot),\quad\mathcal{ B}_{2}(\cdot)=\partial_{x}(m\cdot)+m\partial_{x}(\cdot)\] are the Hamiltonian operators, and the functionals \(\mathcal{H}_{1}\) and \(\mathcal{H}_{2}\) are \[\mathcal{H}_{1}=\frac{1}{2}\int_{\mathbb{R}}(u^{2}+u_{x}^{2})dx,\quad\mathcal{H}_ {2}=\frac{1}{2}\int_{\mathbb{R}}(u^{3}+uu_{x}^{2})dx. \tag{2.0.2}\] As a consequence of its bi-Hamiltonian structure, (2.0.1) has also a recursion operator \(\mathcal{R}=\mathcal{B}_{2}\mathcal{B}_{1}^{-1}\) and infinitely many symmetries as well, being also integrable in this sense. The reader is referred to [54, Chapter 7] or [53] for further details about recursion operators and integrability. It is still worth of mention that Camassa and Holm showed a Lax formulation [6] for (1.0.1) \[\psi_{xx}=\Big{(}\frac{1}{4}-\frac{m}{2\lambda}\Big{)}\psi,\ \ \psi_{t}=-(\lambda+u)\psi_{x}+\frac{1}{2}u_{x}\psi \tag{2.0.3}\] as well as continuous, piecewise soliton like solutions, called peakons. For a review on the Camassa-Holm and related equations, see [25]. ### Wave breaking of solutions Soon after the seminal work [6], the interest and relevance of the equation spread out the field of integrable equations and arrived at the lands of applied analysis. Solutions emanating from Cauchy problems involving initial data in certain Banach spaces were proved to be locally well-posed, being global under additional conditions [12]. Even more interestingly, depending on the slope of the initial datum there exists a finite value \(T>0\) (the lifespan of the solution), such that \[\liminf_{t\to T}\big{(}\inf_{x\in\mathbb{R}}u_{x}(x,t)\big{)}=-\infty. \tag{2.1.1}\] This fact was first observed in [6] and its rigorous demonstration and dependence on the initial datum was shown by Constantin and Escher [13], see also the review [23]. The Hamiltonian \(\mathcal{H}_{1}\) in (2.0.2) is equivalent to the square of the Sobolev \(H^{1}(\mathbb{R})-\)norm of the solution, meaning that solutions with enough decaying at infinity (as those emanating from initial data \(u_{0}\in H^{s}(\mathbb{R})\), \(s>3/2\)) remain uniformly bounded by the norm of the initial datum as long as they exist [12, 23, 60]. On the other hand, (2.1.1) says that the first singularity (blow up) of a solution, whether it occurs, is manifested by the non-existence of any lower bound of \(u_{x}\) as \(t\) approaches a finite time \(T\), at least for some point \(x\in\mathbb{R}\)[13, Theorem 4.2]. The sudden steepening shown in (2.1.1), but preserving the shape of the solution, is better known as wave breaking (of \(u\)). ### Geometric aspects of the CH equation The Camassa-Holm (CH) equation, or its solutions, can also be studied from geometric perspectives [14, 15, 16, 56]. We shall briefly discuss [14, 56] which are the main inspirations for this paper, the first being concerned with infinite dimensional Riemmanian geometry, whereas the latter is concerned with an abstract two-dimensional Riemannian manifold, whose importance for this paper is crucial. Equation (1.0.1) can be associated with the geometric flow in an infinite dimensional manifold \(\mathcal{D}^{3}(\mathbb{R})\) modelled by a Hilbert space in which we can endow a (weak) Riemannian metric [14]. The geodesics in \(\mathcal{D}^{3}(\mathbb{R})\) can either exist globally [14, Theorem 6.1] or breakdown in finite time [14, Theorems 6.3 and 6.4] and, in particular, geodesics starting, at the identity, with initial velocity corresponding to initial datum leading to breaking solutions will also develop singularities at finite time [14, Theorem 6.3]. A different geometric perspective for the CH equation was given by Reyes [56], who showed it describes pseudospherical surfaces [56, Theorem 1] _a la_ Chern and Tenenblat [10], e.g. see [7, Definition 2.1]. **Definition 2.1**.: _A pseudospherical surface \((PSS)\) is a two-dimensional Riemannian manifold whose Gaussian curvature is constant and negative._ For now it suffices saying that an equation describes pseudospherical surfaces, or is of the pseudospherical type, henceforth referred as PSS equation, when the equation is the compatibility condition of the structure equations \[d\omega_{1}=\omega_{3}\wedge\omega_{2},\quad d\omega_{2}=\omega_{1}\wedge \omega_{3},\quad d\omega_{3}=-\mathcal{K}\omega_{1}\wedge\omega_{2}, \tag{2.2.1}\] for a PSS. That said, in section 2 we show how from a given solution of the CH equation we can construct intrinsically a two-dimensional manifold having Gaussian curvature \(\mathcal{K}=-1\). The impossibility of a complete realisation of these surfaces in the three-dimensional space is a consequence of Hilbert theorem, which states one cannot immerse isometrically a complete surface with negative curvature into \(\mathbb{R}^{3}\)[51, page 439], [46]. See also [21, section 5-11] for a proof of the Hilbert theorem. This explains the adjective _abstract_ often used to qualify a PSS. In his work Reyes showed that if \(u\) is a solution of the CH equation, \(m\) is its corresponding momentum, then the one-forms \[\omega_{1} = \Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}-m\Big{)}dx+\Big{(}um +\frac{\lambda}{2}u-\frac{u}{2\lambda}-\frac{1}{2}-\frac{\lambda^{2}}{2}\Big{)}dt,\] \[\omega_{2} = -u_{x}dt, \tag{2.2.2}\] \[\omega_{3} = \Big{(}m+\frac{1}{2\lambda}-\frac{\lambda}{2}\Big{)}dx+\Big{(} \frac{\lambda^{2}}{2}-\frac{1}{2}-\frac{u}{2\lambda}-\frac{\lambda}{2}u-um \Big{)}dt,\] satisfy (2.2.1), for any \(\lambda\in\mathbb{R}\setminus\{0\}\) and \(\mathcal{K}=-1\). This implies that the domain of the solution \(u\), under certain circumstances, can be endowed with a Riemannian metric \(g=\omega_{1}^{2}+\omega_{2}^{2}\) of a PSS, also known as first fundamental form of the surface. From (2.2.2), the corresponding metric is \[g = \Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}-m\Big{)}^{2}dx^{2}+2 \Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}-m\Big{)}\Big{(}um+\frac{\lambda} {2}u-\frac{u}{2\lambda}-\frac{1}{2}-\frac{\lambda^{2}}{2}\Big{)}dxdt\] \[+ \Big{[}u_{x}^{2}+\Big{(}um+\frac{\lambda}{2}u-\frac{u}{2\lambda} -\frac{1}{2}-\frac{\lambda^{2}}{2}\Big{)}^{2}\Big{]}dt^{2}=:g_{11}dx^{2}+2g_{12 }dxdt+g_{22}dt^{2}.\] More precisely, the work of Reyes showed that, in fact, the Camassa-Holm equation is geometrically integrable, in the sense that its solutions may describe a one-parameter family of non-trivial pseudospherical surfaces [56, Corollary 1]. This is a reflection that the parameter \(\lambda\) in (2.2.2) cannot be removed under a gauge transformation1 While in [14] the influence of solutions emanating from Cauchy problems is crucial in the study of the existence and formation of singularities of geodesics, this is a point usually not considered in the literature of PSS equations, see [8, 9, 10, 17, 18, 20, 22, 23, 41, 55, 56, 40, 41, 57, 58, 59, 63] and references therein. Moreover, in the study of PSS and PDEs the solutions are assumed to be smooth, very often implicitly, but sometimes clearly mentioned [8, page 2] and [41, page 2]. A smooth solution of a PSS equation leads to smooth one-forms \(\omega_{1},\omega_{2},\omega_{3}\) and then the corresponding first fundamental form will inherit the same regularity. The solutions considered by Constantin [14], on the contrary, are not necessarily \(C^{\infty}\), showing an enormous difference among [14] and [7, 8, 55, 56, 57, 58, 59, 63] in terms of the regularity of the objects considered. Additionally, in the context of the literature of PDEs and PSS, the problem of uniqueness of solutions is not usually discussed, then the question of whether a given first fundamental form could be associated to one or more solutions of the CH equation has not been yet considered. Therefore, a situation like the one shown in Figure 2 is very likely to happen: how can one study the intrinsic geometry associated to the solutions of the CH equation when our only information is a known curve on the boundary of its graph? ## 3 Notation, notions and main results Throughout this paper \(u=u(x,t)\) denotes a function depending on the variables \(x\) and \(t\), whose physical meaning, when considering the model (1.0.1), are height of the free surface of water above a flat bottom, space and time, respectively. From a geometric point of view, \(x\) and \(t\) are coordinates of a domain in \(\mathbb{R}^{2}\) in which the function \(u\) is defined. We denote by \(u(x,\cdot)\) and \(u(\cdot,t)\) the functions \(t\mapsto u(x,t)\), for fixed \(x\), and \(x\mapsto u(x,t)\), for fixed \(t\), respectively. For given two non-empty and connected subsets \(I,J\subseteq\mathbb{R}\), the notation \(u\in C^{0}(I\times J)\) means that \(u=u(x,t)\) is continuous with respect to both variables in \(I\times J\). By \(u_{x}\) or \(\partial_{x}u\) we denote partial derivative of \(u\) with respect to its first argument, while similarly \(u_{t}\) or \(\partial_{t}u\) will denote partial derivative with respect to the second argument. We can also consider higher order derivatives using similar convention. Figure 1: Graphs of solutions \(u(x,t)=e^{x-ct}\) of the CH equation for different values of \(c\). The curve \(x\mapsto(x,0,e^{x})\), highlighted in red, belongs to the graph of this function for any value of \(c\). The set of ordered \(n-th\) derivatives of \(u\), \(n\in\mathbb{N}\), is denoted by \(u_{(n)}\). By convention, \(u_{(0)}=u\). Whenever \(u\) and its all derivatives up to order \(k\in\mathbb{N}\cup\{0\}\) are continuous on the domain of \(u\), we then write \(u\in C^{k}\). The sets of smooth functions defined on a domain \(\Omega\subseteq\mathbb{R}^{2}\) is denoted by \(C^{\infty}(\Omega)\). Given \(n\in\mathbb{N}\), a non-empty set \(I\subseteq\mathbb{R}\) and a Banach space \(X\), we say that \(u\in C^{n}(X,I)\) whenever \(\partial_{x}^{k}u(\cdot,t)\in C^{0}(X,I)\), \(0\leq k\leq n\). Moreover, \(u\in C^{0}(X,I)\) means \(u(\cdot,t)\in X\) and \(\|u\|_{C^{0}}=\sup_{t\in I}\|u(\cdot,t)\|_{X}\). ### Sobolev spaces The Sobolev spaces \(H^{s}(\mathbb{R})\) and the \(L^{p}(\mathbb{R})\) spaces, \(s\in\mathbb{R}\) and \(1\leq p\leq\infty\), are the most relevant Banach spaces used throughout this work. Familiarity with \(L^{p}(\mathbb{R})\) spaces is presupposed, whereas we opt to revisit some basic facts about Fourier analysis and Sobolev spaces due to their importance for our developments. For further details, see [64, Chapter 4]. The set of smooth rapidly decaying functions (Schwartz space) is denoted by \(\mathcal{S}(\mathbb{R})\), whereas its dual space is denoted by \(\mathcal{S}^{\prime}(\mathbb{R})\). Elements of \(\mathcal{S}(\mathbb{R})\) are called _test functions_, while those lying in its dual are known as _tempered distributions_. The Fourier transform of a test function \(\phi\) is denoted and given, respectively, by \(\hat{\phi}\) and \[\mathcal{F}(\phi)(\xi)=\hat{\phi}(\xi):=\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R} }\phi(x)e^{-ix\xi}dx,\] whose inverse is \[\phi(x)=\mathcal{F}^{-1}(\hat{\phi})(x)=\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R }}\hat{\phi}(\xi)e^{ix\xi}d\xi.\] The Fourier transform of a tempered distribution \(\psi\), denoted by \(\hat{\psi}\), can be defined throughout the relation \(\langle\phi,\mathcal{F}(\psi)\rangle=\langle\mathcal{F}(\phi),\psi\rangle\). The _Sobolev space_ of order \(s\in\mathbb{R}\), denoted by \(H^{s}(\mathbb{R})\), is the set of tempered distributions \(f\in\mathcal{S}^{\prime}(\mathbb{R})\) such that \((1+|\xi|^{2})^{s/2}\hat{f}(\xi)\in L^{2}(\mathbb{R})\), that has a natural inner product induced by \(L^{2}(\mathbb{R})\) \[\langle f,g\rangle_{s}=\int_{\mathbb{R}}(1+\xi^{2})^{s}\hat{f}(\xi)\overline{ \hat{g}(\xi)}d\xi. \tag{3.1.1}\] We denote by \(\langle\cdot,\cdot\rangle_{s}\) and \(\|\cdot\|_{s}\), \(s\in\mathbb{R}\), the inner product in \(H^{s}(\mathbb{R})\) and its induced norm, respectively, whereas by \(\|\cdot\|_{L^{p}(\mathbb{R})}\) we denote the norm in the \(L^{p}(\mathbb{R})\) space, for finite \(p\), and \(\|\cdot\|_{\infty}\) otherwise. In particular, \(\mathcal{S}(\mathbb{R})\subset H^{s}(\mathbb{R})\subset H^{t}(\mathbb{R}) \subset\mathcal{S}^{\prime}(\mathbb{R})\), for any \(s\geq t\). The following is a cornerstone result for our developments. **Lemma 3.1**.: (Sobolev Embedding Theorem, [64, Proposition 1.2, page 317]) _If \(s>1/2\), then each \(u\in H^{s}(\mathbb{R})\) is bounded and continuous. In addition, if \(s>1/2+k\), \(k\in\mathbb{N}\), then \(H^{s}(\mathbb{R})\subseteq C^{k}(\mathbb{R})\cap L^{\infty}(\mathbb{R})\)._ As we will soon see, the natural Sobolev space for our purposes is precisely \(H^{4}(\mathbb{R})\), which, in view of the precedent result, is embedded into \(C^{3}(\mathbb{R})\cap L^{\infty}(\mathbb{R})\). Let us recall the isomorphism \(\Lambda^{s}:H^{\sigma}\to H^{\sigma-s}\), with \(s,\;\sigma\in\mathbb{R}\), given by \(\Lambda^{s}f:=\mathcal{F}^{-1}((1+\xi^{2})^{s/2}\hat{f})\). For us, the most relevant members of this family are just \(s=2\) and its inverse. For this reason we shall pay a more detailed attention to the operators \(\Lambda^{2}:H^{4}(\mathbb{R})\to H^{2}(\mathbb{R})\) and \(\Lambda^{-2}:H^{2}(\mathbb{R})\to H^{4}(\mathbb{R})\). It is a well known property of the Fourier transform that \(\mathcal{F}(f^{(n)})(\xi)=(i\xi)^{n}\hat{f}(\xi)\), where we assume \(f\in C^{n}(\mathbb{R})\). Moreover, in view of linearity, we have \[(\Lambda^{2}(f))(x)=(\mathcal{F}^{-1}((1+\xi^{2})\hat{f})(x)=(\mathcal{F}^{-1 }(\hat{f}))(x)-(\mathcal{F}^{-1}(-\xi^{2}\hat{f}))(x)=((1-\partial_{x}^{2})f)( x),\] and then, \(\Lambda^{2}=1-\partial_{x}^{2}\). On the other hand, let us define \(h\) by \(\hat{h}(\xi)=\hat{g}(\xi)\hat{f}(\xi)\). Then \((\mathcal{F}(fg))(\xi)=\sqrt{2\pi}\hat{h}(\xi)\) and \(h(x)=(g*f)(x)\), where \(*\) denotes the usual convolution between two functions. In particular, if we consider \[g(x)=\frac{e^{-|x|}}{2}, \tag{3.1.2}\] we have \[(\mathcal{F}(g))(\xi)=\frac{1}{\sqrt{2\pi}}\frac{1}{1+\xi^{2}},\] and then, \[(\Lambda^{-2}(f))(x)=\mathcal{F}^{-1}((1+\xi^{2})^{-1}\hat{f})(x)=\sqrt{2\pi} (\mathcal{F}^{-1}(\hat{g}\hat{f}))(x)=(g*f)(x).\] In view of the comments above, given a function \(u_{0}\in H^{s}(\mathbb{R})\), then it uniquely defines another function \(m_{0}(x)=\Lambda^{2}(u_{0})=u_{0}(x)-u_{0}^{\prime\prime}(x)\) and vice-versa \[u_{0}(x)=(\Lambda^{-2}m_{0})(y)=\frac{1}{2}\int_{\mathbb{R}}e^{-|x-y|}m_{0}(y )dy.\] Another frequent operator seen in this paper is \[\partial_{x}\Lambda^{-2}=(\partial_{x}g)(x)=-\frac{\text{sgn}\,(x)}{2}e^{-|x|}, \tag{3.1.3}\] that acts on \(f\) through the formula \((\partial_{x}\Lambda^{-2}(f))(x)=-\frac{1}{2}(\text{sgn}\,(\cdot)e^{-|\cdot|}* f(\cdot))(x)\). ### Intrinsic geometry and PSS Let \(\mathbb{E}\) be the usual three-dimensional euclidean space, with canonical inner product \(\langle\cdot,\cdot\rangle\) and \(\mathcal{M}\subseteq\mathbb{E}\) be an open, non-empty set, which we shall henceforth identify with a surface. A one-form \(\omega=f(x,t)dx+g(x,t)dt\) defined on \(\mathcal{M}\) is said to be of class \(C^{k}\) if and only if its coefficients \(f\) and \(g\) are \(C^{k}\) functions. We say that a triad of \(C^{k}\) one forms \(\{\omega_{1},\omega_{2},\omega_{3}\}\) endows \(\mathcal{M}\) with a PSS structure with Gaussian curvature \(\mathcal{K}=-1\), if \(\{\omega_{1},\omega_{2}\}\) is linearly independent, that is expressed through the condition \(\omega_{1}\wedge\omega_{2}\big{|}_{\mathcal{M}}\neq 0\), and the following equations \[d\omega_{1}=\omega_{3}\wedge\omega_{2},\quad d\omega_{2}=\omega_{1}\wedge \omega_{3},\quad d\omega_{3}=\omega_{1}\wedge\omega_{2} \tag{3.2.1}\] are satisfied. The form \(\omega_{3}\) is called _Levi-Civita connection_ and it is completely determined by the other two one-forms [51, Lemma 5.1, page 289], as well as the Gaussian curvature of \(\mathcal{M}\)[51, Theorem 2.1, page 329]. Since the forms \(\omega_{1},\omega_{2}\), for each point \(p\in\mathcal{M}\), are dual elements of the basis of the corresponding tangent space, then they are intrinsic objects associated to the surface, as well as any other geometry object described only by them. **Definition 3.1**.: _Let \(\omega_{1}\) and \(\omega_{2}\) be given one-forms on a surface \(\mathcal{M}\) in \(\mathbb{E}\), such that \(\{\omega_{1},\omega_{2}\}\) is LI, and \(p\in\mathcal{M}\). The first fundamental form of \(\mathcal{M}\) is defined, on each each tangent space \(T_{p}\mathcal{M}\) and for any \(v\in T_{p}\mathcal{M}\), by \(I(v)=\omega_{1}(v)^{2}+\omega_{2}(v)^{2}\)._ Using the convention \(\alpha\beta=\alpha\otimes\beta\) and \(\alpha^{2}=\alpha\alpha\), for any one-forms \(\alpha\) and \(\beta\), we can rewrite the first fundamental form as \[I=\omega_{1}^{2}+\omega_{2}^{2}. \tag{3.2.2}\] ### Main results Let us now introduce important and sensitive notions for our main purposes. **Definition 3.2**.: _Let \(u=u(x,t)\); \(\Omega\subseteq\mathbb{R}^{2}\) a non-empty, open and simply connected set, and consider a differential equation for \(u\). A function \(v:\Omega\to\mathbb{R}\) is said to be a classical, or strong, solution for an equation_ \[\mathcal{E}(x,t,u,u_{(1)},\cdots,u_{(n)})=0, \tag{3.3.1}\] _if:_ * \(v\) _possesses as many as continuous derivatives (pure or mixed) to make the equation well defined;_ * \(v\) _satisfies the equation pointwise, that is_ \[\mathcal{E}(x,t,u,u_{(1)},\cdots,u_{(n)})\Big{|}_{u=v}\equiv 0.\] _In addition, we say that \(u\) is a strong solution of (3.3.1) subject to an initial condition \(u\big{|}_{X}=u_{0}\) if \(u\) is a solution as previously described and \(u\in C^{0}(X\cup\Omega)\)._ **Example 3.1**.: _Since the CH equation (1.0.1) has the terms \(u_{t}\), \(u_{x}\), \(u_{xx}\), \(u_{txx}\) and \(u_{xxx}\), then any strong solution defined on an open set \(\Omega\subseteq\mathbb{R}^{2}\) has to have these derivatives continuous._ **Definition 3.3**.: _In view of the Example 3.1, the set of functions defined on a set \(\Omega\subseteq\mathbb{R}^{2}\) for which \(u\), \(u_{t}\), \(u_{x}\), \(u_{xx}\), \(u_{txx}\) and \(u_{xxx}\) are all continuous is denoted by \(C^{3,1}(\Omega)\)._ In the class of Sobolev spaces (with a suitable order) we can see the CH equation as a non-local evolution equation, or a dynamical system, [12, 13, 14, 60], and a straightforward calculation shows that if \(v\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\), then \(v\in C^{3,1}(\mathbb{R}\times(0,T))\subseteq C^{1}(\mathbb{R}\times(0,T))\), and \[v_{t}-v_{txx}+3vv_{x}-2v_{x}v_{xx}+vv_{xxx}=(1-\partial_{x}^{2})\Big{(}v_{t}+ vv_{x}+\partial_{x}\Lambda^{-2}\Big{(}v^{2}+\frac{v_{x}^{2}}{2}\Big{)}\Big{)}. \tag{3.3.2}\] Suppose that \(v\) is a solution of the CH equation (1.0.1). Then \(v\) is a solution of the non-local (first order) evolution equation \[u_{t}+uu_{x}+\partial_{x}\Lambda^{-2}\Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}=0. \tag{3.3.3}\] Conversely, assuming that \(v\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\) is a solution of (3.3.3), then (3.3.2) tells us that \(v\) is a solution of (1.0.1). **Example 3.2**.: _If we consider the non-local form of the CH equation (3.3.3), then any strong solution \(v:\Omega\to\mathbb{R}\) belongs to \(C^{1}(\Omega)\)._ **Example 3.3**.: _As previously mentioned, the sets of solutions for (3.3.3) and (1.0.1) are not the same, being any \(C^{3,1}\) solution of the latter a solution of the former. The converse, however, is not true._ _Consider the function \(u(x,t)=e^{x+t}\). A straightforward inspection shows that it is a solution of (1.0.1) in the sense of definition 3.2, but \(u(\cdot,t)\notin L^{2}(\mathbb{R})\), for any \(t\), meaning that the convolutions involving \(u^{2}\) and \(u_{x}^{2}\) and (3.1.3) are not defined._ The above examples show that a solution for (3.3.3) is not necessarily a solution of the (1.0.1), although they agree for solutions belonging to \(H^{s}(\mathbb{R})\) for \(s\) sufficiently large. The observations made above are well known facts in the literature of the CH equation, but in view of their importance in the development of this manuscript, we want to give them the needed attention. **Proposition 3.1**.: _Let \(u\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\). Then \(u\) is a classical solution of the CH equation (1.0.1) if and only if \(u\) is a classical solution of the non-local equation (3.3.3). Moreover, in such a class, the Cauchy problem_ \[\left\{\begin{array}{l}m_{t}+2u_{x}m+um_{x}=0,\\ \\ u(x,0)=u_{0}(x)\end{array}\right. \tag{3.3.4}\] _is equivalent to_ \[\left\{\begin{array}{l}u_{t}+uu_{x}+\partial_{x}\Lambda^{-2}\Big{(}u^{2}+ \frac{u_{x}^{2}}{2}\Big{)}=0,\\ \\ u(x,0)=u_{0}(x).\end{array}\right. \tag{3.3.5}\] In other words, proposition 3.1 says that (1.0.1) and (3.3.3) are the same object in the class \(C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\). The Cauchy problem (3.3.5) is more convenient to address the questions raised in the Introduction. In fact, in view of the tools developed by Kato [43], we can establish the existence and uniqueness of a solution \(u\in\mathcal{B}:=C^{0}(H^{s}(\mathbb{R}),[0,T))\cap C^{1}(H^{s-1}(\mathbb{R} ),[0,T))\), \(s>3/2\), for (3.3.5) emanating from an initial datum \(u_{0}\in H^{s}(\mathbb{R})\)[60, Theorem 3.2]. While any function in \(\mathcal{B}\) is \(C^{1}\) with respect to \(t\), its regularity regarding \(x\) is controlled by \(s\). Therefore, taking \(s\) sufficiently large we can reach to a higher regularity of the solution with respect to \(x\), making it also a solution for (3.3.4). See also [30]. It is time to drive back to PSS equations. As we have already pointed out, we must observe that several notions in this field were introduced, and have been used assuming, implicitly or explicitly, smooth solutions see [59, Definition 2.4], [7, page 89][42, page 89], and [8, page 2] and [41, page 2], respectively. On the other hand, our paper aims at seeing (3.3.3) as a PSS equation and thus, we need to look for notions that do not require \(C^{\infty}\) regularity in the studied objects. **Definition 3.4**.: (\(C^{k}\) PSS modelled by \(\mathcal{B}\) and B-PSS equation, [31, Definition 2.1]) _Let \(\mathcal{B}\) be a function space. A differential equation (3.3.1), for a dependent variable \(u\in\mathcal{B}\), is said to describe a pseudospherical surface of class \(C^{k}\) modelled by \(\mathcal{B}\), \(k\in\mathbb{N}\), or it is said to be of pseudospherical type modelled by \(\mathcal{B}\), if it is a necessary and sufficient condition for the existence of functions \(f_{ij}=f_{ij}(x,t,u,u_{(1)},\cdots,u_{(\ell)})\), \(1\leq i\leq 3,\,1\leq j\leq 2\), depending on \(u\) and its derivatives up to a finite order \(\ell\), such that:_ * \(\mathcal{B}\subseteq C^{k}\) * _the functions_ \(f_{ij}\) _are_ \(C^{k}\) _with respect their arguments;_ * _the forms_ \[\omega_{i}=f_{i1}dx+f_{i2}dt,\quad 1\leq i\leq 3,\] (3.3.6) _satisfy the structure equations of a pseudospherical surface (_3.2.1_);_ * _the condition_ \(\omega_{1}\wedge\omega_{2}\not\equiv 0\) _is satisfied._ If the function space is clear from the context and no confusion is possible, we maintain the original terminology introduced in the works by Tenenblat and co-authors and simply say PSS equation in place of \(\mathcal{B}-\)PSS equation. Whichever function space \(\mathcal{B}\) is, the first condition asks it to be a subset of \(C^{k}\), that is the space who utterly controls the regularity of the surface. It is possible to find books in differential geometry requiring \(C^{2}\) metrics for a surface, which would force the one-forms being \(C^{2}\)[44, Theorem 4.24, page 153]. However, [34, Theorems 10-19 and 10-19, page 232] and [34, Theorem 10-18, page 232] require \(C^{1}\) regularity of the one-forms defining a surface (and thus, a \(C^{1}\) metric). It is worth noticing that this is the same regularity required by Hartman and Wintner [35, page 760], who proved a sort of Bonnet theorem requiring \(C^{1}\) metric of a surface defined on a domain in \(\mathbb{R}^{2}\). **Remark 3.1**.: _The third condition in definition 3.4 is satisfied if we are able to find functions \(\mu_{1}\), \(\mu_{2}\) and \(\mu_{3}\), depending on \(u\) and its derivatives up to a finite order, vanishing identically on the solutions of the equation, that is,_ \[d\omega_{1}-\omega_{3}\wedge\omega_{2}=\mu_{1}dx\wedge dt,\ d\omega_{2}-\omega_ {1}\wedge\omega_{3}=\mu_{2}dx\wedge dt,\ d\omega_{3}-\omega_{1}\wedge\omega_{2 }=\mu_{3}dx\wedge dt,\] _and_ \[\mu_{1}\big{|}_{\eqref{eq:C^{k}}}\equiv 0,\ \ \mu_{2}\big{|}_{\eqref{eq:C^{k}}} \equiv 0\ \ \mu_{3}\big{|}_{\eqref{eq:C^{k}}}\equiv 0.\] **Remark 3.2**.: _In practical terms, the components of the functions \(f_{ij}\), jointly with the conditions in Definition 3.4, tells us the regularity we have to ask from the solution of the Cauchy problem in order to define a PSS. The final regularity that can be achieved is dictated by these coefficients and that required to grant the existence of solutions from the available tools for proving their well-posedness._ **Remark 3.3**.: _The fourth condition is present for technical reasons, to avoid the situation \(d\omega_{3}=0\), which would imply that \(\omega_{1}=\alpha\omega_{2}\), for some \(\alpha\in\mathbb{R}\). In practical aspects, this condition has to be verified case by case, depending on the solution. Despite being technical, this requirement truly ensures a surface structure in definition 3.4._ While definition 3.4 of \(\mathcal{B}-\)PSS equation has made only a minor modification in the previous one (that by Chern and Tenenblat), the same cannot be said about our proposed notion for a generic solution. **Definition 3.5**.: (Generic solution, [31, Definition 2.2]) _A classical solution \(u:U\to\mathbb{R}\) of (3.3.1) is called generic solution for a \(C^{k}\) PSS equation (3.3.1) if:_ * \(u\in\mathcal{B}\)_;_ * _It is a strong solution in the sense of definition_ 3.2_;_ * _The one-forms (_3.3.6_) are_ \(C^{k}\) _on_ \(U\) _d) There exists at least a simply connected open set \(\Omega\subseteq U\) such that \(\omega_{1}\wedge\omega_{2}\big{|}_{p}\neq 0\), for any \(p\in\Omega\)._ _Otherwise, \(u\) is said to be non-generic._ Let us show that the CH equation (1.0.1) is a \(C^{0}(H^{s}(\mathbb{R}),[0,T))\cap C^{1}(H^{s-1}(\mathbb{R}),[0,T))-\)PSS equation. **Example 3.4**.: _Let \(\lambda\in\mathbb{R}\setminus\{0\}\); \(\Omega\subseteq\mathbb{R}\times[0,T)=:U\) be an open and simply connected set, \(u\) be a solution of the CH equation defined on \(\Omega\), with either \(u_{x}\big{|}_{\Omega}>0\) or \(u_{x}\big{|}_{\Omega}<0\); and suppose that \(u\) satisfies the CH equation on \(\mathring{U}=\mathbb{R}\times(0,T)\). Consider the triad of one-forms (2.2.2). A straightforward calculation shows that_ \[d\omega_{1}-\omega_{3}\wedge\omega_{2} = \Big{(}m_{t}+2u_{x}m+um_{x}\Big{)}dx\wedge dt,\] \[d\omega_{2}-\omega_{1}\wedge\omega_{3} = 0, \tag{3.3.7}\] \[d\omega_{3}-\omega_{1}\wedge\omega_{2} = -\Big{(}m_{t}+2u_{x}m+um_{x}\Big{)}dx\wedge dt,\] _and_ \[\omega_{1}\wedge\omega_{2}=-\Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}-m \Big{)}u_{x}dx\wedge dt. \tag{3.3.8}\] _Moreover, if \(u\) is a solution of the CH equation, we conclude that \(\omega_{1}\wedge\omega_{2}=0\) if and only if_ \[m=\frac{\lambda}{2}+\frac{1}{2\lambda}\quad\text{or}\quad u_{x}=0,\] _that, substituted into (2.0.1), implies_ \[u(x,t)=c, \tag{3.3.9}\] _for some constant \(c\)._ _The minimum of regularity we can require to define a surface is \(C^{1}\), see [34, Theorems 10-19 and 10-19, page 232]. Therefore, the component functions of the one-forms (2.2.2) have to be of this order, which in particular, implies \(m\in C^{1}\). As such, \(u\) has to be at least \(C^{3}\) with respect to \(x\) and \(C^{1}\) with respect to \(t\), with continuous mixed derivatives. As a result, the CH equation is a PSS equation modelled by the function space \(\mathcal{B}:=C^{3,1}(U)\) and \(u\) is a generic solution for the equation, bringing to \(\Omega\) the structure of a PSS._ Example 3.4 does not necessarily show that (3.3.3) can be seen as a PSS equation. However, if we restrict the solutions of the CH equation (1.0.1) to the class \(\mathcal{B}=C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T)) \subseteq C^{3,1}(\mathbb{R}\times[0,T))\) as in proposition 3.1, then the _same_ one-forms (2.2.2) give \[d\omega_{1}-\omega_{3}\wedge\omega_{2} = (1-\partial_{x}^{2})\Big{(}u_{t}+uu_{x}+\partial_{x}\Lambda^{-2} \Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}\Big{)}dx\wedge dt,\] \[d\omega_{2}-\omega_{1}\wedge\omega_{3} = 0, \tag{3.3.10}\] \[d\omega_{3}-\omega_{1}\wedge\omega_{2} = -(1-\partial_{x}^{2})\Big{(}u_{t}+uu_{x}+\partial_{x}\Lambda^{-2} \Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}\Big{)}dx\wedge dt,\] and thus (3.3.3) is a PSS equation in the sense of definition 3.4. In fact, we have the following result. **Theorem 3.1**.: _Let \(T>0\) and consider the function space \(\mathcal{B}=C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T)) \subseteq C^{3,1}(\mathbb{R}\times[0,T))\). Then the CH equation (1.0.1) is a PSS equation modelled by \(\mathcal{B}\) if and only if the non-local evolution equation (3.3.3) is a PSS equation modelled by \(\mathcal{B}\). Moreover, they describe exactly the same PSS, in the sense that \(u\in\mathcal{B}\) is a generic solution of (1.0.1) if and only if it is a generic solution of (3.3.3)._ While theorem 3.1 tells us that the geometric object described by (3.3.7) is identical to that given by (3.3.10), it does not say when or how we can determine whether we really have a PSS from a solution. Moreover, finding a solution of a highly non-linear equation like (1.0.1) is a rather non-trivial task. One of the advantages of the modern methods for studying evolution PDEs is the fact that we can extract much information about properties of solutions, that we do not necessarily know explicitly, from the knowledge of an initial datum. The equivalence between Cauchy problems given by proposition 3.1 and theorem 3.1 suggest that we could have qualitative information from the surface provided that we know an initial datum. In geometric terms, an initial datum uniquely defines a curve. The tools from analysis tell us that this curve, which we know, uniquely determines a solution. Ultimately, the curve then provides in a unique way a surface determined by the graph of the solution (in the sense that the one-forms (3.2.1) are uniquely described by \(u\)). Our goal now is to study qualitatively this (unique) graph from the point of view of PSS framework. **Theorem 3.2**.: _Let \(u_{0}\in H^{4}(\mathbb{R})\) be a non-trivial initial datum, and consider the Cauchy problem (3.3.4). Then there exists a value \(T>0,\) uniquely determined by \(u_{0}\), and an open strip of height \(T\ \mathcal{S}=\mathbb{R}\times(0,T)\), such that the forms (2.2.2) are uniquely determined by \(u_{0}\), defined on \(\mathcal{S}\), and of class \(C^{1}\). Moreover, the Hamiltonian \(\mathcal{H}_{1}\), given in (2.0.2), provides a conserved quantity on the solutions of problem (3.3.4)._ By a non-trivial function we mean one that is not identically zero. The geometric meaning of theorem 3.2 is the following: given a regular curve \[\gamma(x)=(x,0,u_{0}(x)),\quad u_{0}\in H^{4}(\mathbb{R}), \tag{3.3.11}\] let \(\Gamma:=\{\gamma(x),\,x\in\mathbb{R}\}\). Then we can uniquely determine a solution \(u(x,t)\) of the CH equation such that \(\Gamma\subseteq\overline{\text{Gr}(u)})\), where \[\text{Gr}(u)=\{(x,t,u(x,t)),\,x\in\mathbb{R},\,t>0\}\] and \(\overline{\text{Gr}(u)}\) denotes the closure of \(\text{Gr}(u)\). Even though the existence of the forms (2.2.2) over a domain \(\mathcal{S}\neq\emptyset\) is a necessary condition for endowing \(\mathcal{S}\) with the structure of a PSS, it is not sufficient, since the condition \(\omega_{1}\wedge\omega_{2}\neq 0\) is fundamental for such, and theorem 3.2 says nothing about it. It is worth mentioning that a solution \(u\) of the CH equation subject to an initial datum in \(H^{4}(\mathbb{R})\) is unique and its domain is determined by the initial datum [12, Proposition 2.7] and it has to be considered intrinsically with its domain. Moreover, the invariance of the conserved quantity \(\mathcal{H}_{1}\) in (2.0.2) implies \(u_{x}(\cdot,t)\in L^{2}(\mathbb{R})\), for each \(t\) for which the solution exists. Let us fix \(t_{0}\in(0,T)\). Then \(u_{x}(x,t_{0})\to 0\) as \(|x|\to\infty\). Since \(\mathcal{H}_{1}(0)>0\), then \(u(\cdot,t_{0})\not\equiv 0\) and cannot be constant. Therefore, \(u_{x}(\cdot,t_{0})\) cannot be constant either. As a result, we conclude the existence of two points \(x_{0}\) and \(x_{1}\) such that the mean value theorem implies \(u_{x}(x_{0},t_{0})=0\), whereas for the other we have \(u_{x}(x_{1},t_{0})\neq 0\), say \(u_{x}(x_{1},t_{0})>0\). The continuity of \(u_{x}\) then implies the existence of an open and simply connected set \(\Omega\) such that \(u_{x}(\cdot,\cdot)\big{|}_{\Omega}>0\). These comments prove the following result. **Corollary 3.1**.: _Assume that \(u_{0}\) is a solution satisfying the conditions in theorem 3.2 and let \(u\) be the unique solution of (3.3.5). Then \(u_{x}(\cdot,\cdot)\) vanishes at a non-countable number of points of \(\mathcal{S}\). Moreover, there exist open and simply connected subsets \(\Omega\subseteq U\) such that \(u_{x}(x,t)\) does not vanish for any \((x,t)\in\Omega\)._ We have an even stronger result coming from the precedent lines. **Corollary 3.2**.: _Any solution of (3.3.5), emanating from a non-trivial initial datum \(u_{0}\in H^{4}(\mathbb{R})\), is a generic solution in the sense of definition 3.4._ Theorem 3.2 and its corollaries show that any non-trivial initial datum determines a PSS, compare with [31, Theorem 2.2], and their proof is given in subsection 5.2. Due to [31, Theorem 2.2], these results are somewhat expected. The same, however, cannot be said about our next proclamation. **Theorem 3.3**.: _Assume that \(u_{0}\in H^{4}(\mathbb{R})\) is a non-trivial, compactly supported initial datum, with \([a,b]=\text{supp}(u_{0})\) and \(u\) be the corresponding solution of (3.3.4). Then there exists two \(C^{1}\) curves \(\gamma_{+},\gamma_{-}:[0,T)\to\overline{\mathcal{S}}\), and two \(C^{1}\) functions \(E_{+},\,E_{-}:[0,T)\to\mathbb{R}\), where \(T\in\mathbb{R}\) and \(\mathcal{S}\subseteq\mathbb{R}^{2}\) are given in Theorem 3.2, such that:_ * \(\pi_{1}(\gamma_{-}(t))<\pi_{1}(\gamma_{+}(t))\)_, for any_ \(t\in[0,T)\)_, where_ \(\pi_{1}:\mathbb{R}^{2}\to\mathbb{R}\) _is the canonical projection_ \(\pi_{1}(x,t)=x\)_;_ * \(\gamma_{\pm}^{\prime}(t)\neq 0\)_, for any_ \(t\in(0,T)\)_;_ * _On the left of_ \(\gamma_{-}\)_, the first fundamental form is given by_ \[\begin{array}{rcl}g&=&\frac{1}{4}\Big{(}\lambda+\frac{1}{ \lambda}\Big{)}dx^{2}+2\Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}\Big{)} \Big{[}\Big{(}\frac{\lambda}{2}-\frac{1}{2\lambda}\Big{)}E_{-}(t)e^{x}-\frac{ 1}{2}-\frac{\lambda^{2}}{2}\Big{]}dxdt\\ &&+\Big{[}E_{-}(t)^{2}e^{2x}+\Big{(}\Big{(}\frac{\lambda}{2}-\frac{1}{2 \lambda}\Big{)}E_{-}(t)e^{x}-\frac{1}{2}-\frac{\lambda^{2}}{2}\Big{)}^{2} \Big{]}dt,\end{array}\] (3.3.12) * _On the right of_ \(\gamma_{+}\)_, the first fundamental form is given by_ \[\begin{array}{rcl}g&=&\frac{1}{4}\Big{(}\lambda+\frac{1}{ \lambda}\Big{)}dx^{2}+2\Big{(}\frac{\lambda}{2}+\frac{1}{2\lambda}\Big{)} \Big{[}\Big{(}\frac{\lambda}{2}-\frac{1}{2\lambda}\Big{)}E_{+}(t)e^{-x}-\frac {1}{2}-\frac{\lambda^{2}}{2}\Big{]}dxdt\\ &&+\Big{[}E_{+}(t)^{2}e^{-2x}+\Big{(}\Big{(}\frac{\lambda}{2}-\frac{1}{2 \lambda}\Big{)}E_{+}(t)e^{-x}-\frac{1}{2}-\frac{\lambda^{2}}{2}\Big{)}^{2} \Big{]}dt.\end{array}\] (3.3.13) If we denote by \((g)\) the matrix of the first fundamental form and fix \(t\in(0,T)\), then the metrics (3.3.12) and (3.3.13) can be written in a unified way, that is, \[(g)=\begin{pmatrix}\frac{1}{4}\Big{(}\lambda+\frac{1}{\lambda}\Big{)}&-\frac {1}{4}(1+\lambda^{2})\Big{(}\lambda+\frac{1}{\lambda}\Big{)}\\ -\frac{1}{4}(1+\lambda^{2})\Big{(}\lambda+\frac{1}{\lambda}\Big{)}&\frac{1}{4 }\Big{(}1+\frac{1}{2\lambda}\Big{)}\end{pmatrix}+O(e^{-|x|})=:(g_{0})+O(e^{-|x|}),\] as \(|x|\to\infty\), meaning that the matrix \((g)\) is an \(O(e^{-|x|})\) perturbation of the singular matrix \((g_{0})\) as \(|x|\to+\infty\). Therefore, the metric determined by a compactly supported initial datum becomes asymptotically singular, for each fixed \(t\in(0,T)\). Hence, for \(|x|\gg 1\) and \(t\) fixed, the components of the metric behave like the famous peakon solutions of the CH equation. **Theorem 3.4**.: _If \(u_{0}\in H^{4}(\mathbb{R})\) and for some \(x_{0}\in\mathbb{R}\), we have_ \[u_{0}^{\prime}(x_{0})<-\frac{\|u_{0}\|_{1}}{\sqrt{2}}, \tag{3.3.14}\] _then there exists \(0<T_{m}<\infty\) such that the metric (2.2.3), determined by the solution \(o\) (3.3.4), blows up as \(t\to T_{m}\). More precisely, the coefficients \(g_{11}\) and \(g_{12}\) are uniformly bounded whereas_ \[\liminf_{t\to T_{m}}\Big{(}\sup_{x\in\mathbb{R}}g_{22}(x,\tau)\Big{)}=+\infty. \tag{3.3.15}\] Expression (3.3.15) says that the metric blows up for a finite value of \(t\) and then, the surface can only be defined on a proper subset of \(\mathbb{R}^{2}\). While Theorem 3.3 tells us that the metric determined by an initial datum becomes asymptotically singular for each fixed \(t\) as long as the solution exists, theorem 3.4 shows us a different sort of singularity, in which the metric blows up over a strip of finite height. Our next result, however, informs us that a compactly supported initial datum actually leads to a singularity of the metric similar to that established in Theorem 3.4. **Theorem 3.5**.: _If \(u_{0}\in H^{4}(\mathbb{R})\) is a non-trivial, compactly supported initial datum, then the metric (2.2.3), determined by the solution \(o\) (3.3.4), blows up within a strip of finite height._ Theorems 3.4 and 3.5 tell us the existence of a height for which the co-frame of dual forms \(\omega_{1}\) and \(\omega_{2}\) are well defined, but their corresponding metric becomes unbounded near some finite height, meaning that the metric, and the forms as well, are only well defined on a certain strip with infinite length, but finite height. A completely different scenario is given by our next result. **Theorem 3.6**.: _Let \(m_{0}\in H^{2}(\mathbb{R})\cap L^{1}(\mathbb{R})\) and \(u\) be the corresponding solution of (3.3.4). If \(m_{0}(x)\geq 0\) or \(m_{0}(x)\leq 0\), then (2.2.2) are \(C^{1}\) one-forms defined on \(\mathcal{S}=\mathbb{R}\times(0,\infty)\). Moreover, for any \(R>0\), there exists a simply connected set \(\mathcal{R}\subseteq\mathbb{R}^{2}\) such that \(\sqrt{x^{2}+t^{2}}>R\), for any \((x,t)\in\mathcal{R}\), and \(u_{x}\big{|}_{\mathcal{R}}>0\) or \(u_{x}\big{|}_{\mathcal{R}}<0\)._ Theorem 3.6 says that subsets of the domain of the solution of the CH equation that can be endowed with a PSS structure cannot be contained in any compact set. In view of this result, regions arbitrarily far away from the origin may be endowed with the structure of a PSS. ## 4 Preliminaries In this section we present auxiliary results that will help us to prove technical theorems and will be of vital importance in order to establish our main results. ### Conserved quantities The topics we discuss make implicit or explicit use of certain quantities that are conserved for solutions having enough decaying at infinity. For this reason we recall them from a geometric perspective. A differential form \(\alpha\) is said to be _closed_ if \(d\alpha=0\), whereas it is called _exact_ when \(\alpha=d\beta\). Two closed forms are said to be equivalent if their difference is an exact one. Given a differential equation (3.3.1), we say that a one-form \(\alpha=C^{0}dx+C^{1}dt\), whose coefficients depend on \((x,t,u)\) and derivatives of \(u\) up to a certain order, is a _conserved current_ if it is closed on the solutions \(u\) of the equation. In particular, note that \[d\alpha=(\partial_{x}C^{1}-\partial_{t}C^{0})dx\wedge dt\] and if \(\alpha\) is closed on the solution of the equation, then \[(\partial_{x}C^{1}-\partial_{t}C^{0})\Big{|}_{\mathcal{E}=0}=0,\] which is a conservation law for the equation. That said, it is a structural property of the equation, e.g. see [62, section 3]. Conserved currents differing from another one by an exact differential form determine exactly the same conservation law. **Example 4.1**.: _Consider the following one-forms_ \[\alpha=u\,dx+\Big{(}\frac{3}{2}u^{2}-u_{tx}-uu_{xx}-\frac{1}{2}u_{x}^{2} \Big{)}\,dt\] _and_ \[\beta=\frac{u^{2}+u_{x}^{2}}{2}\,dx+\Big{(}u^{3}-u^{2}u_{xx}-uu_{tx}+\Big{)}\,dt.\] _A straightforward calculation shows that_ \[d\alpha=(m_{t}+2u_{x}m+um_{x})dx\wedge dt\ \ \text{and}\ \ d\beta=u(m_{t}+2u_{x}m+um_{x})dx \wedge dt.\] _It is easy to see that on the solutions of the CH equation (2.0.1) these two one-forms are closed. Finally, observe that_ \[\tilde{\alpha}=(u-u_{xx})\,dx+\Big{(}\frac{3}{2}u^{2}-uu_{xx}-\frac{1}{2}u_{x} ^{2}\Big{)}\,dt\] _is equivalent to the one-form \(\alpha\), since \(\tilde{\alpha}=\alpha-d(u_{x})\)._ Integrating the conservation law, we obtain \[\frac{d}{dt}\int_{\mathbb{R}}C^{0}dx=\int_{\mathbb{R}}\Big{(}\frac{d}{dx}C^{1} \Big{)}dx=C^{1}\Big{|}_{-\infty}^{+\infty}.\] If the quantity \(C^{1}\) has enough decaying at infinity and the integral in the left hand side of the equation above converges, then we have \[\frac{d}{dt}\int_{\mathbb{R}}C^{0}dx=0.\] Let \[Q(t):=\int_{\mathbb{R}}C^{0}dx.\] Assuming that \(Q\) is defined for \(t\in I\), where \(I\) is a connected set, then it is called _conserved quantity_. In particular, if \(0\in I\), we have \(Q(t)=Q(0)\) for any other \(t\in I\). Returning to our example 4.1, from the forms \(\alpha\) and \(\tilde{\alpha}\) we conclude that \[\mathcal{H}_{0}(t)=\int_{\mathbb{R}}u(x,t)dx=\int_{\mathbb{R}}m(x,t)dx \tag{4.1.1}\] is a conserved quantity, whereas the first Hamiltonian in (2.0.2) is the conserved quantity emanating from the conserved current \(\beta\). Note that these quantities are only conserved for solutions decaying sufficiently fast as \(|x|\to\infty\). While a conservation law is a structural property of the equation, the same cannot be said for the conserved quantity, e.g, see [62] for a better discussion. However, given a solution of the CH equation for which either (4.1.1) or (2.0.2) is conserved, if the corresponding functional exists for the initial datum, then this property persists and, even stronger, it remains invariant as long as the solution exists. ### Auxiliary and technical results **Lemma 4.1**.: ([12, Proposition 2.7]) _If \(u_{0}\in H^{4}(\mathbb{R})\), then there exists a maximal time \(T=T(u_{0})>0\) and a unique solution \(u\) to the Cauchy problem (3.3.5) such that \(u=u(\cdot,u_{0})\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\). Moreover, the solution depends continuously on the initial data, in the sense that the mapping \(u_{0}\mapsto u(\cdot,u_{0}):H^{4}(\mathbb{R})\to C^{0}(H^{4}(\mathbb{R}),[0,T) )\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\) is continuous._ **Remark 4.1**.: _We observe that if, instead of \(u_{0}\in H^{4}(\mathbb{R})\), we assume \(u_{0}\in H^{s}(\mathbb{R})\), \(s>3/2\), we would then conclude that \(u\in C^{0}(H^{s}(\mathbb{R}),[0,T))\cap C^{1}(H^{s-1}(\mathbb{R}),[0,T))\), for the same \(T\), see [60, Theorem 3.2]._ **Lemma 4.2**.: ([30, Theorem 1.1]) _Assume that \(m_{0}\in H^{2}(\mathbb{R})\cap L^{1}(\mathbb{R})\). If \(m_{0}(x)\geq 0\) or \(m_{0}(x)\leq 0\), for any \(x\in\mathbb{R}\), then the corresponding solution \(u\) of the CH equation exists globally. In other words, the solution \(u\) of the CH equation belongs to the class \(C^{0}(H^{4}(\mathbb{R}),[0,\infty))\cap C^{1}(H^{3}(\mathbb{R}),[0,\infty))\)._ **Lemma 4.3**.: ([14, Theorem 3.1]) _Let \(u_{0}\in H^{3}(\mathbb{R})\) and \([0,T)\) be the maximal interval of existence of the corresponding solution of (3.3.5). Then_ \[\left\{\begin{array}{rcl}q_{t}(x,t)&=&u(q,t),\\ \\ q(x,0)&=&x,\end{array}\right. \tag{4.2.1}\] _has a unique solution \(q\in C^{1}(\mathbb{R}\times[0,T),\mathbb{R})\). Moreover, for every fixed \(t\in[0,T)\), the function \(q(\cdot,t)\) is an increasing diffeomorphism of the line._ **Lemma 4.4**.: ([13, Theorem 4.2]) _Given an initial datum \(u_{0}\in H^{3}(\mathbb{R})\) satisfying (3.3.15), then the corresponding solution \(u\) of the CH equation subject to \(u(x,0)=u_{0}(x)\) breaks at finite time, that is, there exists a finite time \(T_{m}>0\) such that_ \[\lim_{t\to T_{m}}\inf\Big{(}\inf_{x\in\mathbb{R}}u_{x}(t,x)\Big{)}=-\infty. \tag{4.2.2}\] **Lemma 4.5**.: ([13, Theorem 2.1]) _Let \(T>0\) and \(v\in C^{1}(H^{2}(\mathbb{R}),[0,T))\) be a given function. Then, for any \(t\in[0,T)\), there exists at least one point \(\xi(t)\in\mathbb{R}\) such that_ \[y(t):=\inf_{x\in\mathbb{R}}v_{x}(x,t)=v_{x}(\xi(t),t) \tag{4.2.3}\] _and the function \(y\) is almost everywhere differentiable in \((0,T)\), with \(y^{\prime}(t)=v_{tx}(\xi(t),t)\) almost everywhere in \((0,T)\)._ **Lemma 4.6**.: ([37, Theorem 1.4]) _If \(u_{0}\in H^{4}(\mathbb{R})\), is compactly supported, then there exist \(C^{1}\) real valued functions \(E_{\pm}\) such that_ \[u(x,t)=\left\{\begin{array}{ll}E_{+}(t)e^{-x},&\mbox{for}\quad x>q(b,t),\\ \\ E_{-}(t)e^{x},&\mbox{for}\quad x<q(a,t),\end{array}\right.\] _where \(q(\cdot,\cdot)\) is the function given in Lemma 4.3, for any \(t>0\) such that the solution exists._ The original statement of Lemma 4.6 says that \(s>5/2\) and the functions \(E_{\pm}\) are continuous. It is immediate then its validity for \(s=4\), that is our case, and a careful analysis on the proof of [37, Theorem 1.4] reveals that the functions are continuously differentiable. ## 5 Proof of the main results ### Proof of theorem 3.1 From (3.3.2), \(u\in\mathcal{B}\) is a solution of (1.0.1) in the sense of definition 3.2 if and only if it is a solution of (3.3.3) in the same sense. Let \(w_{0}\in H^{4}(\mathbb{R})\), \(u_{1}\) and \(u_{2}\) be the corresponding solutions of (1.0.1) and (3.3.3), respectively, subject to the same initial condition \(u_{1}(x,0)=u_{2}(x,0)=w_{0}(x)\). Proposition 3.1 combined with lemma 4.1 inform us that \(u_{1}=u_{2}\) and this is the only solution for both equations satisfying the given initial condition. As a result, they determine the same forms \(\omega_{1},\omega_{2},\omega_{3}\), and the same PSS as well. ### Proof of theorem 3.2 Lemma 4.1, jointly with remark 4.1 and Theorem 3.1, assures that (3.3.5) has a unique solution \(u\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\subseteq C ^{3,1}(\mathbb{R}\times[0,T))\), for a \(T\) uniquely determined by \(u_{0}\). We then conclude that the one-forms (2.2.2) are \(C^{1}\) and defined on the open and connected set \(\mathcal{S}=\mathbb{R}\times(0,T)\). Due to \(u_{0}\in H^{4}(\mathbb{R})\), then \(\|u_{0}\|_{1}<\infty\). Moreover, the functional \(\mathcal{H}_{1}(t)\), given in (2.0.2), is constant, that is, \(\mathcal{H}_{1}(t)=\mathcal{H}_{1}(0)\), \(t\in(0,T)\). Given that \(t\mapsto\mathcal{H}(t)=\|u\|_{1}^{2}/2\) is invariant, we conclude \(\|u\|_{1}=\|u_{0}\|_{1}\). ### Proof of Theorem 3.3 Let \(u\) be the corresponding solution of the CH equation subject to \(u(x,0)=u_{0}(x)\) and \(q\) be the function given by Lemma 4.3. Define \(\varphi(x,t):\mathbb{R}\times[0,T)\rightarrow\mathbb{R}\times[0,T)\) by \(\varphi(x,t)=(q(x,t),t)\). Then \(\varphi\) is a bijection fixing \(\mathbb{R}\times\{0\}\) and \(\varphi\big{|}_{\mathbb{R}\times(0,T)}\) is a \(C^{1}\) diffeomorphism, see [31, Theorem 3.1]. Let \(\gamma_{\pm}:[0,T)\rightarrow\overline{\mathcal{S}}\) be given by \(\gamma_{-}(t)=\varphi(a,t)\) and \(\gamma_{+}(t)=\varphi(b,t)\). Then \(\gamma_{-}^{\prime}(t)=(u(\varphi(a,t)),1)\) and \(\gamma_{+}^{\prime}(t)=(u(\varphi(b,t)),1)\). Again, by Lemma 4.3 we have \[\pi_{1}(\gamma_{-}(t))=q(a,t)<q(b,t)=\pi_{1}(\gamma_{+}(t)),\] for each \(t\in(0,T)\). Let \(p\in\mathcal{S}\) be a point on the left of \(\gamma_{-}\). This then implies that \[x:=\pi_{1}(p)<\pi_{1}(\gamma_{-}(t))=q(a,t).\] By Lemma 4.6 we have \(u(x,t)=E_{-}(t)e^{x}\), that substituted into (2.2.3) gives (3.3.12). To get (3.3.13) we proceed mimetically as before and for this reason is omitted. ### Proof of theorem 3.4 Let us define \[y(t)=\inf_{x\in\mathbb{R}}u_{x}(x,t). \tag{5.4.1}\] By lemma 4.5 we can find \(\xi(t)\) (despite the notation, it is not a function, see [13, Theorem 2.1]) such that \(y(t)=u_{x}(\xi(t),t)\) and it is an a.e. \(C^{1}\) function. Moreover, [13, Theorem 4.2] shows in its demonstration that \(y\) is Lipschitz and \(y(0)\leq u_{0}^{\prime}(x_{0})<0\). Differentiating (3.3.3) with respect to \(x\) and using \(y(t)\) above, we obtain \[y^{\prime}(t)+\frac{y(t)^{2}}{2}=u(\xi(t),t)^{2}-\Big{(}\partial_{x}\Lambda^{- 2}\Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}\Big{)}(\xi(t),t).\] In [12, page 240] it was proved that \(y(t)\) satisfies the differential inequality \[y^{\prime}(t)\leq-\frac{\epsilon}{4}y(t)^{2},\] for some \(\epsilon\in(0,1)\), implying that it is a negative and non-increasing function satisfying the inequality \[\frac{\epsilon}{4}t+\frac{1}{y(0)}\leq\frac{1}{y(t)}. \tag{5.4.2}\] Since \(y(t)<y(0)<0\), then (5.4.2) is only valid for a finite range of values for \(t\). As a result, we conclude the existence of \(T_{m}\) such that (5.4.2) holds for \(t\in(0,T_{m})\), and then, the solution \(u\), as a function of \(t\), is only defined on \((0,T_{m})\). On the other hand, (5.4.2) can be seen in a slightly different way, since it implies \[0\leq\frac{\epsilon}{4}t-\frac{1}{y(t)}\leq-\frac{1}{y(0)},\] which tells us that \(y(t)\to-\infty\) before \(t\) reaches \(-4/(\epsilon y(0))\) (which gives an upper bound to \(T_{m}\)). As a result, if \((t_{k})_{k}\subseteq(0,T_{m})\) is a convergent sequence to \(T_{m}\), we then have \(y(t_{k})\to-\infty\) as \(k\to\infty\). This, in particular, is nothing but (4.2.2). Let us evaluate the coefficients \(g_{ij}\) of the metric (2.2.3) at \(x=\xi(t)\). The Sobolev Embedding Theorem (see lemma 3.1) implies that \(u\) is uniformly bounded in \((0,T_{m})\) by \(\|u_{0}\|_{1}\). Since \(x=\xi(t)\) is a point of minima of the function \(u_{x}(\cdot,t)\), we conclude that \(u_{xx}(\xi(t),t)=0\) and thus, \(m(\xi(t),t)=u(\xi(t),t)\) is bounded as well. As a result, we conclude that both \(g_{11}(\xi(t),t)\) and \(g_{12}(\xi(t),t)\) are uniformly bounded for \(t\in(0,T_{m})\). A different situation occurs with \(g_{22}\). The previous arguments show that \(g_{22}(\xi(t),t)=u_{x}(\xi(t),t)^{2}+B(u(\xi(t),t))\), where \(B(u(\xi(t),t))\) are the uniformly bounded remaining terms of the metric in \((0,T_{m})\). For any sequence \((t_{k})_{k}\subseteq(0,T_{m})\) convergent to \(T_{m}\), we have \[\sup_{x\in\mathbb{R}}g(x,t_{k})\geq g_{22}(\xi(t_{k}),t_{k})=u_{x}(\xi(t_{k} ),t_{k})^{2}+B(u(\xi(t_{k}),t_{k}))\to+\infty\] as \(k\to\infty\), showing that \[\sup_{(x,t)\in\mathbb{R}\times[0,T_{m})}g_{22}(x,t)=\lim_{t\to T_{m}}\inf_{\tau\geq t }\Big{(}\sup_{x\in\mathbb{R}}g_{22}(x,\tau)\Big{)}=+\infty.\] ### Proof of theorem 3.5 From (2.2.2) we have \[f_{32}(x,t)=-u_{x}(x,t),\] and, as a result, \[\|f_{32}(\cdot,t)\|_{\infty}=\|u_{x}(\cdot,t)\|_{\infty}. \tag{5.5.1}\] Therefrom, for each \(t\) such that the solution exist, we have \[\int_{0}^{t}\|f_{32}(\cdot,\tau)\|_{\infty}\,d\tau=\int_{0}^{t}\|u_{x}(\cdot, \tau)\|_{\infty}\,d\tau. \tag{5.5.2}\] By Theorem 3.1 and the conditions on the initial datum, we conclude that the function defined in (5.5.2) is continuous. Let us prove the existence of a height \(T_{m}<\infty\) such that \(\|f_{32}(\cdot,t)\|_{\infty}\to\infty\) as \(t\to T_{m}\). The maximal height \(T_{m}\) corresponds to the maximal time of existence of the solution. Following [37, Corollary 1.1] or [3, Theorem 6.1], the conditions on the initial datum in Theorem 3.5 imply that the solution \(u\) can only exist for a finite time \(T_{m}\), implying on the existence of a maximal height \(T_{m}\) for the strip in Theorem 3.2. By [37, Corollary 1.1, Eq. (1.20)] we then have \[\int_{0}^{T_{m}}\|u_{x}(\cdot,\tau)\|_{\infty}\,d\tau=\infty.\] On the other hand, the singularities of the solution arise only in the form of wave breaking. Moreover, we have the equivalence (e.g, see [50, page 525, Eq. (3.7)]) \[\int_{0}^{T_{m}}\|u_{x}(\cdot,\tau)\|_{\infty}\,d\tau=\infty\Longleftrightarrow \int_{0}^{T_{m}}\|y(\tau)\|_{\infty}\,d\tau=\infty, \tag{5.5.3}\] where \(y(\cdot)\) is given by (5.4.1). Let \((t_{k})_{k\in\mathbb{N}}\subseteq(0,T_{m})\) be any sequence convergent to \(T_{m}\). By (5.5.3), (5.5.2), (5.4.1) and Lemma 4.5, we have \(y(t_{k})=u(\xi(t_{k}),t_{k})<\infty\) and \[\int_{0}^{t_{k}}\|f_{32}(\cdot,\tau)\|_{\infty}\,d\tau<\infty,\] for any \(k\in\mathbb{N}\), but \[\lim_{k\to\infty}\int_{0}^{t_{k}}\|f_{32}(\cdot,\tau)\|_{\infty}\,d\tau=\infty,\] meaning that \(|f_{32}(x,t)|\) becomes unbounded near some point of the line \(\mathbb{R}\times\{T_{m}\}\). Since \(g_{22}(x,t)\geq f_{32}(x,t)^{2}\), we have \[\sup_{x\in\mathbb{R}}g(x,t_{k})\geq f_{32}(\xi(t_{k}),t_{k})\to\infty\] as \(k\to\infty\), and we then get again \[\sup_{(x,t)\in\mathbb{R}\times[0,T_{m})}g_{22}(x,t)=\lim_{t\to T_{m}}\inf_{\tau\geq t }\Big{(}\sup_{x\in\mathbb{R}}g_{22}(x,\tau)\Big{)}=+\infty, \tag{5.5.4}\] which proves the result. We can give a slightly different prove starting from (5.5.3). In fact, that condition implies on the wave breaking of the solution. According to McKean [48, 49], this only happens if and only if the points for which \(m_{0}(x)\) is positive lies to the left of those that \(m_{0}(x)\) is negative, see also [38, Theorem 1.1]. In other words, for some \(x_{0}\in\mathbb{R}\), we have \(m_{0}(x_{0})\geq 0\), for \(x\leq x_{0}\), whereas for \(x\geq x_{0}\) we have \(m_{0}(x_{0})\leq 0\). By [31, Theorem 3.3], we get back to (5.5.4). ### Proof of theorem 3.6 By lemma 4.2, \(u\) is a global solution in the class \(C^{0}(H^{4}(\mathbb{R}),[0,\infty))\cap C^{0}(H^{3}(\mathbb{R}),[0,\infty))\). In particular, it is defined on \(\mathcal{S}=\mathbb{R}\times(0,\infty)\) and, therefore, the coefficients \(f_{ij}\), \(1\leq i\leq 3\), \(1\leq j\leq 2\), of the one-forms (2.2.2) belong to the class \(C^{3,1}(\mathbb{R}\times(0,\infty))\subseteq C^{1}(\mathbb{R}\times(0,\infty))\), and then, \(g_{kl}\in C^{1}(\mathbb{R}\times(0,\infty))\), \(1\leq k,l\leq 2\). By corollary 3.1 we know that \(\{\omega_{1},\omega_{2}\}\) cannot be linearly independent everywhere. Let \(R>0\), \(\overline{B}_{R}(0);=\{(x,t)\in U;\ x^{2}+t^{2}\leq R^{2}\}\), and \(W_{R}:=U\setminus\overline{B}_{R}(0)\). Suppose that for some \(R>0\) we had \(u_{x}\big{|}_{W_{R}}=0\). Then \(u\big{|}_{W_{R}}=c\), for some \(c\in\mathbb{R}\), and since \(u\in L^{2}(\mathbb{R})\), we would conclude that \(c=0\), resulting in \(u\big{|}_{\mathcal{R}}=0\), for any open set \(\mathcal{R}\subseteq W_{R}\). Therefore, we can find numbers \(t_{0}>R\) and \(b>a>R\) such that \([a,b]\times\{t_{0}\}\subseteq\mathcal{R}\), \(u(x,t_{0})=u_{t}(x,t_{0})=0\), \(a\leq x\leq b\). From (3.3.3) we obtain \[\partial_{x}\Lambda^{-2}\Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}(x,t)=-\Big{(} u_{t}+uu_{x}\Big{)}(x,t).\] Evaluating at \(t=t_{0}\) and letting \(x\in(a,b)\), we conclude that \[F(x):=\partial_{x}\Lambda^{-2}\Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}(x,t_{0}) =-\Big{(}u_{t}+uu_{x}\Big{)}(x,t_{0})\equiv 0,\] implying \(F^{\prime}(x)=0\), \(x\in(a,b)\). Since \(\partial_{x}^{2}\Lambda^{2}=\Lambda^{-2}-1\), we get \[0=F^{\prime}(x)=\Lambda^{-2}\Big{(}u^{2}+\frac{u_{x}^{2}}{2}\Big{)}(x,t_{0})= \frac{1}{2}\int_{\mathbb{R}}\frac{e^{-|x-y|}}{2}\Big{(}u^{2}+\frac{u_{x}^{2}} {2}\Big{)}(y,t_{0})dy,\quad x\in(a,b),\] wherefrom we arrive at the conclusion \(u(x,t_{0})\equiv 0\), \(x\in\mathbb{R}\). This would then imply \(\|u\|_{1}=0\) at \(t=t_{0}\). The invariance of \(\|u\|_{1}\) implies \(u\equiv 0\), that conflicts with \(u_{0}\) being a non-trivial initial datum. The contradiction above forces us to conclude that, for any \(R>0\), we can find \((x_{R},t_{R})\in W_{R}\) such that \(u_{x}(x_{R},t_{R})\neq 0\), meaning that we either have \(u_{x}(x_{R},t_{R})>0\) or \((x_{R},t_{R})<0\). Since \(u_{x}\) is continuous, we can find a neighborhood \(V_{R}\) of \((x_{R},t_{R})\) such that \(u_{x}\big{|}_{V_{R}}\) has the same sign. ## 6 Finite height vs finite time of existence The results proved in [31] and those in theorems 3.4 and 3.5 suggest that the metric blows up as long as the solution develops a wave breaking. This is, indeed, the case. **Theorem 6.1**.: _Let \(u\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\) be a solution of the CH equation and \(g_{22}\) be the corresponding component of the metric tensor given in (2.2.3). Then \(g_{22}\) blows up within a strip of finite height if and only if \(u\) breaks in finite time._ Proof.: Let \(q\) be the function given in Lemma 4.3 and \(\varphi(x,t)=(q(x,t),t)\) be the bijection given in the proof of Theorem 3.3 (see subsection 5.3). As long as the solution exists for \(t>0\) and taking (2.0.1) into account, we have \[\frac{d}{dt}m(\varphi(x,t))=(m_{t}+um_{x})(\varphi(x,t))=-2(u_{x}m)(\varphi(x, t)),\] that is, \[m(\varphi(x,t))=m_{0}(x)e^{-2\int_{0}^{t}u_{x}(\varphi(x,\tau))d\tau}. \tag{6.0.1}\] From (2.2.3) we obtain \[u_{x}(x,t)^{2}\leq g(x,t)\leq u_{x}(x,t)^{2}+(um)(x,t)^{2}+\text{lower order terms},\] that, after neglecting the lower order terms, implies \[|u_{x}(x,t)|\leq\sqrt{g(x,t)}\leq|u_{x}(x,t)|+|(um)(x,t)|. \tag{6.0.2}\] Since \(u\in C^{0}(H^{4}(\mathbb{R}),[0,T))\cap C^{1}(H^{3}(\mathbb{R}),[0,T))\) and \(m_{0}(x)=m(x,0)=u(x,0)-u_{xx}(x,0)\), Lemma 3.1 implies that \(\|m_{0}\|_{L^{\infty}}<\infty\). Moreover, we also have \[\|u(\cdot,t)\|_{L^{\infty}}<\|u\|_{1}=\sqrt{2\mathcal{H}_{1}(0)},\] where \(\mathcal{H}_{1}(t)\) is given by (2.0.2). These two estimates, put together into (6.0.2), imply \[|u_{x}(x,t)|\leq\sqrt{g(x,t)}\leq|u_{x}(x,t)|+\sqrt{2\mathcal{H}_{1}(0)}\|m_{ 0}\|_{L^{\infty}}e^{2\int_{0}^{t}\|u_{x}(\cdot,\tau)\|_{L^{\infty}}d\tau}, \tag{6.0.3}\] where we used that \(\|u_{x}(\cdot,\tau)\|_{L^{\infty}}=\|u_{x}(\varphi(x,\tau))\|_{L^{\infty}}\). Inequalities (6.0.3) combined with (5.5.3) show that \(g_{22}\) blows up in a strip of finite height if and only if \(u_{x}\) blows up in finite time. Hence, we have \[\sup_{(x,t)\in\mathbb{R}\times[0,T)}g_{22}(x,t)=\infty\Longleftrightarrow\liminf _{t\to T}\big{(}\inf_{x\in\mathbb{R}}u_{x}(x,t)\big{)}=-\infty.\] In particular, the maximal height of the strip coincides with the maximal time of existence of the solutions. ## 7 Examples We give two examples illustrating qualitative aspects of the surfaces determined by solutions of the CH equation once an initial datum is known. **Example 7.1**.: _Let us consider \(m_{0}(x)=e^{-x^{2}}\). As a consequence of (5.4.2), \(m>0\) and so does the corresponding solution \(u\). As a result of theorem 3.1 and its corollaries, \(u\) is a generic solution of the CH equation in the sense of definition 3.4._ _By theorem 3.6, the one-forms (2.2.2) are defined on \(\mathcal{S}=\mathbb{R}\times(0,T)\) and they endow an infinite number of simply connected open sets \(\Omega\subseteq U\) with the structure of a PSS._ **Example 7.2**.: _Let us now consider the family of functions \(\phi_{n}(x)=e^{-nx^{2}}\), \(n\in\mathbb{N}\) and \(x\in\mathbb{R}\). As pointed out in [13, Example 4.3], for \(n\) sufficiently large, we have_ \[\phi_{n}^{\prime}(x)<-\frac{\|\phi_{n}\|_{1}}{\sqrt{2}}. \tag{7.0.1}\] _Fix \(n\) large enough so that (7.0.1) and choose \(u_{0}=\phi_{n}\). As a consequence of theorem 3.4, we know that \(g_{22}\) blows up for some \(x\in\mathbb{R}\) as long as \(t\) approaches some value \(T_{m}\) determined by the initial datum._ We close this section with some words about the maximal time \(T_{m}\) of existence (lifespan) of a solution of the CH equation emanating from an initial datum in Sobolev spaces. From theorem 3.2 we know that \(u\), and the metric as well, will become unbounded before reaching a certain value determined by the initial datum. The question is: do we have any sort of information about how it is determined? An answer for this question is provided by [19, Theorem 0.1], which shows a lower bound for it: \[T_{m}\geq T(u_{0}):=-\frac{2}{\|u_{0}\|_{1}}\arctan\Big{(}\frac{\|u_{0}\|_{1} }{\inf_{x\in\mathbb{R}}u_{0}^{\prime}(x)}\Big{)}.\] For the initial datum \(u_{0}(x)=e^{-nx^{2}}\) considered in example 7.2, we have \[T(u_{0})=2\sqrt[4]{\frac{2n}{\pi(n+1)^{2}}}\arctan\Big{(}\sqrt[4]{\frac{\pi e^ {2}(n+1)^{2}}{8n^{3}}}\Big{)}.\] In particular, for \(n\gg 1\), we have \[T(u_{0})=\sqrt{\frac{2e}{n}}+O(n^{-1}).\] As a consequence of the quantities shown above, for the given initial datum in example 7.2 we can surely guarantee that only certain open, properly and simply connected sets contained in \[\mathcal{S}=\mathbb{R}\times\Big{(}0,2\sqrt[4]{\frac{2n}{\pi(n+1)^{2}}}\arctan \Big{(}\sqrt[4]{\frac{\pi e^{2}(n+1)^{2}}{8n^{3}}}\Big{)}\Big{)}\] can be endowed with a PSS structure. ## 8 Discussion The connection between surfaces of constant Gaussian curvature \(\mathcal{K}=-1\) has a long history in differential geometry, dating back to the first half part of century XIX [61, page 17], see also [11, chapter 9] and [65, chapter 1]. Roughly half a century ago, a hot topic in mathematical physics emerged after certain hydrodynamics models, more precisely, the KdV equation, was shown to have remarkable properties [33]. In [47] there is a survey of results about the KdV equation and its importance for nourishing a new-born field whose most well known representative is just itself. An explosion of works was seen during the 60 and 70's after [33] exploring properties of the KdV equation, while other quite special equations were also discovered sharing certain properties with the KdV. In this context was proposed the AKNS method [1], which reinvigorated and boosted the field emerged after the KdV, currently called _integrable equations_ (very roughly and naively speaking, an equation sharing properties with the KdV equation). By that time, the interest on this sort of equations spred out fields, attracting people more inclined to analysis of PDEs and geometric aspects of these equations. By the end of the 70's, [63] Sasaki showed an interesting connection between equations described by the AKNS method [1] and surfaces of Gaussian curvature \(\mathcal{K}=-1\), culminating in the seminal work by Chern and Tenenblat [10, section 1] who established the basis for what today is known as _PSS equations_. These works are roots for what Reyes called _geometric integrability_, see [55, 57, 58, 59]. Equation (1.0.1) was discovered in [24], but became famous after its derivation as a hydrodynamic model in a paper by Camassa and Holm [6], and named after then, see also the review [25]. Despite its physical relevance, like other integrable models physically relevant, it attracted the interests of different areas. Probably one of the most impacted was just analysis of PDEs. In particular, the works by Constantin and co-workers [5, 12, 13, 14, 15] payed a crucial role, creating and developing new tools for tackling the CH equation that would later be explored not only to the CH equation itself, but also for other similar models, see [19, 26, 27, 29, 36, 37, 45] to name a few. Most of these works, not to say all, deal with solutions of the CH equation with finite regularity. Apparently, Constantin [14] was the first showing connections between the CH equation and the geometry of manifolds. However, it was not before the fundamental work by Reyes [56] that it was recognised as a PSS equation. Even though these two works are concerned with the same object (the CH equation), they are completely different in nature. In fact, the results reported by Constantin [14] are intrinsically related to Cauchy problems involving the CH equation, whereas those shown by Reyes are concerned with structural aspects of the equation itself, such as integrability and abstract two-dimensional surfaces. The work by Reyes was followed by a number of works dealing with geometric aspects of CH type equations \(\dot{a}\)_la_ Chern and Tenenblat, see [7, 8, 17, 18, 28, 57, 58, 59, 62] and references therein. Despite a tremendous research carried out since the works [5, 12, 13, 60, 14] and [55, 57], it is surprising that until now very few attention has been directed to geometric aspects of well-posed solutions and PSS equations. As far as I know, the first paper trying to make such a connection is [62], where qualitative analysis of a certain PSS equation was used to describe aspects of the corresponding metric. However, even this reference considered an analytic solution. A second attempt addressing this topic is [22], where Cauchy problems involving the equation considered in [28, 62] were studied. In spite of the efforts made in [22, 62], these works do not deal with solutions blowing up, which was first considered in [31]. In [31] the notions of \(C^{k}-\)PSS and generic solutions were first considered and the blow up of metrics determined by the solutions of the CH equation were shown for two situations, depending on how the sign of the momentum behaves, see [31, theorems 3.2 and 3.4]. However, no problems related to global nature, i.e, circumstances in which the co-frame can be defined on \(\mathbb{R}\times(0,\infty)\) or asymptotic behaviors of metrics, were considered. The notions of generic solutions and PSS equations used in the current literature carry intrinsically \(C^{\infty}\) regularity and this brings issues in the study of surfaces in connection with Cauchy problems. This is even more dramatic for equations like the CH because they have different representations depending on the sort of solutions one considers and they only coincide on certain Banach spaces. This explains why in [31] it was needed to step forward and introduce definitions 3.4 and 3.5. Another important aspect of the connections between geometry and analysis made in the present paper is the condition \(\omega_{1}\wedge\omega_{2}\neq 0\). Whenever \(\omega_{1}\wedge\omega_{2}=0\) we have (3.3.9) holding on an open set \(\Omega\). This is a problem of unique continuation of solutions, whose answer would be impossible quite few time ago. For \(c=0\), the answer for arbitrary open sets was given very recently in [45], see also [26, 27, 29]. As long as \(u\big{|}_{\Omega}=0\), for some open set \(\Omega\), then \(u\equiv 0\), see [45, Theorem 1.3]. Our solutions emanate from a non-trivial initial datum, and then, we cannot have \(u\equiv 0\) on an open set \(\Omega\) contained in the domain of \(u\). For \(c\neq 0\), it is unclear if we might have \(u\big{|}_{\Omega}=c\) since this unique continuation problem is still an open question, see [31, Discussion]. The proof of Corollary 3.1 shows that \(u_{x}(x,t)\) vanishes at least once, for each \(t\) as long as \(u\) is defined, see also [31, Theorem 2.3]. As a result, the domain of \(u\) cannot be wholly endowed with a PSS structure. The answer for the open question mentioned above would only clarify whether we may have open sets that cannot be endowed with a PSS structure (those in which \(u\) is constant). If its answer is that \(c=0\) (which I conjecture, it is the case), then Corollary 3.1 would imply that the domain of the solution has a non-countable set of points in which we loss the structure of a PSS equation, but such a set would be of zero measure. On the other hand, if the answer to the question is that we may have \(c\neq 0\), then we would have a situation whose geometric implication should be better understood, but would surely imply on the existence of subsets of the domain of \(u\), with positive measure, in which a PSS structure is not allowed. Event though the ideas developed and better explored in this paper are mostly concerned with the CH equation, they can used to other PSS equations. The main point is that the techniques to deal with Cauchy problems may vary depending on the equation, and this will then impact in how to address the geometric problem. This can be seen by comparing the results established in the present paper and those in [22, 62]. In a subsequent paper [32], qualitative aspects of PSS equations determined by the solutions of the Degasperis-Procesi equation will be reported. As will be shown, the ideas introduced in the present manuscript can be applied to this model, but paying the price that they have to be customised in view of the different nature of the equation and the corresponding Cauchy problems. ## 9 Conclusion In this paper we studied the influence of Cauchy problems and the PSS surfaces defined by the corresponding solutions. To this end, we had to propose a formulation of the notion of PSS equation and generic solutions. Our main results are reported in subsection 3.3, including the new already mentioned definitions. A remarkable fact reported is that any non-trivial initial datum gives rise to a PSS equation. In particular, we showed solutions breaking in finite time lead to metrics having blow up either. ## Acknowledgements I am thankful to the Department of Mathematical Sciences of the Loughborough University for the warm hospitality and amazing work atmosphere. I am grateful to Jenya Ferapontov, Keti Tenenblat, Artur Sergyeyev and Sasha Veselov for stimulating discussions, as well as many suggestions given. I am also indebted to Priscila Leal da Silva for her firm encouragement, support, suggestions and patience to read the first draft of this manuscript. Last but not least, I want to thank CNPq (grant n\({}^{\text{o}}\) 310074/2021-5) and FAPESP (grants n\({}^{\text{o}}\) 2020/02055-0 and 2022/00163-6) for financial support.

2306.05990

Novikov type inequalities for orbifolds

We show a natural extension of the Novikov numbers associated to the basic cohomology class of a closed $1$-form on an orbifold, thus proving corresponding Novikov inequalities for the compact case.

Fabricio Valencia

2023-06-09T15:59:35Z

http://arxiv.org/abs/2306.05990v1

# Novikov type inequalities for orbifolds ###### Abstract. We show a natural extension of the Novikov numbers associated to the basic cohomology class of a closed 1-form on an orbifold, thus proving corresponding Novikov inequalities for the compact case. 2020 Mathematics Subject Classification: 22A22, 57R18, 57R70 ###### Contents * 1 Introduction * 2 Some algebraic topology ingredients * 2.1 Hurewicz \(G\)-homomorphism and coverings * 2.2 Groupoid homology with local coefficients * 3 Closed basic 1-forms * 3.1 G-homomorphisms of periods * 3.2 Closed basic 1-forms of Morse type * 4 Novikov numbers for orbifolds * 4.1 Novikov inequalities ## 1. Introduction In the seminal works [17, 18] Novikov started a generalization of Morse theory in which instead of critical points of smooth functions he dealt with closed 1-forms and their zeros, thus obtaining inequalities that generalize the well known Morse inequalities. In order to do so Novikov defined numbers \(b_{j}(\xi)\) and \(q_{j}(\xi)\) which depend on a real cohomology class \(\xi\in H^{1}(M,\mathbb{R})\). Such numbers are known in the literature as the Novikov Betti and torsion numbers, respectively. It is worth mentioning that in the special case \(\xi=0\), which corresponds to classical Morse theory, the numbers \(b_{j}(\xi)\) recover the usual Betti numbers of the ambient manifold \(M\) and the numbers \(q_{j}(\xi)\) agree with the minimal number of generators of the torsion subgroup \(H^{1}(M,\mathbb{Z})\) inside \(H^{1}(M,\mathbb{R})\). The classical Novikov inequalities state that any closed 1-form \(\omega\) with Morse-type zeros has at least \(b_{j}(\xi)+q_{j}(\xi)+q_{j-1}(\xi)\) zeros of index \(j\) where \(\xi=[\omega]\in H^{1}(M,\mathbb{R})\) is the de Rham cohomology class of \(\omega\). Nowadays, Novikov theory is widely known, as it provides several applications in geometry, topology, analysis, and dynamics. The reader is recommended to visit [11] in order to get in touch with the fundamentals of Novikov theory as well as with some of the classical results around this beautiful subject. The main purpose of this paper is to extend some of the ingredients of the classical Novikov theory to the context of orbifolds, thus providing a way to generalize the Novikov inequalities for the compact ones. We think of orbifolds as proper and etale groupoids up to Morita equivalence. In this spirit, we deal with cohomology classes of closed basic \(1\)-forms on Lie groupoids whose critical orbits are nondegenerate in the sense of Morse-Bott theory. Let \(G\rightrightarrows M\) be a Lie groupoid and let \(\Pi_{1}(G)\rightrightarrows M\) denote its fundamental groupoid of \(G\)-homotopy classes of \(G\)-paths. The fundamental group \(\Pi_{1}(G,x_{0})\) of \(G\) with respect to a base-point \(x_{0}\in M\) is defined to be isotropy group \(\Pi_{1}(G)_{x_{0}}\) which consists of \(G\)-homotopy classes of \(G\)-loops at \(x_{0}\). Let us denote by \(H_{\bullet}(G,\mathbb{Z})\) the total singular homology of \(G\). There is a canonical way to define a Hurewicz \(G\)-homomorphism \(h:\Pi_{1}(G,x_{0})\to H_{1}(G,\mathbb{Z})\) which restricts as an isomorphism to the abelianization of \(\Pi_{1}(G,x_{0})\). Let \(\omega\) be a closed basic \(1\)-form on \(M\). That is, a closed \(1\)-form satisfying \((t^{*}-s^{*})(\omega)=0\). For each smooth \(G\)-path \(\sigma=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) in \(M\) we can naturally define the \(G\)-path integral \(\int_{\sigma}\omega=\sum_{k=0}^{n}\int_{\sigma_{k}}\omega\). It canonically determines a group homomorphism \(l_{\omega}:\Pi_{1}(G,x_{0})\to(\mathbb{R},+)\) which factors through the Hurewicz \(G\)-homomorphism \(h\) by a uniquely determined group homomorphism \(\operatorname{Per}_{\xi}:H_{1}(G,\mathbb{Z})\to\mathbb{R}\) which only depends on the basic cohomology class \(\xi:=[\omega]\). This will be called the \(G\)-homomorphism of periods of \(\xi\). Let \(\mathbf{Nov}\) denote the Novikov ring of \(\mathbb{R}\) and let us further assume that our Lie groupoid \(G\rightrightarrows M\) is etale and proper with compact orbit space \(M/G\). Thus, we may define a ring homomorphism \(\phi_{\xi}:\mathbb{Z}(\Pi_{1}(G,x_{0}))\to\mathbf{Nov}\) by setting \(\phi_{\xi}([\sigma]):=\tau^{\operatorname{Per}_{\xi}(h([\sigma]))}\) for all \([\sigma]\in\Pi_{1}(G,x_{0})\). It yields a local system \(\mathcal{L}_{\xi}\) of left \(\mathbf{Nov}\)-modules over \(G\rightrightarrows M\) whose total homology groups \(H_{j}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\) are called Novikov homology groups of \(\xi\). It follows that the homology \(H_{j}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\) is a finitely generated module over the ring \(\mathbf{Nov}\). Since \(\mathbf{Nov}\) is a principal ideal domain we have that the module \(H_{j}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\) is a direct sum of a free submodule with a torsion submodule. The Novikov Betti number \(b_{j}(\xi)\) is defined to be the rank of the free summand of \(H_{j}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\) and the Novikov torsion number \(q_{j}(\xi)\) is defined to be the minimal number of generators of the torsion submodule of \(H_{j}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\). The critical point set \(\operatorname{Crit}(\omega)\) of a closed basic \(1\)-form \(\omega\) on \(M\) is saturated in \(M\) so that it is formed by a disjoint union of groupoid orbits. Therefore, we say that a critical orbit \(\mathcal{O}_{x}\) of \(\omega\) is nondegenerate if its normal Hessian is a nondegenerate fiberwise bilinear symmetric form on \(\nu(O_{x})\). Accordingly, \(\omega\) is said to be of Morse type if all of its critical orbits are nondegenerate. Let us fix a groupoid metric on \(G\rightrightarrows M\) in the sense of del Hoyo and Fernandes, visit [9]. The nondegeneracy requirement over \(\mathcal{O}_{x}\) imposed above allows us to use the groupoid metric to split \(\nu(\mathcal{O}_{x})\) into the Whitney sum of two subbundles \(\nu_{-}(\mathcal{O}_{x})\oplus\nu_{+}(\mathcal{O}_{x})\) such that normal Hessian of \(\omega\) around \(\mathcal{O}_{x}\) is strictly negative on \(\nu_{-}(\mathcal{O}_{x})\) and strictly positive on \(\nu_{+}(\mathcal{O}_{x})\). Let \(G_{x}\) be the isotropy group at \(x\). From [19] we know that the normal Hessian is invariant with respect to the normal representation \(G_{x}\curvearrowright\nu(\mathcal{O}_{x})\) so that it preserves the splitting above since the normal representation is by isometries in this case. In consequence, we get a normal subrepresentation \(G_{x}\curvearrowright\nu_{-}(\mathcal{O}_{x})\). The stacky index of \(\mathcal{O}_{x}\) in \(M/G\) is defined to be \(\dim\nu_{-}(\mathcal{O}_{x})/G_{x}=\dim\nu_{-}(\mathcal{O}_{x})-\dim G_{x}\). Our main result can be stated in the following way: **Theorem**.: _Let \(G\rightrightarrows M\) be an etale and proper Lie groupoid such that the orbit space \(M/G\) is compact. Let \(\omega\) be a closed basic \(1\)-form on \(M\) of Morse type. If \(c_{j}(\omega)\) denotes the number of critical points in \(M/G\) having stacky Morse index \(j\) then_ \[c_{j}(\omega)\geq b_{j}(\xi)+q_{j}(\xi)+q_{j-1}(\xi),\] _where \(\xi=[\omega]\) is the basic cohomology class of \(\omega\)._ It is worth mentioning that the natural generalization of the Novikov theory we have described above provides a tool for using topological methods to study zeros of symplectic vector fields on symplectic orbifolds. The paper is structured as follows. In Section 2 we define the fundamental Lie groupoid of \(G\)-homotopy classes of \(G\)-paths and briefly recall the definition of the total singular homology and cohomology groups of a topological groupoid. We show a \(G\)-version of the Hurewicz homomorphism and exhibit how to construct a groupoid covering space over \(G\rightrightarrows M\) out of a fixed subgroup in \(\Pi_{1}(G,x_{0})\). This is the content of Propositions 2.2 and 2.4, respectively. Motivated by the notion of local coefficients system for topological spaces we introduce a corresponding notion as well as its associated homology in the groupoid setting. Indeed, we define a double chain complex out of a local system of modules over a Lie groupoid, thus proving that its associated total homology is Morita invariant, compare Lemma 2.7 and Proposition 2.8. We also show in Proposition 2.9 a \(G\)-version of the Eilenberg homology isomorphism. In Section 3 we study closed basic \(1\)-forms. We define the \(G\)-homomorphism of periods associated to the basic cohomology class \(\xi\) of a closed basic \(1\)-form \(\omega\) on \(M\) and explore some of its elementary properties. In particular, we characterize when \(\xi\) is an integral class and describe the groupoid covering space associated to the kernel of the \(G\)-homomorphism of periods of \(\xi\), see Propositions 3.4 and 3.5. We introduce closed basic \(1\)-forms of Morse type and prove that such a notion is Morita invariant, see Proposition 3.8. This gives rise to a notion of stacky closed \(1\)-form of Morse type over the differentiable stack \([M/G]\) presented by \(G\rightrightarrows M\). Finally, in Section 4 we present three different manners to define the Novikov numbers \(b_{j}(\xi)\) and \(q_{j}(\xi)\) associated to the basic cohomology class \(\xi\) over an orbifold. As a result we show the Novikov inequalities for compact orbifolds, compare Theorem 4.6, and quickly explain how to use this result in order to find a lower bound for the amount of zeros of certain symplectic vector fields on symplectic orbifolds. **Acknowledgments:** Part of this work was carried out during a visit to the Dipartimento di Matematica, Universita degli Studi di Salerno, Fisciano, Italy in 2023. I am very thankful for the hospitality and support that the Geometry Group gave me while being there. I have benefited from several conversations with Juan Camilo Arias, Antonio Maglio, Cristian Ortiz and Luca Vitagliano, so that I am grateful for all their comments and suggestions that improved this work. Valencia was supported by Grants 2020/07704-7 and 2022/11994-6 Sao Paulo Research Foundation - FAPESP. ## 2. Some algebraic topology ingredients Before defining the \(G\)-homomorphism of periods of a closed basic \(1\)-form as well as its corresponding Novikov numbers we have to fill out some algebraic topology gaps concerning the fundamental Lie groupoid of \(G\)-homotopy classes of \(G\)-paths which, to our knowledge, do not seem to have been described before in the literature; see for instance [16, s. 3.3] and [4, c. G; s. 3]. Let \(G\rightrightarrows M\) be a Lie groupoid [8, 16]. Throughout this paper the structural maps of \(G\) will be denoted by \((s,t,m,u,i)\) where \(s,t:G\to M\) are the maps respectively indicating the source and target of the arrows, \(m:G^{(2)}\to G\) stands for the partial composition of arrows, \(u:M\to G\) is the unit map, and \(i:G\to G\) is the map determined by the inversion of arrows. A \(G\)_-path_ in \(M\) is a sequence \(\sigma:=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) where \(\sigma_{0},\cdots,\sigma_{n}:[0,1]\to M\) are (piecewise smooth) paths in \(M\) and \(g_{1},\cdots,g_{n}\) are arrows in \(G\) such that \(g_{j}:\sigma_{j-1}(1)\to\sigma_{j}(0)\) for all \(j=1,\cdots,n\). We shall say that \(\sigma\) is a \(G\)-path of _order_\(n\) from \(\sigma_{0}(0)\) to \(\sigma_{n}(1)\). Our groupoid \(G\) is said to be \(G\)_-connected_ if for any two points \(x,y\in M\) there exists a \(G\)-path from \(x\) to \(y\). We will always assume that the groupoids we are working with are \(G\)-connected unless otherwise stated. If \(\sigma^{\prime}:=\sigma^{\prime}_{n}g^{\prime}_{n}\sigma^{\prime}_{n-1}\cdots \sigma^{\prime}_{1}g^{\prime}_{1}\sigma^{\prime}_{0}\) is another \(G\)-path with \(\sigma^{\prime}_{0}(0)=\sigma_{n}(1)\) then we can _concatenate_\(\sigma\) and \(\sigma^{\prime}\) into a new \(G\)-path \[\sigma*\sigma^{\prime}=\sigma^{\prime}_{n}g^{\prime}_{n}\sigma^{\prime}_{n-1} \cdots\sigma^{\prime}_{1}g^{\prime}_{1}\sigma^{\prime}_{0}1_{\sigma_{n}(1)} \sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}.\] We define an equivalence relation in the set of \(G\)-paths which is generated by the following _multiplication equivalence_ \[\sigma_{n}g_{n}\cdots\sigma_{j+1}g_{j+1}\sigma_{j}g_{j}\sigma_{j-1}\cdots g_{1 }\sigma_{0}\quad\sim\quad\sigma_{n}g_{n}\cdots\sigma_{j+1}g_{j+1}g_{j}\sigma_ {j-1}\cdots g_{1}\sigma_{0},\] if \(\sigma_{j}\) is the constant path for any \(0<j<n\), and _concatenation equivalence_ \[\sigma_{n}g_{n}\cdots g_{j+1}\sigma_{j}g_{j}\sigma_{j-1}g_{j-1}\cdots g_{1} \sigma_{0}\quad\sim\quad\sigma_{n}g_{n}\cdots g_{j+1}\sigma_{j}\cdot\sigma_{j -1}g_{j-1}\cdots g_{1}\alpha_{0},\] if \(g_{j}=1_{\alpha_{j-1}(1)}\) for any \(0<j<n\) where \(\sigma_{j}\cdot\sigma_{j-1}\) stands to denote the standard concatenation of the paths \(\sigma_{j}\) and \(\sigma_{j-1}\). A _deformation_ between two \(G\)-paths \(\sigma\) and \(\sigma^{\prime}\) of the same order \(n\) from \(x\) to \(y\) consist of homotopies \(D_{j}:[0,1]\times[0,1]\to M\) from \(\sigma_{j}\) to \(\sigma^{\prime}_{j}\) for \(j=0,1,\cdots,n\) and paths \(d_{i}:[0,1]\to G\) from \(g_{j}\) to \(g^{\prime}_{j}\) for \(j=1,\cdots,n\) such that \(s\circ d_{j}=D_{j-1}(\cdot,1)\) and \(t\circ d_{j}=D_{j}(\cdot,0)\) for \(j=1,\cdots,n\) verifying \(D_{0}([0,1],0)=x\) and \(D_{n}([0,1],1)=y\). That is, a deformation as a continuous family of \(G\)-paths of order \(n\) from \(x\) to \(y\) which may be written as \(D_{n}(\tau,\cdot)d_{n}(\tau)\cdots d_{1}(\tau)D_{0}(\tau,\cdot)\) for \(\tau\in[0,1]\). Accordingly, two \(G\)-paths with fixed endpoints in \(M\) are said to be \(G\)_-homotopic_ if it is possible to pass from one to another by a sequence of equivalences and deformations. With the multiplication induced by concatenation it follows that the \(G\)-homotopy classes of \(G\)-paths form a Lie groupoid over \(M\) which is called the _fundamental groupoid_ of \(G\). This shall be denoted by \(\Pi_{1}(G)\rightrightarrows M\), see [16, s. 3.3]. The _fundamental group_ of \(G\) with respect to a base-point \(x_{0}\in M\) is the isotropy group \(\Pi_{1}(G,x_{0}):=\Pi_{1}(G)_{x_{0}}\). Note that it consists of \(G\)-homotopy classes of \(G\)-loops at \(x_{0}\) which are by definition the \(G\)-homotopy classes of \(G\)-paths from \(x_{0}\) to \(x_{0}\). It is simple to check that for any two different points \(x_{0},y_{0}\in M\) it holds that \(\Pi_{1}(G,x_{0})\) and \(\Pi_{1}(G,y_{0})\) are isomorphic by \(G\)-connectedness. Every Lie groupoid morphism \(\phi_{1}:G\to G^{\prime}\) covering \(\phi_{0}:M\to M^{\prime}\) induces another Lie groupoid morphism \((\phi_{1})_{*}:\Pi_{1}(G)\rightarrow\Pi_{1}(G^{\prime})\) by mapping \([\sigma]\) to \([(\phi_{1})_{*}(\sigma)]\), where \((\phi_{1})_{*}(\sigma)=(\phi_{0})_{*}(\sigma_{n})\phi_{1}(g_{n})(\phi_{0})_{*} (\sigma_{n-1})\cdots(\phi_{0})_{*}(\sigma_{1})\phi_{1}(g_{0})(\phi_{0})_{*}( \sigma_{0})\). This also covers \(\phi_{0}:M\to M^{\prime}\) so that it induces a Lie group homomorphism between the fundamental groups \((\phi_{1})_{*}:\Pi_{1}(G,x_{0})\rightarrow\Pi_{1}(G^{\prime},\phi_{0}(x_{0}))\). As an important feature we have that Morita equivalent groupoids have isomorphic fundamental groupoids, see [16, p. 195]. For specific details and results concerning the fundamental groupoid of a Lie groupoid the reader is recommended to visit [16, s. 3.3] and [4, c. G; s. 3]. Let us now consider the nerve of \(G\rightrightarrows M\). This is formed by manifolds \(\{G^{(n)}\}_{n\in\mathbb{N}}\) with \(G^{(0)}=M\), \(G^{(1)}=G\) and \(G^{(n)}=\{(g_{1},\cdots,g_{n}):s(g_{j})=t(g_{j+1})\}\). We also have the face maps \(d_{k}^{n}:G^{(n)}\to G^{(n-1)}\) given by the surjective submersions \(d_{0}^{1}=t\), \(d_{1}^{1}=s\) and \[d_{k}^{n}(g_{1},\cdots,g_{n})=\left\{\begin{array}{lll}(g_{2},\cdots,g_{n})& \mbox{if}&k=0\\ (g_{1},\cdots,g_{i}g_{i+1},\cdots,g_{n})&\mbox{if}&0<k<n\\ (g_{1},\cdots,g_{n-1})&\mbox{if}&k=n.\end{array}\right.\] These maps satisfy the simplicial relations \(d_{k}^{n-1}\circ d_{k^{\prime}}^{n}=d_{k^{\prime}-1}^{n-1}\circ d_{k}^{n}\) for \(k<k^{\prime}\). We define the singular double chain complex of \(G\rightrightarrows M\) as follows. This is determined by the data \(\{C_{\bullet}(G^{(\bullet)}),\partial,d\}\) where \(C_{q}(G^{(n)})\) is given by the set of singular \(q\)-chains on \(G^{(n)}\), \(\partial:C_{q}(G^{(n)})\to C_{q}(G^{(n-1)})\) is the simplicial boundary operator \(\partial=\sum_{k=0}^{n}(-1)^{k}d_{k}^{n}\), and \(d:C_{q}(G^{(n)})\to C_{q-1}(G^{(n)})\) is the usual singular boundary operator \(d=\sum_{j=0}^{q}(-1)^{j}d_{j}\). We can then consider the associated total complex \(\{\tilde{C}_{\bullet}(G),\delta\}\) where \(\tilde{C}_{v}(G)=\bigoplus_{q+n=v}C_{q}(G^{(n)})\) with total differential operator \(\delta:\partial\to G^{(n)}\). \(\tilde{C}_{v}(G)\to\tilde{C}_{v-1}(G)\) defined as \[\delta(\gamma)=(-1)^{q+n}\partial(\gamma)+(-1)^{q}d(\gamma),\qquad\gamma\in C_{q} (G^{(n)}).\] The associated total singular homology will be denoted by \(H_{\bullet}(G,\mathbb{Z})\). It is worth mentioning that Morita equivalent groupoids have isomorphic total singular homologies, visit for instance [1, 7]. By following a similar approach it may be easily verified that the total singular cohomology \(H^{\bullet}(G,\mathbb{Z})\) of \(G\rightrightarrows M\) can be obtained by dualizing the total singular complex \(\tilde{C}_{\bullet}(G)\) in order to define another total complex \(\tilde{C}^{\bullet}(G)\) and then considering its associated cohomology. Also, by tensoring \(C_{\bullet}(G)\otimes_{\mathbb{Z}}\mathbb{R}\) and \(C^{\bullet}(G)\otimes_{\mathbb{Z}}\mathbb{R}\) we get the total singular homology and cohomology groups \(H_{\bullet}(G,\mathbb{R})\) and \(H^{\bullet}(G,\mathbb{R})\), respectively. Furthermore, via a spectral sequence argument it follows that the classical de Rham isomorphism for manifolds extends to define an isomorphism between the de Rham cohomology \(H^{\bullet}_{dR}(G)\) of \(G\) defined through the Bott-Shulman-Stasheff total complex and the singular cohomology \(H^{\bullet}(G,\mathbb{R})\). Here, the cohomology \(H^{\bullet}_{dR}(G)\) is defined as the total cohomology obtained from the double cochain complex \(\{\Omega_{\bullet}(G^{(\bullet)}),\overline{\partial},d_{dR}\}\) where \(\Omega_{q}(G^{(n)})\) is given by the differential \(q\)-forms on \(G^{(n)}\), \(\overline{\partial}:\Omega_{q}(G^{(n)})\to\Omega_{q}(G^{(n+1)})\) is the simplicial boundary operator \(\overline{\partial}=\sum_{k=0}^{n}(-1)^{k}(d_{k}^{n})^{*}\), and \(d_{dR}:\Omega_{q}(G^{(n)})\to\Omega_{q+1}(G^{(n)})\) is the usual the de Rham differential. As expected, all of these (co)homologies are Morita invariant. For more details visit [1]. ### Hurewicz \(G\)-homomorphism and coverings Let \(G\rightrightarrows M\) be a \(G\)-connected Lie groupoid and fix \(x_{0}\in M\). Consider \(\Pi_{1}(G,x_{0})\) and \(H_{1}(G,\mathbb{Z})\) the fundamental group of \(G\) with respect to the base-point \(x_{0}\) and the first total singular homology group of \(G\), respectively. It is simple to check that every \(G\)-loop at \(x_{0}\) canonically defines a \(1\)-cycle in \(\tilde{C}_{1}(G)=C_{0}(G)\oplus C_{1}(M)\). This suggests us to define the map \(h:\Pi_{1}(G,x_{0})\to H_{1}(G,\mathbb{Z})\) which sends the \(G\)-homotopy class \([\sigma]\) of a \(G\)-loop \(\sigma\) at \(x_{0}\) to its corresponding homology class \(|\sigma|\). **Lemma 2.1**.: _Let \(\sigma\) and \(\sigma^{\prime}\) be two \(G\)-paths. The following assertions hold true:_ * \(\sigma+\sigma^{-1}\) _is equivalent to a boundary in_ \(\tilde{C}_{1}(G)\)_, and_ * _if_ \(\sigma^{\prime}_{0}(0)=\sigma_{n}(1)\) _then_ \(\sigma*\sigma^{\prime}-\sigma-\sigma^{\prime}\) _is also equivalent to a boundary in_ \(\tilde{C}_{1}(G)\)_._ Proof.: Firstly, for every path \(\sigma_{j}\) in \(\sigma\) we may argue as in Lemma 3.2 from [5, p. 173] and for every arrow \(g_{j}\) in \(\sigma\) we consider the pair of composable arrows \((g_{j},g_{j}^{-1})\). Secondly, observe that the expression \(\sigma*\sigma^{\prime}-\sigma-\sigma^{\prime}\) equals \[\sigma_{n}*\sigma^{\prime}_{0}-\sigma_{n}-\sigma^{\prime}_{0}+\text{boundary}.\] So, the last assertion follows directly as a consequence of Lemma 3.1 in [5, p. 173]. **Proposition 2.2** (Hurewicz \(G\)-homomorphism).: _The map \(h\) is a well defined group homomorphism which restricts to the abelianization \(\Pi_{1}^{ab}(G,x_{0})=\Pi_{1}(G,x_{0})/[\Pi_{1}(G,x_{0}),\Pi_{1}(G,x_{0})]\) as an isomorphism._ Proof.: We shall follow the usual strategy used to prove the classical version of this result, see for instance [5, c.4; s.3]. Throughout the proof we shall fix \(\sigma=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) a \(G\)-loop at \(x_{0}\). In order to see that \(h\) is well defined let us consider \(\sigma^{\prime}=\sigma^{\prime}_{n}g^{\prime}_{n}\sigma^{\prime}_{n-1}\cdots \sigma^{\prime}_{1}g^{\prime}_{1}\sigma^{\prime}_{0}\) another \(G\)-loop at \(x_{0}\) which is in the \(G\)-homotopy class of \(\sigma\). If \(\sigma\) and \(\sigma^{\prime}\) are equivalent then the assertion is trivial. Let us suppose then that there is a \(G\)-homotopy \(D(\tau,\cdot)=D_{n}(\tau,\cdot)d_{n}(\tau)\cdots d_{1}(\tau)D_{0}(\tau,\cdot)\) from \(\sigma\) to \(\sigma^{\prime}\). We split the square \([0,1]\times[0,1]\) into two \(2\)-simplices with vertices \(\{(0,0),(0,1),(1,1)\}\) and \(\{(0,0),(1,0),(1,1)\}\), respectively, so that every homotopy \(D_{j}\) can be regarded as the sum of two singular \(2\)-simplices over \(M\). The diagonal from \((0,0)\) to \((1,1)\) will be denoted by \(\lambda\). Therefore, by using Lemma 2.1 we get: \[\delta D = \sum_{j=1}^{n}(t\circ d_{j}-s\circ d_{j})+\sum_{j=0}^{n}(s\circ d_{ j+1}-D_{j}|_{\lambda}-\sigma_{j})-\sum_{j=0}^{n}(\sigma^{\prime}_{j}-D_{j}|_{ \lambda}-t\circ d_{j})\] \[- \sum_{j=1}^{n}(g^{\prime}_{j}-g_{j})+\mbox{constant}=\sigma- \sigma^{\prime}+\mbox{boundary},\] where \(s\circ d_{n+1}:=D_{n}|_{[0,1]\times 1}=x_{0}\) and \(t\circ d_{0}:=D_{0}|_{[0,1]\times 0}=x_{0}\). That is, \(\sigma\) and \(\sigma^{\prime}\) are homologous which means that \(h\) is well defined. Again, as consequence of Lemma 2.1 we have that if \([\sigma],[\sigma^{\prime}]\in\Pi_{1}(G,x_{0})\) then \[h([\sigma][\sigma^{\prime}])=h([\sigma*\sigma^{\prime}])=|\sigma*\sigma^{ \prime}|=|\sigma|+|\sigma^{\prime}|=h([\sigma])+h([\sigma]),\] thus obtaining that \(h\) is a group homomorphism. Note that if \(\sigma^{-1}\) denotes the inverse \(G\)-loop of \(\sigma\) then \(h([\sigma^{-1}])=-|\sigma|\) since \(h\) is a homomorphism. Hence, by Lemma 2.1 it follows that the commutator subgroup of \(\Pi_{1}(G,x_{0})\) lies inside \(\ker(h)\) so that we get another well defined group homomorphism \(\tilde{h}:\Pi_{1}^{ab}(G,x_{0})\to H_{1}(G,\mathbb{Z})\). Since our groupoid is \(G\)-connected we have that for each \(x\in M\) there exists some \(G\)-path \(\lambda_{x}\) from our base-point \(x_{0}\) to \(x\). The constant path at \(x\) is denoted by \(c_{x}\). Let us construct the inverse group homomorphism of \(\tilde{h}\). It is clear that we may think of every arrow \(g\in G\) as a \(G\)-path connecting the constant paths at \(s(g)\) and \(t(g)\). Also, it is obvious that every path \(\gamma:[0,1]\to M\) is equivalent to the \(G\)-path \(c_{\gamma(0)}1_{\gamma(0)}\gamma 1_{\gamma(1)}c_{\gamma(1)}\). In consequence, after possibly reversing orientations and fixing an order, each singular groupoid \(1\)-chain \(\alpha\in\tilde{C}_{1}(G)\) can be thought of as a formal sum of not necessarily different \(G\)-paths \(\alpha=\sum\alpha_{j}\). We define the homomorphism \(l:\tilde{C}_{1}(G)\to\Pi_{1}^{ab}(G,x_{0})\) as \(l(\alpha)=[\sum_{*}\lambda_{\alpha_{j}(1)}^{-1}*\alpha_{j}*\lambda_{\alpha_{j }(0)}]\), where \(\sum_{*}\) denotes the concatenation of all the \(G\)-loops at \(x_{0}\) given by \(\lambda_{\alpha_{j}(1)}^{-1}*\alpha_{j}*\lambda_{\alpha_{j}(0)}\). Note that the definition of \(l\) does not depend on the choice of the \(G\)-paths \(\lambda_{\alpha_{j}(0)}\) and \(\lambda_{\alpha_{j}(1)}\) since we are going into \(\Pi_{1}^{ab}(G,x_{0})\) instead of \(\Pi_{1}(G,x_{0})\). We claim that \(l\) takes the boundaries in \(\tilde{C}_{1}(G)\) into \(1\in\Pi_{1}^{ab}(G,x_{0})\). Indeed, it is clear that it suffices to check this assertion when we apply the boundary operator \(\delta\) over a pair of composable arrows \((g,g^{\prime})\in G^{(2)}\), a path \(\tilde{\sigma}:[0,1]\to G\) and a singular \(2\)-simplex \(\Sigma:\triangle_{2}\to M\). Firstly, \(\delta(g,g^{\prime})=g^{\prime}-g^{\prime}g+g\) so that \[l(\delta(g,g^{\prime})) = l(g^{\prime})l(g)l(g^{\prime}g)^{-1}=[\lambda_{s(g^{\prime})}^{ -1}*g^{\prime}*\lambda_{t(g)}][\lambda_{t(g)}^{-1}*g*\lambda_{s(g)}][\lambda_{ s(g^{\prime})}^{-1}*g^{\prime}g*\lambda_{s(g)}]^{-1}\] \[= [\lambda_{s(g^{\prime})}^{-1}*g^{\prime}g(g^{\prime}g)^{-1}* \lambda_{s(g^{\prime})}]]=[\mbox{constant}]=1.\] Secondly, \(\delta\tilde{\sigma}=t\circ\tilde{\sigma}-s\circ\tilde{\sigma}-(\tilde{ \sigma}(1)-\tilde{\sigma}(0))\). Note that \[l(t\circ\tilde{\sigma}+\tilde{\sigma}(0)-s\circ\tilde{\sigma}-\tilde{\sigma}(1 ))=l(t\circ\tilde{\sigma})l(\tilde{\sigma}(0))l(s\circ\tilde{\sigma})^{-1}l( \tilde{\sigma}(1))^{-1}\] \[=[\lambda_{t\circ\tilde{\sigma}(1)}^{-1}*t\circ\tilde{\sigma}*\lambda_{t\circ \tilde{\sigma}(0)}][\lambda_{t\circ\tilde{\sigma}(0)}^{-1}*\tilde{\sigma}(0)* \lambda_{s\circ\tilde{\sigma}(0)}][\lambda_{s\circ\tilde{\sigma}(1)}^{-1}*s \circ\tilde{\sigma}*\lambda_{s\circ\tilde{\sigma}(0)}]^{-1}[\lambda_{t\circ \tilde{\sigma}(1)}^{-1}*\tilde{\sigma}(1)*\lambda_{s\circ\tilde{\sigma}(1)}]^{ -1}\] \[=[\lambda_{t\circ\tilde{\sigma}(1)}^{-1}*(t\circ\tilde{\sigma})\tilde{\sigma}(0) (s\circ\tilde{\sigma})^{-1}\tilde{\sigma}(1)^{-1}*\lambda_{t\circ\tilde{ \sigma}(1)}]=1,\] since \(\tilde{\sigma}(1)\) is \(G\)-homotopic to the \(G\)-path \((t\circ\tilde{\sigma})\tilde{\sigma}(0)(s\circ\tilde{\sigma})^{-1}\), see [16, p. 191]. Thirdly, the case of the singular \(2\)-simplex \(\Sigma\) over \(M\) follows directly by Lemma 3.5 in [5, p. 174]. The fact we just proved implies that \(l\) descends to a well defined group homomorphism \(\tilde{l}:H_{1}(G,\mathbb{Z})\to\Pi_{1}^{ab}(G,x_{0})\). Furthermore, if \(\sigma\) is a \(G\)-loop at \(x_{0}\) then \[(\tilde{l}\circ\tilde{h})([\sigma])=\tilde{l}(|\sigma|)=[\lambda_{x_{0}}^{-1}* \sigma*\lambda_{x_{0}}]=[\sigma],\] since \(\lambda_{x_{0}}\) may be chosen to be just the constant \(G\)-path at \(x_{0}\). Let us now look at the opposite composition. Observe that the assignment \(x\mapsto\lambda_{x}\) allows us to send singular groupoid \(0\)-simplices into singular groupoid \(1\)-simplices and this can be clearly extended to a homomorphism \(\lambda:\tilde{C}_{0}(G)\to\tilde{C}_{1}(G)\). Suppose that \(\alpha=\sum\alpha_{j}\) is a singular groupoid \(1\)-chain. Then, by Lemma 2.1 it holds that \[(\tilde{h}\circ l)(\alpha) = \tilde{h}([\sum_{*}\lambda_{\alpha_{j}(1)}^{-1}*\alpha_{j}* \lambda_{\alpha_{j}(0)}])=\sum|\lambda_{\alpha_{j}(1)}^{-1}*\alpha_{j}* \lambda_{\alpha_{j}(0)}|\] \[= \sum|\lambda_{\alpha_{j}(1)}^{-1}+\alpha_{j}+\lambda_{\alpha_{j} (0)}|=\sum|\lambda_{\alpha_{j}(0)}+\alpha_{j}-\lambda_{\alpha_{j}(1)}|\] \[= |\sum\alpha_{j}+\lambda_{\sum(\alpha_{j}(1)-\alpha_{j}(0))}|=| \alpha+\lambda_{\delta\alpha}|.\] Therefore, if \(\alpha\) is a singular groupoid \(1\)-cycle it follows that \((\tilde{h}\circ l)(\alpha)=|\alpha|\). In particular, \((\tilde{h}\circ\tilde{l})(|\alpha|)=|\alpha|\), as desired. Let us now consider the notion of covering space over a groupoid as defined in [16, s. 3.3] and [4, c. G; s. 3]. A _covering space_\(E\) over \(G\rightrightarrows M\) is a covering space \(p:E\to M\) equipped with a right \(G\)-action \(E\times_{M}G\to E\) along \(p\). Morphisms between two covering spaces \(E\) and \(F\) over \(G\) are equivariant maps \(f:E\to F\). It is clear that any such morphism is necessarily a covering projection. We will denote by \(\Gamma^{G}(E)\) the set of _equivariant automorphisms_ of \(E\). The map \(p\) extends to a Lie groupoid morphism \(p:E\rtimes G\to G\) from the action groupoid \(E\rtimes G\) into \(G\) defined by \(p(e,g)=g\), which covers the covering map \(E\to M\). We say that \(E\) is _universal_ if the action groupoid \(E\rtimes G\rightrightarrows E\) is \((E\rtimes G)\)-connected and the fundamental groupoid \(\Pi_{1}(E\rtimes G,e_{0})\) is trivial for one (hence all) base-point \(e_{0}\in E\). The latter kind of Lie groupoids shall be called _simply \(G\)-connected_. **Example 2.3**.: There is an explicit way to construct the universal covering space over \(G\rightrightarrows M\) when is \(G\)-connected. Namely, the target fiber \(\Pi_{1}(G)(-,x_{0})\) at \(x_{0}\in M\) of the fundamental groupoid \(\Pi_{1}(G)\) of \(G\) becomes a covering space over \(M\) by restricting the source map \(s:\Pi_{1}(G)(-,x_{0})\to M\). In fact, this map is a left principal \(\Pi_{1}(G,x_{0})\)-bundle. Also, there is a natural right \(G\)-action on \(\Pi_{1}(G)(-,x_{0})\) along \(s\) given by \(([\sigma],g)\to[\sigma gc_{s(g)}]\), where \(c_{s(g)}\) is the constant map at \(s(g)\), which makes it into a covering space over \(G\). This is actually a universal covering space over \(G\), see [4, p. 612-613]. It is well known that \(G\)-paths have a unique path lifting property, visit [16, p. 191]. Let \(e_{0}\) be a base-point in \(E\), denote \(x_{0}=p(e_{0})\), and suppose that \(\sigma=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) is a \(G\)-path from \(x_{0}\) to \(x\). Then there are unique paths \(\tilde{\sigma_{n}},\sigma_{n-1}^{-},\cdots,\tilde{\sigma_{0}}\) in \(E\) with \(p_{*}(\tilde{\sigma_{j}})=\sigma_{j}\), \(\tilde{\sigma_{0}}(0)=e_{0}\) and \(\tilde{\sigma_{i}}(0)g_{j}=\sigma_{j-1}^{-}(1)\). By setting \(\tilde{\sigma_{j}}(0)=e_{j}\) and \((e_{j},g_{j}):e_{j}g_{j}\to e_{j}\) the arrows in \(E\rtimes G\) it follows that \(\tilde{\sigma}=\tilde{\sigma_{n}}(e_{n},g_{n})\sigma_{n-1}^{-}\cdots(e_{1},g_{1 })\tilde{\sigma_{0}}\) is the unique \((E\rtimes G)\)-path starting at \(e_{0}\) which projects onto \(\sigma\). Since \(G\)-homotopic paths are in this way lifted to \((E\rtimes G)\)-homotopic paths, it holds that any covering space over \(G\) with a base-point \(e_{0}\) as above has a natural fiber-wise action of \(\Pi_{1}(G,x_{0})\). Indeed, if \([\sigma]\) is the \(G\)-homotopy class of a \(G\)-loop at \(x_{0}\) then we define \(e_{0}\cdot[\sigma]:=\tilde{\sigma_{n}}(1)\). This defines a right action of \(\Pi_{1}(G,x_{0})\) on \(p^{-1}(x_{0})\). As a consequence of the previous lifting property we get that the induced group homomorphism \(\Pi_{1}(E\rtimes G,e_{0})\to\Pi_{1}(G,x_{0})\) is injective, see [4, p. 611]. Let us assume that the action groupoid \(E\rtimes G\rightrightarrows E\) is \((E\rtimes G)\)-connected. Thus, it is simple to check that the action of \(\Pi_{1}(G,x_{0})\) on \(p^{-1}(x_{0})\) is transitive which implies that \(p^{-1}(x_{0})\) is homogeneous. That is, it is isomorphic to the quotient \(\Pi_{1}(G,x_{0})/p_{*}(\Pi_{1}(E\rtimes G,e_{0}))\) since the isotropy at \(e_{0}\) agrees with \(p_{*}(\Pi_{1}(E\rtimes G,e_{0}))\) by definition. Observe that every element \(f\in\Gamma^{G}(E)\) induces an automorphism of \(p^{-1}(x_{0})\) when seeing it as a \(\Pi_{1}(G,x_{0})\)-space. Indeed, if \(F:E\rtimes G\to E\rtimes G\) is the groupoid automorphism \(F(e,g)=(f(e),g)\) induced by \(f\) then note that the \((E\rtimes G)\)-path \(F_{*}(\tilde{\sigma})\) has initial point \(f(e_{0})\) and final point \(f(e_{0}\cdot[\sigma])\) and satisfies \(p_{*}(F_{*}(\tilde{\sigma}))=\sigma\), which means that it is also a lift of \(\sigma\). Therefore, \(f(e_{0})\cdot[\sigma]=f(e_{0}\cdot[\sigma])\) by uniqueness. Just as in the classical case it follows that if our covering space \(E\) over \(G\) is universal then \(\Gamma^{G}(E)\cong\Pi_{1}(G,x_{0})\), see for instance [4, p. 612]. This important fact together with the previous comments allow us to prove the following result. **Proposition 2.4**.: _If \(H\) is a subgroup of \(\Pi_{1}(G,x_{0})\) then there exists a covering \(r:F\to M\) over \(G\) such that \(r_{*}(\Pi_{1}(F\rtimes G,e_{0}))=H\)._ Proof.: Let us fix a universal covering \(p:E\to M\) over \(G\), see Example 2.3. As \(\Gamma^{G}(E)\cong\Pi_{1}(G,x_{0})\) and \(\Gamma^{G}(E)\) acts on the fibers \(p^{-1}(x_{0})\) without any fixed points then we may assume that \(\Pi_{1}(G,x_{0})\) acts on it without fixed points, see [4, p. 612]. Choose \(e_{0}\in p^{-1}(x_{0})\) and define the subgroup \(H^{\prime}\) of \(\Gamma^{G}(E)\) as follows: \(f\in H^{\prime}\) if and only if there exists \([\sigma]\in H\) such that \(f(e_{0})=e_{0}\cdot[\sigma]\). It is simple to check that \(H\) and \(H^{\prime}\) are isomorphic. As \(H^{\prime}\) is a subgroup of \(\Gamma^{G}(E)\) it is actually a properly discontinuous subgroup of diffeomorphisms of \(E\). Let us define \(F\) as the quotient manifold \(E/H^{\prime}\) and denote by \(q:E\to F\) the corresponding canonical projection. It is well known that this is also a covering space. Denote by \(r:F\to M\) the induced smooth map defined as \(r([e])=p(e)\). As \(r\circ q=p\) it follows that \(r\) is also a covering space. The right action of \(G\) along \(p:E\to M\) induces a well defined right action of \(G\) along \(q:E\to F\) as \([e]\cdot g:=[e\cdot g]\) which, in turn, induces another right action of \(G\) along \(r:F\to M\) by the same expression since the elements of \(\Gamma^{G}(E)\) are \(G\)-equivariant. Therefore, we get that \(\Pi_{1}(G,x_{0})\) acts transitively on the right of the fibers \(r^{-1}(x_{0})\). Let \(\tilde{x_{0}}=q(e_{0})\in r^{-1}(x_{0})\). Finally, by the construction of \(F\) it holds that the isotropy group of the right \(\Pi_{1}(G,x_{0})\)-action on \(r^{-1}(x_{0})\) corresponding to \(\tilde{x_{0}}\) is precisely the subgroup \(H\). That is, \(r_{*}(\Pi_{1}(F\rtimes G,e_{0}))=H\). ### Groupoid homology with local coefficients Motivated by the notion of local coefficients system for topological spaces (see for instance [26, c. 6]), we introduce a corresponding notion as well as its associated homology in the groupoid setting. Let \(R\) be a ring. A _local system of \(R\)-modules_ over \(G\rightrightarrows M\) is defined as a function which assigns to any point \(x_{0}\in M\) a left \(R\)-module \(\mathcal{L}_{x}\) and to any continuous \(G\)-path \(\sigma\) from \(x\) to \(y\) an \(R\)-homomorphism \(\sigma_{\#}:\mathcal{L}_{y}\to\mathcal{L}_{x}\) such that the following conditions are satisfied: * if \(\sigma\) and \(\sigma^{\prime}\) are \(G\)-homotopic then \(\sigma_{\#}=\sigma^{\prime}_{\#}\), * if \(\sigma\) is the constant \(G\)-path at \(x_{0}\) then \(\sigma_{\#}:\mathcal{L}_{x_{0}}\to\mathcal{L}_{x_{0}}\) is the identity map, and * if \(\sigma^{\prime}_{0}(0)=\sigma_{n}(1)\) then \((\sigma*\sigma^{\prime})_{\#}=\sigma_{\#}\circ\sigma^{\prime}_{\#}\). Compare with [6, 22]. Note that for any two point \(x,y\in M\) it holds that \(\mathcal{L}_{x}\) and \(\mathcal{L}_{y}\) are isomorphic by \(G\)-connectedness. In particular, any \(G\)-loop \(\sigma\) at \(x_{0}\) determines an automorphism \(\sigma_{\#}:\mathcal{L}_{x}\to\mathcal{L}_{x}\) which depends only on its \(G\)-homotopy class. This implies that the correspondence \([\sigma]\mapsto\sigma_{\#}\) induces an action of \(\Pi_{1}(G,x_{0})\) on \(\mathcal{L}_{x_{0}}\), thus turning \(\mathcal{L}_{x_{0}}\) into a left module over the group ring \(R[\Pi_{1}(G,x_{0})]\). A _homomorphism_ between two local systems of \(R\)-modules \(\mathcal{L}\) and \(\mathcal{L}^{\prime}\) over \(G\) is just natural transformation \(\Phi:\mathcal{L}\to\mathcal{L}^{\prime}\). Namely, for each \(x_{0}\in M\) we have a homomorphism \(\Phi_{x_{0}}:\mathcal{L}_{x_{0}}\to\mathcal{L}^{\prime}_{x_{0}}\) such that for each \(G\)-path \(\sigma\) from \(x\) to \(y\) it holds \((\sigma_{\#})^{\prime}\circ\Phi_{y}=\Phi_{x}\circ\sigma_{\#}\). If each \(\Phi_{x}\) is an isomorphism then we say that \(\mathcal{L}\) and \(\mathcal{L}^{\prime}\) are _isomorphic_. **Lemma 2.5**.: _Suppose that our groupoid is \(G\)-connected._ * _Let_ \(\mathcal{L}\) _and_ \(\mathcal{L}^{\prime}\) _be two local systems of_ \(R\)_-modules over_ \(G\rightrightarrows M\) _and let_ \(\phi:\mathcal{L}_{x_{0}}\to\mathcal{L}^{\prime}_{x_{0}}\) _be an isomorphism. Then there exists a unique isomorphism_ \(\Phi:\mathcal{L}\to\mathcal{L}^{\prime}\) _such that_ \(\Phi_{x_{0}}=\phi\) * _Let_ \(\mathcal{L}_{0}\) _be an_ \(R\)_-module acted upon by_ \(\Pi_{1}(G,x_{0})\)_. Then there exists a local system of_ \(R\)_-modules_ \(\mathcal{L}\) _such that_ \(\mathcal{L}_{x_{0}}=\mathcal{L}_{0}\) _and which induces the given action of_ \(\Pi_{1}(G,x_{0})\) _on_ \(\mathcal{L}_{x_{0}}\)_._ Proof.: The proofs of these statements are straightforward adaptations of the proofs of Theorems 1.11 and 1.12 in [26, p. 263], so that they are left as an exercise to the reader. **Example 2.6**.: Suppose that \(\alpha:\mathbb{Z}[\Pi_{1}(G,x_{0})]\to R\) is a ring homomorphism. We may view \(R\) as a left \(\mathbb{Z}[\Pi_{1}(G,x_{0})]\)-module with the action \([\sigma]\cdot r:=r\alpha([\sigma])\). Such an action commutes with the standard left \(R\)-module structure on \(R\). Therefore, by Lemma 2.5 it follows that \(\alpha\) determines a local system of \(\mathcal{L}_{\alpha}\) of \(R\)-modules over \(G\rightrightarrows M\). Note that each module \((\mathcal{L}_{\alpha})_{x}\) is isomorphic to \(R\). Let \(\mathcal{L}\) be a local system of \(R\)-modules over \(G\) and consider the nerve structure \(\{G^{(n)},d_{k}^{n}\}\) of \(G\). Let \(\lambda_{n}:G^{(n)}\to G^{(0)}\) denote the "last vertex" map \(\lambda_{n}(g_{1},\cdots,g_{n})=s(g_{n})\) with \(\lambda_{0}=\operatorname{id}_{G^{(0)}}\). We define \(C^{q}(G^{(n)},\mathcal{L})\) as the set of all functions \(c\) with the following properties: * for any \(n\)-singular \(q\)-simplex \(\Sigma:\triangle_{q}\to G^{(n)}\) the value \(c(\Sigma)\) is defined and belongs to the \(R\)-module \(\mathcal{L}_{\lambda_{n}(\Sigma(a_{0}))}\). Here we assume that \(a_{0}\) is the first vertex of the standard simplex \(\triangle_{q}\) with vertices \(a_{0},a_{1},\cdots,a_{q}\), and \(b\). the set of \(n\)-singular \(q\)-simpieces \(\Sigma\) such that \(c(\Sigma)\neq 0\) is finite. The elements of \(C_{q}(G^{(n)},\mathcal{L})\) are called \(n\)_-singular \(q\)-chains with coefficients in \(\mathcal{L}\)_. Note that any chain \(c\in C_{q}(G^{(n)},\mathcal{L})\) can be formally written as a finite sum of the form \(c=\sum_{j}v_{j}\cdot\Sigma_{j}\) where \(\Sigma_{j}:\triangle_{q}\to G^{(n)}\) are \(n\)-singular \(q\)-simplices and \(v_{j}\in\mathcal{L}_{\lambda_{n}(\Sigma_{j}(a_{0}))}\). That is, \(c(\Sigma_{j})=v_{j}\) and \(c(\Sigma)=0\) for any \(n\)-singular \(q\)-simplex \(\Sigma\) different from \(\Sigma_{j}\). Let \(c=v\cdot\Sigma\) be an elementary \(n\)-chain (i.e. the sum above contains only one term). On the one hand, we define the boundary homomorphism \(\partial:C_{q}(G^{(n)},\mathcal{L})\to C_{q}(G^{(n-1)},\mathcal{L})\) as \[\partial c=\partial(v\cdot\Sigma)=\sum_{k=0}^{n-1}(-1)^{k}v\cdot(d_{k}^{n} \circ\Sigma)+(-1)^{n}(g_{n}^{\Sigma})_{\#}^{-1}(v)\cdot(d_{n}^{n}\circ\Sigma).\] Here \(g_{n}^{\Sigma}\) denotes the arrow determined by the \(n\)-projection of \(\Sigma(a_{0})\in G^{(n)}\) onto \(G^{(1)}\) and \((g_{n}^{\Sigma})_{\#}^{-1}\) is the inverse of the isomorphism \((g_{n}^{\Sigma})_{\#}:\mathcal{L}_{t(g_{n}^{\Sigma})}\to\mathcal{L}_{s(g_{n}^ {\Sigma})}\) which is induced by the arrow \(g_{n}^{\Sigma}\), viewed as a \(G\)-path from \(s(g_{n}^{\Sigma})=\lambda_{n}(\Sigma(a_{0}))\) to \(t(g_{n}^{\Sigma})=s(g_{n-1}^{\Sigma})=\lambda_{n-1}(d_{n}^{n}\circ\Sigma(a_{0 }))\). It is well known that \(\partial^{2}=0\), compare [6, 7, 22]. On the other hand, we also define the boundary operator \(\overline{d}:C_{q}(G^{(n)},\mathcal{L})\to C_{q-1}(G^{(n)},\mathcal{L})\) as \[\overline{d}c=\overline{d}(v\cdot\Sigma)=\sigma_{\#}(v)\cdot d_{0}(\Sigma)+ \sum_{j=1}^{q}(-1)^{j}v\cdot d_{j}(\Sigma).\] In this case \(d\) denotes the usual homology boundary operator and \(\sigma:[0,1]\to G^{(0)}\) is the path from \(\lambda_{n}(\Sigma(a_{1}))\) to \(\lambda_{n}(\Sigma(a_{0}))\) defined by \(\sigma(\tau)=\lambda_{n}(\Sigma((1-\tau)a_{1}+\tau a_{0}))\), which has corresponding isomorphism \(\sigma_{\#}:\mathcal{L}_{\lambda_{n}(\Sigma(a_{0}))}\to\mathcal{L}_{\lambda_{ n}(\Sigma(a_{1}))}\). This operator also satisfies \(\overline{d}^{2}=0\), see [26, p. 266]. **Lemma 2.7**.: _The boundary operators \(\partial\) and \(\overline{d}\) commute._ Proof.: On the one side, \[\overline{d}(\partial(c)) = \sum_{k=0}^{n-1}(-1)^{k}(\sigma_{k})_{\#}(v)\cdot d_{0}(d_{k}^{n} \circ\Sigma))+\sum_{k=0}^{n-1}\sum_{j=1}^{q}(-1)^{k+j}v\cdot d_{j}(d_{k}^{n} \circ\Sigma))\] \[+ (-1)^{n}((\sigma_{n})_{\#}((g_{n}^{\Sigma})_{\#}^{-1}(v)))(d_{0}( d_{n}^{n}\circ\Sigma))+\sum_{j=1}^{q}(-1)^{j+n}(g_{n}^{\Sigma})_{\#}^{-1}(v) \cdot d_{j}(d_{n}^{n}\circ\Sigma)),\] where \(\sigma_{k}(\tau)=(\lambda_{n-1}\circ d_{k}^{n})(\Sigma((1-\tau)a_{1}+\tau a_{ 0}))\) for all \(k=0,1,\cdots,n\). On the other side, \[\partial(\overline{d}(c)) = \sum_{k=0}^{n-1}(-1)^{k}\sigma_{\#}(v)\cdot d_{k}^{n}\circ d_{0} (\Sigma)+\sum_{j=1}^{q}\sum_{k=0}^{n-1}(-1)^{j+k}v\cdot d_{k}^{n}\circ d_{j} (\Sigma)\] \[+ (-1)^{n}(g_{n}^{d_{0}(\Sigma)})_{\#}^{-1}(\sigma_{\#}(v))\cdot d _{n}^{n}\circ d_{0}(\Sigma)+\sum_{j=1}^{q}(-1)^{j+n}(g_{n}^{d_{j}(\Sigma)})_{ \#}^{-1}(v)\cdot d_{n}^{n}\circ d_{j}(\Sigma).\] Firstly, note that for all \(k=0,1,\cdots,n-1\) it holds that \(\lambda_{n-1}\circ d_{k}^{n}=\lambda_{n}\) so that the first two expressions in the right hand side of the above equalities agree. The second two expressions are exactly the same. Secondly, as \(d_{j}(\Sigma)(a_{0})=\Sigma(a_{0})\) for all \(j=1,\cdots,q\) then \(g_{n}^{d_{j}(\Sigma)}=g_{n}^{\Sigma}\) which implies that the last expressions also agree. It remains to prove that \((\sigma_{n})_{\#}\circ(g_{n}^{\Sigma})_{\#}^{-1}=(g_{n}^{d_{0}(\Sigma)})_{\#} ^{-1}\circ(\sigma)_{\#}\). Observe that \(\lambda_{n-1}(d_{n}^{n}\circ\Sigma(a_{1}))=\lambda_{n-1}(\Sigma(a_{1}))\) so that \(\sigma_{n}*(g_{n}^{\Sigma})^{-1}\) and \((g_{n}^{d_{0}(\Sigma)})^{-1}*\sigma\) are two \(G\)-paths from \(\lambda_{n-1}(\Sigma(a_{1}))\) to \(\lambda_{n}(\Sigma(a_{0}))\). Define the path \(\gamma:[0,1]\to G^{(1)}\) as \(\gamma(\tau)=(\text{pr}_{n}(\Sigma((1-\tau)a_{1}+\tau a_{0})))^{-1}\). This is such that \(\gamma(0)=(g_{n}^{d_{0}(\Sigma)})^{-1}\) and \(\gamma(1)=(g_{n}^{\Sigma})^{-1}\). Also, \(t\circ\gamma=\sigma\) and \(s\circ\gamma=\sigma_{n}\). Therefore, since \(\gamma(1)\) is \(G\)-homotopic to the \(G\)-path \((s\circ\gamma)^{-1}\gamma(0)(t\circ\gamma)\) (see [16, p. 191]), it holds that the \(G\)-paths above are \(G\)-homotopic and the result follows as desired. The total homology associated to the double chain complex \(\{C_{\bullet}(G^{(\bullet)},\mathcal{L}),\partial,\overline{d}\}\) will be denoted by \(H_{\bullet}^{\mathsf{tot}}(G,\mathcal{L})\) and called the _groupoid homology of \(G\rightrightarrows M\) with local coefficients in \(\mathcal{L}\)_. Let \(\phi:(G\rightrightarrows M)\to(G^{\prime}\rightrightarrows M^{\prime})\) be a Lie groupoid morphism and suppose that \(\mathcal{L}\) is a local system of \(R\)-modules over \(G^{\prime}\). By using the induced Lie groupoid morphism \(\phi_{*}:(\Pi_{1}(G)\rightrightarrows M)\to(\Pi_{1}(G^{\prime}) \rightrightarrows M^{\prime})\) it is possible to define another local system of \(R\)-modules \(\phi^{*}\mathcal{L}\) over \(G\) by setting \((\phi^{*}\mathcal{L})_{x}=\mathcal{L}_{\phi_{0}(x)}\). Let \(c=\sum_{j}v_{j}\cdot\Sigma_{j}\) be an \(n\)-singular \(q\)-chain in \(C_{q}(G^{(n)},\phi^{*}\mathcal{L})\). As \(v_{j}\in(\phi^{*}\mathcal{L})_{\lambda_{n}(\Sigma_{j})(a_{0})}=\mathcal{L}_{ \lambda_{n}^{\prime}(\phi_{n}\circ\Sigma_{j})(a_{0})}\) then we have a well defined map \(\phi_{*}:C_{q}(G^{(n)},\phi^{*}\mathcal{L})\to C_{q}(G^{\prime(n)}, \mathcal{L})\) given as \(\phi_{*}(\sum_{j}v_{j}\cdot\Sigma_{j})=\sum_{j}v_{j}\cdot(\phi_{n}\circ\Sigma _{j})\). It is simple to check that \(\phi_{*}\) commutes with both pairs of boundary operators \(\partial,\overline{d}\) and \(\partial^{\prime},\overline{d}^{\prime}\) so that we have a well defined group homomorphism \(\phi_{*}:H_{\bullet}^{\mathsf{tot}}(G,\phi^{*}\mathcal{L})\to H_{ \bullet}^{\mathsf{tot}}(G^{\prime},\mathcal{L})\). It is well known that if \(\phi\) is a Morita map then horizontal homologies defined by \(\partial\) and \(\partial^{\prime}\) are isomorphic, see [6, 7] and [16, p. 214]. Thus, by using a usual argument of spectral sequences (see for instance [3, p. 108]), we conclude: **Proposition 2.8**.: _A Morita map \(\phi:G\to G^{\prime}\) induces an isomorphism \(\phi_{*}:H_{\bullet}^{\mathsf{tot}}(G,\phi^{*}\mathcal{L})\to H_{ \bullet}^{\mathsf{tot}}(G^{\prime},\mathcal{L})\)._ Let us fix a universal covering \(p:E\to M\) over \(G\) and consider the action groupoid \(E\rtimes G\rightrightarrows E\), see Example 2.3. Recall that \(p\) extends to a Lie groupoid morphism \(p:E\rtimes G\to G\), defined by \(p(e,g)=g\), which covers the covering map \(E\to M\). Pick \(x_{0}\in M\). We know that \(\Gamma^{G}(E)\cong\Pi_{1}(G,x_{0})\) where the isomorphism is as follows. The group \(\Gamma^{G}(E)\) acts simply transitively on the fiber \(p^{-1}(x_{0})\). Take \(e_{0}\in p^{-1}(x_{0})\). Given \(f\in\Gamma^{G}(E)\), let \(\tilde{\sigma}\) be a \(E\rtimes G\)-path joining \(f(e_{0})\) with \(e_{0}\). Define the map \(\beta:\Gamma^{G}(E)\to\Pi_{1}(G,x_{0})\) as \(\beta(f)=[p_{*}(\tilde{\sigma})]\). Observe that \(\beta\) does not depend on \(\tilde{\sigma}\) since \(\Pi_{1}(E\rtimes G,e_{0})\) is trivial. This is the isomorphism we are interested in. For every \([\sigma]\in\Pi_{1}(G,x_{0})\) we denote by \(\beta_{[\sigma]}=\beta^{-1}([\sigma])\) the corresponding element in \(\Gamma^{G}(E)\). That is, \([\sigma]\) can be presented by \(p_{*}(\tilde{\sigma})\) where \(\tilde{\sigma}\) is any \(E\rtimes G\)-path from \(e_{0}\) to \(\beta_{[\sigma]}(e_{0})\). Let us now consider nerve structure \(\{E^{(n)},(d_{k}^{n})^{E}\}\) of the action groupoid \(E\rtimes G\rightrightarrows E\) as well as its total singular groupoid homology. The group \(\Gamma^{G}(E)\) acts naturally on \(C_{q}(E^{(n)},\mathbb{Z})\) as follows. Let \(f_{1}:E\rtimes G\to E\rtimes G\) be the groupoid automorphism \(f_{1}(e,g)=(f_{0}(e),g)\) induced by \(f_{0}=f\in\Gamma^{G}(E)\). So, \(f\cdot\Sigma=f_{n}\circ\Sigma\) for \(\Sigma:\triangle_{q}\to E^{(n)}\) where \(f_{n}:E^{(n)}\to E^{(n)}\) is the induced map along the nerve. Let \(\mathcal{L}\) be a local system of \(R\)-modules over \(G\). Denote by \(\mathcal{L}_{0}:=\mathcal{L}_{x_{0}}\). It is clear that \(\Pi_{1}(G,x_{0})\) acts on \(\mathcal{L}_{0}\) from the left. We transform this action into a right action of \(\Gamma^{G}(E)\) on \(\mathcal{L}_{0}\) by setting \(v\cdot\beta_{[\sigma]}=(\sigma_{\#})^{-1}(v)\). Let us denote by \(\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n)},\mathbb{Z})\) the quotient group of the tensor product \(\mathcal{L}_{0}\otimes C_{q}(E^{(n)},\mathbb{Z})\) by the subgroup \(Q_{q}(\mathcal{L}_{0},E^{(n)})\) generated by all elements of the form \(v\cdot f\otimes\Sigma-v\otimes f\cdot\Sigma\) with \(v\in\mathcal{L}_{0}\), \(f\in\Gamma^{G}(E)\), and \(\Sigma\in C_{q}(E^{(n)},\mathbb{Z})\). Observe that, after naturally extending them to the tensor product, the simplicial boundary operator \(\partial\) maps \(Q_{q}(\mathcal{L}_{0},E^{(n)})\) onto \(Q_{q}(\mathcal{L}_{0},E^{(n-1)})\) and the homology boundary operator \(d\) maps \(Q_{q}(\mathcal{L}_{0},E^{(n)})\) onto \(Q_{q-1}(\mathcal{L}_{0},E^{(n)})\) so that they pass to the quotient, thus giving rise to well defined boundary operators \(\partial:\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n)},\mathbb{Z}) \to\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n-1)},\mathbb{Z})\) and \(d:\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n)},\mathbb{Z})\to \mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q-1}(E^{(n)},\mathbb{Z})\). Hence, we have obtained a double chain complex \(\{\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{\bullet}(E^{(\bullet)},\mathbb{Z}),\partial,d\}\) whose associated total cohomology will be denoted by \(H_{\bullet}(\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}E,\mathbb{Z})\). We are now in conditions to prove our \(G\)-version of the well known _Eilenberg homology isomorphism_, namely: **Proposition 2.9** (Eilenberg \(G\)-isomorphism).: _There exists a chain isomorphism between \(\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{\bullet}(E^{(\bullet)},\mathbb{Z})\) and \(C_{\bullet}(G^{(\bullet)},\mathcal{L})\). In particular, the total homologies \(H_{\bullet}(\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}E,\mathbb{Z})\) and \(H_{\bullet}^{\rm tot}(G,\mathcal{L})\) are isomorphic._ Proof.: Let us first exhibit isomorphisms \(\tilde{p}:\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n)},\mathbb{Z}) \to C_{q}(G^{(n)},\mathcal{L})\). It is clear that the simply \(E\rtimes G\)-connectedness implies that for each \(e\in E\) there exists a unique \(E\rtimes G\)-homotopy class \([\xi_{e}]\) presented by any \(E\rtimes G\)-path \(\xi_{e}\) from \(e\) to \(e_{0}\). Denote by \(p_{n}:E^{(n)}\to G^{(n)}\) the map induced by \(p\) along the nerves. Let \(v_{0}\in\mathcal{L}_{0}\), \(\Sigma:\triangle_{q}\to E^{(n)}\), \(\Lambda=p_{n}\circ\Sigma\), \(e=\lambda_{n}^{E}(\Sigma(a_{0}))\), \(x=\lambda_{n}^{G}(\Lambda(a_{0}))=p(e)\) and define \(\tilde{p}_{o}:\mathcal{L}_{0}\otimes C_{q}(E^{(n)},\mathbb{Z})\to C_{q}(G^{(n )},\mathcal{L})\) by \[\tilde{p}_{o}(v_{0}\otimes\Sigma)=v\cdot\Lambda,\] where \(v=(p_{*}(\xi_{e}))_{\#}(v_{0})\in\mathcal{L}_{x}\). Analogously to how it was proven in Theorem 3.4 from [26, p. 279] it is simple to check that \(\tilde{p}_{o}Q_{q}(\mathcal{L}_{0},E^{(n)})=0\), so that \(\tilde{p}_{o}\) induces a homomorphism \(\tilde{p}:\mathcal{L}_{0}\otimes_{\Gamma^{G}(E)}C_{q}(E^{(n)},\mathbb{Z}) \to C_{q}(G^{(n)},\mathcal{L})\) which is actually an isomorphism. It remains to verify that these maps determine a chain map between the double chain complexes involved. Firstly, recall that \(d_{j}(\Sigma(a_{0}))=\Sigma(a_{0})\) for all \(1\leq j\leq q\) so that \[\tilde{p}_{o}(d_{j}(v_{0}\otimes\Sigma))=\tilde{p}_{o}(v_{0}\otimes d_{j}( \Sigma))=v\cdot(p_{n}\circ(d_{j}(\Sigma)))=v\cdot d_{j}(\Lambda)=\overline{d}_{ j}(\tilde{p}_{o}(v_{0}\otimes\Sigma)).\] As usual, denote by \(\sigma(\tau)=\lambda_{n}^{E}(\Sigma((1-\tau)a_{1}+\tau a_{0}))\). Note that \(e^{\prime}=\lambda_{n}^{E}(d_{0}(\Sigma)(a_{0}))=\lambda_{n}^{E}(\Sigma(a_{1}))\), thus obtaining \(\xi_{e^{\prime}}=\sigma*\xi_{e}\). Hence \[\tilde{p}_{o}(d_{0}(v_{0}\otimes\Sigma))=\tilde{p}_{o}(v_{0}\otimes d_{0}( \Sigma))=v^{\prime}\cdot(p_{n}\circ(d_{0}(\Sigma)))=v^{\prime}\cdot d_{0}( \Lambda),\] where \[v^{\prime}=(p_{*}(\xi_{e^{\prime}}))_{\#}(v_{0})=(p_{*}(\sigma)*p_{*}(\xi_{e}))_{ \#}(v_{0})=(p_{*}(\sigma))_{\#}(v).\] But \(p\circ\lambda_{n}^{E}=\lambda_{n}^{G}\circ p_{n}\) so that \(p_{*}(\sigma)(\tau)=\lambda_{n}^{G}(\Lambda((1-\tau)a_{1}+\tau a_{0}))\). That is, \(\tilde{p}_{o}(d_{0}(v_{0}\otimes\Sigma))=\overline{d}_{0}(\tilde{p}_{o}(v_{0} \otimes\Sigma))\). Secondly, observe that for \(0\leq k\leq n-1\) we have \[\tilde{p}_{o}(\partial_{k}^{E}(v_{0}\otimes\Sigma)) = \tilde{p}_{o}(v_{0}\otimes(d_{k}^{n})^{E}\circ\Sigma)=v\cdot(p_{n-1 }\circ(d_{k}^{n})^{E}\circ\Sigma)\] \[= v\cdot((d_{k}^{n})^{G}\circ p_{n}\circ\Sigma)=v\cdot(d_{k}^{n})^ {G}\circ\Lambda=\partial_{k}^{G}(\tilde{p}_{o}(v_{0}\otimes\Sigma)).\] We denote by \(e_{n}^{\Sigma}\) the arrow determined by the \(n\)-projection of \(\Sigma(a_{0})\in E^{(n)}\) onto \(E\rtimes G\). Observe that \(e^{\prime}=\lambda_{n-1}^{E}((d_{n}^{n})^{E}\circ\Sigma)(a_{0}))=s(e_{n-1}^{ \Sigma})=t(e_{n}^{\Sigma})\) which implies that \(\xi_{e^{\prime}}=(e_{n}^{\Sigma})^{-1}*\xi_{e}\). Thus, \[\tilde{p}_{o}(\partial_{n}^{E}(v_{0}\otimes\Sigma))=\tilde{p}_{o}(v_{0} \otimes(d_{n}^{n})^{E}\circ(\Sigma))=v^{\prime}\cdot(p_{n-1}\circ((d_{n}^{n}) ^{E}\circ(\Sigma)))=v^{\prime}\cdot(d_{n}^{n})^{G}\circ\Lambda,\] where \[v^{\prime}=(p_{*}(\xi_{e^{\prime}}))_{\#}(v_{0})=(p_{*}((e_{n}^{\Sigma})^{-1}) *p_{*}(\xi_{e}))_{\#}(v_{0})=(p(e_{n}^{\Sigma})^{-1})_{\#}(v)=((g_{n}^{\Lambda })^{-1})_{\#}(v),\] since \(p\circ\mathrm{pr}_{n}^{E}=\mathrm{pr}_{n}^{G}\circ p_{n}\). In consequence, \(\tilde{p}_{o}(\partial_{n}^{E}(v_{0}\otimes\Sigma))=\partial_{n}^{G}(\tilde{ p}_{o}(v_{0}\otimes\Sigma))\). This completes the proof. _Remark 2.10_.: It is worth mentioning that similarly to how they were defined the total homology groups \(H^{\mathrm{tot}}_{\bullet}(G,\mathcal{L})\) it is possible to define total cohomology groups \(H^{\bullet}_{\mathrm{tot}}(G,\mathcal{L})\) as well as to prove a \(G\)-version of the Eilenberg cohomology isomorphism. This can be done by mimicking our approach together with the classical constructions, see [11, p. 14]. ## 3. Closed basic \(1\)-forms We say that a Lie groupoid \(G\rightrightarrows M\) is _proper_ if the source/target map \((s,t):G\to M\times M\) is proper. In this case the groupoid orbits \(\mathcal{O}_{x}\) are embedded in \(M\), the isotropy groups \(G_{x}\) are compact, and the orbit space \(M/G\) is Hausdorff, second-countable, and paracompact, visit [8]. Let \(G\rightrightarrows M\) be a proper groupoid and denote by \(X=M/G\) its corresponding orbit space. A differential form \(\omega\) on \(M\) is said to be _basic_ if \(s^{*}\omega=t^{*}\omega\), see [20, 25]. The set of basic forms will be denoted by \(\Omega^{\bullet}_{\mathrm{bas}}(G)\). It is clear that the de Rham exterior differential on \(\Omega^{\bullet}(M)\) restricts to \(\Omega^{\bullet}_{\mathrm{bas}}(G)\), thus yielding the so-called _basic cohomology_\(H^{\bullet}_{\mathrm{bas}}(G,\mathbb{R})\) of \(G\). Such a cohomology is Morita invariant and also satisfies that \(H^{\bullet}(X,\mathbb{R})\cong H^{\bullet}_{\mathrm{bas}}(G,\mathbb{R})\), where \(H^{\bullet}(X,\mathbb{R})\) denotes the singular cohomology of \(X\). ### \(\mathbf{G}\)-homomorphisms of periods Let \(\omega\) be a closed basic \(1\)-form on \(G\) and let \(\xi\) denote the basic cohomology class \([\omega]\in H^{1}_{\mathrm{bas}}(G,\mathbb{R})\). We are interested in studying some features of \(\xi\) by defining its \(G\)-homomorphism of periods as well as its corresponding covering space. In order to do so we will make use of the topological ingredients described in the previous section. For each smooth \(G\)-path \(\sigma=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) in \(M\) we define the \(G\)-path integral \[\int_{\sigma}\omega=\sum_{k=0}^{n}\int_{\sigma_{k}}\omega.\] **Lemma 3.1**.: _Let \([\sigma]\) denote the \(G\)-homotopy class of \(\sigma\). Then the expression_ \[\int_{[\sigma]}\omega=\int_{\sigma}\omega,\] _is well defined._ Proof.: Let us pick another \(G\)-path \(\sigma^{\prime}\) being \(G\)-homotopic to \(\sigma\). If \(\sigma\) and \(\sigma^{\prime}\) are equivalent then the assertion is trivial. Suppose then that there is a smooth \(G\)-homotopy \(D_{n}(\tau,\cdot)d_{n}(\tau)\cdots d_{1}(\tau)D_{0}(\tau,\cdot)\) from \(\sigma\) to \(\sigma^{\prime}\) with fixed endpoints \(x\) and \(y\). It suffices to check that the expression \[I_{\tau}=\int_{D(\tau,\cdot)}\omega=\sum_{k=0}^{n}\int_{0}^{1}\omega_{D_{k}(\tau,\nu)}(D_{k}^{\prime}(\tau,\nu))d\nu,\] does not depend on \(\tau\) by differentiating it with respect to \(\tau\). However, this computation can be verified in local coordinates as in the classical case by using the fact that \(\omega\) is a closed basic 1-form and the following identities are satisfied: \(D_{k}(\tau,0)=t(d_{k}(\tau))\), \(D_{k-1}(\tau,1)=s(d_{k}(\tau))\) for all \(k=1,\cdots,n\) and \(D_{0}(\tau,0)=x\), \(D_{n}(\tau,1)=y\) for all \(\tau\in[0,1]\). _Remark 3.2_.: Suppose that \(\omega\) and \(\omega^{\prime}\) are basic-cohomologous closed basic 1-forms. That is, there is a basic smooth function \(f:M\to\mathbb{R}\) such that \(\omega-\omega^{\prime}=df\). By arguing as in Lemma 3.5 below it is simple to check that \(\int_{\sigma}(\omega-\omega^{\prime})=\int_{\sigma}df=f(\sigma_{n}(1))-f( \sigma_{0}(0))\) since \(f\) is basic. In consequence, the expression \(\int_{[\sigma]}\xi=\int_{\sigma}\omega\) is also well defined when \(\sigma\) is a \(G\)-loop. The first interesting consequence of the previous result is the following. **Proposition 3.3**.: _If \(G\rightrightarrows M\) is simply \(G\)-connected then \(H^{1}_{\rm bas}(G,\mathbb{R})=0\)._ Proof.: Let \(\omega\) be a closed basic 1-form and define \(f(x)=\int_{\lambda_{x}}\omega\) where \(\lambda_{x}\) is any \(G\)-path joining \(x_{0}\) with \(x\). This function is smooth and well defined since \(\Pi_{1}(G,x_{0})\) is trivial. Let \(\gamma:[1,3]\to M\) be a smooth path such that \(\gamma(2)=x\) and \(\gamma^{\prime}(2)=X_{x}\in T_{x}M\). Let \(\lambda_{\gamma(1)}\) be a fixed \(G\)-path from \(x_{0}\) to \(\gamma(1)\) and consider the \(G\)-path \(\lambda_{\gamma(\tau)}=\gamma|_{[1,\tau]}1_{\gamma(1)}\lambda_{\gamma(1)}\) for \(1\leq\tau\leq 3\). Observe that \(f(\gamma(\tau))=f(\gamma(1))+\int_{1}^{\tau}\omega_{\gamma(\nu)}(\gamma^{ \prime}(\nu))d\nu\). Thus \[df_{x}(X_{x})=\frac{d}{d\tau}\left(f(\gamma(1))+\int_{1}^{\tau}\omega_{\gamma( \nu)}(\gamma^{\prime}(\nu))d\nu\right)|_{\tau=2}=\frac{d}{d\tau}\left(\int_{1 }^{\tau}\omega_{\gamma(\nu)}(\gamma^{\prime}(\nu))d\nu\right)|_{\tau=2}=\omega _{x}(X_{x}).\] Let us now check that \(f\) is basic. If \(g\in G\) then it follows that \(\lambda_{t(g)}=c_{t(g)}g\lambda_{s(g)}\), where \(c_{t(g)}\) denotes the constant path at \(t(g)\) and \(\lambda_{s(g)}\) is any \(G\)-path from \(x_{0}\) to \(s(g)\), is a \(G\)-path joining \(x_{0}\) with \(t(g)\). Note that by definition \(\int_{\lambda_{t(g)}}\omega=\int_{\lambda_{s(g)}}\omega\) since the \(G\)-path \(c_{t(g)}g\) does not contribute to the \(G\)-path integral of the left hand side. Hence, \(f(s(g))=f(t(g))\) as desired. If \(\Pi_{1}(G,x_{0})\) is the fundamental group of \(G\) at the base-point \(x_{0}\in M\) then from Lemma 3.1 we get a well defined group homomorphism \(l_{\omega}:\Pi_{1}(G,x_{0})\to(\mathbb{R},+)\) by sending \([\sigma]\mapsto\int_{\sigma}\omega\). Since \(\mathbb{R}\) is abelian it follows that \(l_{\omega}\) factors through the Hurewicz \(G\)-homomorphism \(h\) from Proposition 2.2 by a uniquely determined group homomorphism \(\operatorname{Per}_{\xi}:H_{1}(G,\mathbb{Z})\to\mathbb{R}\) which only depends on the cohomology class \([\omega]\in H^{1}_{\rm bas}(G,\mathbb{R})\), see Remark 3.2. This will be called the \(G\)_-homomorphism of periods_ of \(\omega\). Let \(d\theta\) denote the angle form on \(S^{1}\). Here \(\theta=\frac{1}{2\pi}\phi\) where \(\phi\) is the angle multi-valued function on \(S^{1}\). This is a closed 1-form without zeroes which can not be presented as the differential of a smooth function. The latter fact is consequence of the Stokes Theorem and the fact that \(\int_{S^{1}}d\theta=1\). It is clear that if \(f:M\to S^{1}\) is basic then \(f^{*}(d\theta)\) becomes a closed basic 1-form on \(M\). Let us characterize the closed basic 1-forms that can be obtained in this way. **Proposition 3.4**.: _Let \(\omega\) be a closed basic 1-form on \(M\). Then \(\omega=f^{*}(d\theta)\) where \(f:M\to S^{1}\) is a smooth basic function if and only if the cohomology class \(\xi=[\omega]\in H^{1}_{\rm bas}(G,\mathbb{R})\) is integral, that is, \(\xi\in H^{1}_{\rm bas}(G,\mathbb{Z})=H^{1}_{\rm bas}(G,\mathbb{R})\cap H^{1}(M, \mathbb{Z})\)._ Proof.: We will mainly follow the classical proof of this result as in [11, p. 37]. Suppose that \(\omega=f^{*}(d\theta)\) with \(f:M\to S^{1}\) basic. Note that \(f\) induces a Lie groupoid morphism \(F:(G\rightrightarrows M)\to(S^{1}\rightrightarrows S^{1})\) where \(F\) is either given by \(s^{*}f\) or \(t^{*}f\). Therefore, if \(\sigma=\sigma_{n}g_{n}\sigma_{n-1}\cdots\sigma_{1}g_{1}\sigma_{0}\) is a \(G\)-loop then \[f_{*}(\sigma)=f_{*}(\sigma_{n})1_{f_{*}(\sigma_{n})(0)}f_{*}(\sigma_{n-1}) \cdots f_{*}(\sigma_{1})1_{f_{*}(\sigma_{1})(0)}f_{*}(\sigma_{0}),\] turns out to be equivalent to a usual loop on \(S^{1}\) (actually, we obtain a branch of loops formed by \(f_{*}(\sigma_{j})\) with \(j=0,1,\cdots,n\)). In consequence, the number \[\int_{\sigma}\omega=\sum_{k=0}^{n}\int_{\sigma_{k}}f^{*}(d\theta)=\sum_{k=0}^{ n}\int_{f_{*}(\sigma_{k})}d\theta=\int_{f_{*}(\sigma)}d\theta\in\mathbb{Z},\] is an integer since it agrees with the sum of the degrees of the loops \(f_{*}(\sigma_{j})\) on \(S^{1}\), which clearly computes the degree of the whole branch loop \(f_{*}(\sigma)\). Thus, we get that the \(G\)-homomorphism of periods of any closed basic \(1\)-form \(\omega=f^{*}(d\theta)\) with \(f:M\to S^{1}\) basic takes integral values so that its associated cohomology class lies inside \(H^{1}_{\mathrm{bas}}(G,\mathbb{Z})\). Conversely, let us now suppose that all the \(G\)-periods associated to \(\xi\) are integral. Fix a base point \(x_{0}\in M\) and define \(f(x)=\exp\left(2\pi\sqrt{-1}\int_{\lambda_{x}}\omega\right)\), where \(\lambda_{x}\) is any \(G\)-path joining \(x_{0}\) with \(x\). Note that the definition of \(f\) does not depend on \(\lambda_{x}\) since if \(\lambda_{x}^{\prime}\) is another \(G\)-path from \(x_{0}\) to \(x\) then for the \(G\)-loop at \(x_{0}\) we get \(\sigma=(\lambda_{x}^{\prime})^{-1}*\lambda_{x}\) and \(\int_{\lambda_{x}}\omega-\int_{\lambda_{x}^{\prime}}\omega=\int_{\sigma}\omega \in\mathbb{Z}\). By performing similar computations as those in the proof of Proposition 3.3 it is simple to check that \(f\) is a smooth basic function satisfying \(\omega=f^{*}(d\theta)\). It follows from Proposition 2.4 that we can consider a covering space \(p:M_{\xi}\to M\) over \(G\) which corresponds to the kernel of the \(G\)-homomorphism of periods \(\Pi_{1}(G,x_{0})\to\mathbb{R}\) of the cohomology class \(\xi\). Every \((M_{\xi}\rtimes G)\)-loop in \(M_{\xi}\) project by \(p\) to a \(G\)-loop in \(M\) with trivial periods with respect to \(\omega\). Therefore, it follows that the pullback basic \(1\)-form \(p^{*}\omega\) is basic exact. That is, there is a basic function \(f:M_{\xi}\to\mathbb{R}\) such that \(p^{*}\omega=df\). Recall that the free abelian group of equivariant covering transformations \(\Gamma^{G}(E)\) acts by the left on \(M_{\xi}\). Thus, the cohomology class \(\xi\) determines an injective group homomorphism \(\alpha_{\xi}:\Gamma^{G}(E)\to\mathbb{R}\) with image equal to the group of periods. Indeed, we define \(\alpha_{\xi}\) through the composition \(\mathrm{Per}_{\xi}\circ\beta^{-1}\) where \(\beta:\Gamma^{G}(E)\to\Pi_{1}(G,x_{0})\) as \(\beta(f)=[p_{*}(\tilde{\sigma})]\) is the isomorphism previously described. Therefore, for \(\varphi\in\Gamma^{G}(E)\) we choose \(x\in M_{\xi}\) and any \((M_{\xi}\rtimes G)\)-path \(\tilde{\sigma}\) from \(x\) to \(\varphi(x)\). Note that \(p_{*}(\tilde{\sigma})\) defines a \(G\)-loop at \(p(x)\), so that it also defines a homology class in \(|p_{*}(\tilde{\sigma})|\in H_{1}(G,\mathbb{Z})\). Hence, since \(H^{1}_{dR}(G)\cong\mathrm{Hom}(H_{1}(G,\mathbb{R}),\mathbb{R})\) and \(H^{1}_{\mathrm{bas}}(G,\mathbb{R})\hookrightarrow H^{1}_{dR}(G)\) it makes sense to set \[\alpha_{\xi}(\varphi)=\langle\xi,|p_{*}(\tilde{\sigma})|\rangle=\int_{p_{*}( \tilde{\sigma})}\omega\in\mathbb{R}. \tag{1}\] Such an expression does not depend on \(x\) nor \(\tilde{\sigma}\). It is clear that \(df=p^{*}\omega\) is invariant by the action of \(\Gamma^{G}(E)\) but \(f\) is not, in fact: **Lemma 3.5**.: _The following formula holds true_ \[f(\varphi(x))=f(x)+\alpha_{\xi}(\varphi),\] _for all \(x\in M_{\xi}\) and \(\varphi\in\Gamma^{G}(E)\)._ Proof.: Pick a \((M_{\xi}\rtimes G)\)-path \(\tilde{\sigma}=\tilde{\sigma}_{n}\tilde{g}_{n}\tilde{\sigma}_{n-1}\cdots \tilde{\sigma}_{1}\tilde{g}_{1}\tilde{\sigma}_{0}\) from \(x\) to \(\varphi(x)\). Then, since \(f\) is basic we get \[\int_{\tilde{\sigma}}p^{*}\omega = \int_{\tilde{\sigma}}df=\sum_{k=0}^{n}(f(\tilde{\sigma_{k}}(1))-f( \tilde{\sigma_{k}}(0)))\] \[= f(\tilde{\sigma_{n}}(1))+\sum_{k=1}^{n}(f(t(\tilde{g}_{k}))-f(s( \tilde{g}_{k})))-f(\tilde{\sigma_{0}}(0))=f(\varphi(x))-f(x).\] But \(\int_{\tilde{\sigma}}p^{*}\omega=\int_{p_{*}\tilde{\sigma}}\omega=\alpha_{ \xi}(\varphi)\), so that the formula follows. ### Closed basic 1-forms of Morse type Let \(\omega\) be a closed basic 1-form on \(G\) and let \(\xi\) denote the basic cohomology class \([\omega]\in H^{1}_{\rm bas}(G,\mathbb{R})\). Note that \(\xi=0\) if and only if there exists a basic smooth function \(f:M\to\mathbb{R}\) such that \(\omega=df\). Therefore, in this case the Morse theoretical features of \(\omega\) on \(X\) are the same as those described by means of \(f\) in [19]. We will be mainly interested in studying the case \(\xi\neq 0\). **Lemma 3.6**.: _The critical point set of \(\omega\in\Omega^{1}_{\rm bas}(G)\) is saturated in \(M\). In particular, if \(\omega_{1}=s^{*}\omega=t^{*}\omega\) then we have a topological subgroupoid \({\rm Crit}(\omega_{1})\rightrightarrows{\rm Crit}(\omega)\) of \(G\rightrightarrows M\)._ Proof.: This easily follows from the fact that both \(t\) and \(s\) are surjective submersions and \({\rm Crit}(\omega_{1})=s^{-1}{\rm Crit}(\omega)=t^{-1}{\rm Crit}(\omega)\). If \(\omega\) is a closed basic 1-form then by the Poincare Lemma [20, Lem. 8.5] it follows that for each groupoid orbit \(\mathcal{O}\) there exists an open neighborhood \(\mathcal{O}\subset U\subset M\) and a basic smooth function \(f_{U}\in\Omega^{0}_{\rm bas}(G|_{U})\) such that \(\omega|_{U}=df\). If \(U\) is connected then the function \(f_{U}\) is determined by \(\omega|_{U}\) uniquely up to a constant. In particular, \({\rm Crit}(\omega|_{U})={\rm Crit}(f_{U})\). The _normal Hessian_ of \(\omega\) along a critical orbit \(\mathcal{O}\) is defined to the the normal Hessian of \(f_{U}\) along \(\mathcal{O}\). Thus: **Definition 3.7**.: A critical orbit \(\mathcal{O}\) of \(\omega\) is said to be _nondegenerate_ if and only if \(f_{U}\) is nondegenerate in the sense of Bott. Accordingly, we say that \(\omega\) is _Morse_ if all of its critical orbits are nondegenerate. The notion of Morse-Bott function is classical and was initially introduced by Bott in [2]. Our first key observation is that the previous notion is Morita invariant. **Proposition 3.8**.: _Suppose that \(G\) an \(G^{\prime}\) are Morita equivalent Lie groupoids. If \(G^{\prime}\) admits a Morse closed basic \(1\)-form then so does \(G\)._ Proof.: Let \(P\) be a principal bi-bundle between \(G\) and \(G^{\prime}\) with anchor maps \(a_{l}:P\to M\) and \(a_{r}:P\to M^{\prime}\), see [8]. If \(\omega^{\prime}\) is a Morse closed basic \(1\)-form on \(G^{\prime}\) then there exists a unique closed basic \(1\)-form \(\omega\) on \(G\) such that \(a_{l}^{*}(\omega)=a_{r}^{*}(\omega^{\prime})\), see for instance [20, 25]. This establishes a correspondence between critical orbits since both \(a_{l}\) and \(a_{r}\) are surjective submersions, compare Lemma 5.11 in [21]. Furthermore, if \(\mathcal{O}^{\prime}\) and \(\mathcal{O}\) are related critical orbits then there are connected neighborhoods \(\mathcal{O}^{\prime}\subseteq U^{\prime}\subseteq M^{\prime}\) and \(\mathcal{O}\subseteq U\subseteq M\) together with basic functions \(f_{U^{\prime}}^{\prime}\) and \(f_{U}\) such that \(a_{l}^{*}(f_{U})=a_{r}^{*}(f_{U^{\prime}}^{\prime})\). Hence, if \(\mathcal{O}^{\prime}\) is nondegenerate then so is \(\mathcal{O}\) since both \(a_{l}\) and \(a_{r}\) are surjective submersions. Observe that if \(\omega\) is a basic \(1\)-form on \(G\) then the expression \(\overline{\omega}([x])=\omega(x)\) for \([x]\in X\) is well defined. In particular, this fact allows us to define a _stacky closed \(1\)-form_ on the differentiable stack \([M/G]\) presented by \(G\rightrightarrows M\) as an element \(\overline{\omega}\) presented by a closed basic \(1\)-form \(\omega\) on \(G\). In consequence, we say that \([x]\) is a critical point of \(\overline{\omega}\) if and only if \(\mathcal{O}_{x}\) is a critical orbit of \(\omega\). Also, a critical point \([x]\) is nondegenerate if and only if \(\mathcal{O}_{x}\) is nondegenerate for \(\omega\). **Definition 3.9**.: A stacky closed \(1\)-form \(\overline{\omega}\) on \([M/G]\) is _Morse_ if all of its critical points are nondegenerate. That is, it is presented by a Morse closed basic \(1\)-form on \(G\). It is well known that if \(G\rightrightarrows M\) is a Lie groupoid then its _tangent groupoid_\(TG\rightrightarrows TM\) is obtained by applying the tangent functor to each of its structural maps. If \(\mathcal{O}_{x}\subset M\) is an orbit then we can restrict the groupoid structure to \(G_{\mathcal{O}_{x}}=s^{-1}(\mathcal{O}_{x})=t^{-1}(\mathcal{O}_{x})\), thus obtaining a Lie subgroupoid \(G_{\mathcal{O}_{x}}\rightrightarrows\mathcal{O}_{x}\) of \(G\rightrightarrows M\). Furthermore, the Lie groupoid structure of \(TG\rightrightarrows TM\) induces a Lie groupoid \(\nu(G_{\mathcal{O}_{x}})\rightrightarrows\nu(\mathcal{O}_{x})\) on the normal bundles, having the property that all of its structural maps are fiberwise isomorphisms. In particular, we have that \(\overline{dt}\circ\overline{ds}^{-1}:s^{*}\nu(\mathcal{O}_{x})\to t ^{*}\nu(\mathcal{O}_{x})\) defines a representation \((G_{\mathcal{O}_{x}}\rightrightarrows\mathcal{O}_{x})\curvearrowright(\nu( \mathcal{O}_{x})\to\mathcal{O}_{x})\). As a consequence, for every \(x\in M\) the isotropy group \(G_{x}\) has a canonical representation on the normal fiber \(\nu_{x}(\mathcal{O}_{x})\) called the _normal representation_ of \(G_{x}\) on the normal direction. Let \(\mathcal{O}_{x}\) be a nondegenerate critical orbit of \(\omega\) and let \(\mathcal{O}_{x}\subset U\subset M\) and \(f_{U}:U\to\mathbb{R}\) respectively be an open neighborhood and basic smooth function such that \(\omega|_{U}=f_{U}\). Let us also fix a groupoid metric on \(G\rightrightarrows M\) in the sense of del Hoyo and Fernandes, visit [9]. Since the normal Hessian \(\mathrm{Hess}(f_{U})\) is nondegenerate it follows that by using the groupoid metric the normal bundle \(\nu(\mathcal{O}_{x})\) splits into the Whitney sum of two subbundles \(\nu_{-}(\mathcal{O}_{x})\oplus\nu_{+}(\mathcal{O}_{x})\) such that \(\mathrm{Hess}(f_{U})\) is strictly negative on \(\nu_{-}(\mathcal{O}_{x})\) and strictly positive on \(\nu_{+}(\mathcal{O}_{x})\). Let \(G_{x}\) be the isotropy group at \(x\). From [19] we know that \(\mathrm{Hess}(f_{U})\) is invariant with respect to the normal representation \(G_{x}\curvearrowright\nu(\mathcal{O}_{x})\) so that it preserves the splitting above since the normal representation is by isometries in this case. In consequence, we get a normal sub-representation \(G_{x}\curvearrowright\nu_{-}(\mathcal{O}_{x})\). The _stacky index_ of \([x]\) is defined to be \(\dim\nu_{-}(\mathcal{O}_{x})/G_{x}=\dim\nu_{-}(\mathcal{O}_{x})-\dim G_{x}\). ## 4. Novikov numbers for orbifolds A Lie groupoid \(G\rightrightarrows M\) is said to be _etale_ if either \(s\) or \(t\) is a local diffeomorphism. From now on we assume that our Lie groupoid \(G\rightrightarrows M\) is etale and proper (i.e. it presents an orbifold) and that the orbit space \(X\) is compact. From [20] we know that \(X\) can be triangulated so that the basic cohomology \(H^{\bullet}_{\mathrm{bas}}(G,\mathbb{R})\) becomes a finite dimensional vector space. Moreover, in this specific case we also get that \(H^{\bullet}_{dR}(G)\cong H^{\bullet}_{\mathrm{bas}}(G,\mathbb{R})\), see [23]. Thus, \(H^{\bullet}_{dR}(G)\cong H^{\bullet}(X,\mathbb{R})\). In particular, it follows that we may identify the total singular homology \(H_{1}(G,\mathbb{Z})\) of \(G\) with the singular homology \(H_{\bullet}(X,\mathbb{Z})\) of \(X\). That is, we may assume from now on that \(H_{1}(G,\mathbb{Z})\) is a finitely generated \(\mathbb{Z}\)-module since \(X\) is compact. Let \(\xi\in H^{1}_{\mathrm{bas}}(G,\mathbb{R})\) be the cohomology class of a closed basic \(1\)-form \(\omega\) on \(M\) and let \(\mathrm{Per}_{\xi}\) denote its corresponding \(G\)-homomorphism of periods. Observe that in the particular case of proper and etale Lie groupoids any group homomorphism \(H_{1}(G,\mathbb{Z})\to\mathbb{R}\) can be realized as the homomorphism of periods of a closed basic \(1\)-form. We define the _rank_ of the basic cohomology class \(\xi\) as the rank of the image of \(\mathrm{Per}_{\xi}\). This number will be denoted by \(\mathrm{rank}(\xi)\). Note that by arguing as in the proof Proposition 3.3 we may prove that \(\mathrm{rank}(\xi)=0\) if and only if there is a basic function \(f:M\to\mathbb{R}\) such that \(\omega=df\). That is, \(\mathrm{rank}(\xi)=0\) if and only if \(\xi=0\). _Remark 4.1_.: If we consider a covering space \(p:M_{\xi}\to M\) over \(G\) which corresponds to the kernel of the \(G\)-homomorphism of periods then the expression (1) also shows that the rank of the group \(\Gamma^{G}(M_{\xi})\) equals the rank of the cohomology class \(\xi\). **Proposition 4.2**.: _The set of classes in \(H^{1}_{\mathrm{bas}}(G,\mathbb{R})\) having rank \(1\) is dense._ Proof.: The proof of this result is similar to [11, Cor. 2.2; p. 38] when considering instead \(H^{1}_{\mathrm{bas}}(G,\mathbb{R})\) and \(H^{1}_{\mathrm{bas}}(G,\mathbb{Z})\) We are now in conditions to define the Novikov Betti and torsion numbers associated to the basic cohomology class \(\xi\) by mimicking the three definitions introduced in [11, s. 1.5]. Let \(\Gamma\subset\mathbb{R}\) be an additive subgroup. The _Novikov ring_\(\mathbf{Nov}(\Gamma)\) consists of formal power series of the form \(x=\sum_{j=0}^{\infty}n_{j}\tau^{\gamma_{j}}\) where the coefficients are integers \(n_{j}\in\mathbb{Z}\) and the exponents \(\gamma_{j}\in\Gamma\) are real numbers forming a decreasing sequence converging to \(-\infty\), visit [11, c. 1] for further details concerning the ring structure as well as the properties of \(\mathbf{Nov}(\Gamma)\). When \(\Gamma=\mathbb{R}\) we denote \(\mathbf{Nov}(\mathbb{R})\) just by \(\mathbf{Nov}\). For any basic cohomology class \(\xi\in H^{1}_{\mathrm{bas}}(G,\mathbb{R})\) we may define a ring homomorphism \(\phi_{\xi}:\mathbb{Z}(\Pi_{1}(G,x_{0}))\to\mathbf{Nov}\) by setting \(\phi_{\xi}([\sigma]):=\tau^{\langle\xi,|\sigma|\rangle}\) for all \([\sigma]\in\Pi_{1}(G,x_{0})\). As consequence of Example 2.6 we get that \(\phi_{\xi}\) determines a local system \(\mathcal{L}_{\xi}\) of left \(\mathbf{Nov}\)-modules over \(G\rightrightarrows M\). The groups \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\) are called _Novikov homology groups_ of \(\xi\). It follows that the homology \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\) is a finitely generated module over the ring \(\mathbf{Nov}\). Since \(\mathbf{Nov}\) is a principal ideal domain we have that the module \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\) is a direct sum of a free submodule with a torsion submodule. The _Novikov Betti number_\(b_{j}(\xi)\) is defined to be the rank of the free summand of \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\) and the _Novikov torsion number_\(q_{j}(\xi)\) is defined to be the minimal number of generators of the torsion submodule of \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\). _Remark 4.3_.: On the one hand, if \(G\rightrightarrows M\) is etale and proper and \(p:E\to M\) is a covering space over \(G\) (e.g. the universal covering from Example 2.3), then the action groupoid \(E\rtimes G\rightrightarrows E\) is etale and proper as well. That is, it also represents an orbifold. On the other hand, note that if \(\xi=0\) then \(\mathrm{Per}_{\xi}=0\) which implies that the homomorphism \(\phi_{\xi}\) takes values within the subring \(\mathbb{Z}\subset\mathbf{Nov}\), meaning that \(\mathcal{L}_{\xi}\) is trivial. Therefore, by looking at the groupoid homology with local coefficients as the groupoid equivariant homology described in Proposition 2.9 it follows that the total chain complex \(\mathbf{Nov}\otimes_{\mathbb{Z}(\Pi_{1}(G,x_{0}))}\bar{C}_{\bullet}(E)\) agrees in this case with \(\mathbf{Nov}\otimes_{\mathbb{Z}}\bar{C}_{\bullet}(G)\). In consequence, from [11, Lem. 1.12] we get that \(b_{j}(\xi)\) coincides with the Betti number \(\mathrm{rank}H_{j}(G,\mathbb{Z})\) and \(q_{j}(\xi)\) equals the minimal number of generators of the torsion subgroup of \(H_{j}(G,\mathbb{Z})\). But \(H_{j}(G,\mathbb{Z})\cong H_{j}(X,\mathbb{Z})\), so that we recover the corresponding numerical invariants of the orbit space \(X\). There are other two possible ways to define the Novikov numbers associated to \(\xi\). Let \(\mathbb{Z}[\mathbf{Nov}(\Gamma)]\) be the group ring consisting just of finite sums \(x\) as above. This is a subring of \(\mathbf{Nov}(\Gamma)\). Let \(S\subset\mathbb{Z}[\mathbf{Nov}(\Gamma)]\) denote the subset consisting of the elements in \(\mathbb{Z}[\mathbf{Nov}(\Gamma)]\) with leading term \(1\), so that we have a canonical inclusion of the localized ring \(\mathcal{R}(\Gamma):=S^{-1}\mathbb{Z}[\mathbf{Nov}(\Gamma)]\) into the Novikov ring \(\mathbf{Nov}(\Gamma)\). The ring \(\mathcal{R}(\Gamma)\) will be called the _rational part_ of \(\mathbf{Nov}(\Gamma)\). Firstly, consider the \(G\)-homomorphism of periods \(\mathrm{Per}_{\xi}:H_{1}(G,\mathbb{Z})\to\mathbb{R}\) and denote by \(\Gamma_{\xi}\) its image inside \(\mathbb{R}\). It follows that \(\Gamma_{\xi}\) is a finitely generated free abelian group. The class \(\xi\) determines a ring homomorphism \(\psi_{\xi}:\mathbb{Z}(\Pi_{1}(G,x_{0}))\to\mathcal{R}(\Gamma_{\xi})\) defined as \(\psi_{\xi}([\sigma]):=\tau^{\langle\xi,|\sigma|\rangle}\) for all \([\sigma]\in\Pi_{1}(G,x_{0})\). As above, the homomorphism \(\psi_{\xi}\) gives rise to a local system of left \(\mathcal{R}(\Gamma_{\xi})\)-modules \(\mathcal{M}_{\xi}\) over \(G\rightrightarrows M\) and the homology \(H^{\mathrm{tot}}_{j}(G,\mathcal{M}_{\xi})\) is a finitely generated module over the principal ideal \(\mathcal{R}(\Gamma_{\xi})\). From [11, Cor. 1.12] we obtain that \(\mathbf{Nov}(\Gamma_{\xi})\) is flat over \(\mathcal{R}(\Gamma_{\xi})\) so that we may get an isomorphism \(H^{\mathrm{tot}}_{j}(G,\mathcal{L}_{\xi})\cong\mathbf{Nov}(\Gamma_{\xi}) \otimes_{\mathcal{R}(\Gamma_{\xi})}H^{\mathrm{tot}}_{j}(G,\mathcal{M}_{\xi})\). This immediately implies that \(\mathrm{rank}H^{\mathrm{tot}}_{j}(G,\mathcal{M}_{\xi})\) equals \(b_{j}(\xi)\) and the minimal number of generators of the torsion submodule of \(H^{\mathrm{tot}}_{j}(G,\mathcal{M}_{\xi})\) agrees with \(q_{j}(\xi)\). Secondly, consider the covering space \(p:M_{\xi}\to M\) over \(G\) which corresponds to the kernel of the \(G\)-homomorphism of periods \(\mathrm{Per}_{\xi}\), see Proposition 2.4. It is simple to see that a \(G\)-loop \(\sigma\) in \(M\) lifts to another \((M_{\xi}\rtimes G)\)-loop in \(M_{\xi}\) if and only if \(\mathrm{Per}_{\xi}(|\sigma|)=\langle\xi,|\sigma|\rangle=0\), where \(|\sigma|\in H_{1}(G,\mathbb{R})\) denotes the corresponding homology class of the \(G\)-loop \(\sigma\). Thus, by the isomorphism theorem it follows that the group of covering transformations \(\Gamma^{G}(M_{\xi})\) can be naturally identified with \(L_{\xi}=H_{1}(G,\mathbb{Z})/\ker(\xi)\) and \(\operatorname{Per}_{\xi}\) yields an isomorphisms between the groups \(L_{\xi}\) and \(\Gamma_{\xi}\). Observe that after fixing a base for the free abelian group \(L_{\xi}\) we may identify the group ring \(\Lambda_{\xi}=\mathbb{Z}[L_{\xi}]\) with the ring of Laurent integral polynomials \(\mathbb{Z}[T_{1},\cdots,T_{r},T_{1}^{-1},\cdots,T_{r}^{-1}]\). Let us denote by \(w_{1},\cdots,w_{r}\) the weights of the variables \(T_{1},\cdots,T_{r}\) which are determined by the \(G\)-homomorphism of periods \(\operatorname{Per}_{\xi}:L_{\xi}\to\Gamma_{\xi}\). These weights are linearly independent over \(\mathbb{Z}\) so that we may define the weight of a monomial \(T_{1}^{n_{1}}\cdots T_{r}^{n_{r}}\) as \(\sum n_{j}w_{j}\). Denote by \(S_{\xi}\subset\Lambda_{\xi}\) the set consisting of the Laurent polynomials such that the monomial of maximal weight appearing in them has coefficient \(1\). This is a multiplicative subset and the localized ring \(S_{\xi}^{-1}\Lambda_{\xi}\) is a principal ideal domain since it is isomorphic to the rational subring \(\mathcal{R}(\Gamma_{\xi})\) of the Novikov ring, see [11, s. 1.3] for specific details. Let \(p_{1}:E\to M\) be the universal covering over \(G\) (see Example 2.3) and let \(p_{2}:F\to M\) be the covering space over \(G\) corresponding to the kernel \(\ker(\xi)\subset H_{1}(G,\mathbb{Z})\), after mapping it to \(\Pi_{1}(G,x_{0})\) by using the Hurewicz \(G\)-homomorphism (see Proposition 2.4). These covering spaces give rise to the action groupoids \(E\rtimes G\rightrightarrows E\) and \(F\rtimes G\rightrightarrows F\) which are also etale and proper. By viewing at the groupoid homology with local coefficients in \(\mathcal{M}_{\xi}\) as the groupoid equivariant homology described in Proposition 2.9 we get isomorphisms among the total chain complexes \[\mathcal{R}(\Gamma_{\xi})\otimes_{\mathbb{Z}(\Pi_{1}(G,x_{0}))}\tilde{C}_{ \bullet}(E)\cong S_{\xi}^{-1}\Lambda_{\xi}\otimes_{\Lambda_{\xi}}\tilde{C}_{ \bullet}(F)\cong S_{\xi}^{-1}\tilde{C}_{\bullet}(F).\] It is important to notice that in the previous identifications we used the isomorphism \(\operatorname{Per}_{\xi}:L_{\xi}\to\Gamma_{\xi}\). As localization is an exact functor we obtain that \(H_{\bullet}^{\operatorname{tot}}(G,\mathcal{M}_{\xi})\cong S_{\xi}^{-1}H_{ \bullet}(F,\mathbb{Z})\). Therefore, the Novikov Betti number \(b_{j}(\xi)\) coincides with \(\operatorname{rank}\!H_{\bullet}(F,\mathbb{Z})\) and the Novikov torsion number \(q_{j}(\xi)\) equals the minimal number of generators of the torsion submodule of the \(S_{\xi}^{-1}\Lambda_{\xi}\)-submodule \(S_{\xi}^{-1}H_{\bullet}(F,\mathbb{Z})\). _Remark 4.4_.: The reader probably already noticed that the definitions of the Novikov numbers in the context of orbifolds provided above became both natural and straightforward after having described the algebraic topology notions from Sections 2 and 3. It is left as an exercise to the reader to verify that similar results as those in Sections 1.5 and 1.6 from [11] may be adapted in our context without so many changes along the proofs. In particular, we have that if \(\xi_{1},\xi_{2}\in H_{\operatorname{bas}}^{1}(G,\mathbb{R})\) are two basic cohomology classes such that \(\ker(\xi_{1})=\ker(\xi_{2})\) then \(b_{j}(\xi_{1})=b_{j}(\xi_{2})\) for all \(j\). Also, \(q_{j}(\xi_{1})=q_{j}(\lambda\xi_{2})\) for all \(\lambda\in\mathbb{R}\) with \(\lambda>0\). _Remark 4.5_.: Since the Lie groupoids we are working with above are all of them etale and proper it follows that after naturally adapting Corollaries 4.13 and 4.14 from [16, p. 224] to the homology case we may think of the total homologies \(H_{\bullet}^{\operatorname{tot}}(G,\mathcal{L}_{\xi})\), \(H_{\bullet}^{\operatorname{tot}}(G,\mathcal{M}_{\xi})\), \(H_{\bullet}(E,\mathbb{Z})\) and \(H_{\bullet}(F,\mathbb{Z})\) as being respectively identified with the usual homologies of the orbit spaces \(H_{\bullet}(X,\pi_{*}(\mathcal{L}_{\xi}))\), \(H_{\bullet}(X,\pi_{*}(\mathcal{M}_{\xi}))\), \(H_{\bullet}(E/E\rtimes G,\mathbb{Z})\) and \(H_{\bullet}(F/F\rtimes G,\mathbb{Z})\) where \(\pi:M\to X\) denotes the canonical orbit projection. ### Novikov inequalities Let us now prove the Novikov inequalities for orbifolds. This can be done by following the strategy described in [11, s. 2.3] step by step. It is worth mentioning that the inequalities below depend at some point of the usual Morse inequalities for orbifolds which were already proven in [12] (see also [19]). Although the ideas of the proof are natural and straightforward adaptations of the classical ones, we will bring enough details in order to use most of the machinery introduced in the previous sections. **Theorem 4.6**.: _Let \(G\rightrightarrows M\) be an etale and proper Lie groupoid such that the orbit space \(X\) is compact. Let \(\omega\) be a Morse closed basic \(1\)-form on \(M\). If \(c_{j}(\omega)\) denotes the number of critical points in \(M/G\) having stacky Morse index \(j\) then_ \[c_{j}(\omega)\geq b_{j}(\xi)+q_{j}(\xi)+q_{j-1}(\xi), \tag{2}\] _where \(\xi=[\omega]\in H^{1}_{\rm bas}(G,\mathbb{R})\) is the basic cohomology class of \(\omega\)._ Proof.: It suffices to prove the inequalities under the additional assumption that the basic cohomology class \(\xi\) is integral. That is, \(\xi\in H^{1}_{\rm bas}(G,\mathbb{Z})\). The latter requirement is equivalent to asking that \(\xi\) has rank 1. Indeed, basic cohomology classes \(\xi\) of rank 1 are real multiples of integral basic cohomology classes, namely, \(\xi=\lambda\xi_{0}\) where \(\xi\in H^{1}_{\rm bas}(G,\mathbb{Z})\) and \(\lambda\) is a nonzero real number. It is because in this specific case the image of the \(G\)-homomorphism of periods \({\rm Per}_{\xi}\) is a cyclic subgroup in \(\mathbb{R}\) so that all periods are integral multiples of a minimal period \(\lambda\in\mathbb{R}_{>0}\). In other words, the \(\lambda^{-1}\xi\) has all integral periods and it belongs to \(H^{1}_{\rm bas}(G,\mathbb{Z})\). Therefore, if \(\xi=\lambda\xi_{0}\) has rank 1 and \(\omega\) is a Morse closed basic 1-form in the class \(\xi\) then \(\omega_{0}=\lambda^{-1}\omega\) is another Morse closed basic 1-form in the class \(\xi_{0}\) having the same zeroes. In consequence, \(b_{j}(\xi)=b_{j}(\xi_{0})\) and \(q_{j}(\xi)=q_{j}(\xi_{0})\) by Remark 4.4. Assume for a moment that the Novikov inequalities (2) hold true for basic cohomology classes of rank 1. The argument to prove that the previous assumption is enough to ensure that the inequalities hold true for every basic cohomology class of rank \(>1\) is similar to that in Lemma 2.5 from [11] after considering instead basic cohomology. We sketch the argument here for the sake of completeness as well as the role it plays in our proof. Suppose that \(\xi\in H^{1}_{\rm bas}(G,\mathbb{R})\) is a basic cohomology class of rank \(>1\) and let \(\omega\) be a Morse closed basic 1-form in \(\xi\). Let \(S\subset X\) denote the set of zeroes of \(\overline{\omega}\). This is a finite set since \(X\) is compact. Consider the vector subspace \(N_{\xi}=\{\eta\in H^{1}_{\rm bas}(G,\mathbb{R}):\eta|_{\ker(\xi)}=0\}\). By similar arguments as those used in the proof of Theorem 1.44 from [11] and by Proposition 4.2 it follows that there exists a sequence of rank 1 basic cohomology classes \(\xi_{n}\in N_{\xi}\) such that \(\xi_{n}\to\xi\) as \(n\) goes to \(\infty\), \(b_{j}(\xi_{n})=b_{j}(\xi)\) and \(q_{j}(\xi_{n})=q_{j}(\xi)\) for all \(j\) and \(n\). Let us fix a basis \(\eta_{1},\cdots,\eta_{r}\) of \(N_{\xi}\) whose elements are respectively represented by closed basic 1-forms \(\omega_{1},\cdots,\omega_{r}\). Note that since \(G\) is proper and \(\eta_{k}|_{\ker(\xi)}=0\) we may ensure that there are open neighborhoods \(U_{k}\subset M\) such that \(U_{k}/G_{U_{k}}\subset X\) are open neighborhoods of \(S\) and \(\overline{\omega}_{k}\) vanishes identically on \(U_{k}/G_{U_{k}}\) for all \(k=1,\cdots,r\). It is clear that we can rewrite \(\xi=\sum_{k=1}^{r}a_{k}\eta_{k}\) and \(\xi_{n}=\sum_{k=1}^{r}a_{k,n}\eta_{k}\) with \(a_{k},a_{n,k}\in\mathbb{R}\) such that \(a_{n,k}\to a_{k}\) as \(n\) goes to \(\infty\). Let us define \(\omega_{n}=\omega-\sum_{k=1}^{r}(a_{k}-a_{n,k})\omega_{k}\). This is a basic closed 1-form for which there is an open neighborhood \(U\subset M\) made out of the \(U_{k}^{\prime}s\) above such that \(S\subset U/G_{U}\) and \(\omega-\overline{\omega}_{n}\) vanishes identically on \(U/G_{U}\). Furthermore, for \(n\) large enough it follows that \(\omega_{n}\) has no critical points outside \(U\). That is, \(c_{j}(\omega_{n})=c_{j}(\omega)\) for any \(j\) provided that \(n\) is large enough. Note that from the defining formula above it follows that the basic cohomology class \([\omega_{n}]\) agrees with \(\xi_{n}\) and they have rank 1. In consequence, we get that if the Novikov inequalities hold true for \([\omega_{n}]\) then they must hold true also for \(\xi\). Suppose that \(\xi\) has rank 1. Let us take a covering space \(p:M_{\xi}\to M\) over \(G\) which corresponds to the kernel of the \(G\)-homomorphism of periods \(H_{1}(G,\mathbb{Z})\to\mathbb{R}\) of the cohomology class \(\xi\), see Proposition 2.4. In this case \(M_{\xi}\) has an infinite cyclic group of equivariant covering transformations \(\Gamma^{G}(M_{\xi})\) whose generator is denoted by \(T\). We already know that the pullback \(p^{*}(\omega)\) is a closed basic-exact 1-form so that there exists a basic function \(f:M_{\xi}\to\mathbb{R}\), uniquely determined up to constant, such that \(p^{*}(\omega)=df\). From Lemma 3.5 it follows that \(f(Tx)-f(x)=c\) is a constant for all \(x\in M_{\xi}\). The number \(c\) equals the minimal period of \(\omega\) so that we may assume \(c=\pm 1\) in our case since \(\xi\) is integral. Assume that the generator \(T\) is chosen so that \(f(Tx)-f(x)=-1\) for all \(x\in M_{\xi}\). Otherwise, we may take \(T^{-1}\) instead of \(T\). Let us consider the action groupoid \(M_{\xi}\rtimes G\rightrightarrows M_{\xi}\), the stacky function \(\overline{f}:[M_{\xi}/M_{\xi}\rtimes G]\to\mathbb{R}\) determined by \(f\) and the Lie groupoid morphism \(p:M_{\xi}\rtimes G\to G\) induced by the covering space \(M_{\xi}\to M\) over \(G\). We denote by \(X_{\xi}\) the orbit space \(M_{\xi}/M_{\xi}\rtimes G\) and by \(\overline{T}:X_{\xi}\to X_{\xi}\) and \(\overline{p}:X_{\xi}\to X\) the induced maps between the orbit spaces. On the one hand, observe that the formula \(\overline{f}(T[x])-\overline{f}([x])=-1\) holds true for all \([x]\in X_{\xi}\). On the other hand, the critical points of \(f\) are precisely the elements in \(p^{-1}(x)\) of the zeroes \(x\in M\) of \(\omega\) so that the critical points of the stacky function \(\overline{f}\) in \(X_{\xi}\) are given by the elements in \(\overline{p}^{-1}([x])\) of the zeroes \([x]\) of \(\overline{\omega}\) in \(X\). Pick a regular value \(b\in\mathbb{R}\) of \(\overline{f}\) and set \(V=\overline{f}^{-1}(b)\), \(N=\overline{f}^{-1}([b,b+1])\), and \(Y=\overline{f}^{-1}((-\infty,b+1])\). Note that the projection \(\overline{p}\) determines a one-to-one correspondence between the critical points of \(\overline{f}|_{N}\) and the zeroes of \(\overline{\omega}\). Furthermore, \(\overline{f}|_{N}:N\to[b,b+1]\) is a stacky Morse function since \(p_{\xi}\) is a local diffeomorphism and \(\omega\) is of Morse type. Therefore, \(c_{j}(\overline{f}|_{N})=c_{j}(\omega)\) for all \(j=0,1,2,\cdots\). The homeomorphism \(\overline{T}\) maps \(Y\) into itself. As \(X_{\xi}\) can be triangulated (see [20]), it follows that we may fix a triangulation of \(V\) which in turn induces another triangulation of \(\overline{T}^{-1}V\), thus obtaining a simplicial isomorphism \(\overline{T}:\overline{T}^{-1}V\to V\). Let us choose a triangulation of \(N\) in such a way \(V\) and \(\overline{T}^{-1}V\) are sub-complexes. So, after applying the homeomorphism \(\overline{T}\) we can get a triangulation of the whole \(Y\) so that \(\overline{T}:Y\to Y\) is represented by a simplicial map. In other words, we have obtained a chain complex \(\overline{C}_{\bullet}(Y)\) of simplicial chains which actually is a complex of finitely generated \(\mathbb{Z}[\overline{T}]\)-modules. The standard Morse inequalities for orbifolds were proved in Theorem 7.11 from [12]. Hence, by mimicking the analysis of the Betti numbers \(b_{j}(\overline{C},\mathfrak{p})\) associated to different prime ideals \(\mathfrak{p}\subset\mathbb{Z}[\overline{T}]\), exactly to how it is done in the remaining part of the proof of Theorem 2.4 in [11], we can get the inequalities \[\sum_{k=0}^{j}(-1)^{k}c_{j-k}(\omega)\geq q_{j}(\xi)+\sum_{k=0}^{j}(-1)^{k}b_{ j-k}(\xi),\] which are slightly stronger than the Novikov inequalities we wanted to prove. Let us quickly explain how to apply Theorem 4.6 to find a lower bound for the numbers of zeros of certain symplectic vector fields on symplectic orbifolds. Let \(G\rightrightarrows M\) be an etale and proper Lie groupoid with compact orbit space \(X\) and let \(\rho:A\to TM\) denote its Lie algebroid. The set of _basic vector fields_\(\mathfrak{X}_{\mathrm{bas}}(G)\) is by definition the quotient \[\mathfrak{X}_{\mathrm{bas}}(G)=\frac{\{(v_{1},v_{0})\in\mathfrak{X}(G)\times \mathfrak{X}(M):ds(v_{1})=v_{0}\circ s,\ dt(v_{1})=v_{0}\circ t\}}{\{(v_{1},v_{0 }):ds(v_{1})=v_{0}\circ s,\ dt(v_{1})=v_{0}\circ t,\ v_{1}\in(\ker(ds)+\ker( dt))\}}.\] That is, a basic vector field is not strictly a vector field, but a pair of equivalence classes of vector fields. It is simple to see that a basic vector field \(v=(v_{1},v_{0})\) is determined by its 2st component \(v_{0}\). By Proposition 5.3.12 from [13] we know that Morita equivalent etale a proper groupoids have isomorphic spaces of basic vector fields so that we may think of \(\mathfrak{X}_{\mathrm{bas}}(G)\) as the space of vector field on the orbifold presented by \(G\rightrightarrows M\). Note that if we identify the basic forms \(\Omega^{\bullet}_{\mathrm{bas}}(G)\) with the set of pairs \(\{(\theta_{1},\theta_{0})\in\Omega^{\bullet}(G)\times\Omega^{\bullet}(M):s^{ \ast}(\theta_{0})=\theta_{1}=t^{\ast}(\theta_{0})\}\) then we have contraction operations and Lie derivatives \(\iota:\mathfrak{X}_{\mathrm{bas}}(G)\times\Omega^{\bullet}_{\mathrm{bas}}(G) \to\Omega^{\bullet-1}_{\mathrm{bas}}(G)\) and \(\mathcal{L}:\mathfrak{X}_{\mathrm{bas}}(G)\times\Omega^{\bullet}_{\mathrm{bas} }(G)\to\Omega^{\bullet}_{\mathrm{bas}}(G)\) respectively defined by \[\iota_{v}\theta=(\iota_{\vec{v_{1}}}\theta_{1},\iota_{\vec{v_{0}}}\theta_{0}) \quad\text{and}\quad\mathcal{L}_{v}\theta=(\mathcal{L}_{\vec{v_{1}}}\theta_{1 },\mathcal{L}_{\vec{v_{0}}}\theta_{0}),\] where \((\vec{v_{1}},\vec{v_{0}})\in\mathfrak{X}(G)\times\mathfrak{X}(M)\) is a representative of \(v\). These expressions do not depend on the choice of \((\vec{v_{1}},\vec{v_{0}})\), see [13]. By following [13, 14] we have that a _symplectic form_ on \(G\rightrightarrows M\) is by definition a closed basic 2-form \(\Omega\) on \(M\) which is _nondegenerate_ in the sense that \(\ker(\Omega)=\mathrm{im}(\rho)\). This nondegeneracy requirement implies that the contraction with a symplectic form \(\Omega\) induces a linear isomorphism \(\Omega^{\flat}:\mathfrak{X}_{\mathrm{bas}}(G)\to\Omega^{1}_{\mathrm{bas}}(G)\). Such a notion is also Morita invariant so that it yields a well defined notion of symplectic form over an orbifold. We say that a basic vector field \(v\) is _symplectic_ if \(\mathcal{L}_{\tilde{v_{0}}}\Omega=0\). Note that after using the Cartan formula for the Lie derivative the latter requirement is equivalent to asking that \(\omega=\iota_{\tilde{v_{0}}}\Omega\) is a closed basic 1-form. But we already know that if \(\omega\) is a closed 1-form then the formula \(\omega=\iota_{\tilde{v_{0}}}\Omega\) defines a basic vector field \(v\) which must be symplectic since \(\omega\) is closed. Therefore, it follows that there is a one-to-one correspondence between the closed basic 1-forms and basic symplectic vector fields. Motivated by Proposition 2.6 in [14] we define the _critical point set of a basic vector field_\(v\) as the critical point set of \(\tilde{v_{0}}\) viewed as a section of the vector bundle \(TM/\mathrm{im}(\rho)\to M\). It is simple to check that such a definition does not depend on \(\tilde{v_{0}}\). Because of the nondegeneracy condition imposed over \(\Omega\) it holds automatically that the critical points of \(v\) and \(\omega=\iota_{\tilde{v_{0}}}\Omega\) agree. Hence, on symplectic orbifolds, the problem of estimating from below the number of zeros of closed basic 1-forms is equivalent to finding a lower bound for the numbers of zeros of basic symplectic vector fields. In consequence, the natural generalization of the Novikov theory we have developed in this paper provides a tool for using topological methods to study zeros of symplectic vector fields on symplectic orbifolds. Furthermore, it also opens new research directions for many important physical models which can be described by the Hamiltonian formalism over orbifolds allowing closed basic 1-forms as their Hamiltonians. This can be done in the same spirit that it was studied by Novikov in [18].

2308.13486

On the Practicality of Dynamic Updates in Fast Searchable Encryption

Searchable encrypted (SE) indexing systems are a useful tool for utilizing cloud services to store and manage sensitive information. However, much of the work on SE systems to date has remained theoretical. In order to make them of practical use, more work is needed to develop optimal protocols and working models for them. This includes, in particular, the creation of a working update model in order to maintain an encrypted index of a dynamic document set such as an email inbox. I have created a working, real-world end-to-end SE implementation that satisfies these needs, including the first empirical performance evaluation of the dynamic SE update operation. In doing so, I show a viable path to move from the theoretical concepts described by previous researchers to a future production-worthy implementation and identify issues for follow-on investigation.

Steven Willoughby

2023-08-25T16:50:02Z

http://arxiv.org/abs/2308.13486v1

# On the Practicality of Dynamic Updates in Fast Searchable Encryption ###### Abstract Searchable encrypted (SE) indexing systems are a useful tool for utilizing cloud services to store and manage sensitive information. However, much of the work on SE systems to date has remained theoretical. In order to make them of practical use, more work is needed to develop optimal protocols and working models for them. This includes, in particular, the creation of a working update model in order to maintain an encrypted index of a dynamic document set such as an email inbox. I have created a working, real-world end-to-end SE implementation that satisfies these needs, including the first empirical performance evaluation of the dynamic SE update operation. In doing so, I show a viable path to move from the theoretical concepts described by previous researchers to a future production-worthy implementation and identify issues for follow-on investigation. ## 1 Introduction There are many situations and contexts wherein users of information systems need to collect, store, search, and retrieve large amounts of information. When the collection of data is large enough or needs to be available to multiple geographically-separated users, an attractive option may be to host the document repository on a cloud service provided by a third party. While this allows the users to utilize the service's data centers and network connections to provide a robust platform to host their data, it opens a number of very serious security and privacy concerns if the data being hosted are in any way sensitive, since the hosting service may not necessarily be trusted to protect that information from their own personnel or others. Consider, for example, an organization which uses such an externally-hosted searchable repository to manage confidential pre-release product design documentation, or financial information belonging to the organization. Worse, consider if the data were to contain personal information about employees or customers which would have expensive and disruptive effects on people's lives if it were to be leaked to unauthorized parties. The obvious solution is to encrypt the data, so that they may be stored on the untrusted server in a form that cannot be understood by anyone but authorized personnel. This solves the problem of protecting the data at rest on the server. However, since the index must be decrypted in order to search within it, we must take one of two approaches: either provide decryption keys to the server in order to decrypt and search server-side, or download the entire index to the client for the decryption and search to be performed there. The former approach is not desirable because we have already established that the hosting provider may not be authorized to see the data nor trusted to protect it from unauthorized access. The latter is less than practical due to the amount of data which must be copied to users' local systems. These client systems may not have sufficient storage or processing power1 and the data may well be unreasonably large to transmit repeatedly--it may be hundreds of megabytes, gigabytes, or terabytes depending on the amount of indexed data. Footnote 1: We must accept in these modern times that client computing platforms may well include cell phones and low-power notebooks in addition to more traditional computing platforms. Ideally, we desire to have a method whereby the server can facilitate searches within an index of interesting keywords from the document repository, then report back with a list of documents containing the requested keywords (allowing the user to then retrieve those documents, locally decrypt them, and make use of their contents), all without the server having the ability to actually read the document index itself (since that provides a great deal of insight into the contents of each indexed document). In fact, the server should not even be able to understand what keywords it was searching for (since that provides insight into the nature of the documents and what the users are looking for), or what documents were in the result list of each search. While that may seem to be an impossible expectation, in reality we can find an acceptable middle ground which allows efficient server-side searching without divulging any direct information about the details of the search terms or results. The price paid for this, however, is that a determined hostile observer (perhaps the hosting provider themselves) could analyze patterns of input and output over time which will "leak" useful information from which some amount of the protected data may be inferred. Building on the foundational work of previous researchers in this field, I have created a dynamic update capability which allows an SE index to accumulate new documents over time, whereas previous implementations were primarily focused on a one-time generation of an SE index for a static document set. I also moved beyond the previous theoretical treatments of this subject by adding empirical performance evaluation of my new update mechanism using a typical TCP/IP client-server architecture. Based on this work I identified some considerations for future optimization work. ## 2 Definitions and Nomenclature In this paper I will use the terminology set out by Curtmola, et al. [1] which is also used by other authors, notably Demertzis and Papamanthou, [2] for the sake of consistency with established work on this topic. Basic notation and symbology is summarized in Table 1. Central to this topic is the notion of a collection of documents for which the user wishes to maintain a searchable encrypted index. Following Curtmola, et al.'s nomenclature, let \(\Delta\) be a dictionary of all "interesting" words in all documents, i.e., \(\Delta=\{w_{1},\ldots,w_{d}\}\) where \(d\) is the number of unique interesting words. If \(2^{\Delta}\) is the power set of all possible documents containing words \(w\in\Delta\), then we will consider a set \(\mathcal{D}\subseteq 2^{\Delta}\) which is the specific set of \(n\) documents being indexed in some particular instance of searchable encrypted index being discussed. Each such document has a unique identifier by which it can be fetched from its storage location. Let \(\mathsf{id}(D)\) be the identifier for some arbitrary document \(D\). Further, let \(\mathcal{D}(w)=\{\mathsf{id}(D)\ \forall D\in\mathcal{D}\mid w\in D\}\) be the set of unique identifiers for all documents in our indexed collection which contain the word \(w\). (Curtmola, et al. use the notation \(\mathbf{D}\) and \(\mathbf{D}(w)\) instead of \(\mathcal{D}\) and \(\mathcal{D}(w)\) respectively). For my work which builds primarily on the work by Demertzis and Papamanthou, [2] I will also use the following nomenclature from their work: Let \(\lambda\) be the security parameter (in practical terms, the encryption key length in bits), such that each key \(k_{i}\) is generated using a cryptographically secure random number source, i.e., \(k_{i}\xleftarrow{\$}\{0,1\}^{\lambda}\). Also let \(N\) be the number of entries stored in the SE, where _entry_ is a word which here means a unique tuple \((w,\mathsf{id}(D))\) mapping an indexed keyword \(w\) to the identifier of a document \(D\) containing that word. Thus, we have \[N=\sum_{\forall w\in\Delta}\left|\mathcal{D}(w)\right|.\] As we shall see, Demertzis and Papamanthou [2] posit a storage array arranged in tiered _levels_ of varying sized storage _buckets_. Let \(\ell=\lceil\log_{2}N\rceil\) be the number of levels of index storage which would be employed in this model. Let \(s\leq\ell\) be a configurable number of tiers which will actually be stored on the server (to save space since not all indexes will have values actually assigned to all possible levels), and \(\mathcal{L}\) be the set of \(s\) storage levels allocated for the SE. This SE model supports the notion of _locality_ where data associated with the same keyword are placed in 1 or more co-located areas in the data store. Let \(L\) be the user-configurable locality such that specifying \(L>1\) allows each indexed term to be stored in multiple non-contiguous storage areas, facilitating parallelization of search operations within the index. These levels of storage are implemented in storage arrays \(A_{i}\) where \(i\in\mathcal{L}\). Each level is further partitioned into _buckets_. Bucket \(x\) of array \(A_{i}\) is denoted \(A_{i}[x]\). I will refer to a few standard functions, as follows. Let \(\mathsf{F}\) be a pseudo-random function (prf) \(\mathsf{F}:\{0,1\}^{*}\times\{0,1\}^{*}\rightarrow\{0,1\}^{*}\), which emits a deterministic pattern of bits based on the values of its two inputs (key and data), but whose output is indistinguishable from random bits if those inputs are not known. Let \(\mathsf{Enc}\) and \(\mathsf{Dec}\) be \(\mathsf{cpa}\)-secure2 symmetric encryption functions \(\mathsf{Enc}:\{0,1\}^{*}\times\{0,1\}^{\lambda}\rightarrow\{0,1\}^{*}\) and \(\mathsf{Dec}:\{0,1\}^{*}\times\{0,1\}^{\lambda}\rightarrow\{0,1\}^{*}\) (such that \(\mathsf{Dec}=\mathsf{Enc}^{-1}\)) which take \(\lambda\)-bit keys to transform arbitrary-length bit strings to another arbitrary-length ciphertext and back again. Finally, let \(\mathsf{H}:\{0,1\}^{*}\rightarrow\{0,1\}^{b}\) be a cryptographically strong one-way hash function which outputs \(b\) bits of digest from its input data of arbitrary length. This function must be collision resistant. To the above notation I add the concept of the _order_ of an index, which gives us a useful way to organize a collection of various-size SE indexes. For this research, I chose to assume the order \(o\) of an index to be \(o=\ell=\lceil\log_{2}N\rceil\) with the intention that it would yield a reasonable pattern of varying sizes of indexes to avoid expensive large-index merge operations as long as reasonably possible. ## 3 Basic Principles of SE Here, and throughout the rest of this paper, the term _client_ shall refer to the system a user of the SE system employs to initiate searches or add new documents to the SE index. It is a trusted system under the control of an authorized user. Encryption keys may be employed on it, and plain-text search terms and results may be known to it. The term _server_ shall refer to the remote system on which the encrypted documents and the SE indexes are stored. This system is not allowed to see any decryption keys nor to see the plaintext search terms nor results. The essential principle on which SE is based is that, given an index \(\mathcal{I}\) mapping a set \(\Delta\) of interesting keywords from a document repository \(\mathcal{D}\), we must represent \(\mathcal{I}\) in some opaque fashion such that it can be stored on an untrusted server without anyone being able to glean information about \(\mathcal{D}\) by examining \(\mathcal{I}\), even given an arbitrarily large amount of time to analyze \(\mathcal{I}\). This implies the use of a one-way cryptographically strong hash function, since that will provide a way to derive an opaque value to represent a value in the index without a reliable way to reverse the encoding function to obtain the original value again. If we can then use the same hash function to encode the client-side search terms we can match them on the server to the encoded entries in \(\mathcal{I}\) without revealing the original search terms directly. To illustrate this concept, consider a document repository which contains five documents, specifically, the first five volumes of Douglas Adams' magnum opus _The Hitchhiker's Guide to the Galaxy_. These volumes are separately encrypted and stored on the server. Each is assigned a document ID as shown in Table 2. We identify a set \(\Delta\) of all \begin{table} \begin{tabular}{l l} \hline \hline **Notation** & **Meaning** \\ \hline \(a\parallel b\) & Concatenation of strings \(a\) and \(b\) \\ \(|X|\) & Cardinality of set \(X\) \\ \(x\oplus y\) & Bitwise exclusive-or of \(x\) and \(y\) \\ \(x\xleftarrow{\$}X\) & Element \(x\) sampled uniformly from set \(X\) \\ \(x\leftarrow{\mathcal{A}}\) or \({\mathcal{A}}\to x\) & Output \(x\) from algorithm or function \({\mathcal{A}}\) \\ \(\Delta=\{w_{1},w_{2},\ldots,w_{d}\}\) & Dictionary of \(d\) words in an index \\ \({\mathcal{D}}=\{D_{1},D_{2},\ldots,D_{n}\}\) & Set of \(n\) documents whose words are indexed \\ \({\mathcal{D}}(w)\) & List of all documents containing word \(w\) \\ \(A_{i}[x]\) & Bucket \(x\) of level \(i\) in index storage array \(A\) \\ \(\lambda\) & Bit length of encryption keys \\ \(L\) & Locality of the index \\ \({\mathcal{L}}=\{i_{1},i_{2},\ldots,i_{s}\}\) & Set of \(s\) storage levels in use for the index \\ \(N\) & Number of stored \((w,\mathsf{id}(D))\) tuples in index \\ \(o\) & Order of a SE index, related to its storage capacity \\ \(s\) & Number of actually stored index levels \\ \(c\leftarrow\mathsf{Enc}(K,m)\) & Encryption function with key \(K\) and plaintext message \(m\) \\ \(m\leftarrow\mathsf{Dec}(K,c)\) & Decryption function with key \(K\) and ciphertext message \(c\) \\ \(y\leftarrow\mathsf{F}(K,x)\) & Pseudo-random function with key \(K\) and data \(x\) \\ \(x^{\prime}\leftarrow\mathsf{H}(x)\) & Collision-resistant hash function taking data \(x\) \\ \(\mathsf{id}(D)\) & Unique identifier for document \(D\) \\ \(\varepsilon\) & Empty string or unused storage location \\ 000C & Hexadecimal values are shown in fixed-width type \\ \hline \hline \end{tabular} \end{table} Table 1: Summary of Notation Used in This Paper we find interesting for our purposes. Say, for example, \(\Delta=\{\)Arthur, dolphin, Fenchurch, hooloovoo, krikkit, Zaphod \(\}\). (Obviously, in a full production repository the list of interesting words would be orders of magnitude greater than this trivial example.) If we make a list of all documents in \(\mathcal{D}\) in which each of the words in \(\Delta\) appear, we find the following associations of each keyword \(w\in\Delta\) to a set of document IDs \(D(w)\): \[\texttt{Arthur} \rightarrow\{3,5,8,12,15\}\] \[\texttt{dolphin} \rightarrow\{3,12\}\] \[\texttt{Fenchurch} \rightarrow\{12,15\}\] \[\texttt{hooloovoo} \rightarrow\{3\}\] \[\texttt{krikkit} \rightarrow\{8,12\}\] \[\texttt{Zaphod} \rightarrow\{3,5,8,12,15\}\] From these associations we generate an index \(\mathcal{I}\) which is a collection of tuples \((w,\mathsf{id}(D))\). Specifically, we get: \((\texttt{Arthur},3)\), \((\texttt{Arthur},5)\), \((\texttt{Arthur},8)\), \((\texttt{Arthur},12)\), \((\texttt{Arthur},15)\), \((\texttt{dolphin},3)\), \((\texttt{dolphin},12)\), \((\texttt{Fenchurch},12)\), \((\texttt{Fenchurch},15)\), \((\texttt{hooloovoo},3)\), \((\texttt{krikkit},8)\), \((\texttt{krikkit},12)\), \((\texttt{Zaphod},3)\), \((\texttt{Zaphod},5)\), \((\texttt{Zaphod},8)\), \((\texttt{Zaphod},12)\), and \((\texttt{Zaphod},15)\). We store \(\mathcal{I}\) on disk in two parts: a storage array which holds the actual tuples, and a hash table which associates each search term \(w\) with the location in storage holding its set of tuples. Setting aside for the moment the finer points of storage optimization so that we may focus just on the encryption aspect, let us visualize the storage arrangement of our index \(\mathcal{I}\) as shown in Figure 1. With such a storage arrangement, if the client wishes to search for keyword \(w=\texttt{dolphin}\), the server looks that up in the hash table, finding that the tuples to satisfy the search are contained in storage array level \(1\), bucket \(0\) (which we will designate \(A_{1}[0]\)). Looking in that bucket, we find (among other things that happen to be stored there as well) the tuples \((\texttt{dolphin},3)\) and \((\texttt{dolphin},12)\). From this the server reports the result set \(\{3,12\}\) as the set of document IDs where the word "dolphin" is found. ### Encrypting the Index To the above trivial storage arrange we now need to add a layer of encryption to obscure the meaning of the information in \(\mathcal{I}\) beyond the ability of the server to understand, but in such a way that the client can use it to get the same search results. For this encryption, we generate a secret key known only to authorized clients. This key \(K=(k_{1},k_{2},k_{3})\) has three parts, each of which is created from \(\lambda\) random bits (i.e., \(k_{i}\xleftarrow{8}\{0,1\}^{\lambda}\)). First, given a cryptographically strong one-way hash function \(\mathsf{H}\), pseudo-random function \(\mathsf{F}\), and encryption function \(\mathsf{Enc}\) as described above, we encode the tuples stored in the array \(A\) by encrypting the value \(\mathsf{id}(D)\parallel 0^{\lambda}\) using the encryption key \(\mathsf{F}(k_{3},w)\). In our example, assuming for simplicity that document IDs are \(16\) bits and \(\lambda=16\), the tuple \((\texttt{dolphin},3)\) is encoded by calculating \(\mathsf{Enc}(\mathsf{F}(k_{3},\texttt{dolphin}),0003000)\). Likewise, the tuple \((\texttt{dolphin},12)\) is encoded by calculating \(\mathsf{Enc}(\mathsf{F}(k_{3},\texttt{dolphin}),0000C000)\). Assuming these two calculations produce the hex values A462910E and 07B422A7, and that we carried out corresponding encodings with the other tuples, we would now have the encrypted storage array shown in Figure 2. Note that we also filled the empty storage locations with random bits to further obfuscate the index. It is important to note that each tuple is encrypted with a key that is based on the search term to which it belongs, so the data there is only recoverable if one is in possession of that secret key \(k_{3}\) and the search term \(w\). Now that the tuples are encoded, we must encrypt the hash table's keys and values in a similar fashion. The keys (the search terms) are simply replaced with the results of hashing them with another secret key: \(\mathsf{H}(\mathsf{F}(k_{1},w))\). Thus, search term "dolphin" would be replaced by \(\mathsf{H}(\mathsf{F}(k_{1},\texttt{dolphin}))\), say 38A9039C. The value associated with "dolphin", is the tuple \begin{table} \begin{tabular}{r l} \hline \hline **ID** & **Document Title** \\ \hline 3 & _The Hitchhiker’s Guide to the Galaxy_ \\ 5 & _The Restaurant at the End of the Universe_ \\ 8 & _Life, the Universe, and Everything_ \\ 12 & _So Long, and Thanks for All the Fish_ \\ 15 & _Mostly Harmless_ \\ \hline \hline \end{tabular} \end{table} Table 2: Example Document Repository \(\mathcal{D}\) Figure 1: Example Index \(\mathcal{I}\) Storage (unencrypted) Figure 2: Example Index \(\mathcal{I}\) Storage (encrypted) \((1,0)\) which means that the entries for that keyword are to be found in \(A_{1}[0]\) (storage level 1, bucket 0). We represent location \(A_{i}[x]\) as a single numeric value \(i\,\|\,x\) (in this case the hex value 00010000). This is encoded in the hash table as \([i\,\|\,x]\oplus\mathsf{H}(\mathsf{F}(k_{2},w))\). Again, note that the search term \(w\) and a secret key are part of this encryption scheme. Supposing this gives the result 6BF86758, and continuing this for the rest of the table, we get the completely encrypted index shown in Figure 2. The values where the entries for our example term "dolphin" are encoded in \(\mathcal{I}\) are highlighted in red. Now if we wish to search for a word like "dolphin", we generate a _search token_\(T=(t_{1},t_{2},t_{3})\) by providing the portion of the encoding operations requiring knowledge of the secret values, sending to the server only the output from \(\mathsf{F}\) which it can use to complete the hashing and decryption without divulging the actual keys or search terms: \(t_{1}=\mathsf{F}(k_{1},w)\), \(t_{2}=\mathsf{F}(k_{2},w)\), and \(t_{3}=\mathsf{F}(k_{3},w)\). The server, upon receiving the client's search token \(T\), calculates \(\mathsf{H}(t_{1})\) and gets 38A9039C. Looking at the hash table in Figure 2 we see that this is a key stored there, associated with value 6BF86758. The server then calculates 6BF86758 \(\oplus\,\mathsf{H}(t_{2})\) to get the result 00010000. Although the server never knew the search term \(w\), it was given just enough information in \(T\) to determine that the answer to that query is to be found in storage location \(A_{1}[0]\). \(T\) does not provide any information to decode any other hash table entries since they were encoded using different values of \(w\). Now the server knows that some of the values stored in \(A_{1}[0]\) can be decrypted using the key \(\mathsf{H}(t_{3})\). Running the contents of \(A_{1}[0]\) through this decryption, it gets the results 00030000, 1AED5898, EF00F293, 000C000, and 923BF508. Since any valid entry has \(0^{\lambda}\) bits appended, the server knows that only the first and fourth values were correctly decrypted by the key it was given, so the result reported to the client is the set of document IDs \(\{3,12\}\). Note that when we set locality \(L>1\), we must allow for multiple buckets to hold the tuple lists for any given keyword, so the actual calculations for the hash table keys and values includes a counter \(c\in\{0,1,\ldots,L-1\}\). The key is actually encoded as \(\mathsf{H}(\mathsf{F}(k_{1},w)\|\,c)\) and the value as \([i\|x]\oplus\mathsf{H}(\mathsf{F}(k_{2},w)\|\,c)\). ## 4 Prior Work In their seminal work on the subject, Song, Wagner, and Perrig [3] laid out the essential idea for SE indexing for the first time. From this beginning almost two decades ago, other researchers have further developed and extended these initial concepts in order to improve functionality, security, and performance. One approach explored by Pinkas and Reinman [4] as an alternative to SE was to leverage the concept of oblivious ram (oram)--a specially-arranged memory system originally proposed by Goldreich and Ostrovsky [5] which has the property that "the sequence of memory accesses... reveals no information about the input..., beyond the running-time for the input." Pinkas and Reinman sought to use this aspect of oram to hide the nature of the calculations used to search through an encrypted index to thwart attempts at cryptanalysis or other means of obtaining confidential details of the index. Unfortunately, this approach is very expensive compared to more practical software-only solutions described here. As these software SE systems were developed, they were primarily implemented as in-memory schemes. Cash, et al. [6] note that this approach did not scale effectively as the repository size expanded into the near-terabyte range and beyond. As index entries may be scattered throughout the index, the amount of data transmitted back to the user for local decryption multiplies with the database size. Cash and his co-authors proposed refinements which resulted in greater locality of the encrypted index entries, guaranteeing that entries matching a given search term cluster near each other in the index, thus reducing the number of encrypted index blocks which must be sent. Cash and Tessaro [7] continued improving their previous SE schemes, working on maximizing data locality--the number of non-contiguous storage groups from which the server reads data to satisfy a given search query. They note--and go on to formally prove--how this optimization runs counter to the need to reduce the size of the index storage on the server. Building further on that research, Asharov, et al. [8] created SE schemes with improved read efficiency (they report \(O(\log n)\) and \(O(\log\log n)\)) and demonstrated that it will always be the case that to achieve either maximal read efficiency and locality it will be necessary to sacrifice one to achieve the other. Finally, Demertzis and Papamanthou [2] improved on these earlier efforts by developing a scheme which provides reasonable locality, including controls the repository owner may adjust to "tune" the storage to be maximally efficient for the type of data being indexed. My research is directly based on the work of Demertzis and Papamanthou, whose scheme I extended to include multiple index collections and dynamic up dates. ### Security of SE Systems The observation above that SE systems will "leak" information over time from which an observer can infer confidential information raises the question of how much information leakage is acceptable. This issue has been explored at length by previous researchers. Song, Demertzis, and their colleagues who developed their respective SE implementation models (e.g., [3, 2]) provided with them formal proofs of the security of their encryption schemes. This was important to establish the trustworthiness of the SE concept in general. Following on from this foundation, Naveed, Kamara, and Wright [9] along with Zhang, Katz, and Papamanthou [10] studied various attack scenarios and found that it was possible for a determined observer to eventually decrypt a significant amount of information out of an encrypted database stored on an untrusted server. These findings helped drive Demertzis and Papamanthou to develop more cryptographically robust encryption schemes which I also used for my work, and prompted me to seek a means of periodically invalidating accumulated inferences an observer may have gleaned as part of my update procedure. ### Locality Optimization As noted above, early SE research posited in-memory solutions for the sake of theoretical exploration, but this presented a roadblock to adapting SE systems for real-world applications as it didn't allow the indexes to scale up to the data sizes needed in the real world. To address this, a number of storage strategies were proposed by Cash, et al., [6, 7] but these often ran into difficulties. For example, the practice of obfuscating the layout of the index by permuting index data throughout the storage area came at the expense of having related data clustered to make reads more efficient. Demertzis and Papamanthou [2] proposed one improvement which I found advantageous enough to base my own work upon. Given some array \(A\) of storage locations on the server, this is organized into tiered _levels_\(A_{0},A_{1},\ldots,A_{\ell}\), where each level \(A_{i}\) consists of a number of _buckets_\(A_{i}[0],A_{i}[1],\ldots,A_{i}[q_{i}]\) in which we will store document IDs. At each level, the bucket sizes increase exponentially. For example, one level would hold buckets containing 2 references, the next level would hold buckets of size 4, the next 8, the next 16, and so forth. The documents themselves are stored as encoded (\(w,\mathsf{id}(D)\)) tuples as described above on p. 4. This arrangement nicely facilitates our need to populate the index with document IDs where the number of IDs matching any given keyword varies widely from the number matching another keyword, while allowing us to co-locate these tuples within \(L\) buckets for efficiency. By adjusting the value of \(L\) at index creation time, the SE administrator can reduce the locality of the tuple storage but gain the ability to split up searches into parallel tasks. They also introduced the optimization parameter \(s\) which allows an index to be built with only a subset of levels actually used. Specifically, for \(s=\ell\), all levels are utilized, with each level \(i\) containing buckets sized to hold \(2^{i}\) tuples. If \(s\) is reduced to some value \(1\leq s\leq\ell\), however, the set of actual levels utilized will be \[\mathcal{L}=\{\ell,\ell-p,\ldots,\ell-(s-1)p\}\] and the tuples will be stored in the nearest actual level to the one it would have been assigned if all were allocated. ## 5 My Contributions I focused my work in two specific areas: to create a working production-scale SE implementation based on Demertzis and Papamanthou's model, [2] and then to develop a system to add more information to the SE index over time. Their procedure for building a new SE index is summarized in Algorithm 1. ### Real-World Implementation I investigated two avenues for implementing a remotely hosted SE indexing system. The first was to implement an indexing scheme that maintained its indexes in local files. This was done with some straightforward Python programs: * genkeys generates a set of cryptographic keys \(K=(k_{1},k_{2},k_{3})\) for use by the client to encrypt and decrypt index information. * buildindex reads a collection of documents, extracting a list of words from each. These words are then encoded into an encrypted index stored as a simple dbm database. * search takes a search query from the user, looks that up in the SE index, and reports the matching list of document IDs to the user. Since these operate on local files, all scripts are considered "client-side" in terms of having access to secret keys. In this model, I had in mind an implementation where another operating system layer--transparent to the SE code--handles remote hosting of the files. My choice for this was the InterPlanetary File System (ipfs), [11] which provides a distributed filesystem between clients, so each user sees a copy of the same underlying files, allowing a purely localized operation. However, while that provides for simplicity of SE implementation, it comes at too high a cost for the widest audience since it requires substantial local data storage to be available on every client. It did, however, serve to demonstrate the correctness of the basic SE operations themselves before adding the extra complexity of network operations. From there I switched to a traditional client-server model, defining a hard separation of duties between the data host (which may be remote and untrusted) and the local client. This is implemented in a new set of Python programs: * fseserver runs on the server system to carry out requests on behalf of clients. This manages the dbm databases which comprise the SE index. * buildindex_client works as buildindex does but rather than building local database files, it encrypts a new index and sends it to the server for remote storage. * search_client takes a search query from the user, computes a search token \(T\) as described on p. 6, and passes that to the server. It relays the search results from the server back to the local user. When designing the client-sever protocol for my implementation, one of my significant design goals was to allow large blocks of raw binary data since so much of the index is encrypted and needs to be sent as whole buckets at a time. This way I would not waste resources encoding and decoding the binary blocks as, e.g., base-64. ### Dynamic SE The key intention of my work on SE systems was to find a practical implementation of _dynamic_ SE indexing. In today's world it seems quite likely that an SE indexing system would be gainfully employed to help users search through a fluid stream of communication, such as an email inbox or the conversation history of an official chat service such as those used for customer support by some companies. Most of the work to this point has referenced the idea of building an index \(\mathcal{I}\) from a set of documents \(\mathcal{D}\) and set of search terms \(\Delta\) in a single indexing operation. Once \(\mathcal{I}\) is built, it is then searched any number of times but there is no notion of \(\mathcal{I}\) changing over time. Indeed, the way \(\mathcal{I}\) is constructed in the first place depends on values such as \(N\) (the number of \((w,\mathsf{id}(D))\) tuples stored) and the distribution of words \(w\in\Delta\) throughout the data set and the number of documents \(\mathcal{D}(w)\) for each. If those values change, the internal arrangement of the whole index may be different. This implies that updating \(\mathcal{I}\) over time is necessarily a matter of rebuilding it again from scratch. However, this is obviously untenable for large indexes (e.g., when adding a single email message to an existing index holding 10 million messages already). Cash, et al. [6] discuss their own approach to dynamic SE schemes and the limitations they and others encountered. They note that prior schemes "had an impractically large index or leaked [information allowing the server to learn] the pattern of which keywords appear in which documents... which is a severe form of leakage." They go on to improve on that but make the assumption that full index rebuilds will be performed periodically and that deletions from the index will be rare. However, the approach they took does not lend itself to the tiered architecture I am working with. This prompted me to implement a new dynamic update system which is compatible with the tiered organization so I can retain the optimizations afforded by that structure. Demertzis and Papamanthou [2] do discuss dynamic updates to SE systems, but only to a limited extent. While acknowledging the shortcomings of previous attempts, their proposal was sketched out in basic terms: "The main idea is that we organize \(n\) sequential updates to a collection of... independent encrypted indexes.... [For each \((w,\mathcal{D}(w))\) tuple mapping a search word to a document ID,] the data owner initializes a new SE scheme by creating a new SE index that contains only the specific tuple, [that] is subsequently uploaded to the untrusted server. Whenever two indexes of the same size \(t\) are detected there [sic] are downloaded by the data owner, decrypted and merged to form a sew SE index of size \(2t\), again with a fresh secret key. The new index is then used to replace the two indexes of size \(t\)." To provide real-world practicality to the scheme, I chose to modify this to avoid unnecessarily creating many tiny indexes which will trigger many rapid successions of cascading merges with existing indexes. My design introduced the notion of an SE index _order_\(o=\lceil\log N\rceil\). Rather than storing a single index \(\mathcal{I}\), the server will maintain a collection of indexes. Let \(\mathcal{O}\) be the set of orders of indexes currently stored on a server. Then \(\mathcal{C}=\{\mathcal{I}_{o}\ \forall o\in\mathcal{O}\}\) is the collection of all SE indexes which may logically appear to the client as "the SE index." When asked to perform a search, the server will perform the same operation on all indexes in \(\mathcal{C}\), returning the union of their results. For this scheme to work, I further impose the restriction that no two indexes of the same order may exist at the same time in a collection (i.e., all elements of \(\mathcal{O}\) are unique). When new data are to be added to the SE index, a new index is built for the new data, resulting in a new index \(\mathcal{I}_{o}\). If no other index of order \(o\) exists in \(\mathcal{C}\), then \(\mathcal{I}_{o}\) is added to \(\mathcal{C}\). Otherwise, the contents of the existing \(\mathcal{I}_{o}\) are merged with the new data to create a new index \(\mathcal{I}_{p}\). Then \(\mathcal{I}_{p}\)_replaces_ the old \(\mathcal{I}_{o}\) in \(\mathcal{C}\). It may or may not be the case that \(o=p\). (If, at this point, there is an existing \(p\)-order index in \(\mathcal{C}\), then this process is repeated until all the cascading merges have resulted in an index of an order not currently on the server.) By using this exponential progression in index sizes, I seek to minimize the amount of index data rebuilt at any given time. The larger-order indexes (which will have more data in them) are merged less frequently than the smaller, lower-order ones. Implementing this feature required a compromise to be added to the original scheme proposed by Demertzis and Papamantou--I added an encrypted list \(\Delta\) of all indexed search terms. Only the client can decrypt this list, and it is never referenced during normal operations. It is only accessed and decrypted by the client during merge operations. Doing this is necessary because the SE index is set up to be useful only if a client already knows what search terms they're looking for, so there was no previous reason to store \(\Delta\) inside the index. Thus, no means were provided to reconstruct the original set \(\Delta\) of search terms that was used. Without that information, the rest of the index cannot be decoded back into the original set of tuples. My updated index-building process (replacing Algorithm 1) is summarized in Algorithm 2. ``` 1:\(\mathcal{I}_{o}\) = \(\mathcal{I}_{o}\) \(\forall o\in\ ``` procedureIndexGen(\(k,\mathcal{D}\))\(\triangleright\)Modifies and extends original Setup(\(k,\mathcal{D}\)) Let \(\Delta\) be the list of all "interesting" keywords in document set \(\mathcal{D}\). \(N\leftarrow\sum_{\forall w\in\Delta}|\mathcal{D}(w)|\) \(\ell\leftarrow\lceil\log N\rceil\) \(\mathcal{L}\leftarrow\{\ell,\ell-p,\ldots,\ell-(s-1)p\}\). Let order \(o=\ell\). We will now build a new order-\(o\) index \(\mathcal{I}\). if\(L>1\)then \(\mathcal{L}\leftarrow\mathcal{L}\cup\{0\}\) endif if An order-\(o\) index already exists on the server then Retrieve and decrypt \(\Delta_{o}\) from server from existing order-\(o\) index \(\mathcal{I}_{o}\) Retrieve and decrypt all storage buckets holding actual data from existing index \(\mathcal{I}_{o}\) to \(\mathcal{D}_{o}\) \(\mathcal{D}^{\prime}\leftarrow\mathcal{D}\cup\mathcal{D}_{o}\) Delete old index \(\mathcal{I}_{o}\) returnIndexGen(\(k,\mathcal{D}^{\prime}\)) endif \(p\leftarrow\lceil\ell/s\rceil\) \(\forall i\in\mathcal{L}\) organize storage level \(A_{i}\), divided into buckets \(A_{i}[x]\). for each keyword \(w\in\Delta\) in random order do Find adjacent \(i,j\in\mathcal{L}:L2^{j}<|\mathcal{D}(w)|\leq L2^{i}\). Split \(\mathcal{D}(w)\) into a set of chunks \(C_{w}\). Set \(c=0\). for each chunk \(v\in C_{w}\)do \(c\gets c+1\). Let \(A\) be buckets in \(A_{i}\) able to hold chunk \(v\). Pick one bucket \(A_{i}[x]\) from \(A\) randomly; store \(v\) in it. Add \(H(F(k_{1},w)\mathbin{\|}c)\Rightarrow[i\mathbin{\|}x]\oplus H(F(k_{2},w) \mathbin{\|}c)\) to hash table. endfor endfor Encrypt \(\Delta\) and store in hash table in blocks of 100 words. Permute and encrypt entries in \(\mathcal{L}\); fill HT with random data. Upload \(\mathcal{I}\) to server. endprocedure ``` **Algorithm 2** Create or Add to a Dynamic Index search operations. The remainder of the evaluation is focused on performance characteristics of the update function itself. ### Experimental Methodology To evaluate the efficiency of my implementation, I set up test cases based on real-world data samples representative of highly dynamic document sets. I specifically chose email to approximate casual conversations which include both message text and metadata headers. Indexing archives of chat rooms, text messages, and other similar communication would be analogous. My dataset was taken from the Enron email archive, [12] which is a collection of 517,401 actual email messages belonging to 150 users.3 Since this is a collection of actual communication between users, it provides a valuable test case to simulate how my SE implementation would fare in a real application. As noted by Klimt and Yang, "It can be used both as a large sample of real life email users and as a standard corpus for comparison of results using different methods." [13] Footnote 3: I used a privacy-cleaned list which is slightly smaller than the original release of the dataset. For the purposes of the evaluation, each email message is stored in an individual disk file (in Maildir format). The document ID (\(\mathsf{id}(D)\)) for each is the system's inode number for the message file. This provides a guaranteed unique ID for each file without the overhead of assigning custom IDs. One experimental run of my dynamic SE implementation looked at the case of maintaining a comprehensive index of all messages. To simulate a realistic flow of new messages coming into the repository, I batched up the messages according o the original Date: header lines. I ran three separate experiments, using batch sizes of 1, 7, and 30 days, to measure the performance of the implementation if it were to be re-indexing on a daily, weekly, or monthly basis respectively. Each of these operated on five subsets of the Enron data corpus, organized as shown in Table 3 by the recipient's initials. This is meant to simulate an arbitrary partitioning of a workforce into "departments" which may have slightly different input patterns. Figure 3 shows the day-by-day intake of \((w,\mathsf{id}(D))\) tuples for each of the departments. We see from this that although there are differences in activity day-to-day, the overall pattern of activity was similar, giving us five sample sets to compare and contrast. I observed consistent behavior among all five departments as I examined the server resource usage as each of their document indexes expanded over time. With the input data thus broken into batches of various sizes, I ran each set of updates while measuring the following performance characteristics after each update operation: * Size \(N\) of each index \(\mathcal{I}_{o}\) in terms of \(N\) * Disk storage on the server for each \(\mathcal{I}_{o}\) * Full set of words \(\Delta\) added in that update * Contents of all storage locations in \(A_{i}\) * Histogram of distribution of each keyword \(w\) within the input document set \(\mathcal{D}\) * Wall-clock time in seconds taken to build the new index (including any needed merges with existing ones) * Number of network transactions required to perform the update operation * Number of bytes exchanged between client and server during the update operation * Number of input documents added during that update * If merging, how many words and tuples from previous indexes were downloaded to be merged into the new index * Which orders of indexes were merged at each update ### Results Using Department 1 as a representative example of the results obtained, we see in Figure 4 how the processing time varied as a function of the frequency of updates. At first glance, it is apparent that we can get an overall savings in processing work by doing updates less frequently. For example, in the close-up view in Figure 5 we see that over the same period of time the daily and weekly updates saw spikes of activity as they had to merge multiple indexes but the monthly updates (blue line) did not, since it had the advantage of performing the update operation with more data at a time locally in a single step rather than making more incremental updates which needed to be downloaded again to be merged later. This prompted me to seek a predictor for these merge events, as a potential avenue to optimize the merge by anticipating when merges would be necessary. While there are some straightforward (if naive) indicators we could use, such as the number of incoming documents \(|\mathcal{D}|\) or the number of input words \begin{table} \begin{tabular}{c c c c c c} \hline \hline **Department** & **Initial** & **People** & **Messages** & **Data (kB)** \\ \hline 1 & A–F & 33 & 119,477 & 667,632 \\ 2 & G–K & 29 & 143,652 & 760,388 \\ 3 & L–P & 32 & 103,354 & 504,668 \\ 4 & Q–S & 38 & 105,930 & 526,084 \\ 5 & T–Z & 18 & 44,988 & 225,600 \\ \hline \hline \end{tabular} \end{table} Table 3: Arrangement of Users into Departments Figure 3: Incoming Keyword Tuples For All Departments Figure 4: Processing Time (Dept. 1) by Batch Size Figure 5: Processing Time (Dept. 1) by Batch Size, Detail View \(|\Delta|\) or even the number \(N\) of incoming tuples, none of these is a completely accurate predictor of either the time a merge will happen, or of the magnitude of each merge event. The reason for this is that the conditions which trigger a merge event depend on the exact distribution of keywords (\(w\)) in each existing index \(\mathcal{I}_{o}\in\mathcal{C}\) as well as how the specific incoming data will alter that distribution. As we see in the data sample in Figure 6, a given input batch may have a few messages with a high diversity of input words (driving \(N\) significantly higher than \(|\mathcal{D}|\)) or vice versa. If we want a loose correlation to track with the performance over time of the SE updates, \(N\) would still seem the best reasonable value that is easily at hand, which may be useful for long-range statistical analysis including the prediction of incoming message volume to the system overall, from which merge probability may be inferred. ### Issues I discovered a pathological condition as an index hit certain sizes requiring large-scale cascading re-indexing operations. Most updates were completing in seconds, but these exceptional cases were taking many hours. In in some cases they took days to complete. For example, among the daily updates for Dept. 1 there were 3 out of the total 995 update batches (0.3% of the total) which took more than one day to complete. These took 1 day, 1 hour; 1 day, 14 hours; and 4 days, 15 hours. Further investigation led to an initial diagnosis that this was likely caused by a combination of resource scarcity on the server and inefficiencies in the Python implementation which still need to be optimized further. For example, at some points an index collection included an order-24 index. This would contain a small number of buckets of size 33,554,432 words. For 64-bit words (as my implementation uses to store the \(i\|\,x\) encoded values), that's approximately a quarter-gigabyte string value to be assembled, copied, transmitted, received, buffered, copied, and decoded. In a runtime system such as Python's, that is not necessarily going to be handled as efficiently as it might be. ### Summary of Results Overall, the results indicate that even this experimental Python implementation performed well enough to be of practical use. Updates finished in time for the index to be used for a reasonable time before the next update batch needed to start. The rare exceptions to this were due to the pathological case described in the previous section. Table 4 summarizes the runtime results of each of the experiments. ## 7 Future Work Given the success of the experimental Python implementation, it makes sense to continue optimizing this design by coding it in a more efficient runtime system based on a language such as C or Go, as well as to continue looking for ways to optimize the server protocols and possibly the storage system itself. Specifically, the cause of the occasional very-long update operations should be investigated more. This system also needs to be expanded to include the concept of deletion of messages from the index. Finally, a formal evaluation of the cryptographic strength, including the likelihood and impact of potential information leakage over time when using this design compared to other SE schemes. I did not examine the effects of changing the locality and storage variables \(L\) and \(s\) since that was already thoroughly treated by Demertzis and Papamanthou [2] in their proposal for this SE architecture initially. However, it would be interesting to come back to that once my dynamic changes to their design have matured and evolved into a fully workable model with deletion support, to see if then the effect of adjusting those variables is different from what was found previously. ## 8 Conclusions I have expanded on the work of previous SE researchers to implement an experimental yet functional dynamic SE indexing service. With that in place, I have analyzed the runtime performance with the update batch size as the independent variable for my experiments. I concluded from those experiments that the scheme as described in this paper is practical for real-world applications. Further, I identified how more efficient updates (requiring less overall work) are achieved by delaying updates for longer periods of time (e.g., weekly or monthly rather than hourly or daily). However, this comes at the cost of not having that new data available for users of the index. It is necessary for a server administrator to determine what update frequency serves the needs of their users best. This work contributes to the future implementation of real-world encrypted document indexing systems which can be employed to organize repositories \begin{table} \begin{tabular}{l c c c c c c c c c c} \hline \hline & \multicolumn{3}{c}{**Dept. 1**} & \multicolumn{3}{c}{**Dept. 2**} & \multicolumn{3}{c}{**Dept. 3**} & \multicolumn{3}{c}{**Dept. 4**} & \multicolumn{3}{c}{**Dept. 5**} \\ **Sample** & **Avg** & **SD** & **Avg** & **SD** & **Avg** & **SD** & **Avg** & **SD** & **Avg** & **SD** \\ \hline Daily & 21.14 & 240.75 & 4.75 & 47.87 & 11.07 & 111.57 & 12.49 & 119.33 & 0.99 & 11.20 \\ Weekly & 110.92 & 576.32 & 25.28 & 117.14 & 53.82 & 242.18 & 72.53 & 301.46 & — & — \\ Monthly & 322.31 & 855.13 & — & — & 197.31 & 504.38 & 254.61 & 583.68 & 96.26 & 341.74 \\ \hline \hline \end{tabular} \end{table} Table 4: Summary of Experimental Results (batch wall-clock time in minutes) Figure 6: Incoming Tuple Count vs. Message Count (Dept. 1) of chat logs, emails, discussion forums, or other dynamic document collections while protecting the confidentiality of the content being served. ## 9 Acknowledgements I wish to express gratitude to the guidance provided by Dr. Charles V. Wright, my Ph.D. advisor, who supervised this research and provided valuable feedback along the way. Also to Dr. David Maier, for his advice and instruction on writing styles for the earliest drafts of this paper.

2305.02577

Text Reading Order in Uncontrolled Conditions by Sparse Graph Segmentation

Text reading order is a crucial aspect in the output of an OCR engine, with a large impact on downstream tasks. Its difficulty lies in the large variation of domain specific layout structures, and is further exacerbated by real-world image degradations such as perspective distortions. We propose a lightweight, scalable and generalizable approach to identify text reading order with a multi-modal, multi-task graph convolutional network (GCN) running on a sparse layout based graph. Predictions from the model provide hints of bidimensional relations among text lines and layout region structures, upon which a post-processing cluster-and-sort algorithm generates an ordered sequence of all the text lines. The model is language-agnostic and runs effectively across multi-language datasets that contain various types of images taken in uncontrolled conditions, and it is small enough to be deployed on virtually any platform including mobile devices.

Renshen Wang, Yasuhisa Fujii, Alessandro Bissacco

2023-05-04T06:21:00Z

http://arxiv.org/abs/2305.02577v1

# Text Reading Order in Uncontrolled Conditions by Sparse Graph Segmentation ###### Abstract Text reading order is a crucial aspect in the output of an OCR engine, with a large impact on downstream tasks. Its difficulty lies in the large variation of domain specific layout structures, and is further exacerbated by real-world image degradations such as perspective distortions. We propose a lightweight, scalable and generalizable approach to identify text reading order with a multi-modal, multi-task graph convolutional network (GCN) running on a sparse layout based graph. Predictions from the model provide hints of bidimensional relations among text lines and layout region structures, upon which a post-processing cluster-and-sort algorithm generates an ordered sequence of all the text lines. The model is language-agnostic and runs effectively across multi-language datasets that contain various types of images taken in uncontrolled conditions, and it is small enough to be deployed on virtually any platform including mobile devices. Keywords:Multi-modality, bidimensional ordering relations, graph convolutional networks. ## 1 Introduction Optical character recognition (OCR) technology has been developed to extract text reliably from various types of image sources [4]. Key components of an OCR system include text detection, recognition and layout analysis. As machine learning based digital image processing systems are nowadays ubiquitous and widely applied, OCR has become a crucial first step in the pipeline to provide text input for downstream tasks such as information extraction, text selection and screen reading. Naturally, most image-to-text applications require very accurate OCR results to work well. This requirement is not only on text recognition -- reading order among the recognized text lines is almost always as important as the recognition quality. The reason is self-evident for text selection (copy-paste) and text-to-speech tasks. And for structured document understanding like LayoutLM [33], DocFormer [3], FormNet [18], etc., the order of the input text also has a profound effect as most of these models have positional encoding attached to input text features, and a sequential labeling task for output. Input text order can sometimes be the key factor for the successful extraction of certain entities. Depending on the text layout, the difficulty of deciding its reading order varies greatly. It can be as simple as sorting all the text lines by y-coordinates, but can also be hard like the images in Figure 1. Even if we exclude corner cases like these, there are still complexities brought by the diversity of layout structures which are often domain specific. Previous studies have tackled the problem in different ways. Rule based approaches like [1, 27, 9] usually aim at one specific domain, while learning based approaches like [6, 21, 32] are more general but have scalability issues (more discussions in the following section). In this paper, we propose a composite method that uses both machine learning model and rule based sorting to achieve best results. It is based on the observation from [1] that most reading order sequences are in one of the two patterns -- column-wise and row-wise -- as illustrated in Figure 2. We use a graph convolutional network that takes spatial-image features from the input layout and image, and segments the layout into two types of regions where the paragraphs can be properly sorted by the type of their patterns. A \(\beta\)-skeleton graph built on boxes [31] enables efficient graph convolutions while also providing edge bounding boxes for RoI (regions of interest) pooling from the image feature map. A post-processing cluster-and-sort algorithm finalizes the overall reading order based on model predictions. This unique combination gives us an effective, lightweight, scalable and generalizable reading order solution. ## 2 Related Work Two types of related work are discussed in this section. The first subsection includes previous reading order efforts, and the second subsection discusses other multi-modal image-text-spatial models that share some of the components with our approach. ### Reading Order Detection Previous studies have tackled the reading order problem in various ways. We roughly categorize them into rule based sorting [5, 1, 27, 9] and machine-learning based sequence prediction [6, 21, 32, 29], etc. Figure 1: Hard examples for text reading order. (a) A cropped image of a menu with dish names and prices, where a correct reading order necessarily needs correct association between each dish name and its price, which is a hard task for humans without the full image context due to the perspective distortion in the image. (b) A text layout intentionally made to have two different reading order interpretations, both valid, but with completely opposite meanings. Topological sort was proposed in [5] for document layout analysis where partial orders are based on x/y interval overlaps among text lines. It can produce reading order patterns like Figure 2 (a) for multi-column text layouts. A bidimensional relation rule proposed in [1] provides similar topological rules, and in addition provides a row-wise rule by inverting the x/y axes from column-wise. An argumentation based approach in [9] works on similar rules derived from text block relations. For large text layout with hierarchies, XY-Cut [27; 13] can be an effective way for some layout types to order all the text blocks top-to-bottom and left-to-right. These rule based approaches can work accurately for documents in certain domains. But without extra signals, they will fail for out-of-domain cases like Figure 2 (b). Machine learning based approaches are designed to learn from training examples across different domains to enable a general solution. The data mining approach in [6] learns partial order among text blocks from their spatial features and identifies reading order chains from the partial orders. A similar approach in [29] trains a model to predict pairwise order relations among text regions and curves for handwritten documents. The major limitation is that any partial order between two entities are derived from their own spatial features without the layout structure information in their neighborhood. So these models may not be able to identify the layout structure among a group of text lines and therefore fail to find the correct pattern. Graph convolutional networks and transformer models provide mechanisms for layout-aware signals by interactions between layout entities. A text reorganization model introduced in [21] uses a graph convolutional encoder and a pointer network decoder to reorder text blocks. With a fully-connected graph at its input, the graph encoder functions similarly as a transformer encoder. Image features are added to graph nodes by RoI pooling on node boxes with bilinear interpolation. Another work LayoutReader [32] uses a transformer based architecture on spatial-text features instead of spatial-image features to predict reading order sequence on words. The text features enable it to use the power Figure 2: Two major patterns of reading order. (a) Column-wise order, most common in printed media like newspapers and magazines. (b) Row-wise order, usually in receipts, forms and tabular text blocks. ful LayoutLM [34] model, but also make it less generalizable. These models are capable of predicting reading order within complex layout structures. However, there are scalability issues in two aspects: * Run time scales quadratically with input size. Whether in the graph convolutional encoder with full connections or the sequence pointer decoder, most of the components have \(O(n^{2})\) time complexity, and may become too slow for applications with dense text. * Accuracy scales inversely with input size. The fully-connected self-attention mechanism in the encoder takes all the text entities to calculate a global attention map, which introduces noises to the reading order signals that should be decidable from local layout structures. The sequence decoder uses softmax probabilities to determine the output index for each step, where the output range increases with input size, and so does the chance of errors. Figure 10 illustrates this limitation from our experiments. To summarize briefly, there are multiple effective ways to order OCR text by rule based or machine learning based methods, and in both categories there is room for improvement in generalizability and scalability. ### Spatial, Image Features and Multi-Modality Multi-modal transformer models have become mainstream for document or image understanding tasks. Related work include LayoutLM [34, 33, 15, 13], DocFormer [3], SelfDoc [22], UDoc [12], StrucText [23], TILT [28], LiLT [30], FormNet [18], PaLI [7], etc. Document image understanding starts with an OCR engine that provides text content as the main input for the language model. Alongside, the text bounding boxes associated with the words and lines provide important spatial features (sometimes called layout features or geometric features). Additionally, since not all visual signals are captured by the OCR engine, an image component in the model can help cover the extra contextual information from the input. Thus, a model to achieve best results should take all of the three available modalities. For image features, most previous studies use RoI pooling [8] by the text bounding boxes from OCR, and the pooled features are attached to the corresponding text entity. It is effective for capturing text styles or colors, but less so for visual cues out of those bounding boxes, such as the curly separation lines in Figure 3. While it is possible to use an image backbone with large receptive fields, like ResNet50 used in the UDoc model or U-Net used in the TILT model, it is not an ideal solution for two reasons: * In sparse documents, useful visual cues can be far from any text on the page. * Large receptive fields bring in extra noise from regions irrelevant to the features we need. Thus, it will be more effective to have image RoI boxes that cover pairs of text bounding boxes. A sparse graph like \(\beta\)-skeleton used in [31] can provide the node pairs for such RoI pooling without significantly increasing the model's memory footprint and computational cost. ## 3 Proposed Method Based on previous studies, we design a lightweight machine learning based approach with a model that is small in size, fast to run, and easy to generalize in uncontrolled conditions. ### Strong Patterns of Reading Order From a set of real-world images annotated with reading order, we have an observation that matches very well with the bidimensional document encoding rules in [1] -- column-wise text usually has a zigzag pattern of Figure 2 (a), and row-wise text has a similar but transposed zigzag like Figure 2 (b). Some images may contain both types of text, which makes the pattern more complex. But once the column-wise/row-wise type of a text region is decided, the reading order in this region mostly follows the pattern and can be determined with a topological sort according to the bidimensional rules. Figure 7 (a) shows an example of an annotated reading order sequence. Based on this observation, learning text reading order becomes an image segmentation problem, as opposed to learning arbitrary global sequences of text entities. Instead of predicting the next entity in the entire image, we do a binary classification for each text entity on whether it's in a column-wise or row-wise pattern. Moreover, the pattern classification for a text line can be decided by local layout structures, and global attention maps are therefore unnecessary. ### Model Architecture We use a graph convolutional network (GCN) with a sparse graph construction because of the three major advantages listed here: * GCN models are equivariant to input order permutations. It is natural to assume that a model deciding reading order should not depend on the order of its input. * With a sparse graph like \(\beta\)-skeleton, GCN computation scales linearly with input size. * Graph edges constructed from text boxes can provide edge bounding boxes, which are better for image feature RoI pooling (Figure 3, Table 2). As illustrated in Figure 4, we use an MPNN [11] variant of GCN as the main model backbone, and a \(\beta\)-skeleton graph [17] constructed with text line boxes as nodes. Similar configurations have been applied to other layout problems [19, 31, 25, 18], and graph construction details are available in [31]. The main GCN input is from the spatial features of text line bounding boxes as node features, including \(x\), \(y\) coordinate values of the box corners, and the coordinate values multiplied by rotation angle coefficients \(\cos\alpha\), \(\sin\alpha\). The spatial features go through \(T\) steps of graph convolution layers, each containing a node-to-edge "message passing" layer and edge-to-node aggregation layer with attention weighted pooling. Besides the main input from nodes, we add a side input of edge features from edge box RoI pooling on an image feature map to help capture potential visual cues surrounding text boxes. We use MobileNetV3-Small [14] as the image backbone for its efficiency. Note that the purpose of this image backbone is not for a major task like object detection, but to look for auxiliary features like separation lines and color changes, so a small backbone is capable enough for our task. For the same reason, we reduce the MobileNetV3 input image size to 512\(\times\)512 to speed up training and inference. The details of the image processing are illustrated in Figure 5. In most cases, the text content is no longer Figure 4: Overview of the reading order multi-classifier model. Node classification predicts the reading order patterns, and edge classification predicts paragraph clustering. Figure 3: A cropped example of a \(\beta\)-skeleton graph [31] constructed from text line boxes. Graph node boxes are shown in green and edge lines in cyan. The three orange colored text lines demonstrate how image features can help — the 2nd and 3rd boxes are closer in distance, so spatial features may indicate they are in the same section, but the curly separation line between them indicates otherwise. The yellow box at the bottom is the minimum containing box of the two line boxes inside, where the RoI pooling can cover image features between these lines. recognizable after such downsizing, but the auxiliary image features can be well preserved. We also make sure that the entire layout is contained in a circle of diameter 512 within the processed image, which enables random rotations during model training -- a key augmentation for our model to work in all conditions. Language features are not included in order to keep the model minimal in size and independent of domain knowledge. Also, our annotated reading order data is limited in English only, upon which we try to train a universal model. The GCN is a multi-task model that outputs both node and edge predictions. At node level, it predicts the reading order pattern on each line box (column-wise or row-wise). These predictions are essentially a segmentation for text regions where the lines can be sorted accordingly. At edge level, the model predicts whether the two lines connected by an edge belong to the same paragraph. Thus, it works like the edge clustering models in [31, 25], and we can improve the final reading order by grouping lines together within each paragraph. The reading order estimation by the grouping usually do not affect column-wise order among text lines, but can be critical in row-wise regions such as tables or forms with multi-line cells, e.g. Figure 9 (d). It may be considered that a fully convolutional network can do similar segmentation tasks like [26, 16] on the input image. However, we have observed that such models are less effective for certain types of text content -- e.g. in Figure 2 (b), similar lines in the left column are grouped into a large paragraph, disrupting the row-wise reading order. Figure 5: Image processing for the MobileNetV3 input. The inner yellow box is the minimum containing box of all the text lines in the image. If its diagonal \(d\) is larger than 512, we scale down the image by \(\frac{512}{d}\) so that all the text bounding boxes are contained in the white circle of diameter 512, and then we crop (maybe also pad) around this circle to get the final processed image. This process ensures that both the image and the layout can be randomly rotated during training without any line box moved out of boundary. ### Recovering Reading Order from Model Predictions With the \(\beta\)-skeleton graph that provides local connections among dense text boxes, the GCN model predicts on _local_ properties of the text, which can be aggregated to give us a _global_ reading order. To handle mixed column-wise and row-wise predictions as well as potential text rotations and distortions in the input image, we extend the rule based sorting in [1, 5] and propose a hierarchical cluster-and-sort algorithm to recover the global reading order from line-level pattern predictions and clustered paragraphs. The following Algorithm 1 generates a set of clusters, each cluster \(c_{i}\) contains a non-empty set of paragraphs and maybe a set of child clusters. Each cluster is also assigned a reading order pattern \(R(c_{i})\in\{\mathit{col},\mathit{row}\}\), with _col_ for column-wise and _row_ for row-wise. Row-wise text often involves sparse tables with components not directly connected by \(\beta\)-skeleton edges, so the hop edges like in [25] can be helpful in step 4 of algorithm 1. More details can be added, e.g. setting an edge length threshold in step 3 to avoid merging distant clusters. ``` 1. Cluster lines into paragraphs \(p_{1},...,p_{n}\) from edge predictions. 2. Each paragraph is initialized as a cluster, \(c_{i}=\{p_{i}\}\). Reading order pattern \(R(c_{i})\) is the majority vote from the paragraph's line predictions. 3. For each edge \((i,j)\in G\), find cluster \(c_{a}\) containing line \(i\) and \(c_{b}\) containing line \(j\); if \(R(c_{a})=R(c_{b})=col\), merge \(c_{a}\) and \(c_{b}\) into a bigger column-wise cluster. 4. For each edge \((i,j)\in G\) or hop edge \((i,j)\) (\(\exists k\) that \((i,k)\in G\) and \((k,j)\in G\)), find cluster \(c_{a}\) containing line \(i\) and \(c_{b}\) containing line \(j\); if \(R(c_{a})=R(c_{b})=\mathit{row}\), merge \(c_{a}\) and \(c_{b}\) into a bigger row-wise cluster. 5. Calculate the containing box for each cluster. The rotation angle of the box is the circular mean angle of all the paragraphs in the cluster. 6. Sort the clusters by ascending area of their containing boxes. 7. For each cluster \(c_{i}\), if its containing box \(B(c_{i})\) overlaps with \(B(c_{j})\) by area greater than \(T\times Area(B(c_{i}))\), set \(c_{i}\) as a child cluster of \(c_{j}\). 8. Create a top level cluster with all the remaining clusters as its children. ``` **Algorithm 1**Hierarchical Clustering Once the regions of reading order patterns are decided by the hierarchical clusters, we can use topological sort within each cluster as in Algorithm 2. ``` 1. Cluster lines into paragraphs \(p_{1},...,p_{n}\) from edge predictions. 2. Each paragraph is initialized as a cluster, \(c_{i}=\{p_{i}\}\). Reading order pattern \(R(c_{i})\) is the majority vote from the paragraph's line predictions. 3. For each edge \((i,j)\in G\), find cluster \(c_{a}\) containing line \(i\) and \(c_{b}\) containing line \(j\); if \(R(c_{a})=R(c_{b})=col\), merge \(c_{a}\) and \(c_{b}\) into a bigger column-wise cluster. 4. For each edge \((i,j)\in G\), find cluster \(c_{a}\) containing line \(i\) and \(c_{b}\) containing line \(j\); if \(R(c_{a})=R(c_{b})=\mathit{row}\), merge \(c_{a}\) and \(c_{b}\) into a bigger row-wise cluster. 5. Calculate the containing box for each cluster. The rotation angle of the box is the circular mean angle of all the paragraphs in the cluster. 6. Sort the clusters by ascending area of their containing boxes. 7. For each cluster \(c_{i}\), if its containing box \(B(c_{i})\) overlaps with \(B(c_{j})\) by area greater than \(T\times Area(B(c_{i}))\), set \(c_{i}\) as a child cluster of \(c_{j}\). 8. Create a top level cluster with all the remaining clusters as its children. ``` **Algorithm 2**Hierarchical Clustering With all the clusters sorted, an ordered traversal of the cluster hierarchy can give us the final reading order among all the paragraphs. Figure 6 shows the reading order on a packaging box at different camera angles. Note that the algorithms are not sensitive to bounding box angles, and the model is trained with randomly augmented data, so the rotation has minimal effect on the final result. It can even handle vertical text lines in Chinese/Japanese with the vertical lines regarded as rotated horizontal lines. ### Data Labeling We prepared a dataset with human annotated layout data, including paragraphs as polygons and reading order groups where each group is an ordered sequence of paragraphs. Figure 7 (a) shows a set of paragraphs, where the reading order starts with the green paragraph and follows the jagged line. Figure 6: Reading order example at different angles. Paragraphs with column-wise pattern predictions are shown in yellow, row-wise in pink. The dark blue line shows the overall reading order among all paragraphs. ``` Input: A sequence of ground truth paragraphs \(p_{1},p_{2},p_{3},\cdots,p_{n}\) represented as rectangular boxes. 1. Between each consecutive pair of paragraphs \((p_{i},p_{i+1})\), we categorize their geometrical relation \(R_{i,i+1}\) as one of \(\{\textit{vertical},\textit{horizontal},\textit{unknown}\}\). 1. Calculate \(\alpha\), the circular mean angle of the two boxes' rotation angles. 2. Rotate the boxes of \(p_{i}\) and \(p_{i+1}\) around (0, 0) by \(-\alpha\), denoted as \(b_{i}\) and \(b_{i+1}\). 3. Axis aligned box \(c\) is the minimum containing box of both \(b_{i}\) and \(b_{i+1}\). 4. if \(y_{\textit{overlap}}(b_{i},b_{i+1})<0.1\cdot\textit{height}(c)\) and \(y_{\textit{center}}(b_{i})<y_{\textit{center}}(b_{i+1})\) \(R_{i,i+1}=\textit{vertical}\) 5. else if \(x_{\textit{overlap}}(b_{i},b_{i+1})<0.1\cdot\textit{width}(c)\) and \(x_{\textit{center}}(b_{i})<x_{\textit{center}}(b_{i+1})\) if \(c\) does not cover paragraphs other than \(p_{i}\), \(p_{i+1}\) \(R_{i,i+1}=\textit{horizontal}\) # mostly tabular structures else \(R_{i,i+1}=\textit{vertical}\) # mostly multi-column text 6. In other conditions, \(R_{i,i+1}=\textit{unknown}\) 2. Decide the reading order pattern for paragraph \(p_{i}\) from \(R_{i-1,i}\) and \(R_{i,i+1}\). 1. (_unknown_, _unknown_) \(\rightarrow\)_unknown_ 2. In case of one unknown, the other one decides the pattern: \(\textit{vertical}\rightarrow\) column-wise, _horizontal_\(\rightarrow\) row-wise,. 3. If neither is unknown, \((\textit{vertical},\textit{vertical})\rightarrow\) column-wise, otherwise it is row-wise. ``` **Algorithm 3**Pattern Labeling from Annotated Reading Order Figure 7: Labeling reading order patterns from annotations. (a) Ground truth from human annotated paragraphs and reading order. (b) Reading order pattern inferred from the annotated sequence — column-wise indicated by a vertical/purple line in each paragraph and row-wise by a horizontal/green line. While the edge clustering labels are straightforward from the paragraph polygons, the reading order pattern labeling is less trivial because we need to derive binary labels from ground truths of paragraph ordering. We decide the pattern of a paragraph by comparing its position with its predecessor and successor. Figure 7 (b) shows an example, and detailed logic is elaborated in Algorithm 3. ### Limitations The node-edge classification model can produce reasonable reading order in most cases, but may fail for complex layouts with multiple tabular sections placed closely, like the cross section errors in Figure 11 (a). The root cause is the lack of higher level layout structure parsing with the two classification tasks. Data annotation at section level is generally hard because there is no universal agreement on the exact definition of sections among text. Figure 11 (b) shows the result with extra section level clustering trained on a domain specific dataset. There is significant improvement, yet cross domain generalization is not guaranteed, and we can still see imperfections in the multi-section reading order due to section prediction errors. Another limitation is that our model is not a reliable source for parsing table structures like [24]. Figure 8 shows the reading order result of the image in Figure 1 (a). Note that in the sorting algorithm, we rotate all the bounding boxes to zero out their mean angle. But when the boxes are at different angles due to distortions, there will still be slanted line boxes and misaligned table rows after all the rotations, so the topological sort on the axis-aligned containing boxes cannot guarantee the right order. In presence of tables, a separate model with structure predictions will likely perform better. ## 4 Experiments We experiment with the GCN model with predictions on reading order pattern and paragraph clustering, together with the cluster-and-sort algorithms. Figure 8: Cluster-and-sort result on the cropped menu from Fig. 1 (a). Although the model correctly predicts the row-wise pattern, reading order is still incorrect due to the perspective distortion and the unusually large spacing between the two columns. ### Datasets and Evaluation Metrics Various metrics have been used to evaluate reading order, such as Spearman's footrule distance, Kendall's Tau rank distance used in [29] and BLEU scores in [21]. These metrics can accurately measure order mismatches, but also require full length ground truth order for comparison. We created an annotated layout dataset where reading order ground truths are partially annotated, i.e. some subsets of paragraphs form reading order groups with annotated order, and the order among groups is undefined. This makes it more flexible to match realistic user requirements and less suitable for full ranking metrics. So instead, we use a normalized Levenshtein distance [20] which measures the minimum number of word operations (insertions and deletions) needed to equalize two lists. For each reading order group, we take the ordered list of paragraphs and find all the OCR words \(W\) contained in these polygons. The word order within each paragraph is taken directly from OCR (mostly accurate for a single paragraph). Then we find the shortest subsequence of the serialized OCR output that contains all the words in \(W\), compute its Levenshtein distance to \(W\), and multiply it by the normalization factor \(1/|W|\). Besides our annotated set, we test the model with PubLayNet [35] because of its variety on layout components with different reading order patterns. Although there is no ground truth of reading order, we take "text" instances as paragraphs with column-wise pattern, and "table"/"figure" types as containers of text lines with row-wise pattern. Thus, we are able to train the same multi-task GCN model. The annotated set contains 25K text images in English for training and a few hundred test images for each of the available languages, and PubLayNet contains 340K training images and 12K validation images all in English. ### Model Setup The model is built as shown in Figure 4, with the OCR engine from Google Cloud Vision API producing text lines and their spatial features. Edge image features Figure 9: Reading order results from (a) PubLayNet [35], (b) PRIMA Contemporary dataset [2], (c) the ambiguous example from Figure 1 with a positive interpretation, and (d) our evaluation set. are from a bi-linear interpolation on the MobileNetV3 output with \(16\times 3\) points each box and dropout rate 0.5. The TF-GNN [10] based GCN backbone uses 10 steps of weight-sharing graph convolutions, with node feature dimension 32 and message passing hidden dimension 128. Edge-to-node pooling uses a 4-head attention with 3 hidden layers of size 16 and dropout rate 0.5. Total number of parameters is 267K including 144K from MobileNetV3-Small. We train the model for 10M steps with randomized augmentations including rotation and scaling, so the model can adapt to a full range of inputs. The OCR boxes are transformed together with the image in each training example, resulting in better robustness than previous approaches (Figure 6). ### Baselines Most commercial OCR systems use a topological sort like in [1] with one of the two patterns. We use column-wise pattern in the basic baseline as it produces better scores than row-wise in our evaluations, and is close to the default output order from the OCR engine we use. In addition, we implement a GCN model that directly predicts edge directions on a fully connected graph similar to the model in [21]. Figure 10 shows two examples with comparison between this baseline and our approach, with supports the scalability discussion in subsection 2.1. ### Results We train the multi-task model with PubLayNet and our paragraph reading order set added with the menu photos labelled from human annotations. From Table 1, we can see the difference in the difficulty between the two sets. Real-world images from our dataset have much larger variations on layout styles and image degradations that make the same tasks much harder to learn. \begin{table} \begin{tabular}{c c c c c c c} \hline \hline Dataset & \multicolumn{3}{c}{Reading order pattern} & \multicolumn{3}{c}{Paragraph clustering} \\ & Precision & Recall & F1 & Precision & Recall & F1 \\ \hline PubLayNet & 0.998 & 0.995 & 0.997 & 0.994 & 0.996 & 0.995 \\ Annotated ordered paragraphs & 0.828 & 0.805 & 0.819 & 0.895 & 0.909 & 0.902 \\ \hline \hline \end{tabular} \end{table} Table 1: Scores of the two classification tasks on PubLayNet and our labelled paragraph reading order dataset. \begin{table} \begin{tabular}{c c c c c c c} \hline \hline Boxes for image & \multicolumn{3}{c}{Reading order pattern} & \multicolumn{3}{c}{Paragraph clustering} \\ feature RoI pooling & Precision & Recall & F1 & Precision & Recall & F1 \\ \hline n/a & 0.800 & 0.803 & 0.802 & 0.887 & 0.895 & 0.891 \\ Node boxes & 0.819 & 0.781 & 0.800 & 0.870 & 0.903 & 0.886 \\ Edge boxes & 0.828 & 0.805 & 0.819 & 0.895 & 0.909 & 0.902 \\ \hline \hline \end{tabular} \end{table} Table 2: F1 scores from the image feature ablation test. We also test the effectiveness of the edge box RoI pooling by an image feature ablation test, where the baseline is the model with all image features removed, compared against ones with node box RoI pooling and edge box RoI pooling. Table 2 shows that node box RoI does not help at all, even with a slight accuracy drop compared with the baseline. These results confirm our previous hypothesis that the image backbone mainly helps the model by discovering visual cues out of text bounding boxes, and edge boxes are much more effective for this purpose. Finally, we measure the normalized Levenshtein distance for reading order produced by the GCN and the cluster-and-sort algorithm, and compare it against the two baseline methods in subsection 4.3. As in Table 3, our algorithm can greatly improve reading order quality across all Latin languages, even though the training data is only available in English. The model also works well for examples out of our datasets. Figure 9 includes images from various sources, demon Figure 10: Comparison between the fully-connected graph model and our approach on two receipt examples. The full graph predictions perform well on the sparse example, but fail on the dense one. \begin{table} \begin{tabular}{l|c|c|c|c|c} \hline \hline Language & Training & Test set & All-column-wise & Fully-connected & 2-task GCN \\ & set size & size & baseline & graph baseline & cluster-and-sort \\ \hline English & 25K & 261 & 0.146 & 0.126 & 0.098 \\ French & & 218 & 0.184 & 0.144 & 0.119 \\ Italian & & 189 & 0.172 & 0.145 & 0.122 \\ German & & 196 & 0.186 & 0.162 & 0.112 \\ Spanish & n/a & 200 & 0.183 & 0.103 & 0.097 \\ Russian & & 1003 & 0.202 & 0.159 & 0.148 \\ Hindi & & 990 & 0.221 & 0.181 & 0.152 \\ Thai & & 951 & 0.131 & 0.111 & 0.104 \\ \hline \hline \end{tabular} \end{table} Table 3: Normalized Levenshtein distance (lower is better) on a multi-language reading order evaluation set. Training data is only available in English. starting the effectiveness of our model with inputs ranging from digital/scanned documents to scene images. ## 5 Conclusions and Future Work We show that GCN is highly efficient at predicting reading order patterns and various layout segmentation tasks, which is further enhanced with a small image backbone providing edge RoI pooled signals. Our model is small in size and generalizes well enough to be deployable on any platform to improve OCR quality or downstream applications. In addition, the GCN model has the potential to handle more than two tasks. We tried an extra edge prediction task trained with a dataset of menu photos with section level polygon annotations. Unlike general document or scene text images, menus like Figure 3 usually have clearly defined sections like main dishes, side dishes, drinks, etc. Therefore, the menu dataset has accurate and consistent section level ground truth for model training. The 3-task GCN model provides higher-level layout information to the clustering algorithm and helps produce Figure 11 (b), a major improvement on reading order. Still, there is domain specific knowledge on menu sections that does not always generalize well. And because most evaluation examples have relatively simple layouts, the 3-task model has not produced better results than the 2-task model in our experiments. Nevertheless, we think section level ground truth or higher-level layout structural information will be valuable for further reading order improvements. Future work will explore the possibilities of both data and modeling approaches for parsing layout structures. Figure 11: A multi-section example. (a) Paragraphs with row-wise pattern are clustered into overly large regions, causing incorrect cross-section reading order. (b) With section level clusters (shown in orange) added into Algorithm 1, multi-table results can be improved. #### Acknowledgements The authors would like to thank Ashok C. Popat and Chen-Yu Lee for their valuable reviews and feedback.

2306.00630

Class Anchor Margin Loss for Content-Based Image Retrieval

The performance of neural networks in content-based image retrieval (CBIR) is highly influenced by the chosen loss (objective) function. The majority of objective functions for neural models can be divided into metric learning and statistical learning. Metric learning approaches require a pair mining strategy that often lacks efficiency, while statistical learning approaches are not generating highly compact features due to their indirect feature optimization. To this end, we propose a novel repeller-attractor loss that falls in the metric learning paradigm, yet directly optimizes for the L2 metric without the need of generating pairs. Our loss is formed of three components. One leading objective ensures that the learned features are attracted to each designated learnable class anchor. The second loss component regulates the anchors and forces them to be separable by a margin, while the third objective ensures that the anchors do not collapse to zero. Furthermore, we develop a more efficient two-stage retrieval system by harnessing the learned class anchors during the first stage of the retrieval process, eliminating the need of comparing the query with every image in the database. We establish a set of four datasets (CIFAR-100, Food-101, SVHN, and Tiny ImageNet) and evaluate the proposed objective in the context of few-shot and full-set training on the CBIR task, by using both convolutional and transformer architectures. Compared to existing objective functions, our empirical evidence shows that the proposed objective is generating superior and more consistent results.

Alexandru Ghita, Radu Tudor Ionescu

2023-06-01T12:53:10Z

http://arxiv.org/abs/2306.00630v2

# Class Anchor Margin Loss for Content-Based Image Retrieval ###### Abstract. The performance of neural networks in content-based image retrieval (CBIR) is highly influenced by the chosen loss (objective) function. The majority of objective functions for neural models can be divided into metric learning and statistical learning. Metric learning approaches require a pair mining strategy that often lacks efficiency, while statistical learning approaches are not generating highly compact features due to their indirect feature optimization. To this end, we propose a novel repeller-attractor loss that falls in the metric learning paradigm, yet directly optimizes for the \(L_{2}\) metric without the need of generating pairs. Our loss is formed of three components. One leading objective ensures that the learned features are attracted to each designated learnable class anchor. The second loss component regulates the anchors and forces them to be separable by a margin, while the third objective ensures that the anchors do not collapse to zero. Furthermore, we develop a more efficient two-stage retrieval system by harnessing the learned class anchors during the first stage of the retrieval process, eliminating the need of comparing the query with every image in the database. We establish a set of four datasets (CIFAR-100, Food-101, SVHN, and Tiny ImageNet) and evaluate the proposed objective in the context of few-shot and full-set training on the CBIR task, by using both convolutional and transformer architectures. Compared to existing objective functions, our empirical evidence shows that the proposed objective is generating superior and more consistent results. 2023 Contrastive learning, contrastive loss, class anchors, content-based image retrieval, object retrieval + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote † †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: leftmargin=*] + Footnote †: thanks: [leftmargin=*] + Footnote †: thanks: [leftmargin=*] with image embeddings assigned to the nearest class anchors. By harnessing the learned class anchors, our retrieval process becomes more efficient and effective than the brute-force search. We carry out experiments on four datasets (CIFAR-100 (CIFAR-100, 2009), Food-101 (Friedman et al., 2010), SVHN (Krizhevsky et al., 2012), and Tiny ImageNet (Vaswani et al., 2017)) to compare the proposed loss function against representative statistical and deep metric learning objectives. We evaluate the objectives in the context of few-shot and full-set training on the CBIR task, by using both convolutional and transformer architectures, such as residual networks (ResNets) (Krizhevsky et al., 2012) and shifted windows (Swin) transformers (Krizhevsky et al., 2012). Moreover, we test the proposed losses on various embedding space dimensions, ranging from 32 to 2048. Compared to existing loss functions, our empirical results show that the proposed objective is generating higher and more consistent performance levels across the considered evaluation scenarios. Furthermore, we conduct an ablation study to demonstrate the influence of each loss component on the overall performance. In summary, our contribution is threefold: * We introduce a novel repeller-attractor objective function that directly optimizes for the \(L_{2}\) metric, alleviating the need to generate pairs via hard example mining or alternative mining strategies. * We propose a two-stage retrieval system that leverages the use of the learned class anchors in the first stage of the retrieval process, leading to significant speed and performance gains. * We conduct comprehensive experiments to compare the proposed loss function with popular loss choices in multiple evaluation scenarios. ## 2. Related Work As related work, we refer to studies introducing new loss functions, which are related to our first contribution, and new content-based image retrieval methods, which are connected to our second contribution. **Loss functions.** The problem of generating features that are tightly clustered together for images representing the same objects and far apart for images representing distinct objects is a challenging task usually pursued in the context of CBIR. For retrieval systems based on neural networks, the choice of the objective function is the most important factor determining the geometry of the resulting feature space (Zhu et al., 2017). We hereby discuss related work proposing various loss functions aimed at generating effective embedding spaces. Metric learning objective functions directly optimize a desired metric and are usually based on pairs or tuples of known data samples (Beng et al., 2016; Liu et al., 2017; Wang et al., 2017; Wang et al., 2017). One of the earliest works on metric learning proposed the contrastive loss (Krizhevsky et al., 2012), which was introduced as a method to reduce the dimensionality of the input space, while preserving the separation of feature clusters. The idea behind contrastive loss is to generate an attractor-repeller system that is trained on positive and negative pairs generated from the available data. The repelling can happen only if the distance between a negative pair is smaller than a margin \(m\). In the context of face identification, another successful metric learning approach is triplet loss (Wang et al., 2017). It obtains the desired properties of the embedding space by generating triplets of anchor, positive and negative examples. Each image in a batch is considered an anchor, while positive and negative examples are selected online from the batch. For each triplet, the proposed objective enforces the distance between the anchor and the positive example to be larger than the distance between the anchor and the negative example, by a margin \(m\). Other approaches introduced objectives that directly optimize the AUC (Chen et al., 2018), recall (Liu et al., 2018) or AP (Krizhevsky et al., 2012). The main issues when optimizing with loss functions based on metric learning are the usually slow convergence (Wang et al., 2017) and the difficulty of generating useful example pairs or tuples (Wang et al., 2017). In contrast, our method does not require mining strategies and, as shown in the experiments, it converges much faster than competing losses. The usual example mining strategies are hard, semi-hard, and online negative mining (Wang et al., 2017; Wang et al., 2017). In hard negative mining, for each anchor image, we need to construct pairs with the farthest positive example and the closest negative example. This adds an extra computational step at the beginning of every training epoch, extending the training time. Similar problems arise in the context of semi-hard negative mining, while the difference consists in the mode in which the negatives are sampled. Instead of generating pairs with the farthest example, one can choose a negative example that is slightly closer than a positive sample, thus balancing hardness and stability. A more efficient approach is online negative mining, where negative examples are sampled during training in each batch. In this approach, the pair difficulty adjusts while the model is training, thus leading to a more efficient strategy, but the main disadvantage is that the resulting pairs may not be the most challenging for the model, due to the randomness of samples in a batch. Statistical learning objective functions indirectly optimize the learned features of the neural network. Popular objectives are based on some variation of the cross-entropy loss (Chen et al., 2018; Chen et al., 2018; Chen et al., 2018; Chen et al., 2018; Wang et al., 2017; Wang et al., 2017) or the cosine loss (Chen et al., 2018). By optimizing such functions, the model is forced to generate features that are close to the direction of the class center. For example, ArcFace (Chen et al., 2018) reduces the optimization space to an \(n\)-dimensional hypersphere by normalizing both the embedding generated by the encoder, and the corresponding class weight from the classification head, using their Euclidean norm. An additive penalty term is introduced, and with the proposed normalization, the optimization is performed on the angle between each feature vector and the corresponding class center. Hybrid objective functions promise to obtain better embeddings by minimizing a statistical objective function in conjunction with a metric learning objective (Krizhevsky et al., 2012; Krizhevsky et al., 2012). For example, Center Loss (Wang et al., 2017) minimizes the intra-class distances of the learned features by using cross-entropy in conjunction with an attractor for each sample to its corresponding class center. During training, the class centers are updated to become the mean of the feature vectors for every class seen in a batch. Another approach (Wang et al., 2017) similar to Center Loss (Wang et al., 2017) is based on predefined evenly distributed class centers. A more complex approach (Chen et al., 2018) is to combine the standard cross-entropy with a cosine classifier and a mean squared error regression term to jointly enhance global and local features. Both contrastive and triplet loss objectives suffer from the need of employing pair mining strategies, but in our case, mining strategies are not required. The positive pairs are built online for each batch, between each input feature vector and the dedicated class anchor, the number of positive pairs thus being equal to the number of examples. The negative pairs are constructed only between the class centers, thus alleviating the need of searching for a good negative mining strategy, while also significantly reducing the number of negative pairs. To the best of our knowledge, there are no alternative objective functions for CBIR that use dedicated self-repelling learnable class anchors acting as attraction poles for feature vectors belonging to the respective classes. **Content-based image retrieval methods.** CBIR systems are aimed at finding similar images with a given query image, matching the images based on the similarity of their scenes or the contained objects. Images are encoded using a descriptor (or encoder), and a system is used to sort a database of encoded images based on a similarity measure between queries and images. In the context of content-based image retrieval, there are two types of image descriptors. On the one hand, there are general descriptors (Kumar et al., 2017), where a whole image is represented as a feature vector. On the other hand, there are local descriptors (Kumar et al., 2017) where portions of an image are represented as feature vectors. Hybrid descriptors (Beng et al., 2017) are also used to combine both global and local features. To improve the quality of the results retrieved by learned global descriptors, an additional verification step is often employed. This step is meant to re-rank the retrieved images by a precise evaluation of each candidate image (Kumar et al., 2017). The re-ranking step is usually performed with the help of an additional system, and in some cases, it can be directly integrated with the general descriptor (Kumar et al., 2017). In the CBIR task, one can search for visually similar images as a whole, or search for images that contain similar regions (Kumar et al., 2017) of a query image. In this work, we focus on building global descriptors that match whole images. Further approaches based on metric learning, statistical learning, or hand-engineered features are discussed in the recent survey of Dubey et al. (Dubey et al., 2018). Different from other CBIR methods, we propose a novel two-stage retrieval system that leverages the use of the class anchors learned through the proposed loss function to make the retrieval process more efficient and effective. ## 3. Method **Overview.** Our objective function consists of three components, each playing a different role in obtaining a discriminative embedding space. All three loss components are formulated with respect to a set of learnable class anchors (centroids). The first loss component acts as an attraction force between each input embedding and its corresponding class anchor. Its main role is to draw the embeddings representing the same object to the corresponding centroid, thus creating embedding clusters of similar images. Each center can be seen as a magnet with a positive charge and its associated embeddings as magnets with negative charges, thus creating attraction forces between anchors and data samples of the same class. The second loss component acts as a repelling force between class anchors. In this case, the class centers can be seen as magnets with similar charges. If brought together, they will repel each other, and if they lie at a certain distance, the repelling force stops. The last component acts similarly to the second one, with the difference that an additional magnet is introduced and fixed at the origin of the embedding space. Its main effect is to push the class centers away from the origin. **Notations.** Let \(\mathbf{x}_{i}\in\mathbb{R}^{h\times w\times c}\) be an input image and \(y_{i}\in\mathbb{N}\) its associated class label, \(\forall i\in\{1,2,...,l\}\). We aim to optimize a neural encoder \(f_{\theta}\) which is parameterized by the learnable weights \(\theta\) to produce a discriminative embedding space. Let \(\mathbf{e}_{i}\in\mathbb{R}^{n}\) be the \(n\)-dimensional embedding vector of the input image \(\mathbf{x}_{i}\) generated by \(f_{\theta}\), i.e. \(\mathbf{e}_{i}=f_{\theta}(\mathbf{x}_{i})\). In order to employ our novel loss function, we need to introduce a set of learnable class anchors \(C=\{\mathbf{c}_{1},\mathbf{c}_{2},...,\mathbf{c}_{l}\}\), where \(\mathbf{c}_{j}\in\mathbb{R}^{n}\) resides in the embedding space of \(f_{\theta}\), and \(t\) is the total number of classes. **Loss components.** With the above notations, we can now formally define our first component, the attractor loss \(\mathcal{L}_{A}\), as follows: \[\mathcal{L}_{A}(\mathbf{x}_{i},C)=\frac{1}{2}\|\mathbf{e}_{i}-\mathbf{c}_{y_{i }}\|_{2}^{2}. \tag{1}\] The main goal of this component of the objective is to cluster feature vectors as close as possible to their designated class anchor by minimizing the distance between \(\mathbf{e}_{i}\) and the corresponding class anchor \(\mathbf{c}_{y_{i}}\). Its effect is to enforce low intra-class variance. However, obtaining low intra-class variance is only a part of what we aim to achieve. The first objective has little influence over the inter-class similarity, reducing it only indirectly. Therefore, another objective is required to repel samples from different classes. As such, we introduce the repeller loss \(\mathcal{L}_{R}\), which is defined as follows: \[\mathcal{L}_{R}(C)=\frac{1}{2}\sum_{y,y^{\prime}\in\mathcal{V},y^{\prime}y^{ \prime}}\left\{\max\left(0,2\cdot m-\|\mathbf{c}_{y}-\mathbf{c}_{y^{\prime}} \|\right)\right\}^{2}, \tag{2}\] where \(y\) and \(y^{\prime}\) are two distinct labels from the set of ground-truth labels \(Y\), and \(m>0\) is a margin representing the radius of an \(n\)-dimensional sphere around each anchor, in which no other anchor should lie. The goal of this component is to push anchors away from Figure 1. An example showing the behavior of the attractor-repeller loss components for three classes. The stars represent the class anchors \(C\). Faded circles around class anchors represent the sphere of radius \(m\) around each anchor. Solid circles represent feature vectors generated by the encoder \(f_{\theta}\). Dashed arrows between feature vectors and class anchors represent the attraction of the force generated by the attractor \(\mathcal{L}_{A}\). The solid red arrow between class anchors represents the repelling force generated by the repeller \(\mathcal{L}_{R}\). Best viewed in color. each other during training to ensure high inter-class distances. The margin \(m\) is used to limit the repelling force to an acceptable margin value. If we do not set a maximum margin, then the repelling force can push the anchors too far apart, and the encoder could struggle to learn features that satisfy the attractor loss defined in Eq. (1). A toy example of the attractor-repeller mechanism is depicted in Figure 1. Notice how the optimization based on the attractor-repeller objective tends to pull data samples from the same class together (due to the attractor), and push samples from different classes away (due to the repeller). However, when the training begins, all data samples start from a location close to the origin of the embedding space, essentially having a strong tendency to pull the class anchors to the origin. To ensure that the anchors do not collapse to the origin (as observed in some of our preliminary experiments), we introduce an additional objective that imposes a minimum norm on the class anchors. The minimum norm loss \(\mathcal{L}_{N}\) is defined as: \[\mathcal{L}_{N}(C)=\frac{1}{2}\sum_{y\in Y}\left\{\max\left(0,p-\left\|\mathbf{ c}_{y}\right\|\right)\right\}^{2}, \tag{3}\] where \(p\) is the minimum norm that each anchor must have. This objective contributes to our full loss function as long as at least one class anchor is within a distance of \(p\) from the origin. Figure 2 provides a visual interpretation of the effect induced by the minimum norm loss. Notice how the depicted class anchor is pushed away from the origin (due to the minimum norm loss), while the data samples belonging to the respective class move along with their anchor (due to the attractor loss). Assembling the three loss components presented above into a single objective leads to the proposed class anchor margin (CAM) loss \(\mathcal{L}_{\text{CAM}}\), which is formally defined as follows: \[\mathcal{L}_{\text{CAM}}(\mathbf{x},C)=\mathcal{L}_{A}(\mathbf{x},C)+\mathcal{ L}_{R}(C)+\mathcal{L}_{N}(C). \tag{4}\] Notice that only \(\mathcal{L}_{A}\) is directly influenced by the training examples, while \(\mathcal{L}_{R}\) and \(\mathcal{L}_{N}\) operate only on the class anchors. Hence, negative mining strategies are not required at all. **Gradients of the proposed loss.** We emphasize that the \(L_{2}\) norm is differentiable. Moreover, the updates for the weights \(\theta\) of the encoder \(f_{\theta}\) are provided only by the attractor loss \(\mathcal{L}_{A}\). Hence, the weight updates (gradients of \(\mathcal{L}_{\text{CAM}}\) with respect to \(\theta\)) for some data sample \(\mathbf{x}_{i}\) are computed as follows: \[\frac{\partial\mathcal{L}_{\text{CAM}}(\mathbf{x}_{i},C)}{\partial\theta}= \frac{\partial\mathcal{L}_{A}(\mathbf{x}_{i},C)}{\partial\theta}=\left(f_{ \theta}(\mathbf{x}_{i})-\mathbf{c}_{yi}\right)\cdot\frac{\partial f_{\theta} (\mathbf{x}_{i})^{T}}{\partial\theta}. \tag{5}\] The class anchors receive updates from all three loss components. For a class anchor \(\mathbf{c}_{y}\), the update received from the component \(\mathcal{L}_{A}\) of the joint objective is given by: \[\frac{\partial\mathcal{L}_{A}(\mathbf{x},C)}{\partial\mathbf{c}_{y}}=\mathbf{ c}_{y}-f_{\theta}(\mathbf{x}). \tag{6}\] The contribution of the repeller to a class anchor \(\mathbf{c}_{y}\) is null when the \(m\)-margin sphere of the respective anchor is not intersecting another class anchor sphere, i.e.: \[\frac{\partial\mathcal{L}_{R}(C)}{\partial C}=0, \tag{7}\] when \(2\cdot m-\left\|\mathbf{c}_{y}-\mathbf{c}_{y^{\prime}}\right\|>0\), \(\forall y,y^{\prime}\in Y,y\neq y^{\prime}\). To simply the notation, let \(d=\left\|\mathbf{c}_{y}-\mathbf{c}_{y^{\prime}}\right\|\) be the distance between \(\mathbf{c}_{y}\) and another class center \(\mathbf{c}_{y^{\prime}}\). The update for a class center \(\mathbf{c}_{y}\) of the repeller is given by: \[\frac{\partial\mathcal{L}_{R}(C)}{\partial\mathbf{c}_{y}}=\sum_{y,y^{\prime} \in Y,y\neq y^{\prime}}\delta(d<2\cdot m)\cdot\frac{-(2\cdot m-d)\cdot(\mathbf{ c}_{y}-\mathbf{c}_{y^{\prime}})}{d}, \tag{8}\] where \(\delta(*)=1\) when \(*\) is satisfied, and \(0\) otherwise. Similarly, the contribution of the minimum norm loss is given by: \[\frac{\partial\mathcal{L}_{N}(C)}{\partial\mathbf{c}_{y}}=\sum_{y\in Y}\delta \left(\left\|\mathbf{c}_{y}\right\|<p\right)\cdot\left(1-\frac{p}{\left\| \mathbf{c}_{y}\right\|}\right)\cdot\mathbf{c}_{y}. \tag{9}\] The final update of the learnable class centers \(C\) is given by summing the gradients given in Eq. (6), Eq. (8) and Eq. (9): \[\frac{\partial\mathcal{L}(\mathbf{x},C)}{\partial\mathbf{c}_{y}}=\frac{ \partial\mathcal{L}_{A}(\mathbf{x},C)}{\partial\mathbf{c}_{y}}+\frac{\partial \mathcal{L}_{R}(C)}{\partial\mathbf{c}_{y}}+\frac{\partial\mathcal{L}_{N}(C)}{ \partial\mathbf{c}_{y}}. \tag{10}\] **Fast retrieval via two-stage system.** During inference, we can employ the \(L_{2}\) measure between query and database embeddings to retrieve images similar to the query. Aside from the brute-force search that compares the query with every embedding in the database, we can harness the geometric properties of the resulting embedding space and use the class anchors \(C\) to improve the speed of the retrieval system. Instead of computing the \(L_{2}\) distances between a query feature vector and all the embedding vectors stored in the database, we propose to employ a two-stage system. In the first stage, distances are computed between the query feature vector and each class anchor. Upon finding the closest class anchor, in the second stage, distances are computed between the query feature vector and all the embeddings associated with the established class. Distances are then sorted and the closest corresponding items are retrieved. The main advantage of this approach is an improved retrieval time due to the reduced number of required comparisons. A possible disadvantage consists in retrieving wrong items when the retrieved class anchor is not representative for the query image. Figure 2. Contribution of the minimum norm loss \(\mathcal{L}_{N}\) imposed on the class anchors. The blue star represents a class anchor. Solid circles represent embedding vectors generated by the encoder \(f_{\theta}\). Dashed arrows represent the attraction force generated by the attractor \(\mathcal{L}_{A}\). The solid gray line represents the direction in which the anchor is pushed away from the origin due to the component \(\mathcal{L}_{N}\). Best viewed in color. We present results with our faster alternative in the comparative experiments. **Visualizing the embedding space.** To emphasize the geometric effect of the proposed objective, we have modified the final encoder layer of a ResNet-18 model to output 2D feature vectors without the final non-linearity. The model is trained from scratch on the SVHN dataset by alternatively using the cross-entropy loss, the contrastive loss, and the proposed class anchor margin loss. The resulting embeddings, together with the distribution of the embeddings for each of the models, are presented in Figure 3. For the proposed objective, we have set \(m=4\) and \(p=1\). From the distribution of the embeddings, it can be noted that the proposed objective generates tighter and clearly separable clusters. **Application to classification tasks.** An important remark is that we can use the class centers \(C\) to apply the proposed system to classification tasks. In this case, the predicted label \(\hat{y}\) for a test sample \(\mathbf{x}\) can be obtained as follows: \[\hat{y}=\arg\min_{j}\|\hat{f}_{\theta}(\mathbf{x})-\mathbf{c}_{j}\|. \tag{11}\] We perform additional experiments to demonstrate the application of our class anchor margin loss to classification tasks, with competitive results. ## 4. Experiments In the experiments, we compare our class anchor margin loss with the cross-entropy and the contrastive learning losses on four datasets, considering both convolutional and transformer models. ### Datasets We perform experiments on four datasets: CIFAR-100 (Krizhevsky et al., 2015), Food-101 (Krizhevsky et al., 2015), SVHN (Krizhevsky et al., 2015), and Tiny ImageNet (Vaswani et al., 2017). CIFAR-100 contains 50,000 training images and 10,000 test images belonging to 100 classes. Food-101 is composed of 101,000 images from 101 food categories. The official split has 750 training images and 250 test images per category. SVHN contains 73,257 digits for training, 26,032 digits for testing. Tiny ImageNet is a subset of ImageNet-1K, which contains 100,000 training images, 25,000 validation images and 25,000 test images from 200 classes. ### Experimental setup As underlying neural architectures, we employ three ResNet (He et al., 2016) model variations (ResNet-18, ResNet-50, ResNet-101) and a Swin transformer (Krizhevsky et al., 2015) model (Swin-T). We rely on the PyTorch (Krizhevsky et al., 2015) library together with Hydra (Krizhevsky et al., 2015) to implement and test the models. We apply random weight initialization for all models, except for the Swin transformer, which starts from the weights pre-trained on the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) dataset (Vaswani et al., 2017). We employ the Adam (Kingmare et al., 2014) optimizer to train all models, regardless of their architecture. For the residual neural models, we set the learning rate to \(10^{-3}\), while for the Swin transformer, we use a learning rate of \(10^{-4}\). The residual nets are trained from scratch for 100 epochs, while the Swin transformer is fine-tuned for 30 epochs. For the lighter models (ResNet-18 and ResNet-50), we use a mini-batch size of 512. Due to memory constraints, the mini-batch size is set to 64 for the Swin transformer, and 128 for ResNet-101. Residual models are trained with a linear learning rate decay with a factor of 0.5. We use a patience of 6 epochs for the full-set experiments, and a patience of 9 epochs for the few-shot experiments. Fine-tuning the Swin transformer does not employ a learning rate decay. Input images are normalized to have all pixel values in the range of \([0,1]\) by dividing the values by 256. The inputs of the Swin transformer are standardized with the image statistics from ILSVRC (Vaswani et al., 2017). We use several data augmentation techniques such as random crop with a padding of 4 pixels, random horizontal flip (except for SVHN, where flipping the digits horizontally does not make sense), color jitter, and random affine transformations (rotations, translations). Moreover, for the Swin transformer, we added the augmentations described in (Vaswani et al., 2017). For the models optimized either with the cross-entropy loss or our class anchor margin loss, the target metric is the validation Figure 3. Embedding vectors (left) and their distribution (right) for a ResNet-18 model modified to output 2D features and trained with \((a)\) cross-entropy loss, \((b)\) contrastive loss and \((c)\) class anchor margin loss (ours). The top row represents training embeddings from the SVHN dataset, while the bottom row shows test embeddings. Best viewed in color. accuracy. When using our loss, we set the parameter \(m\) for the component \(\mathcal{L}_{R}\) to 2, and the parameter \(p\) for the component \(\mathcal{L}_{N}\) to 1, across all datasets and models. Since the models based on the contrastive loss optimize in the feature space, we have used the 1-nearest neighbors (1-NN) accuracy, which is computed for the closest feature vector retrieved from the gallery. As evaluation measures for the retrieval experiments, we report the mean Average Precision (mAP) and the precision@\(k\) on the test data, where \(k\in\{20,100\}\) is the retrieval rank. For the classification experiments, we use the classification accuracy on the test set. We run each experiment in 5 trials and report the average score and the standard deviation. ### Retrieval results with full training data **Results with various architectures.** We first evaluate the performance of the ResNet-18, ResNet-50, ResNet-101 and Swin-T models on the CBIR task, while using the entire training set to optimize the respective models with three alternative losses, including our own. The results obtained on the CIFAR-100 (Krizhevsky et al., 2015), Food-101 (Krizhevsky et al., 2015), SVHN (Krizhevsky et al., 2015), and Tiny ImageNet (Vaswani et al., 2017) datasets are reported in Table 1. First, we observe that our loss function produces better results on the majority of datasets and models. Furthermore, as the rank \(k\) increases from 20 to 100, we notice that our loss produces more consistent results, essentially maintaining the performance level as \(k\) increases. \begin{table} \begin{tabular}{l l r r r r r r r r r r r} \hline \hline \multicolumn{1}{c}{\multirow{2}{*}{Loss}} & \multicolumn{3}{c}{CIFAR-100} & \multicolumn{3}{c}{Food-101} & \multicolumn{3}{c}{SVHN} & \multicolumn{3}{c}{Tiny ImageNet} \\ & & mAP & P@20 & P@100 & mAP & P@20 & P@100 & mAP & P@20 & P@100 & mAP & P@20 & P@100 \\ \hline \multirow{3}{*}{\begin{tabular}{l} \(\mathcal{L}_{N}\) \\ \end{tabular} } & CE & \(.249_{\pm 0.003}\) & \(.473_{\pm 0.005}\) & \(.396_{\pm 0.005}\) & \(.234_{\pm 0.002}\) & \(.547_{\pm 0.002}\) & \(.459_{\pm 0.001}\) & \(.895_{\pm 0.013}\) & \(.954_{\pm 0.003}\) & \(.952_{\pm 0.003}\) & \(.130_{\pm 0.001}\) & \(.340_{\pm 0.001}\) & \(.262_{\pm 0.001}\) \\ & CL & \(.220_{\pm 0.13}\) & \(.341_{\pm 0.10}\) & \(.303_{\pm 0.11}\) & \(.025_{\pm 0.001}\) & \(.042_{\pm 0.004}\) & \(.038_{\pm 0.003}\) & \(.116_{\pm 0.000}\) & \(.115_{\pm 0.002}\) & \(.116_{\pm 0.001}\) & \(.070_{\pm 0.12}\) & \(.139_{\pm 0.17}\) & \(.117_{\pm 0.16}\) \\ & CAM (ours) & \(.622_{\pm 0.005}\) & \(.560_{\pm 0.007}\) & \(.553_{\pm 0.007}\) & \(.751_{\pm 0.001}\) & \(.676_{\pm 0.003}\) & \(.669_{\pm 0.003}\) & \(.967_{\pm 0.001}\) & \(.954_{\pm 0.000}\) & \(.953_{\pm 0.001}\) & \(.508_{\pm 0.004}\) & \(.425_{\pm 0.004}\) & \(.418_{\pm 0.005}\) \\ \hline \multirow{3}{*}{\begin{tabular}{l} \(\mathcal{L}_{N}\) \\ \end{tabular} } & CE & \(.211_{\pm 0.006}\) & \(.454_{\pm 0.007}\) & \(.366_{\pm 0.006}\) & \(.158_{\pm 0.004}\) & \(.471_{\pm 0.005}\) & \(.370_{\pm 0.005}\) & \(.909_{\pm 0.010}\) & \(.958_{\pm 0.001}\) & \(.956_{\pm 0.002}\) & \(.088_{\pm 0.003}\) & \(.292_{\pm 0.006}\) & \(.209_{\pm 0.005}\) \\ & CL & \(.164_{\pm 0.16}\) & \(.271_{\pm 0.17}\) & \(.240_{\pm 0.16}\) & \(.019_{\pm 0.000}\) & \(.030_{\pm 0.000}\) & \(.028_{\pm 0.001}\) & \(.372_{\pm 0.310}\) & \(.447_{\pm 3.561}\) & \(.437_{\pm 3.564}\) & \(.005_{\pm 0.000}\) & \(.008_{\pm 0.001}\) & \(.007_{\pm 0.000}\) \\ & CAM (ours) & \(.640_{\pm 0.008}\) & \(.581_{\pm 0.009}\) & \(.578_{\pm 0.009}\) & \(.765_{\pm 0.005}\) & \(.697_{-0.008}\) & \(.697_{-0.009}\) & \(.969_{\pm 0.001}\) & \(.956_{\pm 0.001}\) & \(.957_{\pm 0.001}\) & \(.543_{\pm 0.003}\) & \(.472_{\pm 0.002}\) & \(.468_{\pm 0.002}\) \\ \hline \multirow{3}{*}{\begin{tabular}{l} \(\mathcal{L}_{N}\) \\ \end{tabular} } & CE & \(.236_{\pm 0.009}\) & \(.482_{\pm 0.008}\) & \(.397_{\pm 0.009}\) & \(.160_{\pm 0.003}\) & \(.479_{\pm 0.008}\) & \(.376_{\pm 0.007}\) & \(.936_{\pm 0.003}\) & \(.958_{\pm 0.001}\) & \(.957_{\pm 0.001}\) & \(.093_{\pm 0.002}\) & \(.299_{\pm 0.002}\) & \(.216_{\pm 0.002}\) \\ & CL & \(.034_{\pm 0.028}\) & \(.069_{\pm 0.051}\) & \(.056_{\pm 0.044}\) & \(.014_{\pm 0.002}\) & \(.018_{\pm 0.004}\) & \(.017_{\pm 0.003}\) & \(.595_{\pm 2.420}\) & \(.700_{\pm 2.95}\) & \(.696_{\pm 2.03}\) & \(.006_{\pm 0.001}\) & \(.007_{\pm 0.002}\) & \(.007_{\pm 0.002}\) \\ & CAM (ours) & \(.629_{\pm 0.006}\) & \(.575_{\pm 0.008}\) & \(.573_{\pm 0.008}\) & \(.758_{\pm 0.007}\) & \(.690_{\pm 0.007}\) & \(.693_{\pm 0.006}\) & \(.969_{\pm 0.001}\) & \(.956_{\pm 0.001}\) & \(.956_{\pm 0.001}\) & \(.529_{\pm 0.005}\) & \(.458_{\pm 0.006}\) & \(.455_{\pm 0.006}\) \\ \hline \multirow{3}{*}{ \begin{tabular}{l} \(\mathcal{L}_{N}\) \\ \end{tabular} } & CE & \(.661_{\pm 0.004}\) & \(.808_{\pm 0.001}\) & \(.770_{\pm 0.001}\) & \(.617_{\pm 0.006}\) & \(.817_{\pm 0.002}\) & \(.777_{\pm 0.003}\) & \(.963_{\pm 0.001}\) & \(.968_{\pm 0.000}\) & \(.968_{\pm 0.000}\) & \(.560_{\pm 0.006}\) & \(.743_{\pm 0.001}\) & \(.707_{\pm 0.003}\) \\ & CL & \(.490_{\pm 0.10}\) & \(.629_{\pm 0.007}\) & Results with various embedding dimensions.We train a ResNet-50 encoder from scratch, where the last encoder layer is modified to output embeddings of various dimensions from the set \(\{32,64,128,256,512,1024,2048\}\) on the CIFAR-100 (Krizhevsky et al., 2015) dataset. We illustrate the results with the cross-entropy, contrastive learning and class anchor margin losses in Figure 4. We observe that the performance is degrading as the embedding size gets larger for models trained with cross-entropy or contrastive losses. In contrast, the presented results show that the performance is consistent for the ResNet-50 based on our objective. This indicates that our model is more robust to variations of the embedding space dimension \(n\). Convergence speed.We have also monitored the convergence speed of ResNet-50 models, while alternatively optimizing with cross-entropy, contrastive learning and class anchor margin losses. In Figure 5, we show how the ResNet-50 models converge over 100 training epochs. The model trained with contrastive loss exhibits a very slow convergence speed. The model trained with cross-entropy achieves its maximum performance faster than the model trained with contrastive-loss, but the former model reaches a plateau after about 60 epochs. The model trained with our CAM loss converges at a faster pace compared with the other models. ### Few-shot retrieval results We next evaluate the neural models on the few-shot object retrieval task. For each dataset, we sample a certain number of training images from each class, starting from one example per class. We gradually increase the number of training samples, doubling the number of training examples per class in each experiment, until we reach the maximum amount of available images. In Figure 6, we present the corresponding results on CIFAR-100 (Krizhevsky et al., 2015), Food-101 (Friedman et al., 2015), SVHN (Krizhevsky et al., 2015), and Tiny ImageNet (Zhu et al., 2016). For all three ResNet models, our CAM loss leads to better results, once the number of training samples becomes higher or equal to 4 per class. In all cases, the contrastive loss obtains the lowest performance levels, being surpassed by both cross-entropy and class anchor margin losses. For the Swin-T model, cross-entropy usually leads to better results when the number of samples per class is in the lower range (below 64). As the number of training samples increases, our loss recovers the gap and even surpasses cross-entropy after a certain point (usually when the number of samples per class is higher than 128). In general, all neural models obtain increasingly better performance levels when more data samples are available for training. With some exceptions, the models trained with our CAM loss achieve the best performance. In summary, we conclude that the class anchor margin loss is suitable for few-shot retrieval, but the number of samples per class should be generally higher than 4 to obtain optimal performance. ### Qualitative results We choose the ResNet-50 model and inspect the retrieved images for each of the three loss functions. In Figure 7, we show a set of eight randomly sampled queries from the four datasets (CIFAR-100, Food-101, SVHN, and Tiny ImageNet). The model based on our loss seems to return more consistent results, with the majority of images representing the same object as the query. In contrast, models trained with the other losses can sometimes retrieve items that do Figure 6. Few-shot retrieval performance (mAP) of four models (ResNet-18, ResNet-50, ResNet-101 and Swin-T) on four datasets (CIFAR-100, Food-101, SVHN and Tiny ImageNet). On each dataset, the results are shown from one sample per class (one-shot learning) to the maximum number of samples per class, by doubling the number of training samples in each trial. not always belong to the query category. Overall, the qualitative results confirm the superiority of our class anchor margin loss. ### Ablation study **Ablating the loss.** In Table 3, we demonstrate the influence of each additional loss component on the overall performance of ResNet-18 on the SVHN dataset, by ablating the respective components from the proposed objective. We emphasize that the component \(\mathcal{L}_{A}\) is mandatory for our objective to work properly. Hence, we only ablate the other loss components, namely \(\mathcal{L}_{R}\) and \(\mathcal{L}_{N}\). In addition, we investigate different class center initialization heuristics. As such, we conduct experiments to compare the random initialization of class centers and the base vector initialization. The latter strategy is based on initializing class anchors as scaled base vectors, such that each class center has no intersection with any other class center in the \(n\)-dimensional sphere of radius \(m\), where \(n\) is the size of the embedding space. As observed in Table 3, the class center initialization has a major impact on the overall performance. For each conducted experiment, we notice a significant performance gain for the base vector initialization strategy. Regarding the loss components, we observe that removing both \(\mathcal{L}_{R}\) and \(\mathcal{L}_{N}\) from the objective leads to very low performance. Adding only the component \(\mathcal{L}_{N}\) influences only the overall accuracy, but the mAP is still low, since \(\mathcal{L}_{N}\) can only impact each anchor's position with respect to the origin. Adding only the component \(\mathcal{L}_{R}\) greatly improves the performance, proving that \(\mathcal{L}_{R}\) is crucial for learning the desired task. Using both \(\mathcal{L}_{R}\) and \(\mathcal{L}_{N}\) further improves the results, justifying the proposed design. **Ablating the two-stage retrieval system.** As discussed in Section 3, we employ a two-stage retrieval system to speed up the retrieval process. We hereby ablate the proposed two-stage approach that leverages the class anchors, essentially falling back to the brute-force retrieval process, in which the query is compared with every Figure 7. Top 6 retrieved items by a ResNet-50 model trained with one of three losses: cross-entropy (CE), contrastive (CL), and class anchor margin (CAM). We randomly selected two queries per dataset. Best viewed in color. embedding vector from the database. We compare the ablated (brutefore) retrieval with the two-stage retrieval in Table 2. Remarkably, we observe that our two-stage retrieval system not only improves the retrieval speed, but also the mAP. In terms of time, the two-stage retrieval improves the speed by a factor ranging between 2\(\times\) and 3\(\times\). In terms of performance, the gains are higher than 10% in 9 out of 16 cases. These results illustrate the benefits of our two-stage retrieval process based on leveraging the class anchors. ### Classification results As earlier mentioned, we can leverage the use of the learned class centers and the predictions generated with Eq. (11) to classify objects into the corresponding categories. We present the classification accuracy rates of the ResNet-18 and ResNet-50 models in Table 4, while alternating between the cross-entropy and the class anchor margin losses. Our CAM loss provides competitive results, surpassing cross-entropy in 6 out of 8 cases. This confirms that our loss is also suitable for the classification task, even though its performance gains are not as high as for the retrieval task. ## 5. Conclusion In this paper, we proposed (\(i\)) a novel loss function based on class anchors to optimize convolutional networks and transformers for object retrieval in images, as well as (\(ii\)) a two-stage retrieval system that leverages the learned class anchors to speed up and increase the performance of the system. We conducted comprehensive experiments using four neural models on four image datasets, demonstrating the benefits of our loss function against conventional losses based on statistical learning and contrastive learning. We also performed ablation studies to showcase the influence of the proposed components, empirically justifying our design choices. In future work, we aim to extend the applicability of our approach to other data types, beyond images. We also aim to explore new tasks and find out when our loss is likely to outperform the commonly used cross-entropy.

2304.01357

Excavation Problems in Elamite Mathematics

In this article, we study the problems found in the Susa Mathematical Texts No.\,24 and No.\,25 (\textbf{SMT No.\,24} and \textbf{SMT No.\,25}) which concern excavation projects such as canals and holes. We also examine certain Elamite structures, such as the canal systems serving Susa and a reservoir at the ziggurat of Chogha Zanbil, in whose construction geometry might well have played an important role.

Nasser Heydari, Kazuo Muroi

2023-02-09T17:58:09Z

http://arxiv.org/abs/2304.01357v1

# Excavation Problems in Elamite Mathematics ###### Abstract In this article, we study the problems found in the Susa Mathematical Texts No. 24 and No. 25 (**SMT No. 24** and **SMT No. 25**) which concern excavation projects such as canals and holes. We also examine certain Elamite structures, such as the canal systems serving Susa and a reservoir at the ziggurat of Chogha Zanbil, in whose construction geometry might well have played an important role. ## 1 Introduction **SMT No. 24** and **SMT No. 25** are two of the texts inscribed on 26 clay tablets excavated from Susa in southwest Iran by French archaeologists in 1933. The texts of all the Susa mathematical texts (**SMT**) along with their interpretations were first published in 1961 (see [1]). **SMT No. 24** and **SMT No. 25** are on a one-column clay tablet1. Following Bruins in [1], we treat these two texts separately. Footnote 1: The reader can see the new photos of this tablet on the website of the Louvre’s collection. Please see [https://collections.louvre.fr/en/ark:/53355/cl010186434](https://collections.louvre.fr/en/ark:/53355/cl010186434) for obverse and reverse. **SMT No. 24** contains two problems. The first, an excavation problem, which leads to a complicated quadratic equation, is found on the obverse of the tablet. The second problem, concerning the digging of a canal, is presented on the reverse. The text of **SMT No. 25**, also on the reverse of the tablet, is another belonging to the category of excavation problems. There is only one problem treated in this text, which concerns digging a canal. Although the entire problems are unavailable because of damage to the tablet, considerable understanding of the mathematical calculations utilized in solving these problems can be derived from a careful analysis of the text that remains. Excavation Problems in the SMT In Babylonian mathematics, there are many problems which concern digging a hole, a cistern, a canal or a foundation for a building or a wall. Such problems are often referred as _excavation problems_. Among the Babylonian mathematical texts, there are several which address various quantities concerning canals, although from the mathematical point of view they are relatively simple. By way of examples of such texts, we mention **YBC 4666**, **YBC 7164**, **YBC 9874**, **VAT 7528**, and **BM 85196** (please also see [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 111, 112, 113, 114, 115, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 140, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 212, 223, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 310, 320, 331, 334, 34, 351, 352, 353, 354, 355, 356, 357, 358, 359, 401, 411, 42, 436, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 72, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 111, 109, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 123, 124, 125, 126, 127, 128, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 151, 152, 153, 154, 156, 157, 158, 159, 161, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 201, 212, 213, 214, 215, 216, 217, 218, 219, 222, 223, 217, 219, 223, 219, 224, 225, 216, 218, 219, 226, 227, 228, 229, 230, 231, 232, 234, 235, 236, 237, 238, 239, 240, 241, 243, 244, 245, 246, 247, 249, 250, 252, 254, 256, 257, 258, 259, 260, 261, 262, 263, 265, 266, 267, 269, 270, 272, 274, 275, 276, 277, 278, 279, 281, 282, 283, 285, 286, 287, 289, 290, 292, 293, 294, 295, 296, 297, 298, 299, 301, 31, 320, 331, 334, 351, 352, 353, 354, 356, 357, 358, 359, 401, 41, 42, 436, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 73, 74, 75, 76, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 112, 109, 112, 113, 114, 115, 116, 117, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 151, 152, 153, 154, 156, 157, 159, 161, 179, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 209, 209, 210, 212, 223, 214, 215, 216, 217, 219, 220, 221, 223, 217, 228, 229, 230, 231, 233, 233, 234, 235, 236, 237, 238, 239, 240, 241, 258, 259, 261, 259, 270, 281, 293, 294, 295, 296, 297, 298, 299, 301, 301, 312, 301, 323, 302, 334, 351, 352, 353, 354, 356, 357, 358, 359, 401, 41, 42, 436, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 68, 69, 70, 73, 74, 75, 76, 78, 79, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 109, 112, 109, 112, 113, 114, 115, 116, 117, 119, 120, 121, 131 (L27) 34,41,15 nigni 20:3,[13,21,3]3,45 _ta-mar_ (L28) _a-na_ 20:3,13,21,33,[45] 1,5,55,4,41,15 dah (L29) 21:9,8,26,15 [_ta-mar_] _m[_i-_n_] \(a\) fb-si 35,37,30 fb-si (L30) 34,41,15 _ta-ki-i[_[_ta-_ta_]-ka [_a-na_ 3]5,37,30 dah (L31) 1,10:18,45 _ta-m[_ar_] _m[_i-na_ _a-na_ 1] 14:3,45 (L32) a-sa _sa-ar-ri_ gar _sa_ [1,10:1]8,45 _i-na-ad-di-na_ (L33) 5 gar 5 dagal an-ta 1/2 [5 _he-pe_ 2,30 _ta-mar_ 2,30 _a-na_ ] (L34) 30 dirig dah 3 _ta-mar_ 3 dagal ki-ta [igi-12 _sa_ dagal an-ta] (L35) ugu dagal ki-ta _i-te-ru le-q[_\(\acute{e}\)_] 10 _ta-mar_ [30 \(u\) 10 ul-gar] (L36) 40 _a-na_ 12 _su-up-li_ _i-st_ 8 _ta-mar_ [5 dagal an-ta] (L37) \(u\) 3 dagal ki-ta ul-gar 8 _ta-mar_ [1/2 8 _he-pe_] (L38) 4 _ta-mar_ 4 _a-na_ 8 _su-up-li_ _i-st_-[_ma_] (L39) 32 _ta-mar_igi-32 _pu-tu-ur_ 1,5[2,30 _ta-mar_ ] (L40) 1,52,30 _a-na_ 24 sahar _i-\(\acute{}\)_[_i_ 45 _ta-mar_ 45 u\(\acute{}\)s] Reverse, Lines 1-24 (L1) za-e(?) \(\cdots\) [\(\cdots\) \(\cdots\) \(\cdots\) ] (L2) _u-sa-pi-i[_] _i_(?)-[_na_(?) _k_]_a-la-ak-[_ki-im_ \(\cdots\) \(\cdots\) ] (L3) 2-kam _ta-a[_[_d_]-di-in_ 2_-tu_ [\(\cdots\) \(\cdots\) \(\cdots\) ] (L4) a-sa _ka-la-ak-ki_ gal ul-[gar \(\cdots\) \(\cdots\) \(\cdots\) ] (L5) _a-na_ tun _sa_ _a-ta-ap_ pa-[_n_]_a-_n_[_im_] da[h\(\cdots\) ] (L6) _a-na_ tun _a_s-[_lu-u_]_t_ (?) ul-gar sahar \(\cdots\) [\(\cdots\) \(\cdots\) 1,15] (L7) za-e 1,15 ul-gar _a-na_ 13 _sa_-[_la-\(\acute{}\)_s-ea-ra_-_t_] _i-st-ma_ 16,[15] (L8) 10 _sa ka-la-ak-ku_ ugu _ka-la-a[_kki _i-t_]_e-ru nigin 1,40 _ta-mar_ (L9) 1,40 _i-na_ 16,15 zi 16,[13,20] _ta-mar_ igi-10 dirig _pu-tu-ir_ (L10) 6 _ta-mar_ igi-12 _su-up-li pu-tu-ur_ 5 _ta-mar_ 5 _a-na_ 6 _i-st_ (L11) 30 _ta-mar_ 30 _ta-lu-ku_ 30 _ta-lu-ka a-na_ 16,13,20 _i-si-ma_ (L12) 8,6,40 _ta-mar_ 10 [dir]ig nigni 1,40 _ta-mar_ 1,40 _a-na_ 13 _sa-la-\(\acute{}\)_s-ea-ra-ti_ (L13) _i-st-ma_ 21,[40 _ta-mar_] 21,40 _i-na_ 8,6,40 zi (L14) 7,45 _ta-[_mar re-is-k_] _la-ki-ti_ 3[0 _ta-lu-k_]_a_ (L15) _a-na_ 13 [_sa-la-\(\acute{}\)_s-ea-ru-ti_] _i-si_ 6,30 _t_[_a-mar_ 30 _t_]_a-lu-ka_ (L16) _a-na ka-a-aia-ma_[_ni_] 2 tab-ba 1 _ta-mar_ 1 _a-na_ 6,30 dah (L17) [7],30 _ta-mar_ 1[3 _sa-l_]_a-as-\(\acute{}\)_s-\(\acute{}\)_e-ra-ti_ _a-na_ 3-su_ _a-na_ ka-aiaia-ma_-ni_ (L18) _a-li-ik-ma_ 3[9] _la-mar_ 7,30 _a-na_ 39 dah 46,30 _ta-mar_ (L19) _mi-na_ _a-na_ 46,30 gar _sa_ 7,45 _sa_ _r_[_e-is_]_-ka _u-ki-il-lu_ (L20) [_i-n_]_a-ad-di-na_ [10] gar _re-is-ka_ _li-ki-il_ 1/2 10 dirig _he-pe_ (L21) [5 _ta_]_-mar_ [5(?) gar(?)] 5 nigni 25 _ta-mar_ 25 _a-na_ 10 (L22) [_sa_ re-is-ka _u-ki-il-lu_] dah 10,25 _ta-mar_ mi-na_ ib-si (L23) [25 _ib-si_ 25(?) gar(?)] 5 _a-na_ 25 _is-te-en_ dah 30 _ta-mar_ (L24) [_i-na_ 25 2-kam zi 20 _t_]_a-mar_ 30 gal 20 tur **Translation** Obverse, Lines 1-40 (L1) The canal that I constructed in reed beds and woods, \(\cdots\)\(\cdots\). (L2) \(\cdots\) 0;30 of the excess, the lower breadth. One third of \(\cdots\), (L3) that I constructed, the depth. The volume 24,0(**volume-sar**) that I constructed \(\cdots\). (L4) What are \(\cdots\) and the depth? You, 1 breadth that you do not know \(\cdots\), (L5) you see 0;30. Put down 0;30 of the excess. 1 regular number \(\cdots\), (L6) 0;30 of the excess. Take one third of 0;30, (and) you see 0;10. \(\cdots\). (L7) Multiply (it by 12), and you see 2. Put down 2 as the width. 1, the upper breadth \(\cdots\), (L8) you see 1,30. Halve 1,30, (and) you see 45. 45 is the length. \(\cdots\). (L9) Return. Take one third of 0;30 of the excess, (and) you see 0;10. \(\cdots\). (L10) Multiply (it by 12), and you see 2. \(\cdots\), (L11) you see 15. 15, \(\cdots\). Put down the length. \(\cdots\). (L12) Let it hold your head. \(\cdots\). Make the reciprocal of 45 of the length, (and 0;1,20). (L13) Multiply (it) by \(\cdots\) of the area, and you see 9;22,30. Multiply \(\cdots\). (L14) \(\cdots\) 24,0(?), and 0;2 is the ratio (of the width to the length). (L15) \(\cdots\) \(\cdots\) width(?), subtract 0;2(?), and the area. (L16) \(\cdots\) \(\cdots\), and the length. 15, the one which is to be added to the length. (L17) \(\cdots\) \(\cdots\). Multiply 0;30 by 9;22,30, and (L18-19) you see 4;41,15. Return. Multiply 45 of the length by 0;2 of (the ratio of) the width (to the length), and you see 1;30. Multiply (it) by 9;22,30, and you see 14;3,45. (L20-21) 14;3,45 is the false area. Multiply 14;3,45 by 4;41,15 of the area, and you see 1,5;55,4,41,15. Let it hold your head. (L22) Multiply 15, the one which is to be added to the length, by 0;2 of (the ratio of) the width (to the length), and you see 0;30. (L23) Multiply 0;30 by 3, and you see 1;30. Subtract 0;30 from 1;30, (and) (L24) you see 1. Multiply 1 by 9;22,30, and you see 9;22,30. (L25) Since \(<1,0>\) as the length is said to you, add 1,0, the factor, to 9;22,30, (and) (L26) you see 1,9;22,30. Halve 1,9;22,30, (and) you see 34;41,15. (L27) Square 34;41,15, (and) you see 20,3;13,21,33,45. (L28) Add 1,5;55,4,41,15 to 20,3;13,21,33,45, (and) (L29) you see 21,9;8,26,15. What is the square root? 35;37,30 is the square root. (L30) Add 34;41,15 (which was used in) your completing the square to 35;37,30, (and) (L31-32) you see 1,10;18,45. What should I put down to 14;3,45, the false area, which will give me 1,10;18,45? (L33) Put down 5. 5 is the upper breadth. Halve 5, (and) you see 2;30. (L34) Add 2;30 to 0;30 of the excess, (and) you see 3. 3 is the lower breadth. * (L35) Take 1/12 of the amount by which the upper breadth exceeded the lower breadth, (and) you see 0;10. Add 0;30 and 0;10 together, (and the result is 0;40). * (L36) Multiply 0;40 by 12 of the (constant of the) depth, (and) you see 8. * (L37) Add together 5 of the upper breadth and 3 of the lower breadth, (and) you see 8. Halve 8, (and) * (L38) you see 4. Multiply 4 by 8, of the depth, and * (L39) you see 32. Make the reciprocal of 32, (and) you see 0;1,52,30. * (L40) Multiply 0;1,52,30 by 24,0 of the volume, (and) you see 45. 45 is the length. Reverse, Lines 1-24 * (L1) You(?),......... * (L2)... I excavated. In(?) the hole...... * (L3) you gave the second...... A second time...... * (L4) I added...... and the area of the large hole together,...... * (L5) I added...... to the depth of the former canal,...... * (L6) I cut off (?)... for the small. The sum of the volume and... is 1;15. * (L7) You, multiply 1;15 of the sum by 13 of one thirteenth, and (you see) 16;15. * (L8) Square 0;10 of the amount by which (the length of) the (large) hole exceeded (the length of) the (small) hole, (and) you see 0;1,40. * (L9) Subtract 0;1,40 from 16;15, (and) you see 16;13,20. Make the reciprocal of 0;10 of the excess, * (L10) (and) you see 6. Make the reciprocal of 12 of the depth, (and) you see 0;5. Multiply 0;5 by 6, * (L11) (and) you see 0;30. 0;30 is the product. Multiply 0;30 of the product by 16;13,20, and * (L12) you see 8;6,40. Square 0;10 of the excess, (and) you see 0:1,40. Multiply 0;1,40 by 13 of one thirteenth, * (L13) and you see 0;21,40. Subtract 0;21,40 from 8;6,40, (and) * (L14) you see 7;45. Let it hold your head. * (L15) Multiply 0;30 of the product by 13 of one thirteenth, (and) you see 6;30. * (L16) Multiply 0;30 of the product by regular (number) 2, (and) you see 1. Add 1 to 6;30, (and) * (L17) you see 7;30. Multiply 13 of the one thirteenth by 3, by regular (number three), * (L18) and you see 39. Add 7;30 to 39, (and) you see 46;30. * (L19) What should I put to 46;30 which gives me 7;45 that held your head? * (L20) Put down 0;10. Let it hold your head. Halve 0;10 of the excess, (and) * (L21) you see 0;5. Put down 0;5(?). Square 0;5, (and) you see 0;0,25. * (L22) Add 0;0,25 to 0;10 that held your head, (and) you see 0;10,25. What is the square root? * (L23) 0;25 is the square root. Put down 0;25(?). On the one hand add 0;5 to 0;25, (and) you see 0;30. * (L24) On the other hand subtract (0;5) from 0;25, (and) you see 0;20. 0;30 is the large, (and) 0;20 is the small. ### Mathematical Calculations The two problems in this text deal with constructing canals and computing their dimensions. The general shape of a canal is a prism with trapezoidal bases as shown in Figure 1 along with its reserved water. Denote the lower breadth (width), the upper breadth (width), the length and the depth (height) of the canal by \(v\), \(u\), \(x\) and \(z\) respectively. Also denote the height of the reserved water by \(z^{\prime}\). The area \(S\) of the trapezoidal base and the volume \(V\) of the trapezoidal canal are obtained by \[S=\frac{1}{2}z(u+v), \tag{1}\] and \[V=xS=\frac{1}{2}xz(u+v). \tag{2}\] Although formula (2) gives us the whole capacity of a canal, it is possible to compute the capacity of its reserved water \(V^{\prime}\) by using the constant of a canal given in **SMT No. 3**, line 33. In this line we read: 48 igi-gub _sa_ pa\({}_{5}\)-sig "0;48 is the constant of a small canal" This suggest that the ratio of the depth of canal to that of its reserved water is \(0;48=\frac{4}{5}\). In other words, we have \[\frac{z^{\prime}}{z}=\frac{4}{5}.\] Thus, \(z^{\prime}=\frac{4}{5}z\) and the volume of the reserved water should be \[V^{\prime}=\frac{4}{5}V.\] So it follows from (2) that \[V^{\prime}=\frac{2}{5}xz(u+v). \tag{3}\] Figure 1: The general shape of a canal and its dimensions #### First Problem Due to damage to the tablet, we are unable to establish with certainty the meanings of the technical expressions and calculations found in lines 1-25. We enumerate some of these ambiguities below: Line 18: "0;2 is the ratio \(\cdots\)" If we denote the width by \(y\), this probably means \(\frac{y}{x}=0;2\). In this case, according to lines 22-23, the value of \(y\) is obtained by \[y=(0;2)x=(0;2)\times 45=1;30\] assuming that \(x=45\). But this only adds to the confusion because similar terminologies (such as the upper breadth and the lower breadth) also occur in the text. Line 22: "15, the one which is to be added to the length" If we assume that \(x=45\), this probably concerns the calculation \[x+15=45+15=1,0\] whose result is mentioned in line 25: "Since 1,0 as the length is said to you". Lines 18-19:"\((0;30)\times(9;22,30)=4;41,45\)" The result of this multiplication is called "the area" in lines 20-21. Lines 20-21: "\((1;30)\times(9;22,30)=14;3,45\)" The result of this multiplication is called "the false area" in lines 20-21. Lines 22-24: "\(15\times(0;2)\times 3-0;30=(0;30)\times 3-0;30=1;30-0;30=1\)" The number 0;30 is called "the one which is subtracted from the width" in line 23. Line 24: "\(1\times(9;22,30)=9;22,30\)" We have been unable to reach a conclusion concerning this multiplication. Line 25: "\(1,0+9;22,30=1,9;22,30\)" We have also ben unable to reach a conclusion concerning this addition. Figure 2: Cross-section of a trapezoidal canal In spite of these uncertainties, the scribe of this text is able to compute both the dimensions of a canal in general and also solve a quadratic equation to find the upper breadth of the canal. We have shown the trapezoidal cross-section of a general trapezoidal canal in Figure 2. Note that the calculations in lines 34-40 show that the scribe is assuming the following relations between the dimensions of the canal: \[\begin{cases}v=\dfrac{u}{2}+0;30\\ z=12\left(0;30+\dfrac{1}{12}(u-v)\right).\end{cases} \tag{4}\] By looking carefully at the scribe's calculations in lines 26-33, it can be seen that he has solved the following quadratic equation: \[(14;3,45)u^{2}-(1,9;22,30)u=4;41,15. \tag{5}\] This equation (5) has been solved by the usual method of completing the squares-a standard method used by Babylonian and Elamite scribes to solve quadratic equations. This method was called _Takiltum_ in Babylonian texts (see [13, 14], for a discussion on this topic). Now, let us use this method and solve the quadratic equation (5) as follows: \[(14;3,45)u^{2}-(1,9;22,30)u=4;41,15\] \[\implies (14;3,45)\times(14;3,45)u^{2}-(14;3,45)\times(1,9;22,30)u=(14;3,45 )\times(4;41,15)\] \[\implies (14;3,45)^{2}u^{2}-(1,9;22,30)\times\Big{(}(14;3,45)u\Big{)}=1,5 ;55,4,41,15\] \[\implies \Big{(}(14;3,45)u\Big{)}^{2}-2\times(34;41,15)\times\Big{(}(14;3,4 5)u\Big{)}=1,5;55,4,41,15\] \[\implies \Big{(}(14;3,45)u\Big{)}^{2}-2\times(34;41,15)\times\Big{(}(14;3,4 5)u\Big{)}+(34;41,15)^{2}\] \[=1,5;55,4,41,15+(34;41,15)^{2}\] \[\implies \Big{(}(14;3,45)u-34;41,15\Big{)}^{2}=1,5;55,4,41,15+20,3;13,21,3,45\] \[\implies \Big{(}(14;3,45)u-34;41,15\Big{)}^{2}=21,9;8,26,15\] \[\implies (14;3,45)u-34;41,15=\sqrt{21,9;8,26,15}\] \[\implies (14;3,45)u-34;41,15=\sqrt{(35;37,30)^{2}}\] \[\implies (14;3,45)u-34;41,15=35;37,30\] \[\implies (14;3,45)u=35;37,30+34;41,15\] \[\implies (14;3,45)u=1,10;18,45\] \[\implies u=\dfrac{1}{(14;3,45)}\times(1,10;18,45)\] \[\implies u=(0;4,16)\times(1,10;18,45)\] \[\implies u=5.\] \[u=5. \tag{6}\] Now, according to lines 33-36, we can find the values of \(v\) and \(z\). On the one hand, by (4) and (6), we have \[v =\frac{u}{2}+0;30\] \[=\frac{5}{2}+0;30\] \[=2;30+0;30\] \[=3.\] Hence, \[v=3. \tag{7}\] On the other hand, it follows from (4) and (7) that \[z =12\left(0;30+\frac{1}{12}(u-v)\right)\] \[=12\left(0;30+\frac{1}{12}(5-3)\right)\] \[=12\left(0;30+\frac{2}{12}\right)\] \[=12\times(0;30+0;10)\] \[=12\times(0;40)\] which implies that \[z=8. \tag{8}\] Finally, a verification process seems to begin at line 37. First, the area of the trapezoidal cross-section \(S\) is obtained by using (1), (6), (7) and (8) as follows: \[S=\frac{z(u+v)}{2}=\frac{8(5+3)}{2}=4\times 8=32. \tag{9}\] Then, the length \(x\) of the canal is obtained by using the known volume \(V=24,0\) (mentioned in line 3). In fact, it follows from (2) and (9) that \[x =\frac{V}{S}\] \[=\frac{24,0}{32}\] \[=\frac{1}{32}\times(24,0)\] \[=(0;1,52,30)\times(24,0)\] \[=45.\] #### Second Problem In the lines regarding the second problem, we can recognize several typical expressions found in excavation problems as follows: (1) Reverse, line 2: "I excavated". (2) Reverse, line 4: "the area of the large hole". (3) Reverse, line 5: "the depth of the former canal". (4) Reverse, line 6: "I cut off \(\cdots\) for the small". (5) Reverse, line 8: "the amount by which (the length of) the (large) hole exceeded (the length of) the (small) hole". (6) Reverse, line 10: "the reciprocal of a number of 12 of the depth". Although we cannot restore the statement of this problem either, it seems that a canal has been enlarged, that is, the bottom of a canal has been deepened. Let \(x\), \(y\) and \(z\) denote the length, the width and the depth of a canal whose cross-section is rectangular (see Figure **3**). Judging from the calculations performed in the text, the equations dealt with in this problem are: \[\begin{cases}x-y=0;10\\ z=12(x-y)\\ z(x^{2}+y^{2})+xy(z+1)+\frac{1}{13}\Big{(}x^{2}+y^{2}\Big{)}=1;15.\end{cases} \tag{10}\] Before discussing the geometrical significance of these equations and the nature of the hole (_kalakkum_), we analyze the solution given in lines 7-24. According to line 7, we can multiply both sides of the third equation in (10) by 13. Since \(13\times(1;15)=16;15\), we get \[13z(x^{2}+y^{2})+13xy(z+1)+x^{2}+y^{2}=16;15. \tag{11}\] Figure 3: A canal with rectangular cross-section At this point, the scribe appears to have used the algebraic identity \[x^{2}+y^{2}=(x-y)^{2}+2xy. \tag{12}\] According to lines 8-9, since \(x-y=0;10\), we can use (11) and (12) to write \[13z(x^{2}+y^{2})+13xy(z+1)+x^{2}+y^{2}=16;15\] \[\implies 13z\Big{(}(x-y)^{2}+2xy\Big{)}+13xy(z+1)+(x-y)^{2}+2xy=16;15\] \[\implies 13z(x-y)^{2}+26xyz+13xyz+13xy+(x-y)^{2}+2xy=16;15\] \[\implies 13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy=16;15-(x-y)^{2}\] \[\implies 13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy=16;15-(0;10)^{2}\] \[\implies 13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy=16;15-0;1,40\] thus \[13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy=16;13,20. \tag{13}\] (Following the scribe's calculations, we did not substitute \(x-y=0;10\) in term \(13z(x-y)^{2}\) to simplify (13) more.) In lines 10-11, the scribe calculates the reciprocal of \(z\). It follows from the first two equations in (10) that \[z=12(x-y)\] \[\implies \frac{1}{z}=\frac{1}{12}\times\frac{1}{x-y}\] \[\implies \frac{1}{z}=(0;5)\times\frac{1}{(0;10)}\] which gives us \[\frac{1}{z}=0;30. \tag{14}\] Next, according to lines 11-20, we multiply both sides of (13) by \(1/z\) and then use (13) to find the value of \(xy\) as follows: \[13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy=16;13,20\] \[\implies \frac{1}{z}\times\Big{(}13z(x-y)^{2}+(3\times 13)xyz+(13+2)xy\Big{)}= \frac{1}{z}\times(16;13,20)\] \[\implies 13(x-y)^{2}+(3\times 13)xy+(13+2)\Big{(}\frac{1}{z}\times xy \Big{)}=\frac{1}{z}\times(16;13,20)\] \[\implies 13(x-y)^{2}+(3\times 13)xy+\Big{(}(13+2)\times(0;30)\,\Big{)}xy= (0;30)\times(16;13,20)\] \[\implies 13\times(0;10)^{2}+\Big{(}(3\times 13)+13\times(0;30)+2 \times(0;30)\,\Big{)}xy=8;6,40\] \[\implies 13\times(0;1,40)+(39+6;30+1)xy=8;6,40\] \[\implies 0;21,40+(46;30)xy=8;6,40\] \[\implies (46;30)xy=8;6,40-0;21,40\] \[\implies (46;30)xy=7;45\] \[\implies xy=\frac{1}{(46;30)}\times(7;45)\] \[\implies xy=\frac{1}{6\times(7;45)}\times(7;45)\] which implies that \[xy=0;10. \tag{15}\] The last part of the text proceeds with the common Babylonian method (completing the square). According to lines 20-21, by (10) and (15), we can write: \[\frac{x+y}{2} =\sqrt{\left(\frac{x-y}{2}\right)^{2}+xy}\] \[=\sqrt{\left(\frac{0;10}{2}\right)^{2}+0;10}\] \[=\sqrt{\left(0;5\right)^{2}+0;10}\] \[=\sqrt{0;0,25+0;10}\] \[=\sqrt{0;10,25}\] \[=\sqrt{(0;25)^{2}}\] So \[\frac{x+y}{2}=0;25. \tag{16}\] Now, we are in the position to compute the values of \(x\) and \(y\). According to lines 23-24, it follows from (10) and (16) that \[x =\frac{x+y}{2}+\frac{x-y}{2}\] \[=0;25+0;5\] \[=0;30\] and \[y =\frac{x+y}{2}-\frac{x-y}{2}\] \[=0;25-0;5\] \[=0;20.\] Therefore, the required values of the length and the width are \[x=0;30\quad\text{and}\quad y=0;20.\] Although we can now understand the mathematical meaning of the second problem, difficult questions still remain: What are really \(x\) and \(y\) in this problem?; and What is the relation between "a canal" (_atappum_) and "a hole" (_kalaktum_)? To answer these questions, we may need to consider a hypothetical situation where two canals intersect as shown in Figure 4. In this situation, an old canal of depth \(z\) and width \(x\) is joined by a new canal of depth \(z\) and width \(y\) at right angles. The intersection between the old and new canals is a rectangle with sides \(x\) and \(y\) and deepened by \(1\) kus, probably to allow the deposit of silt2. Footnote 2: In modern Japan, people sometimes see the same device for irrigation canals. The top view of this intersection is pictured with related dimensions in Figure 5. The rectangle in the east-west direction is the old canal and the one in the north-south direction is the new canal. It is clear from the figure that the junction of the two canals is a cube of length \(x\), width \(y\) and depth \(1+z\). According to Figure 4: Intersection of two canals holes adjoining to this intersection which we have named the large hole and the small hole. Note that the large hole (_kalakki_ gal) has dimensions \(x\), \(x\), \(z\) and the small hole (_kalakki_ tur) has dimensions \(y\), \(y\), \(z\). The unknown quantities asked for in this problem are the length \(x\) of the large hole and the length \(y\) of the small hole. One might expect here that \(x\) is called, for example, us_kalakki_ gal "the length of the large hole" instead of _kalakki_ gal "(the length) of the large hole". Since \(x\) is originally the width of the old canal, the Susa scribe might have omitted us "the length" in order to avoid confusion. In fact, neither us "the length" nor sag "the width" occurs in the text at all. Now, let us return to the third equation in (10). We interpret the first and the second terms of the left-hand side, i.e., \(z(x^{2}+y^{2})=zx^{2}+zy^{2}\) and \(xy(z+1)\), as the sum of the volumes of the large hole, the small hole and the junction of the two canals respectively. However, as to the third term \(\frac{1}{13}(x^{2}+y^{2})\), we are of the view that it does not have any geometrical meaning and is added to these volumes to make the equation more complicated. There are examples of similar equations in the so-called _series texts_, that is, the ancient drill books in mathematics. For example, the following system of equations is given in the text of **VAT 75373**: Footnote 3: Tablet **VAT 7537** is a Babylonian mathematical text in the Berlin Museum which originally was published by Neugebauer in [18]. For more information about this tablet and its text, see [19, 20]. \[\begin{cases}xy=10,0\\ x^{2}+\frac{1}{11}\bigg{(}2\left(\frac{1}{7}\Big{(}(2y+3(x-y)\Big{)}\right)^{2}+ x^{2}\bigg{)}=16,40.\end{cases}\] ### SMT No. 25 As mentioned earlier, the problem in **SMT No. 25** concerns the dimensions of a canal. Figure 5: Top view of two intersecting canals and their junction **Transliteration** Reverse, Lines 25-34 (L25) [\(\cdots\)\(\cdots\)\(\cdots\)\(me\)]-\(e\)_sa_giis-g[i\(\cdots\)] 2 _sussar_ (SARxDIS) _sa_pa\({}_{5}\) GIG(?) (L26) [\(\cdots\)\(\cdots\)] \(\cdots\) [\(\cdots\)] u s 5-ta-am _sa_6 _s_[_a_ sar] \(\cdots\)_-ma_ (L27) i[gi-5] _pu-tu-ur_12 _ta-mar_1_2 _a-na_6 _sa_ sar _i-si-ma_ (L28) 1,12 _ta-mar_1 sar \(u\)12 _su-si mu-tu a-na_40 erin-hi-a [gar] (L29) igi-40 erin _pu-tu-ur_1,30 _ta-mar_1,30 _a-na_1,12 _i-st-ma_ (L30) 1,48 _ta-mar mu-u Sa_ erin 1-k[am us] 1 nindan 30 dagal (L31) [1],48 _mu-u_su-up-lu _mi-nu_igi-48 igi-guib pa\({}_{5}\)_pu-tu-\([\)ur_] (L32) [1,15 _t_]_a-mar_1,15 _a-na_1,48 _me-e Sa_ [erin 1-kam _i-si-ma_] (L33) [2,15 _t_]_a-mar_igi-30 dagal _pu-tu-ur_2 _ta-mar_ (L34) [2,15 _a-n_]_a_2 _i-st-ma_4,30 _ta-mar_4,30 _su-up-lu_ **Translation** Reverse, Lines 25-34 (L25) \(\cdots\)\(\cdots\) water of reed bed \(\cdots\). 2,0 saros (of water) of a canal. \(\cdots\). (L26) \(\cdots\)\(\cdots\) the length 5 (nindan) each that of 6 saros \(\cdots\)\(\cdots\). (L27) Make the reciprocal of 5, (and) you see 0;12. Multiply 0;12 by 6 saros, and (L28) you see 1,12,0. 1 saros and 12 sixties is the (volume of) water. Put (it) down for 40,0 workers. (L29) Make the reciprocal of 40,0 of the workers, (and) you see 0;0,1,30. Multiply 0;0,1,30 by 1,12,0, and (L30) you see 1;48. (This is the volume of) the water (per 1 nindan in length) of the first(sic) worker. (If) the length is 1 nindan, the breadth is 0;30 (nindan), (and) (L31) the (volume of) water is 1;48, what is the depth? Make the reciprocal of 0;48, the constant of a canal, (and) (L32) you see 1;15. Multiply 1;15 by 1;48 of the water of a worker, and (L33) you see 2;15. Make the reciprocal of 0;30 of the breadth, (and) you see 2. (L34) Multiply 2;15 by 2, and you see 4;30. 4;30 (kus) is the depth. **Mathematical Calculations** The statement of the problem is almost entirely broken, and we have been unable to locate a similar problem to assist us in restoring it. However, the text calculates the depth of a canal using several conditions which may well have been provided in the statement. Judging from the calculations performed in the text, the missing conditions might have been as follows: **Problem:** Two thousand four hundred workers (40,0 erin-hi-a) built a canal whose width is 0;30 nindan, and whose reserved water is 6 sar4 (that is, 6,0,0 volume-sar). The workers were each assigned to dig a part of the canal 5 **nindan** in length. What is the depth of the canal? Consider a canal of length \(x\), width \(y\), and depth \(z\) and denote the depth of its reserved water by \(z^{\prime}\) as is shown in Figure 6. As we said before, we must have \[\frac{z^{\prime}}{z}=0;48. \tag{17}\] Let \(V\) be the volume of the canal, \(V^{\prime}\) the volume of its reserved water, \(S\) the area of its cross-section and \(S^{\prime}\) the area of a part of the cross-section submerged in water. It is clear from the figure that \[\begin{cases}V=xyz\\ V^{\prime}=xyz^{\prime}\\ S=yz\\ S^{\prime}=yz^{\prime}.\end{cases} \tag{18}\] Note that it follows from (17) and (18) that \[\frac{V^{\prime}}{V}=\frac{S^{\prime}}{S}=0;48. \tag{19}\] Figure 6: A canal and its reserved water Figure 7: The cross-section of a canal and the level of its reserved water In Figure 7, we have shown the cross-section of the canal and its reserved water. The height of canal is \(z^{\prime}\) and that of reserved water is \(z\). In lines 26-30, the scribe has calculated the volume of the water per 1 **nindan** and per one worker, say \(V_{1}^{\prime}\). To do so, he first obtains the the volume \(V_{0}^{\prime}\) of water per 1 **nindan** and all workers as follows: \[V_{0}^{\prime} =\frac{V^{\prime}}{5}\] \[=\frac{1}{5}\times(6\ \mbox{\sout{s}ar})\] \[=(0;12)\times(6,0,0)\] \[=1,12,0\] \[=1\ \mbox{\sout{s}ar}\ 12\ \mbox{\sout{s}isi}.\] (Note that 1 **sisi** is equal to \(1,0=60\).) Then, he divides this volume \(V_{0}^{\prime}=1,12,0\) by the number of all workers to find \(V_{1}^{\prime}\), which is also called "the water of the first worker" in line 30: \[V_{1}^{\prime} =\frac{V_{0}^{\prime}}{(40,0)}\] \[=\frac{1}{(40,0)}\times(1,12,0)\] \[=(0;0,1,30)\times(1,12,0)\] \[=1;48.\] Note that the value of \(V_{1}^{\prime}=1;48\) is the volume of the reserved water of the part of canal with length 1. That is, \(V_{1}^{\prime}\) is obtained by assuming \(x=1\) in the second equation of (18): \[V_{1}^{\prime}=1yz^{\prime}=yz^{\prime}=S^{\prime}.\] This means \(V_{1}^{\prime}\) is equal to the area of the cross-section of the lower part of the canal submerged in water: \[S^{\prime}=1;48. \tag{20}\] From (18) and (20), (according to lines 31-33) we can obtain \(S\): \[S =\frac{S^{\prime}}{(0;48)}\] \[=\frac{1}{(0;48)}\times(1;48)\] \[=(1;15)\times(1;48)\] which implies that \[S=2;15. \tag{21}\] Since by the assumption \(y=0;30\), we can (according to lines 33-34) use (18) and (21) to find the depth of the canal, i.e., \(z\), as follows: \[z=\frac{S}{y}=\frac{1}{(0;30)}\times(2;15)=2\times(2;15)=4;30.\] Therefore the depth of the canal is \[z=4;30. \tag{22}\] Note that the depth of water can be easily computed by using (17) and (22) as follows: \[z^{\prime}=(0;48)\times(4;30)=3;36.\] ## 3 Applications of Mathematics in Ancient Elam In this section, we consider the possible practical applications of mathematics both in construction projects and in Elamite architecture. The reader will note that the mathematical skills demonstrated in **SMT No. 24** and **SMT No. 25** may well have been of considerable assistance to those charged with the construction of Elamite infrastructure and other substantial buildings at the behest of the Elamite rulers of the time. ### Canals As for any ancient civilization, water played a decisive role in where the people of ancient Elam decided to settle. The Elamite people were among the very first farmers in the Ancient Near East (see [1]), utilizing the nearby rivers to irrigate their lands and farms. For example, the ancient capital city of Susa was founded in a region watered by at least two rivers5 : the Karkheh and the Dez (see Figure 8). Footnote 5: Professor Daniel Potts has made a strong case that there was only one river at Susa and that the modern Shavur river was the ancient course of the Karkheh river. See [12] for more details. Figure 8: A satellite view of modern Susa (Google Map) Although the course of the Karkheh in ancient Elam has been the subject of considerable scholarly debate, it has been suggested that since the Karkheh river is at a higher level than the fields between the Karkheh and the Dez, the people of Susa may have used the Karkheh's water to irrigate the land around the city or save it for future purposes (probably for drinking or making mud bricks). According to some research (see [11, 12]), there were at least four agricultural sections near ancient Susa which contained more than 40 small irrigation canals. Based on the analysis of other legal, economic and administrative clay tablets excavated from Susa, scholars have been able to determine the names and locations of many of these canals in each of these agricultural sections. These tablets date to the Sukkalmah Dynasty 1900-1500 BC, as do the **SMT**. We have shown three of these sections and the number of canals they contained in Figure 9. The location of the fourth section, described as being "on the other bank of the _Zamun_", cannot be established as it is not known which river was at that time known as the _Zamun_. In addition to this network of irrigation canals around ancient Susa, there was also the ancient Harmushi canal system, which consisted of a group of connected canals watering the vast area of lands between the Karkheh and the Dez river. Archaeological evidence show that some branches of this network date to the Middle Elamite period (see [1, 21, 22]). The main branch of this canal system was called the _Harmushi_ and connected to the Karkheh river at the point called _Pay-e Pol_ (literally, "at the foot of the bridge"), where there are still the ruins of an ancient dam which regulated the river into different branches. The length of this canal is thought to have been approximately \(50km\) in which case it could have irrigated nearly \(200km^{2}\) of the land between the Karkheh and the Dez rivers. Figure 10 shows a hypothetical course of the Harmushi canal network which has been adopted from Adams [1, Fig. 4]. The Harmushi system had several subbranches which is thought to provide water for the livestock of nomadic tribes. Many western travelers and scholars, who visited Kluzistan during last two centuries, mentioned the Harmushi canal in their reports and asserted that it was the Figure 9: Possible distribution of irrigation canals near Susa (adapted from [11]) main source for irrigating a large area around Susa (see [14, 15, 16, 17]). Although this ancient canal system had been irrigating the Susa area for more than 3000 years, it was finally replaced by the modern canal network after the construction of a concrete dam built on the Dez river in 1963. To understand the historical significance of this canal and its importance to the local population, even though there is almost no physical trace of the original system, the name of Harmushi still lives on in the surname of many people whose ancestors lived alongside it for centuries.6 Footnote 6: The complete name of the first author used to be “Nasser Heydari Harmushi”. In addition to canals, there are archaeological sites near modern Susa containing the remains of other water structures one of which is located in the Elamite holy city of Chogha Zanbil7. There is a hydraulic complex consisting of canals and a water reservoir of length \(10.70m\), width \(7.25m\) and depth \(4.35m\) whose capacity is about \(332m^{3}\) (see Figure 11). Footnote 7: The holy city of Chogha Zanbil (also known as _Dur-Untash_, or City of Untash, in Elam) located at \(38km\) to the southeast of Susa was founded by the Elamite king _Untash-Napirisha_ (1275–1240 BCE) to be the religious center of Elam. The principal element of this complex is an enormous ziggurat dedicated to the Elamite divinities Inshushinak and Napirisha. For more information, see [https://whc.unesco.org/en/list/113/](https://whc.unesco.org/en/list/113/) or [http://tchoghazanbil.com/](http://tchoghazanbil.com/). Although as yet there is no satisfactory explanation for how the hydraulic mechanism of this sophisticated piece of engineering worked, scholars have suggested different hypotheses. Some scholars including Ghirshman believed that this reservoir was fed Figure 10: The hypothetical course of the Harmushi canal network by the Karkheh river via the Harmushi canal, which was specially built to supply the new holy city (see [14, 15]). Drinking water was then thought to be distributed throughout the city by a system of smaller canals. While Ghirshman's opinion prevailed for decades, modern archaeological research has raised serious doubts about his analysis (see [16]). Mofidi-Nasrabadi in [17] has suggested that this water reservoir is actually a part of a drainage system devised by the Elamite engineers to remove floodwater from the holy city during the rainy season. In either case, the mathematical knowledge of Elamite scribes might have been of assistance in the construction of such a complex structure. Their mathematical skills could have been applied to estimate the amount of time and number of workers needed for such a large engineering project. ## 4 Conclusion The interpretation of problems contained in **SMT No. 24** and **SMT No. 25** taken together with the archaeological research confirming the existence of a substantial canal network at the time the problems were inscribed on the tablet, strongly infer that the Susa scribes were interested in the practical application of mathematics to the challenges faced by those living alongside them. These texts suggest that not only was mathematics taught in a "Susa School of Mathematics", but also that the scribes used their mathematical skills to address issues arising in the design and construction of structures, such as canals, which facilitated both the agricultural and economic development of ancient Elam. Figure 11: A water reservoir at the holy city of Chogha Zanbil (Credit: The World Heritage of Chogha Zanbil)

2304.11144

Multifractal Properties of Tribonacci Chains

We introduce two 1D tight-binding models based on the Tribonacci substitution, the hopping and on-site Tribonacci chains, which generalize the Fibonacci chain. For both hopping and on-site models, a perturbative real-space renormalization procedure is developed. We show that the two models are equivalent at the fixed point of the renormalization group flow, and that the renormalization procedure naturally gives the Local Resonator Modes. Additionally, the Rauzy fractal, inherent to the Tribonacci substitution, is shown to serve as the analog of conumbering for the Tribonacci chain. The renormalization procedure is used to repeatedly subdivide the Rauzy fractal into copies of itself, which can be used to describe the eigenstates in terms of Local Resonator Modes. Finally, the multifractal dimensions of the energy spectrum and eigenstates of the hopping Tribonacci chain are computed, from which it can be concluded that the Tribonacci chains are critical.

Julius Krebbekx, Anouar Moustaj, Karma Dajani, Cristiane Morais Smith

2023-04-21T17:40:46Z

http://arxiv.org/abs/2304.11144v2

# Multifractal Properties of Tribonacci Chains ###### Abstract We introduce two 1D tight-binding models based on the Tribonacci substitution, the hopping and on-site Tribonacci chains, which generalize the Fibonacci chain. For both hopping and on-site models, a perturbative real-space renormalization procedure is developed. We show that the two models are equivalent at the fixed point of the renormalization group flow, and that the renormalization procedure naturally gives the Local Resonator Modes. Additionally, the Rauzy fractal, inherent to the Tribonacci substitution, is shown to serve as the analog of conumbering for the Tribonacci chain. The renormalization procedure is used to repeatedly subdivide the Rauzy fractal into copies of itself, which can be used to describe the eigenstates in terms of Local Resonator Modes. Finally, the multifractal dimensions of the energy spectrum and eigenstates of the hopping Tribonacci chain are computed, from which it can be concluded that the Tribonacci chains are critical. Aperiodic; Quasicrystal; Multifractal Spectrum; Tribonacci; Rauzy Fractal pacs: 03.65.-b, 03.65.-b, 03.65.Jb ## I Introduction The description of electrons in solids using Bloch's theorem has allowed for a profound understanding of the electronic band structure of regular crystalline materials [1]. The discovery of quasicrystals [2], aperiodic structures that break translational symmetry, has pushed the field forward. The Penrose tilings [3] or the aperiodic mono-tile discovered recently by Smith et al. [4] are some of the typical examples that have fascinated physicists and mathematicians for years. Quasi-crystalline lattices have been also experimentally realized using different quantum-simulator platforms, such as ultracold atoms [5] or photonics [6]. The advent of topological insulators has reiterated the importance of periodicity in solids because translation invariance is at the core of the topological classification of these materials [7; 8; 9]. It remains an open question how the notion of topology translates to aperiodic structures such as quasicrystals, where translation invariance is often replaced by scale invariance [10]. The topological aspects of quasicrystals have been recently investigated [11; 12; 13; 14; 10], but methods are often tailored to each model, and a general framework to study topology in these systems is lacking. Arguably the most investigated quasicrystal is the Fibonacci chain [15], a one-dimensional model based on the Su-Schrieffer-Heeger (SSH) model [16]. The latter is a tight-binding model in which alternating weak and strong hopping parameters lead to a topological or trivial phase, depending on whether the last bond in the chain corresponds to a weak or strong hopping, respectively. The Fibonacci chain is a natural extension of the SSH model to the aperiodic domain [17], in which the weak and strong hopping parameters are distributed according to a Fibonacci word. This 1D tight-binding chain hosts many interesting properties, such as a multifractal energy spectrum and eigenstates [18; 19; 20]. In addition, it was shown to be equivalent to the Harper model [21], from which it inherits its topological properties. In particular, a description of the system in terms of conumbers [22] has revealed hidden symmetries in Hilbert space and allowed for a systematic prediction of the influence of random disorder based on a renormalization group (RG) scheme [17]. The interpretation of the system in terms of local symmetries has also led to a more profound understanding of its physical properties [23]. In this paper, we go beyond the realm of dimerized models, such as the SSH and Fibonacci chain, and introduce a quantum chain based on the Tribonacci substitution. Two tight-binding chains, the hopping Tribonacci Chain (HTC) and the on-site Tribonacci Chain (OTC), are defined analogously to the Fibonacci chain. These chains are closely linked to the Rauzy fractal, a well-known compact domain with fractal boundary [24]. An RG scheme for the HTC and OTC is developed along the lines proposed by Niu and Nori [17]. This allows for the same interpretation of the spectrum as a multifractal set as for the Fibonacci chain [18]. The RG scheme is also used to render the HTC and OTC equivalent at the RG fixed point. We show how the Rauzy fractal orders the lattice points according to their local environment, in analogy with the conumbering scheme. Furthermore, the RG procedure provides a natural way to enumerate all structures in the Local Resonator Mode (LRM) framework [23]. Finally, we compute the multifractal dimensions of the energy spectrum and eigenstates of the HTC, and compare them with the Fibonacci chain. From these results, it can be concluded that the Tribonacci chains are critical in terms of Anderson localization. The paper is structured as follows. In section II we introduce the HTC, the OTC, and all elements that are needed to define the model, such as the Tribonacci word and the Rauzy fractal. Section III is devoted to the RG scheme for the HTC and OTC, and how the two models can be considered equivalent in the infinite RG limit. In Section IV, the Rauzy fractal is proposed as the analog of conumbering for the HTC and OTC. Multifractal prop erties of the spectrum and wavefunction of the HTC are computed in Section V, and compared to the Fibonacci chain. Finally, the conclusion and outlook are presented in Section VI. ## II The model In this section, we introduce the elements needed to define the Tribonacci chain. The main element is the Tribonacci word, which determines the quasiperiodic modulation in the tight-binding chains. ### The Tribonacci Word #### ii.1.1 The Tribonacci Sequence Analogous to the Fibonacci sequence, one can define the Tribonacci sequence recursively as \[T_{N+1}=T_{N}+T_{N-1}+T_{N-2}, \tag{1}\] with initial values \(T_{-2}=0,T_{-1}=T_{0}=1\). The Tribonacci constant \(\beta\), the analog of the golden ratio, is obtained as the limit \[\beta=\lim_{N\to\infty}\frac{T_{N+1}}{T_{N}}\approx 1.8392\ldots, \tag{2}\] which is also the unique real root of the polynomial \[P(x)=x^{3}-x^{2}-x-1. \tag{3}\] The other two roots \(\omega,\bar{\omega}\) are complex and satisfy \(|\omega|<1\). #### ii.1.2 The Tribonacci Substitution The Tribonacci substitution is the substitution \(\rho\) on the alphabet \(\mathcal{A}=\{0,1,2\}\) that reads \[\rho:\begin{cases}0\mapsto 01,\\ 1\mapsto 02,\\ 2\mapsto 0.\end{cases} \tag{4}\] The Tribonacci word is obtained by repeatedly applying \(\rho\) to the seed \(W_{0}=0\). The resulting word after \(N\) applications \(W_{N}:=\rho^{N}(W_{0})\) is called the \(N\)th Tribonacci approximant. The Tribonacci word is the limit \(W:=\lim_{N\to\infty}W_{N}\). The first few approximants read \[W_{0} =0,\] \[W_{1} =01,\] \[W_{2} =0102,\] \[W_{3} =0102010,\] \[W_{4} =0102010010201.\] An alternative way to generate the Tribonacci word is by concatenating the previous three approximants \[W_{N+1}=W_{N}W_{N-1}W_{N-2}, \tag{5}\] which is reminiscent of Eq. (1). Therefore, the Tribonacci constant is equivalently obtained by the limit \[\beta=\lim_{N\to\infty}\frac{|W_{N+1}|}{|W_{N}|},\] where \(|\cdot|\) denotes the length of the word. Another important tool when dealing with any substitution \(\rho\) is the incidence matrix \(\mathbf{M}=[m_{ij}]\), where \(m_{ij}=|\rho(j)|_{i}\) and \(|w|_{k}\) denotes the number of occurrences of the letter \(k\) in the word \(w\). The incidence matrix is used in the relation \[\mathbf{N}^{(N+1)}=\mathbf{M}\cdot\mathbf{N}^{(N)},\] where \(\mathbf{N}^{(N)}:=(|W_{N}|_{0},|W_{N}|_{1},|W_{N}|_{2})^{T}\) is the vector that counts how often each letter occurs in the approximant \(W_{N}\). If \(\mathbf{M}\) has precisely one eigenvalue \(\lambda\) with \(|\lambda|>1\) and all other eigenvalues have modulus strictly less than 1, the substitution is called Pisot. The incidence matrix for the Tribonacci substitution and its characteristic polynomial read \[\mathbf{M}=\begin{pmatrix}1&1&1\\ 1&0&0\\ 0&1&0\end{pmatrix},\qquad\det(\lambda\mathbf{I}-\mathbf{M})=\lambda^{3}- \lambda^{2}-\lambda-1, \tag{6}\] which is identical to the Tribonacci polynomial Eq. (3). Hence, it is immediate that the Tribonacci substitution is Pisot. The eigenvalues are \(\lambda=\beta>1\) and \(\lambda=\omega,\bar{\omega}\) where \(|\omega|<1\). One can also define the bi-infinite Tribonacci word \(W|W\) in a consistent way (see Ch. 4 of Ref. [25]). Take the seed \(\rho^{-1}(0)|0=2|0\) and apply \(\sigma=\rho^{3}\) infinitely often to the seed. This results in the approximants \(W_{3N-1}|W_{3N}\) and the limit \[W|W:=\lim_{N\to\infty}W_{3N-1}|W_{3N}=\cdots w_{-2}w_{-1}|w_{0}w_{1}\cdots. \tag{7}\] #### ii.1.3 The Rauzy Fractal In 1982, Gerard Rauzy used the Tribonacci substitution to define a 2D compact domain with fractal boundary, called the Rauzy fractal [24] (see Fig. 1). The Rauzy fractal is obtained as a subset of \(\mathbb{C}\) via the valuation map. Let \([W]_{m}\) denote the first \(m\) letters of the Tribonacci word and take the left eigenvector \(\mathbf{v}=(v_{0},v_{1},v_{2})\) of \(\mathbf{M}\) in Eq. (6), corresponding to the eigenvalue \(\omega\). Then, the \(m\)th point in the Rauzy fractal is given by \[z_{m}=E([W]_{m})=\sum_{i\in\{0,1,2\}}|[W]_{m}|_{i}v_{i}\in\mathbb{C}, \tag{8}\] where \(E\) is the valuation map and \(m\geq 0\). Enumerating the letters of \(W=w_{0}w_{1}w_{2}\cdots\), each point can be assigned a color defined by the \(w_{m}\in\{0,1,2\}\), the \((m+1)\)th letter [26]. The Rauzy fractal is the compact set \(\mathfrak{R}=\overline{\{z_{m}\mid m\geq 0\}}\). Another way to obtain Fig. 1 is by starting at the origin in \(\mathbb{R}^{3}\), and for each letter in \(W\), taking a unit step in the \(x,y\) or \(z\)-direction if the letter is \(0\), \(1\) or \(2\), respectively [27]. This will create a staircase in \(\mathbb{R}^{3}\) in the direction of \(\mathbf{v}_{\beta}=(\beta^{2},\beta,1)^{T}\), which spans the expanding eigenspace \(L\) of Eq. (6). Denote \(\pi_{\text{int}}\) the projection along \(\mathbf{v}_{\beta}\) onto the 2D contracting eigenspace. Then, the \(m\)th point for \(m\geq 0\) is given by \[\mathbf{x}_{m}=\pi_{\text{int}}\sum_{i=0}^{m-1}\mathbf{e}_{w_{i}}\in\mathbb{R }^{2}, \tag{9}\] where \(\mathbf{e}_{i}\) are the canonical basis vectors of \(\mathbb{R}^{3}\). The Rauzy fractal is the compact set \(\mathfrak{R}^{\prime}=\overline{\{\mathbf{x}_{m}\mid m\geq 0\}}\), which is not precisely \(\mathfrak{R}\), but related by an affine transformation (see Appendix A for details of this transformation). #### ii.1.4 Cut-and-Project Sets Bearing the Rauzy fractal \(\mathfrak{R}^{\prime}\) in mind, one can view the Tribonacci word as a quasicrystal. Consider again the Tribonacci staircase, which are the points \(\mathbf{y}_{m}=\sum_{i-0}^{m-1}\mathbf{e}_{w_{i}}\). Using the bi-infinite word \(W|W\), the staircase can also be defined for \(m<0\) by \(\mathbf{y}_{m}=\sum_{i=m}^{-1}\mathbf{e}_{w_{i}}\). From the bi-infinite staircase, one can construct a 1D Tribonacci quasicrystal \[\Lambda_{\text{trib}}=\{\pi\mathbf{y}_{m}\mid m\in\mathbb{Z}\}, \tag{10}\] by projecting all staircase points along the stable eigenspace onto the line spanned by \(\mathbf{v}_{\beta}\), where this projection is denoted \(\pi\). One can see that \(\Lambda_{\text{trib}}\) is a cut-and-project set in the following sense. Take a cubic lattice in \(\mathbb{R}^{3}\) and trace out a volume by sliding the set \(\mathfrak{R}^{\prime}\), the acceptance set of the cut-and-project scheme, along the space \(L\). Note that all lattice points lying in the traced out volume are exactly the staircase points \(\mathbf{y}_{m}\), which constitute \(\Lambda_{\text{trib}}\) upon projecting onto \(L\). A key result is that any cut-and-project set has a point diffraction pattern [25], which leads to the conclusion that the aperiodic lattice \(\Lambda_{\text{trib}}\) is a quasicrystal. Finally, we would like to point out that there exists a quasiperiodic 2D tiling, the Rauzy tiling, which is based on the Tribonacci numbers and is cut-and-project set from a 3D space [28]. Several physical properties of tight-binding models on these lattices have been studied [29; 30], in particular the effect of a magnetic field [11; 12; 31]. The generalized Rauzy tiling extends this construction to arbitrary dimension, and this family of tilings can be viewed as a generalization of the Fibonacci chain [28]. #### ii.1.5 Recurrence Properties Another key property of the Tribonacci \(W\) word is its self-similarity [32]. Take any finite string \(s=s_{1}\cdots s_{N}\) of length \(N\) that occurs somewhere in \(W\). We say that \(s\) occurs at position \(i\) in \(W\) if \(s_{1}\cdots s_{N}=w_{i}\cdots w_{i+N}\). Let \(i_{1},i_{2},\dots\) denote the places where \(s\) occurs in \(W\). Then, the words \(r_{j}=w_{i_{j}}\cdots w_{i_{j+1}}\) between occurrences of \(s\) have useful properties. Firstly, for any choice \(s\), the word \(r_{j}\in\{r^{(0)},r^{(1)},r^{(2)}\}\) takes one of three values. Secondly, if we label \(r^{(i)}\) such that \(r^{(0)}\) occurs most often, \(r^{(1)}\) second most often and \(r^{(2)}\) least often, then the map \(\kappa:r^{i}\mapsto i\) maps the string \(r_{1}r_{2}\cdots\) back to \(W\). In other words \[\kappa(r_{1})\kappa(r_{2})\cdots=W, \tag{11}\] where \(r_{i}\) are the words between subsequent occurrences of \(s\) in \(W\). This also works if \(s\) occurs in a Tribonacci approximant \(W_{N}\). By applying periodic boundary conditions when determining \(r_{j}\), the map \(\kappa\) results in \[\kappa(r_{1})\kappa(r_{2})\cdots=W_{N-k}, \tag{12}\] where \(k\) depends on the choice of \(s\). Eqs. (11) and (12) are the foundation of the perturbative RG scheme in Section III. We would like to emphasise that there are other quantum chains based on three-letter substitutions that are Pisot with the same dominant \(\lambda=\beta\). One such example is the system studied by Ali et al. [33]. This is fundamentally different from our work, since in their case there is not a natural RG scheme and our connection to the Rauzy fractal is entirely new. Figure 1: (Color online) A Rauzy fractal with \(T_{14}=5768\) points. Each region corresponds to a symbol: red (0), green (1) and blue (2). ### Tribonacci Tight-Binding Models The definition of the Tribonacci chain, with aperiodic hopping and on-site energy, generalizes the work by Niu and Nori [17] on the Fibonacci chain to the HTC and OTC. #### ii.2.1 Hopping Model The infinite HTC is defined as a 1D tight-binding chain with no on-site potentials and hopping parameters that are modulated according to the Tribonacci word \(W|W\). The Hamiltonian reads \[H=\sum_{n\in\mathbb{Z}}t_{w_{n}}\left|n+1\right\rangle\left\langle n\right|+H.c., \tag{13}\] where \(w_{n}\in\{0,1,2\}\) are the letters of \(W|W\) in Eq. (7) and the model is parameterized by one parameter \(\rho\in[0,1]\) as \(t_{0}/t_{1}=t_{1}/t_{2}=\rho\). Note that Eq. (13) possesses chiral symmetry \(\Gamma H\Gamma=-H\), where \(\Gamma^{2}=1\) and \[\Gamma=\sum_{n\in Z}\left|2n\right\rangle\left\langle 2n\right|-\sum_{n\in Z} \left|2n+1\right\rangle\left\langle 2n+1\right|.\] A direct consequence of chiral symmetry is a symmetric spectrum around \(E=0\). The model is studied in the regime where \(\rho\ll 1\), i.e. \(0<t_{0}\ll t_{1}\ll t_{2}\), such that there is a hierarchy of bond strengths, analogous to Ref. [17]. #### ii.2.2 On-Site Model The OTC is defined by the Hamiltonian \[H=\sum_{n\in\mathbb{Z}}\epsilon_{w_{n}}\left|n\right\rangle\left\langle n \right|-t\sum_{n\in\mathbb{Z}}\left|n+1\right\rangle\left\langle n\right|+H.c., \tag{14}\] where now the hopping parameters \(t\) are constant, and the on-site potential \(\epsilon_{i}\) is modulated according to the Tribonacci word \(W|W\). This model is generally parameterized by two parameters, \(c_{1}=(\epsilon_{1}-\epsilon_{0})/t\) and \(c_{2}=(\epsilon_{2}-\epsilon_{0})/t\). Analogous to Ref. [17], we demand \(|c_{1}|,|c_{2}|,|c_{2}-c_{1}|\gg 1\), which physically means that the on-site potentials dominate and are weakly coupled. One particular choice is \(c_{1}=c_{2}/2=c\gg 1\), which will be used when comparing to the HTC. ## III Perturbative Renormalization of the Tribonacci Chain We now present the perturbative RG scheme for the HTC and OTC. The scheme is possible due to the self-similar recurrence properties of the Tribonacci word (see Section II.1.5), and is analogous to the RG of the Fibonacci chain proposed by Niu and Nori [17] (see the review by Jagannathan [15] for more details on the Fibonacci chain). ### The Renormalization Scheme #### iii.1.1 Hopping Model For the RG scheme, it is convenient to consider the \(N\)th HTC approximant \[H_{N}=\sum_{n=0}^{T_{N}-1}t_{w_{n}}\left|n+1\bmod T_{N}\right\rangle\left \langle n\right|+H.c., \tag{15}\] where periodic boundary conditions are enforced. Furthermore, the Hamiltonian is split up in two parts \[H_{N}=H_{0,N}+H_{1,N}, \tag{16}\] where \(H_{1,N}\) contains only the terms with a \(t_{0}\) hopping, such that \(H_{0,N}\) can be regarded as the unperturbed Hamiltonian. Note that \(H_{0,N}\) has only five highly degenerate energy levels \(E=0,\pm t_{1},\pm t_{2}\). The \(E=0\) states are the atoms, which are isolated sites, corresponding to \(00\) in \(W\). Type-1 molecules are the \(E=\pm t_{1}\) states, corresponding to \(010\) in \(W\). These are isolated dimers consisting of two neighboring sites, coupled by a \(t_{1}\) bond, which can either bond or anti-bond. Similarly, the \(E=\pm t_{2}\) states correspond to \(020\) in \(W\), and are called type-2 molecules. Upon setting \(t_{0}\) nonzero, the atoms/molecules start to interact. If one considers one type of atom or molecule as a lattice site, one can compute the effective coupling between subsequent sites using Brillouin-Wigner perturbation theory. Fig. 2 depicts the spectrum of Eq. (15), where one can see five branches around \(E=0,\pm t_{1},\pm t_{2}\) that would become fully degenerate upon setting \(t_{0}=0\). Now, we explain the simplest case, the type-1 molecule, in detail. The procedure for the other bonds is exactly the same, but with longer computations. Consider the Figure 2: (Color online) The energy spectrum of the HTC Eq. (15) with \(\rho=0.2\) and \(T_{13}=3136\) sites. The five main bands are located around \(E=0,\pm t_{1},\pm t_{2}\). The inset shows a zoom in on the top band, which exhibits a similar spectrum, but with seemingly different \(t_{0},t_{1}\). Tribonacci approximant \[\begin{array}{l}W_{6}=\\ 01\underline{020}100102010101020100102010201001020100102010010201010102010 2010.\end{array} \tag{17}\] The first step is to tabulate all words \(r_{i}\) occurring between \(1\)'s in \(W_{6}\), starting after the first occurrence of \(1\), and considering periodic boundary conditions. The possibilities are \(020,00\) and \(0\), which occur \(7,4\) and \(2\) times, respectively. Therefore \[\begin{array}{l}\{r\}=\{r_{1}=r^{(0)},r_{2}=r^{(1)},r_{3}=r^{(0)},\ldots,r_{ 13}=r^{(1)}\},\\ r^{(0)}=020,r^{(1)}=00,r^{(2)}=0.\end{array} \tag{18}\] Finally, upon applying the map \(\kappa:r^{(i)}\mapsto i\), the Tribonacci approximant \(W_{4}\) is obtained as \[W_{4}=\kappa(r_{1})\kappa(r_{2})\cdots\kappa(r_{13})=0102010010201, \tag{19}\] which has \(k=2\) in Eq. (12). The procedure in Eqs. (17), (18), and (19), which is the \(s=1\) case, can be carried out for any \(s\). This is done for \(s=0,1,2,00\) in Table 1. The procedure outlined in Eqs. (17), (18) and (19) is applied to the HTC Hamiltonian in Eq. (15) as follows. Consider the approximant in Fig. 3. Each dimer of two sites coupled by \(t_{1}\) is considered a lattice site in the renormalized chain, on which a (anti-)bonding state \(\left|\pm\right\rangle_{i}\) can sit. Using perturbation theory, the effective coupling between neighboring sites \[t_{i}^{\prime}=\left\langle\pm\right|_{i}H_{1,N}\left|\pm\right\rangle_{i+1},\] is computed. The perturbation theory framework is explained in Appendix B, for the chain shown in Fig. 3. The computations and results for the hopping and on-site chain are presented in Appendix B.1 and B.2, respectively. The main result of the RG scheme can now be stated. Denote \(H_{N}^{(p,q)}\) given by Eq. (15) with \(t_{0}/t_{1}=\rho^{p},t_{1}/t_{2}=\rho^{q}\), where \(p,q\in[0,\infty)\). Setting \(t_{0}=0\), the HTC Hamiltonian \(H_{0,N}\) has \(T_{N-3},T_{N-2},T_{N-4},T_{N-2}\) and \(T_{N-3}\) states with \(E=-t_{2},E=-t_{1},E=0,E=t_{1}\) and \(E=t_{2}\), respectively. To each of these five energies, we associate an atomic (\(s=00\)), bonding or anti-bonding chain (\(s=1,2\)). The result of the perturbative calculation (see Appendix B.1) is \[H_{N}^{(p,q)}\approx(z_{2}H_{N-3}^{(p+q,p+2q)}-t_{2})\oplus(z_{1}H_{N-2}^{(q,p )}-t_{1})\oplus(z_{0}H_{N-4}^{(p,2p+q)})\oplus(z_{1}H_{N-2}^{(q,p)}+t_{1}) \oplus(z_{2}H_{N-3}^{(p+q,p+2q)}+t_{2}), \tag{20}\] where the parameters read \(z_{0}=\rho^{4p+2q},z_{1}=\rho^{p+q}/2\) and \(z_{2}=\rho^{2p+3q}/2\). The computation of the parameters \(z_{i}\) and the \(p,q\) exponents in each of the five blocks is identical to Ref. [17], and is repeated in detail in Appendix B.1. The HTC in Eq. (15) realizes the case \(p=q=1\). From the result Eq. (20), it is clear that the spectrum consists not simply of scaled and shifted versions of itself, but rather related spectra of chains with various \(p,q\) values. Since one can identify each of the five quasibands in Fig. 2 to a block in Eq. (20), the spectrum can be interpreted as a multifractal set as Zheng [18] did for the Fibonacci chain. The words \(r^{(i)}\) in Table 1 are longer than those in the RG scheme for the Fibonacci chain, which requires higher orders of perturbation theory to yield a nonzero result. This has the advantage that the error made in the approximate RG Eq. (20) is smaller than the RG scheme for the Fibonacci chain. Figure 3: The 3rd approximant HTC, with Hamiltonian \(H_{3}\) (see Eq. (15)) and periodic boundary conditions. The single line denotes a \(t_{0}\) bond, the double line a \(t_{1}\) bond and a triple line a \(t_{2}\) bond. The chain is renormalized by considering the type-1 molecules as the new lattice sites, and the chain between these molecules as the new bonds, which are \(t_{0}^{\prime},t_{1}^{\prime}\). The figure is inspired by Ref. [17]. \begin{table} \begin{tabular}{||l|c c c c||} \hline \(s\) & \(r^{(0)}\) & \(r^{(1)}\) & \(r^{(2)}\) & maps \(W_{N}\) to \\ \hline \hline 0 & 1 & 2 & \(\emptyset\) & \(W_{N-1}\) \\ \hline 1 & 020 & 00 & 0 & \(W_{N-2}\) \\ \hline 2 & 010010 & 01010 & 010 & \(W_{N-3}\) \\ \hline 00 & 10201010201 & 102010201 & 10201 & \(W_{N-4}\) \\ \hline \end{tabular} \end{table} Table 1: For \(W_{N}\) and particular strings \(s=0,1,2,00\), the occurrences between \(s\) can be one of \(r^{(i)}\), and map to \(W_{N-k}\) under the map \(\kappa:r^{(i)}\mapsto i\), \(i=0,1,2\). On-Site Model The \(N\)th approximant of the OTC is defined as \[H_{N}^{o}=\sum_{n=0}^{T_{N}-1}\epsilon_{w_{n}}\ket{n}\bra{n}-t\big{(}\ket{n+1\bmod T _{N}}\bra{n}+H.c.\big{)}, \tag{21}\] where periodic boundary conditions are enforced. When writing \[H_{N}^{o}=H_{0,N}^{o}+H_{1,N}^{o}, \tag{22}\] the part \(H_{1,N}^{o}\) consists of all \(t\) bonds and \(H_{0,N}^{o}\) only the on-site energies. At \(t=0\), the chain consists of \(T_{N-1},T_{N-2}\) and \(T_{N-3}\) isolated sites with energy \(E=0,\epsilon_{1},\epsilon_{2}\), respectively. When \(t\) is nonzero, the degeneracy is lifted and the spectrum consists of three bands, as depicted in Fig. 4. The analysis in Section III.1.1 can be immediately carried over to the three atomic chains of the on-site model, to approximate each of the three bands as a general HTC \(H_{N-k}^{(p,q)}\). For a general OTC parameterized by \(c_{1}\) and \(c_{2}\), the result of the perturbation theory (see Appendix B.1) reads \[H_{N}^{o}\approx(z_{0}H_{N-1}^{(p_{0},q_{0})}+\epsilon_{0})\oplus(z_{1}H_{N-2 }^{(p_{1},q_{1})}+\epsilon_{1})\oplus(z_{2}H_{N-3}^{(p_{2},q_{2})}+\epsilon_{2}), \tag{23}\] where \(z_{0}=t,z_{1}=t/c_{1},z_{2}=t/[c_{2}^{2}(c_{2}-c_{1})]\) and \(p_{i}=\log a_{i}/\log\rho,q_{i}=\log b_{i}/\log\rho\) for \(i=0,1,2\) where \(a_{i},b_{i}\) are given in Table 3 in the appendix. As a final remark, by the recurrence property Eq. (11) of the infinite word \(W\), the approximations Eqs. (20) and (23) are also valid in the infinite limit, where the subscripts \(N,N-k\) are dropped. ### Hopping and On-Site Equivalence The RG scheme of the HTC Eq. (20) can be repeatedly applied to the five HTC Hamiltonians in the direct sum. For the OTC, the same is true after one application of the RG Eq. (23). Considering the infinite HTC and OTC, the Hamiltonian after \(m\) applications of the RG is described by \(5^{m}\) and \(3\cdot 5^{m-1}\) pairs of \(p,q\) values, respectively. We will show that the HTC and OTC are equivalent, in the sense that for both models, the fraction of \(p,q\) values that escape to infinity tends to one as \(m\rightarrow\infty\). For the HTC, define \(I_{m}=\{p_{i},q_{i}\mid i=1,\ldots,5^{m}\}\), the set of \(p,q\) values in the direct sum after \(m\) RG applications. Define the probability measure on the measurable space \((I_{m},2^{I_{m}})\) as \[\mu_{m}(A):=|A|/|I_{m}|, \tag{24}\] where \(2^{I_{m}}\) denotes the powerset, \(A\subset I_{m}\) and \(|\cdot|\) denotes the cardinality of the set. To study the divergence of \(p,q\) values, define the set of values smaller than \(m\) as \[J_{m}:=\{x\in I_{m}\mid x\leq m\}. \tag{25}\] For the OTC, all objects \(I_{m}^{o},J_{m}^{o},\mu_{m}^{o}\) are similarly defined. The mathematical statement of the equivalence, as proven in Appendix C, reads \[\lim_{m\rightarrow\infty}\mu_{m}(J_{m})=\lim_{m\rightarrow\infty}\mu_{m}^{o} (J_{m}^{o})=0. \tag{26}\] This proves that for both the HTC and OTC, the set of \(p\) and \(q\) values that remain finite can be at most a set of measure zero. This means that both the HTC and OTC are described by an infinite direct sum of \(H^{(p,q)}\) Hamiltonians with \(p=q=\infty\), which are Tribonacci chains where only the \(t_{2}\) bonds are nonzero. The similarity discussed in this work is a different notion of similarity than Niu and Nori [17] proved for the Fibonacci chain. In their case, all values would read \(p=1\) and \(q=1\) in Eqs. (20) and (23). Since the Fibonacci chain perturbatively renormalizes to exact scaled copies of itself, it can be viewed as a critical model. The Tribonacci chains renormalize perturbatively to different kinds of HTCs, viz. HTCs with \(p\neq 1\) and/or \(q\neq 1\). The limit of the RG procedure for the HTC yields infinitely many copies of the HTC with \(p=q=\infty\), which is quite different from the original model where \(p=q=1\). In this way, the HTC and OTC can be viewed as less critical than the Fibonacci chain. Regardless of this fact, in Section V it will be shown that the eigenstates of the HTC are critical. Figure 4: (Color online) The energy spectrum of the OTC Eq. (21) with \(c=5\) and \(T_{13}=3136\) sites. The three main bands sit around \(E/t=0,c,2c\). The inset shows that the main bands further split up like HTC Hamiltonians with certain \(p,q\) values. Eigenstates on the Rauzy fractal Considering the Hamiltonian \(H_{N}\) in Eq. (15) (or Eq. (21)), the Schrodinger equation \(H_{N}\left|\psi\right\rangle_{i}=E_{i}\left|\psi\right\rangle_{i}\) will have \(T_{N}\) solutions labeled by \(i=0,\ldots,T_{N}-1\). Each eigenstate has the form \(\left|\psi\right\rangle_{i}=\sum_{n=0}^{T_{N}}\psi_{i}(n)\left|n\right\rangle\), where \(\psi_{i}(n)\in\mathbb{C}\). The eigenstate \(\left|\psi\right\rangle_{i}\) can be plotted on the Rauzy fractal by identifying each point \(\mathbf{x}_{n}\) in Eq. (9) with the probability \(|\psi_{i}(n)|^{2}\), which determines the size of a black triangle at that point. ### Hopping Model When associating HTC lattice points \(\left|n\right\rangle\) with Rauzy fractal points \(\mathbf{x}_{n}\), one has to apply a different coloring of the Rauzy fractal. Each site has no on-site energy, and can have the local environments * Red: 01 (\(t_{0}\) on the left and \(t_{1}\) on the right) or 10, * Green: 02 or 20, * Blue: 00. The eigenstate \(\left|\psi\right\rangle_{0}\) of the HTC \(H_{13}\) is plotted on the Rauzy fractal in Fig. 5(a). Since the energy \(E_{0}\) comes from the bottom branch of the spectrum in Fig. 2, it should be a state that antibonds on sites connected with \(t_{2}\) bonds. This is precisely reflected by the plot on the Rauzy fractal in Fig. 5(a), since the eigenstate is mainly localized in the green region, corresponding to a site neighboring a \(t_{2}\) bond. Generally, a state from branch 1 (or \(2,3,4,5\)), in this case from the lowest set of eigenvalues at \(E=-t_{2}\), in Fig. 2 is primarily localized in the green (or red, blue, red, green) region(s) in Fig. 5(a) (see Appendix D for more examples). Finally, for any \(H_{N}\), each red, green and blue region contains exactly \(T_{N-2},T_{N-3}\) and \(T_{N-4}\) points, matching the amount of points in each branch of the spectrum. For each local structure, there are again exactly five distinct environments around that structure. For example, the environment of 01 or 10 is always \(x010y\) where \(xy=02,20,21,12\) or \(22\). It turns out that these correspond exactly to the local structures \(01,10,02,20,00\) of the type-1 molecular chain. The subdivisions of the Rauzy fractal are carried out in Fig. 5(b). We have shown that if one is interested in all possible environments of a lattice site, it is enough to consider only the nearest-neighbor environments and the RG scheme. Using the RG scheme, next variations on the nearest-neighbor environments of a lattice site are given by the nearest-neighbor environments of the renormalized chain to which that lattice site belongs. ### On-Site Model When plotting the eigenstates of the OTC onto the Rauzy fractal, the original coloring can be used, since each lattice site \(\left|n\right\rangle\) corresponds to some \(\epsilon_{w_{n}}\). The state with index \(i=0\) is plotted on the Rauzy fractal in Fig. 5(c). Since \(E_{0}\) comes from the bottom branch of the spectrum in Fig. 4, the eigenstate is localized on the red part of the Rauzy fractal which corresponds to 0 in \(W\). It is again a general feature that states from some branch in the spectrum localize on the corresponding part of the Rauzy fractal (see Appendix D for more examples). Since the on-site model renormalizes to three hopping models in Eq. (23), additional subdivision of the Rauzy fractal based on next local environments yield a similar subdivision as for the hopping model. This is displayed in Fig. 5(d). We would like to point out the similarity between the eigenstates \(\left|\psi\right\rangle_{0}\) in Fig. 5(a) and in the red region in Fig. 5(d). This can be understood by the fact that the eigenstate \(\left|\psi\right\rangle_{0}\) of the OTC \(H_{13}^{\prime}\) is approximately the eigenstate of the first block of Eq. (23), which is a HTC. Another observation is the self-similar structure of the eigenstates on the Rauzy fractals in Fig. 5. This is a signature of critical eigenstates [34], which are also characteristic of the Fibonacci chain [15]. For the Tribonacci chains, fractality is discussed in Section V. ### Equivalence Local Environment and Local Resonator Modes It is an interesting fact that all local environments are known from only the nearest-neighbor structures and the RG Eq. (20). This fact can be applied to elegantly categorize all LRMs of the HTC and OTC. This LRM framework was developed by Rontgen et al. [23], and applied to the Fibonacci chain. In Figs. 6 and 7, the eigenstate magnitude on each lattice site with \(T_{7}=81\) sites is plotted for every energy level. The green lines define regions that precisely correspond to the diagonal blocks in Eqs. (20) and (23), so they correspond to one application of the RG scheme. By applying Eq. (20) again to each of these blocks of the Hamiltonian at hand, the regions subdivide again into five smaller ones (see black horizontal lines). The connection with the LRM framework is that the subsequent subdivisions order the eigenstates according to their local structure, i.e. where they are mostly localized. This classification is an essential step in the application of the LRM framework, which is a naturally carried out by the RG equations. The RG scheme naturally gives all environments of a lattice site, and at the same time categorizes the LRMs. This simplification of the analysis is founded on the self-similarity of the Tribonacci word. ## V Multifractality The perturbative RG scheme for the Fibonacci chain provided a natural way of explaining the multifractal properties of the spectrum and of the eigenstates. Since an analogous RG scheme is derived for the Tribonacci chains, multifractality is expected to be present. Since the multifractal properties of the HTC are compared to the Fibonacci chain, the definition of the Fibonacci chain is briefly reviewed here. The Fibonacci word \(W^{F}=w_{0}^{F}w_{1}^{F}\cdots\) is the fixed point of the binary substitution \(\rho_{F}:0\to 01,1\to 0\). The Fibonacci approximants are given by \(W_{N}^{F}:=\rho_{F}^{N}(1)\). The length of the approximants is given by the Fibonacci numbers \(F_{N}=F_{N-1}+F_{N-2}\), where \(F_{0}=F_{1}=1\). The Hamiltonian for the periodic hopping Fibonacci chain reads \[H_{N}^{F}=\sum_{n=0}^{F_{N}-1}t_{w_{n}^{F}}\left|n+1\bmod F_{N}\right\rangle \left\langle n\right|+H.c., \tag{27}\] where the hopping parameters \(t_{0},t_{1}\) are related by \(t_{0}/t_{1}=\rho\). To study the multifractal properties of the spectrum of any Hamiltonian, we compute the multifractal dimensions \(D_{q}\), also known as the multifractal spectrum, introduced by Halsey et al. [35]. The multifractal spectrum is a family of dimensions that is continuously parameterized by \(q\in\mathbb{R}\), where \(D_{0}\) recovers the box-counting dimension. For the energy spectrum, the multifractal dimensions are computed as follows. First, cover the energy spectrum with a compact interval \(C\subset\mathbb{R}\). Then, partition \(C\) into intervals \(K_{i}\) of length \(l\). Let the measure \(\mu(K_{i})\) denote the fraction of points that lie in \(K_{i}\) Figure 5: (Color online) The eigenstate \(\left|\psi\right\rangle_{0}\) on the Rauzy fractal of \(T_{13}=3136\) points. The regions are colored according to the local environment of a lattice site \(n\) in the HTC (or OTC), and the length of the black triangles are proportional to \(\left|\psi_{0}(n)\right|^{2}\). (a) \(\left|\psi\right\rangle_{0}\) of the HTC \(H_{13}\) with coupling \(\rho=0.2\) and coloring according to nearest-neighbor bonds. (b) \(\left|\psi\right\rangle_{0}\) of the HTC \(H_{13}\) with coupling \(\rho=0.2\) and coloring according to the five possible environments of the local structures in a). (c) \(\left|\psi\right\rangle_{0}\) of the OTC \(H_{13}^{\prime}\) with coupling \(c=5\) and and coloring according the on-site potential of a lattice site. (d) \(\left|\psi\right\rangle_{0}\) of the OTC \(H_{13}^{\prime}\) with coupling \(c=5\) and coloring according to the five possible environments of the lattice sites in c). The multifractal dimensions are then given by \[D_{q}=\lim_{l\downarrow 0}\frac{1}{q-1}\frac{\log\sum_{i}\mu(K_{i})^{q}}{\log l}. \tag{28}\] The result is shown in Fig. 8(a), where the multifractal dimensions of the HTC \(H_{13}\) and the Fibonacci chain \(H_{19}^{F}\) are plotted. One can see that the HTC energy spectrum is a multifractal, since the spectrum \(D_{q}\) is a smooth curve of \(q\). Moreover, the multifractal dimensions of the HTC are strictly smaller than that of the Fibonacci chain. For the eigenstates, the average multifractal dimension is computed. The average multifractal dimension of the eigenstates is defined as [20; 36] \[\overline{D_{q}^{\psi}}=\frac{1}{q-1}\frac{\log\frac{1}{N}\sum_{i}\sum_{n}| \psi_{i}(n)|^{2q}}{\log 1/N}, \tag{29}\] where the sum over \(i\) ranges over all eigenstates, \(N\) denotes the amount of eigenstates and \(n\) ranges over the lattice sites. The numerical results for the Fibonacci chain and HTC are displayed in Fig. 8(b). The average multifractal dimension of the HTC is lower than of the Fibonacci chain. This is to be expected, since \(\overline{D_{q}^{\psi}}\) is related to diffusive properties of the system [36]. The weakest bonds in the HTC are \(\mathcal{O}(\rho^{2})\), whereas in the Fibonacci they are \(\mathcal{O}(\rho)\). This makes it more difficult for a particle to diffuse in the HTC than in the Fibonacci chain, which is in accordance with the fact that the average multifractal dimension for the HTC is lower than of the Fibonacci chain. Additionally, a lower average multifractal dimension indicates that the wavefunctions are more localized, which is a consequence of the weaker bonds in the HTC. In fact, the HTC is a critical chain in terms of Anderson localization, since the eigenstates are multifractal with \(0<\overline{D_{q}^{\psi}}<1\)[34]. Finally, because the OTC is approximately a direct product of HTC Hamiltonians in Eq. (23), the multifractal properties perturbatively carry over to the OTC. ## VI Conclusion In this work, we introduced two tight-binding chains Eqs. (13) and (14), based on the Tribonacci substitution, which generalizes the Fibonacci chain. One of the first Figure 7: (Color online) The OTC eigenstates \(\ket{\psi}_{i}\) of \(H_{7}^{\sigma}\). The colors and green/black lines have the same meaning as in Fig. 6. Figure 6: (Color online) The HTC eigenstates \(\ket{\psi}_{i}\) of \(H_{7}\), ordered such that \(E_{i}<E_{i+1}\). The sign and magnitude on each site is represented by a color. The green lines denote the splitting after one RG step, the black lines denote two RG steps. Note that the states between two subsequent lines localize on similar local environments, which is more accurate for the black lines than for the green lines. steps towards understanding these models are the RG Eqs. (20) and (23), which are more accurate than those for the Fibonacci chain due to the higher orders of perturbation theory required. As shown in Section III.2, the two models can be regarded as equivalent at the RG fixed point. The Rauzy fractal, which is inherent to the Tribonacci word, is shown to serve as the analog of the conumbers for the HTC and OTC, since it orders the sites according to their local environment. The structure of eigenstates, when plotted on the Rauzy fractal, shows self-similar properties, which reflect the fractal nature of the eigenstates. These self-similar structures can be systematically enumerated using the RG scheme, and are exactly the LRMs within the framework proposed by Rontgen et al. [23]. Finally, the multifractal dimensions of both the energy spectrum and the eigenstates of the HTC have been computed, and compared to those of the Fibonacci chain. The multifractal properties are qualitatively similar to those of the Fibonacci chain, whereas the multifractal dimensions of the HTC are generally smaller than those of the Fibonacci chain. Furthermore, the HTC is shown to be a critical model in terms of Anderson localization, since the wavefunctions exhibit multifractal properties with a dimension between zero and one. This work opens some interesting topics for further research. First of all, it would be interesting to identify an equivalence between the HTC and another model, such as the one by Kraus and Zilberberg [21] for the Fibonacci chain. Such an equivalence would be key to understanding the topological properties of the HTC. One could also generalize the substitution to any Pisot substitution, or consider the general \(k\)-bonacci substitution \(0\to 01,1\to 02,\ldots,(k-1)\to 0\). The latter would make the generalization of the Fibonacci chain as complete as the complementary generalization in Refs.[28; 29; 30]. Yet another proposition to check is whether quasicrystals can generally be studied via their internal space, which is conumbering for the Fibonacci chain and the Rauzy fractal for the HTC, and how the RG scheme can be applied in the internal space to understand the eigenstates. Since the RG scheme originates from the self-similar structure, it could be interesting to study if self-similarity can replace translational invariance in the topological classification of quasicrystals and/or fractals. Finally, experimental realizations, such as polaritonic waveguides [37] and dielectric resonators [38] for the Fibonacci chain, can be realized to probe the electronic and multifractal properties of the HTC and OTC. ###### Acknowledgements. This publication is part of the project TOPCORE with Project No. OCENW.GROOT.2019.048 which is financed by the Dutch Research Council (NWO).

2310.13814

The Turán and Laguerre inequalities for quasi-polynomial-like functions

This paper deals with both the higher order Tur\'an inequalities and the Laguerre inequalities for quasi-polynomial-like functions -- that are expressions of the form $f(n)=c_l(n)n^l+\cdots+c_d(n)n^d+o(n^d)$, where $d,l\in\mathbb{N}$ and $d\leqslant l$. A natural example of such a function is the $A$-partition function $p_{A}(n)$, which enumerates the number of partitions of $n$ with parts in the fixed finite multiset $A=\{a_1,a_2,\ldots,a_k\}$ of positive integers. For an arbitrary positive integer $d$, we present efficient criteria for both the order $d$ Tur\'an inequality and the $d$th Laguarre inequality for quasi-polynomial-like functions. In particular, we apply these results to deduce non-trivial analogues for $p_A(n)$.

Krystian Gajdzica

2023-10-20T21:05:21Z

http://arxiv.org/abs/2310.13814v1

# The Turan and Laguerre Inequalities for Quasi-Polynomial-like Functions ###### Abstract. This paper deals with both the higher order Turan inequalities and the Laguerre inequalities for quasi-polynomial-like functions -- that are expressions of the form \(f(n)=c_{l}(n)n^{l}+\cdots+c_{d}(n)n^{d}+o(n^{d})\), where \(d,l\in\mathbb{N}\) and \(d\leqslant l\). A natural example of such a function is the \(A\)-partition function \(p_{A}(n)\), which enumerates the number of partitions of \(n\) with parts in the fixed finite multiset \(A=\{a_{1},a_{2},\ldots,a_{k}\}\) of positive integers. For an arbitrary positive integer \(d\), we present efficient criteria for both the order \(d\) Turan inequality and the \(d\)th Laguerre inequality for quasi-polynomial-like functions. In particular, we apply these results to deduce non-trivial analogues for \(p_{A}(n)\). Key words and phrases:integer partition, \(A\)-partition function, quasi-polynomial, log-concavity, higher order Turan inequalities, Laguerre inequalities 2020 Mathematics Subject Classification: Primary 11P82, 11P84; Secondary 05A17 ## 1. Introduction A partition of a non-negative integer \(n\) is a weakly-decreasing sequence of positive integers \(\lambda=(\lambda_{1},\lambda_{2},\ldots,\lambda_{j})\) such that \[n=\lambda_{1}+\lambda_{2}+\cdots+\lambda_{j}.\] The numbers \(\lambda_{i}\) are called parts of the partition \(\lambda\). The partition function \(p(n)\) enumerates all partitions of \(n\). For instance, there are \(5\) partitions of \(4\), namely, \((4)\), \((3,1)\), \((2,2)\), \((2,1,1)\) and \((1,1,1,1)\) -- in other words \(p(4)=5\). We do not know any easy formula for \(p(n)\). However, Euler proved that its generating function takes the form \[\sum_{n=0}^{\infty}p(n)x^{n}=\prod_{i=1}^{\infty}\frac{1}{1-x^{i}}.\] The partition theory plays a crucial rule in many parts of mathematics and other sciences. In statistical mechanics, the well-known Rogers-Ramanujan identities are related to the solution of the hard hexagon model, see [3, 7]. Further, partitions have applications in molecular chemistry, crystallography and quantum mechanics, as a consequence of the fact that all irreducible representations of the permutation group \(S_{n}\) and the unitary group \(U(n)\) might be labelled by them. It is also worth noting that partitions appear in genetics in the so-called Ewens's sampling formula, see [24, 34]. There is a plethora of works devoted to the theory of partitions. For a general introduction to the topic, we encourage the reader to see Andrews' books [4, 5] as well as [1, 31, 45]. Now, let us assume that \(A=\{a_{1},a_{2},\ldots,a_{k}\}\) is a finite multiset of positive integers. By an \(A\)-partition of a non-negative integer \(n\), we mean any partition \(\lambda=(\lambda_{1},\lambda_{2},\ldots,\lambda_{j})\) of \(n\) with parts in \(A\). For the sake of clarity, we additionally assume that two \(A\)-partitions are considered the same if there is only a difference in the order of their parts. The \(A\)-partition function \(p_{A}(n)\) enumerates all \(A\)-partitions of \(n\). In particular, we have that \(p_{A}(n)=0\) whenever \(n\) is a negative integer and \(p_{A}(0)=1\) with \(\lambda=()\). The generating function for \(p_{A}(n)\) is given by \[\sum_{n=0}^{\infty}p_{A}(n)x^{n}=\prod_{a\in A}\frac{1}{1-x^{a}}. \tag{1.1}\] For example, if \(A=\{1,2,2,3,3,3,4,4\}\), then we have that \(p_{A}(4)=11\), namely: (4), (4), (3,1), (3,1), (3,1), (2,2), (2,2), (2,2), (2,1,1), (2,1,1) and \((1,1,1)\). There is an abundance of literature devoted to \(A\)-partition function when \(\#A<\infty\). We refer the reader to, for instance, [2, 10, 15, 21, 25, 37, 44, 49]. It turns out that \(p_{A}(n)\) is a quasi-polynomial whenever \(A\) is a finite set or a multiset of positive integers. More precisely, if \(\#A=k\), then the \(A\)-partition function is an expression of the form \[p_{A}(n)=b_{k-1}(n)n^{k-1}+b_{k-2}(n)n^{k-2}+\cdots+b_{0}(n), \tag{1.2}\] where the coefficients \(b_{0}(n),b_{1}(n),\ldots,b_{k-1}(n)\) depend on the residue class of \(n\,(\mathrm{mod\,\,lcm}A)\). The first proof of the above fact is probably due to Bell [10]. We encourage the reader to see Stanley's book [46, Section 4.4] for more information about quasi-polynomials. On the other hand, a quasi-polynomial-like function \(f(n)\) is a function which asymptotically behaves like a quasi-polynomial. More specifically, \(f(n)\) can be written as \[f(n)=c_{l}(n)n^{l}+c_{l-1}(n)n^{l-1}+\cdots+c_{r}(n)n^{r}+o(n^{r}), \tag{1.3}\] where \(r,l\in\mathbb{N}\), \(l\geqslant r\), the coefficients \(c_{r}(n),c_{r+1}(n),\ldots,c_{l}(n)\) depend on the residue class of \(n\,(\mathrm{mod\,\,}M)\) for some positive integer \(M\geqslant 2\). In particular, we see that \(p_{A}(n)\) is a quasi-polynomial-like function. This paper deals with two problems. The first of them concerns the so-called higher order Turan inequalities for quasi-polynomial-like functions. Let us recall that a sequence \(\left(\omega_{i}\right)_{i=0}^{\infty}\) of real numbers satisfies the second order Turan inequality if we have that \[\omega_{n}^{2}\geqslant\omega_{n+1}\omega_{n-1}\] for all \(n\geqslant 1\). Further, it fulfills the third order Turan inequality if the following \[4(\omega_{n}^{2}-\omega_{n-1}\omega_{n+1})(\omega_{n+1}^{2}-\omega_{n}\omega_ {n+2})\geqslant(\omega_{n}\omega_{n+1}-\omega_{n-1}\omega_{n+2})^{2}\] is true for every \(n\geqslant 1\). More generally, if \(J_{\omega}^{d,n}(x)\) are the Jensen polynomials of degree \(d\) and shift \(n\) associated to the sequence \(\omega:=(\omega_{i})_{i=0}^{\infty}\), defined by \[J_{\omega}^{d,n}(x):=\sum_{i=0}^{d}\binom{d}{i}\omega_{n+i}x^{i},\] then it is known that \((\omega_{i})_{i=0}^{\infty}\) satisfies the order \(d\) Turan inequality at \(n\) if and only if \(J_{\omega}^{d,n}(x)\) is hyperbolic, i.e. all of its roots are real numbers (see, [17, 18, 19, 28]). In 2015 DeSalvo and Pak [20] reproved the result (obtained independently by Nicolas [38] in the '70s) that the partition function \(p(n)\) satisfies the second order Turan inequality for all \(n>25\). Afterwards, Chen [13] conjectured that the third order Turan inequality for \(p(n)\) is valid for all \(n\geqslant 94\). The problem was solved by Chen, Jia and Wang [14] and motivated them to state another conjecture that for each \(d\geqslant 1\) there is some integer \(N_{p}(d)\) such that the associated Jensen polynomial \(J_{p}^{d,n}(X)\) is hyperbolic for all \(n\geqslant N_{p}(d)\). That conjecture, on the other hand, was established by Griffin et al. [28]. It is worth pointing out that Larson and Wagner [35] discovered efficient upper bound for the value of \(N_{p}(d)\) for any \(d\). The aforementioned results have initiated vast research related to discovering similar properties for other variations of the partition function. Iskander et al. [33] proved that for every \(d\geqslant 2\) the fractional partition function \(p_{\alpha}(n)\), which is defined for \(\alpha\in\mathbb{Q}\) in terms of the following generating function \[\sum_{n=0}^{\infty}p_{\alpha}(n)x^{n}:=\prod_{i=1}^{\infty}\frac{1}{(1-x^{i})^{ \alpha}}\] (for more information, see [12]), satisfies the order \(d\) Turan inequality for all but finitely many values of \(n\). Further, Craig and Pun [16] investigated the so-called \(k\)-regular partition function \(p_{k}(n)\) (i.e. \(p_{k}(n)\) enumerates only those partitions of \(n\) whose parts are not divisible by \(k\)) in that context. They obtained that for every \(k\geqslant 2\) and \(d\geqslant 1\) the associated Jensen polynomial \(J_{p_{k}}^{d,n}(X)\) is hyperbolic for all sufficiently large numbers \(n\). Heim, Neuhauser and Troger [30] investigated the plane partition function \(PL(n)\) (see Andrews [4, Chapter 11] or [5, Chapter 10]) and its polynomization in this direction. They conjectured that for any \(d\geqslant 1\) the plane partition function fulfills the order \(d\) Turan inequality for all large enough numbers \(n\). That conjecture was solved by Ono, Pujahari and Rolen in [41] with explicit bounds provided by Ono's PhD student Pandey [42]. Further, Baker and Males [8] showed that the number \(\overline{p}_{j}(n)\) of partitions with BG-rank \(j\), and the number \(\overline{p}_{j}(a,b;n)\) of partitions with BG-rank \(j\) and \(2\)-quotient rank congruent to \(a\,(\mathrm{mod}\;b)\) satisfy (asymptotically) all higher order Turan inequalities for even values of \(j\) and \(n\). We refer the reader to Berkovich and Garvan's paper [11] for additional information about \(\overline{p}_{j}(n)\) and \(\overline{p}_{j}(a,b;n)\). Finally, Dong, Ji and Jia [23] discovered that the Jensen polynomial corresponding to \(d\geqslant 1\) and the Andrews and Paule's broken \(k\)-diamond partition function \(\Delta_{k}(n)\), namely \(J_{\Delta_{k}}^{d,n}(X)\), is hyperbolic for \(k=1\) or \(2\) and all but finitely many positive integers \(n\). The explicit definition of broken \(k\)-diamond partitions (for any \(k\geqslant 1\)) together with some properties of \(\Delta_{k}(n)\) might be found in Andrews and Paule's paper [6]. The above-mentioned results have been our motivation to study the higher order Turan inequalities for both quasi-polynomial-like functions in general and \(A\)-partition functions in particular. The second issue which this paper deals with concerns the so-called Laguerre inequalities for quasi-polynomial-like functions. Once again, let us assume that \(\omega=(\omega_{i})_{i=0}^{\infty}\) is a sequence of real numbers. For a fixed non-negative integer \(d\), we say that \(\omega\) satisfies the Laguerre inequality of order \(d\) at \(n\) if \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}\omega_{n+j}\omega_{n+2d-j}\geqslant 0. \tag{1.4}\] The discrete Laguerre inequalities (1.4) were firstly introduced by Wang and Yang [52]. It is also worth noting that Wagner [51, Theorem 1.4] defined them equivalently by dividing (1.4) by \((2d)!\). For \(d=1\), one can easy observe that (1.4) reduces to the second order Turan inequality. If \(d=2\), then (after simplification) we get \[3\omega_{n+2}^{2}-4\omega_{n+1}\omega_{n+3}+\omega_{n}\omega_{n+4}\geqslant 0.\] Further, the order \(3\) Laguerre inequality might be written equivalently as follows: \[10\omega_{n+3}^{2}-15\omega_{n+2}\omega_{n+4}+6\omega_{n+1}\omega_{n+5}- \omega_{n}\omega_{n+6}\geqslant 0,\] and so on. Wang and Yang [52, 53] investigated Lagurre inequalities for many combinatorial sequences. In particular, they showed that the partition function, the overpartition function, the Motzkin numbers, the Fine numbers, the Domb numbers and the distinct partition function satisfy the order \(2\) Laguerre inequality. More recently, Yang [55] also proved that the broken \(k\)-diamond partition function fulfills the second order Laguerre inequality. On the other hand, Wagner [51] showed that the partition function satisfies the inequality (1.4) for every non-negative integer and all sufficiently large values of \(n\). The aforementioned results have motivated us to investigate the issue in the case of quasi-polynomial-like functions. At the end of Introduction, it needs to be pointed out that studying both the higher order Turan inequalities and the Laguerre inequalities is not only art for art's sake. Let us recall that a real entire (i.e. analytic at all points of the complex plane \(\mathbb{C}\)) function \[f(x)=\sum_{n=0}^{\infty}a_{n}\frac{x^{n}}{n!}\] is in the \(\mathcal{LP}\) (Laguerre-Polya) class if it may be written as \[f(x)=cx^{k}e^{-ax^{2}+bx}\prod_{n=1}^{\infty}\left(1+\frac{x}{x_{n}}\right)e^{ -\frac{x}{x_{n}}},\] where \(a,b,c,x_{1},x_{2},\ldots\) are all real numbers with \(a\geqslant 0\), \(k\) is a non-negative integer and \(\sum_{n=1}^{\infty}x_{n}^{-2}<\infty\). For the background of the theory of the \(\mathcal{LP}\) functions, we encourage the reader to see [36, 43]. It turns out that the Riemann hypothesis is equivalent to the statement that the Riemann \(\Xi\)-function \[\Xi(z):=\frac{1}{2}\left(-z^{2}-\frac{1}{4}\right)\pi^{\frac{iz}{2}-\frac{1}{ 4}}\Gamma\left(-\frac{iz}{2}+\frac{1}{4}\right)\zeta\left(-iz+\frac{1}{2}\right)\] is in the \(\mathcal{LP}\) class, where \(\Gamma\) is the gamma function, and \(\zeta\) denotes the Riemann zeta function. There is a necessary condition for the Riemann \(\Xi\)-function to be in the Laguerre-Polya class which states that the Maclaurin coefficients of the \(\Xi\)-function have to fulfill the order \(d\) Turan inequality as well as the Laguerre inequality of order \(d\) for every positive integer \(d\). For additional information, we refer the reader to [22, 40, 47]. This manuscript is organized as follows. Section 2 delivers necessary concepts, notations and properties which are used throughout the paper. Section 3 studies the higher order Turan inequalities for both quasi-polynomial-like functions and \(A\)-partition functions. In Section 4, on the other hand, we deal with the Laguerre inequalities. Finally, Section 5 contains some concluding remarks and open problems. ## 2. Preliminaries At first, we fix some notation. The set of non-negative integers is denoted by \(\mathbb{N}\). Further, we put \(\mathbb{N}_{+}:=\mathbb{N}\setminus\{0\}\) and \(\mathbb{N}_{\geqslant k}:=\mathbb{N}\setminus\{0,1,\ldots,k-1\}\). For a finite multiset \(A=\{a_{1},a_{2},\ldots,a_{k}\}\) of positive integers, we associate the \(A\)-partition function \(p_{A}(n)\), which was defined in Introduction. Due to Bell's theorem [10], we know that \(p_{A}(n)\) is a quasi-polynomial given by the equality (1.2), where the coefficients \(b_{0}(n),b_{1}(n),\ldots,b_{k-1}(n)\) depend on the residue class of \(n\,(\mathrm{mod}\ \mathrm{lcm}A)\). It turns out that under some additional assumptions on \(A\), we may determine some of the coefficients \(b_{i}(n)\). That is a result obtained by several authors, among others, Almkvist [2], Beck et al. [9] or Israailov [32]. We present the theorem due to Almkvist [2]. In order to do that, let us define symmetric polynomials \(\sigma_{i}(x_{1},x_{2},\ldots,x_{k})\) in terms of the power series expansion \[\sum_{m=0}^{\infty}\sigma_{m}(x_{1},x_{2},\ldots,x_{k})t^{m}:=\prod_{i=1}^{k} \frac{x_{i}t/2}{\sinh(x_{i}t/2)}.\] Now, we have the following. **Theorem 2.1** (Almkvist).: _Let \(A=\{a_{1},a_{2},\ldots,a_{k}\}\) be fixed and put \(s_{1}:=a_{1}+a_{2}+\cdots+a_{k}\). For a given integer \(1\leqslant j\leqslant k\), if \(\gcd B=1\) for every \(j\)-element multisubset (\(j\)-multisubset) \(B\) of \(A\), then_ \[p_{A}(n)=\frac{1}{\prod_{i=1}^{k}a_{i}}\sum_{i=0}^{k-j}\sigma_{i}(a_{1},a_{2}, \ldots,a_{k})\frac{(n+s_{1}/2)^{k-1-i}}{(k-1-i)!}+O(n^{j-2}).\] One can check that \(\sigma_{i}=0\) if \(i\) is odd. Furthermore, if we set \(s_{m}:=a_{1}^{m}+a_{2}^{m}+\cdots+a_{k}^{m}\), then \[\sigma_{0}=1,\ \sigma_{2}=-\frac{s_{2}}{24},\ \sigma_{4}=\frac{5s_{2}^{2}+2s_{4 }}{5760},\ \sigma_{6}=-\frac{35s_{2}^{3}+42s_{2}s_{4}+16s_{6}}{2903040}.\] Essentially, Theorem 2.1 maintains that if \(\gcd B=1\) for every \((k-j)\)-multisubset \(B\) of \(A\), then the coefficients \(b_{k-1}(n),b_{k-2}(n),\ldots,b_{k-1-j}(n)\) in the equality (1.2) are independent of the residue class of \(n\,(\text{mod lcm}A)\), i.e. they are constants and can be explicitly calculated. Moreover, it is noteworthy that the \(A\)-partition function is a non-trivial example of a quasi-polynomial-like function -- that is an expression of the form (1.3). Now, let us recall some terminology related to higher order Turan inequalities. Instead of repeating the discussion from Introduction, we directly explain how the order \(d\) Turan inequality arises from the hyperbolicity of the Jensen polynomial \(J_{\omega}^{d,n}(x)\) has to be hyperbolic. Let \[g(x)=c_{s}x^{s}+c_{s-1}x^{s-1}+c_{s-2}x^{s-2}+\cdots+c_{0}\] be a fixed polynomial with real coefficients and denote all its complex roots by \(\alpha_{1},\alpha_{2},\ldots,\alpha_{s}\). By \(P_{m}\), we mean the \(m\)-th Newton's sum of \(g(x)\), which is given by \[P_{m}=\begin{cases}s,&\text{if }m=0,\\ \alpha_{1}^{m}+\alpha_{2}^{m}+\cdots+\alpha_{s}^{m},&\text{if }m=1,2,3,4, \ldots.\end{cases}\] Further, for the sums \(P_{0},\ldots,P_{2s-2}\), we associate the Hankel matrix \(H(g)\), namely \[H(g):=\begin{bmatrix}P_{0}&P_{1}&P_{2}&\cdots&P_{s-1}\\ P_{1}&P_{2}&P_{3}&\cdots&P_{s}\\ P_{2}&P_{3}&P_{4}&\cdots&P_{s+1}\\ \vdots&\vdots&\vdots&\vdots&\vdots\\ P_{s-2}&P_{s-1}&P_{s}&\cdots&P_{2s-3}\\ P_{s-1}&P_{s}&P_{s+1}&\cdots&P_{2s-2}\end{bmatrix}.\] The classical Hermit's theorem [39] states that \(g(x)\) is hyperbolic if and only if the matrix \(H(g)\) is positive semi-definite. Since each of the Newton's sums might be expressed in terms of the coefficients \(c_{s},c_{s-1},\ldots,c_{0}\), Hermit's result provides a set of inequalities on them by \[\det\Big{[}P_{0}\Big{]}\geqslant 0,\det\begin{bmatrix}P_{0}&P_{1}\\ P_{1}&P_{2}\end{bmatrix}\geqslant 0,\ldots,\det\begin{bmatrix}P_{0}&P_{1}&P_{2}& \cdots&P_{s-1}\\ P_{1}&P_{2}&P_{3}&\cdots&P_{s}\\ P_{2}&P_{3}&P_{4}&\cdots&P_{s+1}\\ \vdots&\vdots&\vdots&\vdots&\vdots\\ P_{s-2}&P_{s-1}&P_{s}&\cdots&P_{2s-3}\\ P_{s-1}&P_{s}&P_{s+1}&\cdots&P_{2s-2}\end{bmatrix}\geqslant 0.\] Now, if we assign the Jensen polynomial \(J_{\omega}^{d,n}(x)\) for an arbitrary sequence \(\omega=(w_{i})_{i=1}^{\infty}\), then the corresponding inequality for the determinant of the main minor \(l\times l\) of \(H(J_{\omega}^{d,n})\): \[\det\begin{bmatrix}P_{0}&P_{1}&P_{2}&\cdots&P_{l-1}\\ P_{1}&P_{2}&P_{3}&\cdots&P_{l}\\ P_{2}&P_{3}&P_{4}&\cdots&P_{l+1}\\ \vdots&\vdots&\vdots&\vdots&\vdots\\ P_{l-2}&P_{l-1}&P_{l}&\cdots&P_{2l-3}\\ P_{l-1}&P_{l}&P_{l+1}&\cdots&P_{2l-2}\end{bmatrix}\geqslant 0\] is called the order \(l\) Turan inequality for the sequence \(\omega\). In particular, it means that \(J_{\omega}^{d,n}(x)\) is hyperbolic if and only if the sequence \(\omega_{n}=(w_{n+j})_{j=1}^{\infty}\) satisfies the order \(l\) Turan inequality for every \(l\in\{1,2,\ldots,d\}\). From the above discussion, we see that investigating the higher order Turan inequalities does not seem to be an easy challenge. However, there is a paper due to Griffin, Ono, Rolen and Zagier [28], which delivers an efficient criterion to deal with that issue. **Theorem 2.2** (Griffin, Ono, Rolen, Zagier).: _Let \((\omega_{n})_{n=0}^{\infty}\) be a sequence of real numbers. Suppose further that \((E(n))_{n=0}^{\infty}\) and \((\delta(n))_{n=0}^{\infty}\) are sequences of positive real numbers with \(\lim_{n\to\infty}\delta(n)=0\), and that \(F(t)=\sum_{i=0}^{\infty}c_{i}t^{i}\) is a formal power series with complex coefficients. For a fixed \(d\geqslant 1\), suppose that there are sequences \(\left(C_{0}(n)\right)_{n=0}^{\infty},\left(C_{1}(n)\right)_{n=0}^{\infty}, \ldots,\left(C_{d}(n)\right)_{n=0}^{\infty}\) of real numbers, with \(\lim_{n\to\infty}C_{i}(n)=c_{i}\) for \(0\leqslant i\leqslant d\), such that for \(0\leqslant j\leqslant d\), we have_ \[\frac{\omega_{n+j}}{\omega_{n}}E(n)^{-j}=\sum_{i=0}^{d}C_{i}(n)\delta(n)^{i}j ^{i}+o\left(\delta(n)^{d}\right)\qquad\text{ as }n\to\infty.\] _Then, we have_ \[\lim_{n\to\infty}\left(\frac{\delta(n)^{-d}}{\omega_{n}}J_{\omega}^{d,n}\left( \frac{\delta(n)x-1}{E(n)}\right)\right)=H_{F,d}(x),\] _uniformly for \(x\) in any compact subset of \(\mathbb{R}\), where the polynomials \(H_{F,m}(x)\in\mathbb{C}[x]\) are defined either by the generating function \(F(-t)e^{xt}=\sum_{m=0}^{\infty}H_{F,m}(x)t^{m}/m!\) or in closed form by \(H_{F,m}(x):=m!\sum_{l=0}^{m}(-1)^{m-l}c_{m-l}x^{l}/l!\)._ It is not clear how one can apply the above result in practice. In fact, Griffin et al. use the criterion to prove that for every positive integer \(d\) the partition function \(p(n)\) fulfills the order \(d\) Turan inequality for all but finitely many values of \(n\). More precisely, they obtain the Hermite polynomials \(H_{m}(x)\) as the polynomials \(H_{F,m}(x)\) in Theorem 2.2. Let us recall that they define the Hermit polynomials via the generating function \[\sum_{j=0}^{\infty}H_{j}(x)\frac{t^{j}}{j!}=e^{-j^{2}+jx}=1+jx+\frac{j^{2}}{2! }(x^{2}-2)+\cdots.\] Since these polynomials have only distinct real roots, and since the property of a polynomial with only real roots is invariant under small deformation, the required phenomenon for \(p(n)\) follows. On the other hand, we investigate the higher order Turan inequalities for quasi-polynomial-like functions \[f(n)=c_{l}(n)n^{l}+c_{l-1}(n)n^{l-1}+\cdots+c_{r}(n)n^{r}+o(n^{r}),\] where \(r,l\in\mathbb{N}\), \(l\geqslant r\) and the coefficients \(c_{r}(n),c_{r+1}(n),\ldots,c_{l}(n)\) depend on the residue class of \(n\,(\mathrm{mod}\;M)\) for some positive integer \(M\geqslant 2\). Therefore, we will probably get another family of orthogonal polynomials in Theorem 2.2. The generalized Laguerre polynomials \(L_{n}^{(\alpha)}(x)\) for \(\alpha>-1\) are defined via the following conditions of orthogonality and normalization \[\int_{0}^{\infty}e^{-x}x^{\alpha}L_{n}^{(\alpha)}(x)L_{m}^{(\alpha)}(x)dx= \Gamma(\alpha+1)\binom{n+\alpha}{n}\delta_{n,m},\] where \(\Gamma\) denotes the Euler gamma function, \(\delta_{i,j}\) is the Kronecker delta and \(n,m=0,1,2,\ldots.\) Moreover, we demand that the coefficient of \(x^{n}\) in the polynomial \(L_{n}^{(\alpha)}(x)\) of degree \(n\) have the sign \((-1)^{n}\). One can figure out the explicit representation of these polynomials, namely, \[L_{n}^{(\alpha)}(x)=\sum_{j=0}^{n}\binom{n+\alpha}{n-j}\frac{(-x)^{j}}{j!}.\] Hence, we have that \[L_{0}^{(\alpha)}(x) =1,\] \[L_{1}^{(\alpha)}(x) =-x+(\alpha+1),\] \[L_{2}^{(\alpha)}(x) =\frac{x^{2}}{2}-(\alpha+2)x+\frac{(\alpha+1)(\alpha+2)}{2},\] \[L_{3}^{(\alpha)}(x) =\frac{-x^{3}}{6}+\frac{(\alpha+3)x^{2}}{2}-\frac{(\alpha+2)( \alpha+3)x}{2}+\frac{(\alpha+1)(\alpha+2)(\alpha+3)}{6},\] and so on. It is well-known that if \(\alpha\) is non-negative, then \(L_{n}^{(\alpha)}(x)\) has exactly \(n\) positive real roots. For more information about both the Hermite polynomials and the Laguerre polynomials we encourage the reader to see [48]. Finally, instead of repeating the text from Introduction related to the Laguerre inequalities, we just recall that for an arbitrary sequence \(\omega=\left(\omega_{i}\right)_{i=0}^{\infty}\) of real numbers the Laguerre inequality of order \(d\) at \(n\) is defined via \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}\omega_{n+j}\omega_{n+2d-j}\geqslant 0.\] In order to deal with this issue for quasi-polynomial-like functions we will need some basic identities involving binomial coefficients, which are omitted here and collected in Section 4. Now, we are ready to proceed to the main part of the manuscript. ## 3. The higher order Turan inequalities for quasi-polynomial-like functions The main goal of this section is to prove the following characterization. **Theorem 3.1**.: _Let \(f(n)\) be a quasi-polynomial-like function of the form_ \[f(n)=c_{l}n^{l}+c_{l-1}n^{l-1}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d}),\] _for some \(1\leqslant d\leqslant l\). Then, for every \(1\leqslant j\leqslant d\) the sequence \((f(n))_{n=0}^{\infty}\) satisfies the order \(j\) Turan inequality for all but finitely many values of \(n\)._ Proof.: At first, let us fix \(0\leqslant j\leqslant d\) and expand \(f(n+j)/f(n)\). We have that \[\frac{f(n+j)}{f(n)} =\frac{c_{l}(n+j)^{l}+c_{l-1}(n+j)^{l-1}+\cdots+c_{l-d}(n+j)^{l-d} +o((n+j)^{l-d})}{c_{l}n^{l}+c_{l-1}n^{l-1}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}\] \[=\frac{c_{l}n^{l}+c_{l-1}n^{l-1}\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}{ c_{l}n^{l}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}\] \[+j\cdot\frac{lc_{l}n^{l-1}+(l-1)c_{l-1}n^{l-2}+\cdots+(l-d)c_{l-d }n^{l-d-1}+o(n^{l-d})}{c_{l}n^{l}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}\] \[+j^{2}\cdot\frac{\binom{l}{2}c_{l}n^{l-2}+\binom{l-1}{2}c_{l-1}n ^{l-3}+\cdots+\binom{l-d}{2}c_{l-d}n^{l-d-2}+o(n^{l-d})}{c_{l}n^{l}+\cdots+c_{ l-d}n^{l-d}+o(n^{l-d})}\] \[\vdots\] \[+j^{d}\cdot\frac{\binom{l}{d}c_{l}n^{l-d}+\binom{l-1}{d}c_{l-1}n ^{l-1-d}+\cdots+\binom{l-d}{d}c_{l-d}n^{l-2d}+o(n^{l-d})}{c_{l}n^{l}+\cdots+c_{ l-d}n^{l-d}+o(n^{l-d})}\] \[+\frac{o((n+j)^{l-d})}{c_{l}n^{l}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}\] \[=\frac{c_{l}n^{l}+c_{l-1}n^{l-1}\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}{ c_{l}n^{l}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}\] \[+j\cdot\frac{1}{n}\frac{lc_{l}n^{l-1}+(l-1)c_{l-1}n^{l-2}+\cdots+ (l-d)c_{l-d}n^{l-d-1}+o(n^{l-d})}{c_{l}n^{l-1}+\cdots+c_{l-d-1}n^{l-d-1}+o(n^{ l-d})}\] \[+j^{2}\cdot\frac{1}{n^{2}}\frac{\binom{l}{2}c_{l}n^{l-2}+\binom{ l-1}{2}c_{l-1}n^{l-3}+\cdots+\binom{l-d}{2}c_{d-d}n^{l-d-2}+o(n^{l-d})}{c_{l}n^{l-2 }+\cdots+c_{l-d}n^{l-d-2}+o(n^{l-d-2})}\] \[\vdots\] \[+j^{d}\cdot\frac{1}{n^{d}}\frac{\binom{l}{d}c_{l}n^{l-d}+\binom{ l-1}{d}c_{l-1}n^{l-1-d}+\cdots+\binom{l-d}{d}c_{l-d}n^{l-2d}+o(n^{l-d})}{c_{l}n^{ l-d}+\cdots+c_{l-2d}n^{l-2d}+o(n^{l-2d})}\] \[+\frac{o((n+j)^{l-d})}{c_{l}n^{l}+\cdots+c_{l-d}n^{l-d}+o(n^{l-d})}.\] Now, it is not difficult to see that we can apply Theorem 2.2 with \(\omega_{n}=f(n)\), \(E(n)=1\), and \(\delta(n)=n^{-1}\). Indeed, we get \[\frac{f(n+j)}{f(n)}=\sum_{i=0}^{d}C_{i}(n)\left(\frac{1}{n}\right)^{i}j^{i}+o \left(\left(\frac{1}{n}\right)^{d}\right)\qquad\text{ as }n\to\infty,\] where \[C_{s}(n)=\frac{\binom{l}{s}c_{l}n^{l-s}+\binom{l-1}{s}c_{l-1}n^{l-1-s}+\cdots +\binom{l-d}{s}c_{l-d}n^{l-d-s}+o(n^{l-d})}{c_{l}n^{l-s}+\cdots+c_{l-d}n^{l-d -s}+o(n^{l-d-s})}\] for any \(0\leqslant s\leqslant d\). Hence, it is clear that \[\lim_{n\to\infty}C_{s}(n)=\binom{l}{s}\] for every \(0\leqslant s\leqslant d\), and we obtain that \[H_{F,m}(x)=m!\sum_{j=0}^{m}(-1)^{m-j}\binom{l}{m-j}\frac{x^{j}} {j!} =(-1)^{m}m!\sum_{j=0}^{m}\binom{m+(l-m)}{m-j}\frac{(-x)^{j}}{j!}\] \[=(-1)^{m}m!L_{m}^{(l-m)}(x)\] for each \(0\leqslant m\leqslant d\), where \(L_{m}^{(l-m)}(x)\) is the generalized Laguerre polynomial. Since \(l-m\geqslant 0\), the polynomials \(L_{m}^{(l-m)}(x)\) have only positive real roots (see, the antepenultimate paragraph of Section 2). Finally, Theorem 2.2 asserts that \[\lim_{n\to\infty}\left(\frac{n^{s}}{f(n)}J_{f}^{s,n}\left(\frac{x}{n}-1\right) \right)=(-1)^{s}s!L_{s}^{(l-s)}(x),\] uniformly for \(x\) in any compact subset of \(\mathbb{R}\) for every \(1\leqslant s\leqslant d\). However, we know that the property of a polynomial with real coefficients is invariant under small deformation. Thus, the required phenomenon for \(f(n)\) follows. Theorem 2.1 and Theorem 3.1 deliver an interesting criterion for the order \(d\) Turan inequality for the \(A\)-partition function. **Theorem 3.2**.: _Let \(A\) be a finite multiset (or set) of positive integers with \(\#A=k\), and let \(1\leqslant d<k\) be fixed. Suppose further that \(\gcd B=1\) for every \((k-d)\)-multisubset \(B\subset A\). Then, for any \(1\leqslant j\leqslant d\) the sequence \((p_{A}(n))_{n=0}^{\infty}\) fulfills the order \(j\) Turan inequality for all sufficiently large values of \(n\)._ Proof.: That is a direct consequence of both Theorem 2.1 and Theorem 3.1. An interesting question arises whether Theorem 3.1 and Theorem 3.2 present also the necessary conditions for order \(d\) Turan inequality for both quasi-polynomial-like functions and \(A\)-partition functions, respectively. It is true for the order \(2\) Turan inequality which follows directly from Gajdzica's papers [26, 27]. However, it is not true in general as the forthcoming examples show. **Example 3.3**.: _Let us investigate the order \(3\) Turan inequality for the function_ \[f(n)=\begin{cases}n^{15}+n^{14}+n^{13}+n^{12}+n^{11}+o(n^{11}),&\text{if }n \equiv 0\,(\mathrm{mod}\ 4),\\ n^{15}+n^{14}+n^{13}+2n^{12}+n^{11}+o(n^{11}),&\text{if }n\equiv 0\,(\mathrm{mod }\ 4).\end{cases}\] _It is easy to see that the assumptions from Theorem 3.1 are not satisfied. Nevertheless, it turns out that the function which directly corresponds to the third order Turan inequality takes the form_ \[4\left(f(n)^{2}-f(n-1)f(n+1)\right)\left(f(n+1)^{2}-f(n)f(n+2)\right)\] \[-(f(n)f(n+1)-f(n-1)f(n+2))^{2}=\begin{cases}12411n^{54}+o(n^{54}),& \text{if }n\equiv 0\,(\mathrm{mod}\ 4),\\ 12771n^{54}+o(n^{54}),&\text{if }n\equiv 1\,(\mathrm{mod}\ 4),\\ 12539n^{54}+o(n^{54}),&\text{if }n\equiv 2\,(\mathrm{mod}\ 4),\\ 12659n^{54}+o(n^{54}),&\text{if }n\equiv 3\,(\mathrm{mod}\ 4);\end{cases}\] _and is positive for all sufficiently large values of \(n\). Hence, we conclude that Theorem 1.3 is not an optimal criterion._ In the case of the \(A\)-partition function, we present the following counterexample. **Example 3.4**.: _Let us assume that \(A=\{1,1,1,\,\text{\scriptsize$1$},300\}\), and examine the order \(4\) Turan inequality. In the fashion of Example 3.3, we wish to define a function \(f(n)\) which directly corresponds to that issue. It is tedious but elementary to show that a sequence \(\omega=\left(\omega_{i}\right)_{i=0}^{\infty}\) fulfills the fourth Turan inequality if the following_ \[54\omega_{n}\omega_{n+1}^{2}\omega_{n+2}\omega_{n+4}^{2}+\omega_ {n}^{3}\omega_{n+4}^{3}+108\omega_{n+1}^{3}\omega_{n+2}\omega_{n+3}\omega_{n+ 4}+36\omega_{n+1}^{2}\omega_{n+2}^{2}\omega_{n+3}^{2}\] \[-12\omega_{n}^{2}\omega_{n+1}\omega_{n+3}\omega_{n+4}^{2}-54 \omega_{n+1}^{2}\omega_{n+2}^{3}\omega_{n+4}-6\omega_{n}\omega_{n+1}^{2} \omega_{n+3}^{2}\omega_{n+4}-54\omega_{n}\omega_{n+2}^{3}\omega_{n+3}^{2}\] \[-27\omega_{n+1}^{4}\omega_{n+4}^{2}-180\omega_{n}\omega_{n+1} \omega_{n+2}^{2}\omega_{n+3}\omega_{n+4}-27\omega_{n}^{2}\omega_{n}^{4}+3108 \omega_{n}\omega_{n+1}\omega_{n+2}\omega_{n+3}^{3}\] \[-64\omega_{n+1}^{3}\omega_{n+3}^{3}-18\omega_{n}^{2}\omega_{n+2}^{2 }\omega_{n+4}^{2}+81\omega_{n}\omega_{n+2}^{4}\omega_{n+4}+54\omega_{n}^{2} \omega_{n+2}\omega_{n+3}^{2}\omega_{n+4}\geqslant 0\] _is satisfied for every \(n\in\mathbb{N}\). Therefore, let us put \(\omega_{n}:=p_{A}(n)\) and denote the left hand side of the above inequality by \(f_{A}(n)\). Since \(\gcd(300)\neq 1\), we see that the assumptions from Theorem 3.2 do not hold if \(d=4\). Notwithstanding, one can carry out the appropriate computations in Mathematica [54] and check that \(f_{A}(n)\) is a quasi-polynomial of degree \(12\) with coefficients depending on \(n\,(\mathrm{mod}\;300)\). It might be also verified that the leading coefficient of \(f_{A}(n)\) attains the smallest value whenever \(n\not\equiv 296\,(\mathrm{mod}\;300)\) -- in all of these cases, we have_ \[f_{A}(n)=\frac{n^{12}}{2^{18}\cdot 3^{9}\cdot 5^{12}}+o\left(n^{12}\right).\] _The above discussion agrees with the plot of \(f_{A}(n)\) for \(1\leqslant n\leqslant 10^{4}\), see Figure 1._ _Hence, we conclude that Theorem 3.2 is not optimal, as well._ At the end of this section, let us exhibit two other examples. Sometimes we can not conclude the appropriate order \(d\) Turan inequality if the requirements from Theorem 3.1 or Theorem 3.2 are not satisfied. **Example 3.5**.: _Let us consider a quasi-polynomial-like function of the form_ \[f(n)=\begin{cases}n^{15}+n^{14}+n^{13}+n^{12}+n^{11}+o(n^{11}),&\text{if }n \not\equiv 0\,(\mathrm{mod}\;4),\\ n^{15}+n^{14}+n^{13}+500n^{12}+n^{11}+o(n^{11}),&\text{if }n\equiv 0\,( \mathrm{mod}\;4).\end{cases}\] _We would like to investigate the order \(3\) Turan inequality. However, it is clear that the assumptions from Theorem 3.1 are not satisfied; and one may calculate that_ \[4\left(f(n)^{2}-f(n-1)f(n+1)\right)\left(f(n+1)^{2}-f(n)f(n+2)\right)\] \[-\left(f(n)f(n+1)-f(n-1)f(n+2)\right)^{2}=-266341n^{54}+o(n^{54}),\] _whenever \(n\equiv 2\,(\mathrm{mod}\;4)\). Hence, \(f(n)\) can not satisfy the order \(3\) Turan inequality for all sufficiently large values of \(n\), as required._ As an instance for an \(A\)-partition function, we take a finite analogue of the partition function \(p(n)\). **Example 3.6**.: _For any positive integer \(m\), let us put \(A_{m}:=\{1,2,\ldots,m\}\). We want to consider the third order Turan inequality for \(p_{A_{6}}(n)\) and \(p_{A_{7}}(n)\). In order to make the text more transparent, we set_ \[g_{A_{m}}(n) :=4\left(p_{A_{m}}^{2}(n)-p_{A_{m}}(n-1)p_{A_{m}}(n+1)\right) \left(p_{A_{m}}^{2}(n+1)-p_{A_{m}}(n)p_{A_{m}}(n+2)\right)\] \[\quad-\left(p_{A_{m}}(n)p_{A_{m}}(n+1)-p_{A_{m}}(n-1)p_{A_{m}}(n+2 )\right)^{2}\] _Thus, \(g_{A_{m}}(n)\) directly corresponds to the order \(3\) Turan inequality. It is clear that the demands from Theorem 3.2 are not true for \(A_{6}\). In fact, it turns out that, for instance,_ \[g_{A_{6}}(n)=-\frac{2069n^{14}}{2^{24}\cdot 3^{12}\cdot 5^{6}}+o\left(n^{14} \right),\] _whenever \(n\equiv 2\,(\mathrm{mod}\ 60)\). On the other hand, one can check that the equality_ \[g_{A_{7}}(n)=\frac{n^{18}}{2^{28}\cdot 3^{14}\cdot 5^{7}\cdot 7^{4}}+o\left(n^{18}\right)\] _is valid for every positive integer \(n\), as required. The above discussion agrees with the plots of \(g_{A_{6}}(n)\) and \(g_{A_{7}}(n)\), see Figure 3 and Figure 3, respectively._ ## 4. The Laguerre inequalities for quasi-polynomial-like functions Now, we focus on the Laguerre inequalities for quasi-polynomial-like functions. As it was mentioned at the end of Section 2, we need to use a few binomial coefficient identities to deal with the issue. The first of them arises from comparing the coefficients of the expansions of both \((1-z)^{s}(1+z)^{s}\) and \((1-z^{2})^{s}\) (see, [29, Section 5.4]). **Lemma 4.1**.: _Let \(s\in\mathbb{N}\) be fixed. Then for every even integer \(0\leqslant n\leqslant s\), we have_ \[\sum_{j=0}^{n}(-1)^{j}\binom{s}{j}\binom{s}{n-j}=(-1)^{\frac{n}{2}}\binom{s}{ n/2}.\] To present the second one, we need to recall that the Stirling number of the second kind \(\genfrac{\{}{\}}{0.0pt}{}{n}{k}\) enumerates the number of ways to partition a set of \(n\) labelled objects into \(k\) non-empty unlabelled subsets. Equivalently, it is the number of different equivalence relations with exactly \(k\) equivalence classes that may be defined on an \(n\) element set. It is worth noting that the following identities \[\genfrac{\{}{\}}{0.0pt}{}{n}{1}=1,\quad\genfrac{\{}{\}}{0.0pt}{}{n}{n}=1, \quad\genfrac{\{}{\}}{0.0pt}{}{n}{0}=0\quad\text{and}\quad\genfrac{\{}{\}}{0}{0} {0}=1\] hold for every positive integer \(n\) as well as \(\genfrac{\{}{\}}{0.0pt}{}{m}{k}=0\) whenever \(0\leqslant m<k\). The succeeding lemma, together with the general introduction to the Stirling numbers, might be found in [29, Section 6.1]. **Lemma 4.2**.: _Let \(u\) and \(v\) be arbitrary non-negative integers. Then, we have_ \[u!\genfrac{\{}{\}}{0.0pt}{}{v}{u}=\sum_{k=0}^{u}(-1)^{u-k}\binom{u}{k}k^{v}.\] Now, we are ready to state and prove the main result of this section. **Theorem 4.3**.: _Let \(f(n)\) be a quasi-polynomial-like function of the form_ \[f(n)=c_{l}n^{l}+c_{l-1}n^{l-1}+\cdots+c_{l-2d}n^{l-2d}+o(n^{l-2d}),\] _for some non-negative integer \(d\) such that \(2d\leqslant l\). Then, for every \(0\leqslant j\leqslant d\) the sequence \((f(n))_{n=0}^{\infty}\) satisfies the Laguerre inequality of order \(j\) for all but finitely many values of \(n\). In particular, we have that_ \[\sum_{i=0}^{2d}(-1)^{i+d}\binom{2d}{i}f(n+i)f(n+2d-i)=(2d)!\genfrac{\{}{\}}{0.0 pt}{}{l}c_{l}^{2}n^{2(l-d)}+o\left(n^{2(l-d)}\right).\] Proof.: Let us fix a quasi-polynomial-like function \(f(x)\) as in the statement, and expand the left hand side of the inequality (1.4) with \(\omega_{n}=f(n)\). We have that \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}f(n+j)f(n+2d-j)\] \[= \sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}\left[c_{l}(n+j)^{l}+\cdots +c_{l-2d}(n+j)^{l-2d}+o\left((n+j)^{l-2d}\right)\right]\] \[\times\left[c_{l}(n+2d-j)^{l}+\cdots+c_{l-2d}(n+2d-j)^{l-2d}+o \left((n+2d-j)^{l-2d}\right)\right]\] \[= \sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}\left[c_{l}\sum_{i=0}^{2d} \binom{l}{i}j^{i}n^{l-i}+\cdots+c_{l-2d}n^{l-2d}+o(n^{l-2d})\right]\] \[\times\left[c_{l}\sum_{i=0}^{2d}(-1)^{i}\binom{l}{i}j^{i}(n+2d)^{ l-i}+\cdots+c_{l-2d}(n+2d)^{l-2d}+o(n^{l-2d})\right].\] Since we are interested in the asymptotic behavior of the above expression, we need to determine the leading coefficient of its polynomial part. It is not difficult to notice that whenever we multiply a summand \(\gamma_{i_{0},k_{0}}j^{k_{0}}n^{l-i_{0}}\) from the first square bracket with a summand \(\delta_{i_{1},k_{1}}j^{k_{1}}(n+2d)^{l-i_{1}}\) from the second one (where the coefficients \(\gamma_{i_{0},k_{0}}\) and \(\delta_{i_{1},k_{1}}\) are independent of both \(j\) and \(n\)), we can obtain at most \(j\) to the power \(i_{0}+i_{1}\geqslant k_{0}+k_{1}\). More precisely, we get an expression of the form \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}\gamma_{i_{0},k_{0}}\delta_{i_{1},k_{1} }n^{l-i_{0}}(n+2d)^{l-i_{1}}j^{k_{0}+k_{1}}, \tag{4.1}\] where \(0\leqslant k_{0}+k_{1}\leqslant i_{0}+i_{1}\). Therefore if \(i_{0}+i_{1}<2d\), then (4.1) might be rewritten as \[(-1)^{d}\gamma_{i_{0},k_{0}}\delta_{i_{1},k_{1}}n^{l-i_{0}}(n+2d)^{l-i_{1}} \sum_{j=0}^{2d}(-1)^{2d-j}\binom{2d}{j}j^{k_{0}+k_{1}}=0,\] where the equality follows from Lemma 4.2 with \(k=j\), \(u=2d\) and \(v=k_{0}+k_{1}\). Hence, our task boils down to finding the coefficient of \(n^{2(l-d)}\). Repeating the above discussion, one can observe that the only possible non-zero term takes the form \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}c_{l}^{2}\sum_{i=0}^{2d}(-1)^{i }\binom{l}{i}\binom{l}{2d-i}j^{2d}n^{2(l-d)}\] \[=(-1)^{d}c_{l}^{2}\times\sum_{i=0}^{2d}(-1)^{i}\binom{l}{i}\binom{l }{2d-i}\times\sum_{j=0}^{2d}(-1)^{2d-j}\binom{2d}{j}j^{2d}.\] Now, Lemma 4.1 asserts that \[\sum_{i=0}^{2d}(-1)^{i}\binom{l}{i}\binom{l}{2d-i}=(-1)^{d}\binom{l}{d}.\] On the other hand, Lemma 4.2 maintains that \[\sum_{j=0}^{2d}(-1)^{2d-j}\binom{2d}{j}j^{2d}=(2d)!\left\{\begin{matrix}2d\\ 2d\end{matrix}\right\}=(2d)!.\] In conclusion, we obtain that \[\sum_{j=0}^{2d}(-1)^{j+d}\binom{2d}{j}f(n+j)f(n+2d-j)=(2d)!\left(\begin{matrix} l\\ d\end{matrix}\right)c_{l}^{2}n^{2(l-d)}+o\left(n^{2(l-d)}\right),\] which was to be demonstrated. As an immediate consequence of Theorem 4.3, we get an analogue characterization to that one from Theorem 3.2. **Theorem 4.4**.: _Let \(A\) be a finite multiset (or set) of positive integers with \(\#A=k\), and let \(1\leqslant 2d<k\) be fixed. Suppose further that \(\gcd B=1\) for every \((k-2d)\)-multisubset \(B\subset A\). Then, for each \(1\leqslant j\leqslant d\) the sequence \((p_{A}(n))_{n=0}^{\infty}\) satisfies the Laguerre inequality of order \(j\) for all but finitely many values of \(n\)._ Proof.: The criterion easily follows from both Theorem 2.1 and Theorem 4.3. Analogously to Section 3, we present a few examples showing that Theorem 4.3, as well as Theorem 4.4, does not deliver us a necessary condition for the Laguerre inequality of order \(d\) for \(d\geqslant 2\). **Example 4.5**.: _Let us assume that \(f(n)\) is a quasi-polynomial-like function of the form_ \[f(n)=n^{10}+n^{9}+n^{8}+n^{7}+(n\:(\mathrm{mod}\:5))\cdot n^{6}+o(n^{6}).\] _It is not difficult to see that the assumption from Theorem 4.3 for \(d=2\) does not hold. Nevertheless, one can calculate that_ \[3f(n+2)^{2}-4f(n+1)f(n+3)+f(n)f(n+4)\geqslant 525n^{16}+o(n^{16})\] _for all sufficiently large values of \(n\), and observe that the second Laguerre inequality is asymptotically satisfied for \(f(n)\). Thus, Theorem 4.3 is not an optimal criterion._ As a counterexample for Theorem 4.4, we exhibit the following. **Example 4.6**.: _We put \(A=\{1,1,1,\downarrow 1,300\}\) and consider the order \(2\) Laguerre inequality. It is clear that the assumptions from Theorem 4.4 are not satisfied for \(d=2\). Notwithstanding, if we set_ \[h_{A}(n):=3p_{A}^{2}(n+2)-4p_{A}(n+1)p_{A}(n+3)+p_{A}(n)p_{A}(n+4),\] _then it turns out that_ \[h_{A}(n)\geqslant\frac{n^{4}}{2^{3}\cdot 3^{3}\cdot 5^{4}}+o(n^{4})\] _with the equality whenever \(n\not\equiv 297\,(\mathrm{mod}\ 300)\), which agrees with Figure 4._ _In conclusion, we see that Theorem 4.4 is not an optimal criterion, as well._ At the end of this section, we present an example showing that, in general, it might be difficult to derive an optimal criterion for the Laguerre inequality of order \(d\geqslant 2\) for quasi-polynomial-like functions. **Example 4.7**.: _Let us consider the \(A_{m}\)-partition function defined in Example 3.6. For instance, we may examine the Laguerre inequality of order \(2\) for both \(p_{A_{8}}(n)\) and \(p_{A_{9}}(n)\). For the sake of clarity, let us put_ \[h_{A_{m}}(n):=3p_{A_{m}}^{2}(n+2)-4p_{A_{m}}(n+1)p_{A_{m}}(n+3)+p_{A_{m}}(n)p_{ A_{m}}(n+4).\] _In other words, \(h_{A_{m}}(n)\) corresponds to the second order Laguerre inequality for \(p_{A_{m}}(n)\). We see that the assumptions from Theorem 4.4 do not hold for \(p_{A_{8}}(n)\) and \(d=2\). Moreover, one can determine that_ \[h_{A_{8}}(n)=-\frac{349n^{10}}{2^{20}\cdot 3^{6}\cdot 5^{4}\cdot 7^{3}}+o\left(n ^{10}\right),\] _whenever \(n\equiv 0,2,\ldots,838\,(\mathrm{mod}\ 840)\). In the case of \(h_{A_{9}}(n)\), on the other hand, we get that the equality_ \[h_{A_{9}}(n)=\frac{n^{12}}{2^{24}\cdot 3^{11}\cdot 5^{4}\cdot 7^{3}}+o\left(n ^{12}\right)\] _holds for each positive integer \(n\), which agrees with Theorem 4.4. Figure 5 and Figure 6 exhibit the plots of \(h_{A_{8}}(n)\) and \(h_{A_{9}}(n)\), respectively._ _The above discussion asserts that it might be difficult to find out an easy description of all quasi-polynomial-like functions which (asymptotically) fulfill the Laguerre inequality of order \(d\) for any \(d\geqslant 2\)._ Figure 4. The values of \(h_{A}(n)\) for \(A=\{1,1,1,1,300\}\) and \(1\leqslant n\leqslant 10^{4}\). ## 5. Concluding remarks It is quite unfortunate that neither Theorem 3.1 nor Theorem 3.2 delivers necessary conditions for the order \(d\) Turan inequality. Analogously, neither Theorem 4.3 nor Theorem 4.4 contains necessary conditions for the Laguerre inequality of order \(d\). It is worth pointing out that we have such a result in the case of the \(r\)-log-concavity problem for quasi-polynomial-like functions (and, in particular, \(A\)-partition functions) [27]. Recall that a sequence of real numbers \(\omega=\left(w_{i}\right)_{i=0}^{\infty}\) is called (asymptotically) \(r\)-log-concave for \(r\in\mathbb{N}_{+}\), if there exists an integer \(N\) such that all terms of the sequences \[\widehat{\mathcal{L}}\omega,\widehat{\mathcal{L}}^{2}\omega,\ldots,\widehat{ \mathcal{L}}^{r}\omega\] are positive for every \(i\geqslant N\), where \[\widehat{\mathcal{L}}\omega=\left(w_{i+1}^{2}-w_{i}w_{i+2}\right)_{i=0}^{ \infty}\text{ and }\widehat{\mathcal{L}}^{k}\omega=\widehat{\mathcal{L}}\left(\widehat{ \mathcal{L}}^{k-1}\omega\right)\] for \(k\in\{2,3,\ldots,r\}\). We have the following characterization for that issue. **Theorem 5.1** (Gajdzica).: _Let \(l\) and \(r\) be arbitrary positive integers such that \(l\geqslant 2r\). Suppose further that we have_ \[f(n)=a_{l}(n)n^{l}+a_{l-1}(n)n^{l-1}+\cdots+a_{l-2r}(n)n^{l-2r}+o\left(n^{l-2r }\right),\] _where the coefficients \(a_{l-2r}(n),\ldots,a_{l}(n)\) might depend on the residue class of \(n\left(\mathrm{mod}\ M\right)\) for some positive integer \(M\geqslant 2\). Then the sequence \(\left(f(n)\right)_{n=0}^{\infty}\) is asymptotically \(r\)-log-concave if and only if all the numbers \(a_{l-2r}(n),\ldots,a_{l}(n)\) are independent of the residue class of \(n\left(\mathrm{mod}\ M\right)\)._ Unfortunately, the analogous descriptions are impossible for both the higher order Turan inequalities and the Laguerre inequalities. For the former, if we assume that \[f(n)=c_{l}n^{l}+c_{l-1}n^{l-1}+\cdots+c_{l-d+1}n^{l-d+1}+o(n^{l-d+1})\] for some \(1\leqslant d\leqslant l\). Then, the leading coefficient of the (quasi-polynomial-like) function which corresponds to the \(d\)th Turan inequality may heavily depend on the residue class of \(n\) modulo some positive integer, see Example 3.4. To visualize the issue more accurately, let us consider the third order Turan inequality for \[f(n)=c_{l}n^{l}+c_{l-1}n^{l-1}+c_{l-2}n^{l-2}+c_{l-3}(n)n^{l-3}+o(n^{l-3}),\] where \(l\geqslant 3\) and \(c_{l-3}(n)\) depends on the residue class of \(n\pmod{M}\) for some \(M\geqslant 2\). It is tiresome but elementary to show that we have the following: \[4\left(f(n)^{2}-f(n-1)f(n+1)\right)\left(f(n+1)^{2}-f(n)f(n+2) \right)-\left(f(n)f(n+1)-f(n-1)f(n+2)\right)^{2}\] \[=\big{[}-6c_{l-3}(n-1)c_{l-3}(n+1)+2c_{l-3}(n-1)c_{l-3}(n+2)+4lc_ {lcl-3}(n-1)-c_{l-3}^{2}(n-1)\] \[\quad+6c_{l-3}(n-1)c_{l-3}(n)+6c_{l-3}(n+1)c_{l-3}(n+2)+12lc_{l}c_ {l-3}(n+1)-9c_{l-3}^{2}(n+1)\] \[\quad+18c_{l-3}(n)c_{l-3}(n+1)-4lc_{l}c_{l-3}(n+2)-c_{l-3}^{2}(n+ 2)-6c_{l-3}(n)c_{l-3}(n+2)\] \[\quad+4l^{3}c_{l}^{2}-4l^{2}c_{l}^{2}-12lc_{l}c_{l-3}(n)-9c_{l-3} ^{2}(n)\big{]}c_{l}^{2}n^{4l-6}+o(n^{4l-6}).\] Thus, it is easy to see that the leading coefficient of that expression intensely depends on the residue class of \(n\pmod{M}\), which coincides with our discussion above. One can demonstrate a parallel reasoning to deduce that the same problem plagues us if we deal with the Laguerre inequality of order \(d\) for \(d\geqslant 2\), which Example 4.6 indicates. At the end of the manuscript, we state a few open problems. The first of them encourages us to deal with the higher order Turan inequalities for some particular \(A\)-partition functions. **Problem 5.2**.: _Fix a set (or multiset) \(A\) of positive integers and investigate the higher order Turan inequalities for the \(A\)-partition function._ **Remark 5.3**.: For instance, if we set \(A=A_{m}=\{1,2,\ldots,m\}\) in Problem 5.2, then it is known [26] that the second order Turan inequality for \(p_{A_{m}}(n)\) begins to hold for \(m=5\). In that case we have that \[p_{A_{5}}^{2}(n)\geqslant p_{A_{5}}(n-1)p_{A_{5}}(n+1)\] for all \(n>37\). One can extend the above and deal with the problem for the \(A_{m}^{(l)}\)-partition function, where \(A_{m}^{(l)}=\{1^{l},2^{l},\ldots,m^{l}\}\) and \(m\in\mathbb{N}_{+}\cup\{\infty\}\). The case of \(m=\infty\) and \(l=1\) has been investigated by Griffin et al. [28] and Larson and Wagner [35]. For more information about the general setting (when \(m=\infty\)), we refer the reader to Ulas' paper [50]. We also hope that there is a chance to discover a more efficient criterion than Theorem 3.1, and state the following. **Problem 5.4**.: _Let \(d\geqslant 3\) be arbitrary. Find a more effective criterion than Theorem 3.1 (or Theorem 3.2) for the order \(d\) Turan inequality for quasi-polynomial-like functions. Alternatively, do that for small values of the parameter \(d\)._ Finally, we formulate the analogues of both Problem 5.2 and Problem 5.4 in the context of Laguerre inequalities for quasi-polynomial-like functions. **Problem 5.5**.: _Fix a set (or multiset) \(A\) of positive integers and investigate the Laguerre inequalities for the \(A\)-partition function._ **Problem 5.6**.: _Let \(d\geqslant 2\) be arbitrary. Derive a more efficient criterion than Theorem 4.3 (or Theorem 4.4) for the \(d\)th Laguerre inequality for quasi-polynomial-like functions. Alternatively, do that for small values of the parameter \(d\)._ ## Acknowledgements I wish to express my sincere thanks to Piotr Miska and Maciej Ulas for their time and helpful suggestions. I am also grateful to Ian Wagner for his additional comments. This research was funded by both a grant of the National Science Centre (NCN), Poland, no. UMO-2019/34/E/ST1/00094 and a grant from the Faculty of Mathematics and Computer Science under the Strategic Program Excellence Initiative at the Jagiellonian University in Krakow.

2305.10173

Quantum theory without the Axiom of choice, and Lefschetz Quantum Physics

In this conceptual paper, we discuss quantum formalisms which do not use the famous Axiom of Choice. We also consider the fundamental problem which addresses the (in)correctness of having the complex numbers as the base field for Hilbert spaces in the K{\o}benhavn interpretation of quantum theory, and propose a new approach to this problem (based on the Lefschetz principle). Rather than a Theorem--Proof--paper, this paper describes two new research programs on the foundational level, and focuses on fundamental open questions in these programs which come along the way.

Koen Thas

2023-05-17T12:57:19Z

http://arxiv.org/abs/2305.10173v1

# Quantum theory without the axiom of choice, and Lefschetz quantum physics ###### Abstract. In this conceptual paper, we discuss quantum formalisms which do not use the famous Axiom of Choice. We also consider the fundamental problem which addresses the (in)correctness of having the complex numbers as the base field for Hilbert spaces in the Kobenhavn interpretation of quantum theory, and propose a new approach to this problem (based on the Lefschetz principle). Rather than a Theorem-Proof-paper, this paper describes two new research programs on the foundational level, and focuses on fundamental open questions in these programs which come along the way. ###### Contents * 1 Introduction * 2 Axiom of Choice: mathematical discussion * 3 Kobenhavn interpretations beyond \(\mathbb{C}\) * 4 Quantum Lefschetz Principle A * 5 Automorphisms of \(\mathbb{C}\) and codes * 6 Eigenvalues, eigenvectors and probabilities * 6.1 The \(\mathbb{C}\)-algebra \(\mathbb{C}\) and the \(\mathbb{C}\)-algebra \(\mathbb{C}\) [MISSING_PAGE_POST] This very discussion certainly is one of the leading threads in this paper, and motivates us to introduce a conjectural Lefschetz quantum principle below. Secondly, _if_ we assume to work over \(\mathbb{C}\), then often implicitly the infamous Axiom of Choice is used in the mathematical machinery used to describe quantum physics. But since the philosophical discussion underlying quantum physics -- including its many famous thought experiments -- is extremely important for the theory, and since mathematics -- even basic linear algebra -- is so different without the Axiom of Choice, we want to investigate what happens to quantum theory if one does not rely on the Axiom of Choice. We refer to the essay [3] for a deeper philosophical discussion on this matter. ### Plan of the present paper In section 2, we will mathematically discuss the Axiom of Choice. In section 3, we tersely describe the author's approach to general quantum theories (over division rings) and we pay particular attention to finite fields and the minimal model. In section 4, we develop a conjectural Lefschetz principle for quantum theory, and to that end we first mathematically discuss the classical Lefschetz principle. The next section is the short section 5, in which we develop some basic mechanisms to develop quantum codes in models with the Axiom of Choice. In the final section, we discuss the impact of not accepting the Axiom of Choice on measurements and probabilities, and devise a number of thought experiments. ### Acknowledgments The author wishes to thank Karl Svozil and Bill Wootters for a number of interesting and helpful communications. ## 2. Axiom of Choice: mathematical discussion We start this section with a first formal formulation of the Axiom of Choice (AC): "For every indexed family \(\left(S_{i}\right)_{i\in I}\) of nonempty sets, there exists an indexed family \(\left(x_{i}\right)_{i\in I}\) such that \(x_{j}\in S_{j}\) for every \(j\in I\)." Let us look at a first illuminating example. Suppose each \(S_{i}\) is a subset of the positive integers \(\mathbb{N}\); then we could define \(x_{i}\) as the smallest number in \(S_{i}\). The function which assigns to each \(S_{j}\) the element \(x_{j}\) is called a _choice function_. In this case, we do not need to invoke the Axiom of Choice in order to fulfil the desired property. But in case we define \(\left\{S_{i}\mid i\in I\right\}\) as the set of nonempty subsets of the reals \(\mathbb{R}\), no such a choice function is known. ### Russel's socks Bertrand Russel gave a particularly entertaining example where one has to invoke the Axiom of Choice in order to define a choice function. Suppose one has an infinite collection of pairs of socks, where we assume that in one pair of socks there is no way to distinguish between the socks. Then in order to select one sock from each pair, we need to invoke AC. Note that if we had started from pairs of shoes instead of socks, we _could_ have defined a choice function ("select the left shoe"). ### Choice functions and Cartesian products In terms of choice functions, we can formulate AC as follows: "For any set \(\mathcal{X}\) of nonempty sets, there exists a choice function \(f\) defined on \(\mathcal{X}\) which maps each set of \(\mathcal{X}\) to an element of that set." Obviously, each such choice function \(f\) defines an element of the Cartesian product of the sets in \(\mathcal{X}\), so that we can give another equivalent statement: "For any set \(\mathcal{X}\) of nonempty sets, the Cartesian product of the sets in \(\mathcal{X}\) is nonempty." ### Vector spaces The equivalent formulation which interests us the most in the context of the present paper, is the following. "Every vector space has a basis." Obviously, in the context of quantum theory such results are extremely important! Note that upon accepting the Axiom of Choice, one can show that the size of a basis of a given vector space is independent of the choice of basis, and this size yields a notion of _dimension_. ## 3. Kobenhavn interpretations beyond \(\mathbb{C}\) In classical quantum theory following the Kobenhavn interpretation -- in some papers called "Actual Quantum Theory" (AQT) -- the state space is a Hilbert space (foreseen with the standard inner product). More precisely: * a physical quantum system is represented by a Hilbert space \(\mathcal{H}=\Big{(}(\mathbb{C}^{\omega},+,\cdot),\langle\cdot,\cdot\rangle \Big{)}\), with \(\langle\cdot,\cdot\rangle\) the standard inner product and \(\omega\) allowed to be non-finite; * the standard inner product \(\langle\cdot,\cdot\rangle\) sends \(\Big{(}(x_{1},\ldots,x_{\omega}),(y_{1},\ldots,y_{\omega})\Big{)}\) to \(\overline{x_{1}}y_{1}+\cdots+\overline{x_{\omega}}y_{\omega}\) (or \(x_{1}\overline{y_{1}}+\cdots+x_{\omega}\overline{y_{\omega}}\)), where \(\overline{c}\) is the complex conjugate of \(c\in\mathbb{C}\); complex conjugation is an involutary automorphism of the field \(\mathbb{C}\); * up to complex scalars, pure states (wave functions) are represented by nonzero vectors in \(\mathbb{C}^{\omega}\); usually, one considers normalized vectors; * time evolution operators are represented by linear operators of \(\mathbb{C}^{\omega}\) that preserve \(\langle\cdot,\cdot\rangle\), that is, _unitary operators_. If \(\omega\) is finite, unitary operators correspond to nonsingular complex \((\omega\times\omega)\)-matrices \(U\) such that \(UU^{*}=\mathrm{id}\); * measuring an observable \(A\) in a system described by the wave function \(|\psi\rangle\), amounts to collapsing \(|\psi\rangle\) into one of the orthogonal eigenvectors \(|\psi_{i}\rangle\) of the Hermitian operator \(A\), yielding as measurement the corresponding eigenvalue \(\lambda_{i}\); * composite product states correspond to tensor products \(|\psi_{1}\rangle\otimes|\psi_{2}\rangle\in\mathcal{H}_{1}\otimes\mathcal{H}_{2}\); if a state in \(\mathcal{H}_{1}\otimes\mathcal{H}_{2}\) is not a product state, it is entangled; * one follows Born's rule, which says that \(|\langle\psi,\psi_{i}\rangle|^{2}\) is the probability that the measurement \(\lambda_{i}\) will be made; * (\(\cdots\)). In the rest of this section we tersely explain our approach of [21], which unifies all known modal quantum theories (over finite fields \(\mathbb{F}_{q}\), algebraically closed fields, general division rings with involution). ### The general setting A _division ring_ is a field for which the multiplication is not necessarily commutative. Sometimes they are called "skew fields," but we prefer the name division ring. In [21] we described a general Kobenhavn approach in which all known classical and modal quantum theories are united in one and the same framework. The main philosophy is that instead of the field of complex number or finite fields, the underlying coordinatizing structures are generalized to division rings, so that we consider generalized Hilbert spaces over division rings instead of complex Hilbert spaces or their finite field analogons. Of course, one has to have a good alternative for the classical inproducts, and this is perfectly possible if we foresee the division rings with a so-called _involution_ (cf. the next subsection). The details can be found in the next subsection. ### Standard \((\sigma,1)\)-Hermitian forms A "division ring with involution" is a division ring with an involutory anti-automorphism. If \(k\) is a division ring with involution \(\sigma\), the _standard \((\sigma,1)\)-Hermitian form_ on the right vector space \(V(d,k)\), is given by \[\Big{\langle}x,y\Big{\rangle}\ :=\ x_{1}^{\sigma}y_{1}+\cdots+x_{d}^{\sigma}y_{d}, \tag{2}\] where \(x=(x_{1},\ldots,x_{d})\) and \(y=(y_{1},\ldots,y_{d})\). **Remark 3.1**.: In the case that \(\sigma={\rm id}\), we obtain a form which is usually called _symmetric_; it is not a proper Hermitian form, but still comes in handy in some situations (for example in cases of field reduction: "real Hilbert spaces" have often been considered in quantum theory; see e.g. the work of Wootters et al. [24, 25]). We propose to describe all classical and modal quantum theories in one and the same framework, under the umbrella of "General Quantum Theories" (GQTs), as follows: \[\left\{\begin{array}{l}\mbox{From now on, we propose to depict a physical quantum system in a general Hilbert space}\\ \mbox{\Large$\frac{\mathcal{H}=\left((V(\omega,k),+,\cdot),\langle\cdot,\cdot \rangle\right)$}{\mbox{form.}}$}\end{array}\right\}\] Following [21], we speak of a _standard GQT_ if given an involution \(\sigma\), the general Hilbert space comes with the standard \((\sigma,1)\)-Hermitian form. As some fields such as the real numbers and the rational numbers do not admit nontrivial involutions, they only can describe "improper" quantum systems. ### Algebraically closed fields Let \(k\) be any algebraically closed field in characteristic \(0\). It is well known that upon accepting the Axiom of Choice, there exists an involution \(\gamma\) in \({\rm Aut}(k)^{\times}\), where \({\rm Aut}(k)\) denotes the automorphism group of \(k\). Now consider the set \[k_{\gamma}:=\{\kappa\in k\mid\kappa^{\gamma}=\kappa\}. \tag{3}\] One easily shows that \(k_{\gamma}\), endowed with the addition and multiplication coming from \(k\), is also a field. There is however more by [2]: **Theorem 3.2** (\((\mathbb{C},\mathbb{R})\)-Analogy -- algebraically closed version in char. \(0\)).: _Let \(k\) be any algebraically closed field in characteristic \(0\). Let \(\gamma\) be an involution in \({\rm Aut}(k)^{\times}\). Then \(-1\) is not a square in \(k_{\gamma}\). Suppose \(i\in k\) is such that \(i^{2}=-1\). Then \(k=k_{\gamma}+i\cdot k_{\gamma}\) and \([k:k_{\gamma}]=2\)._ So each element of \(k\) has a unique representation as \(a+bi\), with \(a,b\in k_{\gamma}\) and \(i\) a fixed solution of \(x^{2}=-1\). Fields which have index \(2\) in their algebraic closure are called _real-closed fields_, and can always be constructed as a \(k_{\gamma}\) of some involution \(\gamma\). Real-closed fields share many properties with the reals \(\mathbb{R}\): each such field is _elementarily equivalent_ to the reals, which by definition means that it has the same first-order properties as the reals. We call a GQT _complex-like_ if it is defined over an algebraically closed field \(k\) with nontrivial involution \(\gamma\), where the elements of \(k\) are represented in Theorem 3.2 with respect to the field \(k_{\gamma}\). **Remark 3.3**.: The analogy goes even further: once we have defined \(k_{\sigma}\) as above, and we represent each element \(x\) in \(k\) as \(x=u+iv\) (which can be done by Theorem 3.2), it can be shown that the automorphism \(\sigma\) is given by \[\sigma:\;k\mapsto k:\;u+iv\mapsto u-iv\;\;\;\mbox{(complex conjugation)}. \tag{4}\] ### Extension of quantum theories If we consider a GQT \(\mathcal{T}\) over a field \(k\) in characteristic \(0\), the following fundamental question arises: Is \(\mathcal{T}\)_embeddable_ in a complex-like theory (or in any other GQT, for that matter)? Here, the notion of "embeddable" is obvious: if \(k\) comes with the involution \(\gamma\), we want to have a field extension \(\ell\Big{/}k\) for which \(\ell\) is algebraically closed, and an involution \(\overline{\gamma}\) of \(\ell\) for which the restriction of \(\overline{\gamma}\) to \(k\) precisely is \(\gamma\). Since any GQT depends only on the Hermitian matrix of the \((\sigma,1)\)-Hermitian form with respect to a chosen basis (with suitable adaptation to the infinite-dimensional case), it is clear that if the aforementioned GQT comes with matrix \(A\) over \(k\), then the same matrix \(A\) defines a \((\overline{\gamma},1)\)-Hermitian form over \(\ell\) which induces the initial form over \(k\). **Observation 3.4**.: _If we fix the dimension of the Hilbert space, then any GQT over \(k\) (and with respect to \(\gamma\)) is part of the GQT over \(\ell\) (with involution \(\overline{\gamma}\))._ The reason why a good extension theory is desired, is explained in the next paragraph. **COMPARISON THEORY**.: Any two fields \(k\) and \(k^{\prime}\) of the same characteristic are contained (as a subfield) in a field \(\ell\). The following construction is simple: let \(\wp\) be the prime field in both \(k\) and \(k^{\prime}\) (isomorphic to \(\mathbb{Q}\) in characteristic \(0\) and to \(\mathbb{F}_{p}\) in characteristic \(p>0\)), generated by \(0\) and \(1\). Then put \(k=\wp(S)\) and \(k^{\prime}=\wp(S^{\prime})\), with \(S\) (resp. \(S^{\prime}\)) a basis over \(\wp\) of \(k\) (resp. \(k^{\prime}\)) consisting of algebraic and transcendental elements over \(\wp\). Then \(\wp(S\cup S^{\prime})\) is "the" desired field. Obviously such a field \(\ell\) with the extension property is not unique, since \(\ell\) can be extended indefinitely (for instance, by adding transcendental elements to \(\ell\)). If we have formulated a good extension formalism for general quantum theories (over algebraically closed fields), we would be able to evaluate problems formulated over \(k\) and \(k^{\prime}\) in one and the same theory formulated over \(\ell\) (and any of its extensions). In that way, if we fix the characteristic of the fields, we could look at a quantum theoretical setting prepared over different fields \(k\) and \(k^{\prime}\) as being two guises of the same story: just extends the GQTS over \(k\) and \(k^{\prime}\) to the appropriate GQT over \(\ell\). Since it is sometimes preferable to work over algebraically closed fields, we want essentially that the following diagram commutes, after the map \(\mathsf{GQT}\) is applied which associates with each couple \((\rho,\phi)\) (where \(\rho\) is a field or division ring, and \(\phi\) an involution of \(\rho\)) the corresponding standard general quantum theory. (Of course, the same ideas can be expressed for non-standard GQTs as well.) **SCHNOR'S RESULT ON INVOLUTIONS**.: The bad news is that a general comparison dream cannot hold, as was shown in [21]. However, the result is true if we suppose \(k\) to _be algebraically closed to begin with_, by a result of Schnor [16] (which says that if \(\ell\big{/}k\) is a field extension of algebraically closed fields, and \(\gamma\) is an involution of \(k\), then there exists an involution \(\overline{\gamma}\) of \(\ell\) which fixes \(k\) and induces \(\gamma\) in \(k\)). **Theorem 3.5** (Embedding Theorem of quantum theories [21]).: _Any GQT over an algebraically closed field \(k\) with involution \(\gamma\) is embeddable in a GQT over \(\ell\), where \(\ell\) is any algebraically closed field extension of \(k\)._ In particular, this is also true for AQT: we can embed AQT in an infinite number of "nonisomorphic" GQTs over algebraically closed fields in characteristic \(0\), and this aspect adds an infinite number of layers to the theory which can be used in various situations (such as quantum coding schemes). ### The minimal model: \(\overline{\mathbb{Q}}\) Recall the following basic result: **Theorem 3.6**.: _Let \(k\) be any field. If \(k\) is not finite, then \(|\overline{k}|=|k|\), where \(\overline{k}\) is an algebraic closure of \(k\); if \(k\) is finite, \(\overline{k}\) is countable._ It's important to note that the statement relies on the Axiom of Choice! Since \(\mathbb{Q}\) is the unique prime field in characteristic \(0\), each field of characteristic \(0\) contains \(\mathbb{Q}\), and hence each algebraically closed field \(k\) in characteristic \(0\) contains the algebraically closed field \(\overline{\mathbb{Q}}\) as well. By Theorem 3.6, \(\overline{\mathbb{Q}}\) is countable, and hence it is also minimal with respect to being algebraically closed. **Observation 3.7**.: _The set of GQTs over \(\overline{\mathbb{Q}}\) can be considered as the set of minimal models (over algebraically closed fields) in characteristic \(0\)._ The idea is that by Theorem 3.5, we know that any minimal GQT can be embedded in a GQT over any given algebraically closed field in characteristic \(0\). And also, if \(k\) is algebraically closed in characteristic \(0\) and we consider a GQT over \(k\) with involution \(\sigma\), one observes that \(\sigma\) fixes the prime field \(\mathbb{Q}\), and so it also fixes the algebraic closure \(\overline{\mathbb{Q}}\).1 Footnote 1: Note that the induced action might be trivial, in which case the induced quantum theory over \(\overline{\mathbb{Q}}\) is orthogonal/symmetric. In fact, we can do a bit better: **Observation 3.8**.: _Each general quantum theory \(\mathcal{T}\) over an algebraically closed field \(k\) in characteristic \(0\) induces a general quantum theory over \(\overline{\mathbb{Q}}\)._ Proof.: Suppose \(\sigma\) is the involutory automorphism which comes with \(\mathcal{T}\). As we have seen, it fixes \(\overline{\mathbb{Q}}\). If \(\sigma\) would fix \(\overline{\mathbb{Q}}\) elementwise, then we obtain a nontrivial element of order \(2\) in \(\mathsf{Gal}(\mathbb{C}/\overline{\mathbb{Q}})\) (where the latter denotes the automorphism group of \(\mathbb{C}\) which fixes its subfield \(\overline{\mathbb{Q}}\) elementwise), which is a contradiction since this Galois group is known to be torsion-free (that is, contains no nontrivial elements of finite order). \[\begin{array}{c}\mbox{GQT over $\overline{\mathbb{Q}}$}\quad\xrightarrow{ \longrightarrow}\quad\mbox{GQT over $k$}\end{array} \tag{5}\] From this important point of view, it feels more natural to consider quantum theory coordinatized by the algebraic closure \(\overline{\mathbb{Q}}\) of the rational numbers. In fact, as the rational numbers are dense in the real numbers, the expression \[\mathbb{Q}=\mathbb{Q}+i\mathbb{Q}\ \subset\ \overline{\mathbb{Q}}+i\overline{ \mathbb{Q}}\ \subset\ \mathbb{R}+i\mathbb{R}\ =\ \mathbb{C} \tag{6}\] shows that every element of \(\mathbb{C}\) can be seen as a limit of Cauchy sequences in \(\overline{\mathbb{Q}}\), in other words: **Observation 3.9** (Universality of the minimal model).: _Classical quantum theory can be perfectly approximated by modal quantum theory over \(\overline{\mathbb{Q}}\), while the latter is countable, also algebraically closed and contained in every division ring (and in particular: field) in characteristic \(0\). _ ### Finite fields In [17], Schumacher and Westmoreland introduced modal quantum theory (MQT) as a finite "toy model" for AQT, in which the underlying field \(\mathbb{C}\) is replaced by a finite field \(\mathbb{F}_{q}\) (where \(q\) is an arbitrary prime power). Inner products in the usual formal sense are not defined on vector spaces over a finite field, and hence this aspect is not covered in [17]. This means that the very notion of "orthogonal states" does not occur in their approach. In [14], vector spaces are considered over finite fields \(\mathbb{F}_{p}\) with \(p\) a prime, for which the following property holds: \[-1\mbox{ is not a square in $\mathbb{F}_{p}$, but it is in $\mathbb{F}_{p^{2}}$.}\] The reason is obvious: besides the many similarities between \(\left(\mathbb{C},\mathbb{R}\right)\) and \(\left(\mathbb{F}_{p^{2}},\mathbb{F}_{p}\right)\), one disposes of a natural Hermitian bilinear form which shares important aspects with the inner product \(\left\langle\cdot,\cdot\right\rangle\). In [21] we showed that there is no need at all for restricting the theory to primes with the aforementioned property. Here is a quick overview. * Let \(q\) be any prime power; then up to isomorphism \(\mathbb{F}_{q}\) has a unique extension of degree \(2\), namely \(\mathbb{F}_{q^{2}}\). The map (7) \[\gamma:\mathbb{F}_{q^{2}}\mapsto\mathbb{F}_{q^{2}}:a\mapsto a^{q}\] sends each element of \(\mathbb{F}_{q}\) to itself, while being an involutory automorphism of \(\mathbb{F}_{q^{2}}\). * Let \(n\) be any positive integer different from \(0\); then if \(V=V(n,q^{2})\) is the \(n\)-dimensional vector space over \(\mathbb{F}_{q^{2}}\), define for \(x=(x_{1},\ldots,x_{n})\) and \(y=(y_{1},\ldots,y_{n})\) in \(V\), (8) \[\left\langle x,y\right\rangle:=x_{1}^{\gamma}y_{1}+\cdots+x_{n}^{\gamma}y_{n}.\] * For \(\rho,\rho^{\prime}\in\mathbb{F}_{q^{2}}\) we have that (9) \[\left\langle\rho x,\rho^{\prime}y\right\rangle=\rho^{\gamma}\Big{\langle}x,y \Big{\rangle}\rho^{\prime},\text{ and }\Big{\langle}x,y\Big{\rangle}^{\gamma}=\Big{\langle}y,x \Big{\rangle}.\] The following observation is taken from [21]. **Observation 3.10**.: _The linear \((n\times n)\)-matrices \(U\) which preserve the form \(\left\langle\cdot,\cdot\right\rangle\) precisely are unitary matrices: \((n\times n)\)-matrices \(U\) for which \(U^{*}U\) is the \((n\times n)\)-identity matrix, where \(U^{*}:=(U^{\gamma})^{T}\)._ Classical quantum theory compares with MQT over finite fields as follows. **Remark 3.11** (\((\mathbb{C},\mathbb{R})\)-Analogy -- finite fields version).: In this model of QT, \(\mathbb{F}_{q^{2}}\) plays the role of \(\mathbb{C}\), \(\mathbb{F}_{q}\) the role of \(\mathbb{R}\), \(\gamma\) the role of complex conjugation, and \(\left\langle\cdot,\cdot\right\rangle\) the role of inner product. By choosing any element \(\kappa\) in \(\mathbb{F}_{q^{2}}\setminus\mathbb{F}_{q}\), we can represent each element of \(\mathbb{F}_{q^{2}}\) uniquely as (10) \[a+\kappa b,\] with \(a,b\in\mathbb{F}_{q}\). So viewed from this representation, the situation at least looks "a little classical." ### The base field in fixed characteristic: quantum Lefschetz If we agree that in the Kobenhavn interpretation we need an algebraically closed field \(\ell\) for describing a. o. a theory of observables based on eigenvalues and eigenvectors, we still need to decide in which _characteristic_ it lives. Once a characteristic \(p\geq 0\) is fixed, another fundamental question emerges: Which base field do we select among the division rings of characteristic \(p\)? The very question as of which base field is needed in the Kobenhaving interpretation has a long history, and has been considered in many papers. Here is a very short overview of a number of authors which consider the "base field question" in recent work, as we compare it to our own GQT-approach in [21]. **BARRET/HARDY**: \(\bigcirc\) Both Hardy and Barret see states as probability vectors in some vector space \(V\), and as the probability entries are real numbers (contained in the interval \((0,1)\)), \(V\) is assumed to be a real vector space [6, 10]. This means that the underlying algebraic structure is assumed to contain the field of real numbers. In the unifying viewpoint of GQTs [21], probabilities are manifestations of the Hermitian form (through the generalized Born rule, e.g.), and the field or division ring one uses as underlying algebraic structure (whatever it is). In [10], two integer parameters \(K\) and \(N\) emerge for which the identity \(K=N^{2}\) holds. If one considers underlying algebraic structures such as the real numbers \(\mathbb{R}\), the complex numbers \(\mathbb{C}\) or the quaternions \(\mathbb{H}\), only \(\mathbb{C}\) confirms the aforementioned identity. Hardy concludes that this -- at least intuitively -- points towards the complex numbers, without providing a formal proof [11]. But the identity \(K=N^{2}\) was not considered in the entire realm of fields and division rings in characteristic \(0\) -- only over the set \(\{\mathbb{C},\mathbb{R},\mathbb{H}\}\). (On the other hand, assuming the probabilities to be rational numbers in \((0,1)\) would also yield more flexibility for \(V\).) Barret's generalized probabilistic theories are based on Hardy's axiomatic approach, so he ends up with the complex numbers as well. In our approach of GQTs [21], we also work with vector spaces, but any division ring (with involution) is allowed to be the coordinatizing agent, so as to find unifying behavior in this universe of quantum theories. The no-cloning result of [21], for instance, solely follows from the concept of linearity/superposition and works for all division rings -- hence also fields and algebraically closed fields, such as in particular \(\mathbb{C}\). It shows that no-cloning is not a particular instance at all of quantum theory represented in the framework of complex numbers. **CASSINELLI We also refer to [1, 5, 8, 19] (from that paper) for relevant papers. The discussion of the current subsection motivates us to introduce a _quantum Lefschetz formalism_ in the next section. ## 4. Quantum Lefschetz Principle A In this section we introduce the quantum Lefschetz principle. We first explain the _mathematical_ Lefschetz principle. ### Lefschetz principle The principle in its most naive but perhaps clearest form states the following: "Every true statement about an algebraic variety over the complex numbers \(\mathbb{C}\) is also true for a algebraic variety over _any_ algebraically closed field \(\ell\)." In fact, the naive formulation of the main principle is even stronger: to check whether a statement about an algebraic variety over an algebraically closed field \(\ell\) in characteristic \(0\) is true, it is sufficient to check it over \(\mathbb{C}\) (whatever that means). Lefschetz, in his book [13], states: "In a certain sense, algebraic geometry over a ground field of characteristic \(0\) may be reduced to complex algebraic geometry." (Recall the discussion in the introduction concerning the "choice of \(\mathbb{C}\).") As Seidenbergh remarks in [18], the situation is not quite as simple as Lefschetz claims. He describes the following beautiful yet simple example: consider two curves defined by equations \(f(X,Y,Z)=0\) and \(g(X,Y,Z)=0\) in the projective plane over the algebraically closed field \(k\) of characteristic \(0\). In the complex case, it is well known that such curves meet in at least one point: there exist numbers \(a,b,c\in\mathbb{C}\) such that \[f(a,b,c)\;=\;0\;=\;g(a,b,c). \tag{11}\] According to Lefschetz, we could conclude the same for \(k\). Indeed, assume that \(k\) is the algebraic closure of a field which is finitely generated over \(\mathbb{Q}\), and which is contained in \(\mathbb{C}\), and hence that \(k\) is a subfield of \(\mathbb{C}\) (cf. Theorem 4.1 below). We then conclude that the curves have a point in common (because they have a point in common over \(\mathbb{C}\)). Only now, the situation has changed: the property we were set to obtain was that the curves _have_ a point in common _over_\(k\), but now, we can only conclude that they have a point in common in some extension field of \(k\) inside of \(\mathbb{C}\)! So we obviously need more precise statements in order to grasp its depth. The essence of the principle becomes much clearer if we first look at the--precise-- baby version first: **Theorem 4.1** (Baby Lefschetz Principle).: _Let \(k\) be a field of characteristic \(0\) which is finitely generated over the field of rationals \(\mathbb{Q}\). Then there is an isomorphism of fields_ \[\gamma:\;k\;\mapsto\;\gamma(k)\] _from \(k\) to a subfield \(\gamma(k)\) of the field of complex numbers \(\mathbb{C}\)._ This statement--although simple-- is extremely powerful once we start thinking about its consequences. Lefschetz observed that every equation over an algebraically closed field \(\ell\) of characteristic \(0\) is defined by a finite number of coefficients, which generate a field \(\widetilde{\ell}\) which is finitely generated over \(\mathbb{Q}\), and whence Theorem 4.1 applies. The idea is very deep: although we start in any algebraically closed field of characteristic \(0\), the typical problems which occur in the theory of algebraic varieties involve only finite data, and using Theorem 4.1 we obtain a principle which _transfers_ the problem to the complex numbers. Unfortunately, as we have seen, Lefschetz's initial version was not precise. Tarski came up with a solution in [20]: **Theorem 4.2** (Minor Lefschetz Principle).: _The theory of algebraically closed fields of characteristic \(0\) admits quantifier elimination, and therefore all models are elementary equivalent._ To be clear, we also provide the following equivalent formulation. **Theorem 4.3** (Minor Lefschetz Principle, version 2).: _If an elementary sentence holds for one algebraically closed field of characteristic \(0\), then it holds for every algebraically closed field of characteristic \(0\)._ We recall the notion of _elementary sentence_ for the convenience of the reader. Let \(\ell\) be any field. An _atomic formula_ relative to \(\ell\) is an expression of the form \(f=0\), where \(f\) is a polynomial with coefficients in \(\ell\). By a _formula_ (again relative to \(\ell\)) we mean an expression built up in a finite number of steps from atomic formulae by means of conjunction, negation and quantifiers of the form "there exists an \(x\) such that," where \(x\) varies over an algebraically closed field \(L\) containing \(\ell\). An _elementary sentence_ (relative to \(\ell\)) then is a formula involving no free parameters. Another very interesting variation of Lefschetz's principle is the following. **Theorem 4.4** (Algebraic Lefschetz Principles).: _Let \(\Phi\) be a sentence in the language \(\mathcal{L}_{r}=\{0,1,+,-,\cdot\}\) for rings, where \(0,1\) are constants and \(+,-,\cdot\) are binary functions. The following are equivalent:_ * \(\Phi\) _is true over every algebraically closed field in characteristic_ \(0\)_;_ * \(\Phi\) _is true over some algebraically closed field in characteristic_ \(0\)_;_ * \(\Phi\) _is true over algebraically closed fields in characteristic_ \(p\neq 0\) _for arbitrarily large primes_ \(p\)_;_ * \(\Phi\) _is true over algebraically closed fields in characteristic_ \(p\neq 0\) _for sufficiently large primes_ \(p\)_;_ Unfortunately, although the Minor Lefschetz Principle is very general (and even still more general versions are known), not every statement carries over just like that. For example, the statement that the cardinality (number of rational points) of every variety is at most that of the continuum is true over \(\mathbb{C}\), but obviously this is not true over algebraically closed fields with greater cardinality than \(\mathbb{C}\). ### Algebraically closed fields in quantum theory Even if we agree to work with algebraically closed fields as the base field to describe quantum theory in the Kobenhavn interpretation, why would we have to use the complex numbers? For every prime number \(p\) (including \(0\)) and every cardinal number \(\kappa\), there is an algebraically closed field \(\ell\) which lives in characteristic \(p\) and for which \(|\ell|=\kappa\). And even if the field is not algebraically closed, there are a number of valuable quantum theories around which have much in common with classical quantum theory, but which also still behave differently. For instance, as William Wootters explained to me, quantum theory over the reals contains mysterious features which have not been explained to this day. If we write a prepared state \(|\psi\rangle\) relative to an othonormal eigenbase as \((c_{1},c_{2},\ldots,c_{d})\), and each \(c_{k}\) as \(r_{k}e^{i\phi_{k}}\), then only the real vector \((r_{1},r_{2},\ldots,r_{d})\) contains the information about probabilities. Is there an underlying physical reason? **Remark 4.5**.: Suppose we write each \(c_{k}\) as \(a_{k}+ib_{k}\) (\(a_{k},b_{k}\) real). Then all states which "probability project" on \((r_{1},r_{2},\ldots,r_{d})\) are precisely of the form \((a_{1},b_{1},\ldots,a_{d},b_{d})\) for which \(a_{k}^{2}+b_{k}^{2}=r_{k}^{2}\) for all \(k\) (while the \(r_{k}^{2}\) sum up to 1). So they are precisely the points of a so-called \(2d\)-dimensional Clifford torus. In each of the approaches we have seen in subsection 3.7, either it is (sometimes implicitly) assumed that the reals are contained in \(\ell\), or a number of properties are assumed which in the end hold for a Kobenhavn theory over the complex numbers (but not necessarily characterize the field uniquely as the complex numbers). We propose a totally different approach. If we start from any given algebraically closed field \(\ell\)--say, in characteristic \(0\) to fix ideas--then maybe a reasoning similar to that of Lefschetz might enable us to transfer the entire quantum theory described in the language of the Kobenhavn interpretation over \(\ell\), to the Kobenhavn interpretation over \(\mathbb{C}\). So a sufficiently subtle Lefschetz theory in the spirit of the previous theorems might give us an answer. **Question 4.6**.: _Can we develop a Lefschetz principle for quantum theory, so as to show that one can indeed consider the complex numbers as a suitable field of coordinates?_ An answer would largely settle the base field question in quantum theory. For instance, how much of complex quantum theory can be described in first order logic over \(\mathbb{C}\) (plus some appropriate induction principle)? Currently we are developing an answer to this fundamental question. ## 5. Automorphisms of \(\mathbb{C}\) and codes The most commonly used automorphism of \(\mathbb{C}\) in quantum theory is complex conjugation, but there are many others in all models of Zermelo-Fraenkel set theory plus AC. In fact, so many that the structure of \(\operatorname{Aut}(\mathbb{C})\) is very hard to understand. On the other hand, in models without AC, it is consistent to say that \(|\operatorname{Aut}(\mathbb{C})|=2\) -- that is, that standard complex conjugation is the only nontrivial automorphism of \(\mathbb{C}\)1 Strangely enough, the immense size and complexity of \(\operatorname{Aut}(\mathbb{C})\) (upon accepting AC) is virtually never used in quantum theory, while for instance in quantum coding theory it would be a powerful tool. But even if \(|\operatorname{Aut}(\mathbb{C})|=2\), nice things can happen. The projective semi-linear group \(\operatorname{\sf PTL}_{N}(\mathbb{C})\) acts on the (\((N-1)\)-dimensional) state space \(\operatorname{\sf PG}(N-1,\mathbb{C})\), and can select states based on the occurrence of fixed points; if \(|\operatorname{Aut}(\mathbb{C})|=2\), one can understand and control the action much better in this context. If we work in such a model, it is easy to show that for each automorphism \(\varphi\) of the state space (that is, each element of \(\operatorname{\sf PTL}_{N}(\mathbb{C})\)), we have that at least one of \(\varphi\) or \(\varphi^{2}\) is an element of the projective general linear group, and so has fixed points as the field \(\mathbb{C}\) is algebraically closed. This is a very interesting property in the context of selection processes, and hence also of quantum codes. We will come back to these codes in a future paper [22]. ## 6. Eigenvalues, eigenvectors and probabilities We start this section with a construction taken from Brunner et al. [4], of weird vector spaces upon not accepting AC. ### A particular example from Brunner et al. Let \(S\) be a set. Then \(\ell_{2}(S)\) is defined as \[\ell_{2}(S)\ =\ \{x\in\mathbb{C}^{S}\ |\ \parallel x\parallel_{2}<\infty\}, \tag{12}\] where \(\parallel x\parallel_{2}=\sup_{E\subseteq S\text{ finite}}\sqrt{\sum_{s\in E}|x(s)|^{2}}\), and where we have identified \(x\in\mathbb{C}^{S}\) with a map \(x:S\mapsto\mathbb{C}\). Now let \(\mathcal{R}:=\{P_{n}=\{a_{n},b_{n}\}\ |\ n\in\omega\}\) be a collection of Russel's socks (upon not accepting AC). (In fact, we assume a stronger version of Russel's socks, as in SS1.1 of [4].) Let \(\Omega:=\cup_{n\in\omega}P_{n}\) be the set of all socks, and define \[\mathcal{L}\ :=\ \{x\in\ell_{2}(\Omega)\ |\ (\forall n\in\omega)\ x(a_{n})=-x(b _{n})\}. \tag{13}\] In [4] it is shown that \(\mathcal{L}\) is an irreflexive complex Hilbert space, so that an operator on \(\mathcal{L}\) cannot be equal to its adjoint and the usual Hilbert space formalism of quantum theory fails to work. Brunner, Svozil and Baaz argue in [4] that such Hilbert spaces have to be taken into account, through the following thought experiment. ### Identical particles Before describing the thought experiment of [4], we recall some theory about identical particles. We say that two particles are _identical_ if all their intrinsic properties such as mass, spin, charge, etc. are exactly the same. The configuration space of \(N\) identical particles is defined by \[\mathcal{C}(d,N)\ :=\ \Big{(}\times_{N}\mathbb{R}^{d}\setminus\Delta\Big{)} \Big{/}\mathrm{Sym}(N). \tag{14}\] Here, the particles live in \(\mathbb{R}^{d}\), \(\times_{N}\mathbb{R}^{n}\) denotes the cartesian product \(\underbrace{\mathbb{R}^{n}\times\cdots\times\mathbb{R}^{n}}_{N\text{ times}}\); \(\Delta\) is the subspace of points for which at least two "coordinates" (in \((\mathbb{R}^{d})^{N}\)) are the same (_mathematical explanation_: to remove singularities; _physical explanation_: identical particles cannot occupy the same "location" in space, in the sense that no projection on coordinate axes may coincide); and finally, \(\mathrm{Sym}(N)\) is the symmetric group on \(N\) letters (which has to be divided out since we cannot distinguish between the particles). The fundamental group of \(\mathcal{C}(d,N)\) gives information about the continuous trajectories between the particles. _Example._ Let \(d=1\) and \(N=2\) (two particles moving on a line). Then \(\mathcal{C}(1,2)\) is homeomorphic to the space \(\Big{(}\mathbb{R}\times\mathbb{R}\setminus\Delta\Big{)}\Big{/}\mathrm{Sym}(2)\), in which particles \((u,v)\) and \((v,u)\) are identified, and where \(\Delta\) is defined by the line \(u=v\). So we obtain the half-plane defined by \(u>v\). Finally, a particle that follows Fermi-Dirac statistics is called a _fermion_; generally such particles have a half-odd integer spin. **Thought experiment.** View \(\{a_{n},b_{n}\}\) as an assembly of identical noninteracting spin-\(\frac{1}{2}\) particles which obey the Fermi-Dirac statistics. Its Hilbert space \(\mathcal{H}_{n}\) is defined as \[\mathcal{H}_{n}\ :=\ \Big{\langle}e_{1}(a_{n})\otimes e_{2}(b_{n})-e_{2}(a_{n}) \otimes e_{1}(b_{n})\Big{\rangle}, \tag{15}\] and is isomorphic to \(\mathcal{L}_{n}=\{x\in\ell_{2}\Big{(}\{a_{n},b_{n}\}\Big{)}\ |\ x(a_{n})+a(b_{n})=0\}\). The family of all socks is viewed as the compound system of the distinguishable assemblies. Their Fock space is \[\mathcal{F}\ =\ \oplus_{N\in\omega}\Big{(}\otimes_{n\in N}\mathcal{H}_{n} \Big{)}. \tag{16}\] The space \(\mathcal{F}\) and \(\mathcal{L}=\oplus_{n}\mathcal{L}_{n}\) are counter examples to several assertions in Hilbert space theory [4]: * both spaces do not admit an infinite orthonormal (Schauder, cf. the next subsection) eigenbase, so there is no way to choose a mode of observation (in the sense of Bohr's complementarity interpretation); there is no Hamel base either; * as the vector space duals of \(\mathcal{F}\) and \(\mathcal{L}\) are different from \(\mathcal{F}\) and \(\mathcal{L}\), there is no notion of self-adjoint operator in both \(\mathcal{F}\) and \(\mathcal{L}\). ### Schauder bases In infinite-dimensional Hilbert spaces, quantum theorists usually work with Schauder bases instead of Hamel bases. Hamel bases are the usual bases considered in linear algebra, but Schauder bases are altogether somewhat different. We say that \(\mathcal{B}\) is a _Schauder basis_ of the infinite-dimensional Hilbert space \(\mathcal{H}\), if each vector can be represented as a tuple in \(\mathcal{B}\) with at most a countable number of nonzero coefficients. (Each vector can be obtained as a convergent series of vectors generated by \(\mathcal{B}\) seen as a Hamel base of a subspace of \(\mathcal{H}\), so that the subspace generated by \(\mathcal{B}\) as a Hamel base is dense in \(\mathcal{H}\).) We say that a Hilbert space is _separable_ if it contains a countable Schauder basis. It can be shown that all infinite-dimensional separable Hilbert spaces are isometrically isomorphic to \(\ell^{2}\). Note also that by Baire categoricity, one can show that the Hamel dimension of a complex Hilbert space is always finite or uncountable! Here, "Hilbert" is important. Suppose \((e_{i})\) is an orthonormal basis of \(\mathcal{H}\) and let \(\{e_{j(t)}\}_{i\in\mathbb{N}}\) be a subset. Then \(\sum_{n=1}^{\infty}\frac{1}{n}e_{j(n)}\in\mathcal{H}\), but it is not expressible relative to \((e_{i})\) as a Hamel basis. In spaces \(\mathcal{H}\) with a Schauder base \(\mathcal{B}\) Born's probability formalism works perfectly. Note that if \(\mathcal{H}\) is an infinite-dimensional Hilbert space, the dimension refers to its dimension with respect to a Hamel basis, and that the "Schauder dimension" can be different than the usual Hamel dimension. If one considers an observable \(A\) in an infinite-dimensional Hilbert space, the orthonormal eigenbase of \(A\) is also considered to be a Schauder base. ### Projector operators A Hermitian operator \(\mathbb{P}\) of a Hilbert space \(\mathcal{H}\) is called a _projector (operator)_ if \(\mathbb{P}^{2}=\mathbb{P}\). (It is a dichotomy operator since it only allows two possible outcomes.) Now consider a Hilbert space \(\mathcal{H}\), and let \(\mathbb{P}\) be the observable which projects any vector onto the subspace \(A\) of \(\mathcal{H}\). Let \(\mathcal{B}\) be a Schauder eigenbase of \(\mathbb{P}\) (taken that it exists). Then \(\mathcal{H}=A\oplus B\), where \(A\) is generated by the eigenvectors in \(A\) and \(B\) is generated by the eigenvectors in \(\mathcal{B}\setminus A\). Let a quantum system be prepared in the state \(|\Psi\rangle\). Upon projecting \(|\Psi\rangle\) onto \(A\), respectively \(B\), we obtain vectors \(|\Psi\rangle_{A}\), respectively \(|\Psi\rangle_{B}\), and we can write \[|\Psi\rangle\ =\ |\Psi\rangle_{A}\ +\ |\Psi\rangle_{B}. \tag{17}\] Now expand \(|\Psi\rangle_{A}\), respectively \(|\Psi\rangle_{B}\), in the Schauder base \(\mathcal{B}_{A}\) of \(A\), respectively \(\mathcal{B}_{B}\) of \(\mathcal{B}\), induced by \(\mathcal{B}\) to obtain \(|\Psi\rangle_{A}=\sum_{|\Psi_{i}\rangle\in\mathcal{B}_{A}}a_{i}|\Psi_{i}\rangle\) and \(|\Psi\rangle_{B}=\sum_{|\Psi_{i}\rangle\in\mathcal{B}_{B}}b_{i}|\Psi_{i}\rangle\). Then the probability \(P_{A}\) to measure the eigenvalue \(\lambda=1\) ("YES") and the probability of measuring \(\lambda=0\) ("NO") are given by \[P_{A}\ :=\ \sum_{|\Psi_{i}\rangle\in\mathcal{B}_{A}}|a_{i}|^{2},\ \ \ P_{B}\ :=\ \sum_{|\Psi_{i}\rangle\in\mathcal{B}_{B}}|b_{i}|^{2}. \tag{18}\] Note that if \(\mathbb{P}_{i}\) is the projector operator onto the eigenvector \(|\Psi_{i}\rangle\in\mathcal{B}_{A}\), then \(\mathbb{P}\) can be easily described as \[\mathbb{P}\ =\ \sum_{|\Psi_{i}\rangle\in\mathcal{B}_{A}}|\Psi_{i}\rangle \langle\Psi_{i}|\ =\ \sum_{|\Psi_{i}\rangle\in\mathcal{B}_{A}}\mathbb{P}_{i}. \tag{19}\] ### Double slit experiments: a variation in quantum theory without AC Young's double slit experiment does not need an introduction: a light beam is fired in a straight line through a panel with two disjoint rectangular slits on a screen. Classical Physics would expect a pattern which corresponds to the size and shape of the slits, but that is not what happens. Instead, an interference pattern occurs. This even happens when the experiment involves single particles: through a double split apparatus, particles are sent one at a time and the interference pattern emerges eventually as well. Although the particles are measured as a single pulse in a single position, a probability wave describes the probability of observing the particle at a specific point \((x,y)\) in the plane of the screen. Born's rule gives us the probability distribution for finding an electron at specific places of the screen. Once an electron hits the screen and is detected, the wave function collapses and the outcome can be described through the eigenvalues of an observable matrix. Although the following thought experiment is not directly related to the double slit experiment, it still shares some of it (weird) characteristics. So let \(\mathcal{H}\) be a (generalized) Hilbert space over some field \(k\) in a model of Zermelo-Fraenkel without AC, which allows (Schauder) bases \(\mathcal{B}_{1}\) and \(\mathcal{B}_{2}\) of different cardinalities. This field may not be the complex numbers, but in the context of modal quantum theories, we should still consider this possibility. In any case we know that \(\mathcal{H}\) exist by [12]. Note that \(\mathcal{H}\) necessarily is infinite-dimensional (in the sense that it is not finite-dimensional). Now let \(\tilde{\mathcal{H}}\) be a second Hilbert space over \(k\), and consider \(\widehat{\mathcal{H}}:=\mathcal{H}\oplus\widetilde{\mathcal{H}}\). Let \(\mathbb{P}_{\mathcal{H}}\) be the projector operator of \(\widehat{\mathcal{H}}\) onto \(\mathcal{H}\); this is an observable with eigenvalues \(0\) and \(1\), and acts on \(\mathcal{H}\) as the identity. Now consider a state \(|\Psi\rangle\) in \(\widehat{H}\). After measuring \(\mathbb{P}_{\mathcal{H}}\), \(|\Psi\rangle\) collapses into a state in \(\mathcal{H}\) which is a superposition. One question now is: **Question 6.1**.: _What is the probability that a state is contained in \(\mathcal{H}\)? As \(\mathcal{H}\) has no well-defined dimension, we have no idea how "big" \(\mathcal{H}\) is with respect to \(\widehat{\mathcal{H}}\)._ But we can do better. **Thought experiment: AC black box measurements.** Suppose a quantum system is prepared in a state \(|\Psi\rangle\) in the (possibly generalized) Hilbert space \(\mathcal{H}\) over the field \(k\). We assume that upon not accepting the Axiom of Choice, \(\mathcal{H}\) admits Schauder bases of different cardinalities. Now we are going to perform a measurement corresponding to the Hermitian operator \(A\). Before making the measurement, we do not know whether the Axiom of Choice holds true or not; this can only -- in principle -- be observed after the measurement has been made. (There is a black box which returns the value "\(0\)" or "\(1\).") After the measurement, we obtain an eigenvalue \(\lambda\) with probability \(p_{\lambda}\). If AC holds in the underlying mathematical theory, this outcome is standard. If AC would not hold, \(\lambda\) could have been measured with a different probability (by the formalism below, for instance, \(p_{\lambda}\neq 0\) could be infinitely small). ### New Born formalism and higher Schauder bases Let \(S\) be an uncountable set, and suppose \(\mathcal{C}:=\{r_{s}\mid s\in S\}\) is a set of positive real numbers. Suppose \(\sum_{s\in S}r_{s}=R\) as a limit is also real. Then in any model of Zermelo-Fraenkel with AC, it can be shown that only at most a countable number of elements in \(\mathcal{C}\) are different from zero. The proof uses the fact that a countable union of finite sets is also countable -- a fact which fails miserably when AC is not around. So it makes sense to define _higher Schauder bases_ as follows. We say that \(\mathcal{B}\) is a _higher Schauder basis_ of the infinite-dimensional Hilbert space \(\mathcal{H}\) if all vectors of \(\mathcal{H}\) can be represented by a unique \(|\mathcal{B}|\)-tuple in which an uncountable number of nonzero entries is allowed and so that such vectors also occur. It follows that \(\mathcal{B}\) is not countable. Consider a state \(|\Psi\rangle=(a_{b})_{b\in\mathcal{B}}\). For Born's formalism to work, we want \[\sum_{b\in\mathcal{B}}|a_{b}|^{2}\;=\;1. \tag{20}\] Upon accepting the Axiom of Choice it is easy to show that the latter expression implies that only a countable number of entries in \((a_{b})_{b\in\mathcal{B}}\) is nonzero, and so in \(\mathcal{H}\) we can still consider those states which makes sense in the quantum-theoretical setting. If we work in Zermelo-Fraenkel set theory without choice however, there are models in which (20) is true, with an uncountable number of entries nonzero [15]! In this formalism, we only consider state vectors \(|\Psi\rangle=\left(a_{b}\right)_{b\in\mathcal{B}}\) for which \(\sum_{b\in\mathcal{B}}|a_{b}|^{2}\in\mathbb{R}\) (before normalization). (If one would work over an algebraically closed field \(\ell\) of characteristic \(0\) which is different from \(\mathbb{C}\), then we ask that \(\sum_{b\in\mathcal{B}}|a_{b}|^{2}\) is contained in the real-closed subfield which is defined relative to the choice of complex conjugation.) **Question 6.2**.: _Does this formalism cover quantum-theoretical situations which are not possible over classical Schauder bases?_ We suspect that the answer of this question, and also of the next question is "yes." **Question 6.3**.: _Does it make sense to introduce nonstandard probabilities (e. g., infinitely small probabilities) in this context?_ ### Blass's Theorem and Ineffective observables In [7], Blass showed in Zermelo-Fraenkel set theory that if every vector space has a basis, then AC also holds. He starts with a set \(\mathcal{X}\) of disjoint nonempty sets \(X_{i}\) (\(i\in I\)), picks an arbitrary field \(k\), and constructs the field extension \(k(X)\), where \(X=\cup_{X_{i}\in\mathcal{X}}X_{i}\). Then he constructs a particular subfield \(K\) of \(k(X)\), and interprets \(k(X)\) as a vector space over \(K\). In \(k(X)\) he considers the \(K\)-subspace \(V\) spanned by \(X\). Assuming that \(V\) has a basis, he then constructs a choice function on \(\mathcal{X}\). Unfortunately, it does not follow that AC is deducible from the statement that _every_\(\ell\)-vector space has a basis, with \(\ell\) a specified, fixed field. On the other hand, obviously \(k\), \(k(X)\) and \(K\) live in the same characteristic, so we have the following stronger statement. **Theorem 6.4** (Blass's Theorem, version 2).: _Let \(p\) be any prime, or \(0\). Then in Zermelo-Fraenkel set theory, AC is deducible from the assertion that every vector space over a field of characteristic \(p\) has a basis._ For quantum theory the importance is obvious: in the classical Kobenhavn formalism, observables (Hermitian operators) collapse into one of the vectors of an orthogonal eigenbase, and the corresponding eigenvalue is the resulting observed value. Upon not accepting AC and working in models of ZF-theory without AC, it could very well happen that some Hilbert spaces \(\mathcal{H}\) over \(k=\mathbb{C}\) (or some other field) do not have a base, so that the formalism of Hermitian observables fails, or needs to be adapted at the very least. In any case, by Theorem 6.4, we may suppose that the characteristic of \(k\) is \(0\). For instance, let \(\mathcal{B}\) be the set of orthonormal eigenvectors of some given observable \(B\) (which cannot be maximal by assumption), and let \(\langle\mathcal{B}\rangle\) be the subspace of \(\mathcal{H}\) generated by \(\mathcal{B}\) over \(k\) (either as a Hamel base or as a Schauder base). By taking a state \(|\Psi\rangle\) outside of \(\langle\mathcal{B}\rangle\), we cannot perform a measurement using the state \(|\Psi\rangle\). **Remark 6.5** (Quantum Lefschetz Principle B).: Is \(\mathbb{C}\) a candidate? If not, in view of first order logic we may switch to an other algebraically closed field for which Blass's Theorem does work (as such ending up with ineffective observables).

2307.11414

The Derived Deligne Conjecture

Derived $A_\infty$-algebras have a wealth of theoretical advantages over regular $A_\infty$-algebras. However, due to their bigraded nature, in practice they are often unwieldy to work with. We develop a framework involving brace algebras on operads which allows us to study derived $A_\infty$ algebras in a new conceptual context. One particular advantage is that this construction allows us to generalize the Lie algebra structure on the Hochschild complex of an $A_\infty$-algebra, obtaining new and rigorous versions of the Deligne conjecture.

Javier Aguilar Martín, Constanze Roitzheim

2023-07-21T08:16:23Z

http://arxiv.org/abs/2307.11414v3

# The derived Deligne conjecture ###### Abstract. This paper is based on the author's PhD Thesis under the supervision of Constanze Roitzheim. This research was founded by a GTA scholarship at the University of Kent. The author would also like to thank Sarah Whitehouse for her contributions. Key words and phrases:Operads, Derived \(A_{\infty}-\)algebras, Enriched categories, Deligne conjecture 2020 Mathematics Subject Classification: 18M70, 18M60, 18N70 ###### Contents * 1 Introduction * 2 Background and conventions * 2.1 \(A_{\infty}\)-algebras * 2.2 For the study of derived \(A_{\infty}\)-algebras * 3 Operadic suspension * 3.1 Functorial properties of operadic suspension * 4 Brace algebras * 4.1 Brace algebra structure on an operad * 4.2 Reinterpretation of \(\infty\)-morphisms * 5 \(A_{\infty}\)-algebra structures on operads * 5.1 Iterating the process * Commutativity of the right blue square * Commutativity of the left blue square * 5.2 Explicit \(A_{\infty}\)-algebra structure and Deligne conjecture * 6 Derived \(A_{\infty}\)-algebras and filtered \(A_{\infty}\)-algebras * 6.1 Derived \(A_{\infty}\)-algebras * 6.2 Filtered \(A_{\infty}\)-algebras * 7 Operadic totalization and vertical operadic suspension ###### Contents * 1 Introduction * 2 Preliminaries * 3 The derived Deligne conjecture * 4 The derived Deligne conjecture * 5 The derived Deligne conjecture * 6 The derived Deligne conjecture * 7 The derived Deligne conjecture * 8 The derived Deligne conjecture * 9 Future research * 9 Boundedness conditions * 9.2 Hochschild Cohomology * A Combinatorics * B Koszul sign on operadic suspension * C Sign of the braces ## 1. Introduction There are a number of mathematical fields in which \(A_{\infty}\)-structures arise, ranging from topology to mathematical physics. To study these structures, different interpretations of \(A_{\infty}\)-algebras have been given. From the original definition in 1963 [10], to alternative definitions in terms of tensor coalgebras [11], [12], many approaches use the machinery of operads [13], [14] or certain Lie brackets [15] to obtain these objects. Another technique to describe \(A_{\infty}\)-structures comes from brace algebras [10],[17], which often involves unwieldy sign calculations that are difficult to describe in a conceptual way. Here we used an operadic approach to obtain these signs in a more conceptual and consistent way. As a consequence, we will generalize the Lie bracket used in [15] and will give a very simple interpretation of \(A_{\infty}\)-algebras. The difference between our operadic approach and others mentioned before is that ours uses much more elementary tools and can be used to talk about \(A_{\infty}\)-structrures on any operad. We hope that this provides a useful way of thinking about \(A_{\infty}\)-structures. A first application of this simple formulation is the generalization of the Deligne conjecture. The classical Deligne conjecture states that the Hochschild complex of an associative algebra has the structure of a homotopy \(G\)-algebra [11]. This result has its roots in the theory of topological operads [12]. Since \(A_{\infty}\)-generalize associative algebras, it is natural to ask what sort of algebraic structure arises on their Hochschild complex. Thanks to the tools we develop, we are able to answer this question. Later in 2009, derived \(A_{\infty}\)-algebras were introduced by Savage [14] as a bigraded generalization of \(A_{\infty}\)-algebras in order to bypass the projectivity assumptions that are often required when working with classical \(A_{\infty}\)-algebras. We generalize the operadic description of classical \(A_{\infty}\)-algebras to the derived case by means of an operadic totalization inspired by the totalization functor described in [10]. This way we obtain an operation similar to the star operation in [11] and generalize the construction that has been done for \(A_{\infty}\)-algebras to more general derived \(A_{\infty}\)-algebras. The text is organized as follows. In Section 2 we recall some basic definitions and establish some conventions for both the classical and the derived cases. In Section 3 we define a device called _operadic suspension_ that will help us obtain the signs that we want and link this device to the classical operadic approach to \(A_{\infty}\)-algebras. We also take this construction to the level of the underlying collections of the operads to also obtain a nice description of \(\infty\)-morphisms of \(A_{\infty}\)-algebras. We then explore the functorial properties of operadic suspension, being monoidality (Proposition 3.14) the most remarkable of them. In Section 4 we study the brace algebra induced by operadic suspension and obtain a relevant result, Proposition 4.3, which establishes a relation between the canonical brace structure on an operad and the one induced by its operadic suspension. We show that as a particular case of this result we obtain the Lie bracket from [11]. Following the terminology of [10], if \(\mathcal{O}\) is an operad with an \(A_{\infty}\)-multiplication \(m\in\mathcal{O}\), it is natural to ask whether there are linear maps \(M_{j}:\mathcal{O}^{\otimes j}\to\mathcal{O}\) satisfying the \(A_{\infty}\)-algebra axioms. In Section 5 we use the aforementioned brace structure to define such linear maps on a shifted version of the operadic suspension. We then iterate this process in Section 5.1 to define an \(A_{\infty}\)-structure on the Hochschild complex of an operad with \(A_{\infty}\)-multiplication. This iteration process was inspired by the work of Getzler in [12]. Next, we prove our first main result, Theorem 5.7, which relates the \(A_{\infty}\)-structure on an operad with the one induced on its Hochschild complex. More precisely, we have the following. **Theorem A**.: _There is a morphism of \(A_{\infty}\)-algebras \(\Phi:S\mathfrak{s}\mathcal{O}\to S\mathfrak{s}\operatorname{End}_{S \mathfrak{s}\mathcal{O}}\)._ This result was hinted at by Gerstenhaber and Voronov in [10], but here we introduce a suitable context and prove it as Theorem 5.7. We also draw a connection between our framework and the one from Gerstenhaber and Voronov. As a consequence of this theorem, if \(A\) is an \(A_{\infty}\)-algebra and \(\mathcal{O}=\operatorname{End}_{A}\) its endomorphism operad, we obtain the following \(A_{\infty}\)-version of the Deligne conjecture in Corollary 5.12. **Theorem B**.: _The Hochschild complex \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) of an operad with an \(A_{\infty}\)-multiplication has a structure of a \(J\)-algebra._ In the above theorem, \(J\)-algebras play the role of homotopy \(G\)-algebras in the classical case [10]. After this, we move to the bigraded case. The goal for the bigraded section is showing that an operad \(\mathcal{O}\) with a derived \(A_{\infty}\)-multiplication \(m\in\mathcal{O}\) can be endowed with the structure of a derived \(A_{\infty}\)-algebra, just like in the classical case. We start recalling some definitions of derived \(A_{\infty}\)-algebras and filtered \(A_{\infty}\)-algebras in Section 6. In Section 7, we define the totalization functor for operads and then the bigraded version of operadic suspension. We combine these two constructions to define an operation that allows us to understand a derived \(A_{\infty}\)-multiplication as a Maurer-Cartan element. As a consequence we obtain the star operation that was introduced in [11], which also defines a Lie Bracket. From this, we obtain in Section 7.3 a brace structure from which we can obtain a classical \(A_{\infty}\)-algebra on the graded operad \(S\mathrm{Tot}(\mathfrak{s}\mathcal{O})\). Finally, in Section 8, we prove our main results about derived \(A_{\infty}\)-algebras. The first one is Theorem 8.3, which shows that, under mild boundedness assumptions, the \(A_{\infty}\)-structure on totalization is equivalent to a derived \(A_{\infty}\)-algebra on \(S\mathfrak{s}\mathcal{O}\). **Theorem C**.: _For any operad \(\mathcal{O}\) with a derived \(A_{\infty}\)-multiplication there are linear maps \(M_{ij}:(S\mathfrak{s}\mathcal{O})^{\otimes j}\to S\mathfrak{s}\mathcal{O}\), satisfying the derived \(A_{\infty}\)-algebra axioms._ The next result is Theorem 8.8, which generalizes Theorem 5.7 to the derived setting. More precisely, **Theorem D**.: _There is a morphism of derived \(A_{\infty}\)-algebras \(\Phi:S\mathfrak{s}\mathcal{O}\to S\mathfrak{s}\operatorname{End}_{S \mathfrak{s}\mathcal{O}}\)._ As a consequence of this theorem we obtain a new version of the Deligne conjecture, Corollary 8.10, this time in the setting of derived \(A_{\infty}\)-algebras. For this we also introduce a derived version of \(J\)-algebras. **Theorem E**.: _The Hochschild complex \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) of an operad with a derived \(A_{\infty}\)-multiplication has a structure of derived \(J\)-algebra._ We finish the article in Section 9 by outlining some open question about derived \(A_{\infty}\)-algebras and their Hochschild cohomology that arise from our research. ## 2. Background and conventions This section includes all necessary background and the conventions we use throughout the article. It is divided in two parts, one corresponding to classical \(A_{\infty}\)-algebras and another one for derived \(A_{\infty}\)-algebras. We will also discuss these topics in the language of operads, so we assume the reader to be familiar with them. For a full survey of operads we refer the reader to [13]. ### \(A_{\infty}\)-algebras Let us start by recalling some background definitions and results that we will need to study \(A_{\infty}\)-algebras, as well as stating some conventions. We assume that the reader is familiar with the basic definitions regarding \(A_{\infty}\)-algebras and operads, but we are going to briefly recall some of them in this section to establish notation and assumptions. For more details and intuition, the reader is referred to [10] and [11, SS9.2]. Our base category is the category of graded \(R\)-modules and linear maps, where \(R\) is a fixed commutative ring with unit of characteristic not equal to \(2\). All tensor products are taken over \(R\). We denote the \(i\)-th degree component of \(A\) as \(A^{i}\). If \(x\in A^{i}\) we write \(\deg(x)=i\) and we use cohomological grading. The symmetry isomorphism is given by the following Koszul sign convention. \[\tau_{A,B}:A\otimes B\to B\otimes A,\,x\otimes y\mapsto(-1)^{\deg(x)\deg(y)}y \otimes x\] A map \(f:A\to B\) of degree \(i\) satisfies \(f(A^{n})\subseteq B^{n+i}\) for all \(n\). The \(R\)-modules \(\operatorname{Hom}_{R}(A,B)\) are naturally graded by \[\operatorname{Hom}_{R}(A,B)^{i}=\prod_{k}\operatorname{Hom}_{R}(A^{k},B^{k+i}).\] We also adopt the following Koszul sign convention: for \(x\in A\), \(y\in B\), \(f\in\operatorname{Hom}_{R}(A,C)\) and \(g\in\operatorname{Hom}_{R}(B,D)\) \[(f\otimes g)(x\otimes y)=(-1)^{\deg(x)\deg(g)}f(x)\otimes g(y).\] homomorphisms that raise the degree by \(\mathrm{i}\) **Definition 2.1**.: _For a graded \(R\)-module \(A\) the shift or suspension of \(A\) is the graded \(R\)-module \(SA\) with \(SA^{i}=A^{i-1}\)._ **Definition 2.2**.: _An \(A_{\infty}\)-algebra is a graded \(R\)-module \(A\) together with a family of maps \(m_{n}:A^{\otimes n}\to A\) of degree \(2-n\) such that for all \(n\geq 1\)_ \[\sum_{r+s+t=n}(-1)^{rs+t}m_{r+t+1}(1^{\otimes r}\otimes m_{s}\otimes 1^{ \otimes t})=0. \tag{1}\] The above equation will sometimes be referred to as the \(A_{\infty}\)_-equation_. **Definition 2.3**.: _An \(\infty\)-morphism of \(A_{\infty}\)-algebras \(A\to B\) is a family of maps_ \[f_{n}:A^{\otimes n}\to B\] _of degree \(1-n\) such that for all \(n\geq 1\)_ \[\sum_{r+s+t=n}(-1)^{rs+t}f_{r+1+t}(1^{\otimes r}\otimes m_{s}^{A}\otimes 1^{ \otimes t})=\sum_{i_{1}+\cdots+i_{k}=n}(-1)^{s}m_{k}^{B}(f_{i_{1}}\otimes \cdots\otimes f_{i_{k}}),\] _where \(s=\sum_{\alpha<\beta}i_{\alpha}(1-i_{\beta})\). The composition of \(\infty\)-morphisms \(f:A\to B\) and \(g:B\to C\) is given by_ \[(gf)_{n}=\sum_{r}\sum_{i_{1}+\cdots+i_{r}=n}(-1)^{s}g_{r}(f_{i_{1}}\otimes\cdots \otimes f_{i_{r}}).\] **Definition 2.4**.: _A morphism of \(A_{\infty}\)-algebras is a map of \(R\)-modules \(f:A\to B\) such that_ \[f(m_{j}^{A})=m_{j}^{B}\circ f^{\otimes j}.\] ### For the study of derived \(A_{\infty}\)-algebras Now we move to the categories and conventions that we need in order to study derived \(A_{\infty}\)-algebras. The idea is that we would like to apply what we obtain in the setting of \(A_{\infty}\)-algebras on the derived setting. In order to do that, we need a way to connect a single graded category with a bigraded category. This is usually done through totalization. But in order to properly translate \(A_{\infty}\)-algebras into totalized derived \(A_{\infty}\)-algebras we need to go through several suitably enriched categories that are defined in this section. Most of the definitions come from [1, SS2] but we adapt them here to our conventions. Let \(\mathcal{C}\) be a category and let \(A\), \(B\) be arbitrary objects in \(\mathcal{C}\). We denote by \(\operatorname{Hom}_{\mathcal{C}}(A,B)\) the set of morphisms from \(A\) to \(B\) in \(\mathcal{C}\). If \((\mathcal{C},\otimes,1)\) is a closed symmetric monoidal category, then we denote its internal hom-object by \([A,B]\in\mathcal{C}\). #### 2.2.1. Filtered Modules and complexes First, we collect some definitions about filtered modules and filtered complexes. Filtrations will allow to add an extra degree to single-graded objects that will be necessary in order to connect them with bigraded objects. **Definition 2.5**.: _A filtered \(R\)-module\((A,F)\) is given by a family of \(R\)-modules\(\{F_{p}A\}_{p\in\mathbb{Z}}\) indexed by the integers such that \(F_{p}A\subseteq F_{p-1}A\) for all \(p\in\mathbb{Z}\) and \(A=\bigcup_{p}F_{p}A\). A morphism of filtered modules is a morphism \(f:A\to B\) of \(R\)-modules which is compatible with filtrations: \(f(F_{p}A)\subset F_{p}B\) for all \(p\in\mathbb{Z}\)._ We denote by \(\mathrm{C}_{R}\) the category of cochain complexes of \(R\)-modules. **Definition 2.6**.: _A filtered complex\((K,d,F)\) is a complex \((K,d)\in\mathrm{C}_{R}\) together with a filtration \(F\) of each \(R\)-module \(K^{n}\) such that \(d(F_{p}K^{n})\subset F_{p}K^{n+1}\) for all \(p,n\in\mathbb{Z}\). Its morphisms are given by morphisms of complexes \(f:K\to L\) compatible with filtrations._ We denote by \(\operatorname{fMod}_{R}\) and \(\operatorname{fC}_{R}\) the categories of filtered modules and filtered complexes of \(R\)-modules, respectively. **Definition 2.7**.: _The tensor product of two filtered \(R\)-modules\((A,F)\) and \((B,F)\) is the filtered \(R\)-module with_ \[F_{p}(A\otimes B):=\sum_{i+j=p}\operatorname{Im}(F_{i}A\otimes F_{j}B\to A \otimes B).\] _This makes the category of filtered \(R\)-modules into a symmetric monoidal category, where the unit is given by \(R\) with the trivial filtration \(0=F_{1}R\subset F_{0}R=R\)._ **Definition 2.8**.: _Let \(K\) and \(L\) be filtered complexes. We define \(\underline{\mathrm{Hom}}(K,L)\) to be the filtered complex whose underlying cochain complex is \(\mathrm{Hom}_{\mathrm{C}_{R}}(K,L)\) and the filtration \(F\) given by_ \[F_{p}\underline{\mathrm{Hom}}(K,L)=\{f:K\to L\mid f(F_{q}K)\subset F_{q+p}L\text { for all }q\in\mathbb{Z}\}.\] _In particular, \(\mathrm{Hom}_{\mathrm{Mod}_{R}}(K,L)=F_{0}\underline{\mathrm{Hom}}(K,L)\)._ #### 2.2.2. Bigraded modules, vertical bicomplexes, twisted complexes and sign conventions We collect some basic definitions of bigraded categories that we need to use and we establish some conventions. **Definition 2.9**.: _We consider \((\mathbb{Z},\mathbb{Z})\)-bigraded \(R\)-modules \(A=\{A_{i}^{j}\}\), where elements of \(A_{i}^{j}\) are said to have bidegree \((i,j)\). We sometimes refer to \(i\) as the horizontal degree and \(j\) the vertical degree. The total degree of an element \(x\in A_{i}^{j}\) is \(i+j\) and is denoted \(|x|\)._ **Definition 2.10**.: _A morphism of bidegree \((p,q)\) maps \(A_{i}^{j}\) to \(A_{i+p}^{j+q}\). The tensor product of two bigraded \(R\)-modules \(A\) and \(B\) is the bigraded \(R\)-module \(A\otimes B\) given by_ \[(A\otimes B)_{i}^{j}\coloneqq\bigoplus_{p,q}A_{p}^{q}\otimes B_{i-p}^{j-q}.\] We denote by \(\mathrm{bgMod}_{R}\) the category whose objects are bigraded \(R\)-modules and whose morphisms are morphisms of bigraded \(R\)-modules of bidegree \((0,0)\). It is symmetric monoidal with the above tensor product. We introduce the following scalar product notation for bidegrees: for \(x\), \(y\) of bidegree \((x_{1},x_{2})\), \((y_{1},y_{2})\) respectively, we let \(\langle x,y\rangle=x_{1}y_{1}+x_{2}y_{2}\). The symmetry isomorphism is given by \[\tau_{A\otimes B}:A\otimes B\to B\otimes A,\ x\otimes y\mapsto(-1)^{\langle x,y\rangle}y\otimes x.\] We follow the Koszul sign rule: if \(f:A\to B\) and \(g:C\to D\) are bigraded morphisms, then the morphism \(f\otimes g:A\otimes C\to B\otimes D\) is defined by \[(f\otimes g)(x\otimes z)\coloneqq(-1)^{\langle g,x\rangle}f(x)\otimes g(z).\] **Definition 2.11**.: _A vertical bicomplex is a bigraded \(R\)-module \(A\) equipped with a vertical differential \(d^{A}:A\to A\) of bidegree \((0,1)\). A morphism of vertical bicomplexes is a morphism of bigraded modules of bidegree \((0,0)\) commuting with the vertical differential._ We denote by \(\mathrm{vbC}_{R}\) the category of vertical bicomplexes. The tensor product of two vertical bicomplexes \(A\) and \(B\) is given by endowing the tensor product of underlying bigraded modules with vertical differential \[d^{A\otimes B}:=d^{A}\otimes 1+1\otimes d^{B}:(A\otimes B)_{u}^{v}\to(A \otimes B)_{u}^{v+1}.\] This makes \(\mathrm{vbC}_{R}\) into a symmetric monoidal category. The symmetric monoidal categories \((\mathrm{C}_{R},\otimes,R)\), \((\mathrm{bgMod}_{R},\otimes,R)\) and \((\mathrm{vbC}_{R},\otimes,R)\) are related by embeddings \(\mathrm{C}_{R}\to\mathrm{vbC}_{R}\) and \(\mathrm{bgMod}_{R}\to\mathrm{vbC}_{R}\) which are monoidal and full. **Definition 2.12**.: _Let \(A,B\) be bigraded modules. We define \([A,B]_{*}^{*}\) to be the bigraded module of morphisms of bigraded modules \(A\to B\). Furthermore, if \(A,B\) are vertical bicomplexes, and \(f\in[A,B]_{u}^{v}\), we define_ \[\delta(f):=d_{B}f-(-1)^{v}fd_{A}.\] **Lemma 2.13**.: _If \(A\), \(B\) are vertical bicomplexes, then \(([A,B]_{*}^{*},\delta)\) is a vertical bicomplex._ Proof.: Direct computation shows \(\delta^{2}=0\). **Definition 2.14**.: _A twisted complex \((A,d_{m})\) is a bigraded \(R\)-module \(A=\{A_{i}^{j}\}\) together with a family of morphisms \(\{d_{m}:A\to A\}_{m\geq 0}\) of bidegree \((m,1-m)\) such that for all \(m\geq 0\),_ \[\sum_{i+j=m}(-1)^{i}d_{i}d_{j}=0.\] **Definition 2.15**.: _A morphism of twisted complexes \(f:(A,d_{m}^{A})\to(B,d_{m}^{B})\) is given by a family of morphisms of \(R\)-modules \(\{f_{m}:A\to B\}_{m\geq 0}\) of bidegree \((m,-m)\) such that for all \(m\geq 0\),_ \[\sum_{i+j=m}d_{i}^{B}f_{j}=\sum_{i+j=m}(-1)^{i}f_{i}d_{j}^{A}.\] _The composition of morphisms is given by \((g\circ f)_{m}:=\sum_{i+j=m}g_{i}f_{j}\)._ _A morphism \(f=\{f_{m}\}_{m\geq 0}\) is said to be strict if \(f_{i}=0\) for all \(i>0\). The identity morphism \(1_{A}:A\to A\) is the strict morphism given by \((1_{A})_{0}(x)=x.\) A morphism \(f=\{f_{i}\}\) is an isomorphism if and only if \(f_{0}\) is an isomorphism of bigraded \(R\)-modules._ Note that if \(f\) is an isomorphism, then an inverse of \(f\) is obtained from an inverse of \(f_{0}\) by solving a triangular system of linear equations. Denote by \(\mathrm{tC}_{R}\) the category of twisted complexes. The following construction endows \(\mathrm{tC}_{R}\) with a symmetric monoidal structure, see [1, Lemma 3.3] for a proof. **Lemma 2.16**.: _The category \((\mathrm{tC}_{R},\otimes,R)\) is symmetric monoidal, where the monoidal structure is given by the bifunctor_ \[\otimes:\mathrm{tC}_{R}\times\mathrm{tC}_{R}\to\mathrm{tC}_{R}.\] _On objects it is given by \(((A,d_{m}^{A}),(B,d_{m}^{B}))\to(A\otimes B,d_{m}^{A}\otimes 1+1\otimes d_{m}^{ B})\) and on morphisms it is given by \((f,g)\to f\otimes g\), where \((f\otimes g)_{m}:=\sum_{i+j=m}f_{i}\otimes g_{j}\). In particular, by the Koszul sign rule we have that_ \[(f_{i}\otimes g_{j})(x\otimes z)=(-1)^{\langle g_{j},x\rangle}f_{i}(x)\otimes g _{j}(z).\] _The symmetry isomorphism is given by the strict morphism of twisted complexes_ \[\tau_{A\otimes B}\colon A\otimes B\to B\otimes A,\ x\otimes y\mapsto(-1)^{ \langle x,y\rangle}y\otimes x.\] The internal hom on bigraded modules can be extended to twisted complexes via the following lemma whose proof is in [1, Lemma 3.4]. **Lemma 2.17**.: _Let \(A,B\) be twisted complexes. For \(f\in[A,B]_{u}^{v}\), setting_ \[(d_{i}f):=(-1)^{i(u+v)}d_{i}^{B}f-(-1)^{v}fd_{i}^{A}\] _for \(i\geq 0\) endows \([A,B]_{*}^{*}\) with the structure of a twisted complex._ #### 2.2.3. Totalization Here we recall the definition of the totalization functor from [1] and some of the structure that it comes with. This functor and its enriched versions are key to establish a correspondence between \(A_{\infty}\)-algebras and derived \(A_{\infty}\)-algebras. **Definition 2.18**.: _The totalization \(\operatorname{Tot}(A)\) of a bigraded \(R\)-module \(A=\{A_{i}^{j}\}\) the graded \(R\)-module is given by_ \[\operatorname{Tot}(A)^{n}\coloneqq\bigoplus_{i<0}A_{i}^{n-i}\oplus\prod_{i \geq 0}A_{i}^{n-i}.\] _The column filtration of \(\operatorname{Tot}(A)\) is the filtration given by_ \[F_{p}\operatorname{Tot}(A)^{n}\coloneqq\prod_{i\geq p}A_{i}^{n-i}.\] Given a twisted complex \((A,d_{m})\), define a map \(d:\operatorname{Tot}(A)\to\operatorname{Tot}(A)\) of degree \(1\) by letting \[d(x)_{j}\coloneqq\sum_{m\geq 0}(-1)^{mn}d_{m}(x_{j-m})\] for \(x=(x_{i})_{i\in\mathbb{Z}}\in\operatorname{Tot}(A)^{n}\). Here \(x_{i}\in A_{i}^{n-i}\) denotes the \(i\)-th component of \(x\), and \(d(x)_{j}\) denotes the \(j\)-th component of \(d(x)\). Note that, for a given \(j\in\mathbb{Z}\) there is a sufficiently large \(m\geq 0\) such that \(x_{j-m^{\prime}}=0\) for all \(m^{\prime}\geq m\). Hence \(d(x)_{j}\) is given by a finite sum. Also, for negative \(j\) sufficiently large, one has \(x_{j-m}=0\) for all \(m\geq 0\), which implies \(d(x)_{j}=0\). Given a morphism \(f:(A,d_{m})\to(B,d_{m})\) of twisted complexes, let the _totalization of \(f\)_ be the map \(\operatorname{Tot}(f):\operatorname{Tot}(A)\to\operatorname{Tot}(B)\) of degree \(0\) defined by \[(\operatorname{Tot}(f)(x))_{j}\coloneqq\sum_{m\geq 0}(-1)^{mn}f_{m}(x_{j-m})\] for \(x=(x_{i})_{i\in\mathbb{Z}}\in\operatorname{Tot}(A)^{n}\). The following is [1, Theorem 3.8]. **Theorem 2.19**.: _The assignments \((A,d_{m})\mapsto(\operatorname{Tot}(A),d,F)\), where \(F\) is the column filtration of \(\operatorname{Tot}(A)\), and \(f\mapsto\operatorname{Tot}(f)\) define a functor \(\operatorname{Tot}:\operatorname{tC}_{R}\to\operatorname{fC}_{R}\) which is an isomorphism of categories when restricted to its image._ For a filtered complex of the form \((\operatorname{Tot}(A),d,F)\) where \(A=\{A_{i}^{j}\}\) is a bigraded \(R\)-module, we can recover the twisted complex structure on \(A\) as follows. For all \(m\geq 0\), let \(d_{m}:A\to A\) be the morphism of bidegree \((m,1-m)\) defined by \[d_{m}(x)=(-1)^{nm}d(x)_{i+m},\] where \(x\in A_{i}^{n-i}\) and \(d(x)_{k}\) denotes the \(k\)-th component of \(d(x)\). Note that \(d(x)_{k}\) lies in \(A_{k}^{n+1-k}\). We will consider the following bounded categories since the totalization functor has better properties when restricted to them. **Definition 2.20**.: _We let \(\operatorname{tC}_{R}^{b}\), \(\operatorname{vbC}_{R}^{b}\) and \(\operatorname{bgMod}_{R}^{b}\) be the full subcategories of horizontally bounded on the right graded twisted complexes, vertical bicomplexes and bigraded modules respectively. This means that if \(A=\{A_{i}^{j}\}\) is an object of any of this categories, then there exists \(i\) such that \(A_{i^{\prime}}^{j}=0\) for \(i^{\prime}>i\)._ _We let \(\operatorname{fMod}_{R}^{b}\) and \(\operatorname{fC}_{R}^{b}\) be the full subcategories of bounded filtered modules, respectively complexes, i.e. the full subcategories of objects \((K,F)\) such that there exists some \(p\) with the property that \(F_{p^{\prime}}K^{n}=0\) for all \(p^{\prime}>p\). We refer to all of these as the bounded subcategories of \(\operatorname{tC}_{R}\), \(\operatorname{vbC}_{R}\), \(\operatorname{bgMod}_{R}\), \(\operatorname{fMod}_{R}\) and \(\operatorname{fC}_{R}\) respectively._ The following is [1, Proposition 3.11]. **Proposition 2.21**.: _The functors \(\operatorname{Tot}:\operatorname{bgMod}_{R}\to\operatorname{fMod}_{R}\) and \(\operatorname{Tot}:\operatorname{tC}_{R}\to\operatorname{fC}_{R}\) are lax symmetric monoidal with structure maps_ \[\epsilon:R\to\operatorname{Tot}(R)\text{ and }\mu=\mu_{A,B}:\operatorname{ Tot}(A)\otimes\operatorname{Tot}(B)\to\operatorname{Tot}(A\otimes B)\] _given by \(\epsilon=1_{R}\). For \(x=(x_{i})_{i}\in\operatorname{Tot}(A)^{n_{1}}\) and \(y=(y_{j})_{j}\in\operatorname{Tot}(B)^{n_{2}}\),_ \[\mu(x\otimes y)_{k}\coloneqq\sum_{k_{1}+k_{2}=k}(-1)^{k_{1}n_{2}}x_{k_{1}} \otimes y_{k_{2}}. \tag{2}\] _When restricted to the bounded case, \(\operatorname{Tot}:\operatorname{bgMod}_{R}^{b}\to\operatorname{fMod}_{R}^{b}\) and \(\operatorname{Tot}:\operatorname{tC}_{R}^{b}\to\operatorname{fC}_{R}^{b}\) are strong symmetric monoidal functors._ _Remark 2.22_.: There is a certain heuristic to obtain the sign appearing in the definition of \(\mu\) in Proposition 2.21. In the bounded case, we can write \[\operatorname{Tot}(A)=\bigoplus_{i}A_{i}^{n-i}.\] As direct sums commute with tensor products, we have \[\operatorname{Tot}(A)\otimes\operatorname{Tot}(B)=(\bigoplus A_{i}^{n-i})\otimes \operatorname{Tot}(B)\cong\bigoplus_{i}(A_{i}^{n-i}\otimes\operatorname{Tot}(B)).\] In the isomorphism we can interpret that each \(A_{i}^{n-i}\) passes by \(\operatorname{Tot}(B)\). Since \(\operatorname{Tot}(B)\) used total grading, we can think of this degree as being the horizontal degree, while having \(0\) vertical degree. Thus, using the Koszul sign rule we would get precisely the sign from Proposition 2.21. This explanation is just an intuition, and opens the door for other possible sign choices: what if we decide to distribute \(\operatorname{Tot}(A)\) over \(\bigoplus_{i}B_{i}^{n-i}\) instead, or if we consider the total degree as the vertical degree? These alternatives lead to other valid definitions of \(\mu\), and we will explore the consequences of some of them in Remark 7.7. **Lemma 2.23**.: _In the conditions of Proposition 2.21 for the bounded case, the inverse_ \[\mu^{-1}:\operatorname{Tot}(A_{(1)}\otimes\cdots\otimes A_{(m)})\to \operatorname{Tot}(A_{(1)})\otimes\cdots\otimes\operatorname{Tot}(A_{(m)})\] _is given on pure tensors (for notational convenience) as_ \[\mu^{-1}(x_{(1)}\otimes\cdots\otimes x_{(m)})=(-1)^{\sum_{j=2}^{m}n_{j}\sum_{ i=1}^{j-1}k_{i}}x_{(1)}\otimes\cdots\otimes x_{(m)}, \tag{3}\] _where \(x_{(l)}\in(A_{(m)})_{k_{l}}^{n_{l}-k_{l}}\)._ Proof.: For the case \(m=2\), \[\mu^{-1}:\operatorname{Tot}(A\otimes B)\to\operatorname{Tot}(A)\otimes \operatorname{Tot}(B)\] is computed explicitly as follows. Let \(c\in\operatorname{Tot}(A\otimes B)^{n}\). By definition, we have \[\operatorname{Tot}(A\otimes B)^{n}=\bigoplus_{k}(A\otimes B)_{k}^{n-k}= \bigoplus_{k}\bigoplus_{\begin{subarray}{c}k_{1}+k_{2}=k\\ n_{1}+n_{2}=n\end{subarray}}A_{k_{1}}^{n_{1}-k_{1}}\otimes B_{k_{2}}^{n_{2}- k_{2}}.\] And thus, \(c=(c_{k})_{k}\) may be written as a finite sum \(c=\sum_{k}c_{k}\), where \[c_{k}=\sum_{\begin{subarray}{c}k_{1}+k_{2}=k\\ n_{1}+n_{2}=n\end{subarray}}x_{k_{1}}^{n_{1}-k_{1}}\otimes y_{k_{2}}^{n_{2}- k_{2}}.\] Here, we introduced superscripts to indicate the vertical degree, which, unlike in the definition of \(\mu\) (Equation (2)), is not solely determined by the horizontal degree since the total degree also varies. However we are going to omit them in what follows for simplicity of notation. Distributivity allows us to rewrite \(c\) as \[c=\sum_{k}\bigoplus_{\begin{subarray}{c}k_{1}+k_{2}=k\\ n_{1}+n_{2}=n\end{subarray}}x_{k_{1}}\otimes y_{k_{2}}=\sum_{n_{1}+n_{2}=n} \sum_{k_{1}}\sum_{k_{2}}(x_{k_{1}}\otimes y_{k_{2}})=\sum_{n_{1}+n_{2}=n} \left(\sum_{k_{1}}x_{k_{1}}\right)\otimes\left(\sum_{k_{2}}y_{k_{2}}\right).\] Therefore, \(\mu^{-1}\) can be defined as \[\mu^{-1}(c)=\sum_{n_{1}+n_{2}=n}\left(\sum_{k_{1}}(-1)^{k_{1}n_{2}}x_{k_{1}} \right)\otimes\left(\sum_{k_{2}}y_{k_{2}}\right).\] The general case follows inductively. #### 2.2.4. Enriched categories and enriched totalization Monoidal categories over a baseWe collect some notions and results about enriched categories from [11] and [12, SS4.2] that we will need as a categorical setting for our results on derived \(A_{\infty}\)-algebras. **Definition 2.24**.: _Let \((\mathscr{V},\otimes,1)\) be a symmetric monoidal category and let \((\mathcal{C},\otimes,1)\) be a monoidal category. We say that \(\mathcal{C}\) is a monoidal category over \(\mathscr{V}\) if we have an external tensor product \(*:\mathscr{V}\times\mathcal{C}\to\mathcal{C}\) such that we have the following natural isomorphisms._ * \(1*X\cong X\) _for all_ \(X\in\mathcal{C}\)_,_ * \((C\otimes D)*X\cong C*(D*X)\) _for all_ \(C,D\in\mathscr{V}\) _and_ \(X\in\mathcal{C}\)_,_ * \(C*(X\otimes Y)\cong(C*X)\otimes Y\cong X\otimes(C*Y)\) _for all_ \(C\in\mathscr{V}\) _and_ \(X,Y\in\mathcal{C}\)_._ _Remark 2.25_.: We will also assume that there is a bifunctor \(\underline{\mathscr{C}}(-,-):\mathcal{C}^{op}\times\mathcal{C}\to\mathscr{V}\) such that we have natural bijections \[\operatorname{Hom}_{\mathcal{C}}(C*X,Y)\cong\operatorname{Hom}_{\mathscr{V}}( C,\underline{\mathscr{C}}(X,Y)).\] Under this assumption we get a \(\mathscr{V}\)-enriched category \(\underline{\mathscr{C}}\) with the same objects as \(\mathcal{C}\) and with hom-objects given by \(\underline{\mathscr{C}}(-,-)\). The unit morphism \(u_{A}:1\to\underline{\mathscr{C}}(A,A)\) corresponds to the identity map in \(\mathcal{C}\) under the adjunction, and the composition morphism is given by the adjoint of the composite \[(\underline{\mathscr{C}}(B,C)\otimes\underline{\mathscr{C}}(A,B))*A\cong \underline{\mathscr{C}}(B,C)*(\underline{\mathscr{C}}(A,B)*A)\xrightarrow{ id*ev_{AB}}\underline{\mathscr{C}}(B,C)*B\xrightarrow{ev_{BC}}C,\] where \(ev_{AB}\) is the adjoint of the identity \(\underline{\mathscr{C}}(A,B)\to\underline{\mathscr{C}}(A,B)\). Furthermore, \(\underline{\mathscr{C}}\) is a monoidal \(\mathscr{V}\)-enriched category, namely we have an enriched functor \[\underline{\otimes}:\underline{\mathscr{C}}\times\underline{\mathscr{C}}\to \underline{\mathscr{C}}\] where \(\underline{\mathscr{C}}\times\underline{\mathscr{C}}\) is the enriched category with objects \(\operatorname{Ob}(\mathcal{C})\times\operatorname{Ob}(\mathcal{C})\) and hom-objects \[\underline{\mathscr{C}}\times\underline{\mathscr{C}}((X,Y),(W,Z))\coloneqq \underline{\mathscr{C}}(X,W)\otimes\underline{\mathscr{C}}(Y,Z).\] In particular we get maps in \(\mathscr{V}\) \[\underline{\mathscr{C}}(X,W)\otimes\underline{\mathscr{C}}(Y,Z)\to\underline {\mathscr{C}}(X\otimes Y,W\otimes Z),\] given by the adjoint of the composite \[(\underline{\mathscr{C}}(X,W)\otimes\underline{\mathscr{C}}(Y,Z))*(X\otimes Y )\cong(\underline{\mathscr{C}}(X,W)*X)\otimes(\underline{\mathscr{C}}(Y,Z)* Y)\xrightarrow{ev_{XW}\otimes ev_{YZ}}W\otimes Z.\] **Definition 2.26**.: _Let \(\mathcal{C}\) and \(\mathcal{D}\) be monoidal categories over \(\mathscr{V}\). A lax functor over \(\mathscr{V}\) consists of a functor \(F:\mathcal{C}\to\mathcal{D}\) together with a natural transformation_ \[\nu_{F}:-\ast_{\mathcal{D}}F(-)\Rightarrow F(-\ast_{\mathcal{C}}-)\] _which is associative and unital with respect to the monoidal structures over \(\mathscr{V}\) of \(\mathcal{C}\) and \(\mathcal{D}\). See [14, Proposition 10.1.5] for explicit diagrams stating the coherence axioms. If \(\nu_{F}\) is a natural isomorphism we say \(F\) is a functor over \(\mathscr{V}\). Let \(F,G:\mathcal{C}\to\mathcal{D}\) be lax functors over \(\mathscr{V}\). A natural transformation over \(\mathscr{V}\) is a natural transformation \(\mu:F\Rightarrow G\) such that for any \(C\in\mathscr{V}\) and for any \(X\in\mathcal{C}\) we have_ \[\nu_{G}\circ(1\ast_{\mathcal{D}}\mu_{X})=\mu_{C\ast_{\mathcal{C}}X}\circ\nu_{F}.\] \(A\) lax monoidal functor over \(\mathscr{V}\) is a triple \((F,\epsilon,\mu)\), where \(F:\mathcal{C}\to\mathcal{D}\) is a lax functor over \(\mathscr{V}\), \(\epsilon:1_{\mathcal{D}}\to F(1_{\mathcal{C}})\) is a morphism in \(\mathcal{D}\) and \[\mu:F(-)\otimes F(-)\Rightarrow F(-\otimes-)\] is a natural transformation over \(\mathscr{V}\) satisfying the standard unit and associativity conditions. If \(\nu_{F}\) and \(\mu\) are natural isomorphisms then we say that \(F\) is monoidal over \(\mathscr{V}\). The following is [18, Proposition 4.11]. **Proposition 2.27**.: _Let \(F,G:\mathcal{C}\to\mathcal{D}\) be lax functors over \(\mathscr{V}\). Then \(F\) and \(G\) extend to \(\mathscr{V}\)-enriched functors_ \[\underline{F},\underline{G}:\underline{\mathscr{C}}\to\underline{\mathscr{Q}}\] _where \(\underline{\mathscr{C}}\) and \(\underline{\mathscr{Q}}\) denote the \(\mathscr{V}\)-enriched categories corresponding to \(\mathcal{C}\) and \(\mathcal{D}\) as described in Remark 2.25. Moreover, any natural transformation \(\mu:F\Rightarrow G\) over \(\mathscr{V}\) also extends to a \(\mathscr{V}\)-enriched natural transformation_ \[\underline{\mu}:\underline{F}\Rightarrow\underline{G}.\] _In particular, if \(F\) is lax monoidal over \(\mathscr{V}\), then \(\underline{F}\) is lax monoidal in the enriched sense, where the monoidal structure on \(\underline{\mathscr{C}}\times\underline{\mathscr{C}}\) is described in Remark 2.25._ **Lemma 2.28**.: _Let \(F,G:\mathcal{C}\to\mathcal{D}\) lax functors over \(\mathscr{V}\) and let \(\mu:F\Rightarrow G\) a natural transformation over \(\mathscr{V}\). For every \(X\in\mathcal{C}\) and \(Y\in\mathcal{D}\) there is a map_ \[\underline{\mathscr{Q}}(GX,Y)\to\underline{\mathscr{Q}}(FX,Y)\] _that is an isomorphism if \(\mu\) is an isomorphism._ Proof.: By Proposition 2.27 there is a \(\mathscr{V}\)-enriched natural transformation \[\underline{\mu}:\underline{F}\to\underline{G}\] that at each object \(X\) evaluates to \[\underline{\mu}_{X}:1\to\underline{\mathscr{Q}}(FX,GX)\] defined to be the adjoint of \(\mu_{X}:FX\to GX\). The map \(\underline{\mathscr{Q}}(GX,Y)\to\underline{\mathscr{Q}}(FX,Y)\) is defined as the composite \[\underline{\mathscr{Q}}(GX,Y)\cong\underline{\mathscr{Q}}(GX,Y)\otimes 1 \xrightarrow{\ 1\otimes\underline{\mu}_{X}\ }\underline{\mathscr{Q}}(GX,Y)\otimes\underline{\mathscr{Q}}(FX,GX) \xrightarrow{c}\underline{\mathscr{Q}}(FX,Y), \tag{4}\] where \(c\) is the composition map in the enriched setting. When \(\mu\) is an isomorphism we may analogously define the following map \[\underline{\mathscr{Q}}(FX,Y)\cong\underline{\mathscr{Q}}(FX,Y)\otimes 1 \xrightarrow{\ 1\otimes\underline{\mu}_{X}^{-1}\ }\underline{\mathscr{Q}}(FX,Y)\otimes\underline{\mathscr{Q}}(GX,FX) \xrightarrow{c}\underline{\mathscr{Q}}(GX,Y). \tag{5}\] We show that this map is the inverse of the map in Equation (4). (6) In the above diagram (6), \(\alpha_{X}\) is adjoint to \(1_{GX}:GX\to GX\). Diagrams (1) and (2) clearly commute. Diagram (3) commutes by associativity of \(c\). Diagram (4) commutes because \(\underline{\mu}_{X}^{-1}\) and \(\underline{\mu}_{X}\) are adjoint to mutual inverses, so their composition results in the adjoint of the identity. Finally, diagram (5) commutes because we are composing with an isomorphism. In particular, diagram (5) is a decomposition of the identity map on \(\underline{\mathscr{Q}}(GX,Y)\). By commutativity, this means that the overall diagram composes to the identity, showing that the maps (4) and (5) are mutually inverse. The following is [1, Lemma 4.15]. **Lemma 2.29**.: _The category \(\operatorname{fC}_{R}\) is monoidal over \(\operatorname{vbC}_{R}\). By restriction, \(\operatorname{fMod}_{R}\) is monoidal over \(\operatorname{bgMod}_{R}\)._ Enriched categories and totalization.: Here, we define some useful enriched categories and results from [1, SS4.3 and 4.4]. Some of them had to be modified to adjust them to our conventions. **Definition 2.30**.: _Let \(A,B,C\) be bigraded modules. We denote by \(\underline{\mbox{\sf gzd\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\! _._ 3. _The composition morphism_ \(c:\underline{\mbox{\sf gzdMod}}_{R}(B,C)\otimes\underline{\mbox{\sf gzdMod}}_{R}(A,B) \rightarrow\underline{\mbox{\sf gzdMod}}_{R}(A,C)\) _is given by Definition_ 2.30_._ 4. _The unit morphism_ \(R\rightarrow\underline{\mbox{\sf gzdMod}}_{R}(A,A)\) _is given by the morphism of bigraded modules that sends_ \(1\in R\) _to_ \(1_{A}:A\to A\)_, the strict morphism given by the identity of_ \(A\)_._ **Definition 2.35**.: _The \(\mathrm{vbc}_{R}\)-enriched category of twisted complexes \(\underline{tC}_{R}\) is the enriched category given by the following data._ 1. _The objects of_ \(\underline{tC}_{R}\) _are twisted complexes._ 2. _For_ \(A,B\) _twisted complexes the hom-object is the vertical bicomplex_ \(\underline{tC}_{R}(A,B)\)_._ 3. _The composition morphism_ \(c:\underline{tC}_{R}(B,C)\otimes\underline{tC}_{R}(A,B)\rightarrow\underline{ tC}_{R}(A,C)\) _is given by Definition_ 2.30_._ 4. _The unit morphism_ \(R\rightarrow\underline{tC}_{R}(A,A)\) _is given by the morphism of vertical bicomplexes sending_ \(1\in R\) _to_ \(1_{A}:A\to A\)_, the strict morphism of twisted complexes given by the identity of_ \(A\)_._ The next tensor corresponds to \(\underline{\otimes}\) in the categorical setting of Remark 2.25, see [1, Lemma 4.27]. **Lemma 2.36**.: _The monoidal structure of \(\underline{tC}_{R}\) is given by the following map of vertical bicomplexes._ \[\begin{split}\underline{\otimes}:\underline{tC}_{R}(A,B)\otimes \underline{tC}_{R}(A^{\prime},B^{\prime})\rightarrow\underline{tC}_{R}(A \otimes A^{\prime},B\otimes B^{\prime})\\ (f,g)\rightarrow(f\underline{\otimes}g)_{m}:=\sum_{i+j=m}(-1)^{ ij}f_{i}\otimes g_{j}\end{split}\] _The monoidal structure of \(\underline{\mbox{\sf gzdMod}}_{R}\) is given by the restriction of this map._ **Definition 2.37**.: _The \(\mathrm{bgMod}_{R}\)-enriched category of filtered modules \(\underline{\mbox{\sf gMod}}_{R}\) is the enriched category given by the following data._ 1. _The objects of_ \(\underline{\mbox{\sf fMod}}_{R}\) _are filtered modules._ 2. _For filtered modules_ \((K,F)\) _and_ \((L,F)\)_, the bigraded module_ \(\underline{\mbox{\sf fMod}}_{R}(K,L)\) _is given by_ \[\underline{\mbox{\sf fMod}}_{R}(K,L)_{u}^{v}:=\{f:K\to L\mid f(F_{q}K^{m}) \subset F_{q+u}L^{m+u+v},\forall m,q\in\mathbb{Z}\}.\] 3. _The composition morphism is given by_ \(c(f,g)=(-1)^{u|g|}fg\)_, where_ \(f\) _has bidegree_ \((u,v)\)_._ 4. _The unit morphism is given by the map_ \(R\rightarrow\underline{\mbox{\sf fMod}}_{R}(K,K)\) _given by_ \(1\to 1_{K}\)_._ **Definition 2.38**.: _Let \((K,d^{K},F)\) and \((L,d^{L},F)\) be filtered complexes. We define \(\underline{\mbox{\sf fC}}_{R}(K,L)\) to be the vertical bicomplex whose underlying bigraded module is \(\underline{\mbox{\sf fMod}}_{R}(K,L)\) with vertical differential_ \[\delta(f):=c(d^{L},f)-(-1)^{\langle f,d^{K}\rangle}c(f,d^{K})=d^{L}f-(-1)^{v+u }fd^{K}=d^{L}f-(-1)^{|f|}fd^{K}\] _for \(f\in\underline{\text{Mod}}_{\underline{R}}(K,L)_{u}^{v}\), where \(c\) is the composition map from Definition 2.37._ **Definition 2.39**.: _The \(\operatorname{vbC}_{R}\)-enriched category of filtered complexes \(\underline{\text{fC}}_{\underline{R}}\) is the enriched category given by the following data._ 1. _The objects of_ \(\underline{\text{fC}}_{\underline{R}}\) _are filtered complexes._ 2. _For_ \(K,L\) _filtered complexes the hom-object is the vertical bicomplex_ \(\underline{\text{fC}}_{\underline{R}}(K,L)\)_._ 3. _The composition morphism is given as in_ \(\underline{\text{fMod}}_{\underline{R}}\) _in Definition_ 2.37_._ 4. _The unit morphism is given by the map_ \(R\to\underline{\text{fC}}_{\underline{R}}(K,K)\) _given by_ \(1\to 1_{K}\)_. We denote by_ \(\underline{\text{sfC}}_{\underline{R}}\) _the full subcategory of_ \(\underline{\text{fC}}_{\underline{R}}\) _whose objects are split filtered complexes._ The enriched monoidal structure is given as follows and can be found in [11, Lemma 4.36]. **Definition 2.40**.: _The monoidal structure of \(\underline{\text{fC}}_{\underline{R}}\) is given by the following map of vertical bicomplexes._ \[\underline{\otimes}:\underline{\text{fC}}_{\underline{R}}(K,L) \otimes\underline{\text{fC}}_{\underline{R}}(K^{\prime},L^{\prime})\to \underline{\text{fC}}_{\underline{R}}(K\otimes K^{\prime},L\otimes L^{\prime}),\] \[(f,g)\mapsto f\underline{\otimes}g:=(-1)^{u|g|}f\otimes g\] _Here \(u\) is the horizontal degree of \(f\)._ The proof of the following lemma is included in the proof of [11, Lemma 4.35]. **Lemma 2.41**.: _Let \(A\) be a vertical bicomplex that is horizontally bounded on the right and let \(K\) and \(L\) be filtered complexes. There is a natural bijection_ \[\operatorname{Hom}_{\text{fC}_{R}}(\operatorname{Tot}(A)\otimes K,L)\cong \operatorname{Hom}_{\operatorname{vbC}_{R}}(A,\underline{\text{fC}}_{ \underline{R}}(K,L))\] _given by \(f\mapsto\tilde{f}:a\mapsto(k\mapsto f(a\otimes k))\)._ We now define an enriched version of the totalization functor. **Definition 2.42**.: _Let \(A,B\) be bigraded modules and \(f\in\underline{\text{bgMod}}_{\underline{R}}(A,B)_{u}^{v}\) we define_ \[\operatorname{Tot}(f)\in\underline{\text{fMod}}_{\underline{R}}(\operatorname {Tot}(A),\operatorname{Tot}(B))_{u}^{v}\] _to be given on any \(x\in\operatorname{Tot}(A)^{n}\) by_ \[(\operatorname{Tot}(f)(x)))_{j+u}:=\sum_{m\geq 0}(-1)^{(m+u)n}f_{m}(x_{j-m}) \in B_{j+u}^{n-j+v}\subset\operatorname{Tot}(B)^{n+u+v}.\] _Let \(K=\operatorname{Tot}(A)\), \(L=\operatorname{Tot}(B)\) and \(g\in\underline{\text{fMod}}_{\underline{R}}(K,L)_{u}^{v}\). We define_ \[f:=\operatorname{Tot}^{-1}(g)\in\underline{\text{bgMod}}_{\underline{R}}(A,B )_{u}^{v}\] _to be \(f:=(f_{0},f_{1},\dots)\) where \(f_{i}\) is given on each \(A_{j}^{m+j}\) by the composite_ \[f_{i}:A_{j}^{m-j}\hookrightarrow\prod_{k\geq j}A_{k}^{m-k} =F_{j}(\operatorname{Tot}(A)^{m})\xrightarrow{g}F_{j+u}( \operatorname{Tot}(B)^{m+u+v})\] \[=\prod_{l\geq j+u}B_{l}^{m+u+v-l}\xrightarrow{\times(-1)^{(i+u)m} }B_{j+u+i}^{m-j+v-i},\] _where the last map is a projection and multiplication with the indicated sign._ The following is [1, Theorem 4.39]. **Theorem 2.43**.: _Let \(A\), \(B\) be twisted complexes. The assignments \(\mathfrak{Ind}(A):=\operatorname{Tot}(A)\) and_ \[\mathfrak{Ind}_{A,B}:\underline{t}\mathcal{C}_{R}(A,B)\to\underline{f} \mathcal{C}_{R}(\operatorname{Tot}(A),\operatorname{Tot}(B)),\,f\mapsto \operatorname{Tot}(f)\] _define a \(\operatorname{vbC}_{R}\)-enriched functor \(\mathfrak{Ind}:\underline{t}\mathcal{C}_{R}\to\underline{f}\mathcal{C}_{R}\) which restricts to an isomorphism onto its image. Furthermore, this functor restricts to a \(\operatorname{bgMod}_{R}\)-enriched functor_ \[\mathfrak{Ind}:\underline{\operatorname{bgMod}_{R}}\to\underline{f} \underline{\operatorname{Mod}_{R}}\] _which also restricts to an isomorphism onto its image._ We now define an enriched endomorphism operad. **Definition 2.44**.: _Let \(\underline{\mathscr{C}}\) be a monoidal \(\mathscr{V}\)-enriched category and \(A\) an object of \(\underline{\mathscr{C}}\). We define \(\underline{\operatorname{End}}_{A}\) to be the collection in \(\mathscr{V}\) given by_ \[\underline{\operatorname{End}}_{A}(n):=\underline{\mathscr{C}}(A^{\otimes n},A) \text{ for }n\geq 1.\] The following contains several results results from [1]. **Proposition 2.45**.: * _The enriched functors_ \[\mathfrak{Ind}:\underline{\operatorname{bgMod}_{R}}\to\underline{f} \underline{\operatorname{Mod}_{R}},\hskip 28.452756pt\mathfrak{Ind}:\underline{t} \mathcal{C}_{R}\to\underline{f}\mathcal{C}_{R}\] _are lax symmetric monoidal in the enriched sense and when restricted to the bounded case they are strong symmetric monoidal in the enriched sense._ * _For_ \(A\in\underline{\mathscr{C}}\)_, the collection_ \(\underline{\operatorname{End}}_{A}\) _defines an operad in_ \(\mathscr{V}\)_._ * _Let_ \(\mathcal{C}\) _and_ \(\mathcal{D}\) _be monoidal categories over_ \(\mathscr{V}\)_. Let_ \(F:\mathcal{C}\to\mathcal{D}\) _be a lax monoidal functor over_ \(\mathscr{V}\)_. Then for any_ \(X\in\mathcal{C}\) _there is an operad morphism_ \[\underline{\operatorname{End}}_{X}\to\underline{\operatorname{End}}_{F(X)}.\] **Lemma 2.46**.: _Let \(A\) be a twisted complex. Consider \(\underline{\operatorname{End}}_{A}(n)=\underline{t}\mathcal{C}_{R}(A^{\otimes n },A)\) and \(\underline{\operatorname{End}}_{\operatorname{Tot}(A)}(n)=\underline{f} \mathcal{C}_{R}(\operatorname{Tot}(A)^{\otimes n},\operatorname{Tot}(A))\). There is a morphism of operads_ \[\underline{\operatorname{End}}_{A}\to\underline{\operatorname{End}}_{ \operatorname{Tot}(A)},\] which is an isomorphism of operads if \(A\) is bounded. The same holds true if \(A\) is just a bigraded module. In that case, we use the enriched operads \(\underline{\mathcal{F}\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\! Putting all this together, we get the map \[\underline{\mathcal{E}nd}_{\operatorname{Tot}(A)}\to\underline{\mathcal{E}nd}_{A} \text{, }f\mapsto\operatorname{Tot}^{-1}(c(f,\mu^{-1}))\text{.}\] Since the total degree of \(\mu^{-1}\) is \(0\), composition reduces to \(c(f,\mu^{-1})=f\circ\mu^{-1}\) and we get the desired map. ## 3. Operadic suspension In this section we define an operadic suspension, which is a slight modification of the one found in [10]. This construction will help us define \(A_{\infty}\)-multiplications in a simple way. The motivation to introduce operadic suspension is that signs in \(A_{\infty}\)-algebras and related Lie structures are know to arise from a sequence of shifts. In order to discuss derived structures later, we need to pin this down more generally and rigorously. We are going to work only with non-symmetric operads, although most of what we do is also valid in the symmetric case. Let \(\Lambda(n)=S^{n-1}R\), where \(S\) is the shift of graded modules, so that \(\Lambda(n)\) is the ring \(R\) concentrated in degree \(n-1\). This module can be realized as the free \(R\)-module of rank one spanned by the exterior power \(e^{n}=e_{1}\wedge\dots\wedge e_{n}\) of degree \(n-1\), where \(e_{i}\) is the \(i\)-th element of the canonical basis of \(R^{n}\). By convention, \(\Lambda(0)\) is one-dimensional concentrated in degree \(-1\) and generated by \(e^{0}\). Let us define an operad structure on \(\Lambda=\{\Lambda(n)\}_{n\geq 0}\) via the following insertion maps (7) We are inserting the second factor onto the first one, so the sign can be explained by moving the power \(e^{m}\) of degree \(m-1\) to the \(i\)-th position of \(e^{n}\) passing by \(e_{n}\) through \(e_{i+1}\). More compactly, \[e^{n}\circ_{i}e^{m}=(-1)^{(n-i)(m-1)}e^{n+m-1}\text{.}\] The unit of this operad is \(e^{1}\in\Lambda(1)\). It can be checked by direct computation that \(\Lambda\) satisfies the axioms of an operad of graded modules. In a similar way we can define \(\Lambda^{-}(n)=S^{1-n}R\), with the same insertion maps. **Definition 3.1**.: _Let \(\mathcal{O}\) be an operad. The operadic suspension \(\mathfrak{s}\mathcal{O}\) of \(\mathcal{O}\) is given arity-wise by \(\mathfrak{s}\mathcal{O}(n)=\mathcal{O}(n)\otimes\Lambda(n)\) with diagonal composition. Similarly, we define the operadic desuspension arity-wise as \(\mathfrak{s}^{-1}\mathcal{O}(n)=\mathcal{O}(n)\otimes\Lambda^{-}(n)\)._ Even though the elements of \(\mathfrak{s}\mathcal{O}\) are tensor products of the form \(x\otimes e^{n}\), we may identify the elements of \(\mathcal{O}\) with the elements the elements of \(\mathfrak{s}\mathcal{O}\) and simply write \(x\) as an abuse of notation. **Definition 3.2**.: _For \(x\in\mathcal{O}(n)\) of degree \(\deg(x)\), its natural degree \(|x|\) is the degree of \(x\otimes e^{n}\) as an element of \(\mathfrak{s}\mathcal{O}\), namely, \(|x|=\deg(x)+n-1\). To distinguish both degrees we call \(\deg(x)\) the internal degree of \(x\), since this is the degree that \(x\) inherits from the grading of \(\mathcal{O}\)._ If we write \(\circ_{i}\) for the operadic insertion on \(\mathcal{O}\) and \(\tilde{\circ}_{i}\) for the operadic insertion on \(\mathfrak{s}\mathcal{O}\), we may find a relation between the two insertion maps in the following way. **Lemma 3.3**.: _For \(x\in\mathcal{O}(n)\) and \(y\in\mathcal{O}(m)\) we have_ \[x\tilde{\circ_{i}}y=(-1)^{(n-1)(m-1)+(n-1)\deg(y)+(i-1)(m-1)}x \circ_{i}y. \tag{8}\] Proof.: Let \(x\in\mathcal{O}(n)\) and \(y\in\mathcal{O}(m)\), and let us compute \(x\tilde{\circ}y\). \[\mathfrak{s}\mathcal{O}(n)\otimes\mathfrak{s}\mathcal{O}(m) =(\mathcal{O}(n)\otimes\Lambda(n))\otimes(\mathcal{O}(m)\otimes \Lambda(m))\cong(\mathcal{O}(n)\otimes\mathcal{O}(m))\otimes(\Lambda(n) \otimes\Lambda(m))\] \[\xrightarrow{\circ_{i}\otimes\circ_{i}}\mathcal{O}(m+n-1)\otimes \Lambda(n+m-1)=\mathfrak{s}\mathcal{O}(n+m-1).\] The symmetric monoidal structure produces the sign \((-1)^{(n-1)\deg(y)}\) in the isomorphism \(\Lambda(n)\otimes\mathcal{O}(m)\cong\mathcal{O}(m)\otimes\Lambda(n)\), and the operadic structure of \(\Lambda\) produces the sign \((-1)^{(n-i)(m-1)}\), so \[x\tilde{\circ_{i}}y=(-1)^{(n-1)\deg(y)+(n-i)(m-1)}x\circ_{i}y.\] Now we can rewrite the exponent using that we have mod \(2\) \[(n-i)(m-1)=(n-1-i-1)(m-1)=(n-1)(m-1)+(i-1)(m-1)\] so we conclude \[x\tilde{\circ_{i}}y=(-1)^{(n-1)(m-1)+(n-1)\deg(y)+(i-1)(m-1)}x \circ_{i}y.\] _Remark 3.4_.: The sign from Lemma 3.3 is exactly the sign in [11] from which the sign in the equation defining \(A_{\infty}\)-algebras (eq. (1)) is derived. This means that if \(m_{s}\in\mathcal{O}(s)\) has degree \(2-s\) and \(m_{r+1+t}\in\mathcal{O}(r+1+t)\) has degree \(1-r-t\), abusing of notation we get \[m_{r+1+t}\tilde{\circ}_{r+1}m_{s}=(-1)^{rs+t}m_{r+1+t}\circ_{r+1}m_{s}.\] Next, we are going to use the above fact to obtain a way to describe \(A_{\infty}\)-algebras in simplified operadic terms. We are also going to compare this description with a classical approach that is more general but requires heavier operadic machinery. **Definition 3.5**.: _An operad \(\mathcal{O}\) has an \(A_{\infty}\)-multiplication if there is a map \(\mathcal{A}_{\infty}\to\mathcal{O}\) from the operad of \(A_{\infty}\)-algebras._ Therefore, we have the following. **Lemma 3.6**.: _An \(A_{\infty}\)-multiplication on an operad \(\mathcal{O}\) is equivalent to an element \(m\in\mathfrak{s}\mathcal{O}\) of degree 1 concentrated in positive arity such that \(m\tilde{\circ}m=0\), where \(x\tilde{\circ}y=\sum_{i}x\tilde{\circ}_{i}y\)._ Proof.: By definition, an \(A_{\infty}\)-multiplication on \(\mathcal{O}\) corresponds to a map of operads \[f:\mathcal{A}_{\infty}\to\mathcal{O}.\] Such a map is determined by the images of the generators \(\mu_{i}\in\mathcal{A}_{\infty}(i)\) of degree \(2-i\). Whence, \(f\) it is determined by \(m_{i}=f(\mu_{i})\in\mathcal{O}(i)\). Let \(m=m_{1}+m_{2}+\cdots\). Since \[\deg(m_{i})=\deg(\mu_{i})=2-i,\] we have that the image of \(m_{i}\) in \(\mathfrak{s}\mathcal{O}\) is of degree \(2-i+i-1=1\). Therefore, \(m\in\mathfrak{s}\mathcal{O}\) is homogeneous of degree 1. The fact that \(m\tilde{\circ}m=0\) follows from Remark 3.4 and \(f\) being a map of operads. Conversely, if \(m\in\mathfrak{s}\mathcal{O}\) of degree 1 such that \(m\tilde{\circ}m=0\), let \(m_{i}\) be the component of \(m\) lying in arity \(i\). We have \(m=m_{1}+m_{2}+\cdots\). By the usual identification, \(m_{i}\) has degree \(1-i+1=2-i\) in \(\mathcal{O}\). Now we can use Equation (8) to conclude that \(m\tilde{\circ}m=0\) implies \[\sum_{\begin{subarray}{c}r+s+t\\ r,t\geq 0,\ s\geq 1\end{subarray}}\,(-1)^{rs+t}m_{r+1+t}\circ_{r+1}m_{s}=0.\] This shows that the elements \(m_{i}\) determine a map \(f:\mathcal{A}_{\infty}\to\mathcal{O}\) defined on generators by \(f(\mu_{i})=m_{i}\), as desired. _Remark 3.7_.: An \(A_{\infty}\)-multiplication on the operad \(\operatorname{End}_{A}\) is equivalent to an \(A_{\infty}\)-algebra structure on \(A\). Recall that the Koszul dual cooperad \(\mathcal{A}s^{\mathrm{i}}\) of the associative operad \(\mathcal{A}s\) is \(k\mu_{n}\) in arity \(n\), where \(\mu_{n}\) has degree \(n-1\) for \(n\geq 1\). Thus, for a graded module \(A\), we have the following operad isomorphisms, where the notation \((\geq 1)\) means that we are taking the reduced sub-operad with trivial arity 0 component. \[\operatorname{Hom}(\mathcal{A}s^{\mathrm{i}},\operatorname{End}_{A})\cong \operatorname{End}_{S^{-1}A}(\geq 1)\cong\mathfrak{s}\operatorname{End}_{A}(\geq 1).\] The first operad is the convolution operad, see [12, SS6.4.1], for which \[\operatorname{Hom}(\mathcal{A}s^{\mathrm{i}},\operatorname{End}_{A})(n)= \operatorname{Hom}_{R}(\mathcal{A}s^{\mathrm{i}}(n),\operatorname{End}_{A}(n)).\] Explicitly, for \(f\in\operatorname{End}_{A}(n)\) and \(g\in\operatorname{End}_{A}(m)\), the convolution product is given by \[f\star g=\sum_{i=1}^{n}(-1)^{(n-1)(m-1)+(n-1)\deg(b)+(i-1)(m-1)}f\circ_{i}g= \sum_{i=1}^{n}f\tilde{\circ}_{i}g=f\tilde{\circ}g.\] It is known that \(A_{\infty}\)-structures on \(A\) are determined by elements \(\varphi\in\operatorname{Hom}(\mathcal{A}\mathfrak{s}^{\mathrm{i}},\operatorname{ End}_{A})\) of degree \(1\) such that \(\varphi\star\varphi=0\)[12, Proposition 10.1.3]. Since the convolution product coincides with the operation \(\tilde{\circ}\), such an element \(\varphi\) is sent via the above isomorphism to an element \(m\in\mathfrak{s}\operatorname{End}_{A}(\geq 1)\) of degree \(1\) satisfying \(m\tilde{\circ}m=0\). Therefore, we see that this classical interpretation of \(A_{\infty}\)-algebras is equivalent to the one that Lemma 3.6 provides in the case of the operad \(\operatorname{End}_{A}\). See [12, Proposition 10.1.11] for more details about convolution operads and the more classical operadic interpretion of \(A_{\infty}\)-agebras, taking into account that in the dg-setting the definition has to be modified slightly (also the difference in sign conventions arise from the choice of the isomorphism \(\operatorname{End}_{SA}\cong\mathfrak{s}^{-1}\operatorname{End}_{A}\), see Theorem 3.9). What is more, replacing \(\operatorname{End}_{A}\) by any operad \(\mathcal{O}\) and doing similar calculations to [12, Proposition 10.1.11], we retrieve the notion of \(A_{\infty}\)-multiplication on \(\mathcal{O}\) given by Definition 3.5. _Remark 3.8_.: Above we needed to specify that only positive arity was considered. This is the case in many situations in literature, but for our purposes, we cannot assume that operads have trivial component in arity \(0\) in general, and this is what forces us to specify that \(A_{\infty}\)-multiplications are concentrated in positive arity. Let us expose the relation between operadic suspension and the usual suspension or shift of graded modules. **Theorem 3.9**.: _([13, Chapter 3, Lemma 3.16]) Given a graded \(R\)-module \(A\), there is an isomorphism of operads \(\sigma^{-1}:\operatorname{End}_{SA}\cong\mathfrak{s}^{-1}\operatorname{End}_{A}\)._ The original statement is about vector spaces, but it is still true when \(R\) is not a field. In the case of the operadic suspension defined above, the isomorphism is given by \(\sigma^{-1}(F)=(-1)^{\binom{n}{2}}S^{-1}\circ F\circ S^{\otimes n}\) for \(F\in\operatorname{End}_{SA}(n)\). The symbol \(\circ\) here is just composition of maps. Note that we are using the identification of elements of \(\operatorname{End}_{A}\) with those in \(\mathfrak{s}^{-1}\operatorname{End}_{A}\). The notation \(\sigma^{-1}\) comes from [11], where a twisted version of this map is the inverse of a map \(\sigma\). Here, we define \(\sigma:\operatorname{End}_{A}(n)\to\operatorname{End}_{SA}(n)\) as the map of graded modules given by \(\sigma(f)=S\circ f\circ(S^{-1})^{\otimes n}\). In [11] the sign for the insertion maps was obtained by computing \(\sigma^{-1}(\sigma(x)\circ_{i}\sigma(y))\). This can be interpreted as sending \(x\) and \(x\) from \(\operatorname{End}_{A}\) to \(\operatorname{End}_{SA}\) via \(\sigma\) (which is a map of graded modules, not of operads), and then applying the isomorphism induced by \(\sigma^{-1}\). In the end this is the same as simply sending \(x\) and \(y\) to their images in \(\mathfrak{s}^{-1}\operatorname{End}_{A}\). Even though \(\sigma\) is only a map of graded modules, it can be shown in a completely analogous way to Theorem 3.9 that \(\overline{\sigma}=(-1)^{\binom{n}{2}}\sigma\) induces an isomorphism of operads \[\overline{\sigma}:\operatorname{End}_{A}\cong\mathfrak{s}\operatorname{End}_{ SA}. \tag{9}\] This isomorphism can also be proved using the isomorphism \(\mathfrak{s}\mathfrak{s}^{-1}\mathcal{O}\cong\mathcal{O}\) from Lemma 3.10, namely, since \(\operatorname{End}_{SA}\cong\mathfrak{s}^{-1}\operatorname{End}_{A}\), we have \(\mathfrak{s}\operatorname{End}_{SA}\cong\mathfrak{s}\mathfrak{s}^{-1} \operatorname{End}_{A}\cong\operatorname{End}_{A}.\) In this case, the isomorphism map that we obtain goes in the opposite direction to \(\overline{\sigma}\), and it is precisely its inverse. **Lemma 3.10**.: _There are isomorphisms of operads \(\mathfrak{s}^{-1}\mathfrak{s}\mathcal{O}\cong\mathcal{O}\cong\mathfrak{s} \mathfrak{s}^{-1}\mathcal{O}\)._ Proof.: We are only showing the first isomorphism since the other one is analogous. Note that as graded \(R\)-modules, \[\mathfrak{s}^{-1}\mathfrak{s}\mathcal{O}(n)=\mathcal{O}(n)\otimes S^{1-n}R \otimes S^{n-1}R\cong\mathcal{O}(n),\] and any automorphism of \(\mathcal{O}(n)\) determines such an isomorphism. Therefore, we are going to find an automorphism \(f\) of \(\mathcal{O}(n)\) such that the above isomorphism induces a map of operads. Observe that the insertion in \(\mathfrak{s}^{-1}\mathfrak{s}\mathcal{O}\) differs from that of \(\mathcal{O}\) in just a sign. The insertion on \(\mathfrak{s}^{-1}\mathfrak{s}\mathcal{O}\) is defined as the composition of the isomorphism \[(\mathcal{O}(n)\otimes\Lambda(n)\otimes\Lambda^{-}(n))\otimes( \mathcal{O}(m)\otimes\Lambda(m)\otimes\Lambda^{-}(m))\cong\] \[(\mathcal{O}(m)\otimes\mathcal{O}(m))\otimes(\Lambda(n)\otimes \Lambda(m))\otimes(\Lambda^{-}(n)\otimes\Lambda^{-}(m))\] with the tensor product of the insertions corresponding to each operad. After cancellations, the only sign left is \((-1)^{(n-1)(m-1)}\). So we need to find an automorphism \(f\) of \(\mathcal{O}\) such that, for \(x\in\mathcal{O}(n)\) and \(y\in\mathcal{O}(m)\), \[f(x\circ_{i}y)=(-1)^{(n-1)(m-1)}f(x)\circ_{i}f(y).\] By Lemma A.1, \(f(x)=(-1)^{\binom{n}{2}}x\) is such an automorphism. ### Functorial properties of operadic suspension Here we study operadic suspension at the level of the underlying collections as an endofunctor. Recall that a collection is a family \(\mathcal{O}=\{\mathcal{O}(n)\}_{n\geq 0}\) of graded \(R\)-modules. We define the suspension of a collection \(\mathcal{O}\) as \(\mathfrak{s}\mathcal{O}(n)=\mathcal{O}(n)\otimes S^{n-1}R\), where \(S^{n-1}R\) is the ground ring concentrated in degree \(n-1\). We first show that \(\mathfrak{s}\) is a functor both on collections and on operads. Given a morphism of collections \(f:\mathcal{O}\to\mathcal{P}\), there is an obvious induced morphism \[\mathfrak{s}f:\mathfrak{s}\mathcal{O}\to\mathfrak{s}\mathcal{P},\ \mathfrak{s}f(x\otimes e^{n})=f(x)\otimes e^{n}. \tag{10}\] Since morphisms of collections preserve arity, this map is well defined because \(e^{n}\) is the same for \(x\) and \(f(x)\). Note that if \(f\) is homogeneous, \(\deg(\mathfrak{s}f)=\deg(f)\). **Lemma 3.11**.: _The assignment \(\mathcal{O}\mapsto\mathfrak{s}\mathcal{O}\) and \(f\mapsto\mathfrak{s}f\) is a functor on both the category \(\operatorname{Col}\) of collections and the category \(\operatorname{Op}\) of operads._ Proof.: The assignment preserves composition of maps. Indeed, given \(g:\mathcal{P}\to\mathcal{C}\), by definition \(\mathfrak{s}(g\circ f)(x\otimes e^{n})=g(f(x))\otimes e^{n}\), and also \[(\mathfrak{s}g\circ\mathfrak{s}f)(x\otimes e^{n})=\mathfrak{s}g(f(x)\otimes e^{ n})=g(f(x))\otimes e^{n}.\] This means that \(\mathfrak{s}\) defines an endofunctor on the category \(\operatorname{Col}\) of collections. We know that when \(\mathcal{O}\) is an operad, \(\mathfrak{s}\mathcal{O}\) is again an operad. What is more, if \(f\) is a map of operads, then the map \(\mathfrak{s}f\) is again a map of operads, since for \(a\in\mathcal{O}(n)\) and \(b\in\mathcal{O}(m)\) we have \[\mathfrak{s}f(x\tilde{\circ}_{i}y) =\mathfrak{s}f((x\otimes e^{n})\tilde{\circ}_{i}(y\otimes e^{m}))\] \[=(-1)^{(n-1)\deg(y)+(n-i)(m-1)}\mathfrak{s}f((x\circ_{i}y) \otimes e^{n+m-1})\] \[=(-1)^{(n-1)\deg(y)+(n-i)(m-1)}f(x\circ_{i}y)\otimes e^{n+m-1}\] \[=(-1)^{(n-1)\deg(y)+(n-i)(m-1)+\deg(f)\deg(x)}(f(x)\circ_{i}f(y) )\otimes e^{n+m-1}\] \[=(-1)^{(n-1)\deg(y)+(n-1)(\deg(y)+\deg(f))+\deg(f)\deg(x)}(f(x) \otimes e^{n})\tilde{\circ}_{i}(f(y)\otimes e^{m})\] \[=(-1)^{\deg(f)(\deg(x)+n-1)}\mathfrak{s}f(x)\tilde{\circ}_{i} \mathfrak{s}f(y).\] Note that \(\deg(x)+n-1\) is the degree of \(x\otimes e^{n}\) and as we said \(\deg(\mathfrak{s}f)=\deg(f)\), so the above relation is consistent with the Koszul sign rule. In any case, recall that a morphism of operads is necessarily of degree \(0\), but the above calculation hints at some monoidality properties of \(\mathfrak{s}\) that we will study afterwards. Clearly \(\mathfrak{s}f\) preserves the unit, so \(\mathfrak{s}f\) is a morphism of operads. The fact that \(\mathfrak{s}\) is a functor allows to describe algebras over operads using operadic suspension. For instance, an \(A_{\infty}\)-algebra is a map of operads \(\mathcal{O}\to\mathcal{P}\) where \(\mathcal{O}\) is an operad with \(A_{\infty}\)-multiplication. Since \(\mathfrak{s}\) is a functor, this map corresponds to a map \(\mathfrak{s}\mathcal{O}\to\mathfrak{s}\mathcal{P}\). Since in addition the map \(\mathfrak{s}\mathcal{O}\to\mathfrak{s}\mathcal{P}\) is fully determined by the original map \(\mathcal{O}\to\mathcal{P}\), this correspondence is bijective, and algebras over \(\mathcal{O}\) are equivalent to algebras over \(\mathfrak{s}\mathcal{O}\). In fact, using Lemma 3.10, it is not hard to show the following. **Proposition 3.12**.: _The functor \(\mathfrak{s}\) is an equivalence of categories both at the level of collections and at the level of operads. _ In particular, for \(A_{\infty}\)-algebras it is more convenient to work with \(\mathfrak{s}\mathcal{O}\) since the formulation of an \(A_{\infty}\)-multiplication on this operad is much simpler but we do not lose any information. #### 3.1.1. Monoidal properties of operadic suspension Now we are going to explore the monoidal properties of operadic suspension. Since operads are precisely monoids on the category \(\operatorname{Col}\) of collections, we have the following. **Proposition 3.13**.: _The endofunctor \(\mathfrak{s}:\operatorname{Col}\to\operatorname{Col}\) induces a well-defined endofunctor on the category of monoids of collections \(\operatorname{Mon}(\operatorname{Col})\). _ In fact, we can show a stronger result. **Proposition 3.14**.: _The functor \(\mathfrak{s}:\operatorname{Col}\to\operatorname{Col}\) defines a lax monoidal functor. When restricted to the subcategory of reduced operads, it is strong monoidal._ Proof.: First, we need to define the structure maps of a lax monoidal functor. Namely, we define the unit morphism \(\varepsilon:I\to\mathfrak{s}I\) to be the map \(\varepsilon(n):I(n)\to I(n)\otimes S^{n-1}R\) as the identity for \(n\neq 1\) and the isomorphism \(R\cong R\otimes R\) for \(n=1\). We also need to define a natural transformation \(\mu:\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{P}\to\mathfrak{s}( \mathcal{O}\circ\mathcal{P})\). To define it, observe that for \(\mathcal{P}=\mathcal{O}\) we would want the map \[\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{O}\xrightarrow{\mu} \mathfrak{s}(\mathcal{O}\circ\mathcal{O})\xrightarrow{\mathfrak{s}\gamma} \mathfrak{s}\mathcal{O}\] to coincide with the operadic composition \(\tilde{\gamma}\) on \(\mathfrak{s}\mathcal{O}\), where \(\gamma\) is the composition on \(\mathcal{O}\). We know that \(\mathfrak{s}\gamma\) does not add any signs. Therefore, if \(\tilde{\gamma}=(-1)^{\eta}\gamma\), with \(\eta\) explicitly computed in Proposition 4.3, the sign must come entirely from the map \(\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{O}\to\mathfrak{s}( \mathcal{O}\circ\mathcal{O})\). Thus, we define the map \[\mu:\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{P}\to\mathfrak{s}( \mathcal{O}\circ\mathcal{P})\] as the map given by \[x\otimes e^{N}\otimes x_{1}\otimes e^{a_{1}}\otimes\cdots\otimes x_{N} \otimes e^{a_{N}}\mapsto(-1)^{\eta}x\otimes x_{1}\otimes\cdots\otimes x_{N} \otimes e^{n},\] where \(a_{1}+\cdots+a_{N}=n\) and \[\eta=\sum_{j<l}a_{j}\deg(b_{l})+\sum_{j=1}^{N}(a_{j}+\deg(b_{j})-1)(N-j),\] which is the case \(k_{0}=\cdots=k_{n}=0\) in Proposition 4.3. Note that \((-1)^{\eta}\) only depends on degrees and arities, so the map is well defined. Another way to obtain this map is using the associativity isomorphisms and operadic composition on \(\Lambda\) to obtain a map \(\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{P}\to\mathfrak{s}( \mathcal{O}\circ\mathcal{P})\). We now show that \(\mu\) is natural, or in other words, for \(f:\mathcal{O}\to\mathcal{O}^{\prime}\) and \(g:\mathcal{P}\to\mathcal{P}^{\prime}\), we show that the following diagram commutes. Let \(c=x\otimes e^{N}\otimes x_{1}\otimes e^{a_{1}}\otimes\cdots\otimes x_{N} \otimes e^{a_{N}}\in\mathfrak{s}\mathcal{O}\circ\mathfrak{s}\mathcal{P}\) and let us compute \(\mathfrak{s}(f\circ g)(\mu(c))\). One has \[\mathfrak{s}(f\circ g)(\mu(c)) =\mathfrak{s}(f\circ g)((-1)^{\sum_{j<l}a_{j}\deg(x_{l})+\sum_{j=1} ^{N}(a_{j}+\deg(x_{j})-1)(N-j)}x\otimes x_{1}\otimes\cdots\otimes x_{N}\otimes e ^{n})\] \[=(-1)^{\nu}f(x)\otimes g(x_{1})\otimes\cdots\otimes g(x_{N}) \otimes e^{n}\] where \[\nu=\sum_{j<l}a_{j}\deg(x_{l})+\sum_{j=1}^{N}(\deg(x_{j})+a_{j}-1)(N-j)+N\deg( g)\deg(x)+\deg(g)\sum_{j=1}^{N}\deg(x_{j})(N-j).\] Now let us compute \(\mu((\mathfrak{s}f\circ\mathfrak{s}g)(c))\). We have \[\mu((\mathfrak{s}f\circ\mathfrak{s}g)(c))=(-1)^{\sigma}f(x)\otimes g(x_{1}) \otimes\cdots\otimes g(x_{N})\otimes e^{n},\] where \[\sigma=N\deg(g)(\deg(x)+N-1)+\deg(g)\sum_{j=1}^{N}(\deg(x_{j})+a_{j}-1)(N-j)+\] \[\sum_{j<l}a_{j}(\deg(x_{j})+\deg(g))+\sum_{j=1}^{N}(a_{j}+\deg(x_{j})+\deg(g) -1)(N-j).\] Now we compare the two signs by computing \(\nu+\sigma\mod 2\). After some cancellations of common terms and using that \(N(N-1)=0\mod 2\) we get \[\deg(g)\sum_{j=1}^{N}(a_{j}-1)(N-j)+\sum_{j<l}a_{j}\deg(g)+\sum_{j=1}^{N}\deg( g)(N-j)=\] \[\deg(g)\sum_{j=1}^{N}a_{j}(N-j)+\deg(g)\sum_{j<l}a_{j}=\] \[\deg(g)\left(\sum_{j=1}^{N}a_{j}(N-j)+\sum_{j=1}^{N}a_{j}(N-j)\right)=0\mod 2.\] This shows naturality of \(\mu\). Unitality follows directly from the definitions by direct computation. In the case of associativity, observe that by the definition of \(\mu\), the associativity axiom for \(\mu\) is equivalent to the associativity of the operadic composition \(\tilde{\gamma}\), which we know to be true. This shows that \(\mathfrak{s}\) is a lax monoidal functor. In the case where the operads have trivial arity \(0\) component, we may define an inverse to the operadic composition on \(\Lambda\) from Section 3. Namely, for \(n>0\), we may define \[\Lambda(n)\to\bigoplus_{N\geq 0}\Lambda(N)\otimes\left(\bigoplus_{a_{1}+ \cdots+a_{N}=n}\Lambda(a_{1})\otimes\cdots\otimes\Lambda(a_{N})\right)\] as the map \[e^{n}\mapsto\sum_{a_{1}+\cdots+a_{N}=n}(-1)^{\delta}e^{N}\otimes e^{a_{1}} \otimes\cdots\otimes e^{a_{N}},\] where \(\delta\) is just the same sign that shows up in the operadic composition on \(\Lambda\) (see Proposition 4.3) and \(a_{1},\ldots,a_{k}>0\). Since there are only finitely many ways of decomposing \(n\) into \(N\) positive integers, the sum is finite and thus the map is well defined. In fact, this map defines a cooperad structure on the reduced sub-operad of \(\Lambda\) with trivial arity \(0\) component. This map induces the morphism \(\mu^{-1}:\mathfrak{s}(\mathcal{O}\circ\mathcal{P})\to\mathfrak{s}\mathcal{O} \circ\mathfrak{s}\mathcal{P}\) that we are looking for. The unit morphism \(\varepsilon\) is always an isomorphism, so this shows \(\mathfrak{s}\) is strong monoidal in the reduced case. _Remark 3.15_.: If we decide to work with symmetric operads, we just need to introduce the sign action of the symmetric group on \(\Lambda(n)\), turning it into the sign representation of the symmetric group. The action on tensor products is diagonal, and the results we have obtained follow similarly replacing \(\operatorname{Col}\) by the category of \(\mathbb{S}\)-modules. ## 4. Brace algebras Brace algebras appear naturally in the context of operads when we fix the first argument of operadic composition [10]. This simple idea gives rise to a very rich structure that is the building block of the derived \(A_{\infty}\)-structures that we are going to construct. In this section we define a brace algebra structure for an arbitrary operad using operadic suspension. The use of operadic suspension will have as a result a generalization of the Lie bracket defined in [14]. First recall the definition of a brace algebra. **Definition 4.1**.: _A brace algebra on a graded module \(A\) consists of a family of maps_ \[b_{n}:A^{\otimes 1+n}\to A\] _called braces, that we evaluate on \((x,x_{1},\ldots,x_{n})\) as \(b_{n}(x;x_{1},\ldots,x_{n})\). They must satisfy the brace relation_ \[b_{m}(b_{n}(x;x_{1},\ldots,x_{n});y_{1},\ldots,y_{m})=\] \[\sum_{\begin{subarray}{c}i_{1},\ldots,i_{n}\\ j_{1}\ldots,j_{n}\end{subarray}}(-1)^{\varepsilon}b_{l}(x;y_{1},\ldots,y_{i_{ 1}},b_{j_{1}}(x_{1};y_{i_{1}+1},\ldots,y_{i_{1}+j_{1}}),\ldots,b_{j_{n}}(x_{n} ;y_{i_{n}+1},\ldots,y_{i_{n}+j_{n}}),\ldots,y_{m})\] _where \(l=n+\sum_{p=1}^{n}i_{p}\) and \(\varepsilon=\sum_{p=1}^{n}\deg(x_{p})\sum_{q=1}^{i_{p}}\deg(y_{q})\), i.e. the sign is picked up by the \(x_{i}\)'s passing by the \(y_{i}\)'s in the shuffle._ _Remark 4.2_.: Some authors might use the notation \(b_{1+n}\) instead of \(b_{n}\), but the first element is usually going to have a different role from the others, so we found \(b_{n}\) more intuitive. A shorter notation for \(b_{n}(x;x_{1},\ldots,x_{n})\) found in the literature ([10], [11]) is \(x\{x_{1},\ldots,x_{n}\}\). We will also see a bigraded version of this kind of map in Section 4. ### Brace algebra structure on an operad Given an operad \(\mathcal{O}\) with composition map \(\gamma:\mathcal{O}\circ\mathcal{O}\to\mathcal{O}\) we can define a brace algebra on the underlying module of \(\mathcal{O}\) by setting \[b_{n}:\mathcal{O}(N)\otimes\mathcal{O}(a_{1})\otimes\cdots\otimes\mathcal{O}(a _{n})\to\mathcal{O}(N-n+\sum a_{i})\] \[b_{n}(x;x_{1},\ldots,x_{n})=\sum\gamma(x;1,\ldots,1,x_{1},1,\ldots,1,x_{n},1, \ldots,1),\] where the sum runs over all possible order-preserving insertions. The brace \(b_{n}(x;x_{1},\ldots,x_{n})\) vanishes whenever \(n>N\) and \(b_{0}(x)=x\). The brace relation follows from the associativity axiom of operads. This construction can be used to define braces on \(\mathfrak{s}\mathcal{O}\). More precisely, we define maps \[b_{n}:\mathfrak{s}\mathcal{O}(N)\otimes\mathfrak{s}\mathcal{O}(a_{1})\otimes \cdots\otimes\mathfrak{s}\mathcal{O}(a_{n})\to\mathfrak{s}\mathcal{O}(N-n+ \sum a_{i})\] using the operadic composition \(\tilde{\gamma}\) on \(\mathfrak{s}\mathcal{O}\) as \[b_{n}(x;x_{1},\ldots,x_{n})=\sum\tilde{\gamma}(x;1,\ldots,1,x_{1},1,\ldots,1,x _{n},1,\ldots,1).\] We have the following relation between the brace maps \(b_{n}\) defined on \(\mathfrak{s}\mathcal{O}\) and the operadic composition \(\gamma\) on \(\mathcal{O}\). **Proposition 4.3**.: _For \(x\in\mathfrak{s}\mathcal{O}(N)\) and \(x_{i}\in\mathfrak{s}\mathcal{O}(a_{i})\) of internal degree \(q_{i}\) (\(1\leq i\leq n\)), we have_ \[b_{n}(x;x_{1},\ldots,x_{n})=\sum_{N-n=k_{0}+\cdots+k_{n}}(-1)^{n}\gamma(x \otimes 1^{\otimes k_{0}}\otimes x_{1}\otimes\cdots\otimes x_{n}\otimes 1^{ \otimes k_{n}}),\] _where_ \[\eta=\sum_{0\leq j<l\leq n}k_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}+\sum_{j =1}^{n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)k_{l}.\] Proof.: To obtain the signs that make \(\tilde{\gamma}\) differ from \(\gamma\), we must first look at the operadic composition on \(\Lambda\). We are interested in compositions of the form \[\tilde{\gamma}(x\otimes 1^{\otimes k_{0}}\otimes x_{1}\otimes 1^{\otimes k_{1}} \otimes\cdots\otimes x_{n}\otimes 1^{\otimes k_{n}})\] where \(N-n=k_{0}+\cdots+k_{n}\), \(x\) has arity \(N\) and each \(x_{i}\) has arity \(a_{i}\) and internal degree \(q_{i}\). Therefore, let us consider the corresponding operadic composition \[\Lambda(N)\otimes\Lambda(1)^{k_{0}}\otimes\Lambda(a_{1})\otimes\Lambda(1)^{ \otimes k_{1}}\otimes\cdots\otimes\Lambda(a_{n})\otimes\Lambda(1)^{k_{n}} \rTo\Lambda(N-n+\sum_{i=1}^{n}a_{i}).\] The operadic composition can be described in terms of insertions in the obvious way, namely, if \(f\in\mathfrak{s}\mathcal{O}(N)\) and \(h_{1},\ldots,h_{N}\in\mathfrak{s}\mathcal{O}\), then we have \[\tilde{\gamma}(x;y_{1},\ldots,y_{N})=(\cdots(x\tilde{\circ}_{1}y_{1})\tilde{ \circ}_{1+a(y_{1})}y_{2}\cdots)\tilde{\circ}_{1+\sum a(y_{p})}y_{N},\] where \(a(y_{p})\) is the arity of \(y_{p}\) (in this case \(y_{p}\) is either \(1\) or some \(x_{i}\)). So we just have to find out the sign iterating the same argument as in the \(i\)-th insertion. In this case, each \(\Lambda(a_{i})\) produces a sign given by the exponent \[(a_{i}-1)(N-k_{0}+\cdots-k_{i-1}-i).\] For this, recall that the degree of \(\Lambda(a_{i})\) is \(a_{i}-1\) and that the generator of this space is inserted in the position \(1+\sum_{j=0}^{i-1}k_{j}+\sum_{j=1}^{i-1}a_{j}\) of a wedge of \(N+\sum_{j=1}^{i-1}a_{j}-i+1\) generators. Therefore, performing this insertion as described in the previous section yields the aforementioned sign. Now, since \(N-n=k_{0}+\cdots+k_{n}\), we have that \[(a_{i}-1)(N-k_{0}+\cdots+k_{i-1}-i)=(a_{i}-1)(n-i+\sum_{l=i}^{n}k_{l}).\] Now we can compute the sign factor of a brace. For this, notice that the isomorphism \((\mathcal{O}(1)\otimes\Lambda(1))^{\otimes k}\cong\mathcal{O}(1)^{\otimes k} \otimes\Lambda(1)^{\otimes k}\) does not produce any signs because of degree reasons. Therefore, the sign coming from the isomorphism \[\mathcal{O}(N)\otimes\Lambda(N)\otimes(\mathcal{O}(1)\otimes\Lambda(1))^{ \otimes k_{0}}\otimes\bigotimes_{i=1}^{n}(\mathcal{O}(a_{i})\otimes\Lambda(a_ {i})\otimes(\mathcal{O}(1)\otimes\Lambda(1))^{\otimes k_{i}}\cong\] \[\mathcal{O}(N)\otimes\mathcal{O}(1)^{\otimes k_{0}}\otimes(\bigotimes_{i=1}^ {n}\mathcal{O}(a_{i})\otimes\mathcal{O}(1)^{\otimes k_{i}})\otimes\Lambda(N) \otimes\Lambda(1)^{\otimes k_{0}}\otimes(\bigotimes_{i=1}^{n}\Lambda(a_{i}) \otimes\Lambda(1)^{\otimes k_{i}})\] is determined by the exponent \[(N-1)\sum_{i=1}^{n}q_{i}+\sum_{i=1}^{n}(a_{i}-1)\sum_{l>i}q_{l}.\] This equals \[\left(\sum_{j=0}^{n}k_{j}+n-1\right)\sum_{i=1}^{n}q_{i}+\sum_{i=1}^{n}(a_{i}- 1)\sum_{l>i}q_{l}.\] After doing the operadic composition \[\mathcal{O}(N)\otimes(\bigotimes_{i=1}^{n}\mathcal{O}(a_{i}))\otimes\Lambda (N)\otimes(\bigotimes_{i=1}^{n}\Lambda(a_{i}))\longrightarrow\mathcal{O}(N-n +\sum_{i=1}^{n}a_{i})\otimes\Lambda(N-n+\sum_{i=1}^{n}a_{i})\] we can add the sign coming from the suspension, so all in all the sign \((-1)^{\eta}\) we were looking for is given by \[\eta=\sum_{i=1}^{n}(a_{i}-1)(n-i+\sum_{l=i}^{n}k_{l})+(\sum_{j=0}^{n}k_{j}+n-1 )\sum_{i=1}^{n}q_{i}+\sum_{i=1}^{n}(a_{i}-1)\sum_{l>i}q_{l}.\] It can be checked that this can be rewritten modulo \(2\) as \[\eta=\sum_{0\leq j<l\leq n}k_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}+\sum_{ j=1}^{n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)k_{l}\] as we stated. Notice that for \(\mathcal{O}=\operatorname{End}_{A}\), the brace on operadic suspension is precisely \[b_{n}(f;g_{1},\dots,g_{n})=\sum(-1)^{\eta}f(1,\dots,1,g_{1},1,\dots,1,g_{n},1, \dots,1).\] Using the brace structure on \(\mathfrak{s}\operatorname{End}_{A}\), the sign \(\eta\) gives us in particular the the same sign of the Lie bracket defined in [10]. More precisely, we have the following. **Corollary 4.4**.: _The brace \(b_{1}(f;g)\) is the operation \(f\circ g\) defined in [10] that induces a Lie bracket on the Hochschild complex of an \(A_{\infty}\)-algebra via_ \[[f,g]=b_{1}(f;g)-(-1)^{|f||g|}b_{1}(g;f).\] For this reason may use the notation \(f\bar{\circ}g\) instead of \(b_{1}(f;g)\), keeping the notation \(f\circ g\) whenever the insertion maps are denoted by \(\circ_{i}\). In [10], the sign is computed using a strategy that we generalize in Appendix C. The approach we have followed here has the advantage that the brace relation follows immediately from the associativity axiom of operadic composition. This approach also works for any operad since the difference between \(\gamma\) and \(\tilde{\gamma}\) is going to be the same sign. ### Reinterpretation of \(\infty\)-morphisms As we mentioned before, we can show an alternative description of \(\infty\)-morphisms of \(A_{\infty}\)-algebras and their composition in terms of suspension of collections, recall Definition 2.3 for the definition of these morphisms. Defining the suspension \(\mathfrak{s}\) at the level of collections as we did in Section 3.1 allows us to talk about \(\infty\)-morphisms of \(A_{\infty}\)-algebras in this setting, since they live in collections of the form \[\operatorname{End}_{B}^{A}=\{\operatorname{Hom}(A^{\otimes n},B)\}_{n\geq 1}.\] More precisely, there is a left module structure on \(\operatorname{End}_{B}^{A}\) over the operad \(\operatorname{End}_{B}\) \[\operatorname{End}_{B}\circ\operatorname{End}_{B}^{A}\to\operatorname{End}_{B} ^{A}\] given by compostion of maps \[f\otimes g_{1}\otimes\dots\otimes g_{n}\mapsto f(g_{1}\otimes\dots\otimes g_{ n})\] for \(f\in\operatorname{End}_{B}(n)\) and \(g_{i}\in\operatorname{End}_{B}^{A}\), and also an infinitesimal right module structure over the operad \(\operatorname{End}_{A}\) \[\operatorname{End}_{B}^{A}\circ_{(1)}\operatorname{End}_{A}\to\operatorname{ End}_{B}^{A}\] given by insertion of maps \[f\otimes 1^{\otimes r}\otimes g\otimes 1^{\otimes n-r-1}\mapsto f(1^{\otimes r }\otimes g\otimes 1^{\otimes n-r-1})\] for \(f\in\operatorname{End}_{B}^{A}(n)\) and \(g\in\operatorname{End}_{A}\). In addition, we have a composition \(\operatorname{End}_{C}^{B}\circ\operatorname{End}_{B}^{A}\to\operatorname{ End}_{C}^{A}\) analogous to the left module described above. They induce maps on the respective operadic suspensions which differ from the original ones by some signs that can be calculated in an analogous way to what we do on Proposition 4.3. These induced maps will give us the characterization of \(\infty\)-morphisms in Lemma 4.5. For these collections we also have \(\mathfrak{s}^{-1}\operatorname{End}_{B}^{A}\cong\operatorname{End}_{SB}^{SA}\) in analogy with Theorem 3.9, and the proof is similar but shorter since we do not need to worry about insertions. **Lemma 4.5**.: _An \(\infty\)-morphism of \(A_{\infty}\)-algebras \(A\to B\) with respective structure maps \(m^{A}\) and \(m^{B}\) is equivalent to an element \(f\in\mathfrak{s}\operatorname{End}_{B}^{A}\) of degree 0 concentrated in positive arity such that_ \[\rho(f\circ_{(1)}m^{A})=\lambda(m^{B}\circ f), \tag{11}\] _where_ \[\lambda:\mathfrak{s}\operatorname{End}_{B}\mathfrak{s}\operatorname{End}_{B}^ {A}\to\mathfrak{s}\operatorname{End}_{B}^{A}\] _is induced by the left module structure on \(\operatorname{End}_{B}^{A}\) and_ \[\rho:\mathfrak{s}\operatorname{End}_{B}\circ_{(1)}\mathfrak{s}\operatorname{ End}_{B}^{A}\to\mathfrak{s}\operatorname{End}_{B}^{A}\] _is induced by the right infinitesimal module structure on \(\operatorname{End}_{B}^{A}\)._ _In addition, the composition of \(\infty\)-morphisms is given by the natural composition_ \[\mathfrak{s}\operatorname{End}_{C}^{B}\mathfrak{s}\operatorname{End}_{B}^{A} \to\mathfrak{s}\operatorname{End}_{C}^{A}.\] Proof.: From the definitions of the operations in Equation (11), we know that this equation coincides with the one defining \(\infty\)-morphisms of \(A_{\infty}\)-algebras (Definition 2.3) up to sign. The signs that appear in the above equation are obtained in a similar way to that on \(\tilde{\gamma}\), see the proof of Proposition 4.3. Thus, it is enough to plug into the sign \(\eta\) from Proposition 4.3 the corresponding degrees and arities to obtain the desired result. The composition of \(\infty\)-morphisms follows similarly. Notice the similarity between this definition and the definitions given in [12, SS10.2.4] taking into account the minor modifications to accommodate the dg-case. In the case that \(f:A\to A\) is an \(\infty\)-endomorphism, Equation (11) can be written in terms of operadic composition as \(f\tilde{\gamma}m=\tilde{\gamma}(m\circ f)\). ## 5. \(A_{\infty}\)-algebra structures on operads In this section we use the previously described brace structures to prove Theorem 5.7, which was originally claimed by Gerstenhaber and Voronov [10]. This leads us to our first new version of the Deligne conjecture, that we prove in Corollary 5.12. Let \(\mathcal{O}\) be an operad of graded \(R\)-modules and \(\mathfrak{s}\mathcal{O}\) its operadic suspension. Let us consider the underlying graded module of the operad \(\mathfrak{s}\mathcal{O}\), which we call \(\mathfrak{s}\mathcal{O}\) again by abuse of notation, i.e. \(\operatorname{\mathfrak{s}\mathcal{O}}=\prod_{n}\operatorname{\mathfrak{s} \mathcal{O}}(n)\) with grading given by its natural degree, i.e. \(|x|=n+\deg(x)-1\) for \(x\in\operatorname{\mathfrak{s}\mathcal{O}}(n)\), where \(\deg(x)\) is its internal degree. Following [10] and [11], if we have an \(A_{\infty}\) multiplication \(m\in\mathcal{O}\), one would define an \(A_{\infty}\)-algebra structure on \(\operatorname{\mathfrak{s}\mathcal{O}}\) using the maps \[M^{\prime}_{1}(x) \coloneqq[m,x]=m\tilde{\circ}x-(-1)^{|x|}x\tilde{\circ}m,\] \[M^{\prime}_{j}(x_{1},\dots,x_{j}) \coloneqq b_{j}(m;x_{1},\dots,x_{j}), j>1.\] The prime notation here is used to indicate that these are not the definitive maps that we are going to take. Getzler shows in [11] that \(M^{\prime}=M^{\prime}_{1}+M^{\prime}_{2}+\cdots\) satisfies the relation \(M^{\prime}\circ M^{\prime}=0\) using that \(m\circ m=0\), and the proof is independent of the operad in which \(m\) is defined, so it is still valid if \(m\tilde{\circ}m=0\). But we have two problems here. The equation \(M^{\prime}\circ M^{\prime}=0\) does depend on how the circle operation is defined. More precisely, this circle operation in [11] is the natural circle operation on the endomorphism operad, which does not have any additional signs, so \(M^{\prime}\) is not an \(A_{\infty}\)-structure under our convention. The other problem has to do with the degrees. We need \(M^{\prime}_{j}\) to be homogeneous of degree \(2-j\) as a map \(\operatorname{\mathfrak{s}\mathcal{O}}^{\otimes j}\to\operatorname{\mathfrak{s }\mathcal{O}}\), but we find that \(M^{\prime}_{j}\) is homogeneous of degree \(1\) instead, as the following lemma shows. **Lemma 5.1**.: _For \(x\in\operatorname{\mathfrak{s}\mathcal{O}}\) we have that the degree of the map \(b_{j}(x;-):\operatorname{\mathfrak{s}\mathcal{O}}^{\otimes j}\to\operatorname {\mathfrak{s}\mathcal{O}}\) of graded modules is precisely \(|x|\)._ Proof.: Let \(a(x)\) denote the arity of \(x\), i.e. \(a(x)=n\) whenever \(x\in\operatorname{\mathfrak{s}\mathcal{O}}(n)\). Also, let \(\deg(x)\) be its internal degree in \(\mathcal{O}\). The natural degree of \(b_{j}(x;x_{1},\dots,x_{j})\) for \(a(x)\geq j\) as an element of \(\operatorname{\mathfrak{s}\mathcal{O}}\) by definition is \[|b_{j}(x;x_{1},\dots,x_{j})|=a(b_{j}(x;x_{1},\dots,x_{j}))+\deg(b_{j}(x;x_{1}, \dots,x_{j}))-1.\] We have \[a(b_{j}(x;x_{1},\dots,x_{j}))=a(x)-j+\sum_{i}a(x_{i})\] and \[\deg(b_{j}(x;x_{1},\dots,x_{j})=\deg(x)+\sum_{i}\deg(x_{i}),\] so \[a(b_{j}(x;x_{1},\dots,x_{j}))+\deg(b_{j}(x;x_{1},\dots,x_{j}))-1 =\] \[a(x)-j+\sum_{i}a(x_{i})+\deg(x)+\sum_{i}\deg(x_{i})-1 =\] \[a(x)+\deg(x)-1+\sum_{i}a(x_{i})+\sum_{i}\deg(x_{i})-j =\] \[a(x)+\deg(x)-1+\sum_{i}(a(x_{i})+\deg(x_{i})-1) =\] \[|x|+\sum_{i}|x_{i}|.\] This means that the degree of the map \(b_{j}(x;-):\mathfrak{s}\mathcal{O}^{\otimes j}\to\mathfrak{s}\mathcal{O}\) equals \(|x|\). **Corollary 5.2**.: _The maps_ \[M^{\prime}_{j}:\mathfrak{s}\mathcal{O}^{\otimes j}\to\mathfrak{s}\mathcal{O}, \,(x_{1},\dots,x_{j})\mapsto b_{j}(m;x_{1},\dots,x_{j})\] _for \(j>1\) and the map_ \[M^{\prime}_{1}:\mathfrak{s}\mathcal{O}\to\mathfrak{s}\mathcal{O},\,x\mapsto b _{1}(m;x)-(-1)^{|x|}b_{1}(m;x)\] _are homogeneous of degree 1._ Proof.: For \(j>1\) it is a direct consequence of Lemma 5.1. For \(j=1\) we have the summand \(b_{1}(m;x)\) whose degree follows as well from Lemma 5.1. The degree of the other summand, \(b_{1}(x;m)\), can be computed in a similar way as in the proof Lemma 5.1, giving that \(|b_{1}(x;m)|=1+|x|\). This concludes the proof. The problem we have encountered with the degrees can be resolved using shift maps as the following proposition shows. Recall that we have shift maps \(A\to SA\) of degree 1 given by the identity. **Proposition 5.3**.: _If \(\mathcal{O}\) is an operad with an \(A_{\infty}\)-multiplication \(m\in\mathcal{O}\), then there is an \(A_{\infty}\)-algebra structure on the shifted module \(\SS\mathcal{O}\)._ Proof.: Note in the proof of Lemma 5.1 that a way to turn \(M^{\prime}_{j}\) into a map of degree \(2-j\) is introducing a grading on \(\mathfrak{s}\mathcal{O}\) given by arity plus internal degree (without subtracting 1). This is equivalent to defining an \(A_{\infty}\)-algebra structure \(M\) on \(\SS\mathcal{O}\) shifting the map \(M^{\prime}=M^{\prime}_{1}+M^{\prime}_{2}+\cdots\), where \(S\) is the shift of graded modules. Therefore, we define \(M_{j}\) to be the map making the following diagram commute. In other words, \(M_{j}=\overline{\sigma}(M_{j}^{\prime})\), where \(\overline{\sigma}(F)=S\circ F\circ(S^{\otimes n})^{-1}\) for \(F\in\operatorname{End}_{\mathfrak{s}\mathcal{O}}(n)\) is the map inducing an isomorphism \(\operatorname{End}_{\mathfrak{s}\mathcal{O}}\cong\mathfrak{s}\operatorname{ End}_{S\mathcal{O}}\), see Equation (9). Since \(\overline{\sigma}\) is an operad morphism, for \(M=M_{1}+M_{2}+\cdots\), we have \[M\widehat{\circ}M=\overline{\sigma}(M^{\prime})\widehat{\circ}\overline{\sigma }(M^{\prime})=\overline{\sigma}(M^{\prime}\circ M^{\prime})=0.\] So now we have that \(M\in\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) is an element of natural degree \(1\) concentrated in positive arity such that \(M\widehat{\circ}M=0\). Therefore, in light of Remark 3.7, \(M\) is the desired \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\mathcal{O}\). Notice that \(M\) is defined as an structure map on \(S\mathfrak{s}\mathcal{O}\). This kind of shifted operad is called _odd operad_ in [11]. This means that \(S\mathfrak{s}\mathcal{O}\) is not an operad anymore, since the associativity relation for graded operads involves signs that depend on the degrees, which are now shifted. ### Iterating the process We have defined \(A_{\infty}\)-structure maps \(M_{j}\in\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\). Now we can use the brace structure of the operad \(\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) to get a \(A_{\infty}\)-algebra structure given by maps \[\overline{M}_{j}:(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}})^ {\otimes j}\to S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}} \tag{12}\] by applying \(\overline{\sigma}\) to maps \[\overline{M}_{j}^{\prime}:(\mathfrak{s}\operatorname{End}_{S\mathfrak{s} \mathcal{O}})^{\otimes j}\to\mathfrak{s}\operatorname{End}_{S\mathfrak{s} \mathcal{O}}\] defined as \[\overline{M}_{j}^{\prime}(f_{1},\dots,f_{j})=\overline{B}_{j}(M;f_ {1},\dots,f_{j}) j>1,\] \[\overline{M}_{1}^{\prime}(f)=\overline{B}_{1}(M;f)-(-1)^{|f|} \overline{B}_{1}(f;M),\] where \(\overline{B}_{j}\) denotes the brace map on \(\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\). We define the Hochschild complex as done by Ward in [11]. **Definition 5.4**.: _The Hochschild cochains of a graded module \(A\) are defined to be the graded module \(S\mathfrak{s}\operatorname{End}_{A}\). If \((A,d)\) is a cochain complex, then \(S\mathfrak{s}\operatorname{End}_{A}\) is endowed with a differential_ \[\partial(f)=[d,f]=d\circ f-(-1)^{|f|}f\circ d\] _where \(|f|\) is the natural degree of \(|f|\) and \(\circ\) is the plethysm operation given by insertions._ In particular, \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) is the module of Hochschild cochains of \(S\mathfrak{s}\mathcal{O}\). If \(\mathcal{O}\) has an \(A_{\infty}\)-multiplication, then the differential of the Hochschild complex is \(\overline{M}_{1}\) from Equation (12). _Remark 5.5_.: The functor \(S\mathfrak{s}\) is called the "oddification" of an operad in the literature [12]. The reader might find odd to define the Hochschild complex in this way instead of just \(\operatorname{End}_{A}\). The reason is that operadic suspension provides the necessary signs and the extra shift gives us the appropriate degrees. In addition, this definition allows the extra structure to arise naturally instead of having to define the signs by hand. For instance, if we have an associative multiplication \(m_{2}\in\operatorname{End}_{A}(2)=\operatorname{Hom}(A^{\otimes 2},A)\), the element \(m_{2}\) would not satisfy the equation \(m_{2}\circ m_{2}=0\) and thus cannot be used to induce a multiplication on \(\operatorname{End}_{A}\) as we did above. A natural question to ask is what relation there is between the \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\mathcal{O}\) and the one on \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\). In [10] it is claimed that given an operad \(\mathcal{O}\) with an \(A_{\infty}\)-multiplication, the map \[\mathcal{O}\to\operatorname{End}_{\mathcal{O}},\,x\mapsto\sum_{n\geq 0}b_{n}(x;-)\] is a morphism of \(A_{\infty}\)-algebras. In the associative case, this result leads to the definition of homotopy \(G\)-algebras, which connects with the classical Deligne conjecture. We are going to adapt the statement of this claim to our context and prove it. This way we will obtain an \(A_{\infty}\)-version of homotopy \(G\)-algebras and consequently an \(A_{\infty}\)-version of the Deligne conjecture. Let \(\Phi^{\prime}\) the map defined as above but on \(\mathfrak{s}\mathcal{O}\), i.e. \[\Phi^{\prime}\colon\mathfrak{s}\mathcal{O}\to\operatorname{End}_{\mathfrak{s }\mathcal{O}},\,x\mapsto\sum_{n\geq 0}b_{n}(x;-).\] Let \(\Phi:S\mathfrak{s}\mathcal{O}\to S\mathfrak{s}\operatorname{End}_{S \mathfrak{s}\mathcal{O}}\) the map making the following diagram commute (13) where the isomorphism \(\operatorname{End}_{\mathfrak{s}\mathcal{O}}\cong\mathfrak{s}\operatorname{End }_{S\mathfrak{s}\mathcal{O}}\) is given in Equation (9). Note that the degree of the map \(\Phi\) is zero. _Remark 5.6_.: Notice that we have only used the operadic structure on \(\mathfrak{s}\mathcal{O}\) to define an \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\mathcal{O}\), so the constructions and results in these sections are valid if we replace \(\mathfrak{s}\mathcal{O}\) by any graded module \(A\) such that \(SA\) is an \(A_{\infty}\)-algebra. **Theorem 5.7**.: _The map \(\Phi\) defined in diagram (13) above is a morphism of \(A_{\infty}\)-algebras, i.e. for all \(j\geq 1\) the equation_ \[\Phi(M_{j})=\overline{M}_{j}(\Phi^{\otimes j})\] _holds, where the \(M_{j}\) is the \(j\)-th component of the \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\mathcal{O}\) and \(\overline{M}_{j}\) is the \(j\)-th component of the \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\)._ Proof.: Let us have a look at the following diagram (14) where the diagonal red arrows are shifts of graded \(R\)-modules. We need to show that the diagram defined by the external black arrows commutes. But these arrows are defined so that they commute with the red and blue arrows, so it is enough to show that the inner blue diagram commutes. The blue diagram can be split into two different squares using the dashed arrow \(\mathcal{M}_{j}\) that we are going to define next, so it will be enough to show that the two squares commute. The map \(\mathcal{M}_{j}:(\operatorname{End}_{\mathfrak{s}\mathcal{O}})^{\otimes j} \rightarrow\operatorname{End}_{\mathfrak{s}\mathcal{O}}\) is defined by \[\mathcal{M}_{j}(f_{1},\dots,f_{j}) =B_{j}(M^{\prime};f_{1},\dots,f_{j}) \text{for }j>1,\] \[\mathcal{M}_{1}(f) =B_{1}(M^{\prime};f)-(-1)^{|f|}B_{1}(f;M^{\prime}),\] where \(B_{j}\) is the natural brace structure map on the operad \(\operatorname{End}_{\mathfrak{s}\mathcal{O}}\), i.e. for \(f\in\operatorname{End}_{\mathfrak{s}\mathcal{O}}(n)\), \[B_{j}(f;f_{1},\dots,f_{j}) =\sum_{k_{0}+\dots+k_{j}=n-j}f(1^{\otimes k_{0}}\otimes f_{1} \otimes 1^{\otimes k_{1}}\otimes\dots\otimes f_{j}\otimes 1^{\otimes k_{j}}).\] The \(1\)'s in the braces are identity maps. In the above definition, \(|f|\) denotes the degree of \(f\) as an element of \(\operatorname{End}_{\mathfrak{s}\mathcal{O}}\), which is the same as the degree \(\overline{\sigma}(f)\in\mathfrak{s}\operatorname{End}_{\mathfrak{s}\mathcal{O}}\) because \(\overline{\sigma}\) is an isomorphism, as mentioned in Equation (9). The inner square of diagram (14) is divided into two halves, so we divide the proof into two as well, showing the commutativity of each half independently. **Commutativity of the right blue square.** Let us show now that the right square commutes. Recall that \(\overline{\sigma}\) is an isomorphism of operads and \(M=\overline{\sigma}(M^{\prime})\). Then we have for \(j>1\) \[\overline{M}^{\prime}_{j}(\overline{\sigma}(f_{1}),\dots,\overline{\sigma}(f_ {j}))=\overline{B}_{j}(M;\overline{\sigma}(f_{1}),\dots,\overline{\sigma}(f_ {j}))=\overline{B}_{j}(\overline{\sigma}(M^{\prime});\overline{\sigma}(f_{1}), \dots,\overline{\sigma}(f_{j})).\] Now, since the brace structure is defined as an operadic composition, it commutes with \(\overline{\sigma}\), so \[\overline{B}_{j}(\overline{\sigma}(M^{\prime});\overline{\sigma}(f_{1}), \dots,\overline{\sigma}(f_{j}))=\overline{\sigma}(B_{j}(M^{\prime};f_{1}, \dots,f_{j}))=\overline{\sigma}(\mathcal{M}_{j}(f_{1},\dots,f_{j})),\] and therefore the right blue square commutes for \(j>1\). For \(j=1\) the result follows analogously. The proof that the left blue square commutes consists of several lengthy calculations so we are going to devote the next section to that. However, it is worth noting that the commutativity of the left square does not depend on the particular operad \(\mathfrak{s}\mathcal{O}\), so it is still valid if \(m\) satisfies \(m\circ m=0\) for any circle operation defined in terms of insertions. This is essentially the original statement in [10]. ### Commutativity of the left blue square We are going to show here that the left blue square in diagram (14) commutes, i.e. that \[\Phi^{\prime}(M^{\prime}_{j})=\mathcal{M}_{j}((\Phi^{\prime})^{\otimes j}) \tag{15}\] for all \(j\geq 1\). First we prove the case \(j>1\). Let \(x_{1},\ldots,x_{j}\in\mathfrak{s}\mathcal{O}^{\otimes j}\). We have on the one hand \[\Phi^{\prime}(M^{\prime}_{j}(x_{1},\ldots,x_{j})) =\Phi^{\prime}(b_{j}(m;x_{1},\ldots,x_{j}))=\sum_{n\geq 0}b_{n}(b_{j} (m;x_{1},\ldots,x_{j});-)\] \[=\sum_{n}\sum_{l}\sum b_{l}(m;-,b_{i_{1}}(x_{1};-),\cdots,b_{i_{ j}}(x_{j};-),-)\] where \(l=n-(i_{1}+\cdots+i_{j})+j\). The sum with no subindex runs over all the possible order-preserving insertions. Note that \(l\geq j\). Evaluating the above map on elements would yield Koszul signs coming from the brace relation. Also recall from Lemma 5.1 that \(|b_{j}(x;-)|=|x|\). Now, fix some value of \(l\geq j\) and let us compute the \(M^{\prime}_{l}\) component of \[\mathcal{M}_{j}(\Phi^{\prime}(x_{1}),\ldots,\Phi^{\prime}(x_{j}))=B_{j}(M^{ \prime};\Phi^{\prime}(x_{1}),\ldots,\Phi^{\prime}(x_{j}))\] that is, \(B_{j}(M^{\prime}_{l};\Phi^{\prime}(x_{1}),\ldots,\Phi^{\prime}(x_{j}))\). By definition, this equals \[\sum M^{\prime}_{l}(-,\Phi^{\prime}(x_{1}),\cdots,\Phi^{\prime}( x_{j}),-) =\sum_{i_{1},\ldots,i_{j}}\sum M^{\prime}_{l}(-,b_{i_{1}}(x_{1};-),\cdots,b_{i_{j}}(x_{j};-),-)\] \[=\sum_{i_{1},\ldots,i_{j}}\sum b_{l}(m;-,b_{i_{1}}(x_{1};-), \cdots,b_{i_{j}}(x_{j};-),-).\] We are using hyphens instead of \(1\)'s to make the equality of both sides of the Equation (15) more apparent, and to make clear that when evaluating on elements those are the places where the elements go. For each tuple \((i_{1},\ldots,i_{j})\) we can choose \(n\) such that \(n-(i_{1}+\cdots+i_{j})+j=l\), so the above sum equals \[\sum_{\begin{subarray}{c}n,i_{1},\ldots,i_{j}\\ n-(i_{1}+\cdots+i_{j})+j=l\end{subarray}}b_{l}(m;-,b_{i_{1}}(x_{1};-),\cdots,b _{i_{j}}(x_{j};-),-).\] So each \(M^{\prime}_{l}\) component for \(l\geq j\) produces precisely the terms \(b_{l}(m;\dots)\) appearing in \(\Phi^{\prime}(M^{\prime}_{j})\). Conversely, for every \(n\geq 0\) there exists some tuple \((i_{1},\dots,i_{j})\) and some \(l\geq j\) such that \(n\) is the that \(n-(i_{1}+\dots+i_{j})+j=l\), so we do get all the summands from the left hand side of Equation (15), and thus we have the equality \(\Phi^{\prime}(M^{\prime}_{j})=\mathcal{M}_{j}((\Phi^{\prime})^{\otimes j})\) for all \(j>1\). It is worth treating the case \(n=0\) separately since in that case we have the summand \(b_{0}(b_{j}(m;x_{1},\dots,x_{j}))\) in \(\Phi^{\prime}(b_{j}(m;x_{1},\dots,x_{j}))\), where we cannot apply the brace relation. This summand is equal to \[B_{j}(M^{\prime}_{j};b_{0}(x_{1}),\dots,b_{0}(x_{j}))=M^{\prime}_{j}(b_{0}(x_{ 1}),\dots,b_{0}(x_{j}))=b_{j}(m;b_{0}(x_{1}),\dots,b_{0}(x_{j})),\] since by definition \(b_{0}(x)=x\). Now we are going to show the case \(j=1\), that is \[\Phi^{\prime}(M^{\prime}_{1}(x))=\mathcal{M}_{1}(\Phi^{\prime}(x)). \tag{16}\] This is going to be divided into two parts, since \(M^{\prime}_{1}\) has two clearly distinct summands, one of them consisting of braces of the form \(b_{l}(m;\cdots)\) (insertions in \(m\)) and another one consisting of braces of the form \(b_{l}(x;\cdots)\) (insertions in \(x\)). We will therefore show that both types of braces cancel on each side of Equation (16). #### Insertions in \(m\) Let us first focus on the insertions in \(m\) that appear in Equation (16). Recall that \[\Phi^{\prime}(M^{\prime}_{1}(x))=\Phi^{\prime}([m,x])=\Phi^{\prime}(b_{1}(m;x ))-(-1)^{|x|}\Phi^{\prime}(b_{1}(x;m)) \tag{17}\] so we focus on the first summand \[\Phi^{\prime}(b_{1}(m;x))= \sum_{n}b_{n}(b_{1}(m;x);-)=\sum_{n}\sum_{\begin{subarray}{c}i \\ n\geq i\end{subarray}}\sum b_{n-i+1}(m;-,b_{i}(x;-),-)\] \[= \sum_{\begin{subarray}{c}n,i\\ n-i+1>0\end{subarray}}\sum_{b_{n-i+1}(m;-,b_{i}(x;-),-)\] where the sum with no indices runs over all the positions in which \(b_{i}(x;-)\) can be inserted (from \(1\) to \(n-i+1\) in this case). On the other hand, since \(|\Phi^{\prime}(x)|=|x|\), the right hand side of Equation (16) becomes \[\mathcal{M}_{1}(\Phi^{\prime}(x))=B_{1}(M^{\prime};\Phi^{\prime}(x))-(-1)^{|x |}B_{1}(\Phi^{\prime}(x);M^{\prime}). \tag{18}\] Again, we are focusing now on the first summand, but with the exception of the part of \(M^{\prime}_{1}\) that corresponds to \(b_{1}(\Phi(x);m)\). From here the argument is a particular case of the proof for \(j>1\), so the terms of the form \(b_{l}(m;\cdots)\) are the same on both sides of the Equation (16). _Insertions in \(x\)._ And now, let us study the insertions in \(x\) that appear in Equation (16). We will check that insertions in \(x\) from the left hand side and right hand side cancel. Let us look first at the left hand side. From \(\Phi^{\prime}(M_{1}^{\prime}(x))\) in Equation (17) we had \[-(-1)^{|x|}\Phi^{\prime}(b_{1}(x;m))=-(-1)^{|x|}\sum_{n}b_{n}(b_{1}(x;m);-).\] The factor \(-(-1)^{|x|}\) is going to appear everywhere, so we may cancel it. Thus we just have \[\Phi^{\prime}(b_{1}(x;m))=\sum_{n}b_{n}(b_{1}(x;m);-).\] We are going to evaluate each term of the sum, so let \(z_{1},\ldots,z_{n}\in\mathfrak{s}\mathcal{O}\). We have by the brace relation that \[b_{n}(b_{1}(x;m);z_{1},\ldots,z_{n})=\sum_{l+j=n+1}\sum_{i=1}^{n -j+1}(-1)^{\varepsilon}b_{l}(x;z_{1},\ldots,b_{j}(m;z_{i},\ldots,z_{i+j}), \ldots,z_{n}) \tag{19}\] \[+\sum_{i=1}^{n+1}(-1)^{\varepsilon}b_{n+1}(x;z_{1},\ldots,z_{i- 1},m,z_{i},\ldots,z_{n}),\] where \(\varepsilon\) is the usual Koszul sign with respect to the grading in \(\mathfrak{s}\mathcal{O}\). We have to check that the insertions in \(x\) that appear in \(\mathcal{M}_{1}(\Phi^{\prime}(x))\) (right hand side of the eq. (16)) are exactly those in Equation (19) above (left hand side of eq. (16)). Therefore let us look at the right hand side of Equation (16). Here we will study the cancellations from each of the two summands that naturally appear. From Equation (18), i.e. \(\mathcal{M}_{1}(\Phi^{\prime}(x))=B_{1}(M^{\prime};\Phi^{\prime}(x))-(-1)^{| x|}B_{1}(\Phi^{\prime}(x);M^{\prime})\) we have \[-(-1)^{|x|}b_{1}(\Phi^{\prime}(x);m)=-(-1)^{|x|}\sum_{n}b_{1}(b_{n}(x;-);m)\] coming from the first summand since \(B_{1}(M_{1}^{\prime};\Phi^{\prime}(x))=M_{1}^{\prime}(\Phi^{\prime}(x))\). We are now only interested in insertions in \(x\). Again, cancelling \(-(-1)^{|x|}\) we get \[b_{1}(\Phi^{\prime}(x);m)=\sum_{n}b_{1}(b_{n}(x;-);m).\] Each term of the sum can be evaluated on \((z_{1},\ldots,z_{n})\) to produce \[b_{1}(b_{n}(x;z_{1},\ldots,z_{n});m)=\] \[\sum_{i=1}^{n}(-1)^{\varepsilon+|z_{i}|}b_{n}(x;z_{1},\ldots,b_{ 1}(z_{i};m),\ldots,z_{n})+\sum_{i=1}^{n+1}(-1)^{\varepsilon}b_{n+1}(x;z_{1}, \ldots,z_{i-1},m,z_{i},\ldots,z_{n}) \tag{20}\] Note that we have to apply the Koszul sign rule twice: once at evaluation, and once more to apply the brace relation. Now, from the second summand of \(\mathcal{M}_{1}(\Phi^{\prime}(x))\) in the right hand side of eq. (18), after cancelling \(-(-1)^{|x|}\) we obtain \[B_{1}(\Phi^{\prime}(x);M^{\prime})= \sum_{l}B_{1}(b_{l}(x;-);M^{\prime})=\sum_{l}\sum b_{l}(x;-,M^{ \prime},-)\] \[= \left(\sum_{j>1}\sum_{l}\sum b_{l}(x;-,b_{j}(m;-),-)+\sum_{l}\sum b _{l}(x;-,b_{1}(-;m),-)\right).\] We are going to evaluate on \((z_{1},\ldots,z_{n})\) to make this map more explicit, giving us \[\sum_{l+j=n+1}\ \sum_{i=1}^{n-j+1}(-1)^{\varepsilon}b_{l}(x;z_{1}, \ldots,b_{j}(m;z_{i},\ldots,z_{i+j}),\ldots,z_{n})\] \[\qquad\qquad-\sum_{i=1}^{n}(-1)^{\varepsilon+|z_{i}|}b_{n}(x;z_{ 1},\ldots,b_{1}(z_{i};m),\ldots,z_{n}). \tag{21}\] The minus sign comes from the fact that \(b_{1}(z_{i};m)\) comes from \(M^{\prime}_{1}(z_{i})\), so we apply the signs in the definition of \(M^{\prime}_{1}(z_{i})\). We therefore have that the right hand side of eq. (18) is the result of adding equations (20) and (21). After this addition we can see that the first sum of eq. (20) cancels the second sum of eq. (21). We also have that the second sum in eq. (20) is the same as the second sum in eq. (19), so we are left with only the first sum of eq. (21). This is the same as the first sum in eq. (19), so we have already checked that the equation \(\Phi^{\prime}(M^{\prime}_{1})=\mathcal{M}_{1}(\Phi^{\prime})\) holds. In the case \(n=0\), we have to note that \(B_{1}(b_{0}(x);m)\) vanishes because of arity reasons: \(b_{0}(x)\) is a map of arity \(0\), so we cannot insert any inputs. And this finishes the proof. ### Explicit \(A_{\infty}\)-algebra structure and Deligne conjecture We have given an implicit definition of the components of the \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\mathcal{O}\), namely, \[M_{j}=\overline{\sigma}(M^{\prime}_{j})=(-1)^{\binom{j}{2}}S\circ M^{\prime}_ {j}\circ(S^{-1})^{\otimes j},\] but it is useful to have an explicit expression that determines how it is evaluated on elements of \(S\mathfrak{s}\mathcal{O}\). We will need these explicit expressions to describe \(J\)-algebras, which are \(A_{\infty}\)-version of homotopy \(G\)-algebras. This way we can state the \(A_{\infty}\)-Deligne conjecture in a more precise way. This explicit formulas will also clear up the connection with the work of Gerstenhaber and Voronov. We hope that these explicit expression can be useful to perform calculations in other mathematical contexts where \(A_{\infty}\)-algebras are used. **Lemma 5.8**.: _For \(x,x_{1},\ldots,x_{n}\in\mathfrak{s}\mathcal{O}\), we have the following expressions._ \[M_{n}(Sx_{1},\ldots,Sx_{n})=(-1)^{\sum_{i=1}^{n}(n-i)|x_{i}|}Sb_{ n}(m;x_{1},\ldots,x_{n})\qquad\qquad n>1,\] \[M_{1}(Sx)=Sb_{1}(m;x)-(-1)^{|x|}Sb_{1}(x;m).\] _Here \(|x|\) is the degree of \(x\) as an element of \(\mathfrak{s}\mathcal{O}\), i.e. the natural degree._ Proof.: The deduction of these explicit formulas is done as follows. Let \(n>1\) and \(x_{1},\ldots,x_{n}\in\mathsf{s}\mathcal{O}\). Then \[M_{n}(Sx_{1},\ldots,Sx_{n}) =SM_{n}^{\prime}((S^{\otimes n})^{-1})(Sx_{1},\ldots,Sx_{n})\] \[=(-1)^{\binom{n}{2}}SM_{n}^{\prime}((S^{-1})^{\otimes n})(Sx_{1}, \ldots,Sx_{n})\] \[=(-1)^{\binom{n}{2}+\sum_{i=1}^{n}(n-i)(|x_{i}|+1)}SM_{n}^{\prime }(S^{-1}Sx_{1},\ldots,S^{-1}Sx_{n})\] \[=(-1)^{\binom{n}{2}+\sum_{i=1}^{n}(n-i)(|x_{i}|+1)}SM_{n}^{\prime }(x_{1},\ldots,x_{n})\] Now, note that \(\binom{n}{2}\) is even exactly when \(n\equiv 0,1\mod 4\). In these cases, an even amount of \(|x_{i}|\)'s have an odd coefficient in the sum (when \(n\equiv 0\mod 4\) these are the \(|x_{i}|\) with even index, and when \(n\equiv 1\mod 4\), the \(|x_{i}|\) with odd index). This means that \(1\) is added on the exponent an even number of times, so the sign is not changed by the binomial coefficient nor by adding \(1\) on each term. Similarly, when \(\binom{n}{2}\) is odd, i.e. when \(n\equiv 2,3\mod 4\), there is an odd number of \(|x_{i}|\) with odd coefficient, so the addition of \(1\) an odd number of times cancels the binomial coefficient. This means that the above expression equals \((-1)^{\sum_{i=1}^{n}(n-i)|x_{i}|}SM_{n}^{\prime}(x_{1},\ldots,x_{n})\), which by definition equals \[(-1)^{\sum_{i=1}^{n}(n-i)|x_{i}|}Sb_{n}(m;x_{1},\ldots,x_{n}).\] The case \(n=1\) is analogous since \(\overline{\sigma}\) is linear. It is possible to show that the maps defined explicitly as we have just done satisfy the \(A_{\infty}\)-equation without relying on the fact that \(\overline{\sigma}\) is a map of operads, but it is a lengthy and tedious calculation. _Remark 5.9_.: In the case \(n=2\), omitting the shift symbols by abuse of notation, we obtain \[M_{2}(x,y)=(-1)^{|x|}b_{2}(m;x,y).\] Let \(M_{2}^{GV}\) be the product defined in [10] as \[M_{2}^{GV}(x,y)=(-1)^{|x|+1}b_{2}(m;x,y).\] We see that \(M_{2}=-M_{2}^{GV}\). Since the authors of [10] work in the associative case \(m=m_{2}\), this minus sign does not affect the \(A_{\infty}\)-relation, which in this case reduces to the associativity and differential relations. This difference in sign can be explained by the difference between \((S^{\otimes n})^{-1}\) and \((S^{-1})^{\otimes n}\), since any of these maps can be used to define a map \((S\mathsf{s}\mathcal{O})^{\otimes n}\to\mathsf{s}\mathcal{O}^{\otimes n}\). Now that we have the explicit formulas for the \(A_{\infty}\)-structure on \(S\mathsf{s}\mathcal{O}\) we can state and prove an \(A_{\infty}\)-version of the Deligne conjecture. Let us first re-adapt the definition of homotopy \(G\)-algebra from [10, Definition 2] to our conventions. **Definition 5.10**.: _A homotopy \(G\)-algebra is differential graded algebra \(V\) with a differential \(M_{1}\) and a product \(M_{2}\) such that the shift \(S^{-1}V\) is a brace algebra with brace maps \(b_{n}\). The product differential and the product must satisfy the following compatibility identities. Let \(x,x_{1},x_{2},y_{1},\ldots,y_{n}\in S^{-1}V\). We demand_ \[Sb_{n}(S^{-1}M_{2}(Sx_{1},Sx_{2});y_{1},\ldots,y_{n})=\] \[\sum_{k=0}^{n}(-1)^{(|x_{2}|+1)\sum_{i=1}^{k}|y_{i}|}M_{2}(b_{k}( x_{1};y_{1},\ldots,y_{k}),b_{n-k}(x_{2};y_{k+1},\ldots,y_{n}))\] _and_ \[Sb_{n}(S^{-1}M_{1}(Sx);y_{1},\ldots,y_{n})-M_{1}(Sb_{n}(x;y_{1}, \ldots,y_{n}))\] \[-(-1)^{|x|+1}\sum_{p=1}^{n}(-1)^{\sum_{i=1}^{p}|y_{i}|}Sb_{n}(x;y _{1},\ldots,M_{1}(Sy_{p}),\ldots,y_{n})\] \[= -(-1)^{(|x|+1)|y_{i}|}M_{2}(Sy_{1},Sb_{n-1}(x;y_{2},\ldots,y_{n}))\] \[+(-1)^{|x|+1}\sum_{p=1}^{n-1}(-1)^{n-1+\sum_{i=1}^{p}|y_{i}|}Sb_{ n-1}(x;y_{1},\ldots,M_{2}(Sy_{p},Sy_{p+1}),\ldots y_{n})\] \[-(-1)^{|x|+\sum_{i=1}^{n-1}|y_{i}|}M_{2}(Sb_{n-1}(x;y_{1},\ldots, y_{n-1}),Sy_{n})\] Notice that our signs are slightly different to those in [10] as a consequence of our conventions. Our signs will be a particular case of those in Definition 5.11, which are set so that Corollary 5.12 holds in consistent way with operadic suspension and all the shifts that the authors of [10] do not consider. We now introduce \(J\)-algebras as an \(A_{\infty}\)-generalization of homotopy \(G\)-algebras. This will allow us to generalize the Deligne conjecture to the \(A_{\infty}\)-setting. **Definition 5.11**.: _A \(J\)-algebra \(V\) is an \(A_{\infty}\)-algebra with structure maps \(\{M_{j}\}_{j\geq 1}\) such that the shift \(S^{-1}V\) is a brace algebra. Furthermore, the braces and the \(A_{\infty}\)-structure satisfy the following compatibility relations. Let \(x,x_{1},\ldots,x_{j},y_{1},\ldots,y_{n}\in S^{-1}V\). For \(n\geq 0\) we demand_ \[(-1)^{\sum_{i=1}^{n}(n-i)|y_{i}|}Sb_{n}(S^{-1}M_{1}(Sx);y_{1}, \ldots,y_{n})=\] \[\sum_{\begin{subarray}{c}l+k-1=n\\ 1\leq i_{1}\leq n-k+1\end{subarray}} (-1)^{\varepsilon}M_{l}(Sy_{1},\ldots,Sb_{k}(x;y_{i_{1}}, \ldots),\ldots,Sy_{n})\] \[-(-1)^{|x|}\hskip-14.226378pt\sum_{\begin{subarray}{c}l+k-1=n\\ 1\leq i_{1}\leq n-k+1\end{subarray}} (-1)^{\eta}Sb_{k}(x;y_{1},\ldots,S^{-1}M_{l}(Sy_{i_{1}}, \ldots),\ldots,y_{n})\] _where_ \[\varepsilon=\sum_{v=1}^{i_{1}-1}|y_{v}|(|x|-k+1)+\sum_{v=1}^{k}|y_{i_{1}+v-1}|(k- v)+(l-i_{1})|x|.\] _and_ \[\eta=\sum_{v=1}^{i_{1}-1}(k-v)|y_{v}|+\sum_{v=1}^{i_{1}-1}l|y_{v}|+\sum_{v=i_{1 }}^{i_{1}+l-1}(k-i_{1})|y_{v}|+\sum_{v=i_{1}}^{n-l}(k-v)|y_{v+l}|\] _For \(j>1\) we demand_ \[(-1)^{\sum_{i=1}^{n}(n-i)|y_{i}|}Sb_{n}(S^{-1}M_{j}(Sx_{1},\ldots,Sx _{j});y_{1},\ldots,y_{n})=\] \[\sum(-1)^{\varepsilon}M_{l}(Sy_{1},\ldots,Sb_{k_{1}}(x_{1};y_{i_{ 1}},\ldots),\ldots,Sb_{k_{j}}(x_{j};y_{i_{j}},\ldots),\ldots,Sy_{n}).\] _The unindexed sum runs over all possible choices of non-negative integers that satisfy \(l+k_{1}+\cdots+k_{j}-j=n\) and over all possible ordering-preserving insertions. The right hand side sign is given by_ \[\varepsilon= \sum_{\begin{subarray}{c}1\leq t\leq j\\ 1\leq v\leq k_{t}\end{subarray}}|y_{i_{t}+v-1}|(k_{t}-v)+\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\! theory where projectivity cannot be guaranteed. In 2008, Sagave introduced the notion of derived \(A_{\infty}\)-algebras, providing a framework for not necessarily projective modules over an arbitrary commutative ground ring [10]. In this section we recall some definitions and results about derived \(A_{\infty}\)-algebras and present some new ways of interpreting them in terms of operads and collections. We also recall the notion of filtered \(A_{\infty}\)-algebra, since it will play a role in obtaining derived \(A_{\infty}\)-algebras from \(A_{\infty}\)-algebras on totalization. ### Derived \(A_{\infty}\)-algebras In the following definition we use the notation in [11]. **Definition 6.1**.: _A derived \(A_{\infty}\)-algebra on a \((\mathbb{Z},\mathbb{Z})\)-bigraded \(R\)-module \(A\) consist of a family of \(R\)-linear maps_ \[m_{ij}:A^{\otimes j}\to A\] _of bidegree \((i,2-(i+j))\) for each \(j\geq 1\), \(i\geq 0\), satisfying the equation_ \[\sum_{\begin{subarray}{c}u=i+p,v=j+q-1\\ j=r+1+t\end{subarray}}(-1)^{rq+t+pj}m_{ij}(1^{\otimes r}\otimes m_{pq}\otimes 1 ^{\otimes t})=0 \tag{22}\] _for all \(u\geq 0\) and \(v\geq 1\)._ According to the above definition, there are two equivalent ways of defining the operad of derived \(A_{\infty}\)-algebras \(d\mathcal{A}_{\infty}\) depending on the underlying category. One of them works on the category of bigraded modules \(\mathrm{bgMod}_{R}\) and the other one is suitable for the category of vertical bicomplexes \(\mathrm{vbC}_{R}\). We give the two of them here as we are going to use both. **Definition 6.2**.: _The operad \(d\mathcal{A}_{\infty}\) in \(\mathrm{bgMod}_{R}\) is the operad generated by \(\{m_{ij}\}_{i\geq 0,j\geq 1}\) subject to the derived \(A_{\infty}\)-relation_ \[\sum_{\begin{subarray}{c}u=i+p,v=j+q-1\\ j=r+1+t\end{subarray}}(-1)^{rq+t+pj}\gamma(m_{ij};1^{r},m_{pq},1^{t})=0\] _for all \(u\geq 0\) and \(v\geq 1\)._ _The operad \(d\mathcal{A}_{\infty}\) in \(\mathrm{vbC}_{R}\) is the quasi-free operad generated by \(\{m_{ij}\}_{(i,j)\neq(0,1)}\) with vertical differential given by_ \[\partial_{\infty}(m_{uv})=-\sum_{\begin{subarray}{c}u=i+p,v=j+q-1\\ j=r+1+t,(i,j)\neq(0,1)\neq(p,q)\end{subarray}}(-1)^{rq+t+pj}\gamma(m_{ij};1^{ r},m_{pq},1^{t}).\] **Definition 6.3**.: _Let \(A\) and \(B\) be derived \(A_{\infty}\)-algebras with respective structure maps \(m^{A}\) and \(m^{B}\). An \(\infty\)-morphism of derived \(A_{\infty}\)-algebras \(f:A\to B\) is a family of maps \(f_{st}:A^{\otimes t}\to B\) of bidegree \((s,1-s-t)\) satisfying_ \[\sum_{\begin{subarray}{c}u=i+p,v=j+q-1\\ j=r+1+t\end{subarray}}(-1)^{rq+t+pj}f_{ij}(1^{\otimes r}\otimes m_{pq}^{A} \otimes 1^{\otimes s})=\sum_{\begin{subarray}{c}u=i+p_{1}+\dots+p_{j}\\ v=q_{1}+\dots+q_{j}\end{subarray}}(-1)^{\epsilon}m_{ij}^{B}(f_{p_{1}q_{1}} \otimes\dots\otimes f_{p_{j}q_{j}}) \tag{23}\] _for all \(u\geq 0\) and \(v\geq 1\), where_ \[\epsilon=u+\sum_{1\leq w<l\leq j}q_{w}(1-p_{l}-q_{l})+\sum_{w=1}^{j}p_{w}(j-w).\] **Example 6.4**.: 1. _An_ \(A_{\infty}\)_-algebra is the same as a derived_ \(A_{\infty}\)_-algebra such that_ \(m_{ij}=0\) _for all_ \(i>0\)_._ 2. _One can check that, on any derived_ \(A_{\infty}\)_-algebra_ \(A\)_, the maps_ \(d_{i}=(-1)^{i}m_{i1}\) _define a twisted complex structure. This leads to the possibility of defining a derived_ \(A_{\infty}\)_-algebra as a twisted complex with some extra structure, see Remark_ 8.4_._ Analogously to Definition 3.5, we have the following. **Definition 6.5**.: _A derived \(A_{\infty}\)-multiplication on a bigraded operad \(\mathcal{O}\) is a map of operads \(d\mathcal{A}_{\infty}\to\mathcal{O}\)._ ### Filtered \(A_{\infty}\)-algebras We will make use of the filtration induced by the totalization functor in order to relate classical \(A_{\infty}\)-algebras to derived \(A_{\infty}\)-algebras. For this reason, we recall the notion of filtered \(A_{\infty}\)-algebras. **Definition 6.6**.: _A filtered \(A_{\infty}\)-algebra is an \(A_{\infty}\)-algebra \((A,m_{i})\) together with a filtration \(\{F_{p}A^{i}\}_{p\in\mathbb{Z}}\) on each \(R\)-module \(A^{i}\) such that for all \(i\geq 1\) and all \(p_{1},\dots,p_{i}\in\mathbb{Z}\) and \(n_{1},\dots,n_{i}\geq 0\),_ \[m_{i}(F_{p_{1}}A^{n_{1}}\otimes\dots\otimes F_{p_{i}}A^{n_{i}})\subseteq F_{ p_{1}+\dots+p_{i}}A^{n_{1}+\dots+n_{i}+2-i}.\] _Remark 6.7_.: Consider \(\mathcal{A}_{\infty}\) as an operad in filtered complexes with the trivial filtration and let \(K\) be a filtered complex. There is a one-to-one correspondence between filtered \(A_{\infty}\)-algebra structures on \(K\) and morphisms of operads in filtered complexes \(\mathcal{A}_{\infty}\to\underline{\mathrm{End}}_{K}\) (recall \(\underline{\mathrm{Hom}}\) from Definition 2.8). To see this, notice that if one forgets the filtrations such a map of operads gives an \(A_{\infty}\)-algebra structure on \(K\). The fact that this is a map of operads in filtered complexes implies that all the \(m_{i}\)'s respect the filtrations. Since the image of \(\mathcal{A}_{\infty}\) lies in \(\mathrm{End}_{K}=F_{0}\underline{\mathrm{End}}_{K}\), if we regard \(\mathcal{A}_{\infty}\) as an operad in cochain complexes, then we get a one-to-one correspondence between filtered \(A_{\infty}\)-algebra structures on \(K\) and morphisms of operads in cochain complexes \(\mathcal{A}_{\infty}\to\mathrm{End}_{K}\). **Definition 6.8**.: _A morphism of filtered \(A_{\infty}\)-algebras from \((A,m_{i},F)\) to \((B,m_{i},F)\) is an \(\infty\)-morphism \(f:(A,m_{i})\to(B,m_{i})\) of \(A_{\infty}\)-algebras such that each map \(f_{j}:A^{\otimes j}\to A\) is compatible with filtrations, i.e._ \[f_{j}(F_{p_{1}}A^{n_{1}}\otimes\cdots\otimes F_{p_{j}}A^{n_{j}})\subseteq F_{p _{1}+\cdots+p_{j}}B^{n_{1}+\cdots+n_{j}+1-j},\] _for all \(j\geq 1\), \(p_{1},\ldots p_{j}\in\mathbb{Z}\) and \(n_{1},\ldots,n_{j}\geq 0\)._ We will study the notions from this section from an operadic point of view. For this purpose we introduce some useful constructions in the next section. ## 7. Operadic totalization and vertical operadic suspension In this section we apply the totalization functor defined in Section 2.2.3 to operads, defining a functor from operads in brigraded modules (resp. twisted complexes) to operads in graded modules (resp. cochain complexes). We also define a bigraded version of operadic suspension. The combination of these two devices provides the signs required to encode derived \(A_{\infty}\)-algebras in a very concise and practical way, similar to what we achieve for classical \(A_{\infty}\)-algebras in Section 3. ### Operadic totalization We use Proposition 2.21 and the fact that the image of an operad under a lax monoidal functor is also an operad [11, Proposition 3.1.1(a)] to guarantee that applying totalization on an operad will result again in an operad. Therefore, let \(\mathcal{O}\) be either a bigraded operad, i.e. an operad in te category of bigraded \(R\)-modules or an operad in twisted complexes. We define \(\operatorname{Tot}(\mathcal{O})\) as the operad of graded \(R\)-modules (or cochain complexes) for which \[\operatorname{Tot}(\mathcal{O}(n))^{d}=\bigoplus_{i<0}\mathcal{O}(n)_{i}^{d-i }\oplus\prod_{i\geq 0}\mathcal{O}(n)_{i}^{d-i}\] is the image of \(\mathcal{O}(n)\) under the totalization functor, and the insertion maps are given by the composition \[\operatorname{Tot}(\mathcal{O}(n))\otimes\operatorname{Tot}(\mathcal{O}(m)) \xrightarrow{\mu}\operatorname{Tot}(\mathcal{O}(n)\otimes\mathcal{O}(m)) \xrightarrow{\operatorname{Tot}(\circ_{r})}\operatorname{Tot}(\mathcal{O}(n +m-1)), \tag{24}\] that is explicitly \[(x\bar{\circ}_{r}y)_{k}=\sum_{k_{1}+k_{2}=k}(-1)^{k_{1}d_{2}}x_{k_{1}}\circ_{r }y_{k_{2}}\] for \(x=(x_{i})_{i}\in\operatorname{Tot}(\mathcal{O}(n))^{d_{1}}\) and \(y=(y_{j})_{j}\in\operatorname{Tot}(\mathcal{O}(m))^{d_{2}}\). More generally, operadic composition \(\bar{\gamma}\) is defined by the composite \[\operatorname{Tot}(\mathcal{O}(N))\otimes\operatorname{Tot}( \mathcal{O}(a_{1}))\otimes\cdots\otimes\operatorname{Tot}(\mathcal{O}(a_{N}))\] \[\xrightarrow{\mu}\operatorname{Tot}(\mathcal{O}(N)\otimes \mathcal{O}(a_{1})\otimes\cdots\otimes\mathcal{O}(a_{N}))\xrightarrow{ \operatorname{Tot}(\gamma)}\operatorname{Tot}\left(\mathcal{O}\left(\sum a _{i}\right)\right),\] This map can be computed explicitly by iteration of the insertion \(\bar{\circ}\), giving the following. **Lemma 7.1**.: _The operadic composition \(\bar{\gamma}\) on \(\operatorname{Tot}(\mathcal{O})\) is given by_ \[\bar{\gamma}(x;x^{1},\dots,x^{N})_{k}=\sum_{k_{0}+k_{1}+\dots+k_{N}=k}(-1)^{ \varepsilon}\gamma(x_{k_{0}};x^{1}_{k_{1}},\dots,x^{N}_{k_{N}})\] _for \(x=(x_{k})_{k}\in\operatorname{Tot}(\mathcal{O}(N))^{d_{0}}\) and \(x^{i}=(x^{i}_{k})_{k}\in\operatorname{Tot}(\mathcal{O}(a_{i}))^{d_{i}}\), where_ \[\varepsilon=\sum_{j=1}^{m}d_{j}\sum_{i=0}^{j-1}k_{i} \tag{25}\] _and \(\gamma\) is the operadic composition on \(\mathcal{O}\)._ Notice that the sign is precisely the same appearing in Equation (3). ### Vertical operadic suspension and totalization On a bigraded operad we can use operadic suspension on the vertical degree with analogue results to those of the graded case that we explored in Section 3. We define \(\Lambda(n)=S^{n-1}R\), where \(S\) is a vertical shift of degree so that \(\Lambda(n)\) is the underlying ring \(R\) concentrated in bidegree \((0,n-1)\). As in the single graded case, we express the basis element of \(\Lambda(n)\) as \(e^{n}=e_{1}\wedge\dots\wedge e_{n}\). The operad structure on the bigraded \(\Lambda=\{\Lambda(n)\}_{n\geq 0}\) is the same as in the graded case, see Equation (7). **Definition 7.2**.: _Let \(\mathcal{O}\) be a bigraded linear operad. The vertical operadic suspension \(\mathfrak{s}\mathcal{O}\) of \(\mathcal{O}\) is given arity-wise by \(\mathfrak{s}\mathcal{O}(n)=\mathcal{O}(n)\otimes\Lambda(n)\) with diagonal composition. Similarly, we define the vertical operadic desuspension \(\mathfrak{s}^{-1}\mathcal{O}(n)=\mathcal{O}(n)\otimes\Lambda^{-}(n)\)._ We may identify the elements of \(\mathcal{O}\) with the elements of \(\mathfrak{s}\mathcal{O}\). **Definition 7.3**.: _For \(x\in\mathcal{O}(n)\) of bidegree \((k,d-k)\), its natural bidegree in \(\mathfrak{s}\mathcal{O}\) is the pair \((k,d+n-k-1)\). To distinguish both degrees we call \((k,d-k)\) the internal bidegree of \(x\), since this is the degree that \(x\) inherits from the grading of \(\mathcal{O}\)._ If we write \(\circ_{r+1}\) for the operadic insertion on \(\mathcal{O}\) and \(\tilde{\circ}_{r+1}\) for the operadic insertion on \(\mathfrak{s}\mathcal{O}\), we may find a relation between the two insertion maps in a completely analogous way to Lemma 3.3. **Lemma 7.4**.: _For \(x\in\mathcal{O}(n)\) and \(y\in\mathcal{O}(m)^{q}_{l}\) we have_ \[x\tilde{\circ}_{r+1}y=(-1)^{(n-1)q+(n-1)(m-1)+r(m-1)}x\circ_{r+1}y. \tag{26}\] _Remark 7.5_.: As can be seen, this is the same sign as the single-graded operadic suspension but with vertical degree. In particular, this operation leads to the Lie bracket from [11], which implies that \(m=\sum_{i,j}m_{ij}\) is a derived \(A_{\infty}\)-multiplication if and only if for all \(u\geq 0\) \[\sum_{i+j=u}\sum_{l,k}(-1)^{i}m_{jl}\tilde{\circ}m_{ik}=0. \tag{27}\] In [11, Proposition 2.15] this equation is described in terms of a sharp operator \(\sharp\). We also get the bigraded version of Theorem 3.9 and the functorial properties that we studied for the single-graded case in Section 3.1 and Proposition 3.14. Now we are going to combine vertical operadic suspension and totalization. More precisely, the _totalized vertical suspension_ of a bigraded operad \(\mathcal{O}\) is the graded operad \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\). This operad has an insertion map explicitly given by \[(x\star_{r+1}y)_{k}=\sum_{k_{1}+k_{2}=k}(-1)^{(n-1)(d_{2}-k_{2}-m+1)+(n-1)(m-1 )+r(m-1)+k_{1}d_{2}}x_{k_{1}}\circ_{r+1}y_{k_{2}} \tag{28}\] for \(x=(x_{i})_{i}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(n))^{d_{1}}\) and \(x=(x_{j})_{j}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(m))^{d_{2}}\). As usual, denote \[x\star y=\sum_{r=0}^{m-1}x\star_{r+1}y.\] This star operation is precisely the star operation from [11, SS5.1], i.e. the convolution operation on \(\operatorname{Hom}((dAs)^{\operatorname{i}},\operatorname{End}_{A})\). In particular, we can recover the Lie bracket from in [11]. We will do this in Corollary 7.11. Before continuing, let us show a lemma that allows us to work only with the single-graded operadic suspension if needed. **Proposition 7.6**.: _For a bigraded operad \(\mathcal{O}\) we have an isomorphism \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\cong\mathfrak{s}\operatorname{ Tot}(\mathcal{O})\), where the suspension on the left hand side is the bigraded version and on the right hand side is the single-graded version._ Proof.: Note that we may identify each element \(x=(x_{k}\otimes e^{n})_{k}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(n))\) with the element \(x=(x_{k})_{k}\otimes e^{n}\in\mathfrak{s}\operatorname{Tot}(\mathcal{O}(n))\). Thus, for an element \((x_{k})_{k}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(n))\) the isomorphism is given by \[f:\operatorname{Tot}(\mathfrak{s}\mathcal{O}(n))\cong\mathfrak{s} \operatorname{Tot}(\mathcal{O}(n)),\,(x_{k})_{k}\mapsto((-1)^{kn}x_{k})_{k}\] Clearly, this map is bijective so we just need to check that it commutes with insertions. Recall from Equation (28) that the insertion on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) is given by \[(x\star_{r+1}y)_{k}=\sum_{k_{1}+k_{2}=k}(-1)^{(n-1)(d_{2}-k_{2}-n+1)+(n-1)(m-1) +r(m-1)+k_{1}d_{2}}x_{k_{1}}\circ_{r+1}y_{k_{2}}\] for \(x=(x_{i})_{i}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(n))^{d_{1}}\) and \(y=(y_{j})_{j}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(m))^{d_{2}}\). Similarly, we may compute the insertion on \(\mathfrak{s}\mathrm{Tot}(\mathcal{O})\) by combining the sign produced first by \(\mathfrak{s}\). This results in the following insertion map \[(x\star_{r+1}^{\prime}y)_{k}=\sum_{k_{1}+k_{2}=k}(-1)^{(n-1)(d_{2}-n+1)+(n-1)( m-1)+r(m-1)+k_{1}(d_{2}-m+1)}x_{k_{1}}\circ_{r+1}y_{k_{2}}.\] Now let us show that \(f(x\star y)=f(x)\star f(y)\). We have that \(f((x\star_{r+1}y))_{k}\) equals \[\sum_{k_{1}+k_{2}=k}(-1)^{k(n+m-1)+(n-1)(d_{2}-k_{2}-n+1)+(n-1)(m- 1)+r(m-1)+k_{1}d_{2}}x_{k_{1}}\circ_{r+1}y_{k_{2}}\] \[=\sum_{k_{1}+k_{2}=k}(-1)^{(n-1)(d_{2}-n+1)+(n-1)(m-1)+r(m-1)+k_ {1}(d_{2}-m+1)}f(x_{k_{1}})\circ_{r+1}f(y_{k_{2}})\] \[=(f(x)\star_{r+1}f(y))_{k}\] as desired. _Remark 7.7_.: As we mentioned in Remark 2.22, there exist other possible ways of totalizing by varying the natural transformation \(\mu\). For instance, we can choose the totalization functor \(\operatorname{Tot}^{\prime}\) which is the same as \(\operatorname{Tot}\) but with a natural transformation \(\mu^{\prime}\) defined in such a way that the insertion on \(\operatorname{Tot}^{\prime}(\mathcal{O})\) is defined by \[(x\diamond y)_{k}=\sum_{k_{1}+k_{2}=k}(-1)^{k_{2}n_{1}}x_{k_{1}}\circ y_{k_{2 }}.\] This is also a valid approach for our purposes and there is simply a sign difference, but we have chosen our convention to be consistent with other conventions, such as the derived \(A_{\infty}\)-equation. However, it can be verified that \(\operatorname{Tot}^{\prime}(\mathfrak{s}\mathcal{O})=\mathfrak{s}\mathrm{ Tot}^{\prime}(\mathcal{O})\). With the original totalization we have a non identity isomorphism given by Proposition 7.6. Similar relations can be found among the other alternatives mentioned in Remark 2.22. Using the operadic structure on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\), we can describe derived \(A_{\infty}\)-multiplications in a new conceptual way as we did in Lemma 3.6, with analogous proof. **Lemma 7.8**.: _A derived \(A_{\infty}\)-multiplication on a bigraded operad \(\mathcal{O}\) is equivalent to an element \(m\in\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) of degree 1 concentrated in positive arity such that \(m\star m=0\). _ From Lemma 7.8, we can proceed as in the proof of Proposition 5.3 to show that \(m\) determines an \(A_{\infty}\)-algebra structure on \(S\mathrm{Tot}(\mathfrak{s}\mathcal{O})\cong S\mathfrak{s}\mathrm{Tot}( \mathcal{O})\). The goal now is showing that this \(A_{\infty}\)-structure on \(S\mathrm{Tot}(\mathfrak{s}\mathcal{O})\) is equivalent to a derived \(A_{\infty}\)-structure on \(S\mathfrak{s}\mathcal{O}\) and compute the structure maps explicitly. We will do this in Section 8. Before that, let us explore the brace structures that appear from this new operadic constructions and use them to reinterpret derived \(\infty\)-morphisms and their composition. ### Bigraded braces and totalized braces We are going to define a brace structure on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) using totalization. First note that one can define bigraded braces just like in the single-graded case, only changing the sign \(\varepsilon\) in Definition 4.1 to be \(\varepsilon=\sum_{p=1}^{n}\sum_{q=i}^{i_{p}}\langle x_{p},y_{q}\rangle\) according to the bigraded sign convention. As one might expect, we can define bigraded brace maps \(b_{n}\) on a bigraded operad \(\mathcal{O}\) and also on its operadic suspension \(\mathfrak{s}\mathcal{O}\), obtaining similar signs as in the single-graded case, but with vertical (internal) degrees, see Proposition 4.3. We can also define braces on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) via operadic composition. In this case, these are usual single-graded braces. More precisely, we define the maps \[b_{n}^{\star}:\operatorname{Tot}(\mathfrak{s}\mathcal{O}(N))\otimes \operatorname{Tot}(\mathfrak{s}\mathcal{O}(a_{1}))\otimes\cdots\otimes \operatorname{Tot}(\mathfrak{s}\mathcal{O}(a_{n}))\to\operatorname{Tot}( \mathfrak{s}\mathcal{O}(N-\sum a_{i}))\] using the operadic composition \(\gamma^{\star}\) on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) as \[b_{n}^{\star}(x;x_{1},\dots,x_{n})=\sum\gamma^{\star}(x;1,\dots,1,x_{1},1, \dots,1,x_{n},1,\dots,1),\] where the sum runs over all possible ordering preserving insertions. The brace map \(b_{n}^{\star}(x;x_{1},\dots,x_{n})\) vanishes whenever \(n>N\) and \(b_{0}^{\star}(x)=x\). Operadic composition can be described in terms of insertions in the obvious way, namely \[\gamma^{\star}(x;y_{1},\dots,y_{N})=(\cdots(x\star_{1}y_{1})\star_{1+a(y_{1})} y_{2}\cdots)\star_{1+\sum a(y_{p})}y_{N}, \tag{29}\] where \(a(y_{p})\) is the arity of \(y_{p}\). If we want to express this composition in terms of the composition in \(\mathcal{O}\) we just have to find out the sign factor applying the same strategy as in the single-graded case. In fact, as we said, there is a sign factor that comes from vertical operadic suspension that is identical to the graded case, but replacing internal degree by internal vertical degree. This is the sign that determines the brace \(b_{n}\) on \(\mathfrak{s}\mathcal{O}\). Explicitly, it is given by the following lemma, whose proof is identical to the single-graded case, see Proposition 4.3. **Lemma 7.9**.: _For \(x\in\mathfrak{s}\mathcal{O}(N)\) and \(x_{i}\in\mathfrak{s}\mathcal{O}(a_{i})\) of internal vertical degree \(q_{i}\) (\(1\leq i\leq n\)), we have_ \[b_{n}(x;x_{1},\dots,x_{n})=\sum_{N-n=h_{0}+\dots+h_{n}}(-1)^{\eta}\gamma(x \otimes 1^{\otimes h_{0}}\otimes x_{1}\otimes\cdots\otimes x_{n}\otimes 1^{ \otimes h_{n}}),\] _where_ \[\eta=\sum_{0\leq j<l\leq n}h_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}+\sum_{j =1}^{n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)h_{l}.\] The other sign factor is produced by totalization. This was computed in Lemma 7.1. Combining both factors we obtain the following. **Lemma 7.10**.: _We have_ \[b_{j}^{\star}(x;x^{1},\dots,x^{N})_{k}=\sum_{\begin{subarray}{c}k_{0}+k_{1}+ \dots+k_{N}=k\\ h_{0}+h_{1}+\dots+h_{N}=j-N\end{subarray}}\hskip-14.226378pt(-1)^{\eta+\sum_{j=1 }^{m}d_{j}\sum_{i=0}^{j-1}k_{i}}\gamma(x_{k_{0}};1^{h_{0}},x^{1}_{k_{1}},1^{h_{ 1}},\dots,x^{N}_{k_{N}},1^{h_{N}}) \tag{30}\] _for \(x=(x_{k})_{k}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(N))^{d_{0}}\) and \(x^{i}=(x^{i}_{k})_{k}\in\operatorname{Tot}(\mathfrak{s}\mathcal{O}(a_{i}))^{d _{i}}\), where \(\eta\) is defined in Lemma 7.9._ **Corollary 7.11**.: _For \(\mathcal{O}=\operatorname{End}_{A}\), the endomorphism operad of a bigraded module, the brace \(b_{1}^{\star}(f;g)\) is the operation \(f\star g\) defined in [10] that induces a Lie bracket. More precisely,_ \[[f,g]=b_{1}(f;g)-(-1)^{NM}b_{1}(g;f)\] _for \(f\in\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{A})^{N}\) and \(g\in\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{A})^{M}\), is the same bracket that was defined in [10]._ Notice that in [10] the sign in the bracket is \((-1)^{(N+1)(M+1)}\), but this is because their total degree differs by one with respect to ours. ### Reinterpretation of derived \(\infty\)-morphisms Just like we did for graded modules on Section 4.2, for bigraded modules \(A\) and \(B\) we may define the collection \(\operatorname{End}_{B}^{A}=\{\operatorname{Hom}(A^{\otimes n},B)\}_{n\geq 1}\). Recall that this collection has a left module structure \(\operatorname{End}_{B}\circ\operatorname{End}_{B}^{A}\to\operatorname{End}_{B} ^{A}\) over \(\operatorname{End}_{B}\) given by composition of maps. Similarly, given a bigraded module \(C\), we can define composition maps \(\operatorname{End}_{C}^{B}\circ\operatorname{End}_{B}^{A}\to\operatorname{End}_ {C}^{A}\). The collection \(\operatorname{End}_{B}^{A}\) also has an infinitesimal right module structure \(\operatorname{End}_{B}^{A}\circ_{(1)}\operatorname{End}_{A}\to\operatorname{ End}_{B}^{A}\) over \(\operatorname{End}_{A}\) given by insertion of maps. Similarly to the single-graded case, we may describe derived \(\infty\)-morphisms in terms of the above operations, with analogous proof to Lemma 4.5. **Lemma 7.12**.: _A derived \(\infty\)-morphism of \(A_{\infty}\)-algebras \(A\to B\) with respective structure maps \(m^{A}\) and \(m^{B}\) is equivalent to an element \(f\in\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{B}^{A})\) of degree 0 concentrated in positive arity such that_ \[\rho(f\circ_{(1)}m^{A})=\lambda(m^{B}\circ f),\] _where_ \[\lambda:\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{B})\circ \operatorname{Tot}(\mathfrak{s}\operatorname{End}_{B}^{A})\to\operatorname{Tot }(\mathfrak{s}\operatorname{End}_{B}^{A})\] _is induced by the left module structure on \(\operatorname{End}_{B}^{A}\), and_ \[\rho:\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{B})\circ_{(1)} \operatorname{Tot}(\mathfrak{s}\operatorname{End}_{B}^{A})\to\operatorname{Tot }(\mathfrak{s}\operatorname{End}_{B}^{A})\] _is induced by the right infinitesimal module structure on \(\operatorname{End}_{B}^{A}\)._ _In addition, the composition of \(\infty\)-morphisms is given by the natural composition_ \[\operatorname{Tot}(\mathfrak{s}\operatorname{End}_{C}^{B})\circ\operatorname{ Tot}(\mathfrak{s}\operatorname{End}_{B}^{A})\to\operatorname{Tot}(\mathfrak{s} \operatorname{End}_{C}^{A}).\] In the case that \(f:A\to A\) is an \(\infty\)-endomorphism, the above definition can be written in terms of operadic composition as \(f\star m=\gamma^{\star}(m\circ f)\), where \(\gamma^{\star}\) is the composition map derived from the operation \(\star\), see Equation (29). ## 8. The derived \(A_{\infty}\)-structure on an operad In this section we finally establish the connection between classical and derived \(A_{\infty}\)-algebras through Theorem 8.1. From this, in Theorem 8.3 we are able to obtain explicit derived \(A_{\infty}\)-maps on \(A=S\mathfrak{s}\mathcal{O}\) for a sufficiently bounded operad \(\mathcal{O}\) with a derived \(A_{\infty}\)-multiplication. This opens the door to the formulation and proof on a new version of the Deligne conjecture in Corollary 8.10. We are going to follow a strategy inspired by the proof of the following theorem to show that there is a derived \(A_{\infty}\)-structure on \(A=S\mathfrak{s}\mathcal{O}\). The proof can be found in [1, Proposition 4.55]. We refer the reader to Section 2.2 to recall the definitions of the categories used. **Theorem 8.1**.: _Let \((A,d^{A})\in\operatorname{tC}_{R}^{b}\) be a twisted complex horizontally bounded on the right and \(A\) its underlying cochain complex. We have natural bijections_ \[\operatorname{Hom}_{\operatorname{vbOp},d^{A}}(d\mathcal{A}_{ \infty},\operatorname{End}_{A}) \cong\operatorname{Hom}_{\operatorname{vbOp}}(\mathcal{A}_{ \infty},\underline{\underline{\operatorname{\operatorname{\operatorname{ \operatorname{\operatorname{\operatorname{\operatorname{\operatorname{\operatorname{\operatorname{ \operatorname{\operatorname{\operatorname{\operatorname{\operatorname{\operatorname{ \operatorname{\ for each \(j>0\) we can only have finitely many non-zero components \(m_{ij}\). This situation happens in practice in all known examples of derived \(A_{\infty}\)-algebras so far, some of them are in [12, Remark 6.5], [13], and [1, SS5]. Under this assumption we may replace all direct products by direct sums. We also need to provide \(A\) with a twisted complex structure. The reason for this is that Theorem 8.1 uses the definition of derived \(A_{\infty}\)-algebras on an underlying twisted complex, see Remark 8.4. This follows from Corollary 8.6, which is a consequence of another version of this theorem that works for bigraded modules, Corollary 8.5. With these assumption, by Theorem 8.1 we can guarantee the existence of a derived \(A_{\infty}\)-algebra structure on \(A\) provided that \(\operatorname{Tot}(A)\) has an \(A_{\infty}\)-algebra structure. **Theorem 8.3**.: _Let \(A=S\mathfrak{s}\mathcal{O}\) where \(\mathcal{O}\) is an operad horizontally bounded on the right with a derived \(A_{\infty}\)-multiplication \(m=\sum_{ij}m_{ij}\in\mathcal{O}\). Let \(x_{1}\otimes\cdots\otimes x_{j}\in(A^{\otimes j})_{k}^{d-k}\) and let \(x_{v}=Sy_{v}\) for \(v=1,\ldots,j\) and \(y_{v}\) be of bidegree \((k_{v},d_{v}-k_{v})\). The following maps \(M_{ij}\) for \(j\geq 2\) determine a derived \(A_{\infty}\)-algebra structure on \(A\)._ \[M_{ij}(x_{1},\ldots,x_{j})=(-1)^{\sum_{v=1}^{j}(j-v)(d_{v}-k_{v})}\sum_{l}Sb_{ j}(m_{il};y_{1},\ldots,y_{j}).\] Note that we abuse of notation and identify \(x_{1}\otimes\cdots\otimes x_{j}\) with an element of \(\operatorname{Tot}(A^{\otimes j})\) with only one non-zero component. For a general element, extend linearly. Proof.: Since \(m\) is a derived \(A_{\infty}\)-multiplication \(\mathcal{O}\), we have that \(m\star m=0\) when we view \(m\) as an element of \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\). By Proposition 5.3, this defines an \(A_{\infty}\)-algebra structure on \(S\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) given by maps \[M_{j}:(S\operatorname{Tot}(\mathfrak{s}\mathcal{O}))^{\otimes j}\to S \operatorname{Tot}(\mathfrak{s}\mathcal{O})\] induced by shifting brace maps \[b_{j}^{\star}(m;-):(\operatorname{Tot}(\mathfrak{s}\mathcal{O}))^{\otimes j} \to\operatorname{Tot}(\mathfrak{s}\mathcal{O}).\] The graded module \(S\operatorname{Tot}(\mathfrak{s}\mathcal{O})\) is endowed with the structure of a filtered complex with differential \(M_{1}\) and filtration induced by the column filtration on \(\operatorname{Tot}(\mathfrak{s}\mathcal{O})\). Note that \(b_{j}^{\star}(m;-)\) preserves the column filtration since every component \(b_{j}^{\star}(m_{ij};-)\) has positive horizontal degree. Since \(S\operatorname{Tot}(\mathfrak{s}\mathcal{O})\cong\operatorname{Tot}(S \mathfrak{s}\mathcal{O})\), we can view \(M_{j}\) as the image of a morphism of operads of filtered complexes \(f:\mathcal{A}_{\infty}\to\operatorname{End}_{\operatorname{Tot}(S\mathfrak{s} \mathcal{O})}\) in such a way that \(M_{j}=f(\mu_{j})\) for \(\mu_{j}\in\mathcal{A}_{\infty}(j)\). We now work our way backwards using the strategy also employed by the proof of Theorem 8.1. The isomorphism \[\operatorname{Hom}_{\operatorname{vbOp}}(\mathcal{A}_{\infty},\underline{ \mathcal{End}}_{\operatorname{Tot}(A)})\cong\operatorname{Hom}_{\operatorname{ COOp}}(\mathcal{A}_{\infty},\operatorname{End}_{\operatorname{Tot}(A)})\] does not modify the map \(M_{j}\) at all but allows us to see it as a element of \(\underline{\mathcal{F}_{\underline{n}\underline{d}}}_{\mathrm{Tot}(A)}\) of bidegree \((0,2-j)\). The isomorphism \[\mathrm{Hom}_{\mathrm{vbOp}}(\mathcal{A}_{\infty},\underline{\mathcal{F}_{ \underline{n}\underline{d}}}_{A})\cong\mathrm{Hom}_{\mathrm{vbOp}}(\mathcal{A }_{\infty},\underline{\mathcal{F}_{\underline{n}\underline{d}}}_{\mathrm{Tot} (A)})\] in the direction we are following is the result of applying \(\mathrm{Hom}_{\mathrm{vbOp}}(\mathcal{A}_{\infty},-)\) to the map described in Lemma 2.47. Under this isomorphism, \(f\) is sent to the map \[\mu_{j}\mapsto\mathfrak{Grad}^{-1}\circ c(M_{j},\mu^{-1})=\mathfrak{Grad}^{-1} \circ M_{j}\circ\mu^{-1},\] where \(c\) is the composition in \(\underline{fC}_{\underline{R}}\). The functor \(\mathfrak{Grad}^{-1}\) decomposes \(M_{j}\) into a sum of maps \(M_{j}=\sum_{i}\widetilde{M}_{ij}\), where each \(\widetilde{M}_{ij}\) is of bidegree \((i,2-j-i)\). More explicitly, let \(A=S\mathfrak{sO}\) and let \(x_{1}\otimes\cdots\otimes x_{j}\in(A^{\otimes j})_{k}^{d-k}\). We abuse of notation and identify \(x_{1}\otimes\cdots\otimes x_{j}\) with an element of \(\mathrm{Tot}(A^{\otimes j})\) with only one non-zero component. For a general element, extend linearly. Then we have \[\mathfrak{Grad}^{-1}(M_{j}(\mu^{-1}(x_{1}\otimes\cdots\otimes x_ {j}))) =\] \[\mathfrak{Grad}^{-1}(Sb_{j}^{\star}(m;(S^{-1})^{\otimes j}(\mu^{- 1}(x_{1}\otimes\cdots\otimes x_{j})))) =\] \[\sum_{i}(-1)^{id}\sum_{l}Sb_{j}^{\star}(m_{il};(S^{-1})^{\otimes j} (\mu^{-1}(x_{1}\otimes\cdots\otimes x_{j}))) =\] \[\sum_{i}(-1)^{id}\sum_{l}(-1)^{\varepsilon}Sb_{j}(m_{il};(S^{-1}) ^{\otimes j}(\mu^{-1}(x_{1}\otimes\cdots\otimes x_{j}))) = \tag{31}\] \[\sum_{i}\sum_{l}(-1)^{id+\varepsilon}Sb_{j}(m_{il};(S^{-1})^{ \otimes j}(\mu^{-1}(x_{1}\otimes\cdots\otimes x_{j})))\] so that \[\widetilde{M}_{ij}(x_{1},\ldots,x_{j})=\sum_{l}(-1)^{id+\varepsilon}Sb_{j}(m_ {il};(S^{-1})^{\otimes j}(\mu^{-1}(x_{1}\otimes\cdots\otimes x_{j}))),\] where \(b_{j}\) is the brace on \(\mathfrak{sO}\) and \(\varepsilon\) is given in Lemma 7.1. According to the isomorphism \[\mathrm{Hom}_{\mathrm{vbOp},d^{A}}(d\mathcal{A}_{\infty},\mathrm{End}_{A}) \cong\mathrm{Hom}_{\mathrm{vbOp}}(\mathcal{A}_{\infty},\underline{\mathcal{F} _{\underline{n}\underline{d}}}_{A}), \tag{32}\] the maps \(M_{ij}=(-1)^{ij}\widetilde{M}_{ij}\) define an \(A_{\infty}\)-structure on \(S\mathfrak{sO}\). Therefore we now just have to work out the signs. Notice that \(d_{v}\) is the total degree of \(y_{v}\) as an element of \(\mathfrak{sO}\) and recall that \(d\) is the total degree of \(x_{1}\otimes\cdots\otimes x_{j}\in A^{\otimes j}\). Therefore, \(\varepsilon\) can be written as \[\varepsilon=i(d-j)+\sum_{1\leq v<w\leq j}k_{v}d_{w}.\] The sign produced by \(\mu^{-1}\), as we saw in Lemma 2.23, is precisely determined by the exponent \[\sum_{w=2}^{j}d_{w}\sum_{v=1}^{w-1}k_{v}=\sum_{1\leq v<w\leq j}k_{v}d_{w},\] so this sign cancels the right hand summand of \(\varepsilon\). This cancellation was expected since this sign comes from \(\mu^{-1}\), and operadic composition is defined using \(\mu\), see Equation (24). Finally, the sign \((-1)^{i(d-j)}\) left from \(\varepsilon\) cancels with \((-1)^{id}\) in Equation (31) and \((-1)^{ij}\) from the isomorphism (32). This means that we only need to consider signs produced by vertical shifts. This calculation has been done previously in Lemma 5.8 and as we claimed the result is \[M_{ij}(x_{1},\dots,x_{j})=(-1)^{\sum_{v=1}^{j}(j-v)(d_{v}-k_{v})}\sum_{l}Sb_{j} (m_{il};y_{1},\dots,y_{j}).\] _Remark 8.4_.: Note that as in the case of \(A_{\infty}\)-algebras in \(\mathrm{C}_{R}\) we have two equivalent descriptions of \(A_{\infty}\)-algebras in \(\mathrm{tC}_{R}\). 1. A twisted complex \((A,d^{A})\) together with a morphism \(\mathcal{A}_{\infty}\to\underline{\mathpzc{T}\mathpzc{nd}}_{A}\) of operads in \(\mathrm{vbC}_{R}\), which is determined by a family of elements \(M_{i}\in\underline{t\mathcal{C}_{R}}(A^{\otimes i},A)_{0}^{2-i}\) for \(i\geq 2\) for which the \(A_{\infty}\)-relations holds for \(i\geq 2\), Equation (1). The composition is the one prescribed by the composition morphisms of \(\underline{t\mathcal{C}_{R}}\). 2. A bigraded module \(A\) together with a family of elements \(M_{i}\in\underline{\mathpzc{g}\mathpzc{M}\mathpzc{od}}_{R}(A^{\otimes i},A)_{ 0}^{2-i}\) for \(i\geq 1\) for which all the \(A_{\infty}\)-relations hold, see Equation (1). The composition is prescribed by the composition morphisms of \(\underline{\mathpzc{g}\mathpzc{M}\mathpzc{od}}_{R}\). Since the composition morphism in \(\underline{\mathpzc{g}\mathpzc{M}\mathpzc{od}}_{R}\) is induced from the one in \(\underline{t\mathcal{C}_{R}}\) by forgetting the differential, these two presentations are equivalent, see [10]. This equivalence allows us to formulate the following alternative version of Theorem 8.1. **Corollary 8.5**.: _Given a bigraded module \(A\) horizontally bounded on the right we have isomorphisms_ \[\mathrm{Hom}_{\mathrm{bgOp}}(d\mathcal{A}_{\infty},\mathrm{End} _{A}) \cong\mathrm{Hom}_{\mathrm{bgOp}}(\mathcal{A}_{\infty},\underline{ \mathpzc{T}\mathpzc{nd}}_{A})\] \[\cong\mathrm{Hom}_{\mathrm{bgOp}}(\mathcal{A}_{\infty},\underline{ \mathpzc{T}\mathpzc{nd}}_{\mathrm{Tot}(A)})\] \[\cong\mathrm{Hom}_{\mathrm{fOp}}(\mathcal{A}_{\infty},\underline {\mathrm{End}}_{\mathrm{Tot}(A)}),\] _where \(\mathrm{bgOp}\) is the category of operads of bigraded modules and \(\mathrm{fOp}\) is the category of operads of filtered modules._ Proof.: Let us look at the first isomorphism \[\operatorname{Hom}_{\operatorname{bgOp}}(\mathcal{A}_{\infty},\underline{\underline{ \mathcal{F}_{\underline{n}\underline{d}}}}_{A})\cong\operatorname{Hom}_{ \operatorname{bgOp}}(d\mathcal{A}_{\infty},\operatorname{End}_{A}).\] Let \(f:\mathcal{A}_{\infty}\to\underline{\underline{\mathcal{F}_{\underline{n} \underline{d}}}}_{A}\) be a map of operads in \(\operatorname{bgOp}\). This is equivalent to maps in \(\operatorname{bgOp}\) \[\mathcal{A}_{\infty}(j)\to\underline{\underline{\mathcal{F}_{\underline{n} \underline{d}}}}_{A}(j)\] for each \(j\geq 1\), which are determined by elements \(M_{j}\coloneqq f(\mu_{j})\in\underline{\underline{\mathcal{F}_{\underline{n} \underline{d}}}}_{A}(j)\) for \(v\geq 1\) of bidegree \((0,2-j)\) satisfying the \(A_{\infty}\)-equation with respect to the composition in \(\underline{\underline{\mathcal{F}_{\underline{n}\underline{d}}}}_{R}\). Moreover, \(M_{j}\coloneqq(\tilde{m}_{0j},\tilde{m}_{1j},\dots)\) where \(\tilde{m}_{ij}\coloneqq(M_{j})_{i}:A^{\otimes n}\to A\) is a map of bidegree \((i,2-i-j)\). Since the composition in \(\underline{\underline{\mathcal{F}_{\underline{n}\underline{d}}}}_{R}\) is the same as in \(\underline{\mathcal{U}_{R}}\), the computation of the \(A_{\infty}\)-equation becomes analogous to the computation done in [2, Prop 4.47], showing that the maps \(m_{ij}=(-1)^{i}\tilde{m}_{ij}\) for \(i\geq 0\) and \(j\geq 0\) define a derived \(A_{\infty}\)-algebra structure on \(A\). The second isomorphism \[\operatorname{Hom}_{\operatorname{bgOp}}(\mathcal{A}_{\infty},\underline{ \underline{\mathcal{F}_{\underline{n}\underline{d}}}}_{A})\cong\operatorname{ Hom}_{\operatorname{bgOp}}(\mathcal{A}_{\infty},\underline{\underline{\mathcal{F}_{ \underline{n}\underline{d}}}}_{\operatorname{Tot}(A)})\] follows from the bigraded module case of Lemma 2.46. Finally, the isomorphism \[\operatorname{Hom}_{\operatorname{bgOp}}(\mathcal{A}_{\infty},\underline{ \underline{\mathcal{F}_{\underline{n}\underline{d}}}}_{\operatorname{Tot}(A) })\cong\operatorname{Hom}_{\operatorname{Hop}}(\mathcal{A}_{\infty},\underline {\underline{\operatorname{F}_{\underline{n}\underline{d}}}}_{\operatorname{Tot }(A)})\] is analogous to the last isomorphism of Theorem 8.1, replacing the quasi-free relation by the full \(A_{\infty}\)-algebra relations. According to Corollary 8.5, if we have an \(A_{\infty}\)-algebra structure on \(A=S\mathfrak{s}\mathcal{O}\), we can consider its arity \(1\) component \(M_{1}\in\underline{\operatorname{End}}_{\operatorname{Tot}(A)}\) and split it into maps \(M_{i1}\in\operatorname{End}_{A}\). Since these maps must satisfy the derived \(A_{\infty}\)-relations, they define a twisted complex structure on \(A\). The next corollary describes the maps \(M_{i1}\) explicitly. **Corollary 8.6**.: _Let \(\mathcal{O}\) be a bigraded operad with a derived \(A_{\infty}\)-multiplication and let \(M_{i1}:S\mathfrak{s}\mathcal{O}\to S\mathfrak{s}\mathcal{O}\) be the arity 1 derived \(A_{\infty}\)-algebra maps induced by Corollary 8.5 from \(M_{1}:\operatorname{Tot}(S\mathfrak{s}\mathcal{O})\to\operatorname{Tot}(S \mathfrak{s}\mathcal{O})\). Then_ \[M_{i1}(x)=\sum_{l}(Sb_{1}(m_{il};S^{-1}x)-(-1)^{\langle x,m_{il}\rangle}Sb_{1}( S^{-1}x;m_{il})),\] _where \(x\in(S\mathfrak{s}\mathcal{O})_{k}^{d-k}\) and \(\langle x,m_{il}\rangle=ik+(1-i)(d-1-k)\)._ Proof.: Notice that the proof of Corollary 8.5 is essentially the same as the proof Theorem 8.1. This means that the proof of this result is an arity \(1\) restriction of the proof of Theorem 8.3. Thus, we apply Equation (31) to the case \(j=1\). Recall that for \(x\in(S\mathfrak{s}\mathcal{O})_{k}^{d-k}\), \[M_{1}(x)=b_{1}^{\star}(m;S^{-1}x)-(-1)^{n-1}b_{1}^{\star}(S^{-1}x;m).\] In this case, there is no \(\mu\) involved. Therefore, introducing the final extra sign \((-1)^{i}\) from the proof of Theorem 8.3, we get from Equation (31) that \[\widetilde{M}_{i1}(x)=(-1)^{i}\sum_{l}((-1)^{id+i(d-1)}Sb_{1}(m_{il};S^{-1}x)-( -1)^{d-1+id+k}Sb_{1}(S^{-1}x;m_{il})),\] where \(b_{1}\) is the brace on \(\mathfrak{s}\mathcal{O}\). Simplifying signs we obtain \[\widetilde{M}_{i1}(x)=\sum_{l}Sb_{1}(m_{il};S^{-1}x)-(-1)^{\langle m_{il},x \rangle}Sb_{1}(m_{il};S^{-1}x))=M_{i1}(x),\] where \(\langle m_{il},x\rangle=ik+(1-i)(d-1-k)\), as claimed. ### Derived Deligne conjecture Note that the maps given by Theorem 8.3 and Corollary 8.6 formally look the same as their single-graded analogues in Lemma 5.8 but with an extra index that is fixed for each \(M_{ij}\). This means that we can follow the same procedure as in Section 5.1 to define higher derived \(A_{\infty}\)-maps on the Hochschild complex of a derived \(A_{\infty}\)-algebra. More precisely, given an operad \(\mathcal{O}\) with a derived multiplication and \(A=S\mathfrak{s}\mathcal{O}\), we will define a derived \(A_{\infty}\)-algebra structure on \(S\mathfrak{s}\operatorname{End}_{A}\). We will then connect the algebraic structure on \(A\) with the structure on \(S\mathfrak{s}\operatorname{End}_{A}\) through braces. This connection will allow us to formulate and show a new version of the Deligne conjecture. Let \(\overline{B}_{j}\) be the bigraded brace map on \(\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) and consider the maps \[\overline{M}^{\prime}_{ij}:(\mathfrak{s}\operatorname{End}_{S\mathfrak{s} \mathcal{O}})^{\otimes j}\to\mathfrak{s}\operatorname{End}_{S\mathfrak{s} \mathcal{O}} \tag{33}\] defined as \[\overline{M}^{\prime}_{ij}(f_{1},\ldots,f_{j})=\overline{B}_{j}(M_ {i\bullet};f_{1},\ldots,f_{j}) j>1,\] \[\overline{M}^{\prime}_{i1}(f)=\overline{B}_{1}(M_{i\bullet};f)-( -1)^{ip+(1-i)q}\overline{B}_{1}(f;M_{i\bullet}),\] for \(f\) of natural bidegree \((p,q)\), where \(M_{i\bullet}=\sum_{j}M_{ij}\). We define \[\overline{M}_{ij}:(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}} )^{\otimes j}\to S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}},\ \overline{M}_{ij}=\overline{\sigma}(M^{\prime}_{ij})=S\circ M^{\prime}_{ij} \circ(S^{\otimes n})^{-1}.\] As in the single-graded case we can define a map \(\Phi:S\mathfrak{s}\mathcal{O}\to S\mathfrak{s}\operatorname{End}_{S \mathfrak{s}\mathcal{O}}\) as the map making the following diagram commute (34) where \[\Phi^{\prime}\colon\mathfrak{s}\mathcal{O}\to\operatorname{End}_{s\mathcal{O}},\,x\mapsto\sum_{n\geq 0}b_{n}(x;-).\] In this setting we have the bigraded version of Theorem 5.7. But before stating the theorem, for the sake of completeness let us state the definition of the Hochschild complex of a bigraded module. **Definition 8.7**.: _We define the Hochschild cochain complex of a bigraded module \(A\) to be the bigraded module \(S\mathfrak{s}\operatorname{End}_{A}\). If \((A,d)\) is a vertical bicomplex, then the Hochschild complex has a vertical differential given by \(\partial(f)=[d,f]=d\circ f-(-1)^{q}f\circ d\), where \(f\) has natural bidigree \((p,q)\) and \(\circ\) is the plethysm corresponding to operadic insertions._ In particular, \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) is the Hochschild cochain complex of \(S\mathfrak{s}\mathcal{O}\). If \(\mathcal{O}\) has a derived \(A_{\infty}\)-multiplication, then the differential of the Hochschild complex \(S\mathfrak{s}\operatorname{End}_{S\mathfrak{s}\mathcal{O}}\) is given by \(\overline{M}_{01}\) from Equation (33). The following is the same as Theorem 5.7 but carrying the extra index \(i\) and using the bigraded sign conventions. **Theorem 8.8**.: _The map \(\Phi\) defined in diagram (34) above is a morphism of derived \(A_{\infty}\)-algebras, i.e. for all \(i\geq 0\) and \(j\geq 1\) the equation_ \[\Phi(M_{ij})=\overline{M}_{ij}(\Phi^{\otimes j})\] _holds. _ Now that we have Theorem 8.8 and the explicit formulas for the derived \(A_{\infty}\)-structure on \(S\mathfrak{s}\mathcal{O}\), we can deduce the derived version of the Deligne conjecture in an analogous way to how we obtained the \(A_{\infty}\)-version in Corollary 5.12. In order to do that, we need to first introduce the derived \(A_{\infty}\)-version of homotopy \(G\)-algebras. To have a more succinct formulation we use the notation \(\operatorname{vdeg}(x)\) for the vertical degree of \(x\). **Definition 8.9**.: _A derived \(J\)-algebra\(V\) is a derived \(A_{\infty}\)-algebra with structure maps \(\{M_{ij}\}_{i\geq 0,j\geq 1}\) such that the shift is \(S^{-1}V\) a brace algebra. Furthermore, the braces and the derived \(A_{\infty}\)-structure satisfy the following compatibility relations. Let \(x,x_{1},\dots,x_{j},y_{1},\dots,y_{n}\in S^{-1}V\). For all \(n,i\geq 0\) we demand_ \[(-1)^{\sum_{i=1}^{n}(n-v)\operatorname{vdeg}(y_{v})}Sb_{n}(S^{-1}M _{i1}(Sx);y_{1},\dots,y_{n})=\] \[\sum_{\begin{subarray}{c}l+k-1=n\\ 1\leq i_{1}\leq n-k+1\end{subarray}}\hskip-14.226378pt(-1)^{\sigma}M_{il}(Sy_{1 },\dots,Sb_{k}(x;y_{i_{1}},\dots),\dots,Sy_{n})\] \[-(-1)^{\langle x,M_{il}\rangle}\hskip-14.226378pt\sum_{ \begin{subarray}{c}l+k-1=n\\ 1\leq i_{1}\leq n-k+1\end{subarray}}\hskip-14.226378pt(-1)^{\eta}Sb_{k}(x;y_{1},\dots,S^{-1}M_{il}(Sy_{i_{1}},\dots,),\dots,y_{n})\] _where_ \[\varepsilon=\sum_{v=1}^{i_{1}-1}\langle Sy_{v},S^{1-k}x\rangle+\sum_{v=1}^{k} \operatorname{vdeg}(y_{i_{1}+v-1})(k-v)+(l-i_{1})\operatorname{vdeg}(x).\] _and_ \[\eta =\sum_{v=1}^{i_{1}-1}(k-v)\mathrm{vdeg}(y_{v})+l\sum_{v=1}^{i_{1}-1} \mathrm{vdeg}(y_{v})\] \[\quad+\sum_{v=i_{1}}^{i_{1}+l-1}(k-i_{1})\mathrm{vdeg}(y_{v})+\sum_ {v=i_{1}}^{n-l}(k-v)\mathrm{vdeg}(y_{v+l})\] _For \(j>1\) we demand_ \[(-1)^{\sum_{i=1}^{n}(n-v)\mathrm{vdeg}(y_{v})}Sb_{n}(S^{-1}M_{ij} (Sx_{1},\dots,Sx_{j});y_{1},\dots,y_{n})=\] \[\sum(-1)^{\varepsilon}M_{il}(Sy_{1},\dots,Sb_{k_{1}}(x_{1};y_{i_{1 }},\dots),\dots,Sb_{k_{j}}(x_{j};y_{i_{j}},\dots),\dots,Sy_{n}).\] _The unindexed sum runs over all possible choices of non-negative integers that satisfy \(l+k_{1}+\dots+k_{j}-j=n\) and over all possible ordering preserving insertions. The right hand side sign is given by_ \[\varepsilon =\sum_{\begin{subarray}{c}1\leq t\leq j\\ 1\leq v\leq k_{t}\end{subarray}}\mathrm{vdeg}(y_{i_{t}+v-1})(k_{v}-v)+\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \!\! known examples of derived \(A_{\infty}\)-algebras. These examples usually come as minimal models of dgas. So a first question that arises is the following. **Question 1**.: _Are there any conditions on a dga that guarantee that its minimal model is horizontally bounded on the right?_ An answer to this question would give us a better understanding of how general our results are. In fact, it is open whether a derived \(A_{\infty}\)-structure can be obtained for a more general operad. Even though we needed to use some monoidality results that require boundedness, the explicit maps that we obtain in Theorem 8.3 can be defined for any operad with a derived \(A_{\infty}\)-multiplication. A first idea would be attempting a direct computation to see if they satisfy the derived \(A_{\infty}\)-equation, see Equation (22). Of course, we would like to use a more conceptual approach. So more generally the question would be the following. **Question 2**.: _Can we define a derived \(A_{\infty}\)-structure on any operad with a derived \(A_{\infty}\)-multiplication?_ ### Hochschild Cohomology The classical Deligne conjecture states that the Hoschschild complex of an associative algebra has a structure of homotopy \(G\)-algebra [10]. This has implications on the Hochschild cohomology of the associative algebra. Namely, the homotopy \(G\)-algebra structure on the Hoschschild complex induces a Gerstenhaber algebra structure on cohomology. We would like to extend this result to derived \(A_{\infty}\)-algebras. Let us review the structure on the Hochschild complex of an associative operad in order to understand the question that we will be asking about the derived \(A_{\infty}\)-case. Let \(\mathcal{O}\) be an operad with an associative multiplication \(m\), i.e. an \(A_{\infty}\)-multiplication \(m\) such that \(m=m_{2}\), see Definition 3.5. In this case, as a consequence of Proposition 5.3 or by [10, Proposition 2], we have a dg-algebra structure on \(S\mathfrak{s}\mathcal{O}\) given by the differential \[d(Sx)=Sb_{1}(m;x)-(-1)^{|x|}Sb_{1}(x;m)\] and the multiplication \[m(Sx,Sy)=Sb_{2}(m;x,y).\] In particular, if \(\mathcal{O}=\operatorname{End}_{A}\) is the endomorphism operad of an associative algebra \(A\), these maps provide a dg-algebra structure on the Hochschild complex of \(A\). But this is not all the structure that we get. Since any operad is a brace algebra, we have an interaction between the dg-algebra and the brace structure. More precisely, \(\mathcal{O}\) has a structure of _homotopy \(G\)-algebra_, see Definition 2 and Theorem 3 of [10]. Given the algebraic structure described above on the Hochschild complex of an associative algebra, we can then take cohomology with respect to \(d\) to compute the Hochschild cohomology of \(A\), denoted by \(HH^{*}(A)\). It is known that \(m\) and the bracket \[[x,y]=Sb_{1}(x;y)-(-1)^{|x||y|}Sb_{1}(y;x)\] induce a structure of a Gerstenhaber algebra on \(HH^{*}(A)\)[13, Corollary 5]. The proof relies on some identities that can be deduced from the definition of homotopy \(G\)-algebra, such as graded homotopy commutativity. If we try to replicate this argument for \(A_{\infty}\)-algebras, the structure we get on the Hochschild complex is that of a \(J\)-algebra, see Definition 5.11. In this case, we have to compute cohomology with respect to \(M_{1}\), see Lemma 5.8. In the definition of \(J\)-algebras, we encounter an explosion in the number and complexity of relations and maps involved with respect to homotopy \(G\)-algebras. Therefore, the resulting structure has not been feasible to manipulate and it is not very clear what kind of algebraic structure is induced on cohomology. The derived case is of course even more difficult to handle as we would need to work with the even more complex derived \(J\)-algebras, see Definition 8.9. In addition, it is possible to consider vertical and horizontal cohomologies [16, SS1.2]. These should be taken with respect to \(M_{01}\) and \(M_{11}\) respectively, see Corollary 8.6. So the natural question to ask is the following. **Question 3**.: _What algebraic structure do derived \(J\)-algebras induce on the vertical and horizontal cohomologies of a derived \(A_{\infty}\)-algebra?_ ## Appendix A Combinatorics **Lemma A.1**.: _For any integers \(n\) and \(m\), the following equality holds mod 2._ \[\binom{n+m-1}{2}+\binom{n}{2}+\binom{m}{2}=(n-1)(m-1).\] Proof.: Let us compute \[\binom{n+m-1}{2}+\binom{n}{2}+\binom{m}{2}+(n-1)(m-1)\mod 2.\] By definition, this equals \[\frac{(n+m-1)(n+m-2)}{2}+\frac{n(n-1)}{2}+\frac{m(m-1)}{2}+(n-1)(m -1)\] \[=\frac{(n^{2}+2nm-2n+m^{2}-2m-n-m+2)+(n^{2}-n)+(m^{2}-m)+2(nm-n-m +1)}{2}\] \[=n^{2}+2nm-3n+m^{2}-3m+2=n^{2}+m+m^{2}+m=0\mod 2\] as desired, because \(n^{2}=n\mod 2\) ## Appendix B Koszul sign on operadic suspension The purpose of this section is to clear up the procedure to apply the Koszul sign rule in situations in which operadic suspension is involved. Let \(\operatorname{End}_{A}\) be the endomorphism operad of some \(R\)-module \(A\) and consider the operadic suspension \(\mathfrak{s}\operatorname{End}_{A}\). Let \(f\otimes e^{n}\in\mathfrak{s}\operatorname{End}_{A}(n)\) be of degree \(\deg(f)+n-1\). For \(a\in A^{\otimes n}\) we have \[(f\otimes e^{n})(a)=(-1)^{\deg(a)(n-1)}f(a)\otimes e^{n}\] because \(\deg(e^{n})=n-1\). Note that \(f\otimes e^{n}=g\otimes e^{n}\) if and only if \(f=g\). In addition, it is not possible that \(f\otimes e^{n}=g\otimes e^{m}\) for \(n\neq m\). If we take the tensor product of \(f\otimes e^{n}\in\mathfrak{s}\operatorname{End}_{A}(n)\) and \(g\otimes e^{m}\in\mathfrak{s}\operatorname{End}_{A}(m)\) and apply it to \(a\otimes b\in A^{\otimes n}\otimes A^{\otimes m}\), we have \[((f\otimes e^{n})\otimes(g\otimes e^{m}))(a\otimes b)= (-1)^{\deg(a)(\deg(g)+m-1)}(f\otimes e^{n})(a)\otimes(g\otimes e^{ m})(b)\] \[= (-1)^{\varepsilon}(f(a)\otimes e^{n})\otimes(f(b)\otimes e^{m}),\] where \(\varepsilon=\deg(a)(\deg(g)+m-1)+\deg(a)(n-1)+\deg(b)(m-1)\). The last remark that we want to make is the case of a map of the form \[f(1^{\otimes k-1}\otimes g\otimes 1^{\otimes n-k})\otimes e^{m+n-1}\in \mathfrak{s}\operatorname{End}_{A}(n+m-1),\] such as those produced by the operadic insertion \(\mathfrak{s}f\tilde{\circ}_{k}\mathfrak{s}g\). In this case, this map applied to \(a_{k-1}\otimes b\otimes a_{n-k}\in A^{\otimes k-1}\otimes A^{\otimes m} \otimes A^{\otimes n-k}\) results in \[(f(1^{\otimes k-1}\otimes g\otimes 1^{\otimes n-k})\otimes e^{m+n-1})( a_{k-1}\otimes b\otimes a_{n-k})=\] \[(-1)^{(m+n)(\deg(a_{k-1})+\deg(b)+\deg(a_{n-k}))}(f(1^{\otimes k- 1}\otimes g\otimes 1^{\otimes n-k}(a_{k-1}\otimes b\otimes a_{n-k}))\otimes e^{m+n-1}\] To go from the first line to the second, we switch \(e^{m+n-1}\) of degree \(m+n-2\) with \(a_{k-1}\otimes b\otimes a_{n-k}\). If we apply the usual sign rule for graded maps we obtain \[(-1)^{(m+n)(\deg(a_{k-1})+\deg(b)+\deg(a_{n-k}))+\deg(a_{k-1})\deg(g)}f(a_{k-1 }\otimes g(b)\otimes a_{n-k})\otimes e^{m+n-1}.\] The purpose of this last remark is not only review the Koszul sign rule but also remind that the insertion \(\mathfrak{s}f\tilde{\circ}_{k}\mathfrak{s}g\) is of the above form, so that the \(e^{m+n-1}\) component is always at the end and does not play a role in the application of the sign rule with the composed maps. In other words, it does not affect their individual degrees, just the degree of the overall composition. ## Appendix C Sign of the braces In order to find the sign of the braces on \(\mathfrak{s}\operatorname{End}_{A}\), let us use an analogous strategy to the one used in [11, Appendix] to find the signs of the Lie bracket \([f,g]\) on \(\operatorname{End}_{A}\). Let \(A\) be a graded module. Let \(SA\) be the graded module with \(SA^{v}=A^{v+1}\), and so the _suspension_ or _shift_ map \(S:A\to SA\) given by the identity map has degree \(-1\). Let \(f\in\operatorname{End}_{A}(N)^{i}=\operatorname{Hom}_{R}(A^{\otimes N},A)^{i}\). Recall that \(\sigma\) is the inverse of the map from Theorem 3.9, so that \(\sigma(f)=S\circ f\circ(S^{-1})^{\otimes N}\in\operatorname{End}_{A}(N)^{i+N-1}\). _Remark C.1_.: In [10] there is a sign \((-1)^{N+i-1}\) in front of \(f\), but it seems to be irrelevant for our purposes. Another fact to remark on is that the suspension of graded modules used here and in [10] is the opposite that we have used to define the operadic suspension in the sense that in Section 3 we used \(SA^{v}=A^{v-1}\). This does not change the signs or the procedure, but in the statement of Theorem 3.9, operadic desuspension should be changed to operadic suspension. Notice that by the Koszul sign rule \[(S^{-1})^{\otimes N}\circ S^{\otimes N}=(-1)^{\sum_{j=1}^{N-1}j}1=(-1)^{\frac{ N(N-1)}{2}}1=(-1)^{\binom{N}{2}}1,\] so \((S^{-1})^{\otimes N}=(-1)^{\binom{N}{2}}(S^{\otimes N})^{-1}\). For this reason, given \(F\in\operatorname{End}_{S(A)}(m)^{j}\), we have \[\sigma^{-1}(F)=(-1)^{\binom{m}{2}}S^{-1}\circ F\circ S^{\otimes m}\in \operatorname{End}_{A}(m)^{j-m+1}.\] For \(g_{j}\in\operatorname{End}_{A}(a_{j})^{q_{j}}\), let us write \(f[g_{1},\dots,g_{n}]\) for the map \[\sum_{k_{0}+\dots+k_{n}=N-n}f(1^{\otimes k_{0}}\otimes g_{1}\otimes 1^{ \otimes k_{1}}\otimes\dots\otimes g_{n}\otimes 1^{\otimes k_{n}})\in \operatorname{End}_{A}(N-n+\sum a_{j})^{i+\sum q_{j}}.\] We define \[b_{n}(f;g_{1},\dots,g_{n})=\sigma^{-1}(\sigma(f)[\sigma(g_{1}),\dots,\sigma(g_ {n})])\in\operatorname{End}_{A}(N-n+\sum a_{j})^{i+\sum q_{j}}.\] With this the definition we can prove the following. **Lemma C.2**.: _We have_ \[b_{n}(f;g_{1},\dots,g_{n})=\sum_{N-n=k_{0}+\dots+k_{n}}(-1)^{\eta}f(1^{\otimes k _{0}}\otimes g_{1}\otimes\dots\otimes g_{n}\otimes 1^{\otimes k_{n}}),\] _where_ \[\eta=\sum_{0\leq j<l\leq n}k_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}+\sum_{j =1}^{n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)k_{l}.\] Proof.: Let us compute \(\eta\) using the definition of \(b_{n}\). \[\sigma^{-1}(\sigma(f)[\sigma(g_{1}),\ldots,\sigma(g_{n})])\] \[=(-1)^{\binom{N-n+\sum a_{j}}{2}}S^{-1}\circ(\sigma(f)(1^{\otimes k_ {0}}\otimes\sigma(g_{1})\otimes 1^{\otimes k_{1}}\otimes\cdots\otimes\sigma(g_{n}) \otimes 1^{\otimes k_{n}}))\circ S^{\otimes N-n+\sum a_{j}}\] \[=(-1)^{\binom{N-n+\sum a_{j}}{2}}S^{-1}\circ S\circ f\circ(S^{-1 })^{\otimes N}\circ\] \[(1^{\otimes k_{0}}\otimes(S\circ g_{1}\circ(S^{-1})^{\otimes a_{ 1}})\otimes 1^{\otimes k_{1}}\otimes\cdots\otimes(S\circ g_{n}\circ(S^{-1})^{ \otimes a_{n}})\otimes 1^{\otimes k_{n}}))\circ S^{\otimes N-n+\sum a_{j}}\] \[=(-1)^{\binom{N-n+\sum a_{j}}{2}}f\circ((S^{-1})^{k_{0}}\otimes S ^{-1}\otimes\cdots\otimes S^{-1}\otimes(S^{-1})^{k_{n}})\] \[\circ(1^{\otimes k_{0}}\otimes(S\circ g_{1}\circ(S^{-1})^{ \otimes a_{1}})\otimes\cdots\otimes(S\circ g_{n}\circ(S^{-1})^{\otimes a_{n}} )\otimes 1^{\otimes k_{n}}))\circ S^{\otimes N-n+\sum a_{j}}.\] Now we move each \(1^{\otimes k_{j-1}}\otimes S\circ g_{j}\circ(S^{-1})^{a_{j}}\) to apply \((S^{-1})^{k_{j-1}}\otimes S^{-1}\) to it. Doing this for all \(j=1,\ldots,n\) produces a sign \[(-1)^{(a_{1}+q_{1}-1)(n-1+\sum k_{l})+(a_{2}+q_{2}-1)(n-2+\sum_{ 2}^{n}k_{l})+\cdots+(a_{n}+q_{n}-1)k_{n}}\] \[=(-1)^{\sum_{j=1}^{n}(a_{j}+q_{j}-1)(n-j+\sum_{j}^{n}k_{l})},\] and we denote the exponent by \[\varepsilon=\sum_{j=1}^{n}(a_{j}+q_{j}-1)(n-j+\sum_{j}^{n}k_{l}).\] So now we have that, decomposing \(S^{\otimes N-n+\sum a_{j}}\), the last map up to multiplication by \((-1)^{\binom{N-n+\sum a_{j}}{2}+\varepsilon}\) is \[f\circ((S^{-1})^{k_{0}}\otimes g_{1}\circ(S^{-1})^{\otimes a_{1}}\otimes \cdots\otimes g_{n}\circ(S^{-1})^{\otimes a_{n}}\otimes(S^{-1})^{k_{n}}) \circ(S^{\otimes k_{0}}\otimes S^{\otimes a_{1}}\otimes\cdots\otimes S^{ \otimes a_{n}}\otimes S^{\otimes k_{n}}).\] Now we turn the tensor of inverses into inverses of tensors by introducing the appropriate signs. More precisely, we introduce the sign \[(-1)^{\delta}=(-1)^{\binom{k_{0}}{2}+\sum\left(\binom{a_{j}}{2}+\binom{k_{j}}{ 2}\right)}. \tag{35}\] Therefore we have up to multiplication by \((-1)^{\binom{N-n+\sum a_{j}}{2}+\varepsilon+\delta}\) the map \[f\circ((S^{k_{0}})^{-1}\otimes g_{1}\circ(S^{\otimes a_{1}})^{-1}\otimes \cdots\otimes g_{n}\circ(S^{\otimes a_{n}})^{-1}\otimes(S^{k_{n}})^{-1}) \circ(S^{\otimes k_{0}}\otimes S^{\otimes a_{1}}\otimes\cdots\otimes S^{ \otimes a_{n}}\otimes S^{\otimes k_{n}}).\] The next step is moving each component of the last tensor product in front of its inverse. This will produce the sign \((-1)^{\gamma}\), where \[\gamma =-k_{0}\sum_{1}^{n}(k_{j}+a_{j}+q_{j})-a_{1}\left(\sum_{1}^{n}k_ {j}+\sum_{2}^{n}(a_{j}+q_{j})\right)-\cdots-a_{n}k_{n}\] \[=\sum_{j=0}^{n}k_{j}\sum_{l=j+1}^{n}(k_{l}+a_{l}+q_{l})+\sum_{j=1} ^{n}a_{j}\left(\sum_{l=j}^{n}k_{l}+\sum_{l=j+1}^{n}(a_{l}+q_{l})\right)\mod 2. \tag{36}\] So in the end we have \[b_{n}(f;g_{1},\dots,g_{n})=\sum_{k_{0}+\dots+k_{n}=N-n}(-1)^{\binom{N-n+\sum a_{j} }{2}+\varepsilon+\delta+\gamma}f(1^{\otimes k_{0}}\otimes g_{1}\otimes 1^{ \otimes k_{1}}\otimes\dots\otimes g_{n}\otimes 1^{\otimes k_{n}}).\] This means that \[\eta=\binom{N-n+\sum a_{j}}{2}+\varepsilon+\delta+\gamma.\] Next, we are going to simplify this sign to get rid of the binomial coefficients. _Remark C.3_.: If the top number of a binomial coefficient is less than \(2\), then the coefficient is \(0\). In the case of arities or \(k_{j}\) this is because \((S^{-1})^{\otimes 1}=(S^{\otimes 1})^{-1}\), and if the tensor is taken \(0\) times then it is the identity and the equality also holds, so there are no signs. We are now going to simplify the sign to obtain the desired result. Notice that \[N-n+\sum_{j}a_{j}=\sum_{i}k_{i}+\sum_{j}a_{j}.\] In general, consider a finite sum \(\sum_{i}b_{i}\). We can simplify the binomial coefficients mod \(2\) in the following way. \[\binom{\sum_{i}b_{i}}{2}+\sum_{i}\binom{b_{i}}{2}=\sum_{i<j}b_{i}b_{j}\mod 2.\] The result of applying this to \(\binom{N-n+\sum a_{j}}{2}\) and adding \(\delta\) from eq. (35) in our sign \(\eta\) is \[\sum_{0\leq i<l\leq n}k_{i}k_{l}+\sum_{1\leq j<l\leq n}a_{j}a_{l}+\sum_{i,j}k_ {i}a_{j}. \tag{37}\] Recall \(\gamma\) from Equation (36). \[\gamma=\sum_{j=0}^{n}k_{j}\sum_{l=j+1}^{n}(k_{l}+a_{l}+q_{l})+\sum_{j=1}^{n}a_ {j}\left(\sum_{l=j}^{n}k_{l}+\sum_{l=j+1}^{n}(a_{l}+q_{l})\right).\] As we see, all the sums in the previous simplification appear in \(\gamma\) so we can cancel them. Let us rewrite \(\gamma\) in a way that this becomes more clear: \[\sum_{0\leq j<l\leq n}k_{j}k_{l}+\sum_{0\leq j<l\leq n}k_{j}a_{l}+\sum_{0\leq j <l\leq n}k_{j}q_{l}+\sum_{1\leq j\leq l\leq n}a_{j}k_{l}+\sum_{1\leq j<l\leq n }a_{j}a_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}.\] So after adding the expression (37) modulo \(2\) we have only the terms that include the internal degrees, i.e. \[\sum_{0\leq j<l\leq n}k_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}. \tag{38}\] Let us move now to the \(\varepsilon\) term in the sign to rewrite it. \[\varepsilon=\sum_{j=1}^{n}(a_{j}+q_{j}-1)(n-j+\sum_{j}^{n}k_{l})=\sum_{j=1}^ {n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)k_{l}\] We may add this to Equation (38) in such a way that the brace sign becomes \[\eta=\sum_{0\leq j<l\leq n}k_{j}q_{l}+\sum_{1\leq j<l\leq n}a_{j}q_{l}+\sum_{j=1}^ {n}(a_{j}+q_{j}-1)(n-j)+\sum_{1\leq j\leq l\leq n}(a_{j}+q_{j}-1)k_{l}. \tag{39}\] as announced at the end of Section 4.

2306.06241

Almost paratopological groups

A class of almost paratopological groups is introduced, which (1) contains paratopological groups and Hausdorff quasitopological groups; (2) is closed under products; (3) subgroups. Almost paratopological $T_1$ groups $G$ are characterized by the fact that $\{(x,y)\in G^2: xy=e\}$ is closed in $G^2$. A compact almost paratopological group is topological. A regular $\Sigma$-space with a countable extend and a separately continuous Mal'tsev operation is $\omega$-cellular (and ccc). A $\sigma$-compact regular almost paratopological group is ccc. In particular, a $\sigma$-compact regular quasitopological group is ccc.

Evgenii Reznichenko

2023-06-09T20:27:33Z

http://arxiv.org/abs/2306.06241v2

# Almost paratopological groups ###### Abstract A class of almost paratopological groups is introduced, which (1) contains paratopological groups and Hausdorff quasitopological groups; (2) is closed under products; (3) subgroups. Almost paratopological \(T_{1}\) groups \(G\) are characterized by the fact that \(\{(x,y)\in G^{2}\,:\,xy=e\}\) is closed in \(G^{2}\). A compact almost paratopological group is topological. A regular \(\Sigma\)-space with countable extend and a separately continuous Mal'tsev operation is \(\omega\)-cellular (and ccc). A \(\sigma\)-compact regular almost paratopological group is ccc. In particular, a \(\sigma\)-compact regular quasitopological group is ccc. keywords: almost paratopological groups, topological groups, paratopological groups, semitopological groups, quasitopological groups, Mal'tsev spaces, ccc spaces, \(\omega\)-cellular spaces + Footnote †: journal: ## 1 Introduction One of the most important concepts in mathematics is the concept of a topological group. A topological group is a group with a topology with respect to which the operations of product and inverse are continuous. Other groups with topology are also important. Groups \(G\) with topology is called * _paratopological_ if multiplication in \(G\) is continuous; * _semitopological_ if multiplication in \(G\) is separately continuous; * _quasitopological_ if \(G\) is a semitopological group and the operation taking the inverse element \(x\mapsto x^{-1}\) is continuous. In the book [1] and the survey [9] one can find further information about the enumerated classes of groups. This article introduces the concept of almost paratopological group, a class of semitopological groups that contains paratopological groups and Hausdorff quasitopological groups. This class of semitopological groups is closed under the product and subgroups (Theorem 5). For them, assertions are proved that were previously known for paratopological groups. A compact almost paratopological group is topological (Theorem 6). In particular, a compact paratopological group is topological (5, Lemma 6). A mapping \(M:X^{3}\to X\) is called a _Mal'tsev operation_ if \[M(x,y,y)=M(y,y,x)=x\] for all \(x,y\in X\). There is a natural Mal'tsev operation on groups: \[M_{G}(x,y,z)=xy^{-1}z\] for \(x,y,z\in G\). On retracts of sets with the Mal'tsev operation, there is a Mal'tsev operation: if \(r:X\to X\) is a retraction, then we set \(M^{\prime}(x,y,z)=r(M(x,y,z))\) for \(x,y,z\in r(X)\). A topological space with a continuous Mal'tsev operation is called a _Mal'tsev spaces_. On topological groups, the Mal'tsev operation \(M_{G}\) is continuous, and on quasitopological groups \(M_{G}\) is separately continuous. In [6] it was proved that if \(X\) is a Lindelof \(\Sigma\) or a countably compact Tychonoff space with a separately continuous Mal'tsev operation, then \(X\) is an \(\omega\)-cellular space (in particular, \(X\) is ccc). It was also noted there that this assertion can be proved for Tychonoff \(\Sigma\)-spaces with countable extend. This article proves that if \(X\) is a regular \(\Sigma\)-space with countable extend with a separately continuous Mal'tsev operation, then \(X\) is an \(\omega\)-cellular space (Theorem 7). The above facts imply that a regular quasitopological \(\Sigma\) group with countable extend is \(\omega\)-cellular. Based on this fact, it is proved that the Lindelof \(\Sigma\) almost paratopological group is \(\omega\)-cellular (Theorem 10). In particular, an \(\sigma\)-compact regular quasitopological or almost paratopological group is ccc. ## 2 Definitions and notation The identity in the group \(G\) will be denoted as \(e\), the family of open neighborhoods \(e\) as \(\mathcal{N}_{e}\). Denote the multiplication in the group \[\mathfrak{m}:G\times G\to G,\ (x,y)\mapsto xy\] and operation taking the inverse element \[\mathfrak{i}:G\to G,\ x\mapsto x^{-1}.\] Let \(X\) be a topological space. Then two points \(x\) and \(y\) in \(X\) are _topologically indistinguishable_ if they have exactly the same neighborhoods; that is, \(x\in\overline{\{y\}}\) and \(y\in\overline{\{x\}}\). Points \(x\) and \(y\) in \(X\) are _topologically distinguishable_ if they are not topologically indistinguishable. The space \(X\) has _countable extend_ if every discrete closed subset is at most countable. A space \(X\) is _ccc_, or has the _Suslin property_, if every disjoint family of nonempty open sets in \(X\) is at most countable. Let us say that a space \(X\) is \(\omega\)_-cellular_ if for any family \(\lambda\) of \(G_{\delta}\)-subsets of \(X\) there exists a countable subfamily \(\mu\subset\lambda\) such that \(\overline{\bigcup\lambda}=\overline{\bigcup\mu}\). Clearly \(\omega\)-cellular spaces are ccc. A space \(X\) is called a _\(\Sigma\)-space_ if there exists a \(\sigma\)-locally finite family \(\mathcal{S}\) and the covering of \(\mathcal{C}\) by closed countably compact sets, such that if \(C\in\mathcal{C}\) and \(U\supset C\) is open, then \(C\subset S\subset U\) for some \(S\in\mathcal{S}\). Lindelof \(\Sigma\)-spaces are exactly the regular continuous images of perfect inverse images of metrizable separable spaces. \(\sigma\)-countably compact spaces are \(\Sigma\)-spaces. Regular \(\sigma\)-compact spaces are Lindelof \(\Sigma\)-spaces. **Statement 1**.: _A space \(X\) is a \(\Sigma\)-space with countable extend if and only if there is a countable family \(\mathcal{S}\) and the covering of \(\mathcal{C}\) by closed countably compact sets, such that if \(C\in\mathcal{C}\) and \(U\supset C\) are open, then \(C\subset S\subset U\) for some \(S\in\mathcal{S}\)._ Proof.: Let \(\mathcal{S}\) be a \(\sigma\)-discrete family and \(\mathcal{C}\) be a covering of closed countably compact sets, such that if \(C\in\mathcal{C}\) and \(U\supset C\) are open, then \(C\subset S\subset U\) for some \(S\in\mathcal{S}\). Since \(X\) has a countable extend, then \(\mathcal{S}\) is countable. Conversely, it suffices to note that a countable family is a \(\sigma\)-discrete family. A space is _\(\sigma\)-countably compact_ if it is the union of a countable family of countably compact subspaces. Let us call a space _closely \(\sigma\)-countably compact_ if it is the union of a countable family of closed countably compact subspaces. Any closely \(\sigma\)-countably compact space is a \(\Sigma\)-space with countable extend. ## 3 \(R_{0}\) spaces and groups Let \(X\) and \(Y\) be topological spaces. We say that a map \(f:X\to Y\)_preserves the topology_ (\(f\) is _topology-preserving_) if * the mapping \(f\) is surjective, continuous, open and closed, and the set \(U\subset X\) is open if and only if \(U=f^{-1}(f(U))\) and \(f(U)\) open. It is easy to verify the following assertion. **Proposition 1**.: _Let \((X,\mathcal{T}_{X})\) and \((Y,\mathcal{T}_{Y})\) be topological spaces, \(f:X\to Y\) be a surjective mapping. The following conditions are equivalent:_ 1. \(f\) _preserves the topology;_ 2. _mapping_ \[f^{\#}:\mathcal{T}_{Y}\to\mathcal{T}_{X},\ U\mapsto f^{-1}(U)\] _is a bijection;_ 3. \(f\) _is quotient and the subspace_ \(f^{-1}(y)\) _is antidiscrete for every_ \(y\in Y\)_._ The relation '\(x\) and \(y\) are topologically indistinguishable' on \(X\) is an equivalence relation, denote this equivalence as \(\sim_{T}\): \(x\!\sim_{T}y\) if and only if \(x\in\overline{\{y\}}\) and \(y\in\overline{\{x\}}\). The space quotient by this equivalence relation is denoted as \(X_{T_{0}}\) and \(\pi_{T_{0}}:X\to X_{T_{0}}\) is the quotient mapping. The space \(X_{T_{0}}\) is a \(T_{0}\) space and the mapping \(f=\pi_{T_{0}}:X\to Y=X_{T_{0}}\) preserves the topology. If \(X\) is a \(T_{0}\) space, then \(\pi_{T_{0}}\) is a homeomorphism. Recall the axiom of separation of \(R_{0}\) topological spaces. The space \(X\) is \(R_{0}\), or _symmetric_, if \(x\in\overline{\{y\}}\) implies \(y\in\overline{\{x\}}\) (that is, \(x\!\sim_{T}y\)) for any \(x,y\in X\)[7]. The space \(X\) is \(R_{0}\) if and only if \(X_{T_{0}}\) is \(T_{1}\). See Section 16.2 in [7]. ### Arkhangel'skii-Motorov theorem In 1984 D.B. Motorov constructed a famous example of a compact space that is not a retract of any compact space. This is the closure on the plane of the graph of the function \(\sin\frac{1}{x}\) with domain \((0,1]\) ([4]). A little later, A.V. Arkhangelsky published Motorov's improved results [2]. We need the following two statements from their papers. **Theorem 1** (Arkhangel'skii [2, Corollary 1]).: _Let \(X\) be a homogeneous space and \(\overline{\{x\}}\) is compact for \(x\in X\). Then \(X\) is a \(R_{0}\) space._ **Theorem 2** (Motorov).: _If \(X\) is a homogeneous compact space, then \(X\) is an \(R_{0}\) space._ ### Topology-preserving group homomorphisms **Proposition 2**.: _Let \(G\) be a semitopological group, \(H\) a group with topology, and \(\varphi:G\to H\) a surjective quotient homomorphism. Then \(\varphi\) is open and \(H\) is a semitopological group._ Proof.: Let \(U\) be an open subset of \(X\). Because \[\varphi^{-1}(\varphi(U))=U\ker\varphi\] and right shifts are homeomorphisms, then \(\varphi^{-1}(\varphi(U))\) and hence \(\varphi(U)\) are open. Let \(V\) be an open subset of \(H\), \(h\in H\) and \(g\in\varphi^{-1}(h)\). Because \[\varphi^{-1}(hV)=g\varphi^{-1}(V)\qquad\qquad\text{and}\qquad\quad\varphi^{-1 }(Vh)=\varphi^{-1}(V)g,\] then the sets \(hV\) and \(Vh\) are open. Propositions 1 and 2 imply the following assertion. **Proposition 3**.: _Let \(G\) be a semitopological group, \(H\) a group with topology, and \(\varphi:G\to H\) a surjective homomorphism. The following conditions are equivalent:_ 1. \(\varphi\) _preserves the topology;_ 2. _the mapping_ \(\varphi\) _is quotient and_ \(\ker\varphi\) _is antidiscrete._ _If \(\varphi\) preserves the topology, then \(H\) is a semitopological group._ The following theorem is practically the same as (8, Theorem 3.1). **Theorem 3**.: _Let \(G\) be a semitopological group, \(H=\overline{\{e\}}\cap\overline{\{e\}}^{-1}\). Then \(H\) is a normal antidiscrete subgroup, the quotient group \(G/H\) coincides with \(G_{T_{0}}\), \(G/H\) is a semitopological \(T_{0}\) group, and the quotient mapping \(G\to G/H\) is topology-preserving._ Proof.: Let \(x\in G\). Then \(x\in H\Leftrightarrow x\in\overline{\{e\}}\) and \(x^{-1}\in\overline{\{e\}}\Leftrightarrow x\in\overline{\{e\}}\) and \(e\in\overline{\{x\}}\Leftrightarrow e\sim_{T}x\). Hence \(H=\pi_{T_{0}}^{-1}(\pi_{T_{0}}(e))\). The equivalence relation \(\sim_{T}\) is invariant under right and left shifts, so the quotient set by this equivalence relation coincides with the right (and left) cosets with respect to the normal subgroup \(H\). It remains to apply Proposition 3. Theorems 3 and 1 imply the following assertion. **Theorem 4**.: _Let \(G\) be a semitopological group, \(H=\overline{\{e\}}\) be compact. Then \(G\) is the \(R_{0}\) space, \(H\) is a normal closed antidiscrete subgroup, the quotient group \(G/H\) coincides with \(G_{T_{0}}\), \(G/H\) is a semitopological \(T_{1}\) group, and the quotient mapping \(G\to G/H\) is topology-preserving._ ## 4 Almost paratopological groups **Definition 1**.: A semitopological group \(G\) is called _almost paratopological_, if for any \(g\in G\) such that \(e\notin\overline{\{g\}}\) there exists a neighborhood \(U\) of \(e\) such that \(g\notin U^{2}\). **Theorem 5**.: _Let \(G\) be a semitopological group. If any of the following conditions is satisfied, then \(G\) is almost paratopological._ 1. \(G\) _is a subgroup of an almost paratopological group;_ 2. \(G\) _is the product of almost paratopological groups;_ 3. \(G\) _is a paratopological group;_ 4. \(G\) _is a Hausdorff quasitopological group;_ 5. _there exists a continuous isomorphism of_ \(G\) _onto a_ \(T_{1}\) _almost paratopological group._ Proof.: Conditions (1) and (3) are obvious. (2) Let \(G=\prod_{\alpha\in A}G_{\alpha}\), where \(G_{\alpha}\) is an almost paratopological group for all \(\alpha\in A\). Let \(g=(g_{\alpha})_{\alpha\in A}\in G\) and \(e=(e_{\alpha})_{\alpha\in A}\notin\overline{\{g\}}\). Then \(e_{\alpha}\notin\overline{\{g_{\alpha}\}}\) for some \(\alpha\in A\). There is a neighborhood \(U_{\alpha}\) of \(e_{\alpha}\) such that \(g_{\alpha}\notin U_{\alpha}^{2}\). Let \(U\) be the inverse image of \(U_{\alpha}\) under the projection of \(G\) onto \(G_{\alpha}\). Then \(g\notin U^{2}\). (4) Let \(g\in G\setminus\{e\}\). There is \(U\in\mathcal{N}_{e}\) for which \(U=U^{-1}\) and \(U\cap gU=\varnothing\). Then \(g\notin U^{2}\). (5) Let \(\varphi:G\to H\) be a continuous isomorphism of the group \(G\) onto a \(T_{1}\) almost paratopological group \(H\). Let \(g\in G\setminus\{e\}\). There is a neighborhood \(V\subset H\) of \(\varphi(e)\) such that \(\varphi(g)\notin V^{2}\). Then \(g\notin U^{2}\), where \(U=\varphi^{-1}(V)\). **Example 1**.: Let \(G\) be the group of integers with topology consisting of co-finite sets. The group \(G\) is a quasitopological compact \(T_{1}\) group which is not almost paratopological. **Proposition 4**.: _Let \(G\) be a semitopological group, \(M\subset G\). Then_ \[\overline{M}=\bigcap_{U\in\mathcal{N}_{e}}MU^{-1}=\bigcap_{U\in\mathcal{N}_{e }}U^{-1}M.\] Proof.: Let \(g\in\overline{M}\) and \(U\in\mathcal{N}_{e}\). Then \(gU\cap M\neq\varnothing\) and \(g\in MU^{-1}\). Hence \(\overline{M}\subset\bigcap_{U\in\mathcal{N}_{e}}MU^{-1}\). Let \(g\in\bigcap_{U\in\mathcal{N}_{e}}MU^{-1}\). Then \(gU\cap M\) for any \(U\in\mathcal{N}_{e}\). Hence \(\bigcap_{U\in\mathcal{N}_{e}}MU^{-1}\subset\overline{M}\). \(\overline{M}=\bigcap_{U\in\mathcal{N}_{e}}U^{-1}M\) is proved similarly. For a group \(G\) we denote \[S_{G}=\{(x,y)\in G\times G\,:\,xy=e\},\qquad\qquad E_{G}=\bigcap_{U\in\mathcal{N}_ {e}}\overline{U^{-1}}.\] **Proposition 5**.: _Let \(G\) be a semitopological group. Then_ 1. \[\overline{\{e\}}\subset E_{G}=\bigcap_{U\in\mathcal{N}_{e}}\left(U^{-1}\right) ^{2};\] 2. _the group_ \(G\) _is almost paratopological if and only if_ \(E_{G}=\overline{\{e\}}\)_;_ 3. \(\overline{S_{G}}=\mathfrak{m}^{-1}(E_{G})\)_;_ 4. _the following conditions are equivalent:_ 1. \(G\) _is an almost paratopological_ \(T_{1}\) _group;_ 2. \(E_{G}=\{e\}\)_;_ 3. \(S_{G}\) _is closed in_ \(G^{2}\)_._ Proof.: Denote \(Q=\bigcap_{U\in\mathcal{N}_{e}}\left(U^{-1}\right)^{2}\). (1). Proposition 4 implies that \(\overline{\{e\}}\subset E_{G}\subset Q\). Let \(x\in G\setminus E_{G}\). Then \(xU\cap U^{-1}=\varnothing\) for some \(U\in\mathcal{N}_{e}\). Hence \(x\notin(U^{-1})^{2}\supset Q\). We get \(Q\subset E_{G}\). (2). Assume that \(G\) is an almost paratopological group. It follows from (1) that \(\overline{\{e\}}\subset E_{G}\). Let \(g\in G\setminus\overline{\{e\}}\). Then \(e\notin\overline{\{g^{-1}\}}\). Since \(G\) is almost paratopological, it follows that \(g^{-1}\notin U^{2}\) for some \(U\in\mathcal{N}_{e}\). Then \(g\notin(U^{-1})^{2}\supset E_{G}\). Hence \(E_{G}\subset\overline{\{e\}}\). Suppose \(E_{G}=\overline{\{e\}}\). Let \(g\in G\) and \(e\notin\overline{\{g\}}\). Then \(g^{-1}\notin\overline{\{e\}}=E_{G}\). It follows from (1) that \(g^{-1}\notin(U^{-1})^{2}\) for some \(U\in\mathcal{N}_{e}\). We get, \(g\notin U^{2}\). (3). Let \((x,y)\in G\). Then \((x,y)\in\overline{S_{G}}\Leftrightarrow(Ux\times Uy)\cap S_{G}\) for all \(U\in\mathcal{N}_{e}\Leftrightarrow e\in UxUy\) for all \(U\in\mathcal{N}_{e}\Leftrightarrow e\in UUxy\) for all \(U\in\mathcal{N}_{e}\Leftrightarrow xy\in(U^{-1})^{2}\) for all \(U\in\mathcal{N}_{e}\Leftrightarrow xy\in Q\). It follows from (1) that \((x,y)\in\overline{S_{G}}\Leftrightarrow xy\in E_{G}\). (4). From (2) follows (a)\(\Leftrightarrow\)(b). Since \(S_{G}=\mathfrak{m}^{-1}(e)\), then (3) implies (b)\(\Leftrightarrow\)(c). For a group with the topology \((G,\mathcal{T})\), we denote \[\mathcal{T}_{Sym}=\{U\cap V^{-1}\,:\,U,V\in\mathcal{T}\}\] and \(\operatorname{Sym}G=(G,\mathcal{T}_{Sym})\). Obviously, \(\mathcal{T}_{Sym}\) is a topology that is stronger than \(\mathcal{T}\). **Proposition 6**.: _Let \(G\) be a group with a topology. Then_ 1. _if_ \(G\) _is a semitopological group, then_ \(\operatorname{Sym}G\) _is a quasitopological;_ 2. _if_ \(G\) _is a paratopological group, then_ \(\operatorname{Sym}G\) _is topological;_ 3. \(\operatorname{Sym}G\) _is homeomorphic to_ \(S_{G}\)_;_ 4. _if_ \(G\) _is an almost paratopological_ \(T_{1}\) _group, then_ 1. \(\operatorname{Sym}G\) _embeds closed in_ \(G^{2}\)_;_ 2. \(\operatorname{Sym}G\) _is a Hausdorff quasitopological group._ Proof.: Let \(\mathcal{T}\) be the topology of \(G\) and \(\mathcal{T}_{Sym}\) be the topology of \(\operatorname{Sym}G\). (1). If \(g\in G\), \(U,V\in\mathcal{T}\) and \(W=U\cap V^{-1}\in\mathcal{T}_{Sym}\) then \[gW =gU\cap(Vg^{-1})^{-1}\in\mathcal{T}_{Sym},\] \[Wg =Ug\cap(g^{-1}V)^{-1}\in\mathcal{T}_{Sym},\] \[W^{-1} =V\cap U^{-1}\in\mathcal{T}_{Sym}.\] Hence \(\operatorname{Sym}G\) is quasitopological. (2). Follows from (1). (3). Let's put \[\mathfrak{j}:\operatorname{Sym}G\to S_{G},\ x\mapsto(x,x^{-1}).\] The mapping \(\mathfrak{j}\) is a homeomorphism since \[\mathfrak{j}^{-1}(S_{G}\cap(U\times V))=U\cap V^{-1}.\] for \(U,V\in\mathcal{T}\). (4). From (3) and Proposition 5(4) follows (a). It follows from (1) that \(G\) is a quasitopological group. Let us show that \(\operatorname{Sym}G\) is a Hausdorff space. Let \(e\neq g\in G\). Since \(G\) is an almost paratopological \(T_{1}\) group, then \(g\notin U^{2}\) for some \(U\in\mathcal{N}_{e}\). Let \(S=U\cap U^{-1}\). Then \(S\) is an open neighborhood of the identity in \(\operatorname{Sym}G\) and \(eS\cap gS=\varnothing\). **Proposition 7**.: _Let \(G\) and \(H\) be groups with topology and \(\varphi:G\to H\) be a topology-preserving homomorphism. Let \(\mathcal{P}\) be one of the enumerated classes of groups: semitopological; quasitopological; paratopological; almost paratopological; compact. Then \(G\in\mathcal{P}\) if and only if \(H\in\mathcal{P}\)._ **Theorem 6**.: _A compact almost paratopological group is a topological group._ Proof.: Let \(Q\) be a compact almost paratopological group. We set \(H=\overline{\{e\}}\). Theorem 4 implies that \(H\) is a normal closed antidiscrete subgroup, the quotient group \(G=Q/H\) is a \(T_{1}\) compact semitopological group, and the quotient mapping \(\varphi:Q\to G\) is a topology-preserving homomorphism. The Proposition 7 implies that \(G\) is a compact almost paratopological \(T_{1}\) group and \(G\) is a topological group if and only if \(Q\) is a topological group. Therefore, to prove the theorem, it suffices to prove that \(G\) is a topological group. It follows from the Proposition 6(1) that \(\operatorname{Sym}G\) is quasitopological. The Proposition 6(4)(a) implies that \(\operatorname{Sym}G\) embeds in \(G^{2}\) in a closed manner, and hence the group \(\operatorname{Sym}G\) is compact. It follows from the Proposition 6(4)(b) that \(\operatorname{Sym}G\) is Hausdorff. Hence \(\operatorname{Sym}G\) is a compact Hausdorff semitopological group. It follows from the Ellis theorem [3, Theorem 2] that \(\operatorname{Sym}G\) is a topological group. Let \(\mathcal{T}\) be the topology of \(G\) and \(\mathcal{T}_{Sym}\) be the topology of \(\operatorname{Sym}G\). Let us show that \(G\) is a Hausdorff space. Let \(e\neq g\in G\). Since \(G\) is an almost paratopological \(T_{1}\) group, it follows from Proposition 5(4) that \(E_{G}=\{e\}\) and hence \(g\notin\overline{U^{-1}}\) for some \(U\in\mathcal{N}_{e}\). **Claim**.: \(e\in\operatorname{Int}\overline{U^{-1}}\)_._ Proof.: Since \(\operatorname{Sym}G\) is a topological group and \(U^{-1}\in\mathcal{T}_{Sym}\), then \(S^{2}\subset U^{-1}\) for some \(S\in\mathcal{T}_{Sym}\) for which \(e\in S=S^{-1}\). Since \(\operatorname{Sym}G\) is compact, then \(G=\bigcup_{i=1}^{n}x_{i}S\) for some \(x_{1},x_{2},...,x_{n}\in G\). A topological space cannot be the union of a finite number of nowhere dense sets, so \(\operatorname{Int}\overline{x_{i}S}\neq\varnothing\) for some \(i\). Hence \(\operatorname{Int}\overline{S}\neq\varnothing\). Let \(q\in S\cap\operatorname{Int}\overline{S}\). Then \(e\in\operatorname{Int}\overline{q^{-1}S}\subset\overline{S^{2}}\subset \overline{U^{-1}}\). Hence \(e\in\operatorname{Int}\overline{U^{-1}}\). Set \(U_{g}=G\setminus\overline{U^{-1}}\) and \(U_{e}=\operatorname{Int}\overline{U^{-1}}\). Then \(g\in U_{g}\in\mathcal{T}\), \(e\in U_{e}\in\mathcal{T}\) and \(U_{g}\cap U_{e}=\varnothing\). Thus, the space \(G\) is Hausdorff. The group \(G\) is a Hausdorff compact semitopological group. It follows from Ellis [3, Theorem 2] that \(G\) is a topological group. Theorems 5 and 6 imply the following assertion. **Corollary 1** ([5, Lemma 6]).: _A compact paratopological group is a topological group._ ## 5 Spaces with separately continuous Mal'tsev operation For the space \(X\) we define the following property: 1. Let \(\{x_{\alpha}\,:\,\alpha<\omega_{1}\}\subset X\) and for \(\alpha<\omega_{1}\) let \(\mathcal{F}_{\alpha}\) be at most a countable family of closed subsets of \(X\). Then there exists \(\beta<\omega_{1}\) for which the following condition is satisfied: 1. exists \(y\in\overline{\{x_{\alpha}\,:\,\alpha<\beta\}}\) such that if \(\gamma<\beta\), \(F\in\mathcal{F}_{\gamma}\) and \(x_{\beta}\in F\), then \(y\in F\). The (U\({}_{\text{s}}\)) property is a strengthening of the (P) property from [6, 10] in the class of Tikhonov spaces. The (U\({}_{\text{s}}\)) property can be used for regular spaces. **Proposition 8**.: _Let \(X\) be a regular \(\Sigma\)-space with countable extend. Then (U\({}_{\text{s}}\)) is true for \(X\)._ Proof.: Let \(\mathcal{S}\) and \(\mathcal{C}\) be as in Statement 1. We can assume that the family \(\mathcal{S}\) is closed under finite intersections. Let \(\{x_{\alpha}\,:\,\alpha<\omega_{1}\}\subset X\) and for \(\alpha<\omega_{1}\) let \(\mathcal{F}_{\alpha}\) be at most a countable family of closed subsets of \(X\). Denote \[F_{\gamma}^{*} =\{\bigcap\mathcal{F}\,:\,\mathcal{F}\subset\bigcup_{\alpha< \gamma}\mathcal{F}_{\alpha},\ |\mathcal{F}|<\omega,\bigcap\mathcal{F}\neq\varnothing\},\] \[X_{\gamma} =\{x_{\alpha}\,:\,\alpha<\gamma\}\] for \(\gamma\leq\omega_{1}\). By induction on \(n\in\omega\) we construct an increasing sequence of countable ordinals \((\beta_{n})_{n\in\omega}\) such that for \(n>0\) the following condition is satisfied: 1. if \(x_{\gamma}\in S\cap F\) for \(\gamma<\omega_{1}\), \(S\in\mathcal{S}\) and \(F\in\mathcal{F}_{\beta_{n-1}}^{*}\), then there exists \(y\in\overline{X_{\beta_{n}}}\) such that \(y\in S\cap F\). Let \(\beta_{0}=\omega\). Suppose that \(n>0\) and \(\beta_{0}<\beta_{1}<...<\beta_{n-1}<\omega_{1}\) are constructed. For \(S\in\mathcal{S}\) and \(F\in\mathcal{F}_{\beta_{n-1}}^{*}\) we denote \[A_{S,F}=\{\alpha<\omega_{1}\,:\,x_{\alpha}\in S\cap F\}.\] If \(A_{S,F}\neq\varnothing\), then we denote \(\alpha_{S,F}=\min A_{S,F}\). Let us put \[\beta_{n}=\sup\{\alpha_{S,F}\,:\,S\in\mathcal{S},F\in\mathcal{F}_{\beta_{n-1}} ^{*}\text{ and }A_{S,F}\neq\varnothing\}+1.\] The sequence \((\beta_{n})_{n\in\omega}\) is constructed. Let \(\beta=\sup\{\beta_{n}\,:\,n\in\omega\}\). Let us check \((*)\) in \(\mathrm{(U_{s})}\) definition. There is \(C\in\mathcal{C}\) for which \(x_{\beta}\in C\). Let us put \[\mathcal{S}^{\prime} =\{S\in\mathcal{S}\,:\,C\subset S\}=\{S_{n}^{\prime}\,:\,n\in \omega\},\] \[\mathcal{F}^{\prime} =\{F\in\mathcal{F}_{\beta}^{*}\,:\,x_{\beta}\in F\}=\{F_{n}^{ \prime}\,:\,n\in\omega\}.\] Let \(n\in\omega\). Let us put \[S_{n} =\bigcap_{i=0}^{n}S_{i}^{\prime}, F_{n} =\bigcap_{i=0}^{n}F_{i}^{\prime},\] \[\alpha_{n} =\alpha_{S_{n},F_{n}}, y_{n} =x_{\alpha_{n}}.\] Since \(\mathcal{F}_{\beta}^{*}=\bigcap_{m\in\omega}\mathcal{F}_{\beta_{m}}^{*}\), then \(F_{n}\in\mathcal{F}_{\beta_{m}}^{*}\) for some \(m\in\omega\). Since \(\beta\in A_{S_{n},F_{n}}\neq\varnothing\), then \(\alpha_{n}=\alpha_{S_{n},F_{n}}\leq\beta_{m+1}<\beta\). Hence \(y_{n}\in X_{\beta}\). It follows from the definition of the families \(\mathcal{S}\) and \(\mathcal{C}\) that the sequence \((y_{n})_{n\in\omega}\) accumulates to some point \(y\in C\cap\bigcap_{n\in\omega}F_{n}\). Since \((y_{n})_{n\in\omega}\subset X_{\beta}\), then \(y\in\overline{X_{\beta}}\). Let \(F\in\mathcal{F}_{\gamma}\) for \(\gamma<\beta\) and \(x_{\beta}\in F\). Then \(F\in\mathcal{F}^{\prime}\) and \(F=F_{m}^{\prime}\supset F_{m}\) for some \(m\in\omega\). Hence, \[y\in\bigcap_{n\in\omega}F_{n}\subset F_{m}\subset F.\] **Proposition 9**.: _Let \(X\) be a regular space with a separately continuous Mal'tsev operation and let \(X\) satisfy \(\mathrm{(U_{s})}\). Then \(X\) is an \(\omega\)-cellular space._ Proof.: Let us assume the opposite. Then there is a family \(\{K_{\alpha}^{\prime}\,:\,\alpha<\omega_{1}\}\) of non-empty sets of type \(G_{\delta}\), such that \[K_{\beta}^{\prime}\not\subset\overline{\bigcup_{\alpha<\beta}K_{\alpha}^{ \prime}}\] for \(\beta<\omega_{1}\). Let us choose \[x_{\beta}\in K_{\beta}^{\prime}\setminus\overline{\bigcup_{\alpha<\beta}K_{ \alpha}^{\prime}}\] and a sequence \((U_{\beta,n})_{n\in\omega}\) of open sets \(X\), such that \[x_{\beta}\in U_{n+1}\subset\overline{U_{n+1}}\subset U_{n}\] for \(n\in\omega\). Then \[x_{\beta}\in K_{\beta}=\bigcap_{n\in\omega}U_{\beta,n}\subset K^{\prime}_{\beta}\] And \[x_{\beta}\notin\overline{\bigcup_{\alpha<\beta}K_{\alpha}}. \tag{1}\] For \(\alpha,\gamma<\omega_{1}\) we put \[h_{\alpha,\gamma}:X\to X,\ x\mapsto M(x,x_{\alpha},x_{\gamma}).\] Note that \[h_{\alpha,\gamma}(x_{\alpha})=x_{\gamma}, h_{\alpha,\alpha}=\operatorname{id}_{X}.\] Let us put \[\mathcal{P}_{\beta} =\{h_{\alpha,\gamma}^{-1}(X\setminus U_{\gamma,n})\,:\,\alpha, \gamma<\beta\text{ and }n<\omega\},\] \[\mathcal{F}_{\beta} =\mathcal{P}_{\beta+1}.\] Since the condition (\(\operatorname{U_{s}}\)) is satisfied for \(X\), then there exists \(\beta<\omega_{1}\) and \[y\in\overline{\{x_{\alpha}\,:\,\alpha<\beta\}},\] such that if \(\gamma<\beta\), \(F\in\mathcal{F}_{\gamma}\) and \(x_{\beta}\in F\), then \(y\in F\). Then \[\text{if }x_{\beta}\in F\in\mathcal{P}_{\beta},\text{ then }y\in F. \tag{2}\] Let us put \[y_{\gamma}=M(x_{\beta},y,x_{\gamma})\] for \(\gamma<\beta\). **Claim.**\(y_{\gamma}\in K_{\gamma}\). Proof.: Suppose \(y_{\gamma}\notin K_{\gamma}\). Then \(y_{\gamma}\notin U_{\gamma,n}\) for some \(n\in\omega\). Let us put \[U_{1} =\{x\in X\,:\,M(y,x,x_{\gamma})\in U_{\gamma,n+1}\},\] \[U_{2} =\{x\in X\,:\,M(x_{\beta},x,x_{\gamma})\in X\setminus\overline{U _{\gamma,n+1}}\}.\] The sets \(U_{1}\) and \(U_{2}\) are open. Because \[x_{\gamma}=M(y,y,x_{\gamma})\in U_{\gamma,n+1},\] \[y_{\gamma}=M(x_{\beta},y,x_{\gamma})\notin U_{\gamma,n}\supset \overline{U_{\gamma,n+1}},\] then \(y\in U_{1}\cap U_{2}\). The set \(U_{1}\cap U_{2}\) is an open neighborhood of \(y\). Since \(y\in\overline{\{x_{\alpha}\,:\,\alpha<\beta\}}\), then \(x_{\alpha}\in U_{1}\cap U_{2}\) for some \(\alpha<\beta\). Then \[h_{\alpha,\gamma}(y) =M(y,x_{\alpha},x_{\gamma})\in U_{\gamma,n+1},\] \[h_{\alpha,\gamma}(x_{\beta}) =M(x_{\beta},x_{\alpha},x_{\gamma})\in X\setminus\overline{U_{ \gamma,n+1}}\subset X\setminus U_{\gamma,n+1}.\] Hence \[x_{\beta} \in h_{\alpha,\gamma}^{-1}(X\setminus U_{\gamma,n+1})=F\in\mathcal{P }_{\beta},\] \[y \in h_{\alpha,\gamma}^{-1}(U_{\gamma,n+1})=X\setminus F.\] Contradiction with (2). Let us put \[h:X\to X,\ x\mapsto M(x_{\beta},y,x).\] Since \(h\) is continuous, \(y_{\gamma}=h(x_{\gamma})\) for \(\gamma<\beta\) and \(y\in\overline{\{x_{\gamma}\,:\,\gamma<\beta\}}\), then \[x_{\beta}=M(x_{\beta},y,y)=h(y)\in\overline{h(\{x_{\gamma}\,:\,\gamma<\beta\} )}=\overline{\{y_{\gamma}\,:\,\gamma<\beta\}}.\] It follows from the claim that \[x_{\beta}\in\overline{\bigcup_{\gamma<\beta}K_{\gamma}}.\] Contradiction with (1). Propositions 8 and 9 imply the following sentence. **Theorem 7**.: _Let \(X\) be a regular \(\Sigma\)-space with countable extend and separately continuous Mal'tsev operation. Then \(X\) is an \(\omega\)-cellular space._ **Corollary 2**.: _Let \(X\) be a regular (closely \(\sigma\)-)countably compact space with separately continuous Mal'tsev operation. Then \(X\) is an \(\omega\)-cellular space._ ## 6 ccc in groups Since the Mal'tsev operation \(M_{G}(x,y,z)=xy^{-1}z\) on a quasitopological group is separately continuous, then Theorem 7 implies the following assertion. **Theorem 8**.: _Let \(G\) be a regular \(\Sigma\) quasitopological group with countable extend. Then \(G\) is an \(\omega\)-cellular space._ **Corollary 3**.: _Let \(G\) be a regular quasitopological group. If any of the following conditions is satisfied for \(G\), then \(G\) is an \(\omega\)-cellular space:_ 1. \(G\) _is closely_ \(\sigma\)_-countably compact;_ 2. \(G\) _is a Lindelof_ \(\Sigma\)_-space;_ 3. \(G\) _is_ \(\sigma\)_-compact._ **Theorem 9**.: _Let \(G\) be a regular almost paratopological group and \(G^{2}\) be a \(\Sigma\)-space with countable extend. Then \(G\) is an \(\omega\)-cellular space._ Proof.: The Proposition 6 implies that \(\operatorname{Sym}G\) embeds closed in \(G^{2}\) and is a quasitopological group. Hence \(\operatorname{Sym}G\) is a regular \(\Sigma\) quasitopological group with countable extend. Theorem 8 implies that \(\operatorname{Sym}G\) is \(\omega\)-cellular. Since \(G\) is a continuous image of \(\operatorname{Sym}G\), then \(G\) is \(\omega\)-cellular. Since the square of a Lindelof \(\Sigma\)-space is a Lindelof \(\Sigma\)-space, then the Theorem 9 implies the following assertion. **Theorem 10**.: _Let \(G\) be a regular Lindelof \(\Sigma\) almost paratopological group. Then \(G\) is an \(\omega\)-cellular space._ **Corollary 4**.: _Let \(G\) be a regular almost paratopological group. If any of the following conditions is satisfied for \(G\), then \(G\) is an \(\omega\)-cellular space:_ 1. \(G^{2}\) _is closely_ \(\sigma\)_-countably compact;_ 2. \(G\) _is a Lindelof_ \(\Sigma\)_-space;_ 3. \(G\) _is_ \(\sigma\)_-compact._ **Question 1**.: Let \(G\) be a semitopological group. Which of the following conditions imply that \(G\) is an \(\omega\)-cellular space (\(G\) is ccc)? 1. \(G\) is a \(\sigma\)-countably compact (regular) (almost paratopological, paratopological, quasitopological) group; 2. \(G\) is a closely \(\sigma\)-countably compact regular (almost paratopological, paratopological) group; 3. \(G\) is a (closely \(\sigma\)-)countably compact (almost paratopological, paratopological) group; 4. \(G\) is a \(\sigma\)-compact (almost paratopological, paratopological, quasitopological) group; 5. \(G\) is a Lindelof \(\Sigma\) group; 6. \(G\) is a \(\sigma\)-compact regular group. The author thanks the referee for useful comments.

2301.12293

ACL-Fig: A Dataset for Scientific Figure Classification

Most existing large-scale academic search engines are built to retrieve text-based information. However, there are no large-scale retrieval services for scientific figures and tables. One challenge for such services is understanding scientific figures' semantics, such as their types and purposes. A key obstacle is the need for datasets containing annotated scientific figures and tables, which can then be used for classification, question-answering, and auto-captioning. Here, we develop a pipeline that extracts figures and tables from the scientific literature and a deep-learning-based framework that classifies scientific figures using visual features. Using this pipeline, we built the first large-scale automatically annotated corpus, ACL-Fig, consisting of 112,052 scientific figures extracted from ~56K research papers in the ACL Anthology. The ACL-Fig-Pilot dataset contains 1,671 manually labeled scientific figures belonging to 19 categories. The dataset is accessible at https://huggingface.co/datasets/citeseerx/ACL-fig under a CC BY-NC license.

Zeba Karishma, Shaurya Rohatgi, Kavya Shrinivas Puranik, Jian Wu, C. Lee Giles

2023-01-28T20:27:35Z

http://arxiv.org/abs/2301.12293v1

# ACL-Fig: A Dataset for Scientific Figure Classification ###### Abstract Most existing large-scale academic search engines are built to retrieve text-based information. However, there are no large-scale retrieval services for scientific figures and tables. One challenge for such services is understanding scientific figures' semantics, such as their types and purposes. A key obstacle is the need for datasets containing annotated scientific figures and tables, which can then be used for classification, question-answering, and auto-captioning. Here, we develop a pipeline that extracts figures and tables from the scientific literature and a deep-learning-based framework that classifies scientific figures using visual features. Using this pipeline, we built the first large-scale automatically annotated corpus, ACL-Fig consisting of 112,052 scientific figures extracted from \(\approx 56\)K research papers in the ACL Anthology. The ACL-Fig-pilot dataset contains 1,671 manually labeled scientific figures belonging to 19 categories. The dataset is accessible at [https://huggingface.co/datasets/citeseerx/ACL-fig](https://huggingface.co/datasets/citeseerx/ACL-fig) under a CC BY-NC license. ## 1 Introduction Figures are ubiquitous in scientific papers illustrating experimental and analytical results. We refer to these figures as _scientific figures_ to distinguish them from natural images, which usually contain richer colors and gradients. Scientific figures provide a compact way to present numerical and categorical data, often facilitating researchers in drawing insights and conclusions. Machine understanding of scientific figures can assist in developing effective retrieval systems from the hundreds of millions of scientific papers readily available on the Web [1]. The state-of-the-art machine learning models can parse captions and shallow semantics for specific categories of scientific figures. [2] However, the task of reliably classifying general scientific figures based on their visual features remains a challenge. Here, we propose a pipeline to build categorized and contextualized scientific figure datasets. Applying the pipeline on 55,760 papers in the ACL Anthology (downloaded from [https://aclanthology.org/](https://aclanthology.org/) in mid-2021), we built two datasets: ACL-Fig and ACL-Fig-pilot. ACL-Fig consists of 112,052 scientific figures, their captions, inline references, and metadata. ACL-Fig-pilot (Figure 1) is a subset of unlabeled ACL-Fig, consisting of 1671 scientific figures, which were manually labeled into 19 categories. The ACL-Fig-pilot dataset was used as a benchmark for scientific figure classification. The pipeline is open-source and configurable, enabling others to expand the datasets from other scholarly datasets with pre-defined or new labels. ## 2 Related Work Scientific Figures ExtractionAutomatically extracting figures from scientific papers is essential for many downstream tasks, and many frameworks have been developed. A multi-entity extraction framework called PDFMEF incorporating a figure extraction module was proposed [3]. Shared tasks such as ImageCLEF [4] drew attention to compound figure detection and separation. [5] proposed a framework called PDFFigures that extracted figures and captions in research papers. The authors extended their work and built a more robust framework called PDFFigures2 [6]. DeepFigures was later proposed to incorporate deep neural network models [2]. Scientific Figure ClassificationScientific figure classification [7; 8] aids machines in understanding figures. Early work used a visual bag-of-words representation with a support vector machine classifier [7]. [9] applied hough transforms to recognize bar charts in document images. [10] used handcrafted features to classify charts in scientific documents. [11] combined convolutional neural networks (CNNs) and the deep belief networks, which showed improved performance compared with feature-based classifiers. Figure classification DatasetsThere are several existing datasets for figure classification such as DocFigure [12], FigureSeer [10], Revision [7], and datasets presented by [13] (Table 1). FigureQA is a public dataset that is similar to ours, consisting of over one million question-answer pairs grounded in over 100,000 synthesized scientific images [14] with five styles. Our dataset is different from FigureQA because the figures were directly extracted from research papers. Especially, the training data of DeepFigures are from arXiv and PubMed, labeled with only "figure" and "table", and does not include fine-granular labels. Our dataset contains fine-granular labels, inline context, and is compiled from a different domain. \begin{table} \begin{tabular}{l|l|l|l} \hline \hline Dataset & **Labels** & **\#Figures** & **Image Source** \\ \hline Deepchart & 5 & 5,000 & Web Image \\ Figreseer1 & 5 & 30,600 & Web Image \\ Prasad et al. & 5 & 653 & Web Image \\ Revision & 10 & 2,000 & Web Image \\ FigureQA3 & 5 & 100,000 & Synthetic figures \\ \hline DeepFigures & 2 & 1,718,000 & Scientific Papers \\ DocFigure2 & 28 & 33,000 & Scientific Papers \\ ACL-Fio-pilot & **19** & **1,671** & Scientific Papers \\ ACL-Fio (inferred)4 & - & **112,052** & Scientific Papers \\ \hline \hline \end{tabular} * Only 1000 images are public. * Not publicly available. * Scientific-style synthesized data. * ACL-Fio-inferred does not contain human-assigned labels. \end{table} Table 1: Scientific figure classification datasets. Figure 1: Example figures of each type in ACL-Fio-pilot. ## 3 Data Mining Methodology The ACL Anthology is a sizable, well-maintained PDF corpus with clean metadata covering papers in computational linguistics with freely available full-text. Previous work on figure classification used a set of pre-defined categories (e.g., [14], which may only cover some figure types. We use an unsupervised method to determine figure categories to overcome this limitation. After the category label is assigned, each figure is automatically annotated with metadata, captions, and inline references. The pipeline includes 3 steps: figure extraction, clustering, and automatic annotation (Figure 2). ### Figure Extraction To mitigate the potential bias of a single figure extractor, we extracted figures using pdffigures2[6] and deepfigures[2] which work in different ways. PDFFigures2 first identifies captions and the body text because they are identified relatively accurately. Regions containing figures can then be located by identifying rectangular bounding boxes adjacent to captions that do not overlap with the body text. DeepFigures uses the distant supervised learning method to induce labels of figures from a large collection of scientific documents in LaTeX and XML format. The model is based on TensorFlow, applying the Overfeat detection architecture to image embeddings generated using ResNet-101 [2]. We utilized the publicly available model weights1 trained on 4M induced figures and 1M induced tables for extraction. The model outputs the bounding boxes of figures and tables. Unless otherwise stated, we collectively refer to figures and tables together as "figures". We used multi-processing to process PDFs. Each process extracts figures following the steps below. The system processed, on average, 200 papers per minute on a Linux server with 24 cores. Footnote 1: [https://github.com/allenai/deepfigures-open](https://github.com/allenai/deepfigures-open) 1. Retrieve a paper identifier from the job queue. 2. Pull the paper from the file system. 3. Extract figures and captions from the paper. 4. Crop the figures out of the rendered PDFs using detected bounding boxes. 5. Save cropped figures in PNG format and the metadata in JSON format. Figure 2: Overview of the data generation pipeline. ### Clustering Methods Next, we use an unsupervised method to label extracted figures automatically. We extract visual features using VGG16 [15], pretrained on ImageNet [16]. All input figures are scaled to a dimension of \(224\times 224\) to be compatible with the input requirement of VGG16. The features were extracted from the second last hidden (dense) layer, consisting of 4096 features. Principal Component Analysis was adopted to reduce the dimension to 1000. Next, we cluster figures represented by the 1000-dimension vectors using \(k\)-means clustering. We compare two heuristic methods to determine the optimal number of clusters, including the Elbow method and the Silhouette Analysis [17]. The Elbow method examines the _explained variation_, a measure that quantifies the difference between the between-group variance to the total variance, as a function of the number of clusters. The pivot point (elbow) of the curve determines the number of clusters. Silhouette Analysis determines the number of clusters by measuring the distance between clusters. It considers multiple factors such as variance, skewness, and high-low differences and is usually preferred to the Elbow method. The Silhouette plot displays how close each point in one cluster is to points in the neighboring clusters, allowing us to assess the cluster number visually. ### Linking Figures to Metadata This module associates figures to metadata, including captions, inline reference, figure type, figure boundary coordinates, caption boundary coordinates, and figure text (text appearing on figures, only available for results from PDFFigures2). The figure type is determined in the clustering step above. The inline references are obtained using GROBID (see below). The other metadata fields were output by figure extractors. PDFFigures2 and DeepFigures extract the same metadata fields except for "image text" and "regionless captions" (captions for which no figure regions were found), which are only available for results of PDFFigures2. An inline reference is a text span that contains a reference to a figure or a table. inline references can help to understand the relationship between text and the objects it refers to. After processing a paper, GROBID outputs a TEI file (a type of XML file), containing marked-up full-text and references. We locate inline references using regular expressions and extract the sentences containing reference marks. ## 4 Results ### Figure Extraction The numbers of figures extracted by PDFFigures2 and DeepFigures are illustrated in Figure 3, which indicates a significant overlap between figures extracted by two software packages. However, either package extracted (\(\approx 5\%\)) figures that were not extracted by the other package. By inspecting a random sample of figures extracted by either software package, we found that DeepFigures tended to miss cases in which two figures were vertically adjacent to each other. We took the union of all figures extracted by both software packages to build the ACL-Fig dataset, which contains a total of 263,952 figures. All images extracted are converted to 100 DPI using standard OpenCV libraries. The total size of the data is \(\sim 25\)GB before compression. inline references were extracted using GROBID. About 78% figures have inline references. ### Automatic Figure Annotation The extraction outputs 151,900 tables and 112,052 figures. Only the figures were clustered using the \(k\)-means algorithm. We varied \(k\) from 2 to 20 with an increment of 1 to determine the number of clusters. The results were analyzed using Figure 3: Numbers of extracted images. the Elbow method and Silhouette Analysis. No evident elbow was observed in the Elbow method curve. The Silhouette diagram, a plot of the number of clusters versus silhouette score exhibited a clear turning point at \(k=15\), where the score reached the global maximum. Therefore, we grouped the figures into 15 clusters. To validate the clustering results, 100 figures randomly sampled from each cluster were visually inspected. During the inspection, we identified three new figure types: _word cloud_, _pareto_, and _venn diagram_. The ACL-Fig-pilot dataset was then built using all manually inspected figures. Two annotators manually labeled and inspected these clusters. The consensus rate was measured using Cohen's Kappa coefficient, which was \(\kappa-0.78\) (substantial agreement) for the ACL-Fig-pilot dataset. For completeness, we added 100 randomly selected tables. Therefore, the ACL-Fig-pilot dataset contains a total of 1671 figures and tables labeled with 19 classes. The distribution of all classes is shown in Figure 4. ## 5 Supervised Scientific Figure Classification Based on the ACL-Fig-pilot dataset, we train supervised classifiers. The dataset was split into a training and a test set (8:2 ratio). Three baseline models were investigated. Model 1 is a 3-Layer CNN, trained with a categorical cross-entropy loss function and the Adam optimizer. The model contains three typical convolutional layers, each followed by a max-pooling and a drop-out layer, and three fully-connected layers. The dimensions are reduced from \(32\times 32\) to \(16\times 16\) to \(8\times 8\). The last fully connected layer classifies the encoded vector into 19 classes. This classifier achieves an accuracy of 59%. Model 2 was trained based on the VGG16 architecture,except that the last three fully-connected layers in the original network were replaced by a long short-term memory layer, followed by dense layers for classification. This model achieved an accuracy of \(\sim 79\%\), 20% higher than Model 1. Model 3 was the Vision Transformer (ViT) [18], in which a figure was split into fixed-size patches. Each patch was then linearly embedded, supplemented by position embeddings. The resulting sequence of vectors was fed to a standard Transformer encoder. The ViT model achieved the best performance, with 83% accuracy. ## 6 Conclusion Based on the ACL Anthology papers, we designed a pipeline and used it to build a corpus of automatically labeled scientific figures with associated metadata and context information. This corpus, named ACL-Fig, consists of \(\approx 250\) Figure 4: Figure class distribution in the ACL-Fig-pilot dataset. objects, of which about 42% are figures and about 58% are tables. We also built ACL-Fig-pilot, a subset of ACL-Fig, consisting of 1671 scientific figures with 19 manually verified labels. Our dataset includes figures extracted from real-world data and contains more classes than existing datasets, e.g., DeepFigures and FigureQA. One limitation of our pipeline is that it used VGG16 pre-trained on ImageNet. In the future, we will improve figure representation by retraining more sophisticated models, e.g., CoCa, [19], on scientific figures. Another limitation was that determining the number of clusters required visual inspection. We will consider density-based methods to fully automate the clustering module.

2303.04288

Polynomial Time and Private Learning of Unbounded Gaussian Mixture Models

We study the problem of privately estimating the parameters of $d$-dimensional Gaussian Mixture Models (GMMs) with $k$ components. For this, we develop a technique to reduce the problem to its non-private counterpart. This allows us to privatize existing non-private algorithms in a blackbox manner, while incurring only a small overhead in the sample complexity and running time. As the main application of our framework, we develop an $(\varepsilon, \delta)$-differentially private algorithm to learn GMMs using the non-private algorithm of Moitra and Valiant [MV10] as a blackbox. Consequently, this gives the first sample complexity upper bound and first polynomial time algorithm for privately learning GMMs without any boundedness assumptions on the parameters. As part of our analysis, we prove a tight (up to a constant factor) lower bound on the total variation distance of high-dimensional Gaussians which can be of independent interest.

Jamil Arbas, Hassan Ashtiani, Christopher Liaw

2023-03-07T23:24:27Z

http://arxiv.org/abs/2303.04288v2

# Polynomial Time and Private Learning of Unbounded Gaussian Mixture Models ###### Abstract We study the problem of privately estimating the parameters of \(d\)-dimensional Gaussian Mixture Models (GMMs) with \(k\) components. For this, we develop a technique to reduce the problem to its non-private counterpart. This allows us to privatize existing non-private algorithms in a blackbox manner, while incurring only a small overhead in the sample complexity and running time. As the main application of our framework, we develop an \((\varepsilon,\delta)\)-differentially private algorithm to learn GMMs using the non-private algorithm of Moitra and Valiant [14] as a blackbox. Consequently, this gives the first sample complexity upper bound and first polynomial time algorithm for privately learning GMMs without any boundedness assumptions on the parameters. ## 1 Introduction The problem of learning the parameters of a Gaussian Mixture Model (GMM) is a fundamental problem in statistics, dating back to the early work of Karl Pearson in 1894 [10]. A GMM with \(k\) components in \(d\) dimensions can be represented as \((w_{i},\mu_{i},\Sigma_{i})_{i=1}^{k}\), where \(w_{i}\) is a mixing weight (\(w_{i}\geq 0\), and \(\sum_{i\in[k]}w_{i}=1\)), \(\mu_{i}\in\mathbb{R}^{d}\) is a mean, and \(\Sigma_{i}\in\mathbb{R}^{d\times d}\) is a covariance matrix (of the \(i\)-th Gaussian component). To draw a random instance from this GMM, one first samples an index \(i\in[k]\) (with probability \(w_{i}\)) and then returns a random sample from the Gaussian distribution \(\mathcal{N}(\mu_{i},\Sigma_{i})\). In this work we consider the problem of parameter estimation in the probably approximately correct (PAC) model, where the goal is to "approximately recover"1 the parameters of an unknown GMM given only independent samples from it. Footnote 1: See Definition 1.4 for the precise notion of distance. The sample complexity and computational complexity of learning the parameters of GMMs has been studied extensively. A notable breakthrough in this line of work was the development of polynomial-time methods (with respect to \(d\)) for learning GMMs [14, 2]. The running time and sample complexity of these methods is exponential \(k\), which is necessary for parameter estimation [14]. The above approaches, however, may not maintain privacy of the individuals whose data has been used for the estimation. To address this issue, we adopt the rigorous and widely accepted notion of differential privacy (DP) [13]. At a high-level, DP ensures that the contribution of each individual has only a small (indistinguishable) effect on the output of the estimator. The classical notion of \(\varepsilon\)-DP (pure DP) is, however, quite restrictive. For instance, even estimating the mean of an unbounded univariate Gaussian random variable in this model is impossible. Therefore, in line with recent work on private estimation in unbounded domains, we consider the \((\varepsilon,\delta)\)-DP (i.e. approximate differential privacy [1]) model. For the simpler case of multivariate Gaussians (without any boundedness assumptions on the parameters), it has been shown that learning with a finite number of samples is possible in the \((\varepsilon,\delta)\)-DP model [1]. More recently, computationally efficient estimators have been devised for the same task [1, 2, 3]. This begs answering the corresponding question for GMMs. Is there an \((\varepsilon,\delta)\)-DP estimator with a bounded sample complexity for learning unbounded GMMs? Is there a polynomial time estimator (in terms of \(d\)) for the same task? Note that if additional boundedness2 and strong separation3 assumptions are made about the GMM, then the work of [3] offers a positive answer to the above question in the \(\varepsilon\)-DP model. Our aim is, however, learning unbounded GMMs and also with minimal separation assumptions. Footnote 2: They assume there are known quantities \(R,\sigma_{max},\sigma_{min}\) such that \(\forall i\in[k],\|\mu_{i}\|_{2}\leq R\) and \(\sigma_{min}^{2}\leq||\Sigma_{i}||\leq\sigma_{max}^{2}\). Footnote 3: They assume \(\forall i\neq j,||\mu_{i}-\mu_{j}||_{2}\geq\widehat{\Omega}\left(\sqrt{k}+ \sqrt{\frac{1}{w_{i}}+\frac{1}{w_{j}}}\right)\cdot\max\left\{||\Sigma_{i}^{1/ 2}||,||\Sigma_{j}^{1/2}||\right\}\). To approach this problem, it is natural to ask if there is a general reduction from the private learning of GMMs to its non-private counterpart. If so, this would enable us to easily reuse existing results for non-private learning of GMMs. Is there a reduction from private to non-private learning of GMMs that incurs only a polynomial time and polynomial sample overhead? The main result of this paper is the existence of such a reduction; see Theorem 6.2 for a rigorous version. **Theorem 1.1** (**Private to Non-private Reduction for GMMs, Informal**).: _There is a reduction from learning the parameters of a GMM in the \((\varepsilon,\delta)\)-DP model to its non-private counterpart. Moreover, this reduction adds only polynomial time and sample overhead in terms of \(d\) and \(k\)._ This reduction, along with the non-private learner of [14] gives the first finite sample complexity upper bound for learning the parameters of unbounded GMMs in the \((\varepsilon,\delta)\)-DP model. Moreover, the resulting estimator essentially inherits all the properties of the non-private estimator of [14]; it runs in time that is polynomial in \(d\) (for fixed \(k\)) and shares the advantage of requiring provably minimal separability assumptions on the components of the GMM. Concurrent work.In an independent work, [11] offer an \((\varepsilon,\delta)\)-DP method for learning GMMs, removing the boundedness and strong separation requirements of [3]. However, they assume Gaussian components are spherical. We do not make that assumption, and learn the covariance matrices as well. ### Related Work Private Learning of a Single Gaussian.Karwa and Vadhan [10] established polynomial time and sample efficient methods for learning the mean and variance of a univariate Gaussian in both the pure and approximate-DP settings. Namely, in the \((\varepsilon,\delta)\)-DP setting, they can recover the mean and variance of the Gaussian without any boundedness assumption on the parameters. This result can be generalized to the multivariate setting [1, 2], where one finds Gaussians that approximate the underlying Gaussian in terms of total variation distance. However, the sample complexity of these methods depends on the condition number of the covariance matrix, and requires a priori bounds on the range of the parameters. The first finite sample complexity bound for privately learning unbounded Gaussians appeared in [1], nearly matching the sample complexity lower bound of [3]. The work of [1] relies on a private version of the minimum distance estimator [23] and is based on ideas from the private hypothesis selection method [16]. However, this method is not computationally efficient. Recently, several papers offered \((\varepsilon,\delta)\)-DP and computationally efficient algorithms for learning unbounded Gaussians [1, 15, 14], where the work of [1] achieved a near-optimal sample complexity for this task. Part of the approach of [1] is a sub-sample-and-aggregate scheme which we modify and use in this paper. FriendlyCore [13] is an alternative sample-and-aggregate framework that can be used for privately learning unbounded Gaussians. It is noteworthy that the approaches of [1, 12] work in the robust setting as well albeit with sub-optimal sample complexities. The recent work of [1] offers a robust and private learner with near-optimal sample requirements in terms of dimension. Finally, [11] ticks all the boxes by offering a sample near-optimal, robust, and efficient learner for unbounded Gaussians. Another related result is a sample-efficient and computationally efficient method for learning bounded and high-dimensional Gaussians in the \(\varepsilon\)-DP model [11]. There is also work on the problem of private mean estimation with respect to Mahalanobis distance [1, 12]. Finding private and robust estimators [13] and also the interplay between robustness and privacy [1, 1, 1, 15, 16] are subjects of a few recent papers. Parameter Learning for GMMs with PAC Guarantees.Given i.i.d. samples from a GMM, can we approximately recover its parameters? There has been an extensive amount of research in developing sample efficient and computationally efficient methods for learning the parameters of a GMM [1, 1, 1, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 222, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 287, 288, 289, 291, 292, 293, 294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 323, 334, 335, 35, 36, 37, 38, 39, 31, 33, 34, 36, 38, 39, 32, 33, 34, 36, 39, 33, 35, 37, 38, 39, 39, 30, 31, 32, 33, 34, 36, 39, 33, 36, 39, 31, 33, 32, 35, 37, 38, 39, 32, 36, 39, 33, 37, 39, 38, 39, 39, 31, 32, 33, 34, 36, 39, 32, 35, 38, 39, 33, 36, 39, 37, 38, 39, 31, 33, 34, 36, 39, 32, 37, 39, 33, 38, 39, 32, 39, 33, 34, 35, 36, 39, 37, 38, 39, 39, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 42, 44, 49, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 74, 75, 76, 77, 78, 79, 80, 81, 82, 82, 83, 84, 85, 86, 87, 88, 89, 90, 82, 84, 85, 86, 87, 88, 89, 91, 80, 83, 84, 85, 87, 89, 80, 84, 86, 88, 89, 80, 85, 87, 88, 89, 81, 82, 84, 85, 86, 88, 87, 89, 82, 88, 89, 80, 83, 84, 85, 88, 89, 80, 84, 86, 89, 81, 83, 85, 87, 88, 89, 80, 84, 88, 89, 82, 85, 86, 87, 88, 89, 82, 89, 83, 84, 85, 88, 89, 80, 86, 87, 88, 88, 89, 80, 87, 88, 89, 80, 88, 89, 81, 82, 83, 84, 85, 86, 88, 89, 82, 87, 88, 88, 89, 80, 88, 89, 82, 89, 83, 84, 85, 86, 87, 88, 89, 80, 88, 89, 81, 84, 86, 88, 89, 82, 85, 87, 88, 82, 86, 89, 83, 88, 89, 80, 84, 88, 85, 89, 86, 87, 88, 89, 82, 89, 83, 84, 86, 88, 87, 89, 80, 88, 85, 89, 80, 86, 87, 88, 89, 82, 89, 80, 88, 89, 80, 89, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 82, 89, 80, 83, 84, 87, 88, 85, 89, 86, 88, 89, 80, 87, 88, 89, 82, 89, 83, 85, 89, 80, 86, 88, 89, 80, 87, 88, 89, 80, 88, 89, 80, 89, 81, 80, 82, 83, 84, 85, 86, 87, 88, 89, 82, 83, 88, 89, 80, 84, 88, 85, 89, 86, 87, 88, 88, 89, 80, 88, 89, 82, 89, 80, 83, 84, 86, 88, 89, 85, 87, 88, 89, 80, 88, 89, 80, 89, 82, 89, 80, 81, 83, 84, 85, 86, 87, 88, 89, 82, 83, 85, 88, 89, 83, 86, 89, 84, 87, 88, 89, 85, 89, 86, 89, 87, 88, 89, 80, 89, 80, 88, 89, 82, 89, 80, 89, 80, 89, 80, 81, 82, 83, 84, 86, 88, 89, 80, 81, 82, 84, 85, 86, 87, 88, 89, 83, 82, 85, 87, 89, 84, 88, 89, 85, 86, 89, 87, 88, 89, 80, 89, 80, 82, 89, 80, 83, 8 ### Preliminaries We use \(\|v\|_{2}\) to denote the Euclidean norm of a vector \(v\in\mathbb{R}^{d}\) and \(\|A\|_{F}\) (resp. \(\|A\|\)) to denote the Frobenius (resp. spectral) norm of a matrix \(A\in\mathbb{R}^{d\times d}\). In this paper, we write \(\mathcal{S}^{d}\) to denote the positive-definite cone in \(\mathbb{R}^{d\times d}\). Let \(\mathcal{G}(d)=\{\mathcal{N}(\mu,\Sigma)\,:\,\mu\in\mathbb{R}^{d},\Sigma\in \mathcal{S}^{d}\}\) be the family of \(d\)-dimensional Gaussians. We can now define the class \(\mathcal{G}(d,k)\) of mixtures of Gaussians as follows. **Definition 1.2** (Gaussian Mixtures).: The class of mixtures of \(k\) Guassians in \(\mathbb{R}^{d}\) is defined by \(\mathcal{G}(d,k)\coloneqq\left\{\sum\limits_{i=1}^{k}w_{i}G_{i}\,:\,G_{i}\in \mathcal{G}(d),w_{i}\geq 0,\sum_{i=1}^{k}w_{i}=1\right\}\). We represent the Gaussian Mixture Model (GMM) by a set of \(k\) tuples \(\left(w_{i},\mu_{i},\Sigma_{i}\right)_{i=1}^{k}\), where each tuple represents the mean, covariance matrix, and mixing weight of one of its components. Note that the order of the components is important in our notation, since the order of the output may have an impact on the privacy. In the following definition and the remainder of the paper, we may abuse terminology and refer to a distribution via its probability density function (p.d.f.). **Definition 1.3** (Total Variation Distance).: Given two absolutely continuous probability measures \(f(x),g(x)\) on \(\mathbb{R}^{d}\), the total variation (TV) distance between \(f\) and \(g\) is defined as \(d_{\mathrm{TV}}\left(f(x),g(x)\right)=\frac{1}{2}\int_{\mathbb{R}^{d}}|f(x)-g (x)|\,\mathrm{d}x\). A standard way to define the distance between two GMMs is as follows [10][Definition 2]. **Definition 1.4** (The Distance between Two GMMs).: The \(\mathrm{dist}_{\mathrm{GMM}}\) distance between two GMMs is defined by \[\mathrm{dist}_{\mathrm{GMM}}\left(\left(w_{i},\mu_{i},\Sigma_{i}\right)_{i=1}^ {k},\left(w^{\prime}_{i},\mu^{\prime}_{i},\Sigma^{\prime}_{i}\right)_{i=1}^{k }\right)\,=\min_{\pi}\max_{i\in[k]}\max\left\{\left|w_{i}-w^{{}^{\prime}}_{ \pi(i)}\right|,d_{\mathrm{TV}}\left(\mathcal{N}(\mu_{i},\Sigma_{i}),\mathcal{ N}(\mu^{{}^{\prime}}_{\pi(i)},\Sigma^{{}^{\prime}}_{\pi(i)})\right)\,\right\}\] where \(\pi\) is chosen from the set of all permutations over \([k]\). If \(X\) (resp. \(Y\)) is a random variable distributed according to \(f\) (resp. \(g\)), we write \(d_{\mathrm{TV}}\left(X,Y\right)=d_{\mathrm{TV}}\left(f,g\right)\). We drop the reference to the p.d.f. of the random variable when it is clear or implicit from context. ### Differential Privacy Basics At a high-level, an algorithm is differentially private if, given two datasets that differ only in a single element, the output distribution of the algorithm are nearly the same4. Footnote 4: For sake of simplicity, we consider data sets to be ordered and therefore the neighboring data sets are defined based on their Hamming distances. However, one can easily translate guarantees proven for the ordered setting to the unordered one; see Proposition D.6 in [1]. **Definition 1.5** (Neighbouring Datasets).: Let \(\mathcal{X},\mathcal{Y}\) denote sets and \(n\in\mathbb{N}\). Two datasets \(D=(X_{1},\ldots,X_{n}),D^{\prime}=(X_{1},\ldots,X_{n})\in\mathcal{X}^{n}\) are said to be _neighbouring_ if \(d_{H}(D,D^{\prime})\leq 1\) where \(d_{H}\) denotes Hamming distance, i.e., \(d_{H}(D,D^{\prime})=|\{i\in[n]\,:\,X_{i}\neq X^{\prime}_{i}\}|\). **Definition 1.6** (\((\varepsilon,\delta)\)-Indistinguishable).: Let \(D,D^{\prime}\) be two distributions defined on a set \(\mathcal{Y}\). Then \(D,D^{\prime}\) are said to be \((\varepsilon,\delta)\)-indistinguishable if for all measurable \(S\subseteq\mathcal{Y}\), \(\mathbb{P}_{Y\sim D}\left[Y\in S\right]\leq e^{\varepsilon}\mathbb{P}_{Y\sim D ^{\prime}}\left[Y\in S\right]+\delta\) and \(\mathbb{P}_{Y\sim D^{\prime}}\left[Y\in S\right]\leq e^{\varepsilon}\mathbb{P} _{Y\sim D}\left[Y\in S\right]+\delta\). **Definition 1.7** (\((\varepsilon,\delta)\)-Differential Privacy [16]).: A randomized mechanism \(\mathcal{M}\colon\mathcal{X}^{n}\to\mathcal{Y}\) is said to be \((\varepsilon,\delta)\)-differentially private if for all neighbouring datasets \(D,D^{\prime}\in\mathcal{X}^{n}\), \(\mathcal{M}(D)\) and \(\mathcal{M}(D^{\prime})\) are \((\varepsilon,\delta)\)-indistinguishable. ### Techniques The techniques in this paper are inspired by the techniques in [1] which are based on the Propose-Test-Release framework [14] and the Subsample-And-Aggregate framework [20]. Given a dataset \(D\), we first split \(D\) into \(t\) sub-datasets and run a non-private algorithm \(\mathcal{A}\) on each of the sub-datasets. Next, we privately check if most of the outputs of \(\mathcal{A}\) are "well-clustered" (i.e., are close to each other). If not, then the algorithm fails as this suggests that the outputs of the non-private algorithm is not very stable (either due to lack of data or simply that the non-private algorithm is sensitive to its input). On the other hand, if most of the outputs are well-clustered then we can aggregate these clustered outputs and release a noisy version of it. There are, however, multiple additional technical challenges that need to be addressed. One core difficulty is the issue of the ordering of the Gaussian components. Namely, the non-private GMM learners may output GMM components in different orders. Therefore, aggregating these non-private solutions (e.g., by taking their weighted average in the style of [1]) seems impossible. We therefore propose to skip the aggregation step all together by simply picking an arbitrary solution from the cluster. Therefore, our private populous mean estimator (PPE) simplifies and generalizes the private populous mean estimator (PPME) framework of [1], making it applicable to general semimetric spaces (and therefore GMMs). A precise discussion of this framework is presented in Subsection 2.1. Another challenge is designing an appropriate mechanism for adding noise to GMMs. As discussed above, our framework requires that we are able to release a noisy output of a candidate output. More precisely, given two neighbouring datasets \(Y_{1},Y_{2}\), we want to design a mechanism \(\mathcal{B}\) such that \(\mathcal{B}(Y_{1})\), \(\mathcal{B}(Y_{2})\) are indistinguishable whenever \(Y_{1},Y_{2}\) are sufficiently close. As in [1], we refer to such a mechanism as a "masking mechanism". In the context of mixture distributions with \(k\) components, a candidate output corresponds to a \(k\)-tuple where each element of the tuple contain the parameters and the mixing weight of a single component. We prove that, if one can design a masking mechanism for a _single_ component then it is possible to use this masking mechanism as a blackbox to design a masking mechanism for the \(k\)-tuple with only a \(\text{poly}(k)\) overhead in the running time. One important ingredient is that we randomly shuffle the components, making the output invariant to the order of the components. Another challenge related to the order of components is that computing the distance between two GMMs based on Definition 1.4 requires minimizing over all permutations. A naive method for computing this distance could require exponential time but we show this task can be done in polynomial time using a simple reduction to bipartite matching. To showcase the utility of the above framework, we show that it is straightforward to apply the framework to privately learning mixtures of Gaussians. We design a masking mechanism of a single Gaussian component which consists of mixing the weight, the mean, and the covariance matrix. Masking the mixing weight is fairly standard while masking the mean and the covariance matrix can be done using known results (e.g. by using [1][Lemma 5.2] for the covariance matrix and a similar technique for the mean). Finally, we note that, in some of the literature for Gaussian mixtures, the results usually assert that for each Gaussian component \(\mathcal{N}(\mu,\Sigma)\), the algorithm returns \(\hat{\mu},\hat{\Sigma}\) such that \(\mathcal{N}(\mu,\Sigma)\) and \(\mathcal{N}(\hat{\mu},\hat{\Sigma})\) are close in _total variation_ distance (e.g. [13]). Our framework requires that \(\hat{\mu}\) (resp. \(\hat{\Sigma}\)) is close to \(\mu\) (resp. \(\Sigma\)) for some appropriate norm. Intuitively, this ought to be the case but no tight characterization was previously known unless the Gaussians had the same mean [13][Theorem 1.1]. In this paper, we prove the following tight characterization between the TV distance of a Gaussian and its parameters. We believe that such a result may be of independent interest. **Theorem 1.8**.: _Let \(\mu_{1},\mu_{2}\in\mathbb{R}^{d}\) and \(\Sigma_{1},\Sigma_{2}\) be \(d\times d\) positive-definite matrices. Suppose that we have \(d_{\mathrm{TV}}(\mathcal{N}(\mu_{1},\Sigma_{1}),\mathcal{N}(\mu_{2},\Sigma_{ 2}))<\frac{1}{600}\). Let_ \[\Delta=\max\left\{\|\Sigma_{1}^{-1/2}\Sigma_{2}\Sigma_{1}^{-1/2}-I_{d}\|_{F}, \|\Sigma_{1}^{-1}(\mu_{1}-\mu_{2})\|_{2}\right\}.\] _Then_ \[\frac{1}{200}\Delta\leq d_{\mathrm{TV}}\left(\mathcal{N}(\mu_{1},\Sigma_{1}), \mathcal{N}(\mu_{2},\Sigma_{2})\right)\leq\frac{1}{\sqrt{2}}\Delta.\] ## 2 Private Populous Estimator In this section, we describe our main framework which we call the "private populous estimator" (PPE). Before that, we need a few definitions. Semimetric spaces.In our application, we need to deal with distance functions which only satisfy an _approximate_ triangle inequality that hold only when the points are sufficiently close together. To that end, we first define the notion of a semimetric space. **Definition 2.1** (Semimetric Space).: We say \((\mathcal{F},\mathrm{dist})\) is a semimetric space if for every \(F,F_{1},F_{2},F_{3}\in\mathcal{F}\), the following conditions hold. 1. **Non-negativity.**\(\mathrm{dist}(F,F)=0\); \(\mathrm{dist}(F_{1},F_{2})\geq 0\). 2. **Symmetry.**\(\mathrm{dist}(F_{1},F_{2})=\mathrm{dist}(F_{2},F_{1})\). 3. \(z\)**-approximate \(r\)-restricted triangle inequality.** Let \(r>0\) and \(z\geq 1\). If \(\mathrm{dist}(F_{1},F_{2}),\mathrm{dist}(F_{2},F_{3})\leq r\) then \(\mathrm{dist}(F_{1},F_{3})\leq z\cdot(\mathrm{dist}(F_{1},F_{2})+\mathrm{dist} (F_{2},F_{3}))\). Masking mechanism.Intuitively, a masking mechanism \(\mathcal{B}\) is a random function that returns a noisy version of its input, with the goal of making close inputs indistinguishable. Formally, we define a masking mechanism as follows. **Definition 2.2** (Masking Mechanism [2], Definition 3.3).: Let \((\mathcal{F},\mathrm{dist})\) be a semimetric space. A randomized function \(\mathcal{B}\colon\mathcal{F}\to\mathcal{F}\) is a \((\gamma,\varepsilon,\delta)\)-masking mechanism for \((\mathcal{F},\mathrm{dist})\) if for all \(F,F^{\prime}\in\mathcal{F}\) satisfying \(\mathrm{dist}(F,F^{\prime})\leq\gamma\), we have that \(\mathcal{B}(F),\mathcal{B}(F^{\prime})\) are \((\varepsilon,\delta)\)-indistinguishable. Further, \(\mathcal{B}\) is said to be \((\alpha,\beta)\)-concentrated if for all \(F\in\mathcal{F}\), \(\mathbb{P}[\mathrm{dist}(\mathcal{B}(F),F)>\alpha]\leq\beta\). ### The Private Populous Estimator (PPE) In this section, we define the PPE framework which allows us to use non-private algorithms to design private algorithms. We represent the non-private algorithm by \(\mathcal{A}\colon\mathcal{X}^{*}\to\mathcal{Y}\) which takes elements from a dataset as inputs and outputs an element in \(\mathcal{Y}\). PPE requires two assumptions. Firstly, we assume that \((\mathcal{Y},\mathrm{dist})\) is a semimetric space. Secondly, we assume that we have access to an efficient masking mechanism for \((\mathcal{Y},\mathrm{dist})\). The PPE framework we introduce in this section can be seen as a somewhat generalized version of the framework used in [2] and requires fewer assumptions. Given a dataset \(D\) as inputs, we partition \(D\) into \(t\) disjoint subsets. Next, we run the non-private algorithm \(\mathcal{A}\) on each of these subsets to produce \(t\) outputs \(Y_{1},\ldots,Y_{t}\). We then privately check if most of the \(t\) outputs are close to each other. If not, PPE fails. Otherwise, it chooses a \(Y_{j}\) that is close to more than \(60\%\) of other \(Y_{i}\)'s. It then adds noise to \(Y_{j}\) using a masking mechanism \(\mathcal{B}\), and returns the masked version of \(Y_{j}\). The formal details of the algorithm can be found in Algorithm 1. The following theorem establishes the privacy and accuracy of Algorithm 1. The proof can be found in Appendix D.1. **Theorem 2.3**.: _Suppose that \((\mathcal{Y},\mathrm{dist})\) satisfies a \(z\)-approximate \(r\)-restricted triangle inequality. Further, suppose that \(\mathcal{B}\) is a \((r,\varepsilon,\delta)\)-masking mechanism._ * **Privacy.** _For_ \(t>5\)_, Algorithm_ 1 _is_ \((2\varepsilon,4e^{\varepsilon}\delta)\)_-DP._ * **Utility.**_Suppose \(\alpha\leq r/2z\) and \(t\geq(\frac{20}{\varepsilon}\ln\left(1+\frac{\varepsilon^{\varepsilon}-1}{2 \delta}\right)\). Let \(\mathcal{B}\) be \((\alpha/2z,\beta)\)-concentrated. If there exists \(Y^{*}\) with the property that for all \(i\in[t],\mathrm{dist}(Y^{*},Y_{i})<\alpha/2z\), then \(\mathbb{P}\left[\mathrm{dist}(\widetilde{Y},Y^{*})>\alpha\right]\leq\beta\)._ The utility guarantee asserts that if the outcome of all non-private procedures are close to each other, then the output of the PPE will be close to those non-private outcomes. **Remark 2.4**.: _Let \(T_{\mathcal{A}}\) be the running time of the algorithm \(\mathcal{A}\) in Line 2, \(T_{\mathrm{dist}}\) be the time to compute \(\mathrm{dist}(Y_{i},Y_{j})\) for any \(Y_{i},Y_{j}\in\mathcal{Y}\) in Line 3, and \(T_{\mathcal{B}}\) be the time to compute \(\widetilde{Y}\) in Line 9. Then Algorithm 1 runs in time \(O(t\cdot T_{\mathcal{A}}+t^{2}\cdot T_{\mathrm{dist}}+T_{\mathcal{B}})\). We will see that \(T_{\mathcal{A}}\), \(T_{\mathcal{B}}\), and \(T_{\mathrm{dist}}\) can be polynomially bounded for GMMs._ To apply Algorithm 1 for private learning of GMMs, we need to introduce a masking mechanism for them. In order to do that, we start by defining a masking mechanism for a single Gaussian component (presented in Section 4). We then show how one can convert a masking mechanism for a component to one for mixtures (Section 3). Finally, we apply this to come up with a masking mechanism for GMMs as shown in Section 5. ## 3 Masking Mixtures The goal of this section is to show how to "lift" a masking mechanism for a single component to a masking mechanism for mixtures. We can do this by adding noise to each of the components and randomly permute the output components. Formally, let \(\mathcal{F}\) denote a space and let \(\mathcal{F}^{k}=\mathcal{F}\times\ldots\times\mathcal{F}\) (\(k\) times). The following definition is useful in defining the distance between two mixtures, as it is invariant to the order of components. **Definition 3.1**.: Let \(\mathrm{dist}\) denote a distance function on \(\mathcal{F}\). We define \(\mathrm{dist}^{k}\colon\mathcal{F}^{k}\times\mathcal{F}^{k}\to\mathbb{R}_{ \geq 0}\) as \[\mathrm{dist}^{k}((F_{1},\ldots,F_{k}),(F_{1}^{\prime},\ldots,F_{k}^{\prime}) )\coloneqq\min_{\pi}\max_{i\in[k]}\mathrm{dist}(F_{i},F_{\pi(i)}^{\prime}),\] where the minimization is taken over all permutations \(\pi\). Note that computing \(\mathrm{dist}^{k}\) requires computing a minimum over all permutations \(\pi\). Naively, one might assume that this requires exponential time to try all permutations. However, it turns out that one can reduce the problem of computing \(\mathrm{dist}^{k}\) to deciding whether a perfect matching exists in a weighted bipartite graph. The details of this argument can be found in Appendix E.3. **Lemma 3.2**.: _If \(T_{\rm dist}\) is the running time to compute \({\rm dist}\) then \({\rm dist}^{k}\) can be computed in time \(O(k^{2}T_{\rm dist}+k^{3}\log k)\)._ The following definition is useful for extending a masking mechanism for a component to a masking mechanism for a mixture. The important thing is that the components are shuffled randomly in this mechanism, making the outcome independent of the original order of the components. **Definition 3.3**.: Suppose that \(\mathcal{B}\) is a \((\gamma,\varepsilon,\delta)\)-masking mechanism for \(\mathcal{F}\). We define the mechanism \(\mathcal{B}^{k}_{\sigma}\) as \(\mathcal{B}^{k}_{\sigma}(F_{1},\ldots,F_{k})=(\mathcal{B}(F_{\sigma(1)}), \ldots,\mathcal{B}(F_{\sigma(k)}))\), where \(\sigma\) is a uniform random permutation. We also note that \(\mathcal{B}^{k}_{\sigma}\) can be computed with only polynomial overhead. The proof can be found in Appendix E.2. **Lemma 3.4**.: _If \(T_{\mathcal{B}}\) is the running time of \(\mathcal{B}\) then \(\mathcal{B}^{k}_{\sigma}\) can be computed in time \(O(k\cdot T_{\mathcal{B}}+k\log k)\)._ The next lemma shows that \(\mathcal{B}^{k}_{\sigma}\) is indeed a masking mechanism w.r.t. \((\mathcal{F}^{k},{\rm dist}^{k})\) and that \(\mathcal{B}^{k}_{\sigma}\) is accurate provided that \(\mathcal{B}\) is accurate. **Lemma 3.5**.: _If \(\mathcal{B}\) is an \((\alpha,\beta)\)-concentrated \((\gamma,\varepsilon,\delta)\)-masking mechanism for \((\mathcal{F},{\rm dist})\) then, for any \(\delta^{\prime}>0\), \(\mathcal{B}^{k}_{\sigma}\) is an \((\alpha,k\beta)\)-concentrated \((\gamma,\varepsilon^{\prime},k\delta+\delta^{\prime})\)-masking mechanism for \((\mathcal{F}^{k},{\rm dist}^{k})\) where_ \[\varepsilon^{\prime}=\sqrt{2k\ln(1/\delta^{\prime})}\varepsilon+k\varepsilon( e^{\varepsilon}-1).\] Proof.: First, we prove privacy. Let \(F=(F_{1},\ldots,F_{k})\in\mathcal{F}^{k}\) and \(F^{\prime}=(F^{\prime}_{1},\ldots,F^{\prime}_{k})\in\mathcal{F}_{k}\) be such that \({\rm dist}^{k}(F,F^{\prime})\leq\gamma\). In other words, there exists a permutation \(\pi\) such that \({\rm dist}(F_{i},F^{\prime}_{\pi(i)})\leq\gamma\) for all \(i\in[k]\). Since \(\mathcal{B}\) is a \((\gamma,\varepsilon,\delta)\)-masking mechanism, we know that \(\mathcal{B}(F_{i}),\mathcal{B}(F^{\prime}_{\pi(i)})\) are \((\varepsilon,\delta)\)-indistinguishable. Thus, by advanced composition (see Theorem C.7), \((\mathcal{B}(F_{1}),\ldots,\mathcal{B}(F_{k}))\) and \((\mathcal{B}(F^{\prime}_{\pi(1)}),\ldots,\mathcal{B}(F^{\prime}_{\pi(k)}))\) are \((\varepsilon^{\prime},k\delta+\delta^{\prime})\)-indistinguishable with \(\varepsilon^{\prime}\) as stated in the lemma. Since \(\mathcal{B}^{k}_{\sigma}((F^{\prime}_{1},\ldots,F^{\prime}_{k}))\) has the same distribution has \(\mathcal{B}^{k}_{\sigma}((F^{\prime}_{\pi(1)},\ldots,F^{\prime}_{\pi(k)}))\), we conclude, using the fact that permutation preserves privacy (see Lemma C.8), that \(\mathcal{B}^{k}_{\sigma}(F)\) and \(\mathcal{B}^{k}_{\sigma}(F^{\prime})\) are \((\varepsilon^{\prime},k\delta+\delta^{\prime})\)-indistinguishable. Finally, it remains to prove accuracy (i.e. that \(\mathcal{B}^{k}_{\sigma}\) is \((\alpha,k\beta)\)-concentrated). Indeed, given \(F=(F_{1},\ldots,F_{k})\in\mathcal{F}^{k}\), we know that \({\rm dist}(\mathcal{B}(F_{i}),F_{i})\leq\alpha\) with probability at least \(1-\beta\). Thus, by a union bound \({\rm dist}(\mathcal{B}(F_{i}),F_{i})\leq\alpha\) for all \(i\in[k]\) with probability at least \(1-k\beta\). We conclude that \({\rm dist}(\mathcal{B}(F),F)\leq\alpha\) with probability at least \(1-k\beta\). Recall that Theorem 2.3 requires that the distance function satisfies an \(r\)-restricted \(z\)-approximate. The following lemma shows that \({\rm dist}^{k}\) indeed does satisfy this property provided that \({\rm dist}\) does. The proof can be found in Appendix E.1. **Lemma 3.6**.: _If \({\rm dist}\) satisfies an \(r\)-restricted \(z\)-approximate triangle inequality then so does \({\rm dist}^{k}\)._ ## 4 Masking a Single Gaussian Component In this section, we develop a masking mechanism for a single Gaussian component. In the following section, we utilize this masking mechanism combined with the general purpose mechanism in Section 3 to develop a masking mechanism for GMMs. Let \(\mathcal{F}_{\textsc{Comp}}=\mathbb{R}\times\mathbb{R}^{d}\times\mathbb{R}^{d \times d}\) (corresponding to the weight \(w\), mean \(\mu\), and covariance matrix \(\Sigma\), respectively). Define \({\rm dist}_{\textsc{Comp}}\colon\mathcal{F}_{\textsc{Comp}}\times\mathcal{F}_ {\textsc{Comp}}\to\mathbb{R}_{\geq 0}\) as \[{\rm dist}_{\textsc{Comp}}((w_{1},\mu_{1},\Sigma_{1}),(w_{2},\mu_{2},\Sigma_{2}) )=\max\{|w_{1}-w_{2}|,{\rm dist}_{\textsc{Mean}}((\mu_{1},\Sigma_{1}),(\mu_{2},\Sigma_{2})),{\rm dist}_{\textsc{Cov}}(\Sigma_{1},\Sigma_{2})\},\] where \[\operatorname{dist}_{\textsc{Conv}}(\Sigma_{1},\Sigma_{2})=\max\{\|\Sigma_{1}^{1/ 2}\Sigma_{2}^{-1}\Sigma_{1}^{1/2}-I_{d}\|_{F},\|\Sigma_{2}^{1/2}\Sigma_{1}^{-1} \Sigma_{2}^{1/2}-I_{d}\|_{F}\}\] and \[\operatorname{dist}_{\textsc{Mean}}((\mu_{1},\Sigma_{1}),(\mu_{2},\Sigma_{2}) )=\max\{\|\mu_{1}-\mu_{2}\|_{\Sigma_{1}},\|\mu_{1}-\mu_{2}\|_{\Sigma_{2}}\}.\] First, we show that \(\operatorname{dist}_{\textsc{Comp}}\) satisfies an approximate triangle inequality; this is useful in order to use Theorem 2.3. **Lemma 4.1**.: \(\operatorname{dist}_{\textsc{Comp}}\) _satisfies a \(1\)-restricted \((3/2)\)-approximate triangle inequality._ Proof.: For any positive-definite matrix \(\Sigma\), \(\|\cdot\|_{\Sigma}\) is a metric and thus, \(\operatorname{dist}_{\textsc{Mean}}\) is a metric (and therefore satisfies the \(1\)-restricted \((3/2)\)-approximate triangle inequality). Next, \(\operatorname{dist}_{\textsc{Cov}}\) satisfies the \(1\)-restricted \((3/2)\)-approximate triangle inequality (see Lemma A.8). A straightforward calculation concludes that, as a result, \(\operatorname{dist}_{\textsc{Comp}}\) also satisfies a \(1\)-restricted \((3/2)\)-approximate triangle inequality. The following lemma gives a masking mechanism for a single Gaussian mechanism. The proof can be found Appendix F. The mechanism essentially noises the mixing weight, the mean, and the covariance matrix separately. For noising the mixing weight, one can do this using the Gaussian mechanism. Care must be taken to noise the mean and the covariance matrix. In both cases, we use the empirical covariance matrix itself to re-scale both the mean and the covariance matrix. Pseudocode for the various pieces can be found in Algorithm 2 (the first four functions). Note that the parameters \(\eta_{\textsc{N}},\eta_{\textsc{Mean}},\eta_{\textsc{Cov}}\) must be set correctly to ensure privacy and accuracy but these details are relegated to Appendix F. **Lemma 4.2**.: _For \(\gamma\leq\frac{\varepsilon\alpha}{C_{2}\sqrt{d(d+\ln(4/\beta))\cdot\ln(2/ \delta)}}\), there exists a \((\gamma,3\varepsilon,3\delta)\)-masking mechanism, \(\mathcal{B}_{\textsc{Comp}}\), for \((\mathcal{F}_{\textsc{Comp}},\operatorname{dist}_{\textsc{Comp}})\) that is \((\alpha,3\beta)\)-concentrated, where \(C_{2}\) is a universal constant._ ## 5 A Masking Mechanism for GMMs In this section, we show how to mask a mixture of \(k\) Gaussians. Let \(\mathcal{F}_{\textsc{Gmm}}=\mathcal{F}_{\textsc{Comp}}\times\ldots\times \mathcal{F}_{\textsc{Comp}}\) (\(k\) times). Note we drop \(k\) from \(\mathcal{F}_{\textsc{Gmm}}\) (and related notation below) since \(k\) is fixed and implied from context. Let \(\operatorname{dist}_{\textsc{Comp}}\) be as defined in Eq. (4) and define the distance \[\operatorname{dist}_{\textsc{Param}}(\{(w_{i},\mu_{i},\Sigma_{i})\}_{i\in[k]}, \{(w^{\prime}_{i},\mu^{\prime}_{i},\Sigma^{\prime}_{i})\}_{i\in[k]})=\min_{ \pi}\max_{i\in[k]}\operatorname{dist}_{\textsc{Comp}}((w_{\pi(i)},\mu_{\pi(i)}, \Sigma_{\pi(i)}),(w^{\prime}_{i},\mu^{\prime}_{i},\Sigma^{\prime}_{i})),\] where \(\pi\) is chosen from the set of all permutations over \([k]\). Now define the masking mechanism \[\mathcal{B}_{\textsc{Gmm}}(\{(w_{i},\mu_{i},\Sigma_{i})\}_{i\in[k]})=\{ \mathcal{B}_{\textsc{Comp}}(w_{\sigma(i)},\mu_{\sigma(i)},\Sigma_{\sigma(i)}) \}_{i\in[k]},\] where \(\mathcal{B}_{\textsc{Comp}}\) is the masking mechanism from Lemma 4.2 and \(\sigma\) is a permutation chosen uniformly at random from the set of all permutations over \([k]\). In words, \(\mathcal{B}_{\textsc{Gmm}}\) applies the masking mechanism \(\mathcal{B}_{\textsc{Comp}}\) from Section 4 to each component separately and then permutes the components. To summarize the entire masking mechanism for GMMs, we provide pseudocode in Algorithm 2. The following lemma asserts that \(\mathcal{B}_{\textsc{Gmm}}\) is indeed a masking mechanism. At a high-level, it follows by combining Lemma 4.2 with Lemma 3.5. The details can be found in Appendix G.1. **Lemma 5.1**.: _Let \(\varepsilon<\ln(2)/3\). There is a sufficiently large constant \(C_{2}\) such that for \(\gamma\leq\frac{\varepsilon\alpha}{C_{2}\sqrt{k\ln(2/\delta)}\sqrt{d(d+\ln(12k/ \beta))\cdot\ln(12k/\delta)}}\), \(\mathcal{B}_{\textsc{Gmm}}\) is a \((\gamma,\varepsilon,\delta)\)-masking mechanism with respect to \((\mathcal{F}_{\textsc{Gmm}},\operatorname{dist}_{\textsc{Param}})\). Moreover, \(\mathcal{B}_{\textsc{Gmm}}\) is \((\alpha,\beta)\)-concentrated._ Note that \(\operatorname{dist}_{\textsc{Param}}\) also satisfies a \(1\)-restricted \((3/2)\)-approximate triangle inequality since \(\operatorname{dist}_{\textsc{Comp}}\) does (see Appendix G.2 for a proof). **Lemma 5.2**.: \(\operatorname{dist}_{\textsc{Param}}\) _satisfies a \(1\)-restricted \((3/2)\)-approximate triangle inequality._ **Input:** GMM given by \(\{(w_{i},\mu_{i},\Sigma_{i})\}_{i\in[k]}\) and parameters \(\eta_{\textsc{W}},\eta_{\textsc{Mean}},\eta_{\textsc{Cov}}>0\) ``` 1:function\(\mathcal{R}_{\textsc{W}}(w)\)\(\triangleright\) Noise mixing weights 2:Return\(\max(0,w+\eta_{\textsc{W}}g)\) where \(g\sim\mathcal{N}(0,1)\). 3:endfunction 4:function\(\mathcal{R}_{\textsc{Mean}}(\mu,\Sigma)\)\(\triangleright\) Noise mean 5:Return\(\mu+\eta_{\textsc{Mean}}g\) where \(g\sim\mathcal{N}(0,\Sigma)\) 6:endfunction 7:function\(\mathcal{R}_{\textsc{Cov}}(\Sigma)\)\(\triangleright\) Noise covariance 8: Let \(G\in\mathbb{R}^{d\times d}\) matrix with independent \(\mathcal{N}(0,1)\) entries. 9:Return\(\Sigma^{1/2}(I_{d}+\eta_{\textsc{Cov}}G)(I_{d}+\eta_{\textsc{Cov}}G)^{\top} \Sigma^{1/2}\) 10:endfunction 11:function\(\mathcal{B}_{\textsc{Comp}}(w,\mu,\Sigma)\)\(\triangleright\) Mask component 12:Return\((\mathcal{R}_{\textsc{W}}(w),\mathcal{R}_{\textsc{Mean}}(\mu,\Sigma), \mathcal{R}_{\textsc{Cov}}(\Sigma))\) 13:endfunction 14:function\(\mathcal{B}_{\textsc{Gmm}}(\{(w_{i},\mu_{i},\Sigma_{i})\}_{i\in[k]})\)\(\triangleright\) Mask GMM 15: Let \(\sigma\) be uniformly random permutation. 16:\(\{(\hat{w}_{i},\hat{\mu},\hat{\Sigma}_{i})\}\leftarrow\{\mathcal{B}_{\textsc{ Comp}}(w_{\sigma(i)},\mu_{\sigma(i)},\Sigma_{\sigma(i)})\}\). 17: Normalize: \(\hat{w}_{i}\leftarrow\hat{w}_{i}/\sum_{i\in[k]}\hat{w}_{i}\). 18:Return\(\{(\hat{w}_{i},\hat{\mu},\hat{\Sigma}_{i})\}_{i\in[k]}\). 19:endfunction ``` **Algorithm 2** GMM Masking Mechanism ## 6Privately Learning GMMs At this point, we have everything we need to develop a private algorithm for learning the parameters of a GMM. First, we define the problem more formally. **Definition 6.1** (PAC Learning of Parameters of GMMs).: Let \(\mathcal{F}=\left\{\left(w^{j}_{i},\mu^{j}_{i},\Sigma^{j}_{i}\right)_{i=1}^{k} \right\}^{j}\) be a class of \(d\)-dimensional GMMs with \(k\) components5. Let \(\mathcal{A}\) be function that receives a sequence \(S\) of instances in \(\mathbb{R}^{d}\) and outputs a mixture \(\hat{F}=(\hat{w}_{i},\hat{\mu}_{i},\hat{\Sigma}_{i})_{i=1}^{k}\). Let \(m\colon(0,1)^{2}\times\mathbb{N}^{2}\to\mathbb{N}\). We say \(\mathcal{A}\) learns the parameters of \(\mathcal{F}\) with \(m\) samples if for every \(\alpha,\beta\in(0,1)\) and every \(F\in\mathcal{F}\), if \(S\) is an i.i.d. sample of size \(m(\alpha,\beta,k,d)\) from \(F\), then \(\operatorname{dist}_{\textsc{Gmm}}(F,\hat{F})<\alpha\) with probability at least \(1-\beta\). Footnote 5: For examples, it is standard to pick \(\mathcal{F}\) to be those GMMs that are separable/identifiable. Plugging the masking mechanism developed in Section 5 (in particular, Lemma 5.1 and Lemma 5.2) into PPE (Theorem 2.3) gives a private to non-private reduction for GMMs. **Theorem 6.2** (Private to Non-Private Reduction).: _Let \(\mathcal{F}\) be a subclass of GMMs with \(k\) components in \(\mathbb{R}^{d}\). Let \(\mathcal{A}\) be a non-private Algorithm that PAC learns the parameters of \(\mathcal{F}\) with respect to \(\operatorname{dist}_{\textsc{Gmm}}\) using \(m_{\textsc{non-private}}(\alpha,\beta,k,d)\) samples. Then for every \(\varepsilon<\ln(2)/3\), \(\delta\in(0,1)\), \(\gamma\leq\frac{\varepsilon\alpha}{C_{2}\sqrt{k\ln(2/\delta)}\sqrt{d(d+\ln(12k /\beta))\cdot\ln(12k/\delta)}}\) for a sufficiently large constant \(C\) and \(t=\max\{5,\lceil\frac{20}{\varepsilon}\ln(1+\frac{\varepsilon^{\prime}-1}{2 \delta})\rceil\}\), there is a learner \(\mathcal{A}_{\textsc{private}}\) with the following properties:_ 1. \(\mathcal{A}_{\textsc{private}}\) _is_ \((2\varepsilon,4e^{\varepsilon}\delta)\)_-DP._ 2. \(\mathcal{A}_{\textsc{private}}\) _PAC learns the parameters of_ \(\mathcal{F}\) _using_ \(O(m_{\textsc{non-private}}(\gamma,\beta/2t,k,d)\log(1/\delta)/\varepsilon)\) _samples._ 3. \(\mathcal{A}_{\textsc{private}}\) _runs in time_ \(O((\log(1/\delta)/\varepsilon)\cdot T_{\mathcal{A}}+(\log(1/\delta)/ \varepsilon)^{2}\cdot(k^{2}d^{3}+k^{3}\log k))\)_, where_ \(T_{\mathcal{A}}\) _is the running time for the non-private algorithm._ To prove Theorem 6.2, we require the following lemma whose proof can be found in Appendix H. **Lemma 6.3**.: _Let \(F=(w_{i},\mu_{i},\Sigma_{i})_{i=1}^{k}\) and \(F^{\prime}=(w^{\prime}_{i},\mu^{\prime}_{i},\Sigma^{\prime}_{i})_{i=1}^{k}\) be two \(d\)-dimensional GMMs where \(\Sigma_{i}\) and \(\Sigma^{\prime}_{i}\) are positive-definite matrices. Suppose that \(\operatorname{dist}_{\textsc{GMM}}\left(F,F^{\prime}\right)<\frac{1}{600}\). Then \(\frac{1}{200}\operatorname{dist}_{\textsc{Param}}(F,F^{\prime})\leq\frac{1}{ \sqrt{2}}\operatorname{dist}_{\textsc{Param}}(F,F^{\prime})\)._ Proof of Theorem 6.2.: Let \(z=3/2\), \(r=1\), and \(t\geq\frac{20}{\varepsilon}\ln\left(1+\frac{e^{\varepsilon}-1}{2\delta} \right)=O(\log(1/\delta)/\varepsilon)\). We run Algorithm 1 with the following. * For the non-private algorithm \(\mathcal{A}\), we use the algorithm from Theorem 6.5 with accuracy parameter \(\alpha/2z\) and failure probability \(\beta/2t\). * For the masking mechanism, we use the \((r,\varepsilon,\delta)\)-masking mechanism \(\mathcal{B}_{\textsc{GMM}}\) which is defined in Lemma 5.1. Further, this mechanism is \((\alpha/2z,\beta/2)\)-concentrated. * Finally, note that the distance function \(\operatorname{dist}_{\textsc{Param}}\) satisfies the \(z\)-approximate \(r\)-restricted triangle inequality (Lemma 5.2). Let \(F^{*}\) be the true GMM. Let \(F_{i}\) be the estimated GMMs computed by \(\mathcal{A}\) in Line 2 of Algorithm 1. Then the first item above guarantees that \(\operatorname{dist}_{\textsc{Param}}(F^{*},F_{i})\leq\alpha/2z\) for all \(i\in[t]\) with probability at least \(1-\beta/2\). We thus conclude that we have a private algorithm for learning GMMs that is \((2\varepsilon,4e^{\varepsilon}\delta)\)-DP and that returns \(\widetilde{F}\) satisfying \(\operatorname{dist}_{\textsc{Param}}(\widetilde{F},F^{*})\leq\alpha\) with probability \(1-\beta\). By Lemma 6.3, we further conclude that \(\operatorname{dist}_{\textsc{GMM}}(\widetilde{F},F^{*})\leq O(\alpha)\) with probability \(1-\beta\). It remains to check the sample complexity and computational complexity of our algorithm. Since we run \(t\) independent instances of the non-private algorithm \(\mathcal{A}\), we require \(t\cdot m_{\textsc{private}}(\alpha/2z,\beta/2t,k,d)=O(m_{\textsc{private}}( \alpha/2z,\beta/2t,k,d)\cdot\log(1/\delta)/\varepsilon)\) samples. Finally, we bound the running time. Lemma 3.4 shows that the running time to apply the masking mechanism is \(O(k\cdot d^{3}+k\log k)\) and Lemma 3.2 shows that the running time to compute dist is \(O(k^{2}d^{3}+k^{3}\log k)\). The claimed running time now follows from Remark 2.4. ### Application As a concrete application, we apply Theorem 6.2 with the algorithm of [14] to obtain the first private algorithm for learning the parameters of a GMM with sample and computational complexity that is polynomial in \(d\) (for a fixed \(k\)) with minimal separation assumptions. Note that our algorithm does not require any boundedness assumptions on the parameters. **Definition 6.4** (\(\gamma\)-Statistically Learnable [14]).: We say a GMM \(F=(w_{i},\mu_{i},\Sigma_{i})_{i=1}^{k}\) is \(\gamma\)-statistically learnable if (i) \(\min_{i}w_{i}\geq\gamma\) and (ii) \(\min_{i\neq j}d_{\mathrm{TV}}\left(\mathcal{N}(\mu_{i},\Sigma_{i}),\mathcal{N} (\mu_{j},\Sigma_{j})\right)\geq\gamma\). If a GMM is \(\gamma\)-statistically learnable, we will be able to recover its components accurately. **Theorem 6.5** (Non-private Learning of GMMs [14]).: _There exists an algorithm \(\mathcal{A}\) and a function \(m_{\mathcal{A}}(d,k,\alpha,\beta)\) with the following guarantee. Fix \(\alpha,\beta\in(0,1)\), \(k,d\in\mathbb{N}\)._ * _For fixed_ \(k\)_, the sample complexity_ \(m_{\mathcal{A}}(d,k,\alpha,\beta)\) _is polynomial in_ \(d/\alpha\beta\)_._ * _For fixed_ \(k\)_,_ \(\mathcal{A}\) _runs in time_ \(\operatorname{poly}(d/\alpha\beta)\)_._ * _Let_ \(\mathcal{F}^{*}\) _be an_ \(\alpha\)_-statistically learnable subclass of GMMs with_ \(k\) _components in_ \(\mathbb{R}^{d}\) _and let_ \(F^{*}\in\mathcal{F}^{*}\)_. Given an i.i.d. sample_ \(D\) _of size_ \(m_{\mathcal{A}}(d,k,\alpha,\beta)\) _drawn from_ \(F^{*}\)_, with probability at least_ \(1-\beta\)_,_ \(\mathcal{A}\) _return_ \(\hat{F}\) _such that_ \(\operatorname{dist}_{\textsc{GMM}}(\hat{F},F^{*})\leq\alpha\)_._ The following corollary follows immediately by plugging Theorem 6.5 into Theorem 6.2. **Corollary 6.6**.: _There exists an algorithm \(\mathcal{A}\) and a function \(m_{\mathcal{A}}(d,k,\alpha,\beta,\varepsilon,\delta)\) with the following guarantee. Fix \(\alpha,\beta,\varepsilon,\delta\in(0,1)\), \(k,d\in\mathbb{N}\)._ * \(\mathcal{A}\) _is_ \((\varepsilon,\delta)\)_-DP._ * _For fixed_ \(k\)_, the sample complexity_ \(m_{\mathcal{A}}(d,k,\alpha,\beta,\varepsilon,\delta)\) _is polynomial in_ \(d\log(1/\delta)/\alpha\beta\varepsilon\)_._ * _For fixed_ \(k\)_,_ \(\mathcal{A}\) _runs in time_ \(\operatorname{poly}(d\log(1/\delta)/\alpha\beta\varepsilon)\)_._ * _Let_ \(\mathcal{F}^{*}\) _be an_ \(\alpha\)_-statistically learnable subclass of GMMs with_ \(k\) _components in_ \(\mathbb{R}^{d}\) _and let_ \(F^{*}\in\mathcal{F}^{*}\)_. Given an i.i.d. sample_ \(D\) _of size_ \(m_{\mathcal{A}}(d,k,\alpha,\beta,\varepsilon,\delta)\) _drawn from_ \(F^{*}\)_, with probability at least_ \(1-\beta\)_,_ \(\mathcal{A}\) _return_ \(\hat{F}\) _such that_ \(\operatorname{dist}_{\textsc{GMM}}(\hat{F},F^{*})\leq\alpha\)_._

2308.11291

Improving Knot Prediction in Wood Logs with Longitudinal Feature Propagation

The quality of a wood log in the wood industry depends heavily on the presence of both outer and inner defects, including inner knots that are a result of the growth of tree branches. Today, locating the inner knots require the use of expensive equipment such as X-ray scanners. In this paper, we address the task of predicting the location of inner defects from the outer shape of the logs. The dataset is built by extracting both the contours and the knots with X-ray measurements. We propose to solve this binary segmentation task by leveraging convolutional recurrent neural networks. Once the neural network is trained, inference can be performed from the outer shape measured with cheap devices such as laser profilers. We demonstrate the effectiveness of our approach on fir and spruce tree species and perform ablation on the recurrence to demonstrate its importance.

Salim Khazem, Jeremy Fix, Cédric Pradalier

2023-08-22T09:12:11Z

http://arxiv.org/abs/2308.11291v1

# Improving Knot Prediction in Wood Logs with Longitudinal Feature Propagation ###### Abstract The quality of a wood log in the wood industry depends heavily on the presence of both outer and inner defects, including inner knots that are a result of the growth of tree branches. Today, locating the inner knots require the use of expensive equipment such as X-ray scanners. In this paper, we address the task of predicting the location of inner defects from the outer shape of the logs. The dataset is built by extracting both the contours and the knots with X-ray measurements. We propose to solve this binary segmentation task by leveraging convolutional recurrent neural networks. Once the neural network is trained, inference can be performed from the outer shape measured with cheap devices such as laser profilers. We demonstrate the effectiveness of our approach on fir and spruce tree species and perform ablation on the recurrence to demonstrate its importance. Keywords:Knot segmentation Outer-Inter relationship prediction ConvLSTM ## 1 Introduction Distribution of knots within logs is one of the most important factor in wood processing chain since it determines how the log will be sliced and used. A knot is defined as a piece of a branch that is lodged in a stem and often starts at the stem path. Knots come in various dimensions, shapes and trajectories inside the trunk, these characteristics often depend on tree specie and environmental factors [15]. In wood processing, the knots are considered as defects that affect the quality of logs; hence, detecting their features such as position, size and angle of inclination are relevant and crucial for foresters and sawyers. Knowing these characteristics before the tree processing could generate a relative gain of 15-18% in value of products [2]. Nowadays, internal prediction of tree trunk density from bark observation is a complex and tedious task that requires a lot of human expertise or cannot be performed without expensive X-rays machines. In recent years, with the advent and success of deep learning, convolutional neural networks have achieved great performances on a variety of tasks such as object detection and image classification due to their strong features extraction capabilities [7, 13]. Compared to traditional methods, the data driven deep learning based approaches learn discriminative characteristics from annotated data automatically instead of human engineering. While the era of deep learning led to significant improvements in several areas of computer vision and natural language processing, there are still only a few paper which study the interest of these approaches for forestry and wood processing industry. This is due to the lack of open data, but also due to the lack of transfer of architectures that have demonstrated their efficiency in computer vision to the specific tasks of the forestry and wood processing industry. In this paper, we propose to explore an original task that does not seem to bear resemblance with a task in another domain: predicting the inner structure from the outer appearance of a wood log. The internal knots of a wood log are a consequence of the growth of branches of the tree and there is therefore, at least for some species, a causality between the presence of an inner knot and the growth or scar of an external branch. As we will demonstrate in the paper, the deformation of the outer surface of the tree, which is the consequence of the presence of branches, allows inferring the location and shape of inner knots. Our experiments are carried on conifers for which there is a clear relationship between the growth of branch and the knots. However, for other species such as deciduous trees, this relationship is unclear, and the task remains challenging. To solve the task of predicting the inner knots from the outer contour, we consider convolutional neural networks of the encoder-decoder family, where the encoder extracts features for the contour which are then used to decode the presence of a knot as a binary mask. Regularly spaced contour slices of the tree are provided as input to the network. As the presence of a knot is causally linked with a deformation of the contour due to a branch, inferring a knot needs to integrate features from the contour slices further away up or down the tree. To propagate these features between different slices, we consider convolutional LSTM, which are convolutional bidirectional recurrent neural networks [19]. A convolutional recurrent network keeps the spatial structure of the representation and extracts features along the recurrent paths by applying convolutions rather than dense matrix products. This has the benefit of reducing the cost of the net Figure 1: The recurrent neural network involves a recurrent encoder and feedforward decoder. The context along the slice dimension is propagated with convolutional LSTMs. work. In our task, this makes sense because a knot progressively diffuses within the wood as one moves along the longitudinal axis of the tree. That progressive diffusion induces that relevant features can be extracted locally, without having to resort to longer range interactions. Finally, given knots have various shapes and diffuses along a varying number of slices, using LSTMs lets the neural network learn how many slices need to be integrated to properly recover the shape of the knot. In summary, the main contribution of our work lies in two parts: * we propose to address an original machine learning task that is also valuable for the forestry industry, namely, the prediction of inner defects given observations of the outer deformation of a wood log, * we demonstrate the efficiency of integrating recurrent connections in the segmentation network to solve this task. The code used for running all the experiments of this paper are available on the following github repository: [https://github.com/jeremyfix/icvs2023](https://github.com/jeremyfix/icvs2023). ## 2 Related Work **Semantic segmentation** is a fundamental task in computer vision where the goal is to predict the label of each pixel in an image. Deep learning architectures for this task are typically based on the auto-encoder architecture. An autoencoder consists of an encoder and a decoder. The encoder maps the input data to a lower-dimensional latent space representation, while the decoder maps the latent space back to the original input data dimension [20]. In semantic segmentation, the decoder decodes the target labels instead of reconstructing the input. Fully Convolutional Networks (FCN) [14] is an important approach in semantic segmentation and has influenced the design of modern segmentation network. Other refinements of the encoder-decoder structure, such as U-Net and SegNet, have also been proposed in the literature [18, 1]. **Recurrent Neural Networks** have been introduced to deal with sequence data. They can learn the required size of the temporal window to gather the context required for taking a decision at any given time. The difficulty to integrate and propagate information through time, which is the foundation of the fundamental deep learning problem [10] of the vanishing/exploding gradient, has led authors to design dedicated memory units. Representatives of this family are the Long Short Term Memory networks (LSTMs) [6, 8] and Gated Recurrent Units networks (GRUs) [3]. **Convolutional LSTM** preserves the convolutional nature of the data [19]. Indeed, the recurrent weights in the LSTMs involve dense connections and do not exploit the spatial structure of the data they process. Convolution LSTM, by considering convolutional recurrent ways, do preserve the spatial nature of data and reduces the number of parameters required in the recurrent connections. In the original paper, the convolutional LSTMs have been successfully applied to spatio-temporal sequences for weather forecasting. In our work, we use an encoder-decoder architecture to predict the knot distribution (binary mask) from the slices of contours of the tree. To propagate encoder features through the slices, the encoder involves recurrent connections. In order to keep the convolutional nature of the representations, the encoder involves convolutional LSTM networks. Alternatively, we could have considered a 3D convolutional encoder, but this would have fixed the size of the slice context necessary to form a prediction. Using LSTMs let the network learn which contour features influence which other contour features. ## 3 Methodology ### Data Preprocessing In order to learn to transform the knot distribution from the contour of trees, we need aligned pairs of contours and knot masks. To build the input and target, we considered the pipelines of [11] for segmenting knots and identifying the contours by using X-rays data. Note that even though the pipelines of [11] are used to acquire data from X-rays images. The main objective of our approach is to avoid X-ray scanners and recover the external geometry from other modalities such as vision camera or laser profilers. The dataset is built from 27 fir trees and 15 spruce trees, with slices every 1.25 mm for tree sections of 1 meter long in average, which makes a total of 30100 slices. Each image is an \(512\times 512\) that is downscaled to \(256\times 256\) for the extraction of the contour and knot segmentation, and further downscaled to \(192\times 192\) for the sequence models presented in this paper. Every tree is sliced in blocks of 40 consecutive slices. In the following of the paper, the axis along which the slices are stacked will be referred as either the longitudinal axis or the z-axis for short. In the experiments, we used 18 fir tree and 8 spruce tree for the training set, we used 4 fir tree and 2 spruce tree for the validation and 5 tree of each specie for the test set. Note that, each tree is represented with by 800 slices. ### Neural network architectures without recurrent connections We trained two feedforward neural networks based on U-Net [18] and SegNet [1] in order to obtain a baseline to compare with the architecture involving recurrent connections along the z-axis. Although the U-Net and SegNet do not involve recurrent connections, these have been trained on the same data as the recurrent networks, e.g., stacks of slices. This allows to guarantee that training has been performed on the same data and the metrics are computed the same way. The U-Net encoder involves fewer channels than the original network to fit with the input data. The upsampling along the decoder path is performed using a nearest-pixel policy. Along the decoding path, the encoder features are concatenated with the decoder features. The SegNet encoder involves less channels and convolutional layers than the original network. The number of blocks and channels is reduced with respect to the original SegNet because our inputs are smaller. ### Neural network architectures with recurrent connections In order to propagate the contextual features of the contours in the encoder, we also consider neural network architectures with recurrent connections along the slice dimension (longitudinal axis of the tree). Recurrent connections are implemented with convolutional LSTMs which allow the network to learn which slice is impacting the features of another slice. We remind that the knots within a log can be causally linked to the presence of a branch. Instead of a fully connected LSTM, the convolutional LSTM involves fewer parameters by exploiting the spatial structure of the input data. In this paper, we consider recurrent connections only in the encoder part and not in the decoder part. The rationale is that introducing recurrent connections in the encoder allows the network to propagate contour features through the slices, and our experiments show that this is already sufficient to get good performances. These recurrent connections are bidirectional to allow information to propagate in both directions along the longitudinal axis. For the decoder, we do not add recurrent connections. That could be helpful but at a higher computational cost, and our experiments already demonstrated good performances with recurrent connections only in the encoder. The neural network architecture is depicted on Figure 1. The recurrent encoder is built from 3 consecutive ConvLSTMs bidirectional blocks. Every block has the same number of memory cells than the size of the spatial dimensions times the channel dimension. The input, output, and forget gates compute their values from a convolution with kernel size 3 from the "previous" sequence index (here, previous is to be considered along the longitudinal z-axis and be either following upward or downward directions given we consider bidirectional LSTMs). We use the same representation depth than for the SegNet with 32, 48 and 64 channels and a maxpooling layer is placed after every ConvLSTM layer to downscale spatially the representation by a factor of 2. The decoder is not recurrent and is the same as for our SegNet, namely 3 consecutive blocks with an upsampling (nearest) followed by a \(2\times[Conv2D(3\times 3)-BatchNorm-ReLU]\) block. The final layer is a \(Conv(1\times 1)\) to output the unnormalized scores for the classification of every pixel. ### Evaluation metrics Our experiments use different quantitative metrics to evaluate the quality and the performance of our method. For the segmentation task, the ground truth output is usually very sparse and there are much more negatives than positives. Hence, we need to use evaluation metrics that are not biased due to this class imbalance. We used the Dice similarity coefficient (Dice) [5], which is also known as F1-score as overlap metric, the Hausdorff Distance (HD) [9] as distance-based metric, and the Cohen's Kappa \(\kappa\)[4, 17] as counting-based metric to evaluate the segmentation results. The Hausdorff distance complements the Dice similarity because it indicates if false positives are close to a patch of positives or further away, while the Cohen's Kappa indicates the agreement between ground truth and the prediction. For each pixel, Cohen's Kappa compares the labels assigned by the model with the ground truth and measures the degree of agreement between them. The Cohen's Kappa ranges from -1 to 1 where a value of 1 indicates perfect agreement between the prediction and ground truth, whereas 0 indicates a prediction which is not better than random guessing and a negative value indicates less agreement than expected by chance. The advantage of using Cohen's Kappa is that it takes into account the possibility of chance agreement and provides a more accurate measure of agreement between prediction and ground truth, this is important in cases where the number of pixels assigned to each class is imbalanced. For the different equations, we denote FN, FP, TP, TN respectively the number of false negatives, false positives, true positives and true negatives, where \(\hat{y}\) is defined as final prediction computed by thresholding the output probability computed by the network (the threshold is set to 0.5 for all the experiments), and \(y\) the true value to be predicted (a mask, made of either 1 for a pixel belonging to a knot, or 0 otherwise). The metrics are always evaluated on the whole volume of 40 slices. As mentioned in section 3.2, even the feedforward neural networks (SegNet and UNet) are trained on the volumes. Although these networks do not propagate informations throught the longitudinal axis, training and evaluating these networks on the volume allow to have comparable measures (averaged on the same data). The value of the Hausdorff Distance is reported in millimeters. The metrics reported in the result section are averaged over the total number of volumes in the considered fold. \begin{table} \begin{tabular}{l||c c} Method & Dice/F1 \(\uparrow\) HD \(\downarrow\) \\ \hline \hline SegNet & 0.68 & 26.18 \\ U-Net & 0.72 & 47.80 \\ \hline ConvLSTM & **0.84** & **17.34** \\ \end{tabular} \end{table} Table 1: Left) Comparison of the segmentation methods on Dice score and HD using the validation fold. Right) Results of the SegNet and ConvLSTM models for a Fir tree specie. The first row corresponds to the input images, the second row is the associated ground truth and the bottom ones are the predictions. These samples all belong to the validation fold. Every column corresponds to one of 5 slices from different volumes. ### Other experimental hyperparameters For all the experiments presented in the paper, the optimization followed the same schedule. The networks have been trained for 150 epochs with a batch size of either 10 for U-Nets and ConvLSTMs, reduced to 4 for SegNet. The parameters have been optimized with Adam [12], a base learning rate of 0.0001. The loss is the binary cross entropy. ConvLSTMs trained for one week, the U-Net and SegNet trained for almost 10 days, using two RTX 3090. The experiments were coded either with Tensorflow 2.45 or Pytorch 1.9. We used Tensorboard6 to track the experiments and log the curves (loss and the different metrics). For regularizing the ConvLSTM encoder-decoder, a dropout layer is inserted between the encoder and decoder parts with a probability of 10% to mask a neuron. Following the original papers of U-Net and SegNet, we did not insert dropout layers in these networks. In all the trainings, data augmentation is applied to the input data with a random rotation out of 8 possible angles, and horizontal flip with a probability of 0.5. Footnote 5: [https://www.tensorflow.org](https://www.tensorflow.org) Footnote 6: [https://www.tensorflow.org/tensorboard](https://www.tensorflow.org/tensorboard) ## 4 Results In this section, we present both quantitatively and qualitatively the performances of the various models on the prediction of knots. The results on the validation fold and test folds are provided respectively in table 1 and table 2. For all the metrics, the ConvLSTM model performs better than the neural networks without recurrent connections. Looking only at the DICE and HD metrics, it seems that even without the recurrent connections, both the SegNet and U-Net perform reasonably well on the task. However, we observed qualitatively that this is not really the case as several knots are not predicted by these models. In that respect, the kappa metric seems to reflect more the difference in performance between the feedforward and recurrent networks. Including the context with the recurrent connections in the encoder provides a boost in performance. The quality of the segmentation of the recurrent network is better if we look at the Hausdorff distance, which means that the predicted masks with the ConvLSTM are closer in distance to the ground truth than with the non-recurrent segmentation networks. The Hausdorff distance is given in millimeters, and we remind that the slices are \(192\times 192\) pixels which correspond to \(192\mathrm{mm}\times 192\mathrm{mm}\). Additionally, we computed on the test set the Cohen's Kappa to evaluate the agreement between the predicted masks and the ground truth. The results show that the ConvLSTM achieves a score of 0.41 for fir trees and 0.21 for spruce indicating respectively moderate agreement and fair agreement, while the non-recurrent networks score lower with Kappa values between 0.05 and 0.12 for both species indicating very weak agreement. These findings demonstrate the boost provided by the recurrent networks. In table 2, right, we provide the metrics of the ConvLSTM model on the different trees of the test fold, either fir or spruce. The metrics computed on individual trees are consistent with the averaged metrics computed over all the volumes and reported in table 2, left. However, some spruce trees are particularly challenging. That's the case for example for the trees 4327 and 4948 which have a really unconventional contours, strongly distorted for some reason unknown to the authors. This out-of-distribution contours probably explains why the model fails to correctly predict all the knots. In addition to these averaged metrics, we provide in Figure 3 the distribution of Cohen's Kappa metric computed on the test fold for both fir and spruce trees, for both the ConvLSTM and SegNet networks. We observe that the ConvLSTM model outperforms the SegNet for all trees for both species, with a clear separation between the distributions. Specifically, the ConvLSTM model achieves nearly a twofold improvement over the SegNet for almost all trees. As the SegNet performs better on the test set than the U-Net, further comparison will only be made between SegNet and the ConvLSTM network. To better appreciate the difference in segmentation quality between the SegNet and ConvLSTM networks, the prediction masks of both networks on individual slices from different volumes are given in Table 1, right. On this figure, every column is a slice from a different volume of a fir tree and consecutive rows represent the input contour, the ground truth, the prediction of SegNet and the prediction of the ConvLSTM. From these 5 samples, it appears that SegNet usually underestimates knots and sometimes, knots may be even not predicted at all. For the ConvLSTM, most knots are predicted, although the knots might be overestimated in shape. \begin{table} \begin{tabular}{l c||c c c} \hline \hline \multirow{2}{*}{Specie} & Tree & \multicolumn{4}{c}{Metrics} \\ & ID & Dice \(\uparrow\) & HD \(\downarrow\) & Kappa \(\uparrow\) \\ \hline \multirow{4}{*}{**Fir**} & 4392 & 0.72 & 14.6 & 0.28 \\ & 4394 & 0.75 & 16.3 & 0.29 \\ & 4396 & 0.78 & 8.0 & 0.52 \\ & 5027 & 0.84 & 6.5 & 0.50 \\ & 5028 & 0.78 & 8.4 & 0.53 \\ \hline \multirow{4}{*}{**Spruce**} & 4327 & 0.70 & 29.0 & 0.12 \\ & 4328 & 0.72 & 19.2 & 0.12 \\ \cline{1-1} & 4329 & 0.73 & 9.1 & 0.25 \\ \cline{1-1} & 4948 & 0.70 & 31.0 & 0.11 \\ \cline{1-1} & 4990 & 0.73 & 13.6 & 0.26 \\ \hline \hline \end{tabular} \end{table} Table 2: Left) Comparison of the segmentation methods on Dice, HD and Kappa metrics on the test fold. Right) Quantitative results of the ConvLSTM model for the different trees of the test set. These are the same trees than the ones used for table on the left. The metrics are averaged over all the volumes of the same tree. All the trees had almost the same number of slices (from 800 to 810 slices). The predictions on some consecutive slices of the same volume of a tree are shown on Figures 2 for respectively a fir tree and a spruce tree. On the fir tree (left), we see part of the branch getting out from the tree, which is the anchor feature from which a network could learn the presence of an inner knot. Indeed, the ConvLSTM seems to be able to propagate information through the slices with its recurrent connections, as it is able to predict the location of a knot on the first of the five slices. It seems unlikely a network could predict the presence of a knot solely based on the first slice, given the deformation of the latter is barely visible on the contour of this first slice. Figure 3: Distribution of the Kappa metric on the fir and spruce trees of the test fold for both the SegNet and ConvLSTM neural networks. These are the same trees than used in table 2. Figure 2: Results of the SegNet and ConvLSTM models for a fir tree (left) or spruce tree specie (right) on 5 consecutive slices from the same volume. The first row corresponds to the input contours, the second row is the associated ground truth, and the two bottom rows are the predictions. These slices belong to a tree from the test set. The figure 4 shows a 3D representation of the contour of a fir tree from the test set, as well as the ground truth and the prediction produced by the ConvLSTM. The full tree represents a set of 803 slices, and all these slices are processed by sequences of 40 slices, with a stride of 1. From this full tree 3d representation, we observe that every knot present in the ground truth is also predicted by the ConvLSTM. It seems also that some knots may not have been correctly labeled as knots in the ground truth. This 3D representation also highlights the consistency of the knot predictions. From this representation, we also better see that there are various types of branch scars, some being clearly visible while others are more like little bumps on the surface of the tree. The smallest scars are certainly the ones for which it is the most challenging for the network to infer the location of knots, but even in some of these difficult cases, we see that the ConvLSTM model succeeds in predicting knots. ## 5 Discussion In this paper, we investigated a machine learning task that is highly valuable for the forestry industry : predicting the location of inner defects, knots, from the outside appearance of the tree. From the machine learning perspective, this task is original. We addressed this problem by training various neural network architectures of the encoder/decoder family and the most promising tested architectures are the convolutional LSTMs which benefit from recurrent connections along the longitudinal axis of the tree to propagate contour features reflecting the scars of a branch to the slices where the knots must be predicted. Although from the averaged metrics (DICE and Hausdorff), the feedforward networks (SegNet, U-Net) seem to perform well, it turns out that their predictions are pretty bad when we observe them qualitatively. This is not the case for the convolutional LSTM model, which have better metrics and clearly better segmentation of the knots when we check them visually. This discrepancy needs further investigation, and it is unclear why good classification metrics would lead to bad segmentation. The performances of the networks appear to be more contrasted by the Cohen's Kappa. The data used by the proposed machine learning pipeline relies on the work of [11] that extract contour and inner knots of tree logs from X-ray scans. X-ray Figure 4: 3D representation of the ground truth (left) and prediction of the ConvLSTM (right) viewed from the side of the tree or the top on both sides. Generated with Paraview. scans are only essential to produce the targets but are not used by our proposed approach. The required contours for the model can be obtained using laser scanners. We have a work in progress to create a platform with calibrated lasers to extract the contour of a tree log. From a machine learning perspective, the input contours are sparse, but dense representations are used for encoding. There is room for improvement in encoding and decoding methods. Instead of using a binary mask, encoding the contour as a polygon and utilizing graph neural networks for differentiable feature learning could be more efficient. Furthermore, recent research on neural radiance fields [16] suggests the possibility of encoding a 3D volume as a parameterized function, eliminating the need to explicitly construct the 3D volume of knots. Although these ideas require experimentation, a lightweight recurrent encoding of contours that parameterizes a 3D knot density function holds promise. ## Acknowledgment This research was made possible with the support from the French National Research Agency, in the framework of the project WoodSeer, ANR-19-CE10-011.

2310.02919

Attention-based Multi-task Learning for Base Editor Outcome Prediction

Human genetic diseases often arise from point mutations, emphasizing the critical need for precise genome editing techniques. Among these, base editing stands out as it allows targeted alterations at the single nucleotide level. However, its clinical application is hindered by low editing efficiency and unintended mutations, necessitating extensive trial-and-error experimentation in the laboratory. To speed up this process, we present an attention-based two-stage machine learning model that learns to predict the likelihood of all possible editing outcomes for a given genomic target sequence. We further propose a multi-task learning schema to jointly learn multiple base editors (i.e. variants) at once. Our model's predictions consistently demonstrated a strong correlation with the actual experimental results on multiple datasets and base editor variants. These results provide further validation for the models' capacity to enhance and accelerate the process of refining base editing designs.

Amina Mollaysa, Ahmed Allam, Michael Krauthammer

2023-10-04T16:01:06Z

http://arxiv.org/abs/2310.02919v2

# Attention-based Multi-task Learning for Base Editor Outcome Prediction ###### Abstract Human genetic diseases often arise from point mutations, emphasizing the critical need for precise genome editing techniques. Among these, base editing stands out as it allows targeted alterations at the single nucleotide level. However, its clinical application is hindered by low editing efficiency and unintended mutations, necessitating extensive trial-and-error experimentation in the laboratory. To speed up this process, we present an attention-based two-stage machine learning model that learns to predict the likelihood of all possible editing outcomes for a given genomic target sequence. We further propose a multi-task learning schema to jointly learn multiple base editors (i.e. variants) at once. Our model's predictions consistently demonstrated a strong correlation with the actual experimental results on multiple datasets and base editor variants. These results provide further validation for the models' capacity to enhance and accelerate the process of refining base editing designs. Machine Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning Multi-task Learning, Multi-task Learning, Multi-task Learning, Multi-task Learning Multi-task where we directly learn the probability distribution over all possible outcome sequences for a given target sequence. The second one is a two-stage model where we first estimate the probability of the given target sequence being edited, acknowledging that in many cases, the editor fails and no changes are observed which is often referred to as wild-type outcome. We then proceed to estimate the probability distribution of edited outcomes. Different editors exhibit varying behaviors on the same target sequences due to factors like binding affinities and editing window sizes, introducing distributional shifts. In response to this challenge, we introduce a multi-task learning framework. Rather than training individual models for each editor, as current models do, we propose a unified model capable of simultaneously accommodating multiple editors. In this work, we study the different modeling strategies for training machine learning models for the base editor outcome prediction task. We explore the spectrum of modeling choices evaluated on multiple datasets and base editors. A key highlight is the proposed unified multi-task model that is capable of learning from various base editors without necessitating training separate models for each setup. We train our models on six libraries corresponding to the outcomes of six base editors applied on thousands of target sites (Table 2). Our models' predictions show a good correlation with the ground truth across all datasets demonstrating the potential of machine learning in guiding and exploring genome editing space. ## 2 Related Work In recent years, the intersection of deep learning and CRISPR-Cas9 systems has witnessed substantial interest from the bioinformatics community. Researchers have explored the applications of deep learning in predicting various aspects of CRISPR-Cas9 systems, including predicting gRNA activities (Amenen et al., 2021; Xie et al., 2023; Zhang et al., 2021) and editing outcomes for both base editing and prime editing scenarios (Mathis et al., 2023). Among those, one notable approach is the BE-Hive proposed by (Arbab et al., 2020), which aims to predict base editing outcomes and efficiencies while considering sequence context, PAM compatibility, and cell-type-specific factors. The model employs a gradient boosting tree for predicting overall editing efficiency and a deep conditional autoregressive model for predicting probability of edited outcome sequences (denoted by bystander efficiency). Similarly, (Song et al., 2020) presented DeepABE and DeepCBE, that is based on convolutional neural networks to model both overall editing efficiency and bystander efficiency of adenine and cytosine base editors. Recently, Marquart et al. (2021) proposed BE-DICT, which predicts per-base editing efficiency (i.e. editing efficiency of each target base in a sequence) and bystander base-editing efficiency using attention-based deep learning. In a latest comprehensive study, (Kim et al., 2023) developed DeepCas9variants and DeepBEs to predict editing efficiencies and outcomes of various BEs, taking into account different Cas9 variants. They build on and adapt the models proposed in (Song et al., 2020) (i.e. convolutional networks) to generate predictions for a range of CRISPR-Cas9 systems. While the surge of interest in applying machine learning to CRISPR-Cas9 systems is clear in recent literature, it's noteworthy that many of these works have a primary emphasis on designing CRISPR-Cas9 systems under various conditions and less focused on the analysis of ML models without offering a holistic and systematic analysis of model design. Given the intricate nature of CRISPR-Cas9 systems and the multitude of model paradigms adopted, deriving concrete conclusions about optimal model design strategies remains elusive. In this context, our work aims to serve as model-first work that presents the base editing outcome prediction through a modeling lens. We focus on model development and provide a systematic analysis of each component of the models, providing a structured framework for problem formulation and model design specifically tailored to the prediction of base editing outcomes. Through this structured examination of these critical aspects, our aim is to lay the groundwork for more informed and refined approaches for using deep learning models to assist the design of base editors. ## 3 Method Base editor and related conceptsBase editors (BEs) are created by fusing the Cas9 protein with DNA-modifying enzymes. They are directed by a 20-base pair guiding RNA molecule (sgRNA) that acts as a GPS to locate and bind to a matching DNA segment known as the _protospacer_. The effectiveness of BEs largely depends on the composition of this protospacer sequence. BEs, in tandem with the sgRNA, can only bind to the DNA if there's a _protospacer adjacent motif_ (PAM) - a sequence consisting of 2-6 nucleotides - present adjacent to the protospacer. This PAM sequence further influences the activity of BEs. There are two primary types of base editors: adenine base editors (ABEs) (presented in figure 1), which convert adenine (A) to guanine (G), and cytosine base editors (CBEs) that chemically convert cytosine (C) to thymine (T). A detailed description of the base editor is provided in the appendix section 6.1.1. ### Data representation Assume we have a target (reference) DNA sequence denoted as \(\mathbf{x}_{\text{ref}}=[x_{1},x_{2},\ldots,x_{T}]\) where \(x_{i}\in\{A,C,G,T\}\), and a set of DNA sequences \(\mathbf{X}_{\text{out}}=[\mathbf{x}_{\text{out},1},\mathbf{x}_{\text{out},2}, \ldots,\mathbf{x}_{\text{out},M}]\in\mathbb{R}^{T}\). The _target_ of the DNA sequence \(\mathbf{x}_{\text{out},1}\) is the target of the DNA sequence \(\mathbf{x}_{\text{out},M}\), and the target of the DNA sequence \(\mathbf{x}_{\text{out},M}\) is the target of the DNA sequence \(\mathbf{x}_{\text{out},M}\). The target of the DNA sequence \(\mathbf{x}_{\text{out},M}\) is the target of the DNA sequence \(\mathbf{x}_{\text{out},M}\). \(\mathbb{R}^{M\times T}\) representing corresponding outcomes when a specific base editor is applied to the reference sequence \(\mathbf{x}_{\text{ref}}\). The associated probabilities for these outcomes are given by \(\mathbf{y}=[y_{1},y_{2},\dots,y_{M}]\) where \(y_{i}=P(\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}})\in[0,1]\), _for_\(i=1,2,\dots,M\), indicating the likelihood of obtaining outcome \(\mathbf{x}_{\text{out},i}\) through editing of \(\mathbf{x}_{\text{ref}}\). Here, \(T\) is the length of the reference sequence, and \(M\) represents the total number of possible outcomes for a given reference sequence. The count of outcomes can vary depending on the reference sequence. An example of a reference sequence and associated outcome sequences is represented in Figure 2. In this paper, we use bold uppercase letters for matrices (\(\mathbf{X}\)), bold lowercase letters for vectors or sequences (\(\mathbf{x}\)), and regular non-bold letters for scalars or tokens. We use \(P\) for probability distributions and non-bold uppercase letters (\(X\)) for random variables. To represent the reference sequence, we consider protospacer, PAM, and overhangs. Here, "overhangs" refer to adjacent nucleotides on both sides of the protospacer.. To declutter the notation, we will mainly use \(\mathbf{x}_{\text{ref}}\) to denote the reference sequence which could refer to one of these representations: (a) protospacer, (b) protospacer + PAM, or a (c) left overhangs + protospacer + PAM + right overhangs where + is the concatenation operator. Respectively, the outcome sequences are the DNA sequences with the same length as the reference sequence and with a modification of the target bases at the protospacer. The outcome sequence identical to the reference sequence (no edits) is referred as the wild-type. The training dataset comprises \(N\) pairs, each containing a reference sequence, its associated outcomes, and the corresponding probabilities, denoted as \(D=\{\mathbf{x}_{\text{ref}}^{i},\mathbf{X}_{\text{out}}^{i},\mathbf{y}^{i}\}_ {i=1}^{N}\). For simplicity, when referring to a specific reference sequence and its outputs, we omit the instance-level indexing and use only \(\mathbf{x}_{\text{ref}}\). ### Problem formulation Our objective is to predict the likelihood of potential outcomes resulting from a specific base editor applied to a reference sequence.One approach would be formulating it as a generative model where we directly model the condition distribution \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}})\) that we can both sample different outcomes for a given reference sequence and calculate the probability of each outcome. However, unlike typical generative models that must learn to generate entire output sequences, our scenario benefits from already knowing a portion of the output sequences. Due to the base editor's specific targeting of A-to-G or C-to-T transformations, a substantial portion of the output sequence remains consistent with the reference sequence, with only a few positions undergoing alteration. In the inference phase, for a given reference sequence, we can efficiently generate all possible outcomes by considering only the edit combination of target bases (A/G) within the protospacer. By traversing through a range of possible edits, we cover the entire landscape of potential outcome sequences. Therefore, we only need to learn the distribution \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}})\) such that we can evaluate the probability of a specific outcome for a given reference sequence \(P(X_{\text{out}}=\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}})\). One-stage ModelIn this setup, we tackle the problem by learning a function \(f(\mathbf{x}_{\text{ref}},\mathbf{x}_{\text{out},i})\rightarrow\hat{y}_{i}\) where \(i=1,\dots,M\), and \(\sum_{i=1}^{M}\hat{y}_{i}=1\), that takes as input the reference sequence and one of its corresponding outcome and learns to approximate the probability of obtaining that specific outcome. Notably, this function \(f\) characterizes a categorical distribution \(P(X_{\text{out}}=\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}})\sim Cat(M, \hat{\mathbf{y}})\), where \(\hat{\mathbf{y}}\) is the vector containing probabilities for M outcomes. To learn the function \(f\), we propose to use attention-based encoder blocks to learn the encoding of both the Figure 1: Adenine base editor Figure 2: An example of a reference sequence of 20 bases (i.e. nucleotides) and associated outcome sequences when applying ABEmax base editor. The first row represents the reference (target) sequence, and the second row is the outcome sequence with no modification (i.e. wild-type) with a probability of occurrence of 0.52. The third row represents a possible outcome sequence where the letter A is changed to G at position 5 with a probability of 0.35. The rest of the rows represent all possible changes of the reference sequence targeting letters A to G with their associated probabilities. reference sequence and output sequence. Subsequently, we apply a prediction model on the learned encoded representation to output the probability of obtaining the outcome. The network architecture to learn \(f\) is reported in figure 3 (B: proportion model). However, there is a relatively higher probability often associated with the wild-type outcome (\(\mathbf{x}_{\text{out},i}=\mathbf{x}_{\text{ref}}\)), while the probabilities linked to the edited outcome sequences are often very small. This situation presents a challenge when directly modeling \(P(X_{\text{out}}|x_{\text{ref}})\)--as the model might easily learn the wild-type probability but struggle with outcomes that have extremely low probabilities. ### Two-stage model To address this, we propose a two-stage model where we break down \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}})\) as the product of two probabilities: \[P(\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}})=\begin{cases}P(\mathbf{x} _{\text{out},i}|\mathbf{x}_{\text{ref}},\text{edited})P(\text{edited}|\mathbf{ x}_{\text{ref}}),\\ \text{if }\mathbf{x}_{\text{out},i}\neq\mathbf{x}_{\text{ref}}\\ 1-P(\text{edited}|\mathbf{x}_{\text{ref}}),\text{if }\mathbf{x}_{\text{out},i}= \mathbf{x}_{\text{ref}}\end{cases} \tag{1}\] For a given reference sequence, we first predict the probability of overall efficiency which is defined in Eq. 2. It provides the probability of the target sequence being edited, \(P(edited|\mathbf{x}_{\text{ref}})\), which in turn gives the probability of the wild-type. Next, we predict the probability of all possible edited outcomes, \(P(\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}},edited)\). We refer to the first as _overall efficiency_ and the second as _proportion_ \[P(edited|\mathbf{x}_{\text{ref}})=\frac{\text{Sum of the read count of all edited reads for the target}}{\text{Total read count of the target sequence}} \tag{2}\] We estimate the overall efficiency of the given reference sequence using \(f_{\theta_{1}}(\mathbf{x}_{\text{ref}})\) (Eq. 3), denoted by the overall efficiency model, and subsequently, we predict the conditional probabilities of all non wild-type outcomes using \(f_{\theta_{2}}(\mathbf{x}_{\text{ref}},\mathbf{x}_{\text{out},i})\) (Eq. 4) which we denote by the proportion model. \[f_{\theta_{1}}(\mathbf{x}_{\text{ref}})=P(edited|\mathbf{x}_{\text{ref}}) \tag{3}\] where \(P(\text{\emph{wild-type}}|\mathbf{x}_{\text{ref}})=1-P(edited|\mathbf{x}_{ \text{ref}})\) \[f_{\theta_{2}}(\mathbf{x}_{\text{ref}},\mathbf{x}_{\text{out},i})=P(\mathbf{ x}_{out,i}|\mathbf{x}_{\text{ref}},\text{\emph{edited}}), \tag{4}\] where \(\mathbf{x}_{\text{out},i}\neq\mathbf{x}_{\text{ref}}\) Once \(f_{\theta_{1}}\) and \(f_{\theta_{2}}\) are learned, we can calculate \(P(X=\mathbf{x}_{\text{out},i}|\mathbf{x}_{\text{ref}})\) where \(i=1,\ldots M\) for all outcome sequences, including wild-type and edited sequences using Eq 1. #### 3.3.1 Overall efficiency model We formulate the overall efficiency model as a probabilistic classification task where \(f_{\theta_{1}}\) parameterizes a binomial distribution \(P(C|\mathbf{x}_{\text{ref}})\) of a random variable \(C\in\{\textit{edited},\textit{not edited}\}\) with the aim to learn to output the \(P(C=edited|\mathbf{x}_{\text{ref}})\) for a given reference sequence. To learn \(f_{\theta_{1}}\), we first computed the overall editing efficiency for each reference sequence by summing all probabilities attributed to the non wild-type outcomes as given in Eq 2, or equivalently, \(1-P(\textit{wild-type}|\mathbf{x}_{ref})\). Then we use multiple 1D-Convolutional layers (LeCun et al., 1995) on the one-hot-encoded representation of \(\mathbf{x}_{ref}\) to learn discriminative feature embedding that is passed to the multi-layer perceptron (MLP) layer to approximate the distribution \(P(C|\mathbf{x}_{\text{ref}})\). The model architecture is presented in Figure 3 (A). We trained \(f_{\theta_{1}}\) using KL-divergence loss that is applied on the true distribution \(P(C|\mathbf{x}_{\text{ref}})\) and learned distribution \(\hat{P}_{\theta_{1}}(C|\mathbf{x}_{\text{ref}})\) for each reference sequence. \[\mathcal{L}_{\textit{efficiency}}(\theta_{1},D)=\sum_{i=1}^{N}D_{kl}(P(C| \mathbf{x}_{\text{ref}}^{i})\|\hat{P}_{\theta_{1}}(C|\mathbf{x}_{\text{ref}}^ {i})) \tag{5}\] #### 3.3.2 Proportion model This model is designed to approximate the conditional distribution \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}},\textit{edited})\). To achieve this, we first remove the wild-type from each reference sequence's corresponding output \(X_{\text{out}}\). Then, we normalize the probabil Figure 3: Two-stage Model overview ities of the remaining outcomes to ensure a valid distribution effectively converting \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}})\) into the distribution \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}},\textit{edited})\). The proportion model \(f_{\theta_{2}}\) is designed to learn the parameters governing the distribution \(P(X_{\text{out}}|\mathbf{x}_{\text{ref}},\textit{edited})\). Similar to the one-stage model, \(f_{\theta_{2}}\) is provided with both the reference sequence \(\mathbf{x}_{\text{ref}}\) and its associated outcome sequence \(\mathbf{x}_{\text{out},i}\). The model is then trained to estimate the likelihood \(P(\mathbf{x}_{\text{out},i}\mid\mathbf{x}_{\text{ref}},\textit{edited})\), representing the probability of reference sequence being edited, and result in the outcome sequence \(\mathbf{x}_{\text{out},i}\). As illustrated in Figure 3 (B), \(f_{\theta_{2}}\) uses attention-based model comprised of two encoder networks, \(\text{Enc}^{1}(\mathbf{x}_{\text{ref}})\), \(\text{Enc}^{2}(\mathbf{x}_{\text{out}})\), and one output network \(g\). The design of the encoder networks adapts the transformer encoder blocks architecture (Vaswani et al., 2017), characterized by multiple layers of multi-head self-attention modules. The two encoder networks process the reference sequence and one of its corresponding output sequence \(\mathbf{x}_{\text{out},i}\), leading to the extraction of their respective latent representations, namely \(\mathbf{Z}_{\text{ref}}\in\mathbb{R}^{T\times d}\) and \(\mathbf{Z}_{\text{out}}\in\mathbb{R}^{T\times d}\). Both vectors are then concatenated to form a unified learned representation \(\mathbf{Z}\in\mathbb{R}^{T\times 2d}\). Subsequently, the output network \(g\) embeds this unified representation \(\mathbf{Z}\) to compute the probability of obtaining the output sequence given the reference sequence, \(P(\mathbf{x}_{\text{out},i}\mid\mathbf{x}_{\text{ref}},\textit{edited})\). Precisely, the output network \(g(\mathbf{Z})\) takes as input the final representation \(\mathbf{Z}\in\mathbb{R}^{T\times 2d}\) and performs an affine transformation followed by softmax operation to compute the probability of conversion of every target base (i.e. base A or C depending on the chosen base editor) as it is shown below: \[\hat{y}_{it}=\sigma(\mathbf{W}\mathbf{z}_{it}+\mathbf{b}_{t}) \tag{6}\] where \(\mathbf{W}\in\mathbb{R}^{2\times 2d}\), \(\mathbf{b}_{t}\in\mathbb{R}^{2}\) and \(\sigma\) is softmax function. \(\hat{y}_{it}\) represents the probability of editing occurring at the \(t\)-th position in the \(i\)-th outcome sequence. The un-normalized probability for the whole \(i\)-th output sequence \(\mathbf{x}_{\text{out},i}\) given its reference sequence is computed by \(\hat{y}_{i}=\prod_{t=1}^{T}\hat{y}_{i,t}\), which is then normalized across all the outcomes to make it valid probability distribution (Eq. 7). Therefore, the approximated probability for obtaining \(i\)-th edited (non-wild type) outcome sequence is given by: \[\hat{P}(\mathbf{x}_{\text{out},i}\mid\mathbf{x}_{\text{ref}},\textit{edited} )=\frac{\hat{y}_{i}}{\sum_{i=1}^{M}\hat{y}_{i}} \tag{7}\] Objective FunctionWe used the Kullback-Leibler (KL) divergence on the model's estimated distribution over all outcome sequences for a given reference sequence \(\mathbf{x}_{\text{ref}}^{i}\) and the actual distribution: \[D_{\text{KL}}^{i}(P(X_{\text{out}}|\mathbf{x}_{\text{ref}}^{i}, \textit{edited})||\hat{P}_{\theta_{2}}(X_{\text{out}}|\mathbf{x}_{\text{ref}}^ {i},\textit{edited})) \tag{8}\] \[=\sum_{j=1}^{M_{i}}P(\mathbf{x}_{\text{out},j}|\mathbf{x}_{\text{ ref}}^{i},\textit{edited})\log\frac{P(\mathbf{x}_{\text{out},j}|\mathbf{x}_{ \text{ref}}^{i},\textit{edited})}{\hat{P}_{\theta_{2}}(\mathbf{x}_{\text{out},j }|\mathbf{x}_{\text{ref}}^{i},\textit{edited})}\] Lastly, the objective function for the whole training set is defined by the average loss across all the reference sequences as follows: \[\mathcal{L}_{\text{proportion}}(\theta_{\mathbf{2}};D)= \tag{9}\] \[\sum_{i=1}^{N}D_{KL}^{i}(P(X_{\text{out}}\mid\mathbf{x}_{\text{ ref}}^{i},\textit{edited})||\hat{P}_{\theta_{2}}(X_{\text{out}}\mid\mathbf{x}_{ \text{ref}}^{i},\textit{edited})\] The final objective is composed of both the overall efficiency model loss and the proportion model loss with a weight regularization term (i.e. \(l_{2}\)-norm regularization) applied to the model parameters represented by \(\theta=\{\theta_{\mathbf{1}},\theta_{\mathbf{2}}\}\) (Eq. 10) \[\mathcal{L}_{\text{proportion}}(\theta_{\mathbf{1}};D)+\mathcal{L}_{\text{ efficiency}}(\theta_{\mathbf{2}},D)+\frac{\lambda}{2}\|\theta\|_{2}^{2} \tag{10}\] ### Multi-task learning with multiple base editors Multi-task learningThere is a diverse set of base editors, each distinguished by its unique design attributes. These distinctions, including variations in binding affinities, editing window sizes, and deaminase activities, result in differing editing efficiencies even when targeting the same sequence. The variations in editing efficiency across different editors emphasize the complexity of the base editing landscape. Conventional approaches have often proposed training separate models for each editor. However, this approach not only demands additional effort but also fails to leverage the shared structural similarities among editors. To leverage common patterns and relationships present across various libraries derived from different base editors, and optimize predictive capability while also reducing computational time, we propose a more efficient solution based on multi-task learning. Instead of training separate models for each editor, we train a single model capable of predicting the efficiency of all editors when applied to the same reference sequence. Given a total number of \(D\) editors where each editor has its own dataset \(B_{i}\) (denoted by screening libraries), we developed a multi-task learning model that uses shared encoding Figure 4: Multi-task learning model overview layers to extract a common representation across all the libraries as well as individual branches that fine-tune the model specifically for each library, ensuring a better fit to their respective data. This approach implicitly models \(P(X_{\text{out}}\mid\mathbf{x}_{\text{ref}},B_{i})\) where \(B_{i}\) represents the base editor type applied on the reference sequence. To implement one universal model across all datasets, we extend our proposed two-stage model architecture (illustrated in Figure 3) for multi-task learning, as depicted in Figure 4. Specifically, we modify the overall efficiency model by initially employing two convolutional layers as a shared building block across all datasets/editors, enabling the learning of a common representation for the reference sequence. Then a set of output blcoks is used to represent editor-specific transformation. Each editor type has its own output block consisting of two layers of convolutional network followed by MLP layers to predict the probability \(P(edited|\mathbf{x}_{\text{ref}})\) for each editor/dataset accordingly. We adapt the proportion model by using a common encoder network across all editors/datasets to establish a unified representation \(\mathbf{Z}_{\text{ref}}\) for the reference sequence while using separate encoders and output blocks for each distinct editor. To counterbalance any bias towards larger datasets, we implemented a data loader that uniformly samples the same number of data samples in each mini-batch throughout the training phase. ## 4 Experiments ### Dataset and Experiment setup DatasetTo comprehensively assess base editors' efficiency across thousands of genomic sequences, we conducted high-throughput screening, resulting in the creation of six distinct datasets. Each dataset corresponds to the application of one of the following base editors: SpRY-ABESe, SpCas9-ABESe SpG-ABE8e, SpRY-ABEmax, SpCas9-ABEmax, and SpG-ABEmax, as listed in Table 1. Detailed descriptions of the used editors are provided in Appendix Section 6.2. Each dataset encompasses approximately 11,000 reference sequences and their corresponding output sequences. In each dataset, we leveraged 193 distinct PAM sites, each comprising four nucleotide bases. Experiment setupWe divided every dataset into training, testing, and validation sets, maintaining a ratio of 80%, 10%, and 10%. This procedure is repeated three times to ensure robust performance reporting. All reported results are based on the average performance over the three runs(indicated by \(mean\pm std\)). First, we use the one-stage model to identify the best features to represent reference sequence for predicting the base editing outcomes (i.e. determine reference sequence representation option as explained in section 3.2). Using the selected features (i.e., protospacer + PAM), we proceed to evaluate the performance between the one-stage and two-stage models. Finally, using the two-stage model, we compare the multi-task learning (i.e. a unified model training for all editors) to the single-task learning setup where separate models are trained for the different editors. Throughout model training, we track the epoch at which the best validation scores are attained. Evaluation of the trained models for each base editor is based on their average performance on the test sets across the three runs. Pearson and Spearman correlation were used as performance measures for all tested models. More details about the network structure, optimization, and hyperparameters are presented in the Appendix Section 6.3. ### Experiment results Reference sequence representationExisting models have explored different factors that could affect the base editor's efficiency, which we categorize into three scenarios: 1) the protospacer, 2) the protospacer along with its PAM, and 3) an extended range including left overhangs, protospacer, PAM, and right overhangs. We investigate all three scenarios with the one-stage model to identify the best features to represent the reference sequence. As shown in Table 2, we observe that incorporating PAM information significantly enhances performance, whereas the inclusion of overhangs demonstrates minimal impact. Besides, adding overhangs increases the computational complexity drastically. Consequently, we opt to employ protospacer and PAM information to represent reference sequences in all the subsequent model results presented below. Comparing One-stage with Two-stage ModelAs detailed in Section 3.2, our model can be conceptualized as either a one-stage model, directly capturing the distribution across all potential outcomes for a given reference, or as a two-stage model. The latter approach involves initially \begin{table} \begin{tabular}{l c c c c c} \hline Editor & \#ins & \#refseq & \#outcome & mean & std \\ \hline SpRY-ABESe & 110141 & 11291 & 9.7 & 0.102 & 0.211 \\ SpCas9-ABESe & 43054 & 11337 & 4.6 & 0.217 & 0.323 \\ SpG-ABESe & 80873 & 11307 & 7.1 & 0.139 & 0.263 \\ SpRY-ABEmax & 70851 & 11347 & 6.2 & 0.159 & 0.301 \\ SpCas9-ABEmax & 39606 & 11302 & 3.5 & 0.285 & 0.417 \\ SpG-ABEmax & 70851 & 11347 & 6.2 & 0.159 & 0.301 \\ \hline \end{tabular} \end{table} Table 1: Data statistics: “#ins” refers to the number of reference and output sequence pairs, “#refseq” denotes the number of distinct reference sequences, “#outcome” denotes the average number of outcomes per reference sequence, the mean and std refers to the mean and standard deviation of the probability across all the outcomes. predicting the probability of an edit occurring in the reference sequence, followed by predicting the probabilities of individual edited outcomes. In this section, we present results for both models to illustrate the advantages of the two-stage approach over the one-stage counterpart. For the one-stage model, we use exactly the same architecture as the proportion model from the two-stage model on the original data without preprocessing where we remove the wild type and renormalize the probability for each reference. Table 3 shows the two-stage model has slightly superior results (Spearman correlation) over the one-stage model. This improvement can be attributed to the model's two-step prediction approach, which first predicts the wild-type alone and subsequently refines predictions for various edited outcomes. To better understand the difference between the two models' ability to predict the wild-type and edited outcome sequences, we rigorously evaluated each model's performance separately on both types of outcome. The two-stage model outperforms the one-stage model in most of the datasets when considering both wild type and edited outcomes as presented in Table 4 and 5. Multi-task learningGiven this conclusion, we proceed with the multi-task learning with the two-stage mode (see Figure 4). We compared the performance of multi-task learning across all the datasets/editors with a single-task setup where we trained one model per dataset/editor. Table 6 reports similar performance for both models. Although there wasn't a substantial performance difference, adopting a unified multi-task model offers advantages such as reduced run-time (for training and inference) and smaller model size (fewer parameters) while maintaining consistent performance across all datasets. Moreover, with a unified model, we can simultaneously predict the editing outcomes of all six editors at once for a given target sequence. Comparing to baselines in the literatureWe also compared our model with BE-DICT (Marquart et al., 2021) which is one of the relevant existing models that tackle base editing outcome prediction. BE-DICT is a sequence-to-sequence model where the decoding happens in an auto-regressive manner, it is computationally heavy compared to our proposed method. Moreover, it is trained as single-task model (i.e. one model for each editor) and use only the protospacer to represent the target sequence. We extended and retrained BE-DICT on two of the datasets (randomly chosen) and compared the prediction results with ours. For a fair comparison, we first used our one-stage model trained in the single task setting (one model per dataset) using only the protospacer. The results of this experiment reveal the advantages of the architectural changes, particu \begin{table} \begin{tabular}{l c c c c} \hline \hline \multicolumn{2}{c}{Single task learning} & \multicolumn{2}{c}{Multi task learning} \\ Libraries & Spearman & Pearson & Spearman & Pearson \\ \hline SpKP-ABEmax & \(0.872\pm 0.001\) & \(0.986\pm 0.001\) & \(0.872\pm 0.002\) & \(0.986\pm 0.0002\) \\ SpCap-ABEmax & \(0.879\pm 0.004\) & \(0.989\pm 0.001\) & \(0.864\pm 0.0019\) & \(0.992\pm 0.0001\) \\ SpCap-ABEmax & \(0.882\pm 0.001\) & \(0.991\pm 0.0006\) & \(0.889\pm 0.0016\) & \(0.992\pm 0.0004\) \\ SpKP-ABEbe & \(0.861\pm 0.0029\) & \(0.974\pm 0.001\) & \(0.863\pm 0.0011\) & \(0.975\pm 0.001\) \\ SpCap-ABEbe & \(0.856\pm 0.008\) & \(0.938\pm 0.0005\) & \(0.852\pm 0.002\) & \(0.937\pm 0.003\) \\ SpG-ABEbe & \(0.856\pm 0.004\) & \(0.980\pm 0.0008\) & \(0.871\pm 0.003\) & \(0.979\pm 0.001\) \\ \hline \hline \end{tabular} \end{table} Table 6: Performance comparison between the multi-task and single task learning models \begin{table} \begin{tabular}{l c c c c c} \hline \hline \multicolumn{2}{c}{} & \multicolumn{2}{c}{One-stage Model} & \multicolumn{2}{c}{Two-stage Model} \\ Libraries & Spearman & Pearson & Spearman & Pearson \\ \hline SpKP-ABEmax & \(0.745\pm 0.015\) & \(0.711\pm 0.011\) & \(0.798\pm 0.007\) & \(0.872\pm 0.012\) \\ SpCap-ABEmax & \(0.82\pm 0.003\) & \(0.851\pm 0.014\) & \(0.838\pm 0.009\) & \(0.890\pm 0.030\) \\ SpG-ABEmax & \(0.807\pm 0.003\) & \(0.752\pm 0.014\) & \(0.845\pm 0.011\) & \(0.822\pm 0.014\) \\ SpKP-ABEbe & \(0.393\pm 0.021\) & \(0.508\pm 0.025\) & \(0.547\pm 0.056\) & \(0.699\pm 0.051\) \\ SpCap-ABEbe & \(0.855\pm 0.007\) & \(0.840\pm 0.003\) & \(0.806\pm 0.002\) & \(0.858\pm 0.0021\) \\ SpCap-ABEbe & \(0.712\pm 0.002\) & \(0.732\pm 0.004\) & \(0.774\pm 0.005\) & \(0.810\pm 0.009\) \\ \hline \hline \end{tabular} \end{table} Table 3: Prediction performance all outcomes (i.e. including wild-type sequences). \begin{table} \begin{tabular}{l c c c c} \hline \hline \multicolumn{2}{c}{} & \multicolumn{2}{c}{Protospacer \& PAM} & \multicolumn{2}{c}{Protospacer \& PAM \& Overhang} \\ Libraries & Spearman & Pearson & Spearman & Pearson & Pearson \\ \hline SpKP-ABEmax & \(0.835\pm 0.007\) & \(0.981\pm 0.001\) & \(0.854\pm 0.006\) & \(0.983\pm 0.001\) & \(0.854\pm 0.003\) & \(0.983\pm 0.002\) \\ SpCap89-ABEmax & \(0.786\pm 0.003\) & \(0.978\pm 0.002\) & \(0.881\pm 0.001\) & \(0.989\pm 0.0005\) & \(0.891\pm 0.002\) & \(0.989\pm 0.001\) \\ SpG-ABEmax & \(0.841\pm 0.002\) & \(0.985\pm 0.0007\) & \(0.866\pm 0.004\) & \(0.989\pm 0.0003\) & \(0.878\pm 0.008\) & \(0.991\pm 0.0009\) \\ SpRY-ABEbe & \(0.776\pm 0.019\) & \(0.965\pm 0.001\) & \(0.779\pm 0.0036\) & \(0.968\pm 0.002\) & \(0.803\pm 0.008\) & \(0.967\pm 0.0003\) \\ SpCas9-ABEbe & \(0.762\pm 0.007\) & \(0.883\pm 0.005\) & \(0.857\pm 0.007\) & \(0.945\pm 0.0006\) & \(0.862\pm 0.003\) & \(0.945\pm 0.003\) \\ SpG-ABEbe & \(0.803\pm 0.005\) & \(0.963\pm 0.002\) & \(0.820\pm 0.005\) & \(0.974\pm 0.0009\) & \(0.819\pm 0.006\) & \(0.9771\pm 0.0008\) \\ \hline \hline \end{tabular} \end{table} Table 2: Pearson and Spearman correlation using one-stage Model across the three different reference sequence representations. In our experiment, we chose 5 neighboring nucleotides for both sides to represent the overhangs. \begin{table} \begin{tabular}{l c c c c} \hline \hline \multicolumn{2}{c}{} & \multicolumn{2}{c}{One-stage Model} & \multicolumn{2}{c}{Two-stage Model} \\ Libraries & Spearman & Pearson & Spearman & Pearson \\ \hline SpKP-ABEmax & \(0.851\pm 0.006\) & \(0.983\pm 0.001\) & \(0.873\pm 0.001\) & \(0.986\pm 0.001\) \\ SpCap-ABEmax & \(0.881\pm 0.006\) & \(0.989\pm 0.0005\) & \(0.879\pm 0.004\) & \(0.991\pm 0.001\) \\ SpG-ABEmax & \(0.865\pm 0.004\) & \(0.989\pm 0.0003\) & \(0.887\pm 0.003\) & \(0.991\pm 0.0006\) \\ SpK-ABEbe & \(0.779\pm 0.003\) & \(0.968\pm 0.002\) & \(0.862\pm 0.003\) & \(0.974\pm 0.001\) \\ SpCap-ABEbe & \(0.857\pm 0.007\) & \(0.945\pm 0.0006\) & \(0.856\pm 0.003\) & \(0.597\pm 0.002\) \\ SpG-ABEbe & \(0.820\pm 0.005\) & \(0.974\pm 0.0009\) & \(0.865\pm 0.004\) & \(0.978\pm 0.0008\) \\ \hline \hline \end{tabular} \end{table} Table 3: Prediction performance all outcomes (i.e. including wild-type sequences). \begin{table} \begin{tabular}{l c c c} \hline \hline \multicolumn{2}{c}{} & \multicolumn{ larly in adopting an encoder-encoder architecture over the traditional sequence-to-sequence (encoder-decoder) model. Then we extended their model to use both the protospacer and PAM as the reference sequence and compared it with our proposed multi-task model (trained on eight data sets using protospacer and PAM information as target sequence). Results in Table 7 show that our model consistently outperforms BE-DICT. Furthermore, considering computational efficiency during model training (on SpRY-ABE8e data using the protospacer and PAM reference sequence representation), BE-DICT takes in the order of minute per epoch (wall time), while our single-task model accomplishes the same task in the order of seconds ( 15 seconds). Notably, the multi-task learning model trained jointly on all six datasets takes 21 seconds per epoch. This highlights the benefits of avoiding the complex sequence-to-sequence architecture in favor of a streamlined encoder-encoder structure. This choice not only improves the computational efficiency but also leads to model predictive performance improvements. Moreover, the introduction of a two-stage model and a multi-task framework amplifies these performance gains even further. We present additional results for comparisons with other baselines in Table 8 in the appendix. To assess our model's performance against other state-of-the-art models, we conducted evaluations using the test sets provided by these models. Table 8 displays our findings, which include three most recent models: BE-HIVE (Arbab et al., 2020), DeepABE (Song et al., 2020), and BEDICT (Marquart et al., 2021), along with their respective test sets labeled as A. et al., S. et al., and M. et al. The idea is to take the published trained model and evaluate their performance on those various test sets. For the three baseline models, we refer to the results reported in the BEDICT paper. As for our model, to ensure fairness in comparison, we used our single-stage model trained on SpG-ABEmax libraries since most baselines, except DeepABE, do not incorporate the PAM as input. The results correspond to two scenarios: 1) considering all possible outcomes, and 2) only considering non-wild type outcomes. The results for the non-wild type outcomes correspond to the model prediction where we only consider non-wild outcomes. In the case of non-wild-type outcome prediction, we mention that other models were trained exclusively on non-wild outcomes, with outcomes per sequence being renormalized. Our one-stage model, however, was trained on data encompassing all outcomes, so we report non-wild-type results with outcomes renormalized for a fair comparison. ## 5 Conclusion Our work provides a detailed assessment of the modeling approaches for base editor outcome prediction. Through the development of a unified model, we transcend the limitations of single-editor models and pave the way for more versatile and comprehensive tools. By combining self-attention mechanisms and multi-task learning, we capture the nuances of editing outcomes across various editors, enhancing the accuracy and applicability of our predictions. As the first machine learning-focused paper in the domain of base editor outcome prediction, our work represents a stepping stone toward a more systematic and informed modeling approach to genome editing. We explored the different modeling decisions from one-stage to two-stage models, and from single-task to multi-task learning. We evaluated the different sequence representations and benchmarked our best model with one of the main models developed for base editing outcome prediction. We believe that further work studying systematically the different modeling decisions for genome editing will help guide researchers toward more promising editing strategies that in turn will bring advancements in gene therapy and disease modeling. For the future, given the current absence of standardized and systematic benchmark datasets in the field, we aim to bridge this gap by creating standard benchmark datasets, establishing baseline models, and proposing better performance metrics. This initiative will provide the machine-learning community with a solid foundation for testing a wide array of innovative ideas and approaches. ## Acknowledgements We thank G. Schwank, K. Marquart, L. Kissling and S. Janjuha for input on the CRISPR-Cas and base editing technology and for data sharing and preprocessing. This work was supported by the URPP 'Human Reproduction Reloaded' and 'University Research Priority Programs'. \begin{table} \begin{tabular}{l l l l l l} \hline \hline \multicolumn{2}{c}{} & \multicolumn{3}{c}{BE-DICT} & \multicolumn{3}{c}{Ours} \\ reference sequence & Libraries & Spearman & Person & Spearman & Person \\ \hline protospacer & SpRY-ABE8e & 0.801 & 0.943 & 0.835 & 0.981 \\ & SpRY-ABE8e & 0.746 & 0.861 & 0.776 & 0.965 \\ & SpRY-ABE8e & 0.804 & 0.951 & 0.870 & 0.987 \\ & SpRY-ABE8e & 0.762 & 0.850 & 0.860 & 0.975 \\ \hline \hline \end{tabular} \end{table} Table 7: Performance comparison with the baselines \begin{table} \begin{tabular}{l c c c c c c} \hline \hline & \multicolumn{3}{c}{All Outcomes} & \multicolumn{3}{c}{Non wild-types} \\ Datasets & A.et al & S. et al & M. et al & A.et al & S. et al & M. et al \\ \hline BEDICT & 0.96 & 0.94 & 0.86 & 0.81 & 0.90 & 0.82 \\ DeepABE & 0.86 & 0.93 & 0.8 & 0.86 & 0.96 & 0.84 \\ BE-HIVE & 0.71 & 0.88 & 0.74 & 0.92 & 0.93 & 0.81 \\ Our model & 0.972 & 0.974 & 0.972 & 0.939 & 0.945 & 0.953 \\ \hline \hline \end{tabular} \end{table} Table 8: Model performance on the test set from the different published studies. Columns represent test sets, rows represent models used