Datasets:

liyucheng

/

arxiv-march-2023

like 0

http://arxiv.org/abs/2303.06806v1

20230313012855

Neural Diarization with Non-autoregressive Intermediate Attractors

eess.AS

Boosting Source Code Learning with Data Augmentation: An Empirical Study Jianjun Zhao ======================================================================== End-to-end neural diarization (EEND) with encoder-decoder-based attractors (EDA) is a promising method to handle the whole speaker diarization problem simultaneously with a single neural network. While the EEND model can produce all frame-level speaker labels simultaneously, it disregards output label dependency. In this work, we propose a novel EEND model that introduces the label dependency between frames. The proposed method generates non-autoregressive intermediate attractors to produce speaker labels at the lower layers and conditions the subsequent layers with these labels. While the proposed model works in a non-autoregressive manner, the speaker labels are refined by referring to the whole sequence of intermediate labels. The experiments with the two-speaker CALLHOME dataset show that the intermediate labels with the proposed non-autoregressive intermediate attractors boost the diarization performance. The proposed method with the deeper network benefits more from the intermediate labels, resulting in better performance and training throughput than EEND-EDA. end-to-end neural diarization, intermediate objective, self-conditioning, attractor, speaker diarization § INTRODUCTION Speaker diarization is the task of detecting multi-speaker speech activity in audio recordings. It has been actively studied as an essential component for conversational speech understanding <cit.>. The task has been evaluated in telephone conversations (CALLHOME <cit.>), meetings (ICSI <cit.>, AMI <cit.>), web videos (VoxConverse <cit.>) and various hard scenarios (DIHARD Challenges <cit.>). A standard approach to speaker diarization is speaker embedding clustering <cit.>, which first extracts speaker-discriminative embeddings like x-vectors <cit.> and d-vectors <cit.> for fixed-length speech segments, and then merges homogeneous segments to the same speaker by applying a clustering algorithm such as spectral clustering <cit.> and agglomerative hierarchical clustering <cit.>. For speaker embedding clustering, voice activity detection should be done in advance to determine speech/non-speech boundaries, and overlapping speech segments have to be eliminated with some pre- or post-processing methods<cit.>. As an alternative to speaker embedding clustering, end-to-end neural diarization (EEND) <cit.> has been proposed. EEND learns a neural network to directly produce full speaker diarization labels containing speech/non-speech boundaries and overlapping speech segments, as well as speaker assignments for the detected speech segments. Since the speaker diarization labels have permutation ambiguity due to the arbitrary order of speaker indices, the EEND network is trained with permutation-invariant training objectives. Various network architectures have been investigated for EEND including Transformer <cit.>, time-dilated convolutional neural network <cit.>, and Conformer <cit.>. Encoder-decoder-based attractor (EDA) <cit.> is a promising network architecture for EEND, which first generates an attractor vector for each speaker from an embedding sequence, and then generates the speaker's activity by measuring the similarity between the embedding sequence and the attractor vector. Although the EEND-EDA model is found to be effective compared with speaker embedding clustering, there are two points of improvement. First, the model assumes conditional independence between frame-level speaker labels, which limits their performance. The EEND-EDA model runs in a non-autoregressive manner and produces all frame-level speaker labels in parallel. However, such non-autoregressive models disregard the potential benefit of the output label dependency between frames. Second, the long-short term memory (LSTM)-based encoder in EDA receives the frame-level embeddings recursively. The well-known vanishing gradient problem of LSTM hinders the optimization of the lower layers. In related fields, researchers have studied the use of “intermediate” labels to relax the conditional independence assumption in such non-autoregressive models. For non-autoregressive automatic speech recognition (ASR) based on connectionist temporal classification (CTC), Intermediate CTC <cit.> is proposed to introduce auxiliary tasks of predicting labels inside the network by inserting the same CTC losses to the intermediate predictions. Self-conditioned CTC <cit.> further utilizes the intermediate labels at the lower layer as conditions for enhancing the predictions at the upper layers. This self-conditioning technique achieves the best performance among the non-autoregressive ASR systems <cit.>. For EEND, a similar intermediate speaker label prediction technique <cit.> is proposed, which uses the same permutation-invariant training objectives for the intermediate speaker labels. However, the intermediate speaker labels are not utilized for the self-conditioning features. These prior studies motivate us to introduce the self-conditioning technique in EEND models. In this paper, we propose a novel network architecture for EEND that uses intermediate speaker labels to condition the subsequent network layers. For producing the intermediate speaker labels, the proposed method extracts intermediate attractors at every encoder layer. The auxiliary permutation-invariant training losses are introduced for optimizing the intermediate labels. For conditioning the subsequent network layers, the proposed method adds the weighted intermediate attractors to the frame-level embeddings. While the proposed network still works in a non-autoregressive manner, the speaker labels are iteratively refined by referring to the whole sequence of intermediate speaker labels. For the vanishing gradient problem in the EDA's LSTM, we adopt the attention mechanism <cit.>. Unlike LSTM, it does not suffer from vanishing gradients, thus facilitating optimization of the lower layers. Another advantage of the attention mechanism is that it is non-autoregressive, so training throughput is much higher than LSTM when interacting with the intermediate training objectives. The experimental results with the two-speaker CALLHOME dataset show that the intermediate labels with the proposed non-autoregressive intermediate attractors boost the diarization performance while the original EDA cannot get benefit from the intermediate labels. The proposed method with the deeper network benefits more from the intermediate labels, resulting in better performance and training throughput than EEND-EDA. § METHOD This section briefly introduces EEND-EDA as our baseline system, followed by our proposed method: intermediate attractors. §.§ End-to-end Neural Diarization EEND formulates the speaker diarization problem as a frame-wise multi-label classification task <cit.>. In this paper, we denote X ∈ℝ^D× T as a T-length sequence of D-dimensional audio features. A neural network accepts X and produces the same-length sequence of speaker label posteriors Y ∈ [0,1]^C× T, where C is the number of speakers and [Y]_c,t is the probability that c-th speaker is speaking at time t. The network is trained to minimize the binary cross-entropy loss between the ground-truth speaker labels Y^* ∈{0,1}^C× T and the estimated label posteriors Y: ℒ_𝖯𝖨𝖳(Y^*, Y) = 1/CTmin_ϕ∈𝒫(C)∑_c=1^C ∑_t=1^T 𝖡𝖢𝖤([Y^*]_ϕ_c,t, [Y]_c,t), where 𝖡𝖢𝖤(y^*,y) = - y^*log y - (1-y^*)log(1-y), 𝒫(C) is the set of all permutations of a sequence {1,…,C}. This permutation-invariant training scheme <cit.> correctly handles the label ambiguity caused by the arbitrary order of speaker indices. §.§ EEND with Encoder-decoder-based Attractors The neural network for EEND adopted in <cit.> comprises of a stack of Transformer encoders: E_l = 𝖤𝗇𝖼𝗈𝖽𝖾𝗋𝖫𝖺𝗒𝖾𝗋_l(E_l-1) ∈ℝ^D× T (1 ≤ l ≤ L), where L is the number of Transformer layers and E_0 = X [In this paper, we assume X is subsampled beforehand.]. Whereas the vanilla EEND <cit.> simply transforms E_L to Y by a linear layer followed by a sigmoid function, EDA <cit.> first generates speaker-wise attractor vectors A = [a_1, …, a_C] ∈ℝ^D× C: A = 𝖤𝖣𝖠(E_L). Here, this EDA function is implemented using long short-term memory (LSTM) layers: (h_t, c_t) = 𝖫𝖲𝖳𝖬_𝖾𝗇𝖼(h_t-1, c_t-1, [E_L]_:,t) (1 ≤ t ≤ T), (a_c, d_c) = 𝖫𝖲𝖳𝖬_𝖽𝖾𝖼(a_c-1, d_c-1, 0) (1 ≤ c ≤ C), where 𝖫𝖲𝖳𝖬_𝖾𝗇𝖼() is an unidirectional LSTM layer that sequentially reads an embedding vector for time t, [E_L]_:,t is the embedding vector from the t-th column of E_L, h_t ∈ℝ^D is a hidden state, c_t ∈ℝ^D is a cell state. 𝖫𝖲𝖳𝖬_𝖽𝖾𝖼() is another unidirectional LSTM layer with the initial hidden state a_0 = h_T and the initial cell state d_0 = c_T. The decoder LSTM receives zero vector C times to produce the speaker-wise attractor vector a_c for C speakers. EEND-EDA estimates the speaker label by comparing the embedding sequence E_L with the speaker-wise attractors A: Y = 𝖲𝗂𝗀𝗆𝗈𝗂𝖽(A^⊤E_L). As the number of attractor vectors can vary with the number of iterations C in Eq. <ref>, EDA can handle an unknown number of speakers by jointly estimating the number of iterations. In this paper, we fix the number of iterations to two since we only evaluate the method on two-speaker conversations. With EEND-EDA described above, the speaker labels at all frames are estimated in parallel. This parallel estimator lacks a mechanism handling the label dependency between frames. In the next subsection, we consider the label dependency between frames by using intermediate speaker labels. §.§ Intermediate Attractors Considering the label dependency between frames, we extract “intermediate” speaker labels and feed them into the subsequent network layers. Because the upper layer can refer to the whole sequence of intermediate speaker labels from the lower layer, conditional independence between frames is relaxed. The overview of the proposed method with the four-layer EEND-EDA model is depicted in Fig. <ref>. For producing the intermediate speaker labels, intermediate attractors are generated with EDA. Using the same EDA components in Eq. <ref>, intermediate attractors for the l-th layer are calculated as follows: A_l = 𝖤𝖣𝖠(E_l). Then, intermediate speaker labels Y_l for each layer are estimated using the intermediate attractors A_l, similar to Eq. <ref>: Y_l = 𝖲𝗂𝗀𝗆𝗈𝗂𝖽(A_l^⊤E_l) (1 ≤ l ≤ L-1) The auxiliary permutation-invariant training loss is introduced for the intermediate speaker labels, and the model is trained with the summation of the original loss (Eq. <ref>) and the intermediate losses [We can take the weighted loss by introducing a mixing ratio. However, we ignore the hyperparameter in this work.]: ℒ_𝗂𝗇𝗍𝖾𝗋 = ℒ_𝖯𝖨𝖳(Y^*, Y) + 1/L-1∑_l=1^L-1ℒ_𝖯𝖨𝖳(Y^*, Y_l). For conditioning the subsequent network layers, the weighted intermediate attractors are added to the input embeddings. Eq. <ref> is modified by inserting the conditioning function for the input embedding: E_l = 𝖤𝗇𝖼𝗈𝖽𝖾𝗋𝖫𝖺𝗒𝖾𝗋_l(𝖢𝗈𝗇𝖽𝗂𝗍𝗂𝗈𝗇(E_l-1)), 𝖢𝗈𝗇𝖽𝗂𝗍𝗂𝗈𝗇(E_l) = E_l + W A_l Y_l, where W ∈ℝ^D× D is learnable parameters that control the weights of intermediate predictions.W is the only additional parameter introduced for the proposed method and is shared among L-1 layers. A_l Y_l in Eq. <ref> can be interpreted as a T-length sequence of weighted averages of attractor vectors, and the weights are determined by the intermediate speaker labels. Through the intermediate attractors, the subsequent layers are conditioned on the intermediate speaker labels. §.§ Non-autoregressive Attractor Extraction For training efficiency with the intermediate speaker labels, we propose a non-autoregressive extraction of speaker-wise attractors instead of the LSTM-based autoregressive extractor used in EDA <cit.>. The difference is depicted in Fig. <ref>. We prepare query vectors Q ∈ℝ^C× D for the attractors as learnable parameters through training. A cross-attention module extracts the attractors by using the query vector and the frame-wise embeddings E_L as keys and values: A = 𝖠𝗍𝗍𝗇(Q, E_L, E_L), where 𝖠𝗍𝗍𝗇 is a multi-head attention layer used in Transformer decoders <cit.>. The intermediate attractors are extracted by using the same attention layer: A_l = 𝖠𝗍𝗍𝗇(Q, E_l, E_l) (1 ≤ l ≤ L-1). Speaker labels and their intermediates are estimated using Eqs. <ref> and <ref>, respectively. The intermediate labels are utilized to condition the subsequent layers using Eqs.<ref> and <ref>. Note that the original EDA can work with an unknown number of speakers, while our non-autoregressive extraction generates a fixed number of attractors. Our method can be extended to an unknown number of speakers by adding a speaker-counting objective for the attractors, similar to EDA <cit.>. § EXPERIMENTS §.§ Data The statistics of the datasets are listed in Table <ref>. We followed the family of EEND works <cit.> to prepare test data for two-speaker telephone conversations extracted from CALLHOME [The data preparation code is available at <https://github.com/hitachi-speech/EEND>]. We call CH-adapt for the adaptation dataset and CH-test for the test dataset. Simulated two-speaker audio mixtures were used for a training set. The source audio corpora were Switchboard-2 (PhaseI, II, III), Switchboard Cellular(Part1, 2), and the NIST Speaker Recognition Evaluation (2004, 2005, 2006, 2008). MUSAN corpus <cit.> was used for adding noise. With a recently proposed mixture simulation algorithm <cit.>, simulated conversations (SimConv) were prepared using the statistics of the CH-adapt dataset. Note that reverberation was not applied, which was the best configuration in <cit.>. §.§ Model Configurations EEND-EDA: We used PyTorch-based implementation of EEND-EDA [<https://github.com/BUTSpeechFIT/EEND>] used in <cit.>. Our baseline EEND-EDA model was exactly the same configuration as described in <cit.>, which is also the same as the original EEND-EDA paper <cit.>. Audio features were 23-dimensional log-scale Mel-filterbanks computed every 10 msec. For input to the neural network, 15 consecutive audio features were stacked to form 345-dimensional vectors every 100 msec. We used four Transformer encoder layers with 256 attention units containing four heads. Training configuration was also the same as <cit.>. The training batch was a collection of 50-sec segments in the training set. The batch size was 32. Adam optimizer was used with the Noam learning scheduler and 200k linear warm-up steps. The training was run for 100 epochs with the training set. For adaptation, we run 100 epochs using the adaptation set with a learning rate of 10^-5. After training, the last 10 model checkpoints were averaged. EEND-EDA+InterLoss: The model is built with additional computations (Eqs. <ref>-<ref>) introducing intermediate losses to the EEND-EDA model as described in Section <ref>. EEND-EDA+SelfCond is built with additional computations (Eqs. <ref>-<ref>) to apply the self-conditioning technique with the intermediate attractors. EEND-NA: We built a model named EEND-NA as EEND with the proposed non-autoregressive attractor extraction as described in Sec. <ref>. An +InterLoss model and a +SelfCond model were built on top of the EEND-NA model, similar to the EEND-EDA models. EEND-EDA-deep and EEND-NA-deep: We investigated the effect of deeper networks by increasing the number of Transformer blocks from four to eight. §.§ Metrics The diarization error rates (DERs) were evaluated with a forgiveness collar of 0.25s. We also showed DER breakdown into miss, false alarm (FA), and confusion (CF) error rates, followed by speech activity detection (SAD) miss and false alarm errors. The number of parameters (#Params) and training throughput (Tp) for each model were shown to discuss the results with training efficiency. The training throughput was calculated on a Tesla V100 32G GPU. §.§ Results Table <ref> shows the DERs on CH-test. The first three rows show that the conventional EEND-EDA model could not benefit from the proposed intermediate labels and the self-conditioning technique. Training throughput was down to one-third. The results indicate that the LSTM-based autoregressive attractors cannot optimize the intermediate frame-level embeddings. On the other hand, the proposed EEND-NA model showed performance improvement with the intermediate labels and the self-conditioning technique. The training throughput of the proposed model was higher than that of the EEND-EDA model, and the slowdown by introducing the intermediate labels and self-conditioning was in an acceptable range. The EEND-NA+SelfCond model reached a similar performance to the EEND-EDA model, while the proposed model has less number of parameters and higher throughput than the EEND-EDA model. The results suggest that the non-autoregressive attractors can help optimize the intermediate frame-level embeddings, unlike EDA. The EEND-NA model itself was worse than EEND-EDA, although the intermediate labels reduce the difference. The disadvantage of the non-autoregressive attractors may come from the conditional independence between speakers. Using conditional inference on previously estimated speakers, like in the decoder part of EDA, may improve performance. Table <ref> also shows the results of deeper (eight-layer) models. The proposed EEND-NA-deep+SelfCond model achieved the best performance among the evaluated models. While the conventional EEND-EDA-deep model achieved better performance than the original four-layer EEND-EDA model, we could not get the results with intermediate labels because of the slow training. We expect that the EEND-EDA-deep model cannot benefit from the intermediate labels as with the baseline EEND-EDA models. §.§ Results with Intermediate Speaker Labels Table <ref> shows diarization error rates with intermediate speaker labels produced by the EEND-NA-deep+SelfCond model. The errors were consistently reduced layer by layer. The results indicate that optimizing the speaker label at the lower layers improves the final diarization performance. Unexpectedly, the seventh-layer results were better than the last eighth-layer results. We think that conditioning with the intermediate labels could be enhanced by selecting the set of intermediate layers: not all the layers. § CONCLUSION We proposed an end-to-end diarization model that uses intermediate speaker labels to condition the subsequent network layers. The experiments showed that the intermediate labels with the non-autoregressive intermediate attractors boosted the diarization performance. The proposed method with the deeper network benefits more from the intermediate labels, resulting in better performance and training throughput than EEND-EDA. IEEEbib

http://arxiv.org/abs/2303.06776v1

20230312233142

Robot Health Indicator: A Visual Cue to Improve Level of Autonomy Switching Systems

cs.RO

Axr1050@student.bham.ac.uk Extreme Robotics Laboratory, University of Birmingham Edgbaston Birmingham England B15 2TT Maen0191@student.umu.se Andreas.Theodorou@umu.se Umeå university MIT-huset Umeå Sweden 901 87 R.Stolkin@bham.ac.uk M.Chiou@bham.ac.uk Extreme Robotics Laboratory, University of Birmingham Edgbaston Birmingham England B15 2TT Using different Levels of Autonomy (LoA), a human operator can vary the extent of control they have over a robot's actions. LoAs enable operators to mitigate a robot's performance degradation or limitations in the its autonomous capabilities. However, LoA regulation and other tasks may often overload an operator's cognitive abilities. Inspired by video game user interfaces, we study if adding a Robot Health Bar to the robot control UI can reduce the cognitive demand and perceptual effort required for LoA regulation while promoting trust and transparency. This Health Bar uses the robot vitals and robot health framework to quantify and present runtime performance degradation in robots. Results from our pilot study indicate that when using a health bar, operators used to manual control more to minimise the risk of robot failure during high performance degradation. It also gave us insights and lessons to inform subsequent experiments on human-robot teaming. Robot Health Indicator: A Visual Cue to Improve Level of Autonomy Switching Systems Manolis Chiou =================================================================================== § INTRODUCTION AND RELATED WORK During mobile robot navigation tasks, factors like uneven terrain, obstacles, sensor noise etc. can degrade a robot's performance. If left unattended, performance degradation can cause robots to fail or perform sub-optimally <cit.>. Timely operator intervention or triggering of recovery behaviours can mitigate this. Using Levels of Autonomy (LoA) <cit.>, the extent of control that a human operator has over the robot's actions can be varied. For example during high performance degradation, an autonomous robot can be switched to manual control by an experienced human operator. The LoA can be switched back once the robot is capable of functioning autonomously again. In Human Initiative (HI) LoA switching systems, the robot operator is in charge of LoA regulation during the task. Compared to fully autonomous robots, manually controlled robots, and other implementations of variable autonomy systems, HI-LoA shows better task performance during remote navigation tasks in unknown environments <cit.>. However, the additional perceptual and mental effort required to monitor robot operation data and determine if LoA switching is needed, imposes comparatively higher levels of cognitive workload on the operator. While overloading an operator's cognitive abilities for prolonged periods can reduce task performance <cit.> due to stress, fatigue and varying levels of trust in the system <cit.>, low workload can cause out-of-loop performance problems like complacency and over-trusting the system <cit.>. Effective HI-LoA system design remains an open problem in existing literature <cit.>. A well-designed system should keep operator cognitive workload within a sweet spot <cit.> (i.e., acceptable levels). LoA switches by operators should be due to a clear understanding of the robot's capabilities and limitations, not due to trust issues. Transparent design, i.e., where the operator can accurately interpret the robot's capabilities, goals, and its progress, have reported effective calibration of trust <cit.>, even with increasing levels of automation <cit.>. Systems where information about the robot is presented over the graphic/visual modalities are trusted more than the audio and textual modalities <cit.>. Independent of systems design considerations, significant differences still exist in people's perceived reliability of variable autonomy systems. However, they can be minimised through standardised training <cit.>. Video games serve as inspiration to understand how real-time information can be presented in a transparent manner, while keeping the operator cognitive workload in the sweet spot. Video game players report that they are receptive to information presented through visual cues as long as they are clear and consistent <cit.>. Game visual cues are usually colour coded UI elements overlayed on the interface during gameplay, e.g. the Health Bar. In popular online games[See https://overwatch.blizzard.com/en-us/], remotely situated players use these cues to get their teammates attention. Similarly, robot operators can use colour coded visual cues to acquire quick situational awareness <cit.> while operating a robot with minimum perceptual effort. For robot interfaces, Murphy et al. <cit.> recommend that visual cues may be added to the robot control UI, without blocking useful information. The existing literature shows that effective trust calibration, system transparency and standardised training are crucial for a well designed HI-LoA switching system. Information to assist the operator with making decisions about LoA regulation can be presented through properly designed visual cues. Therefore, in this study, we present our pilot experiments to explore how visual cues about a robot's performance degradation can help LoA regulation decisions during high cognitive workload situations. Based on the Robot Vitals and Robot Health framework <cit.> to quantify performance degradation, we design a colour coded Robot Health Bar UI element. We hypothesise that using the Health Bar will reduce the perceptual effort required by operators to make LoA switching decisions, improve task performance, and result in a lower overall cognitive workload. We also took feedback from the participants about their overall trust in the system, its transparency and recommendation for improving the Health Bar design. § SYSTEM DESIGN The experiment consisted of two tasks, similar to the experiments setup by Chiou et al. <cit.>. The primary task was a mobile robot navigation task using the Clearpath Husky Robot, simulated on Gazebo. As shown in Figure <ref>, these arenas were designed to mimic urban search and rescue scenarios, and were populated with performance degrading factors commonly found in them like - obstacles, uneven terrain and laser noise. A 2D laser scan of an empty map (Figure <ref> - Right) was generated before the performance degrading factors were added. The difference between the map used for robot navigation and the actual arena affects robot navigation planning, thereby adding another performance degrading factor. Two different arenas were created to compare operator performance with (condition A) and without the Health Bar (condition B). RViz was used to visualise the robot map, sensor data, and give commands to the robot. The standard RViz interface in Figure <ref> was used for Condition B, and the interface with a Robot Health Bar shown in Figure <ref> was used for Condition A. The Husky robot used two LoAs - 1) Waypoint-based navigation and 2) Manual Control by an operator using a Joystick. The Robot Health at each instant was calculated using the Robot Vitals and Robot Health framework <cit.>, and the Health Bar was created and displayed using the JSK Visualisations[http://wiki.ros.org/jsk_visualization/] ROS Package. Here, Robot Health is defined as an overall scalar estimate of a robot’s ability to carry out its tasks without its capabilities being impaired by any performance degrading factors . Therefore, an operator can monitor the health to detect situations where a robot is likely to fail, and trigger LoA switches to assist the robot. To improve readability, the Robot Health was standardised to the range [0,1]. The Red-Amber-Green colour coding convention was used for the Health Bar as shown in Figure <ref>. Based on preliminary experiments, health above 0.7 was decided as Healthy , and the colour of the Health Bar was green. Values between 0.5-0.7 were coloured in amber, and values below 0.5 were coloured red. Instead of sharp colour changes, the Health Bar gradually went from green to amber to red. This was done to minimise rapid colour changes due to small fluctuations in health near the threshold values. The secondary task was a 3D object rotation task <cit.>, used to induce additional cognitive workload for the operator. Here participants were successively presented with two 3D objects. They had to determine whether the objects were the same or different. Instead of having a fixed time limit like in Chiou et al.<cit.>, here the participants did this task throughout the runtime. Both the primary and secondary tasks were displayed on adjacent screens. Participants were explicitly instructed to prioritise the robot navigation task and minimise the likelihood of robot failure . They were also told to simultaneously do the secondary task to the best of their ability, and that they were not being evaluated on their performance in it. § EXPERIMENT METHODOLOGY Eight test subjects participated in the experiment and performed both conditions (i.e. within subjects design). The order of the conditions was counterbalanced to minimise learning and fatigue effects. First, participants had to fill out background information (see appendix <ref>). Then they were introduced to basic robot navigation in a training arena similar to the one used in Chiou et al.<cit.>. All participants trained on this arena till they were able to demonstrate a minimum proficiency in robot navigation. This ensured that confounding factors due to a variation in skill levels were minimised. Next, they were shown the 2 different LoA, how to switch between them and were given some time to practice LoA switching on the training arena. Participants were then introduced to the interface with the Health Bar and asked open-ended questions without prior explanation (see appendix <ref>). After answering the questions, they were given the following instructions regarding the Health Bar: “The Robot Health Bar indicates how much a robot’s performance is degraded by environmental factors. These environmental factors can be anything ranging from bad terrain to laser noise. During low health, the Health Bar will become redder. During high health, the Health Bar will become greener. The lower the robot's health, the more likely it will fail. You may use the Health Bar to help you determine when the robot requires a LoA switching”. The effect of performance degrading factors on the Health Bar was demonstrated by introducing obstacles and laser noise in the training arena. Participants were also given time to familiarise themselves with the interface before starting the navigation task. Next participants were familiarised only with the secondary task, and allowed multiple practice runs till they felt comfortable with it. Their baseline on the secondary task alone was measured before the experiment. Before each condition, participants were shown the start and finish points in the navigation task and told that the next waypoint would be assigned automatically on reaching the current one. Then the participants carried out each experimental condition followed by a NASA-TLX form to evaluate the perceived cognitive workload during the task. Lastly, participants completed open-ended questions and a transparency/trust questionnaire after the experimental trials and the NASA-TLX forms (see <ref> and <ref>). § RESULTS Robot health values under 0.7 were classified as unhealthy . This criterion was heuristically determined based on previous studies on the robot vitals and robot health framework <cit.>. Results from the experiments conducted on 8 participants were computed and tested for statistical differences using two-tailed pairwise T-Tests. These results are summarised in table <ref>. The total percentage of runtime that the robot was autonomous, manually controlled and unhealthy was calculated for each experimental condition. The percentage of run time that the robot health was unhealthy showed no statistical differences between the two conditions. Similarly, no significant differences were observed in the percentage of run time the robot was manually controlled. However, the percentage of run time that the robot was manually controlled when it was unhealthy, showed a significant (p <0.05, t = -2.36) difference between both conditions. That is, operators manually controlled unhealthy robots for longer when the health of the robot was displayed (i.e., Condition A). As a result, the robot was unhealthy for an average of 61.91% of the runtime in condition A, but 78.03% of the runtime in condition B. The change in perceptual effort required to carry out LoA switching tasks was measured using the operator's accuracy on the secondary task. When the health bar was not displayed (Condition B), operators showed significantly higher levels of accuracy (p <0.05, t = 2.36) on the secondary task. This indicates the perceptual effort required to use the interface with the robot health bar, is higher. However, the NASA-TLX scores showed no significant differences in the overall cognitive workload imposed by both conditions. Open-ended questions asked before the experiment showed that some participants were slightly confused about what the Robot Health Bar was. One participant said - "The health bar looks like a timer, because it has a number. I thought it meant seconds". Others successfully grasped the idea behind the UI element. Another participant thought "When the number goes down the robot dies", indicating a game-like perception of the Health Bar's function. However, all participants understood it was there to assist or alert an operator performance degradation. Most participants preferred the interface of condition A (with the Robot Health bar ), finding it intuitive for regulating LoA - "When the health was low, it was better to manually control the robot". One participant said - "Without the health bar I had to actively use my brain to detect when the robot needed help". Three participants, however, felt that the colours were sometimes inaccurate and did not match the robot's state. One of them pointed out that while red and green clearly helped indicate which LoA was better, amber was confusing because they did not know what to do. Although not significant, operators that had experience with operating robots (according to the background information) seemed to prefer more transparency in the UI, while the rest preferred less transparency. Novice operators (as reported by the background information), instead, stated that they responded to the colours of the UI to a greater extent than the participants with more experience in operating robots. Pearson's correlation showed a significant negative trend between how often the participants operated remote-controlled vehicles and how easy it was to know when to change LoA for alternative 1 (Figure <ref>) in the questionnaire (p <0.01). Significant positive trends were also observed for how often they used AI for work and how easy it was to know when to change LoA for alternative 3 in the questionnaire (Figure <ref>) (p <0.05), and how easy it was to understand why to change LoA for alternative 2 (Figure <ref>) (p <0.01) and alternative 3 in the questionnaire(Figure <ref>) (p <0.01) (see Table <ref>). In terms of determining when to change LoA for the alternatives in the questionnaire, four participants rated either the highest or both the lowest and highest levels of transparency equally highly on the Likert scale (Figure <ref>). Conversely, the remaining four participants rated the lowest level of transparency highest in comparison to the other two levels. Overall, the medium level of transparency was rated the lowest. As for why to change LoA, six participants either increased or maintained their rating for each increase in the level of transparency, resulting in the highest level of transparency being rated the highest (Figure <ref>). However, two participants decreased their rating on the Likert scale for each increase in the level of transparency. § DISCUSSION AND INSIGHTS Our experimental results illustrate how informing an operator about a robot's performance degradation through visual cues can affect the operator's driving and LoA regulation style. When shown the health bar, operators triggered LoA switches to mitigate situations where the robot was unhealthy . This style of HI-LoA regulation reduces the aggregate risk of robot failure. The interface with the health bar did not impose a significant additional cognitive workload on the operator during the experiment, and did not significantly change the task completion time. Contrary to our hypothesis, LoA regulation using the Health Bar required significantly (p<0.05) higher perceptual effort. In condition A, Participants carried out fewer rotations and gave less accurate answers. One likely explanation for this is that adding a Health Bar increased the number of points on the UI the operator had to focus on, which increased the perceptual effort for the primary task. Table <ref> shows that novice robot prefer significantly less transparency than experienced robot operators. Participant feedback revealed that while novice operators responded more to the colours of the health bar, experienced operators preferred having more information displayed to inform their LoA switching decisions. This indicates that participants value transparency only when they understand the necessity of transparency. People with different levels of experience have varied mental models about the Robot, leading them to perceive the system and the Health Bar differently. This makes it difficult to standardise explanations. Depending on their experience, each participant may want to clarify different aspects of the Human-Robot System, leading to differences in the way they are primed for the experiment. Therefore, rigorous training coupled with detailed explanations about the different components of the system are required to minimise the differences in perception. Finally, the major limitation of this study is the small sample size. Therefore all the results and insights presented in this study can require validation with a larger set of participants. §.§.§ Participant Recommendations to improve the Robot Health Bar and User Interface Design: * Make the Health Bar more Salient, so that it attract attention when necessary * Do not include a drop down-menu in the Health Bar as shown in the Questionnaire * Use a percentage to the health value instead of a value between [0,1] * To attract attention to 'low health' use multi-modal awareness cues like sound alerts or UI elements e.g. By Shaking or Blinking the health bar after a threshold value. § CONCLUSION AND FUTURE WORK In this study, we explored if visual cues about a robot's performance degradation can reduce the perceptual effort required to make LoA switching decisions for remote mobile robot navigation tasks. Inspired by video games, we designed a 'Robot Health Bar'. This health bar displayed the total runtime performance degradation the robot is facing. A total of 8 participants carried out a mobile robot navigation task with and without the health bar UI element under high cognitive workload, and their performance was measured. Adding a Robot Health Bar to the robot control UI significantly changed how the operator makes Level of Autonomy Switching Decisions. When the Health Bar was displayed, operators were more attentive and took control of the robot more to minimise the risk of the robots failing. Visual cues that indicate a robot's health can serve as an effective way of ensuring safer control of robots, especially in extreme environments where robot missions have high levels of risk and environmental adversities. In the future, we aim to explore how the insights gained from this study can scale to multi-robot systems. unsrtnat § OPEN-ENDED QUESTIONS AND QUESTIONNAIRES §.§ Background information questionnaire The background questionnaire consisted of the following questions: Q1: How often do you operate or use to operate remote controlled vehicles (e.g. Robots, Drones, Heavy machinery) Q2: How often do you play or used to play video games involving driving, flight simulation and third person shooters, RPG and sports? Q3: Do you use AI (e.g. Autonomous Robots, Machine Learning Algorithms, AI Tools ) for work? Q4: Do you use AI (e.g. Personal Assistant) in your personal life? The possible answers ranged between 1-5; i.e., least to most often. §.§ Open-ended questions Questions asked before participants had performed the experiment under both conditions are listed below: Q1: What are you thinking as you look at this? Q2: What is your first impression of this UI element? Q3: What do you think this UI element does or will do? All questions asked after participants had performed the experiment under both conditions are listed below: Q1: Was anything surprising or did not perform as expected in either of the interfaces? Q2: Was the interface without the Robot Health Bar easy to understand? Q3: Was the interface with the health bar easy to understand? Q4: What did you think about the colours used in the health bar when the health changed? Q5: Was the LoA switching behaviour of the interface without the Robot Health Bar transparent? Q6: Was the LoA switching behaviour of the interface with the Robot Health Bar transparent? Q7: Is there anything else that you can think of, specific to the UI and UX to improve the use of LoA switching Robots? §.§ Transparency/trust questionnaire The transparency/trust questionnaire consisted of images of three different alternatives of a health bar ranging from low to high transparency and a 7-point Likert scale for each image (Figure <ref>). The questions were the following pair, for each alternative: Q1: It is easy to understand WHEN to change Level of Autonomy (LoA) based ONLY on alternative 1? Q2: It is easy to understand WHY to change or not change Level of Autonomy (LoA) based ONLY on alternative 1? § RESULTS tablesection figuresection

http://arxiv.org/abs/2303.07208v1

20230313154006

Social network analysis of Japanese manga: similarities to real-world social networks and trends over decades

physics.soc-ph

APS/123-QED Department of Transdisciplinary Science and Engineering, Tokyo Institute of Technology, 152-8550 Tokyo, Japan naokimas@buffalo.edu Department of Mathematics, State University of New York at Buffalo, Buffalo, NY 14260-2900, USA Computational and Data-Enabled Science and Engineering Program, State University of New York at Buffalo, Buffalo, NY 14260-5030, USA Manga, Japanese comics, has been popular on a global scale. Social networks among characters, which are often called character networks, may be a significant contributor to their popularity. We collected data from 162 popular manga that span over 70 years and analyzed their character networks. First, we found that many of static and temporal properties of the character networks are similar to those of real human social networks. Second, the character networks of most manga are protagonist-centered such that a single protagonist interacts with the majority of other characters. Third, the character networks for manga mainly targeting boys have shifted to denser and less protagonist-centered networks and with fewer characters over decades. Manga mainly targeting girls showed the opposite trend except for the downward trend in the number of characters. The present study, which relies on manga data sampled on an unprecedented scale, paves the way for further population studies of character networks and other aspects of comics. Social network analysis of Japanese manga: similarities to real-world social networks and trends over decades Naoki Masuda March 30, 2023 ============================================================================================================= § INTRODUCTION The global comic market has been rapidly growing. The market size was valued at USD 14.7 billion in 2021 and is expected to expand at a compound annual growth rate of 4.8% from 2022 to 2029 <cit.>. A driving force underlying the rapid expansion of the market is the widespread availability of e-books, which enables us to easily access comics around the world. The COVID-19 pandemic also helped the growth of the comic market because the lockdown in many countries increased the demand for comics. Manga, Japanese comics, has been especially popular on a global scale. Many manga have been translated into different languages. For example, over 510 million copies of One Piece have circulated worldwide; One Piece is recorded in the Guinness Book of Records as the best-selling comic in the history <cit.>. The history of the development of manga culture in Japan, its social background, and its global prevalence have been studied <cit.>. Our social networks are complex but characterized by common properties such as heterogeneity and community structure <cit.>. Because literally all storylines of manga rely on interactions among characters, social networks among characters, which are often called character networks, may provide a backbone of the story of manga and influence its popularity. More generally, character networks have been analyzed for quantitatively characterizing fiction works <cit.>. A number of problems can be addressed through analyses of character networks, such as summarization <cit.>, classification <cit.>, and role detection <cit.>. Character networks have been studied for a variety of fiction such as novels <cit.>, plays <cit.>, movies <cit.>, and TV series <cit.>. However, character networks for comics have been rarely explored except for Marvel comics <cit.> and a few manga titles <cit.>. One possible reason for this is the difficulty of automating data collection for comics <cit.>. In the present study, we analyze character networks of 162 popular Japanese manga that span decades and test the following two hypotheses. First, we expect that popular manga tend to produce realistic social networks. Therefore, we hypothesize that structural and temporal properties of the character networks of manga are similar to those of human social networks. Second, our data enables us to investigate how trends of manga have changed over approximately 70 years. Therefore, we hypothesize that character networks of modern manga and old ones are systematically different. We also examine differences in the network structure between manga mainly targeting boys versus girls. § RESULTS §.§ Similarities to Empirical Social Networks Our original data are equivalent to a temporal bipartite graph in which the two types of nodes are characters and pages, and the edges connect characters to the pages in which they appear. The pages are ordered in time. We show the degree distribution for the character nodes for One Piece and the Poisson distribution with the same mean in Fig. <ref>A. The coefficient of variation (CV), which is defined as the standard deviation divided by the average, of the original degree distribution is 1.66. This value is approximately 11.1 times larger than that of the Poisson distribution. Therefore, similar to human social networks <cit.>, the characters in manga have heterogeneous numbers of connections, and there are a small number of characters that appear disproportionately frequently on various pages. We obtained similar results for the other manga (see https://github.com/KS-92/MangaSI file). In contrast, the CV of the degree distribution for the page nodes is 0.462 for One Piece. The CV for the Poisson distribution with the same mean is 0.580. Therefore, we argue that the characters do not appear uniformly randomly over the pages. As we show in Fig. <ref>B, there are typically two or three characters on a page in One Piece, and pages containing none or just one character are relatively rare. We obtained similar results for the other manga (see https://github.com/KS-92/MangaSI file). These results suggest that the high heterogeneity of the characters in terms of the frequency of appearance on pages and the tendency of a page typically containing two or three characters are two common properties of manga. Therefore, in the following analyses, we consider the bipartite configuration model, in which the degrees of all character and page nodes are preserved and the edges are otherwise randomly placed, as a null model. Then, we examine properties of character networks that we can explain by the null model versus those we cannot. We show the weighted character network, which is the projection of the bipartite graph onto the space of character nodes, for One Piece in Fig. <ref> (see https://github.com/KS-92/MangaSI Fig. S1 for the networks for all 162 manga). By definition, the edge weight is equal to the number of pages in which the two characters simultaneously appear. The character networks vary widely in size from manga to manga. Specifically, the number of nodes, N, ranges from 8 to 124, with mean± standard deviation=32.6 ± 16.6. The number of edges, M, ranges from 19 to 950, with mean± standard deviation=156.1 ± 128.9. We define the protagonist as the node with the largest node strength (i.e., weighted degree) in the character network. The protagonist is the character that appears in the largest number of pages for most manga (97.5%). In fact, the protagonist for One Piece, indicated by an arrow in Fig <ref>, is Monkey D. Luffy, who is generally known as the most central character in One Piece. We compared various properties of character networks between the empirical character networks and their randomization obtained by the one-mode projection of the bipartite network generated by the bipartite configuration model. We first compare three strength-related indices, i.e., the average strength, ⟨ s ⟩, the CV of the strength, s_ CV, and the strength of the protagonist, s_P, of the character networks for all 162 manga between the original and randomized networks in Figs. <ref>A-C. These figures suggest that randomized character networks well explain the strength-related indices of the original character networks. Figure <ref>A shows the strength distribution for One Piece. We find that the strength obeys a heavy-tailed distribution over some scale. This result is consistent with Fig. <ref>B, which shows that the CV of the strength for most manga is substantially larger than 1; the CV for One Piece is 1.48. Note that the exponential distribution, which lacks a heavy tail, yields CV=1. The heavy-tailed strength distributions shown in Fig. <ref>A for One Piece and implied in Fig. <ref>B for a majority of manga are consistent with those for empirical social networks <cit.>. In contrast to the strength-related indices, the randomized character networks do not explain the degree-related indices of the original character networks. The randomized networks tend to overestimate the average degree, ⟨ k ⟩, of the original character networks (see Fig. <ref>D). This result suggests that characters tend to repeat interacting with relatively few characters. The CV of the degree, k_ CV, for the original networks tends to be larger than that for the randomized networks (see Fig. <ref>E). The fraction of the other characters that the protagonist is adjacent to, k_ P/(N-1), is 0.900 ± 0.103 ( mean± standard deviation based on the 162 manga) and 0.971 ± 0.0413 for the original and randomized character networks, respectively (see Fig. <ref>F). Therefore, the protagonist is adjacent to most of the other characters in both original and randomized networks. Based on these results, we conclude that the character networks are strongly protagonist-centered, in which the protagonist interacts with most of the other characters, while other characters tend to interact only with fewer characters than expected for the randomized networks, but including the protagonist. This interpretation is consistent with the aforementioned observation that ⟨ k⟩ and k_ CV are smaller and larger for the original than randomized networks, respectively. We show the degree distribution for One Piece in Fig. <ref>B. The CV of the degree, k_ CV, for One Piece is 0.77. Figure <ref>E shows that the CV of the degree for most of the manga is smaller than 1. In fact, this result is consistent with that for empirical social networks with similar number of nodes <cit.>, while large empirical social networks tend to have heavy-tailed degree distributions implying a large CV <cit.>. We found that the strength is super-linearly scaled with the degree, i.e., s_i ∝ k_i^α with α>1, in the character networks, where ∝ indicates “proportional to”. Note that the absence of correlation between the strength and degree would yield α=1 <cit.>. Specifically, in Fig. <ref>C, we show the relationship between k_i and s_i for One Piece. We obtained α=1.94 with the coefficient of determination R^2=0.847. We obtained similar results for the other manga (see Fig. <ref>G). These results are consistent with power-law relationships between the strength and degree with α > 1, which has been observed in empirical social networks <cit.>. Note that the randomized character networks also yield α > 1 and that α for the original and randomized networks are highly correlated with a small number of exceptions. The character networks are disassortative. The degree assortativity coefficient, denoted by r, of the character network is -0.345 on average. In fact, the values of r for all but two manga are negative (see Fig. <ref>H). This result is in stark contrast with various observations that social networks are more often than not assortative with positive r <cit.>. We argue that the character networks are disassortative because they are strongly protagonist-centered. In fact, star graphs are disassortative with large negative values of r <cit.>. The degree disassortativity is also observed for empirical egocentric social networks <cit.>. Figure <ref>H indicates that r is also negative for the randomized character networks although the correlation between r for the original and randomized networks is low. Therefore, we conclude that the degree disassortativity in our character networks is a consequence of the protagonist-centered nature of the original character-page bipartite network. The character networks are highly clustered. The clustering coefficient, denoted by C, over the different manga is 0.790 with the minimum value of 0.598 (see Fig. <ref>I). This result is consistent with the observations that empirical social networks have high clustering coefficients <cit.>. In fact, randomized character networks also have similarly large C although the spread is large between the empirical and randomized networks. Therefore, we conclude that a high clustering coefficient is a consequence of one-mode projection of the bipartite graph, which is known <cit.>. We found that heavy-tailed distributions of interevent times (IETs) are simultaneously present for nodes and edges in the character networks. Copresence of two characters on a page defines a time-stamped event on the edge, where we identify the page number as discrete time. A time-stamped event for a given node is the presence of the character on a page. We show the survival functions of the IETs (i.e., probability that the IET, denoted by τ, is larger than the specified value) for individual nodes and edges in One Piece in Figs. <ref>A and <ref>B, respectively. The relatively slow decay in Fig. <ref> suggests heavy-tailed distributions for both nodes and edges across some scales of τ. The CV values for IETs on nodes and edges for One Piece are 1.72 ± 0.697 and 1.92 ± 0.535, respectively. We obtained similar results for the other manga, as we show in Figs. <ref>J and <ref>K. These figures also indicate that the randomization of the bipartite network does not preserve this feature, yielding CV values close to 1 regardless of the CV value for the original temporal character networks. It should be noted that a Poisson process produces an exponential IET distribution, which yields CV = 1. These results support that IETs for nodes and edges in the original character networks are non-Poissonian and heterogeneously distributed, which is consistent with properties of empirical social networks <cit.>. Related to heavy-tailed IET distributions, we found that epidemic spreading occurs more slowly in the temporal character networks than in randomized counterparts. We ran the susceptible-infectious (SI) model. By assumption, an infectious character independently infects each susceptible character coappearing on the same page with probability β=0.2. We ran the SI dynamics N times, with the ith character being the only initially infectious one in the ith run. For One Piece, we show in Fig. <ref> the time course of the fraction of the infectious characters averaged over the N runs. We also show the corresponding averaged time courses for each of the 1,000 randomized temporal networks by the blue lines. Figure <ref> indicates that the infection occurs more slowly in the original temporal network than in the randomized temporal networks. We obtained similar results for the other manga (see Fig. <ref>L). These results are qualitatively the same as those observed for empirical social temporal networks <cit.>. §.§ Trends of Network Structure over Decades The 162 manga span nearly 70 years, from Astro Boy with the first volume being published in 1952 to Jujutsu Kaisen in 2018 (see https://github.com/KS-92/MangaSI file for detailed information on the 162 manga). The trend of the structure of manga character networks may have changed over the ≈ 70 years, reflecting the transitions of Japanese society including in economics and fashion. In this section, we explore possible existence of such trends. We first calculated the Pearson correlation coefficient between the year of publication of the first volume, denoted by y, and various indices of network structure on the basis of all 162 manga. We show the correlation coefficients and their 95% confidence intervals (CIs) in Fig. <ref>A. We find that the number of characters, N, tends to decrease over decades (i.e., negative correlation between N and y). The CV of the node degree, k_ CV, and that of the node strength, s_ CV, are negatively correlated with y with moderate effect sizes. In general, these and other network indices may be affected by the number of nodes, N <cit.>. Therefore, we also examined the partial correlation coefficient between y and each index by partialing out the effect of N. The results of the partial correlation coefficient for k_ CV and s_ CV, shown by the lines with squares in Fig. <ref>A, are consistent with those of the Pearson correlation coefficient, although the partial correlation is closer to 0 than the Pearson correlation. We also find that the average degree, ⟨ k ⟩, is positively correlated with y in terms of the partial correlation. These results suggest that more recent manga tend to be denser and more homogeneous in the node's connectivity, such that various characters, not just the protagonist, tend to have more connections in more recent manga. How the trend of the character network changes over time may depend on genres of manga. A most major categorization of Japanese manga is the one based on the gender of their main readership <cit.>. Therefore, we classify the 162 manga into 124 boys' and 38 girls' manga based on the genre section on their Wikipedia pages (see https://github.com/KS-92/MangaSI file for the genre of each manga), while their classification is necessarily subjective. We show the Pearson and partial correlation coefficients between each index and y, separately for the boys' and girls' manga in Figs. <ref>B and <ref>C, respectively. We find that there tend to be less characters for more recent manga for both boys' and girls' manga. The boy's and girls' manga are different in the following aspects in terms of the partial correlation coefficients, i.e., when we control for the number of characters. The trends that we identified for all the 162 manga are even more pronounced in the boys' manga. The effect sizes of the partial correlation for ⟨ k ⟩, k_ CV, and s_ CV for the boys' manga are larger than those for the 162 manga. In addition, the degree assortativity coefficient, r, is positively correlated with y. Because r is negative for most manga, this result implies that r tends to be closer to 0 for more recent boys' manga. On the other hand, the two protagonist-related indices, i.e., k_ P/(N-1) and s_ P, are not correlated with y. These results suggest that the character networks in more recent boys' manga tend to be less protagonist-centered, in which non-protagonist characters have more connections among them. The trend for the girls' manga is opposite to that for the boys' manga except that the downward trend in the number of characters is common. The two indices for which the partial correlation was negative for the boys' manga, i.e., k_ CV and s_ CV, are positively correlated with y for the girls' manga in terms of the partial correlation. On the other hand, the two indices for which the partial correlation was positive for the boys' manga, i.e., ⟨ k ⟩ and r, are negatively correlated with y for the girls' manga. Furthermore, the normalized degree of the protagonist, k_ P/(N-1), for which we did not confirm a correlation for the boys' manga, is positively correlated with y. Overall, these results indicate that the character networks in the girls' manga have shifted toward more protagonist-centered networks. § DISCUSSION We found that the empirical character networks have various features consistent with empirical social networks. The bipartite configuration model, which is a standard random bipartite graph, has been shown to be reasonably accurate at explaining some features of the empirical character networks including statistics of the node's strength, α values, disassortativity, and the high clustering coefficients. In contrast, there are other features of the empirical character networks that deviate from the expectation from the bipartite configuration graph. Such features include statistics of the degree, the IET distributions, and speeds of epidemic spreading. We suggested that part of these differences originates from the protagonist-centered nature of the character networks. In other words, the protagonist interacts with most characters, while non-protagonist characters tend to interact only with the protagonist and a smaller number of other characters than expected by the configuration model. There are positive support of protagonist-centered social networks in the real world when the network is egocentric <cit.>. Similarities between character networks and human social networks have been investigated for some fictions such as Marvel comics <cit.>, Shakespeare's plays <cit.>, and myths <cit.>. Our results of the high clustering coefficient and the heavy-tailed distribution of the strength in the character networks are consistent with these previous studies. However, we ascribe the high clustering coefficient to the effect of one-mode projection rather than to similarity to empirical social networks because the randomized character networks also have high clustering. Our results of the disassortativity is also consistent with the results for Marvel comics <cit.> and myths <cit.>. Therefore, we infer that the heavy-tailed distribution of the strength and disassortativity in the character networks are common properties for various types of fictions. Although the aforementioned previous studies investigated only the static properties of the character networks, we further revealed temporal properties of the character networks, such as the long-tailed IET distributions, which are in fact consistent with empirical social networks <cit.>. We also found that the trend of the manga character networks has shifted over decades. Specifically, more modern manga tend to have fewer characters, denser and be less protagonist-centered for the boys’ manga, which may reflect a modern change in the society that places more emphasis on diversity and teamwork <cit.>. In contrast, the character networks in the girls' manga have shifted toward more protagonist-centered, although the downward trend in the number of characters is common. In general, girls' manga in Japan tend to revolve around issues of love and friendship with a focus on inner feelings of the protagonist <cit.>. Our results suggest that more recent girls' manga may describe the relationships between the protagonist and a few other characters in depth. There are many future directions of investigation. First, we focused on 162 manga with high circulations. By analyzing manga with low circulation as well, we may be able to discover static and temporal properties of character networks that readers favor, contributing to understanding why some manga are more popular than others. Second, we analyzed only the first three volumes of each manga. Examining all volumes would allow us to understand the narrative structure, which have been analyzed for novels <cit.>, movies <cit.>, and TV series <cit.>. For this purpose, a wide variety of time series and temporal network analysis tools, such as change-point detection, temporal centrality, and temporal community structure, may be useful <cit.>. Third, character networks are probably signed in most cases, connecting characters by positive or negative ties. It is worth to deploy sentiment analysis based on text <cit.> and facial expressions <cit.> to construct and analyze signed networks of characters in manga. Temporal network analysis of signed character networks may also reveal complicated dynamics of relationships among characters (e.g., an enemy later becomes an ally). Fourth, manga in different genres such as action, adventure, sports, and comedy, may have different structures of character networks. In addition, we have only analyzed the manga that have been published in the paperback pocket edition, which is the most common for the boys' and girls' manga in Japan. Analyzing manga of other sizes may help us understand variation of character networks over a wider range of genres. Fifth, studying character networks in comics in countries other than Japan warrants future work. For example, character networks may reflect societal differences from country to country. In conclusion, to the best of our knowledge, the present study of character networks for 162 manga titles is unprecedented in scale in quantitative studies of comics and has enabled us to discover their general features. We also introduced new tools and views to the analysis of fiction character network data, such as the bipartite configuration model, the concept of protagonist-centeredness, and temporal network analysis. We hope that this study triggers further quantitative studies of character networks in manga and other types of fictions. § MATERIALS AND METHODS §.§ Data Collection We focus on manga of which more than 10 million copies have been published in Japan as of January 2021 <cit.>. Since the physical size of the manga may affect the number of panels per page and we adopted the page as the unit of analysis, we excluded the manga that had not been published in the paperback pocket edition (112 mm× 174 mm). Note that the paperback pocket edition is the most common for the boys' and girls' manga in Japan. As a result, we included 162 manga (see https://github.com/KS-92/MangaSI file for the list of the 162 manga). We prepared a data table of time-stamped copresence of characters for each manga as follows. For each manga, we manually examined volumes one through three. All the selected 162 manga had at least three volumes. Then, we first extracted all the characters whose name, blood relation to a named character, or job title that uniquely identifies the character, is known. To ensure the reproducibility of this work, we avoided to use any other information sources (e.g., other volumes of the same manga, the Internet, or anime) to attempt to identify more characters. In practice, such additional information sources would not contribute to identifying many more characters. Second, we recorded the copresence of characters on each page as interaction between the characters. Note that the copresence is the most common definition of interaction in constructing character networks in fictional works <cit.>. We used copresence on a page rather than in a single panel within a page because it is common that characters appearing in different panels on the same page have some interactions (e.g., two characters talk to each other by using alternately occupying successive panels). §.§ Bipartite Configuration Model The original data from which we construct the temporal and static character networks are equivalent to a temporal bipartite graph in which the two types of nodes are characters and pages, and an edge connects a character and a page in which the character appears. To generate a randomized bipartite network, we choose a pair of edges in this bipartite graph uniformly at random, denoted by (c, p) and (c', p'), where c and c' are characters and p and p' are pages. If c and c' are the same or p and p' are the same, we discard the pairs and redraw them. Then, we rewire the two original edges to (c, p') and (c', p). We repeat this procedure 1,000 times excluding the discarded edge pairs to generate a randomized temporal bipartite network. We construct a static randomized character network from the generated randomized bipartite graph by one-mode projection. In other words, we connect two characters by an edge whose weight is the number of pages on which the two characters are copresent. §.§ SI Model We use the SI model <cit.> to numerically investigate contagion. We assume that just one character is initially infectious and that the other N-1 characters are initially susceptible. If a susceptible character co-appears with an infectious character on a page, the susceptible character becomes infectious with probability β. Different infectious characters appearing on the same page independently attempt to infect each susceptible character on the page. Once a character contracts infection, it stays infectious forever. We run the SI dynamics with each of the N characters as the sole character that is initially infectious. §.§ Temporal Correlation Coefficient We quantify the persistence of edges over time by the temporal correlation coefficient <cit.>. First, we calculate the topological overlap for node i at time t by O_it = ∑_j=1^N A_ij^t A_ij^t+1/√(∑_j=1^N A_ij^t∑_j=1^N A_ij^t+1) , where A_ij^t is the adjacency matrix of an unweighted network at time step t. In other words, A_ij^t = 1 if characters i and j are copresent on page t, and A_ij^t = 0 otherwise. We then define the average topological overlap for node i by O_i = 1/T-1∑_t=1^T-1 O_it , where T is the number of pages. The temporal correlation coefficient for the entire temporal network is given by O = 1/N∑_i=1^N O_i . We thank Juyong Park for valuable discussion. K.S. thanks the financial support by the Japan Society for the Promotion of Science (under Grant No. 19K23531 and 22K14438). N.M. thanks the financial support by AFOSR European Office (under Grant No. FA9550-19-1-7024), the Japan Science and Technology Agency (JST) Moonshot R&D (under Grant No. JPMJMS2021), and the National Science Foundation (under Grant No. 2052720 and 2204936).

http://arxiv.org/abs/2303.06772v1

20230312231542

Matrices whose field of values is inscribed in a polygon

math.RA

[label1]Supported by NSF Award https://www.nsf.gov/awardsearch/showAward?AWD_ID=2150511DMS-2150511. bgsu]Matthew J. Fyfe mjfyfe@bgsu.edu [bgsu] organization=Bowling Green State University, city=Bowling Green, postcode=43403, state=OH, country=U.S.A. bmc]Yesenia Hernandez yhernandez@brynmawr.edu [bmc] organization=Bryn Mahr College, city=Bryan Mahr, postcode=19010, state=PA, country=U.S.A. uwb]Pietro Paparellacor1 pietrop@uw.edu [url]http://faculty.washington.edu/pietrop/ [cor1]Corresponding author. [uwb] organization=Division of Engineering & Mathematics, addressline=University of Washington Bothell, city=Bothell, postcode=98011-8246, state=WA, country=U.S.A. bmc]Malini Rajbhandari mrajbhanda@brynmawr.edu In this work, it is shown that if A is an n-by-n convexoid matrix (i.e., its field of values coincides with the convex hull of its eigenvalues), then the field of any (n-1)-by-(n-1) principal submatrix of A is inscribed in the field of A, i.e., the field is tangent to every side of the polygon corresondping to the boundary of the field of A. This result generalizes a special case established by Johnson and Paparella [Amer. Math. Monthly 127 (2020), no. 1, 45–53]. convexoid matrix field of values numerical range polygon principal submatrix [2020] 15A60 § INTRODUCTION The field of values (or numerical range) of a matrix A is the image of the two-norm unit-sphere in complex Euclidean space with respect to the map x ⟼ x^∗ A x. Recently, Johnson and Paparella <cit.> used various concepts from matrix analysis, including the discrete Fourier transform matrix, the field of values, trace vectors, and differentiators, to provide a framework that admits short proofs of the Gauss–Lucas and Bôcher–Grace–Marden theorems (the latter is often simply referred to as Marden's theorem), which are classical results in the geometry of polynomials. In particular, and germane to what follows, Johnson and Paparella <cit.> proved that if A = FDF^∗, where D = (λ_1,…,λ_n) and F is the n-by-n discrete Fourier transform matrix, then the principal submatrix F(A_(1)), obtained by deleting the first-row and first column of A, is tangent to the midpoints of every side of the polygon ∂ F(A) = ∂(λ_1,…,λ_n). In this work, this result is generalized to the fullest extent possible—in particular, it is shown that if A is convexoid, i.e., F(A) coincides with the convex hull of its eigenvalues, then F(A_(k)) is inscribed in the polygon ∂ F(A), ∀ k ∈{1,…,n}. § NOTATION AND BACKGROUND The set of m-by-n matrices with entries over ℂ is denoted by 𝖬_m × n(ℂ); when m = n, 𝖬_n × n(ℂ) is abbreviated to 𝖬_n. The set of all n-by-1 column vectors is identified with the set of all ordered n-tuples with entries in ℂ and thus denoted by ℂ^n. If x ∈ℂ^n, then x_i denotes the ith entry of x. The n-by-n identity matrix is denoted by I = I_n and e_k denotes the kth column of I. Given A ∈𝖬_n, we let * σ(A) denote the spectrum (i.e., multiset of eigenvalues) of A; * A_(k) denote the (n-1)-by-(n-1) principal submatrix obtained by deleting the kth row and kth column of A); and * A^∗ denotes the conjugate transpose of A. If A ∈𝖬_n and B ∈𝖬_m, then the direct sum of A and B, denoted by A ⊕ B, is defined by A ⊕ B =[ A 0; 0 B ]. If U ∈𝖬_n, then U is called unitary if U^* U = I. If A ∈𝖬_n, then A is called normal if A^* A = A A^*. A matrix A is normal if and only if there is a unitary matrix U and a diagonal matrix D such that A = U D U^* <cit.>. The field (of values) or numerical range of A ∈𝖬_n), denoted by F(A), is defined by F(A) = { x^*A x | x^*x = 1 }⊆ℂ. A general reference for the field is <cit.>. If S = {λ_1,…,λ_n }⊂ℂ (repetitions allowed), then the convex hull of S is denoted by S = (S). The following well-known properties will be useful in the sequel: fvprops If A = [a_ij] ∈𝖬_n and B ∈𝖬_m, then: * σ(A) ⊆ F(A) <cit.>; * F(A) = (σ(A)), whenever A is normal <cit.>; * ds F ( A⊕ B ) = (F(A) ∪ F(B)) <cit.>; * F(A_(k)) ⊆ F(A), ∀ k ∈{1,…, n } <cit.>; and * F(A) is convex <cit.>. For completeness, we give a proof of Property 1.2.11 <cit.>, which generalizes part <ref>, given that it is ubiquitous in the literature; the proof-strategy suggested by Horn and Johnson is tedious; and ideas presented in the demonstration will be used in the sequel. To this end, let α = {α_1, …, α_m } be a nonempty subset of {1, …, n} (if α = ∅, then F(A[α]) = ∅⊆ F(A)) and denote by A[α] the m-by-m matrix whose (i,j) entry is a_α_i, α_j, 1 ≤ i,j ≤ m. If P:= [ e_α_1 ⋯ e_α_m ], then a_α_i, α_j = e_α_i^⊤ A e_α_j = [ P^⊤ A P]_ij, i.e., A[α] = P^⊤ A P. If z ∈ F(A[α]), then z = x^*F(A[α])x, where x^* x = 1. If y:=Px, then y^*y=x^*P^⊤ Px = x^*I_m x= x^*x=1. Furthermore, z = x^* A[α] x = x^* P^⊤ A P x=(Px)^*APx = y^*Ay ∈ F(A), i.e, F(A[α]) ⊆ F(A). If A ∈𝖬_n(ℂ), then A is called convexoid if F(A) = (σ(A)). Johnson <cit.> gave the following characterization of convexoid matrices. convexoid If A ∈𝖬_n(ℂ), then A is convexoid if and only if A is normal or there is a unitary matrix U such that U^∗ A U = [ A_1 0; 0 A_2 ], where A_1 is normal and F(A_2) ⊆ F(A_1). By fvprops(<ref>), if A is normal and λ_1,…,λ_n are its eigenvalues (repetitions included), then F(A) = (λ_1,…,λ_n). Without loss of generality, we may label the vertices as λ_1,…, λ_d, 1 ≤ d ≤ n. Notice that ∂ F(A) = ⋃_k=1^d (λ_k,λ_k+1), where, for convenience, d + 1 := 1. We say that F(A_(k)) is inscribed in F(A) if F(A_(k)) ∩(λ_k,λ_k+1)∅, ∀ k ∈{1,…,d}. § MAIN RESULT rotevec Let Av= λ v, where v ≠ 0 and v_k ≠ 0. If v_k = rexp(i θ), where θ∈ (-π,π], then w:= exp(-iθ)v is an eigenvector such that v _2 = w _2 and w_k > 0. The conclusion that w is an eigenvector with the same length as v follows from the fact that |exp(-iθ)| = 1. Lastly, notice that w_k = exp(-iθ) (rexp(i θ)) = r > 0. projection Let α = {α_1, …, α_m } be a nonempty subset of {1, …, n}, let P:= [ e_α_1 ⋯ e_α_m ], and let y ∈ℂ^n be any vector such that y_k = 0 whenever k ∉α. If x:= P^⊤ y ∈ℂ^m, then y = Px and y^∗ y = x^∗ x. Since PP^⊤ = ∑_k=1^m e_α_ke_α_k^⊤, it follows that Px = P(P^⊤ y) = (P P^⊤) y = ( ∑_k=1^m e_α_ke_α_k^⊤) y = ∑_k=1^m e_α_k y_α_k. If z := ∑_k=1^m e_α_k, then z_k = 0 whenever k ∉α. Thus, y = Px and y^∗ y = (Px)^∗ (Px) = x^∗ P^⊤ P x = x^∗ I_m x = x^∗ x, as desired. main If A is convexoid, then F(A_(k)) is inscribed in the polygon ∂ F(A), ∀ k ∈{1,…,n}. In view of fvprops<ref> and convexoid, it suffices to consider the case when A is normal. To this end, if A is normal, then there is a diagonal matrix D = (λ_1,…,λ_n) and a unitary matrix U such that A = UDU^*. Let λ_i and λ_j be adjacent vertices of the polygon ∂(λ_1,…,λ_n), let v := U e_i, and let w := Ue_j. Note that v^∗ v = w^∗ w = 1 and v^∗ w = w^∗ v = 0. We distinguish the following cases: * v_k = 0 or w_k =0. If γ := {1,…,n }\{k} = {γ_1,…, γ_n-1}, P := [ e_γ_1 ⋯ e_γ_n-1 ], and x := P^⊤ v, then v = Px and x^∗ x = 1 by projection. Thus, λ_i = v^∗ Av = (Px)^∗ A (Px) = x^∗ P^⊤ A P x = x^∗ A_(k) x ∈ F(A_(k)), i.e., F(A_(k)) ∩(λ_i-1,λ_i)∅ and F(A_(k)) ∩(λ_i,λ_i+1)∅ (if i=1, then i-1 := d). Similarly, if w_k = 0, then λ_j ∈ F(A_(k)). If v_k = w_k = 0, then the line-segment (λ_i,λ_j)⊆ F(A_(k)) by fvprops<ref>. * v_k 0 and w_k 0. By rotevec, it may be assumed, without loss of generality, that v_k > 0 and w_k > 0. If α := ∓ w_k/√(v_k^2 + w_k^2) and β := ± v_k/√(v_k^2 + w_k^2), then α and β are nonzero reals such that α^2 + β^2 = 1. If u := α v + β w, then u_k = α v_k + β w_k = ∓ v_k w_k/√(v_k^2 + w_k^2) + ± v_k w_k/√(v_k^2 + w_k^2) = 0 and, since v^∗ v = w^∗ w = 1 and v^∗ w = w^∗ v = 0, it follows that u^∗ u = (α v^∗ + β w^∗)(α v + β w) = α^2v^∗ v + αβ v^∗ w + βα w^∗ v + β^2 w^∗ w = α^2 + β^2 = 1 and u^∗ A u = (α v^∗ + β w^∗) (α A v + β Aw) = (α v^* + β w^*)(αλ_i v + βλ_j w) = α^2 λ_i v^∗ v + αβ v^∗ w + βα w^∗ v + β^2 λ_jw^∗ w = α^2 λ_i + β^2 λ_j. Because α^2 + β^2 = 1, α^2 > 0, and β^2 > 0, it follows that u^∗ A u lies in the interior of the line-segment (λ_i,λ_j). Finally, if x := P^⊤ u ∈ℂ^n-1, where γ and P are defined as in (<ref>) and (<ref>), respectively, then u = Px and x^∗ x = 1 (by projection) and α^2 λ_i + β^2 λ_j = u^∗ A u = x^∗ P^⊤ A P x = x^∗ A_(k) x ∈ F(A_(k)). Thus, F(A_(k)) ∩(λ_i,λ_j)∅. In all cases, F(A_(k)) ∩(λ_k,λ_k+1)∅, ∀ k ∈{1,…,d}. Although there are two possible choices for α and β, the convex combination α^2 + β^2 = 1 is unique. Therefore, in the case that v_k 0 and w_k 0, F(A_(k)) intersects the interior of the line segment (λ_i,λ_j) at a single point. If U = [ 1/2 √(3)/2 0 0; 1/2 -√(3)/6 0 √(6)/3; 1/2 -√(3)/6 √(2)/2 -√(6)/6; 1/2 -√(3)/6 -√(2)/2 -√(6)/6; ], then U^⊤ U = I_4. If A = UDU^⊤, where D = (-1-5i,-2,3-2i,2+5i), then A is normal. Figure <ref> illustrates main. dftcor If λ_1,…,λ_n are complex numbers, F denotes the n-by-n discrete Fourier transform matrix, D=(λ_1,…,λ_n), and A = FDF^∗, then F(A_(k)) is inscribed in the polygon ∂(λ_1,…,λ_n) and the points of tangency occur at the midpoints of sides of ∂(λ_1,…,λ_n). Since f_ij = ω^(i-1)(j-1)/√(n), where ω := exp(-2 π i/n), it follows that α = -1/√(n)/√(1/n + 1/n) = -1/√(n)/√(2)/√(n) = -1/√(2), β = 1/√(n)/√(1/n + 1/n) = 1/√(n)/√(2)/√(n) = 1/√(2), and α^2 = β^2 = 1/2, where α and β are defined as in (<ref>). § ACKNOWLEDGEMENTS The authors thank the National Science Foundation for funding and the University of Washington Bothell for hosting REU Site: Tiling Theory, Knot Theory, Optimization, Matrix Analysis, and Image Reconstruction. In addition, we thank Casey Mann and Milagros Loreto for their efforts. abbrv

http://arxiv.org/abs/2303.06950v1

20230313093519

Reconfigurable Distributed Antennas and Reflecting Surface (RDARS): A New Architecture for Wireless Communications

cs.IT

Reconfigurable Distributed Antennas and Reflecting Surface (RDARS): A New Architecture for Wireless Communications Chengzhi Ma, Xi Yang, Jintao Wang, Guanghua Yang, and Shaodan Ma, Senior Member, IEEE C. Ma, J. Wang and S. Ma are with the State Key Laboratory of Internet of Things for Smart City and the Department of Electrical and Computer Engineering, University of Macau, Macao SAR, China (e-mails: yc07499@um.edu.mo; wang.jintao@connect.um.edu.mo; shaodanma@um.edu.mo). X. Yang is with the Shanghai Key Laboratory of Multidimensional Information Processing, East China Normal University, Shanghai 200241, China (e-mail: xyang@cee.ecnu.edu.cn). G. Yang is with the School of Intelligent Systems Science and Engineering and GBA and B&R International Joint Research Center for Smart Logistics, Jinan University, Zhuhai 519070, China (e-mail: ghyang@jnu.edu.cn). ============================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================ Reconfigurable intelligent surfaces (RISs) have drawn much attention recently for their appealing advantages in shaping wireless channels to improve the spectral and energy efficiencies of wireless communications. However, conventional fully-passive RISs generally suffer from the so-called “multiplicative fading” effect and necessitate a large number of passive reflecting units to provide satisfying performance, which thereby limits RISs' practicability and manufacturability. In this paper, a novel architecture of “Reconfigurable Distributed Antennas and Reflecting Surfaces (RDARS)” is first proposed to overcome this limitation from the “multiplicative fading” effect. Specifically, unlike existing active RIS variants, RDARS inherits the low-cost and low-energy-consumption benefits of fully-passive RISs by default configuring all the elements as passive to perform the reflection mode. On the other hand, based on the design of the additional direct-through state, any element of the RDARS can be dynamically programmed to connect with the base station (BS) via fibers and perform the connected mode as remote distributed antennas of the BS to receive signals. Consequently, a controllable trade-off between the reflection gain and the distribution gain can be achieved via RDARS at the BS. To unveil the system behavior of the RDARS-aided system, we analyze the received signal-to-noise ratio (SNR) under maximum ratio combining (MRC) at BS. Closed-form outage probability and ergodic achievable rate are also provided and are verified through extensive simulations. To demonstrate the superiority of the proposed RDARS, we further compare the performance of RDARS with the fully-passive RIS-aided systems under different system settings, and experiments are carried out using a prototype of RDARS with a total number of 256 elements. Experiment results revealed that, compared to the fully-passive RIS, an extra 76% throughput improvement could be achieved by deploying RDARS with only three elements performing connected mode. This thus confirms the effectiveness of the proposed RDARS and envisions it as a promising candidate for future 6G wireless systems. Reconfigurable distributed antennas and reflecting surfaces (RDARS), reconfigurable intelligent surfaces (RIS), distributed antenna systems (DAS), outage probability, ergodic achievable rate. § INTRODUCTION The development of 5G cellular systems accelerates the invention of new technologies to meet high-demand communications needs, including massive multiple-input multiple-out (massive MIMO), ultra-dense network (UDN), and millimeter wave (mmWave) communication, to name a few<cit.>. The success of 5G commercial deployment has motivated both industry and academia to look for new disruptive technologies to achieve unprecedented requirements aiming at ubiquitous coverage, ultra high-rate throughput, and so on <cit.>. To accomplish these ambitious goals, reconfigurable intelligent surface (RIS), also known as intelligent reflecting surface (IRS), has been proposed as a promising technology to extend the coverage while maintaining a low cost and energy consumption for improving spectrum efficiency (SE) and energy efficiency (EE) of the communication network <cit.>. Specifically, RIS is a planar array that is composed of a large number of passive elements, where each of them imposes an independent phase shift on the incident signals. By smartly tuning the phase shifts, the wireless propagation environment becomes controllable and programmable, which thus provides an extra degree of freedom to optimize communication performance. Thanks to the above advantages, RIS has been widely investigated w.r.t. its fundamental performance limits <cit.>, channel modeling <cit.> as well as prototype implementation <cit.>. In particular, in <cit.>, the joint active and passive beamforming design for a RIS-aided system is proposed, and the asymptotic performance of RIS passive beamforming with an infinitely large number of reflecting elements is presented, where the "squared power gain" brought by deploying RIS is unveiled. Also, it is shown that even with discrete phase shifts, through proper optimization, the RIS is still capable of boosting system performance<cit.>. Moreover, as demonstrated in <cit.>, 26 dB power gain is introduced by the RIS prototype with 1100 elements working at 5.8 GHz. Though RIS has revealed its ability to extend the coverage area by creating an extra link between the source-destination when the direct link is blocked. A large number of reflecting elements is generally needed for RIS to achieve the same performance even with ideal phase-shifting as compared to other technologies, e.g., decode-and-forward (DF) relaying <cit.>. The reason lies in the fact that the channel gain of the RIS reflective link suffers from the “multiplicative fading” effect, i.e., the equivalent channel gain of the transmitter-RIS-receiver link is the product of the transmitter-RIS and RIS-receiver channel gains <cit.>. Therefore, the path loss over the RIS reflective link may be high due to the "multiplicative fading", which limits the practical contribution of RIS. Moreover, the fully passive nature of conventional RIS imposes challenges in channel estimations since the number of channel coefficients to be estimated scales linearly with the product of the numbers of antennas at the BS, UE and the elements on RIS <cit.>. Even with the proper design of the channel estimation schemes and leveraging the channel characteristics, e.g., sparsity, the pilot overhead is still large and becomes unaffordable when deploying a large number of passive elements for performance enhancement<cit.>. To overcome these challenges, several variants of RIS architectures are proposed, namely, active RIS, hybrid relay-reflecting intelligent surface (HR-RIS), and so on <cit.>. In particular, different from fully-passive RIS, the elements of active RIS are integrated with amplifiers to amplify the incident signal. Instead of enabling all the elements with the capability to amplify the signal, the HR-RIS proposed replacing only a few passive elements with active ones to relay the signal. To some extent, these variants can partially compensate for the “multiplicative fading” effect by introducing active elements for amplifying or relaying. However, they also introduce additional power consumption and require a much more complicated implementation of the surface. Moreover, the practicability and manufacturability of these variants are still debatable. To amend the aforementioned limitations of fully-passive RIS while maintaining its appealing advantages, we start to think outside the box and revisit the successful technologies proposed in communication society in the past to gain inspiration. While RIS is recognized as a promising technology to combat signal blockages and extend the coverage area, the effort toward dealing with coverage problems is not new. Among all of them, one of the successful and revolutionary examples is the Distributed Antenna System (DAS) <cit.>. In DAS, instead of deploying co-located antennas in the base station (BS), antenna modules are geographically distributed and connected to a central processor by fiber or coaxial cable.[Here, antenna modules refer to entities with sophisticated signal processing capabilities and energy-demanding components, e.g., RF chains and power amplifiers. ] In this way, DAS enables the efficient utilization of space resources by potentially shorting the distance between BS and users and also providing extra spatial diversity. Though the original DAS is introduced to cover the blind spots in wireless communications <cit.>, studies have also identified its potential in improving system energy efficiency and system capacity <cit.>. However, to acquire a satisfactory distributed antenna gain with DAS, many distributed antenna modules should be well allocated in the entire cell, and large antenna arrays are preferred to be deployed in each module to obtain sufficient antenna gain and alleviate interference. This undoubtedly introduces tremendous hardware cost and energy consumption, thus limiting its applications. Motivated by DAS, a natural question arises: Can we introduce the concepts of DAS into the current passive RIS architecture to boost the system performance while still maintaining RIS's low-cost and low-energy consumption advantages? In this paper, we propose, for the first time, a novel “Reconfigurable Distributed Antennas and Reflecting Surfaces (RDARS),” and present both theoretical analysis and experiment results to answer this question. The main contributions of this paper can be summarized as follows: * We propose a novel RDARS architecture, which preserves the appealing advantages of both DAS and the fully-passive RIS. Specifically, each element of RDARS can be dynamically programmed to perform either two modes, i.e., reflect signals as a conventional fully-passive RIS element under reflection mode or perform as a remote antenna for BS to receive signals under connected mode. By introducing the connected mode, the “multiplicative fading” effect in conventional RIS-aided systems can be well compensated without extra power and hardware complexity for amplifying or relaying at the surface. Moreover, a controllable trade-off between the reflection gain and the distribution gain can be achieved via RDARS at the BS. * We analyze the performance of the RDARS-aided system under MRC. Closed-form expressions for bounds of ergodic achievable rate and outage probability are provided. Besides, by setting all elements at RDARS as reflection mode or connected mode, existing results for fully-passive RIS-aided systems or DAS consisting of two distributed antenna sets can be regarded as special cases of our proposed RDARS-aided system. * We investigate the impacts of the number of elements performing different modes and the deployment location of RDARS on its performance. Furthermore, to unveil the superiority of the proposed RDARS, we compare the outage probability and the ergodic achievable rate of RDARS with the fully-passive RIS under various scenarios via extensive simulation results. * We conduct practical experiments to compare the performance of the proposed RDARS and the passive RIS (with all the elements on RDARS performing reflection mode) by establishing an RDARS-aided communication system prototype. The experiment results demonstrate the superiority of the proposed RDARS in terms of performance gain and manufacturability. The rest of the paper is orginized as follows. In Section II and In Section III, we propose the idea of the RDARS architecture and the system model of the RDARS-aided system. Then the performance of RDARS-aided system is analyzed in Section IV. Simulation results are given in Section V, followed by experimental results in Section VI. Further discussion and conclusion are drawn in Section VII and VIII. Notations: Throughout the paper, numbers, vectors, and matrices are denoted by lower-case, bold-face lower-case, and bold-face upper-case letters, respectively. (·)^T, (·)^H denote the transpose and conjugate transpose of a matrix or vector. diag(𝐯) denotes a diagonal matrix with each diagonal element being the corresponding element in 𝐯. Furthermore, |·| denotes modulus of complex number, and 𝔼[·] represent the expected value of a random variable. § PROPOSED RDARS The architecture of the RDARS-aided system is presented in Fig. 1. The total number of elements in RDARS is N, and each element can switch between two modes, namely, the reflection mode and the connected mode. The RDARS controller can dynamically configure the specific mode of each element. When working under reflection mode, the elements work as passive reflecting units similar to the elements in fully-passive RIS. When working under connected mode, the elements act as a remote antenna connected to the BS via dedicated wires or fibers and can receive the incoming wireless signal. Denote the number of elements working on connected mode as a. Then, for a = 0, the RDARS-aided system becomes a fully-passive RIS-aided system; for a = N, the RDARS-aided system can be interpreted as a DAS which has two distributed antenna sets, and one set has N antennas located at a remote place while the other set of antennas is deployed at the BS <cit.>. To maintain the appealing advantages of fully-passive RIS, we only consider the case where the number of elements performing connected mode is very small, i.e., a << N. To better characterize the configuration of this architecture, we introduce a diagonal indicating matrix 𝐀, i.e., 𝐀 is a diagonal matrix with element 𝐀(i,i) ∈{0,1}. Specifically, 𝐀(i,i) = 1 means the i^th element works at connected mode, while 𝐀(j,j) = 0 indicates that the j^th element operates at reflection mode, i.e., the j^th element can only alter the phase of incoming signals. Let 𝐯 = [v_1, ⋯, v_N] denote the RDARS reflection-coefficient vector where v_i = e^jθ_i, ∀ i, and θ_i is the phase shift the i^th element will induce. As such, the equivalent RDARS reflection-coefficient matrix can be modeled as (𝐈-𝐀) diag(𝐯). Unlike active RIS <cit.> or HR-RIS <cit.>, which requires additional reflection-type amplifiers for elements, the realization of the element of RDARS can be built merely upon phase-shift circuits with a slight modification without much additional hardware complexity. However, the detailed implementation of RDARS is out of the scope of this paper. We focus on presenting the idea of the RDARS architecture, analyzing the performance, and further demonstrating its superiority with simulation and experiment results. § SYSTEM MODEL In this paper, we consider RDARS-aided uplink communications where an RDARS is deployed to assist the communications from the user equipment (UE) to the BS. To gain more insight into the system performance of this novel architecture, we consider single-antenna UE and single-antenna BS. The channels between UE-BS, RDARS-BS, and UE-RDARS are all i.i.d. Rayleigh channel, i.e., h_UB∼𝒞𝒩(0, ρ_UB^2) ∈ℂ, 𝐡_RB∼𝒞𝒩(0,ρ_RB^2𝐈_N) ∈ℂ^N× 1 and 𝐡_UR∼𝒞𝒩(0, ρ_UR^2𝐈_N) ∈ℂ^N× 1, where 𝒞𝒩(·,·) denote the complex univariate or multivariate Gaussian distribution. §.§ Signal Model Armed with the notation of indicating matrix 𝐀, the received signal y ∈ℂ from the single antenna of BS and the received signal 𝐮∈ℂ^N × 1 from the a elements acting with connected mode at the RDARS can be expressed as[Here we only consider single-bounce signal reflected by the passive reflecting units since the multi-bounce signal has much small power and thus is omitted here <cit.>.] [ y^T𝐮^T ]^T_∈ℂ^(1+N) × 1 = [ h_UB + 𝐡_RB^H𝐁𝐡_UR ^ ∈ℂ^1 × 1; 𝐀𝐡_UR_∈ℂ^N × 1 ]_∈ℂ^(1+N) × 1x_∈ℂ + [ n_BS; 𝐀𝐧_R ]_∈ℂ^(1+N) × 1 = 𝐡 x + 𝐧 where x ∈ℂ is the signal transmitted from the UE to BS with total power 𝔼[|x|^2] ≤ P. n_BS∼𝒞𝒩(0, σ_BS^2) is the additive white Gaussian noise (AWGN) received at BS. 𝐀𝐧_R is the equivalent noise vector, i.e., [𝐀𝐧_R]_i ∼𝒞𝒩(0, σ_R^2), ∀𝐀(i,i) ≠ 0. It should be noted that if the 𝐀(i,i) = 0, then [𝐀𝐧_R]_i = 0. Since the received signal is jointly processed at BS and advanced technology can be applied to pre-process signal, we assume the noises at RDARS and BS have equal power after some basic pre-processing, i.e., σ_BS = σ_R = σ. Also, 𝐁 in (<ref>) is defined as 𝐁 = (𝐈 - 𝐀) diag(𝐯) to simplify the notation. By defining 𝐡 as the equivalent channel as above, the whole system can be regarded as a special SIMO system. To effectively combine the received signal, the maximum-ratio-combining (MRC), which is a low-complexity combining scheme, is adopted in this paper. As such, the signal after MRC is written as 𝐡^H[ y^T 𝐮^T ]^T_∈ℂ^(1+N) × 1 = 𝐡^H𝐡 x + 𝐡^H𝐧 . Accordingly, with some simple algebra manipulation, the received SNR γ is given by γ = γ ( | h_UB + 𝐡_RB^H𝐁𝐡_UR |^2 + 𝐡_UR^H𝐀^H𝐀𝐡_UR ) , where γ = P/σ^2 is the transmit SNR. It is clear that the indicator matrix 𝐀 and the reflection coefficient vector 𝐯 affect the value of γ. Some insightful observations can be found by considering different system parameter configurations as follows: * When a = 0 (or 𝐀 = 0_N), γ becomes the received SNR of a fully-passive RIS-aided SISO system with a direct link <cit.>. * When a = N (or 𝐀 = 𝐈_N), γ is the received SNR of a DAS with two distributed antenna sets where one set is the home-based station with one antenna and the other deployed at the remote RDARS consists of N co-located antennas <cit.>. The optimal phase shift for the passive reflecting units with perfect channel state information (CSI) at BS in the above system setup can be obtained in closed-form as: θ_i = arg (h_UB) - arg( h_RB,i h_UR,i ) , ∀𝐀(i,i) = 0, where v_i = exp(jθ_i), and arg(·) retrieves the phase of the input complex number. Also, h_RB,i and h_UR,i are the i^th element of the channel vectors 𝐡_RB and 𝐡_UR, respectively. The proof can be obtained by leveraging the triangle inequality similar to <cit.> thus omitted here for brevity. It is clear that the optimal phase shift design for elements performing reflection mode is the same as the traditional RIS system<cit.>. This indicates that the optimal phase shift design for the above RDARS-aided system does not require additional complexity in phase shift design in this system setup. With optimal phase shift, the received SNR can be further formulated as γ = γ [ (|h_UB| + ∑_i=1^N(1-a_i)|h_RB,i||h_UR,i| )^2 + 𝐡_UR^H𝐀^H𝐀𝐡_UR ) ], where a_i = 𝐀(i,i). Here we define (|h_UB| + |𝐡_RB^H𝐁𝐡_UR| )^2 as the reflection gain provided by elements in RDARS performing reflection mode similar to the conventional fully-passive RIS-aided system, while 𝐡_UR^H𝐀^H𝐀𝐡_UR as the distribution gain provided by the elements in RDARS working under connected mode. It can be deduced that the reflection gain of RDARS should be smaller or equal (when a = 0) to the conventional fully-passive RIS-aided systems. However, with the introduced distribution gain, the proposed RDARS is anticipated to greatly outperform the corresponding fully-passive RIS-aided systems. To fully uncover the benefits of the proposed RDARS, we conduct the performance analysis by first analyzing the distribution of γ, comparing the received signal power of RDARS-aided systems with that of traditional fully-passive RIS-aided systems, and deriving closed-form expressions for ergodic achievable rate and outage probability in the following section. § PERFORMANCE ANALYSIS We assume the BS has perfect CSI and the passive reflecting units on RDARS are able to alter the signal with continuous phase shifts. Hence, the results obtained under these assumptions serve as performance upper bounds. To fully unveil the potential of RDARS, we consider the optimal phase shifts configuration for elements performing reflection mode, i.e., θ_i = arg (h_UB) - arg( h_RB,i h_UR,i ) , ∀𝐀(i,i) = 0, and then derive the distribution of γ for a given indicating matrix 𝐀.[ For simplicity, we consider given indicating matrix 𝐀 for analysis. However, the design of 𝐀 provides an extra degree of freedom in optimization for RDARS-aided systems and will be our future research direction.] It can be seen from (<ref>) that γ involves the sum of random variables (RVs). In particular, the reflection gain in γ provided by the passive reflecting units and the direct link involves the square of the sum of RVs, i.e., (|h_UB| + ∑_i=1^N(1-a_i)|h_RB,i||h_UR,i| )^2. Therefore, obtaining the exact distribution of γ involves multiple integrals, convolution, and transformation of RVs and thus makes γ intractable. To obtain more insight into γ and acquire a tractable analytical expression for the distribution of γ to reveal the system behavior, we resort to moment matching based approximation. Specifically, by approximating γ as a Gamma RV via matching their first and second moments, i.e., 𝔼[γ] and 𝔼[γ^2], the result is formalized in the following Theorem. With the optimal phase shift configuration, the distribution of γ given an indicator matrix 𝐀 can be approximated as γ approx∼Γ(k, p ), with shape and scale parameters obtained as k = 𝔼[γ]^2/𝔼[γ^2] - 𝔼[γ]^2, p = 𝔼[γ^2] - 𝔼[γ]^2/𝔼[γ], where the first and second moments of γ, i.e., 𝔼[γ] and 𝔼[γ^2], can be evaluated as 𝔼[γ] = γ[ N^2 ( π^2/16ρ_UI^2ρ_IB^2 ) + ρ_UI^2ρ_IB^2 ( N(1 - π^2/8a^2 - π^2/16) + π^2/16(a^2 + a - 1) ) + ρ_UBρ_UIρ_IB( π/4√(π)(N-a) ) + aρ_UI^2 + ρ_UB^2 ], 𝔼[γ^2] = γ^2 𝔼[ ( (γ_1 + γ_2)^2 + γ_3)^2 ], = γ [2ρ_UB^4 + ρ_UB^2 *((N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2) + 2* 3/4√(π)ρ_UB^3 *((N-a) π/4ρ_URρ_RB) + ρ_UB^2 *((N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2) + C_4 ρ_UR^4ρ_RB^4 + 2*1/2√(π)ρ_UB *(C_3 ρ_UR^3 ρ_RB^3) + 2*3/4√(π)ρ_UB^3 * ((N-a) π/4ρ_URρ_RB) + 2*1/2√(π)ρ_UB*(C_3 ρ_UR^3 ρ_RB^3) + 4*ρ_UB^2* ((N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2) + 2*ρ_UB^2 *a ρ_UR^2 + 2 ((N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2) a ρ_UR^2 + 4 1/2√(π)ρ_UB (N-a) π/4ρ_URρ_RB a ρ_UR^2 + a(a+1) ρ_UR^4 ]. The proof can be found in Appendix A. The Gamma moments matching is shown to be accurate by well capturing the probability mass of γ. This has been validated by the simulation results in Section 5. Moreover, the average received power after MRC combining P_RDARS can be obtained as P_RDARS = P [ ρ_UB^2 + (N-a) π/4√(π)ρ_UBρ_URρ_RB + (N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2 + aρ_UR^2 ]. Note that P_RDARS covers the existing results for the conventional RIS-aided SISO systems (a=0) with or without (ρ_UB^2 ≠ 0 or ρ_UB^2 = 0) direct link as special cases <cit.>. It can also be seen that P_RDARS is affected by the large-scale channel parameters as well as the total number of elements N and the number of elements performing connected mode a. As such, to unveil the benefits of the proposed RDARS, we compare the average received power of the RDARS-aided system with that of the conventional RIS-aided system and determine the condition under which the RDARS-aided system performs better than the conventional RIS-aided system in terms of the average received power in the following theorem. Regardless of the number of elements performing connected mode a (∀ a ≤ N and a>0), the RDARS-aided system will provide higher average received power than the corresponding conventional RIS-aided system if the following condition is satisfied: N < 16/π^2(1/ρ_RB^2 -1) + 1 - 4√(π)/πρ_UB/ρ_RBρ_UR . The received powers of the RDARS-aided system and the RIS-aided system are obtained as P_RDARS = P [ ρ_UB^2 + (N-a) π/4√(π)ρ_UBρ_URρ_RB + (N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2 + aρ_UR^2 ], P_RIS = P [ ρ_UB^2 + Nπ/4√(π)ρ_UBρ_URρ_RB + (N)(1+ π^2/16(N-1)) ρ_UR^2ρ_RB^2]. Let P_RDARS> P_RIS, we have N < 8/π^2(1/ρ_RB^2 -1) + a+1/2 - 2√(π)/πρ_UB/ρ_RBρ_UR. Since a ≤ N, the right-hand-side (RHS) of (<ref>) should be greater than a. As a result, we have the condition for a as a < 16/π^2(1/ρ_RB^2 -1) + 1 - 4√(π)/πρ_UB/ρ_RBρ_UR. Plugging (<ref>) into the RHS of (<ref>), we have (<ref>). Thus we complete the proof. Moreover, when the direct link between the user and BS is blocked (i.e., ρ_UB^2 = 0), we have the following corollary. The RDARS-aided system (without a direct link) will provide higher average received power as compared to the corresponding conventional RIS-aided system (without a direct link), regardless of the number of elements performing connected mode a (∀ a ≤ N and a>0), if the following condition of the total number of elements N is satisfied: N < 16/π^2(1/ρ_RB^2 -1) + 1 . By setting ρ_UB^2 = 0 in (<ref>), the result is obtained. From Theorem <ref> and Corollary <ref>, the condition that the RDARS-aided system outperforms the corresponding conventional RIS-aided system involves an upper bound of N for any given a. This can be interpreted by comparing the reflection gain and distribution gain: When N is extremely large, the reflection gain will dominate for a given transmit power P. Since the reflection gain of the RDARS-aided system is less than the RIS-aided system given an exact total number of N, the corresponding conventional RIS-aided system will perform slightly better than the RDARS-aided system. However, for a practical number of N, the reflection gain is much smaller due to the “multiplicative fading” effect and thus becomes subordinate when a direct link exists <cit.>. In addition, since there is an extra distribution gain in the RDARS-aided system, a point where the average received powers of these two systems are equal to each other should exist. This thus leads to the above upper bounds. Moreover, it should be noted that the average received power of the RDARS-aided system is always higher or equal (when a=N) to DAS which has no aid of passive reflecting units. This can be easily proved by inspecting (<ref>) with ρ_RB=0 and finding that the average received power of the RDARS-aided system has more positive terms than that of the DAS (under the assumption of a ≤ N). §.§ Outage probability Since the SNR γ is approximated as Gamma RV, the closed-form outage probability, which is defined as P_o(x) = Pr(γ < x), can be obtained as P_o(x) = γ(k,x/p)/Γ(k) = (x/p)^k/kΓ(k)_1F_1 (k,k+1,-x/p) , where _1F_1(·,·,·) is the confluent hypergeometric function of the first kind <cit.>. §.§ Ergodic Achievable Rate Here we present the closed-form ergodic achievable rate and its upper bounds to provide insights. §.§.§ Gamma Moment Matching Since γ is approximated as a Gamma RV, the closed-form ergodic achievable rate can be obtained as R = 𝔼[(log_2(1+γ))] = 1/Γ(k)ln(2) H_3, 2^1, 3[ p | (1,1) , (1,1) , (-k+1,1) (1,1) , (0,1) ] , where H(z) = H_p, q^m, n[ p | (a_i,A_i)_1,p (b_i,B_i)_1,q], is the Fox's H-function (FHF) defined via the Mellin-Barnes type integral <cit.>. Define 1/pγ = y, we have R = ∫_0^∞log_2(1+γ) f_γ(γ) dγ, = 1/Γ(k)∫_0^∞ y^k-2log_2(1+py) y exp(-y) dy . Then by expressing log_2(1+py) and y exp(-y) as FHF, we have R = 1/Γ(k) ln(2) ∫_0^∞ y^k-2 H_2, 2^1, 2[ py | (1,1) (1,1) (1,1) (0,1) ] H_0, 1^1, 0[ y | – (1,1) ] dy. By using the integral involving the product of two FHFs from <cit.>, we obtain (<ref>). §.§.§ Upper bound Though the ergodic achievable rate using (<ref>) is accurate which has been verified by the simulation results, calculating the FHF still involves high computational complexity and also hinders the extraction of meaningful insights. As such, we derive the upper bound for the ergodic achievable rate by using Jensen's inequalities as follows: 𝔼[(log_2(1+γ))] ≤ log_2(1+𝔼[γ]) = R^U. Hence, by applying the first and second moments of γ given by (<ref>) and (<ref>), the upper bound R^U obtained as R^U = log_2 [ 1 + N^2 ( π^2/16ρ_UI^2ρ_IB^2 ) + ρ_UI^2ρ_IB^2 ( N(1 - π^2/8a^2 - π^2/16) + π^2/16(a^2 + a - 1) ) + ρ_UBρ_UIρ_IB( π/4√(π)(N-a) ) + aρ_UI^2 + ρ_UB^2 ]. It can be seen that the upper bound R^U in (<ref>) is scaling up with N^2. Though this is obtained for a limited number of N, the result coincides with the power scaling law for conventional RIS <cit.>. This indicates that the RDARS-aided system (using MRC) and the corresponding RIS-aided system perform similarly in the asymptotic regime (with optimal phase shifts), i.e., N →∞. However, as compared to the recently proposed active RIS-aided system <cit.>[The performance of HR-RIS has not been thoroughly investigated yet <cit.>. Therefore, we only mention active RIS as a representative variant.], where the SNR only scales linearly with N in the asymptotic regime due to the additional noises introduced by the use of active components, the RDARS-aided system is still scaling up with N^2 without additional power for amplification. Moreover, as shown in Theorem <ref> and the simulation results provided in Section 5, the RDARS-aided system will perform better than the RIS-aided system for the practical number of N by compensating the “multiplicative fading” effect with distribution gain provided by the elements performing connected mode. These emphasize the practical application value of the proposed RDARS architecture. § SIMULATION RESULTS In this section, numerical results are presented to validate the effectiveness of the proposed architecture and the correctness of the corresponding analysis. To fully unveil the superiority of the proposed architecture, the optimal phase shift as in (<ref>) for passive reflecting units is adopted for both RDARS and RIS-aided systems. The system layout used in this paper is similar to <cit.>, i.e., the BS and the UE are located at points (0,0) and (90,0), respectively, as shown in Fig. 2. The parameter h is chosen to be 10 m (if not specified otherwise), whereas d is varied across simulations. The distance-dependent path loss model is considered and is given by L(d) = C_0 (d/D_0)^-α, where C_0 is the path loss at the reference distance D_0 = 1 m, d denotes the link distance, and α represents the path loss exponent. The path loss exponent of UE-RDARS(RIS) is set to be α_UR = 2.1 while the path loss exponent of RDARS(RIS)-BS will be specified later to study its effects on the system performance. Due to the relatively large distance between the UE-BS link, we set the path loss exponent α_UB = 3.2. All the theoretical results are verified via Monte-Carlo simulations, with each running 30000 channel realizations. Furthermore, we set γ = 60 dB (or P = -20 dBm and σ^2 = -80 dBm when comparing the average received power) and C_0 = -30 dB unless specified. In the following, we first verify the correctness of the analysis in Section IV and then study the performance of the proposed RDARS-aided system under various system setups. §.§ Verification of Analysis In this subsection, we first validate the accuracy of the proposed approximation and bounds by comparing the result of Theorem 2 in (<ref>), the outage probability in (<ref>), and the ergodic achievable rate in (<ref>) and (<ref>) with the Monte-Carlo simulation results. Note that we set d = 20 m and α_RB = 2.2, while the parameters N and a are specified in each figure. In Fig. 3, the PDFs of γ using Monte-Carlo simulations and the proposed analytical approximation are presented. It is clear that the approximation proposed in Theorem 1 is valid for different settings of a. Also, the proposed approximation becomes more accurate with an increasing number of a as shown in Fig. 3.(a)-Fig. 3.(d) (and Fig. 3.(e)-Fig. 3.(h)). This is because the distribution gain becomes dominant with the increasing a. Next, in Fig. 4, the outage probability versus the outage threshold with different system settings is presented. Consistently to Fig. 3, the analytical approximation becomes more and more accurate with the increase of a under various numbers of N. Though some mismatches between the approximation and the simulated results exist in the tail of the outage probability, the derived approximation still provides a good fit by capturing most of the probability mass of γ <cit.>. Finally, in Fig. 5, the ergodic achievable rate versus the total number of elements N (Fig. 5.(a), Fig. 5.(b)) and transmit power (Fig. 5.(c), Fig. 5.(d)) with the different number of a are presented. For thorough comparison, the lower bound based on the method from <cit.> is also provided. It is observed that the proposed ergodic achievable rate approximation (<ref>) matches well with the simulation results under various settings for both scenarios with and without a direct link. Moreover, the lower <cit.> and the upper bounds (<ref>) become tighter as a increases. §.§ Comparison between RDARS and RIS In this subsection, we first validate the proposed theorem and corollary in Section IV, which determine the condition when RDARS outperforms the conventional RIS in different system setups. Then, we investigate the impact of the deployment location of RDARS and compare it with the conventional fully-passive RIS-aided system. Finally, the outage probability and ergodic achievable rate achieved by the RDARS-aided system are compared with that of the corresponding RIS-aided system to unveil the superiority of the proposed RDARS. §.§.§ Is RDARS promising? First, we verify the condition presented in Theorem <ref>. Since (<ref>) is valid regardless of the number of elements performing connected mode, we choose two representative scenarios, i.e., a = 1 and a = N, to verify Theorem <ref>. Intuitively, using a large value of d generally results in small ρ_RB, and thus the condition of N in (<ref>) becomes large. To ease the computation burden, we set d = 10 m, h = 10 m, and ρ_RB = 2 in Fig. 6. To highlight the correctness of our analysis, instead of calculating the difference of average received powers between the RDARS-aided system and the conventional RIS-aided system, we plot the difference of the average SNR[Here, the average SNR refers to 𝔼[γ] where the expectation is taken w.r.t. different channel realizations. Therefore, we have 𝔼[γ_RDARS] = P_RDARS/σ^2 which does not affect our analysis.], i.e., 𝔼[γ_RDARS] - 𝔼[γ_RIS] versus the total number of N for a = 1 and a = N as shown in Fig. 6. It can be seen that the difference of the average SNR between the RDARS-aided and the RIS-aided systems is positively large for small N and decreases with the number of N. The SNR difference approaches zero with N ≈ 4800 when a = 1 and N ≈ 9700 when a = N. This coincides with the results obtained by directly calculating (<ref>) and (<ref>), which are 9713 and 4857 as presented in TABLE 1 and TABLE 2. Moreover, since the large-scale channel coefficient is affected by the distance d and the path loss exponent α_RB, the condition of N for the required number of elements of RIS to outperform RDARS in terms of average received power is also influenced by d and α_RB. For fixed a, larger values of d and α_RB lead to larger 1/ρ_RB^2 and consequently make 1/ρ_RB^2 dominate in (<ref>) and (<ref>), thus calling for more elements in RIS to achieve comparable performance to RDARS. Also, for fixed d and α_RB (or equivalently, fixed ρ_RB), the condition of N only increases linearly with a, which thus confirms the correctness of (<ref>). Notably, as shown in TABLEs 1 and 2, the required N is generally large, e.g., N≥ 4000. Thus we can conclude that for most of the practical number of N, deploying RDARS will provide more benefits than the conventional RIS-aided system in terms of the average received power. In Fig. 7, the average received power for the RDARS(RIS)-aided system versus BS-RDARS(RIS) horizontal distance is presented with N = 256. It is obvious that the RDARS-aided system significantly outperforms the RIS-aided system with different d in the above system setup. Note that the average received power of the RDARS-aided and the RIS-aided systems varies differently with the BS-RDARS(RIS) horizontal distance when N and a are fixed. For example, defining P_RDARS(d) ( or P_RIS(d) ) as the average received power for RDARS-aided system (or RIS-aided system) along d, we have P_RDARS(30) < P_RDARS(60) and P_RIS(30) ≈ P_RIS(60) when the other parameters are fixed. Research in <cit.> has pointed out that deployment of RIS generally should be either near the BS or UE to boost the system performance, and the average received power provided by RIS is almost symmetrical about the midpoint of the position of UE and BS when considering single-user scenarios as in <cit.>. However, as demonstrated in Fig. 7, for a moderate size of the surface (i.e., having a limited number of total elements, e.g., N = 256 in our setting), a large performance gain can be achieved by deploying RDARS near the UE via the obtained distribution gain. Nonetheless, if d is extremely large (i.e., RDARS is far from the receiver (BS) ), the cost for connecting the elements performing connected mode with BS may be unaffordable. Fortunately, as shown in Fig. 7, deploying RDARS near the BS (or receiver side) still seems like a promising choice as it also provides substantial gain over conventional RIS-aided systems and has an acceptable cost for connecting elements performing connected mode on RDARS at the meantime. §.§.§ Outage Probability Here, we investigate the outage probability of the RDARS-aided and the RIS-aided systems under different RDARS(RIS) locations. We set d = 30 m and d =45 m in Fig. 8.(a) and Fig. 8.(b), respectively. As can be seen from Fig. 8.(a) and Fig. 8.(b), for fixed a, the SNR to achieve the same outage probability of the RDARS-aided system decreases with the increasing N thanks to the increasing reflection gain brought by the passive units of RDARS. Besides, when comparing the RDARS-aided and the conventional RIS-aided systems with N = 256 in Fig. 8, a substantial SNR improvement can be achieved with only one element on RDARS performing connected mode, i.e., the 4 dB and 8 dB SNR gains are attained at the outage probability of 0.1 for d = 30, 45 m, respectively. It can also be observed that the gain introduced by increasing the number of elements performing connected mode exhibits a diminishing marginal effect when the total number of elements N is large. This indicates that the performance enhancement may be fulfilled by programming only a few elements in RDARS to perform connected mode. §.§.§ Ergodic Achievable Rate In Fig. 9.(a), the ergodic achievable rate versus the total number of elements N for both RIS and RADRS-aided systems with different a is presented. Here we set γ = 60 dB and α_RB = 2.2, and the scenario where the RIS and the RDARS are both deployed near the receiver (BS) for lowering the cost of connection for elements performing connected mode, i.e., d = 30 m, is considered. As compared to the traditional RIS-aided system, the RDARS-aided system brings 0.8, 1.5, 1.9 bps/Hz rate improvements with only 1, 2, 3 elements performing connected mode, respectively, when N = 256. Note that the rate gain provided by RDARS also vanishes with the increasing N as shown in Fig. 9.(a). This is because the reflection gain will become dominant for relatively large N, and from Remark 2, the ergodic achievable rate of the RDARS-aided system (using MRC combiner) and the conventional RIS-aided system perform similarly in the asymptotic regime of N under optimal phase shift configurations. It is worth mentioning that, although the passive reflecting elements for RIS are designed in sub-wavelength size, achieving a satisfying performance with a RIS-aided system usually requires a large N and thus may result in an overall large size reflecting surface <cit.>. In contrast, by replacing RIS with RDARS, the total size of the surface can be greatly reduced, which makes it more flexible and convenient to be deployed on the surface of walls or buildings to assist practical communications. In Fig. 9.(b), we compare the ergodic achievable rate versus the transmit power for both RIS and RDARS-aided systems with different N and a. It is clear that the RDARS with 3, 2, 1 elements performing connected mode brings 11, 10, 6 dB gains, respectively, to obtain the 6 bps/Hz ergodic achievable rate as compared to the conventional RIS-aided system. This thereby demonstrates the superiority of RDARS in saving the transmit power. § EXPERIMENTAL RESULT In this section, experiment results are presented using a fabricated prototype of RDARS to verify the performance of this proof-of-concept. Fig. 10 presents the picture of the proposed RDARS-aided wireless communication system. The system is composed of one transmitter (UE), one receiver (BS), and one RDARS. Both the transmitter and receiver consist of a computer, a Universal Software Radio Peripheral (USRP), and an antenna.[Note that only one antenna is connected to the USRP at the receiver (BS) as shown in Fig. 10.(b).] The detailed system parameters are presented in TABLE 3. Specifically, the RDARS is made of 16 unit cells per column and per row, i.e., 16× 16=256 total elements. Each element of RDARS has 2-bit phase resolution when performing reflection mode, and will switch to the connected mode when has been programmed and cable is connected to its port in advance at the back of the surface as shown in Fig. 11.(b). The mode and the phase shift of each element can be dynamically and independently controlled by the RDARS controller with a total 512-bit control signal delivered through the control link using User Datagram Protocol (UDP). Since this experiment aims to verify the proposed architecture's effectiveness, we mainly focus on the performance comparison of the RDARS-aided and the conventional RIS-aided systems under the same system configurations. As shown in Fig. 10.(c), RDARS is deployed to assist in uplink communication between UE and BS in a relatively static scenario. Since the RDARS is designed to have a Uniform Planer Array (UPA) form, the optimal phase shift is determined by searching through the codebook and finding the codeword that gives the maximum received power, as mentioned in<cit.>. Once determined, the optimal phase shift is fixed for all schemes during the experiment conducted in a short time period for a fair comparison. The experimental results are presented in Table 4 with three performance metrics, i.e., the received SNR[ Here the received SNR refers to the symbol SNR in dB for the given sequance of IQ samples and is calculated based on Error-Vector-Magnitude (EVM), i.e., SNR [dB] = -20 log_10(EVM).], the data transmission throughput, and the block error rate (BLER). The throughput improvement in Table 4 is calculated between the RDARS-aided and the RIS-aided systems based on the optimal phase shift. One example of the throughput measurements in the experiments is shown in Fig. 12 for clear illustration. As seen from Table 4, for optimal phase shift design, when RIS is deployed, the throughput is improved, i.e., from 2.5 Mbps to 15.0 Mbps. The received SNR is improved by 0.76 dB, i.e., from 10.09 dB to 10.85 dB, and the BLER is reduced by 46%, i.e., from 0.911% to 0.451%. This validated the positive impact of RIS in improving the system performance to a certain extent. However, compared to RIS, RDARS is shown to greatly improve system performance as clearly demonstrated in Table 4, e.g., the RDARS with three elements performing connected mode brought a 76% throughput improvement, i.e., from 15.0 Mbps to 26.4 Mbps, as compared to the RIS-aided system. Also, the received SNR is improved by 2.26 dB, i.e., from 10.85 dB to 13.11 dB, and the BLER is reduced by 41.7%, i.e., from 45.1% to 3.4%. This demonstrates the superiority of RDARS by well compensating the “multiplicative fading" effect. Notably, the gain introduced by increasing the number of elements performing connected mode shows a diminishing marginal effect, e.g., 4.4 Mbps and 1 Mbps throughput improvement from a = 1 to a = 2 and from a = 2 to a = 3, respectively. This indicates that for practical deployment of RDARS, only a few elements performing connected mode are sufficient to greatly improve performance with moderate cost. The experiment results have validated the effectiveness of the proposed RDARS, and its potential in assisting wireless communications systems by greatly boosting the performance as compared to RIS. In fact, in the conventional fully-passive RIS setting, a control link is needed for delivering the control signal for phase shift design. Though our prototype adopts two separate links for RDARS-controller (or RIS-controller by setting all the elements under reflection mode) and the connection of elements performing connected mode with BS (using dedicated cable), we envision that these two links could be integrated together. Research in <cit.> has pointed out that the design of the RIS phase shift based on statistical CSI can provide similar gain as compared to instantaneous CSI. As such, the phase shift information delivery through the controller only needs to be updated on a large-time scale where the remaining time block could be utilized to deliver the signal received at elements performing connected mode on RDARS. This could potentially boost the performance of the RDARS-aided system in terms of achievable rate further. § FURTHER DISCUSSION From the numerical and experiment results, we reveal the effectiveness of the proposed RDARS architecture. However, the potential of RDARS is far beyond what is mentioned in this paper. For example, obtaining global CSI is possible with RDARS whereas only cascaded CSI is available for conventional RIS <cit.>. In fact, the proposed channel estimation methods for “semi-passive RIS” could be utilized as a potential channel estimation scheme for the RDARS-assisted system<cit.>. It should be noted that the original “semi-passive RIS” was proposed merely for channel estimation and the received pilot signals were mainly processed locally on RIS with low-cost signal processing units, this may limit the estimation accuracy. On the contrary, the RDARS-assisted system has the freedom to jointly process signals at BS with more advanced signal processing units while still maintaining low-cost and low-energy-consumption for RDARS. Also, though we focus on unveiling the positive impact of RDARS-aided uplink communication in this paper, RDARS has the potential to assist downlink communication, e.g., the joint active and passive beamforming design for the antenna at BS, the elements acting as remote antennas at RDARS, and the elements performing reflection mode on RDARS. As such, it is worthy of further unveiling the potential of RDARS in various scenarios and combining RDARS with other advanced technologies for future wireless communications. § CONCLUSION This paper proposed the architecture of “Reconfigurable distributed antennas and reflecting surfaces (RDARS)” to overcome the “multiplicative fading” effect, one of the bottlenecks of the conventional fully-passive RIS. Specifically, the elements of RDARS can perform either reflection mode, which acts as passive reflecting units, or connected mode, which performs as remote antennas for BS collecting signal. Therefore, with joint processing of signals at the receiver, a significant gain can be achieved with the controllable trade-off between the reflection gain and the distribution gain in an RDARS-aided system. To characterize the performance of the proposed RDARS-aided system, we analyzed the received SNR under MRC with the system configuration of SISO. Closed-form expressions of outage probability and ergodic achievable rate were also derived. Moreover, the performance of the RDARS-aided and the fully-passive RIS-aided systems was compared under various system parameters to unveil the superiority of the RDARS-aided systems. Finally, to demonstrate the practicability and manufacturability of the proposed RDARS, we conducted experiments by establishing an RDARS prototype. Both simulation and experimental results confirmed the effectiveness of the proposed RDARS architecture, thus envisioning RDARS as a promising candidate for future 6G wireless systems. § THE FIRST AND THE SECOND MOMENTS OF Γ Recall that γ = γ ( (γ_1 + γ_2)^2 + γ_3 ), where we define γ_1 = |h_UB|, γ_2 = ∑_i=1^N(1-a_i)|h_RB,i||h_UR,i|, and γ_3 = 𝐡_UR^H𝐀^H𝐀𝐡_UR. To obtain the first and second moments of γ, we need the following statistical information: the first to the fourth moments of γ_1, γ_2, and the first and the second moments of γ_3. To simplify the notation, we define x_i=(1-a_i)|h_UR,i||h_RB,i|. Below we derive the aforementioned required statistics. §.§ γ_1 The first to fourth moments of h_UB can be easily obtained as 𝔼[γ_1] = 1/2√(π)ρ_UB, 𝔼[γ_1^2] = ρ_UB^2, 𝔼[γ_1^3] = 3/4√(π)ρ_UB^3 and 𝔼[γ_1^4] = 2ρ_UB^4. §.§ γ_2 §.§.§ 𝔼[γ_2] 𝔼[γ_2] = ∑_i=1^N(1-a_i) 𝔼[|h_UR,i|] 𝔼[|h_RB,i|]= (N-a) π/4ρ_URρ_RB. §.§.§ 𝔼[γ_2^2] 𝔼[γ_2^2] = 𝔼[∑_i=1^N (1-a_i)^2 |h_UR,i|^2 |h_RB,i|^2 + ∑_i ≠ j^N (1-a_i) (1-a_j) |h_UR,i| |h_RB,i| |h_UR,j| |h_RB,j|], = (N-a)(1+ π^2/16(N-a-1)) ρ_UR^2ρ_RB^2. §.§.§ 𝔼[γ_2^3] 𝔼[γ_2^3] = 𝔼[∑_i=1^N x_i^3 + ∑_i≠ j^N x_i^2 x_j + ∑_i ≠ j ≠ k^N x_i x_j x_k ], we have 𝔼[∑_i=1^N x_i^3] = (N-a) 9/16πρ_UR^3 ρ_RB^3, 𝔼[∑_i≠ j^N x_i^2 x_j] = C_3^1C_N-a^2C_2^11/4πρ_UR^3 ρ_RB^3, and 𝔼 [∑_i ≠ j ≠ k^N x_i x_j x_k] = =C_N-a^3C_3^1C_2^11/64π^3 ρ_UR^3 ρ_RB^3 where C_N^M = N!/M!(N-M)!. Therefore we have 𝔼 [γ_2^3] = C_3 ρ_UR^3 ρ_RB^3, where C_3 = (N-a) 9/16π + 3(N-a)(N-a-1) 1/4π + (N-a)(N-a-1)(N-a-2) 1/64π^3. §.§.§ 𝔼[γ_2^4] 𝔼[γ_2^4] = 𝔼[ ∑_i=1^N x_i^4 + ∑_i ≠ j^N x_i^3 x_j^1 + ∑_i ≠ j^N x_i^2 x_j^2 + ∑_i ≠ j ≠ k ^N x_i^2 x_j x_l + ∑_i ≠ j ≠ k ≠ l^N x_i^2 x_j^2 x_k^2 x_l^2 ], We have 𝔼 [ ∑_i=1^N x_i^4] = 4(N-a) ρ_UR^4ρ_RB^4, 𝔼 [ ∑_i ≠ j^N x_i^3 x_j^1] = C_4^1C_N-a^2C_2^19/64π^2ρ_UR^4ρ_RB^4, 𝔼 [ ∑_i ≠ j^N x_i^2 x_j^2] = C_4^2C_N-a^2ρ_UR^4ρ_RB^4, 𝔼[∑_i ≠ j ≠ k ^N x_i^2 x_j x_l] = C_N-a^3C_4^2C_3^1C_2^1C_2^1π^2/16ρ_UR^4ρ_RB^4, and 𝔼 [ ∑_i ≠ j ≠ k ≠ l^N x_i x_j x_k x_l] = C_N-a^4C_4^1C_3^1C_2^11/256π^4ρ_UR^4ρ_RB^4. Thus it yields 𝔼 [γ_2^4] = C_4 ρ_UR^4ρ_RB^4, where C_4 = 4(N-a) + 9/16(N - a)(N-a-1) π^2 + 3 (N-a) (N-a-1) + 6(N-a)(N-a-1)(N-a-2) π^2/16 + 1/256(N-a)(N-a-1)(N-a-2)(N-a-3) π^4. §.§ γ_3 𝐡_UI^H𝐀^H𝐀𝐡_UI = ∑_i=1^N a_i^2 |h_UR,i|^2 = ∑_i=1^N a_i^2 y_i , where a_i = 𝐀(i,i) and y_i = |h_UR,i|^2. Note that y_i/1/2ρ_UI^2 follows Chi-square distribution with two degrees of freedom, i.e., y_i/1/2ρ_UR^2 ∼χ^2(2). And y_i = ρ_UR^2/2y_i/1/2ρ_UR^2 follows Gamma distribution, i.e., y_i ∼Γ(1, ρ_UR^2). Therefore we have 𝐡_UR^H𝐀^H𝐀𝐡_UR∼Γ(a, ρ_UR^2), and the first and second moments are obtained as 𝔼[γ_3] = 𝔼[𝐡_UR^H𝐀^H𝐀𝐡_UR] = a ρ_UR^2, 𝔼[γ_3^2] = 𝔼[(𝐡_UR^H𝐀^H𝐀𝐡_UR)^2] = ∑_i=1^N a_i^2 𝔼[|h_UR,i|^4] + ∑_i ≠ j^N a_i a_j 𝔼[|h_UR,i|^2] 𝔼[|h_UR,j|^2] = a(a+1) ρ_UR^4 Armed with the above statistics, the first and second moments of γ are derived. IEEEtran

http://arxiv.org/abs/2303.07161v1

20230313143615

Quantum estimation of a charge distance from a hole-spin qubit in Si quantum dots

cond-mat.mes-hall

^1Università di Modena e Reggio Emilia, I-41125 Modena, Italy ^2Centro S3, CNR-Istituto di Nanoscienze, I-41125 Modena, Italy Hole-spin qubits in semiconductors represent a mature platform for quantum technological applications. Here we consider their use as sensors of the electrostatic environment, and specifically for estimating the distance from the qubit of an external charge. Different approaches are considered — based on the use of single or double quantum dots, ground and out-of-equilibrium states, Rabi and Ramsey measurements — and comparatively analyzed by means of the classical and quantum Fisher information. Detailed quantitative aspects result from the multiband character of the hole states, which we account for within a multiband approach. Beyond this, general conclusions can be drawn on the relative efficiency of the above options, and analytical expressions are derived for the Fisher information of a generic qubit within the Rabi and Ramsey schemes. Quantum estimation of a charge distance from a hole-spin qubit in Si quantum dots Filippo Troiani^2 March 30, 2023 ================================================================================= § INTRODUCTION Spin qubits in silicon quantum dots are the subject of an intense research activity <cit.>. In recent years, high fidelities have been achieved in all fundamental primitive operations, including single- <cit.> and two-qubit gates <cit.>, state preparation <cit.> and readout <cit.>. The suitability of silicon as a host material for spin qubits <cit.> is related to the long coherence times <cit.>, resulting from the natural abundance of nonmagnetic isotopes and from the possibility of further enhancing such fraction through isotopic purification <cit.>. Moreover, the manufacturing processes of silicon devices are well established in the modern industry <cit.>. Thus, dense arrays of spin qubits can be realised <cit.> and possibly operated at temperatures as high as 4K <cit.>. Silicon quantum dots can host both electron- and hole-spin based qubits. An appealing feature of the hole-based implementation lies in the strong atomic spin-orbit coupling, which offers the opportunity to perform all-electrical manipulation of the qubit states <cit.>, without the inclusion of micromagnets in the device <cit.>. A double-edged sword in the use of hole spins is represented by the dependence of the qubit properties and of its coupling to external fields on the band mixing, which can in turn be rather sensitive to the electrostatic environment. On the one hand, this makes hole-spin qubits highly susceptible to electric noise <cit.>. On the other hand, it allows for a strong tunability and dynamic control of the qubit properties <cit.>. Here we explore the possibility of exploiting such properties, along with the maturity of hole-spin qubits, in the context of quantum parameter estimation <cit.>. Specifically, we aim at estimating the distance D from the qubit of a point charge: as shown in Fig. <ref>, different distances D_1 and D_2 result in different spatial distributions of the hole ground state [panels (a,b)], and consequently in different values of the Larmor (ω_ L) and Rabi (ω_ R) frequencies. This, in turn, translates into different rotation angles of the qubit state for given free precession times τ [panels (c,d)] or manipulation pulses [panels (e,f)]. The inference of the unknown parameter D can thus be based either on the properties of the hole ground state (static approach), or on the final out-of-equilibrium state |ψ⟩ of the spin qubit resulting from a given pulse sequence (dynamic approach). We assess and compare the quality of the parameter estimation based on these approaches by computing the classical and quantum Fisher information, the former one being specifically related to the Rabi and to the Ramsey measurement schemes. This approach thus combines aspects of remote charge sensing <cit.> and of quantum parameter estimation. As in remote charge sensing, a property of a target charge is inferred from its effect on a sensor dot. However, unlike in remote charge sensing, the unknown quantity is a continuous one (the charge distance, rather than its value) and is mapped onto the state of the sensor qubit, rather than on the measured position of its Coulomb peaks. Hole eigenstates are computed within a six-band 𝑘· 𝑝 approach <cit.>. The Coulomb interaction between the hole and the external charge contributes to the total electrostatic potential acting on the hole, under the assumption that the external charge is localized enough for its state to be unaffected by the interaction with the hole. This allows for a realistic simulation of the interplay between the shape of the confining electrostatic potential, the Coulomb hole-charge interaction, the atomic spin-orbit coupling and an external magnetic field. The role of the confinement potential clearly emerges from the comparison between single quantum dots (SQDs) and double quantum dots (DQDs). The latter ones are characterized by a wider polarizability of the hole ground state, and thus a stronger dependence of its properties on the distance of the external charge. Moreover, the Rabi and Ramsey frequencies of DQDs have a stronger dependence on D than those of SQDs, eventually resulting in a large enhancement of the precision in the estimate. Overall, the dynamic approach based on the Ramsey measurement with a DQD allows one to achieve values of the Fisher information that are several orders of magnitude larger than those obtained with the other considered approaches. This emerges as a robust conclusion, irrespective of the detailed properties of the quantum dot systems. The rest of the paper is organized as follows. In Section II we introduce the multiband model of the hole-spin qubit, along with the Hubbard model that qualitatively reproduces the main features of the DQD implementation. Section III discusses the static estimation strategy, based on the dependence of the hole ground state on the charge distance. Section IV is devoted to the dynamic strategy, where the parameter estimation is based on the Rabi and Ramsey measurements. Finally, in Section V we report the conclusions and the outlook. § THE MODEL SYSTEM The multiband character of the hole eigenstates plays a crucial role in determining the relevant qubit properties, specifically the Larmor and Rabi frequencies. In fact, these result from the interplay between the atomic spin-orbit coupling, the confining potential, and the external magnetic field. In order to account for these features, we model the hole-spin qubit by means of a six-band Luttinger-Kohn Hamiltonian <cit.> (Subsec. <ref>), with an external confining potential that models either a SQD or a DQD. In the DQD case, the interplay between interdot tunneling and charge-induced bias can also be described by means of a simpler and more intuitive Hubbard model, which confirms the independence of some reported behaviors on the details of the DQD system, and additionally allows for the derivation of analytical results. Such model is introduced in Subsec. <ref>. §.§ Multiband model of the single and double quantum dots The six-band Luttinger-Kohn envelope-function Hamiltonian is appropriate for the description of the low-energy hole states close to the maximum of the valence band in silicon (Γ point) <cit.>. There, the atomic spin-orbit coupling provides an energy splitting Δ_ SO = 44 meV between the j = 3/2 quartet (heavy and light hole bands) and the j = 1/2 doublet (split-off band). The single-hole Hamiltonian defined in such combined position-spin space and with the inclusion of external (electric and magnetic) fields, is diagonalized numerically in order to derive the hole energies and eigenstates <cit.>. Single and double quantum dots. The hole confinement is induced by a combination of band offset and electrostatic coupling to the metal gates. In the presence of an additional external charge, the total potential acting on the hole is given by: V( r) = V_ conf( r) + V_ c( r), where V_ conf( r) is the confinement potential, and V_ c( r) accounts for the Coulomb interaction with the charge. In the case of the single (SQD) and double quantum dots (DQD) considered hereafter, V_ conf( r) coincides with V_ SQD( r) = κ/2( x^2 + y^2 ) + V_∥(z) , and V_ DQD( r) = κ/2[ (x^2-a^2)^2/4a^2 +y^2] + V_∥(z) , respectively. In both cases, the last contribution accounts for the confinement in the vertical direction and reads V_∥(z) = { - E_∥ z |z| < L_z / 2 V_ bo |z| > L_z / 2 . , where L_z is the channel width, V_ bo is the band offset between the well and the barriers along the z direction (here V_ bo = 4000meV and L_z = 5nm), and E_∥ is an electric field applied along the z direction <cit.> (E_∥ is set to zero in Sec. <ref> and to -50mV/nm in Sec. <ref>). The strength of the in-plane confinement is determined by κ and, close to the minima of V_ DQD( r), it approaches that of the SQD, both in the x and y directions (here κ = 7.6566 × 10^-2meV/nm^2). The interdot distance is identified with the distance 2a between the minima of V_ DQD( r). The Coulomb interaction between the confined hole and the external charge (assumed to be equal to - |e|) is V_ c( r) = - e^2/ε√((x-x_ c)^2+(y-y_ c)^2+(z-z_ c)^2) , where ε = 11.68 is the dielectric constant of bulk silicon <cit.>. In the following, we specifically consider the case where the charge is positioned along the x axis, r_ c = (D,0,0), in order to highlight the interplay between the interdot tunneling and the Coulomb interaction. Finally, the coupling between the hole and the static magnetic field is accounted for by the Zeeman-Bloch Hamiltonian and the substitution k→ k + (e / ħ) A in the 𝑘· 𝑝 Hamiltonian, with A being the vector potential associated to the magnetic field <cit.>. Larmor and Rabi frequencies. In order to realize a hole-spin qubit, it is crucial to split the Kramers degeneracy in the ground doublet by means of a static magnetic field B=B (sinθ cosϕ,sinθ sinϕ,cosθ). The resulting energy splitting between the lowest energy eigenvalues e_1 and e_2 defines the Larmor angular frequency ω_ L≡1/ħ (e_2-e_1) , whose value depends on the field intensity (B) and orientation (θ and ϕ), on the detailed composition of the hole eigenstates, and (therefore) on the parameters that define the confinement potential. Hole-spin qubits can be manipulated electrically, by applying voltage pulses to a neighboring metal gate. Here, we assume that such voltage induces a time-dependent potential perturbation δ V( r,t) = - δ E_0 cos (ω t) z, corresponding to an oscillating, homogeneous electric field (here δ E_0 = 3meV/nm). The resulting Rabi angular frequency is given by: ω_ R≡1/ħ δ E_0 |⟨ψ_1| ẑ |ψ_2⟩| , where |ψ_1⟩ and |ψ_2⟩ are the hole eigenstates belonging to the ground doublet. Both ω_ L and ω_ R are affected by the Coulomb interaction between the hole and the external charge [Eq. (<ref>)], and depend on the distance D between this and the center of the (double) quantum dot. The figures of merit that determine the accuracy in the estimation of D (Section <ref>) are functions of both these angular frequencies and their derivatives with respect to D, ω_ L'≡∂_D ω_ L and ω_ R'≡∂_D ω_ R. These are here calculated through the extended Hellmann-Feynman theorem, which allows to avoid the finite-differences approach and the related numerical noise (see Appendix <ref> for further details). §.§ Hubbard model of the double quantum dot The coupling of the hole to the external (electric and magnetic) fields, and the resulting accuracy in the parameter estimation achieved through the dynamic approach (Sec. <ref>) depends on the detailed composition of the hole eigenstates, and thus on the specific geometry of the host device and on the resulting shape of the confining potential. The properties of the hole ground state that determine the performance of the static approach (Sec. <ref>) are less dependent on such details. It is therefore possible to identify general trends and properties, which can be effectively described in terms of a simpler model. In fact, as shown below, the simplest model of a DQD system, namely, the biased two-site single-band Hubbard Hamiltonian, qualitatively accounts for several features of the more realistic multiband system, including the dependence on D of the quantum Fisher information. The biased DQD Hubbard Hamiltonian, for a single hole, reads: Ĥ^ Hub=∑_α = ⇑, ⇓( ϵ_1 n̂_1, α+ϵ_2 n̂_2, α- t ĉ_1, α^†ĉ_2, α- t^* ĉ_2, α^†ĉ_1, α) . Here, α∈{⇑, ⇓} is a pseudospin quantum number, distinguishing time-reversal conjugated single-particle states <cit.>; ĉ^†_i, α creates a hole in a state localized in dot i with pseudospin α, and n̂_i,α = ĉ^†_i, αĉ_i, α; t is the inter-dot tunneling parameter, and ϵ_i is the onsite energy for site i. The eigenvalues are independent of α, and read e^ Hub_± = 1/2[ ϵ_1 + ϵ_2 ±√((ϵ_1 - ϵ_2)^2 + 4 |t|^2 )] . The parameters introduced in Eq. (<ref>) can be related to the energies characterizing the multiband system. In particular, the two dots are assumed to be symmetric, so that the difference between the onsite energies ϵ_i is entirely due to the presence of the external charge. As a simple approximation, we then take ϵ_i to coincide with V_ c( r_i), where r_1/2 = (± a,0,0) correspond to the minima of V_ DQD( r): ϵ_1 = V_c ( r_1) = -e^2/ε (D+a), ϵ_2 = V_c ( r_2) = -e^2/ε (D-a) . The bias δ_ c between the two dots, induced by the charge, is then given by δ_ c = ϵ_1 - ϵ_2 = 2ae^2/ε(D^2-a^2). As to the energy separation δ_3 , 1 ≡ e_3 - e_1 between the ground and first excited doublets, obtained from the multiband calculations at zero magnetic field and for V_ c = 0, this is identified with 2 |t|, corresponding to the energy gap e^ Hub_+-e^ Hub_- of the Hubbard model in the absence of external charges. As discussed in Subsec. <ref> and Appendix <ref>, the hopping parameter in the DQD depends on a according to |t| = t_0 e^- γ a^2 to a very good approximation. § STATIC APPROACH TO THE ESTIMATE OF THE CHARGE DISTANCE The presence of a charge in the vicinity of the DQD affects the eigenstates of the system. In particular, the dependence of the hole ground state on the distance of the charge can be used to estimate its value [Fig. <ref> (a,b)]. The highest achievable precision in such estimate is here quantified by means of the quantum Fisher information (QFI). Its maxima, corresponding to optimal working points, are investigated through the dependence on the charge position of the hole density distribution and of the energy gap between the ground and first excited doublets. §.§ Charge distribution and energy eigenvalues If the confinement potential V_ DQD( r) is symmetric with respect to the plane x=0, the same applies to the hole density of the eigenstates. This symmetry is broken by the presence of the charge positioned away from the yz plane (Fig. <ref>). In particular, a negative charge at r_ c =(D,0,0) (with D>0) tends to localize the hole density n_k( r)=|⟨ r|ψ_k⟩|^2 for the ground (k=1,2, red curves) and first excited doublets (k=3,4, cyan curves) respectively in the right (x>0) and left (x<0) dots. The degree of localization results from the competition between the Coulomb interaction with the charge, which induces a bias between the two dots and thus tends to localize the hole, and the interdot tunneling, which tends to delocalize the eigenstates over the DQD. A transition between a delocalized and a localized-hole regime is observed for increasing values of a at a given charge distance D [panels (a-c)], and for decreasing values of D at a given interdot half distance a [panels (a-c)]. Such transition clearly shows up in the behavior of the hole energy eigenvalues at zero magnetic field. In particular, we consider the energies e_1=e_2 and e_3=e_4 of the Kramers-degenerate ground and first-excited doublets as a function of a [Fig. <ref>(a)]. In the absence of an external charge, the energy gap δ_3,1≡ e_3-e_1 decreases monotonically in the considered range of values [Fig. <ref>(b), dashed grey line]. This is because δ_3,1 is related to the tunneling amplitude between the two dots, which decreases approximately as e^-γ a^2 (for further details, see Appendix <ref>). This behavior changes in the presence of the external charge (solid lines), where δ_3,1 displays a minimum for a≈ a_ min. There, a transition takes place from a regime where the interdot tunneling is larger than the charge-induced bias and the hole state is delocalized, to the opposite for a≳ a_min. Correspondingly, the energy levels e_1 and e_3 undergo an avoided crossing, as can be seen by comparing the curves corresponding to the same value of D in Fig. <ref>(a). This interplay is qualitatively reproduced by the Hubbard model described in Subsec. <ref>. In fact, the bias between the two dots, induced by the external charge, concurs with quantum tunneling to determine the avoided crossing between the energy levels e_-^ Hub and e_+^ Hub [Eq. (<ref>)], corresponding to e_1 and e_3, respectively (see Appendix <ref>). §.§ Quantum Fisher information of the hole ground state The dependence on the interdot distance of the hole localization and of the energy gap anticipate and explain that of the QFI. The QFI is a key quantity in quantum parameter estimation <cit.>, for it provides — via the quantum Cramér-Rao bound — the highest achievable precision in the estimate of the unknown value of the parameter. For a generic state |ψ⟩ depending on a parameter λ, the QFI reads as H = 4 ( ⟨∂_λψ|∂_λψ⟩ + ⟨∂_λψ|ψ⟩^2 ) . The QFI related to the parameter λ=D is computed hereafter in order to further characterize the dependence of the hole ground state on the position of the charge and to allow for a quantitative comparison between the potentialities of the static and dynamic approaches to the estimate of the charge distance. The dependence of H on the half interdot distance a, for different distances D of the charge from the DQD, is characterized by a maximum for a≈ a_ min, where δ_3,1 is minimal (Fig. <ref>, solid lines). In agreement with what was previously observed for the energy eigenvalues, the maxima occur at values of the interdot distance that increase with the distance D between the charge and the DQD. The behavior obtained from the six-band 𝑘· 𝑝 Hamiltonian is qualitatively reproduced by the Hubbard model (dashed lines), especially for values of D that are sufficiently larger than 2a. Besides, in the case of the Hubbard model, one can derive the analytical expression of the QFI: H^ Hub = [ 2e^2 a |t| ε D /a^2 e^4+ |t|^2ε^2(D^2-a^2)^2]^2 , which depends on the half interdot distance a also through the tunneling parameter t (see Appendix <ref>). At a given value of a, the maximum value of H^ Hub as a function of D (in the case of D > a) is obtained at D_max H^ Hub = √(a/3( a + √(4 a^2 + 3 e^4/ε^2 |t|^2) ) ) . The above expression thus gives the value of D at which the hole ground state is most sensitive to small variations in D. The problem of identifying the value of a that maximizes H^ Hub, for a given D, is not analytically solvable. However, we find that such optimal value of a is close to satisfying the condition σ≈ 1 (see Fig. <ref>), where σ≡ 2 |t| / δ_ c is the ratio between the tunneling energy gap and the charge-induced bias. The Hubbard model thus provides a clear indication that the maximal sensitivity of the hole ground state to the position of the external charge is achieved when the competition between interdot tunneling and charge-induced bias is balanced. § DYNAMIC APPROACH TO THE ESTIMATE OF THE CHARGE DISTANCE In the dynamic approach, the position of the charge is inferred from the outcome of the qubit readout, performed after a coherent control scheme. When the hole is operated as a qubit, the only degree of freedom which is accessed is the particle spin, defined by the lowest-energy doublet: {|0⟩≡|ψ_1⟩,|1⟩≡|ψ_2⟩}. The particular qubit state |ψ⟩ that is generated after manipulating the hole depends on the Larmor and Rabi frequencies and, through these, on the distance D between the (double) quantum dot and the external charge. §.§ Larmor and Rabi frequencies The presence of the external charge affects the Larmor frequency in ways that depend on the orientation of the magnetic field 𝐁. As the charge distance D from the DQD increases, f_L≡ω_ L/2π monotonically increases [Fig. <ref>(a-c), solid lines], and eventually tends to the value obtained in the absence of the charge (dashed lines). While this behavior is qualitatively independent of the magnetic-field orientation, the sensitivity of f_L to D is maximal when B is oriented along either the x or y directions. The situation is analogous for a SQD, although in this case the curves corresponding to 𝐁∥𝐱 and 𝐁∥𝐲 have opposite curvatures (i.e. opposite values of f_ L'). Both the value of the Larmor frequency (ω_ L) and its change (Δω_ L) in the presence of an external charge display a non-monotonic dependence on the interdot distance [panels (d-e), 𝐁∥𝐱]. The position of the maxima of both ω_ L and Δω_ L depends on the distance D. The role played by spin-orbit coupling, and specifically by the value of the splitting between the j=3/2 and the j=1/2 bands, is demonstrated by the comparison between the above mentioned results and those obtained by a four-band calculation (corresponding to the limit where Δ→∞). In the opposite limit with no spin-orbit coupling (Δ_ SO=0) the Larmor frequency is entirely independent of the orbital degrees of freedom and, thus, of the electrostatic potential. In this case f_ L is given by the Zeeman splitting, hf^ Z_L=g_ sμ_BB=115.77 μeV, where μ_B is the Bohr magneton and g_ s≈ 2 is the electron-spin g-factor. This implies that the presence of Δ_ SO≠ 0 is fundamental, since the dependence of the Larmor frequency on D plays a crucial role in the estimation of the parameter D performed both in the Rabi and in the Ramsey measurement schemes (see Subsec. <ref>). The Rabi frequency for a hole-spin qubit in a Si SQD has been previously studied for different devices and confining potentials. A crucial role in the tuning of ω_ R is played by the orientation of the magnetic field with respect to the crystal and dot axes. The dependence of the Rabi frequency on the angles θ and φ, that we obtain here, qualitatively coincides with the results reported elsewhere for SQDs <cit.>. In the following, we focus of the field orientation defined by θ=60^∘ and ϕ=90^∘, which represents a good compromise between the conflicting requirements of obtaining large Larmor and Rabi frequencies. Both ω_ R and its derivative with respect to D decrease monotonically with the distance D [Fig. <ref>(c,d)], in analogy to what is obtained for the Larmor frequency [panels (a,b)]. The comparison between the two quantities shows that the Larmor frequency, besides being much larger than the Rabi frequency in absolute terms, displays a much stronger dependence on the position of the charge [panels (e,f)]. This will have important implications in the identification of the optimal parameter estimation protocols. §.§ Fisher information of the qubit state Once the dependence of the Larmor and Rabi frequencies on D has been derived, one can devise a strategy for estimating the charge distance. This consists in identifying a manipulation protocol that results in a strong dependence on D of the qubit state |ψ⟩, and specifically of the measurement-outcome probabilities p_0=|⟨ 0 | ψ⟩|^2 and p_1=|⟨ 1 | ψ⟩|^2. In the following, we focus on the Rabi and Ramsey measurement schemes. In both cases, the key figure of merit is the Fisher information (FI) associated with the observable σ_Z. For a qubit in a generic state |ψ⟩, the FI related to the parameter D reads: F(|ψ⟩,σ_Z) = ∑_k=0,1(∂_D p_k)^2/p_k = (∂_D p_1)^2/(1-p_1) p_1 . The above expression also applies to the case of a mixed state ρ̂, where p_1=⟨ 1|ρ̂|1⟩. Decoherence tends to erase the information on the value of D encoded in the qubit state, and thus to reduce the FI. In order to account for this effect, in particular for that induced by the hyperfine interactions <cit.>, we include a simple decoherence model. This corresponds to a depolarizing channel with a Gaussian decay of the pure-state component on a characteristic time T_2^* (see Appendix <ref>), and essentially plays a role during the free evolution of the spin qubit in the Ramsey measurement. Rabi measurement. In the Rabi scheme and under the rotating-wave approximation, the occupation probability of state |1⟩, for a qubit initialized in |0⟩ and addressed by a rectangular pulse of frequency ω and duration T, is given by <cit.>: p_1=ω_ R^2/Ω^2sin^2(T Ω/2) , where Ω = √(ω_ R^2 + Δ^2), and Δ≡ω - ω_ L. We assume that T ≪ T_2^*, so that the effect of decoherence on the occupation probabilities p_0 and p_1 can be neglected. Combining Eqs. (<ref>-<ref>), one obtains the expression of the FI, which reads: F_ Rabi = [ T cos(Ω T/2) ω_ RΩ( ω_ Rω'_ R - Δω'_ L) + 2 Δsin(Ω T /2) ( ω_ Rω'_ L + Δω'_ R) ]^2/Ω^4 [ Δ^2 + ω_ R^2 cos^2( Ω T/2) ] . For each pulse duration (T) and central frequency (ω), the qubit read out thus allows the estimate of D with a precision whose upper bound is fixed by the above F_ Rabi. We stress that the above expression is valid for a generic Rabi measurement, irrespective of the involved physical system. In local quantum parameter estimation, the goal is the precise estimate of a parameter, whose possible values are already restricted to a limited range by an a priori knowledge. In order to account for this, we consider hereafter values of the charge distance from the (double) quantum dot ranging from D_ min = 140nm to D_ max = 160nm. A suitable choice of the pulse is crucial in order to optimize the estimation process. The uncertainty on the value of D also implies an uncertainty on ω_ R and ω_ L, and thus on the pulse duration and frequency that are required to generate a given qubit state |ψ⟩. However, if the variation of the Larmor and Rabi frequencies within the range of possible values of D is sufficiently small, then a given pulse leads to the implementation of well-defined rotations. More specifically, we consider a value of the pulse frequency equal to ω̅_ L, defined as the integral average of the Larmor frequency in the interval [D_ min;D_ max]. By doing so, we effectively show in Fig. <ref> that the final occupation probability of the qubit state |1⟩ displays a negligible dependence on the exact value of D for both the double [panel (a)] and single dot [panel (b)]. We also evaluate the integral average of the Rabi frequency, ω̅_ R, which allows us to define a reference time for the duration of the pulse, T_ Rabi=π/ω̅_R. We note that the value of ω̅_ R is approximately 25 times larger for the DQD than for the SQD, which determines a much shorter pulse duration T_ Rabi in the former case (0.038 vs 0.95 μs). As to the FI, the regions of interest clearly coincide with those where ω̅_ RT≈ (2n+1)π [panels (c,d)]. In order to gain further insight into the parameter dependence of F_ Rabi, we focus hereafter on the case n=0 [panels (e,f)], where the assumption that T ≪ T_2^* is better satisfied. The FI, plotted as a function of D for slightly different values of the pulse duration, displays a non-trivial behavior. It generally tends to decrease with the distance between the hole and the charge. However, the curves also display a marked minimum in correspondence of the resonance condition (Δ=0), where the FI approaches zero. Besides, such dip becomes increasingly narrow as ω̅_ RT approaches the value of π. While such complex behavior can only be described by the full expression of the FI [Eq. (<ref>)], the trend obtained away from the resonance condition at T=T_ Rabi is reflected by the simplified expression: F_ Rabi≈4 ω_ R^2(ω_ L')^2/Ω^4 . This is obtained from Eq. (<ref>) if Ω T_ Rabi≈π for any charge distance within the considered range, and exploiting the fact that |ω_ L'|≫ |ω_ R'| (see Subsec. <ref>). Overall, the values of the FI are comparable to those obtained for the QFI in the static approach, and larger for the single than for the double quantum dot. On the other hand, the pulse duration T_ Rabi required for the DQD is about 25 times shorter than that for the SQD, which might represent a significant advantage in dots with shorter decoherence times. Ramsey measurement. In the Ramsey measurement, one ideally applies two π/2 pulses to the system, separated by a waiting time τ. The final occupation probability of the state |1⟩ depends on the phase difference between the two basis states accumulated during the waiting time τ, and thus on ω_ L. The inference of D is based on such dependence, but also on the dependence on ω_ L and ω_ R of the rotations implemented by the two pulses. Within the RWA, and in the absence of decoherence, the expression of the final p_1 reads <cit.>: p_1, coh = 4 | ω_ R|^2 /Ω^2sin^2( Ω T /2) [ cos( Ω T /2) cos( Δτ/2) - Δ/Ωsin( Ω T /2) sin( Δτ/2) ]^2 , where T is the duration of each pulse. We report hereafter the analytical expression of the FI. Even though it is quite complex, it applies to any qubit within the Ramsey measurement scheme. The FI is given by F_ Ramsey, coh= 4 B^2/1-X^2 A^2 , where the terms A and B read A= cos(Δτ/2) sin(Ω T )-Ysin(Δτ/2) [1-cos(Ω T )] , and B= A ω_ R'-XΩ'/Ω + A_1 ω_ L'τ/2 + A_2 Ω' T + A_3 ω_ L'+YΩ'/Ω . In the above equations, we have introduced the dimensionless quantities X≡ω_ R^X/Ω and Y≡Δ/Ω, defined such that X^2+Y^2=1. As usual the prime denotes the derivative with respect to the unknown parameter λ = D . Finally, the three terms A_1, A_2, and A_3, which appear in the expression of B, are given by: A_1 = sin(Δτ/2) sin(Ω T )+Ycos(Δτ/2) [1-cos(Ω T )] , A_2 = cos(Δτ/2) cos(Ω T )-Ysin(Δτ/2) sin(Ω T ) , A_3 = sin(Δτ/2) [1-cos(Ω T )] . The sensitivity allowed by the Ramsey measurement can ideally increase with the duration of the waiting time τ, but it is in practice limited by decoherence. In order to account for such effect, and being in general τ≫ T, we assume that the effect of decoherence is mainly manifest during the free evolution phase (see Appendix <ref> for details). As a result, the final occupation probability p_1, coh [Eq. (<ref>)], obtained for the pulse sequence in the absence of decoherence, is replaced by the probability p_1 = e^- ( τ / T_2^* )^2 p_1, coh + 1/2[1- e^- ( τ / T_2^* )^2] . Combining the above equation with Eq. (<ref>), one obtains the FI in the presence of decoherence, which reads: F_ Ramsey =16X^2A^2B^2/[e^2(τ/T_2^*)^2-1]+4X^2A^2(1-X^2A^2) =F_ Ramsey,coh1-(1-2p_1)^2/e^2(τ/T_2^*)^2-(1-2p_1)^2 . The latter equation above explicits the relation with the FI in the coherent case, and shows how the ratio F_ Ramsey/F_ Ramsey,coh varies from 1 in the limit τ≪ T_2^*, to the exponentially decaying function e^-2(τ/T_2^*)^2 in the opposite limit. The general expressions of the qubit occupation probabilities and of the corresponding FI are finally applied to the case of the hole-spin qubit in the single and double quantum dots, where we include the specific values of the Larmor and Rabi frequencies and of their derivatives with respect to λ=D obtained from our 𝑘· 𝑝 simulations. In particular, we plot p_1 and F_ Ramsey as a function of D and of the waiting time τ, for T_2^*=2.5 μs (Fig. <ref>). We fix the pulse duration to a reference time T=T_ Ramsey=π/2ω̅_ R. In the case of the DQD, the dependence of p_1 on τ and on D [panel (a)] recalls the one typically obtained in Ramsey spectroscopy <cit.>, with D modulating ω_ L and thus being the cause of detuning. In both cases, the oscillations as a function of D are damped for values of the waiting time τ that are much larger than T_2^*. The FI displays analogous oscillations as a function of D and τ, with a shorter periodicity and a similar damping at long waiting times [panel (c)]. Further details can be appreciated from the sections of the contour plots, corresponding to given values of τ [panel (e)]. The dependence of the frequency in the oscillations on τ reflects the fact that the sensitivity of the accumulated phase on the exact value of ω_ L (and thus of D) increases with τ. This should also imply a monotonic increase of the local maxima with τ, which however does not occur in the presence of decoherence. The effect of decoherence can be appreciated by comparing these plots with those obtained for larger values of T_2^* and for the coherent case (Appendix <ref>). The difference with respect to the SQD case [panels (b,d,f)] is striking. On the one hand, in the SQD case one hardly observes an oscillatory behavior for either p_1 and F, because ω_ L varies much less than in the case of the DQD in the considered range of values of D. On the other hand, the maximal values of the Fisher information in the DQD are two orders of magnitude larger than in the SQD.For optimal values of τ, F_ Ramsey is roughly 10 times larger than F_ Rabi for the SQD, while it is 1000 times larger for the DQD. In fact, since F_ Rabi is similar in both systems, the discriminating factor in F_ Ramsey is the value of ω_ R, which we found to be an order of magnitude larger in the DQD with respect to the SQD. This conclusion results from a simplified expression of the Fisher information, derived for Ω T_ Ramsey≈π/2 and | ω_ L' τ| ≫ |X'|, |Y'|. In this case, one has that F_ Ramsey ≈ C 4ω_R^2(ω_L')^2/Ω^4 (ω_Rτ)^2 . where C is a factor that oscillates as a function of Δ (and thus of D), whose expression in reported in Appendix <ref>. The second factor in the above formula coincides with the approximate expression of F_ Rabi given in Eq. (<ref>). The third factor increases quadratically with the free evolution time τ. Since ω_ R T_ Ramsey≈π/2, when τ≫ T_ Ramsey the factor (ω_ Rτ)^2 might become very large, allowing the maximal values of F_ Ramsey to be much larger than those of F_ Rabi. On the other hand, τ cannot be chosen to be large at will, since C also includes an exponentially decreasing dependence on τ/T_2^*, which suppresses F_ Ramsey at τ≫ T_2^*. Therefore, the optimal value of τ should represent a compromise between these conflicting requirements. The strongly oscillating behavior of F_ Ramsey with respect to both D and τ should be taken into account in devising the measurement strategy. The oscillations as a function of the charge distance imply no problem if their period is larger than the a priori uncertainty on the value of D. If this is not the case, one can either enhance the period by reducing the free evolution time τ, or make use of a composite measure, which combines two Ramsey measurements, obtained with different values of the pulse frequency ω, and thus with different positions of the minima. This approach would lead to a smearing of the oscillating features, and thus to an enhancement of the minimal precision, at the cost of reducing the maxima. § CONCLUSIONS AND OUTLOOK The purpose of this paper is to explore the potentialities of hole-spin qubits for performing quantum sensing of the electrostatic environment. These potentialities ultimately lie in the multiband character of the eigenstates, which can be significantly influenced by the electrostatic environment. This, in turn, affects its coupling to the external (electric and magnetic) fields, and thus the qubit state |ψ⟩ that is generated by a given pulse sequence. We have specifically considered different approaches for estimating the distance D from the qubit of a point charge, and quantified the estimate precision in terms of the classical and quantum Fisher information. Beyond the dependence on the specific properties of the quantum dots, some general aspects emerge from our investigation. First, the use of double — rather than single — quantum dots allows for a significant enhancement of the precision in the estimate. In fact, at the static level, the hole ground state is more polarizable, and thus more sensitive to the position of the external charge, especially in the range where the charge-induced bias is comparable to the interdot tunneling amplitude. Besides, at the dynamic level, when the hole is localized in the double quantum dot, its Larmor and Rabi frequencies, ω_ L and ω_ R, are more sensitive to the charge position. Second, we saw that a dynamic approach, where the value of D is encoded in the statistics of a Rabi or a Ramsey measurement, allows for better estimates than the static approach, where the distance of the charge determines the statistics of an arbitrary measurement performed on the hole ground state. Third, among the dynamic approaches, the Ramsey measurement performs better than the Rabi scheme, because it allows for more efficient exploitation of the dependence of ω_ L on the charge distance, which is stronger than that of ω_ R. While the dependence on D of the ground state, ω_ L, and ω_ R is specific to the hole-spin qubit, the analytical expressions of the Fisher information corresponding to the Rabi and Ramsey measurements that we have derived are completely general and apply to any physical implementation of the qubit. Possible generalizations of the presented results can be envisaged. These include the inclusion of a multidimensional estimate, where more than one parameter needs to be simultaneously inferred, and less constrained spatial distributions of the point charge. On the other hand, the downside of the sensitivity of the hole-spin qubit to the electrostatic environment is represented by its fragility to electrical noise. The present approach and figures of merit can in principle also be used to investigate and quantify such fragility, and devise strategies for reducing or minimizing the effects of electrical-noise-induced decoherence. The authors acknowledge financial support from the European Commission through the project IQubits (Call: H2020–FETOPEN–2018–2019–2020–01, Project ID: 829005) and fruitful discussions with M. G. A. Paris. § NUMERICAL EVALUATION OF THE QFI VIA THE EXTENDED HELLMANN-FEYNMAN THEOREM The evaluation of the QFI for a parameter-dependent state | ψ_λ⟩ requires the evaluation of the state derivative, | ∂_λψ_λ⟩. Here, the unknown parameter is the charge distance from the origin (λ=D), but the discussion that follows is more general. Each state | ψ_λ⟩ is obtained by numerically diagonalizing the Hamiltonian, for each value of λ. The state | ψ_λ⟩ is thus given with a global phase factor (gauge) that is fixed randomly, i.e. without any defined or user-controlled relation with the phase assigned to the state | ψ_λ'⟩, resulting from another diagonalization. In order to compute the derivative | ∂_λψ_λ⟩ via, e.g., the finite-differences method, one should perform two numerical calculations at the values λ±δλ / 2, for sufficiently small values of δλ, and then fix the gauge in order to remove the random non-derivable global phase. One possible choice would consist in requiring that the scalar product of the state of interest with a fixed reference (basis) state is real. Then, one can compute the derivative | ∂_λψ_λ⟩≈| ψ_λ+δλ / 2> - | ψ_λ-δλ / 2>/δλ . Although well-defined, this procedure has two drawbacks: (1) it requires two numerical calculations for each value of λ, and (2) in practice, the result might depend on δλ and, ultimately, be inaccurate. In fact, if δλ is too large, then Eq. (<ref>) is not a good approximation of the derivative. Instead, if it is too small, the derivative is given by the ratio between two very small numbers, which might be prone to significant numerical noise. To avoid this problem, we apply the extended Hellmann-Feynman theorem <cit.>, whose statement can be summarized as follows. Let us consider the Schrödinger equation ℋ̂ | ψ_n ⟩ = e_n | ψ_n ⟩, where the Hamiltonian, its eigenstates and eigenvalues depend on the parameter λ. Then, the following relations hold: ∂_λ e_n = ℋ'_n,n , and | ∂_λψ_n ⟩ = i w_n | ψ_n ⟩ + ∑_m ≠ nℋ'_m,n/e_n - e_m | ψ_m ⟩ , where ℋ'_m,n≡⟨ψ_m | ( ∂_λℋ̂ ) | ψ_n ⟩, while w_n is an undetermined real number. By an appropriate choice of the overall phase of state | ψ_n ⟩, w_n can be set to zero, even though this is not required in what follows. The advantage of the present approach lies in the fact that it does not require to perform and combine the results of two independent numerical calculations, and avoids the above mentioned drawbacks. In fact, the derivative with respect to λ is only applied to the Hamiltonian, which is a known analytical function of the parameter, and can thus be computed by performing calculations for a single value of λ. In particular, the numerical evaluation of the QFI for the ground state (n=1) can be performed by combining Eq. (<ref>) and Eq. (<ref>). The result is H = 4 ∑_m ≠ 1ℋ'_1,mℋ'_m,1/(e_m-e_1)^2 , independent of w_n. In calculating the QFI with Eq. (<ref>), one should only take care that the sum converges. While it might seem that the convergence of the sum requires convergence on the calculation of the excited states (and thus the expansion of the Hamiltonian on a large basis set), we have verified that in practice only the convergence of the ground state is required. In order to compute the QFI for the state generated by the Rabi and Ramsey experiments, one needs to numerically evaluate the D-derivatives of the Larmor and Rabi frequencies. We use the extended Hellmann-Feynman theorem for this purpose as well. The derivative of the Larmor frequency is obtained by applying Eq. (<ref>): ∂_D ω_ L = 1/ħ( ℋ'_2,2 - ℋ'_1,1) . The derivative of the Rabi frequency, instead, requires the application of Eq. (<ref>), which yields: ∂_D ω_ R = | δ E_0 | /ħ |z_1,2| Re( z_1,2∑_m ≠ 2ℋ'_2,m/e_2 - e_m z_m,1. . + z_2,1∑_m ≠ 1ℋ'_1,m/e_1 - e_m z_m,2) , where z_m,n≡⟨ψ_m | ẑ | ψ_n ⟩. This procedure can be applied to the numerical calculation of analogous differential quantities, corresponding to the derivative with respect to λ of diagonal or off-diagonal matrix elements. § TUNNELING PARAMETER AND ENERGY LEVELS IN THE HUBBARD MODEL In the absence of external charge and magnetic field, the gap δ_3,1 = e_3 - e_1 between the lowest Kramers doublets is identified with the corresponding energy gap in the solutions of the Hubbard model, given by 2 |t| as mentioned in Subsec. <ref>. In general, the tunneling parameter is expected to decrease exponentially under a linear increase of the barrier height between the two dots, due to the exponential decay of the dot-localized envelope functions at each side of the barrier. For the potential that we have used, Eq. (<ref>), the barrier height is V_ DQD( 0) = κ a^2 / 8, quadratic in a. It is thus reasonable to try and fit the dependence on a of the energy gap obtained from the diagonalization of the 𝑘· 𝑝 Hamiltonian by means of the function f_3,1(a;t_0,γ) = 2 t_0 e^- γ a^2 , where the parameters t_0∈ℝ^+ and γ are determined using the method of non-linear least-squares. As shown in Fig. <ref>, the fitting can indeed reproduce the dependence of the gap on the interdot half distance with a high degree of accuracy in a significant range of values. Given the functional dependence of the tunneling parameter on the interdot half distance, one can also analytically express the energy levels and the quantum Fisher information as a function of a. The QFI derived from the Hubbard model is shown to reproduce qualitatively the behavior of the one derived from the multiband approach (Sec. <ref>). The same applies to the energy levels and to the gap between the ground and first excited levels. Figure <ref>(a) shows the energies e_±^Hub obtained from Eq. <ref> in the presence of an external charge, after identifying the hopping parameter δ_3,1 for each given a with the t_0 e^- γ a^2. The corresponding energy gap e_+^Hub-e_-^Hub is shown in Fig. <ref>(b), for different values of D. The comparison between these results and those reported in Fig. <ref> shows that the Hubbard model also qualitatively reproduces the behavior of the energy levels obtained from the six-band calculations. § EFFECT OF QUBIT DECOHERENCE ON THE RAMSEY MEASUREMENT The precision that one can achieve in the Ramsey scheme is limited by the environment-induced decoherence that affects the qubit during the free evolution time τ. For the sake of simplicity, we assume here that the environment acts as a depolarizing channel <cit.>, with depolarization probability p=1- e^-( τ / T_2^* )^2, and that the pulse duration T is much shorter than T_2^*, so that the time evolution during the application of the pulses can be regarded as coherent. Therefore, if | ψ⟩ is the qubit state in the absence of decoherence, the state density matrix at the end of the free-evolution interval in the presence of decoherence reads ρ̂ = e^- ( τ / T_2^* )^2 | ψ⟩⟨ψ | + 1 - e^- ( τ / T_2^* )^2 /21̂ . The application of the second pulse induces an approximately unitary evolution, which leaves unaffected the contribution proportional to the identity operator. As a result, the final population difference between the two basis states, χ = p_0 - p_1, is entirely due to the first contribution in Eq. (<ref>), and is thus related to the corresponding quantity in the coherent case by the relation: χ = χ_ coh e^-( τ / T_2^* )^2. From this, from Eq. (<ref>), and from the relation p_1=(1-χ)/2, it follows that F = ( χ'_ coh)^2/ e^ 2 ( τ / T_2^* )^2 - χ^2_ coh . If the environment acts as a phase-flip channel (pure dephasing), its effect on the final occupation probabilities of the qubit state, and thus on the Fisher information, is slightly different. In particular, the dephased component of the density operator ρ̂ does not include the dependence on ω_L resulting from the free evolution, but it still contains a dependence on the Rabi and Larmor frequencies related to the qubit time evolution during the application of the pulses. However, as discussed in Section <ref>, this represents a minor contribution to the Fisher information in view of the fact that τ≫ T. In order to highlight the effect of decoherence on the final occupation probability p_1 and on the Fisher information, in Fig. <ref> we plot these quantities in the case of a fully coherent evolution (T_2^*→∞) and for decreasing values of T_2^*. All quantities are shown as a function of the charge distance D and the waiting time τ, for a DQD with half-distance a=12nm, while pulse duration T is fixed in each case to the corresponding value of T_ Ramsey (defined in the main text). The probability p_1, coh [Fig. <ref>(a)] displays similar features to those obtained in the presence of decoherence [Fig. <ref>(c,e,g)], with a contrast in the oscillations which increases with T_2^*. As to the Fisher information, the maxima progressively move from the region of small τ, for the smallest value of the decoherence times [panel (h)], to the one of large τ in the coherent case [panel (b)]. Correspondingly, the evidence of the oscillations as a function of D decreases, resulting in a qualitative difference between the behaviors in the presence and in the absence of decoherence. We finally report the expression of the oscillating prefactor that appears in the approximate expression of F_ Ramsey, given in Eq. (<ref>): C = [cos(Δτ/2) - sin(Δτ/2) Y ]^2 [ cos(Δτ/2) Y + sin(Δτ/2) ]^2 / e^2 ( τ / T_2^* )^2 - { 1-2 X^2 [ cos(Δτ/2) - sin(Δτ/2) Y ]^2}^2 .

http://arxiv.org/abs/2303.06679v1

20230312145048

RotoGBML: Towards Out-of-Distribution Generalization for Gradient-Based Meta-Learning

cs.CV

Journal of Class Files, Vol. 14, No. 8, August 2021 Shell et al.: A Sample Article Using IEEEtran.cls for IEEE Journals RotoGBML: Towards Out-of-Distribution Generalization for Gradient-Based Meta-Learning Min Zhang^1, 2, Zifeng Zhuang^2, Zhitao Wang^2, Donglin Wang^2, †, Wenbin Li^3 ^† Corresponding author, ^1 Zhejiang University, ^2 Westlake University ^3 Nanjing University, liwenbin@nju.edu.cn {zhangmin, zhuangzifeng, wangzhitao, wangdonglin}@westlake.edu.cn March 30, 2023 ========================================================================================================================================================================================================================================================================== Gradient-based meta-learning (GBML) algorithms are able to fast adapt to new tasks by transferring the learned meta-knowledge, while assuming that all tasks come from the same distribution (in-distribution, ID). However, in the real world, they often suffer from an out-of-distribution (OOD) generalization problem, where tasks come from different distributions. OOD exacerbates inconsistencies in magnitudes and directions of task gradients, which brings challenges for GBML to optimize the meta-knowledge by minimizing the sum of task gradients in each minibatch. To address this problem, we propose RotoGBML, a novel approach to homogenize OOD task gradients. RotoGBML uses reweighted vectors to dynamically balance diverse magnitudes to a common scale and uses rotation matrixes to rotate conflicting directions close to each other. To reduce overhead, we homogenize gradients with the features rather than the network parameters. On this basis, to avoid the intervention of non-causal features (e.g., backgrounds), we also propose an invariant self-information (ISI) module to extract invariant causal features (e.g., the outlines of objects). Finally, task gradients are homogenized based on these invariant causal features. Experiments show that RotoGBML outperforms other state-of-the-art methods on various few-shot image classification benchmarks. Out-of-distribution, Few-shot learning, Gradient-based meta-learning, Multi-task learning § INTRODUCTION Deep learning has achieved great success in many real-world applications such as visual recognition <cit.> and natural language processing <cit.>. However, deep learning relies heavily on large-scale training data, showing the limitation of not being able to effectively generalize to small data regimes. To overcome this limitation, researchers have explored and developed a variety of meta-learning algorithms <cit.>, whose goal is to extract the meta-knowledge over the distribution of tasks rather than instances, and hence compensate for the lack of training data. Among the two dominant strands of meta-learning algorithms, we prefer gradient-based <cit.> over metric-based <cit.> for their flexibility and effectiveness. Unfortunately, most of these researches have a restrictive assumption that each task comes from the same distribution (in-distribution, ID). However, distribution shifts among tasks are usually inevitable in real-world scenarios <cit.>. In this paper, we consider a realistic scenario, where each minibatch is constructed of tasks from different distributions or datasets (out-of-distribution, OOD). Surprisingly, through repeated experiments, we found that the OOD tasks seriously affect the performance of GBML, e.g., the performance of MAML drops from 75.75% to 54.29% with CUB dataset under the 5-way 5-shot setting. Intuitively, we explain the possible reason from the optimization objective of the meta-knowledge. Specifically, GBML algorithms learn the meta-knowledge by minimizing the sum of gradients for each minibatch of tasks (see equation (<ref>)). If task gradients have a significant inconsistency, it may cause the learned meta-knowledge to be dominated by certain tasks with large gradient values and fail to generalize to new tasks, affecting the performance of GBML algorithms. The OOD generalization problem exacerbates the phenomenon, where task gradients are inconsistent. To demonstrate the impact of the OOD problem on task gradients in each minibatch, we next give an illustrative example. We take two different distributions to evaluate the impact in Figure <ref>. The optimization process of meta-knowledge in GBML has two loops: an inner-loop at task space (Figure <ref> (a) and (c)) and an outer loop at meta space (Figure <ref> (b) and (d)). We randomly sample two OOD tasks (task1 and task2) from the two different distributions and represent them using green and gray contour lines, respectively. First, in the task space, each task learns a task-specific parameter ψ_1^τ or ψ_2^τ using the same meta-knowledge ψ. Then, in the meta space, the losses of two OOD tasks are calculated using ψ_1^τ and ψ_2^τ and are summed in turn to optimize the meta-knowledge from ψ to ψ^*. From Figure <ref> (a), it clearly shows that in the whole optimization process, task2 has a large gradient value (the length of g_2 > g_1) and the large gradient dominates the learning process, i.e., the updated meta-knowledge ψ^* is close to task2 in Figure <ref> (b). This causes the learning of meta-knowledge to be dominated by task2 and ignore the existence of task1, and eventually GBML cannot fast adapt to new tasks using the learning meta-knowledge. When the gradient directions of OOD tasks are conflicting, the meta-knowledge optimized by summing and averaging the two task gradients may counteract each other. In this paper, to solve inconsistencies in task-gradient magnitudes and directions, we propose a simple yet effective framework, RotoGBML, to simultaneously homogenize magnitudes and directions and boost the learning of meta-knowledge in GBML. Specifically, (1) RotoGBML solves the gradient magnitudes by dynamically reweighting task gradients at each step of the learning process, while encouraging the learning of ignored tasks. (2) Instead of directly modifying gradient directions, RotoGBML smoothly rotates each task space, seamlessly aligning gradient directions in the long run (see Figure <ref> (c)). (3) To reduce overhead, we use the features instead of network parameters to homogenize task gradients and more details can be found in Section <ref>. (4) We also propose an invariant self-information (ISI) module to extract invariant causal features (e.g., object outline), which are used for homogenization. The introduction of ISI is mainly because of the fact that the features learned by neural networks are inevitably interfered with by some non-causal features (e.g., image backgrounds), which in turn affects the homogeneity of task gradients. We theoretically and experimentally demonstrate that the OOD tasks affect meta-knowledge learning process. The main contributions could be briefly summarized as follows: * We consider a real-world OOD scenario and propose a general RotoGBML algorithm to solve the OOD generalization problem. RotoGBML helps GBML algorithms learn good meta-knowledge by homogenizing task-gradient magnitudes and directions. * To reduce memory, we homogenize task gradients at a feature level. Also, we design an invariant self-information (ISI) module to extract the invariant causal features. Homogenizing gradients using these features provides further guarantees for learning robust meta-knowledge. * We theoretically evaluate our motivation and experimentally demonstrate the effectiveness of RotoGBML algorithm in various few-shot image classification benchmarks. § PRELIMINARIES §.§ Task formulation in GBML. In this paper, a supervised learning setting is considered, where each data point is denoted by (x,y) with x∈ X being the input and y∈ Y being its corresponding label. GBML assumes N training tasks {𝒯_i}_i=1^N∼𝒯_tr in each minibatch to be sampled, and an arbitrary testing task 𝒯_i∼𝒯_te is picked. Formally, each n-way k-shot task 𝒯_i consists of a support set 𝒮_i=(x_i^s, y_i^s) and a query set 𝒬_i=(x_i^q, y_i^q). x_i^s∈ℝ^n_s× d, x_i^q∈ℝ^n_q× d, y_i^s∈ℝ^n_s× n and y_i^q∈ℝ^n_q× n, where d=c× h× w indicates the image size, n_s and n_q denote the number of support and query examples, respectively. The label spaces of training, validation and testing tasks are different, i.e., Y_tr∩ Y_val∩ Y_te =∅. Most gradient-based meta-learning (GBML) algorithms have a strict assumption for tasks as: In-distribution (ID): each task is randomly sampled from the same distribution ℙ(𝒯), i.e., 𝒯_i∼ℙ(𝒯) comes from the task set {𝒯_tr, 𝒯_val, 𝒯_te} with ∀ i. The assumption indicates a highly restrictive setting due to the following reasons: (1) It is contradictory that tasks with disjoint classes come from the same distribution. (2) The data collection process is susceptible to some unobservable variables, which can cause distribution shifts even from the same dataset. Improving the generalization of gradient-based meta-learning (GBML) algorithms in the real world is very important. §.§ Task optimization in GBML. GBML aims to learn a good meta-knowledge and fast adapts to new tasks using two optimization loops (inner loop and outer loop). Generally, the network architectures of GBML contain a feature encoder f_θ parameterized by θ, which is used to extracted features f_θ: ℝ^d→ℝ^m, and a classifier c_ϕ parameterized by ϕ, which is used to output predict labels c_ϕ: ℝ^m→ℝ^n. In the inner loop, the model is grounded to an initialization (or meta-knowledge), i.e., (ϕ, θ), which is adapted to the i-th task in a few gradient steps τ (typically 1 ∼ 10 steps) by using its support set 𝒮_i. In the outer loop, the performance of the adapted model, i.e. (ϕ_i^τ, θ_i^τ), is measured on the query set 𝒬_i, and in turn used to optimize the initialization from (ϕ, θ) to (ϕ^*, θ^*). Let ℒ denotes the loss function and the above interleaved process is formulated as a bi-level optimization, (ϕ^*, θ^*) := min_(ϕ, θ)1/N∑_i=1^N𝔼_(x_i^q, y_i^q)∼𝒬_iℒ_i(ϕ_i^τ(θ_i^τ(x_i^q)), y_i^q), s.t. (ϕ_i^τ, θ_i^τ) ←min_(ϕ_i^0, θ_i^0)𝔼_(x_i^s, y_i^s)∼𝒮_iℒ_i(ϕ_i^0(θ_i^0(x_i^s)), y_i^s), where equations (<ref>) and (<ref>) are the outer-loop and inner-loop, respectively. (ϕ_i^0, θ_i^0) = (ϕ, θ) is the initial parameters for the i-th task. The learning of meta-knowledge uses the summed and averaged gradients over all tasks in current batch. § METHODOLOGY To improve the generalization ability of GBML algorithms in real-world scenarios, in this paper, we consider an out-of-distribution (OOD) setting, i.e., all tasks randomly sampled from a distribution set {ℙ_i(𝒯)}_i=1^N. A new assumption on the OOD task distributions is proposed as follows: Out-of-distribution (OOD): each task comes from the different distributions {ℙ_i(𝒯)}_i=1^N, i.e., 𝒯_i∼ℙ_i(𝒯) and 𝒯_j∼ℙ_j(𝒯) with ∀ i, j, i ≠ j and ℙ_i(𝒯) ≠ℙ_j(𝒯). To solve the problem that OOD exacerbates the inconsistencies of task-gradient magnitudes and directions, we introduce RotoGBML, a novel algorithm that consists of two building blocks: (1) Reweighting OOD task-gradient magnitudes. A reweighted vector set 𝐰_ω={ω_i}_i=1^N parameterized by ω is used to normalize the magnitudes to a common scale (see Section <ref>); (2) Rotating OOD task-gradient directions. A rotation matrix set R_γ={γ_i}_i=1^N parameterized by γ is used to rotate the directions close to each other (see Section <ref>). The two blocks complement each other and facilitate meta-knowledge to learn common information of all tasks in each minibatch, thereby fast adapting to new tasks with only a few samples. The optimization of RotoGBML is re-defined as: (ϕ^*, θ^*) := min_(ϕ,θ)1/N∑_i=1^N𝔼_(x_i^q, y_i^q)∼𝒬_iω_iℒ_i(ϕ_i^τ(γ_iθ_i^τ(x_i^q)), y_i^q), s.t. (ϕ_i^τ, θ_i^τ) ←min_(ϕ_i^0, θ_i^0)𝔼_(x_i^s, y_i^s)∼𝒮_iℒ_i(ϕ_i^0(θ_i^0(x_i^s)), y_i^s). We use red to annotate the differences between the generic GBML and RotoGBML. It clearly shows that a set of corresponding ω_i and γ_i for each task at each batch is used to homogenize gradients in the outer loop optimization process. And, if task-gradient magnitudes and directions are consistent, the equation (<ref>) degenerates into the equation (<ref>). It is worth mentioning that our proposed RotoGBML algorithm is a model-agnostic method that can be equipped with arbitrary GBML algorithms (More details is given in Section <ref>). §.§ Reweighting OOD Task-Gradient Magnitudes How to determine the value of 𝐰_ω? This is a key challenge for reweighting OOD task-gradient magnitudes. Most previous works on multi-task learning use static approaches, such as hyperparameters or prior assignment <cit.>. However, these methods are difficult to adapt to the learning process due to the use of fixed values 𝐰_ω={ω_i}_i=1^N for each task. In Table <ref>, the experiments of static reweighted vector evaluate the phenomenon. Instead, we dynamically adjust ω_i to adapt task distribution shifts by using an optimization strategy over the training iterative process. Specifically, we initialize {ω_i=1|i∈ N} for each task in each batch 𝒯_b={𝒯_i|i∈ N}, which aims to treat each task equally at the beginning. Then, the reweighted vector set is dynamically adjusted by using average gradient norms for the current batch during training. The optimization objective is as: ω^* := min_ω∑_i=1^N𝔼_(x_i^q, y_i^q)∼𝒬_iℒ_ω_i(θ_i^τ, x_i^q, y_i^q), s.t. ℒ_ω_i = g_wi - g̅_wb× [I_i]^β_1, where ·_1 is the ℓ_1-norm. g_wi=∇_θω_iℒ_i(θ_i^τ(x_i^q), y_i^q)_2 is the ℓ_2-norm of the gradients for each reweighted task loss ω_iℒ_i in outer loop. g̅_wb=1/N∑_i=1^N g_wi is the average gradient for 𝒯_b. In our experimental settings, we only use the gradients of feature encoder θ to optimize the reweighted vector set 𝐰_ω. This operation not only saves compute times but also is reasonable because the feature knowledge is more helpfule for the generalization of GBML algorithms <cit.>. I_i aims to balance each task gradient and the formulation is as follows: I_i = ℒ̂_i/∑_i=1^Nℒ̂_i, s.t. ℒ̂_i = ℒ_i/ℒ_i^0, where ℒ_i is the cross-entropy loss for query set in outer loop and ℒ_i^0 is the initial loss to be determineed as log(n) and n is classes. Specifically, the vaule of I_i is used to adjust the tasks of learning too fast or too slow, i.e., the lower value of I_i encourages the task to train more slowly, and vice versa. In equation (<ref>), a novel hyperparameter β is introduced to control the learning rate of each task. When the distributions of tasks very different, a higher β should be used to produce the large learning rate. Conversely, a lower β is appropriate for similar tasks. Note that β=0 means the learning rate is equal for all tasks. The dynamically optimized ω_i can be integrated into the learning process to homogenize task-gradient magnitudes. In the next section, we describe how to resolve conflicts of task directions in each batch. §.§ Rotating OOD Task-Gradient Directions How to define and learn the matrix of R_γ? This is a key challege for rotating OOD task-gradient directions. (1) Definition. Motivated by previous work <cit.>, we initialize R_γ∈ SO(M), where SO(M) is special orthogonal group to denote the set of all rotation matrix γ_i with size M (the network parameters). Due to the large size of M, to reduce overhead, we focus on feature-level task gradients ∇_ẑ_i^qℒ_i (rather than ∇_θℒ_i). This is reasonable due to chain rule ∇_θℒ_i = ∇_θẑ_i^q·∇_ẑ_i^qℒ_i, where ẑ_i^q = γ_iz_i^q, z_i^q = θ_i^τ(x_i^q). Finally, R_γ∈ SO(m) is used, where m is the size of feature dimensions. (2) Learning. R_γ aims to reduce the direction conflict of the task gradients in each batch by rotating the feature space. We optimize γ_i to make task spaces closer to each other and the objective is to maximize the cosine similarity or to minimize γ^* := min_γ∑_i=1^N𝔼_(x_i^q, y_i^q)∼𝒬_iℒ_γ_i(θ_i^τ, x_i^q, y_i^q), s.t. ℒ_γ_i = - <γ_ig_i, g̅_rb>, where g_i=∇_θℒ_i(θ_i^τ(x_i^q), y_i^q), g_ri=γ_ig_i=∇_ẑ_i^qℒ_i(γ_iθ_i^τ(x_i^q), y_i^q) and g̅_rb=1/N∑_i∈ Ng_ri is the average gradient for 𝒯_b. The optimization of the reweighted vector set in equation (<ref>), the rotation matrix set in equation (<ref>) and the neural network in equation (<ref>) can be interpreted as a Stackelberg game: a two player-game in which the leader and follower alternately move to minimize their respective losses, and the leader knows how the followers will react to their moves. Such an interpretation allows us to derive simple guidelines to guarantee training convergence, i.e., the network loss does not oscillate as a result of optimizing the two different objectives. We introduce more details in Appendix. §.§ Invariant Self-Information As mentioned above, we homogenize the task gradients from feature level. However, these features learned by neural network are susceptible to some biases, e.g., image backgrounds or textures <cit.>. For example, in an image of a coat shape with a dog skin texture, the CNN tends to classify it as a dog rather than a coat. This is because the neural network trained with empirical risk minimization (ERM) are more prone to inrobust texture features and ignore robust shape features <cit.>. In this section, we further introduce the invariant self-information (ISI) module to extract the robust shape features as invariant causal features. We homogenize OOD task gradients based on these invariant causal features, which helps the reweighted vector set and the rotation matrix set avoid being affected by these biases (e.g., texture features). How to extract the robust shape features? Since there are no existing corresponding shape labels, this is a challenging problem. Given an image, we observe that the shape regions tend to be more pronounced and carry a high information compared with neighboring regions. Therefore, at each layer of the neural network, we learn shape features by saving neurons that extract high-informative regions and zeroing out neurons in low-informative regions. Specifically, we first extract features z_i^l∈ℝ^c^l× k× k from input data x_i∈ℝ^c× h× w in i-th task 𝒯_i, where l is the l-th layer of the neural network. Then a drop coefficient d is proposed to drop these less-information regions: d(z_i,c^l) ∝ e^-ℐ(p_i,k^l-1)/T, s.t. ℐ(p_i,k^l-1) = -log q_i,k^l-1(p_i,k^l-1), where p_i,k^l-1 is the k-th region for c-th channel's in z_i^l-1 and p_i,k^l-1∈ℝ^c_l-1× k× k. p_i,k^l-1 is sampled from the defined distribution q_i,k^l-1. ℐ denotes self-information and T is temperature. When the amount of information of p_i,k^l-1 is low, that is, the value of this part of the self-information ℐ is low, the corresponding neuron is not optimized. To evaluate q_i,k^l-1, we use Manhattan radius C to obtain (2C+1)^2 regions as the neighbourhood of p_i,k^l-1. And, we also assume that p_i,k^l-1 and its neighbourhood regions 𝒩_i,k^l-1 come from the same distribution. We approximate q_i,k^l-1(·) with its kernel desity estimator and these neighbouring regions, the approximate representation q̂_i,k^l-1 is shown as follows: q̂_i,k^l-1(p) = 1/(2C+1)^2∑_p^'∈𝒩_i,k^l-1K(p, p^'), where K(·, ·) is kernel function. A Gaussian kernel is used, i.e., K(p, p^')=1/√(2π)hexp(-1/2h^2p-p^'^2), where h is the bandwidth. Then the information of p_i,k^l-1 is estimated by ℐ̂(p_i,k^l-1) = -log {∑_p^'∈𝒩_i,k^l-1e^-1/2h^2p_i,k^l-1-p^'^2}. If the more information is included, it can be observed that the larger the difference between p_i,k^l-1 and its neighbourhood regions 𝒩_i,k^l-1. In other words, shapes are more unique in their surroundings and thus more informative. §.§ Implementation We summarize the overall proposed approach in Algorithm <ref>. In this paper, we consider a more challenging and real-world setting, i.e., tasks are randomly sampled from different distributions during training and testing phases. In the training phase, we first use the invariant self-information (ISI) module to extract the invariant causal features. Then, based on these invariant features, we use the reweighted vector set 𝐰_ω to reweight task-gradient magnitudes and the rotating matrix R_γ to rotate task-gradient directions. Finally, these homogenized task gradients are used to optimize the meta-knowledge in equation <ref>. In the testing phase, invariant self-information (ISI) module is removed to examine whether our network has truly learned invariant features. We also remove 𝐰_ω and R_γ to evaluate the performance of meta-knowledge learned by gradient-based meta-learning (GBML) algorithms. § THEORETICAL ANALYSIS In this section, we first theoretically investigate why OOD acerbates inconsistencies in task magnitudes and directions. We sample two tasks from different distributions, where 𝒯_i=(𝒮_i, 𝒬_i)∼ℙ_i(𝒯) and 𝒯_j=(𝒮_j, 𝒬_j)∼ℙ_j(𝒯) with i ≠ j. In outer loop, the gradient difference is calculated as: d_ij = ∇_(ϕ,θ)ℒ_i(ϕ_i^τ(θ_i^τ(x_i^q)), y_i^q) - ∇_(ϕ,θ)ℒ_j(ϕ_j^τ(θ_j^τ(x_j^q)), y_j^q), where g_i=∇_(ϕ,θ)ℒ_i(ϕ_i^τ(θ_i^τ(x_i^q)), y_i^q) and g_j=∇_(ϕ,θ)ℒ_j(ϕ_j^τ(θ_j^τ(x_j^q)), y_j^q) are gradients of task 𝒯_i and 𝒯_j, respectively. To bridge the connections of task distributions and task gradients, we introduce Total Variation Distance (TVD) to re-estimate d_ij on the task distribution. The definition of Total Variation Distance (TVD) is shown as follows: (Total Variation Distance (TVD)) For two distributions P and Q, defined over the sample space Ω and σ-field ℱ, the TVD is defined as P-Q_T V:=sup _A ∈ℱ|P(A)-Q(A)|. It is well-known that the total variation distance (TVD) admits the following characterization P-Q_T V=sup _f: 0 ≤ f ≤ 1𝔼_x ∼ P[f(x)]-𝔼_x ∼ Q[f(x)]. The difference of task gradients d_ij=g_i-g_j can be bounded by the TVD as: d_ij≤ 4 η_base GL ℙ_i(𝒯)-ℙ_j(𝒯)_TV, where η_base is the learning rate of the inner loop. The G, L and more proofs are introduced in Appendix. When ℙ_i(𝒯) ≠ℙ_j(𝒯), there is a gradient difference between 𝒯_i and 𝒯_j. Then, we analyze that RotoGBML can narrow the difference between task gradients from Cosine Theorem <cit.>. When viewing gradients as a vector, the value of d_ij is simplified to calculate the length of the third edge d_ij=g_i-g_j using Cosine Theorem in Figure <ref> (left). Based on Cosine Theorem, equation (<ref>) can be re-expressed as √(g_i^2+g_j^2-2g_ig_jcos<g_i, g_j>), where · means the length of the vector and cos<·, ·> is the consine angle between two vectors. RotoGBML homogenize task gradients by reweighting task-gradient magnitudes from g_i, g_j to g_i^*, g_j^*, respectively and rotating task-gradient directions from cos<g_i, g_j> to cos<g_i^*, g_j^*> to narrow the task gradients difference from d_ij to d_ij^* in Figure <ref> (right). § RELATED WORK GBML for few-shot image classification. A plethora of GBMLs have been proposed to solve the few-shot image classification problem. The goal of these algorithms is to fast adapt to new tasks with only a few samples by transferring the meta-knowledge acquired from training tasks <cit.>. However, these works have a strict assumption that training and testing tasks come from the same distribution, which seriously violates the real world. Recently, some works <cit.> have been proposed to study the OOD problem of GBML. However, they mainly focus on the shifts caused by training and testing data (e.g., training from miniImageNet, testing from CUB). This setting follows the definition of distribution shifts in traditional machine learning <cit.>. They do not consider the problem from the GBML optimization itself and proposed appropriate algorithms for GBML frameworks. We consider the distribution shifts at the task level due to GBML uses task information of each minibatch to optimize the meta-knowledge. Task gradients homogenization in other fields. Homogenization of task-gradient magnitudes and directions is also found in multi-task learning <cit.>. Their goal is to simultaneously learn K different tasks, that is, finding K mappings from a common input dataset to a task-specific set of labels (generally K=2). Motivated by <cit.>, we regard GBML meta-knowledge learning as a multi-task optimization process and homogenize task gradients from a multi-task perspective. Our work differ from these works in: (1) the input data is different for different tasks in GBML, but the input data of different tasks in multi-task learning is the same. (2) In multi-task learning, existing works on alleviating gradient conflicts usually assume that tasks come from the same distribution. However, we address the problem of conflicting gradients in different distributions. (3) We propose a new invariant self-information module to extract causal features. § EXPERIMENTS In this section, we conduct comprehensive experiments to evaluate the effectiveness of RotoGBML and compare it with state-of-the-art algorithms. Specifically, we consider two OOD generalization problems in few-shot image classification: The weak OOD generalization problem is training and testing tasks from the same dataset but with disjoint classes <cit.>. The strong OOD generalization problem is training and testing tasks from different datasets and has multiple datasets in training data <cit.>. Unlike these works, we use a more difficult setting, i.e., each minibatch of tasks from different datasets. We aim to answer the following questions: Q1: How does the proposed RotoGBML framework perform for the few-shot image classification task on the weak OOD generalization problem (see Section <ref>)? Q2: Can RotoGBML fast generalize to new tasks from new distributions when faced with the strong OOD generalization problem (see Section <ref>)? Q3: How well each module (𝐰_ω, R_γ and ISI) in our proposed framework performs in learning the meta-knowledge (see Section <ref>)? Datasets. For the weak OOD generalization problem, we adopt two popular benchmarks for image classification, i.e., miniImageNet and CUB-200-2011 (abbr: CUB). For the strong OOD generalization problem, we adopt nine benchmarks, including miniImageNet, CUB, Cars, Places, Plantae, CropDiseases, EuroSAT, ISIC and ChestX following prior works <cit.>. In the strong OOD generalization, miniImageNet is used only as training data because of its diversity, and evaluate the trained model on the other eight datasets using a leave-out-of method, i.e., randomly sampled one dataset as testing data and other datasets as training data. Backbone of GBML. Since our RotoGBML is model-agnostic and can be equipped with arbitrary GBML algorithms, we use five representative and generic GBML algorithms as our GBML backbone including MAML <cit.>, MAML++ <cit.>, ANIL <cit.>, Reptile <cit.>, iMAML <cit.>. In all experimental results, iMAML is used as our GBML backbone due to its good performance, unless otherwise stated. Feature encoders. We follow previous works <cit.> using three general networks as our feature encoder including, i.e. conv4 (filter:64), resnet10 and resnet18. However, it is well known that GBML algorithms (e.g, MAML) under-perform when applied to large networks <cit.>, so we use conv4 to show the main experimental results for the weak OOD generalization. For a fair comparison with some cross-domain methods <cit.>, we follow their setting using resnet10 for the strong OOD generalization. We also show the performance of the three feature encoders in Figure <ref>. Experimental setting. We use the Adam optimizer with the learning rate of inner loop η_base=0.01, outer loop η_meta=0.001 and rotation matrix η_γ=5e-4. We set β=0.1 for weak and β=1.5 for strong OOD. We optimize all models from scratch and process datasets using data augmentation following previous work <cit.>. We evaluate the performance under two generic settings, i.e., 5-way 1-shot and 5-shot and report the average accuracy as well as 95% confidence interval. §.§ Weak OOD Generalization Problem We conduct experiments on two representative few-shot image classification datasets: miniImageNet and CUB. Table <ref> reports the experimental results. From Table <ref>, we have some findings as follows: (1) Compared to Vanilla GBML algorithms (2-nd block), the proposed RotoGBML framework consistently and significantly improves all performances by homogenizing task-gradient magnitudes and directions. Moreover, compared to the 95% confidence interval, it is clear that RotoGBML reduces the uncertainty of model predictions and further reduces the high variances of model learning, thereby improving the robustness and generalization. (2) Compared to some state-of-the-art few-shot image classification methods (1-st block), RotoGBML outperforms all methods, including metric-based model (ProtoNet), fine-tuning model (Linear), pretrain-based model (MTL) and knowledge-distillation-based model (SKD-GEN1). This further demonstrates the effectiveness of our proposed RotoGBML algorithm, while also providing a new solution for existing gradient-based meta-learning (GBML) algorithms to outperform metric-based meta-learning methods. We also compare the performance of the model with different feature encoders (conv4, resnet10 and resnet18) in Figure <ref>. We can find that (1) our proposed RotoGBML algorithm outperforms all existing gradient-based meta-learning (GBML) algorithms in all experimental settings with different network sizes. (2) It's worth noting that RotoGBML can alleviate the performance degradation problem of large networks. This is mainly because our proposed ISI module can extract invariant causal features, i.e., the shapes or outlines of objects, which avoid some non-causal features (e.g., the backgrounds, colors or textures of objects) to affect the generalization of large networks on some new tasks from new distributions. §.§ Strong OOD Generalization Problem In this section, we consider a more challenging OOD setting, i.e., each task of each minibatch (typically the number of tasks N=4) comes from different datasets. We compare our method and some state-of-the-art methods with the strong OOD generalization problem in Table <ref>. From Table <ref>, we have some significant findings as follows: (1) Compared to Vanilla GBML algorithms (2-nd block), RotoGBML outperforms all algorithms on eight datasets with different performance gains. These results again prove that our RotoGBML can learn robust meta-knowledge, which in turn generalizes quickly to new tasks from new distributions. (2) Compared to some cross-domain methods (1-st block), RotoGBML has a good performance in most settings, which demonstrates that we can help GBML algorithms to achieve state of the art by homogenizing task-gradient magnitudes and directions. (3) The improvement of the strong OOD setting is usually larger than that of the weak OOD setting (Table <ref>). The phenomenon justifies our conclusion that the larger the distribution difference, the more obvious inconsistencies in magnitudes and directions of task gradients. In Figure <ref>, the performances of the strong OOD problem are shown with different network sizes. We show the results on Cars and EuroSAT as we have similar observations on other datasets. Compared to Vanilla GBML algorithms, iMAML has the best performance on the weak OOD problem, but significantly degrades on the strong OOD problem. Because it uses a regularization to encourage task parameters to close to meta parameters, large distribution shifts affect the learning of meta parameters. Our RotoGBML can learn good meta parameters by homogenizing task gradients. §.§ Ablation Study The performance of our proposed three modules We use a hierarchical ablation study to evaluate the performance of each module for RotoGBML under the weak and strong OOD problems in Table <ref>. The 1-st row represents the results of Vanilla iMAML. The plus sign means performance increase and the minus sign means performance decrease, where violet is the weak OOD and blue is the strong OOD experiments. To evaluate the performances of a fixed reweighted vector set, e.g., [0.1,0.5,0.3,0.4] is used for the four tasks in each minibatch in our experiments, the results are shown in the 2-nd row. From Table <ref>, the following findings: (1) compared to Vanilla iMAML, the performances of hyperparameters have a significant drop in most settings, but our dynamic adjustment strategy (3-rd row) has a clear improvement. This further evaluates the effectiveness of our learning method to dynamically optimize the reweighted vectors. (2) Each module has different performance gains (3∼5-th rows). Compared to the strong OOD problem, the performances of ISI on the weak OOD problem have a large improvement, because even if invariant features are learned, the strong inconsistencies of task gradients affect the learning of meta-knowledge. (3) Arbitrary combination of two modules has good performance in 6∼8-th rows, especially with ISI. This is because adjusting the gradients based on invariant robust information from different distributions can learn better. We also give some visualization experiments of ISI learning invariant features in Figure <ref>. And all modules are used to have the best performance. The hyperparameter of β The hyperparameter β is introduced in equation (<ref>) in Section <ref>. The β aims to adjust the tasks learning rate and β=0.1 is used for the weak OOD generalization problem and β=1.5 is used for the strong OOD generalization problem. It is because a lower β is appropriate for similar tasks and a higher β should be used for dissimilar tasks. More experimental results for β are shown in Figures <ref> and <ref>, where “None” means the performance of iMAML and the bars with “red” represents the performance of our method (iMAML-). From Figures <ref> and <ref>, β=0 damages the performance of vanilla iMAML in the strong OOD generalization problem. The main reason is that β = 0 tries to pin the backpropagated gradients of each task to be equal at network parameters. For the strong OOD generalization problem, this is unreasonable since each task comes from a different distribution, which varies widely from each other. Moreover, we have two hyperparameters and experiment with a control variable strategy. The hyperparameter of temperature T The hyperparameter of temperature T is used as a “soft threshold" of information in equation (<ref>) in Section <ref>. When T is small, this means more conservative filtering, i.e., only patches with the least information will be dropped and most shape and texture are preserved. When T goes to infinity, all neurons will be dropped with equal probability, and the whole algorithm becomes regular Dropout <cit.>. As mentioned above, we have two hyperparameters, so we adopt a strategy of fixing one to experiment with the other. The experimental results are shown in Figure <ref> under the 5-way 1-shot and Figure <ref> under the 5-way 5-shot setting, where “None” means the performance of iMAML, “INF” means that it degrades to regular Dropout and the bars with “red” represents the performance of our method (iMAML-). From Figures <ref> and  <ref>, we find that the strong OOD generalization problem uses a larger temperature than the weak OOD generalization problem. This phenomenon is consistent with our analysis, because the strong OOD generalization problem has greater distribution differences than the weak OOD generalization problem, and more useless information needs to be dropped. Visualization of the invariant self-information module We propose the invariant self-information (ISI) module to avoid some non-causal features to affect the generalization of the neural network on some new tasks from new distributions in Section <ref>. Motivated by <cit.>, we extract the robust shape features as invariant causal features. To verify the learning of shape features with the ISI module, we visualize gradients of model output using the saliency map. Specifically, we use SmoothGrad <cit.> to calculate the saliency map S(x): S(x) = 1/n∑_i=1^n∂ f(x_i)/∂ x_i, where x_i = x + δ_i is original image x with i.i.d. Gaussian noise δ_i and f(·) is the network. The experiments are shown in Figure <ref>, where these original images are randomly sampled from the nine datasets used in our experiments. We can see that ISI is more human-aligned and sensitive to the shapes of objects. While the Saliency map of the baseline (iMAML) algorithm is filled with noise, which indicates the baseline method is less shape-biased and lacks interpretability. Although on some difficult examples (the 3-rd, 4-th and 5-th columns), our ISI module also has impressive performance. § CONCLUSION In this paper, we address the challenging problem of learning a good meta-knowledge of gradient-based meta-learning (GBML) algorithms under the OOD setting. We show that OOD generalization exacerbates the inconsistencies in task-gradient magnitudes and directions. Therefore, we propose RotoGBML, a general framework to dynamically reweight diverse magnitudes to a common scale and rotate conflicting directions close to each other. Moreover, we also design an invariant self-information (ISI) module to extract the invariant causal features. Homogenizing gradients using these causal features provide further guarantees for learning robust meta-knowledge. Finally, we analyze the feasibility of our method from the theoretical and experimental levels. We believe that our work makes a meaningful step toward adjusting task gradients under the OOD problem for GBML. IEEEtran § THEORETICAL ANALYSIS §.§ RotoGBML as Stackelberg games The optimization objective of learning the meta-knowledge in GBML algorithms has two loops: an inner loop and an outer loop. In the outer loop, we simultaneously optimize the network parameters (ϕ, θ) and the module parameters of (ω, γ). The two optimization objectives can be interpreted as a Stackelberg game <cit.>, which is an asymmetric game with players playing alternately. The neural networks (ϕ, θ) are called as the follower to minimize their loss function. The reweighted vector set and rotation matrix set (ω, γ) are named as the leader. Compared to the follower, the leader attempts to minimize their own loss function, but it does so with the advantage of knowing which will be the best response to their movement by the follower. The problem can be written as: Leader: min_(ω, γ){ℒ_(ω, γ)(ω, γ, ϕ, θ)| (ϕ, θ)∈min_(ϕ, θ)∑_i=1^Nℒ_i(ϕ, θ, ω, γ) }, Follower: min_(ϕ, θ)∑_i=1^Nℒ_i(ϕ, θ, ω, γ) . The leader knows the global objective while the follower only has access to their own local loss function. In gradient-play Stackelberg game, all players perform gradient updates at s-th step: (ω^s, γ^s) = (ω^s-1, γ^s-1) - η_(ω, γ)∇_(ω, γ)ℒ_(ω^s-1, γ^s-1)(ω^s-1, γ^s-1, ϕ^s-1, θ^s-1), (ϕ^s, θ^s) = (ϕ^s-1, θ^s-1) - η_(ϕ, θ)∑_i=1^N∇_(ϕ^s-1, θ^s-1)ℒ_i(ϕ^s-1, θ^s-1, ω^s-1, γ^s-1), where η_(ω, γ) and η_(ϕ, θ) are the learning rates of the RotoGBML and neural network, respectively. Equilibrium is an important concept in game theory, the point at which both players are satisfied with their situation, meaning that there is no action available that can immediately improve any player's score. If the two players can converge to an equilibrium point, this also guarantees the convergence of the training phase. We introduce the definition of Stackelberg equilibrium <cit.>: (Differential Stackelberg equilibrium). If ∇_(ω, γ)ℒ_(ω, γ)(ω, γ, r(ω, γ))=0 and ∇_(ω, γ)^2ℒ_(ω, γ)(ω, γ, r(ω, γ)) is positive definite, the pair of points (ϕ, θ), (ω, γ) is a differential Stackelberg equilibrium point and (ϕ, θ)=r(ω, γ) is implicitly defined by ∇_(ϕ, θ)ℒ_i(ϕ, θ, ω, γ)=0. When the players manage to reach such an equilibrium point, both of them are at a local optimum. Therefore, here we also provide theoretical convergence guarantees for an equilibrium point. In the given setting, if the leader's learning rate η_(ω, γ) goes to zero at a faster rate than the follower's η_(ϕ, θ), that is, they will asymptotically converge to a differential Stackelberg equilibrium point almost surely. As long as the follower learns faster than the leader, they will end up in a situation where both are satisfied. In our experiments, the learning rate of the follower is 1e-3 and that of the leader is 5e-4. §.§ Proof of Theorem 4.1 In our main paper, we propose the Theorem 4.1 to bound the difference of task gradients d_ij=g_i-g_j≤ 4 η_base GL ℙ_i(𝒯)-ℙ_j(𝒯)_TV by using the total variation distance (TVD) based on an assumption of G and L. Motivated by previous work <cit.>, the assumption is defined as follows: For any x∈ X, the loss function ℒ(·, x) is twice continuously differentiable. Furthermore, we assume it satisfies the following properties for any ψ=(ϕ,θ), ψ̅=(ϕ̅, θ̅): * The gradient norm is uniformly bounded by G over ψ, i.e., ∇ℒ(ψ, x)≤ G. * The loss is L-smooth, i.e., ∇ℒ(ψ, x) - ∇ℒ (ψ̅, x) ≤ L ψ- ψ̅. 𝒯_i=(𝒮_i, 𝒬_i)∼ℙ_i(𝒯) and 𝒯_j=(𝒮_j, 𝒬_j)∼ℙ_j(𝒯) with i ≠ j are randomly sampled from the different distributions. The difference of task gradients is definited as follows: d_ij = ∇_ψℒ_i(ψ_i^τ(x_i^q), y_i^q) - ∇_ψℒ_j(ψ_j^τ(x_j^q), y_j^q) , = ∇_ψℒ_i(ψ- in_i, x_i^q, y_i^q) - ∇_ψℒ_j(ψ - in_j, x_j^q, y_j^q), = 𝔼_{(x_i^k, y_i^k)∼𝒬_i∼ℙ_i(𝒯)}_k=1^n_q∇_ψℒ_i(ψ- in_i, x_i^k, y_i^k) - 𝔼_{(x_j^k, y_j^k)∼𝒬_j∼ℙ_j(𝒯)}_k=1^n_q∇_ψℒ_j(ψ- in_j, x_j^k, y_j^k) , where in_i = η_base𝔼_{(x_i^k, y_i^k)∼𝒮_i∼ℙ_i(𝒯)}_k=1^n_s∇_ψ_i^τℒ_i(ψ_i^τ-1(x_i^k), y_i^k), in_j = η_base𝔼_{(x_j^k, y_j^k)∼𝒮_j∼ℙ_j(𝒯)}_k=1^n_s∇_ψ_j^τℒ_j(ψ_j^τ-1(x_j^k), y_j^k), where n_s and n_q are the number of the support set and query set examples, respectively. g_i = ∇_ψℒ_i(ψ_i^τ(x_i^q), y_i^q) and g_j = ∇_ψℒ_j(ψ_j^τ(x_j^q), y_j^q) are the task gradients for 𝒯_i and 𝒯_j using query set in the outer loop, respectively. Hence, with probability n_qt (ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^t (1-ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^n_q-t, we have x_i^k≠ x_j^k for t choices of k (out of 1,...,n_q). In addition, based on the Assumption <ref>, the equation (<ref>) is bounded as follows: d_ij = 𝔼_{(x_i^k, y_i^k)∼𝒬_i∼ℙ_i(𝒯)}_k=1^n_q∇_ψℒ_i(ψ- in_i, x_i^k, y_i^k) - 𝔼_{(x_j^k, y_j^k)∼𝒬_j∼ℙ_j(𝒯)}_k=1^n_q∇_ψℒ_j(ψ- in_j, x_j^k, y_j^k) , ≤ 2 η_base L 𝔼_{(x_i^k, y_i^k)∼𝒬_i∼ℙ_i(𝒯)}_k=1^n_q∇_ψℒ_i(ψ_i^τ(x_i^k), y_i^k) - 𝔼_{(x_j^k, y_j^k)∼𝒬_j∼ℙ_j(𝒯)}_k=1^n_q∇_ψℒ_j(ψ_j^τ(x_j^k), y_j^k) , ≤ 4 η_base GL t/K. As a result, we have d_ij = 𝔼_{(x_i^k, y_i^k)∼𝒬_i∼ℙ_i(𝒯)}_k=1^n_q∇_ψℒ_i(ψ- in_i, x_i^k, y_i^k) - 𝔼_{(x_j^k, y_j^k)∼𝒬_j∼ℙ_j(𝒯)}_k=1^n_q∇_ψℒ_j(ψ- in_j, x_j^k, y_j^k) , ≤∑_t=0^K Kt (ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^t (1-ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^K-t· 4 η_base GL t/K, = 4 η_base GL (ℙ_i(𝒯) - ℙ_j(𝒯) _TV) ∑_t=0^K t/KKt (ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^t-1 (1-ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^K-t, = 4 η_base GL (ℙ_i(𝒯) - ℙ_j(𝒯) _TV), where the last equality follows from the fact that t/KKt (ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^t-1 (1-ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^K-t = K-1t-1 (ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^t-1 (1-ℙ_i(𝒯) - ℙ_j(𝒯) _TV)^K-1-(t-1). Therefore, we can prove the conclusion of d_ij=g_i-g_j≤ 4 η_base GL ℙ_i(𝒯)-ℙ_j(𝒯)_TV.

http://arxiv.org/abs/2303.06933v1

20230313085659

Distributionally Robust Chance-Constrained Optimization for Hierarchical UAV-based MEC

cs.NI

Distributionally Robust Chance-Constrained Optimization for Hierarchical UAV-based MEC Can Cui^†, Ziye Jia^†, Chao Dong^†, Zhuang Ling^∗, Jiahao You^† and Qihui Wu^† ^†The Key Laboratory of Dynamic Cognitive System of Electromagnetic Spectrum Space, Ministry of Industry and Information Technology, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu, 210016, China ^∗College of Communication Engineering, Jilin University, Changchun, Jilin, 130012, China. {cuican020619, jiaziye, dch, yjiahao, wuqihui}@nuaa.edu.cn, lingzhuang@jlu.edu.cn March 30, 2023 ===================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================== empty empty Multi-access edge computing (MEC) is regarded as a promising technology in the sixth-generation communication. However, the antenna gain is always affected by the environment when unmanned aerial vehicles (UAVs) are served as MEC platforms, resulting in unexpected channel errors. In order to deal with the problem and reduce the power consumption in the UAV-based MEC, we jointly optimize the access scheme and power allocation in the hierarchical UAV-based MEC. Specifically, UAVs are deployed in the lower layer to collect data from ground users. Moreover, a UAV with powerful computation ability is deployed in the upper layer to assist with computing. The goal is to guarantee the quality of service and minimize the total power consumption. We consider the errors caused by various perturbations in realistic circumstances and formulate a distributionally robust chance-constrained optimization problem with an uncertainty set. The problem with chance constraints is intractable. To tackle this issue, we utilize the conditional value-at-risk method to reformulate the problem into a semidefinite programming form. Then, a joint algorithm for access scheme and power allocation is designed. Finally, we conduct simulations to demonstrate the efficiency of the proposed algorithm. Unmanned aerial vehicles, multi-access edge computing, distributionally robust optimization, conditional value-at-risk. § INTRODUCTION [lines=2]Due to the rapid development of smart devices and artificial intelligence, there exist a growing number of application data to be processed, which is an outstanding characteristic in the sixth generation of communication system <cit.>. Besides, an emergent architecture, multi-access edge computing (MEC), becomes prevalent in recent years, which contributes to reducing the system delay and improving the quality of service (QoS) <cit.>. Furthermore, unmanned aerial vehicles (UAVs) are well known due to the easy deployment and flexible movement, which attract tremendous attentions in both academics and industries. UAVs can serve as communication relay nodes, mobile base stations (BSs), as well as computing platforms for MEC <cit.>, <cit.>. However, since there exist uncertain environmental factors such as wind, temperature, and airflow, the gains of UAVs are under uncertainty sets due to the environmental factors. For instance, the effect of wind results in the fluctuation of transmission gain within a certain range, and the extreme temperature leads to transmission gain errors. The effects of such unpredictable errors cannot be ignored in practical applications. As illustrated that UAV-enabled MEC system can provide benefits for energy-efficient task offloading, there exist a couple of related works <cit.>. For example, <cit.> studies a UAV-based mobile cloud computing system to minimize total energy consumption. In <cit.>, Zeng et al. propose an algorithm to maximize the throughput by jointly optimizing the trajectory of the UAV swarm along with the transmit power. <cit.> presents a UAV-assisted MEC for task offloading which is solved with a deep reinforcement learning method. In <cit.>, a response delay optimization algorithm is suggested for the deployment of UAVs as well as MEC servers. However, the uncertain errors of transmission gains are ignored or described as a certain distribution in the aforementioned works, which are not realistic in the practical environment. As above, in this paper, we consider the errors without distribution information to minimize the total power consumption in the hierarchical UAV-based MEC. Since there exist unpredictable parameters such as gain errors, which are in an uncertainty set, the proposed optimization problem is still intractable. Hence, a distibutionally robust optimization (DRO) problem considering the uncertainty of gains is further proposed to deal with this issue. Both the stochastic programming and robust optimization can be applied to uncertain optimization problems, and the distribution information of random parameters is necessary, which cannot be obtained in some scenarios<cit.>, <cit.>. However, DRO can be implemented without such detailed statistical information and provide more conservative solutions <cit.>. Besides, without distribution information on the uncertainty, the DRO problems are usually computationally-prohibitive. One approach to solving the DRO problem utilizes the conditional value-at-risk (CVaR) with limited statistical information <cit.>, <cit.>. Utilizing CVaR with historical data, we formulate the chance-constrained problem into a semidefinite programming (SDP) form, which can be handled with feasible solutions <cit.>. In this paper, we build the uncertainty set for the error parameters with historical data, enabling more practical applications. Accordingly, the system model can be formulated as a DRO problem and solved via the CVaR mechanism. In detail, we propose a hierarchical UAV-assisted MEC system to complete data processing and transmit data back to the ground station. The system is applicable to many real scenarios, such as real-time monitoring and disaster rescue for remote areas short of BS coverage <cit.>,<cit.>. Note that the UAVs are always constrained by energy, the total power consumption should be optimized, and the QoS of users also should be guaranteed. Taking into account the errors from antenna gains, we propose a chance-constrained problem with the uncertainty set and handle it by the CVaR mechanism. The main contributions of our work are summarized as follows. * A hierarchical UAV-based MEC system is proposed, including the lower-layer UAVs to collect data and the upper-layer UAV for relay. Besides, both layers of UAVs are equipped with computation resources for MEC. Furthermore, the gain uncertainty from transceivers is considered for practical scenarios, and we present a corresponding system model to minimize the total power consumption under such uncertainty. * The problem is formulated with an uncertainty set for transceiver gain errors, and to handle this issue, we propose a DRO-based mechanism, which is non-convex. Then, we approximate the chance constraints by the CVaR method and reformulate the issue into a tractable SDP form. * To deal with the reformulated SDP problem, we present an algorithm to jointly optimize both the access scheme and power allocation. Finally, we conduct simulations to evaluate the performance of the proposed algorithm, and the results verify the effectiveness. The rest of this paper is organized as follows. Section <ref> presents the system model and corresponding problem formulation. Section <ref> employs the CVaR-based mechanism to reformulate the original problem with chance constraints into a SDP form, and the corresponding algorithm is designed. Simulation results and the analyses are provided in Section <ref>. Finally, conclusions are drawn in Section <ref>. § SYSTEM MODEL AND PROBLEM FROMULATION In this paper, we consider a hierarchical UAV-assisted MEC system, in which lower-layer UAVs are deployed to collect and compute the data produced by ground users and transmit them back to the BS. Due to the limited computation resources of UAVs, a UAV with strong computing ability is deployed in the upper layer to handle the compute-intensive tasks. Lower-layer UAVs can complete the computation tasks locally and transmit the results to the BS, or transmit the data to the upper-layer UAV for further computation to reduce power consumption. Then, the upper-layer UAV transmits the processed results to the ground BS. The hierarchical UAV-based MEC scenario is shown in Fig. <ref>. §.§ System Model §.§.§ Communication Model In this subsection, the communication model of UAV-to-UAV and UAV-to-BS is investigated. We utilize 𝒩={1,2,…,i,…,N} to indicate the set of UAVs of the lower layer. Let p^d_i denote the transmit power of i-th lower-layer UAV. g_i denotes the i-th UAV transceiver antenna gain and B_i indicates the bandwidth. Consequently, the data transmission rate from the i-th UAV in the lower layer to the upper-layer UAV or the BS is expressed as: r_i=B_ilog_2(1+γ_ip^d_i|g_i|^2/σ_s^2),∀ i∈{1,2,…,N}, where σ_s^2 is the variance of white Gaussian noise with a mean value of 0. γ_i is the path loss related to the distance between i-th lower-layer UAV and upper-layer UAV, or between i-th lower-layer UAV and BS. We assume the data transmission adopts the orthogonal multiple access technology in this system, so the interference between different data can be ignored. Moreover, only one powerful UAV is considered in the upper layer. Let B_h denote the bandwidth of the upper-layer UAV. p^d_h is the transmit power, g_h denotes the transceiver antenna gain of the UAV hovering in the upper layer, and γ_h represents the path loss. According to the Shannon formula, the data transmission rate of the upper layer UAV towards the BS is calculated as: r_h=B_hlog_2(1+γ_hp^d_h|g_h|^2/σ_s^2). We assume L_i represents the data length transmitted by the i-th UAV. x_i∈{0,1} is a binary indicator for computing mode. x_i=1 represents the i-th UAV transmits the users' data to the upper-layer UAV for further computation, and x_i=0 denotes the i-th UAV computes collected data locally and transmit them back to the BS. Based on the above formulas, we can obtain the transmission delay as follows: t_i^d=L_i/r_i,∀ i∈{1,2,…,N}, and t_h^d=∑_i^N x_i L_i/r_h, where t_i^d denotes the transmission delay of the i-th UAV in the lower layer, and t_h^d is the transmission delay of the upper-layer UAV, respectively. Further, the total transmission power p_r^d of UAVs in the lower layer is p_r^d=∑_i=1^Np_i^d. §.§.§ Uncertainty Model Take into account the transceiver antenna gain errors influenced by the environmental factors denoted as Δ g_i and Δ g_h, i.e., g_i=g_i+Δ g_i,∀ i∈{1,2,…,N}, and g_h=g_h+Δ g_h, where g_i and g_h indicate the theoretical antenna transmission gain of the i-th lower-layer UAV and upper-layer UAV, respectively. Similarly, Δg_i and Δg_h represent the antenna gain errors produced by actual environmental impacts. The estimation error Δg_i and Δg_h are assumed to follow an unknown distribution ℙ with mean 𝔼_ℙ(Δ g_i)=μ_i and 𝔼_ℙ(Δ g_h)=μ_h, as well as the variance 𝔻_ℙ(Δ g_i)=σ_i^2 and 𝔻_ℙ(Δ g_h)=σ_h^2, respectively. Uncertainty set 𝒫 reprensents for all possible probability distributions of Δg_i and Δg_h, i.e., ℙ∈𝒫. §.§.§ Computation Model Let variable c_i represent the complexity of the computing tasks processed by the i-th UAV. η_i and η_h denote the computational power consumption coefficient of the i-th lower-layer UAV and the upper-layer UAV, respectively, which represent the power required to calculate the data of each cycle. Then we can formulate the computation power of i-th lower-layer UAV and the upper-layer UAV as: p_i^c=η_ic_i,∀ i∈{1,2,…,N}, and p_h^c=η_h∑_i=1^N x_ic_i. With the above formulations we can obtain the total computation power consumption of UAVs in the lower layers as: p_r^c=∑_i=1^N (1-x_i)p_i^c. §.§.§ Power Consumption According to the above discussion, ignoring the power required for UAV hovering and flying, which is a constant value, the total power is expressed as: P_total=p_r^d+p_h^d+p_r^c+p_h^c. §.§ Problem Formulation In this part, we focus on reducing the total power in the system for a better service. The target is to minimize P_total by jointly optimizing the UAV access scheme and power allocation when the QoS meets users' demands. Hence, the problem is formulated as P0: min_x_i,p_i^d,p_h^d P_total, s.t. 𝐶1: 𝐏𝐫{t_i^d≤ t_i,max^d}≥α_1,∀ i, 𝐶2: 𝐏𝐫{t_h^d≤ t_h,max^d}≥α_2, 𝐶3: 0≤∑_i=1^N x_i≤ m, 𝐶4: x_i∈{0,1},∀ i, 𝐶5: p_i^d≥0,∀ i, 𝐶6: p_h^d≥0, where 𝐶1 and 𝐶2 are the chance constraints, indicating that the transmission delays of UAVs are constrained in a probabilistic manner. In other words, the delays should not be larger than the tolerable delay t_i,max^d and t_h,max^d with probabilities of α_1 and α_2 at a minimum, respectively. Futhermore, α_1, α_2 ∈ (0,1). 𝐶3 denotes that the upper-layer UAV with powerful computation ability can at most access 𝑚 UAVs at the same time. 𝐶4 denotes 𝑥_𝑖 is a binary variable. 𝐶5 and 𝐶6 guarantee that p_i^d and p_h^d are non-negative continuous variables. Due to the chance contraints (<ref>) and (<ref>), the transmission delay t_i^d and t_h^d in P0 are difficult to obtain. However, there exists one effective way to handle the problem with distributionally robust method. Specifically, let ℙ∈𝒫inf represent the lower bound of the probability under the probability distribution ℙ, and (<ref>) and (<ref>) can be rewritten by: ℙ∈𝒫inf 𝐏𝐫_ℙ{t_i^d≤ t_i,max^d}≥α_1, ∀ i∈{1,2,…,N}, and ℙ∈𝒫inf 𝐏𝐫_ℙ{t_h^d≤ t_h,max^d}≥α_2 , respectively, which are distributionally robust chance constraints (DRCCs). § REFORMULATION AND ALGORITHM Since the DRCC problems (<ref>) and (<ref>) are intractable with the uncertainty set, in this section, we employ the CVaR mechanism to handle the issue and then design an algorithm to obtain the final solution. §.§ CVaR Based Mechanism Generally, the value-at-risk (VaR) of a variable u with the safety factor α is defined as the minimal value of v, and u is no more than v at a possibility of α <cit.>, i.e., VaR_α(u)=min{v|P(u ≤ v) ≥α}. It is noted that VaR is non-convex and discontinuous. Based on VaR, we can consequently propose the definition of CVaR, which is defined as the conditional expectation of u when u ≥ VaR_α(u), i.e., CVaR_α(u)=E[u|u≥ VaR_α(u)]. The relationship between VaR and CVaR is shown in Fig. <ref>. It is obvious that CVaR_α(u)≥ VaR_α(u) <cit.>. Furthermore, CVaR is a conservative approximate estimation of loss, which is robust. As is proposed in <cit.>, for a given measurable loss function φ(ξ): ℝ^k →ℝ, CVaR under the safety factor α concerning the probability distribution ℙ on ℝ^k is expressed as: ℙ-CVaR_α(φ(ξ))=β∈ℝinf{β+1/1-α𝔼_ℙ[max(0,φ(ξ)-β)]}. As is shown in Fig. <ref>, according to the definition, we can then obtain the following formula: ℙ{φ(ξ)≤ℙ-CVaR_α(φ(ξ))}≥α. As above, the CVaR constraint can be formed as: ℙ∈𝒫sup ℙ-CVaR_α(φ(ξ))≤0, ∀ℙ∈𝒫⇔ ℙ∈𝒫inf ℙ{φ(ξ)≤0}≥α. Formula (<ref>) represents that the conservative approximation of DRCC on the right side can be constituted by the CVaR constraint on the left side. In other words, CVaR represents a convex function of the random parameter. Then, we can reformulate the CVaR constraint into a tractable SDP form. Lemma 1: Let φ(ξ)=ξ^⊤Θξ+θ^⊤ξ+θ^0 for Θ∈𝕊^k, θ∈ℝ^k, and θ^0∈𝐑, and then, the worst-case CVaR ℙ∈𝒫sup ℙ-CVaR_α(φ(ξ)) can be rewritten into a tractable SDP as follows: inf β+1/1-α⟨Ω,H⟩, s.t. H∈𝕊^k+1, β∈ℝ, H≽0, H-[Θ 1/2θ 1/2θ^⊤ θ^0-β ]≽0, where Ω=[Σ+μμ^⊤ μ μ^⊤ 1 ], Σ∈𝕊^k is the covariance of the random vector ξ, and μ∈ℝ^k is the mean matrix. Furthermore, H is an auxiliary matrix and β is an auxiliary variable. ⟨Ω,H⟩=Tr (ΩH) represents the trace scalar product. H≽0 represents that the matrix H is positive semidefinite<cit.>. Therefore, according to Lemma 1, we transform the DRO into the form of SDP by the CVaR mechanism, which is a tractable and computationally efficient approximation technique. This form can be regarded as an extension of linear programming, and the effectiveness of this solution is proved in previous works <cit.>. §.§ Problem Reformulation As for (<ref>), a DRCC problem, the first-order Taylor expansion of the loss function is φ(ξ_1)=L_iσ_s^2-B_iγ_ip^d_it_i,max^dξ_1^2, where ξ_1 denotes the random parameter |g_i|. CVaR can construct a convex approximation for the chance constraints. According to Lemma 1, we transform (<ref>) into a SDP problem, i.e., inf β_1+1/1-α_1⟨Ω_1,H_1⟩≤0, s.t. H_1∈𝕊^2, β_1∈ℝ, H_1≽0, H_1-[ -B_iγ_ip_i^dt_i,max^d 0 0 L_iσ_s^2-β_1 ]≽0, where H_1 and β_1 are both auxiliary variables. Let μ_g_i=g_i+μ_i present the mean of g_i, we have Ω_1=[σ_i^2+μ_g_iμ_g_i^⊤ μ_g_i μ_g_i^⊤ 1 ]. Likewise, let ξ_2 denote the random parameter |g_h|, and the loss function in (<ref>) can be rewritten as: φ(ξ_2)=∑_i=1^N x_iL_iσ_s^2-B_hγ_hp^d_ht_h,max^dξ_2^2, which can be further presented in a SDP form, i.e., inf β_2+1/1-α_2⟨Ω_2,H_2⟩≤0, s.t. H_2∈𝕊^2, β_2∈ℝ, H_2≽0, H_2-[ -B_hγ_hp^d_ht_h,max^d 0 0 ∑_i=1^Nx_iL_iσ_s^2-β_2 ]≽0. Similarly, H_2 is the auxiliary matrix, and β_2 is the auxiliary variable for the CVaR constraint. μ_g_h=g_h+μ_h is the mean value of g_h. Then we obtain Ω_2=[σ_h^2+μ_g_hμ_g_h^⊤ μ_g_h μ_g_h^⊤ 1 ]. Finally, based on above discussions, the chance-constrained problem (<ref>) under the uncertainty set can be reformulated into a SDP form, i.e., P1: min_x_i,p^d_i,p^d_h, H_1,H_2,β_1,β_2P_total, 𝐬.𝐭. (<ref>)-(<ref>), (<ref>), (<ref>), ∀ i∈{1,2,…,N}, which is still non-convex with the binary variable x_i, and continuous variables p^d_i and p^d_h. The problem is in the form of mixed integer non-linear programming. To tackle with this problem, we design a joint optimization algorithm on access scheme and power allocation, as shown in Algorithm <ref>. §.§ Algorithm Design It is observed that due to the two decision variables x and p^d, problem P1 is still non-convex. To solve this problem, we divide it into two subproblems. The first subproblem is concerning the access scheme x, and the second subproblem is related to the transmission power allocation plan p^d. As shown in step 2 of Algorithm <ref>, in the first subproblem, an initial value p^d is assumed to minimize the computation power consumption and ensure the access scheme, i.e., x. In other words, the lower-layer UAVs make a choice to complete the computation work locally or transform it to the upper-layer UAV based on the task complexity. The problem with integers is solved by MOSEK. Then, in the second subproblem, according to the access scheme x^* that is obtained in step 2, we can further substitute it to solve the transmission power allocation program. Specifically, the subproblem can be tackled by fixing the access scheme, i.e., x^*. Then, in step 3 of Algorithm <ref>, the transmission power is allocated and variables p^d_i as well as p^d_h are obtained. Repeat the process until the solutions of two subproblems converge. The final result, i.e., the total power consumption P_total in the hierarchical UAV-assisted MEC system, can be figured out as optval. § SIMULATION RESULTS Numeral simulations are conducted to evaluate the performance of the proposed algorithm. Briefly speaking, the data lengths are assumed to range from 40kbits to 60kbits. B_i and B_h are assumed as 10MHz and 10MHz, respectively. The task complexities range from 40kcycles to 100kcycles. The CPU computational power consumption coefficient η_i is 5×10^-6cycles/W, and η_h is 10^-6cycles/W. Antenna gains |g_i| and |g_h| are 5. The mean value of errors is 0 and the variance is 0.01 for all. Furthermore, noise σ_s^2 is assumed as 10^-12W. The safety factors are α_1=0.95 and α_2=0.95. To verify the effectiveness and feasibility, we compare the results of the CVaR method and non-robust method without uncertainty errors Δg_i and Δg_h. In detail, Fig. <ref> shows the performance of two methods in total power v.s. the number of lower-layer UAVs. As the number of lower-layer UAVs increases, the total power consumption of the MEC system increase as well. It can be explained that as the scale of the whole problem becomes larger, the users' data collected are larger to be computed and transmitted. Comparing the results of the two methods, it is observed the power consumption in the CVaR method is smaller than that in the non-robust method. It is explained that there is a tolerance in the CVaR method, so that the transmission power of each UAV can be reduced within the tolerant range. Moreover, different from the non-robust method, which optimizes the problems under the perfect assumption which is apparently inconsistent with the real circumstance, the designed algorithm takes into account the uncertainties in reality and is applicable to the real situation. Fig. <ref> shows the individual transmission power of each lower-layer UAV when there exist six UAVs in the lower layer. We compare the result with the non-robust method. For tasks varying in data length, it demonstrates that the CVaR method performs better in transmission power consumption. The conclusion can be drawn that the proposed method, which is considered in a more realistic world, can obtain a better result than that in the ideal situation. Specifically, considering the errors between the real situation and the ideal situation, the designed algorithm only needs to obtain the mean and variance of the gain errors from the historical data, instead of obtaining the true distribution or probability function of the errors. The experimental results verify the effectiveness of the proposed joint optimization algorithm on access scheme and power allocation. In Fig. <ref>, we investigate the performance of total power consumption, i,e., the corresponding system power v.s. lower-layer UAV numbers for different maximum transmission delays. It depicts the influence of the maximum tolerance on the total power. With the increment of maximum delay to be satisfied, the transmission power consumption reduces. It is accounted for the fact that transmission rates become lower for all UAVs when the tolerance transmission delay increases. § CONCLUSIONS In this paper, we study a hierarchical UAV-assisted MEC scenario to optimize the power consumption. To tackle with the potential uncertainty impacted by environmental factors, which is not ignorable in practical circumstances, a DRO problem based on the uncertainty set is proposed. Then, with the CVaR mechanism, the original problem is reformulated into a SDP form. To jointly optimize the reformulation on the access scheme and power allocation, we further design an algorithm to handle it and obtain the final solution. By conducting simulation experiments, the robustness and feasibility of the proposed algorithm compared with the non-robust method are verified. IEEEtran

http://arxiv.org/abs/2303.06858v1

20230313050301

Continuous-Time Zeroth-Order Dynamics with Projection Maps: Model-Free Feedback Optimization with Safety Guarantees

math.OC

Continuous-Time Zeroth-Order Dynamics with Projection Maps: Model-Free Feedback Optimization with Safety Guarantees Xin Chen, Jorge I. Poveda, Na Li Xin Chen is with the MIT Energy Initiative at the Massachusetts Institute of Technology, MA, USA. E-mail: xinch512@mit.edu. Jorge I. Poveda is with ECE Department, University of California, San Diego, CA, USA. E-mail: poveda@ucsd.edu. Na Li is with the School of Engineering and Applied Sciences at Harvard University, MA, USA. E-mail: nali@seas.harvard.edu. This work was supported in part by NSF CAREER: ECCS 2305756, NSF CAREER: ECCS-1553407, NSF EAGER: ECCS-1839632, and AFOSR grant FA9550-22-1-0211. Received: date / Accepted: date =================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================== This paper introduces a class of model-free feedback methods for solving generic constrained optimization problems where the specific mathematical forms of the objective and constraint functions are not available. The proposed methods, termed Projected Zeroth-Order (P-ZO) dynamics, incorporate projection maps into a class of continuous-time model-free dynamics that make use of periodic dithering for the purpose of gradient learning. In particular, the proposed P-ZO algorithms can be interpreted as new extremum-seeking algorithms that autonomously drive an unknown system toward a neighborhood of the set of solutions of an optimization problem using only output feedback, while systematically guaranteeing that the input trajectories remain in a feasible set for all times. In this way, the P-ZO algorithms can properly handle hard and asymptotical constraints in model-free optimization problems without using penalty terms or barrier functions. Moreover, the proposed dynamics have suitable robustness properties with respect to small bounded additive disturbances on the states and dynamics, a property that is fundamental for practical real-world implementations. Additional tracking results for time-varying and switching cost functions are also derived under stronger convexity and smoothness assumptions and using tools from hybrid dynamical systems. Numerical examples are presented throughout the paper to illustrate the above results. Model-free control, zeroth-order methods, constrained optimization, extremum seeking, nonsmooth dynamical systems. § INTRODUCTION This paper studies the design of model-free feedback control algorithms for autonomously steering a plant toward the set of solutions of an optimization problem using high-frequency dither signals. This type of feedback control design has recently attracted considerable attention, due to successful applications in power grids, communication networks, and mobile robots; see <cit.> and references therein. The designing of these controllers for practical applications is particularly challenging because of two major obstacles: one is the lack of accurate models of the system, as many real-world systems are too complex to derive tractable mathematical equations that accurately describe their behavior in unknown or dynamic environments; the other obstacle is to meet safety requirements by properly handling various constraints, including physical laws, control saturation, capacity and budget limits, etc. This paper introduces a class of algorithms that can overcome both of these obstacles and are suitable for the solution of model-free constrained optimization problems describing safety-critical applications. §.§ Literature Review To address the problem of unknown system models, real-time model-free control and optimization schemes have been extensively studied. In these approaches, instead of pre-establishing a complex and often static/stationary system model from first principles and historical data, adaptive control algorithms probe the unknown plant and learn its optimal operation points using real-time output feedback. Such techniques, called extremum seeking (ES) controllers, leverage multi-time scale principles in control theory to steer dynamical systems towards an optimal steady state operation, while preserving closed-loop stability guarantees. ES techniques date back to the early 1920s <cit.>. However, the first general stability analysis for nonlinear systems was presented in the 2000s in <cit.> using averaging-based methods, and in <cit.> using sampled-data approaches based on finite-differences approximations. Since these methods make use of only measurements of the objective function, ES exhibits a close connection with discrete-time zeroth-order optimization dynamics <cit.>. In the continuous-time and finite-dimensional setting, ES algorithms have been further advanced during the last two decades using more general analytical and design techniques for continuous-time systems, see <cit.>. However, despite the theoretical advances and practical applications in ES, one of the major challenges of existing schemes is how to guarantee the systematic satisfaction of hard and asymptotical constraints simultaneously. Hard constraints refer to physical or safety-critical constraints that need to be satisfied by the actions of the controller at all times, e.g. saturation or actuator capacity limits, the generation capacity of a power plant, etc. On the other hand, asymptotic constraints refer to soft physical limits or performance requirements that can be violated temporarily during transient processes but should be met in the long-term steady state, e.g., the thermal limits of power lines and voltage limits imposed by industrial standards, the comfortable temperature ranges required in building climate control, etc. Properly handling these two types of constraints is essential to ensure stability and optimality in real-time optimization algorithms. In the context of ES, most of the approaches and stability results have been developed for unconstrained optimization problems. For optimization problems with hard constraints, the majority of the results and schemes have been limited to methods that integrate barrier or penalty functions in the cost <cit.>, which can limit the type and number of constraints that can be handled by the algorithms. In <cit.>, ES algorithms were introduced to solve optimization problems with constraints defined by certain Euclidean smooth manifolds and Lie Groups. These schemes, however, do not incorporate soft constraints in the optimization problem, and can only handle boundaryless manifolds. Anti-windup techniques in ES for problems that involve saturation were studied in <cit.>, and ES with output constraints were studied in <cit.> using boundary tracing techniques. Switching ES algorithms that emulate sliding-mode techniques were also presented in <cit.> to handle hard constraints in time-varying problems. More recently, an innovative approach that combines safety filters and ES was introduced in <cit.> using control barrier functions and quadratic programming. To handle soft constraints, ES approaches based on saddle flows have also been studied in <cit.>. Finally, more closely related to our setting are the works <cit.>, which considered ES algorithms with certain projection maps for scalar problems <cit.>, and numerical studies of Nash-seeking problems with box constraints <cit.>. §.§ Contributions and Organization This paper introduces a class of continuous-time projected zeroth-order (P-ZO) algorithms for solving generic constrained optimization problems with both hard and asymptotical constraints. Based on ES and two different types of projection maps, the proposed P-ZO methods can be interpreted as model-free feedback controllers that steer a plant towards the set of solutions of an optimization problem with hard and soft constraints, using only measurements of the objective and constraint functions. In particular, we explain the main advantages and innovations of the proposed algorithms below: (a) Model-Free Methods: We study a class of optimization algorithms that use only measurements or evaluations of the objective function and the constraints, i.e., zeroth-order information. In this way, the algorithms do not require knowledge of the mathematical forms of the expressions that define the optimization problem, or their gradients. We show that, under suitable tuning of the control parameters, the trajectories of the proposed model-free ZO algorithms can approximate the behavior of continuous and discontinuous first-order model-based dynamics <cit.>. By exploiting real-time output feedback, the proposed algorithms are inherently robust and adaptive to time-varying unknown disturbances, and for a general class of objective functions, potentially time-varying or switching between a finite collection of candidates. (b) Safety and Optimality via Hard and Soft Constraints: The proposed algorithms can satisfy safety-critical constraints at all times by using continuous or discontinuous projection maps. The systematic incorporation of these mappings into ES vector fields remained mostly unexplored in the literature, and our results show that they can be safely used in feedback loops to solve optimization problems with hard constraints. In the context of ES, to allow for enough exploration via dithering, the projection maps are applied to a shrunken feasible set that can be made arbitrarily close to the nominal feasible set by decreasing the amplitude of the dithers. In this way, the algorithms are able to provide suitable evolution directions near the boundary of the feasible set, achieving a property of “practical safety”, similar in spirit to the one studied in <cit.>. In addition to the hard constraints, the proposed controllers are also able to simultaneously handle soft constraints via primal-dual ES vector fields, thus achieving safety and optimality in a variety of model-free optimization problems. (c) Stability and Performance Guarantees: We leverage averaging and singular perturbation theory for non-smooth (and hybrid) systems, as well as Lyapunov-based arguments, to show that the proposed dynamics can guarantee convergence to an arbitrarily small neighborhood of the optimal set, from arbitrarily large compact sets of initial conditions in the feasible set. Moreover, by exploiting the well-posedness of the dynamics and the optimization problem, the algorithms also guarantee suitable robustness properties with respect to arbitrarily small bounded additive disturbances acting on the states and dynamics of the closed-loop system. This is a fundamental property for practical applications and is non-trivial to achieve in model-free algorithms. We also provide tracking bounds for time-varying optimization problems using (practical) input-to-state stability tools, and we provide stability results for a class of ES problems with unknown switching objective functions, which have remained mostly unexplored in the literature. Earlier, partial results of this paper appeared in the conference paper <cit.>. The results of <cit.> are dedicated only to a particular optimal voltage control problem in power systems using only one of the algorithms studied in this paper. In contrast to <cit.>, in this paper, we consider a generic constrained optimization problem and we study two different types of projection maps (continuous and discontinuous), which require different analytical tools and lead to two different algorithms. Additionally, we present novel tracking results for time-varying optimization problems and problems with switching costs, and we establish robustness guarantees for all the algorithms. Unlike <cit.>, we also present the complete proofs of the results, as well as novel illustrative examples. The remainder of this paper is organized as follows: Section <ref> introduces the notation and the preliminaries. Section <ref> presents the problem formulation. Section <ref> introduces the projected ZO dynamics that incorporate Lipschitz projection maps, and establishes results for static maps, time-varying maps, and switching maps. Section <ref> considers projected gradient-based ZO dynamics with discontinuous projections. Section <ref> presents the analysis and proofs. Numerical experiments are presented throughout the paper to illustrate the main ideas and results. The paper ends with conclusions presented in Section <ref>. § NOTATION AND PRELIMINARIES §.§ Notation We use unbolded lower-case letters for scalars and bolded lower-case letters for column vectors. We use ℝ_+:=[0,+∞) to denote the set of non-negative real values and use to denote a closed unit ball of appropriate dimension. We use ||·|| to denote the Euclidean norm of a vector and use [; ] := [^⊤, ^⊤]^⊤ to denote the column merge of column vectors ,. Given a positive integer n, we define the index set [n]:={1,⋯,n}. The distance between a point ∈^n and a nonempty closed convex set ⊆^n is denoted as ||||_:=∈inf ||-||; and the Euclidean projection of onto the set is defined as 𝒫_():= ∈arg inf || -||. The norm cone to a set at a point ∈ is defined as N_():= {∈ℝ^n: ^⊤( -) ≤ 0, ∀ ∈}. The tangent cone to at a point ∈ is defined as T_():= {∈ℝ^n: ^⊤≤ 0, ∀ ∈ N_() }, which is the polar cone of the normal cone N_(). A continuous function β(r,s): _+×_+→_+ is said to be of class-𝒦ℒ if it is zero at zero, non-decreasing in its first argument r, non-increasing in the second argument s, lim_r→0^+β(r,s)=0 for each s, and lim_s→∞β(r,s)=0 for each r <cit.>. In this paper, we will consider constrained dynamical systems in the form: 𝐱∈ C,   𝐱̇∈ F(𝐱), where 𝐱∈ℝ^n is the state, C is the flow set, and F:ℝ^n⇉ℝ^n is the flow map, which can be set-valued. We use 𝐱̇=d𝐱(t)/dt to denote the time derivative of the function t↦𝐱(t). A function is said to be a (Caratheodory) solution to (<ref>) if 1) t↦𝐱(t) is absolutely continuous on each compact sub-interval of its domain dom(𝐱); 2) 𝐱(0)∈C; and 3) 𝐱̇(t)∈ F(𝐱(t)) and 𝐱(t)∈ C for almost all t∈dom(𝐱) <cit.>. The solution 𝐱 is said to be complete if dom(𝐱)=[0,∞). If the flow map F is single-valued, (<ref>) reduces to an ordinary differential equation. If F is also continuous, solutions 𝐱 to (<ref>) are continuously differentiable functions. In addition, if F is locally Lipschitz, then solutions of (<ref>) are unique. §.§ Preliminaries on Extremum Seeking Control Extremum Seeking (ES) control is a type of adaptive control that is able to steer a plant towards a state that optimizes a particular steady-state performance metric using real-time output feedback. These types of controllers can be seen as continuous-time ZO optimization algorithms with (uniform) convergence and stability guarantees. To explain the rationale behind these algorithms, we consider the optimization problem min_x f(x), where f:ℝ→ℝ is a continuously differentiable function. A standard approach to finding the minimizer of f is to use a gradient descent flow in the form ẋ = -k_x·d f(x)/d x, where the gain k_x defines the rate of evolution of the system. However, when the derivative of f is unknown, gradient flows cannot be directly implemented, and instead, model-free techniques are required. To address this issue, ES approximates the behavior of the gradient flow by adding a high-frequency periodic probing signal ε_a μ̂(t) with amplitude ε_a to the nominal input of the plant. The resulting output y = f(x+ ε_a μ̂(t)), which is assumed to be available for measurements, is then multiplied by the same probing signal μ̂(t), and further normalized by the constant 2/ε_a. The loop is closed with an integrator with a negative gain -k_x, leading to the ES dynamics: ẋ = -k_x 2/ε_a f(x+ε_a μ̂(t)) μ̂(t). When the frequency of μ̂(·) is sufficiently large compared to the rate of evolution k_x, the ES dynamics (<ref>) exhibits a time scale separation property that allows to approximate the behavior of x based on the average of the vector field of (<ref>). For example, consider the use of a sinusoidal signal as the probing single, i.e., μ̂(t) := sin (ω t). With large ω>0 and small ε_a, we consider the Taylor expansion of f: f(x+ε_asin(ω t)) = f(x) + ε_asin(ω t)d f(x)/d x + 𝒪(ε_a^2). By computing the average of the vector field of (<ref>) over one period T=2π/ω of the probing signal, one obtains ẋ = 1/T∫_0^T-k_x2/ε_a f(x+ε_asin(ω t))sin(ω t) dt = -k_x/T∫_0^T2sin^2(ω t)d f(x)/d x + (ε_a) dt , = -k_xd f(x)/d x + (ε_a):=h_ave(x) where (ε_a) denotes high-order terms, bounded on compact sets, that vanish as ε_a→0^+. The average system (<ref>) is essentially an (ε_a)-perturbed gradient descent flow. Under suitable assumptions on f, averaging theory and perturbation theory show that the trajectories of (<ref>) will approximate those of (<ref>) (on compact sets and compact time intervals) as ε_a→0^+ and as ω→0^+ <cit.>. Uniform stability properties of gradient flows can then be leveraged to establish stability results for (<ref>) in the infinite horizon <cit.>. This analysis can also be applied to the multi-variable case using an appropriate choice of the (vector) frequencies ω, and to other architectures using Lie-bracket averaging theory that results in a similar average system <cit.>. § PROBLEM FORMULATION In contrast to (<ref>), in this paper, we consider generic constrained optimization problems in the form: Obj. min_ f() s.t. ∈ g_j()≤ 0, j∈[m], where ∈^n is the decision variable, f:^n→ is the objective function, ⊆^n denotes the feasible set of , and the vector-valued function :=[g_1;g_2;⋯;g_m]:^n→^m describes additional inequality constraints on . The set of the optimal solutions of (<ref>) is denoted as 𝒳^*⊂ℝ^n. Information Availability: We consider the problem setting where the feasible set 𝒳 is known but the mathematical forms of f(·) and (·) are unknown. In this case, one can only query (in real-time) the values of f() and () for a given . That is, the optimization solver can only access the zeroth-order information of f(·) and (·), but not their (first-order) gradients or (second-order) Hessian information. The motivation and rationale of the above problem setting are explained below: 1) The above problem is motivated by the feedback control design that seeks to steer an unknown plant in real time to an optimal solution of problem (<ref>). Here, we model the plant using the static input-to-output maps f(·) and (·) to approximate its dynamics. The validity of this approximation lies in the fact that in many applications the plant is a stable dynamical system that converges to a steady state in a much faster time scale compared to the controller. The steady-state approximation of the plant can then be justified using singular perturbation results for nonlinear systems <cit.>, provided that the time-scale separation is sufficiently large. 2) For many complex engineering systems, their models, captured by the maps f(·) and (·), may be unknown, unavailable, or too costly to estimate. On the other hand, the widespread deployment of smart meters and sensors provides real-time measurements of the system outputs. These measurements can be interpreted as the function evaluations of f(·) and (·) and can be used as the system feedback to circumvent the unknown model information. 3) In problem (<ref>), we distinguish hard constraints, modeled by , and asymptotic constraints, modeled by the inequalities (<ref>). Thus, the constraints imposed by (<ref>) must be satisfied at all times, while inequalities (<ref>) may be violated during the transient process but should be satisfied in the steady states. This paper aims to develop model-free feedback optimization algorithms that are able to solve problem (<ref>) using only zeroth-order information, while simultaneously satisfying hard and asymptotic constraints. To achieve these goals, in Sections <ref> and <ref>, we will study a class of ZO feedback optimization algorithms that are based on ES and incorporate two types of projection maps. To guarantee that problem (<ref>) is well-posed, throughout this paper we will make the following assumptions, which, as discussed later, can be used to relax standard global convexity assumptions considered in the literature of ES. The feasible set 𝒳 is nonempty, closed, and convex. The functions f and g_1,⋯,g_m are convex and continuously differentiable on an open set containing 𝒳. The function f is radially unbounded. Problem (<ref>) has a finite optimum and the Slater's conditions hold. Moreover, the set of optimal solutions ^* is compact. § MODEL-FREE FEEDBACK OPTIMIZATION WITH LIPSCHITZ PROJECTIONS In this section, we introduce a class of gradient-based continuous-time ZO algorithms that incorporate Lipschitz continuous projection maps. We term these algorithms as the projected gradient-based zeroth-order (P-GZO) dynamics. We first study a reduced version of (<ref>) that considers only the hard constraint (<ref>), i.e., we consider the problem: min_∈ f(). For this problem, we establish stability, safety, and tracking results under the P-GZO dynamics. After this, we develop results for the case when f is dynamically drawn from a finite collection of cost functions that share the same minimizer, a problem that emerges in systems with switching plants. Lastly, we further incorporate the constraints (<ref>) using a projected primal-dual zeroth-order (P-PDZO) algorithm. §.§ GZO Dynamics with Lipschitz Projection To solve (<ref>), we develop the following dynamics termed projected gradient zeroth-order (P-GZO) dynamics: 𝐱̇ =k_x (𝒫_𝒳(𝐱-α_x ξ)-𝐱), =1/ε_ξ(-+2/ε_af(𝐱̂) ), =1/ε_ωΛ_κ, where k_x,α_x,ε_ξ,ε_a,ε_ω>0 are tunable parameters. The dynamics (<ref>) incorporates a Lipschitz projection map of the form (<ref>) to ensure that stays within the feasible set . The dynamics (<ref>) estimates the gradient ∇ f with a new state ∈^n, whose dynamics depend on the measured output y=f(), where is the perturbed input defined as := + ε_a . In (<ref>), :ℝ_≥0→^n is a vector-valued periodic dither signal that is generated by the linear dynamic oscillator (<ref>). Specifically, the vector collects all the odd entries of the state ∈^2n, i.e., := [ μ_1, μ_3, μ_5,⋯, μ_2n-1]^⊤. The matrix Λ_κ∈ℝ^2n× 2n in (<ref>) is block diagonal, with the i-th diagonal block given by Λ_κ_i=[[ 0 -2πκ_i; 2πκ_i 0 ]]∈^2× 2, i∈[n], which is parameterized by the tunable constant κ_i>0. Hence, (<ref>) describes n autonomous oscillators, whose solutions can be explicitly computed as μ_i(t) = μ_i(0)sin(2πκ_i/ε_ω t) +μ_i+1(0)cos(2πκ_i/ε_ω t), ∀ i = 1, 3, ⋯, 2n-1, and we choose initial conditions that satisfy μ_i(0)^2 + μ_i+1(0)^2=1. For example, when μ_i(0) = 1 and μ_i+1(0) = 0 for i = 1, 3, ⋯, 2n-1, equation (<ref>) becomes (t):= [sin(2πκ_1/ε_ω t), ⋯, sin(2πκ_2n-1/ε_ω t) ]^⊤. In addition to sinusoidal dither signals, other types of dither signals can also be employed to obtain suitable estimations of the gradient, including triangular waves and square waves, see <cit.>. By incorporating the linear dynamic oscillator (<ref>), the P-GZO dynamics (<ref>) becomes an autonomous system, which facilitates the theoretical analysis. The P-GZO dynamics (<ref>), with the overall state :=(, , ), is restricted to evolve on the following flow set 𝐂_1:=𝒳×ℝ^n×𝕋^n, where 𝕋^n:=𝕊×𝕊×…×𝕊 and 𝕊⊂^2 denotes the unit circle centered at the origin. By construction and Assumption <ref>, the set 𝐂_1 is closed, and it enforces condition (<ref>) on the initialization of the state . Note that the P-GZO dynamics (<ref>) has a Lipschitz continuous vector field on the right-hand side due to the use of a Lipschitz projection mapping. Figure <ref> shows a block diagram of the proposed algorithm. The following assumption will be used throughout this paper to distinguish different dither signal components with different frequency parameters. The parameters κ_i>0 in (<ref>) are rational numbers and satisfy κ_i≠κ_j, and κ_i≠ 2κ_j for all i≠ j. We further explain the proposed P-GZO dynamics (<ref>) with the following remarks. The intuition behind the P-GZO dynamics (<ref>) is that, under a sufficiently small ε_ω and ε_a, the term 2/ε_af(𝐱̂) gives, on average, an (ε_a)-approximation of the gradient ∇ f(). The dynamics (<ref>) with a small ε_ξ behaves like a low-pass filter that takes 2/ε_af(𝐱̂) as the input, and proceeds to output a refined signal , which can mitigate oscillations induced by the dither. Moreover, the introduction of the state with dynamics (<ref>) facilitates the analysis of the projected system via averaging theory by removing from the term that explicitly includes the highly oscillatory signal . Otherwise, the projection in (<ref>) may interfere with the computation of the average dynamics of near the boundary of . (Safety and Optimality). As we will show below in Lemma <ref>, the projection map in (<ref>) guarantees that remains always in the feasible set , and thus the actual decision input always remains in a small tunable (ε_a)-inflation of . This property defines a notion of “practical" safety, similar to those studied in using safety filters and smooth orthogonal projections <cit.>. However, in contrast to other constrained model-free algorithms that use barrier functions <cit.>, orthogonal projections <cit.>, or safety filters <cit.>, the state in (<ref>) can actually hit the boundary of 𝒳 in a finite time, a situation that emerges in problems with saturation constraints. On the other hand, for safety-critical applications where the decision input must stay within , the projection map can be applied to a shrunk feasible set 𝒳_ε_a satisfying 𝒳_ε_a+ε_a𝔹⊆𝒳. In addition, as stated below in Theorem <ref>, when ε_ξ and ε_a are also sufficiently small, the trajectory of (<ref>) will converge to a small tunable neighbor of the optimal set ^* that solves problem (<ref>). §.§ Stability Analysis of the P-GZO Dynamics To study the P-GZO dynamics (<ref>), we first establish the following lemma, showing that the solutions of (<ref>) remain in the flow set 𝐂_1. The proof is provided in Appendix <ref>. Suppose that Assumption <ref> holds. Let 𝐳:=(𝐱,ξ,μ) be a solution of P-GZO dynamics (<ref>) with 𝐳(0)∈𝐂_1. Then, 𝐳(t)∈𝐂_1 and 𝐱̂(t)∈𝒳+ε_a𝔹 for all t∈dom(𝐳). We analyze the stability and convergence properties of the P-GZO dynamics (<ref>) based on the properties of a nominal “target system”, given by 𝐩∈, 𝐩̇=k_x(𝒫_𝒳(𝐩-α_x∇ f(𝐩))-𝐩), which has been well studied in the literature <cit.>. The following theorem, which is the first result of this paper, only relies on assuming the well-posedness of (<ref>) and suitable stability properties for (<ref>). Particular cases where these assumptions are satisfied are discussed afterwards. Suppose that Assumptions <ref>-<ref> hold, and * Every solution of (<ref>) with 𝐩(0)∈𝒳 is complete; * System (<ref>) renders the optimal set 𝒳^* forward invariant and uniformly attractive. Then, for any Δ>ν>0 there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ) there exists ε̂_a>0 such that for all ε_a∈(0,ε̂_a), there exists ε̂_ω>0 such that for all ε_ω∈(0,ε̂_ω), every solution 𝐳 of the P-GZO dynamics with 𝐳(0)∈𝐂_1∩(𝒲_1^*+Δ𝔹)×𝕋^n) is complete and satisfies: (Practical Convergence): lim sup_t→∞ ||𝐱(t)||_𝒳^*≤ν, (Practical Safety): 𝐱(t)∈𝒳, 𝐱̂(t)∈𝒳+ε_a𝔹, ∀ t≥ 0, where 𝒲_1^*:={(𝐱,)∈ℝ^2n:𝐱∈𝒳^*, =∇ f(𝐱)}. The complete proof of Theorem <ref> is presented in Section <ref> as a particular case of a more general result presented later in Theorem <ref>. The result of Theorem <ref> establishes two main properties: 1) convergence from arbitrarily large Δ-compact sets of initial conditions to arbitrarily small ν-neighborhoods of the optimal set, which is typical in zeroth-order algorithms; and 2) the safety result (<ref>) for and that holds for all time t≥ 0. The result does not assume that the feasible set 𝒳 is bounded, but, when this is the case, the results become global. The conditions under which the assumptions (a) and (b) in Theorem <ref> hold for the nominal system (<ref>) have been extensively studied in the literature <cit.>. For example, these two assumptions hold when the objective function f is strictly convex <cit.>, in which case 𝒲_1^* is a singleton. In fact, the result of Theorem <ref> holds even when ∇ f in (<ref>) is replaced by a general strictly monotone mapping, since in this case assumptions (a) and (b) of Theorem <ref> also hold <cit.>. This implies that Theorem <ref> can also be used for decision-making problems in games using the pseudo-gradient instead of the gradient. One of the main limitations of traditional zeroth-order algorithms that emulate gradient descent, such as (<ref>), is that the cost f might not be convex (or gradient-dominated) in ℝ^n, precluding global convergence results. In this case, projection maps can be used to restrict the evolution of the algorithm to “safe” zones 𝒳 where suitable convexity/monotonicity properties hold. This observation is illustrated in Figure <ref>, where a non-convex landscape, with multiple local minima, maxima, and saddle points, is safely optimized in a set 𝒳 where the assumptions of Theorems <ref> and <ref> hold. §.§ Tracking Properties of P-GZO Dynamics For many practical applications, the corresponding optimization problem (<ref>) is not static but time-varying, whose objective and constraints may change over time. This subsection considers the time-varying optimization setting by letting the objective f in (<ref>) depend on a time-varying parameter ∈^p, i.e., we now consider mappings (,)↦ f(,). In addition, is assumed to be generated by an (unknown) exosystem in the form: ∈Θ, =ε_θΠ(), where ε_θ>0 is a parameter that describes the rate of change of θ, Θ⊂ℝ^p is a compact set, and Π(·) is a locally Lipschitz map. System (<ref>) is assumed to make the set Θ forward invariant. As a result, the optimizer ^* is also time-varying and forms a trajectory t↦𝐱^*((t)). We study the tracking performance of the P-GZO dynamics for solving the time-varying version of problem (<ref>). We make the following assumption on the variation of f, which guarantees uniqueness of the optimal trajectory t↦𝐱^*(). There exists a continuously differentiable function :ℝ^p→ℝ^n such that 𝐱^*():=()=argmin_𝐱∈𝒳f(𝐱,), for all ∈Θ. Also, there exist ℓ,γ>0 such that ||∇ f(𝐱,)-∇ f(𝐲,)||≤ℓ||𝐱-𝐲||, f(𝐱,)-f(𝐲,)≥∇_ f(𝐲,)(𝐱-𝐲)+γ/2||𝐱-𝐲||^2, for all 𝐱,𝐲∈ℝ^n and ∈Θ. In addition, there exists M>0 such that ||∂/∂∇_ f(𝐱,) ||≤ M, for all 𝐱∈ℝ^n and all ∈Θ. The conditions (<ref>) and (<ref>) imply the smoothness and strong convexity of f with respect to , respectively, uniformly on θ. They are commonly assumed in time-varying optimization problems and will enable an exponential practical input-to-state stability for the P-GZO dynamics (<ref>). The following theorem states the tracking performance of the P-GZO dynamics (<ref>), while preserving the Practical Safety property (<ref>). The proof is presented in Section <ref>. Consider the system dynamics (<ref>), (<ref>) with the flow set 𝐂_1×Θ. Suppose that Assumptions <ref>-<ref> hold. Then, there exists c>0 such that for any Δ>ν>0, there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for all ε_a∈(0,ε̂_a), there exists ε̂_ω>0 such that for all ε_ω∈(0,ε̂_ω), every solution 𝐳 of the P-GZO dynamics with 𝐳(0)∈𝐂_1∩(( 𝐰^*(0)+Δ𝔹)×𝕋^n) is complete and satisfies the Practical Safety property (<ref>), and also: (Practical Tracking): lim sup_t→∞ ||𝐱(t)-𝐱^*((t))||≤ c·sup_t≥0||(t)||+ν. where 𝐰^*(0):= (𝐱^*((0)), ∇ f(𝐱^*((0))) ). The proof of Theorem <ref> relies on input-to-state stability (ISS) tools for perturbed systems, which have been recently exploited to study other model-free optimization problems, e.g., <cit.>. Due to the continuity of Π(·) and the compactness of Θ, the function t↦(t) is uniformly bounded, and thus the term sup_t≥0||(t)|| in (<ref>) is well-defined and bounded by ε_θcmax_∈ΘΠ(). To illustrate the tracking behaviors of the P-GZO dynamics, we consider a simple problem in the plane, where f(𝐱,)=(x_1-x_1^*())^2+(x_2-x_2^*())^2 and the feasible set is the disk :={∈ℝ^2: (x_1-1.5)^2+x_2^2≤9/4}. Let be generated by the dynamics θ̇_1=ε_θsin(2θ_2), θ̇_2=ε_θ/2cos(θ_1), with ε_θ=1× 10^-3 and let x_i^*():=θ_i, i∈[2]. To ensure strict safety, we use an ε_a-shrunk feasible set _ε_a:= {: (x_1-1.5)^2+x_2^2≤ (3/2-ε_a)^2}. Figure <ref> shows the time-varying optimizer trajectory t↦^*(t) and the solution trajectory t↦(t) of the P-GZO dynamics (<ref>) under frequencies that are not necessarily too large, e.g., ε_ω∼𝒪(10^0), and amplitudes that are not necessarily too small, e.g., ε_a∼𝒪(10^-1), which is a situation that is common in practical applications with computational limitations. It is observed that (t) tends to closely track ^*(t) in the interior and the boundary of 𝒳. §.§ Switching Objective Functions This subsection considers the problem setting with switching objective functions. Depending on the information available to the decision-maker, the objective function in (<ref>) is drawn from a finite collection of functions {f_q(𝐱)}_q∈ Q. The selection of the current function to be optimized at each time t might be performed by an external entity, leading to passive switching, or by the decision-maker, leading to active switching. In both cases, we show that provided the minimizers and critical points coincide across functions, the P-GZO dynamics can safely achieve extremum seeking in a model-free way. Under switching objective functions, the dynamics of the low-pass filter (<ref>) become =1/ε_ξ(-+2/ε_af_q(𝐱̂)), where q is a switching signal that selects from the set of indices Q:={1,2,…,q̅}, with q̅<∞, the function f_q to be used in the model-free algorithm at each time t, see Figure <ref>. This switching signal is generated by the following hybrid dynamical system <cit.>: (q,τ)∈ Q×[0,N_0],   q̇=0,    τ̇∈[0,1/τ_d], (q,τ)∈ Q×[1,N_0],   q^+∈ Q\{q},  τ^+=τ-1, where the state τ is a timer indicating when the signal q is allowed to switch via (<ref>). In (<ref>), τ_d>0 is called the dwell-time, and N_0 is the chatter bound. As shown in <cit.>, the hybrid system (<ref>) guarantees that every switching signal q satisfies an average dwell-time (ADT) constraint. In particular, for every pair of times (t_1,t_2) with t_2>t_1, every solution of (<ref>) satisfies: 𝒮(t_1,t_2)≤1/τ_d(t_2-t_1)+N_0, where 𝒮(t_1,t_2) is the number of switches between times t_1 and t_2. The following theorem establishes the convergence and safety properties of the P-GZO dynamics under switching objectives. For simplicity, we consider the static optimization case when the optimizer ^* is not time-varying but remains the same, and we omit the dependence of on discrete-time indices, which is typical in hybrid systems of the form (<ref>). The proof of Theorem <ref> is provided in Section <ref>. Consider the system dynamics (<ref>), (<ref>), (<ref>), (<ref>). Suppose that all functions in {f_q()}_q∈ Q are strongly convex and smooth, Assumptions <ref> and <ref> hold for each of them, and they share (a) common minimizer: ^*=min f_q(), for all q∈ Q; (b) common critical point: ^*=∇ f_q(^*),  for all q∈ Q. Then, for any Δ>ν>0, there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for all ε_a∈(0,ε̂_a), there exists ε̂_ω>0 such that for all ε_ω∈(0,ε̂_ω), every solution (t) of the P-GZO dynamics (<ref>), (<ref>), (<ref>) with 𝐳(0)∈𝐂_1 ∩(((^*,^*)+Δ𝔹)×𝕋^n) is complete and satisfies the Practical Safety property (<ref>), and also: (Practical Stability under ADT Switching): lim sup_t→∞ ||𝐱(t)-𝐱^*||≤ν. For unconstrained optimization problems, it has been shown in <cit.> and <cit.> that switched ES algorithms are stable when each mode is stable and the switching is sufficiently slow. The novelty of Theorem <ref> lies in the incorporation of constraints into the switched zeroth-order dynamics via projection maps. Moreover, as shown in the analysis, the rate of convergence in (<ref>) and (<ref>) is actually exponential. Real-time optimization problems with switching costs and safety constraints emerge in many engineering problems. For example, in the economic dispatch problem in electric power systems, the generation fuel costs and electricity prices can change over time, but sudden small price changes may not lead to different optimal dispatch solutions. §.§ Projected Primal-Dual ZO Dynamics with Lipschitz Projections We now consider the complete optimization problem (<ref>), including the inequality constraints (<ref>). To solve this problem, we first introduce the dual variable :=(λ_j)_j∈[m]∈_+^m for the inequality constraints (<ref>), and we formulate the saddle point problem (<ref>): max_∈_+^m min_∈𝒳 L(,):=f()+ ^⊤(), where L(,) is the Lagrangian function. Denote := [;], define 𝒴:= 𝒳×_+^m as the feasible set of , and denote ^* as the set of the saddle points that solve (<ref>). By strong duality (implied by Assumptions <ref> and <ref>), the -component of any saddle point ^*:= [^*;^*]∈^* of (<ref>) is an optimal solution to problem (<ref>). Similar to the design of P-GZO dynamics (<ref>) in Section <ref>, we develop the projected primal-dual zeroth-order (P-PDZO) dynamics (<ref>) for the solution of problem (<ref>): = k_x[_ ( - α_x _1 ) -], =k_λ[ __+^m( + α_λ_2 ) -], _1 =1/ε_ξ[ -_1 + 2/ε_a(f() + ^⊤ () )], _2 =1/ε_ξ[ -_2 + () ], =1/ε_ωΛ(κ) , where the parameters are defined in the same way as (<ref>), and and are defined as (<ref>) and (<ref>), respectively. Thus, the P-PDZO dynamics (<ref>) is restricted to evolve in the flow set 𝐂_2=𝒳×ℝ_+^m×ℝ^n×ℝ^m×𝕋^n. The P-PDZO dynamics (<ref>) can be regarded as a generalization of the P-GZO dynamics (<ref>) that further incorporates the inequality constraint (<ref>). Hence, the properties of P-GZO are generally applicable to P-PDZO, such as those mentioned in Remarks <ref> and <ref>. The following lemma states the forward invariance of 𝐂_2, which directly follows by Lemma <ref> by replacing 𝒳 with 𝒳×ℝ_+^m, and thus we omit the proof here. Suppose that Assumption <ref> holds. Let 𝐳:=(𝐱,,_1,_2,) be a solution of the P-PDZO dynamics (<ref>). Then, 𝐳(t)∈𝐂_2 and 𝐱̂(t)∈𝒳+ε_a𝔹 for all t∈dom(𝐳). We study the stability properties of the P-PDZO dynamics (<ref>) based on the stability of the nominal target system: 𝐩̇_1 = k_1[_( 𝐩_1 - α_1( ∇ f(𝐩_1) + ∇(𝐩_1)^⊤_2 ) ) -𝐩_1 ], 𝐩̇_2 = k_2 [__+^m( 𝐩_2 + α_2 (𝐩_1) ) -𝐩_2], where ∇:= [∇^⊤ g_1;⋯;∇^⊤ g_m ] is the Jacobian matrix. The nominal system (<ref>) is a well-known projected saddle flow that has been widely studied in the literature <cit.>. The following theorem shows that the component of the solution of (<ref>) will converge to a neighborhood of the saddle-point set ^* using only zeroth-order information of f and , provided ^* is compact and uniformly globally asymptotically stable (UGAS) under the nominal system (<ref>). The proof is presented in Section <ref>. Let :=[_1;_2], and suppose that Assumptions <ref>-<ref> hold, and: * The saddle point set ^* is compact; * Every solution of (<ref>) with (0)∈ is complete; * System (<ref>) renders the saddle point set ^* forward invariant and uniformly attractive. Then, for any Δ>ν>0, there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for all ε_a∈(0,ε̂_a), there exists ε̂_ω>0 such that for all ε_ω∈(0,ε̂_ω), every solution 𝐳(t) of the P-PDZO dynamics (<ref>) with 𝐳(0)∈𝐂_2∩ ( (𝒲^*_2+Δ𝔹)×𝕋^n) is complete and satisfies: (Practical Stability): lim sup_t→∞ ||𝐲(t)||_𝒴^*≤ν, (Practical Safety): 𝐲(t)∈𝒴, 𝐱̂(t)∈𝒳+ε_a𝔹, ∀ t≥ 0, where 𝒲_2^*:={(𝐲,_1,_2)∈^2(n+m):𝐲∈𝒴^*, _1=∇ f(𝐱) + ∇(𝐱)^⊤, _2 = ()}. The assumption of a compact saddle point set ^* in Theorem <ref> is commonly made in order to use the analysis tools of averaging theory and singular perturbation theory. For many practical applications, the feasible set describes physical capacity limits or control saturation bounds, and thus is naturally compact. In addition, we can replace the feasible region _+^m of the dual state by the feasible box set [0,M_λ]^m with a sufficiently large M_λ such that the saddle point set of practical interest is compact. As discussed in Section II-B, the vanilla ES algorithm (<ref>) emulates the behavior of an 𝒪(ε_a)-perturbed gradient flow. Similarly, the P-GZO dynamics (<ref>) emulate the behavior of an 𝒪(ε_a)-perturbed projected gradient flow, and the P-PDZO dynamics (<ref>) emulate the behavior of an 𝒪(ε_a)-perturbed projected saddle flow. While model-based algorithms of this form have been extensively studied in the literature, continuous-time zeroth-order implementations of these dynamics with stability and safety guarantees were mostly unexplored. Since in many cases the stability properties of (27) (see items (a)-(c) of Theorem 5) are established via the Krasovskii-LaSalle invariance principle, the result of Theorem 4 allows us to establish stability properties for the model-free algorithm with similar generality as their model-based counterparts. § MODEL-FREE FEEDBACK OPTIMIZATION WITH DISCONTINUOUS PROJECTIONS In Section <ref> above, all the ZO algorithms employed the Euclidean projection onto the feasible set 𝒳, which leads to ordinary differential equations (ODEs) with Lipschitz continuous vector fields on the right-hand side. This continuity property facilitates the well-posedness and stability analysis of the ZO dynamics, since the existence and uniqueness of solutions are guaranteed by standard results for ODEs <cit.>. In this section, we turn to the study of another class of projected ZO dynamics that enforce the hard constraints (<ref>) via discontinuous projection maps. This type of projection has been extensively studied in the field of (discontinuous) model-based projected dynamical systems <cit.>. To simplify our presentation, we focus on the reduced problem (<ref>), which does not include the inequality constraints (<ref>). As shown in Section <ref>, such inequality constraints can be handled via primal-dual dynamics or additional penalty terms in the objective f; see <cit.> for details. §.§ GZO Dynamics with Discontinuous Projection To solve the reduced problem (<ref>), we consider the following ZO dynamics: 𝐱̇ =k_x𝒫_T_𝒳(𝐱)(- ), =1/ε_ξ(-+2/ε_af(𝐱̂) ), =1/ε_ωΛ_κ, which are restricted to evolve in the flow set 𝐂_1 defined in (<ref>). In (<ref>), the mapping 𝒫_T_𝒳(𝐱)(·) projects the vector - onto the tangent cone of the feasible set 𝒳 at point , i.e., T_𝒳(𝐱). As a result, the right-hand side of (<ref>) is in general discontinuous, but it guarantees that stays within the feasible set. Figure <ref> shows a block diagram of the dynamics (<ref>). In conjunction with the flow set (<ref>), we term the dynamics (<ref>) as discontinuous projected gradient-based zeroth-order (DP-GZO) dynamics. We study its stability and regularity properties using tools from differential inclusions and the notion of Krasovskii regularization <cit.>: Consider the ODE 𝐳̇=(𝐳), where 𝐳∈𝐂⊂ℝ^n and :ℝ^n→ℝ^n is locally bounded. The Krasovskii regularization of this ODE is the differential inclusion 𝐳∈𝐂,  𝐳̇∈ K(𝐳):= ⋂_ϵ>0con ((𝐳+ϵ𝔹)∩𝐂), where for a set , con() denotes its convex hull and denotes its closure. The existence of solutions for the Krasovskii regularization of (<ref>) is guaranteed by well-posedness of (<ref>) and standard viability results <cit.>. Moreover, it can be shown that system (<ref>) accurately captures the limiting behavior of (<ref>) under arbitrarily small additive perturbations on the states and dynamics via the so-called Hermes solutions <cit.>. This suggests that (<ref>) provides a useful characterization of the solutions of the ODE =() under small perturbations, a setting that naturally emerges in the context of ZO dynamics. The solutions of (<ref>) are also solutions of its Krasovskii regularization, but the converse is not always true. However, under mild regularity assumptions on (<ref>), it turns out that every solution of its Krasovskii regularization is also a standard (i.e., Caratheodory) solution of (<ref>) <cit.>. This fact allows us to study the behaviors of the DP-GZO dynamics (<ref>) based on its nominal target system: 𝐩̇= k_x·𝒫_T_𝒳(𝐩)(-∇ f(𝐩)). The following theorem establishes the stability and (practical) safety properties for the DP-GZO dynamics (<ref>). The proof of Theorem <ref> is provided in Section <ref>. Suppose that Assumptions <ref>-<ref> hold and that f is strictly convex. Then, for any Δ>ν>0, there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for all ε_a∈(0,ε̂_a), there exists ε̂_ω>0 such that for all ε_ω∈(0,ε̂_ω), every solution 𝐳(t) of the DP-ZO dynamics (<ref>) with 𝐳(0)∈𝐂_1∩(𝒲_1^*+Δ𝔹)×𝕋^n) is complete and satisfies the Practical Convergence property (<ref>) and the Practical Safety property (<ref>). Although the results of Theorem <ref> for the DP-GZO dynamics (<ref>) are similar to Theorem <ref> for the P-GZO dynamics (<ref>), their analyses and proof ideas are different, as the DP-ZO dynamics are discontinuous; see Section <ref>. With sufficiently small parameters ε_a,ε_ξ,ε_ω, the 𝐱-trajectories of the DP-GZO dynamics (<ref>) emulate the trajectories of (<ref>). Since any closed and convex set 𝒳⊂ℝ^n is Clarke regular <cit.> and prox-regular <cit.>, and since f is locally Lipschitz under Assumption <ref>, the solutions of (<ref>) and its Krasovskii regularization coincide and are unique <cit.>. Nevertheless, since the uniqueness of solutions is not required in our analysis, the convexity of 𝒳 and the strict convexity of f can be relaxed to mere Clarke regularity and the assumption that all first-order critical points of (<ref>) are optimal and also equilibria of (<ref>). To illustrate the behavior of system (<ref>), we consider a simple example of problem (<ref>), where the feasible set := [-1,1]× [-1,1] is a box constraint and the objective function f is the same as the one in Example <ref>, but now with θ̇_1=-ε_pθ_2 and θ̇_2=ε_pθ_1, ε_p=1× 10^-3. As shown in Figure <ref>, this exosystem makes the the minimizer of f in ℝ^n (i.e., min_ f()) a slowly varying signal that forms a circular trajectory, shown in blue color. It can be observed that the trajectory t↦𝐱(t), shown in green, generated by the DP-GZO dynamics (<ref>) tends to closely track the blue minimizer trajectory, but it stays in the feasible set at all times due to the projection map. Similar to Example <ref>, we have replaced with a shrunk set _ε_a:= [-1+ε_a,1-ε_a]× [-1+ε_a,1-ε_a] in (<ref>) to ensure the actual input ∈ all the time. Here, ε_a=1× 10^-2, so the difference between _ε_a and 𝒳 is almost indistinguishable. The right plot shows the trajectory of each of the components of . §.§ Structural Robustness The ZO algorithms proposed in this paper rely heavily on function evaluations (or system output measurements) to steer the decision variable 𝐱̂ to an optimal solution of problem (<ref>) or (<ref>). Hence, suitable robustness properties are necessary to handle small disturbances and noises that are inevitable in practice. The following result, i.e., Corollary <ref>, indicates that, under the corresponding assumptions of Theorems <ref>-<ref>, all the proposed ZO algorithms (<ref>), (<ref>), (<ref>) are structurally robust to small bounded additive perturbations on the states and dynamics. We rewrite these ZO dynamics as 𝐳∈𝐂, 𝐳̇=(𝐳 ), and consider their perturbed dynamics (<ref>): 𝐳+𝐞∈𝐂,   𝐳̇=(𝐳+𝐞)+𝐞, where is the state of the ZO dynamics, (·) denotes the vector field describing the right-hand side of the dynamics, is the additive noise, and 𝐂 denotes the flow set. Under the assumptions and parameters of Theorems <ref>-<ref>, there exists e̅>0 such that for any additive noise 𝐞(t):[0,+∞)→^n with sup_t≥ 0 ||𝐞(t)||≤e̅, the trajectory (t) of the perturbed ZO dynamics (<ref>) satisfies the respective practical convergence bounds in Theorems <ref>-<ref>. The result is a corollary of Theorems <ref>-<ref> because the convergence, well-posedness, and stability properties of the dynamics imply that, for each sufficiently large compact set of initial conditions K, and fixed parameters of the controller that induce the convergence bounds (which are actually uniform, see Section <ref>), there exists a compact set that is locally asymptotically stable under the nominal dynamics (the so-called Omega-limit set of K), and also semi-globally practical asymptotically stable as e̅→0^+ for the perturbed system (<ref>) <cit.>. However, we note that Corollary <ref> only shows the existence of a sufficiently small e̅, such that any additive noise bounded by e̅ does not change drastically the convergence properties of the ZO algorithms. However, in practice, the computation of this robustness bound is important, but difficult and application-dependent. §.§ Non-Sinusoidal Dither Signals The ZO algorithms above use sinusoidal dither signals t↦(t), generated by the linear dynamic oscillator (<ref>), as the perturbation to estimate the gradients, while other types of dither signals can also be employed <cit.>. However, as discussed in <cit.>, these are not the only types of signals that can be used for the purpose of gradient and Hessian estimation via averaging theory. In particular, we could also consider any vector of piece-wise continuous T-periodic probing signals t↦𝐝(μ(t)) that satisfies max_t∈[0,T]𝐝(t) = 1 and ∫_0^T𝐝(μ(t)) dt = 0;   1/T∫_0^T𝐝(μ(t))𝐝(μ(t))^⊤ dt =1/η_d I_n for some known η_d>0. Common choices of the probing signals include sinusoidal, triangular waves, and square waves defined in the scalar case as 𝐝_sq(μ(t)) 1, t∈ [2kπ, (2k+1)π] -1, t∈ [(2k+1)π, 2(k+1)π] , k∈ℤ. Note that (<ref>) holds with η_d = 2 for sinusoidal dithers, while η_d = 1 for square-wave signals d_sq. Similarly, for the square signals we also require κ_i ≠ (2k+1)κ_j for all k∈ℤ and i≠ j to meet the above. As discussed in <cit.>, a variety of these signals can be generated by autonomous (sometimes hybrid) dynamical systems. Indeed, when t↦𝐝(μ(t)) is not periodic or continuous, but still generates a well-defined average map, generalized averaging theorems can be used to establish stability results without changing the stability analysis. The averaging techniques used in this paper can be applied to these and other non-smooth dither signals by making use of hybrid boundary layer dynamics, see <cit.>. § ANALYSIS AND PROOFS In this section, we present the proofs of our main results. Since the result of Theorem <ref> can be seen as a particular case of Theorem <ref> when the set of inequality constraints (<ref>) is empty, we first present the proof of Theorem <ref>. Subsequently, we show how to adapt this proof to Theorem <ref>. The proofs of Theorems <ref>, <ref>, <ref> are based on the construction of suitable Lyapunov functions, and therefore are presented afterwards. §.§ Proof of Theorem <ref> Let :=[;], :=[_1;_2], and :=[;]. The P-PDZO dynamics (<ref>)-(<ref>) can be written in compact form as = [ ; ] = [ 𝐪_1(,); 1/ε_ξ(- + 𝐪_2(,)) ]:= 𝐪(,), where 𝐪_1 captures the dynamics (<ref>) and _2 is given by _2(𝐲,):=[ 2/ε_a(f()+^⊤() ); () ], where is generated by the oscillator (<ref>), and 𝐱̂, are defined in (<ref>) and (<ref>). We analyze the stability properties of this system using multi-time scale techniques based on averaging theory and singular perturbation theory. We divide the analysis into the following three main steps. Step 1) Behavior of system (<ref>) restricted to compact sets: Let Δ>0 and :=𝒳×ℝ^m_+. Consider the compact set [(^*+Δ𝔹)∩]×Δ for the initial condition (0) and any desired Δ>0 . Here, ^*+Δ𝔹 denotes the union of all sets obtained by taking a closed ball of radius Δ around each point in the saddle point set ^*. By items (a)-(c) in Theorem <ref>, the compact set 𝒴^* is uniformly globally asymptotically stable (UGAS) for the model-based dynamics (<ref>) restricted to evolve in , which is also a forward invariant set. Thus, there exists a class-𝒦ℒ function β such that for any initial condition 𝐩(0)∈, the solutions 𝐩 of (<ref>) satisfy ||𝐩(t)||_^*≤β(||𝐩(0)||_^*, t) for all t≥0. Without loss of generality, let ν∈(0,1) and consider the set :={∈: ||𝐲||_^*≤β( max_∈^*+Δ𝔹 ||||_^*, 0 )+1 }. Note that the set is compact under the assumption that ^* is compact. Due to the boundedness of , there exists a positive constant M_1 such that ⊂ M_1. Let ℓ():=[ ∇ f() + ∑_j=1^m λ_j ∇ g_j(); () ], and note that, by continuity of ℓ, there exists a positive constant M_2>max{Δ,1} such that ||ℓ()|| +1≤ M_2 for all ||||≤ M_1. Denote M_3= M_2+1. We then study the behavior of system (<ref>) restricted to evolve in the compact set × M_3. Step 2) Stability properties of the average dynamics of system (<ref>): Since the solutions of the oscillator (<ref>) are given by (<ref>), and μ_i(0)^2+μ_i+1(0)^2=1 for all i∈{1,3,…,2n-1}, system (<ref>) with small values of ε_ω is in standard form for the application of averaging theory along the trajectories of . The following Lemma <ref> characterizes the average map of 𝐪_2. The proof is presented in Appendix <ref>. The average of t↦_2(,(t)) is given by _2():= 1/T∫_0^T _2(𝐲,(t)) dt = ℓ() +(ε_a), where T is the common period of . Using Lemma <ref>, we obtain the average dynamics of (<ref>): = [ ; ]=[ 𝐪_1(,); 1/ε_ξ(- + ℓ() +(ε_a) ) ], where :=[;]. We study (<ref>) restricted to evolve in the compact set × M_3. We treat the right-hand side of (<ref>) as an (ε_a)-perturbation of a nominal system with ε_a=0. This nominal system (i.e., (<ref>) with ε_a=0) is in the standard form for the application of singular perturbation theory <cit.>, with being the slow state, and being the fast state. The boundary layer dynamics of this nominal system, in the time scale τ=t/ε_ξ, are d /d τ = - +ℓ(), where is kept constant. This linear system (<ref>) has a unique exponentially stable equilibrium point ^*=ℓ(). As a result, the associated reduced system is derived as = _1(, ℓ()), which is exactly the nominal target system (<ref>). Under the assumptions of Theorem <ref>, system (<ref>) renders the set ^* UGAS. By using singular perturbation theory <cit.>, we can conclude that, as ε_ξ→ 0^+, the set ^*×M_3𝔹 is semi-globally practically asymptotically stable (SGPAS) for the unperturbed average system (<ref>) with ε_a=0. Since system (<ref>) has a continuous right-hand side , the perturbed average system (<ref>) also renders the set ^*×M_3𝔹 SGPAS as (ε_ξ,ε_a)→ 0^+, which is stated as Lemma <ref>. There exists β∈𝒦ℒ such that for each ν>0, there exists ε̂_ξ>0 such that for any ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for any ε_a∈ (0,ε̂_a), every solution of the average system (<ref>) (restricted in × M_3) with initial condition (0)∈ [(^*+Δ𝔹)∩]×Δ satisfies ||(t)||_^*≤β(||(0)||_^*, t) +ν/4, for all t∈dom(). Since the average system (<ref>) is restricted in × M_3, we have ||(t)||_M_3 = 0 for all t∈dom(), which implies that ||(t)||_^*× M_3 = ||(t)||_^* for all t∈dom(). Hence, it follows that for all t∈dom(): ||(t)||_^*× M_3≤β(||(0)||_^*× M_3, t) +ν/4. Next, we show the completeness of solutions of the average system (<ref>) by leveraging Lemma <ref>, which follows as a special case of <cit.>. Let k,M_2>0 be given and 𝐮:_+ → M_2 be a continuous function of time. Then, the set M_2 is forward invariant under the dynamics = k·(- +𝐮(t)). Under the initial condition (0)∈ [(^*+Δ𝔹)∩]×Δ, by (<ref>), the trajectory 𝐲̅ of (<ref>) satisfies 𝐲̅(t)∈int() for all t∈dom(). This implies that ||𝐲̅(t)||≤ M_1 and thus ||ℓ(𝐲̅(t)) +(ϵ_a))||< M_2 for all t∈dom(), where we take ||(ϵ_a)||<1 for all ε_a∈ (0,ε̂_a) without loss of generality. Using Lemma <ref>, (t) ∈ M_2⊂int(M_3) for all t≥ 0. Thus, under the given initialization, satisfies (t)∈int(× M_3), ∀ t≥ t_0, and thus it has an unbounded time domain. The following lemma proved in Appendix <ref> will be useful for our results. There exists ε̂_ξ>0 such that for any ε_ξ∈(0,ε̂_ξ), there exists ε̂_a>0 such that for any ε_a∈ (0,ε̂_a), there exists a compact set[See <cit.> for the notion of “Omega-limit set of a set".] Ω(× M_3) that is uniformly globally (pre)-asymptotically stable for the average system (<ref>) restricted to × M_3. Step 3) Link the stability property of the average dynamics (<ref>) to the stability property of the original system (<ref>): Since the set ^*× M_3 is SGPAS for the average system (<ref>) (restricted in × M_3) as (ε_ξ,ε_a)→ 0^+, by averaging theory for perturbed systems <cit.>, it follows that for each pair of (ε_ξ,ε_a) inducing the bound (<ref>), there exists ε̂_ω>0 such that for any ε_ω∈(0,ε̂_ω), the solution of the original system (<ref>) (restricted to × M_3) satisfies ||(t)||_^*× M_3≤β(||(0)||_^*× M_3, t) +ν, for all t∈dom(). Since ||(t)||_^*= ||(t)||_^*× M_3 for all t∈dom(), we obtain the bound (<ref>). The only task left is to show the completeness of solutions of the original system (<ref>). This can be done by using the following lemma, corresponding to <cit.>. Let (ε_ξ,ε_a)>0 take sufficiently small values such that Lemmas <ref> and <ref> hold. Then, for each T,δ>0, there exists ε̂_ω>0 such that for all ε_ω∈ (0,ε̂_ω) and all solutions of the original system (<ref>), with 𝐬(0)∈ [(^*+Δ𝔹)∩]×Δ𝔹, there exists a solution 𝐬̅ of the average system (<ref>), with 𝐬̅(0)∈ [(^*+Δ𝔹)∩]×Δ𝔹 such that 𝐬 and 𝐬̅ are (τ,ε) close. By Lemma <ref>, there exists a 𝒦ℒ-class function β̃ such that every solution of the average system (<ref>) satisfies ||(t)||_Ω( × M_3 )≤β̃( ||(0)||_Ω( × M_3 ),t), for all t∈dom(). Moreover, by (<ref>), the trajectories are complete if (0)∈ [(^*+Δ𝔹)∩]×Δ. Using this bound, the closeness of solutions property of Lemma <ref>, and stability results in averaging theory <cit.>, there exists ε̂_ω such that for all ε_ω∈(0,ε̂_ω), every solution of the original system (<ref>) with (0)∈ [(^*+Δ𝔹)∩]×Δ, satisfies ||(t)||_Ω( × M_3 )≤β̃(||(0)||_Ω( × M_3 ),t)+ν/2, for all t∈dom(𝐬). This implies that (t) ∈int(× M_3) for all t≥0. Therefore, every solution of the original system (<ref>), under the given initialization, has an unbounded time domain. Thus Theorem <ref> is proved. §.§ Proof of Theorem <ref> The proof of Theorem <ref> follows the same steps as the proof of Theorem <ref>. In this case, 𝐪_2:=2/ε_af(𝐱̂), and the function ℓ(·) in (<ref>) becomes ℓ(𝐱)=∇ f(𝐱). Therefore, system (<ref>) is precisely the Lipschitz continuous projected gradient flow (<ref>). Under the conditions (a) and (b) in Theorem <ref>, there exists a class 𝒦ℒ function β such that Lemma 6 holds. The rest proof follows the same steps as the proof of Theorem <ref>. §.§ Proof of Theorem <ref> Following similar computations to those in Lemma <ref>, we compute the average dynamics of (<ref>) along the trajectories of and obtain: 𝐱̇̅̇ =k_x 𝒫_𝒳(𝐱̅-α_x )-k_x𝐱̅, =1/ε_ξ(-+∇_ f(𝐱̅,)+𝒪(ε_a)), =ε_θΠ(), which is restricted to evolve in the set 𝐂_3:=𝒳×ℝ^n×Θ. Next, we establish a key lemma for the average system (<ref>). For system (<ref>) with ε_a=0 and ε_θ>0, there exist α̂_x, k̂_x,ε̂_ξ>0, such that for all α_x∈(0,α̂_x), all k_x∈(0,k̂_x) and all ε_ξ∈(0,ε̂_ξ), every solution satisfies ||(t)-^*((t))||≤ c_1||(0) -^*((0))||e^-c_2 t +csup_0≤τ≤ t||(τ)||, where 𝐬̅:=(𝐱̅,), 𝐬^*():=(𝐱^*(),∇_ f(𝐱^*(),)), and c_1,c_2,c>0. Proof: We introduce the error variables 𝐱̃:=𝐱̅-𝐱^*() and :=-∇_f(𝐱̃+𝐱^*,), which leads to the error dynamics 𝐱̇̃̇ =k_x𝒫_𝒳(𝐱̃+𝐱^*-α_x(+∇_ f(𝐱̃+𝐱^*,)))    -k_x (𝐱̃+𝐱^*)-𝐱̇^*, =-1/ε_ξ-d/dt∇_ f(𝐱̃+𝐱^*,). We study the stability properties of (<ref>) with respect to the origin. To this end, we consider the Lyapunov function: W(𝐱̃,)=(1-λ)V_x(𝐱̃)+λ V_ξ(), where λ∈(0,1), V_x(𝐱̃)=1/2||𝐱̃|||^2 and V_ξ()=1/2||||^2. The function (<ref>) is radially unbounded and positive definite, and it satisfies Ẇ =(1-λ)𝐱̃^⊤𝐱̇̃̇+λ^⊤. Let (,𝐱̃):=∇_ f(𝐱̃+𝐱^*,) and thus = + (,𝐱̃). We rewrite the 𝐱̃-error dynamics (<ref>) compactly as 𝐱̇̃̇= _x(𝐱̃,+(,𝐱̃))-𝐱̇^*. It follows that 𝐱̃^⊤𝐱̇̃̇=𝐱̃^⊤ (_x(𝐱̃,+(,𝐱̃))-𝐱̇^*) = 𝐱̃^⊤ (_x(𝐱̃,+(,𝐱̃))-_x(𝐱̃,(,𝐱̃)))    +𝐱̃^⊤( _x(𝐱̃,(,𝐱̃))-𝐱̇^*) ≤ ||𝐱̃||·||𝐱̇^*||+||𝐱̃||·||_x(𝐱̃,+(,𝐱̃))-_x(𝐱̃,(,𝐱̃))||    +𝐱̃^⊤_x(𝐱̃,(,𝐱̃)) ≤ ||𝐱̃||·||𝐱̇^*||+k_xα_x||𝐱̃||·||||+𝐱̃^⊤_x(𝐱̃,(,𝐱̃)), where we use the fact that 𝒫_𝒳(·) is Lipschitz continuous with unitary Lipschitz constant <cit.>, such that: ||_x(𝐱̃,)-_x(𝐱̃,(,𝐱̃))|| ≤ k_xα_x||-h(,𝐱)||=k_xα_x||||. Using the uniform strong convexity and Lipschitz properties of Assumption <ref> and the same steps of the proof of <cit.>, we obtain ^⊤_x(𝐱̃,(,𝐱̃))≤ -k̃_x||||^2, where k̃_x=α_x(γ-α_xℓ^2/4). Thus, for all α_x∈(0, α̂_x) with α̂_x=4γ/ℓ^2 such that k̃_x is positive, we obtain 𝐱̃^⊤𝐱̇̃̇≤ -k̃_x||𝐱̃||^2+k_xα_x||𝐱̃||·||||+m||||·||||, where m:=max_∈Θ||∇()|| and (·) is defined in (<ref>). Note that m<∞ because Θ is compact and is continuously differentiable by Assumption <ref>. From (<ref>), we have ^⊤=-1/ε_ξ||||^2-^⊤d/dt∇_ f(𝐱̃+𝐱^*,), and note that d/dt∇_f(𝐱̃+𝐱^*,) =H_f(,𝐱̃+𝐱^*)(𝐱̇̃̇+𝐱̇^*) +∂/∂∇_ f(𝐱̃+𝐱^*,), where H_f(·,·) is the Hessian matrix of f. By (<ref>) in Assumption <ref>, we have that ||H_f(,𝐱)||≤ℓ for all 𝐱∈ℝ^n and all ∈Θ, and since 𝐱̇̃̇+𝐱̇^*=_x(𝐱̃,+(,𝐱̃)) by (<ref>), we obtain ||H_f(,𝐱̃+𝐱^*)(𝐱̇̃̇+𝐱̇^*)|| ≤ ℓ ||_x(𝐱̃,)-_x(𝐱̃,(,𝐱̃))+_x(𝐱̃,(,𝐱̃))|| ≤ ℓ k_xα_x||||+ℓ||_x(𝐱̃,(,𝐱̃))|| ≤ℓ k_xα_x||||+cℓ||||, where c:=k_x(2+α_xℓ). The second inequality above is due to (<ref>). The last inequality is because _x(0,(,0))=0 and 1/k_x·|| _x(𝐱̃,(,𝐱̃))-_x(0,(,0))|| ≤ ||𝐱̃||+||𝒫_𝒳(𝐱̅-α_x ∇_ f(𝐱̅,))-𝒫_𝒳(𝐱^*-α_x ∇_ f(𝐱^*,))| ≤ ||𝐱̃||+||𝐱̃||+α_x||∇_ f(𝐱̅,)-∇_ f(𝐱^*,)|| ≤ 2||𝐱̃||+α_x ℓ||𝐱̃|| = c ||||. Combining the above inequalities and using (<ref>), we obtain ^⊤≤ -(1/ε_ξ-ℓ k_xα_x)||||^2+cℓ||𝐱̃|| ||||+M|||| ||||. Using (<ref>) and (<ref>): Ẇ ≤ (1-λ)k_xα_x||𝐱̃|| |||+m(1-λ)|||| ||||+Mλ|||| |||| -k̃_x(1-λ)||𝐱̃||^2-λ(1/ε_ξ-ℓ k_xα_x)||||^2+cℓλ||𝐱̃|| |||| =-𝐲̃^⊤ Q 𝐲̃+c̅||𝐲̃|| ||||, where 𝐲̃=(||𝐱̃||,||||) and Q is given by the 2×2 matrix Q=[[ k̃_x(1-λ)m 1/2(1-λ)k_xα_x+cℓλ/2; 1/2(1-λ)k_xα_x+cℓλ/2 (1/ε_ξ-ℓ k_xα_x)λ ]], and c̅=max{m(1-λ),Mλ}. This matrix is positive definite whenever λ(1-λ)k̃_x m(1/ε_ξ-ℓ k_xα_x)>1/4[(1-λ)k_xα_x+λ cℓ]^2, which is equivalent to 1/ε_ξ>0.25 [(1-λ)k_xα_x+λ cℓ]^2/λ(1-λ)k̃_x m+ℓ k_xα_x. This guarantees the existence of ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ), the matrix Q is positive definite, and therefore W in (<ref>) is a smooth input-to-state stability (ISS) Lyapunov function for the error dynamics (<ref>) with respect to the input . This establishes the bound of Lemma <ref>. Since ||||=ε_ ||Π()||≤ε_σ, where σ=max_∈Θ||Π()||, the result also implies uniform ultimate boundedness (UUB) with ultimate bound proportional to ε_0. ▪ Since Lemma <ref> directly establishes ISS for the nominal error averaged dynamics of (<ref>) with ε_a=0, evolving in 𝐂_3, the perturbed error average system (<ref>) renders the origin semi-globally practically ISS as ε_a→0^+. We can now directly link the stability properties of the average dynamics (<ref>) and the original dynamics (<ref>) via standard averaging results for ISS systems <cit.>. The practical safety property follows directly by Lemma <ref> since the dynamics (<ref>) are independent of (see the proof of Lemma <ref>). §.§ Proof of Theorem <ref> For each constant q, the average dynamics of the P-GZO dynamics (<ref>) are obtained as: 𝐱̇̅̇ =k_x 𝒫_𝒳(𝐱̅-α_x )-k_x𝐱̅, =1/ε_ξ(-+∇ f_q(𝐱̅)+𝒪(ε_a)). The following lemma is the main technical result needed to prove Theorem <ref>. Denote :=(,) and 𝒲_1^*:={(𝐱,)∈ℝ^2n:𝐱∈𝒳^*, =∇ f(𝐱)} Consider the system (<ref>) with ε_a=0. Then, there exist a function V_q and k_x,ε̂_ξ,c_1,c_2>0 such that for all ε_ξ∈(0,ε̂_ξ), we have c_1||||^2_𝒲_1^*≤ V_q()≤ c_2||||^2_𝒲_1^* ⟨∇ V_q(), ⟩≤ -λ_q V_q(), for all ∈𝒳×ℝ^n and all q∈ Q. Proof: For each mode q∈ Q, we consider the Lyapunov function V_q()=1/2||-^*||^2+1/2||-∇ f_q()||^2. Since ||||^2_𝒲_1^*=||-^*||^2+||-∇ f_q(^*)||^2 and ∇ f_q(^*)=^*, ||-∇ f_q()|| = ||-∇ f_q(^*)+∇ f_q(^*)-∇ f_q() || ≤ ||-^*||+ℓ_q ||-^*||, where we use the Lipschitz property of ∇ f_q. Hence, there exists c̅_q>0 such that V_q()≤c̅_q||||^2_𝒲_1^* for all ∈𝒳×ℝ^n. Similarly, ||-∇ f_q(^*) || ≤ ||-∇ f_q()||+||∇ f_q()-∇ f_q(^*)|| ≤ ||-∇ f_q()||+ℓ_q||-^*||. Therefore, ||-∇ f_q(^*)||^2 ≤(||-∇ f_q()||+ℓ_q||-^*||)^2 ≤ 2||-∇ f_q()||^2+2ℓ_q^2|| -^*||^2. Adding ||-^*||^2 to both sides and dividing by 2d_q=max{2,2ℓ_q^2+1} leads to 1/2d_q||||^2_𝒲_1^*≤ V_q(),  ∀ ∈𝒳×ℝ^n. It follows that c_2=max_q∈ Qc̅_q, and c_1=1/2max_q∈ Qd_q. Next, note that V̇_q=(-^*)^⊤-(-∇ f_q())^⊤ H_f_q()+(-∇ f_q())^⊤. The first term of V̇_q satisfies (-^*)^⊤=(-^*)^⊤ (- _r + _r) ≤ (-^*)^⊤_r+||-^*|| ||-_r|| ≤ -k̃_x||-^*||^2+k_xα_x||-^*|| ||-∇ f_q()||, where _r denotes the right-hand side of (<ref>) with =∇ f_q() and we used <cit.> and (<ref>) to obtain the last inequality. Next, using (<ref>), the second term of V̇_q can be bounded as -(-∇ f_q())^⊤ H_f_q() ≤ℓ k_xα_x||-∇ f_q()||^2   +cℓ||-∇ f_q()|| ||-^*||. The last term of V̇_q satisfies (-∇ f_q())^⊤=-1/ε_ξ||-∇ f_q()||^2. Therefore, V̇_q satisfies the same upper bound of (<ref>) with =0, and there exists ε̂_ξ>0 such that for all ε_ξ∈(0,ε̂_ξ) the matrix Q is positive definite and V̇_q≤ -k V_q(),  ∀∈𝒳×ℝ^n. This establishes the result of Lemma <ref>. ▪ The result of Lemma <ref>, in conjunction with <cit.>, guarantees the existence of a τ_d sufficiently large such that, the hybrid dynamical system with flows 𝐱̇̅̇ =k_x 𝒫_𝒳(𝐱̅-α_x )-k_x𝐱̅, =1/ε_ξ(-+∇ f_q̅(𝐱̅)) q̇̅̇ =0 τ̇̅̇ ∈[0,1/τ_d] evolving on the flow set C=(𝒳×ℝ^n)× [0,N_0]×𝒬, and jumps 𝐱̅^+=𝐱,  ξ̅^+=ξ̅,  q̅^+∈ Q,  τ̅^+=τ̅-1. evolving on the jump set D=(𝒳×ℝ^n)× [1,N_0]×𝒬, renders the set {z^*}×[0,N_0]×𝒬 UGAS. In turn, by robustness properties of well-posed hybrid systems, the 𝒪(ε_a)-perturbation of this nominal average system renders the same set SGPAS as ε_a→0^+ <cit.>. Then the result of Theorem <ref> follows directly by an application of averaging theory for perturbed hybrid systems <cit.>. §.§ Proof of Theorem <ref> Since the DP-GZO dynamics (<ref>) is a discontinuous ODE, we consider its Krasovskii regularization defined in (<ref>), which only affects the right-hand side of 𝐱̇: 𝐱̇∈ K(𝐳), =1/ε_ξ(-+2/ϵ_af(𝐱̂)), where 𝐳:=(𝐱,). Since 𝒳 is closed and convex, and f is continuously differentiable, by <cit.>, every solution of (<ref>) is also a solution of the DP-GZO dynamics (<ref>), and vice versa. Moreover, since the dynamics of 𝐱 are independent of , system (<ref>) is in standard form for the application of averaging theory <cit.>. In particular, similar to Lemma <ref>, we compute the average dynamics of (<ref>) along t→(t) and obtain ∈ K(), =1/ε_ξ(-+∇ f(𝐱̅)+𝒪(ε_a)). We will study the stability properties of this perturbed system by analyzing the nominal unperturbed dynamics corresponding to ε_a=0. The following lemma will be key for our results. Under the assumptions of Theorem <ref>, system (<ref>) with ε_a=0 restricted to evolve in the set 𝒳×ℝ^n renders the point ^*=(𝐱^*,∇ f(𝐱^*)) UGAS. Proof: Using the equivalence between Krasovskii and Caratheodory solutions for well-posed projected gradient systems <cit.>, we consider the dynamics =k_x𝒫_T_𝒳()(-),  =1/ε_ξ(-+∇ f()), and the Lyapunov function with λ∈(0,1) V(𝐳̅)=(1-λ)(f(𝐱̅)-f(^*))+λ1/2||-∇ f(𝐱̅)||^2, which is continuously differentiable, radially unbounded, and positive definite with respect to ×^n. We proceed to compute the inner product ⟨∇ V, 𝐳̇̅̇⟩, where 𝐳̇̅̇=(𝐱̇̅̇,). To do this, we use the fact that for any regular set 𝒳, and any 𝐱∈𝒳, ν∈ℝ^n, there exists a unique η∈ N_𝒳() such that 𝒫_T_𝒳()(ν)=ν-η, η^⊤ (ν-η)=0, and ν^⊤ (ν-η)=||ν-η||^2, <cit.>. Thus using ν=-, 𝐯̃(𝐱̅):=ν-η, and (𝐱̅)=-∇ f(𝐱̅) we obtain: ⟨∇ V, 𝐳̇̅̇⟩ = k_x(1-λ)∇ f(𝐱̅)^⊤(𝐱̅)-k_xλ(𝐱̅)^⊤ H_f(𝐱̅) (𝐱̅) +λ(𝐱̅)^⊤ = k_x(1-λ)∇ f(𝐱̅)^⊤(𝐱̅)-k_xλ(𝐱̅)^⊤ H_f(𝐱̅) (𝐱̅) -λ/ε_ξ ||(𝐱̅)||^2. To upper-bound the first term, we note that ∇ f(𝐱̅)^⊤(𝐱̅) =(∇ f(𝐱̅)-)^⊤(𝐱̅)+^⊤(𝐱̅) ≤ ||(𝐱̅)|| ||(𝐱̅)||-ν^⊤ (ν-η) ≤ ||(𝐱̅)|| ||(𝐱̅)||-||ν-η||^2 = ||(𝐱̅)|| ||(𝐱̅)|| -||(𝐱̅)||^2. Moreover, since ∇ f is ℓ-globally Lipschitz by assumption, the second term of V̇ satisfies (-∇ f(𝐱̅))^⊤ H_f(𝐱̅) (𝐱̅) ≤ℓ||(𝐱̅)|| ||(𝐱̅)||. Therefore, defining (𝐳̅):=((𝐱̅),(𝐱̅)), we obtain: ⟨∇ V(𝐳̅), 𝐳̇̅̇⟩≤ - (𝐳̅)^⊤ Q 𝐪̃(𝐳̅), where Q=[[ k_x(1-λ) -1/2(k_xδℓ+k_x(1-λ)); -1/2(k_xλℓ+k_x(1-λ)) λ1/ε_ξ ]]. This matrix is positive definite whenever λ (1-λ) k_x/ε_ξ>1/4[k_xλℓ +k_x(1-λ)]^2, which can be satisfied for sufficiently small values of ε_ξ. Since (𝐳)=0 if and only if 𝐳=(𝐱^*,^*), we obtain that ^* is uniformly globally asymptotically stable (UGAS) for (<ref>). ▪ By equivalence between Krasovskii and Caratheodory solutions, the result of Lemma <ref> guarantees UGAS for the Krasovskii regularization of (<ref>), which is precisely (<ref>) with ε_a=0. Since, by construction, this system is well-posed (outer-semi-continuous, locally bounded, and convex-valued), the set ^* is semi-globally practically asymptotically stable for (<ref>) as ε_a→0^+. The stability result of Theorem <ref> follows directly by the application of averaging results for non-smooth system <cit.>. ▪ § CONCLUSION In this paper, we introduce a class of continuous-time projected zeroth-order (P-ZO) dynamics methods for solving generic constrained optimization problems with both hard and asymptotical constraints, where the mathematical forms of the objective and constraint functions in these problems are unknown and only their function evaluations are available. Therefore, the proposed P-ZO methods can be interpreted as model-free feedback controllers that steer a black-box plant towards optimal steady states described by an optimization problem using only measurement feedback. This paper considers both the continuous and discontinuous projection maps and establishes the stability and robustness results for the proposed P-ZO methods. Moreover, their dynamic tracking performance under the time-varying setting and switching cost functions is analyzed. The focus of this paper is to introduce these P-ZO methods and to analyze their theoretical performance. Our future work will delve into the practical implementation of these P-ZO methods in different applications in engineering and sciences. § AUXILIARY LEMMAS §.§ Proof of Lemma <ref> First, we have (0)∈𝕋^n and μ_i(0)^2+μ_i+1(0)^2=1 for all i∈{1,3,…,2n-1}. Since d/dt(μ_i(t)^2+μ_i+1(t)^2) =2μ_iμ̇_i+2μ_i+1μ̇_i+1 =2 (μ_i,μ_i+1)^⊤Λ_i(μ_i,μ_i+1)=0, 𝕋^n is forward invariant for (t) under (<ref>). Then, following the ideas of <cit.>, we define Φ(𝐱):=|𝐱-𝒫_𝒳(𝐱)|^2 and have Φ̇ =2(𝐱-𝒫_𝒳(𝐱))^⊤𝐱̇ =2k_x(𝐱-𝒫_𝒳(𝐱))^⊤(𝒫_𝒳(𝐱-α_xξ)-𝐱) =-2k_x(𝐱-𝒫_𝒳(𝐱))^⊤(𝐱- 𝒫_𝒳(𝐱))    -2k_x(𝐱-𝒫_𝒳(𝐱))^⊤(𝒫_𝒳(𝐱)-𝒫_𝒳(𝐱-α_x)) ≤ -2k_x |𝐱-𝒫_𝒳(𝐱)|^2 = -2k_xΦ(𝐱), for all 𝐱∈ℝ^n, where the first equality follows by <cit.>, and the inequality in the last step used the property (𝐮-P_𝒳(𝐮))^⊤(P_𝒳(𝐮)-𝐯)≥ 0 for all 𝐮∈ℝ^n and all 𝐯∈𝒳. This implies that Φ̇(𝐱(t))≤ -2k_xΦ(𝐱(t))≤ 0, for all t∈dom(𝐳). We next show that (t)∈ for all t∈dom() given (0)∈ using the contradiction argument. Suppose that there exists t_2>t_1 with t_2,t_1∈dom(𝐳) such that 𝐱(t)∈𝒳 for all t∈[0,t_1] and 𝐱(t)∉𝒳 for all t∈(t_1,t_2]. Then, it follows that Φ(𝐱(t_1))=0 and Φ(𝐱(t_2))>0. But the mean value theorem implies the existence of a t̅∈(t_1,t_2) such that Φ̇(t̅)=Φ(t_2)-Φ(t_1)/t_2-t_1>0, which is a contradiction. Therefore, we conclude that if 𝐳(0)∈𝐂_1, then 𝐳(t)∈𝐂_1 for all t∈dom(𝐳). Since the input 𝐱̂ is defined via (<ref>) and |μ_i(t)|≤ 1 for all i and t≥0, it follows that 𝐱̂(t)∈𝒳+ε_a𝔹 for all t∈dom(𝐳). §.§ Proof of Lemma <ref> First, consider the integration on the first part of _2(y,(t)). By the Taylor expansion of f(·), we have (∀ i∈[n]) 1/T∫_0^T 2/ε_a f(+ ε_a (t)) _i(t) dt = 1/T∫_0^T2/ε_a [f()+ε_a ∇ f()^⊤(t) +(ε_a^2)] _i(t) dt = 1/T∫_0^T 2∑_j=1^n [∂ f()/∂ x_j_j(t)_i(t)] dt +(ε_a) = ∂ f()/∂ x_iη_d/ T∫_0^T _i(t)^2 dt +(ε_a)= ∂ f()/∂ x_i +(ε_a). Similarly, we have (∀ j∈[m], i∈[n]) 1/T∫_0^T2/ε_a λ_jg_j(+ε_a (t)) _i(t) dt= λ_j ∂ g_j()/∂ x_i+(ε_a). As for the integration on the second part of _2(y,μ(t)), i.e., ((t)), each component of this integration is (∀ j∈[m]) 1/T∫_0^T g_j(+ ϵ_a(t)) dt = 1/T∫_0^T g_j() + ϵ_a ∇ g_j()^⊤(t) +(ε_a^2) dt = g_j() + (ε_a^2). Combining these two parts, Lemma <ref> is proved. §.§ Proof of Lemma <ref> Take ε_a sufficiently small such that (ϵ_a)< 1 in (<ref>). For each ν∈(0,1), there exists a time T_1>0 such that for any t≥ T_1, β(Δ, t)≤ν/4. Such T_1 always exists because β is a class-𝒦ℒ function, and thus |(t)|_^*≤ν/2 for t≥ T_1 by (<ref>). In addition, by the exponential input-to-output stability of the fast dynamics in (<ref>), there exists a time T_2>0 such that for any t≥ T_2, every solution of (<ref>) with (0)∈ [(^*+Δ𝔹)∩]×Δ satisfies |ξ̅(t)|≤ν/2 + sup_τ≥ t_0 ||ℓ((τ)) +(ε_a)||≤ν/2+ M_2. Thus, for all t≥max{T_1,T_2}, the trajectory converges to a ν/2-neighborhood of ^*× M_2. Therefore, the Omega-limit set from × M_3 is nonempty and Ω(× M_3)⊂ (^*× M_2) + ν/2⊂int(× M_3). By <cit.>, the set Ω(× M_3) is uniformly globally asymptotically stable for the average system (<ref>) restricted to × M_3. ieeetr Xin Chen is a Postdoctoral Associate in the MIT Energy Initiative at Massachusetts Institute of Technology. He received the Ph.D. degree in electrical engineering from Harvard University, and received the double B.S. degrees in engineering physics and economics and the master’s degree in electrical engineering from Tsinghua University, China. His research interests lie in the intersection of control, learning, and optimization for human-cyber-physical systems, with particular applications to power and energy systems. He was a recipient of the IEEE PES Outstanding Doctoral Dissertation Award in 2023, the Outstanding Student Paper Award in the IEEE Conference on Decision and Control in 2021, the Best Student Paper Award Finalist in the IEEE Conference on Control Technology and Applications in 2018, and the Best Conference Paper Award in the IEEE PES General Meeting in 2016. Jorge I. Poveda is an Assistant Professor in the ECE Department at the University of California, San Diego. He received the M.Sc. and Ph.D. degrees in Electrical and Computer Engineering from UC Santa Barbara in 2016 and 2018, respectively. Before joining UCSD, he was an Assistant Professor at CU Boulder from 2019 until 2022, and a Postdoctoral Fellow at Harvard University during part of 2018. He has received the NSF Research Initiation (CRII) and CAREER awards, the Young Investigator Award from the Air Force Office of Scientific Research, the campus-wide RIO Faculty Fellowship at CU Boulder, and the CCDC Outstanding Scholar Fellowship and Best Ph.D. Thesis awards from UCSB. He has co-authored papers selected as finalists for the Best Student Paper Award at the 2017 and 2021 IEEE Conference on Decision and Control. Na Li is the Gordon McKay Professor of Electrical Engineering and Applied Mathematics in the School of Engineering and Applied Sciences at Harvard University. She received her PhD degree in Control and Dynamical systems from the California Institute of Technology in 2013. In 2014, she was a postdoctoral associate of the Laboratory for Information and Decision Systems at Massachusetts Institute of Technology. She has received the Donald P. Eckman Award, ONR and AFOSR Young Investigator Awards, NSF CAREER Award, the Harvard PSE Accelerator Award, and she was also a Best Student Paper Award finalist in the 2011 IEEE Conference on Decision and Control. She has served as Associate Editor for IEEE Transactions on Automatic Control, Systems and Control Letters, and IEEE Control System Letters.

http://arxiv.org/abs/2303.07141v1

20230313140940

An Improved Baseline Framework for Pose Estimation Challenge at ECCV 2022 Visual Perception for Navigation in Human Environments Workshop

cs.CV

headings 16SubNumber*** Beijing University of Posts and Telecommunications, Beijing, China An Improved Baseline Framework for Pose Estimation Challenge at ECCV 2022 Visual Perception for Navigation in Human Environments Workshop Jiajun Fu, Yonghao Dang, Ruoqi Yin, Shaojie Zhang, Feng Zhou, Wending Zhao, Jianqin YinCorresponding Author March 30, 2023 ========================================================================================================================================= This technical report describes our first-place solution to the pose estimation challenge at ECCV 2022 Visual Perception for Navigation in Human Environments Workshop. In this challenge, we aim to estimate human poses from in-the-wild stitched panoramic images. Our method is built based on Faster R-CNN <cit.> for human detection, and HRNet <cit.> for human pose estimation. We describe technical details for the JRDB-Pose dataset, together with some experimental results. In the competition, we achieved 0.303 OSPA_IOU and 64.047% AP_0.5 on the test set of JRDB-Pose. § INTRODUCTION Pose estimation plays an important role in computer vision, which serves as a fundamental task for high-level human action analysis, such as action recognition <cit.>, action assessment <cit.>, and human-robot interaction <cit.>. In recent years, there has been a growing interest in developing pose estimation models that can accurately detect and estimate human poses in wild scenarios. Although many large-scale datasets <cit.> have been introduced in recent years, no dataset primarily targets robotic perception tasks in social navigation environments, and thus rarely reflects the specific challenges found in human-robot interaction and robot navigation in crowded human environments. The JRDB-Pose dataset <cit.> aims at bridging this gap and sets up a new benchmark for pose estimation. This technical report focuses on adopting image-based pose estimation, where bottom-up and top-down methods are two commonly used approaches. In bottom-up methods, the pose of individual body parts is first estimated and then grouped to form a complete pose. On the other hand, top-down methods first detect the entire body and then estimate the pose of each body part. A main strength of top-down methods is their ability to incorporate high-level information about the body, such as prior knowledge about human anatomy and shape. This makes them more robust to occlusion and deformation of body parts, which can be a challenge for bottom-up methods. Moreover, panoramic images contain lots of people with different scales. Compared with bottom-up methods, top-down methods are more robust to multi-scale targets. Therefore, we adopt a classical top-down method-HRNet <cit.> for human pose estimation. Details are given in the remainder of this report. § METHODS We directly adopted Faster R-CNN for human detection and HRNet for pose estimation[For detailed network architectures, please check the original papers <cit.>.]. These two networks rely heavily on pre-trained models. For the HRNet, a standard initialization setting is to load pre-trained weights from the COCO dataset <cit.>. However, the annotated keypoint locations in the COCO dataset are different from the ones in the JRDB-Pose dataset. To handle this keypoint mismatch, we modified the weights of the last convolution (from the feature space to the keypoint space). For a keypoint in the JRDB-Pose dataset, we find the nearest counterpart(s) in the COCO dataset and copy the last convolution weights for this counterpart. If there are multiple counterparts, we utilize the average weights. We name this nearest-match-based weight initialization. The detailed correspondences are shown in Table <ref>. § EXPERIMENTS §.§ Implementation Details §.§.§ Dataset. For 2022's challenge of human pose estimation, the JRDB-Pose dataset <cit.> provides 56,000 stitched panoramic images and 600,000 bounding box annotations, making JRDB-Pose one of the most extensive publicly available datasets that provide ground truth human body pose annotations. Moreover, these annotations come from in-the-wild videos and incorporate heavily occluded poses, which makes JRDB-Pose challenging and an accurate representation of the real-world environment. For the needs of the challenge, we select training images in the "Bytes-coffee", "Huang", "Huang-hall", "Huang-lane", and "Jordan-hall" as the validation set. The rest images are used as the training set. Please refer to the JRDB-Pose paper <cit.> for the detailed calculation of evaluation metrics. §.§.§ Training. We trained the detection model and pose estimation model, separately. Each model is trained in an end-to-end manner. For the detection model, we trained the Faster R-CNN <cit.> using ResNet-50 <cit.> as the backbone. It adopted SGD with a mini-batch of four on four 3090 GPUs and trained it for 50 epochs with a base learning rate of 0.02, which is decreased by a factor of 10 at epochs 30 and 45. We performed a linear warm-up <cit.> during the first 1500 iterations. MMDetection <cit.> was used for implementation. Besides, we directly utilized the stitched panoramic images and extracted ground truth bounding boxes based on corresponding poses. We performed a random left/right shift and removed bounding boxes across the left and right sides of the image. For the pose estimation model, we trained the HRNet-w48 <cit.> using the pre-trained model on COCO <cit.> as initialization weights. It adopted SGD with a mini-batch of 64 on two 3090 GPUs and trained it for 210 epochs with a base learning rate of 0.0005, which is decreased by a factor of 10 at epochs 170 and 200. We performed a linear warm-up <cit.> during the first 500 iterations. MMPose <cit.> was used for implementation. Besides, we cropped the input images based on corresponding poses and resized the images into 384 × 288. §.§.§ Inference. During testing, we first extracted bounding boxes from the detector model and performed pose estimation. Specifically, we directly input the stitched panoramic images to the Faster R-CNN model. Afterward, we utilized non-maximal suppression (NMS) to filter overlapping detections. The rest detections were resized into 384 × 288 and fed into HRNet-w48 to generate pose estimation results. For each model, we utilize the checkpoint with the best performance on the validation set. §.§ Results. We compare our model with the state-of-the-art methods on the JRDB-Pose test set. Table <ref> shows that our framework outperforms existing methods by a large margin. We show some visualization results in Figure <ref>. §.§ Ablation Studies In this section, we show the effectiveness of the weight initialization and the human detector. For the weight initialization, we evaluated the adopted initialization in the validation set. Specifically, we tested three initialization and the results are presented in Table <ref>. The “ImageNet" indicates initializing image backbone weights from ImageNet <cit.>. This initialization performed the worst since the decoders in HRNet were initialized randomly and get no prior keypoint information. The “COCO + Random" represents COCO weight initialization except for the last convolution, whose weights were initialized randomly. The performance is slightly worse than our initialization method (“COCO + Match"). For the human detector, we tested bounding boxes from different approaches on the test set. The results are shown in Table <ref>. The “Baseline" indicates pose estimation results from the official bounding boxes. There are many redundant boxes and some bounding boxes are small to cover enough image cues. With the NMS and implemented human detector, the performance improves a lot on the test set. [t]0.48 captypetable Impact of weight initialization Method AP_0.5↑ ImageNet 92.892% COCO + Random 96.482% COCO + Match 97.968% [t]0.48 captypetable Impact of human detector Name OSPA_IOU↓ AP_0.5↑ Baseline 0.406 34.186% Ours 0.303 64.047% § CONCLUSION We demonstrate our solution to the pose estimation challenge in this technical report. Despite simple, it has achieved promising results and has the potential for further improvements. splncs

http://arxiv.org/abs/2303.06705v1

20230312165408

Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement

cs.CV

Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement Yuanhao Cai ^1,2, Hao Bian ^1,2,, Jing Lin ^1,2, Haoqian Wang ^1,2,, Radu Timofte ^3,4, Yulun Zhang ^4 ^1 Shenzhen International Graduate School, Tsinghua University, ^2 Shenzhen Institute of Future Media Technology, ^3 University of Würzburg, ^4 ETH Zürich March 30, 2023 =============================================================================================================================================================================================================================================================================== empty When enhancing low-light images, many deep learning algorithms are based on the Retinex theory. However, the Retinex model does not consider the corruptions hidden in the dark or introduced by the light-up process. Besides, these methods usually require a tedious multi-stage training pipeline and rely on convolutional neural networks, showing limitations in capturing long-range dependencies. In this paper, we formulate a simple yet principled One-stage Retinex-based Framework (ORF). ORF first estimates the illumination information to light up the low-light image and then restores the corruption to produce the enhanced image. We design an Illumination-Guided Transformer (IGT) that utilizes illumination representations to direct the modeling of non-local interactions of regions with different lighting conditions. By plugging IGT into ORF, we obtain our algorithm, Retinexformer. Comprehensive quantitative and qualitative experiments demonstrate that our Retinexformer significantly outperforms state-of-the-art methods on seven benchmarks. The user study and application on low-light object detection also reveal the latent practical values of our method. Codes and pre-trained models will be released. § INTRODUCTION Low-light image enhancement is an important yet challenging task in computer vision. It aims to improve the poor visibility and low contrast of low-light images and restore the corruptions (e.g., noise, artifact, color distortion, etc.) hidden in the dark or introduced by the light-up process. These issues challenge not only human visual perception but also other vision tasks like nighttime object detection. Hence, a large number of algorithms have been proposed for low-light image enhancement. Plain methods like histogram equalization and gamma correction tend to produce undesired artifacts because they barely consider the illumination factors. Traditional cognition methods rely on the Retinex theory <cit.> that assumes the color image can be decomposed into two components, i.e., reflectance and illumination. Different from plain methods, traditional methods focus on illumination estimation but usually introduce severe noise or distort color locally because these methods assume that the images are noise- and color distortion-free. This is inconsistent with real under-exposed scenes. With the development of deep learning, convolutional neural networks (CNNs) have been applied in low-light image enhancement. These CNN-based methods are mainly divided into two categories. The first category directly employs a CNN to learn a brute-force mapping function from the low-light image to its normal-light counterpart, thereby ignoring human color perception. This kind of methods lack interpretability and theoretically proven properties. The second category is inspired by the Retinex theory. These methods <cit.> usually suffer from a multi-stage training pipeline. They employ different CNNs to decompose the color image, denoise the reflectance, and adjust the illumination, respectively. These CNNs are first trained independently and then connected together to be finetuned end-to-end. The training process is tedious and time-consuming. In addition, these CNN-based methods show limitations in capturing long-range dependencies and non-local self-similarity, which are critical for image restoration. The recently rising deep learning model, Transformer, may provide a possibility to address this drawback of CNN-based methods. However, directly applying original vision Transformers for low-light image enhancement may encounter an issue. The computational complexity is quadratic to the input spatial size. This computational cost may be unaffordable. Due to this limitation, some CNN-Transformer hybrid algorithms like SNR-Net <cit.> only employ a single global Transformer layer at the lowest spatial resolution of a U-shaped CNN. Thus, the potential of Transformer for low-light image enhancement still remains under-explored. To cope with the above problems, we propose a novel method, Retinexformer, for low-light image enhancement. Firstly, we formulate a simple yet principled One-stage Retinex-based Framework (ORF). We revise the original Retinex model by introducing perturbation terms to the reflectance and illumination for modeling the corruptions. Our ORF estimates the illumination information and uses it to light up the low-light images. Then ORF employs a corruption restorer to suppress noise, artifacts, under-/over-exposure, and color distortion. Different from previous Retinex-based deep learning frameworks that suffer from a tedious multi-stage training pipeline, our ORF is trained end-to-end in a one-stage manner. Secondly, we propose an Illumination-Guided Transformer (IGT) to model the long-range dependencies. The key component of IGT is Illumination-Guided Multi-head Self-Attention (IG-MSA). IG-MSA exploits the illumination representations to direct the computation of self-attention and enhance the interactions between regions of different exposure levels. Finally, we plug IGT into ORF as the corruption restorer to derive our method, Retinexformer. As shown in Fig. <ref>, our Retinexformer surpasses state-of-the-art (SOTA) Retinex-based deep learning methods by large margins on various datasets. Especially on SID <cit.>, SDSD <cit.>-indoor, and LOL-v2 <cit.>-synthetic, the improvements are over 6 dB. Our contributions can be summarized as follows: * We propose a novel Transformer-based algorithm, Retinexformer, for low-light image enhancement. * We formulate a one-stage Retinex-based low-light enhancement framework, ORF, that enjoys an easy one-stage training process and models the corruptions well. * We design a new self-attention mechanism, IG-MSA, that utilizes the illumination information as a key clue to guide the modeling of long-range dependences. * Quantitative and qualitative experiments show that our Retinexformer outperforms SOTA algorithms on seven datasets. The results of user study and low-light detection also suggest the practical values of our method. § RELATED WORK §.§ Low-light Image Enhancement Plain Methods. Plain methods like histogram equalization <cit.> and Gama Correction (GC) <cit.> directly amplify the low visibility and contrast of under-exposed images. Yet, these methods barely consider the illumination factors, making the enhanced images perceptually inconsistent with the real normal-light scenes. Traditional Cognition Methods. Different from plain algorithms, conventional methods <cit.> bear the illumination factors into consideration. They rely on the Retinex theory and treat the reflectance component of the low-light image as a plausible solution of the enhanced result. For example, Guo et al. <cit.> propose to refine the initial estimated illumination map by imposing a structure prior on it. Wang et al. <cit.> design an algorithm NPE to preserve the naturalness of details during the enhancement of non-uniform illumination images. However, these methods naively assume that the low-light images are corruption-free, leading to severe noise and color distortion issues during the enhancement process. Besides, these methods rely on hand-crafted priors, usually requiring careful parameter tweaking and suffering from poor generalization ability. Deep Learning Methods. With the rapid progress of deep learning, CNN <cit.> has been widely used in low-light image enhancement. For instance, Lore et al. <cit.> propose a stacked sparse denoising auto-encoder LLNet for joint low-light enhancement and noise suppression. Wei et al. <cit.> and follow-up works <cit.> combine the Retinex decomposition with deep learning. However, these methods usually suffer from a tedious multi-stage training pipeline. Several CNNs are employed to learn or adjust different components of the Retinex model, respectively. Wang et al. <cit.> propose a one-stage Retinex-based CNN, dubbed DeepUPE, to directly predict the illumination map. Nonetheless, DeepUPE does not consider the corruption factors, leading to amplified noise and color distortion when lighting up under-exposed photos. In addition, these CNN-based methods also show limitations in capturing long-range dependencies. §.§ Vision Transformer The natural language processing model, Transformer, is proposed in  <cit.> for machine translation. In recent years, Transformer and its variants have been applied in many computer vision tasks and achieved impressive results in high-level vision (e.g., image classification <cit.>, semantic segmentation <cit.>, object detection <cit.>, etc.) and low-level vision (e.g., image restoration <cit.>, image synthesis <cit.>, etc.). For example, Xu et al. <cit.> propose an SNR-aware CNN-Transformer hybrid network, SNR-Net, for low-light image enhancement. However, SNR-Net only employs a single global Transformer layer at the lowest resolution of a U-shaped CNN due to the enormous computational costs of the vanilla global Transformer. The potential of Transformer has not been fully explored for low-light image enhancement. § METHOD Fig. <ref> illustrates the overall architecture of our method. As shown in Fig. <ref> (a), our Retinexformer is based on our formulated One-stage Retinex-based Framework (ORF). ORF consists of an illumination estimator (i) and a corruption restorer (ii). We design an Illumination-Guided Transformer (IGT) to play the role of the corruption restorer. As depicted in Fig. <ref> (b), the basic unit of IGT is Illumination-Guided Attention Block (IGAB), which is composed of two layer normalization (LN), an Illumination-Guided Multi-head Self-Attention (IG-MSA) module, and a feed-forward network (FFN). Fig. <ref> (c) shows the details of IG-MSA. §.§ One-stage Retinex-based Framework According to the Retinex theory. A low-light image 𝐈∈ℝ^H× W× 3 can be decomposed into a reflectance image 𝐑∈ℝ^H× W× 3 and an illumination map 𝐋∈ℝ^H× W as 𝐈 = 𝐑⊙𝐋, where ⊙ denotes the element-wise multiplication. This Retinex model assumes 𝐈 is corruption-free, which is inconsistent with the real under-exposed scenes. We analyze that the corruptions mainly steam from two factors. Firstly, the high-ISO and long-exposure imaging settings of dark scenes inevitably introduce noise and artifacts. Secondly, the light-up process may amplify the noise and artifacts and also cause under-/over-exposure and color distortion, as illustrated in the zoomed-in patch i and ii of Fig. <ref> (a). To model the corruptions, we reformulate Eq. (<ref>) by introducing a perturbation term for 𝐑 and 𝐋 respectively, as 𝐈 = (𝐑 + 𝐑̂) ⊙ (𝐋 + 𝐋̂) = 𝐑⊙𝐋 + 𝐑⊙𝐋̂ + 𝐑̂⊙ (𝐋 + 𝐋̂), where 𝐑̂∈ℝ^H× W× 3 and 𝐋̂∈ℝ^H× W denote the perturbations. Similar to <cit.>, we regard 𝐑 as a well-exposed image. To light up 𝐈, we element-wisely multiply the two sides of Eq. (<ref>) by a light-up map 𝐋̅ such that 𝐋̅⊙𝐋 = 1 as 𝐈⊙𝐋̅ = 𝐑 + 𝐑⊙ (𝐋̂⊙𝐋̅) + (𝐑̂⊙ (𝐋 + 𝐋̂)) ⊙𝐋̅, where 𝐑̂⊙ (𝐋 + 𝐋̂) represents the noise and artifacts hidden in the dark scenes and are amplified by 𝐋̅. 𝐑⊙ (𝐋̂⊙𝐋̅) indicates the under-/over-exposure and color distortion caused by the light-up process. We simplify Eq. (<ref>) as 𝐈_lu= 𝐈⊙𝐋̅ = 𝐑 + 𝐂, where 𝐈_lu∈ℝ^H× W× 3 represents the lit-up image and 𝐂∈ℝ^H× W× 3 indicates the overall corruption term. Subsequently, we formulate our ORF as (𝐈_lu, 𝐅_lu) = ℰ(𝐈, 𝐋_p),     𝐈_en = ℛ(𝐈_lu, 𝐅_lu), where ℰ denotes the illumination estimator and ℛ represents the corruption restorer. ℰ takes 𝐈 and its illumination prior map 𝐋_p ∈ℝ^H× W as inputs. 𝐋_p= mean_c(𝐈) where mean_c indicates the operation that calculates the mean values for each pixel along the channel dimension. ℰ outputs the lit-up image 𝐈_lu and light-up feature 𝐅_lu∈ℝ^H× W× C. Then 𝐈_lu and 𝐅_lu are fed into ℛ to restore the corruptions and produce the enhanced image 𝐈_en∈ℝ^H× W× 3. The architecture of ℰ is shown in Fig. <ref> (a) (i). ℰ firstly uses a conv1×1 (convolution with kernel size = 1) to fuse the concatenation of 𝐈 and 𝐋_p. We notice that the well-exposed regions can provide semantic contextual information for under-exposed regions. Thus, a depth-wise separable conv9×9 is adopted to model the interactions of regions with different lighting conditions to generate the light-up feature 𝐅_lu. Then ℰ uses a conv1×1 to aggregate 𝐅_lu to produce the light-up map 𝐋̅∈ℝ^H× W× 3. We set 𝐋̅ as a three-channel RGB tensor instead of a single-channel one like  <cit.> to improve its representation capacity in simulating the nonlinearity across RGB channels for color enhancement. Then 𝐋̅ is used to light up 𝐈 in Eq. (<ref>). Discussion. (i) Different from previous Retinex-based deep learning methods <cit.>, our ORF estimates 𝐋̅ instead of the illumination map 𝐋 because if ORF estimates 𝐋, then the lit-up image will be obtained by an element-wise division (𝐈 ./ 𝐋). Computers are vulnerable to this operation. The values of tensors can be very small (sometimes even equal to 0). The division may easily cause the data overflow issue. Besides, small errors randomly generated by the computer will be amplified by this operation and lead to inaccurate estimation. Hence, modeling 𝐋̅ is more robust. (ii) Previous Retinex-based deep learning methods mainly focus on suppressing the corruptions like noise on the reflectance image, i.e., 𝐑̂ in Eq. (<ref>). They overlook the estimation error on the illumination map, i.e., 𝐋̂ in Eq. (<ref>), thus easily leading to under-/over-exposure and color distortion during the light up process. In contrast, our ORF considers all these corruptions and employs ℛ to restore them all. §.§ Illumination-Guided Transformer Previous deep learning methods mainly rely on CNNs, showing limitations in capturing long-range dependencies. Some CNN-Transformer hybrid works like SNR-Net <cit.> only employ a global Transformer layer at the lowest resolution of a U-shaped CNN due to the enormous computational complexity of global multi-head self-attention (MSA). The potential of Transformer has not been fully explored. To fill this gap, we design an Illumination-Guided Transformer (IGT) to play the role of the corruption restorer ℛ in Eq. (<ref>). Network Structure. As illustrated in Fig. <ref> (a) (ii), IGT adopts a three-scale U-shaped architecture <cit.>. The input of IGT is the lit-up image 𝐈_lu. In the downsampling branch, 𝐈_lu undergoes a conv3×3, an IGAB, a strided conv4×4 (for downscaling the features), two IGABs, and a strided conv4×4 to generate hierarchical features 𝐅_i∈ℝ^H/2^i×W/2^i× 2^iC where i = 0, 1, 2. Then 𝐅_2 passes through two IGABs. Subsequently, a symmetrical structure is designed as the upsampling branch. The deconv2×2 with stride = 2 is exploited to upscale the features. Skip connections are used to alleviate the information loss caused by the downsampling branch. The upsampling branch outputs a residual image 𝐈_re∈ℝ^H× W× 3. Then the enhanced image 𝐈_en is derived by the sum of 𝐈_lu and 𝐈_re, i.e., 𝐈_en = 𝐈_lu + 𝐈_re. IG-MSA. As illustrated in Fig. <ref> (c), the light-up feature 𝐅_lu∈ℝ^H× W× C estimated by ℰ is fed into each IG-MSA of IGT. Please note that Fig. <ref> (c) depicts IG-MSA for the largest scale. For smaller scales, conv4×4 layers with stride = 2 are used to downscale 𝐅_lu to match the spatial size, which is omitted in this figure. As aforementioned, the non-trivial computational cost of global MSA limits the application of Transformer in low-light image enhancement. To tackle this issue, IG-MSA treats a single-channel feature map as a token and then computes the self-attention. Firstly, the input feature 𝐅_in∈ℝ^H× W× C is reshaped into tokens 𝐗∈ℝ^HW× C. Then 𝐗 is split into k heads: 𝐗 = [𝐗_1, 𝐗_2, ⋯, 𝐗_k], where 𝐗_i ∈ℝ^HW× d_k, d_k = C/k, and i = 1, 2, ⋯, k. Note that Fig. <ref> (c) shows the situation with k = 1 and omits some details for simplification. For each head_i, three fully connected (fc) layers without bias are used to linearly project 𝐗_i into query elements 𝐐_i∈ℝ^HW × d_k, key elements 𝐊_i ∈ℝ^HW × d_k, and value elements 𝐕_i ∈ℝ^HW × d_k as 𝐐_i = 𝐗_i𝐖_𝐐_i^T,  𝐊_i = 𝐗_i𝐖_𝐊_i^T,  𝐕_i = 𝐗_i𝐖_𝐕_i^T, where 𝐖_𝐐_i, 𝐖_𝐊_i, and 𝐖_𝐕_i∈ℝ^d_k × d_k represent the learnable parameters of the fc layers and T denotes the matrix transpose. We notice that different regions of the same image may have different lighting conditions. Dark regions usually have severer corruptions and are more difficult to restore. Regions with better lighting conditions can provide semantic contextual representations to help enhance the dark regions. Thus, we use the light-up feature 𝐅_lu encoding illumination information and interactions of regions with different lighting conditions to direct the computation of self-attention. To align with the shape of 𝐗, we also reshape 𝐅_lu into 𝐘∈ℝ^HW× C and split it into k heads: 𝐘 = [𝐘_1, 𝐘_2, ⋯, 𝐘_k], where 𝐘_i ∈ℝ^HW× d_k, i = 1, 2, ⋯, k. Then the self-attention for each head_i is formulated as Attention(𝐐_i, 𝐊_i, 𝐕_i, 𝐘_i) = (𝐘_i ⊙𝐕_i)softmax(𝐊_i^T𝐐_i/α_i), where α_i ∈ℝ^1 is a learnable parameter that adaptively scales the matrix multiplication. Subsequently, k heads are concatenated to pass through an fc layer and then plus a positional encoding 𝐏∈ℝ^HW× C (learnable parameters) to produce the output tokens 𝐗_out∈ℝ^HW× C. Finally, we reshape 𝐗_out to derive the output feature 𝐅_out∈ℝ^H× W× C. Complexity Analysis. We analyze that the computational complexity of our IG-MSA mainly comes from the k computations of the two matrix multiplication in Eq. (<ref>), i.e., ℝ^d_k× HW×ℝ^HW× d_k and ℝ^HW× d_k×ℝ^d_k × d_k. Therefore, the complexity 𝒪(IG-MSA) can be formulated as 𝒪(IG-MSA) = k·[d_k·(d_k· HW) + HW·(d_k· d_k)], = 2HWkd_k^2 = 2HWk(C/k)^2 = 2HWC^2/k. While the complexity of the global MSA (G-MSA) used by some previous CNN-Transformer methods like SNR-Net is 𝒪(G-MSA) = 2(HW)^2C. Compare Eq. (<ref>) with Eq. (<ref>). 𝒪(G-MSA) is quadratic to the input spatial size (HW). This burden is expensive and limits the application of Transformer for low-light image enhancement. Therefore, previous CNN-Transformer hybrid algorithms only employ a G-MSA layer at the lowest spatial resolution of a U-shaped CNN to save the computational costs. In contrast, 𝒪(IG-MSA) is linear to the spatial size. This much lower computational complexity enables our IG-MSA to be plugged into each basic unit IGAB of the network. By this means, the potential of Transformer for low-light image enhancement can be further explored. § EXPERIMENT §.§ Datasets and Implementation Details We eveluate our method on LOL (v1 <cit.> and v2 <cit.>), SID <cit.>, SMID <cit.>, and SDSD <cit.> (indoor and outdoor). LOL. The LOL dataset has two versions: v1 and v2. LOL-v1 provides 485 low-/normal-light image pairs for training and 15 image pairs for testing. Each pair contains a low-light image and its corresponding well-exposed reference image. LOL-v2 is divided into two subsets, i.e., LOL-v2-real and LOL-v2-synthetic. The train set of LOL-v2-real includes 689 low-/normal-light image pairs and the test set contains 100 pairs. Most low-light images are captured from a variety of scenes by changing the ISO and exposure time, while other parameters are fixed. LOL-v2-synthetic is created by analyzing the illumination distribution of low-light images and then synthesizing from RAW images. SID. The subset of SID dataset captured by the Sony α7S II camera is adopted for evaluation. There are 2697 short-/long-exposure RAW image pairs. The low-/normal-light RGB images are obtained by using the same in-camera signal processing of SID <cit.> to transfer the RAW images to RGB domain. 2099 low-/normal-light image pairs are used for training and 598 image pairs are selected for testing. SMID. The SMID benchmark collects 20809 short-/long-exposure RAW image pairs. We transfer the RAW data to low-/normal-light RGB image pairs. 15763 pairs are used for training and the left pairs are adopted for testing. SDSD. The static version of the SDSD dataset is adopted for evaluation. It is captured by a Canon EOS 6D Mark II camera with an ND filter. The SDSD dataset contains two subsets, i.e., SDSD-indoor and SDSD-outdoor. SDSD-indoor includes 62 low-/normal-light video pairs for training and 6 pairs for testing. SDSD-outdoor contains 116 low-/normal-light video pairs for training and 10 video pairs for testing. Implementation Details. We implement Retinexformer by PyTorch <cit.>. The model is trained with the Adam <cit.> optimizer (β_1 = 0.9 and β_2 = 0.999) for 2.5 × 10^5 iterations. The learning rate is initially set to 2× 10^-4 and then steadily decreased to 1× 10^-6 by the cosine annealing scheme <cit.> during the training process. Patches at the size of 128×128 are randomly cropped from the low-/normal-light image pairs as training samples. The batch size is 8. The training data is augmented with random rotation and flipping. The training objective is to minimize the mean absolute error (MAE) between the lit-up image and enhanced image. We adopt the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) <cit.> as the evaluation metrics. §.§ Low-light Image Enhancement Quantitative Results. We compare our method with a wide range of SOTA enhancement algorithms. The PSNR and SSIM results are reported in Tab. <ref>. Our Retinexformer significantly outperforms SOTA methods on the seven datasets while requiring moderate computational and memory costs. When compared with the recent best method SNR-Net, our Retinexformer achieves 0.55, 1.32, 1.53, 1.57, 0.66, 0.33, and 1.18 dB improvements on LOL-v1, LOL-v2-real, LOL-v2-synthetic, SID, SMID, SDSD-indoor, and SDSD-outdoor datasets. However, our Retinexformer only costs 40% (1.61 / 4.01) Parmas and 59% (15.57 / 26.35) FLOPS. When compared with SOTA Retinex-based deep learning methods (including DeepUPE <cit.>, RetinexNet <cit.>, RUAS <cit.>, and KinD <cit.>), our Retinexformer yields 4.30, 4.43, 8.54, 6.00, 3.27, 6.60, and 6.00 dB improvements on the seven benchmarks. Especially on SID and SDSD datasets that are severely corrupted by noise and artifacts, the improvements are over 6 dB, as plotted in Fig. <ref>. When compared with SOTA Transformer-based image restoration algorithms (including IPT <cit.>, Uformer <cit.>, and Restormer <cit.>), our Retinexformer gains by 2.73, 2.86, 4.26, 2.17, 1.95, 3.66, and 2.29 dB on the seven datasets. Nonetheless, Retinexformer requires only 1.4% and 6.2% Params, 0.2% and 10.9% FLOPS of IPT and Restormer. All these results clearly suggest the outstanding effectiveness and efficiency advantage of our Retinexformer. Qualitative Results. The visual results of Retinexformer and SOTA algorithms on the seven benchmarks are shown in Fig. <ref>, <ref>, and <ref>. Please zoom in for a better view. Previous methods either cause color distortion like RUAS in Fig. <ref>, or contain over-/under-exposed regions and fail to suppress the noise like RetinexNet and DeepUPE in Fig. <ref>, or generate blurry images like Restormer and SNR-Net in Fig. <ref>, or introduce black spots and unnatural artifacts like DRBN <cit.> and SNR-Net in Fig. <ref>. In contrast, our Retinexformer can effectively enhance the poor visibility and low contrast, reliably remove the noise without introducing spots and artifacts, and robustly preserve the color. User Study Score. We conduct a user study to quantify the human subjective visual perception quality of the enhanced low-light images from the seven datasets. 23 human subjects are invited to score the visual quality of the enhanced results, independently. These testers are told to observe the results from: (i) whether the results contain under-/over-exposed regions, (ii) whether the results contain color distortion, and (iii) whether the results are corrupted by noise or artifacts. The scores range from 1 (worst) to 5 (best). For each low-light image, we display it and the results enhanced by various algorithms but without their names to the human testers. There are 156 testing images in total. The user study scores are reported in Tab. <ref>. Our Retinexformer achieves the highest score on average. Besides, our results are most favored by the human subjects on LOL-v2-real (L-v2-R), LOL-v2-synthetic (L-v2-S), SID, SMID, and SDSD-outdoor (SD-out) datasets and second most favored on LOL-v1 (L-v1) and SDSD-indoor (SD-in) benchmarks. §.§ Low-light Object Detection Experiment Settings. We conduct low-light object detection experiments on the ExDark <cit.> dataset to compare the preprocessing effects of different enhancement algorithms for high-level vision understanding. The ExDark dataset consists of 7363 under-exposed images annotated with 12 object category bounding boxes. 5890 images are selected for training while the left 1473 images are used for testing. YOLO-v3 <cit.> is employed as the detector and trained from scratch. Different low-light enhancement methods serve as the preprocessing modules with fixed parameters. Quantitative Results. The average precision (AP) scores are listed in Tab. <ref>. Our Retinexformer achieves the highest result on average, 66.1 AP, which is 0.5 AP higher than the recent best self-supervised method SCI <cit.> and 0.8 AP higher than the recent best fully-supervised method SNR-Net <cit.>. Besides, Retinexformer yields the best results on five object categories: bicycle, boat, bottle, cat, and dog. Qualitative Results. Fig. <ref> depicts the visual comparisons of detection results. The detector easily misses some target objects or predicts inaccurate locations on the images enhanced by previous algorithms. In contrast, the detector can reliably predict well-placed bounding boxes to cover all the desired objects on the image enhanced by our Retinexformer. These results demonstrate the superiority of the proposed method in benefiting high-level vision understanding. §.§ Ablation Study We conduct ablation study on the SDSD-outdoor dataset for the good convergence and stable performance of Retinexformer on it. The results are reported in Tab. <ref>. Break-down Ablation. We conduct a break-down ablation to study the effect of each component towards higher performance, as shown in Tab. <ref>. Baseline-1 is derived by removing ORF and IG-MSA from Retinexformer. When we respectively apply ORF and IG-MSA, baseline-1 achieves 1.45 and 2.39 dB improvements. When jointly exploiting the two techniques, baseline-1 gains by 3.37 dB. This evidence suggests the effectiveness of our ORF and IG-MSA. One-stage Retinex-based Framework. We conduct an ablation to study ORF. The results are listed in Tab. <ref>. We first remove ORF from Retinexformer and set the input of ℛ as 𝐈_lu = 𝐈. The model yields 28.86 dB. Then we apply ORF but set ℰ to estimate the illumination map 𝐋. The input of ℛ is 𝐈./𝐋 where ./ indicates the element-wise division. To avoid exceptions thrown by computer, we add 𝐋 with a small constant ϵ = 1×10^-4. Yet, as analyzed in Sec. <ref>, the computer is vulnerable to the division of small values. Thus, the model obtains a limited improvement of 0.11 dB. To tackle this issue, we estimate the light-up map 𝐋̅ and set the input of ℛ as 𝐈_lu = 𝐈⊙𝐋̅. The model gains by 0.40 dB. After using 𝐅_lu to direct ℛ, the model continues to achieve an improvement of 0.58 dB in PSNR and 0.007 in SSIM. Self-Attention Scheme. We conduct an ablation to study the effect of the self-attention scheme. The results are reported in Tab. <ref>. Baseline-2 is obtained by removing IG-MSA from Retinexformer. For fair comparison, we plug the global MSA (G-MSA) used by previous CNN-Transformer hybrid methods into each basic unit of ℛ. The input feature maps of G-MSA are downscaled into 1/4 size to avoid out of memory. We also compare our IG-MSA with local window-based MSA (W-MSA) proposed by Swin Transformer <cit.>. As listed in Tab. <ref>, our IG-MSA surpasses G-MSA and W-MSA by 1.41 and 1.34 dB while costing 2.08G and 0.86G FLOPS less. These results demonstrate the cost-effectiveness advantage of the proposed IG-MSA. § CONCLUSION In this paper, we propose a novel Transformer-based method, Retinexformer, for low-light image enhancement. We start from the Retinex theory. By analyzing the corruptions hidden in the under-exposed scenes and caused by the light-up process, we introduce perturbation terms into the original Retinex model and formulate a new Retinex-based framework, ORF. Then we design an IGT that utilizes the illumination information captured by ORF to direct the modeling of long-range dependences and interactions of regions with different lighting conditions. Finally, our Retinexformer is derived by plugging IGT into ORF. Extensive quantitative and qualitative experiments show that our Retinexformer dramatically outperforms SOTA methods on seven datasets. The results of user study and low-light detection also demonstrate the practical values of our method. ieee_fullname

http://arxiv.org/abs/2303.06805v1

20230313012331

Lie algebras of differential operators for Matrix valued Laguerre type polynomials

math.CA

A.L. Gallo, P. Román]Andrea L. Gallo, Pablo Román March 30, 2023 2020 Mathematics Subject Classification. Primary 33C45. Partially supported by CONICET, FONCyT and SECyT-UNC We study algebras of differential and difference operators acting on matrix valued orthogonal polynomials (MVOPs) with respect to a weight matrix of the form W^(ν)_ϕ(x) = x^νe^-ϕ(x) W^(ν)_pol(x), where ν>0, W^(ν)_pol(x) is certain matrix valued polynomial and ϕ an entire function. We introduce a pair differential operators 𝒟, 𝒟^† which are mutually adjoint with respect to the matrix inner product induced by W^(ν)_ϕ(x). We prove that the Lie algebra generated by 𝒟 and 𝒟^† is finite dimensional if and only if ϕ is a polynomial, giving a partial answer to a problem by M. Ismail. In the case ϕ polynomial, we describe the structure of this Lie algebra. The case ϕ(x)=x, is discussed in detail. We derive difference and differential relations for the MVOPs. We give explicit expressions for the entries of the MVOPs in terms of classical Laguerre and Dual Hahn polynomials. Lie algebras of differential operators for Matrix valued Laguerre type polynomials [ ================================================================================== § INTRODUCTION The theory of matrix valued orthogonal polynomials (MVOPs) was initiated by Krein 1940s, and it has since been used in various areas of mathematics and mathematical physics. These areas include spectral theory, scattering theory, tiling problems, integrable systems, and stochastic processes. For further details and insights on these subjects, refer to <cit.>, <cit.>, <cit.>, <cit.>, <cit.>, <cit.>, and the references therein. Significant progress has been made in the past two decades towards understanding how the differential and algebraic properties of classical scalar orthogonal polynomials can be extended to the matrix valued setting. A fundamental role has been played by the connection between harmonic analysis of matrix valued functions on compact symmetric pairs and matrix valued orthogonal polynomials. In <cit.>, A. Durán poses the problem of determining families of MVOPs which are eigenfunctions of a suitable second order differential operator. In the scalar case, the answer to this problem is a classical result due to Bochner <cit.>. The only families with this property are those of Hermite, Laguerre and Jacobi. The matrix valued setting turns out to be much more involved. The first explicit examples appeared in connection with spherical functions of the compact symmetric pair (SU(3),U(2)). Following <cit.>, a direct approach was taken in <cit.>, <cit.> for the case of (SU(2) ×SU(2), diag), leading to a general set-up in the context of multiplicity free pairs <cit.>. In this context, certain properties of the orthogonal polynomials, such as orthogonality, recurrence relations, and differential equations, are understood in terms of the representation theory of the corresponding symmetric spaces. Recently, Casper and Yakimov developed a framework in <cit.> to solve the matrix Bochner problem. This involves the classification of all N × N weight matrix W(x) whose associated MVOPs are eigenfunctions of a second-order differential operator. Given N ∈ℕ we consider a matrix valued function W: [a,b] → M_N(ℂ) such that W(x) is positive definite for all x∈ [a,b] and W has finite moments of all order. In such a case, we say that W is a weight function, which induces matrix valued inner product ⟨ P,Q ⟩=∫_a^b P(x)W(x)Q(x)^*dx ∈ M_N(), such that for all P,Q,R∈ M_N()[x], T∈ M_N(C) and a,b∈ℂ the following properties are satisfied ⟨ aP+bQ,R⟩=a⟨ P,R⟩+b⟨ Q,R⟩, ⟨ TP,Q⟩=T⟨ P,Q⟩, ⟨ P,Q⟩^*=⟨ Q,P⟩. Moreover ⟨ P,P ⟩ = 0 if and only if P=0. Using standard arguments, it can be shown that there exists a unique sequence (P(x,n))_n monic MVOPs with respect to W in the following sense: ⟨ P(x,n), P(x,m)⟩ = ℋ(n) δ_n,m. where the squared norm ℋ(n) is a positive definite matrix. By orthogonality, the polynomials P(x,n)'s satisfy the following three-term recurrence relation: xP(x,n) = P(x,n+1) + B(n) P(x,n) + C(n) P(x,n-1) where B(n), C(n) ∈ M_N(ℂ) and n ≥ 1. Notice that B(n) and C(n) satisfy B(n)=X(n)-X(n+1), C(n)= ℋ(n) ℋ(n-1)^-1, where X(n) is the one-but-leading coefficient of P(x,n) and ℋ(n) as in (<ref>). Moreover, for n ≥ 2, let Y(n) denotes the second-but-leading coefficient of P(x,n). Then Y(n)=Y(n+1)+B(n)X(n)+C(n). In <cit.>, the authors studied difference–differential relations for a specific class of MVOPs associated with the weight W(x)=e^-v(x)e^xAe^xA^∗, where x ∈ℝ, v(x) is a scalar polynomial of even degree, and A is a constant matrix. There is a way of obtaining information about the matrix orthogonal polynomials by investigating two mutually adjoint operators 𝒟 and 𝒟^†. If v(x) is a polynomial of degree two, in addition to 𝒟 and 𝒟^†, there exists a second order differential operator D having the MVOPs as eigenfunctions. It turns out that 𝒟, 𝒟^† and D generate a finite dimensional Lie algebra which is isomorphic to the Lie algebra of the oscillator group. The Casimir operator for this algebra is given explicitly and used to obtain information of the MVOPs. In this work, we solve the analogous problem for Laguerre-type weights. This case is more involved than the previous one due to the structure of the associated Lie algebra and the non-diagonality of certain formulas that involve W. In the scalar case, this problem is closely related to <cit.>. Here Ismail proposed to study the finite dimensionality of certain Lie algebra generated by a pair of differential operators which are mutually adjoint respect to a Laguerre-type weight. More precisely, given the scalar weight w_1(x)=x^αe^-ϕ(x) with x>0, α>1 and differential operators 𝒟_1,n=x∂_x +x B_n(x), 𝒟_2,n=-x∂_x + x B_n(x)+xϕ'(x), where {B_n} is a sequence of scalar polynomials, the problem asks to prove that “The Lie algebra generated by 𝒟_1,n and 𝒟_2,n is finite dimensional if and only if ϕ is a polynomial”. In this paper we provide a partial answer to this problem in the context of matrix valued orthogonal polynomials. We give an explicit matrix valued weight W^(ν)_ϕ(x) =x^ν e^-ϕ(x) W^(ν)_pol(x), where W^(ν)_pol(x) is a matrix polynomial depending on ν, and differential operators 𝒟 = ∂_x x + x(A-1), 𝒟^† = -∂_x x - (1+ν+J)+xϕ'(x)-x. In this case, we prove that the Lie algebra generated by is finite dimensional if and only if ϕ is a polynomial. As a consequence, this solves <cit.> when B_n(x)=-1 for all n∈ℕ. §.§ Outline and main results In Section 2 we recall some preliminaries. In particular, we introduce the left and right Fourier algebras related to the sequence of monic MVOPs. In Section 3 for a given analytic function ϕ on a neighborhood of the interval [0,∞), we introduce a Laguerre type weight W^(ν)_ϕ and the operators 𝒟 , 𝒟^† and prove that they are mutually adjoint with respect to W^(ν)_ϕ(x). For the MVOPs {P_n} respect to W^(ν)_ϕ, we find discrete operators M, M^† associated to 𝒟, 𝒟^† respectively, given by the relations M · P_n = P_n ·𝒟 and M^†· P_n = P_n ·𝒟^†. In Section 4 we study the Lie algebra 𝔤_ϕ generated by the differential operators 𝒟, 𝒟^†. We prove that 𝔤_ϕ is finite dimensional if and only if ϕ is a polynomial. Also, for this family of Lie algebras {𝔤_ϕ} we obtain that 𝔤_ϕ=ℂ^2⊕𝔥 and 𝔥 is a solvable Lie algebra with nilradical of codimension one. Moreover, we obtain a classification of this family of Lie algebra up to isomorphisms. In Section 5 we give an explicitly expression for 𝒟, 𝒟^†, M and M^† in the case ϕ(x)=x. In this case, we also find a symmetric second-order differential operator D which have {P_n} as eigenfunctions. We describe the Lie algebra 𝒜 generated by 𝒟, 𝒟^† and D, 𝒜=𝒵_𝒜⊕ [𝒜,𝒜] where 𝒵_𝒜=2 and [𝒜,𝒜] is isomorphic to SL(2,ℂ). Also, we obtain some relations between ℋ_n, B_n and C_n. In Section 6, we consider the polynomials R(x,n)=K_n P(x,n)e^xA where P(x,n) are the MVOPs associated with the weight W^(ν) and K_n certain lower triangular matrices. Using the operator D, we show that the matrix entries of R_n can be put in terms of generalized Laguerre polynomials and a family of constants ξ(n,i,j)'s. Finally, we give two-terms recursions for the constants ξ(n,i,j)'s and for the squared norm ℋ_n. Finally, in Section 7, in the case A=- ∑_k=1^N-1 E_k+1,k, and δ^k>0 satisfying two non-linear conditions (related to Pearson's equations), we show that the constants ξ(n,i,j)'s are written in terms of dual Hanh polynomials. § PRELIMINARIES This section presents the left and right Fourier algebras associated with the sequence of monic MVOPs, as developed by Casper and Yakimov in <cit.>. The results discussed in this section have been previously covered in a more comprehensive context in <cit.>. Let Q(x,n) be a function Q:ℂ×ℕ_0 → M_N(ℂ) such that Q(x,n) is a rational function of x for fixed n. A differential operator of the form =∑_j=0^n ∂_x^j F_j(x), ∂_x^j := d^jdx^j, where F_j:ℂ→ M_N(ℂ) is a rational function of x, acts on Q from the right by (Q·)(x,n) = ∑_j=0^n (∂_x^jQ)(x,n) F_j(x). The algebra of all differential operators of the form (<ref>) will be denoted by ℳ_N. In addition to the right action by differential operators, we also consider a left action on Q by difference operators on the variable n. For j∈ℤ, let δ^j be the discrete operator which acts on a sequence A:ℕ_0 → M_N(ℂ) by (δ^j · A)(n)=A(n+j). Here we assume that the value of a sequence at a negative integer is equal to zero. For given sequences A_-ℓ,…,A_k, a discrete operator of the form M=∑_j=-ℓ^k A_j(n) δ^j, acts on Q from the left by (M · Q)(x,n) = ∑_j=-ℓ^k A_j(n) (δ^j· Q)(x,n) = ∑_j=-ℓ^k A_j(n) Q(x,n+j). We shall denote the algebra of difference operators (<ref>) by 𝒩_N. As in <cit.> we define: The left and right Fourier algebras are given by: ℱ_L(P) ={ M∈𝒩_N ∃ ∈ℳ_N, M· P = P·}⊂𝒩_N, ℱ_R(P) ={∈ℳ_N ∃ M∈𝒩_N, M· P = P·}⊂ℳ_N. The definition of the Fourier algebras directly implies a connection between the elements of ℱ_L(P) and ℱ_R(P). Moreover, the map φℱ_L(P) →ℱ_R(P), defined by M· P = P ·φ(M), is an algebra isomorphism. In <cit.> this map is called the generalized Fourier map. More precisely, M_1M_2· P = P ·φ(M_1)φ(M_2) for all M_1,M_2∈ℱ_L(P). On the other hand, by the definition of φ, we have that M_1M_2· P = P·φ(M_1M_2). In this context, the three term recurrence relation (<ref>) can be written as xP = P· x = L· P, where L=δ + B(n) + C(n)δ^-1. Therefore x∈ℱ_R, L∈ℱ_L and φ(L)=x. For every polynomial v ∈ℂ[x], we have P· v(x) =P· v(φ(L))= v(L)· P. On of the crucial results from <cit.> is the existence of an adjoint operation † in the Fourier algebras ℱ_L(P) and ℱ_R(P) as described in <cit.>. To define the adjoint operation in ℱ_L(P), we initially observe that the algebra of difference operators 𝒩_N has a ∗-operation defined as follows: ( ∑_j=-ℓ^k A_j(n) δ^j )^∗ = ∑_j=-ℓ^k A_j(n-j)^∗ δ^-j, where A_j(n-j)^∗ is the conjugate transpose of A_j(n-j). Now, the adjoint of M∈𝒩_N is given by M^† = ℋ(n) M^∗ℋ(n)^-1, where ℋ(n) is the squared norm which we view as an difference operator of order zero. The following holds: ⟨ (M· P)(x,n),P(x,m)⟩ = ⟨ P(x,n),(M^†· P)(x,m)⟩. In <cit.> the authors show that every differential operator D∈ℱ_R(P) has a unique adjoint ^†∈ℱ_R(P) with the property ⟨ P·, Q ⟩ = ⟨ P,Q·^†⟩, for all P,Q∈ M_N(ℂ)[x]. Moreover, φ(M^†) = φ(M)^† for all M∈ℱ_L(P). § SEMI-CLASSICAL LAGUERRE TYPE SOLUTIONS In the sequel, we consider the following two matrices A,J∈ M_N(ℝ) which satisfy J=∑_k=1^N k E_k,k A =∑_k=1^N-1 a_k E_k+1,k. Notice that, it is straightforward to show that [J,A]=A and e^xAJ e^-xA=J-Ax. Let us consider the following weight matrix supported on the interval [0,∞): W^(ν)_ϕ(x) = e^Ax T^(ν)_ϕ(x) e^A^∗ x, T^(ν)_ϕ(x) = e^-ϕ(x)∑_k=1^N δ^(ν)_k x^ν+k E_k,k, where δ^(ν)_k is a constant real number for 1 ≤ k ≤ N, and ϕ be an analytic function on a neighborhood of the interval [0,∞). In the sequel, we assume that W(x)P(x) = 0 has vanishing limits at the endpoints of support for any matrix polynomial P. Let A,J∈ M_N(ℂ) as in (<ref>). Then, the first order differential operators 𝒟 = ∂_x x + x(A-1), 𝒟^† = -∂_x x - (1+ν+J)+xϕ'(x)-x, are mutually adjoint. Let P,Q ∈ M_N(ℂ[x]). In order to simplify notation, in the rest of the proof, we denote by W(x):=W^(ν)_ϕ(x) and T(x)=W^(ν)_ϕ(x) ⟨ P ·𝒟, Q ⟩ = ∫_0^∞ (P·𝒟) W(x) Q^∗(x) dx = ∫_0^∞( xP'(x)+xP(x)(A-1) ) W(x) Q^∗(x) dx. Notice that, since W(x)P(x) has vanishing limits at the endpoints x=0, and x=∞, integration by parts implies that ∫_0^∞ xP'(x)W(x)Q^∗(x)dx = -∫_0^∞ P(x) (xW(x)Q^∗(x) )' dx. On the other hand, we have that ∫_0^∞ xP(x)(A-1)W(x) Q^∗(x) dx=∫_0^∞ P(x)W(x) ( W^-1(x) x (A-1)W(x) )Q^∗ (x) dx, by linearity we obtain that ⟨ P ·𝒟, Q ⟩ = - ∫_0^∞ P(x) (xW(x)Q^∗(x) )' dx + ∫_0^∞ P(x)W(x) ( W^-1(x) x (A-1)W(x) )Q^∗ (x) dx = -∫_0^∞ P(x)(W(x)Q^∗(x) +x W'(x) Q^∗+xW(x)(Q^∗(x))') dx + ∫_0^∞ P(x)W(x) ( W^-1(x) x (A-1)W(x) )Q^∗ (x) dx. Notice that since (Q · (1+x∂_x) )^∗(x)=Q^∗(x)+x(Q^∗(x))', we can put ∫_0^∞ P(x)(W(x)Q^∗(x) +W(x)x(Q^∗(x))') dx=∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx. On the other hand, ∫_0^∞ P(x) x W'(x) Q^∗(x) dx=∫_0^∞ P(x) W(x) xW^-1(x) W'(x) Q^∗(x) dx. Hence, we obtain that ⟨ P ·𝒟, Q ⟩ = - ∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx - ∫_0^∞ P(x) W(x) xW^-1(x) W'(x) Q^∗(x) dx + ∫_0^∞ P(x)W(x) ( W^-1(x) x (A-1)W(x) )Q^∗ (x) dx. By (<ref>) we have that W^-1(x) W'(x) =e^-A^∗ x(T^-1(x) A T(x)+T ^-1(x)T'(x)+ A^∗) e^A^∗ x, W^-1(x) x (A-1)W(x) = e^-A^∗x(x T^-1(x) A T(x) - x)e^A^∗x. Thus, we obtain ⟨ P ·𝒟, Q ⟩ = - ∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx - ∫_0^∞ P(x) W(x) e^-A^∗ x(xT^-1(x) A T(x)+x T^-1(x)T'(x)+ x A^∗) e^A^∗ x Q^∗(x) dx + ∫_0^∞ P(x)W(x) e^-A^∗x(x T^-1(x) A T(x) - x)e^A^∗x Q^∗(x)dx = - ∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx - ∫_0^∞ P(x) W(x) e^-A^∗ x(x T^-1(x)T'(x)+ x A^∗+x) e^A^∗ x Q^∗(x) dx. By taking into account that xT'(x)=T(x) (-xϕ'(x)+ν+J), we obtain that ⟨ P ·𝒟, Q ⟩ = - ∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx - ∫_0^∞ P(x) W(x) e^-A^∗ x ( -xϕ'(x)+ν+J+xA^∗+x) e^A^∗x Q^∗(x) dx. Notice that the second expresion of the right hand of the above equality is e^-A^∗ x ( -xϕ'(x)+ν+J+xA^∗+x) e^A^∗x = -xϕ'(x)+ ν+ e^-A^∗ x J e^A^∗x+xA^∗+x. On the other hand, the equation e^xAJ e^-xA=J-Ax implies that e^-A^∗ x J e^A^∗x=J-A^∗x. Hence, we obtain that ⟨ P ·𝒟, Q ⟩ = - ∫_0^∞ P(x)W(x)(Q · (1+x∂_x) )^∗(x) dx - ∫_0^∞ P(x) W(x) (x -xϕ'(x)+(ν+J-A^∗x+xA^∗) Q^∗(x) dx = ∫_0^∞ P(x)W(x)(Q · -(x∂_x+x-xϕ'(x)+(ν+J+1) )^∗(x) dx = ⟨ P,Q·𝒟^†⟩. Therefore, the operators 𝒟 and 𝒟^† are mutually adjoint, as asserted. By the above theorem, since 𝒟 = ∂_x x + x(A-1) and 𝒟^† = -∂_x x - (1+ν+J)+xϕ'(x)-x, then we obtain that 𝒟^† = -𝒟 + (Ax-J) -(1+ν) + xϕ'(x)-2x. Let A,J∈ M_N(ℂ) be matrices as in (<ref>). Then, 𝒞=Ax-J is a symmetric operator respect to the weight W:=W^(ν)_ϕ as in (<ref>). Moreover, if P(x,n)'s are monic MVOPs associated with the weight W, such that P·𝒞=M_𝒞· P, then M_𝒞=∑_i=-1^1 U_j(n)δ^j, with U_1(n):=A, U_0(n):=X(n)A-AX(n+1)-J, U_-1(n):= Y(n)A-AY(n+1) + [J,X(n)] + (AX(n+1)-X(n)A)X(n), where X(n) and Y(n) are the coefficients of the (n-1)-term and (n-2)-term of P(x,n) respectively. In the same manner as Proposition <ref>, it can be shown that 𝒞 is an symmetric operator respect with W. If we put M_𝒞=∑ U_j(n)δ^j such that P·𝒞=M_𝒞· P. By taking into account that 𝒞 increases the degree of any polynomial in 1, we obtain that U_j(n)=0 for j>1. On the other hand, since 𝒞 is an symmetric operator respect to W, we have that M_𝒞=M^†_𝒞 and so U_j(n)=0 for j<-1. The formulas for U_-1(n),U_0(n) and U_1(n) can be obtained by direct computation from P·𝒞=M_𝒞· P. Since 𝒞=Ax-J=(Ax-J)^†, then M_𝒞=M_𝒞^†. Thus, from equations (<ref>) we have that U_1(n) =ℋ(n) U_-1(n+1)^∗ℋ(n+1)^-1, U_0(n) =ℋ(n)U_0(n)^∗ℋ(n)^-1, and so we obtain that A=ℋ(n)(Y(n+1)A-AY(n+2)+[J,X(n+1)]+(AX(n+2)-X(n+1)A)X(n+1))^∗ℋ(n+1)^-1, X(n)A-AX(n+1)-J=ℋ(n) (X(n)A-AX(n+1)-J)^∗ℋ(n)^-1. Let W:=W^(ν)_ϕ be a matrix weight as in (<ref>), with monic MVOPs P(x,n) such that 𝒟= ∂_x x + (A-1)x, 𝒟^†= -𝒟 + 𝒞 + v'(x) for some polynomial v(x) of degree k and 𝒞=Ax-J. If X(n) and Y(n) are the coefficients of the (n-1)-term and (n-2)-term of P(x,n) respectively. Then, the monic polynomials P(x,n) satisfy the following relation P·𝒟=M· P, M= ∑_j=-k+1^1 A_j(n)δ^j with A_1(n)=A-1, A_0(n)=n+X(n)A - A X(n+1)-B(n), A_-1(n)=(n-1)X(n)+Y(n)(A-1)-(A-1)Y(n+1)-A_0(n)X(n), A_j(n)=(v'(L))_j(n), -k+1 <j< -1 where B(n) is given by (<ref>). Clearly, the formulas for A_j(n) with j=-1,0,1 can be derived from the equalities in (<ref>) by using the definition of 𝒟. For j<-1, we have that A_j(n) = ⟨ P·𝒟,δ^j · P ⟩ℋ(n-j)^-1=⟨ P, δ^j · P ·𝒟^†⟩ℋ(n-j)^-1 = ⟨ P, δ^j · P · v'(x)⟩ℋ(n-j)^-1=⟨ P· v'(x), δ^j · P⟩ℋ(n-j)^-1 = ⟨ v'(L)· P, δ^j· P ⟩ℋ(n-j)^-1, where we have used that ⟨ P, δ^j · P·𝒟⟩ and ⟨ P, δ^j · P·𝒞⟩ are both zero for j<-1 in the third equality, and the fact that v'(x) is a scalar funtion in the fourth one. Then, we have that A_j(n)= (v'(L))_j(n) for j<-1. To complete the proof, notice that (v'(L))_j(n)=0 for j≤ -k. As a direct consequence, we obtain the following corollary. In the same hypothesis as in Theorem <ref>. If the polynomial v has degree 1, then the discrete operator M associated with 𝒟 satisfies M= A_0(n)+(A-1)δ with A_0(n) as in Theorem <ref>. Moreover, in this case we have that (n-1)X(n)+Y(n)(A-1)-(A-1)Y(n+1)-( n+X(n)A - A X(n+1)-B(n))X(n) = 0. By Theorem <ref>, since (v) < 2 we obtain that A_j(n)=0 for all j<-1. On the other hand, notice that since v has degree 1, then v'(x) is a constant function and so the operator 𝒟^† does not increase degrees. This implies that A_1^†=0 and therefore A_-1(n)=(n-1)X(n)+Y(n)(A-1)-(A-1)Y(n+1)-A_0(n)X(n) = 0, the formula for A_1(n) and A_0(n) are consequence of Theorem <ref>. § LIE ALGEBRAS ASSOCIATED TO ORTHOGONAL POLYNOMIALS In this section, we solve a particular case of the problem proposed by Ismail in <cit.>, as described in the introduction. For this purpose, we study the structure of a Lie algebra related with the operators 𝒟 and 𝒟^†. Recall that if 𝔤 is a finite dimensional Lie algebra, and if 𝔤^j and 𝔤_j denote the following recursions 𝔤^0=𝔤_0=𝔤, 𝔤^j+1=[𝔤^j,𝔤^j] and 𝔤_j+1=[𝔤,𝔤_j], then 𝔤 is called solvable (nilpotent) if 𝔤^j=0 for some j (if 𝔤_j=0 for some j). Clearly, any nilpotent Lie algebra is solvable. The radical (nilradical) of 𝔤 is its maximal solvable ideal (maximal nilpotent ideal) of 𝔤. We will denote by Rad(𝔤) and Nil(𝔤) to the radical and nilradical of 𝔤, respectively. §.§ Lie algebra generated by 𝒟 and 𝒟^† Let A,J∈ M_N(ℂ) as in (<ref>) and let ϕ an entire function over ℂ, let us consider the operators 𝒟 = ∂_x x + x(A-1), 𝒟^† = -∂_x x - (1+ν+J)+xϕ'(x)-x. If x and x^jϕ^(j)(x) act over matrix valued polynomials by right constant multiplication, then we have that [𝒟,x] = -x, [𝒟^†,x]=x, [𝒟,𝒟^†] = -x^2ϕ^(2)(x)+(2-ϕ'(x))x, [𝒟,ϕ^(j)(x)x^j] = -(jx^jϕ^(j)(x)+x^j+1ϕ^(j+1)(x)) = -[𝒟^†,ϕ^(j)(x)x] for all j≥ 1. In the sequel, given ϕ an entire function over ℂ, we denote by 𝔤_ϕ= ⟨ 1,𝒟, 𝒟^†, x, xϕ'(x), x^2ϕ^(2)(x),…⟩ with bracket as above. We are interested in the case that this Lie algebra is finite dimensional. The following proposition states that this happens if and only if ϕ is a polynomial. We will need the following notation, given ϕ a polynomial over ℂ with ℓ non-zero coefficients k=ℓ+2 if ϕ'(0)=ϕ(0)=0, ℓ+1 if ϕ'(0)=0, ϕ(0)≠ 0, ℓ+1 if ϕ(0)=0, ϕ'(0)≠ 0, ℓ if ϕ(0)≠ 0, ϕ'(0)≠ 0, Let ϕ an analitic function over ℂ and let 𝔤:=𝔤_ϕ its associated Lie algebra as in (<ref>). Then, we have that (𝔤) is finite if and only if ϕ is a polynomial. In such case, if k is as in (<ref>) then (𝔤)= k+2. Clearly, if ϕ is a polynomial, then the dimension of 𝔤 is finite, since if n is the degree of ϕ then ϕ^(m)(x)=0 for all m>n. Conversely, assume now that ϕ is not a polynomial. Since ϕ is analytic, we can express ϕ as follow ϕ(x)= ∑_i=0^∞ a_ix^i. This implies that x^jϕ^(j)(x)= ∑_i=0^∞ b_i,j x^i, with b_i,j= ijj! a_i if j≤ i, 0 if i<j. In particular, if a_i=0 then b_i,j=0 for all j ∈ℕ. Let {i_t}_t∈ℕ be the sequence of non-zero coefficients indices of ϕ, that is i_t<i_t+1 for all t ∈ℕ, such that a_i ≠ 0 if and only if i= i_t for some t ∈ℕ. Claim: The vector space ⟨ x^i_1ϕ^(i_1)(x),…, x^i_ℓϕ^(i_ℓ)(x)⟩ has dimension ℓ. Let c_1,…,c_ℓ∈ℂ such that c_1x^i_1ϕ^(i_1)(x)+⋯+ c_ℓx^i_ℓϕ^(i_ℓ)(x)=0, this induces the following system of equations ∑_t=1^h c_ti_hi_t i_h! a_i_t=0 for h=1,…,ℓ. By taking into account that a_i_1≠ 0, the equation for h=1 implies that c_1=0. In the same way, since c_1=0, the equation for h=2 implies that c_2 a_i_2 i_2!=0 and so c_2=0 since a_i_2≠ 0. Inductively, if c_1=c_2=…=c_ℓ-1=0, then the equation for h=ℓ implies that c_ℓ a_i_ℓ i_ℓ!=0 and so c_ℓ=0 since a_i_ℓ≠ 0, hence we obtain that c_h=0 for all h∈{1,…,ℓ}. By the claim, the space 𝔤 has subspaces of all of the possible dimensions and so is non-finite dimensional, as asserted. Now, assume that ϕ is a polynomial of degree n, in the same notation as above, by (<ref>) we have that ⟨ xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x)⟩⊆⟨ x^i_1,…,x^i_ℓ⟩. The claim and the above statement imply that ⟨ xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x)⟩ has dimension ℓ. Finally, the last assertion follows from the fact that 𝒟,𝒟^† are linearly independent respect to ⟨ 1,x,xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x)⟩ and this vector space has dimension k, with k as in the statement. The above proposition solves the problem proposed by Ismail in <cit.> for the case B_n=-1 for all n natural number. In the notation of <cit.>, the differential operators 𝒟 and 𝒟^† correspond to 𝒟=x L_1,n and 𝒟^†=x L_2,n +(1+ν). Then, the algebra generated by {𝒟, 𝒟^†,1} is isomorphic to the algebra generated by {xL_1,n, xL_2,n,1}. By the proof of the above theorem, if ϕ(x)=a_0+a_1 x +…+ a_nx^n is a polynomial of degree n with ℓ non-zero coefficients. If {i_1,…,i_ℓ}⊆{0,…,n} is the set of indices such that a_i_j≠ 0. then we have that ⟨ xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x)⟩=⟨ x^i_1,…,x^i_ℓ⟩. Let ϕ_1(x)=x^3 and ϕ_2(x)=x^3+x^2, by the above theorem the associated Lie algebras 𝔤_ϕ_1 and 𝔤_ϕ_2 have the dimensions 5 and 6, respectively. Then, the algebras 𝔤_ϕ_1 and 𝔤_ϕ_2 are non-isomorphic. The element z= 𝒟+𝒟^† +2x -xϕ'(x) is a symmetric differential operator which belongs to the center of the Lie algebra 𝔤. It is follows immediately from the definition of the bracket of 𝔤_ϕ. The central element that we found in the above lemma, it is related with the symmetric operator 𝒞 that was considered in Lemma <ref>, we will see this in the following section. In the sequel, given a polynomial ϕ(x)=a_0+a_1 x+ ⋯ +a_n x^n of degree n≥ 2 with ℓ non-zero coefficients and k as in (<ref>), let us consider the following notations. I_ϕ={i ∈{2,…,n}: a_i≠ 0}={ j_1,…,j_k-2}= {i_3,…,i_ℓ} if a_0≠ 0, a_1≠ 0, {i_2,…,i_ℓ} if a_0=0 and a_1≠ 0, {i_2,…,i_ℓ} if a_0≠ 0 and a_1= 0, {i_1,…,i_ℓ} if a_0=0 and a_1=0, with j_t< j_t+1 and i_t<i_t+1. Let ϕ(x)=a_0+a_1 x+ ⋯ +a_n x^n be a polynomial of degree n≥ 2 with ℓ non-zero coefficients and k as in (<ref>). If 𝔤:=𝔤_ϕ is the associated Lie algebra of ϕ as in (<ref>), then we have that 𝔤≅ℂ^2 ⊕𝔥 where 𝔥 is a solvable Lie algebra of dimension k, with an abelian nilradical of dimension k-1. More precisely, if I_ϕ is as in (<ref>) then 𝔥≅⟨ E ⟩⋉⟨ E_1, … E_k-1⟩ where ⟨ E_1, … E_k-1⟩ is abelian and the rest of the brackets satisfy [E,E_1]=E_1 and [E,E_t]= j_t-1 E_t for t=2,…,k-1. By Lemma <ref>, the element z= 𝒟+𝒟^† +2x -xϕ'(x) belongs to the center of 𝔤, and so we obtain an element in the center which does not belong to ⟨ 1⟩, thus if 𝔥=⟨𝒟, x,xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x) ⟩ then we obtain that 𝔤≅ℂ^2 ⊕𝔥, with 𝔥 a Lie algebra of dimension k. Thus, it is enough to show that 𝔥 is solvable with nilradical of dimension k-1. Let us consider 𝔨=⟨ x,xϕ'(x),x^2ϕ^(2)(x),…, x^nϕ^(n)(x)⟩, by definition of the bracket and by taking into account that [x,x^lϕ^(l)(x)]=[x^i ϕ^(i)(x),x^j ϕ^(j)(x)]=0 for all i≠ j, we obtain that [𝔥,𝔥]⊆𝔨. Hence, by (<ref>) we obtain that [[𝔥,𝔥],[𝔥,𝔥]]=0, and so 𝔥 is solvable. Finally, notice that 𝔨 is an abelian ideal of 𝔥 of dimension (𝔨)= (𝔥)-1=k-1, this implies that 𝔨 is the nilradical of 𝔥, as desired. In the same manner as in Remark <ref> we have that 𝔥=⟨𝒟, x, x^j_1,…, x^j_k-2⟩, in this case ⟨ x, x^j_1,…, x^j_k-2⟩ is an abelian subalgebra of dimension k-1. It is enough to see the brackets [𝒟,x^j_t] and [𝒟,x], in this case we obtain that [𝒟,x]=-x and [𝒟,x^j_t]= -j_t x^j_t, so we obtain that ⟨ x, x^j_1,…, x^j_k-2⟩ is an abelian ideal. Finally, if we consider the following correspondence 𝒟⟼ -E, x⟼ E_1, x^j_i⟼ E_i+1 for i=1,…,k-2, is an Lie Algebra isomorphism between 𝔥 and ⟨ E ⟩⋉⟨ E_1, … E_k-1⟩ with brackets given as in (<ref>), as asserted. In the sequel we are going to study the structure of the solvable Lie algebra 𝔥_ϕ. In general, a Lie algebra 𝔥' with an abelian ideal of codimension 1 is called almost abelian. This kind of algebra was studied by V.V. Gorbatsevich in <cit.>. The author asserted that in general this kind of algebra it decomposes as ℂ⋉_ψℂ^k-1, this semidirect product gives a linear transformation ψ: ℂ→ gl_k-1(ℂ), moreover he asserted that the structure of this kind of algebras it determines by the matrix Ψ=ψ(1). More precisely, ℂ⋉_ψℂ^k-1≅ℂ⋉_ψ'ℂ^k-1⟺ Ψ and Ψ' are conformally similar, recall that Ψ and Ψ' are conformally similar if and only if there exist a matrix P∈ Gl_k-1(ℂ) and a non-zero complex number λ∈ℂ∖{0} such that Ψ= λ P Ψ' P^-1. Let k be an integer greater than 1 and let 1<j_1<…<j_k-2 and 1<j'_1<…< j'_k-2 be two sequences of positive integers. Let us consider Ψ = diag(1, j_1,…,j_k-2) and Ψ' = diag(1, j'_1,…,j'_k-2). Then, Ψ and Ψ' are conformally similar if and only if Ψ=Ψ'. Clearly, if Ψ=Ψ' then they are conformally similar trivially. Now, assume that Ψ and Ψ' are conformally similar, so there exist λ∈ℂ^* and P∈ GL_k-1(ℂ) such that Ψ' = λ P Ψ P^-1, by similarity we obtain the following spectral relationship Spec(Ψ')= λ·Spec(Ψ), i.e. all of the eigenvalues of Ψ' can be obtained from the eigenvalues of Ψ by multiplication by λ. Since Ψ and Ψ' are both diagonal, we have that Spec(Ψ)={1,j_1,…,j_k-2} and Spec(Ψ')={1, j'_1,…,j'_k-2}, since 1 belong to both spectra and the rest of the eigenvalues of Ψ and Ψ' are greater than 1, we obtain that λ=1 necessarily. Hence, we obtain that Spec(Ψ')= Spec(Ψ). Therefore, Ψ=Ψ' as asserted. We are in position to give the following theorem, which says when the Lie algebra associated to two different polynomials (as in (<ref>)) are isomorphic. Let ϕ_1(x),ϕ_2(x) be polynomials over ℂ of degree greater or equal than 2. Let 𝔥_ϕ_1 and 𝔥_ϕ_2 be its associated solvable Lie algebras given as in Theorem <ref>. Then, 𝔥_ϕ_1≅𝔥_ϕ_2⟺ I_ϕ_1 = I_ϕ_2, where I_ϕ_1, I_ϕ_2 are as in (<ref>). Moreover, we have that 𝔤_ϕ_1≅𝔤_ϕ_2⟺ I_ϕ_1 = I_ϕ_2, where 𝔤_ϕ_1 and 𝔤_ϕ_2 are the associated Lie algebras of ϕ_1 and ϕ_2, respectively. From Theorem <ref>, we have that 𝔤_ϕ_1≅𝔤_ϕ_2 if and only if 𝔥_ϕ_1≅𝔥_ϕ_2. So, it is enough to see that 𝔥_ϕ_1≅𝔥_ϕ_2⟺ I_ϕ_1 = I_ϕ_2. Now by (<ref>), it enough to see that the associated matrices Φ_1 and Φ_2 are conformally similar. By Theorem <ref> and (<ref>) we have that Φ_1 and Φ_2 are conformally similar to Ψ = diag(1, j_1,…,j_k-2) and Ψ' = diag(1, j'_1,…,j'_k-2) respectively, where {j_1,…,j_k-2} =I_ϕ_1 and {j'_1,…,j'_k-2}=I_ϕ_2. Finally, by Lemma <ref>, we obtain that Ψ and Ψ' are conformally similar if and only if Ψ=Ψ' which is equivalent to say that I_ϕ_1=I_ϕ_2. Therefore 𝔥_ϕ_1≅𝔥_ϕ_2 if and only if I_ϕ_1= I_ϕ_2 as asserted. As a direct consequence, we obtain the following. Let ϕ_1(x),ϕ_2(x) be polynomials over ℂ with degree greater than 2. Then, we have the following cases: * If ϕ_1=ϕ_2=2, then 𝔤_ϕ_1≅𝔤_ϕ_2. * If ϕ_1≠ϕ_2, then 𝔤_ϕ_1≇𝔤_ϕ_2. Notice that if we consider ϕ_m(x)=x^m+ax with m≥ 2, then (𝔥_ϕ_m)=3. The structure of solvable Lie algebras of dimension 3 was studied by J. Patera and H. Zassenhaus in <cit.>. In page 4, the authors define the Lie algebra L_3,6=⟨ a_1,a_2,a_3⟩ with brackets [a_1,a_2]= a_3 [a_1,a_3]= a_3-α· a_2 with parameter α satisfying α≠ 0 and 1-4α≠ 0, this parameter α is in one-to-one correspondence with isomorphism classes of this kind of algebras. Its associated matrix ψ_α(1) is ψ_α(1)= [ 0 -α; 1 1 ] The eigenvalues of ψ_α(1) are λ_0=1-√(1-4α)2 and λ_1=1+√(1-4α)2. On the other hand, if we consider ϕ_m(x)=x^m+ax with m≥ 2, then 𝔥_ϕ_m has dimension 3 and its associated matrix Ψ_m is conformally similar to [ 1 0; 0 m ] thus, Ψ_m is conformally similar to ψ_α(1) if and only if λ_0=r, λ_1=rm for some complex number r, since diagonalizable matrices are similar if and only if its spectrum are equal. This system of equation has a solution r=1m+1 and α= m(m+1)^2. Therefore 𝔥_ϕ_m≅ L_3,6^α with α=m(m+1)^2. § LAGUERRE TYPE SOLUTIONS Let A∈ M_N(ℂ) be a constant matrix and let ν∈ℝ such that ν>0. In this section and the sequel, we are going to consider the weight matrix W given by W^(ν)(x) = e^Ax T^(ν)(x) e^A^∗ x, T^(ν)(x) = e^-x∑_k=1^N δ^(ν)_k x^ν+k E_k,k. If we denote L(x)=e^Ax, then W^(ν)(x)=L(x) T^(ν)(x) L^*(x). Recall that if A,J are as in (<ref>), the equation (<ref>) in terms of L says that L(x)J L^-1(x)=J- Ax. In this case, we obtain the same kind of weight that was consider in section 3, with ϕ(x)=x. The first order differential operators 𝒟 = ∂_x x + x(A-1), 𝒟^† = -∂_x x - (1+ν +J), are mutually adjoint and satisfy M = ψ^-1(𝒟) = (A-1)δ-(n+1+ν)-ℋ(n)Jℋ^-1(n), M^† = ψ^-1(𝒟^†) = -(n+ν+J+1) + ℋ(n)(A-1)^∗ℋ^-1(n-1) δ^-1. The operators 𝒟 and 𝒟^† are mutually adjoint by taking ϕ(x)=x in Proposition <ref>. On the other hand, since v(x) has degree 1 in this case, by Corollary <ref> we obtain that A_j(n)=0 for j≤ -1 and A_1(n)=A-1. Finally, the formula for A^†_0(n) can be obtained directly from the relation P·𝒟^†=M_𝒟^†· P and the term A_0(n) can be obtained from equation (<ref>). As a direct consequence of the above proposition and Corollary <ref> we obtain the following result. Let A,J∈ M_N(ℂ) as in (<ref>) and let W:=W^(ν)_ϕ be a matrix weight as in (<ref>) with ϕ(x)=x and MVOPs P(x,n). If X(n) and Y(n) are the coefficients of the (n-1)-term and (n-2)-term of P(x,n) respectively, then n+X(n)A - A X(n+1)-B(n) = -(n+1+ν)-ℋ(n)Jℋ^-1(n), X(n)+[J,X(n)]=ℋ(n)(A^∗-1) ℋ^-1(n-1), where ℋ(n) and B(n)are as in (<ref>) and (<ref>). It follows from Proposition <ref> and Corollary <ref>, by taking into account the term A_0(n) of 𝒟. On the other hand, from the equality P·𝒟^†= M^†· P with 𝒟^†= -∂_x x - (1+ν +J) and M^†=-(n+ν+J+1)+A^†_-1(n) δ^-1, we can obtained the equation (<ref>) from the above theorem. §.§ Existence of the operator D Families of matrix valued orthogonal polynomials which are eigenfunctions of a second order differential operator are of great importance, see e.g. <cit.>, <cit.>, <cit.>, <cit.>. Using the approach of <cit.>, we get a symmetric second-order differential operator which preserves polynomials and its degree. For this we establish a conjugation with a diagonal matrix differential operator. Let us consider matrix valued polynomials F_2, F_1, F_0 of degrees two, one and zero respectively and let us assume that we have a matrix valued second-order differential operator D such that Q· D=(d^2Qdx^2)(x) F_2(x) + (dQdx)(x) F_1(x) +Q(x) F_0(x). for a matrix valued polynomial Q. It follows from the definition of the matrix valued inner product (<ref>) that a differential operator D is symmetric with respect to W if for all matrix valued polynomials G,H we have ∫_0^∞ (GD)(x)W(x)(H(x))^∗ dx = ∫_0^∞ G(x)W(x)((HD)(x))^∗ dx. By <cit.>, this symmetry condition is equivalent to the following equations F_2(x)W(x) = W(x) ( F_2(x))^∗, 2 d(F_2W)dx(x) - F_1(x)W(x) = W(x) ( F_1(x))^∗, d^2(F_2W)dx^2(x) - d(F_1W)dx(x) + F_0(x) W(x) = W(x) ( F_0(x))^∗, and the boundary conditions lim_x→ 0 F_2(x)W(x) = 0 = lim_x→∞ F_2(x)W(x), lim_x→ 0 F_1(x)W(x) - d(F_2W)dx(x) = 0 = lim_x→∞ F_1(x)W(x) - d(F_2W)dx(x). We have the following lemma. The second order differential operator D_Q=∂_x^2 x +∂_x ( 1+ν-x+J) -J is symmetric respect to the weight T^(ν)(x). By taking into account that F_2(x)=x, F_1(x)= 1+ν -x+J and F_0(x)=-J, it is a straightforward computation see that (<ref>), (<ref>), (<ref>) and (<ref>) are satisfied for D_Q and T^(ν)(x). The second order differential operator D=∂_x^2 x +∂_x ( (A-1)x +1+ν+J) +Aν + JA -J is symmetric respect to the weight W(x). Moreover P· D = Γ· P where Γ(n)=A(n+ν+1+J)-n-J. It follows from Remark 4.1 in <cit.>. §.§ The Lie Algebra associated to W Recall that a Lie algebra 𝔤 is called reductive if its radical is equal to its center. Let A,J∈ M_N(ℂ) as in (<ref>) and let us consider the operators 𝒟 = ∂_x x + x(A-1), 𝒟^† = -∂_x x - (1+ν+J), D=∂_x^2 x +∂_x ( (A-1)x +1+ν+J) +Aν + JA -J. If x acts over matrix valued polynomials by right constant multiplication, then we have that [𝒟,x] = -x, [𝒟^†,x]=x, [𝒟,𝒟^†] = x, [D,x]=-𝒟+𝒟^†, [𝒟,D]= -𝒟+ D - (1+ν), [𝒟^†,D]= 𝒟^†- D+ (1+ν). The subjacent Lie algebra generated by {𝒟, 𝒟^†, D,x, I} is isomorphic to the Lie algebra 𝒜=⟨ x_1,x_2,x_3,x_4,x_5⟩ with brackets [x_1,x_2]=-x_4, [x_1,x_4]=-x_4 , [x_2,x_4]=x_4, [x_3,x_4]=-x_1+x_2, [x_1,x_3]= -x_2+ x_3 +x_4 - (1+ν)x_5 and [x_2,x_3]= x_1 - x_3 -x_4 + (1+ν)x_5, with correspondence: x_1⟼𝒟+x, x_2⟼𝒟^†+x, x_3⟼ D, x_4⟼ x, x_5 ⟼ 1. Let 𝒟'=𝒟+x and 𝒟'^†=𝒟^†+x. It can be shown by direct computation that [𝒟',x]=-x, [𝒟'^†,x]= x, [𝒟',𝒟'^†]=-x, [D,x]=-𝒟'+𝒟'^†, [𝒟',D]= -𝒟'^†+ D +x - (1+ν), [𝒟'^†,D]= 𝒟'- D -x + (1+ν). Clearly, the subjacent Lie algebra associated to this representation, is the Lie algebra 𝒜=⟨ x_1,x_2,x_3,x_4,x_5⟩ with brackets [x_1,x_2]=-x_4, [x_1,x_4]=-x_4 , [x_2,x_4]=x_4, [x_3,x_4]=-x_1+x_2, [x_1,x_3]= -x_2+ x_3 +x_4 - (1+ν)x_5 and [x_2,x_3]= x_1 - x_3 -x_4 + (1+ν)x_5, with correspondence x_1⟼𝒟+x, x_2⟼𝒟^†+x, x_3⟼ D, x_4⟼ x, x_5 ⟼ 1, as desired. We have the following structure result. The Lie algebra 𝒜 defined as above is a 5-dimensional reductive algebra with center of dimension two, given by 𝒵_𝒜=⟨ x_1+x_2- x_4,x_5⟩. Moreover, 𝒜=[𝒜,𝒜]⊕𝒵_𝒜 with [𝒜,𝒜] isomorphic to SL(2,ℂ). In particular, 𝒞_1= -4x_4(x_1-x_3-x_4+(1+ν)x_5)+(x_4-x_1+x_2)^2, 𝒞_2= x_1+x_2- x_4 and 𝒞_3=x_5 are Casimir elements of 𝒜. From Lie algebra theory (see <cit.>), the radical of 𝒜 can be computed from its Killing form, in this case we have that Rad(𝒜)=𝒵_𝒜=⟨ x_1+x_2- x_4,x_5⟩, and so, the Lie algebra 𝒜 is reductive. Now, from general theory, since 𝒜 is reductive, we obtain 𝒜=[𝒜,𝒜] ⊕𝒵_𝒜. In this case, we obtain that [𝒜,𝒜]= ⟨ x_4, x_1 -x_2, x_1-x_3 -x_4 +(1+ν) x_5⟩. By taking a_1=x_4, a_2= x_1-x_2 and a_3=x_1-x_3-x_4+(1+ν)x_5 we obtain that [a_1,a_2]=-2a_1, [a_1,a_3]=a_1-a_2, [a_2,a_3]=-a_3 and so if we take â_2=a_1-a_2 and â_3=-a_3 we have [a_1,â_2]=2a_1, [a_1,â_3]=-â_2, [â_2,â_3]=2â_3, and so [𝒜,𝒜] is isomorphic to SL(2,ℂ) by consider the map a_1↦ e_1 â_2 ↦ e_2 and â_3 ↦ e_3. In particular, the Casimir element of SL(2,ℂ) given by 4e_1e_3+e_2^2 induces a Casimir element of [𝒜,𝒜] 𝒞_[𝒜,𝒜]= -4x_4(x_1-x_3-x_4+(1+ν)x_5)+(x_4-x_1+x_2)^2. By taking into account that 𝒞_[𝒜,𝒜] commutes with the central elements of 𝒜, we obtain that 𝒞_[𝒜,𝒜] commutes with all of the elements of 𝒜 and so is a Casimir element of 𝒜. The last assertion is clear, since the central elements always are Casimir elements of a given Lie algebra. Notice that 𝒞=𝒞_2+ (1+ν )𝒞_3 = 𝒟+𝒟^†+x+(1+ν) it is also a Casimir invariant of 𝒜. Hence, under the representation given by 𝒟,𝒟^†, D,x,1, the image of this Casimir satisfies 𝒞= Ax-J. Thus, in terms of M,M^† and L, we obtain that φ^-1(𝒞)=M+M^†+L+(1+ν). On the other hand, it can be check that 𝒞_[𝒜,𝒜]=18(A^2x^2+ν^2+J^2+Ax-2ν Ax-2xJA+2ν J -1). Thus, 𝒞'=A^2x^2-2xJA+J+J^2 it is also a Casimir since 𝒞'=8𝒞_[𝒜,𝒜]-(1-2ν)𝒞 - (ν^2-1)𝒞_3, moreover, notice that the relation [J,A]=A implies that 𝒞'=𝒞^2-𝒞, and so we obtain that 𝒞_[𝒜,𝒜]=18( 𝒞^2-2ν𝒞 + (ν^2-1)𝒞_3). Therefore, in this representation the Casimir element corresponding to 𝒞_[𝒜,𝒜] does not give more information than 𝒞. Hence, in the rest of the paper we will only consider the Casimir element 𝒞. We can also consider the Lie subalgebra generated by {x_1,x_2,x_4,x_5}. Notice that a representation of this kind of algebra was consider in the above section by taking ϕ(x)=x. In this case, we have the following structure result. The Lie subalgebra 𝒜' = ⟨ x_1,x_2,x_4,x_5 ⟩ of 𝒜 is isomorphic to 𝔤_2⊕ℂ^2 where 𝔤_2 is the 2-dimensional solvable Lie algebra with bracket [e_1,e_2]=e_2. In particular, 𝒜' has no non-central Casimir invariants. By taking the map x_2↦ e_1, x_4 ↦ e_2, x_1+x_2-x_4↦ e_3 and x_5 ↦ e_4. we obtain an isomorphic algebra of 𝒜', in this case the only non-vanishing bracket of ⟨ e_1,e_2,e_3,e_4⟩ is the bracket [e_1,e_2]=e_2 and so 𝒜' is isomorphic to 𝔤_2⊕ℂ^2, as asserted. The last assertion is a consequence of 𝔤_2 has not central elements. Hence, the only Casimir invariants of 𝒜' are the central elements. In the sequel, in order to simplify the notation, we will consider B_n:=B(n), C_n:=C(n), ℋ_n:=ℋ(n), Γ_n= Γ(n). The following proposition is a consequence, of the relations between the brackets of M,M^†,L,Γ. Let A,J∈ M_N(ℂ) be matrices as in (<ref>) and let B_n,C_n,ℋ_n and Γ_n be as in (<ref>). Then, B_n (A-1) - (A-1) B_n+1 = 2 + ℋ_n+1 J ℋ_n+1^-1 - ℋ_n J ℋ_n^-1, B_n=[B_n,J] + ℋ_n (A^t-1) ℋ_n-1^-1 - ℋ_n+1 (A^t-1) ℋ_n^-1, 2 ℋ_n ℋ_n-1^-1= [ℋ_n ℋ_n-1^-1,J] - B_n ℋ_n (A^t-1) ℋ_n-1^-1 - ℋ_n (A^t-1) ℋ_n-1^-1 B_n-1, B_n = -[J,ℋ_n J ℋ_n^-1] - ℋ_n (A^t-1) ℋ_n-1^-1 (A-1) + (A-1) ℋ_n+1 (A^t-1) ℋ_n^-1, [Γ, C_nδ^-1] = ℋ_n(A-1)^*ℋ_n-1^-1, [Γ_n,ℋ_nJ ℋ_n^-1]= n+Γ_n+ℋ_nJ ℋ_n^-1, [Γ,ℋ_n(A-1)^*ℋ_n-1^-1δ^-1]=-ℋ_n(A-1)^*ℋ_n-1^-1. The equation (<ref>) is consequence of seeing the coefficient of δ^1 in the bracket relation [M,L]=L. In the same way, the equations (<ref>), (<ref>) are consequence of seeing the coefficients of δ^0 and δ^-1 in the bracket relation [M^†, L] = L. On the other hand, the equation (<ref>) it follows from the bracket relation [M,M^†] = L. The equation (<ref>) is a consequence of the coefficient of δ^-1 in the bracket [Γ,L] = -M+M^†. The equation (<ref>) is obtained from the coefficient of δ^0 in the bracket [Γ,M] = M-Γ+(1+ν). Finally, the equation (<ref>) can be obtained from the coefficient of δ^-1 in the bracket relation [Γ,M^†] = -M^†+Γ-(1+ν). Let A,J∈ M_N(ℂ) be matrices as in (<ref>). Then, [B_n,J] = B_n - Γ_nC_n + C_n Γ_n-1 + Γ_n+1 C_n+1 - C_n+1Γ_n, [C_n,J] = 2C_n+B_n (Γ_nC_n - C_n Γ_n-1) + (Γ_nC_n - C_n Γ_n-1) B_n-1, where B_n,C_n,ℋ_n and Γ_n be as in (<ref>) By (<ref>) we have that [B_n,J] = B_n - ℋ_n (A^t-1) ℋ_n-1^-1 + ℋ_n+1 (A^t-1) ℋ_n^-1. On the other hand, by (<ref>) and taking into account that ℋ_n∈ M_N(ℝ), we obtain that ℋ_n(A^t-1)ℋ_n-1^-1 = [Γ, C_nδ^-1] = Γ_nC_n - C_n Γ_n-1, ℋ_n+1(A^t-1)ℋ_n^-1 = [Γ, C_nδ^-1] = Γ_n+1 C_n+1 - C_n+1Γ_n, and so we have that [B_n,J] = B_n - Γ_nC_n + C_n Γ_n-1 + Γ_n+1 C_n+1 - C_n+1Γ_n, as asserted. For the second equality, notice that since C_n=ℋ_nℋ_n-1^-1, thus the equation (<ref>) implies that [C_n,J]=2C_n+B_n ℋ_n (A^t-1) ℋ_n-1^-1 + ℋ_n (A^t-1) ℋ_n-1^-1 B_n-1. In the same way as above, we obtain that [C_n,J]=2C_n+B_n (Γ_nC_n - C_n Γ_n-1) + (Γ_nC_n - C_n Γ_n-1) B_n-1, as desired. § MATRIX ENTRIES OF P(X,N) AS CLASSICAL LAGUERRE POLYNOMIALS In this section we will give explicit expressions of the entries of the Laguerre-type MVOPs in terms of the classical scalar Laguerre polynomials. For this we use the approach of <cit.>, <cit.>, <cit.>, which consists in observing that a symmetric second order differential operator can be diagonalized via conjugation with an appropriate matrix valued function. The present situation is more involved than the previous cases because, although the differential operator can be diagonalized, the eigenvalue remains non-diagonal. In the rest of the section, the weight matrix W(x) is as in the previous section. §.§ Step I: Diagonalizing the differential operator: Let A,J∈ M_N(ℂ) be matrices as in (<ref>). We can define the following auxiliary matrix valued polynomials Q(x,n):=P(x,n)L(x), where L(x)= e^xA. The polynomials Q_n satisfy the following relations, Q_n·𝒟_Q = M · Q_n, Q_n· D_Q = Λ_n · Q_n, Q_n·𝒞_Q = M_𝒞· Q_n, where 𝒟_Q= L(x)^-1𝒟 L(x), D_Q=L(x)^-1 D L(x), 𝒞_Q=L(x)^-1𝒞 L(x). Here 𝒟, D are as in Proposition <ref>, Proposition <ref> respectively and 𝒞=Ax-J . In the following lemma, we show that 𝒟_Q, D_Q and 𝒞_Q are simple diagonal operators. The operators 𝒟_Q, D_Q and 𝒞_Q as in (<ref>) are given explicitly as follows: 𝒟_Q= ∂_x x -x, D_Q=∂_x^2 x +∂_x ( 1+ν-x+J) -J, 𝒞_Q=-J. All of the equalities are follow directly from definition and the equation (<ref>). In the following proposition, we will use the expressions of 𝒞_Q, 𝒞 and 𝒟 in order to obtain an equation which relates Q(x,n) and the recurrent matrix H_n(A^∗-1)H_n-1^-1. The auxiliary functions Q(x,n)'s defined in (<ref>), satisfy the following equation which depend on the squared norms: -Q(x,n)J=xQ'(x,n)- (n+J)Q(x,n)+ℋ(n)(A-1)^∗ℋ^-1(n-1)Q(x,n-1) for n≥ 1 and for n=0 -Q(x,0)J=-J-xQ'(x,0). By Lemmas <ref> and <ref> we have that -Q(x,n)J = Q(x,n)·𝒞_Q = (M_𝒞· P(x,n))e^xA where M_𝒞 is the discrete operator M_𝒞=A δ + X(n)A-AX(n+1)-J + (Y(n)A-AY(n+1) + [J,X(n)] + (AX(n+1)-X(n)A)X(n))δ^-1. By equation (<ref>), we have (n-1)X(n)+Y(n)(A-1)-(A-1)Y(n+1)-( n+X(n)A - A X(n+1)-B(n))X(n) = 0, and so Y(n)A-AY(n+1) + (AX(n+1)-X(n)A)X(n)= X(n)+Y(n)-Y(n+1)-B(n)X(n). By taking into account the relation P· x =L· P with L=δ+B(n)+C(n)δ^-1, thus we have that Y(n)=Y(n+1)+B(n)X(n)+C(n). Hence, we obtain that (M_𝒞· P(x,n))e^xA= AQ(x,n+1) + (X(n)A-AX(n+1)-J)Q(x,n)+ (C(n) + [J,X(n)] + X(n))Q(x,n-1). On the other hand, by Corollary <ref> we have that (M_𝒟· P(x,n))e^xA = (A-1) Q(x,n+1)+ (n+X(n)A - A X(n+1)-B(n))Q(x,n), (L · P(x,n))e^xA = Q(x,n+1) + B(n)Q(x,n)+ C(n) Q(x,n-1), where B(n) and C(n) are as in equation (<ref>). Hence, we have that ((M_𝒟+L)· P(x,n))e^xA= A Q(x,n+1)+(n+X(n)A - A X(n+1))Q(x,n)+C(n)Q(x,n-1) and thus (M_𝒞· P(x,n))e^xA= ((M_𝒟+L)· P(x,n))e^xA-(n+J)Q(x,n)+ (X(n)+[J,X(n)])Q(x,n-1). Notice that if we denote L(x)=e^xA, the Lemma <ref> implies that ((M_𝒟+L)· P(x,n))L(x)=(P(x,n)·(𝒟+x))L(x)= Q(x,n)· L^-1(x)(𝒟+x))L(x)= Q(x,n)·∂_xx, and by (<ref>), X(n)+[J,X(n)]=ℋ(n)(A-1)^∗ℋ^-1(n-1). Thus, we obtain that -Q(x,n)J= xQ'(x,n)- (n+J)Q(x,n)+ℋ(n)(A-1)^∗ℋ^-1(n-1), as asserted. Finally, the equation for n=0, it follows from equation (<ref>). §.§ Step II: Diagonalizing the eigenvalue Λ_n: Although the differential operator D_Q is a diagonal operator, the system of equations given by Q_n· D_Q = Λ_n Q_n is not a decoupled system since Λ_n is a lower triangular matrix. However, Λ_n can be diagonalized in a somewhat simple way: Γ_n=A(n+ν+J+1) - (n+J) = K_n Λ_n K_n^-1, where Λ_n is the diagonal matrix Λ_n = -(n+J). We can choose K_n such that is a lower triangular matrix with diagonal elements (K_n)_i,i=1 for all i=1,…,N. Moreover, in this case K_n is the following matrix: (K_n)_i,j= (-1)^i-j(n+ν + j+1)_i-1(i-j)!∏_k=j^i-1a_k i> j, 1 i=j, 0 i<j. Since A(n+ν+J+1) - (n+J) is lower triangular, the its characteristic polynomial satisfy (λ I -(A(n+ν+J+1) - (n+J)))=(λ I +nI+J)=∏_r=1^N(λ+n+r). Then, the eingevalues of A(n+ν+J+1) - (n+J) are λ_1=-(n+1), …, λ_N=-(n+N) with multiplicity 1. Since K_n diagonalizes A(n+ν+J+1) - (n+J), then its r-th column can be obtained from the eigenspace correspond to λ_r=-(n+r), that is Nu(λ_r I-A(n+ν+J+1) + n+J)=Nu(-A(n+ν+J+1) +J-rI), for r=1, …,N. We can obtain (<ref>) by a straightforward computation of these eigenspaces. With the matrix K_n as in (<ref>), we will consider the following matrix polynomial R(x,n):=K_n^-1Q(x,n). The following result that shows a relationship between the non-zero matrix entries of R(x,n) and generalized Laguerre polynomials. Let n∈ℕ and let ν>0. The matrix elements of R(x,n) are multiples of scalar Laguerre functions R(x,n)_i,j= L^(ν + j)_n+i-j(x) ξ(n,i,j) if n+i-j ≥ 0 0 if n+i-j < 0. Notice that the polynomials R(x,n)'s are eigenfunctions of D_Q=∂_x^2 x + ∂_x ( 1+ν-x+J) -J with associated eigenvalues Λ_n=-(n+J) where J is the diagonal matrix diag(1,2,…,N). If we look the (i,j)-entry of xR”(x,n) + R'(x,n) (1+ν-x+J) + (nR(x,n) + JR(x,n) -R(x,n)J)=0, we obtain that the following expression xR”(x,n)_i,j + ∑_k=1^N R'(x,n)_i,k (1+ν-x+J)_k,j + nR(x,n)_i,j + ∑_k=1^N (J_i,kR(x,n)_k,j - R(x,n)_i,kJ_k,j) = 0. Since J=diag(1,2,…,N), the above equality is equivalent to xR”(x,n)_i,j + R'(x,n)_i,j (1+ν-x+j) + (n+i-j)R(x,n)_i,j = 0. Hence, we obtain that R(x,n)_i,j= L^(ν+j)_n+i-j(x) ξ(n,i,j), as asserted. On the other hand, it is well-known that if M<0 the only solution for the differential equation x P”(x) + P'(x) (1+ν-x) + M P(x) = 0, is P(x)=0 and since xR”(x,n)_i,j + R'(x,n)_i,j (1+ν-x+j) + (n+i-j)R(x,n)_i,j = 0, we obtain that R(x,n)_i,j=0 if n+i-j<0, as asserted. Notice that by the above theorem the are only defined for n+i-j ≥ 0, so we extend its definition as follows ξ(n,i,j)=0 if n+i-j < 0. It is well-known that the generalized Laguerre polynomial satisifes L^(α)_N(0)=(α+1)_N/N!, where (a)_N is the Pochhammer symbol defined by (a)_N= 1 if N=0, a(a+1)⋯(a+N-1) if N>0. As a direct consequence of the above theorem we obtain the following corollary. Let n∈ℕ and let ν>0. Then, the coefficients of R(0,n) satisfy R(0,n)_i,j= (ν+j+1)_n+i-j/(n+i-j)!ξ(n,i,j) if n+i-j ≥ 0, 0 if n+i-j < 0. In particular, the (i,n+i)-th and (i,n+i-1)-th coordinates of R(0,n) satisfy R(0,n)_i,n+i=ξ(n,i,n+i) and R(0,n)_i,n+i-1=(n+ν+i)ξ(n,i,n+i-1). Now we need to identify the coefficients ξ(n,i,j). For this, we exploit the relation in Proposition <ref>. In the following lemma, we observe that the factor ℋ(n)(A-1)^∗ℋ^-1(n-1) in this relation is turned into a diagonal matrix via multiplication by appropriate matrices K_n. This will allow us to obtain a simple recursion for ξ(n,i,j). For this purpose, we define the following G(n):= K_n^-1ℋ_n (A^∗-1)ℋ_n-1^-1K_n-1 I(n):= K_n^-1ℋ_n J ℋ_n^-1K_n, where ℋ_n and K_n are as in (<ref>) and (<ref>), respectively. For n≥ 1 and let G(n), I(n) be the matrices as in (<ref>). Then, G(n) is diagonal and I(n) satisfies (I(n))_i,i = i, (I(n))_i,j=0 for i≠ j and i ≠ j-1. Moreover (G(n))_i,i= (ℋ_n (A^∗-1)ℋ_n-1^-1)_i,i and (I(n))_i,i+1=(ℋ_n J ℋ_n^-1)_i,i+1. By equation (<ref>), [Γ_n,ℋ_n(A^∗-1)ℋ_n-1^-1δ^-1]=-ℋ_n(A^∗-1)ℋ_n-1^-1. Then, K_n^-1Γ_n(ℋ_n(A^∗-1)ℋ_n-1^-1)K_n-1-K_n^-1(ℋ_n(A^∗-1)ℋ_n-1^-1)Γ_n-1K_n-1=-G(n). Then, by K_n^-1Γ_nK_n=Λ_n=-(n+J) we obtain that Λ_n G(n)-G(n) Λ_n-1=-G(n). The assertion it follows by observing the (i,j)-entry for i≠ j. On the other hand, from definition we have that K_n G(n) K_n-1^-1= ℋ_n (A^∗-1)ℋ_n-1^-1. Since K_n and K_n-1^-1 are both lower triangular matrices with 1's in its diagonal, then the left term in the above equation is a lower triangular matrix with (i,i)-th coordinate equal to (G(n))_i,i and so (ℋ_n (A^∗-1)ℋ_n-1^-1)_i,i=(K_n G(n) K_n-1^-1)_i,i=(G(n))_i,i, as asserted. Now, for the second assertion, we can proceed as in the same way, in this case by equation (<ref>), we have that [Γ_n,ℋ_nJ ℋ_n^-1]=n+Γ_n + ℋ_nJℋ_n^-1. Then, we obtain that K_n^-1Γ_n ℋ_nJ ℋ_n^-1K_n-K_n^-1ℋ_nJ ℋ_n^-1Γ_n K_n = K_n^-1(n+Γ_n+ℋ_nJ ℋ_n^-1)K_n. Hence, by K_n^-1Γ_nK_n=Λ_n=-(n+J) we obtain that Λ_n I(n)-I(n) Λ_n = n+ Λ_n + I(n)). The proof follows by observing the (i,j)-entry in the above matrix equality. Finally, from definition we have K_nI(n)K_n^-1=ℋ_n J ℋ_n^-1. So, in general we have that (K_nI(n)K_n^-1)_i,j= ∑_k=1^N (K_n)_i,k (I(n)K_n^-1)_k,j and (I(n)K_n^-1)_k,j=(I(n))_k,k(K_n^-1)_k,j+(I(n))_k,k+1(K_n^-1)_k+1,j. By taking into account that (K_n)_i,j=(K_n^-1)_i,j=0 for j>i, if j=i+1 we obtain that (K_nI(n)K_n^-1)_i,i+1=∑_k=1^N (K_n)_i,k (I(n)K_n^-1)_k,j=∑_k=1^i (K_n)_i,k (I(n)K_n^-1)_k,i+1= (I(n))_i,i+1(K_n^-1)_i+1,i+1, by taking into account that (K_n^-1)_i+1,i+1=1 for any i. Therefore, we have that (K_nI(n)K_n^-1)_i,i+1=(I(n))_i,i+1, as desired. The following proposition is a consequence of the relation given by the Casimir operator and Proposition 6.4 Let ξ(n,i,j) as in Theorem <ref>. If n+i-j> 0, then the constants ξ(n,i,j)'s satisfy the following: * ξ(0,i,j)=(K_0^-1)_i,j(i-j)!(ν+j+1)_i-j, * If i=1 and n>0 then ξ(n,1,j) =(G(n))_1,1 n+ν+1ξ(n-1,1,j), * If i>1 and n>0, then ξ(n,i,j) = a_i-1ξ(n,i-1,j) +(G(n))_i,i n+ν+iξ(n-1,i,j), with G(n) as in (<ref>). For the first assertion, recall that R(0,0)=K_0^-1Q(0,0)=K_0^-1P(0,0)=K_0^-1, thus we have that (K_0^-1)_i,j=ξ(0,i,j)L^(ν+j)_i-j(0). By taking into account that L^(ν)_N(0)=(ν+1)_NN! for any N, hence we obtain (K_0^-1)_i,j(i-j)!(ν+j+1)_i-j=ξ(0,i,j), as asserted. For items (b) and (c), recall that Proposition <ref> implies that -Q(0,n)J = -(n+J)Q(0,n) + ℋ(n) (A^∗-1) ℋ(n-1)^-1 Q(0,n-1). Then, -K_n R(0,n)J = -(n+J)K_n R(0,n) + ℋ(n) (A^∗-1) ℋ(n-1)^-1 K_n-1R(0,n-1), and so we obtain that -R(0,n)J = -K_n^-1(n+J)K_n R(0,n) +G(n) R(0,n-1) where G(n)=K_n^-1ℋ(n) (A^∗-1) ℋ(n-1)^-1 K_n-1. By recalling that K_n^-1(n+J)K_n R(0,n)=(n+J)-A(n+ν+J+1), we obtain that -R(0,n)J = -(n+J) R(0,n)+ A(n+ν+J+1) R(0,n)+ G(n)R(0,n-1). Thus, in terms of coordinates we have that -j R(0,n)_i,j = -(n+i)R(0,n)_i,j + ∑_k=1^N(A(n+ν + J+1))_i,kR(0,n)_k,j + ∑_k=1^N (G(n))_i,kR(0,n-1)_k,j. Since A has only non-zero entries a_i's in the place (i,i+1) and n+ν + J+1 is diagonal, we obtain that ∑_k=1^N(A(n+ν + J+1))_i,kR(0,n)_k,j= (a_i-1(n+ν + i))R(0,n)_i-1,j if i>1, 0 if i=1. Thus, for i>1 the equation (<ref>) takes the form -j R(0,n)_i,j= -(n+i)R(0,n)_i,j + a_i-1(n+ν + i)R(0,n)_i-1,j+ (G(n))_i,iR(0,n-1)_i,j where in the last term of the equality, we use Lemma <ref>. By Theorem <ref>, we have that (n+i-j) L^(ν+j)_n+i-j(0) ξ(n,i,j) = L^(ν+j)_n+i-1-j(0) (a_i-1(n+ν+i) ξ(n,i-1,j)+ (G(n))_i,iξ(n-1,i,j)). Therefore, since n+i-j>0 we obtain a similar expression of (<ref>) by multiplication for ((n+i-j) L^(ν+j)_n+i-j(0))^-1, that is ξ(n,i,j) = C_1(n,i,j)ξ(n,i-1,j) +C_2(n,i,j) ξ(n-1,i,j), where C_1(n,i,j) = (a_i-1(n+ν+i)) L^(ν+j)_n+i-1-j(0)(n+i-j) L^(ν+j)_n+i-j(0) and C_2(n,i,j) = (G(n))_i,i L^(ν+j)_n-1+i-j(0)(n+i-j) L^(ν+j)_n+i-j(0). To finish the proof, recall that L^(α)_N(0)=(α+1)_N/N!, where (a)_N is the Pochhammer symbol defined by (a)_N= 1 if N=0, a(a+1)⋯(a+N-1) if N>0. Hence, we have that L^(ν+j)_n+i-1-j(0)/L^(ν+j)_n+i-j(0)= n+i-j/n+ν+i. Therefore we obtain (<ref>), as asserted. Now, for i=1 the equation (<ref>) takes the form -j R(0,n)_1,j = -(n+1)R(0,n)_i,j + (G(n))_1,1R(0,n-1)_1,j. So, the equation (<ref>) it can be prooved in a similar way as in the case i>1. Notice that the item (a) in the above proposition still holds for n+i-j=0 since in this case we just use the definition of ξ's. Now, we are going to study the case n+i-j=0. In this case, we have the following result. Let n∈ℕ_0 and let I(n), G(n) be as in (<ref>). If i≥1, then the constants ξ's satisfy the following: * ξ(0,i,i)= 1 and ξ(1,i,i+1)= I(0)_i,i+1. * If i=1 and n>0, then N_1(n,i) ξ(n,1,n+1)= N_2(n,i) ξ(n-1,2,n+1), with N_1(n,i)=(ν+2n+3) (G(n+1))_1,1 n+ν+2 + (n+2+ν) + I(n)_1,2 (ν+2n+2) a_1, N_2(n,i)=I(n)_1,2 (ν+2n+2)(G(n))_2,2 n+ν+2. * If i>1 and n>0, then M_1(n,i)ξ(n+1,i-1,n+i)= M_2(n,i)ξ(n,i,n+i)+M_3(n,i)ξ(n-1,i+1,n+i), with M_1(n,i)=a_i-1( i(n+1+ν+i)_i-2-(n+ν+i)), M_2(n,i)=(G(n+1))_i,i + (n+ν+i+1), M_3(n,i)=(I(n))_i,i+1 G(n)_i+1,i+1. The first assertion of item (a) it follows from item (a) of <ref> (see Remark <ref>). Indeed, since K_n^-1 has a diagonal of 1's, in this case we have that ξ(0,i,i)= (K_0^-1)_i,i0!(ν+i+1)_0=(K_0^-1)_i,i=1. By Proposition <ref>, we have 𝒟 = ∂_x x + x(A-1) and M = (A-1)δ-(n+1+ν)-ℋ(n)Jℋ^-1(n) satisfies P_n ·𝒟 = M · P_n. By recalling that (P_n ·𝒟)(0)=0, we obtain that (A-1)P(0,n+1)-(n+1+ν)P(0,n)-ℋ_nJℋ^-1_n P(0,n)=0, since P(0,m)=K_m R(0,m), we have that K_n^-1(A-1)K_n+1 R(0,n+1)-(n+1+ν)R(0,n)-I(n) R(0,n)=0 where I(n)=K_n^-1ℋ_n Jℋ_n ^-1K_n. If we consider the expression (<ref>) with n=0, we arrive to 0 = K_0^-1(A-1)K_1 R(0,1)-K_0^-1(1+ν)K_0 R(0,0)-I(0) R(0,0). By taking the (i,i+1)-th coordinate, we obtain that 0 = - R(0,1)_i,i+1-(1+ν)R(0,0)_i,i+1-I(0)_i,i R(0,0)_i,i+1- I(0)_i,i+1R(0,0)_i+1,i+1. From Theorem <ref> we have ξ(1,i,i+1)= I(0)_i,i+1ξ(0,i+1,i+1)=I(0)_i,i+1. as asserted. Now, let us consider n>0 in the equation (<ref>). By taking into account the (i,n+i)-coordinate in (<ref>), ∑_r=1^N(K_n^-1(A-1)K_n+1)_i,rR(0,n+1)_r,n+i-(n+1+ν) R(0,n)_i,n+i-∑_r=1^NI(n)_i,rR(0,n)_r,n+i=0. By Lemma (<ref>) and since R(0,n+1)_r,n+i= 0 for r<i-1 we have that ∑_r=i-1^N(K_n^-1(A-1)K_n+1)_i,rR(0,n+1)_r,n+i-(n+1+ν+i) R(0,n)_i,n+i-I(n)_i,i+1R(0,n)_i+1,n+i=0. On the other hand, since the matrix K_n^-1(A-1)K_n+1 is lower triangular we obtain that ∑_r=i-1^N(K_n^-1(A-1)K_n+1)_i,rR(0,n+1)_r,n+i= ∑_r=i-1^i(K_n^-1(A-1)K_n+1)_i,rR(0,n+1)_r,n+i. By taking into account that (K_n^-1(A-1)K_n+1)_i,i-1= a_i-1 -(K_n^-1K_n+1)_i,i-1 and (K_n^-1(A-1)K_n+1)_i,i=-1, we obtain that 0 = (a_i-1 -(K_n^-1K_n+1)_i,i-1) R(0,n+1)_i-1,n+i- R(0,n+1)_i,n+i - (n+1+ν+i)R(0,n)_i,n+i-I(n)_i,i+1R(0,n)_i+1,n+i. By Corollary <ref> we have 0 = (a_i-1 -(K_n^-1K_n+1)_i,i-1) ξ(n+1,i-1,n+i) - (n+1+ν+i) ξ(n+1,i,n+i) - (n+1+ν+i) ξ(n,i,n+i) -I(n)_i,i+1 (n+ν +i+1)ξ(n,i+1,n+i). On the other hand, by Proposition <ref> if n+i-j>0 we know that ξ(n,i,j) = a_i-1ξ(n,i-1,j) +(G(n))_i,i n+ν+iξ(n-1,i,j). Thus, in particular we obtain ξ(n+1,i,n+i) = a_i-1ξ(n+1,i-1,n+i) +(G(n+1))_i,i n+1+ν+iξ(n,i,n+i), ξ(n,i+1,n+i) = a_iξ(n,i,n+i) + (G(n))_i+1,i+1)n+ν+i+1) ξ(n-1,i+1,n+i) and so we have (a_i-1 -(K_n^-1K_n+1)_i,i-1) ξ(n+1,i-1,n+i) - (ν+n+i+1) (a_i-1ξ(n+1,i-1,n+i) +(G(n+1))_i,i n+1+ν+iξ(n,i,n+i)) = (n+1+ν+i) ξ(n,i,n+i) + I(n)_i,i+1 (n+1+ν+i)( a_iξ(n,i,n+i) + (G(n))_i+1,i+1n+ν+i+1ξ(n-1,i+1,n+i)). Hence M_1(n,i)ξ(n+1,i-1,n+i)= M_2(n,i)ξ(n,i,n+i)+M_3(n,i)ξ(n-1,i+1,n+i), with M_1(n,i)=-(K_n^-1K_n+1)_i,i-1 - a_i-1(ν+n+i), M_2(n,i)=(G(n+1))_i,i + (n+ν+i+1), and M_3(n,i)=(I(n))_i,i+1 G(n)_i+1,i+1. To finish the proof, notice that since K_n^-1 and K_n+1 are both lower triangular matrix with 1's in the diagonal, we have that (K_n^-1K_n+1)_i,i-1=(K_n^-1)_i,i-1+(K_n+1)_i,i-1 and (K_n^-1)_i,i-1=-(K_n)_i,i-1, by definition of K_n's we have that (K_n^-1K_n+1)_i,i-1=a_i-1( (n+ν+i)_i-1-(n+1+ν+i)_i-1)=-a_i-1 i(n+1+ν+i)_i-2. Hence, M_1(n,i)=a_i-1( i(n+1+ν+i)_i-2-(n+ν+i)), as asserted. For the case i=1, j=n+1 in the expression (<ref>) we obtain 0 = - R(0,n+1)_1,n+1-(n+2+ν) R(0,n)_1,n+1-I(n)_1,2R(0,n)_2,n+1, then, from Corollary <ref> we have 0 = - (ν+2n+3)ξ(n+1,1,n+1)-(n+2+ν) ξ(n,1,n+1)-I(n)_1,2 (ν+2n+2) ξ(n,2,n+1). As a consequence of Proposition <ref> we have that ξ(n,i,j) = a_i-1ξ(n,i-1,j) +(G(n))_i,i n+ν+iξ(n-1,i,j) for j<n+i, which implies that ξ(n,2,n+1) = a_1ξ(n,1,n+1) +(G(n))_2,2 n+ν+2ξ(n-1,2,n+1), and ξ(n,1,j) =(G(n))_1,1 n+ν+1ξ(n-1,1,j) for j<n+1. In particular ξ(n+1,1,n+1)=(G(n+1))_1,1 n+ν+2ξ(n,1,n+1). Thus, we obtain N_1(n,i) ξ(n,1,n+1)= N_2(n,i) ξ(n-1,2,n+1), with N_1(n,i)= (ν+2n+3) (G(n+1))_1,1 n+ν+2 + (n+2+ν) + I(n)_1,2 (ν+2n+2) a_1, N_2(n,i)=I(n)_1,2 (ν+2n+2)(G(n))_2,2 n+ν+2 as desired. As a consequence of Propositions <ref> and <ref> we obtain the following result. Let n≥ 0 and let G(n),I(n) as in (<ref>). Then, all of the non-zero entries of R(x,n) can be found in terms of G(ℓ) and I(ℓ) and the generalized Laguerre polynomials L^(α)_ℓ for ℓ=0,…,n. The equality (<ref>) implies that (R(x,n))_i,j=L^(ν+j)_n+i-j(x)ξ(n,i,j) for n+i-j≥ 0. It is enough to show that all of the non-zero constants ξ(n,i,j) can be obtained in terms of G(ℓ) and I(ℓ) for ℓ=0,…,n+1. By items (a)'s of Propositions <ref> and <ref> we obtain the values of ξ(0,i,j). Thus, assume that n>0 and suppose that we want to determine ξ(n,i,j) with n+i-j>0. Notice that each time that we use items (b) and (c) of Proposition <ref>, the value of "n+i-j" it reduces by 1, so if we use these items inductively we obtain that ξ(n,i,j) can be determined by the values of somes ξ(n',i',j')'s with n'+i'-j'=0 and in each step also appear G(ℓ) with ℓ=0,…,n. So, it is enough to see that we can determine ξ(n,i,j) with n+i-j=0, in terms of G(ℓ) and I(ℓ). Notice that by Lemma <ref> and Propositions <ref>, <ref>, we can replace (G(n))_i,i and (I(n))_i,i+1 by the expressions (ℋ_n(A^∗-1)ℋ_n-1^-1))_i,i and (ℋ_n J ℋ_n^-1)_i,i+1, respectively. As a direct consequence of the above results and the equations obtained in Proposition <ref>, we obtain a three-terms non-linear recursion for ℋ_n. With this in mind, we need the following lemma. Let A,J∈ M_N(ℂ) be matrices as in (<ref>). Then, the matrix ℋ_1 can be obtained from ℋ_0. By (<ref>), we have that ℋ_1=(X(1)+[J,X(1)])ℋ_0 (A^*-1)^-1. So, it is enough to show that we can obtain X(1) in terms of ℋ_0. In order to show this, recall that P(x,1)=x+ X(1). Let us put P_1(x):=P(x,1), by applying the operator D=∂_x^2 x +∂_x ( (A-1)x +1+ν+J) +Aν + JA -J to P_1(x), we have that (P_1· D)(x)= (A-1)x +1+ν+J + (x+X(1))(Aν + JA -J), by taking into account that P_1· D= Γ_1 · P_1, where Γ_1 is the discrete constant operator A(ν+2+J)-1-J. So we obtain that (A-1)x +1+ν+J + (x+X(1))(Aν + JA -J)= (A(ν+2+J)-1-J)(x+X(1)). After some computation, we obtain that X(1)(Aν + JA -J) +1+ν+J = (A(ν+2+J)-1-J)X(1), by taking into account that [J,A]=A, we obtain that JA=A+AJ and hence [X(1),A(1+ν+J) ] +X(1)+[J,X(1)] + 1+ν+J -AX(1)=0. Thus, by seeing the (i,j)-th coordinate in the above equality we obtain the following recurrences a_j(ν+j+1) X(1)_1,j+1+ (2-j) X(1)_1,j+ (1+ν+J)_1,j=0, a_j(ν+j+1) X(1)_i,j+1- a_i-1(ν+i+1)X(1)_i-1,j + (1+i-j) X(1)_i,j+ (1+ν+J)_i,j=0 for i≥ 2. Recall that by Theorem <ref> we have that (R(0,1))_i,j=0 if i+1< j. Claim: If i+1<j then X(1)_i,j=0. We are going to prove this assertion by induction on i. Assume first that i=1 and let us consider j>2, in this case 0=(R(0,1))_1,j= (K_1 P(0,1))_1,j=∑_r=1^N(K_1)_1,r X(1)_r,j=X(1)_1,j. Now, let i>1 and assume that the statement is true for r<i, that is X(1)_r,j=0 when r+1<j. Thus, if i+1<j then 0=(R(0,1))_i,j= (K_1 P(0,1))_i,j=∑_r=1^N(K_1)_i,r X(1)_r,j= (K_1)_i,iX(1)_i,j=X(1)_i,j. By induction hypothesis and by taking into account that K_1 is a lower triangular matrix with 1's in its diagonal, we have that ∑_r=1^N(K_1)_i,r X(1)_r,j=∑_r=1^i(K_1)_i,r X(1)_r,j= (K_1)_i,iX(1)_i,j=X(1)_i,j. Therefore, X(1)_i,j=0 when i+1<j as claimed. Hence, by equations (<ref>) and (<ref>), we can observe by a recursive argument, that in order to compute X(1), it is enough to know the value of X(1)_i,i+1 for any i≥ 1. By taking into account that (R(0,1))_i,i+1=(K_1 P(0,1))_i,i+1=∑_r=1^i(K_1)_i,r (X(1))_r,i+1=X(1)_i,i+1, thus by Corollary <ref>, we obtain that X(1)_i,i+1=(R(0,1))_i,i+1=ξ(1,i,i+1). Thus, by item (a) of Proposition <ref>, we obtain that X(1)_i,i+1=ξ(1,i,i+1)=I(0)_i,i+1= (ℋ_0 J ℋ_0^-1)_i,i+1. Let A,J∈ M_N(ℂ) be matrices as in (<ref>). Then, ℋ_n+2 = (A-1)^-2( (-[J,ℋ_n J ℋ_n^-1] - ℋ_n (A^t-1) ℋ_n-1^-1 (A-1) + (A-1) ℋ_n+1 (A^t-1) ℋ_n^-1)(A-1) -2-ℋ_n+1 J ℋ_n+1^-1 + ℋ_n J ℋ_n^-1 +(A-1)[J,ℋ_n+1 J ℋ_n+1^-1]+(A-1)ℋ_n+1(A^t-1)ℋ_n^-1(A-1) )ℋ_n+1 (A^t-1)^-1, where B_n,C_n and ℋ_n be as in (<ref>). Moreover, (ℋ_0)_i,j = Γ(ν) ∑_r=1^min{i,j}δ_r^(ν)/(i-r)!(j-r)!(∏_k=r^i-1 a_k ) (∏_s=r^j-1 a_s) (ν)_i+j-r, where Γ is the Gamma function. By (<ref>), we have that B_n+1 =(A-1)^-1(B_n (A-1) - 2 - ℋ_n+1 J ℋ_n+1^-1 + ℋ_n J ℋ_n^-1). On the other hand, the equation (<ref>) implies B_n = -[J,ℋ_n J ℋ_n^-1] - ℋ_n (A^t-1) ℋ_n-1^-1 (A-1) + (A-1) ℋ_n+1 (A^t-1) ℋ_n^-1. B_n+1 = -[J,ℋ_n+1 J ℋ_n+1^-1] - ℋ_n+1 (A^t-1) ℋ_n^-1 (A-1) + (A-1) ℋ_n+2 (A^t-1) ℋ_n+1^-1 The statement it follows by changing B_n+1, and B_n in (<ref>). For the last assertion, by definition we have that ℋ_0= ∫_0^∞ e^xA T^(ν)(x)e^xA^∗ dx with T^(ν)(x)=e^-x∑_k=1^Nδ_k^(ν)x^ν+kE_k,k. Then, (ℋ_0)_i,j = ∑_r=1^Nδ_r^(ν)∫_0^∞ e^-x x^(ν+r) (e^xA)_i,r (e^xA^∗)_r,j. By taking into account that A^∗_r,j=A_j,r, and (e^xA)_i,r=1(i-r)!((xA)^i-r)_i,r if r< i and 0 otherwise, we have (ℋ_0)_i,j = ∑_r=1^min{i,j}δ_r^(ν)∫_0^∞ e^-x x^(ν+r)((xA)^i-r)_i,r(i-r)!((xA)^j-r)_j,r(j-r)! dx = ∑_r=1^min{i,j}δ_r^(ν)(A^i-r)_i,r(i-r)!(A^j-r)_j,r(j-r)!∫_0^∞ e^-x x^(ν+i+j-r) dx. Notice that (A^i-r)_i,r= ∏_k=r^i-1a_k, and recall that Γ(z+1)=∫_0^∞ e^-x x^z dx, we can write (ℋ_0)_i,j= ∑_r=1^min{i,j}δ_r^(ν)/(i-r)!(j-r)!(∏_k=r^i-1 a_k ) (∏_s=r^j-1 a_s) Γ(ν+i+j-r+1). Taking into account that Γ(z+1)=zΓ(z) we obtain that (ℋ_0)_i,j = Γ(ν) ∑_r=1^min{i,j}δ_r^(ν)/(i-r)!(j-r)!(∏_k=r^i-1 a_k ) (∏_s=r^j-1 a_s) (ν)_i+j-r, as asserted. Notice that we also have the expression ℋ_0 = (L^(ν)_μ(0))^-1ℋ^(ν,ν)_0 (L^(ν)_μ(0))^∗, with L^(ν)_μ(0) and ℋ^(ν,ν)_0 defined in <cit.>, where μ=(μ_1, …, μ_N) such that a_i=μ_i+1μ_i for i=1, …, N-1. This assertion can be deduced from W^ν(x)= (L^(ν)_μ(0))^-1 W^(ν,ν)(x) (L^(ν)_μ(0))^∗, where W^(ν,ν)(x) are as in <cit.>. They also proved that ℋ^(ν,ν)_0 is a diagonal matrix. The following result gives a recursion for X(n)'s in terms of I(n). For n ≥ 1, let I(n) as in (<ref>). If δ_i,j denotes the kronecker delta function, then: * n δ_1,j + X(n)_1,j+1 a_j - X(n)_1,j = -(n+1+ν)δ_1,j - I(n)_1,j. * n δ_i,j + X(n)_i,j+1 a_j - a_i-1 X(n+1)_i-1,j - X(n)_i,j + X(n+1)_i,j = -(n+1+ν)δ_i,j - I(n)_i,j for i≥ 2. By Corollary <ref>, we have that n+X(n)A - A X(n+1)-B(n) = -(n+1+ν)-ℋ(n)Jℋ^-1(n). The result is a direct consequense of taking (i,j)-th coordinate in the above equation. * By item (a) of Corollary <ref> we can obtain that G(n)_i,i=X(n)_i,i for any i ≥ 1. * In order to compute P(x,n), for a given n, we can do the following: * Compute ℋ_n and the explicit form of ℋ_0, ℋ_1 using Lemma <ref> and Proposition <ref>. * Compute G(n) and I(n) using ℋ_n. * Compute the ξ(n,i,j) using the recursions of Propositions <ref> and <ref>. * Then we have R(x,n) and so P(x,n)=K_n R(x,n) e^-xA. § MATRIX ENTRIES OF P(X,N) IN TERMS OF LAGUERRE AND DUAL HAHN POLYNOMIALS. In this section, we will show that under some hypothesis, the ξ(n,i,j)'s can be expressed in terms of dual Hahn polynomials. §.§ Some technical lemmas Let N≥ 1 be a fixed integer and let μ=(μ_1,…,μ_N) be a sequence of non-zero coefficients and α>0. Then L_μ^() is the N× N unipotent lower triangular matrix defined by L^(α)_μ(x)_m,n=μ_m/μ_n L^(α+n)_m-n(x), m≥ n, 0 n<m. For ν>0 we consider the weight matrix W^(α,ν)_μ(x)=L_μ^(α)(x) T^(ν)(x) L_μ^(α)(x)^∗, T^(ν)(x)=e^-x∑_k=1^N x^ν+kδ_k^(ν) E_k,k. It can be showed that W_μ^(α, ν)(x) = L^(α)_μ(0) e^xA_μ T^(ν)(x) e^xA_μ^∗ L^(α)_μ(0)^∗ where A_μ=- ∑_k=1^N-1μ_k+1μ_kE_k+1,k. We impose conditions on the sequence {μ_i}_i=1^N and the coefficients δ_k^(ν). First of all, we assume that the coefficients μ_i are real and non-zero for all i and δ_k^(ν)>0, 1≤ k≤ N, so that the weight matrix is positive definite (see <cit.> for more information about the weight matrix W^(α,ν)_μ). On the other hand, we consider the diagonal matrix Δ^(ν) = diag(δ_1^(ν), …, δ_N^(ν)), so that (T^(ν))_k,k=e^-x x^ν+k (Δ^(ν))_k,k. We assume that there exist coefficients c^(ν) and d^(ν) such that Δ^(ν+1)=(d^(ν) J+c^(ν)) Δ^(ν). We also assume that the coefficients μ_k and δ^(ν)_k satisfy the relation μ_k+1^2μ_k^2=d^(ν)k(N-k) δ_k+1^(ν)δ_k^(ν+1), k=1,…,N-1. Under the above conditions, Propositions 5.1 and 5.2 from <cit.> say that Φ^(α, ν)(x)=(W_μ^(α, ν)(x))^-1W_μ^(α, ν+1)(x) and Ψ^(α, ν)(x)=(W_μ^(α, ν)(x))^-1dW_μ^(α, ν+1)dx(x) are matrix polynomials of degree 2 and 1 respectively. Moreover, Corollary 6.3 from <cit.> asserts that the operator D_2 defined by D_2(x)= d^2dx^2Φ^∗(x) + ddxΨ^∗(x) is symmetric respect to W_μ^(α, ν). We begin with the following technical lemma which relates the matrix polynomials R_n(x) with the constants c^( ν) and d^( ν). The proof of Lemma <ref> and Lemma <ref> can be found in the appendix. Let μ=(μ_1,…,μ_N) and δ^(ν)_k>0 for 1≤ k≤ N, satisfying (<ref>) and (<ref>). Let A:=A_μ as in (<ref>). If R(x,n) are the matrix polynomials defined in (<ref>), then ( R'(0,n)- R(0,n) A)C^(ν)=D^(ν)R(0,n) where C^(ν)= (d^(ν)J+c^(ν))(ν+J+1)+ ((Δ^(ν))^-1A Δ^(ν+1))^∗ and D^(ν)= n(d^(ν)(J-N-1)-c^(ν)) with c^( ν) and d^( ν) as in (<ref>). In the sequel, let 0 ≤ n, 1 ≤ i ≤ N. We consider the sequence ϵ_j:=ϵ_j^(n,i) defined recursively by ϵ_0=1, ϵ_j=(n+i-j+1) d^(ν)((j-1) + c^(ν)d^(ν)) ϵ_j-1 for n+i-j ≥ 0, with c^(ν), d^(ν) defined as in (<ref>). We have the following result. Let 0≤ n and let 1 ≤ N, 1 ≤ i ≤ N integers and let {ϵ_l}_l=0^i+n be the sequence defined as in (<ref>). If ν>0, δ^(ν) satisfies (<ref>) and 0 ≤ j < n+i, we have ϵ_jϵ_j+1M_j=1 where M_j= (n+i-j) (d^(ν) j + c^(ν)) with c^(ν), d^(ν) as in (<ref>). It follows by a simple inductive argument from the definition of ϵ_j. Given i=1,…,N, n ≥ 0 and j a positive integer such that n+i-j ≥ 0, in the sequel we consider the sequence q_j:=q_j^(n,i) by the expression q_j^(n,i) := ϵ_j^(n,i)ξ(n,i,j), where ϵ_j's are as in (<ref>). Let μ=(μ_1,…,μ_N) and δ^(ν)_k>0 for 1≤ k≤ N, satisfying (<ref>) and (<ref>). For 1 ≤ i ≤ N and 0 ≤ n, let {q_l}_l=0^i+n be the sequence as in (<ref>). Then, the sequence {q_l}_l=0^i+n satisfies E_j q_j + F_j q_j-1 + 1d^(ν) q_j+1 = 0, n+i-j > 0, where E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( (i-N-1)-c^(ν)d^(ν)) , F_j =(j-1) (N-j+1) μ_j-1μ_j (n+i-j+1) (d^(ν) (j-1) + c^(ν)) with c^(ν), d^(ν) as in (<ref>). §.§ Dual Hahn polynomials For 0 ≤ n, 1 ≤ i ≤ N let us consider the following sequence q_j^(n,i):= (d^(ν))^-j q_j^(n,i), n+i-j > 0, where q_j's are as in (<ref>) and d^(ν) as (<ref>). Let μ=(μ_1,…,μ_N) and 0 < δ^(ν)_k, for 1≤ k≤ N, satisfying (<ref>) and (<ref>). Let q_l be as in (<ref>). If μ_j=1 for all j then the q_l's satisfy the following equation E_j q_j+ F_j q_j-1 + q_j+1 = 0, n+i-j > 0, where E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( i-N-1-c^(ν)d^(ν)), F_j = (j-1) (N-j+1) (n+i-j+1) (j-1 + c^(ν)d^(ν)) with c^(ν), d^(ν) as in (<ref>). By Lemma <ref>, we have that E_j q_j + F_j q_j-1 + 1d^(ν) q_j+1 = 0 with E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( (i-N-1)-c^(ν)d^(ν)) , F_j =(j-1) (N-j+1) μ_j-1μ_j (n+i-j+1) (d^(ν) (j-1) + c^(ν)). Now, since μ_j=1 for all j and q_j(x):= (d^(ν))^-j q_j(x), we have that ((n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( (i-N-1)-c^(ν)d^(ν)) ) q_j(d^(ν))^j + (j-1) (N-j+1) (n+i-j+1) (d^(ν) (j-1) + c^(ν)) (d^(ν))^j-1q_j-1 + 1d^(ν)(d^(ν))^j+1q_j+1 = 0. By multiplication for (d^(ν))^-j, we obtain that E_j q_j+ F_j q_j-1 + q_j+1 = 0, with E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( i-N-1-c^(ν)d^(ν)), F_j = (j-1) (N-j+1) (n+i-j+1) (j-1 + c^(ν)d^(ν)), as asserted. Recall that if the sequence of polynomials {s_k} satisfies the normalized recurrence relations xs_k(x)= s_k+1(x) -(u_k+v_k) s_k(x)+u_k-1v_k s_k-1(x), with u_k=(k+γ+1)(k-M), v_k=k(k-δ-M-1), then s_k satisfies T_k(λ(x);γ,δ,M) = 1(γ+1)_k (-M)_k s_k(λ(x)) where λ(x)= x(x+γ+δ +1) and {T_k} is the sequence of dual Hahn polynomials defined by T_k(λ(x);γ,δ,M)= _3 F_2(-k, -x, x+γ+δ+1 γ+1, -M ;1) for k=0,1,…,M. Let μ=(μ_1,…,μ_N) and δ^(ν)_k>0 for 1≤ k≤ N, satisfying (<ref>) and (<ref>). Let q_j be as in (<ref>). If μ_ℓ=1 for all ℓ=1,…,N, then the q_j's satisfy q_j= (γ+1)_j-1(-(N-1))_j-1 T_j-1(λ(x^(n,i)); γ,δ, N-1) n+i-j > 0 where γ= c^(ν)d^(ν), δ=n+i-N and x^(n,i)=(γ+1)(N+i-2)-n(N-i). By Lemma <ref>, we have that E_j q_j+ F_j q_j-1 + q_j+1 = 0, n+i-j > 0, where E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( i-N-1-c^(ν)d^(ν)), F_j = (j-1) (N-j+1) (n+i-j+1) (j-1 + c^(ν)d^(ν)) with c^(ν), d^(ν) as in (<ref>). Now, if we consider u_k=(k+γ+1)(k-M) and v_k=k(k-δ-M-1), with γ= c^(ν)d^(ν), δ=n+i-N and M=N-1. It is straightforward to check that F_j = u_j-2 v_j-1 and E_j = -(u_j-1+v_j-1+x^(n,i)) where x^(n,i)=(γ+1)(N+i-2)-n(N-i). Thus, since q_j can be seen as constants polynomial, we have that x^(n,i)q_j(x^(n,i)) = q_j+1(x^(n,i)) -(u_j-1+v_j-1) q_j(x^(n,i))+u_j-2 v_j-1q_j-1(x^(n,i)). Therefore, by definition of dual Hahn polynomials and by taking into account that q_j is a constant polynomial, we obtain (<ref>) as desired. By recalling that q_j=(d^(ν))^-j q_j = ϵ_j (d^(ν))^-jξ(n,i,j) for n+i-j > 0, we obtain the following result Let μ=(μ_1,…,μ_N) and δ^(ν)_k > 0, for 1≤ k≤ N, satisfying (<ref>) and (<ref>). Let ϵ_j as in (<ref>). If μ_ℓ=1 for all ℓ=1,…,N, then the constants ξ(n,i,j)'s satisfy ξ(n,i,j)= (d^(ν))^j(γ+1)_j-1(-(N-1))_j-1ϵ_j T_j-1(λ(x^(n,i)); γ,δ, N-1), n+i-j > 0 where γ= c^(ν)d^(ν), δ=n+i-N and x^(n,i)=(γ+1)(N+i-2)-n(N-i). Moreover, if n+i=j, the constants ξ(n,i,j)'s satisfy the recursion ξ(0,1,1)=1ν +2, ξ(n,i,j-1) = ( n(N+1-i)(i-1)(j-1) (N-j+1) + 1 )ξ(n,i,j), j >1, If n+i-j > 0, by Proposition <ref> we have that ξ(n,i,j)=(d^(ν))^jϵ_jq_j= (d^(ν))^j(γ+1)_j-1(-(N-1))_j-1ϵ_j T_j-1(λ(x^(n,i)); γ,δ, N-1) with γ= c^(ν)d^(ν), δ=n+i-N and x^(n,i)=(γ+1)(N+i-2)-n(N-i) as asserted. Now, if we take j=n+i in the expression (<ref>), we obtain R_n'(0)_i,j (d^(ν) j + c^(ν)) (ν + j +1) + R_n'(0)_i(j-1)((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1)j - (R_n(0)_i,j A_j(j-1)((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1),j = n(d^(ν)(J-N-1)-c^(ν)) R(n,0)_i,j. By taking into account that R_n(x)_i,j=L^(ν+j)_n+i-j(x) ξ(n,i,j), L^α_n(0)=(α+1)_nn! and ∂∂ xL^α_n(x)=(-1) L^α+1_n-1(x), we have - ξ(n,i,j-1) ((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1)j - (ξ(n,i,j) A_j(j-1)((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1),j = n(d^(ν)(i-N-1)-c^(ν)) ξ(n,i,j). Recall that by (<ref>) we have that ((Δ^(ν))^-1 A Δ^(ν+1))^∗_(j-1,j) = d^(ν) (j-1) (N-j+1) μ_j-1μ_j, then, taking in account that μ_k=1 for all k, we have - ξ(n,i,j-1) d^(ν) (j-1) (N-j+1) - (ξ(n,i,j) A_j(j-1)d^(ν) (j-1) (N-j+1) = n(d^(ν)(i-N-1)-c^(ν)) ξ(n,i,j). Thus, using the conditions (<ref>), (<ref>), and taking in account that A_s,s-1=-1, we obtain - d^(ν) (j-1) (N-j+1) ξ(n,i,j-1) = (-nd^(ν)(N+1-i)(i-1) - d^(ν) (j-1) (N-j+1) )ξ(n,i,j), as desired. The last assertion it follows from ξ(0,1,1)(ν +2)=ξ(0,1,1) L^1_1(0)=R(0,0)_1,1=(K_1 P(0,0))_1,1=1 plain § In this apprendix we give the proofs of the Lemmas <ref> and <ref>. Let P_n(x)=P(x,n) be the sequence of monic orthogonal polynomials respect to the weight W^(ν) as in (<ref>). Since W_μ^(α, ν)(x)=L_μ^(α)(0) W^(ν)(x) L_μ^(α)(0)^* we have that ∫_0^∞ P_n(x) (L_μ^(α)(0)^-1)W_μ^(α, ν)(x) (P_m(x)L_μ^(α)(0)^-1)^* dx =∫_0^∞ P_n(x) W^(ν)(x) P_m(x)^* dx= δ_n,mℋ_n. Hence, if P_n^(α,ν)(x)=L_μ^(α)(0) P_n(x) L_μ^(α)(0)^-1 then P^(α,ν)_n(x) is the sequence of monic orthogonal polynomials with respect to the weight W_μ^(α, ν). Now, if D_2 is the operator defined in (<ref>), the Corollary 6.3 of <cit.> implies that P^(α,ν)_n D_2= n K̂_n^(α,ν) P^(α,ν)_n for certain matrix K̂_n^(α,ν) defined recursively in section 6 from <cit.>, and so P_n L_μ^(α)(0)^-1D_2 L_μ^(α)(0)=nL_μ^(α)(0)^-1K̂_n^(α,ν) L_μ^(α)(0) P_n. By Proposition 6.1 of <cit.>, we obtain that (L_μ^(α)(0))^-1K̂_n^(α,ν) L_μ^(α)(0)=d^(ν)(J-(J+ν+n)A-N-1)-c^(ν), and by taking into account that JA-AJ=A, we have that J-JA-(ν+n)A-N-1=J-AJ-A-(ν+n)A-N-1. On the other hand, since Γ_n=A(n+ν+J+1)-(n+J)=-J+(n+ν)A+A+AJ-n, then (L_μ^(α)(0))^-1K̂_n^(α,ν) L_μ^(α)(0)=d^(ν)(-Γ_n-n-N-1)-c^(ν). By (<ref>), we have that K_n R_n e^-xA(L_μ^(α)(0))^-1D_2 L_μ^(α)(0)= n (L_μ^(α)(0))^-1K̂_n^(α,ν) L_μ^(α)(0)K_n R_n e^-xA and thus R_n e^-xA(L_μ^(α)(0))^-1D_2 L_μ^(α)(0)e^xA=n K_n^-1(d^(ν)(-Γ_n-n-N-1)-c^(ν))K_n R_n. By recalling that K_n^-1Γ_n K_n =-n-J, we have that R_n e^-xA(L_μ^(α)(0))^-1D_2 L_μ^(α)(0)e^xA=n(d^(ν)(J-N-1)-c^(ν))R_n. Now, by taking into account that D_2= d^2dx^2Φ^∗(x) + ddxΨ^∗(x) with Φ, Ψ polynomials of degree 2 and 1. In general, if U(x) is a polynomial, we have that U(x) · e^-xA(L_μ^(α)(0))^-1D_2 L_μ^(α)(0)e^xA = d^2dx^2( U(x) e^-xA) (L_μ^(α)(0))^-1Φ^∗(x) L_μ^(α)(0)e^xA + ddx( U(x) e^-xA) (L_μ^(α)(0))^-1Ψ^∗(x) L_μ^(α)(0)e^xA = ( U”(x)-2 U'(x)A+ U(x) A^2) e^-xA(L_μ^(α)0))^-1Φ^∗(x) L_μ^(α)(0)e^xA + ( U'(x)- U(x) A) e^-xA(L_μ^(α)(0))^-1Ψ^∗(x) L_μ^(α)(0)e^xA. Hence, we want to find some easy expression for e^-xA L(0)^-1Φ^∗(x) L(0) e^xA and e^-xA L(0)^-1Ψ^∗(x) L(0) e^xA. Some similar expressions was studied by Koelink and Roman (see <cit.>). By Corollary 5.3 in <cit.>, we have that L_μ^(α)(0)^∗Φ(x) ((L_μ^(α)(0))^∗)^-1 = -d^(ν)x^2 A^∗+x (d^(ν) J +c^(ν)) L_μ^(α)(0)^∗Ψ(x) ((L_μ^(α)(0))^∗)^-1 = x(d^(ν)(J-A^∗(J+ν+1) -N-1)-c^(ν)) ) + (ν+J+1)(d^(ν)J+c^(ν))+(Δ^(ν))^-1A Δ^(ν+1). Hence, we have that e^-xAL_μ^(α)(0)^-1Φ^∗(x)L_μ^(α)(0) e^xA = -d^(ν)x^2 e^-xAA e^xA+xd^(ν)e^-xAJe^xA+x c^(ν) = -d^(ν)x^2 A +xd^(ν)(xA+J)+ xc^(ν) = x(d^(ν)J+c^(ν)). e^-xAL_μ^(α)(0)^-1Ψ^∗(x)L_μ^(α)(0) e^xA = x(d^(ν)(xA+J-(xA+J+ν+1)A-N-1)-c^(ν)) ) + (d^(ν)(xA+J)+c^(ν))(ν+xA+J+1)+e^-xA ((Δ^(ν))^-1A Δ^(ν+1))^∗ e^xA. By evaluation in x=0 in (<ref>) and by taking into account the above expressions, we obtain (<ref>), as desired. By taking into account that (R_n(x))_i,j satisfies the expression given by Theorem <ref>, the (i,j)-coordinate of the matrix in the right hand of (<ref>) is n(d^(ν)(i-N-1)-c^(ν))(R_n(0))_i,j= n(d^(ν)(i-N-1)-c^(ν))L^(ν+j)_n+i-j(0) ξ(n,i,j). On the other hand, for n+i-j > 0 the (i,j)-coordinate of the matrix in the left hand of (<ref>) is R_n'(0)_i,j (d^(ν) j + c^(ν)) (ν + j +1) + R_n'(0)_i(j-1)((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1)j - R_n(0)_i(j+1) (d^(ν) j + c^(ν)) (ν + j +1) A_(j+1)j - (R_n(0)_i,j A_j(j-1)((Δ^(ν))^-1A Δ^(ν+1))^∗_(j-1),j. By taking into account that R_n(x)_i,j=L^(ν+j)_n+i-j(x) ξ(n,i,j), L^α_n(0)=(α+1)_nn! and ∂∂ xL^α_n(x)=(-1) L^α+1_n-1(x), we obtain that the above expression is equivalent to - n (d^(ν) (i-N-1)-c^(ν)) (ν+j+1)_n+i-j(n+i-j)!ξ(n,i,j) = (ν+j+2)_n+i-j-1(n+i-j-1)! (d^(ν) j + c^(ν)) (ν +j +1) ξ(n,i,j) + (ν+j+1)_n+i-j(n+i-j)!( (Δ^(ν))^-1 A Δ^(ν+1))^∗_(j-1,j)ξ(n,i,j-1) + (ν+j+2)_n+i-j-1(n+i-j-1)! (d^(ν) j + c^(ν)) (ν+j+1) ξ(n,i,j+1) + (ν+j+1)_(n+i-j)(n+i-j)! A_j(j-1)((Δ^(ν))^-1 A Δ^(ν))^∗_(j-1)jξ(n,i,j). Thus, by taking into account that (ν+j+2)_n+i-j-1(ν +j +1)= (ν+j+1)_n+i-j-1, and multiplying both sides by (n+i-j-1)!(ν+j+1)_n+i-j-1, we obtain - n (d^(ν) (i-N-1)-c^(ν)) (n+i-j-1)n+i-jξ(n,i,j) = (d^(ν) j + c^(ν)) ξ(n,i,j) + (n+i-j-1)n+i-j( (Δ^(ν))^-1 A Δ^(ν+1))^∗_(j-1,j)ξ(n,i,j-1) + (d^(ν) j + c^(ν)) ξ(n,i,j+1) + (n+i-j-1)n+i-j A_j(j-1)((Δ^(ν))^-1 A Δ^(ν))^∗_(j-1)jξ(n,i,j). Finally, we obtain that ((n+i-j)(d^(ν) j + c^(ν)) + A_j(j-1)((Δ^(ν))^-1 A Δ^(ν))^∗_(j-1)j + n (d^(ν) (i-N-1)-c^(ν)) ) ξ(n,i,j) + ( (Δ^(ν))^-1 A Δ^(ν+1))^∗_(j-1,j)ξ(n,i,j-1) + (n+i-j) (d^(ν) j + c^(ν)) ξ(n,i,j+1) = 0. By (<ref>) we have that δ^(ν)_k+1δ^(ν+1)_k = μ_k+1^2d^(ν) k (N-k) μ_k^2, and then ((Δ^(ν))^-1 A Δ^(ν+1))^∗_(j-1,j) = δ^(ν+1)_j-1δ^(ν)_j A_j(j-1) = d^(ν) (j-1) (N-j+1) μ_j-1^2μ_j^2μ_jμ_j-1 = d^(ν) (j-1) (N-j+1) μ_j-1μ_j. Thus, we can rewrite the above equation as follows ((n+i-j)(d^(ν) j + c^(ν)) + d^(ν) (j-1) (N-j+1) + n (d^(ν) (i-N-1)-c^(ν)) ) ξ(n,i,j) + d^(ν) (j-1) (N-j+1) μ_j-1μ_jξ(n,i,j-1) + (n+i-j) (d^(ν) j + c^(ν)) ξ(n,i,j+1) = 0. Now, if we consider M_j=(n+i-j) (d^(ν) j + c^(ν)). Since ϵ^(n,i)_j ξ(n,i,j) = q_j^(n,i), we can rewrite the above equation as follows ((n+i-j)(d^(ν) j + c^(ν)) + d^(ν) (j-1) (N-j+1) + n (d^(ν) (i-N-1)-c^(ν)) ) q_jϵ_j + d^(ν) (j-1) (N-j+1) μ_j-1μ_jq_j-1ϵ_j-1 + M_j q_j+1ϵ_j+1 = 0. If we multiply by ϵ_j, we have that ((n+i-j)(d^(ν) j + c^(ν)) + d^(ν) (j-1) (N-j+1) + n (d^(ν) (i-N-1)-c^(ν)) ) q_j + d^(ν) (j-1) (N-j+1) μ_j-1μ_jϵ_jϵ_j-1 q_j-1 + M_j ϵ_jϵ_j+1 q_j+1 = 0. By Lemma <ref>, we have that M_j ϵ_jϵ_j+1 =1 and so we obtain d^(ν)((n+i-j)( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( (i-N-1)-c^(ν)d^(ν)) ) q_j + d^(ν) (j-1) (N-j+1) μ_j-1μ_j (n+i-j+1) (d^(ν) (j-1) + c^(ν)) q_j-1 + q_j+1 = 0, which is equivalent to E_j q_j + F_j q_j-1 + 1d^(ν) q_j+1 = 0, with E_j =(n+i-j) ( j + c^(ν)d^(ν)) + (j-1) (N-j+1) + n ( (i-N-1)-c^(ν)d^(ν)), F_j =(j-1) (N-j+1) μ_j-1μ_j (n+i-j+1) (d^(ν) (j-1) + c^(ν)), as asserted.

http://arxiv.org/abs/2303.07222v1

20230313155731

Reconciling rough volatility with jumps

q-fin.MF

Efficient Semantic Segmentation by Altering Resolutions for Compressed Videos Yubin Hu1 Yuze He1 Yanghao Li1 Jisheng Li1 Yuxing Han2 Jiangtao Wen3 Yong-Jin Liu1Corresponding author. 1Department of Computer Science and Technology, Tsinghua University 2Shenzhen International Graduate School, Tsinghua University 3Eastern Institute for Advanced Study {huyb20, hyz22, liyangha18}@mails.tsinghua.edu.cn, jas0n1ee@icloud.com, yuxinghan@sz.tsinghua.edu.cn, jtwen@eias.ac.cn, liuyongjin@tsinghua.edu.cn March 30, 2023 ======================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================= We reconcile rough volatility models and jump models using a class of reversionary Heston models with fast mean reversions and large vol-of-vols. Starting from hyper-rough Heston models with a Hurst index H∈(-1/2,1/2), we derive a Markovian approximating class of one dimensional reversionary Heston-type models. Such proxies encode a trade-off between an exploding vol-of-vol and a fast mean-reversion speed controlled by a reversionary time-scale ϵ>0 and an unconstrained parameter H ∈ℝ. Sending ϵ to 0 yields convergence of the reversionary Heston model towards different explicit asymptotic regimes based on the value of the parameter H. In particular, for H ≤ -1/2, the reversionary Heston model converges to a class of Lévy jump processes of Normal Inverse Gaussian type. Numerical illustrations show that the reversionary Heston model is capable of generating at-the-money skews similar to the ones generated by rough, hyper-rough and jump models. Mathematics Subject Classification (2010): 91G20, 60G22, 60G51 JEL Classification: G13, C63, G10. Keywords: Stochastic volatility, Heston model, Normal Inverse Gaussian, rough Heston model, Ricatti equations § INTRODUCTION Since the 1987 financial crash, financial option markets have exhibited a notable implied volatility skew, especially for short-term maturities. This skew reflects the market's expectation of significant price movements on very short time scales in the underlying asset, which poses a challenge to traditional continuous models based on standard Brownian motion. To address this issue, the literature has developed several classes of models that capture the skewness in implied volatilities. Three prominent approaches are: * conventional one-factor stochastic volatility models boosted with large mean-reversion and vol-of-vol. This class of models have been justified by several empirical studies that have identified the presence of very fast mean-reversion in the S&P volatility time series <cit.> and by the fact that they are able to correct conventional models to reproduce the behavior of the at-the-money (ATM) skew for short maturities <cit.>; * jump diffusion models, especially the class of affine jump-diffusions for which valuation problems become (semi-)explicit using Fourier inversion techniques, see <cit.>. Such class of models incorporates occasional and large jumps to explain the skew observed implicitly on option markets, see <cit.>, and <cit.> for an empirical analysis of the impact of adding jumps to stochastic volatility diffusion on the implied volatility surface; * rough volatility models, where the volatility process is driven by variants of the Riemann-Liouville fractional Brownian motion W_t^H = 1/Γ( H+1/2 )∫_0^t ( t-s)^H-1/2dW_s, t ≥ 0, with W a standard Brownian motion and H ∈(0,1/2) the Hurst index. Such models are able to reproduce the roughness of the spot variance's trajectories measured empirically <cit.> together with the explosive behavior of the ATM-skew <cit.>. So far, in the mathematical finance community, jump diffusion models and rough volatility models have often been treated as distinct approaches, and, in some cases, they have even been opposed to each other, see for instance <cit.>. However, on the one side, connections between rough volatility models and fast mean-reverting factors have been established in <cit.>. On the other side, jump models have been related to fast regimes stochastic volatility models in <cit.>. In parallel, from the empirical point of view, it can be very challenging for the human eye and for statistical estimators to distinguish between roughness, fast mean-reversions and jump-like behavior, as shown in <cit.>. The above suggests that rough volatility and jump models may not be that different after all. Our main motivation is to establish for the fist time in the literature a connection between rough volatility and jump models through conventional volatility models with fast mean-reverting regimes. We aim to reconcile these two classes of models through the use of the celebrated conventional Heston model <cit.> but with a parametric specification which encodes a trade-off between a fast mean-reversion and a large vol-of-vol. We define the reversionary Heston model as follows: dS_t^ϵ = S_t^ϵ√(V_t^ϵ)( ρ dW_t + √(1-ρ^2) dW_t^⊥), S_0^ϵ=S_0, dV^ϵ_t = ( ϵ^H- 1/2θ - ϵ^-1( V_t^ϵ-V_0 ) ) dt + ϵ^H-1/2ξ√(V_t^ϵ)dW_t, V^ϵ_0=V_0, where ( W, W^⊥) is a two-dimensional Brownian motion, θ≥ 0, S_0, ξ, V_0 >0, ρ∈ [-1,1]. The two crucial parameters here are the reversionary time-scale ϵ>0 and H ∈ℝ. Such parametrizations nest as special cases the fast regimes extensively studied by <cit.>, see also <cit.>, which correspond to the case H=0; and also the regimes studied in <cit.> for the case H = -1/2. Letting the parameter H∈(∞, 1/2] free in (<ref>) introduces more flexibility in practice and leads to better fits with stable calibrated parameters across time as recently shown in <cit.>. In theory, it allows for a better understanding of the impact of the scaling in H on the limiting behavior of the model as ϵ→ 0 as highlighted in this paper. In a nutshell, we show that: * for H>-1/2, the reversionary Heston model can be constructed as a proxy of rough and hyper-rough Heston models where H ∈ (-1/2,1/2] plays the role of the Hurst index, * for H≤ -1/2, as ϵ→ 0, the reversionary Heston model converges towards Lévy jump processes of Normal Inverse Gaussian type with distinct regimes for H=-1/2 and H<-1/2 respectively, * the reversionary Heston model is capable of generating implied volatility surfaces and at-the-money (ATM) skews similar to the ones generated by rough, hyper-rough and jump models, and comes arbitrarily close to the ATM skew scaling as τ^-0.5 for small τ that characterizes the market, contrary to widespread understanding. Our results allow for a reconciliation between rough and jump models as they suggest that jump models and (hyper-)rough volatility models are complementary, and do not overlap. For H>-1/2, the reversionary Heston model can be interpreted as a proxy of rough and hyper-rough volatility models, while for H≤ -1/2, it can be interpreted as a proxy of jump models. Jump models actually start at H=-1/2 (and below), the first value for which hyper-rough volatility models can no-longer be defined. More precisely, our argument is structured as follows. First, in Section <ref>, we show how the reversionary Heston model (<ref>)-(<ref>) can be obtained as a Markovian and semimartingale proxy of rough and hyper-rough Heston models <cit.> with Hurst index H ∈ (-1/2,1/2). This is achieved using the resolvent of the first kind of the shifted fractional kernel. Second, in Section <ref>, we derive the joint conditional characteristic functional of the log-price log S^ϵ and the integrated variance V̅^ϵ := ∫_0^· V^ϵ_s ds in the model (<ref>)–(<ref>) in terms of a solution to a system of time-dependent Riccati ordinary differential equations; see Theorem <ref>. Compared to the literature, we provide a novel and concise proof for the existence and uniqueness of a global solution to such Riccati equations using the variation of constant formulas. Finally, in Section <ref>, we establish the convergence of the log-price and the integrated variance (log S^ϵ,V̅^ϵ) in the reversionary Heston model (<ref>)-(<ref>) towards a Lévy jump process (X,Y), as ϵ goes to 0. More precisely, we show that the limit (X, Y) belongs to the class of Normal Inverse Gaussian - Inverse Gaussian (NIG-IG) processes which we construct from its Lévy exponent and we connect such class to first hitting-time representations in the same spirit of <cit.>. Our main Theorem <ref> provides the convergence of the finite-dimensional distributions of the joint process (log S^ϵ, V̅^ϵ) through the study of the limiting behavior of the Riccati equations and hence the characteristic functional given in Theorem <ref>. Interestingly, the limiting behavior disentangles three different asymptotic regimes based on the values of H. The convergence of the integrated variance process is even strengthened to a functional weak convergence on the Skorokhod space of càdlàg paths on [0,T] endowed with the M_1 topology. We stress that the usual J_1 topology is not useful here, since jump processes cannot be obtained as limits of continuous processes in the J_1 topology. Related Literature. Convergence of the reversionary Heston models towards jump processes: our results clarify and extend the results of <cit.>, derived for the case H=-1/2, that establish and make clear the precise limiting connection between the Heston log-price process and the normal inverse-Gaussian (NIG) process of <cit.>. Connections between the long time behavior of the Heston log-price process and NIG distribution were first exposed in <cit.> and were the main motivations behind the work of <cit.>. Relevance of fast regimes in practice: the pricing of options near maturity is challenging because of the very steep slope of smiles observed on the market and <cit.> showed that stochastic volatility should embed both a fast regime Ornstein-Uhlenbeck factor (see Remark <ref> below) from which approximations of option prices can be derived using a singular perturbation expansion, and a slowly varying factor to be able to match options with long maturities. On the other hand, <cit.> considers a Heston model with a fast mean-reverting volatility and uses large deviation theory techniques to derive an approximation price for out-of-the-money vanilla options when the maturity is small, but large compared to the characteristic time-scale of the stochastic volatility factor. More recently an Ornstein-Uhlenbeck process with the same parametrization as in (<ref>) has been used to construct the Quintic stochastic volatility model <cit.> to achieve remarkable joint fits of SPX and VIX implied volatilities, outperforming its rough and path-dependent counterparts as shown empirically in <cit.>. Notations. For p≥ 1, we denote by L^p_loc the space of measurable functions f:_+→ such that ∫_0^T |f(s)|^p ds<∞, for all T>0. We will denote by √(x) the principal square root of x ∈ℂ, i.e. its argument lies within (-π/2,π/2]. § FROM ROUGH HESTON TO REVERSIONARY HESTON In this section, we show how reversionary Heston models (<ref>)-(<ref>) can be seen as proxies of rough and hyper-rough Heston models whenever H>-1/2. §.§ Rough and hyper-rough Heston Let ( W, W^⊥) be a two-dimensional Brownian motion and set B := ρ W + √(1-ρ^2) W^⊥ with ρ∈[ -1, 1 ]. We take as starting point a stochastic volatility model for an underlying asset P in terms of a time-changed Brownian motion: dP_t = P_t dB_U̅_t, P_0>0, for some non-decreasing continuous process U̅. If U̅_t = ∫_0^t U_s ds, then U would correspond to the spot variance and U̅ plays the role of the integrated variance. The hyper-rough Volterra Heston model introduced in <cit.> and studied further in <cit.> assumes that the dynamics of the integrated variance is of the form U̅_t = G̅_0(t) + ξ∫_0^t K_H(t-s) W_U̅_s ds, for a suitable continuous function G̅_0, and ξ > 0, and K_H is the fractional kernel K_H(t) = t^H-1/2, t >0, for H ∈ (-1/2,1/2]. The lower bound H>-1/2 ensures the L^1_loc integrability of the kernel K_H so that the stochastic convolution appearing in (<ref>) is well-defined. Any kernel K only in L^1_loc can be considered for the specification of the integrated variance in (<ref>), and if furthermore the kernel happens to be in L^2_loc, the following lemma ensures the existence of a spot variance process. Let K ∈ L^2_loc and g_0 ∈ L^1_loc. Assume there exists a non-decreasing adapted process U̅ and a Brownian motion W such that U̅_t = ∫_0^t g_0(s)ds + ∫_0^t K(t-s) W_U̅_s ds, with sup_t≤ T𝔼[ | U̅_t | ] < ∞, for all T>0. Then, U̅_t = ∫_0^t U_s ds, where U is a non-negative weak solution to the following stochastic Volterra equation U_t = g_0(t) + ∫_0^t K(t-s) √(U_s)dW_s, ℚ⊗ dt-a.e. Conversely, assume there exists a non-negative weak solution U to the stochastic Volterra equation (<ref>) such that sup_t≤ T𝔼[ U^2_t ] < ∞, for all T>0, then U̅ solves (<ref>). This is obtained by an application of stochastic Fubini's theorem, see <cit.>. Going back to the fractional case, if we restrict H in ( 0, 1/2 ], then we have K_H ∈ L^2_loc. For G̅_0(t) := ∫_0^t g_0(s) ds, a direct application of Lemma <ref> yields that the model (<ref>)-(<ref>) is equivalent to the rough Heston model of <cit.> written in spot-variance form dP_t = P_t√(U_t)dB_t, U_t = g_0(t) + ∫_0^t K_H(t-s) ξ√(U_s) dW_s, for some initial input curve g_0: ℝ_+ →ℝ ensuring the non-negativity of V. Two notable specifications of such admissible input curves are given by <cit.> and read g_0 continuous and non-decreasing with g_0 ≥ 0, or g_0(t) = U_0 + θ∫_0^t K_H(s)ds, for some U_0, θ≥ 0. Moreover, for H∈ (0,1/2] the sample paths of the spot variance U are locally Hölder continuous of any order strictly less than H, and consequently rougher than those of the standard Brownian motion, which corresponds to the case H=1/2, justifying the denomination `rough model'. The hyper-rough appellation corresponds to the case H∈ (-1/2,0] for which the process U̅ is continuous but no longer absolutely continuous. Indeed, in this case, one can show that the trajectories of U are nowhere differentiable, see <cit.>. A key advantage of rough and hyper-rough Heston models is the semi-explicit knowledge of the characteristic function of the log-price modulo a deterministic Riccati Volterra convolution equation, as they belong the class of Affine Volterra processes <cit.>. More precisely, for any u=(u_1,u_2) ∈ℂ^2 satisfying (u_1)=0, (u_2) ≤ 0, the joint Fourier–Laplace transform of (log P, U̅) is given by 𝔼[ exp( u_1log P_T + u_2 U̅_T ) ] = exp( u_1 log P_0 + ∫_0^T R(ψ_H(T-s))dG̅_0(s) ), for all T ≥ 0, where ψ_H is the continuous solution to the following fractional Riccati–Volterra equation ψ_H(t) =∫_0^t K_H(t-s) R(ψ_H(s))ds, t ≥ 0, R(x) =1/2 (u_1^2 -u_1) +u_2 + ρν u_1 x + ν^2/2 x^2, see <cit.>. This allows fast pricing and calibration via Fourier inversion techniques. Compared to the conventional Heston model where the characteristic function is known explicitly, the solution to the Riccati Volterra equation is not explicitly known. §.§ Deriving reversionary Heston as a proxy: ϵ-shifting the singularity In both regimes, rough and hyper-rough, with the exception of H=1/2, the model is non-Markovian, non-semimartingale with singular kernels. From a practitioner standpoint, it is therefore natural to look for Markovian approximations by suitable smoothing of the singularity of the fractional kernel (<ref>) sitting at the origin. In this section, we show how we can build a Markovian semi-martingale proxy of hyper-rough models. This is achieved using a two-step procedure. First step: recover semimartingality by smoothing out the singularity of the fractional kernel K_H. We fix ϵ>0, and we consider the shifted fractional kernel K_H, ϵ(t) := ( t+ ϵ)^H-1/2, t> 0, and the corresponding `integrated variance' U̅^ϵ given by U̅^ϵ_t = ∫_0^t g_0^ϵ(s) ds + ξ∫_0^t K_H, ϵ(t-s) W_U̅^ϵ_s ds, with g_0^ϵ(t) = U_0 + θ∫_0^t K_H,ϵ(s)ds. Note that now K_H, ϵ is in L^2_loc for any value of H, so that an application of Lemma <ref> yields that U̅^ϵ = ∫_0^· U^ϵ_s ds where the spot variance U^ϵ solves the equation dP^ϵ_t = P^ϵ_t √(U^ϵ_t)dB_t U^ϵ_t = U_0 + ∫_0^t K_H,ϵ(t-s) θ ds + ∫_0^t K_H,ϵ(t-s) ξ√(U^ϵ_s) dW_s. Moreover, since K_H,ϵ is continuously differentiable on [0,T], denoting by K'_H,ϵ its derivative, we get that U^ϵ is a semimartingale with the following dynamics dU_t^ϵ = (K_H,ϵ (0)θ + ∫_0^t K'_H,ϵ(t-s) dZ_s^ϵ)dt + K_H,ϵ (0) ξ√(U_t^ϵ) dW_t, with dZ^ϵ_t = θ dt + ξ√(U_t^ϵ) dW_t. Second step: recover a Markovian proxy thanks to the resolvent of the first kind. The only non-Markovian term in (<ref>) is the term ∫_0^t K'_H,ϵ(t-s) dZ_s^ϵ appearing in the drift. Using the resolvent of the first kind of K_H,ϵ we will re-express this term in terms of a functional of the past of the process U^ϵ. For a kernel K, a resolvent of the first kind is a measure L on _+ of locally bounded variation such that ∫_[0,t] K(t-s)L(ds) = 1, t≥ 0, see <cit.>. A resolvent of the first kind does not always exist. We will make use of the notations (f*g)(t)=∫_0^tf(t-s)g(s)ds and (f*L)(t)=∫_[0,t]f(t-s)L(ds). Fix ϵ>0 and H ∈ (-1/2,1/2). The kernel K_H, ϵ admits a resolvent of the first kind L_ϵ of the form L_ϵ (dt)= δ_0 (dt)/K_H,ϵ(0) + ℓ_ϵ(t)dt, with ℓ_ϵ a locally integrable function. Moreover, the function ( K_H, ϵ' * L_ϵ) is continuously differentiable and it holds, for all t≥ 0, that ∫_0^t K'_H,ϵ(t-s) dZ_s^ϵ = - ( 1/2-H ) ϵ^-1( U_t^ϵ-U_0) + ∫_0^t ( K_H, ϵ' * L_ϵ)'(t-s) ( U_s^ϵ - U_0) ds. First, the existence of the resolvent is justified as follows. Given H ∈ (-1/2,1/2), K_H, ϵ is a positive completely monotone kernel[Recall that a function f is completely monotone if it is infinitely differentiable on (0,∞) such that (-1)^n f^(n)≥ 0, for all n ≥ 0.] on [0,T] so that an application of <cit.> yields the existence of a resolvent of the first kind in the form (<ref>) with ℓ_ϵ a completely monotone function. Convolving (<ref>) with K'_H,ϵ one obtains that (K'_H,ϵ * L_ϵ)(t) = K'_H,ϵ(0)/K_H,ϵ(0) + (K'_H,ϵ * ℓ_ϵ)(t). Since K_H,ϵ is twice continuously differentiable on [0,T] and ℓ_ϵ is integrable, it follows that (K'_H,ϵ * ℓ_ϵ) is continuously differentiable and so is (K'_H,ϵ * L_ϵ). Writing ( K_H, ϵ' * L_ϵ) = ( K_H, ϵ' * L_ϵ)(0)+ 1 * ( K_H, ϵ' * L_ϵ)', convolving on the left hand side by by K_H, ϵ combined with the associativity of the convolution operation and the fact that (L_ϵ*K_H, ϵ)=1 yields: K_H, ϵ' * 1 = K_H, ϵ' *( L_ϵ*K_H, ϵ) = (K_H, ϵ' * L_ϵ)*K_H, ϵ = ( K_H, ϵ' * L_ϵ)(0) (1*K_H, ϵ) + 1 * ( K_H, ϵ' * L_ϵ)'*K_H, ϵ. And thus, we obtain almost everywhere with regards to the Lebesgue measure that: K_H, ϵ' = ( K_H, ϵ' * L_ϵ)(0) K_H, ϵ + (K_H, ϵ' * L_ϵ)' * K_H, ϵ. In addition, using (<ref>), we notice that ( K_H, ϵ' * L_ϵ)(0)= K_H, ϵ'(0)/K_H, ϵ(0) = ( H-1/2) ϵ^-1. Combining the above, we obtain that ∫_0^t K'_H,ϵ(t-s) dZ_s^ϵ = ( H-1/2) ϵ^-1∫_0^t K_H,ϵ(t-s) dZ_s^ϵ + ((K_H, ϵ' * L_ϵ)' * (K_H, ϵ*dZ_t^ϵ))_t, which yields (<ref>), after recalling that U^ϵ - U_0 = ∫_0^· K_H,ϵ(·-s) dZ_s^ϵ. With the help of the resolvent of the first kind, we were able to recover in the first term of (<ref>), the first order mean-reversion scale of the fractional kernel, the second term depends on the whole past trajectory of U^ϵ. We can now derive our Markovian proxy of the hyper-rough Heston model as follows: plugging the expression (<ref>) in the drift of (<ref>), recalling that K_H,ϵ(0)=ϵ^H-1/2 and dropping the non-Markovian term ( ( K_H, ϵ' * L_ϵ)' )_. * ( U^ϵ_· - U_0^ϵ) )_t, we arrive to the Markovian process: d Ṽ_t^ϵ = ( ϵ^H-1/2θ - ( 1/2-H ) ϵ^-1( Ṽ_t^ϵ - U_0) ) ) dt + ϵ^H-1/2ξ√(Ṽ_t^ϵ) dW_t, Ṽ_0 = U_0. Finally, re-scaling the mean-reversion speed from ( 1/2-H ) ϵ^-1 to ϵ^-1 leads to our reversionary Heston model (<ref>)–(<ref>) where the parameter H becomes unconstrained. In the following section, we illustrate numerically the fact that such reversionary Heston model can be seen as a proxy of rough and hyper-rough Heston models. Such proxy approximation can directly be applied to the Riemann-Liouville fractional Brownian motion defined in (<ref>) to get the proxy: W_t^H,ϵ = ϵ^H-1/2∫_0^t e^- ( 1/2-H ) ϵ^-1 (t-s) dW_s, which is a mean-reverting Ornstein-Uhlenbeck process as long as H<1/2, while the value H=1/2 yields back the standard Brownian motion. First, such Ornstein-Uhlenbeck process has been recently used to construct the Quintic stochastic volatility model <cit.> to achieve remarkable joint fits of SPX and VIX implied volatilies, outperforming even its rough and path-dependent counterparts as shown empirically in <cit.>. Furthermore, notice that the case H=0 degenerates into the fast scale volatility factor from <cit.>, with m=0, ν=1 and their time-scale is twice the reversionary time-scale ϵ, and whose auto-correlation under the invariant distribution is given by 𝔼[ W_t^0,ϵ W_s^0,ϵ] = e^-|t-s|/2ϵ. Consequently, the reversionary time-scale ϵ sets the speed of decay of the auto-correlation function of W^H,ϵ. §.§ Numerical illustration We now illustrate numerically that the reversionary Heston model (<ref>)–(<ref>) is able to reproduce shapes of implied volatility surfaces and at-the-money skew that are similar to the ones generated by rough and hyper-rough Heston models. For this, we first generate implied volatility surfaces of the hyper-rough and rough Heston model via the Fourier-Cosine series expansion technique from <cit.>, where we used the fractional Adams scheme described in <cit.> on the fractional Riccati equation (<ref>) to compute the characteristic function of the (hyper-)rough Heston models. Three target smiles are generated with a (hyper-)rough Heston having parameters ρ=-0.7, θ=0.02, ξ=0.3, U_0=0.02, for H ∈{ 0.1, 0, -0.05 }. For each of these smiles, we calibrate the parameters ( ϵ̂, Ĥ) of the reversionary Heston model (<ref>)-(<ref>), while fixing the other parameters equal to those of the hyper-rough Heston's, by minimizing a weighted loss ∑_i,j w_i,j( C_rough Heston(T_i,K_j)-C^ϵ̂, Ĥ_reversionary Heston(T_i,K_j) )^2. The reversionary Heston prices are also obtained by Fourier-Cosine expansion of the characteristic function. In contrast to the rough Heston models, the characteristic function is known explicitly, see Corollary <ref> below. After calibration, we obtain the following parameters The resulting At-The-Money (ATM) skews between 1 week and 1 year are shown on Figure <ref>. The implied volatility surfaces for the case H=0.1 is illustrated on Figure <ref>. The fit of the smiles for H=0 and H=-0.05 are deferred to Appendix <ref>, see Figures <ref> and <ref>. The graphs show that the reversionary Heston model seems to be able to generate similar shapes of the implied volatility surfaces of rough and hyper-rough models and very steep skews even in the hyper rough regimes H≤ 0. § THE JOINT CHARACTERISTIC FUNCTIONAL OF REVERSIONARY HESTON The following theorem provides the joint conditional characteristic functional of the log-price log S^ϵ and the integrated variance V̅^ϵ := ∫_0^· V^ϵ_s ds in the model (<ref>)–(<ref>) in terms of a solution to a system of time-dependent Riccati ordinary differential equations. Let f,g: [0,T]→ℂ be measurable and bounded functions such that g + 1/2(( f)^2 - f) ≤ 0. Then, the joint conditional characteristic functional of (log S^ϵ, V̅^ϵ) is given by 𝔼[ . exp(∫_t^T f(T-s) dlog S^ϵ_s + ∫_t^T g(T-s) d V̅_s^ϵ) | ℱ_t ] = exp(ϕ_ϵ( T-t ) + ϵ^1/2-Hψ_ϵ( T-t ) V^ϵ_t ), t ≤ T, where (ϕ_ϵ, ψ_ϵ) is the solution to the following system of time-dependent Riccati equations ϕ_ϵ'(t) = ( θ + ϵ^-H-1/2 V_0 ) ψ_ϵ(t), ϕ_ϵ(0) = 0, ψ_ϵ'(t) = ϵ^H-1/2ξ^2/2ψ_ϵ ^2(t) + ( ρξϵ^H - 1/2 f(t) - ϵ^-1) ψ_ϵ(t) + ϵ^H - 1/2( g(t) + f^2(t)-f(t)/2), ψ_ϵ(0) = 0. The proof is given in Section <ref> below. Before proving the result, we note that in the case f and g are constant, one recovers the usual formula for the characteristic function of the <cit.> model, where the solution (ϕ_ϵ, ψ_ε) of (<ref>)-(<ref>) is explicit as stated in the following corollary. Let u,v ∈ℝ and set f(t) =i u and g(t)=i v, for all t∈ [0,T]. Then, the solution (ϕ_ϵ,ψ_ϵ) to the Riccati equations (<ref>)-(<ref>) is explicitly given by ϕ_ϵ(t) = ( ϵ^-1/2-Hθ + ϵ^-1-2H V_0) ξ^-2( ( 1 - i ρϵ^H+1/2ξ u - d ) t - 2 ϵln( 1 - ge^- ϵ^-1 td/1-g) ), ψ_ϵ(t) = ϵ^-H-1/2ξ^-2( 1 - i ρϵ^H+1/2ξ u - d ) 1 - e^- ϵ^-1 td /1 - ge^- ϵ^-1 td , with g := 1 - i ρϵ^H+1/2ξ u - d/1 - i ρϵ^H+1/2ξ u + d, d := √(( 1 - i ρϵ^H+1/2ξ u )^2 - 2 ( ϵ^H+1/2ξ)^2 ( i v - u^2 + iu/2)), d > 0. Consequently, the conditional joint characteristic function of (log S_T^ϵ, ∫_t^T V^ϵ_s ds ) is given by 𝔼[ . exp( iu logS^ϵ_T + iv∫_t^T V_s^ϵ ds) | ℱ_t ] = exp(iulog S^ϵ_t +ϕ_ϵ( T-t ) + ϵ^1/2-Hψ_ϵ( T-t ) V^ϵ_t ), t ≤ T. For the explicit derivation of the formulas, see for example <cit.>. An application of Theorem <ref> yields the result. Such formulas for ϕ_ϵ avoid branching issues as described in <cit.>. The rest of the section if dedicated to the proof of Theorem <ref>. We first study the existence of a solution to time-dependent Riccati ODEs for which equation (<ref>) is a particular case, and provide some of their properties in Section <ref>. We complete the proof of Theorem <ref> in Section <ref>. §.§ Time-dependent Riccati ODEs: existence and uniqueness In this section, we consider a generic class of time-dependent Ricatti equations that encompass equation (<ref>), in the form ψ'(t) = a(t) ψ^2(t) + b(t) ψ(t) + c(t), ψ(0) = u_0, t≤ T, with u_0 ∈ℂ and a,b,c:[0,T]→ℂ three measurable and bounded functions. We say that ψ:[0,t^*] →ℂ for some t^* ∈( 0,T ] is a local extended solution to (<ref>) with some initial condition ψ(0)=u_0 ∈ℂ if, almost everywhere on [0,t^*], it is continuously differentiable and satisfies the relations in (<ref>). The extended solution is global if t^*=T. The presence of the squared non-linearity in (<ref>) precludes the application of the celebrated Cauchy-Lipschitz theorem and can lead to explosive solutions in finite time. Compared to the related literature on similar Riccati equations <cit.>, we provide a concise and simplified proof for the existence and uniqueness of a global extended solution to the Riccati equation (<ref>) using a variation of constant formula under the following assumption on the coefficients (a,b,c) and the initial condition u_0: (a(t)) = 0, a(t) >0, ( c(t) ) + ( b(t) )^2/4a(t)≤ 0, (u_0)≤ 0, t≤ T. The following theorem gives the existence and uniqueness of a solution to the Riccati equation (<ref>). Let u_0∈ℂ with (u_0)≤ 0 and a,b,c:[0,T]→ℂ be measurable and bounded functions satisfying (<ref>). Then, there exists a unique extended solution ψ:[0,T]→ℂ to the Ricatti equation (<ref>) such that (ψ(t))≤ 0, t≤ T, and sup_t≤ T |ψ(t)|<∞. For the existence part, we proceed in two steps. First, we start by arguing the existence of a local solution using Carathéodory's theorem. For this we rely on <cit.>, using the notations therein (see equation (1.7) for example), we consider the integral equation ψ(t) = ψ(0) + ∫_0^t g (t,s,ψ(s))ds, t ≥ 0, where the operator g is defined by g(t,s,ψ(s)):= a(s) x(s)^2 + b(s) x(s) + c(s). Let D be an open, connected subset of ℝ^+ ×ℂ that contains ( 0, ψ(0) ). Define T_∞ := sup{ t ∈ℝ^+ | C_ψ(0),D( [0,T), ℂ) ≠∅}, where C_ψ(0),D( [0,T), ℂ) := {ϕ∈ C( D, ℂ) | ϕ(0) = ψ(0) and ( t, ϕ(t)) ∈ D for t ∈ [0,T)}. An application of <cit.> yields the existence of a unique non-continuable solution to (<ref>) that satisfies ( t, ψ(t) ) ∈ D on the interval [0,T_∞) such that either T_∞ = T or t → T_∞lim |ψ(t)| = ∞. Indeed, the assumptions (i) to (v) of <cit.> are readily satisfied by boundedness and integrability of a, b and c and the fact that g does not depend on t and satisfies the Carathéodory conditions. For the existence part, assume without loss of generality that u_0=0. Define the local domain D: 0 ≤ t ≤ T, | x | ≤ b, with T and b are positive real numbers. The right-hand function f(t,x) := a(t) x^2 + b(t) x + c(t) satisfies the three Carathéodory conditions on D such that * x ↦ f(t,x) is continuous for almost all t ∈[ 0, T ], * t ↦ f(t,x) is measurable for all x, * | f(t,x) | ≤ A | x |^2 + B | x | + C ≤ A b^2 + B b + C =: K(b) < ∞, with K(b) a constant (which is integrable on [ 0, T ]) depending on b and a,b and c are upper-bounded by A, B and C respectively. Consequently, <cit.> yields the existence of a local extended solution ψ to (<ref>) up to a maximal time 0<t_+≤ T, and one can take an arbitrary number t_+ which satisfies the inequalities 0 ≤ t_+ ≤ T, t_+ ≤b/K(b) C'est un mauvais résultat car la borne supérieure en temps dépend de la borne en espace, et T → 0 pour b → 0... The right-hand function f(t,x) := a(t) x^2 + b(t) x + c(t) satisfies the three Carathéodory conditions and <cit.> yields the existence of a local extended solution ψ to (<ref>) up to a maximal time t_+≤ T such that either t_+ = T or t → t_+lim |ψ(t)| = ∞[How to justify this?]. Second, we argue that sup_t≤ T_∞ |ψ(t)|<∞, which would then yield T_∞=T and the existence of a global solution ψ. Let t< T_∞. We start by showing that (ψ(t)) ≤ 0. Indeed, taking real parts in (<ref>), ψ_ r:=( ψ) satisfies the following equation on [0,T_∞): ψ_ r' (s)= { a(s) ( ψ(s) ) + ( b(s) ) }ψ_ r(s)+ d(s), where d(s) = - a(s) ( ( ψ(s) ) + ( b(s) )/2a(s))^2 + ( c(s) ) + ^2( b(s) )/4a(s)≤ 0 thanks to condition (<ref>), after a completion of squares. The variation of constant for ψ_ r then yields ψ_ r(t) = e^∫_0^t ( a(u) ( ψ(u) ) + ( b(u) ) ) du(u_0) + ∫_0^t d(s) e^∫_s^t ( a(u) ( ψ(u) ) + ( b(u) ) ) duds ≤ 0, since the exponential is positive and d(s)≤ 0, and (u_0)≤ 0 by assumption. This shows that (ψ)≤ 0, on [0,T_∞). Finally, an application of a similar variation of constants formula on equation (<ref>) leads to ψ(t) = e^∫_0^t ( a(u) ( ψ(u) ) + ( b(u) ) ) du u_0 + ∫_0^t e^∫_s^t ( a(u) ( ψ(u) ) + ( b(u) ) ) du c(s) ds, so that taking the module together with the triangle inequality and the fact that (ψ)≤ 0 on [0,T_∞), yields |ψ(t)| ≤ |e^∫_0^t ( a(u) ψ(u) + ( b(u) ) ) du | |u_0| + ∫_0^t | e^∫_s^t ( a(u) ψ(u)+ ( b(u) ) ) du| |c(s)| ds = e^∫_0^t ( a(u) ( ψ(u) ) + ( b(u) ) ) du |u_0| + ∫_0^t e^∫_s^t ( a(u) ( ψ(u) ) + ( b(u) ) ) du |c(s)|ds ≤ e^∫_0^t ( b(u) ) du |u_0| + ∫_0^T e^∫_s^t ( b(u) ) du |c(s)|ds ≤ C( |u_0| + ∫_0^T |c(s)|ds), where C=sup_s,s'∈ [0,T]^2 e^∫_s^s'(b(u)) du does not depend on t and is finite by boundedness of b. This shows (<ref>) as needed. Combining the above we obtain the existence of a solution ψ on [0,T] satisfying (<ref>) and (<ref>). To argue uniqueness, assume there are two such extended solutions ψ_1 and ψ_2 that satisfy (<ref>). Then, (ψ_2-ψ_1)'(t) = ( a(t) (ψ_2+ψ_1)(t) + b(t) ) (ψ_2-ψ_1)(t), (ψ_2-ψ_1)(0) = 0, t≤ T, which yields |ψ_2-ψ_1|(t) ≤∫_0^t | a(s) (ψ_2+ψ_1)(s) + b(s) | |ψ_2-ψ_1|(s) ds ≤ c ∫_0^t |ψ_2-ψ_1|(s) ds , t≤ T, for some c>0 by boundedeness of (ψ_1,ψ_2, a, b) using (<ref>), so that the uniqueness is obtained from Gronwall's lemma. §.§ Proof of Theorem <ref> We first argue the existence of a solution to the system of Riccati equations (<ref>)-(<ref>). Let us rewrite the Ricatti ODE from (<ref>) as ψ_ϵ'(t) = a_ϵψ_ϵ ^2(t) + b_ϵ(t) ψ_ϵ(t) +c_ϵ(t), ψ_ϵ(0) = 0, t ≤ T, where we defined a_ϵ := ϵ^H-1/2ξ^2/2 b_ϵ(t) := ϵ^H-1/2ρξ f(t) - ϵ^-1 c_ϵ(t) := ϵ^H-1/2[ g(t) + f^2(t) - f(t)/2]. Since condition (<ref>) ensures ( c_ϵ) + ( b_ϵ)^2/4a_ϵ = ϵ^H-1/2( g + 1/2( ( f)^2 - f ) + ( ρ^2 - 1 ) ( f )^2 ) ≤ 0, then conditions (<ref>) are readily satisfied and consequently Theorem <ref> yields the existence and uniqueness of a solution ψ_ϵ:[0,T]→ℂ to the Ricatti ODE (<ref>) such that (ψ_ϵ(t)) ≤ 0, t ≤ T. The function ϕ_ϵ defined in integral form as ϕ_ϵ(t) = ( θ + ϵ^-H-1/2 V_0 ) ∫_0^t ψ_ϵ(s)ds, t ≤ T, solves (<ref>). We now prove the expression for the charateristic functional (<ref>). Define the following process M: M_t = exp(U_t), U_t = ϕ_ϵ( T-t ) + ϵ^1/2-Hψ_ϵ( T-t ) V^ϵ_t + ∫_0^t f(T-s) dlog S^ϵ_s + ∫_0^t g(T-s) d V̅_s^ϵ. In order to obtain (<ref>), it suffices to show that M is a martingale. Indeed, if this is the case, and after observing that the terminal value of M is given by M_T = exp( ∫_0^T f(T-s) dlog S^ϵ_s + ∫_0^T g(T-s) d V̅_s^ϵ), recall that ϕ_ϵ( 0 )=ψ_ϵ( 0 )=0, we obtain 𝔼[ exp( ∫_0^T f(T-s) dlog S^ϵ_s + ∫_0^T g(T-s) d V̅_s^ϵ) |ℱ_t] = 𝔼[ M_T | ℱ_t] = M_t = exp(U_t), which yields (<ref>). We now argue that M is a martingale. We first show that M is a local martingale using Itô formula. The dynamics of M read dM_t = M_t (dU_t + 1/2 d⟨ U⟩_t ), with dU_t = {ϕ_ϵ'(T-t) - ( θ + ϵ^-1/2-H V_0 ) ψ_ϵ(T-t) + ( - ϵ^1/2-Hψ_ϵ '(T-t) - ϵ^-1/2-Hψ_ϵ(T-t) + g(T-t) - f(T-t)/2) V^ϵ_t } dt + ( ξψ_ϵ(T-t) + ρ f(T-t) ) √(V^ϵ_t)dW_t+√(1-ρ^2)f(T-t)√(V^ϵ_t)dW^⊥_t. This yields that the drift in dM_t/M_t is given by ϕ_ϵ'(T-t) - ( θ + ϵ^-1/2-H V_0 ) ψ_ϵ(T-t) + ( - ϵ^1/2-Hψ_ϵ'(T-t) + ξ^2/2( ψ_ϵ (T-t) )^2 + ( ρξ f(T-t) - ϵ^-1/2-H) ψ_ϵ(T-t) + g(T-t) + f^2(T-t) - f(T-t)/2) V^ϵ_t which is equal to 0 from the Riccati equations (<ref>) and (<ref>). This shows that M is a local martingale. To argue that M is a true martingale, we note that (ψ_ϵ)≤ 0 which implies (ϕ_ϵ)≤ 0, so that (U_t) ≤∫_0^t (f(T-s)) dlog S^ϵ_s + ∫_0^t (g(T-s)) d V̅_s^ϵ = ∫_0^t ( (g(T-s)) -1/2(f(T-s))) V_s^ϵ ds + ∫_0^t (f(T-s)) √(V_s^ϵ)dB_s ≤ -1/2∫_0^t (f(T-s))^2 V_s^ϵ ds + ∫_0^t (f(T-s)) √(V_s^ϵ)dB_s=:Ũ_t, where the last inequality follows from (<ref>). It follows that |M_t|=exp((U_t)) ≤exp(Ũ_t), where the process exp(Ũ) is a true martingale, see <cit.>. This shows that M is a true martingale, being a local martingale bounded by a true martingale, see <cit.>, which concludes the proof. § FROM REVERSIONARY HESTON TO JUMP PROCESSES In this section, we establish the convergence of the log-price and the integrated variance (log S^ϵ,V̅^ϵ) in the reversionary Heston model (<ref>)-(<ref>) towards a Lévy jump process (X,Y), as ϵ goes to 0. More precisely, the limit (X, Y) belongs to the class of Normal Inverse Gaussian - Inverse Gaussian (NIG-IG) processes defined as follows. Fix α≥ |β| ≥ 0, δ,λ > 0 and μ∈. We say that ( X_t, Y_t )_t ≥ 0 is a Normal Inverse Gaussian - Inverse Gaussian (NIG-IG) process with parameters ( α,β,δ,μ, λ) if it is a two-dimensional homogeneous Lévy process with càdlàg sample paths, starting from (X_0,Y_0)=(0,0) almost surely, with Lévy exponent η defined by η(u,v) := [ i μ u + δ( √(α^2 - β^2) -√(α^2 - 2iλ v -( β + iu )^2)) ], u, v ∈ℝ, i.e. the joint characteristic function is given by 𝔼[ exp( iu X_t + iv Y_t ) ] = exp( η(u,v) t ), u, v ∈ℝ, t ≤ T. In order to justify the existence of such a class of Lévy processes, one needs to justify that η given in (<ref>) is indeed the logarithm of a characteristic function associated to an infinitely divisible distribution, see <cit.>. This is the object of the following lemma, which also provides the link with first-hitting times and subordinated processes. Let α≥ |β| ≥ 0, δ,λ > 0, μ∈ and (W, W^⊥) be a two dimensional Brownian motion. Let (Λ_t)_t∈ [0,T] be the first hitting-time process defined as Λ_t := inf{ s ≥ 0 : √(α^2-β^2)s + W_s≥δ t }, t ∈[ 0,T ], and define Z as the following shifted subordinated process Z_t =μ t + βΛ_t + W^⊥_Λ_t, t ∈[ 0,T ]. Then, 𝔼[ exp( iuZ_t+ iv λΛ_t ) ] = exp(η(u,v)t), u,v ∈ℝ, t ∈[ 0,T ]. In particular, η given by (<ref>) is the logarithm of the characteristic function of the joint random variable (Z_1,λΛ_1) which is infinitely divisible. Fix t ∈[ 0,T ]. By construction, it is well-known that Λ_t has an Inverse Gaussian distribution if α > | β| with parameters IG( δ t/√(α^2 - β^2), δ^2 t ), and in the drift-free case α = | β|, Λ_t follows a Lévy distribution with parameters Lévy( 0, δ^2 t ) (see <cit.> and Definition <ref> in the Appendix). Now conditional on Λ_t, Z_t is Gaussian with parameters 𝒩( μ t + βΛ_t, Λ_t ) and using the tower property of conditional expectation, we get for the first case that 𝔼[exp( iu Z_t + iv λΛ_t ) ] = 𝔼[ 𝔼[ . exp( iu Z_t ) | Λ_t ] exp( iv λΛ_t ) ] = 𝔼[ exp( iu ( μ t + βΛ_t ) - Λ_t u^2/2 + iv λΛ_t ) ] = exp( iu μ t ) 𝔼[ exp( ( iu β - u^2/2 + iv λ) Λ_t ) ] = exp( iu μ t + δ t √(α^2 - β^2)( 1 - √(1-2/α^2 - β^2( iu β - u^2/2 + iv λ))) ) i.e. 𝔼[ exp( iuZ_t+ivλΛ_t ) ] = exp( i u μ t + δ t ( √(α^2 - β^2) -√(α^2 - 2iλ v -( β + iu )^2))) = exp(η(u,v)t), u,v ∈ℝ, where we used Definition <ref> to get the fourth equality, noting that ( i (u β + v λ) - u^2/2) ≤ 0. Similar computations yield the result for the case α = | β|. Furthermore, we will say that that the random variable (Z_1,λΛ_1) follows a NIG-IG distribution with parameters (α,β,μ, δ,λ) (see Definition <ref> in the Appendix). Such distribution is infinitely divisible because if ( X_1, Y_1 ), ⋯, ( X_m, Y_m ) are independent NIG-IG random variables with common parameters ( α, β, λ) and individual ( μ_i, δ_i ), for i =1, ⋯, m, then ( X, Y ) := ( ∑_i=1^mX_i, ∑_i=1^mY_i ) is again NIG-IG-distributed with parameters ( α, β, ∑_i=1^mμ_i, ∑_i=1^mδ_i, λ). The appellation NIG-IG for the couple (X,Y) in Definition <ref> is justified as follows: * Y is an Inverse Gaussian process first derived by <cit.> which can be checked either by recovering the Inverse Gaussian distribution with parameters IG( λδ/√(α^2 - β^2), λδ^2 ) after setting u=0 in (<ref>); or by using the representation as a first passage-time in (<ref>). It is worth pointing that, for α = |β|, one recovers the well-known Lévy distribution for the first-passage of a Brownian motion with parameters Lévy( 0, λδ^2 ). The Lévy distribution can be seen as a special case of the Inverse Gaussian distribution. * X is the celebrated Normal Inverse Gaussian process of <cit.>, with parameters NIG( α, β, μ, δ), which can be checked by setting v=0 in (<ref>) or by using the representation as subordinated Brownian motion with an Inverse Gaussian subordinator as in (<ref>). In addition, we allow in Definition <ref> the parameter α to be equal to ∞ in the following sense: Considering the set of parameters (α,β,δ,μ, λ) = ( α, 0, σ^2 α, μ, 1 ), a second order Taylor expansion, as α→∞, of the square root yields 𝔼[ exp( iu X_t + iv Y_t ) ] = exp( [ i μ u - σ^2 (u^2/2 + iv)] t ), u, v ∈ℝ, t ≤ T, which is equivalent to the normal-deterministic process defined by ( X_t, Y_t )_t ∈ [0,T] = ( μ t + σW_t, σ^2 t )_t ∈ [0,T]d=( μ t + W_σ^2 t, σ^2 t )_t ∈ [0,T]. We will consider that such (degenerate) process is a particular case of Definition <ref> with parameters denoted by . ( α, 0, σ^2 α, μ, 1 )|_α→∞. We are now in place to state our main convergence theorem. Theorem <ref> provides the convergence of the finite-dimensional distributions of the joint process (log S^ϵ, V̅^ϵ) through the study of the limiting behavior of the characteristic functional given in Theorem <ref>. Interestingly, the limiting behavior disentangles three different asymptotic regimes based on the values of H that can be seen intuitively on the level of the Riccati equation (<ref>) as follows. Applying the variation of constants on ψ_ϵ, we get: ψ_ϵ(t) = ϵ^H+1/2∫_0^t K_ϵ(t-s) F(s,ψ_ϵ(s)) ds, t≤ T, F(s,u) := ξ^2/2 u^2 + ρξ f(s) u + g(s) + f^2(s) - f(s)/2, with K_ϵ the kernel defined by K_ϵ(t) = ϵ^-1 e^-ϵ^-1t, t≥ 0. Assuming that ψ_ϵ converges to some ψ_0 and observing that K_ϵ plays the role of the Dirac delta as ϵ→ 0, one expects ∫_0^t K_ϵ(t-s) F(s,ψ_ϵ(s)) ds → F(t,ψ_0(t)) in (<ref>), the pre-factor ϵ^H+1/2 suggests then three different limiting regimes with respect to H that can be characterized through the functions F and ψ_0: ψ_0(t) = 0, H>-1/2, ψ_0(t)= F(t,ψ_0(t)), (ψ_0(t))≤ 0, H=-1/2, 0 = F(t,ψ_0(t)), (ψ_0(t))≤ 0, H<-1/2. The function ψ_0 in (<ref>) is even explicitly given by ψ_0(t)= 0, H>-1/2, ξ^-2( 1 - ρξ f(t) - √(( 1 - ρξ f(t) )^2 - 2 ξ^2 ( g(t) + f^2(t)-f(t)/2))), H=-1/2, - ξ^-1( ρ f(t) + √(f(t) ( 1 - ( 1- ρ^2 ) f(t) ) - 2 g(t))), H<-1/2, see Lemma <ref> below. Furthermore, the convergence of the integrated variance process is strengthened to a functional weak convergence on the Skorokhod space (D(0,T), M_1) of càdlàg paths on [0,T] endowed with the strong M_1 topology, see Section <ref> below. Such topology is weaker and less restrictive than the commonly used uniform or J_1 topologies which share the property that a jump in a limiting process can only be approximated by jumps of comparable size at the same time or, respectively, at nearby times. On the contrary, the M_1 topology of <cit.> captures approximations of unmatched jumps, which in our case, will allow us to prove the convergence of the stochastic process V̅^ϵ with continuous sample trajectories towards a Lévy process with càdlàg sample trajectories. The statement is now made rigorous in the following theorem. Let f,g:[0,T]→ℂ be bounded and measurable such that f = g = 0 and such that ψ_0 defined in (<ref>) has bounded variations. Then, based on the value of H, we obtain different explicit asymptotic formulas for the characteristic functional given in Theorem <ref>: lim_ϵ→ 0𝔼[ exp(∫_0^T f(T-s) dlogS^ϵ_s + ∫_0^T g(T-s) d V̅_s^ϵ) ]=exp( ϕ_0(T)), with ϕ_0(T) := V_0 ∫_0^T h(s) ds, H>-1/2, ( θ + V_0 ) ξ^-2( T - ∫_0^T ( ρξ f(s) + √(( 1 - ρξ f(s))^2 - 2 ξ^2 h(s))) ds ), H=-1/2, -θξ^-1∫_0^T ( ρ f(s) + √(ρ^2 f^2(s) - 2 h(s))) ds, H<-1/2, where h(s): =g(s)+ f^2(s)-f(s)/2. In particular for ρ∈ (-1,1), as ϵ→ 0, the finite-dimensional distributions of the joint process (logS^ϵ/S_0,V̅^ϵ) converge to the finite-dimensional distributions of a NIG-IG process (X,Y) in the sense of Definition <ref> with the following parameters depending on the value of H: (α,β,δ,μ, λ) := .( α, 0, V_0 α, - V_0/2, 1 )|_α→∞, H>-1/2, ( 1/2√((ξ-2ρ)^2 + 4(1-ρ^2))/ξ (1-ρ^2), -1/2ξ - 2 ρ/ξ( 1 - ρ^2 ), √(1-ρ^2) (θ+V_0)ξ^-1, - ρ(θ + V_0) ξ^-1, 1/1-ρ^2) , H=-1/2, ( 1/2(1-ρ^2), - 1/2(1-ρ^2), √(1-ρ^2)θξ^-1, - ρθξ^-1, 1/1-ρ^2), H<-1/2, where θ, S_0, ξ and V_0 are the same from (<ref>)-(<ref>). Furthermore, the process V̅^ϵ converges weakly towards Y on the space (D(0,T), M_1), as ϵ→ 0. The convergence of the characteristic functional in (<ref>) is established in Section <ref> (Lemmas <ref> and <ref>). This implies the convergence of the finite-dimensional distributions of (log S^ϵ,V̅^ϵ) as detailed in Section <ref>. Finally, the weak convergence of V̅^ϵ on ( D(0,T),M_1) is proved in Section <ref>. The interpretation of the convergence results becomes even more interesting when combined with Section <ref>. In Section <ref>, for H>-1/2, the reversionary Heston model (log S^ϵ, V̅^ϵ) is constructed as a proxy of rough and hyper-rough Heston models. Theorem <ref> shows that the limiting regime for H>-1/2 is a (degenerate) Black-Scholes regime, cf.Remark <ref>, whereas, for H≤ -1/2 one obtains the convergence of the reversionary regimes towards (non-degenerate) jump processes with distinct regimes between H=-1/2 and H<-1/2, see Corollary <ref> below. This suggests that jump models and (hyper-)rough volatility models are complementary, and do not overlap. For H>-1/2 the reversionary model can be interpreted as a proxy of rough and hyper-rough volatility models, while for H≤ -1/2 it can be interpreted as a proxy of jump models. Jump models actually start at H=-1/2 (and below), the first value for which hyper-rough volatility models can no-longer be defined. In Figures <ref> and <ref>, we plot respectively the convergence of the smiles and the skew of the reversionary Heston model (log S^ϵ, V̅^ϵ) for the case H=-1/2 towards the Normal Inverse Gaussian model. The volatility surface is obtained by applying Fourier inversion formulas on the corresponding characteristic functions. Similar to Figures <ref> and <ref>, the graphs show that the fast parametrizations introduced in the Heston model are able to reproduce very steep skews for the implied volatility surface. In the case (f,g) = ( iu, iv ), with u, v ∈ℝ, the asymptotic marginals of reversionary Heston expressed in Corollary <ref> below are obtained as a direct consequence of the convergence Theorem <ref>. Based on the value of H ∈ℝ, the pair of normalized log price and integrated variance (logS^ϵ_T/S_0, V̅^ϵ_T ) has distinct asymptotic marginals as the reversionary time-scale ϵ goes to zero given by: * H>-1/2, i.e. Black Scholes-type asymptotic regime (BS regime) 𝔼[ iulog S^ϵ_T/S_0 + ivV̅^ϵ_T ] ϵ→ 0⟶exp{ - V_0/2( u^2 - 2i ( v - u/2) ) T }. * H=-1/2, i.e. Normal Inverse Gaussian-type asymptotic regime (NIG regime) 𝔼[ iulog S^ϵ_T/S_0+ ivV̅^ϵ_T ] ϵ→ 0⟶exp{( θ + V_0 ) ξ^-2( 1 - i ρξ u -√(( 1 - i ρξ u )^2 - 2 ξ^2 ( i v - u^2 + iu/2))) T }. * H<-1/2, i.e. Normal Lévy-type asymptotic regime (NL regime) 𝔼[ iulog S^ϵ_T/S_0 + ivV̅^ϵ_T ] ϵ→ 0⟶exp{ - θξ^-1(i ρ u + √(( 1 - ρ^2 ) u^2 - 2i ( v - u/2))) T }. In Figure <ref>, we illustrate numerically the convergence of the characteristic function in all three regimes. The rest of the section is dedicated to the proof of Theorem <ref>. §.§ Convergence of the joint characteristic functional In this section we prove the convergence of the characteristic functional of (logS^ϵ/S_0, V̅^ϵ) as ϵ goes to 0 stated in Theorem <ref>. For this, we fix f,g:[0,T]→ℂ bounded and measurable such that f = g = 0. We note that (<ref>) is trivially satisfied so that an application of Theorem <ref>, with t=0, yields that 𝔼[ exp(∫_0^T f(T-s) dlog S^ϵ_s + ∫_0^T g(T-s) d V̅_s^ϵ) ] = exp( ϕ_ϵ(T) + ϵ^1/2 - H ψ_ϵ(T) V_0 ), where (ϕ_ϵ,ψ_ϵ) solve (<ref>)-(<ref>). We start by showing that the second term in the exponential ϵ^1/2-Hψ_ϵ goes to 0 for any value of H∈ in the following lemma. For ϵ>0, let ψ_ϵ be a solution the time-dependent Ricatti ODE (<ref>) such that (ψ^ϵ)≤ 0 with f, g:[0,T]→ℂ bounded and measurable functions such that f = g = 0. Then, | ψ_ϵ(t) | ≤ C ϵ^H+1/2( 1 - e^- ϵ^-1t), t≤ T, for some constant C independent of ϵ. In particular, we have the uniform convergence lim_ϵ→ 0sup_t≤ Tϵ^1/2-H|ψ_ϵ(t)| = 0, H ∈. The variation of constants applied to the differential equation (<ref>) yields: ψ_ϵ(t) = ∫_ 0^t c_ϵ(s)e^∫_s^t ( a_ϵψ_ϵ(u) + b_ϵ(u) ) du ds, with a_ϵ, b_ϵ, c_ϵ defined as in (<ref>). Given that f and g are both bounded on [0,T], we fix C ≥ 0 such that | c_ϵ(u) | ≤ C ϵ^H-1/2, u ≤ T. Note that b_ϵ = - ϵ^-1, having also ( ψ_ϵ) ≤ 0 and a_ϵ > 0, we get the desired upper bound on the solution. The convergence result follows immediately. The first term ϕ_ϵ, however, yields different limits based on the value of H. Consequently, we will study in Lemma <ref> the convergence of the following quantity ϕ_ϵ (t) = ∫_0^t ( θ + ϵ^-H-1/2 V_0 ) ψ_ϵ(s) ds, for different regimes of H. We have the convergence lim_ϵ→ 0ϕ_ϵ(T) = ϕ_0(T), where ϕ_0(T) is given by (<ref>). Case H>-1/2. In this case, the solution ψ_ϵ converges uniformly to zero from the upper bound given in Lemma <ref> which, combined with the expression of F in (<ref>), yields lim_ϵ→ 0 F(s,ψ_ϵ(s)) = F(s,0) = g(s) + f^2(s) - f(s)/2, s≤ T. Furthermore, integrating the variation of constants expression (<ref>) leads to ∫_0^t ϵ^-H-1/2ψ_ϵ(u) du = ∫_0^t {∫_0^u K_ϵ(u-s) F(s,ψ_ϵ(s)) ds } du = ∫_0^t {∫_s^t K_ϵ(u-s) du } F(s,ψ_ϵ(s)) ds = ∫_0^t ( 1 - e^-ϵ^-1(t-s)) F(s,ψ_ϵ(s)) ds, where we used Fubini for the second equality as the integrated quantity is bounded and measurable. Now, given the function s ↦( 1 - e^-ϵ^-1(t-s))F(s,ψ_ϵ(s)) is uniformly bounded in ϵ by a constant on [0,T], and that it converges pointwise to s ↦ F(0,s), then an application of Lebesgue's Dominated Convergence Theorem yields ∫_0^t ϵ^-H-1/2ψ_ϵ(u) du ϵ→ 0⟶∫_0^t ( g(s) + f^2(s) - f(s)/2)ds, hence the resulting convergence ϕ_ϵ (t) ϵ→ 0⟶ V_0 ∫_0^t ( g(s) + f^2(s) - f(s)/2) ds. Case H=-1/2. Now fix ϵ>0 and using the second equation in (<ref>), observe that ψ_ϵ'(t) = - ϵ^-1ψ_ϵ(t) + ϵ^-1( F(t,ψ_ϵ(t)) - F(t,ψ_0(t)) + ψ_0(t) ) = - ϵ^-1( ψ_ϵ(t) - ψ_0(t) ) + ϵ^-1β_ϵ(t) ( ψ_ϵ(t) - ψ_0(t) ), with β_ϵ(t) := ξ^2/2( ψ_ϵ(t) + ψ_0(t) ) + ρξ f(T-t). Since ψ_0 has bounded variations by assumption, the complex-valued Riemann-Stieltjes integral on continuous functions f against ψ_0 is well-defined, and satisfies an integration by part formula such that ∫_0^t f(s) d ψ_0(s) = f(t)ψ_0(t)-f(0)ψ_0(0) - ∫_0^t ψ_0(s) df(s), see Theorems A.1 and A.2 from <cit.>. Define Δ_ϵ(t):=(ψ_ϵ(t)-ψ_0(t))e^t/ϵ. Then, it follows that dΔ_ϵ(t) = (ψ'_ϵ(t)dt-dψ_0(t))e^t/ϵ + ϵ^-1(ψ_ϵ(t)-ψ_0(t))e^t/ϵdt = ϵ^-1β_ϵ(t) Δ_ϵ(t) - e^t/ϵdψ_0(t), and applying the variation of constants formula leads to Δ_ϵ(t) = e^ϵ^-1∫_0^tβ_ϵ(r)dr Δ_ϵ(0) - ∫_0^t e^ϵ^-1∫_s^tβ_ϵ(r)dr e^ϵ^-1s dψ_0(s), so that, recalling that ψ_ϵ(0)=0, ψ_ϵ(t) - ψ_0(t) = -e^ϵ^-1∫_0^tβ_ϵ(r)dr e^-ϵ^-1tψ_0(0) - ∫_0^t e^ϵ^-1∫_s^tβ_ϵ(r)dre^-ϵ^-1(t-s) dψ_0(s) =: I_ϵ(t) + II_ϵ(t) . We now prove successively that |∫_0^T I_ϵ(t)dt| → 0 and |∫_0^T II_ϵ(t)dt| → 0. * Given that β_ϵ≤ 0, then | e^ϵ^H-1/2∫_0^tβ_ϵ(r)dr e^-ϵ^-1tψ_0(0) | ≤ e^-ϵ^-1t| ψ_0(0) | → 0, t ∈ (0,T), so that I_ϵ(t) converges pointwise to 0 on (0,T) and is dominated by an integrable function, hence |∫_0^T I_ϵ(t)dt| → 0 by Lebesgue's dominated convergence theorem. * Regarding the second term, we have | ∫_0^T ∫_0^t e^ϵ^H-1/2∫_s^tβ_ϵ(r)dre^-ϵ^-1(t-s) dψ_0(s) dt| ≤∫_0^T ∫_0^t e^-ϵ^-1(t-s)| dψ_0(s) | dt = ∫_0^T ∫_s^T e^-ϵ^-1(t-s) dt | dψ_0(s) | = ∫_0^T ϵ( 1 - e^-ϵ^-1( T-s)) | dψ_0(s) | where the inequality comes from Theorem A.4 from <cit.> and using again that β_ϵ≤ 0, with the positive measure on the right-hand side being the total variation measure defined as in Theorem 6.2 from <cit.>, and we used Fubini-Lebesgue to get the first equality. Noting the point-wise convergence of the function f_ϵ : s ↦ϵ( 1 - e^-ϵ^-1( T-s)) to zero and its uniform boundedness in ϵ by a constant on [0,T], the dominated convergence theorem applied to the total variation measure proves the result. Thus, we obtained ∫_0^T ψ_ϵ(t) dt ϵ→ 0⟶∫_0^T ψ_0(t) dt, where ψ_0 satisfies the second equation in (<ref>), hence we get ϕ_ϵ (T) ϵ→ 0⟶( θ + V_0 ) ∫_0^T ψ_0(s) ds, which is the desired convergence. Case H<-1/2. Define ψ_0 as the root with non-positive real part from the third equation in (<ref>), recall Lemma <ref>. Fixing again ϵ>0, we have ψ_ϵ'(t) = - ϵ^-1ψ_ϵ(t) + ϵ^H-1/2( F(t,ψ_ϵ(t)) - F(t,ψ_0(t)) ) = - ϵ^-1ψ_ϵ(t) + ϵ^H-1/2β_ϵ(t) ( ψ_ϵ(t) - ψ_0(t) ), with β_ϵ given by (<ref>). Similarly the case H<-1/2, computing the differential of Δ_ϵ(t):=(ψ_ϵ(t)-ψ_0(t))e^t/ϵ and applying the variation of constants formula leads to Δ_ϵ(t) = e^ϵ^H-1/2∫_0^tβ_ϵ(r)dr Δ_ϵ(0) - ∫_0^t e^ϵ^H-1/2∫_s^tβ_ϵ(r)dr e^s/ϵ(dψ_0(s) + ϵ^-1ψ_0(s) ds ), so that ψ_ϵ(t) - ψ_0(t) = -e^ϵ^H-1/2∫_0^tβ_ϵ(r)dr e^-ϵ^-1tψ_0(0) - ∫_0^t e^ϵ^H-1/2∫_s^tβ_ϵ(r)dre^-ϵ^-1(t-s)(dψ_0(s) + ϵ^-1ψ_0(s) ds ) = -e^ϵ^H-1/2∫_0^tβ_ϵ(r)dr e^-ϵ^-1tψ_0(0) - ∫_0^t e^ϵ^H-1/2∫_s^tβ_ϵ(r)dre^-ϵ^-1(t-s) dψ_0(s) - ∫_0^t e^ϵ^H-1/2∫_s^tβ_ϵ(r)dre^-ϵ^-1(t-s)ϵ^-1ψ_0(s) ds =: I_ϵ(t) + II_ϵ(t) + III_ϵ (t). We already have from the previous case H=-1/2 that, as ϵ→ 0, both integrals ∫_0^T I_ϵ(t) dt, ∫_0^T II_ϵ(t) dt converge to 0, all that remains to show consequently is that ∫_0^T III_ϵ(t) dt converges to 0 too. * A finer upper bound on β_ϵ is required to deal with this third term. We already know from Theorem <ref> that ψ_ϵ≤ 0, and by definition of ψ_0, we get the following bound β_ϵ = ξ^2/2( ψ_ϵ + ψ_0) ≤ξ^2/2ψ_0≤ 0. Set E := { s ∈ [0,T], ( f(s), g(s) ) ≠( 0, 0 ) }, and from (<ref>), we know that ψ_0=0 on E^C while lemma <ref> yields ψ_0 ≠ 0 on E so that we can bound III_ϵ as follows | III_ϵ (t) | ≤∫_[0,t] ∩ Eϵ^-1 |ψ_0(s)| e^ϵ^H-1/2∫_s^tβ_ϵ(r)dr e^-ϵ^-1(t-s) ds ≤∫_[0,t] ∩ Eϵ^-1|ψ_0(s)|/-ϵ^H-1/2ξ^2/2ψ_0(s)( -ϵ^H-1/2ξ^2/2ψ_0(s) e^∫_s^t ϵ^H-1/2ξ^2/2ψ_0(r) dr) ds ≤∫_[0,t] ∩ Eϵ^-H-1/2 2 ξ^-2|ψ_0|/- ψ_0( -ϵ^H-1/2ξ^2/2ψ_0(s) e^∫_s^t ϵ^H-1/2ξ^2/2ψ_0(r) dr) ds, and an application of lemma <ref> yields the existence of a finite positive constant C>0 such that |ψ_0|/- ψ_0≤ C, ∀ s ∈ E, so that | III_ϵ (t)| ≤ C ϵ^-H-1/2∫_[0,t] ∩ E( -ϵ^H-1/2ξ^2/2ψ_0(s) e^∫_s^t ϵ^H-1/2ξ^2/2ψ_0(r) dr) ds ≤ C ϵ^-H-1/2( 1 - e^∫_0^t ϵ^H-1/2ξ^2/2ψ_0(r) dr) ϵ→ 0⟶ 0, since -H-1/2>0 so that ϵ^-H-1/2→ 0 as ϵ→ 0. Thus III_ϵ is dominated by a finite constant C independent of ϵ (which is integrable on [0,T]) and Lebesgue's dominated convergence theorem yields that ∫_0^T III_ϵ(t) dt converges to 0. Consequently, we obtained ∫_0^T ψ_ϵ(t) dt ϵ→ 0⟶∫_0^T ψ_0(t) dt, which then yields ϵ^-H-1/2∫_0^T ψ_ϵ(t) dt ϵ→ 0⟶ 0, and finally we get ϕ_ϵ (T) ϵ→ 0⟶θ∫_0^T ψ_0(s) ds. §.§ Convergence of the finite-dimensional distributions towards NIG-IG In this section, we prove the second part of Theorem <ref>, that is the convergence of the finite-dimensional distributions of ( logS^ϵ/S_0, V̅^ϵ) towards those of a NIG-IG process ( X, Y ) in the sense of Definition <ref> with parameters (α, β, μ, δ,λ) as in (<ref>) depending on the regime of H. Let d ∈ℕ^* and take 0 =: t_0 < t_1 < … < t_d ≤ T to be d distinct times of the time interval [0, T] and ( u_k, v_k )_k ∈{ 1, …, d }∈(ℝ^2)^d. We will prove that 𝔼[ exp( i∑_k=1^d u_k logS^ϵ_t_k/S_0 + i∑_k=1^d v_k V̅_t_k^ϵ) ] ϵ→ 0⟶𝔼[ exp( i∑_k=1^d u_k X_t_k + i∑_k=1^d v_k Y_t_k) ]. First, we recover the finite-dimensional distributions of (logS^ϵ_T/S_0, V̅^ϵ_T ) from (<ref>) by setting the bounded and measurable functions f and g to be equal to f(s) := i∑_k=1^d 1_[t_k-1, t_k)(T-s) ∑_k=j^d u_j g(s) := i∑_k=1^d 1_[t_k-1, t_k)(T-s) ∑_k=j^d v_j. Notice indeed that i ∑_k=1^d u_k logS^ϵ_t_k/S_0 + i∑_k=1^d v_k V̅_t_k^ϵ = i∑_k=1^d (logS^ϵ_t_k/S_0-logS^ϵ_t_k-1/S_0) ∑_j=k^d u_j + i∑_k=1^d (V̅_t_k^ϵ-V̅_t_k-1^ϵ) ∑_j=k^d v_j = ∫_0^T f(T-s) dlogS^ϵ_s + ∫_0^T g(T-s) d V̅_s^ϵ, and that the corresponding ψ_0 defined in (<ref>) has bounded variations (being piece-wise constant for this choice of f and g), so that an application of the convergence of the characteristic functional in (<ref>) yields 𝔼[ exp( i∑_k=1^d u_k logS^ϵ_t_k/S_0 + i∑_k=1^d v_k V̅_t_k^ϵ) ] ⟶exp( ϕ_0(T) ), with ϕ_0(T) = V_0 ∑_k=1^d (t_k - t_k-1) (iv̅_k- u̅_k^2 + iu̅_k/2) , H>-1/2, ( θ + V_0 ) ξ^-2∑_k=1^d(t_k - t_k-1) ( 1 - ρξ iu̅_k + √(( 1 - ρξ iu̅_k)^2 - 2 ξ^2 (iv̅_k- u̅_k^2+iu̅_k/2))), H=-1/2, -θξ^-1∑_k=1^d (t_k - t_k-1) ( ρ iu̅_k + √(-ρ^2 u̅_k^2 - 2 (iv̅_k- u̅_k^2+iu̅_k/2))), H<-1/2, where we defined u̅_k := ∑_j=k^d u_j, v̅_k := ∑_j=k^d v_j. Second, we identify such ϕ_0(T) with the corresponding finite-dimensional distributions of the NIG-IG process ( X, Y ) with parameters (α, β, μ, δ,λ) as in (<ref>) depending on the regime of H. We denote by η its Lévy exponent, recall (<ref>), and we write 𝔼[ exp( i∑_k=1^d u_k X_t_k + i∑_k=1^d v_k Y_t_k) ] = 𝔼[ exp( i∑_k=1^d (X_t_k-X_t_k-1) ∑_j=k^d u_j + i∑_k=1^d (Y_t_k-Y_t_k-1) ∑_j=k^d v_j) ] = ∏_k=1^d 𝔼[ exp( (X_t_k-X_t_k-1) i∑_j=k^d u_j + (Y_t_k-Y_t_k-1) i∑_j=k^d v_j) ] = ∏_k=1^d 𝔼[ exp( X_t_k-t_k-1 i∑_j=k^d u_j + Y_t_k-t_k-1 i∑_j=k^d v_j) ] = exp( ∑_k=1^d ( t_k - t_k-1) η( u̅_k, v̅_k ) ) using respectively telescopic summation, the independence of increments, the fact that ( X_t_2, Y_t_2) - ( X_t_1, Y_t_1) law=( X_t_2-t_1, Y_t_2-t_1) and the definition of the characteristic function of ( X_t_k-t_k-1, Y_t_k-t_k-1) for all k ∈ 1, …, d to get the successive equalities. Using the parameters (α, β, μ, δ,λ) as in (<ref>) it is immediate to see that ∑_k=1^d ( t_k - t_k-1) η( u̅_k, v̅_k ) = ϕ_0(T), hence the desired convergence (<ref>). §.§ Weak-convergence of the integrated variance process for the M_1 topology In this section, we prove the weak convergence stated in Theorem <ref> of the integrated variance process V̅^ϵ with sample paths in C([0,T], ℝ^+) to the Lévy process V̅^0 whose Lévy exponent is given by v ↦η(0,v) with η defined in (<ref>) and with sample paths valued in the càdlàg functional space D := 𝒟([0,T], ℝ^+)) endowed with Skorokhod's Strong M1 (SM_1) topology. There will be two subsections: first, in Section <ref> we recall briefly the definition of the Strong M1 (SM_1) topology as well as some associated convergence results, and then we prove the tightness of the integrated variance in Section <ref>. §.§.§ Reminder on the SM_1 topology and conditions for convergence We recall succinctly key definitions and convergence theorems for the M_1 topology. We refer the reader to the key reference book <cit.> for more details. For x ∈ D, we define the thin graph of x as Γ_x := { (z,t) ∈ℝ×[ 0, T ]: z ∈[ x(t^-), x(t) ] }, where, for t ∈[ 0,T ], [ x(t^-), x(t) ] denotes the standard segment {α x(t^-) + ( 1 - α) x(t), 0 ≤α≤ 1 }, which is different from a singleton at discontinuity points of the càdlàg sample trajectory x. We denote Disc(x) the set of such instants. Define on Γ_x the strong order relation as follows: (z_1,t_1) ≤(z_2,t_2) if either t_1<t_2 or t_1=t_2 and |x(t_1^-)-z_1| ≤|x(t_1^-)-z_2|. Furthermore define a strong parametric representations of x as a continuous non-decreasing (with respect to the previous order relation) function ( x̂, r̂) mapping [ 0,1 ] into Γ_x such that ( x̂(0), r̂(0) ) = ( x(0),0 ) and ( x̂(T), r̂(T) ) = ( x(T),T ). We say the component r̂ scales the time interval [ 0,T ] to [ 0,1 ] while x̂ time-scales x, and we denote by Π_s(x) the set of all strong parametric representations of x. Finally, the strong M1 topology is the topology induced by the metric d_s defined as d_s( x_1, x_2) := ( x̂^i, r̂^i) ∈Π_s(x^j), j=1,2inf{| x̂^1 - x̂^2| ∨| r̂^1 - r̂^2| }. Below we mention briefly five theorems in a row that eventually yields criteria to prove the desired convergence. D := D([0,T], ℝ^+) endowed with the SM_1 topology is Polish, i.e. metrizable as a complete separable metric space. Let (S,m) be a metric space. If a subset A in 𝒫(S) is tight, then it's relatively compact. On the other hand, if the subset A is relatively compact and the topological space is Polish, then A is tight. The first two theorems ensure that, since (D,SM_1 ) is Polish, proving relative compactness of any family of probability measures on such space is sufficient to ensure the existence of a convergent sub-sequence. Finally the convergence of finite-dimensional laws allows to uniquely determine its limit as formulated in the following theorem. Let ( ( X_n)_n∈ℕ, X ) be random functions defined on a common filtered probability space ( Ω, ( ℱ_t )_t, ℙ). Then, X_n ⇒ X in D for the SM_1 topology if the following conditions hold: * (X_n)_n∈ℕ is tight with regards to the SM_1 topology. * The finite dimensional distributions of X_n converge to those of X on T_X, where: T_X := { t > 0, ℙ(t ∈ Disc(X) ) = 0 }∪{T}. To conclude this section, we recall a characterization of tightness for a sequence of probability measures. The sequence of probability measures {ℙ_n }_n ≥ 1 on (D,SM_1) is tight if and only if: (i) ∀ε̅>0, ∃ c < ∞, ∀ n ≥ 1, ℙ_n ( { x ∈ D: ||x|| > c }) < ε̅ (ii) ∀ε̅>0, ∀η > 0, ∃δ>0, ∀ n ≥ 1, ℙ_n ( { x ∈ D: w'(x,δ) ≥η}) < ε̅ Where we defined for x ∈ D, t∈ [0,T] and δ>0: ||x|| := t ∈ [0,T]sup |x_t|, w'(x,δ) := w(x,δ) ∨v̅(x,0,δ) ∨v̅(x,T,δ) , w(x,δ) := t ∈ [0,T]sup w_S(x,t,δ), w_S(x,t,δ) := 0 ∨ t-δ≤ t_1 < t_2 < t_3 ≤ (t+δ) ∧ Tsup| x(t_2)-[ x(t_1), x(t_3)] |, v̅(x,t,δ) := 0 ∨ t-δ≤ t_1 ≤ t_2 ≤ (t+δ) ∧ Tsup| x(t_1)-x(t_2) |. §.§.§ Convergence of the integrated variance process We already proved in Section <ref> the convergence of finite dimensional distributions as ϵ goes to zero of V̅^ϵ toward those of either the deterministic linear, or the Inverse Gaussian, or the Lévy process denoted Y depending respectively on the value of the Hurst index with Lévy exponent η(0,·) from (<ref>) with the respective parameters given in Theorem <ref>. Consequently, all that remains to prove is the tightness of the family of processes ( V̅^ϵ)_ϵ>0 for the SM_1 topology to get the desired convergence result as a direct consequence of Theorem <ref>. We will apply the characterization Theorem <ref> of tightness in SM_1 to conclude, and more precisely, we will see that the criteria of tightness within the SM_1 topology simplifies greatly for almost surely non-decreasing and continuous stochastic processes in general. Fix ε̅> 0. Since, for all ϵ>0, V̅^ϵ is almost surely non-decreasing and non-negative, we have that, for all ω∈Ω ||V̅^ϵ(ω)|| = V̅_T^ϵ(ω), w'(V̅^ϵ(ω)), δ) = 0∨V̅^ϵ_δ(ω) ∨( V̅^ϵ_T(ω) - V̅^ϵ_T-δ(ω)), δ >0. This yields that, for a threshold c>0 big enough, the probability in condition (i) on the measures ℙ_ϵ := ℙ_(V̅_t^ϵ)^-1 reduces to ℙ_ϵ( { x ∈ D, ||x|| > c }) = ℙ( {ω: ||V̅_.^ϵ(ω)||>c }) = ℙ( V̅_T^ϵ>c ) ≤sup_ϵ>0ℙ( V̅_T^ϵ>c ) < ε̅, where the last inequality is satisfied by tightness of the family ( V̅_T^ϵ)_ϵ>0 of random variables in which is obtained as a direct consequence of Lévy's continuity theorem, recall that ( V̅_T^ϵ)_ϵ>0 has been shown to converge in Section <ref>. This yields (i). In addition, regarding the second condition (ii), set an arbitrary η>0, then the oscillation function w' simplifies to ℙ_ϵ( { x ∈ D: w'(x,δ) ≥η}) = ℙ( 0 ∨|V̅_δ^ϵ| ∨|V̅_T^ϵ - V̅_T-δ^ϵ| ≥η) < ε̅, where we take δ small enough to ensure the last inequality by stochastic continuity of V̅^ϵ. Since log S^ϵ = - 1/2V̅^ϵ + ρ W_V̅^ϵ + √(1-ρ^2) W^⊥_V̅^ϵ, and composition is not continuous in ( D, M_1 ) (see <cit.>), we cannot expect the tightness of the log price within SM_1. § SOME LEMMAS (Uniqueness of the complex root with a non-positive real part) Take ξ>0. For all f, g bounded and measurable with f = g = 0, t ∈[ 0,T ] and ρ∈[ -1,1 ], both polynomials P(X) := ξ^2/2 X^2 - ( 1 - ρξ f(t) ) X + g(t) + f^2(t) - f(t)/2 Q(X) := ξ^2/2 X^2 + ρξ f(t) X + g(t) + f^2(t) - f(t)/2 admit exactly two roots with respective real parts of strict opposite signs if ( f(t), g(t) ) ≠( 0, 0 ), and if ( f(t), g(t) ) = ( 0, 0 ), then the polynomial P has roots 0 and 2/ξ^2, while Q has 0 as a double root. Let us detail the proof for P, similar arguments will apply to Q. By d'Alembert-Gauss theorem, the polynomial P admits exactly two roots expressed as: {ξ^-2( 1 - ρξ f(t) ±√(( 1 - ρξ f(t) )^2 - 2 ξ^2 ( g(t) + f^2(t)-f(t)/2))) }, where we take the principal square root in the expression above, i.e.with non-negative real-part. Consequently, the roots have real parts {ξ^-2( 1 ±(√(( 1 - ρξ f(t) )^2 - 2 ξ^2 ( g(t) + f^2(t)-f(t)/2))) ) }, so that it remains to show | (√(( 1 - ρξ f(t) )^2 - 2 ξ^2 ( g(t) + f^2(t)-f(t)/2))) | > 1. Denote δ=a+ib, a,b ∈ℝ such that δ^2 = ( 1 - ρξ f(t) )^2 - 2 ξ^2 ( g(t) + f^2(t)-f(t)/2), then it follows that a and b satisfy a^2 - b^2 = 1 + ( 1 - ρ^2 ) ( ξ f(t) )^2, ab = - ( ρξ f(t) + ξ^2 ( g(t) - f(t)/2) ). If ρ≠± 1, then the result is immediate from the first inequality in (<ref>), while if ρ = ± 1, then |a| = √(1+b^2) and b cannot be zero, otherwise δ^2 = a^2 = 1 - 2i( ρξ f(t) + ξ^2 ( g(t) - f(t)/2) ) which cannot be the case, since a ∈ℝ and ( f(t), g(t)) ≠( 0, 0 ). Let f and g be bounded measurable functions such that f = g = 0. Then there exists a finite positive constant C such that the ratio γ(s) := |ψ_0(s)|/- ψ_0(s)≤ C, ∀ s ∈ E, where the set E is given by E := { s ∈ [0,T], ( f(s), g(s) ) ≠( 0, 0 ) }, and ψ_0 is given in (<ref>) in the case H<-1/2. We start by explicitly computing the real and imaginary parts of ψ_0 in the case H<-1/2, whose expression is given anew by ψ_0(s) = - ξ^-1( ρ f(s) + √(f(s) ( 1 - ( 1- ρ^2 ) f(s) ) - 2 g(s))), s ∈ [0,T]. Set the real functions a and b such that, for any s ∈ [0,T] a(s) + i b(s) = √(f(s) ( 1 - ( 1- ρ^2 ) f(s) ) - 2 g(s)), square the above equality, identify the real and imaginary parts, find a unambiguously on E (which imposes a ≠ 0) as a root to a fourth-degree polynomial knowing the square roots in (<ref>) are principal (i.e. with positive real parts), then deduce the associated b, such that a = √(1/2( ( 1 - ρ^2 )( f )^2 + √(( 1 - ρ^2 )^2( f )^4 + ( f - 2 g)^2))), b = 1/2 f - g/a. Consequently, we get explicitly on E ψ_0 = ψ_0 + i ψ_0, with ψ_0 = -ξ^-1 a, ψ_0 = -ξ^-1( ρ f + 1/2 f - g/a). Rewrite γ as γ = √( 1 + γ̃^2), γ̃:= ψ_0/ψ_0, and we can readily discard the case ρ=0 and f - 2 g=0, since ψ_0 = 0 in that case and the ratio simplifies into 1 which yields the result. Assume from now on that ρ≠ 0 or f - 2 g ≠ 0. We can write γ̃= A( f, g ) + B( f, g ), with A( f, g ) := ρ f/√(1/2( ( 1 - ρ^2 )( f )^2 + √(( 1 - ρ^2 )^2( f )^4 + ( f - 2 g)^2))), B( f, g ) := 1/2 f - g/1/2( ( 1 - ρ^2 )( f )^2 + √(( 1 - ρ^2 )^2( f )^4 + ( f - 2 g)^2)), so that there are three remaining cases for which it is sufficient to show that both A( f, g ) and B( f, g ) are bounded to conclude the proof. * Case ρ≠ 0 and f - 2 g = 0, then A( f, g ) = ρ/√(1-ρ^2), B( f, g ) = 0, are both bounded, recall ρ∈ (-1,1). * Case ρ = 0 and f - 2 g ≠ 0, then A( f, g ) = 0, B( f, g ) = 1/2 f - g/1/2( ( f )^2 + √(( f )^4 + ( f - 2 g)^2)), and B( f, g ) is bounded since ( x, y ) ↦B( x, y ) is continuous on any compact set of ℝ^2\{(0,0)} (recall both f and g are bounded) and has a finite limit at (0,0), valued 1, indeed B( x, y ) = 1/( 1-ρ^2) x^2/x-2y + √(( 1-ρ^2)^2x^4/( x-2y )^2+1)x-2y ≠ 0(x,y) →(0,0)⟶ 1. * Case ρ≠ 0 and f - 2 g ≠ 0, then A( f, g ) = ρ f/√(1/2( ( 1 - ρ^2 )( f )^2 + √(( 1 - ρ^2 )^2( f )^4 + ( f - 2 g)^2))), B( f, g ) = 1/2 f - g/1/2( ( 1 - ρ^2 )( f )^2 + √(( 1 - ρ^2 )^2( f )^4 + ( f - 2 g)^2)), are both bounded by continuity of ( x, y ) ↦A( x, y ) and ( x, y ) ↦B( x, y ) on any compact set of ℝ^2\{(0,0)} (recall both f and g are bounded) and both functions have a finite limit at (0,0), valued 0 and 1 respectively, obtained with similar arguments as in the previous case. (Inverse Gaussian, Lévy and NIG-IG distributions) * We say X follows a Normal Inverse distribution, denoted X ↪IG( μ, λ) if its probability density writes f(x) = √(λ/2π x^3)exp( -λ( x-μ)^2/2 μ^2 x), where μ∈ℝ, λ>0, or equivalently if the following equality holds true 𝔼[ exp( w X ) ] = exp( λ/μ( 1 - √(1-2μ^2 w/λ)) ), w ∈ℂ, w ≤ 0. * We say τ follows a Lévy distribution, denoted τ↪Lévy( μ, c), if its probability density writes √(c/2π)e^-c/2(x-μ)/( x-μ)^3/2, where μ∈ℝ, c>0, or equivalently if the following equality holds true 𝔼[exp( w τ)] = exp( μ w - √(-2c w)), w ∈ℂ, w ≤ 0. * We say ( X, Y ) follows a Normal Inverse Gaussian - Inverse Gaussian distribution, denoted ( X, Y ) ↪NIG-IG( α, β, μ, δ, λ), if its characteristic function writes 𝔼[ exp( iu X + iv Y ) ] = exp[ i μ u + δ( √(α^2 - β^2) -√(α^2 - 2iλ v -( β + iu )^2)) ], u, v ∈ℝ, § ADDITIONAL PLOTS plainnat

http://arxiv.org/abs/2303.06742v1

20230312200332

An energy-efficient GMRES-Multigrid solver for space-time finite element computation of dynamic poro- and thermoelasticity

math.NA

An energy-efficient GMRES–Multigrid solver for space-time finite element computation of dynamic poro- and thermoelasticity Mathias Anselmann^∗anselmann@hsu-hh.de (corresponding author) , Markus Bause^∗, Nils Margenberg^∗, Pavel Shamko^∗ ^∗ Helmut Schmidt University, Faculty of Mechanical and Civil Engineering, Holstenhofweg 85, 22043 Hamburg, Germany ================================================================================================================================================================================================================================================== Abstract We present families of space-time finite element methods (STFEMs) for a coupled hyperbolic-parabolic system of poro- or thermoelasticity. Well-posedness of the discrete problems is proved. Higher order approximations inheriting most of the rich structure of solutions to the continuous problem on computationally feasible grids are naturally embedded. However, the block structure and solution of the algebraic systems become increasingly complex for these members of the families. We present and analyze a robust geometric multigrid (GMG) preconditioner for GMRES iterations. The GMG method uses a local Vanka-type smoother. Its action is defined in an exact mathematical way. Due to nonlocal coupling mechanisms of unknowns, the smoother is applied on patches of elements. This ensures the damping of error frequencies. In a sequence of numerical experiments, including a challenging three-dimensional benchmark of practical interest, the efficiency of the solver for STFEMs is illustrated and confirmed. Its parallel scalability is analyzed. Beyond this study of classical performance engineering, the solver's energy efficiency is investigated as an additional and emerging dimension in the design and tuning of algorithms and their implementation on the hardware. Keywords. Poro-/thermoelasticity, dynamic Biot model, space-time finite element approximation, geometric multigrid method, GMRES iterations, computational efficiency, energy consumption. § INTRODUCTION §.§ Mathematical model We present and investigate numerically a geometric multigrid preconditioning technique, based on a local Vanka-type smoother, for solving by GMRES iterations the linear systems that arise from space-time finite element discretizations of the coupled hyperbolic-parabolic system of dynamic poroelasticity ρ∂_t^2 u⃗ - ∇·(C⃗ ε⃗(u⃗)) + α∇⃗p = ρf⃗ , in Ω×(0,T] , c_0∂_t p + α∇·∂_t u⃗ - ∇·(K⃗ ∇p) = g , in Ω×(0,T] , u⃗ (0) = u⃗_0 , ∂_t u⃗ (0) = u⃗_1 , p(0) = p_0 , in Ω×{0} , u⃗ = u⃗_D , on Γ_u⃗^D ×(0,T] , -(C⃗ε⃗(u⃗) - αpE⃗) n⃗ = t⃗_N , on Γ_u⃗^N ×(0,T] , p = p_D , on Γ_p^D ×(0,T] , - K⃗ ∇p ·n⃗ = p_N , on Γ_p^N ×(0,T] . In (<ref>), Ω⊂^d, with d∈{2,3}, is an open bounded Lipschitz domain with outer unit normal vector n⃗ to the boundary ∂Ω and T>0 is the final time point. We let ∂Ω = Γ_u⃗^D∪Γ_u⃗^N and ∂Ω = Γ_p^D∪Γ_p^N with closed portions Γ_u⃗^D and Γ_p^D of non-zero measure. Important applications of the model (<ref>), that is studied as a prototype system, arise in poroelasticity; cf. <cit.> and <cit.>. In poroelasticity, Eqs. (<ref>) are referred to as the dynamic Biot model. The system (<ref>) is used to describe flow of a slightly compressible viscous fluid through a deformable porous matrix. The small deformations of the matrix are described by the Navier equations of linear elasticity, and the diffusive fluid flow is described by Duhamel’s equation. The unknowns are the effective solid phase displacement u⃗ and the effective fluid pressure p. The quantity ε⃗(u⃗):= (∇u⃗ + (∇u⃗)^⊤)/2 denotes the symmetrized gradient or strain tensor and E⃗∈^d,d is the identity matrix. Further, ρ is the effective mass density, C⃗ is Gassmann’s fourth order effective elasticity tensor, α is Biot’s pressure-storage coupling tensor, c_0 is the specific storage coefficient and K⃗ is the permeability field. For brevity, the positive quantities ρ>0, α>0 and c_0 >0 as well as the tensors C⃗ and K⃗ are assumed to be constant in space and time. The tensors C⃗ and K⃗ are assumed to be symmetric and positive definite, ∃k_0>0 ∀ξ⃗= ξ⃗^⊤∈^d,d: ∑_i,j,k,l=1^d ξ_ij C_ijkl ξ_kl ≥k_0 ∑_j,k=1^d |ξ_jk|^2 , ∃k_1>0 ∀ξ⃗∈^d: ∑_i,j,=1^d ξ_i K_ij ξ_j ≥k_1 ∑_i=1^d |ξ_i|^2 . In order to enhance physical realism in poroelasticity, generalizations of the model (<ref>) have been developed and investigated in, e.g., <cit.>. We note that the system (<ref>) is also formally equivalent to the classical coupled thermoelasticity system which describes the flow of heat through an elastic structure; cf. <cit.> and <cit.>). In that context, p denotes the temperature, c_0 is the specific heat of the medium, and K⃗ is the conductivity. Then, the quantity α∇ p arises from the thermal stress in the structure, and the term α∇·∂_t u⃗ models the internal heating due to the dilation rate. Well-posedness of (<ref>) is ensured; cf., e.g., <cit.>. This can be shown by different mathematical techniques, by semigroup methods <cit.>, Rothe's method <cit.> and Picard's theorem <cit.>. In these works, boundary conditions different to the ones in (<ref>) to (<ref>) are partly used. §.§ Space-time finite element and multigrid techniques The coupled hyperbolic-parabolic structure of the system (<ref>) of partial differential equations adds an additional facet of complexity onto its numerical simulation. A natural and promising approach for the numerical approximation of coupled systems is given by the application of space-time finite element methods that are based on an uniform treatment of the space and time variables by variational formulations. Therein, the discrete unknown functions are defined on the entire space-time domain Ω× [0,T] and are represented in terms of basis functions. This facilitates the discretization of even complex coupling terms that may involve combinations of temporal and spatial derivatives as in (<ref>). Moreover, space-time finite element methods enable the natural construction of higher order members in the families of schemes. Higher order discretizations offer the potential to achieve accurate results on computationally feasible grids with a minimum of numerical costs. In particular for second-order hyperbolic problems, higher order time discretization schemes have shown their superiority, measured in terms of accuracy over temporal mesh resolution and compute time; cf., e.g., <cit.>. Finally, space-time adaptivity based on a-posteriori error control by duality concepts and multi-rate in time approaches become feasible. For this we refer to <cit.> for the presentation of the abstract framework of duality based error control and, e.g., to <cit.> for its application to physical problems. For the variational time discretization of second-order hyperbolic and parabolic equations, various families have been proposed and investigated recently; cf., e.g., <cit.>. They differ by the approximation order and regularity in time of the discrete solution as well as stability and energy conservation properties. In this work, we use a discontinuous Galerkin time discretization of arbitrary polynomial order and focus on the efficient solution of the resulting complex algebraic block systems. Higher order (variational) time discretizations for three-dimensional applications put a facet of complexity on the algebraic solver and have rarely been studied yet in the literature. For the discretization of the space variables, numerous finite element approaches have been proposed and studied, in particular, for the quasi-static Biot system of flow in deformable porous media. The latter model results from (<ref>) by neglecting the acceleration term ρ∂_t^2 u⃗ in (<ref>). For the quasi-static Biot model, three-field formulations with the fluid flux q⃗ = - K⃗∇ p as an additional explicit variable and the application of mixed finite element methods for the spatial discretization of p and q⃗ attracted strong interest of researchers; cf., e.g., <cit.> and the references therein. In this work, we use an approximation of the variables u⃗ and p by higher order inf-sup stable pairs of finite element spaces with continuous or discontinuous discrete variable p. Precisely, either the well-known Taylor–Hood family of finite element pairs for u⃗ and p or a continuous approximation of u⃗ combined with a discontinuous approximation of p is used; cf. (<ref>). In the literature, the latter option has shown its superiority for the approximation of incompressible viscous flow problems <cit.>. Structure-preserving approximations with the fluid flux and stress tensor as additional variables, that are based on a reformulation of the coupled problem (<ref>) as a first-order in space and time system (cf. <cit.>), remain as a work for the future. In our space-time finite element approach, we choose a discontinuous temporal test basis such a time-marching process is obtained. On each subinterval I_n=(t_n-1,t_n] of the time mesh ℳ_τ := {I_1,…, I_N}, we use a Lagrangian basis with respect to the (k+1) Gauss–Radau quadrature points of I_n for the time interpolation. Two discrete systems differing by the treatment of the term ∇·∂_t u⃗ in (<ref>) are then proposed. Dirichlet boundary conditions are implemented in a weak form by means of Nitsche's method and/or a discontinuous (spatial) approximation of the scalar variable p. Well-posedness of the discrete problems is carefully proved for arbitrary polynomial degrees of the discretization. The degrees of freedom on I_n are associated with the (k+1) Gauss–Radau quadrature nodes of this subinterval; cf. Fig. <ref>. On each subinterval I_n, the resulting algebraic problem is a linear system of equations with a block matrix of (k+1)× (k+1) subsystems where each of the subsystems exhibits a complex structure itself. Here, k denotes the piecewise polynomial order of the time discretization; cf. <cit.>. The solution of these algebraic problems with complex block structure demands for efficient and robust iteration schemes. Geometric multigrid (GMG) methods are known as the most efficient iterative techniques for the solution of large linear systems arising from the discretization of partial differential equations, particularly of elliptic type; cf. <cit.> for a general overview. In contrast to other iterative solvers (like for instance BiCGStab or GMRES), multigrid solvers converge with a rate which is independent of the (spatial) mesh size and require computational costs that are only linearly dependent on the number of unknowns. Massively parallel implementations of GMG methods on modern architectures show excellent scalability properties and their high efficiency has been recognized, e.g., in <cit.>. Today they are widely used and are employed in many variants; cf. <cit.> for flow and saddle point problems. Nowadays, GMG methods are employed as preconditioner in Krylov subspace iterations, for instance GMRES iterations, to enhance their robustness, which is also done here. Analyses of GMG methods (cf., e.g., <cit.>) have been done for linear systems in saddle point form, with matrix 𝒜⃗= [ A⃗ B⃗^⊤; B⃗ - C⃗ ] and symmetric and positive definite submatrices A⃗ and C⃗, arising for instance from mixed discretizations of the Stokes problem. If higher order discontinuous Galerkin time stepping is used, the resulting linear system matrix is a (k+1)× (k+1) block matrix with each of its blocks being of the form 𝒜⃗. This imposes an additional facet of complexity on the geometric multigrid preconditioner. To the best of our knowledge, analyses of GMG methods arising form higher order (variational) time discretizations of the coupled system (<ref>) or related ones are still missing in the literature. GMG methods exploit different mesh levels of the underlying problem in order to reduce different frequencies of the error by employing a relatively cheap smoother on each grid level. Different iterative methods have been proposed in the literature as smoothing procedure; cf. <cit.> and the references therein. They range from low-cost methods like Richardson, Jacobi, and SOR applied to the normal equation of the linear system to collective smoothers, that are based on the solution of small local problems. Here, we use a Vanka-type smoother <cit.> of the family of collective methods. Numerical computations show that an elementwise application of the Vanka smoother fails to reduce the high frequencies of the error on the multigrid levels. The reason for this comes through spatial interelement couplings of degrees of freedom of the scalar variable p in (<ref>). This holds for the discontinuous and continuous spatial approximation of p. For the spatially continuous approximation of p, the coupling of degrees of freedom is obvious. For the spatially discontinuous approximation of p, the interelement coupling of degrees of freedom of p is due to the presence of the diffusion term in (<ref>) and the surface integrals involving averages and jumps of the discrete solution over interelement faces resulting from its discretization. This is in contrast to the Navier–Stokes equations for which elementwise Vanka-type smoother are known to perform excellently if discontinuous pressure approximations are applied; cf. <cit.>. The difference comes through the fact that in case of the Navier–Stokes system no interelement coupling of pressure degrees of freedom arises. As a remedy, we propose the application of the Vanka-type smoother on cell patches that are linked to the grid nodes and built from four neighbored cells in two dimensions and eight neighbored elements in three dimensions, with appropriate adaptations for grid nodes close to or on the domain's boundary. Further we employ some relaxation strategy in the smoothing steps. Then an efficient damping of frequencies in the error on the multigrid hierarchy is obtained. Here, this Vanka-type smoother is presented in an mathematically exact way and its performance properties are investigated by numerical experiments. Our numerical experiments confirm that GMRES iterations that are preconditioned by the proposed multigrid method converge at a rate which is (nearly) independent of the mesh sizes in space and the time step size; cf. also <cit.>. In the smoothing steps, the application of the Vanka-type operator requires the solution of patchwise linear systems of equations. For (<ref>) and three space dimensions the size of the local systems no longer continues to remain negligible, in particular if a higher polynomial order in space and/or time is applied. Their solution deserves attention. In our approach we compute of inverse of the patchwise system matrix by LAPACK routines to solve to local system of equations. Details of an data-structure efficient MPI-based parallel implementation of the GMG algorithm in our in-house high-performance frontend solver for the deal.II finite element library <cit.> can be found in <cit.> where this is presented for the Navier–Stokes system. Multi-field formulations of the system (<ref>), that are not studied in this work, further increase the complexity of the arising linear systems; cf. <cit.>. However, the methodology of the linear solver, that is developed here, offers the potential of its generalization to more advanced problem formulations and their discretizations. §.§ Energy efficiency In the past, performance engineering and hardware engineering for large scale simulations of physical phenomena have been eclipsed by the longing for ever more performance where faster seemed to be the only paradigm. ”Classical” performance engineering has been applied to enhance, firstly, the efficiency of the current method on the target hardware or to find numerical alternatives that might better fit to the hardware in use and/or, secondly, to develop other numerical methods can be found to improve the numerical efficiency. Tuning both simultaneously is called hardware-oriented numerics in the literature; cf. <cit.>. Since recently, a growing awareness of energy consumption in computational science, particularly, in extreme scale computing with a view to exascale computing has raised; cf., e.g., <cit.>. It has been observed that as a consequence of decades of performance-centric hardware development there is a huge gap between pure performance and energy efficiency. An analysis of our algorithm's parallel scaling and energy consumption properties by performance models exceeds the scope of this work and would overburden it. However, since energy consumption of application codes on the available hardware is of growing awareness and a key for future improvements, we study the energy consumption and parallel scaling properties of our algorithm and its implementation by three-dimensional numerical experiments. The development of a proper model that quantifies performance and energy efficiency in some appropriate metric and can be used for a code optimization still deserves research and is left as a work for the future. §.§ Outline of the work This work is organized as follows. In Sec. <ref> we introduce our notation. In Sec. <ref> the space-time finite element approximation of arbitrary order of (<ref>) is derived and well-posedness of the fully discrete problem is proved. Our GMRES–GMG solver is introduced in Sec. <ref>. In Sec. <ref> our performed numerical computations for analyzing the performance properties of the overall approach are presented. In Sec. <ref> we end with a summary and conclusions. In the appendix, supplementary results are summarized. § BASIC NOTATION In this work, standard notation is used. We denote by H^1(Ω) the Sobolev space of L^2(Ω) functions with first-order weak derivatives in L^2(Ω). Further, H^-1(Ω) is the dual space of H^1_0(Ω), with the standard modification if the Dirichlet condition is prescribed on a part Γ^D⊂Γ of the boundary ∂Ω only; cf. (<ref>). The latter is not explicitly borne out by the notation of H^-1(Ω). It is always clear from the context. For vector-valued functions we write those spaces bold. By ⟨·, ·⟩_S we denote the L^2(S) inner product for a domain S. For S=Ω, we simply write ⟨·, ·⟩. For the norms of the Sobolev spaces the notation is · := ·_L^2 , ·_1 := ·_H^1 . For short, we put Q:=L^2(Ω) and V⃗:= (H^1(Ω))^d . For a Banach space B, we let L^2(0,T;B) be the Bochner space of B-valued functions, equipped with its natural norm. For a subinterval J⊆ [0,T], we will use the notation L^2(J;B) for the corresponding Bochner space. In what follows, the constant c is generic and indepedent of the size of the space and time meshes. For the time discretization, we decompose the time interval I:=(0,T] into N subintervals I_n=(t_n-1,t_n], n=1,…,N, where 0=t_0<t_1< ⋯ < t_N-1 < t_N = T such that I=⋃_n=1^N I_n. We put τ := max_n=1,…, Nτ_n with τ_n = t_n-t_n-1. Further, the set ℳ_τ := {I_1,…, I_N} of time intervals is called the time mesh. For a Banach space B and any k∈_0, we let ℙ_k(I_n;B) := {w_τ : I_n → B , w_τ(t) = ∑_j=0^k W^j t^j ∀ t∈ I_n , W^j ∈ B ∀ j } . For k∈_0 we define the space of piecewise polynomial functions in time with values in B by Y_τ^k (B) := {w_τ: I→ B | w_τ_|I_n∈ℙ_k(I_n;B) ∀ I_n∈ℳ_τ, w_τ(0)∈ B }⊂ L^2(I;B) . For any function w: I→ B that is piecewise sufficiently smooth with respect to the time mesh ℳ_τ, for instance for w∈ Y^k_τ (B), we define the right-hand sided and left-hand sided limit at a mesh point t_n by w^+(t_n) := lim_t→ t_n+0 w(t) , for n<N , and w^-(t_n) := lim_t→ t_n-0 w(t) , for n>0 . For the integration in time of a discontinuous Galerkin approach it is natural to use the right-sided (k+1)-point Gauss–Radau quadrature formula. On the subinterval I_n, it reads as Q_n(w) := τ_n/2∑_μ=1^k+1ω̂_μ^GR w(t_n,μ^GR ) ≈∫_I_n w(t) t , where t_n,μ^GR=T_n(t̂_μ^GR), for μ = 1,…,k+1, are the Gauss–Radau quadrature points on I_n and ω̂_μ^GR the corresponding weights. Here, T_n(t̂):=(t_n-1+t_n)/2 + (τ_n/2)t̂ is the affine transformation from Î = [-1,1] to I_n and t̂_μ^GR are the Gauss–Radau quadrature points on Î. Formula (<ref>) is exact for all polynomials w∈ℙ_2k (I_n;). In particular, there holds that t_n,k+1^GR=t_n. For the space discretization, let {𝒯_l}_l=0^L be the decomposition on every multigrid level of Ω into (open) quadrilaterals or hexahedrals, with 𝒯_l = {K_i| i=1,… , N^el_l}, for l=0,…,L. These element types are chosen for our implementation (cf. Sec. <ref>) that is based on the deal.II library <cit.>. The finest partition is 𝒯_h=𝒯_L. We assume that all the partitions {𝒯_l}_l=0^L are quasi-uniform with characteristic mesh size h_l and h_l=γ h_l-1, γ∈ (0,1) and h_0 = 𝒪(1). On the actual mesh level, the finite element spaces used for approximating the unknowns u⃗ and p of (<ref>) are of the form (l∈{0,…,L}) V⃗_h_l^l := {v⃗_h ∈V⃗ ∩C(Ω)^d : v⃗_h_|K∈V⃗(K) for all K ∈𝒯_l} , Q_h_l^l,cont := {q⃗_h ∈Q∩C(Ω) : q⃗_h_|K∈Q(K) for all K ∈𝒯_l} , Q_h_l^l,disc := {q⃗_h ∈Q : q⃗_h_|K∈Q(K) for all K ∈𝒯_l} . By an abuse of notation, we skip the index l of the mesh level when it is clear from the context and put V⃗_h := V⃗_h^l and Q_h := Q_h^l with Q_h^l∈{Q_h^l,cont, Q_h^l,disc} . For the local spaces V⃗(K) and Q(K) we employ mapped versions of the pairs ℚ_r^d/ℚ_r-1 and ℚ_r^d/ℙ_r-1^, for r≥ 2. The pair ℚ_r^d/ℚ_r-1 with a (globally) continuous approximation of the scalar variable p in Q_h^l,cont is the well-known Taylor–Hood family of finite element spaces. The pair ℚ_r^d/ℙ_r-1^ comprises a discontinuous approximation of p in the broken polynomial space Q_h^l,disc. For the Navier–Stokes equations, the multigrid method has shown to work best for higher-order finite element spaces with discontinuous discrete pressure; cf. <cit.> and <cit.>. For a further discussion of mapped and unmapped versions of the pair ℚ_r^d/ℙ_r-1^ we refer to <cit.>. For an analysis of stability properties of (spatial) discretizations for the quasi-static Biot system we refer to, e.g., <cit.>. In particular, both choices of the local finite element spaces, ℚ_r^d/ℚ_r-1 and ℚ_r^d/ℙ_r-1^, satisfy the inf–sup stability condition, inf_q_h ∈ Q_h\{0}sup_v⃗_h∈V⃗_h\{0⃗}b(v⃗_h,q_h)v⃗_h _1 q_h≥β > 0 , for some constant β independent of h; cf. <cit.>. In <cit.> optimal interpolation error estimates for mapped finite elements on quadrilaterals and hexahedra are studied. It turned out that the optimality is given for special families of triangulations. In two and three dimensions, families of meshes, which are obtained by a regular uniform refinement of an initial coarse grid, are among these special families. Such a regular refinement that is natural for the construction of the multigrid hierarchy is used in our computations. Thus, for v⃗∈H⃗^r+1(Ω) and q∈ H^r(Ω) there exist approximations i_h v⃗∈V⃗_h and j_h q ∈ Q_h such that, with some generic constant c>0 independent of h, v⃗ - i_h v⃗ + h ∇ (v⃗-i_h v⃗)≤ c h^r+1 and q - j_h q≤ c h^r . § SPACE-TIME FINITE ELEMENT APPROXIMATION For the discretization we rewrite (<ref>) as a first-order in time system by introducing the new variable v⃗:= ∂_t u⃗. Then, we recover (<ref>) and (<ref>) as ∂_t u⃗ - v⃗ = 0⃗ ρ∂_t v⃗ - ∇·(C⃗ ε⃗(u⃗)) + α∇⃗p = ρf⃗ , c_0∂_t p + α∇·v⃗ - ∇·(K⃗ ∇⃗p) = g along with the initial and boundary conditions <ref> to <ref>. We employ discontinuous Galerkin methods (cf. <cit.>) for the discretization of the time variable and inf-sup stable pairs of finite elements (cf. Sec. <ref>) for the approximation of the space variables in <ref>. The derivation of the discrete scheme, presented below in Problem <ref>, is standard and not explicitly presented here. It follows the lines of <cit.>, where continuous in time Galerkin methods are applied to (<ref>), and <cit.>, where discontinuous in time Galerkin methods are used to discretize the Navier–Stokes system. In contrast to <cit.>, Dirichlet boundary conditions are implemented here by Nitsche’s method <cit.>. This yields a strong link between two different families of inf-stable finite element pairs for the space discretization. The main reason for using Nitsche’s method here comes through our more general software framework. It captures problems on evolving domains solved on fixed computational background grids (cf. <cit.>), where Nitsche’s method is applied. We note that Nitsche’s method does not perturb the convergence behavior of the space-time discretization; cf. Sec. <ref>. For the discrete scheme we need further notation. On the multigrid level l with decomposition 𝒯_l, for w⃗_h, χ⃗_h∈V⃗_h and q_h, ψ_h ∈ Q_h we define A_γ(w⃗_h,χ⃗_h) := ⟨C⃗ ε⃗(w⃗_h),ε⃗(χ⃗_h)⟩- ⟨C⃗ε⃗(w⃗_h) n⃗, χ⃗_h⟩_Γ^D_u⃗+ a_γ(w⃗_h,χ⃗_h) , C (χ⃗_h,q_h) := -α⟨∇·χ⃗_h, q_h⟩+ α⟨χ⃗_h ·n⃗ , q_h ⟩_Γ^D_u⃗ , B_γ(q_h,ψ_h) := {[ ⟨K⃗ ∇q_h, ∇ψ_h ⟩- ⟨K⃗ ∇q_h ·n⃗, ψ_h ⟩_Γ_p^D + b_γ(q_h,ψ_h) , for Q_h = Q_h^l,cont ,; [ ∑_K∈𝒯_l⟨K⃗ ∇q_h, ∇ψ_h ⟩_K - ∑_F∈ℱ_h (⟨K⃗ ∇q_h ·n⃗, ψ_h ⟩_F; + ⟨q_h, K⃗ ∇ψ_h ·n⃗ ⟩_F) + ∑_F∈ℱ_h γ/h_F ⟨q_h , ψ_h⟩_F , ] for Q_h = Q_h^l,disc ,; ]. where, for w⃗∈H⃗^1/2(Γ^D_u⃗) and q∈ H^1/2(Γ^D_p), a_γ(w⃗ ,χ⃗_h) := - ⟨w⃗, C⃗ε⃗(χ⃗_h) n⃗ ⟩_Γ^D_u⃗ + γ_a/h_F ⟨w⃗, χ⃗_h ⟩_Γ^D_u⃗ , b_γ(q, ψ_h) := - ⟨q, K⃗ ∇ψ_h ·n⃗⟩_Γ_p^D + γ_b/h_F ⟨q, ψ_h ⟩_Γ^D_p . The second of the options in (<ref>) amounts to a symmetric interior penalty discontinuous Galerkin discretization of the scalar variable p; cf., e.g., <cit.>. As usual, the average · and jump · of a function w∈ L^2(Ω) on an interior face F between two elements K^+ and K^-, such that F=∂ K^+ ∩∂ K^-, are w := 1/2 (w^++ w^-) and w : = w^+ - w^- . For boundary faces F ⊂∂ K ∩∂Ω, we set w:= w_|K and w := w_|K. The set of all faces (interior and boundary faces) on the multigrid level 𝒯_l is denoted by ℱ_h. In the second of the options in (<ref>), the parameter γ of the last term has to be chosen sufficiently large, such that discrete coercivity on Q_h of B_γ is preserved. The local length h_F is chosen as h_F = h_F := 1/2 (|K^+|_d + |K^-|_d) with Hausdorff measure |· |_d; cf. <cit.>. For boundary faces we set h_F := |K|_d. In (<ref>), the quantities γ_a and γ_b are the algorithmic parameters of the stabilization terms in the Nitsche formulation. To ensure well-posedness of the discrete systems the parameters γ_a and γ_b have to be chosen sufficiently large; cf. Appendix <ref>. Based on our numerical experiments we choose the algorithmic parameter γ, γ_a and γ_b in (<ref>) and (<ref>) as γ_a = 5· 10^4 · r · (r+1) and γ = γ_b = 1/2· r · (r-1) , where r is the polynomial degree of the finite element space (<ref>) for the displacement variable. Finally, for given f⃗∈H⃗^-1(Ω), u⃗_D ∈H⃗^1/2(Γ^D_u⃗), t⃗_N ∈H⃗^-1/2(Γ^N_u⃗) and g∈ H^-1(Ω), p_D ∈ H^1/2(Γ^D_p), p_N∈ H^-1/2(Γ^N_p) for Q_h = Q_h^l,cont, and properly fitted assumptions about the data for Q_h = Q_h^l,disc, we put F_γ(χ⃗_h) := ⟨f⃗ , χ⃗_h⟩- ⟨t⃗_N,χ⃗_h ⟩_Γ^N_u⃗ + a_γ(u⃗_D ,χ⃗_h) , G_γ(ψ_h) := {[ ⟨⟨g,ψ_h ⟩- α⟨v⃗_D ·n⃗ , ψ_h ⟩_Γ^D_u⃗ - ⟨p_N, ψ_h⟩_Γ^N_p + b_γ(p_D, ψ_h) , for Q_h = Q_h^l,cont ,; [ ⟨g,ψ_h ⟩- ∑_F∈ℱ_h^D,u⃗ α⟨v⃗_D ·n⃗ , ψ_h ⟩_F - ∑_F∈ℱ_h^D,p⟨p_D, K⃗ ∇ψ_h ·n⃗ ⟩_F; + ∑_F∈ℱ_h^D,p γ/h ⟨p_D, ψ_h⟩_F - ∑_F⊂ℱ_h^N,p ⟨p_N, ψ_h ⟩_F , ] for Q_h = Q_h^l,disc . ]. In the second of the options in (<ref>), we denote by ℱ_h^D,p⊂ℱ_h and ℱ_h^N,p⊂ℱ_h the set of all element faces on the boundary parts Γ_p^D and Γ_p^N, respectively; cf. (<ref>). The second of the terms on the right-hand side of (<ref>), with v⃗_D = ∂_t u⃗_D, is added to ensure consistency of the form (<ref>) in the fully discrete formulation (<ref>) of (<ref>), i.e., that the discrete equation (<ref>) is satisfied by the continuous solution to (<ref>). We use a temporal test basis that is supported on the subintervals I_n; cf., e.g., <cit.>. Then, a time marching process is obtained. In that, we assume that the trajectories u⃗_τ,h, v⃗_τ,h and p_τ,h have been computed before for all t∈ [0,t_n-1], starting with approximations u⃗_τ,h(t_0) :=u⃗_0,h, v⃗_τ,h(t_0) :=u⃗_1,h and p_τ,h(t_0) := p_0,h of the initial values u⃗_0, v⃗_0 and p_0,h. Then, we consider solving the following local problem on I_n. [Numerically integrated I_n-problem] For given u⃗_h^n-1:= u⃗_τ,h(t_n-1)∈V⃗_h, v⃗_h^n-1:= [4]v⃗_τ,h(t_n-1)∈V⃗_h, and p_h^n-1:= p_τ,h(t_n-1) ∈ Q_h with u⃗_τ,h(t_0) :=u⃗_0,h, v⃗_τ,h(t_0) :=u⃗_1,h and p_τ,h(t_0) := p_0,h, find (u⃗_τ,h,v⃗_τ,h,p_τ,h) ∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h) such that Q_n (⟨∂_t u⃗_τ,h , ϕ⃗_τ,h ⟩- ⟨v⃗_τ,h , ϕ⃗_τ,h ⟩) + ⟨u⃗^+_τ,h(t_n-1), ϕ⃗_τ,h^+(t_n-1)⟩= ⟨u⃗_h^n-1, ϕ⃗_τ,h^+(t_n-1)⟩ , Q_n (⟨ρ∂_t v⃗_τ,h , χ⃗_τ,h ⟩+ A_γ(u⃗_τ,h, χ⃗_τ,h ) + C(χ⃗_τ,h,p_τ,h)) + ⟨ρv⃗^+_τ,h(t_n-1), χ⃗_τ,h^+(t_n-1)⟩ = Q_n (F_γ(χ⃗_τ,h)) + ⟨ρv⃗_h^n-1, χ_τ,h^+(t_n-1)⟩ , Q_n (⟨c_0 ∂_t p_τ,h,ψ_τ,h ⟩- C(v⃗_τ,h,ψ_τ,h)+ B_γ(p_τ,h, ψ_τ,h)) + ⟨c_0 p^+_τ,h(t_n-1), ψ_τ,h^+(t_n-1)⟩ = Q_n ( G_γ(ψ_τ,h)) + ⟨c_0 p_h^n-1, ψ_τ,h^+(t_n-1)⟩ for all (ϕ⃗_τ,h,χ⃗_τ,h,ψ_τ,h)∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h). The trajectories defined by Problem <ref>, for n = 1,…,N, satisfy that u⃗_τ,h,v⃗_τ,h∈ Y_τ^k(V⃗_h) and p_τ,h∈ Y_τ^k(Q_h). The quadrature formulas on the left hand-side of (<ref>) can be rewritten by time integrals since the Gauss–Radau formula (<ref>) is exact for all polynomials w∈ P_2k(I_n;). Well-posedness of Problem <ref> is ensured. Problem <ref> admits a unique solution. We prove Lem. <ref> for Q_h=Q_h^l,cont only, thus assuming the first of the options in (<ref>) and (<ref>). For Q_h=Q_h^l,disc, the proof can be done similarly by using, in addition, standard techniques of error analysis for discontinuous Galerkin methods; cf., e.g., <cit.>. Since Problem <ref> is finite-dimensional, it suffices to prove uniqueness of the solution. Existence of the solution then follows directly from its uniqueness. Let now (u⃗^(1)_τ,h,v⃗^(1)_τ,h,p^(1)_τ,h) ∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h) and (u⃗^(2)_τ,h,v⃗^(2)_τ,h,p^(2)_τ,h) ∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h) denote two triples of solutions to (<ref>). Their differences (u⃗_τ,h,v⃗_τ,h,p_τ,h) =(u⃗^(1)_τ,h,v⃗^(1)_τ,h,p^(1)_τ,h) -(u⃗^(2)_τ,h,v⃗^(2)_τ,h,p^(2)_τ,h) then satisfy the equations Q_n (⟨∂_t u⃗_τ,h , ϕ⃗_τ,h ⟩- ⟨v⃗_τ,h , ϕ⃗_τ,h ⟩) + ⟨u⃗^+_τ,h(t_n-1), ϕ⃗_τ,h^+(t_n-1)⟩= 0 , Q_n (⟨ρ∂_t v⃗_τ,h , χ⃗_τ,h ⟩+ A_γ(u⃗_τ,h, χ⃗_τ,h ) + C(χ⃗_τ,h,p_τ,h)) + ⟨ρv⃗^+_τ,h(t_n-1), χ⃗_τ,h^+(t_n-1)⟩= 0 , Q_n (⟨c_0 ∂_t p_τ,h,ψ_τ,h ⟩- C(v⃗_τ,h,ψ_τ,h)+ B_γ(p_τ,h, ψ_τ,h)) + ⟨c_0 p^+_τ,h(t_n-1), ψ_τ,h^+(t_n-1)⟩= 0 for all (ϕ⃗_τ,h,χ⃗_τ,h,ψ_τ,h)∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h). We let A⃗_γ: V⃗_h ↦V⃗_h be the discrete operator that is defined, for w⃗_h ∈V⃗_h and all ϕ⃗_h∈V⃗_h, by ⟨A⃗_γw⃗_h , ϕ⃗_h ⟩ = A_γ (w⃗_h,ϕ⃗_h) . In (<ref>) we choose ϕ⃗_τ,h=A⃗_γu⃗_τ,h, χ⃗_τ,h=v⃗_τ,h and ψ_τ,h=p_τ,h. Adding the resulting equations yields that Q_n (⟨∂_t u⃗_τ,h , A⃗_γu⃗_τ,h⟩ + ⟨ρ∂_t v⃗_τ,h , v⃗_τ,h⟩ + ⟨ c_0 ∂_t p_τ,h, p_τ,h⟩ + B_γ(p_τ,h,p_τ,h) ) + ⟨u⃗^+_τ,h(t_n-1), A⃗_γ u_τ,h^+(t_n-1)⟩ + ⟨ρv⃗^+_τ,h(t_n-1), v⃗_τ,h^+(t_n-1)⟩ + ⟨ c_0 p^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ = 0 . Recalling the exactness of the Gauss–Radau formula (<ref>) for w∈ℙ_2k(I_n;), Eq. (<ref>) yields that 1/2∫_t_n-1^t_nd/dt( ⟨A⃗_γu⃗_τ,h , u⃗_τ,h⟩ + ⟨ρv⃗_τ,h , v⃗_τ,h⟩ + ⟨ c_0 p_τ,h, p_τ,h⟩ ) t + Q_n(B_γ(p_τ,h,p_τ,h) ) + ⟨u⃗^+_τ,h(t_n-1), A⃗_γ u_τ,h^+(t_n-1)⟩ + ⟨ρv⃗^+_τ,h(t_n-1), v⃗_τ,h^+(t_n-1)⟩ + ⟨ c_0 p^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ = 0 . Using (<ref>), this shows that A_γ (u⃗_τ,h(t_n) , u⃗_τ,h(t_n)) + ⟨ρv⃗_τ,h (t_n), v⃗_τ,h(t_n) ⟩ + ⟨ c_0 p_τ,h(t_n), p_τ,h(t_n)⟩ + 2 Q_n (B_γ(p_τ,h,p_τ,h) ) + A_γ(u⃗^+_τ,h(t_n-1), u⃗_τ,h^+(t_n-1)) + ⟨ρv⃗^+_τ,h(t_n-1), v⃗_τ,h^+(t_n-1)⟩ + ⟨ c_0 p^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ = 0 . From (<ref>) along with the discrete coercivity properties (<ref>) of A_γ and (<ref>) of B_γ we directly deduce that u⃗_τ,h(t_n)= u⃗_τ,h^+(t_n-1)=0⃗ , v⃗_τ,h(t_n)= v⃗_τ,h^+(t_n-1)=0⃗ and p_τ,h(t_n)=p_τ,h^+(t_n-1)= 0 , as well as p_τ,h(t_n,μ^GR) = 0 , for μ = 1,…, k+1 . We note that t_n,k+1^GR=t_n. Relation (<ref>) now implies that p_τ,h≡ 0 on I_n. To prove that u⃗_τ,h≡0⃗ and v⃗_τ,h≡0⃗, by (<ref>) it is sufficient to show that u⃗_τ,h(t_n,μ^G)=0⃗ and v⃗_τ,h(t_n,μ^G) = 0⃗, for μ = 1,…, k, where t_n,μ^G, for μ = 1,…, k, are the nodes of the k-point Gauss quadrature formula on I_n that is exact for all polynomials in ℙ_2k-1(I_n;). This is done now. Recalling (<ref>), we conclude from (<ref>) by a suitable choice of test functions that ∂_t u⃗_τ,h(t_n,μ^GR)=v⃗_τ,h(t_n,μ^GR) , for μ = 1,…, k . Next, choosing χ_τ,h=v⃗_τ,h in (<ref>) and recalling (<ref>) imply that Q_n (⟨ρ∂_t v⃗_τ,h , v⃗_τ,h⟩ + A_γ(u⃗_τ,h, ∂_t u⃗_τ,h ) )= 0 By the exactness of the Gauss–Radau formula (<ref>) for all w∈ℙ_2k(I_n;) we have from (<ref>) that ∫_t_n-1^t_n⟨ρ∂_t v⃗_τ,h , v⃗_τ,h⟩ t + 1/2∫_t_n-1^t_nd/dt A_γ (u⃗_τ,h , u⃗_τ,h ) t = 0 . The second of the terms in (<ref>) vanishes by (<ref>). The stability result of <cit.> then implies that v⃗_τ,h(t_n,μ^G) = 0⃗ , for μ = 1,…, k . From (<ref>) along with (<ref>) we then deduce that v⃗_τ,h≡0⃗ on I_n. Choosing the test function ϕ⃗_τ,h=u⃗_τ_h in (<ref>), using v⃗_τ,h≡0⃗ and applying the stability result <cit.>, it follows that u⃗_τ,h(t_n,μ^G) = 0⃗ , for μ = 1,…, k . From (<ref>) along with (<ref>) we then have that u⃗_τ,h≡0⃗ on I_n. Thus uniqueness of solutions to (<ref>) and, thereby, well-posedness of Problem <ref> is thus ensured. In Appendix <ref> an alternative formulation for the system (<ref>) is still presented. It is based on using the time derivative ∂_t u⃗_τ,h of the primal variable u⃗_τ,h instead of the auxiliary variable v⃗_τ,h in (<ref>). In this case, an additional surface integral has to be included; cf. Eq. (<ref>) . § ALGEBRAIC SOLVER BY GEOMETRIC MULTIGRID PRECONDITIONED GMRES ITERATIONS On the algebraic level, the variational problem (<ref>) leads to linear systems of equations with complex block structure, in particular if higher order (piecewise) polynomial degrees k for the approximation of the temporal variable are involved. This demands for a robust and efficient linear solver, in particular in the three-dimensional case Ω⊂^3. For solving (<ref>) we consider using flexible GMRES iterations <cit.> that are preconditioned by a V-cycle geometric multigrid method (GMG) based on a local Vanka smoother. In <cit.>, the combined GMG–GMRES approach is further embedded in a Newton iteration for solving space-time finite element discretizations of the Navier–Stokes system. Thus, nonlinear extensions of the prototype model (<ref>) are expected to become feasible by our approach as well. To derive the algebraic form of (<ref>), the discrete functions u⃗_τ,h, v⃗_τ,h and p_τ,h are represented in a Lagrangian basis {χ_n,m}_m=1^k+1⊂ℙ_k(I_n;) with respect to the (k+1) Gauss–Radau quadrature points of I_n, such that u⃗_τ,h_|I_n(x⃗,t) = ∑_m=1^k+1 u⃗_n,m(x⃗) χ_n,m(t) , v⃗_τ,h_|I_n(x⃗,t) = ∑_m=1^k+1 v⃗_n,m(x⃗) χ_n,m(t) , p_τ,h_|I_n(x⃗,t) = ∑_m=1^k+1 p_n,m(x⃗) χ_n,m(t) . The resulting coefficient functions (u⃗_n,m,v⃗_n,m,p_n,m)∈V⃗_h×V⃗_h× Q_h, for m=1,…,k+1, are developed in terms of the finite element basis of V⃗_h and Q_h, respectively. Letting V⃗_h= span{ψ⃗_1,…,ψ⃗_R } and Q_h=span{ξ_1,…,ξ_S }, we get that u⃗_n,m(x⃗) = ∑_r=1^R u^(r)_n,mψ⃗_r(x⃗) , v⃗_n,m(x⃗) = ∑_r=1^R v^(r)_n,mψ⃗_r(x⃗) and p_n,m(x⃗) = ∑_s=1^S p^(s)_n,m ξ_s(x⃗) . For the coefficients of the expansions in (<ref>) we define the subvectors U⃗_n,m = (u^(1)_n,m,…,u^(R)_n,m )^⊤ , V⃗_n,m= (v^(1)_n,m,…,v^(R)_n,m )^⊤ and P⃗_n,m = (p^(1)_n,m,…,p^(S)_n,m )^⊤ , for m=1,…,k+1 , of the degrees of freedom for all Gauss–Radau quadrature points and the global solution vector on I_n by X⃗_n^⊤ = ((V⃗_n,1)^⊤,(U⃗_n,1)^⊤,(P⃗_n,1)^⊤,…, (V⃗_n,k+1)^⊤,(U⃗_n,k+1)^⊤,(P⃗_n,k+1)^⊤) . We note that X⃗_n comprises the (spatial) degrees of freedom for all (k+1) Gauss–Radau nodes, representing the Lagrange interpolation points in time , of the subinterval I_n. The approximations at these time points will be computed simultaneously. Substituting (<ref>) and (<ref>) into (<ref>) and choosing in (<ref>) the test basis {χ_n,mψ⃗_r,χ_n,mψ⃗_r,χ_n,mξ_s}, for m=1,…,k+1, r=1,…, R and s=1,…,S, built from the trial basis in (<ref>) and (<ref>), we obtain the following algebraic system. [Algebraic I_n-problem] For the vector X⃗_n, defined in (<ref>) along with (<ref>), of the coefficients of the expansions (<ref>) solve A⃗_n X⃗_n = F⃗_n , where the matrix A⃗_n exhibits the (k+1)× (k+1) block structure A⃗_n = (A⃗_a,b)_a,b=1^k+1 with block submatrices A⃗_a,b, for a,b=0,…,k, defined by A⃗_a,b = ([ - M⃗^0,V⃗_h_a, b M⃗^1,V⃗_h_a,b 0⃗; M⃗^1,V⃗_h_a,b A⃗_a,b + N⃗^A_a,b C⃗^⊤_a,b; 0⃗ - C⃗_a,b M⃗^1,Q_h_a,b + B⃗_a,b + N⃗^B_a,b ]) . For the choice Q_h^l=Q_h^l,cont in (<ref>), the explicit representation of the submatrices in (<ref>) reads as (M⃗^1,V⃗_h_a,b)_i,j := Q_n(⟨ρ∂_t χ_n,aψ⃗_j, χ_n,bψ⃗_i⟩) + ⟨ρχ_n,a(t_n-1^+) ψ⃗_j, χ_n,b(t_n-1^+) ψ⃗_i⟩ , (M⃗^0,V⃗_h_a,b)_i,j := Q_n (⟨ρχ_n,aψ⃗_j, χ_n,bψ⃗_i⟩) , (A⃗_a,b)_i,j : = Q_n (⟨C⃗ε⃗(χ_n,aψ⃗_j), ε⃗(χ_n,bψ⃗_i) ⟩ + ⟨C⃗ε⃗(χ_n,aψ⃗_j) n⃗ , χ_n,bψ⃗_i⟩_Γ^D_u⃗) , (N⃗^A_a,ψ⃗_b)_i,j := Q_n (a_γ(χ_n,aψ⃗_i,χ_n,bψ⃗_i)) as well as (C⃗_a,b)_r,j := Q_n (-α⟨∇· (χ_n,bψ⃗_j), χ_n,aξ_r ⟩ + α⟨χ_n,bψ⃗_j·n⃗ , χ_n,aξ_r ⟩_Γ^D_u⃗) , (M⃗^1,Q_h_p_a,p_b)_r,s := Q_n(⟨ c_0 ∂_t χ_n,aξ_s, χ_n,bξ_r ⟩) + ⟨ c_0 χ_n,a(t_n-1^+) ξ_s, χ_n,b(t_n-1^+) ξ_r⟩) and (B⃗_a,b)_r,s := Q_n (⟨K⃗∇ (χ_n,aξ_s), ∇ (χ_n,bξ_r) ⟩ - ⟨K⃗∇ (χ_n,aξ⃗_s) ·n⃗ , χ_n,bξ_r⟩_Γ^D_p) , (N⃗^B_a,b)_r,s := Q_n (b_γ(χ_n,aξ_s,χ_n,bξ_r)) , with a_γ(·,·) and b_γ(·,·) being defined in (<ref>) and i,j=1,…, R and r,s=1,… , S. The right-hand side vector F⃗_n in (<ref>) is defined similarly, according to (<ref>) along with (<ref>). Its definition is skipped here for brevity. We note that (<ref>) is still multiplied by ρ>0 for the definition of the first row in (<ref>). For the family of finite element pairs, corresponding to Q_h^l=Q_h^l,disc in (<ref>), the definition of the submatrix B_a,b has to be adjusted to the second of the options in (<ref>). Further, the contribution N⃗^B_a,b has to be omitted. Increasing values of the piecewise polynomial degree in time k enhance the complexity of the block structure of the system matrix A⃗_n in (<ref>) along with (<ref>). They impose an additional facet of challenge on the construction of efficient block preconditioners for (<ref>). We solve the linear system (<ref>) for the unknown X⃗_n on the subinterval I_n by flexible GMRES iterations <cit.> that are preconditioned by a V-cycle geometric multigrid (GMG) algorithm <cit.>. Key ingredients of the GMRES–GMG approach are * left preconditioning of each GMRES iteration by a single V-cycle GMG iteration, * prolongation to grids by interpolation and application of the transpose operator for restriction to grids, * J_max pre- and post-smoothing steps on each grid level {𝒯_l} for l=1,…,L and * application of a (parallel) direct solver on the coarsest mesh partition 𝒯_0. In our computational studies presented in Sec. <ref>, we use J_max=4 and the parallel direct solver SuperLU_DIST <cit.>. For the implementation of the restriction and prolongation operators the deal.II classes MultiGrid and MGTransferPrebuilt are used. For the deal.II finite element library we refer to <cit.>. For details of the parallel implementation by the message passing interface (MPI) protocol of our GMG approach we refer to <cit.>. The choice of the smoothing operator in the GMG method is extremely diverse; cf., e.g., <cit.> for a further discussion. We use a collective smoother of Vanka type that is based on the solution of small local problems. Compared to the Navier–Stokes system <cit.>, the dynamic Biot problem (<ref>) requires adaptations in the construction of the local Vanka smoother. In inf-sup stable discretizations of the Navier–Stokes equations by the ℚ_r^d/ℙ_r-1^disc, r≥ 2, family of finite element spaces no coupling between the pressure degrees of freedom is involved due to the discontinuous approximation of the pressure variable. This feature leads to excellent performance properties of the Vanka smoother <cit.>. In contrast to this, the discretization (<ref>) of (<ref>) by ℙ_r-1^disc, r≥ 2, elements involves a coupling between degrees of freedom of the scalar variable p due to the presence of the face integrals over the average and jump in the second of the options in (<ref>) for the definition of the bilinear form B_γ. This coupling reduces the smoothing properties of elementwise Vanka operator. For the ℚ_r-1, r≥ 2, family of elements for the scalar variable p, leading to the first of the options in (<ref>), the coupling of degrees of freedom of p by its continuous in space approximation and the loss of smoothing properties arise likewise. As a remedy, for both families of inf-sup stable approximation in (<ref>) the local Vanka smoother is computed on overlapping patches of adjacent elements. In addition, an averaging of the patchwise updates is done after the Vanka smoother has been applied on all of them. To construct the patchwise Vanka, we let on the mesh partition 𝒯_l, for l=1,…, L, of the multigrid hierarchy the linear system, to be solved, be represented by A⃗_l d⃗_l = b⃗_l , for l=1,…, L . To each grid node ξ⃗_l^m, for m = 1 ,… , M_l, where M_l denotes the total number of grid nodes on the mesh partition 𝒯_l, we built a patch of adjacent elements such that P_l^m := ⋃{ K ∈𝒯_l |ξ_l^m ∈K} , for m=1,…, M_l . In two space dimensions P_l^m is built from four elements, if ξ_l^m ∉∂Ω. In three space dimensions, P_l^m has eight elements, if ξ_l^m ∉∂Ω. If ξ_l^m ∈∂Ω, patches of less elements are obtained. On 𝒯_l, let Z_l denote the index set of all global degrees of freedom with cardinality C_l, C_l := card(Z_l) . Let Z_l(P_l^m) denote the subset of Z_l all global degrees of freedom linked to the patch P_l^m, i.e. degrees of freedom of u⃗_τ,h, v⃗_τ,h and p_τ,h for the (k+1) Gauss–Radau time points of I_n. The cardinality of Z_l(P_l^m) is denoted by C_l^m, C_l^m : = card(Z_l(P_m^l)) , for m=1,…, M_l . Further, we denote the index set of all local degrees of freedom on P_l^m by Ẑ_l(P_m^l) :={0,…,C_l^m-1}. For the notation, we note that the set Ẑ_l(P_m^l) depends on the cardinality of the patch P_l^m. For a given patch P_l^m and a local degree of freedom with number μ̂∈Ẑ_l(P_m^l) let the mapping dof:𝒯_l×Ẑ_l(P_m^l) → Z_l , μ = dof(P_l^m,μ̂)∈ Z_l(P_l^m) , yield the uniquely defined global number μ∈ Z_l. Finally, we put R = dim V⃗_h and S = dim Q_h. We are now in a position to define the local Vanka operator for the patch P_l^m in an exact mathematical way; cf., e.g., <cit.>. [Patchwise Vanka smoother] On patch P_l^m, for m∈{1,…,M_l}, the local Vanka operator S⃗_P_l^m: ^(k+1) · (2R+S)→^C^l_m is defined by S⃗_P_l^m(d⃗) = R⃗_ P_l^md⃗ + ω A⃗_P_l^m^-1 R⃗_P_l^m (b⃗_l-A⃗_l d⃗) , with underrelaxation factor ω>0 and P_l^m-local restriction operator R⃗_K:^(k+1)· (2R+S)→^C_l^m defined by (R⃗_P_l^md⃗)[μ̂] = d⃗[dof(P_l^m,μ̂)] , for μ̂∈Ẑ_l(P_l^m) . Further, the patch system matrix A⃗_P_l^m∈^C_l^m,C_l^m in (<ref>) is defined by A⃗_P_l^m[ν̂][μ̂] := A⃗[dof(P_l^m,ν̂)][dof(P_l^m,μ̂)] , for ν̂,μ̂∈Ẑ_l(P_l^m) . In the numerical experiments of Sec. <ref> we choose ω = 0.7. The P_l^m-local restriction operator (<ref>) assigns to a global defect vector d⃗∈^(k+1)· (2R+S) the local block vector R⃗_K d⃗∈^C_l^m that contains all components of d⃗ that are associated with all degrees of freedom (for all (k+1) Gauss–Radau points of I_n) belonging to the patch P_l^m. For the computation of the inverse (A⃗_P_l^m)^-1 of the patch system matrixA⃗_P_l^m in (<ref>) we use LAPACK routines. The application of the smoother for (<ref>) on the mesh partition 𝒯_l is summarized in Algorithm <ref>. In line <ref> of Alg. <ref> the defect and solution vector is pre-initialized with 0⃗. In line <ref> the loop over all J_max smoothing steps starts. In line <ref> and <ref> the counter vector p⃗ for the number of updates of the degrees of freedom and the auxiliary vector z⃗ are initialized with 0⃗. In line <ref> the loop over all patches P_l^m, for m=1,…,M, starts. In line <ref> the local Vanka smoother is applied on patch P_l^m to the current iterate d⃗ of the defect vector and the image is stored in the local patch vector y⃗. In line <ref> the local vector y⃗ is assigned to an auxiliary global vector z⃗. In line <ref> the components of the counter p⃗ are incremented for the (global) indices associated with the degrees of freedom of the patch P_l^m processed in the loop. Finally, in line <ref> the arithmetic mean of the local (patchwise) updates z⃗ is assigned to the update of the global defect vector d⃗. In line <ref>, the P_l^m-dependent extension operator E⃗_P_l^m:^C_l^m→^(k+1)· (2R+S) is defined by (E⃗_P_l^my⃗)[μ] = {[ y[μ̂] , if ∃μ̂∈Ẑ_l(P_l^m): μ = dof(K,μ̂) ,; 0 , if μ∉Z_l(P_l^m) . ]. Regarding the performance of Alg. <ref> and the overall GMRES–GMG approach we note the following. * Averaging of the patchwise updates implemented in line <ref> and <ref> of Alg. <ref>, that is used instead of overwriting successively the (global) degrees of freedom within the patch loop starting in line <ref>, is essential and ensures the convergence and efficiency of the local Vanka smoother and, thereby, the desired performance of the overall GMRES–GMG linear solver. Without the averaging step we encountered convergence problems of the GMRES–GMG solver for the experiments of Sec. <ref>. * The same holds for the application of the Vanka smoother on the patches (<ref>) instead of using an elementwise Vanka smoother on the element K. The latter would lead to systems (<ref>) of smaller dimension, however causes convergence problems for the GMRES–GMG solver. This is expected to be due coupling of the degrees of freedom of the scalar variable p in the spatial discretizations used here,. § NUMERICAL STUDIES In this section we study numerically the proposed space-time finite element and GMRES–GMG solver approach with respect to its computational and energy efficiency. Firstly, we demonstrate the accuracy of solutions in terms of convergence rates. Further, we illustrate the robustness of the GMRES–GMG solver with respect to the resolution in space and time and the polynomial approximation order. This is done for a challenging three-dimensional test problem of practical interest, in order to demonstrate the feasibility of large scale three-dimensional simulations. The numerical convergence of goal quantities of the benchmark is addressed. The parallel scaling properties of our implementation are also investigated. Beyond these studies of classical performance engineering, the energy efficiency of the approach is considered further. The implementation of the numerical scheme and the GMRES–GMG solver was done in an in-house high-performance frontend solver for the library <cit.>. For details of the parallel implementation of the geometric multigrid solver we refer to <cit.>. In all numerical experiments, the stopping criterion for the GMRES iterations is an absolute residual smaller than . The computations were performed on a Linux cluster with 571 nodes, each of them with 2 CPUs and 36 cores per CPU. The CPUs are Intel Xeon Platinum 8360Y with a base frequency of 2.4, a maximum turbo frequency of 3.5 and a level 3 cache of 54. Each node has 252 of main memory. §.§ Accuracy of the discretization: Experimental order of convergence We study (<ref>) for Ω=(0,1)^2 and I=(1,2] and the prescribed solution u( x, t) = ϕ( x, t) E_2 and p( x, t) = ϕ( x, t) with ϕ( x, t) = sin(ω_1 t^2) sin(ω_2 x_1) sin(ω_2 x_2) and ω_1=ω_2 = π. We put ρ=1.0, α=0.9, c_0=0.01 and K= E_2 with the identity E⃗_2∈^2,2. For the fourth order elasticity tensor C, isotropic material properties with Young's modulus E=100 and Poisson's ratio ν=0.35, corresponding to the Lamé parameters λ = 86.4 and μ = 37.0, are chosen. For an experiment with larger values of λ and μ we refer to Table <ref> in the appendix. In our numerical experiments, the norm of L^∞(I;L^2) is approximated by computing the function's maximum value in the Gauss quadrature nodes t_n,m of I_n, i.e., w_L^∞(I;L^2)≈max{ w_|I_n(t_n,m)| m=1,… ,M , n=1,…,N} , with M=100 . We study the space-time convergence behavior of the scheme (<ref>). For this, the domain Ω is decomposed into a sequence of successively refined meshes of quadrilateral finite elements. The spatial and temporal mesh sizes are halved in each of the refinement steps. The step sizes of the coarsest space and time mesh are h_0=1/(2√(2)) and τ_0=0.1. We choose the polynomial degree k=2 and r=3, such that discrete solutions u⃗_τ,h, v⃗_τ∈ Y_τ^2(V⃗_h), p_τ,h∈ Y_τ^2(Q_h) with local spaces ℚ_3^2/ P_2^disc are obtained, as well as k=3 and r=4 with u⃗_τ,h, v⃗_τ∈ Y_τ^3(V⃗_h) and p_τ,h∈ Y_τ^3(Q_h) and local spaces ℚ_4^2/ℙ_3^disc ; cf. (<ref>), (<ref>) and (<ref>). The calculated errors and corresponding experimental orders of convergence are summarized in Table <ref> and <ref>, respectively. The error is measured in the quantities associated with the energy of the system (<ref>); cf. <cit.> and <cit.>. Table <ref> and <ref> nicely confirm the optimal rates of convergence with respect to the polynomial degrees in space and time of the overall approach. A notable difference in the convergence behavior between the pairs ℚ_r^2/ℙ_r-1^disc and ℚ_r^2/ℚ_r-1 of local spaces for the discretization of the spatial variables is not observed. For completeness, we summarize in Appendix <ref> the convergence results obtained for the pair ℚ_r^2/ℚ_r-1 of spaces of the Taylor–Hood family. A minor superiority of the pair ℚ_r^2/ℙ_r-1^disc over the pair ℚ_r^2/ℚ_r-1 is only seen in the approximation of the scalar variable p. The coincidence of the convergence results also holds for the application of Problem <ref> instead of Problem <ref>. The two discrete problems differ from each other by the discretization of the term α∇·∂_t u⃗ in (<ref>). Next, we show numerically that the time discretization is even superconvergent of order 2k+1 in the discrete time nodes t_n, for n=1,…, N. For this, we introduce the time mesh dependent norm w_l^∞(L^2) := max{ w(t_n)| n=1,…, N} . We prescribe the solution u( x, t) = [ -2 (x-1)^2 x^2 (y-1) y (2 y-1) sin (ω_1 t); -2 (x-1) x (2 x-1) (y-1)^2 y^2 sin (ω_1 t) ] and p( x, t) = -2 (x-1)^2 x^2 (y-1) y (2 y-1) sin (ω_2 t) with ω_1=40·π and ω_2 = 10·π. We put ρ=1.0, α=0.9, c_0=0.01 and K= E_2 with the identity E⃗_2∈^2,2. For the elasticity tensor C, isotropic material properties with Young's modulus E=100 and Poisson's ratio ν=0.35 are used. For the local spaces we choose the pair ℚ_5^2/ℚ_4 such that, for any t∈ [0,T], the solution (<ref>) belongs to the discrete spaces V⃗_h and Q_h, respectively, and its spatial approximation is exact. This simplification is done here since we aim to study the convergence of the temporal discretization only. In the experiment, we choose k=2 such that discrete solutions u⃗_τ,h, v⃗_τ∈ Y_τ^2(V⃗_h) and p_τ,h∈ X_τ^2(Q_h) are obtained. We use a spatial mesh that consists of 16 cells with h=1/(2√(2)) and set τ_0=0.02. The calculated errors and corresponding experimental orders of convergence are summarized in Table <ref>. Superconvergence of order 2k+1 in the discrete time nodes is clearly observed in the second of the tables in Table <ref>. §.§ Computational efficiency: Accuracy of goal quantities and convergence of the GMRES–GMG solver in a 3d benchmark For a three-dimensional problem configuration of practical interest (cf. <cit.>), we study the potential of the proposed approach to compute reliably and efficiently goal quantities of physical interest that are defined in (<ref>). Concurrently, we analyze the convergence and performance properties of the GMRES–GMG solver of Sec. <ref> for our families of space-time finite element methods. More precisely, we consider the L-shaped domain Ω⊂^3 sketched in Fig. <ref> along with the boundary conditions for u⃗ prescribed on the different parts of ∂Ω. Beyond the boundary conditions (<ref>) and (<ref>), we further apply (homogeneous) directional boundary conditions on parts of the boundary for the variable u⃗. This is done for the sake of physical realism. For the portion Γ^d_u⃗⊂∂Ω they read as u⃗·n⃗ = 0 and (σ⃗(u⃗)n⃗) ·t⃗_i = 0 , for i=1,…, d-1 , on Γ^d_u⃗× (0,T] , for the stress tensor σ⃗(u⃗) = C⃗ε⃗(u⃗) and the unit basis vectors t⃗_i, for i=1,…,d-1, of the tangent space at x⃗∈Γ^d_u⃗. In the definition of A_γ in (<ref>) and (<ref>), the conditions (<ref>) still need to be implemented properly. For instance, for the left boundary of the L-shaped domain Ω the first of the equations in (<ref>) implies that u⃗_τ,h·n⃗ = u_τ,h,1 = 0 , for u⃗_τ,h=(u_τ,h,1,u_τ,h,2)^⊤ . By the second of the conditions in (<ref>) we get for the second of the terms on the right-hand of (<ref>) that ⟨C⃗ε(u⃗_τ,h)n⃗, χ⃗_τ,h⟩_Γ_u⃗^l = ⟨C⃗ε(u⃗_τ,h) n⃗·n⃗,n⃗·χ⃗_τ,h⟩_Γ_u⃗^d . For the boundary part Γ^d_u⃗ we then put, similarly to (<ref>), a^d_γ (w⃗,χ⃗_h) := - ⟨w⃗·n⃗, C⃗ε(χ⃗_h) n⃗·n⃗⟩_Γ^d_u⃗ + γ_a h^-1⟨w⃗·n⃗, χ⃗_h ·n⃗⟩_Γ^d_u⃗ , while leaving a_γ unmodified for the part Γ^D_u⃗ where Dirichlet boundary conditions are prescribed for u⃗. In its entirety, we thus have that a_γ(·,·) := a_γ^D (·,·)+a_γ^d (·,·) with a_γ^D (·,·) being defined by the right-hand side of (<ref>). By the arguments of Appendix <ref>, the coercivity of the resulting form A_γ is still ensured. On the right boundary we prescribe a homogeneous Neumann condition for u⃗ with t⃗_N=0⃗ in (<ref>). On the left part of the upper boundary, i.e. for Γ^N_u⃗:=(0,0.5)×{1}× (0,0.5), we impose the traction force (cf. (<ref>)) t⃗_N = [ 0; 5 · 10^9· (32 x z-18 x-16 z+10) sin (8 π t); 0 ] . For the scalar variable p we prescribe a homogeneous Dirichlet condition (<ref>) on the left upper part of ∂Ω, i.e., for (x_1,x_2,x_3)∈Γ_p^D:=[0,0.5]×{1}× [0,0.5]. On Γ_p^N:=∂Ω\Γ_p^D a homogeneous Neumann condition with p_N=0 in (<ref>) is prescribed. We put ρ=1.0, α=0.9, c_0=0.01 and K= E_3 with the identity E⃗_3∈^3,3. For the elasticity tensor C, isotropic material properties with Young's modulus E=20000 and Poisson's ratio ν=0.3, that model the material behavior of human bone (cf. <cit.>), are used. We put T=4. As goal quantities of this benchmark problem, we measure the displacement in the normal direction as well as the pressure at the right part of the boundary by b_u⃗ = ∫_Γ_mu⃗·n⃗ o and b_p = ∫_Γ_m p o for Γ_m:= {1}× (0,0.5)× (0,0.5) . The time step size of the temporal discretization is chosen as τ = 0.01. The calculated profile for t=3.8 of the modulus of the vectorial variable u⃗ and the scalar variable p are illustrated in Fig. <ref>. In Table <ref> we summarize characteristic quantities and results of our simulations for various spatial resolutions and different temporal and spatial polynomial degrees of the STFEMs. In Fig. <ref> we visualize the computed benchmark quantities (<ref>) of some of the simulations over the temporal axis. For the temporal discretization with polynomial degree k = 0 and k=1 the computed quantities differ on the spatial refinement level 4 only by a small shift of the phases of the goal quantifies. This indicates that for the current benchmark setting (choice of the model parameter and boundary conditions) the spatial discretization error dominates the temporal one. Therefore, in order to study numerically the convergence and accuracy of the goal quantities, we computed the benchmark problem for higher spatial resolutions (spatial mesh and approximation order; cf. Table <ref>. The lowest row in Table <ref> contains the results of the finest simulation that we could run on our hardware. Fig. <ref> shows that the solution (i.e., the goal quantities) is not fully converged yet. This underlines the complexity of the benchmark. The final column of Table <ref> shows the convergence statistics of the proposed GMRES–GMG solver. In terms of the average number of iterations per time step the solver is (almost) grid independent. This underlines its capability and robustness for solving efficiently the complex systems arising from space-time finite element discretizations of partial differential equations. scientific-notation = false, round-mode=places, round-precision=2, output-exponent-marker=, table-figures-integer=4, table-figures-decimal=2, table-figures-exponent=1, table-sign-mantissa = false, table-sign-exponent = false, table-number-alignment=right §.§ Parallel scaling and energy efficiency Here, we study briefly the parallel scaling and energy efficiency of our solver. By studying energy consumption, we'd like to draw attention to this emerging dimension in the tuning of algorithms. Energy efficiency broadens the classical hardware-oriented numerics that is applied to enhance the performance of the current method on the target hardware and/or to find other numerical methods to improve the numerical efficiency. For the longer term, energy and power consumption needs to be mapped into a rigorous performance model. Here, we restrict ourselves to illustrate numerically the parallel scaling and energy consumption properties of our implementation that uses Message Passing Interface(MPI) libraries and multi threading parallelism. We perform a strong scaling benchmark for the test problem of Subsec. <ref> with k = 1, r = 2 and ref_n = 5, with 10571532 degrees of freedom in each subinterval I_n on the fine level 𝒯_L (cf. Table <ref>) and 98304 mesh cells on 𝒯_L. Throughout, we assign 18 MPI processes to each of the nodes used for the computations and vary the number n of nodes from n=20 to n=200. For the evaluation of the parallel scaling properties, we compute the parallel speedup of the code (cf. <cit.>) that is approximated by S = t_wall(n=n_min)/t_wall(n) , where t_wall(n) denotes the wall time of the simulation of fixed size on n compute nodes and t_wall(n=n_min) is the wall time of the simulation on the minimum number of nodes involved in the scaling experiment. Secondly, we compute similarly the energy ratio by means of R = E(n)/E(n=n_min) , where E(n) measures the total energy consumption of the simulation on n nodes. The energy consumption is determined by the Linux cluster workload manager slurm <cit.>. The energy consumption data is collected from hardware sensors using Intel's Running Average Power Limit (RAPL) mechanism. It measures the energy consumption of the processor and memory. On our system, the sampling interval of energy consumption is determined by the value of 30 seconds. Fig. <ref> illustrates the results of the performance test. To quantify and evaluate the productivity or resource costs of the algorithm and its implementation, we use a simple model for the Productivity P = Output O/Input I; cf. <cit.>. In an economic sense, all outputs should be the desired ones. Therefore, we use the reciprocal of the wall time t_wall as the output O = 1/t_wall, such that a decrease in t_wall represents an increase of the (abstract) output. As the input I we use the total energy consumption E of the simulation. We scale the result by multiplying P with the constant factor E(n_min) · t_wall(n_min) such that the computation with n = n_min has a productivity of P=1.0: P = 1/t_wall(n)/E(n)· E(n_min) · t_wall(n_min) = S/R , with S and R being defined in (<ref>) and (<ref>), respectively. The resulting productivity curve of our computations is presented in Fig. <ref>. In our model the simulation on 80 compute nodes is the most productive one, that is, the ratio of output (low wall-time) to input (energy) is best. The characteristic quantities of our performance study are also summarized in Table <ref>. scientific-notation = false, round-mode=places, round-precision=2, output-exponent-marker=, table-figures-integer=3, table-figures-decimal=3, table-figures-exponent=0, table-sign-mantissa = false, table-sign-exponent = false, table-number-alignment=right § SUMMARY AND OUTLOOK In this work we presented and analyzed families of space-time finite element discretizations of the coupled hyperbolic-parabolic system (<ref>) of poro- or thermoelasticity. The time discretization uses the discontinuous Galerkin method. The space discretization is based on inf-sup stable pairs of finite element spaces with continuous and discontinuous approximation of the scalar variable p. Well-posedness of the discrete problems is proved. For efficiently solving the arising algebraic systems with complex block structure in the case of increasing polynomial degrees of the time discretization a geometric multigrid preconditioner with a local Vanka smoother on patches of finite elements is proposed and studied. The overall approach is carefully evaluated for a sequence of numerical experiments, in particular for a challenging three-dimensional test setting of practical interest. Parallel scaling and energy consumption is also considered in the experiments. The benchmark problem of Subsec. <ref> deserves further studies. In particular, soft materials with a dominating impact of the hyperbolic equation (<ref>) to the overall system (<ref>) need to be studied as well. The definition of the goal quantities of the proposed benchmark might be reconsidered and revised with regard to either an improved choice of such quantities or their computation with higher rate of convergence. This evaluation remains as a work for the future. Finally, multi-field formulations of (<ref>) with an explicit approximation of the stress tensor σ⃗= C⃗ε⃗ and the flux vector q⃗=-K⃗∇ p might be advantageous for applications of (<ref>) in that these quantities are of interest or desired. By using the concept of the local (patchwise) Vanka smoother GMG preconditioner, that was presented here, multi-field formulations of (<ref>) are expected to be become feasible as well. This also remains as a work for the future. § ACKNOWLEDGEMENTS Computational resources (HPC-cluster HSUper) have been provided by the project hpc.bw, funded by dtec.bw — Digitalization and Technology Research Center of the Bundeswehr. 99 A67 G. M. Amdahl, Validity of the single processor approach to achieving large scale computing capabilities, in Proceedings of the April 18-20, 1967, Spring Joint Computer Conference on — AFIPS ’67 (Spring), Atlantic City, New Jersey: ACM Press, 1967, p. 483. AB21 M. Anselmann, M. Bause, A geometric multigrid method for space-time finite element discretizations of the Navier–Stokes equations and its application to 3d flow simulation, ACM Trans. Math. Softw., accepted (2023), https://doi.org/10.1145/3582492, pp. 1–27; arXiv:2107.10561. AB22 M. Anselmann, M. Bause, Efficiency of local Vanka smoother geometric multigrid preconditioning for space-time finite element methods to the Navier-Stokes equations, PAMM Proc. Appl. Math. Mech., accepted (2022), pp. 1–6; arXiv:2210.02690. AB21_2 M. Anselmann, M. Bause, CutFEM and ghost stabilization techniques for higher order space-time discretizations of the Navier–Stokes equations, Int. J. Numer. Meth. Fluids, 94 (2022), pp. 775–802. ABBM20 M. Anselmann, M. Bause, S. Becher, G. Matthies, Galerkin-collocation approximation in time for the wave equation and its post-processing, ESAIM Math. Model. Numer. Anal., 54 (2020), pp. 2099–2123. Aetal21 D. Arndt, W. Bangerth, B. Blais, M. Fehling, R. Gassmöller, T. Heister, L. Heltai, U. Köcher, M. Kronbichler, M. Maier, P. Munch, J.-P. Pelteret, S. Proell, K. Simon, B. Turcksin, D. Wells, J. Zhang, The deal.II Library, Version 9.3, J. Numer. Math., 29 (2021), pp. 171–186. A02 D. N. Arnold, D. Boffi, R. S. Falk, Approximation by quadrilateral finite elements, Math. Comp., 71 (2002), pp. 909–922. BGR10 W. Bangerth, M. Geiger, R. Rannacher, Adaptive Galerkin finite element methods for the wave equation, Comput. Meth. Appl. Math., 10 (2010), pp. 3–48. BR03 W. Bangerth, R. Rannacher, Adaptive Finite Element Methods for Differential Equations, Birkhäuser, Basel, 2003. BKR22 M. Bause, M. Anselmann, U. Köcher, F. A. Radu, Convergence of a continuous Galerkin method for hyperbolic-parabolic systems, Comput. Math. with Appl., submitted (2022), pp. 1–24; arXiv:2201.12014. BRK17 M. Bause, R. Radu, U. Köcher, Space-time finite element approximation of the Biot poroelasticity system with iterative coupling, Comput. Methods Appl. Mech. Engrg., 320 (2017), pp. 745–768. B02 R. Becker, Mesh adaptation for Dirirchlet flow control via Nitsche's method, Commun. Numer. Meth. Engrg., 18 (2002), pp. 669–680. B41 M. Biot, General theory of three-dimensional consolidation, J. Appl. Phys., 12 (1941), pp. 155–164. B55 M. Biot, Theory of elasticity and consolidation for a porous anisotropic solid, J. Appl. Phys., 26 (1955), pp. 182–185. B72 M. Biot, Theory of finite deformations of porous solids, Indiana Univ. Math. J., 21 (1972), pp. 597–620. BKNR22 J. W. Both, N. A. Barnafi, F. A. Radu, P. Zunino, A. Quarteroni, Iterative splitting schemes for a soft material poromechanics model, Comput. Methods Appl. Mech. Engrg., 388 (2022), 114183. BL11 A. Brandt, O. E. Livne, Multigrid Techniques—1984 Guide with Applications to Fluid Dynamics, SIAM, Philadelphia, 2011. B03 S. C. Brenner, Korn's inequalities for piecewise H⃗^1 vector fields, Math. Comp., 73 (2003), pp. 1067–1087. BKB22 M. P. Bruchhäuser, U. Köcher, M. Bause, On the implementation of an adaptive multirate framework for coupled transport and flow, J. Sci. Comput., 93:59 (2022), https://doi.org/10.1007/s10915-022-02026-z, pp. 1–29. C72 D. E. Carlson, Linear thermoelasticity, Handbuch der Physik V Ia/2, Springer, Berlin, 1972. C00 W. W. Cooper, L. M. Seiford, K. Tone Data Envelopment Analysis, Kluwer Academic Publishers, Dordrecht, 2000. PE12 D. A. Di Pietro, A. Ern, Mathematical Aspects of Discontinuous Galerkin Methods, Springer, Berlin, 2012. DJRWZ18 D. Drzisga, L. John, U. Rüde, B. Wohlmuth, W. Zulehner, On the analysis of block smoothers for saddle point problems, SIAM J. Matrix Anal. Appl., 39 (2018), pp. 932–960. F78 G. Fairweather, Finite Element Galerkin Methods for Differential Equations, Lecture Notes in Pure and Applied Mathematics 34, Marcel Dekker Inc., New York, 1978. FTW19 S. Franz, S. Trostorff, M. Waurick, Numerical methods for changing type systems, IMA J. Numer. Anal., 39 (2019), pp. 1009–1038. GHJRW16 B. Gmeiner, M. Huber, L. John, U. Rüde, B. Wohlmuth, A quantitative performance study for Stokes solvers at the extreme scale, J. Comput. Sci., 17 (2016), pp. 509–521. GRSWW15 B. Gmeiner, U. Rüde, H. Stengel, C. Waluga, B. Wohlmuth, Performance and scalability of hierarchical hybrid multigrid solvers for Stokes systems, SIAM J. Sci. Comput., 37 (2015), pp. C143–C168. GRSWW15_2 B. Gmeiner, U. Rüde, H. Stengel, C. Waluga, B. Wohlmuth, Towards textbook efficiency for parallel multigrid, Numer. Math. Theory Methods Appl., 8 (2015), pp. 22–46. H85 W. Hackbusch, Multigrid methods and applications, Springer, Berlin, 1985. HK18 Q. Hong, J. Kraus, Parameter-robust stability of classical three-field formulation of Biot's consolidation model, Electron. Trans. Numer. Anal., 48 (2018), pp. 202–226. HKXZ16 Q. Hong, J. Kraus, J. Xu, L. Zikatanov, A robust multigrid method for discontinuous Galerkin discretizations of Stokes and linear elasticity equations, Numer. Math., 132 (2016), pp. 23–49. HST14 S. Hussain, F. Schieweck, S. Turek, Efficient Newton-multigrid solution techniques for higher order space-time Galerkin discretizations of incompressible flow, Appl. Numer. Math., 83 (2014), pp. 51–71. HST13 S. Hussain, F. Schieweck, S. Turek, An efficient and stable finite element solver of higher order in space and time for nonstationary incompressible flow, Internat. J. Numer. Methods Fluids, 73 (2013), pp. 927–952. HST11 S. Hussain, F. Schieweck, S. Turek, Higher order Galerkin time discretizations and fast multigrid solvers for the heat equation, J. Numer. Math., 19 (2011), pp. 41–61. JR18 S. Jiang, R. Racke, Evolution equations in thermoelasticity, CRC Press, Boca Raton, 2018. J16 V. John, Finite Element Methods for Incompressible Flow Problems, Springer, Cham, 2016. J02 V. John, Higher order finite element methods and multigrid solvers in a benchmark problem for the 3D Navier-Stokes equations, Int. J. Numer. Meth. Fluids, 40 (2002), pp. 775–798. JM01 V. John, G. Matthies, Higher-order finite element discretizations in a benchmark problem for incompressible flows, Int. J. Numer. Meth. Fluids, 37 (2001), pp. 885–903. JT00 V. John, L. Tobiska, Numerical performance of smoothers in coupled multigrid methods for the parallel solution of the incompressible Navier–Stokes equations, Int. J. Numer. Meth. Fluids, 33 (2000), pp. 453–473. KR18 G. Kanschat, B. Riviere, A finite element method with strong mass conservation for Biot's linear consolidation model, J. Sci. Comput., 77 (2018), pp. 1762–1779. KM04 O. Karakashian, C. Makridakis, Convergence of a continuous Galerkin method with mesh modification for nonlinear wave equations, Math. Comp., 74 (2004), pp. 85–102. KM99 O. Karakashian, C. Makridakis, A space-time finite element method for the nonlinear Schrödinger equation: The continuous Galerkin method, SIAM J. Numer. Anal., 36 (1999), pp. 1779–1807. L86 R. Leis, Initial boundary value problems in mathematical physics, Teubner, Stuttgart, John Wiley & Sons, Chichester, 1986. LD03 X. S. Li, J. W. Demmel, SuperLU_DIST: A scalable distributed-memory sparse direct solver for unsymmetric linear systems, ACM Trans. Math. Softw., 29 (2003), pp. 110–140. LLRS94 J. Linden, G. Lonsdale, H. Ritzdorf, A. Schüller, Scalability aspects of parallel multigrid, Future Gener. Comput. Syst., 10 (1994), pp. 429–439. M06 S. Manservisi, Numerical analysis of Vanka-type solvers for steady Stokes and Navier–Stokes flows, SIAM. J. Numer. Anal., 44 (2006), pp. 2025–2056. M01 G. Matthies, Mapped finite elements on hexahedra. Necessary and sufficient conditions for optimal interpolation errors, Numer. Algor., 27 (2001), pp. 317–327. MT02 G. Matthies, L. Tobiska, The inf-sup condition for the mapped ℚ_k^d/P_k-1^ element in arbitrary space dimensions, Computing, 69 (2002), pp. 119–139. MW12 A. Mikelić, M. F. Wheeler, Theory of the dynamic Biot–Allard equations and their link to the quasi-static Biot system, J. Math. Phys., 53 (2012), 123702:1–15. N71 J. Nitsche, Über ein Variationsprinzip zur Lösung von Dirichlet Problemen bei Verwendung von Teilräumen, die keinen Randbedingungen unterworfen sind (in German), Abh. Math. Sem. Univ. Hamburg, 36 (1971), pp. 9–15. PW07 P. J. Philips, M. F. Wheeler, A coupling of mixed and continuous Galerkin finite element methods for poroelasticity I, II, Comput. Geosci., 11 (2007), 131–158. PW08 P. J. Philips, M. F. Wheeler, A coupling of mixed and discontinuous Galerkin finite element methods for poroelasticity , Comput. Geosci., 12 (2008), 417–435. RHOAGZ17 C. Rodrigo, X. Hu, P. Ohm, J. H. Adler, F. J. Gaspar, L. T. Zikatanov, New stabilized discretizations for poroelasticity and the Stokes' equations, Comput. Methods Appl. Mech. Engrg., 341 (2018), pp. 467–484. R17 U. Rüde, Algorithmic Efficiency and the Energy Wall, 2017. DOI: 10.13140/RG.2.2.33914.18881. STW22 C. Seifert, S. Trostorff, M. Waurick, Evolutionary Equations: Picard's Theorem for Partial Differential Equations, and Applications, Birkhäuser, Cham, 2022. S03 Y. Saad, Iterative Methods for Sparse Linear Systems. 2nd edition. SIAM, Philadelphia, 2003. S00 R. Showalter, Diffusion in poro-elastic media, J. Math. Anal. Appl., 251 (2000), pp. 310–340. S89 M. Slodička, Application of Rothe's method to integrodifferential equation, Comment. Math. Univ. Carolinae, 30 (1989), pp. 57–70. S23 Slurm Workload Manager, Version 22.05, <https://github.com/SchedMD/slurm/tree/master/src/plugins/acct_gather_energy/rapl>, 2023. SZ22 O. Steinbach, M. Zank, A generalized inf–sup stable variational formulation for the wave equation, J. Math. Anal. Appl., 505 (2022), 125457. T06 V. Thomeé, Galerkin Finite Element Methods for Parabolic Problems, Springer, Berlin, 2006. TOS01 U. Trottenberg, C. W. Oosterlee, A. Schüller, Multigrid, Academic Press, San Diego, CA, 2001. T99 S. Turek, Efficient Solvers for Incompressible Flow Problems, Springer, Berlin, 1999. TBK06 S. Turek, C. Becker, D. Kilian, Hardware-oriented numerics and concepts for PDE software, Future Generation Computer Systems, 22 (2006), pp. 217–238. TGBBW10 S. Turek, D. Göddecke, C. Becker, S. Buijssen, H. Wobker, FEAST — realisation of hardware-oriented numerics for HPC simulations with finite elements, Concurrency and Computation: Practice and Experience 6 (May), pp. 2247–2265. Special Issue Proceedings of ISC 2008, doi:10.1002/cpe.1584. V86 S. Vanka, Block-implicit multigrid solution of Navier-–Stokes equations in primitive variables, J. Comput. Phys., 65 (1986), pp. 138–158. WS00 D. C. Wirtz, N. Schiffers, T. Pandorf, K. Rademacher, D. Weichert, R. Forst, Critical evaluation of known bone material properties to realize anisotropic FE-simulation of the proximal femur, J. Biomech., 33 (2000), pp. 1325–1330. § APPENDIX § DISCRETE COERCIVITY Here, we prove the discrete coercivity of the bilinear form A_γ defined in (<ref>) along with (<ref>). For this we introduce the mesh- and parameter-dependent norm v⃗_h _h,γ_a := (ε(v⃗_h) ^2 + γ_a h^-1v⃗_h _Γ^D_u⃗^2)^1/2 , for v⃗_h ∈V⃗_h , with some parameter γ_a> 0. Its choice is addressed below. The norm property of (<ref>) is ensured by a variant of Korn's inequality; cf. <cit.>. We recall the well-known inverse inequality (cf. <cit.>) ε⃗(v⃗_h) n⃗_Γ^D_u⃗≤ c h^-1/2ε⃗(v⃗_h) , for v⃗_h ∈V⃗_h . From (<ref>) along with (<ref>), it follows by the positive definiteness (<ref>) of C⃗, the inequality of Cauchy–Young with a sufficiently small constant δ >0 and the trace inequality (<ref>) that, for all v⃗_h ∈V⃗_h, A_γ( v⃗_h, v⃗_h) = ⟨C⃗ε⃗(v⃗_h),ε⃗(v⃗_h)⟩ - ⟨C⃗ε⃗(v⃗_h) n⃗, v⃗_h⟩_Γ^D_u⃗ - ⟨v⃗_h, C⃗ε⃗(v⃗_h) n⃗⟩_Γ^D_u⃗ + γ_a h^-1⟨v⃗_h, v⃗_h ⟩_Γ^D_u⃗ ≥ c ε⃗(v⃗_h) ^2 - c ε⃗(v⃗_h) n⃗_Γ^D_u⃗v⃗_h_Γ^D_u⃗+ γ_a h^-1v⃗_h ^2_Γ^D_u⃗ ≥ c ε⃗(v⃗_h) ^2 - c δ h ε⃗(v⃗_h) n⃗^2_Γ^D_u⃗ - c δ^-1 h^-1v⃗_h^2_Γ^D_u⃗ + γ_a h^-1v⃗_h ^2_Γ^D_u⃗ ≥ c ε⃗(v⃗_h) ^2 - c δε⃗(v⃗_h) ^2 - c δ^-1 h^-1v⃗_h^2_Γ^D_u⃗ + γ_a h^-1v⃗_h ^2_Γ^D_u⃗ = c( ε⃗(v⃗_h) ^2 + γ_a h^-1v⃗_h ^2_Γ^D_u⃗) , with some constant γ_a >0 for a sufficiently large choice of the algorithmic parameter γ_a in (<ref>), such that γ_a :=γ_a -cδ^-1>0. Thus, there holds for some constant c>0 that A_γ( v⃗_h, v⃗_h) ≥ c v⃗_h _h,γ_a , for all v⃗_h ∈V⃗_h . Secondly, we prove the discrete coercivity of the bilinear form B_γ defined in (<ref>) along with (<ref>). For brevity, we study the case Q_h= Q_h^l,cont only, such that Q_h⊂ H^1(Ω) is satisfied. The case Q_h= Q_h^l,disc can be captured similarly. We introduce the mesh- and parameter-dependent norm q_h _h,γ_b := (∇ q_h ^2 + γ_b h^-1 q_h _Γ^D_p^2)^1/2 , for q_h ∈ Q_h⊂ H^1(Ω) , with some parameter γ_b> 0, where Q_h=Q_h^l,cont; cf. (<ref>). Further, we recall the well-known inverse inequality ∇ q_h ·n⃗_Γ^D_p≤ c h^-1/2∇ q_h , for q_h ∈ Q_h . From (<ref>) along with (<ref>), it follows by the positive definiteness (<ref>) of K⃗, the inequality of Cauchy–Young with a sufficiently small constant δ >0 and the trace inequality (<ref>) that, for all q_h ∈ Q_h, B_γ( q_h, q_h) = ⟨K⃗∇ q_h ,∇ q_h ⟩ - ⟨K⃗∇⃗q_h ·n⃗, q_h⟩_Γ^D_p - ⟨ q_h, K⃗∇ q_h ·n⃗⟩_Γ^D_p + γ_b h^-1⟨ q_h, q_h ⟩_Γ^D_p ≥ c ∇ q_h) ^2 - c ∇ q_h ·n⃗_Γ^D_p q_h_Γ^D_p+ γ_b h^-1 q_h ^2_Γ^D_p ≥ c ∇ q_h ^2 - c δ h ∇ q_h ·n⃗^2_Γ^D_p - c δ^-1 h^-1 q_h^2_Γ^D_p + γ_b h^-1 q_h ^2_Γ^D_p ≥ c ∇ q_h ^2 - c δ∇ q_h ^2 - c δ^-1 h^-1v⃗_h^2_Γ^D_p + γ_b h^-1 q_h ^2_Γ^D_p = c( ∇ q_h ^2 + γ_b h^-1 q_h ^2_Γ^D_p) , with some constant γ_b >0 for a sufficiently large choice of the algorithmic parameter γ_b in (<ref>), such that γ_b :=γ_a -cδ^-1>0. Thus, there holds for some constant c>0 that B_γ( q_h, q_h) ≥ c q_h _h,γ_b , for all q_h ∈ Q_h . § ALTERNATIVE FORMULATION OF THE FULLY DISCRETE PROBLEM Here, we present an alternative formulation for the fully discrete system to (<ref>). The difference to Problem <ref> comes through using ∂_t u⃗_τ,h instead of v⃗_τ,h in (<ref>) by means of (<ref>) and (<ref>), respectively. However, this modification requires the inclusion of an additional boundary integral in the resulting equation (<ref>). This non-obvious adaptation results from the proof of well-posedness of the discrete problem. The following problem is thus considered. [Numerically integrated I_n-problem with ∂_t u⃗_τ,h] For given u⃗_h^n-1:= u⃗_τ,h(t_n-1)∈V⃗_h, v⃗_h^n-1:= v⃗_τ,h(t_n-1)∈V⃗_h, and p_h^n-1:= p_τ,h(t_n-1) ∈ Q_h with u⃗_τ,h(t_0) :=u⃗_0,h, v⃗_τ,h(t_0) :=u⃗_1,h and p_τ,h(t_0) := p_0,h, find (u⃗_τ,h,v⃗_τ,h,p_τ,h) ∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h) such that Q_n (⟨∂_t u⃗_τ,h , ϕ⃗_τ,h ⟩- ⟨v⃗_τ,h , ϕ⃗_τ,h ⟩) + ⟨u⃗^+_τ,h(t_n-1), ϕ⃗_τ,h^+(t_n-1)⟩= ⟨u⃗_h^n-1, ϕ⃗_τ,h^+(t_n-1)⟩ , Q_n (⟨ρ∂_t v⃗_τ,h , χ⃗_τ,h ⟩+ A_γ(u⃗_τ,h, χ⃗_τ,h ) + C(χ⃗_τ,h,p_τ,h)) + ⟨ρv⃗^+_τ,h(t_n-1), χ⃗_τ,h^+(t_n-1)⟩ = Q_n (F_γ(χ⃗_τ,h)) + ⟨ρv⃗_h^n-1, χ⃗_τ,h^+(t_n-1)⟩ , Q_n (⟨c_0 ∂_t p_τ,h,ψ_τ,h ⟩- C(∂_t u⃗_τ,h,ψ_τ,h)+ B_γ(p_τ,h, ψ_τ,h)) + ⟨c_0 p^+_τ,h(t_n-1) +α∇·u⃗^+_τ,h(t_n-1), ψ_τ,h^+(t_n-1)⟩- α⟨u⃗^+_τ,h(t_n-1) ·n⃗ ,ψ_τ,h^+(t_n-1) ⟩_Γ^D_u⃗ = Q_n ( G_γ(ψ_τ,h)) + ⟨c_0 p_h^n-1 + α∇·u⃗_h^n-1, ψ_τ,h^+(t_n-1)⟩- α⟨u⃗_D(t_n-1) ·n⃗ ,ψ_τ,h^+(t_n-1) ⟩_Γ^D_u⃗ for all (ϕ⃗_τ,h,χ⃗_τ,h,ψ_τ,h)∈ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;V⃗_h) ×ℙ_k (I_n;Q_h). Problem <ref> admits a unique solution. The proof follows basically the lines of the proof of Lem. <ref> and is kept short. Only differences to Lem. <ref> are depicted. For the differences (u⃗_τ,h,v⃗_τ,h,p_τ,h) of two triples satisfying (<ref>) and the test functions ϕ⃗_τ,h= A_γu⃗_τ,h, χ⃗_τ,h=v⃗_τ,h and ψ_˝τ,h=p_τ there holds that 1/2∫_t_n-1^t_nd/dt( A_γ (u⃗_τ,h , u⃗_τ,h) + ⟨ρv⃗_τ,h , v⃗_τ,h⟩ + ⟨ c_0 p_τ,h, p_τ,h⟩ ) t + Q_n(B_γ(p_τ,h,p_τ,h) ) + Q_n(C_γ (v⃗_τ,h - ∂_t u⃗_τ,h,p_τ,h) ) + α⟨∇·u⃗^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ - α⟨u⃗^+_τ,h(t_n-1) ·n⃗ ,p_τ,h^+(t_n-1) ⟩_Γ^D_u⃗ + A_γ (u⃗^+_τ,h(t_n-1), u_τ,h^+(t_n-1)) + ⟨ρv⃗^+_τ,h(t_n-1), χ_τ,h^+(t_n-1)⟩ + ⟨ c_0 p^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ = 0 . Now, let l∈{1,…,k+1} be arbitrary but fixed and ϕ⃗_τ,h∈ℙ_k(I_n;V⃗_h) be chosen as ϕ⃗_τ,h(t) := ξ_n,l(t) ϕ⃗_h with ξ_n,l(t) := (∏_i=1 i ≠ l^k+1(t-t_n,i^GR)) (∏_i=1 i ≠ l^k+1(t_n,l^GR -t_n,i^GR))^-1∈ℙ_k(I_n;) , ϕ⃗_h ∈V⃗_h , and the Gauss–Radau quadrature nodes t_n,μ^GR, for μ = 1,… , k+1; cf. (<ref>). By the exactness of the Gauss–Radau quadrature formula (<ref>) for all polynomials in ℙ_2k(I_n;) we deduce from (<ref>) that 0 = τ_n/2∑_μ=1^k+1ω̂_μ^GR(⟨∂_t u⃗_τ,h(t_n,μ^GR), ϕ⃗_τ,h(t_n,μ^GR) ⟩ - ⟨v⃗_τ,h(t_n,μ^GR), ϕ⃗_τ,h(t_n,μ^GR) ⟩) + ⟨u⃗^+_τ,h(t_n-1), ϕ⃗_τ,h^+(t_n-1)⟩ = τ_n/2ω̂_l^GR⟨∂_t u⃗_τ,h(t_n,l^GR) - v⃗_τ,h(t_n,l^GR), ϕ⃗_h ⟩ + ⟨u⃗^+_τ,h(t_n-1), ξ^+_n,l(t_n-1) ϕ⃗_h ⟩ . Thus, we have that (l=1,…, k+1) v⃗_τ,h(t_n,l^GR) - ∂_t u⃗_τ,h(t_n,l^GR) = c_n,l u⃗^+_τ,h(t_n-1) with c_n,l = 2 τ_n^-1 (ω̂_l^GR)^-1 ξ^+_n,l(t_n-1) . Substituting (<ref>) into the third term on the left-hand side of (<ref>), we get that Q_n(C_γ (v⃗_τ,h - ∂_t u⃗_τ,h,p_τ,h) ) = ∑_μ=1^k+1 C_γ (u⃗_τ_h^+(t_n-1),ξ^+_n,μ(t_n-1) p_τ,h(t_n,μ^GR) ) = C_γ (u⃗_τ_h^+(t_n-1), p_τ,h^+(t_n-1)) = - α⟨∇·u⃗_τ_h^+(t_n-1), p_τ,h^+(t_n-1)⟩ + α⟨u⃗_τ_h^+(t_n-1) ·n⃗ , p_τ,h^+(t_n-1) ⟩_Γ^D_u⃗ . Combining (<ref>) with (<ref>) then implies that A_γ (u⃗_τ,h(t_n) , u⃗_τ,h(t_n)) + ⟨ρv⃗_τ,h (t_n), v⃗_τ,h(t_n) ⟩ + ⟨ c_0 p_τ,h(t_n), p_τ,h(t_n)⟩ + 2 Q_n (B_γ(p_τ,h,p_τ,h) ) + A_γ(u⃗^+_τ,h(t_n-1), u⃗_τ,h^+(t_n-1)) + ⟨ρv⃗^+_τ,h(t_n-1), χ_τ,h^+(t_n-1)⟩ + ⟨ c_0 p^+_τ,h(t_n-1), p_τ,h^+(t_n-1)⟩ = 0 . From (<ref>) along with the discrete coercivity properties (<ref>) and (<ref>) we deduce that u⃗_τ,h(t_n)= u⃗_τ,h^+(t_n-1)=0⃗ , v⃗_τ,h(t_n)= v⃗_τ,h^+(t_n-1)=0⃗ and p_τ,h(t_n)=p_τ,h^+(t_n-1)= 0 as well as p_τ,h(t_n,μ^GR) = 0 , for μ = 1,…, k+1 . The rest then follows as in the proof of Lem. <ref>. § ADDITIONAL NUMERICAL EXPERIMENTS Here, we present some additional results of our numerical experiments. Table <ref> shows the results of the convergence study introduced in Subsec. <ref> for the solution (<ref>) of (<ref>). For the computations of Table <ref> the Taylor-Hood pair of finite element spaces ℚ_r^2/ℚ_r-1, with r=4, is used instead of the pair ℚ_r^2/ℙ_r-1^disc chosen for the results of Table <ref>. For the pressure approximation marginally smaller errors are observed. In Table <ref> the computed errors of the approximation of (<ref>) for larger values of the Lamé parameters λ and μ in the elasticity tensor C⃗ are summarized. No significant increase of the errors is observed, indicating the independence of the error constant on the magnitude of C⃗.

http://arxiv.org/abs/2303.07008v2

20230313110646

Status substitution and conspicuous consumption

econ.TH

Appendix Mengchao He March 30, 2023 ================== This paper adapts ideas from social identity theory to set out a new framework for modelling conspicuous consumption. Notably, this approach can explain two stylised facts about conspicuous consumption that initially seem at odds with one another, and to date have required different families of models to explain each: (1) people consume more visible goods when their neighbours' incomes rise, but (2) consume less visible goods when incomes of those with the same race in a wider geographic area rise. The first fact is typically explained by `Keeping up with the Joneses' models, and the second by signalling models. Our model also explains related features of conspicuous consumption: that the rich are more sensitive to others' incomes than the poor, and that the effect of income inequality on consumption differs qualitatively across groups. Importantly, it explains this fourth stylised fact without falling back on differences in preferences across groups, as required in other models. In addition, our model delivers new testable predictions regarding the role of network structure and income inequality for conspicuous consumption. JEL: D63, D85, D91 Keywords: Social identity, Keeping Up with the Joneses, networks, centrality, income inequality Appendix Mengchao He March 30, 2023 ================== Economists widely recognise that people's consumption decisions respond to the consumption decisions of others. Existing literature documents two main stylised facts about these responses. First, an influential paper by <cit.> documents that, holding their own permanent income fixed, people in the US consume less when average income of those of the same race and who live in the same state is higher. Second, work by <cit.> shows that people consume more when average income of their geographic neighbours increases. This is often called a “Keeping up with The Joneses” effect. One the face of it, these two stylised facts may appear to conflict. Plus, the classes of models typically used to rationalise each of them – signalling models for the <cit.> result and social comparisons models for the <cit.> result – cannot capture both simultaneously. We present a simple model of social identity that explains both of these stylised facts as a single phenomenon. In it, people derive status directly from consuming visible goods themselves, and from belonging to an identity that consumes lots of visible goods. In our model, this identity group status depends on average consumption of visible goods by members of the identity group, and we assume there is some substitutability between status from each of these two sources. People also feel pressure to conform to some benchmark for their identity, and we assume it depends only on a person's neighbours rather than their entire identity group. These two features are both core components of social identity theory <cit.>. Our model is also able to speak to the first stylised fact without relying on the characteristics of the very poorest person – which is needed in signalling models. This aligns our model and its predictions more closely with the kind of data typically available. We also explain a third and a fourth stylised fact, both of which are related to the main two facts. Third, <cit.> finds that richer people are more sensitive to changes in the income of others. Our model predicts exactly this behaviour. Fourth, <cit.> find that higher dispersion of income within a person's reference group leads to lower conspicuous consumption among Whites, but higher conspicuous consumption among Blacks. Our model is flexible enough to capture these contrasting behaviours. It can do so because it finds that the overall impact of a change in inequality depends on the network structure. Therefore we can explain both of these findings if Blacks and Whites have different network structures. We believe explaining the fourth stylised fact in this way is a useful contribution because it does not rely on differences in preferences across racial groups. <cit.> stress that an advantage of their signalling model is that they can explain their main empirical findings without falling back on racial differences in preferences. But they rely on exactly these differences to explain their findings regarding inequality. On the way to explaining these four stylised facts, we provide a simple characterisation of behaviour. It shows that consumption depends critically on a measure of network centrality that is closely related to Bonacich centrality <cit.> – a well-known metric in the networks literature. Both an agent's own centrality and the average centrality among her identity group are important. Centrality is increasing in the strength of network connections, in an agent's own income and in the income of those share her identity and whom she is connected to in the network (whether directly or indirectly). Related to this, we find that stronger homophily – when people are more tightly connected to those like themselves – increases consumption. This is because being more focused on people with her own identity increases the pressure an agent feels to conform – which here pushes up consumption. We also study the effect of inequality. Doing so requires some simplifying assumption to keep the model tractable. We approach this from two different angles. The first shuts down some of the complex network effects by assuming that people only interact with others who have the same income. Here, the impact of redistribution depends on the relatively network density of groups whose income change, rather than on whether redistribution raises or lowers inequality. It is this viewpoint that explains our fourth stylised fact. The second approach imposes a regularity condition on the network, but retains interactions between people with different incomes. With this approach, redistribution increases average consumption if (and only if) the redistribution reduces inequality. It is now helpful to discuss social identity models, and the key features of ours, in a little more detail. Social identity theory assumes people have an identity, and that this affects their utility in three ways. First, they gain status benefits associated with their identity – typically based on some summary statistic of all the people who share their identity. Second, there is some pressure to conform – they bear some dissonance costs if they differ from a prototype for their identity. Third, ties to others sharing their identity affect dissonance costs. Status benefits come from average visible goods consumption of people holding a given identity. Our model makes two assumptions that are not standard in existing models of social identity. First, that people derive some status benefit from their own consumption, and that this component of status is a substitute for the status from the social identity. An intuitive way to view this substitutability is to assume that people simply get status from both sources – their own visible consumption and the average income of their identity – but that utility is concave in total status. Second, we assume the status associated with an identity depends on the whole population, while dissonance depends only on a set of neighbours. Intuitively, a person's view of what they should be conforming to depends on who they know. That status depends on a population average is standard. However, the assumption that dissonance costs depend on neighbours' decisions is much richer than normal. Existing models of social identity assume the prototype is exogenous. In contrast, we both demand consistency between actual behaviour and the prototype. By using an explicit network, we also allow this prototype to differ across people.[Using an exogenous fixed prototype will often result in (average) equilibrium consumption for an identity being different from the prototype. However, the prototype is supposed to capture what is `normal' consumption for that identity. We consider this an undesirable tension because it allows permanent differences between what is actually normal behaviour for an identity and what is supposed to be normal behaviour for that identity. The network can of course be asymmetric. For example, people may feel pressure to conform with celebrities they have never met personally, but these celebrities likely do not feel corresponding pressure to conform to them.] In light of this, we make a simplifying assumption in the interest of tractability – we assume the social identity is exogenously fixed. To see how this model explains both main stylised facts, suppose that incomes increase for some people. These people increase their consumption of visible goods – the traditional direct response to an increase in own income. Now let us focus on a single person whose income did not change. The change in other people's income creates two spillover effects. First, it increases the status benefits for the single person, by increasing the average income of those who share her identity. This then pushes her to consume fewer visible goods – because the status from her identity and the status from her consumption of visible goods are substitutes. Second, it increases her neighbours' consumption, which increases the pressure to consume visible goods through the dissonance costs. This pushes her to consume more visible goods. Which of the two forces wins out depends on how closely connected the single person is to the people whose incomes increase. This is what delivers both of the stylised facts. When the people whose incomes increase are close neighbours (as in <cit.>) we should expect the second force to win out – so she consumes more visible goods. In contrast, when these people are not closely connected to her (as in <cit.>) we should expect the first force to win out – delivering the seemingly counter-intuitive result that she consumes fewer visible goods. The rest of the paper is organised in the usual way. <Ref> briefly reviews related literature. <Ref> presents the model. <Ref> presents the results. <Ref> shows how our model explains the stylised facts. Then <Ref> examines the impact of inequality. §.§ CG version The utility individuals receive from consumption of a commodity goes often beyond its egoistic value. This is revealed by a price which is higher than the marginal egoistic utility. A specific consumption behaviour might be the evidence of envy or altruism. It can be associated to the adoption of a social identity, for example an Indian caste. The literature typically uses the notion of status good when the consumption of the good is meant to reveal the social status of the individual, for example her income, cast, or education. In this paper we wish to propose a single model able to capture the consumption behaviour of a status good in a variety of situations, some of which appearing at first sight unrelated. The leading situation we want to capture has been studied in an influential paper by Charles et al. They showed that for individuals of any given permanent income, black people consume more luxury goods than whites. Surprisingly, by comparing different US states, they also find that the lower is the average income in the state the higher is the consumption of luxury goods for a given income, this result holding for both black and whites. Charles et al. propose a signalling model to capture these findings. Signalling models have the feature that the equilibrium is determined by the behaviour of the poorest individual. Charles et al. signalling model capture the data provided that a lower mean is associated to the existence of poorer individuals at the bottom of the distribution. As the data concerns the means, the model does not provide an exhaustive explanation of the data. To capture this stylised fact we adopt the view that consumption of a visible status good is governed by the social identity adopted by the individual. Social Identity theory determines precisely the main factors affecting the utility of a given identification. These are the status associated to the identity, the dissonance to the group and the ties with the group. We make some important assumptions. First, status depends both on the status of the individuals in the society with that identity as well as the status of the individual herself. We assume that status is increasing in both inputs and that these two inputs are substitutes. Second, we assume that status of the identity depends on the whole population while dissonance is evaluated in respect to her neighbours. As status is linked to conspicuous consumption, dissonance gives rise to a conformist action by the individual in response to her neighbours. The mechanism by which the present model is able is able to rationalise the counter intuitive regularity about the effect of the mean on conspicuous conspicuous consumption is as follows. When the mean income of the identity group falls, their consumption of status good falls, and their status follows. An individual with a given income would then substitute the identity status with her own status, rising her own consumption of the status good. Delivering the surprising result. Our model is able also to capture another important stylised fact. Kuhn et al. (2008) have shown that when individuals in a geographical neighbourhood receive a positive unexpected positive shock to their income, as a lottery gain, the other neighbours follow the changes in the visible consumption of the lucky ones. There are several other situations in which this behaviour, known as keep up with the Jones, has been observed. Our model captures this pattern mainly via the dissonance term. We then use our model to analyse the role of inequality on the consumption of status goods. Signalling models have the feature that as the equilibrium is determined by the behaviour of the poorest individual, she is only affected by the income distribution between her and the poorest individual. Consequently, consumption may increase or decrease with income inequality. Consumption would increase with inequality if the Engel curves, that is consumption as a function of income, are convex, but data seems to suggest the opposite, see Heffetz (2011). Our model captures the role of inequality in a more natural way that fits with the stylized facts. Our model is also able to capture the role of the social network. We provide results on the type of networks that are favorable to a large consumption of status goods. To conclude, the substitutability in the two contributors of status and the tendency to conform with neighbours, deliver the necessary forces to obtain both empirical regularities. We believe this is an important contribution as previous explanations not relying on social identity required completely different models. The present paper advances the literature in other three directions. 1) The model is able to capture the link between average income of a group and the individual consumption for a given income. This contrasts with signalling models that rely on the characteristics of the poorest consumers, which is hardly testable. 2) The model allows to make predictions about the role of inequality on conspicuous consumption. 3) The model allows to capture the effect of the social network. § LITERATURE This paper relates to a very large literature on interpersonal comparisons in consumption decisions.[See <cit.> for an interesting survey of the wider topic of social pressure.] Work on this aspect of consumer behaviour span a diverse range of disciplines, including marketing, sociology, politics, social psychology, economics and neuroscience. It is variously referred to as `conspicuous consumption' or `status consumption'.[<cit.> present empirical evidence showing that while `status consumption' and `conspicuous consumption' are not entirely identical in principle, they substantially overlap in practice.] See <cit.> and <cit.> for recent reviews in psychology, and <cit.> for up to date references within economics. There is also an influential literature within economics that focuses on direct income comparisons across individuals, rather than consumption comparisons. This literature shows that people's happiness is negatively affected by others' income. Recent empirical evidence on this front includes <cit.> and <cit.>, and relative income concerns are also found in a range of laboratory experiments <cit.>.[Early contributions to this literature include <cit.> and <cit.>.] Our model is one of conspicuous consumption in the sense that people care directly about others' consumption. But others' income has an indirect effect, via its influence on their consumption. So we will examine how income, and especially income inequality, affects people's consumption. Within this vast literature, our work is most closely related to: the signalling and social comparisons models currently used to explain the main stylised facts, to the social identity theory work we build, and to work on the relationship between inequality and conspicuous consumption. Signalling. At the heart of signalling models is the idea that agents take actions to advertise their underlying type. This provides one clean rationale for consumption of visible and luxury goods – signalling otherwise unobservable income, when income carries social status. <cit.> use a signalling model – which follows <cit.> very closely – to explain the first stylised fact. There, people use consumption of visible goods to signal their (unobservable) income. In this family of models, people signal to differentiate themselves from those who have lower income. Equilibrium is ultimately determined by the behaviour of the poorest person; who never does any signalling in equilibrium. A consequence of this framework is that a person's consumption is unaffected by changes in incomes among those richer than herself. The model therefore cannot provide an exhaustive explanation of the data – which concerns changes in mean incomes, regardless of which part of the distribution is driving the change in the mean.[Changes to the income of the poorest person (i.e. to the support of the income distribution) is what drives changes in visible goods consumption in the <cit.> signalling model. While <cit.> draw an analogy between reducing the minimum income in the distribution and reducing the mean, they need not go together. Strictly speaking, the model and the data do not look at exactly the same quantity.] Additionally, with no explicit notion of `neighbours', these models struggle to explain the second stylised fact – that people consume more when their neighbours' incomes rise. Social Comparisons. In contrast to signalling models, most of the literature on conspicuous consumption assumes that visible and luxury goods carry status directly, and so focuses directly on interpersonal comparisons in the consumption of these goods. Such `social comparisons' models can explain the second stylised fact – that people consume more when their neighbours' incomes increase. This family of models posits that people desire high consumption relative to others, not just in absolute terms. It is often called a `Keeping Up With the Joneses' effect. Our paper is most closely connected to a relatively recent strand that uses an explicit network to allow for very rich patterns of social comparisons. The broader literature on social comparisons in consumption is very old – dating at least to <cit.> and <cit.> in economics – and far too large to survey here. Explicit networks were first introduced into social comparison models by <cit.>.[Before explicit networks were introduced, most authors either assumed that people care about their consumption relative to the population average <cit.> or about their rank within the population <cit.>. More recent contributions include <cit.> and <cit.>.] Other recent contributions include <cit.> and <cit.>.[By also using an explicit network, our paper relates to the literature of network games – <cit.> and <cit.> provide detailed surveys. A paper in this literature for our purposes is <cit.>, which introduces a measure of network centrality – called Bonacich centrality (first defined by <cit.> and <cit.>) – as the Nash equilibrium of a game of social interactions. Our contribution to this literature is to show how network games of this kind can be added to models from social identity theory.] These models can cleanly explain why someone consumes more visible goods if their neighbours' incomes increase. However, they cannot fit the first stylised fact. Because these models embed a desire to `Keep Up With the Joneses', an increase in the incomes of some people must (weakly) increase everyone's consumption of visible goods. Social Identity. Our model builds on existing models of social identity theory in economics. <cit.> and <cit.> make seminal contributions, formalising social identity in economics and introducing the notions status dissonance costs which are core parts of the modelling framework (see <cit.> for a survey). A number of recent papers extend or apply Shayo's approach (see for example, <cit.>, and <cit.>). These papers allow people to choose their identity, but keep each identity's prototype (the benchmark for `normal') exogenously fixed. In contrast, we make the prototype endogenous – in effect demanding that there are not persistent differences between actual behaviour in equilibrium and the benchmark for `normal' behaviour. However, we do not allow people to choose their identity. We keep identity exogenously fixed in the interest of tractability, and because it is not needed to explain the stylised facts we focus on. For each identity, we follow the utility formulation in <cit.> but we make the prototype endogenous, and base our formulation of the prototype on factors <cit.> identifies as key for social identity. Clearly, there are other factors affecting which identity becomes salient including the cognitive cost of switching identities <cit.>. Inequality and conspicuous consumption. Formal models within economics provide ambiguous results on the link between inequality and conspicuous consumption. In <cit.>, equilibrium only depends on the poorest consumer, so redistribution of income from a poorer person to a richer person has an ambiguous effect on total consumption. The poorer person will consume less, and the richer person will consume more. but total consumption only increases if consumption is a convex function of income, which is not a uniform property of the data <cit.>. Similar predictions come out of other versions of the signalling model (for example, <cit.> and <cit.>). Finally, <cit.> confirms that with ordinal utility, inequality reduces conspicuous consumption while with cardinal status inequality may increase conspicuous consumption. The effects of inequality on luxury goods consumption has recently been studied in consumer psychology. <cit.> find that income inequality is associated with high frequency and greater positivity of tweets mentioning luxury brands, and less frequent tweets mentioning low status brands (see also <cit.>). Relatedly, <cit.> find that inequality has an effect on the consumption of luxury goods by poor individuals in India (see also <cit.>). Formal models of this literature also predict a positive link between inequality and conspicuous good consumption, as for example the models on social rank proposed by <cit.> and by <cit.>. The link appears to depend on circumstances. In a laboratory experiment, <cit.> find that reducing inequality reduces conspicuous consumption when poorer consumers do not care about status, but increases their conspicuous consumption when they do. Existing literature also hints at the idea that the link between income inequality and conspicuous consumption depends on the social network; that is, on the granular details of who compares themselves with whom <cit.>. By including an explicit social network that is able to capture very rich patterns of interpersonal comparisons, this is something our setup is able to investigate more formally. NOTE??: Frank, Levine, Dijk (2014) find that higher inequality within a reference group leads to lower savings (i.e. higher consumption) % References * Dubois, D., and N. Ordabayeva, Social Hierarchy, Social Status, and Status Consumption, The Cambridge Handbook of Consumer Psychology, Cambridge Handbooks in Psychology. Ed., M. I. Norton, D. D. Rucker, and C. Lamberton, March 2017, Paperback ISBN: 9781107641426. * Dubois, D., S. Jung, and N. Ordabayeva, The psychology of luxury consumption, Current Opinion in Psychology 2021, 39:82–87. * Costa-Font, J., and F. Cowell, Social identity and redistributive preferences: a survey. Journal of Economic Surveys (2015) Vol. 29, No. 2, pp. 357–374 * Daly, Boyce, and Wood (2015) A social rank explanation of how money influences health. Health Psychology, 34, 222–230. * Walasek and Brown (2016). Income inequality,income, and internet searches for status goods: A cross-national study of the association between inequality and well-being. Social Indicators Research,129, 1001–1014. https://doi.org/10.1007/s11205-015-1158-4 * Walasek L, Brown GD, Income inequality and status seeking: searching for positional goods in unequal US states. Psychol Sci 2015, 26:527-533 http://dx.doi.org/10.1177/ 0956797614567511. * Walasek L, Bhatia S, Brown GD, Positional goods and the social rank hypothesis: income inequality affects online chatter about high-and low-status brands on Twitter. J Consum Psychol 2018, 28:138-148, http://dx.doi.org/10.1002/jcpy.1012.) * Nave G, Nadler A, Dubois D, Zava D, Camerer C, Plassmann H, Single-dose testosterone administration increases men's preference for status goods. Nat Commun 2018, 9:1-8 http://dx. doi.org/10.1038/s41467-018-04923-0. * O'Cass, A., and H. McEwen (2004). Exploring consumer status and conspicuous consumption. Journal of Consumer Behaviour, 4(1), 25–39. https://doi.org/10.1002/cb.155 * Zinn, Anna K., Lavric, Aureliu, Levine, Mark and Koschate, Miriam (2022).Social identity switching: How effective is it?. Journal of Experimental Social Psychology, 101 104309. doi: 10.1016/j.jesp.2022.104309 * Garcia, Stephen M., Kimberlee Weaver, and Patricia Chen. "The status signals paradox." Social Psychological and Personality Science 10.5 (2019): 690-696. * Podolny, J. M. (2010). Status signals: A sociological study of market competition. Princeton, NJ: Princeton University Press. * Bursztyn, L., and R. Jensen, Social Image and Economic Behavior in the Field: Identifying, Understanding, and Shaping Social Pressure, Annual Review of Economics 2017 9:1, 131-153 * Mason, Roger. "Conspicuous consumption: A literature review." European journal of Marketing 18.3 (1984): 26-39. * Jinkins, David (2016) Conspicuous consumption in the United States and China", Journal of Economic Behavior & Organization, 127: 115-132. * Alvarez-Cuadrado, F., Casado, J. M., Labeaga, J. M. (2016) Envy and habits: Panel data estimates of interdependent preferences", Oxford Bulletin of Economics and Statistics, 78 (4), 443-469. * Heffetz, O. (2011). A test of conspicuous consumption: Visibility and income elasticities. Review of Economics and Statistics, 93(4), 1101-1117. * Frank, Robert H. 1985b. “The Demand for Unobservable and Other Nonpositional Goods.” American Economic Review 75 (1): 101–16. * Clark, Andrew E., Paul Frijters, and Michael A. Shields (2008) \Relative income, happiness,and utility: An explanation for the Easterlin paradox and other puzzles", Journal of Economic Literature 46(1), 95-144. * Hopkins, Ed and Tatiana Kornienko (2004) \ Running to Keep in the Same Place: Consumer Choice as a Game of Status", American Economic Review, September 2004, 94 (4), 1085-1107. * Bramoullé, Y and C. Ghiglino (2022), “DP17181 Loss Aversion and Conspicuous Consumption in Networks”, CEPR Press Discussion Paper No. 17181. https://cepr.org/publications/dp17181 * Festinger, L.. 1954. “A Theory of Social Comparison Processes.” Human Relations 7 (2): 117–40. * van de Stadt, H., A. Kapteyn, and S. van de Geer. 1985. “The Relativity of Utility: Evidence from Panel Data.” Review of Economics and Statistics 67 (2): 179–87. * Luttmer, Erzo F. P. 2005. “Neighbors as Negatives: Relative Earnings and Well-Being.” Quarterly Journal of Economics 120 (3): 963–1002. * Frank, Robert H. 1985b. “The Demand for Unobservable and Other Nonpositional Goods.” American Economic Review 75 (1): 101–16. * Clark, A., P. Frijters, and M. Shields (2008) Relative income, happiness, and utility: An explanation for the Easterlin paradox and other puzzles", Journal of Economic Literature 46(1), 95-144. * Brown, G., J. Gardner, A. J. Oswald, and J. Qian (2008), Does wage rank affect employees' well-being?", Industrial Relations: A Journal of Economy and Society, 47(3), 355-389. * Cooper, David J., and John H. Kagel (2016) "Other-regarding preferences", The Handbook of Experimental Economics, Vol 2, J. Kagel and A. Roth (eds.), Princeton:Princeton University Press. * De Giorgi, G., A. Frederiksen, L. Pistaferri, Consumption Network Effects, The Review of Economic Studies, Volume 87, Issue 1, January 2020, Pages 130–163, https://doi.org/10.1093/restud/rdz026 * Hopkins, E., and T. Kornienko (2004) Running to Keep in the Same Place: Consumer Choice as a Game of Status", American Economic Review, September 2004, 94 (4), 1085-1107. * Ireland, N. J. (2001) Optimal income tax in the presence of status effects", Journal of Public Economics, 81.2, 193-212. * Drechsel-Grau, M., Schmid, K. (2014) Consumption savings decisions under upward looking comparisons", Journal of Economic Behavior and Organization, 106, 254-268. * Bilancini, E., and L. Boncinelli (2012) Redistribution and the notion of social status", Journal of Public Economics 96.9: 651-657. * Moav, O., and Z. Neeman (2012) \Saving rates and poverty: The role of conspicuous consumption and human capital", The Economic Journal, 122(563), 933-956. § MODEL Agents and endowments. There are J agents, with typical agent j. Each agent is endowed with income w_j > 0 and a fixed identity θ_j, from a finite set of possible identities Θ. Further, agents are embedded in a network that is represented by a J × J non-negative matrix G. We say that agent k is a neighbour of j if G_jk > 0, and assume that j is not her own neighbour (so G_jj = 0 for all j).[The network is weighted and directed as we allow G_jk∈ℝ^+ and G_jk≠ G_kj, respectively.] Let I(θ) = { k : θ_k = θ} be the set of agents who have identity θ. Then I(θ_j) = { k : θ_k = θ_j } is the set of agents who have the same identity as j (including herself). Consumption and prototypes. Agents simultaneously choose consumption of a visible good, x_j ≥ 0. Each identity θ has a prototype, Y(θ) = 1/|I(θ)|∑_k ∈ I(θ) x_k, which is simply the average consumption of agents who have that identity. Additionally, each agent j has a reference point, Y_j = ∑_k ∈ I(θ_j) G_jk x_k, which is a weighted sum of the consumption of their neighbours who share their identity. The reference point is similar to the prototype, but (a) only includes people j is connected to in the social network, and (b) allows the weights to be heterogeneous. Benefits and costs: status and dissonance. Agent j gets status benefits that is increasing in her identity's prototype, Y(θ_j), and in her own consumption, x_j. We assume that own consumption and prototypical consumption are substitutes. Agent j also experiences dissonance costs when her consumption differs from her reference point. Specifically, we assume dissonance is convex in the difference between her own consumption and her reference point. Note that j's prototype depends on everyone who shares her identity, but her reference point depends only on her neighbours who share her identity. As is standard, there is also a convex cost of consumption. Both absolute and marginal cost of consumption are decreasing in own income. This cost function captures spending on other goods (which are outside of our model) in a reduced form way; implicitly assuming some concave benefits from consuming the unmodelled goods. Preferences. In order to fix ideas and to show the results more clearly, we impose a particular functional form. We assume agents have the following preferences: u_j = α x_j + Y(θ_j) + γ x_j/Y(θ_j)_Status - β (x_j - Y_j)^2 _Dissonance - 1/w_j x_j^2 _Cost, with α > 0, β > 0, γ > 0. This strong functional form assumption allows for more transparent analysis, especially regarding the role of the network. In <Ref>, we replace <Ref> with a far more general functional form and show that there is still a unique equilibrium. The generalisation does not provide new economic insight, but does suggest that our results are driven by the social identity framework rather than by the particulars of the functional form we chose. Recall that Y(θ_j) = 1/|I(θ_j)|∑_k ∈ I(θ_j) x_k and Y_j = ∑_k ∈ I(θ_j) G_jk x_k. With this functional form, our assumption that the status benefits are increasing in Y(θ_j) may not always be met. We will impose an assumption on the network structure that guarantees it is always met in equilibrium. Economically, it amounts to a requirements that the dispersion in network centrality is `not too high'. As is standard, we will also impose an assumption on the network structure needed to guarantee that our measure of network centrality is well-defined. Before stating these assumptions formally, we first need to introduce some machinery. Machinery. First, only the with-identity links in the network will matter. This is because we assumed that in both the status and dissonance functions, an agent only cares about those who share her identity. So we work with a sub-network containing only the within-identity links. Formally, define G such that G_jk = G_jk·1{θ_j = θ_k} for all j,k, where 1{·} is an indicator function. Second, a variant on Bonacich centrality (a well-known measure of network centrality) will play a key role in our results. The definition is a little terse, but is intuitively much the same as a standard measure of Bonacich centrality, except that the network is weighted by the importance of the dissonance function, β, and by agent's income, w_j. For this, define a network H where H_jk = β w_j/β w_j + 1G_jk for all j,k. It will be important to bear in mind that the network H is a function of G, agents' incomes, and of β. The centrality of agent j is: C^b_j = ∑_k[ ∑_t=0^∞ H^t ]_jk·w_k/β w_k + 1 In the interest of brevity, and because it the centrality measure that matters in our model, we will just refer to this as `centrality'. It will also be convenient to define the average centrality of agents with a given identity. Formally, let C̅^b_θ = 1/|I(θ)|∑_k ∈ I(θ) C^b_k be the average centrality of agents with identity θ. Unsurprisingly, we need centrality to exist and be well-defined. A standard result is that centrality is well-defined if and only if the largest eigenvalue of the matrix is strictly less than one <cit.>. This condition ensures that the infinite sum in <Ref> converges. The result extends straightforwardly to our definition of centrality. |λ_1| < 1, where λ_1 is the largest eigenvector modulus of the matrix H. Finally, we need to ensure that the status benefits are increasing in both arguments in equilibrium. A necessary and sufficient condition for this is that no agent's centrality is `too far' above the average. Formally, it amounts the the following restriction. Assume that for all agents j: C^b_j < 2 Z^3/1 + (1 + α) Z, where Z = α/4C̅^b_θ_j + 1/4· (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8 γ/α)^0.5. Intuitively, each agent's centrality must be less than some (increasing) function of the average centrality.[The right-hand side of the restriction in <Ref> is increasing in C^b_j. This is because centralities must always be (weakly) larger than one – so the same must be true of average centrality.] To our knowledge, there is no strong intuition behind the particular functional form on this restriction. Nevertheless, it is clear that the higher the average centrality, the more dispersion in centralities is possible. We impose <Ref> and <Ref> throughout the paper. We will not restate this for each result. Discussion: the prototype and reference point. Before moving the the results, it is helpful to briefly discuss the prototype and reference point in our model. Existing literature on social identity theory in economics assumes the prototype and reference point are exogenous. A limitation of this assumption is that it allows for inconsistency between what is considered `normal' behaviour for an identity group (the prototype) and the actual behaviour of that group. In the short run, this inconsistency may be natural. But in the long run,it is a demanding assumption – in effect it requires that society never learns that its prototype is incorrect. In contrast, we suppose that both the prototype and the reference point are determined endogenously. That is, our model demands consistency between the prototypical consumption of an identity group and the actual consumption of people in that group. We view this consistency as an important in our setting. As a further step, we allow the reference point to depend on an agent's position in a network. This follows the now-common notion that people compare themselves to those they actually interact with (e.g. <cit.>). The trade-off is that we treat identity as fixed in order to keep the model tractable and ensure clean results. § RESULTS This section analyses behaviour in our model. We characterise equilibrium behaviour, present some comparative statics, and show the link between homophily – the tendency for people to be connected to those similar to themselves – and consumption. These results are of independent interest, and are also needed to to show how the model explains the stylised facts (which we show in the next section). §.§ Solving the Game Having set out the necessary machinery and assumptions in <Ref>, we can immediately characterise equilibrium behaviour. There exists a unique equilibrium. For all j equilibrium consumption is: x_j^* = 1/2( α + γ/1/4αC̅^b_θ_j + 1/4 (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8 γ/α)^0.5) C_j^b. The dissonance costs create pressure to keep up with their neighbours. This creates strategic complementarities in consumption. An agent wants to consume more visible goods when her neighbours consume more. Further, agents' best responses are linear in neighbours' consumption. It is easy to see this by differentiating <Ref> with respect to x_j and rearranging. These two features are the common thread to many models where Bonacich centrality (or, in our case, a generalised version of it) plays an critical role. As is standard in these types of games, the restriction on the eigenvalue of the network is there to ensure that these strategic complementarities are not too strong. If it is violated, then successive best responses leads to unbounded growth in consumption, so there would be no equilibrium. Were we to follow <cit.> and assume that the prototype, Y(θ_j), was fixed exogenously, then the actions would be directly proportional to centrality. The average centrality would not appear. We would simply replace the terms containing average centrality with an exogenous term Y(θ_j). This would be mathematically very similar to a range of network games, including models of social comparisons (<cit.> and <cit.> perhaps being the closest). An additional feature of our model, however, is that actions also depend on the average centrality of agents who share the same identity. Importantly, average centrality works in the opposite direction to own centrality. This is because of assumption there is some substitutability between own consumption and average consumption within an identity. Higher average consumption within an identity raises the status of that identity, all else equal reducing the incentive for a given agent to consume. §.§ Comparative Statics With the exact characterisation of equilibrium behaviour provided by <Ref>, the relationship between centrality and consumption is clear. And some links between the primitives of the model and centrality will come from examining our definition of centrality (<Ref>). Before stating the result, it is helpful to recall that agents only care about those who share their identity. So only the within-identity links that matter. This observation allows us to restrict attention to agents who share an identity when thinking about comparative statics results. It is clear that an agent j is unaffected by others who have a different identity.[This is because it is G, not G, that appears in H. So if agents j,k have different identities, there cannot be a walk between them. This is obviously a consequence of our assumption that each agent has only one identity. Allowing for multiple identities, with some held more strongly than others, would relax this stark benchmark, but at a significant cost to tractability.] The first part of the result shows how income and the network each affect centrality. The second part then shows how centrality affects equilibrium consumption. Assume that agents j and ℓ share an identity and that there is a walk from ℓ to j. (i) C_j^b is strictly increasing in w_ℓ, and in G_ℓ m, for all j,ℓ. (ii) x_j^* is strictly increasing in C_j^b, and (iii) strictly decreasing in C̅^b_θ_j, for all j. Higher income for agent j increases her consumption directly. This is because the marginal cost of consumption falls when w_k rises, so re-optimisation leads to an increase in own consumption. This pushes up her neighbours' reference points, in turn increasing their consumption (due to the strategic complementarities). This then pushes up her neighbours' neighbours' reference points, and so on. This process affects everyone who is connected to k in the network G. Mathematically, this appears in two places. An increase in w_ℓ increases elements in the matrix H, and it also increases w_ℓ/(β w_ℓ + 1) (which appears once when summing over k in <ref>). Adding, or strengthening, links in G has much the same effect. It pushes up agent ℓ's reference point, increasing her consumption. This increases her neighbours' reference points, and hence their consumption, and so on. The second part of the result is due to the dissonance function. The higher an agent's own centrality, the higher her reference point, and hence the more she needs to consume in order to conform with those around her. The third part is due to the substitutability between status derived from her own consumption and status derived from her identity. When average centrality in her identity group is high, then average consumption in her identity group is also high. This in turn brings high status benefits from her identity. And this reduces the marginal benefit of status from her own consumption, due to the substitutability between the two sources of status. Overall, higher average centrality in her identity group pushes her consumption down. §.§ Homophily Proposition <ref>(i) shows the impact of adding (or removing) links from the network. While not directly related to the two main stylised facts – which concern the impact of changes in income – we can characterise the impact of somewhat richer changes to the network structure. Specifically, we focus on the impact of changing a link. This amounts to removing a link between agents j and k, and replacing it with a link between agents j and ℓ.[This is, in effect, adding a link and removing a link. Note that this is not covered by <Ref>(i). That result shows the qualitative impact of adding or removing a link, but does not show the net impact of both together.] We noted above that links between agents with different identities behave very differently from links between agents with the same identity. This has important implications for the role of homophily – the tendency for people to be disproportionately connected ti those similar to themselves, and one of the most ubiquitous features of social networks (see, for example, <cit.> or <cit.> for detailed discussions of homophily). Before going further, we now need a formal definition of homophily. There are many in the literature, but we use the homophily index as one of the simplest available. It requires agents one group, from a set of possible groups. The identities, θ, map neatly onto these groups. The homophily index is specific to an agent and simply measures the proportion of total links that are with agents in the same group. For an agent j in group θ: ϕ_j = ∑_k ∈ I(θ) G_jk/∑_k G_jk It is clear from this definition that changing a link G_jk to a link G_j ℓ only affects the homophily index of agent j. So, when examining the effect of changing a link, we can conflate changes in the homophily index for agent j and homophily more broadly without problems. There are four possible cases when changing a link G_jk to a link G_j ℓ: (1) j,k share an identity and ℓ has a different one, (2) j, ℓ share an identity and k has a different one, (3) j has a different identity to both k and ℓ, and (4) j,k and ℓ all share an identity.[In case (3) it does not matter whether or not k and ℓ have the same identity because the link between k and ℓ (if it exists) is not affected.] It follows easily from the definition that homophily falls in case (1), rises in case (2), and does not change in cases (3) and (4). To see this, first notice that the denominator on the right-hand side of <ref> never changes – this is because we are only changing links. So homophily falls if changing the link reduces the number of within-identity links. Similarly, homophily rises if changing the link increases the number of within-identity links. And finally homophily is unaffected if the number of within-identity links is unaffected. Since it is only within-identity links that matter for equilibrium behaviour (recall that it is the matrix G rather than the matrix G that appears in the definition of centrality), changing a link in the network G in a way that increases homophily adds a links in the network G. This observation allows us to say that increasing homophily increases agents' centralities. The reverse is obviously also true: changing a link in the network G in a way that reduces homophily removes a link in the network G, so centralities fall. It also means that swapping a between-identity link for a different between-identity link has no effect. This is because neither link appear in the matrix G – the one that matters for behaviour. The final case, where we swap one between-identity link for an different between-identity link is a little trickier. Here, both adding the link G_j ℓ and removing G_jk have an impact on behaviour. However, high centrality is driven by being linked to people who themselves have high centrality. So j's centrality rises if swapping the link connects her to someone who has higher centrality. Conversely, if the swap connects her to someone with lower centrality, her own centrality falls. Swap a link G_jk for a link G_j ℓ. (i) if homophily rises, then C_m^b rises for all m ∈ I(θ_j), (ii) if homophily falls, then C_m^b falls for all m ∈ I(θ_j), (iii) if homophily does not change and j has a different identity to both k and ℓ, then centralities do not change (iv) if homophily does not change and j,k, ℓ have the same identity, then C_m^b rises for all m ∈ I(θ_j) if and only if C_ℓ^b > C_k^b. § EXPLAINING THE STYLISED FACTS The main aim of this section is to show how the model explains the two stylised facts set out by <cit.> (so far explained by signalling models) and <cit.> (so far explained by social comparisons models). We then also show how the model explains a related stylised fact set out by <cit.>: that people with higher income are more sensitive to changes in other's income.[A large part of the contribution in <cit.> is to extend the analysis in <cit.> to a lower income country – South Africa. He replicates the findings for Black South Africans (who make up the large majority of the South African population), but not for White South Africans. But the finding regarding sensitivity is novel to <cit.>.] While the comparative statics from the previous section link the primitives of the model to centrality, and centrality to behaviour, an important limitation is that they do not cleanly link the primitives directly to behaviour. This is because a change in agents' incomes, or in the network structure, affects both a given agent's centrality and, in turn, the average centrality in their identity group. But own centrality and average centrality in the identity group impact behaviour in different directions. So in general it is not obvious which of these two forces dominates. We therefore focus on a stark benchmark network where agents are split up into disconnected groups – sometimes called an `Islands model'. This allows us to isolate the two effects, and cleanly show how we get behaviour consistent with the first stylised fact when one of these effects wins out, and behaviour consistent with the second stylised fact when the other effect wins out. In this benchmark network, agents with a given identity θ are separated into N different partitions such that there are no links between agents in different partitions. In other words, each agent is on an “island” and there are no links between different islands. In order to clarify the analysis we also assume that each partition is strongly connected within G and that there are many partitions.[A walk in G from j to k is a sequence of agents j, j',...,k',k such that G_j'k'>0 for all j',k' in the sequence. A network is strong connected if there exists a walk from j to k for all j and all k.] Agents are partitioned into N equal-sized, disjoint subsets, called `islands' with typical island n. * There are many islands (N is large). * For any pair of agents j,k with the same identity: (i) if j and k are in different islands, they are not connected (ii) if j and k are in the same island, there is a walk from one to the other in G This stylised benchmark is helpful because changes to primitives affect an agent's own centrality if and only if they occur within her partition. Changes that happen in other partitions affect the average centrality but not her own centrality. And changes that happen within her partition affect only her partition. Because there are, by assumption, many partitions, the impact on her own centrality dominates the effect on average centrality. This feature allows us to isolate the two forces. When a change happens within an agent's partition, the impact on her own centrality with dominate the effect on average centrality. Conversely, when a change happens outside of an agent's partition, there is an impact on average centrality but not on own centrality. Before stating our result, it is helpful to quickly recap our two main stylised facts. This will help make the link between them and the behaviour predicted by our model easier to see. * People in the US consume less visible goods when the average income of others of the same race in their state is higher <cit.>. * People in the Netherlands consume more when the average income of their geographic neighbours increases <cit.>. With these in mind, we can now state our result. Consider an islands network. Suppose income increases for some set of agents with identity θ in partition n. (i) Consumption decreases for all agents with identity θ in partitions other than n, (ii) Consumption increases for all agents with identity θ in partition n. The mapping from <Ref> into the two stylised facts is fairly direct. An increase in income in other partitions corresponds to higher average income of others of the same race in the same state.[We use the term `income' in our model for convenience. Where appropriate, this should be taken to mean the `permanent income' examined in the empirical analysis.] <Ref> then suggests the reason for this observed behaviour is that agents whose identity group has higher average consumption gains more status from their identity. They then feel less need to seek status through their own consumption of visible goods. The mapping is obviously not perfect. There could be changes to the income distribution that affect both the agent's own partition and other partitions. <Ref> is silent in this case because the comparative static is ambiguous. An increase in income within an agent's partition – i.e. amongst her neighbours – causes her to increase her consumption. This is because the higher consumption by her neighbours (which is due to their own higher income) raises the bar for what is considered normal. So dissonance costs push her to consume more herself. Again, the mapping is not perfect. <Ref> only applies to agents who share the same identity, and there is no absolute guarantee that geographic neighbours in fact have the same identity. But there are two main reasons to believe this result is useful. First, there is well-documented prevalence of homophily and sorting along geographic lines <cit.>. So it is likely that an agent's neighbours actually do share her identity. Second, the formal result in fact only requires that an agent only shares her identity with one person whose income increased. §.§ The third stylised fact Closely related to our first stylised fact, <cit.> finds that a given change in reference group income has a larger impact on the conspicuous consumption of richer agents than of poorer agents. In other words, richer agents are more sensitive to changes in others' income. This relationship is monotonic – the richer the agent, the more sensitive they are. Our model is completely consistent with this stylised fact. It is relatively straightforward to see from the characterisation of equilibrium consumption in <Ref>. Changes in the income of agents on other islands only affects average centrality, C̅^b_θ_j. The marginal impact of a change in average centrality, and hence of a change in the income of agents on other islands, is increasing in an agent's own centrality. This is because own centrality, C_j^b, multiplies a function of average centrality in <Ref>. And own centrality is increasing income in own income (<Ref>). Putting this together yields the result: richer agents are more sensitive to changes in the income of other people (for tractability, our formal result restricts this to people on other islands). Suppose j and k are on different islands. The marginal effect of k's income on j's consumption is increasing in j's own income. Intuitively, this is connected to the fact that agents with higher income have a lower marginal cost of consumption, and this marginal cost rises more slowly. When the marginal status benefits of consumption rises, as happens when the prototype Y(θ_j) decreases, it should not come as a surprise that agents increase their consumption in response. But by how much they adjust their consumption depends on how quickly the marginal cost rises as consumption rises. It is precisely higher income agents whose marginal costs rise more slowly, and so who make greater adjustments. Again, it is useful to note that our model does not rely critically on the income of the very poorest agent (as in the signalling model used by <cit.> and <cit.>). Rather, it focuses on the average income more directly. § INEQUALITY We now turn to the role of income inequality – a natural question for a model of social identity and conspicuous consumption. However, a key barrier to tractable results is the multiple ways income affects centrality. Recall that in <Ref> H is a function of the w_j's. This means that higher income of one agent plays an analogous role to having stronger (outward) links. So increasing the income of some agents and lowering the income of others is in a sense equivalent to makes some links stronger and others weaker. It is in general difficult to characterise the overall impact this complex change to the network structure has on centrality. Compounding this, income also appears outside of H as a weighting. In most network games where Bonacich centrality plays a role, only the network structure appears in H, and income does not weight its components.[See, for example, <cit.>.] Given this, we take two different simplifying approaches. They give complementary views of the role of inequality. First, we use the Islands Network from <Ref> and assume that all agents on a given island have the same income. This simplifies the link between income and centrality, while retaining rich network effects. It allows us to focus more clearly on the interaction of network structure and inequality. The second approach imposes a regularity condition on the network. It assumes agents are equally well connected to each different `level' of income. This simplifies the network effects, while retaining a richer link between centrality and income. It allows us to focus on the role of income inequality directly, abstracting from the more complex network effects. §.§ Island Networks The simplifying assumption. Here, we use the islands model from the previous section (<Ref>) and add the assumption that income is homogeneous within each island. This simplifies matters significantly as changes in income on an island must all be in the same direction. It also gives equal weight to all components of the infinite sum.[The islands model is critical here – it creates disconnected components with no interactions between them. When calculating centrality, we can consider each island in isolation – giving us most of the benefits of homogeneous income from a tractability perspective, while retaining significant heterogeneity in society as a whole. The assumption that all islands are the same size also helps make results cleaner here: it allows a transfer of income from one island to another to have a symmetric impact, in the sense of raising per-capita income on one island by the same amount as it reduces per-capita income on the other.] While the assumption shuts down the possibility of inequality on a given island, it still permits significant inequality in society as a whole (as there are many islands). Two building blocks. At this point, it is helpful to define a property of the network and show a key implication of our simplifying assumption. They will be useful for the substantive results. The density of the network on an island n is: D_n = ∑_j,k[ ∑_t=0^∞G^t ]_jk Readers familiar with network games where Bonacich centrality plays a critical role will notice that our definition of `density' is the sum of centralities of all agents on the island (using the `standard' definition of Bonacich centrality that only involves the network, rather than our variant that also includes income). It is not exactly density as conventionally defined. Nevertheless, it captures some measure of how closely connected agents on an island are on aggregate. As some reassurance, adding links within the island always increases our density metric. The islands network guarantees that agents are only connected to – both directly and indirectly – others on the same island (at least amongst with the same identity; as discussed earlier links to those with a different identity are ignored and so play no role). So our assumption that income is homogeneous within each island means that everyone an agent is connected to – both directly and indirectly – has the same income. It turns out that this drastically simplifies the equation for centrality in <Ref>. Suppose the network is an islands network (a la <Ref>), all islands are the same size and all agents on a given island, P_n, have the same income, w_n. Then C_j^b = w_n·∑_k (I - G)^-1_jk for all j ∈ P_n. So in this special case, centrality is linear in income. This result is useful because it separates out the impact of income from the impact of the network. We now know how income affects centrality subject to these benchmark assumptions, and we know from earlier how centrality affects consumption. This allows us to analyse the impact of redistribution in a much more tractable manner. Impact of redistribution. We focus on a particularly simple form of redistribution – a transfer from one island to another, keeping average income in society constant. It is reduces income inequality if the island receiving the transfer is poorer than the island giving the transfer (and similarly increases income inequality if the island receiving is richer).[We will not consider the case where the transfer changes which island is richer. In that case, the impact on inequality is ambiguous.] Richer changes to the income distribution can obviously be built from a series of these changes. Interestingly, the impact of this transfer on overall consumption does not depend on whether it increases or decreases inequality. Rather, it depends on the relative densities of the networks on the two islands that are party to the transfer. If the island receiving the transfer has a denser network, then the increase in centrality there is greater than the fall in centrality on the island giving the transfer. Then, average centrality rises, which in turn reduces consumption on all other islands. Consider a transfer of income from island n to island n'. (i) Consumption falls on island n and rises on island n', (ii) Consumption on all other islands rises if and only if D_n' < D_n. The first part of this result is unsurprising. When everyone on an island has more income, they consume more. There is a direct effect of higher income, due to the lower marginal cost of consumption. And there is an indirect effect through the dissonance cost, as higher consumption by others on the island raises the references level an agent feels pressure to conform to. Recall that there is also a third effect – higher consumption by others in an agent's identity group raises her status, and in doing reduces the marginal status benefit of her own consumption. But with many islands, this effect is dominated by the other two. Everything works in reverse when everyone on an island has less income. For all the islands whose own income remains unchanged, so the only force that applies here is how the transfer from island n to n' affects average centrality. Average centrality – and hence consumption – rises if and only if the (positive) effect on island n' is larger than the (negative) effect on island n. It is the simplified characterisation of centrality in <Ref> allows us to determine this. From it, we can see that the marginal impact of a change in income on an agent's centrality is some measure of their network connectedness. The density measure in <Ref> calculates the total of this `connectedness' across all agents on the island. So the density measure captures the marginal impact on total centrality of changing income on an island. From this it is clear that average centrality rises if and only if the marginal (positive) impact on island n' is larger than the marginal (negative) impact on island n. §.§ The fourth stylised fact The fourth stylised fact also comes from <cit.>, and relates to inequality. They find that higher dispersion in income (i.e. higher inequality) among Whites in a given state leads to lower consumption of visible goods (i.e. conspicuous consumption) by Whites in that state. But they find the reverse for Blacks: higher dispersion in income among Blacks in a given state leads to higher consumption of visible goods by Whites in that state.[Note that there appears to be a typographical error in <cit.>. The sentence states they “find that greater dispersion of reference group income is associated with lower visible spending for minorities”. But this conflicts with the regression results, reported in Table VII (column 2), on which the paragraph is based.] <Ref> can help us understand this seemingly puzzling empirical finding. According to the result, a change in inequality an either raise or lower agents' consumption. The direction of the effect depends critically on the relative network density of the agents whose incomes change. If the agents whose incomes rise have denser social networks than the agents whose income falls, then everyone else's consumption falls (by `everyone else', we mean the agents whose income did not change). And the same applies equally in reverse. Therefore, we can rationalise the empirical finding as being due to differences in network structures across different racial groups. One drawback of this explanation is that it relies on the density of agents' social networks, which is likely to be challenging to measure empirically. Nevertheless, we view it as an advance on the explanation proposed by Charles et al's signalling model, which must appealing to racial differences in preferences to explain this empirical finding. Note that <cit.> repeatedly stress shortcomings of an explanation relies on racial differences in preferences. They say that “[t]his argument is consistent with the basic facts, but is essentially tautological. Moreover, an argument centered on racial differences in preferences yields no prediction that is falsifiable in the data.” So despite empirical challenges, we at least progress beyond a reliance on racial differences in preferences. §.§ A Regularity Condition The simplifying assumption. Our second approach imposes a regularity condition on the network. We need each agent to have the same number of neighbours (in the weighted sense) with those of each different income level. For example, if there are two different level of income, w_1 and w_2, then each agent must have L neighbours with income w_1, and also L neighbours with income w_2.[Note that if every agent in the network has a unique income level, then this assumption requires that the network is complete. But we believe that there being a small number of different levels of income is a reasonable first-order approximation of reality.] Note that this is stronger than assuming the network is regular (which would only require that each agent has L links in total, regardless of the neighbours' incomes). To formalise this, assume there are M different levels of income, with typical income level w_m. Then let ℳ = { j : w_j = w_m } denote the set of agents with income w_m. The assumption is then: Each agent j has exactly L links to agents k ∈ℳ, for each ℳ.[This is in the weighted sense. This is a restriction on the total weight of links from agent j to each group 𝒵. It does not matter if there are a small number of strong links, or a large number of weak links.] Intuitively, <Ref> ensures that how strongly j is connected to others is unrelated to how much income those other people have. Colloquially, she is, on average, connected to “Mr and Mrs Average”. This retains the network effects, but “smooths them out” so that they apply equally throughout the network. A building block. Similar to our use of the islands network with homogeneous income on each islands, <Ref> allows us to find a significantly simplified expression for agents' centralities. In turn, this will allow us to analyse the impact of changing inequality much more precisely. Under <Ref>; C_j^b = βΓ_j / (1 - Γ̅), where Γ_j = w_j/β w_j + 1 Γ̅ = 1/J∑_k Γ_j. In contrast to <Ref>, an agent's centrality here is concave in her own income.[If J=1 (i.e. there is only one agent in total), then an agent's centrality is linear in her own income. But thin this case, there is no network, so this should not be a surprise. And related to <Ref>, centrality is also linear in income if all agents have the same income.] This is because we are allowing agents to interact with others who have different incomes. It shows that interactions between people with different incomes is an important part of the role networks have in determining consumption. Impact of redistribution. Again, we focus on a transfer of income from agents with income w_z to agents with income w_z'. This is the same as our transfer in the previous subsection, except the network structure is different – here agents with different incomes interact with one another in the network. It reduces in equality if the agents receiving the transfer are poorer than the agents giving the transfer (and similarly increasing inequality if the agents receiving are richer). Here, the impact of the transfer depends critically on whether it increases or decreases inequality (i.e. whether it's regressive or progressive). This is closely linked to the concavity we can see in <Ref>. Because Γ_j = w_j/β w_k + 1 is concave in w_j, a given change in income has a larger impact on lower income agents. So the mean, Γ̅, rises when there is a regressive transfer of income (from a richer group to a poorer group). This then pushes the centrality of agents not party to the transfer down. Finally, this pushes average centrality down and hence consumption up. Consider a transfer of income from agents in group ℳ to agents in group ℳ' such that the ranking of the two groups' income does not change. (i) Consumption falls for agents in ℳ and rises for agents in ℳ'. (ii) Consumption of all other agents rises if and only if the transfer increases inequality (i.e. w_m' > w_m). (iii) Average consumption rises if and only if the transfer reduces inequality (i.e. w_m' < w_m). An agent consumes more when her income rises (part (i)). This is because her own income effect dominates the network effects. This is standard. But consumption is concave in her own income.[This is fairly transparent from the equation in <Ref>, but we provide a proof in the appendix for completeness.] Intuitively, this is linked to the convex cost function (<Ref>). So a regressive transfer (one which makes the rich richer and the poor poorer, increasing inequality) reduces average consumption. But this reduction in average consumption has a spillover effect for everyone else: lower average consumption reduces the status from the identity group. This drop in status induces agents who were not party to the transfer to increase their own consumption to compensate – due to the substitutability between status from own consumption and status from the identity (part (ii)). Finally, part (iii) notes that the increase in consumption by the agents not (directly) affected by the transfer cannot dominate the decrease in consumption by the directly affected ones. To see why this must be the case, bear in mind that the agents not party to the transfer are only responding to a fall in average consumption. If their own response led average consumption to be higher than before the transfer, then they would respond to that by reducing their consumption – leading to a contradiction. The key insight is that a transfer which increases inequality has an important impact on agents whose incomes do not change. This is because agents derive status from consumption by others who share their identity. And this `status identity' is substitutable with status derived from their own consumption – so lower status identity leads them to compensate with more of their own consumption. The result apes existing results in the conspicuous consumption literature, but our mechanism is novel. abbrvnat tocsectionReferences equationsection thmsection propsection defnsection remsection lemsection corsection § PROOFS §.§ Proposition 1 First, we find the best response function, holding Y(θ_j) constant (the first derivative is linear and decreasing in x_j, and is positive when x_j = 0, so a unique solution exists). Note that we have substituted in Y_j = ∑_k ∈ I(θ_j) G_jk x_k = ∑_k G_jk x_k. x_j^* = 1/2( β w_j + 1/w_j)^-1 [α + γ Y(θ_j)^-1 + 2 βG_jk x_k] Some rearranging yields: x_j - β w_j/β w_j + 1G_jk x_k = 1/2w_j/β w_j + 1 [α + γ Y(θ_j)^-1] Then recall that we defined the matrix H such that H_jk = β w_j/β w_j + 1G_jk. We now substitute this in to the equation and switch to matrix notation. (In an abuse of notation use the same notation for a vector of constants as for the equivalent scalar, and we let 𝐖 denote the vector where the j^th entry is w_j/β w_j + 1. And recall that ⊙ is the Hadamard, or element-wise, product.) 𝐱 - H 𝐱 = 1/2𝐖⊙ [α + γ Y(θ)^-1]. We have assumed that |λ_1| < 1 (<Ref>),[Where λ_1 is the largest eigenvector modulus (i.e. the largest in magnitude) of the matrix H.], so (I - H) is invertible (<cit.>). Therefore: 𝐱 = (I - H)^-1·1/2𝐖⊙ [α + γ Y(θ)^-1]. Reverting to scalar notation: x_j = ∑_k [(I - H)^-1]_jk·1/2w_j/β w_j + 1 [α + γ Y(θ_k)^-1]. A comment: by construction, G is a block matrix with off-diagonal blocks being all zero. This is because an agent j can only have a link to agent k if j and k share the same identity (i.e. if θ_j = θ_k). Therefore H is also a block matrix. As a result, H_jk will be zero whenever j and k have different identities (i.e. whenever θ_j ≠θ_k). An important consequence of this is that θ_k = θ_j whenever [(I - H)^-1]_jk > 0. This means we can treat the term Y(θ_k)^-1 as a constant an pull it out of the sum. So x_j = 1/2 [α + γ Y(θ_k)^-1] ·∑_k [(I - H)^-1]_jk·w_j/β w_j + 1, and we then using <Ref> gives us: x_j = 1/2 [α + γ Y(θ_j)^-1] C_j^b. This would give equilibrium actions for a fixed value of the prototype. But Y(θ_j) is endogenous. Recall that by definition Y(θ_j) = 1/|I(θ_j)|∑_k ∈ I(θ_j) x_k. Hence, by substituting our equation for consumption into this definition: Y(θ_j) = 1/|I(θ_j)|∑_k ∈ I(θ_j)1/2 [α + γ Y(θ_j)^-1] C_k^b Y(θ_j) = 1/2 [α + γ Y(θ_j)^-1] ·1/|I(θ_j)|∑_k ∈ I(θ_j) C_k^b. The term after the dot is, by definition, C^b_θ_j (the average Bonacich centrality of agents with identity θ_j, see definition immediately after <ref>). Using this definition and rearranging terms yields: 2 Y(θ_j)^2 - αC̅^b_θ_j Y(θ_j) - γC̅^b_θ_j = 0 Then apply the usual quadratic formula and ignore the negative root (because we rule out x_j < 0 by assumption and hence cannot have Y(θ_j)<0). Y(θ_j) = 1/4αC̅^b_θ_j + 1/4 (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8 γ/α)^0.5 Finally, substitute in this equilibrium value of Y(θ_j) into our equation for x_j^*. x_j^* = 1/2( α + γ/1/4αC̅^b_θ_j + 1/4 (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8 γ/α)^0.5) C_j^b §.§ Proposition 2 (i) An increase in w_ℓ strictly increases w_k / (β w_k + 1) when k = ℓ, and leaves it unchanged whenever k ≠ℓ. It also weakly increases all elements of the matrix H. Because there is a walk from ℓ to j by assumption, the (j ℓ)^th entry of the infinite sum is strictly positive. It is then clear from the definition (<Ref>) that centrality strictly increases. An increase in G_ℓ m strictly increases the (ℓ m)^th entry of the infinite sum. Because there is a walk from ℓ to j by assumption, this also increases the (j ℓ)^th entry of the infinite sum. (ii) and (iii) follow straightforwardly from the characterisation of equilibrium consumption in <Ref> (<Ref>). §.§ Corollary 1 Notice that adding/removing/changing links that appear in G but not in G has no impact on consumption. (i) for the link swap to increase homophily, it must be the case that ℓ has the same identity as j, but k does not (i.e. ℓ∈ I(θ_j) and k ∉ I(θ_j)). This means the link G_j ℓ also appears in G, but the link G_jk does not. So removing the link G_jk has no impact, and this becomes equivalent to just adding G_j ℓ. With this observation, the result follows from <Ref>(i). (ii) inverse of part (i). (iii) Follows from the observation that links not in G have no impact, and neither G_jk nor G_j ℓ appear in G. (iv) This result is essentially a restatement of Proposition 3 in <cit.> (and a generalisation, Proposition G.4, ibid). It follows from the proofs of those results. §.§ Corollary 2 The marginal effect of others' income on an agent j's consumption is d x_j^*/d w_k. When j,k are on different islands, then w_k does not affect C_j^b (<Ref>). Using this, plus the equation for x_j^* from <Ref> yields: d x_j^*/d w_k = 1/2 C_j^b ·d/d w_k( γ/1/4αC̅^b_θ_j + 1/4 (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8 γ/α)^0.5). We know from <Ref> that C_j^b is strictly increasing in w_j. And because there are many islands (by assumption, in <Ref>), w_j has a negligible impact on C̅^b_θ_j. Therefore d x_j^*/d w_k is increasing in w_j. §.§ Proposition 3 We prove this in two steps. First, we show the link between income and centrality. Second, we show the link between centrality and consumption. Step 1. Notice that under <Ref>, each island is a disconnected component in the network G. Therefore, we can treat each island separately. So we can replace G with G_n, which is the graph for island n only.[This is because the matrix for an islands network can be arranged as a diagonal block matrix, and it is well known that the inverse of a diagonal block matrix is the inverse of each block arranged along the diagonal.] (i) Then it is clear that wealth of agent j, w_j, does not affect the centrality of agent k, C_k^b, if agents j and k are only different islands. This is because the definition of C_k^b does not contain w_j. (ii) <Ref> shows that C_ℓ^b is increasing in w_k if there is a walk from j to ℓ. For two agents on the same island, this walk exists by assumption (<Ref>). (iii) An increase in w_j increases Bonacich centrality for some agents, and leaves centrality unaffected for the rest. Therefore an increase in w_j must increase average Bonacich centrality of agents with the same identity as j, C^b_θ_j. Step 2. The result then follows from <Ref>. (i) We know from Step 1 that for all agents k on a different island to j, an increase in w_j increases C^b_θ_j = C^b_θ_k, but does not affect C_k^b. It is then clear from <Ref> that x_k^* is decreasing in w_j. (ii) We know from Step 1 that for all agents k on the same island to j, an increase in w_j increases both C^b_θ_j = C^b_θ_k and C_k^b. The assumption (in <Ref>) that there are many islands means that the impact on C^b_θ_j = C^b_θ_k must be small relative to the effect on C_k^b. Therefore the increase in C_k^b (which pushes x_k^* up) must dominate the increase in C^b_θ_j = C^b_θ_k (which pushes x_k^* down). §.§ Lemma 1 Start with the definition of Bonacich centrality (<Ref>). First, notice that under <Ref>, each island is a disconnected component in the network G. Therefore, we can treat each island separately. So we can replace G with G_n, which is the graph for island n only. Second, recall that all agents on island n have the same income, w_n. C^b_j = ∑_k[ ∑_t=0^∞( β w_n/β w_n + 1G_n )^t ]_jk·w_n/β w_n + 1. Now the w_n/β w_n + 1 term at the end (far right) can be pulled out the front, and the same term that appears within the infinite sum can be separated from the G_n term, and can also be pulled out of the summation over k. C^b_j = w_n/β w_n + 1·∑_t=0^∞( β w_n/β w_n + 1)^t ·∑_k[ ∑_t=0^∞G^t ]_jk Then apply the standard closed form solution for the infinite sum of a geometric series (which always exists because w_n/β w_n + 1 < 1 by construction), and notice that the result (partially) cancels with the w_n/β w_n + 1 term. This yields: C^b_j = w_n ·∑_k[ ∑_t=0^∞G^t ]_jk Finally, to complete the proof, the Neumann series representation gives the fact that ∑_k[ ∑_t=0^∞G^t ]_jk = ∑_k (I - G)^-1_jk. §.§ Proposition 4 The proof follows the same structure as the proof to <Ref>. It follows straightforwardly from <Ref> that centrality rises on island n' and falls on island n. Next, we can express average Bonacich centrality as the sum over agents on each island, then summed over islands (rather than the sum over agents directly), finally divided by J (the number of agents in society). That is: C^b_θ = 1/J∑_n ∑_j ∈ P_n w_n ∑_k (I - G_n)^-1_jk. Then, we can pull w_n through the middle summation, and then substitute in for <Ref>. This yields: C^b_θ = 1/J∑_n w_n D_n It is then clear that a transfer from island n to n' increases average Bonacich centrality if and only if D_n' > D_n. (i) For the islands n and n', the effect of own centrality dominates the effect on average centrality (because there are many islands). Centrality, and hence consumption, rises [resp. falls] for island n' [resp. n]. (ii) For the other islands, own centrality does not change. Only average centrality matters. The result then follows from the claims shown above, and the observation that consumption is decreasing in average centrality. <Ref> implies that ∑_k Γ_k G_jk - 1/J∑_k Γ_k ·∑_k G_jk = 0, where Γ_k ≡w_k/β w_k + 1.[This is equivalent to corr(G_jk, Γ_k) = 0.] Begin with ∑_k Γ_k G_jk, and notice that Γ_k is a constant for all k ∈𝒵. Therefore, ∑_k Γ_k G_jk = ∑_z Γ_z ∑_k ∈𝒵G_jk. Then <Ref> requires that each agent j has exactly L links to agents with income w_z, for each z. This means ∑_k ∈𝒵G_jk = L for all j and for all z. Therefore ∑_k Γ_k G_jk = L ∑_z Γ_z. Next, it must be the case that 1/J∑_k Γ_k = 1/Z∑_z Γ_z (these are just two different ways of calculating the average value of Γ, and J ≡ n · Z). Finally, using the assumption that ∑_k G_jk = 1 for all j, and the implication that L = 1/Z. Together, these yield ∑_k Γ_k G_jk = 1/J∑_k Γ_k ·∑_k G_jk §.§ Lemma 2 We begin by restating the definition of Bonacich centrality (<Ref>), with Γ_k ≡w_k/β w_k + 1: 1/β C^b_j = ∑_k[ ∑_t=0^∞( ΓG)^t ]_jk·Γ_k. We have moved the β term to the left-hand side simply to avoid carrying it around in the calculations. Expand out the infinite sum and consider each term. We deal with the first two terms individually, and then all further terms by induction. Clearly ∑_k Γ_k 1{ j=k } = Γ_k. Then ∑_k Γ_k (ΓG)_jk = ∑_k Γ_k Γ_j G_jk = Γ_j ∑_k Γ_k G_jk. Apply <Ref>, and the assumption that ∑_k G_jk = 1. This yields: = Γ_j ·Γ̅. Induction: for t=2; ∑_kΓ_k (ΓG)^2_jk = ∑_kΓ_k ∑_s (ΓG)_js (ΓG)_sk = ∑_kΓ_k ∑_sΓ_jG_jsΓ_sG_sk = Γ_j∑_sΓ_sG_js∑_kΓ_k G_sk By <Ref> and the assumption that ∑_jG_sj = 1: ∑_sΓ_s G_js= Γ∑_sG_js = Γ. Applying this twice (to the sum over k and then to the sum over s) yields: ∑_k (ΓG)^2_jk = α_j·Γ^2. Now assume that for t=τ > 2, ∑_k (ΓG)^τ_jk = Γ_j·Γ^τ. Then show for t=τ+1. By definition ∑_k (ΓG)^τ+1_jk = ∑_j [(ΓG) · (ΓG)^τ]_jk. For clarity of exposition, we let (ΓG)^τ_jk = 𝒜_jk. ∑_k (ΓG)^τ+1_jk = ∑_k∑_sΓ_jG_js𝒜_sk = Γ_j∑_sG_js∑_k𝒜_sk Using the assumption for t=τ, namely ∑_k𝒜_sk = Γ_sΓ^τ, yields: ∑_k (ΓG)^τ+1_jk = Γ_j·Γ^τ∑_sG_isΓ_s. Applying <Ref> and the assumption that ∑_jG_sj = 1 once more yields Γ_j·Γ^τ + 1. By induction, we have calculated each term in the infinite sum. Having calculated each term in the infinite sum, we can now express it more simply: 1/β C_j^b = Γ_j (1 + Γ + Γ^2 + Γ^3 + ...) Therefore, finally: C_j^b = β Γ_j/1 - Γ. §.§ Proposition 5 Consider a transfer of income such that: (i) w_z ↑, w_z'↓, (iii) w_z + w_z' unchanged. Now notice that Γ_j is increasing and concave in w_j (by the definition of Γ_j). Step 1. So the transfer causes: (i) Γ_z ↑, (ii) Γ_z'↓, and (iii) (Γ_z + Γ_z') ↑ if and only if w_z < w_z' [and hence Γ̅↑ if and only if w_z < w_z', because the transfer does not affect Γ_z” for any z”≠ z, z']. Step 2. In turn these imply: (i) C^b_z↑, (ii) C^b_z'↓, and (iii) C̅^b_θ↑ if and only if w_z < w_z'. Note that for (i), C_z^b is increasing in both Γ_z and Γ̅, but these may move in different directions. However, there are many islands by assumption, so the the transfer must have a much larger impact on Γ_z than on Γ̅, so the effect of Γ_z dominates. The same argument applies to (ii). Step 3. Finally, recall the definition of x_j^* from <Ref>. Our findings in Step 2 in turn imply that: (i) x_z^* ↑, (ii) x_z'^* ↓, and (iii) x_z”^* ↑ if and only if w_z < w_z'. Again, point (i) [resp. point (ii)] rely on the assumption that there are many islands to ensure that the effect of the change in C^b_z [resp. C^b_z'] dominates the effect of the change in C^b_z'. §.§ Additional Proofs Here, we formalise the claims we made about the necessity and sufficiency of Assumptions <ref> and <ref>. (i) Bonacich Centrality exists and is well defined if and only if <Ref> holds. (ii) In equilibrium, α x_j + Y(θ_j) + γ x_j/Y(θ_j) is increasing in Y(θ_j) if and only if <Ref> holds. (i) Bonacich centrality is well-defined if and only if the infinite sum ∑_t=0^∞( β/β + 1/𝐰G)^t converges. In this case then, by a Neumann Series representation, it equals (I - β/β + 1/𝐰G)^-1. <cit.> then shows that this matrix inverse is well-defined and non-negative if and only if |λ_1| < 1. (ii) Start with the equilibrium status benefit S = α x_j^* + Y^*(θ_j) + γ x_j^* Y^*(θ_j)^-1. Then substitute in x_j^* = 0.5 (α + Y^*(θ_j)^-1) C_j^b (shown in the proof to <Ref>), and take partial derivatives with respect to Y^*(θ_j). This (plus some rearranging) yields: d S/d Y^*(θ_j) = 1 - 0.5 C_j^b (Y^*(θ_j))^-2 (1 + α + (Y^*(θ_j))^-1) Some further rearranging shows that this is greater than zero if and only if C_j^b < 2 (Y^*(θ_j))^3/1 + (1+α) Y^*(θ_j). Noticing that Y^*(θ_j) = α/4C̅^b_θ_j + 1/4· (αC̅^b_θ_j)^0.5· (αC̅^b_θ_j + 8/α)^0.5 completes the result (this is shown in the proof to <Ref>). § NON-PARAMETRIC MODEL This section generalises the main model from <Ref> by relaxing the functional form assumptions made in <Ref>. It first restates the model in its more general form, and then shows that this generalised model has a unique equilibrium. There is no new economic insight gained by doing this. But it aims to provide reassurance that the results are driven by the general social identity framework and not by the particulars of the functional forms we chose. Model. The model is as in <Ref>, except we replace <Ref> with the following, more general, utility function for a typical agent: u_j = S(Y(θ_j), x_j) - D(x_j - Y_j) - C(x_j, w_j) where all functions are continuously differentiable with S'_1>0, S'_2>0, S”_22≤ 0, S”_12<0, C'_1>0, C”_11 > 0 C'_2<0, C”_12<0, D'>0 if x_j>Y_j, D'<0 if x_j<Y_j, and D”>0. Equilibrium. We now show that the model has a unique equilibrium. Without an explicit functional form, the strategy we used to prove <Ref> will not work here. We prove it in several steps. First, we show every agent always has a unique best response (taking as given the actions of all other agents), and characterise it. Second, we show bounds on the derivatives of the best response with respect to the agent's reference point. Specifically, we show an agent responds less than one-for-one to a change in her reference point. Third, we state a variant of the Mean Value Theorem that applies to vector valued functions. Finally, we use steps two and three to show the best responses are a contraction mapping and use the Banach Fixed Point Theorem to complete the proof. There exists a unique equilibrium. Step 1: Unique Best Responses. For any agent j, a unique best response always exists. Consider the first and second partial derivatives with respect to x_j. d u_j/d x_j≡ Z = S'_2 + S'_1/|I(θ_j)| - D' - C'_1, d^2 u_j/d x_j^2≡d Z/d x_j = S”_22 + 2 S”_12/|I(θ_j)| + S”_11/|I(θ_j)|^2 - D” - C”_11, and recall that S”_22≤ 0, D” > 0, and C”_11 > 0.[Notice that because Y(θ_j) = 1/|I(θ)|∑_k ∈ I(θ) x_k by definition, x_j appears in Y(θ_j). However, as j's identity group becomes large, the impact this has becomes small. In contrast, x_j does not appear in Y_j.] Hence d^2 u_j / d x_j^2 < 0, and so the first derivative is monotonically decreasing in x_j. Therefore there is a unique best response (i.e. x_j^* that solves d u_j / d x_j = 0). Step 2: Bounding responses. An agent's best response changes less than proportionately in the actions of other agents. We want to find d x_j^*/d x_k, but we don't have an expression for x_j^*. Nevertheless, we can find it from some things we do have expressions for. Suppose we start at a best response, and then x_k increases by Δ units (where Δ is small). Then Z increases by Δd Z/d x_k. To restore a best response (i.e. return Z to 0), we need x_j to `undo' that increase in Z. Since Z is decreasing in x_j, it must increase by γΔ units, such that γΔd Z/d x_j + Δd Z/d x_k = 0. Therefore γ = d B_j(𝐱_-j)/d x_k = - d Z/d x_k / d Z/d x_j We have an expression for Z, and hence can express d Z/d x_k and d Z/d x_j as follows: d Z/d x_k = G_jk· D” + S”_12/|I|·S”_11/|I|^2 > 0, d Z/d x_j = - D” + 2 S”_12/|I| + S”_11/|I|^2 - C”_11 < 0. Therefore, for all j,k: d B_j(𝐱_-j)/d x_k = G_jk D” + S”_12/|I|·S”_11/|I|^2/ D” - 2 S”_12/|I| - S”_11/|I|^2 + C”_11 Upper bound. Since S”_12≤ 0 and S”_11≤ 0 by assumption, we must then have d B_j(𝐱_-j)/d x_k≤ G_jk· D”/ D” - 2 S”_12/|I| - S”_11/|I|^2 + C”_11 Then recall that C”_11 > 0 and V” < 0. So D” - 2 S”_12/|I| - S”_11/|I|^2 + C”_11 > D”. Therefore we must have d B_j(𝐱_-j)/d x_k < G_jk Lower bound. Since G_jk≥ 0 by assumption, we must also have d B_j(𝐱_-j)/d x_k ≥S”_12/|I| + S”_11/|I|^2/D” - 2 S”_12/|I| - S”_11/|I|^2 + C”_11 ≥ -1 · - S”_12 + S”_11· |I|^-1/ |I| · (D” + C”_11) - (2 S”_12 + S”_11· |I|^-1) At this point we impose a purely technical assumption. Assume that (D” + C”_11) ≥ - S”_12 + S”_11· |I|^-1. Intuitively, this requires that the substitutability between own action and average action in the identity group is “not too high” in the status benefits function. Mathematically, it is sufficient to guarantee that d B_j(𝐱_-j)/d x_k ≥ - 1/|I| To summarise: we have d B_j(𝐱_-j)/d x_k∈ (-1/|I|, G_jk) for all j,k. Step 3: Mean Value Inequality. Here, we restate a result from <cit.>.[Adjusting the notation to align with our model.] Let U ⊂ℝ^J be open and B:U →ℝ^J be differentiable on U with continuous derivative. Let 𝐱, 𝐲∈ U be such that [𝐱,𝐲] ⊂ U. Then ||B(𝐲) - B(𝐱)|| ≤max{ ||B'(𝐜)|| : c ∈ [𝐱,𝐲] } || 𝐲 - 𝐱 || where ||B'(𝐜)|| = max{||B'(𝐜) 𝐱|| : || 𝐱 || = 1 }, and || · || denotes the Euclidean norm in ℝ^J. Notice that B'(c) is the J × J Jacobian Matrix. Therefore B'(c) 𝐱 is a J × 1 vector, where the k^th entry is x_k ·∑_j d B_j/d x_k. It is then clear that maximising ||B'(𝐜) 𝐱|| subject to || 𝐱 || = 1 involves finding j ∈max_j ∈ J |∑_j d B_j/d x_k | and placing all weight on that entry. We know from earlier that -1I| < d B_j/d x_k < G_jk. So |∑_j d B_j/d x_k| < max{∑_j G_jk, 1 }. By assumption ∑_j G_jk < 1 for all k. Notice that this is also independent of the argument that B(·) takes. Therefore we have max{ ||B'(𝐜)|| : c ∈ [𝐱,𝐲] } < 1. Step 4: Contraction Mapping. It is well-known that (ℝ^J, ||·||), where ||·|| is the Euclidean norm, constitutes a complete metric space. Therefore the function B: ℝ^J →ℝ^J is a contraction mapping if there exists some k ∈ [0,1) such that ||B(𝐲) - B(𝐱)|| ≤ k || 𝐲 - 𝐱 || for all 𝐱,𝐲∈ℝ^J. Let k = max{ ||B'(𝐜)|| : c ∈ [𝐱,𝐲] }, where ||B'(𝐜)|| = max{||B'(𝐜) 𝐱|| : || 𝐱 || = 1 }. We have shown in part 3 above that this is strictly between 0 and 1. So B(·) is a contraction mapping. Therefore, by the Banach Fixed Point Theorem, there exists a unique fixed point 𝐱^*. This completes the proof. w_j ↑ x_j ↑ w_j ↑ C'_1(x_j,w_j) ↓ (for any given x_j) x_j ↑. Suppose α→ 0 and β > 0. Then w_k ↑ for k ∈ B(θ_j) x_j ↑. With α→ 0, the status benefits drop out. Recall that w_k ↑ x_k ↑ (by the remark above). Then x_k ↑ Y_j ↑. (i) If x_j < Y_j then D(·) and D'(·) are decreasing in x_j and increasing in Y_j. Y_j ↑ D' ↑ x_j ↑. (ii) If x_j ≥ Y_j then D(·) and D'(·) are increasing in x_j and decreasing in Y_j. Y_j ↑ D' ↓ x_j ↑. This is a bit hard to follow without some proper explanation of the best response functions. Suppose α > 0 and β→ 0. Then w_k ↑ for k ∈ B(θ_j) x_j ↓. With β→ 0, the dissonance costs drop out. Recall that w_k ↑ x_k ↑ (by the remark above). Then x_k ↑ Y_j ↑ S'_1 ↓ x_j ↓.

http://arxiv.org/abs/2303.06895v1

20230313065748

An Improved Sample Complexity for Rank-1 Matrix Sensing

cs.IT

Intern Project Yichuan Deng. University of Science and Technology of China. Zhihang Li. Huazhong Agriculture University. Zhao Song. Adobe Research ========================================================================================================================================== ???? [ ??? equal* Aeiau Zzzzequal,to Bauiu C. Yyyyequal,to,goo Cieua Vvvvvgoo Iaesut Saoeued Fiuea Rrrrto Tateu H. Yaseheed,to,goo Aaoeu Iasohgoo Buiui Eueued Aeuia Zzzzed Bieea C. Yyyyto,goo Teoau Xxxxed Eee Pppped toDepartment of Computation, University of Torontoland, Torontoland, Canada gooGoogol ShallowMind, New London, Michigan, USA edSchool of Computation, University of Edenborrow, Edenborrow, United Kingdom Cieua Vvvvvc.vvvvv@googol.com Eee Ppppep@eden.co.uk Machine Learning, ICML 0.3in ] Intern Project Yichuan Deng. University of Science and Technology of China. Zhihang Li. Huazhong Agriculture University. Zhao Song. Adobe Research ========================================================================================================================================== Matrix sensing is a problem in signal processing and machine learning that involves recovering a low-rank matrix from a set of linear measurements. The goal is to reconstruct the original matrix as accurately as possible, given only a set of linear measurements obtained by sensing the matrix <cit.>. In this work, we focus on a particular direction of matrix sensing, which is called rank-1 matrix sensing <cit.>. We present an improvement over the original algorithm in <cit.>. It is based on a novel analysis and sketching technique that enables faster convergence rates and better accuracy in recovering low-rank matrices. The algorithm focuses on developing a theoretical understanding of the matrix sensing problem and establishing its advantages over previous methods. The proposed sketching technique allows for efficiently extracting relevant information from the linear measurements, making the algorithm computationally efficient and scalable. Our novel matrix sensing algorithm improves former result <cit.> on in two senses, * We improve the sample complexity from O(ϵ^-2 dk^2) to O(ϵ^-2 (d+k^2)). * We improve the running time from O(md^2 k^2) to O(m d^2 k). The proposed algorithm has theoretical guarantees and is analyzed to provide insights into the underlying structure of low-rank matrices and the nature of the linear measurements used in the recovery process. It advances the theoretical understanding of matrix sensing and provides a new approach for solving this important problem. empty Matrix sensing is a problem in signal processing and machine learning that involves recovering a low-rank matrix from a set of linear measurements. The goal is to reconstruct the original matrix as accurately as possible, given only a set of linear measurements obtained by sensing the matrix <cit.>. In this work, we focus on a particular direction of matrix sensing, which is called rank-1 matrix sensing <cit.>. We present an improvement over the original algorithm in <cit.>. It is based on a novel analysis and sketching technique that enables faster convergence rates and better accuracy in recovering low-rank matrices. The algorithm focuses on developing a theoretical understanding of the matrix sensing problem and establishing its advantages over previous methods. The proposed sketching technique allows for efficiently extracting relevant information from the linear measurements, making the algorithm computationally efficient and scalable. Our novel matrix sensing algorithm improves former result <cit.> on in two senses, * We improve the sample complexity from O(ϵ^-2 dk^2) to O(ϵ^-2 (d+k^2)). * We improve the running time from O(md^2 k^2) to O(m d^2 k). The proposed algorithm has theoretical guarantees and is analyzed to provide insights into the underlying structure of low-rank matrices and the nature of the linear measurements used in the recovery process. It advances the theoretical understanding of matrix sensing and provides a new approach for solving this important problem. § INTRODUCTION The matrix sensing problem is a fundamental problem in signal processing and machine learning that involves recovering a low-rank matrix from a set of linear measurement. This problem arises in various applications such as image and video processing <cit.> and sensor networks <cit.>. Mathematically, matrix sensing can be formulated as a matrix view of compressive sensing problem <cit.>. The rank-1 matrix sensing problem was formally raised in <cit.>. The matrix sensing problem has attracted significant attention in recent years, and several algorithms have been proposed to solve it efficiently. In this paper, we provide a novel improvement over the origin algorithm in <cit.>, with improvement both on running time and sample complexity. Matrix sensing is a fundamental problem in signal processing and machine learning that involves recovering a low-rank matrix from a set of linear measurements. Specifically, given a matrix W_*∈ℝ^d × d of rank k that is not directly accessible, we aim to recover W_* from a set of linear measurements b ∈^n applied to the ground truth matrix W^* where b_i=[A_i^⊤ W_*],   ∀ i=1, …, m, where A_i are known linear operators. The measurements b_i are obtained by sensing the matrix W_* using a set of linear measurements, and the goal is to reconstruct the original matrix W_* as accurately as possible. This problem arises in various applications such as image and video processing, sensor networks, and recommendation systems. The matrix sensing problem is ill-posed since there may exist multiple low-rank matrices that satisfy the given linear measurements. However, the problem becomes well-posed under some assumptions on the underlying matrix, such as incoherence and restricted isometry property (RIP) <cit.> , which ensure unique and stable recovery of the matrix. A well-used method to solve this problem is to use convex optimization techniques that minimize a certain loss function subject to the linear constraints. Specifically, one can solve the following convex optimization problem: min_W_* (W_*) s.t.  [A_i^⊤ W_*] = b_i, ∀ i=1,…,m. However, this problem is NP-hard <cit.> and intractable in general, and hence, various relaxation methods have been proposed, such as nuclear norm minimization and its variants, which provide computationally efficient solutions with theoretical guarantees. In this work, we focus on the rank-one independent measurements. Under this setting, the linear operators A_i can be decomposed into the form of A_i = x_iy_i^⊤, where x_i ∈^d, y_i ∈^d are all sampled from zero-mean multivariate Gaussian distribution N(0, I_d). Our work on improving the matrix sensing algorithm is based on a novel analysis and sketching technique that enables faster convergence rates and better accuracy in recovering low-rank matrices. We focus on developing a theoretical understanding of the proposed algorithm and establishing its advantages over previous methods. Our analysis provides insights into the underlying structure of the low-rank matrices and the nature of the linear measurements used in the recovery process. The proposed sketching technique allows us to efficiently extract relevant information from the linear measurements, making our algorithm computationally efficient and scalable. Overall, our contribution advances the theoretical understanding of matrix sensing and provides a new approach for solving this important problem. §.§ Our Result To summarize, we improve both the running time of original algorithm <cit.> from O(md^2k^2) to O(md^2k), and the sample complexity from O(ϵ^-2dk^2) to O(ϵ^-2(d + k^2)). Formally, we get the following result, Let ϵ_0 ∈ (0,0.1) denote the final accuracy of the algorithm. Let δ∈ (0,0.1) denote the failure probability of the algorithm. Let σ_1^* denote the largest singular value of ground-truth matrix W_* ∈^d × d. Let κ denote the condition number of ground-truth matrix W_* ∈^d × d. Let ϵ∈ (0, 0.001/(k^1.5κ)) denote the RIP parameter. Let m = Θ (ϵ^-2 (d+k^2)log(d/δ)). Let T = Θ(log(k κσ_1^* /ϵ_0)) . There is a matrix sensing algorithm (Algorithm <ref>) that takes O(m T) samples, runs in T iterations, and each iteration takes O(md^2 k) time, finally outputs a matrix W ∈^d × d such that (1-ϵ_0) W_* ≼ W ≼ (1+ϵ_0) W_* holds with probability at least 1-δ. §.§ Related Work Matrix Sensing The matrix sensing problem has attracted significant attention in recent years, and several algorithms have been proposed to solve it efficiently. One of the earliest approaches is the convex optimization-based algorithm proposed by Candès and Recht in 2009 <cit.>, which minimizes the nuclear norm of the matrix subject to the linear constraints. This approach has been shown to achieve optimal recovery guarantees under certain conditions on the linear operators, such as incoherence and RIP. Since then, various algorithms have been proposed that improve upon the original approach in terms of computational efficiency and theoretical guarantees. For instance, the iterative hard thresholding algorithm (IHT) proposed by Blumensath and Davies in 2009 <cit.>, and its variants, such as the iterative soft thresholding algorithm (IST), provide computationally efficient solutions with improved recovery guarantees. In the work by Recht, Fazel, and Parrilo <cit.>, they gave some measurement operators satisfying the RIP and proved that, with O(k d log d) measurements, a rank-k matrix W_* ∈^d × d can be recovered. Moreover, later works have proposed new approaches that exploit additional structure in the low-rank matrix, such as sparsity or group sparsity, to further improve recovery guarantees and efficiency. For instance, the sparse plus low-rank (S + L) approach proposed by <cit.>, and its variants, such as the robust principal component analysis (RPCA) and the sparse subspace clustering (SSC), provide efficient solutions with improved robustness to outliers and noise. More recently, <cit.> considers the non-square matrix sensing under RIP assumptions, and show that matrix factorization does not introduce any spurious local minima under RIP. <cit.> studies the technique of discrete-time mirror descent utilized to address the unregularized empirical risk in matrix sensing. Compressive Sensing Compressive sensing has been a widely studied topic in signal processing and theoretical computer science field <cit.>. <cit.> gave a fast algorithm (runs in time O(klog n log( n/k)) for generall in puts and O(klog n log(n/k)) for at most k non-zero Fourier coefficients input) for k-sparse approximation to the discrete Fourier transform of an n-dimensional signal. <cit.> provided an algorithm such that it uses O_d(k log N loglog N) samples of signal and runs in time O_d(klog^d+3 N) for k-sparse approximation to the Fourier transform of a length of N signal. Later work <cit.> proposed a new technique for analysing noisy hashing schemes that arise in Sparse FFT, which is called isolation on average, and applying it, it achieves sample-optimal results in klog^O(1)n time for estimating the values of a list of frequencies using few samples and computing Sparse FFT itself. <cit.> gave the first sublinear-time ℓ_2/ℓ_2 compressed sensing which achieves the optimal number of measurements without iterating. After that, <cit.> provided an algorithm which uses O(k log k log n) samples to compute a k-sparse approximation to the d-dimensional Fourier transform of a length n signal. Later by <cit.> provided an efficient Fourier Interpolation algorithm that improves the previous best algorithm <cit.> on sample complexity, time complexity and output sparsity. And in <cit.> they presented a unified framework for the problem of band-limited signal reconstruction and achieves high-dimensional Fourier sparse recovery and high-accuracy Fourier interpolation. Recent work <cit.> designed robust algorithms for super-resolution imaging that are efficient in terms of both running time and sample complexity for any constant dimension under the same noise model as <cit.>, based on new techniques in Sparse Fourier transform. Faster Iterative Algorithm via Sketching Low rank matrix completion is a well-known problem in machine learning with various applications in practical fields such as recommender systems, computer vision, and signal processing. Some notable surveys of this problem are provided in <cit.>. While Candes and Recht <cit.> first proved the sample complexity for low rank matrix completion, other works such as <cit.> and <cit.> have provided improvements and guarantees on convergence for heuristics. In recent years, sketching has been applied to various machine learning problems such as linear regression <cit.>, low-rank approximation <cit.>, weighted low rank approximation, matrix CUR decomposition <cit.>, and tensor regression <cit.>, leading to improved efficiency of optimization algorithms in many problems. For examples, linear programming <cit.>, matrix completion <cit.>, empirical risk minimization <cit.>, training over-parameterized neural network <cit.>, discrepancy algorithm <cit.>, frank-wolfe method <cit.>, and reinforcement learning <cit.>. Roadmap. We organize the following paper as follows. In Section <ref> we provide the technique overview for our paper. In Section <ref> we provide some tools and existing results for our work. In Section <ref> we provide the detailed analysis for our algorithm. In Section <ref> we argue that our measurements are good. In Section <ref> we provide analysis for a shrinking step. In Section <ref> we provide the analysis for our techniques used to solve the optimization problem at each iteration. § TECHNIQUE OVERVIEW In this section, we provide a detailed overview of the techniques used to prove our results. Our approach is based on a combination of matrix sketching and low-rank matrix recovery techniques. Specifically, we use a sketching technique that allows us to efficiently extract relevant information from linear measurements of the low-rank matrix. We then use this information to recover the low-rank matrix using a convex optimization algorithm. With these techniques, we are able to improve previous results in both sample complexity and running time. From the two perspective, we give the overview of our techniques here. §.§ Tighter Analysis Implies Reduction to Sample Complexity Our approach achieves this improvement by using a new sketching technique that compresses the original matrix into a smaller one while preserving its low-rank structure. This compressed version can then be used to efficiently extract relevant information from linear measurements of the original matrix. To analyze the performance of our approach, we use tools from random matrix theory and concentration inequalities. Specifically, we use the Bernstein's inequality for matrices to establish bounds on the error of our recovery algorithm. We first define our measurements and operators, for each i ∈ [m], let x_i, y_i denotes samples from (0, I_d). We define * A_i := x_i y_i^⊤; * b_i := x_i^⊤ W_* y_i; * W_0 := 1/m∑_i = 1^m b_i A_i; * B_x := 1/m∑_i = 1^m (y_i^⊤ v)^2 x_ix_i^⊤; * B_y := 1/m∑_i = 1^m (x_i^⊤ v)^2 y_iy_i^⊤; * G_x := 1/m∑_i=1^m (y_i^⊤ v)(y_i^⊤ v_)x_ix_i^⊤; * G_x := 1/m∑_i=1^m (x_i^⊤ v)(x_i^⊤ v_)y_iy_i^⊤. We need to argue that our measurements are good under our choices of m, here the word “good” means that * W_0 - W_*≤ϵ·W_*; * B_x- I≤ϵ and B_y - I≤ϵ; * G_x≤ϵ and G_y≤ϵ. In our analysis we need to first bound Z_i and [Z_iZ_i^⊤], where Z_i := x_ix_i^⊤ U_*Σ_*Y_*^⊤ y_iy_i^⊤. With an analysis, we are able to show that (Lemma <ref> and Lemma <ref>) [Z_i≤ C^2 k^2 log^2(d/δ)σ^4 ·σ_1^*]  ≥ 1 - δ/(d) [Z_iZ_i^⊤]  ≤ C^2k^2σ^4(σ_1^*)^2. Now, applying these two results and by Bernstein's inequality, we are able to show that our operators are all “good” (Theorem <ref>). §.§ Induction Implies Correctness To get the final error bounded, we use an inductive strategy to analyze. Here we let U_* and V_* be the decomposition of ground truth W_*, i.e., W_* = U_* V_*. We show that, when iteratively applying our alternating minimization method, if U_t and V_t are closed to U_* and V_* respectively, then the output of next iteration t+1 is close to U_* and V_*. Specifically, we show that, if (U_t, U_*) ≤1/4·(V_t, V_*), then it yields (V_t+1, V_*) ≤1/4·(U_t, U_*). Similarly, from the other side, if (V_t+1, V_*) ≤1/4·(U_t, U_*), we have (U_t+1, U_*) ≤1/4·(V_t+1, V_*). This two recurrence relations together give the guarantee that, if the starting error U_0 - U_* and V_0 - V_*, the distance from V_t and U_t to V_* and U_*, respectively. To prove the result, we first define the value of ϵ_d as /10. Then, by the algorithm, we have the following relationship between V_t+1 and V_t+1 R^-1, V_t+1 = V_t+1 R^-1 = (W_*^⊤ U_t - F)R^-1, where the second step follows from the definition of V and defining F as Definition <ref>. Now we show that, F and R^-1 can be bound respectively, F  ≤ 2ϵ k^1.5·σ_1^* ·(U_t, U_*) Lemma <ref> R^-1  ≤ 10/σ_k^*  Lemma <ref> Note that the bound of ^-1 need (U_t, U_*) ≤1/4·(V_t, V_*)) With these bounds, we are able to show the bound for (V_t+1, V_*). We first notice that, (V_t+1, V_*) can be represented as (V_*,)^⊤ V_t+1, where V_*,∈^d × (d-k) is a fixed orthonormal basis of the subspace orthogonal to span(V_*). Then we show that (Claim <ref>) (V_*,)^⊤ V_t+1 = -(V_*, )^⊤ FR^-1. Now, by turning (V_t+1, V_*) to the term of F and R, and using the bound for F and R^-1, we are finally able to reach the bound (V_t+1, V_*) =  FR^-1 ≤  F·R^-1 ≤   2ϵ k^1.5·σ_1^* ·(U_t, U_*) ·R^-1 ≤   2ϵ k^1.5·σ_1^* ·(U_t, U_*) · 10/σ_k^* ≤   0.01 ·(U_t, U_*). By a similar analysis, we can show Eq.(<ref>). Now applying them and with a detailed analysis, we have the claimed proved. Finally, when we prove that the initialization of the parameters are good, we can show that, the final output W_T satisfies W_T - W_*≤ϵ_0. §.§ Speeding up with Sketching Technique Now we consider the running time at each iteration. At each iteration of our algorithm, we need to solve the following optimization problem: min_V ∈^d × k∑_i = 1^m ([A_i^⊤ UV^⊤] - b)^2. When this problem is straightforwardly solved, it costs O(md^2k^2) time, which is very expensive. So from another new direction, we give an analysis such that, this problem can be converted to a minimization problem where the target variable is a vector. To be specific, we show that, above optimization question (<ref>) is equivalent to the following (Lemma <ref>), min_v ∈^dkMv - b_2^2, where the matrix M ∈^m × dk is defined to be the reformed matrix of U^⊤ A_i's, i.e., M_i,* := (U^⊤ A_i),   ∀ i ∈ [m]. When working on this form of optimization problem, inspired by a recent work <cit.>, we apply the fast sketch-to-solve low-rank matrix completion method. With this technique, we are able to reduce the running time to O(md^2k) (Theorem <ref>), which is much more acceptable. § PRELIMINARY In this section, we provide preliminaries to be used in our paper. In Section <ref> we introduce notations we use. In Section <ref> and Section <ref> we provide some randomness facts and algebra facts respectively. In Section <ref> we introduce the important definition of restricted isometry property. In Section <ref> we provide results fro rank-one estimation. In Section <ref> we introduce the rank-one independent Gaussian operator. In Section <ref> we state our notations for angles and distances. In Section <ref> we provide some matrix concentration results. §.§ Notations Let x ∈^n and w ∈_≥ 0^n, we define the norm x_w := (∑_i=1^n w_i x_i^2)^1/2. For n > k, for any matrix A ∈^n × k, we denote the spectral norm of A by A, i.e., A := sup_x∈^k A x _2 / x _2. We denote the Frobenius norm of A by A _F, i.e., A _F : = (∑_i=1^n ∑_j=1^k A_i,j^2 )^1/2. For any square matrix A ∈^n × n, we denote its trace by [A], i.e., [A] := ∑_i=1^n A_i,i. For any A ∈^n × d and B ∈^n × d, we denote ⟨ A , B ⟩ = [A^⊤ B]. Let A ∈^n × d and x ∈^d be any matrix and vector, we have that A x _2^2 = ⟨ A x, A x ⟩ = ⟨ x , A^⊤ A x ⟩ = x^⊤ A^⊤ A x. Let the SVD of A ∈^n × k to be UΣ B^⊤, where U ∈^n × k and V ∈^k × k have orthonormal columns and Σ∈^k × k be diagonal matrix. We say the columns of U are the singular vectors of A. We denote the Moore-Penrose pseudoinverse matrix of A as A^†∈k × n, i.e., A^† := VΣ^-1U^⊤. We call the diagonal entries σ_1, σ_2, …, σ_k of Σ to be the eigenvalues of A. We assume they are sorted from largest to lowest, so σ_i denotes its i-th largest eigenvalue, and we can write it as σ_i(A). For A ∈^n_1 × d_1, B ∈^n_2 × d_2. We define kronecker product ⊗ as (A ⊗ B)_i_1+(i_2-1)n_1, j_1 + (j_2-1)n_2 for all i_1 ∈ [n_1], j_1 ∈ [d_1], i_2 ∈ [n_2] and j_2 ∈ [d_2]. For any non-singular matrix A ∈^n × n, we define A=QR its QR-decomposition, where Q ∈^n × n is an orthogonal matrix and R ∈^n × n is an non-singular lower triangular matrix. For any full-rank matrix A ∈^n × m, we define A=QR its QR-decomposition, where Q ∈^m × n is an orthogonal matrix and R ∈^n × n is an non-singular lower triangular matrix. We use R=QR(A) ∈^n × n to denote the lower triangular matrix obtained by the QR-decomposition of A ∈^m × n. Let A ∈^k× k be a symmetric matrix. The eigenvalue decomposition of A is A = UΛ U^⊤, where Λ is a diagonal matrix. If a matrix A is positive semidefinite (PSD) matrix, we denote it as A ≽ 0, which means x^⊤ A x ≥ 0 for all x. Similarly, we say A ≽ B if x^⊤ Ax ≥ x^⊤ B x for all vector x. For any matrix U ∈^n × k, we say U is an orthonormal basis if U_i=1 for all i ∈ [k] and for any i≠ j, we have ⟨ U_i, U_j ⟩ = 0. Here for each i ∈ [k], we use U_i to denote the i-th column of matrix U. For any U ∈^n × k (suppose n > k)which is an orthonormal basis, we define U_∈^n × (n-k) to be another orthonormial basis that, U U^⊤ + U_ U_^⊤ = I_n and U^⊤ U_ = 0^k × (n-k) where we use 0^k × (n-k) to denote a k × (n-k) all-zero matrix. We say a vector x lies in the span of U, if there exists a vector y such that x = U y. We say a vector z lies in the complement of span of U, if there exists a vector w such that z = U_ w. Then it is obvious that ⟨ x,z ⟩ = x^⊤ z =z^⊤ x =0. For a matrix A, we define σ_min(A) := min_x A x _2 / x _2. Equivalently, σ_min(A) := min_x: x _2=1 A x _2. Similarly, we define σ_max(A) := max_x A x _2 / x _2. Equivalently, σ_max(A) := max_x: x _2=1 A x _2 Let A_1, ⋯, A_n denote a list of square matrices. Let S denote a block diagonal matrix S = [ A_1 ; A_2 ; ⋱ ; A_n ]. Then S = max_i∈ [n] A_i. We use [] to denote probability. We use [] to denote expectation. Let a and b denote two random variables. Let f(a) denote some event that depends on a (for example f(a) can be a=0 or a ≥ 10.). Let g(b) denote some event that depends on b. We say a and b are independent if [f(a)  and  g(b)] = [f(a)] ·[g(b)]. We say a and b are not independent if [ f(a)  and  g(b)] ≠[f(a)] ·[g(b)]. Usually if a and b are independent, then we also have [ab] = [a] ·[b]. We say a random variable x is symmetric if [x = u] = [x=-u]. For any random variable x ∼ N(μ,σ^2). This means [x ] = μ and [x^2] = σ^2. We use O(f) to denote f ·(log f). We use (a,b,c) to denote the time of multiplying an a × b matrix with another b × c matrix. We use ω to denote the exponent of matrix multiplication, i.e., n^ω =(n,n,n). §.§ Randomness Facts We have * Part 1. Expectation has linearity, i.e., [ ∑_i=1^n x_i ] = ∑_i=1^n [x_i]. * Part 2. For any random vectors x and y, if x and y are independent, then for any fixed function f, we have _x,y[f(x) f(y)] = _x[f(x) ] ·_y[ f(y)]. * Part 3. Let A∈^d × d denote a fixed matrix. For any fixed function f : ^d →^d × d, we have _x[f(x) · A ] = _x [f(x)] · A. * Part 4. Given n events A_1, A_2, ⋯ A_n. For each i ∈ [n], if [ A_i ] ≥ 1-δ_i. Then taking a union bound over all the n events, we have [ A_1  and  A_2 ⋯ A_n] ≥ 1- ∑_i=1^n δ_i. §.§ Algebra Facts We state some standard facts and omit their proofs, since they're very standard. We have * For any orthonormal basis U ∈^n × k, we have U x _2 = x _2. * For any orthonornal basis U ∈^n × k, we have U _F ≤√(k). * For any diagonal matrix Σ∈^k × k and any vector x ∈^k, we have Σ x _2 ≥σ_min(Σ) x _2. * For symmetric matrix A, we have σ_min(A) = min_z : z _2=1 z^⊤ A z. * For symmetric matrix A, we have σ_min(A) z_2^2 ≤ z^⊤ A z for all vectors z. * For symmetric matrix A, we have σ_max(A) z_2^2 ≥ z^⊤ A z for all vectors z. * For any matrix A, we have A≤ A _F. * For any square matrix A ∈^k × k and vector x ∈^k, we have x^⊤ A x = ∑_i=1^k ∑_j=1^k x_i A_i,j x_j = ∑_i=1^k x_i A_i,i x_i + ∑_i≠ j x_i A_i,j x_j. * For any square and invertible matrix R, we have R^-1 = σ_min(R)^-1 * For any matrix A and for any unit vector x, we have A ≥ A x _2. * For any matrix A, A A^⊤ = A^⊤ A. §.§ Restricted Isometry Property A linear operator 𝒜: ^d× d→^m satisfies RIP iff, for ∀ W ∈^d × d s.t. (W)≤ k, the following holds: (1-ϵ_k) ·W_F^2≤ A(W)_F^2≤(1+ϵ_k) ·W_F^2 where ϵ_k > 0 is a constant dependent only on k. §.§ Rank-one Estimation The goal of matrix sensing is to design a linear operator 𝒜:^d × d→^m and a recovery algorithm so that a low-rank matrix W_*∈^d × d can be recovered exactly using 𝒜(W_*). Given a ground-truth matrix W_* ∈^d × d. Let (x_1, y_1) , ⋯, (x_m, y_m) ∈^d×^d denote m pair of feature vectors. Let b ∈^m be defined b_i = x_i^⊤ W_* y_i,    ∀ i ∈ [m]. The goal is to use b ∈^m and { (x_i,y_i)}_i ∈ [m]⊂^d ×^d to recover W_* ∈^d × d. We propose two different kinds of rank-one measurement operators based on Gaussian distribution. §.§ Rank-one Independent Gaussian Operator We formally define Gaussian independent operator, here. Let (x_1, y_1) , ⋯, (x_m, y_m) ⊂^d ×^d denote i.i.d. samples from Gaussian distribution. For each i ∈ [m], we define A_i ∈^d × d as follows A_i := x_i y_i^⊤ . We define A_GI∈^d × m d as follows: 𝒜_GI := [ A_1 A_2 ⋯ A_m ] . Here GI denotes Gaussian Independent. §.§ Matrix Angle and Distance We list several basic definitions and tools in literature, e.g., see <cit.>. Let X, Y ∈^n × k denote two matrices. For any matrix X, and for orthonormal matrix Y (Y^⊤ Y = I_k) we define * tanθ(Y,X) := Y_^⊤ X ( Y^⊤ X )^-1 For orthonormal matrices Y and X (Y^⊤ Y = I_k and X^⊤ X = I_k), we define * cosθ (Y,X) := σ_min (Y^⊤ X). * It is obvious that cos (Y,X) = 1/ (Y^⊤ X)^-1 and cos(Y,X) ≤ 1. * sinθ(Y,X) := (I - Y Y^⊤) X. * It is obvious that sinθ(Y,X) = Y_ Y_^⊤ X = Y_^⊤ X and sinθ(Y,X) ≤ 1. * From Lemma <ref>, we know that sin^2θ(Y,X) + cos^2θ(Y,X) = 1. * (Y,X) := sinθ(Y,X) Let X, Y∈^n× k be orthogonal matrices, then tanθ(Y,X) = sinθ(Y,X)/cosθ(Y,X). Let X, Y∈^n× k be orthogonal matrices, then sin^2θ(Y, X) + cos^2θ(Y,X) =1. §.§ Matrix Concentration Given a finite sequence { X_1, ⋯ X_m }⊂^n_1 × n_2 of independent, random matrices all with the dimension of n_1 × n_2. Let Z = ∑_i=1^m X_i. Assume that [X_i] = 0, ∀ i ∈ [m], X_i ≤ M, ∀ i ∈ [m] Let [Z] be the matrix variances statistic of sum [Z] = max{∑_i=1^m [X_iX_i^⊤] , ∑_i=1^m [X_i^⊤ X_i] } Then it holds that [ Z ] ≤ (2 [Z] ·log(n_1+n_2))^1/2 + M log(n_1 + n_3) /3 Further, for all t>0 [ Z ≥ t ] ≤ (n_1 + n_2) ·exp( -t^2/2/[Z] + M t/3 ) § ANALYSIS Here in this section, we provide analysis for our proposed algorithm. In Section <ref>, we provide definitions in our algorithm analysis. In Section <ref> we define the operators to be used. In Section <ref> we provide our main theorem together with its proof. In Section <ref> we introduce our main induction hypothesis. §.§ Definitions We define W_* ∈^d × d as follows W_* = U_* Σ_* V_*^⊤ where U_* ∈^n × k are orthonormal columns, and V_* ∈^n × k are orthonormal columns. Let σ_1^*, σ_2^*, ⋯σ_k^* denote the diagonal entries of diagonal matrix Σ_* ∈^d × d. Let W_* be defined as Definition <ref>. We define κ to the condition number of W_*, i.e., κ : = σ_1/σ_k. It is obvious that κ≥ 1. For each i ∈ [m], let x_i,y_i denote samples from N(0,I_d). For each i ∈ [m], we define A_i = x_i y_i^⊤ and b_i = x_i^⊤ W_* y_i. §.§ Operators For each i ∈ [m], let A_i and b_i be defined as Definition <ref>. We define W_0 := 1/m∑_i=1^m b_i A_i. We say initialization matrix W_0 ∈^d × d is an ϵ-good operator if W_0 - W_* ≤ W_* ·ϵ. For any vectors u,v, we define * B_x:=1/m∑_l=1^m(y_l^⊤ v)^2x_lx_l^⊤ * B_y:=1/m∑_l=1^m(x_l^⊤ u)^2 y_ly_l^⊤ We say B = (B_x,B_y) is ϵ-operator if the following holds: * B_x-I≤ϵ * B_y-I≤ϵ For any vectors u,v ∈^d. We define * G_x:=1/m∑_l=1^m(y_l^⊤ v)(y_l^⊤ v_)x_lx_l^⊤ * G_y:=1/m∑_l=1^m(x_l^⊤ u)(x_l^⊤ u_ ) y_ly_l^⊤ u,u_∈^d,v,v_∈^d are unit vectors, s.t., u^⊤ u_=0 and v^⊤ v_=0. We say G = (G_x,G_y) is ϵ-operator if the following holds * G_x≤ϵ, * G_y≤ϵ. §.§ Main Result We prove our main convergence result as follows: Let W_* ∈^d × d be defined as Definition <ref>. Also, let 𝒜:^d × d→^m be a linear measurement operator parameterized by m matrices, i.e., 𝒜={A_1,A_2,⋯,A_m} where A_l=x_l y_l^⊤. Let 𝒜(W) be as given by b=𝒜(W)= [ [ A_1^⊤ W] [ A_2^⊤ W] ⋯ [A_m^⊤ W] ]^⊤ If the following conditions hold * ϵ= 0.001 / (k^1.5κ ) * T = 100log( κ k / ϵ_0) * Let {(b_i,A_i)}_i∈ [m] be an ϵ-init operator (Definition <ref>). * Let B be an ϵ-operator (Definition <ref>). * Let G be an ϵ-operator(Definition <ref>). Then, after T-iterations of the alternating minimization method (Algorithm <ref>), we obtain W_T=U_T V_T^⊤ s.t., W_T-W_*≤ϵ_0. We first present the update equation for V̂_t+1∈^d × k. Also, note that using the initialization property (first property mentioned in Theorem <ref>), we get, W_0 -W_*≤ϵσ_1^* ≤σ_k^*/100 . Now, using the standard sin theta theorem for singular vector perturbation <cit.>, we get: (U_0,U_*) ≤  1/100 (V_0,V_*) ≤  1/100 After T iteration (via Lemma <ref>), we obtain (U_T,U_*) ≤   (1/4)^T (V_T,V_*) ≤   (1/4)^T which implies that W_T - W_* ≤ϵ_0 §.§ Main Induction Hypothesis We define ϵ_d: = 1/10. We assume that ϵ= 0.001 / (k^1.5κ ). For all t ∈ [T], we have the following results. * Part 1. If (U_t,U_*) ≤1/4(V_t, V_*) ≤ϵ_d, then we have * (V_t+1, V_*) ≤1/4(U_t,U_*) ≤ϵ_d * Part 2. If (V_t+1, V_*) ≤1/4(U_t,U_*) ≤ϵ_d, then we have * (U_t+1,U_*) ≤1/4(V_t+1, V_*) ≤ϵ_d Proof of Part 1. Recall that for each i ∈ [n], we have b_i = x_i^⊤ W_* y_i = ⟨ x_i y_i^⊤ , W_* ⟩ = ⟨ A_i, W_* ⟩ = [A_i^⊤ W_*]. Recall that V̂_t+1=  min_V∈^d× k∑_i=1^m(b_i-x_i^⊤ U_t V^⊤ y_i)^2 =  min_V∈^d× k∑_i=1^m(x_i^⊤ W_* y_i-x_i^⊤ U_t V^⊤ y_i)^2 Hence, by setting gradient of this objective function to zero. Let F ∈^d × k be defined as Definition <ref>. We have V_t+1∈^d × k can be written as follows: V̂_t+1 = W_*^⊤ U_t - F where F ∈^d × k is the error matrix F = [ F_1 F_2 ⋯ F_k ] where F_i ∈^d for each i ∈ [k]. Then, using the definitions of F ∈^d × k and Definition <ref>, we get: [ [ F_1; ⋮; F_k ]] =B^-1(BD-C)S·(V_*) where (V_*) ∈^dk is the vectorization of matrix V_* ∈^d × k. Now, recall that in the t+1-th iteration of Algorithm <ref>, V_t+1∈^d × k is obtained by QR decomposition of V̂_t+1∈^d × k. Using notation mentioned above, V̂_t+1=V_t+1R where R ∈^k × k denotes the lower triangular matrix R_t+1∈^k × k obtained by the QR decomposition of V_t+1∈^d × k. We can rewrite V_t+1∈^d × k as follows V_t+1 =  V̂_t+1 R^-1 =   (W_*^⊤ U_t-F)R^-1 where the first step follows from Eq. (<ref>) , and the last step follows from Eq. (<ref>). Multiplying both the sides by V_*,∈^d × (d-k), where V_*,∈^d × (d-k) is a fixed orthonormal basis of the subspace orthogonal to span(V_*), using Claim <ref> (V_*, )^⊤ V_t+1 = -(V_*, )^⊤ FR^-1 Thus, we get: (V_t+1, V_*) =  (V_*, )^⊤ V_t+1 =   (V_*, )^⊤ F R^-1 =   F R^-1 ≤  F·R^-1 ≤   0.001 σ_k^* (U_t, U_*) · R^-1 ≤   0.001 σ_k^* (U_t, U_*) · 2 (σ_k^*)^-1 ≤   0.01 ·(U_t,U_*) where the first step follows from definition of (see Definition <ref>), the second step follows from Eq. (<ref>), the third step follows from V_*, is an orthonormal basis, and the forth step follows from Fact <ref>, the fifth step follows from Lemma. <ref>, the sixth step follows from Lemma <ref> (In order to run this lemma, we need to the condition of Part 1 statement to be holding), the last step follows from simple algebra. Proof of Part 2. Similarly, we can prove this as Part 1. § MEASUREMENTS ARE GOOD OPERATOR In this section, we provide detailed analysis for our operators. First Section <ref> we introduce some standard results for truncated Gaussian. In Section <ref> and Section <ref> we bound the term Z_i and [Z_iZ_i^⊤] respectively. In Section <ref> we state our main lemma. In Section <ref> we show that out initialization is good. In Section <ref> we show our two operators are good. §.§ Tools for Gaussian We state a standard tool from literature, Let X ∼𝒳_k^2 be a chi-squared distributed random variable with k degrees of freedom. Each one has zero means and σ^2 variance. Then it holds that [X - kσ^2 ≥ (2√(kt) + 2t) σ^2] ≤  exp(-t) [kσ^2 - X ≥ 2√(kt)σ^2] ≤  exp(-t) Further if k ≥Ω(ϵ^-2 t) and t ≥Ω(log(1/δ)), then we have [ | X - k σ^2 | ≤ϵ k σ^2 ] ≤δ. We state a standard fact for the 4-th moment of Gaussian distribution. Let x ∼ N(0,σ^2), then it holds that _x ∼ N(0,σ^2)[x^4] = 3 σ^2. Let x ∼ N(0, σ^2 I_d) denote a random Gaussian vector. Then we have * Part 1 [x x^⊤ x x^⊤] = (d+2) σ^4 * Part 2 [ x x^⊤ x x^⊤ ] = (d+2) σ^4 We define A:=xx^⊤ xx^⊤. Then we have A_i,j = x_i ∑_l=1^d x_l x_l x_j For i=j, we have [A_i,i] =  [ x_i ∑_l=1^d x_l x_l x_i ] =  [x_i(∑_l=1^i-1x_l x_l + x_i x_i + ∑_l=i+1^d x_l x_l) x_i] =  [x_i^4] + ∑_l ∈ [d] \ i[x_l^2 x_i^2] =  [x_i^4] + ∑_l ∈ [d] \ i[x_l^2] [x_i^2] =  [x_i^4] + (d-1) σ^4 =   3 σ^4 + (d-1) σ^4 =   (d + 2) σ^4 where the third step follows from linearity of expectation (Fact <ref>), the forth step follows from x_l and x_i are independent, the fifth step follows _z ∼ N(0,σ^2)[z^4] =3 σ^4. For i≠ j, we have [A_i,j] =  [ x_i ∑_l=1^d x_l x_l x_j ] =  [x_i x_j^3] + [x_i^3 x_j] + ∑_l ∈ [d] \ i,j[x_i x_l^2 x_j] =   0 where the second step follows from linearity of expectation (Fact <ref>). [Rotation invariance property of Gaussian] Let A^⊤∈^d × k with k < d denote an orthonormal basis (i.e., AA^⊤ = I_k). Then for a Gaussian x ∼(0, σ^2 I_d), we have Ax ∼(0, σ^2 I_k). Let y := Ax ∈^k, then y_i = ∑_j = 1^dA_ijx_j,   ∀ i ∈ [k]. By definition of Gaussian distribution y_i ∼(0, σ^2∑_j = 1^dA_ij^2). Recall that A^⊤ is an orthonormal basis. We have A_ij^2 = 1. Thus we have y ∼(0, σ^2 I_k), §.§ Bounding Let x_i denote a random Gaussian vector samples from N(0, σ^2 I_d). Let y_i denote a random Gaussian vector samples from N(0, σ^2 I_d). Let U_*, V_* ∈^d × k. We define Z_i := x_i x_i^⊤ U_* Σ_* V_*^⊤ y_i y_i^⊤,    ∀ i ∈ [m] * Part 1. We have [ Z_i ≤ C^2 k^2 log^2(d/δ) σ^4 ·σ_1^* ] ≥ 1-δ/(d). * Part 2. If k ≥Ω(log(d/δ)) We have [ Z_i ≤ C^2 k^2 σ^4 ·σ_1^* ] ≥ 1-δ/(d). Proof of Part 1. We define a_i := U_*^⊤ x_i ∈^k b_i := V_*^⊤ y_i ∈^k Since U_* and V_* are orthornormal basis, due to rotation invariance property of Gaussian (Fact <ref>) , we know that a_i ∼ N(0,σ^2 I_k) and b_i ∼ N(0, σ^2 I_k). We also know that x_i = (U_*^⊤)^† a_i = U_* a_i y_i = (V_*^⊤)^† b_i = V_* b_i Thus, by replacing x_i,y_i with a_i,b_i, we have Z_i =   x_i x_i^⊤ U_* Σ_* V_*^⊤ y_i y_i^⊤ =   U_* a_i a_i^⊤ U_*^⊤ U_* Σ_* V_*^⊤ V_* b_i b_i^⊤ V_*^⊤ =   U_* a_i a_i^⊤Σ_* b_i b_i^⊤ V_*^⊤ ≤   U_* · a_i a_i^⊤·Σ_* · b_i b_i^⊤· V_*^⊤ ≤  σ_1^* · a_i _2^2 · b_i _2^2 where the second step follows from replacing x,y by a,b, the third step follows from U_*^⊤ U_* = I and V_*^⊤ V_* = I, the forth step follows from Fact <ref>. Due to property of Gaussian, we know that [ |a_i,j| > √(Clog(d/δ))σ ] ≤δ/(d) Taking a union bound over k coordinates, we know that [ a_i _2^2 ≤ C k log(d/δ) σ^2 ] ≥ 1-δ /(d) Similarly, we can prove it for b_i _2^2. Proof of Part 2. Since k ≥Ω(log(d/δ)), then we can use Lemma <ref> to obtain a better bound. §.§ Bounding We can show that [ Z_i Z_i^⊤] ≤ C^2 k^2 σ^4 (σ_1^*)^2. Using Lemma <ref> _a ∼ N(0, σ^2 I_k )[ a_i a_i^⊤ a_i a_i^⊤ ] ≤ C k σ^2. Thus, we have [ a_i a_i^⊤ a_i a_i^⊤] ≼ Ck σ^2 · I_k Then, we have [Z_i Z_i^⊤] =  _x,y[ x_i x_i^⊤ U_* Σ_* V_*^⊤ y_i y_i^⊤ y_i y_i^⊤ V_* Σ_* U_*^⊤ x_i x_i^⊤ ] =  _a,b[ U_* a_i a_i^⊤ U_*^⊤ U_* Σ_* V_*^⊤ V_* b_i b_i^⊤ V_*^⊤ V_* b_i b_i^⊤ V_*^⊤ V_* Σ_* U_*^⊤ U_* a_i a_i^⊤ U_*^⊤ ] =  _a,b[ U_* a_i a_i^⊤Σ_* b_i b_i^⊤ V_*^⊤ V_* b_i b_i^⊤Σ_* a_i a_i^⊤ U_*^⊤ ] =  _a,b[ U_* a_i a_i^⊤Σ_* b_i b_i^⊤ b_i b_i^⊤Σ_* a_i a_i^⊤ U_*^⊤ ] ≤  _a,b[ a_i a_i^⊤Σ_* b_i b_i^⊤ b_i b_i^⊤Σ_* a_i a_i^⊤ ] ≤  _a[ a_i a_i^⊤Σ_* _b[ b_i b_i^⊤ b_i b_i^⊤ ] Σ_* a_i a_i^⊤ ] ≤   C^2 k^2 σ^4 (σ_1^*)^2 where the first step follows from the definition of Z_i, the second step follows from replacing x_i,y_i with a_i,b_i, the third step follows from U_*,V_* are orthonormal columns, the fourth step follows from V_* are orthonormal columns, the fifth step follows from U_* ≤ 1 , the sixth step follows from using Lemma <ref> twice. §.§ Main Results We prove our main result for measurements. Let {A_i,b_i}_i∈ [m] denote measurements be defined as Definition <ref>. Assuming the following conditions are holding * k = Ω(log(d/δ)) * m = Ω(ϵ^-2 (d+k^2) log(d/δ)) Then, * The property in Definition <ref>, initialization is a ϵ-operator * The property in Definition <ref>, B are ϵ-operator. * The property in Definition <ref>, G are ϵ-operator. holds with probability at least 1-δ/(d). Using Lemma <ref> and Lemma <ref>, we complete the proof. §.§ Initialization Is a Good Operator We define matrix S ∈^d × d as follows S: = 1/m∑_i=1^m b_i A_i. If the following two condition holds * Condition 1. k = Ω(log(d/δ)), * Condition 2. m = Ω( ϵ^-2 k^2 log(d/δ) ). Then we have [ S - W_* ≤ϵ· W_* ] ≥ 1-δ. (Initialization in Definition <ref>) Now, we have: S =  1/m∑_i=1^m b_i A_i =  1/m∑_i=1^m b_i x_i y_i^⊤ =  1/m∑_i=1^m x_i b_i y_i^⊤ =  1/m∑_i=1^m x_i x_i^⊤ W_* y_i y_i^⊤ =  1/m∑_i=1^m x_i x_i^⊤ U_* Σ_* V_*^⊤ y_i y_i^⊤, where the first step follows from Definition <ref>, the second step follows from A_i = x_i y_i^⊤, the third step follows from b_i is a scalar, the forth step follows from b_i = x_i^⊤ W_* y_i, the fifth step follows from W_* = U_* Σ_* V_*^⊤. For each i ∈ [m], we define matrix Z_i ∈^d × d as follows: Z_i := x_i x_i^⊤ U_* Σ_* V_*^⊤ y_i y_i^⊤, then we can rewrite S ∈^d × d in the following sense, S = 1/m∑_i=1^m Z_i Note that, we can compute [Z_i] ∈^d × d _x_i,y_i[Z_i] =  _x_i, y_i[ x_ix_i^⊤_d × d U_* Σ_* V_*^⊤_ d × d y_i y_i^⊤_d × d ] =  _x_i[ x_ix_i^⊤_d × d U_* Σ_* V_*^⊤_ d × d ] ·_y_i [ y_i y_i^⊤_d × d ] =  _x_i [x_i x_i^⊤ ] · U_* Σ_* V_*^⊤·_y_i[ y_i y_i^⊤] =   U_* Σ_* V_*^⊤ where the first step follows definition of Z_i, the second step follows from x_i and y_i are independent and Fact <ref>, the third step follows from Fact <ref> the forth step follows from [x_ix_i^⊤] = I_d and [y_i y_i^⊤] = I_d. As S ∈^d × d is a sum of m random matrices, the goal is to apply Theorem <ref> to show that S is close to [S] =   W_* =   U_* Σ_* V_*^⊤ for large enough m. Using Lemma <ref> (Part 2) with choosing Gaussian variance σ^2=1, we have [ Z_i ≤ C^2 k^2 σ_1^*, ∀ i ∈ [m] ] ≥ 1-δ/(d) Using Lemma <ref> with choosing Gaussian variance σ^2= 1, we can bound [Z_i Z_i^⊤] as follows [Z_i Z_i^⊤] ≤   C^2 k^2 (σ_1^*)^2 Let Z = ∑_i=1^m (Z_i - W_*). Applying Theorem <ref> we get [ Z ≥ t ] ≤ 2d ·exp( -t^2/2/[Z] + M t/3 ) where Z =   m S - m W_* [Z] =   m · C^2 k^2 (σ_1^*)^2,  by Eq. (<ref>) M=   C^2 k^2 σ_1^*  by Eq. (<ref>) Replacing t= ϵσ_1^* m and Z = mS - mW_* inside [] in Eq. (<ref>), we have [ S - W^* ≥  ϵσ_1^* ] ≤ 2d ·exp( -t^2 /2/[Z] + M t /3) Our goal is to choose m sufficiently large such that the above quantity is upper bounded by 2d ·exp ( - Ω( log(d/δ) )). First, we need t^2/[Z] =  ϵ^2 m^2 (σ_1^*)^2 / m · C^2 k^2 (σ_1^*)^2 =  ϵ^2 m / C^2 k^2 ≥  log(d/δ) where the first step follows from choice of t and bound for [Z]. This requires m ≥ C^2 ϵ^-2 k^2 log(d/δ) Second, we need t^2 / M t =  ϵ m σ_1^* / M =  ϵ m σ_1^* /C^2 k^2 σ_1^* =  ϵ m /C^2 k^2 ≥  log(d/δ) where the first step follows from choice of t and the second step follows from bound on M. This requires m ≥ C^2 ϵ^-2 k^2 log(d/δ) Finally, we should choose m ≥ 10C^2 ϵ^-2 k^2 log(d/δ) , Which implies that [ S - W_* ≤ϵ·σ_1^* ] ≥ 1- δ/(d). Taking the union bound with all Z_i are upper bounded, then we complete the proof. §.§ Operator and is good If the following two conditions hold * Condition 1. d = Ω(log(d/δ)) * Condition 2. m = Ω(ϵ^-2 d log(d/δ)) Then operator B (see Definition <ref>) is ϵ good, i.e., [ B_x - I_d≤ϵ ] ≥   1-δ/(d) [ B_y - I_d ≤ϵ ] ≥   1-δ/(d) Similar results hold for operator G (see Definition <ref>). Recall that B_x:=1/m∑_l=1^m(y_l^⊤ v)^2x_lx_l^⊤. Recall that B_y:=1/m∑_l=1^m(x_l^⊤ u)^2 y_ly_l^⊤. Now, as x_i,y_i are rotationally invariant random variables , wlog, we can assume u=e_1. We use x_i,1∈ to denote the first entry of x_i ∈^d. Thus, (x_i^⊤ u u^⊤ x_i)=x_i,1^2 Then [ (x_i^⊤ u u^⊤ x_i)^2 ] = [x_i,1^4 ] = 3 We define Z_i = (x_i^⊤ u)^2 y_i y_i^⊤ then [Z_i] = I_d Using similar idea in Lemma <ref>, we have [ Z_i ≤ C d , ∀ i ∈ [m] ] ≥ 1- δ/(d) We can bound [ Z_i Z_i^⊤ ] =  _x,y[ (x_i^⊤ u)^2 y_i y_i^⊤ y_i y_i^⊤ (x_i^⊤ u)^2 ] =  _x[ (x_i^⊤ u)^2 _y[y_i y_i^⊤ y_i y_i^⊤ ] (x_i^⊤ u)^2 ] =   (d+2) · | _x[ (x_i^⊤ u)^2 (x_i^⊤ u)^2 ] | =   (d+2) · 3 ≤   C d where the fourth step follows from C ≥ 1 is a sufficiently large constant. Let Z = ∑_i=1^m (Z_i - I_d). Applying Theorem <ref> we get [ Z≥ t] ≤ 2d ·exp(-t^2/2/[Z] + Mt/3), where Z =   m · B - m · I [Z] =   C m d M =   C d Using t = m ϵ and Z = ∑_i=1^m (Z_i - I_d), and B = 1/m∑_i=1^m Z_i, we have [ Z ≥ t] =  [ ∑_i=1^m (Z_i - I_d) ≥ m ϵ ] =  [ 1/m∑_i=1^m Z_i - I_d ≥ϵ ] =  [ B - I_d ≥ϵ ] By choosing t = m ϵ and m = Ω(ϵ^-2 d log(d/δ)) we have [ B - I_d ≥ϵ ] ≤δ/(d). where B can be either B_x or B_y. Similarly, we can prove [G_x≤ϵ] ≥ 1 - δ, [G_y≤ϵ] ≥ 1 - δ. § ONE SHRINKING STEP In this section, we provide a shirking step for our result. In Section <ref> we define the matrices B, C, D ,S to be used in analysis. In Section <ref> we upper bound the norm of BD- C. In Section <ref> we show the update term V_t + 1 can be written in a different way. In Section <ref> and Section <ref> we upper bounded F and R^-1 respectively. §.§ Definitions of For each p ∈ [k], let u_*,p∈^n denotes the p-th column of matrix U_* ∈^n × k. For each p ∈ [k], let u_t,p denote the p-th column of matrix U_t ∈^n × k. We define block matrices B, C, D, S ∈^kd × kd as follows: For each (p,q) ∈ [k] × [k] * Let B_p,q∈^d × d denote the (p,q)-th block of B B_p,q= ∑_i=1^m y_i y_i^⊤_d × d  matrix· (x_i^⊤ u_t,p)_scalar· (x_i^⊤ u_t,q) _scalar * Let C_p,q∈^d × d denote the (p,q)-th block of C, C_p,q= ∑_i=1^m y_i y_i^⊤_ d × d  matrix· (x_i^⊤ u_t,p ) _scalar· (x_i^⊤ u_*q) _scalar * Let D_p,q∈^d × d denote the (p,q)-th block of D, D_p,q= u_t,p^⊤ u_*q I * Let S_p,q∈^d × d denote the (p,q)-th block of S, S_p,q= σ_p^* I , if  p=q;   0, if  p q. Here σ_1^*, ⋯σ_k^* are singular values of W_* ∈^d × d. * We define F ∈^d × k as follows (F) _d × 1 := B^-1_d × d (BD-C) _d × d S _d × d·(V_*) _d × 1. §.§ Upper Bound on Let B, C and D be defined as Definition <ref>. Then we have BD-C≤ϵ·(U,U_*) · k Let z_1, ⋯, z_k ∈^d denote k vectors. Let z = [ z_1; ⋮; z_k ]. We define f(z):=z^⊤ (BD- C)z We define f(z,p,q) = z_p^⊤ (BD-C)_p,q z_q. Then we can rewrite z^⊤ (BD - C) z =  ∑_p=1^k ∑_q=1^k z_p^⊤ (BD-C)_p,q z_q =  ∑_p=1^k ∑_q=1^k z_p^⊤ ( B_p,: D_:,q - C_p,q ) z_q =  ∑_p=1^k ∑_q=1^k z_p^⊤ ( ∑_l=1^k B_p,l D_l,q - C_p,q ) z_q By definition, we know B_p,l =  ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · ( u_t,l^⊤ x_i ) D_l,q =   (u_*,q^⊤ u_t,l ) I_d C_p,q =  ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · ( u_*,q^⊤ x_i ) We can rewrite C_p,q as follows C_p,q = ∑_i=1^m y_i y_i^⊤· (x_i^⊤ u_t,p) · ( u_*,q^⊤ I_d x_i ) Let us compute B_p,l D_l,q =  ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · ( u_t,l^⊤ x_i ) · ( u_*,q^⊤ u_t,l ) =  ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · ( u_*,q^⊤ u_t,l ) · ( u_t,l^⊤ x_i ) where the second step follows from a · b = b · a for any two scalars. Taking the summation over all l ∈ [k], we have ∑_l=1^k B_p,l D_l,q =  ∑_l=1^k ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · ( u_*,q^⊤ u_t,l ) · ( u_t,l^⊤ x_i ) =  ∑_i=1^m y_i y_i^⊤ (x_i^⊤ u_t,p) · u_*,q^⊤∑_l=1^k (u_t,l· u_t,l^⊤ ) x_i =  ∑_i=1^m y_i y_i^⊤_matrix· (x_i^⊤ u_t,p) _scalar· u_*,q^⊤ U_t U_t^⊤ x_i _scalar where first step follows from definition of B and D. Then, we have ∑_l=1^k B_p,l D_l,q - C_p,q =   (∑_i=1^m y_i y_i^⊤_matrix· (x_i^⊤ u_t,p) _scalar· u_*,q^⊤ U_t U_t^⊤ x_i _scalar) - C_p,q =   (∑_i=1^m y_i y_i^⊤_matrix· (x_i^⊤ u_t,p) _scalar· u_*,q^⊤ U_t U_t^⊤ x_i _scalar) - (∑_i=1^m y_i y_i^⊤· (x_i^⊤ u_t,p) · ( u_*,q^⊤ I_d x_i )) =  ∑_i=1^m y_i y_i^⊤_matrix· (x_i^⊤ u_t,p) _scalar· u_*,q^⊤ ( U_t U_t^⊤ - I_d) x_i _scalar where the first step follows from Eq. (<ref>), the second step follows from Eq. (<ref>), the last step follows from merging the terms to obtain (U_t U_t^⊤ - I_d). Thus, f(z,p,q) =  z_p^⊤ ( ∑_l=1^k B_p,l D_l,q - C_p,q ) z_q =  ∑_i=1^m ( z_p^⊤ y_i ) _scalar ( y_i^⊤ z_q ) _scalar· (x_i^⊤ u_t,p) _scalar· u_*,q^⊤ ( U_t U_t^⊤ - I_d) x_i _scalar For easy of analysis, we define v_t:= u_*,q^⊤ ( U_t U_t^⊤ - I_d). This means v_t lies in the complement of span of U_t. Then v_t _2 =   u_*,q^⊤ ( U_t U_t^⊤ - I_d) _2 =   e_q^⊤ U_*^⊤ (U_t U_t^⊤ - I_d) ≤   U_*^⊤ (U_t U_t^⊤ - I_d) =  (U_*,U_t). where the second step follows from u_*,q^⊤ = e_q^⊤ U_*^⊤ (e_q ∈^k is the vector q-th location is 1 and all other locations are 0s), third step follows from Fact <ref>. We want to apply Definition <ref>, but the issue is z_p, z_q and v_t are not unit vectors. So normalize them. Let z_p = z_p / z_p _2 , z_q = z_q / z_q _2 and v_t = v_t/ v_t _2. In order to apply for Definition <ref>, we also need v_t^⊤ u_t,p=0. This is obvious true, since v_t lies in the complement of span of U_t and u_t,p in the span of U_t. We define G := ∑_i=1^m (x_i^⊤ u_t,p) _scalar· (x_i^⊤v_t) _scalar· y_i y_i^⊤_matrix By Definition <ref>, we know that G ≤ϵ. By definition of spectral norm, we have for any unit vector z_p and z_q, we know that |z_p^⊤ G z_q | ≤ G ≤ϵ. where the first step follows from definition of spectral norm (Fact <ref>), and the last step follows from Definition <ref>. Note that f(p,q,z) =  ∑_i=1^m (x_i^⊤ u_t,p) · (x_i^⊤v_t) _scalar· (z_p^⊤ y_i) · (y_i^⊤z_q) _scalar· z_p _2 · z_q _2 · v_t _2 _scalar =  z_p^⊤_ 1 × d·( ∑_i=1^m (x_i^⊤ u_t,p) · (x_i^⊤v_t) _scalar· y_i y_i^⊤_ d × d ) ·z_q _d × 1· z_p _2 · z_q _2 · v_t _2 _scalar =  z_p^⊤_1 × d· G _d × d·z_q _d × 1· z_p _2 · z_q _2 · v_t _2 _scalar where the second step follows from rewrite the second scalar (z_p^⊤ y_i) (y_i^⊤z_q) = z_p^⊤ (y_i y_i^⊤) z_q, the last step follows from definition of G. Then, |f(z,p,q)| =   | ∑_i=1^m z_p^⊤ G z_q | · z_p _2 z_q _2 v_t _2 ≤  ϵ z_p _2 z_q _2 · v_t _2 ≤  ϵ z_p _2 z_q _2 ·(U_t,U_*) where the last step follows from Eq. (<ref>). Finally, we have BD-C =  max_z,z_2=1|z^⊤(BD-C)z| =  max_z,z_2=1|∑_ p ∈ [ k ],q ∈ [k] f(z,p,q)| ≤  max_z,z_2=1∑_ p ∈ [ k ],q ∈ [k] | f(z,p,q)| ≤  ϵ·(U_t,U_*) max_z,z_2=1∑_p∈ [k] , q ∈ [k] z_p_2z_q_2 ≤  ϵ·(U,U_*) · k where the first step follows from Fact <ref>, the last step step follows from ∑_p=1^k z_p _2 ≤√(k) (∑_p=1^k z_p _2^2)^1/2 = √(k). §.§ Rewrite If V_t+1 = (W_*^⊤ U_t-F)R^-1 then, (V_*, )^⊤ V_t+1 = -(V_*, )^⊤ FR^-1 Multiplying both sides by V_*,∈^d × (d-k): V_t+1=   (W_*^⊤ U_t-F)R^-1 (V_*, )^⊤ V_t+1=  (V_*, )^⊤(W_*^⊤ U_t-F)R^-1 (V_*, )^⊤ V_t+1=  (V_*, )^⊤ W_*^⊤ R^-1-(V_*, )^⊤ FR^-1 We just need to show (V_*, )^⊤ W_*^⊤ R^-1=0. By definition of V_*,, we know: V_*,^⊤ V_*= 0_k × (n-k) Thus, we have: (V_*, )^⊤ W_*^⊤ =   V_*,^⊤ V_* Σ_* U_*^⊤ =   0 §.§ Upper bound on Let 𝒜 be a rank-one measurement operator where A_i = x_i u_i^⊤. Let κ be defined as Definition <ref>. Then, we have F ≤ 2 ϵ k^1.5·σ_1^* ·(U_t,U_*) Further, if ϵ≤ 0.001 / ( k^1.5κ ) F ≤ 0.01 ·σ_k^* ·(U_t,U_*). Recall that (F) = B^-1(BD-C)S ·(V_*). Here, we can upper bound F as follows F≤  F_F =  (F) _2 ≤  B^-1·BD-C·S·(V_*)_2 =  B^-1·(BD-C)· S ·√(k) ≤  B^-1·(BD-C)·σ_1^* ·√(k) where the first step follows from ·≤·_F (Fact <ref>), the second step follows vectorization of F is a vector, the third step follows from A x _2 ≤A · x _2, the forth step follows from (V_*) _2 = V_* _F ≤√(k) (Fact <ref>) and the last step follows from S ≤σ_1^* (see Definition <ref>). Now, we first bound B^-1=1/(σ_min(B)). Also, let Z=[ z_1 z_2 ⋯ z_k ] and let z=(Z). Note that B_p,q denotes the (p,q)-th block of B. We define B := { x ∈^kd |  x _2 = 1 }. Then σ_min(B) =  min_z ∈ Bz^⊤ B z =  min_z ∈ B∑_ p ∈ [k], q ∈ [k] z_p^⊤ B_pqz_q =  min_z ∈ B∑_p=1^k z_p^⊤ B_p,pz_p+∑_p≠ qz_p^⊤ B_p,qz_q. where the first step follows from Fact <ref>, the second step follows from simple algebra, the last step follows from (Fact <ref>). We can lower bound z_p^⊤ B_p,pz_p as follows z_p^⊤ B_p,p z_p ≥  σ_min(B_p,p) · z_p _2^2 ≥   (1-ϵ) · z_p _2^2 where the first step follows from Fact <ref> , the last step follows from Definition <ref> . We can upper bound | z^⊤ B_p,q z_q | as follows, |z_p^⊤ B_p,q z_q| ≤   z_p _2 · B_p,q· z_q _2 ≤  ϵ· z_p _2 · z_q _2 where the first step follows from Fact <ref>, the last step follows from Definition <ref> . We have σ_min(B) =  min_z,z_2=1∑_p=1^k z_p^⊤ B_p,pz_p+∑_p≠ qz_p^⊤ B_p,qz_q ≥  min_z,z_2=1 (1-ϵ)∑_p=1^k z_p _2^2 +∑_p≠ qz_p^⊤ B_p,qz_q ≥  min_z,z_2=1(1-ϵ)∑_p=1^kz_p_2^2-ϵ∑_p ≠ qz_p_2z_q_2 =  min_z,z_2=1 (1-ϵ) -ϵ∑_p ≠ qz_p_2z_q_2 =  min_z,z_2=1 (1-ϵ) - k ϵ ≥   1- 2 kϵ ≥   1/2 where the first step follows from Eq. (<ref>), the second step follows from Eq. (<ref>), the third step follows from Eq. (<ref>), the forth step follows from ∑_p=1^k z_p _2^2 = 1(which derived from the z_2=1 constraint and the definition of z_2), the fifth step follows from ∑_p ≠ q z_p _2 z_q _2 ≤ k, and the last step follows from ϵ≤ 0.1/k. We can show that B^-1 = σ_min(B) ≤ 2. where the first step follows from Fact <ref>, the second step follows from Eq. (<ref>). Now, consider BD-C, using Claim <ref>, we have BD-C≤ k·ϵ·(U_t,U_*) Now, we have F ≤   B^-1· (BD - C) ·σ_1^* ·√(k) ≤   2 · (BD - C) ·σ_1^* ·√(k) ≤   2 · k ·ϵ·(U_t,U_*) ·σ_1^* ·√(k) where the first step follows from Eq .(<ref>), the second step follows from Eq. (<ref>), and the third step follows from Eq. (<ref>). §.§ Upper bound on Let 𝒜 be a rank-one measurement operator matrix where A_i=x_i y_i^⊤. Also, let 𝒜 satisfy three properties mentioned in Theorem <ref>. If the following condition holds * (U_t, U_*) ≤1/4≤ϵ_d = 1/10 (The condition of Part 1 of Lemma <ref>) Then, R^-1≤   10 /σ_k^* For simplicity, in the following proof, we use V to denote V_t+1. We use U to denote U_t. Using Fact <ref> R^-1 = σ_min(R)^-1 We can lower bound σ_min(R) as follows: σ_min(R) =  min_z,z_2=1Rz_2 =  min_z,z_2=1VRz_2 =  min_z,z_2=1V_*Σ_*U_*^⊤ Uz-Fz_2 ≥  min_z,z_2=1V_*Σ_*U_*^⊤ Uz_2-Fz_2 ≥  min_z,z_2=1V_*Σ_*U_*^⊤ Uz_2-F where the first step follows from definition of σ_min, the second step follows from Fact <ref>, the third step follows from V = (W_*^⊤ U-F)R^-1 = (V_* Σ_* U_*^⊤ U - F) R^-1 (due to Eq. (<ref>) and Definition <ref>) , the forth step follows from triangle inequality, the fifth step follows from A x _2 ≤ A for all x _2=1. Next, we can show that min_z,z_2=1V_*Σ_*U_*^⊤ Uz_2 =  min_z,z_2=1Σ_*U_*^⊤ Uz_2 ≥  min_z,z_2=1σ_k^* · U_*^⊤ Uz_2 =  σ_k^* ·σ_min(U^⊤ U_*) where the first step follows from Fact <ref>, the second step follows from Fact <ref>, the third step follows from definition of σ_min, Next, we have σ_min(U^⊤ U_*) =  cosθ(U_*, U) =  √(1-sin^2 θ(U_*,U)) ≥  √(1- (U_*,U)^2) where the first step follows definition of cos, the second step follows from sin^2 θ + cos^2 θ =1 (Lemma <ref>), the third step follows from sin≤ (see Definition <ref>). Putting it all together, we have σ_min(R) ≥  σ_k^* √(1-(U_*,U)^2) - F ≥  σ_k^* √(1-(U_*,U)^2) - 0.001 σ_k^* (U_*,U) =  σ_k^* ( √(1-(U_*,U)^2) - 0.001 (U_*,U) ) ≥   0.2 σ_k^* where the second step follows from Lemma <ref>, the last step follows from (U_*,U) < 1/10. § MATRIX SENSING REGRESSION Our algorithm has O(log(1/ϵ_0)) iterations, in previous section we have proved why is that number of iterations sufficient. In order to show the final running time, we still need to provide a bound for the time we spend in each iteration. In this section, we prove a bound for cost per iteration. In Section <ref> we provide a basic claim that, our sensing problem is equivalent to some regression problem. In Section <ref> we show the different running time of the two implementation of each iteration. In Section <ref> we provide the time analysis for each of the iteration of our solver. In Section <ref> shows the complexity for the straightforward solver. Finally in Section <ref> we show the bound for the condition number. §.§ Definition and Equivalence In matrix sensing, we need to solve the following problem per iteration: Let A_1,…,A_m ∈^d× d, U∈^d× k and b∈^m be given. The goal is to solve the following minimization problem min_V∈^d× k∑_i=1^m ([A_i^⊤ U V^⊤]-b_i)^2, We define another regression problem Let A_1,…,A_m ∈^d× d, U∈^d× k and b∈^m be given. We define matrix M∈^m× dk as follows M_i,* :=  (U^⊤ A_i),    ∀ i ∈ [m]. The goal is to solve the following minimization problem. min_v∈^d kMv-b _2^2, We can prove the following equivalence result Let A_1,…,A_m ∈^d× d, U∈^d× k and b∈^m be given. If the following conditions hold * M_i,* := (U^⊤ A_i),    ∀ i ∈ [m]. * The solution matrix V ∈^d × k can be reshaped through vector v ∈^dk, i.e., v = (V^⊤). Then, the problem (defined in Definition <ref>) is equivalent to problem (defined in Definition <ref>) . Let X, Y∈^d× d, we want to show that [X^⊤ Y] =  (X)^⊤(Y). Note that the RHS is essentially ∑_i ∈ [d]∑_j ∈ [d] X_i,jY_i,j, for the LHS, note that (X^⊤ Y)_j,j =  ∑_i∈ [d] X_i,j Y_i,j, the trace is then sum over j. Thus, we have Eq. (<ref>). This means that for each i∈ [d], [A_i^⊤ UV^⊤]=(U^⊤ A_i)^⊤(V^⊤). Set M∈^m × dk be the matrix where each row is (U^⊤ A_i), we see Definition <ref> is equivalent to solve the regression problem as in the statement. This completes the proof. §.§ From Sensing Matrix to Regression Matrix Let A_1,…,A_m ∈^d× d, U∈^d× k . We define matrix M∈^m× dk as follows M_i,* :=  (U^⊤ A_i),    ∀ i ∈ [m]. The naive implementation of computing M ∈^m × dk takes m ·(k,d,d) time. Without using fast matrix multiplication, it is O(md^2k) time. For each i ∈ [m], computing matrix U^⊤∈^k × d times A_i ∈^d × d takes (k,d,d) time. Thus, we complete the proof. The batch implementation takes (k,dm,d) time. Without using fast matrix multiplication, it takes O(md^2 k) time. We can stack all the A_i together, then we matrix multiplication. For example, we construct matrix A ∈^d × dm. Then computing U^⊤ A takes (k,d,dm) time. The above two approach only has difference when we use fast matrix multiplication. §.§ Our Fast Regression Solver In this section, we provide the results of our fast regression solver. Our approach is basically as in <cit.>. For detailed analysis, we refer the readers to the Section 5 in <cit.>. Assume m = Ω(dk). There is an algorithm that runs in time O( m d^2 k + d^3 k^3 ) and outputs a v' such that M v' - b _2 ≤ (1+ϵ) min_v ∈^dk M v - b_2 From Claim <ref>, writing down M ∈^m × dk takes O(md^2 k) time. Using Fast regression resolver as <cit.>, the fast regression solver takes O( ( m· dk + (dk)^3 ) ·log(κ(M)/ϵ) ·log^2(n/δ) ) In each iteration, our requires takes O( m d^2 k) time. Finally, in order to run Lemma <ref>, we need to argue that κ(M) ≤(k,d,κ(W_*)). This is true because κ(U) ≤ O(κ(W_*)) and condition number of random Gaussian matrices is bounded by (k,d). Then applying Lemma <ref>, we can bound κ(M) in each iteration. Eventually, we just run standard error analysis in <cit.>. Thus, we should get the desired speedup. The reason we can drop the (dk)^3 is m ≥ dk^2. §.§ Straightforward Solver Note that from sample complexity analysis, we know that m = Ω(dk). Assume m = Ω(dk). The straightforward implementation of the regression problem (Defintion <ref>) takes O(md^2 k^2) time. The algorithm has two steps. From Claim <ref>, writing down M ∈^m × dk takes O(md^2 k) time. The first step is writing down the matrix M ∈^m × dk. The second step is solving regression, it needs to compute M^† b (where M^†∈^d k × m ) M^† b = ( M^⊤ M )^-1 M b this will take time (dk,m,dk) + (dk,dk,dk) =   m d^2k^2 + (dk)^3 =   m d^2 k^2 the second step follows from m =Ω(dk) . Thus, the total time is m d^2 k + md^2 k^2 = O(m d^2k^2) §.§ Condition Number We define B ∈^m × k as follows B := X U and X ∈^m × d and U ∈^d × k. Then, we can rewrite M ∈^m × dk M_m × dk = B_m × k⊗Y_m × d Then, we know that κ(M) = κ(B) ·κ(Y) ≤κ(U) κ(X) κ(Y). Recall U ∈^d × k. Then we define b_i = U^⊤ x_i for each i ∈ [m]. Then we have M_i,* = ( U^⊤ x_i y_i^⊤ ) = (b_i y_i^⊤ ). Thus, it implies M = B ⊗ Y alpha alpha

http://arxiv.org/abs/2303.07081v1

20230313130640

Evolution of many-body systems under ancilla quantum measurements

quant-ph

[Corresponding author: ]elmer.doggen@kit.edu Department of Condensed Matter Physics, Weizmann Institute of Science, 7610001 Rehovot, Israel Ioffe Institute, 194021 St. Petersburg, Russia We explore the concept of implementing quantum measurements by coupling a many-body lattice system to an ancillary degree of freedom on which projective measurements are performed. We analyze the effect of repeated (“stroboscopic”) measurements on the dynamical correlations of interacting hard-core bosons in a one-dimensional chain. An important distinctive ingredient of the protocol is the fact that the detector ancillas are not re-initialized after each measurement step. The detector thus maintains memory of the accumulated influence by the measured correlated system. Initially, we consider a model in which the ancilla is coupled to a single lattice site. This setup allows obtaining information about the system through Rabi oscillations in the ancillary degrees of freedom, modulated by the ancilla-system interaction. The statistics of quantum trajectories exhibits a “quantum-Zeno-valve effect” that occurs when the measurement becomes strong, with sharp branching between low and high entanglement. We proceed by extending numerical simulations to the case of two ancillas and, then, to measurements on all sites. With this realistic measurement apparatus, we find evidence of a disentangling-entangling measurement-induced transition as was previously observed in more abstract models. The dynamics features a broad distribution of the entanglement entropy. Evolution of many-body systems under ancilla quantum measurements Dmitry G. Polyakov March 30, 2023 ================================================================= § INTRODUCTION The concept of measurements in quantum systems has been a topic of great interest ever since the dawn of quantum mechanics, but the microscopic description of the dynamics of systems under the influence of measurements has remained a challenging problem <cit.>. In recent years, this problem has gained much attention particularly because of its relevance to the stability of quantum computing architectures <cit.> and to the goal of achieving “quantum supremacy” <cit.>. Furthermore, the physics of quantum measurements is closely connected with the study of open quantum systems <cit.>. It is also fundamentally relevant to understanding the link between the microscopic physics and the thermodynamics on macroscopic scales <cit.>. Recently, the notion of a measurement-induced entanglement transition—a dynamical phase transition driven by the strength or frequency of measurements—has been proposed <cit.>. Initially discussed in the context of quantum circuits <cit.>, the idea of measurement-induced transitions has been extended to non-interacting lattice fermions <cit.>, Dirac fermions <cit.>, the quantum Ising model <cit.>, the Sachdev-Ye-Kitaev model <cit.>, as well as a number of other integrable <cit.> and nonintegrable many-body systems <cit.>. Evidence of a measurement-induced transition has also been reported in trapped-ion <cit.> and superconducting <cit.> quantum-processor architectures. An intriguing question is to what extent the essential properties of these measurement-induced transitions are universal <cit.>, and whether a sharp transition exists at all for specific models <cit.>. In particular, it has been debated whether the main features of the transitions are sensitive to the system realization (e.g., quantum circuits vs. Hamiltonians with or without interactions) and to implementation of measurements or monitoring. This includes a distinction between the protocols comprising rare strong measurements and those with frequent (or continuous) weak measurements. These different classes of protocols may possess identical long-time dynamics at the level of averaged density matrix, yet the latter is inappropriate for studying entanglement dynamics. It is worth noting that, in the vast majority of works in the field, either projective local measurements are employed, or an effective model of a continuous monitoring of the system's dynamics is introduced (typically formulated at the level of a stochastic Schrödinger equation). A realistic description of the measurement process <cit.> requires a microscopic consideration of the joint evolution of the measured system and the detectors. In this situation, without resetting (re-initializing) the detectors, not only the backaction of measurement affects the systems, but also an inevitable accumulated feedback of the system on the detector can bias the next measurements. This may influence the classification of generalized measurements into strong- or weak-measurement classes, as the effective measurement strength depends on the system's state along the quantum trajectory. This type of “memory-effect” correlations is absent in conventional models implementing measurements or monitoring. The correlated dynamics in the physical realization of the coupled system-detector setup can therefore be expected to exhibit novel features. A possible interplay of various types of criticality and correlations in measured systems <cit.> is another important facet of the problem. In this context, one may draw a certain analogy between a measurement-induced entanglement transition for a finite density of detectors and disorder-induced localization transition. At the same time, it is known that even a single impurity drastically affects the properties of one-dimensional correlated systems  <cit.>. For example, a weak impurity may cut a wire into two, suppressing transport, similar to localization. This calls for the consideration of the effect of a single detector on the correlated chain, especially when the detector (or a pair of detectors) is located near the bipartition cut that is used for the definition of entanglement. In particular, it is tempting to look for some features characteristic of entanglement transitions already in models with a minimum number of detectors, which would shed more light on the nature of the true transition. Furthermore, one may expect that, in a realistic measurement setup, the ancilla-system coupling could be “renormalized” by the correlation in the main chain, similar to the Kane-Fisher impurity problem <cit.>, which in turn would affect the entanglement transition. The above analogies with correlated wires with impurities serve as an additional motivation for our study. In this work, we investigate the influence of measurements on quantum many-body lattice systems by coupling the system to ancillary degrees of freedom. The projective measurements are performed on the ancillary sites only. A distinguishing feature of our approach is that, after projecting the ancillary sites, the latter are not re-initialized for the next measurement cycle. The effect of the measurements on the dynamics of the “main" system is mediated by interactions between it and the ancilla (“detector”), with no particle exchange between the two. Within the proposed framework, the dynamics of the system as a whole (main system plus ancilla) remains closed, aside from the projections of the detector sites at fixed discrete intervals. The projections are effected through nonunitary operators that implement Born's rule, which is the only assumption we make about the nature of the measurement process. Furthermore, the feedback effect of the main system on the dynamics of ancillary degrees of freedom is automatically taken into account in the course of joint evolution between consecutive projections. This procedure thus brings us a step closer to a realistic description of quantum measurements in interacting many-body systems, with the ancillary sites mimicking a “measurement apparatus.” As such, our protocol is applicable to an arbitrary many-body lattice system and can readily be adapted to current experimental settings of interest, including cold atoms, trapped ions, and superconducting qubits. We first demonstrate the power of this approach by outlining how the “measured” site density of interacting hard-core bosons on a one-dimensional lattice can be reconstructed from the ancilla dynamics. Next, we investigate the effect of a measurement backaction on the density distribution and the entanglement entropy, for one and two sites coupled to the detectors. The feedback of the main system on the detectors gives rise to what we dub a “quantum-Zeno-valve effect” (involving a blockade of the ancilla dynamics), which occurs either when the stroboscopic-projection frequency is commensurate with the ancilla Rabi frequency or at strong system-detector coupling. Finally, we analyze the dynamics of the system for a finite density of measured sites. We observe manifestations of a disentangling-entangling measurement-induced transition in the time- and system size-dependence of the entanglement entropy averaged over realizations of individual quantum trajectories. The dynamics features a broad distribution of the entanglement entropy and density fluctuations. We thus provide evidence for a measurement-induced transition driven by the change of ancilla-system coupling in a correlated quantum system with a realistic measurement setup. The paper is organized as follows. We formulate the model and specify the measurement protocol in Sec. <ref>. In Sec. <ref>, we outline the basic concept of “measuring” the main system by means of performing projections on the ancilla. In Sec. <ref>, we study the effect of backaction from repeated measurements with one or two ancillas, focusing on the density distribution and the entanglement entropy. Two different initial states of the main system are considered: a domain-wall state in Sec. <ref> and the ground state in Sec. <ref>. In Sec. <ref>, we present results for the entanglement and density dynamics in a chain with all sites coupled to the detectors, with a particular focus on the disentangling-entangling transition. Section <ref> provides a summary and conclusions. Some of the technical details of performed numerical simulations and additional benchmarks are described in Appendices. § MODEL, MEASUREMENT PROTOCOL, AND OBSERVABLES §.§ Model For concreteness, we consider a model of hard-core bosons on a one-dimensional lattice at half filling. The Hamiltonian of the chain (main system) is given by ℋ_s = ∑_i=1^L-1[ -J/2(b_i^† b_i+1 + H.c.) + Un̂_in̂_i+1], where b_i (b_i^†) annihilates (creates) a boson on site i, the lattice size is L, n̂_i ≡ b_i^† b_i denotes the density operator on site i, J is the strength of hopping between neighboring sites, and U is the strength of interaction between particles on neighboring sites. This model is equivalent, through the Jordan-Wigner transformation, to the XXZ Heisenberg chain, and its ground-state and dynamical properties have been studied extensively <cit.>. The simplest setup with an ancilla is then realized when the ancilla is represented by a single pair of sites. Taking L to be even, we couple, by means of interaction, site L/2 of the main system and one of the sites of an ancillary pair, as illustrated in Fig. <ref>. The total Hamiltonian ℋ is then written as ℋ = ℋ_s + ℋ_a+ℋ_sa, where ℋ_a = -J/2( a_1^† a_2 + a_2^† a_1 ) describes the ancilla and ℋ_sa = M n̂_L/2 a^†_1 a_1 the coupling between the main system and ancilla. Here, a_1,2 (a^†_1,2) are the annihilation (creation) operators of hard-core bosons on ancilla sites 1 and 2, and M is the strength of interaction between the main system and the ancilla. Note that the choice of interaction in Eqs. (<ref>) and (<ref>) in the form of only particle-particle interaction breaks particle-hole symmetry of the model. Everywhere below, the intra-chain and chain-ancilla interaction is taken to be repulsive (U>0 and M>0). For simplicity we have also chosen, in Eq. (<ref>), the hopping strength in the ancilla to be the same as in the main system. In what follows, we first study the basic setup with a single ancillary pair, as described above. We proceed by including a second ancillary pair, as shown in Fig. <ref>, permitting a more detailed understanding of the influence of measurements on the system. Finally, we employ this approach to the case where the ancillary sites are attached to every site of the chain, see Fig. <ref>. §.§ Ancilla measurement protocol We perform projective density measurements on the lower ancilla site (the one connected to the main chain through interaction), periodically at an interval Δ T. Consider the case of a single ancilla (multiple ancillas are treated similarly). After each projection, the ancilla pair finds itself in a state with either the lower site occupied or the lower site empty, i.e., the whole system is then in the state |Ψ⟩_10⊗ |1 0 ⟩ or |Ψ⟩_01⊗ |0 1⟩, respectively, where |10⟩ and |01⟩ denote the ancilla states, and |Ψ⟩_10 and |Ψ⟩_01 are the corresponding states of the main chain. Because of the backaction of the measurement onto the main chain, the states |Ψ⟩_10 and |Ψ⟩_01 are in general different. The probability of each outcome is determined by the Born rule according to the density n_a = ⟨ a_1^† a_1 ⟩. Note that the projection breaks both unitarity and integrability. We model the projection as an instantaneous event. Importantly, the probability of a particular readout shown by the “measurement apparatus” (the ancilla pair) is dependent on the whole history of the measurement process. Indeed, because of the Born rule, the previous measurement results affect the dynamics in the main chain (backaction). This, in turn, affects the value of n_a at the time when the projection takes place (feedback), and so on. This backaction-feedback memory loop gives rise to system-ancilla correlations. In particular, a self-sustained “blockade” or “freezing" of the ancilla dynamics may occur, resulting in an almost projective measurement of the system site, which leads to a quantum Zeno-like effect. Numerically, the unitary time evolution of the entire system—main chain plus ancilla(s)—is obtained using the time-dependent variational principle <cit.>, see Appendix <ref>. For the model with ancillas attached to every site of the chain (Fig. <ref>), we move beyond exact simulations and utilize the full power of matrix-product-state (MPS) simulations, considering systems of up to 72 sites, including ancillary sites. It is worth mentioning that the implementation of generalized measurements through coupling the system to ancillary degrees of freedom followed by their projections has deep roots in various contexts involving weak measurements (see, e.g., Refs. <cit.>). In particular, this type of measurement allows one to associate the counting statistics <cit.> of discrete outcomes of ancilla projections with the properties of the system without strong backaction. Furthermore, these outcomes can be used in a feedback loop to control the system by active-decision making on further measurements or the unitary evolution (see, e.g., Refs. <cit.> and references therein). Microscopically, perhaps, the closest setup resembling ours was introduced in Refs. <cit.>, where a double-dot (“two-site”) system was electrostatically coupled to a point contact in a conducting channel (see also Refs. <cit.> for related setups). However, the roles of ancilla and system were interchanged in that setup compared to ours (cf. Ref. <cit.>): the transmission of the conducting channel was measured strongly, yielding the information about the occupation of the dot connected to it. In a recent work <cit.>, a setup with ancillas was considered, in the context of measurement-induced transitions, for a quantum circuit. There are, however, essential differences between our framework and the one studied in Ref. <cit.>. A clear difference is that we deal with a Hamiltonian system undergoing real time evolution, and which respects particle number conservation. Another difference is in the dependence of the measurement outcome on the history of the system. As mentioned above, this dependence is an inherent property of our model. By contrast, a newly initialized (in a prescribed state) ancilla is introduced in Ref. <cit.> at every “time step” (at every layer of the quantum circuit), similarly to the steering protocols of Refs. <cit.>. This is an important difference, particularly because adaptive feedback and “preselection” mechanism may prove essential, as argued in Refs. <cit.>, for observing a measurement-induced phase transition. In our framework, the feedback emerges “automatically”, through the mechanism of joint unitary evolution fixed by the Hamiltonian, with no re-initialization of the ancilla after the projections. This results in effective adaptive dynamics of the entire system. A further key difference with previous works is that we explore the density and entanglement dynamics in various setups, with one, two, and many detectors. §.§ Observables We focus on the following observables: (i) the density distribution n_i(t) = ⟨ b_i^† b_i ⟩ (t) of bosons along the main chain, as a function of time t, (ii) the density n_a(t) = ⟨ a_1^† a_1 ⟩ (t) of the ancilla boson at the lower site (the one coupled to the main chain through the interaction with magnitude M), and (iii) the bipartite von Neumann entropy of entanglement S(t). The latter is defined as S(t) = -Tr( ρ_A lnρ_A ), ρ_A = Tr_Bρ , where ρ is the density matrix of the whole system, ρ_A is the reduced density matrix (Tr_B denotes a trace over subsystem B), and we place the bipartition between subsegments A and B of the system in the middle of the main chain. Note that for an ancilla, both its sites also belong to one of the subsystems A or B. Since the ancilla is entangled with the main chain during the unitary evolution between the projections, the ancillary degrees of freedom contribute to the value of S. From the density in the main chain n_i, we can obtain the particle imbalance: ℐ = -L/4 + ∑_i=1^L/2⟨n̂_i ⟩. This quantity measures the bipartite fluctuations with respect to a homogeneous distribution of particles in the main chain, so that ℐ = 0 corresponds to a homogeneous state, and ℐ = ± L/4 is a state where all particles are in the left (+) or right (-) side of the chain respectively. Recall that since we consider the main chain at half filling, there are L/2 bosons present there. All three quantities (i)-(iii) are, in principle, directly measurable in experiment; the imbalance was considered for instance in Ref. <cit.> (defined in a slightly different way). The entanglement can also be measured in state-of-the-art experimental settings <cit.>. Within the framework outlined in Sec. <ref>, a measurement apparatus in the form of an ancilla pair can be efficiently used to “measure” the density distribution by means of projections performed only on the ancilla. § RABI OSCILLATIONS IN THE ANCILLA COUPLED TO THE MAIN CHAIN Before moving on to study the effect of ancilla projections on the main system, let us outline the basic concept of gaining information about the main system from the dynamics of the ancilla. It is based on the use of the fact that the probability of projection to either “particle" or “hole" in the ancilla dynamics is controlled by the expectation value of the ancilla density n_a, introduced in Sec. <ref>, right before the moment of projection. Since the ancilla site at which the projection is made is coupled by particle-particle interaction to the chain, the unitary dynamics of n_a between consecutive projections provides information about the occupation of the chain site to which it is connected. In the following Sec. <ref>, we will proceed to analyze the effect of the ancilla on the main chain, as opposed to the effect of the main chain on the ancilla as we consider in this section. In the absence of coupling between the ancilla and the main chain (M=0), the ancilla dynamics is that of a two-level system subjected to periodic projections. Specifically, the ancilla displays Rabi oscillations, with the frequency given by |J|, and is reset after each projection performed on the lower ancilla site in either the |1 0⟩ or |0 1⟩ state, after which the Rabi oscillations continue with the same frequency. For M≠ 0, the interaction between the main chain and the ancilla modifies both the characteristic Rabi frequency and the oscillation amplitude. As an instructive simple example, consider the static mean-field approximation for the configuration of Fig. <ref>. In this approximation, the density operator n̂_L/2 for site L/2 in the Hamiltonian is replaced by its expectation value n = ⟨n̂_L/2⟩ and its time dependence is neglected. The ancilla boson then “feels” the time-independent potential Mn. The Rabi frequency Ω for the oscillations of n_a(t) becomes Ω = √(J^2+M^2n^2), growing as M is increased, while the amplitude A of the oscillations is reduced: A = 1/2J^2/J^2+M^2n^2. Note that Eqs. (<ref>) and (<ref>) correctly reproduce the limit of M=0, giving, in particular, A=1/2 for the unperturbed two-level system. Henceforth we set ħ = 1, express energies in units where J=1, and time in units of 1/J. The ancilla dynamics in the setup of Fig. <ref> (with the domain-wall initial state of the main chain) for 0<t<Δ T and various choices of M is shown in Fig. <ref>, with the colored lines denoting the numerical results obtained from a full simulation of the entire system (for the case of U=1) and the dashed lines corresponding to Eqs. (<ref>) and (<ref>), for the case of site L/2 fully occupied (n=1) at t=0. The case of M=5 demonstrates a very good agreement between the numerical data and the two-level mean-field approximation. This is because the ancilla site in the limit of M≫ 1 acts as a pinning “impurity" for the main chain. The small-amplitude Rabi oscillations in the ancilla are then a weak fast-oscillating perturbation of the otherwise slow dynamics of the entire system (tunneling processes in the main chain that change the occupation of the pinning site are weak for large M, yielding no appreciable change of n for M=5 within the time interval Δ T in Fig. <ref>). For intermediate M (exemplified by the cases of M=1 and M=3), a noticeable deviation becomes apparent, resulting, apart from fluctuations around the mean-field solution, from the time evolution of n associated with a melting of the domain wall. Qualitatively, however, Eqs. (<ref>) and (<ref>), which become exact in the limits of both small and large M, capture well the leading time-dependent behavior in n_a(t) for arbitrary M. It is worth noting that the expectation value of the ancilla density for M=3 in Fig. <ref> appears to be very close to 1 at the projection time t=Δ T =2, similarly to n_a for M=5, but in a sharp contrast to the case of M=1. This is because of an approximate resonant condition: ΩΔ T=2√(10)≃ 2π for M=3 and n=1. As a result, at the time of projection, the ancilla with an intermediate-strength coupling (M=3) is felt by the system as if the coupling would be extremely strong. In other words, the resonance condition makes the potential of the ancilla exerted on the bosons in the chain similar to the potential of a strong impurity. The probability of having no potential for a quantum trajectory (a_1^† a_1 projected to 0) at t=Δ T vanishes upon approaching the exact resonance. This resonance is, however, in general broadened through the breakdown of the two-level approximation for the Rabi oscillations. The dynamics in the chain breaks down the resonance condition by changing n in Eq. (<ref>). However, as discussed above, a strong impurity (either with M≫ 1 or in the resonant case) slows down this dynamics, resulting in repeated measurements of n=1 (i.e., the quantum Zeno effect). Only rarely the occupations may change for the resonant coupling, which immediately removes the pinning, thus opening the channel underneath the ancilla. This behavior is reminiscent of the operation of a valve, hence we dub it the “quantum-Zeno-valve effect" (see the detailed discussions of implications of this phenomenon in Sec. <ref> below). The procedure involves an inherent trade-off. On the one hand, the longer Δ T, the stronger the effect of the interaction between the main system and the ancilla (the “measurement apparatus”). On the other hand, as Δ T increases, the effect of the main chain on the ancilla is averaged over a longer time interval. This means that if one endeavors to obtain information about the main chain through subsequent measurements of the ancilla, this information will then be “blurred” over the time window Δ T. In Fig. <ref>, this effect is seen through the increasing difference in time between the simulated curves and those for the static mean-field approximation of Eqs. (<ref>) and (<ref>). Another effect of a similar kind concerns the strength of interaction M. The characteristic Rabi frequency increases with growing M, i.e., the effect of interaction on the ancilla dynamics becomes apparent at earlier times for larger M (Fig. <ref>). At the same time, the effect of the ancilla on the main chain also increases as M grows. In the limit of large M, the measurement of a particle at the ancilla site leads, as already mentioned above, to a complete freezing of dynamics in the main chain. Importantly, the backaction, like the projective measurement, is instantaneous and results in a “spooky action at a distance” at every site of the chain. In particular, the (expectation value of the) particle density at each site of the main chain changes discontinuously at each measurement step. Building on the preceding discussion, the remarkably simple setup with a two-level ancilla can be utilized to “measure” n in the main system through a subsequent determination of the ancilla dynamics. The interested reader is referred to Appendix <ref> for a proof-of-principle discussion concerning how such a measurement protocol might be effected in practice. In the following sections, we will focus on the effect of the ancilla projections on the main chain dynamics, and investigate possible signatures of a measurement-induced transition. § EFFECT OF REPEATED, SPATIALLY LOCALIZED MEASUREMENTS ON THE MAIN CHAIN Having discussed the basic concept of coupling an ancilla to a many-body Hamiltonian system in Sec. <ref>, we now turn to an analysis of the effect of projective ancilla measurements on the main chain, where said measurements are performed repeatedly with a given frequency. We initialize the detectors by placing a boson in the ancillary pair at the lower site, with the other site empty. For the main chain, we consider two different initial states, both with half filling (L/2 bosons). We first examine, in Sec. <ref>, the initial state in the form of a product state with a domain wall in the middle, such that the leftmost sites are fully occupied, as depicted in Fig. <ref>. This initial state is especially convenient for exploring the effect of measurements on the relaxation of inhomogeneous density distribution, as characterized, in particular, by the particle imbalance, Eq. (<ref>). For this initial state, the measurement is performed with a single ancillary pair. We then proceed, in Sec. <ref>, to analyze the case where the initial state is the ground state of Eq. (<ref>) (with decoupled ancilla sites, M=0), and use a single ancilla, as well as two ancillary pairs at the neighboring sites of the main chain, as illustrated in Fig. <ref>. Choosing this initial state has the benefit that the system, in the absence of measurements, does not evolve and remains in its weakly entangled state, as opposed to the domain-wall initial state. This is particularly useful for studying the entanglement dynamics. The setting with two ancillas illuminates the nature of the feedback between the detectors in setups with multiple ancillas. Finally, in Sec. <ref>, we will attach the ancillary pairs to every site of the main chain, which will be again initialized in its ground state. §.§ Domain-wall initial state with a single ancilla We begin by analyzing the observables in a single-ancilla setup of Fig. <ref>. Starting from the domain-wall state, we observe a nonmonotonic dependence on the coupling constant M for both the domain-wall melting rate and the rate at which the entanglement grows. This makes evident a competition between different measurement-induced effects. For M=0 the dynamics is integrable, and the domain wall displays relatively slow melting dynamics, as shown for reference in Fig. <ref> in Appendix <ref>. This behavior is similar to that studied in Refs. <cit.> and thought to belong to the Kardar-Parisi-Zhang class (see Refs. <cit.> and references therein). §.§.§ Effect of measurements on averaged properties As M is increased, the melting dynamics first accelerates as a result of integrability breaking induced by measurements, which is exemplified by the case of M=1 in Fig. <ref>(a). Indeed, we observe in the right panel that the number of particles transported through the center of the system has reached a value of almost L/4 for t = 200, indicating the almost complete melting of the domain wall. Transport is therefore faster than in the uncoupled case M = 0. Repeated measurements on the ancilla for M=1 also increase the entanglement growth in the system quenched from the domain-wall state, compared to the case [see Fig. <ref>(c)]. Recently, a similar measurement-induced enhancement of the entanglement growth in a Hamiltonian system was also found in the quantum Ising chain <cit.>. Note that the acceleration of the entanglement growth by switching on coupling of the main chain to ancilla at M=1 is due to the effect of repeated measurements, but not the effect of coupling as such, see Appendix <ref>. Generally, in the limit of large M, strong coupling between the main chain and the ancilla overcomes the effect of integrability breaking and, as already discussed in Sec. <ref>, brings dynamics to a stop. Domain-wall melting is arrested in this limit, and the ancilla boson is prevented from tunneling to the upper ancilla site by forming a repulsively bound pair with the boson on the measured site in the main chain. Projective measurements of the ancilla site only very rarely destroy this correlated state (see discussion below). A trend towards the freezing of the system dynamics is apparent by comparing the cases of M=1 and M=5 in Fig. <ref>. Dynamics slows down for M=5 compared to M=1, as is attested by the results in Figs. <ref>(a), <ref>(b), and <ref>(c) for the particle density in the main chain, the particle density in the ancilla, and the entanglement entropy [see the thick black lines in Fig. <ref>(c) that depict the average over 40 quantum trajectories] respectively. When M is increased to M = 50, we find no appreciable melting of the domain wall or growth of the entropy (see Appendix <ref>). Aside from the general trend of, first, an increase in the melting rate as M is increased to M ≈ 1, and, second, a slowing down as M is further increased to M ≫ 1, the dependence on M is more intricate. Specifically, dynamics of the system shows oscillatory behavior as a function of M between M∼ 1 and M→∞. This is because of a possible commensurability of the Rabi frequency in the ancilla and the measurement rate 2π/Δ T. Although the density oscillations in the ancilla for a given M are not exactly periodic, their spectral weight is peaked around a certain frequency, the existence of which derives from the static mean-field approximation discussed in Sec. <ref>. One of the consequences of this effect is an enhanced likelihood of obtaining the same result for two consecutive projections if Δ T/2π is close to an integer of the inverse peak frequency. For instance, under this condition, if a projection in the ancilla yields 1 for the occupation of the lower ancilla site, then the Born rule for the next projection most likely also yields 1 (quantum Zeno effect). Thus, when the resonant condition is closely satisfied, the effective measurement strength is enhanced, suppressing dynamics. This is the case for M=3, as seen from Fig. <ref>, where dynamics for all three observables (n, n_a, and S) is distinctly slower for M=3 than for either M=1 or M=5. §.§.§ Fluctuations: trajectory branching and the quantum Zeno valve effect The mechanism by which the dynamics of the entanglement entropy S, averaged over realizations, is suppressed with increasing M reveals itself when we study the evolution of the entropy for particular quantum trajectories. These results are shown in Fig. <ref>(c). For M=1, the picture of S resolved with respect to realizations is relatively featureless. For individual quantum trajectories, the entropy grows gradually, without sharp acceleration at any point in time, and closely approaches the thermal value (the Page entropy) already on a timescale on the order of 10^2. The distribution of S over different quantum trajectories for M=1 is rather sharply peaked at any given t, with fluctuations becoming smaller and smaller as S approaches the Page value. This forms a well-defined reference curve for the (“thermalized”) behavior of the entropy. Fluctuations between different realizations grow as a function of M. Of particular interest is the manner in which the fluctuations become more prominent for M=3, when the system, as discussed above, exhibits a resonance between the Rabi frequency in the ancilla and the measurement rate. As a striking feature of the realization-resolved entropy for M=3, a given quantum trajectory exhibits a slow growth of S, identical to many other trajectories, until the trajectory suddenly “branches off”—this process is visualized as kinks on the trajectories in Fig. <ref>(c)—and starts quickly trending towards a highly entangled state that evolves similar to a typical state for M = 1. The delay time t' at which the branching-off from the least entangled state occurs for M=3 is seen to be widely spread in Fig. <ref>(c) within the interval of observation. The effect of quantum-trajectory branching can be understood by considering the probability of measuring a hole on the ancilla site. Since the system for M=3 is close to the resonance (see Sec. <ref> above) and the ancilla is prepared at t=0 in a state with n_a=1, ancilla projections under these conditions measure a particle during a long sequence of measurements. This explains the relatively slow (compared to the case M=1) growth of S for time scales much larger than Δ T. Once a hole is measured, which is a rare event near the resonance, the domain wall starts to melt much more rapidly, which triggers the rapid growth of the entanglement entropy. There hence emerges a valve effect (“quantum-Zeno-valve effect"), in which the measurement of a hole moves the system away from the resonance and the system never evolves back towards it. The quantum trajectories eventually all end up in a thermal state once they have moved away from the resonance. The sharp branching-off of quantum trajectories, associated with the measurement of a hole in the ancilla, is in fact a more robust feature of the entanglement spreading in the setup with a domain wall, not necessarily related to the commensurability of the Rabi frequency and the measurement rate. Namely, it is a feature of the entanglement dynamics also for arbitrary M≫ 1. In this limit, however, the mechanism of branching is different compared to the resonant case. The amplitude of Rabi oscillations given by Eq. (<ref>) is for M≫ 1 strongly suppressed, so that the likelihood of measuring a hole is small irrespective of the resonant condition. This is illustrated in Fig. <ref>(c) by the behavior of individual quantum trajectories for M=5, which is qualitatively similar to that for M=3. Note, however, that for M = 5 a larger fraction of trajectories have moved towards a thermal-type state, compared to the resonant case M = 3. The evolution of the behavior of S for individual trajectories as a function of M involves an interplay of the above two mechanisms of branching, which leads to a nonmonotonic dependence of a typical delay time t' on M. Namely, for M=5, it is much larger than for M=1, but substantially smaller than for M=3. More specifically, the distribution of t' over trajectories for given M is parameterized in Fig. <ref>(c) by plotting the evolution of S(t) for a series of representative trajectories, sorted by the entropy integrated over time within the interval of observation, Ξ = ∫_0^t_fdt S(t), for the final time t_f = 200. For each M, we have shown the curves having the highest Ξ (purple line), the lowest (blue), and the median (green). We have also plotted the curves that correspond to Ξ higher than 25 (orange) and 75 (red) percent of curves respectively. For M = 1, the characteristic t' is of the order of unity, but the delay is seen to dramatically increase by two orders of magnitude to t'∼ 10^2 for M = 3, with the median trajectory (green curve) branching off at t'≈ 175. No branching whatsoever occurs for M=3 within the interval of observation for the least entangled trajectories in the bottom 25th percentile. The fact that the typical delay time is so large for M=3 attests to the resonance as the prime reason for the delayed branching. For M = 5, the delay time for the median trajectory decreases, compared to M=3, to t'≈ 60, pointing to the suppression of Rabi oscillations as a mechanism of the delay. For very strong coupling, M = 50, no branching at all is observed within the time t_f=200, see Fig. <ref> in Appendix <ref>. §.§ Ground state as the initial state In Sec. <ref>, we have considered the case of a single ancilla coupled to the main chain, initialized in a domain-wall state. Let us now turn to the case where the initial state is the ground state of the uncoupled system (M=0) with U = 1/4, as schematically depicted in Fig. <ref> for two ancillas. This initial state is homogeneous (aside from boundary effects), features power-law correlations of the Luttinger-liquid type, and a finite entanglement entropy that depends logarithmically on L <cit.>. A useful feature of this choice is that measurement-induced disentangling behavior of the system can be probed more easily, as the entanglement entropy then decreases with respect to that in the initial state <cit.>. Conversely, we shall find that the quantum-Zeno-valve effect is less pronounced in this case, since domain wall melting is initially constrained by dynamics at the domain wall only, while particles can more easily rearrange themselves starting from the half-filling delocalized ground state. This means that the dynamics also more easily moves away from the resonant condition as discussed in Sec. <ref>. §.§.§ Single ancilla In Fig. <ref>, we show results for the entropy dynamics for a single ancilla, when the main chain is initialized in the ground state. In general, the qualitative behavior of the entropy is analogous to the case of the domain-wall initial condition (Sec. <ref>), see Fig. <ref>(c). For M=1, the entropy quickly grows with relatively small fluctuations that decrease with time. For intermediate coupling, M=5, the quantum-Zeno-valve effect is apparent in the form of a sharp branching-off from the low-entanglement curve with a broad distribution of time t'. For very strong coupling, M=50, we observe a clear bimodal distribution of the entropy. Specifically, for about half of the quantum trajectories, the system quickly “thermalizes,” whereas the other half of trajectories stay at low entanglement close to that of the initial state. This behavior can be understood as follows. The initial quench—coupling the ancilla to the chain—introduces high energy M in the components of the wave function with n_L/2=1. After projecting the ancilla, this component survives in about half of trajectories. After this, a blockade occurs for these trajectories, in a full analogy with the case of the domain wall (where n_L/2=1 initially for all the trajectories), see Sec. <ref> and Fig. <ref>. The stability of such a “repulsive bound state” formed on the ancilla site and the chain site with i=L/2 is explained by the energy conservation constraint. Indeed, this highly excited state cannot easily decay, since M is much larger than the bandwidth. We conclude that for a single ancilla with sufficiently large M, independently of the initial state, thermalization is accompanied by a formation of a bimodal distribution of quantum trajectories. The trajectories show sharp branching-off from the low-entropy to the high-entropy class of states, with a broad distribution of branching times t'. §.§.§ Two ancillas With a view to ultimately connecting to the physics behind measurement-induced transitions <cit.> and as a step towards the finite density of measured sites (Sec. <ref>), we enhance the effect of measurements by considering now two ancilla pairs, one connected to site L/2 and the other to site L/2 + 1 in the middle of the chain (see Fig. <ref>). This setting also allows us to explore possible correlations, mediated by the dynamics of the main chain, between multiple ancillas. The measurement part of the Hamiltonian is a straightforward generalization of Eqs. (<ref>) and (<ref>) with identical interaction (M) and hopping (J) constants for both ancilla pairs. In the measurement protocol, a projective measurement is performed on the left ancilla pair (i = L/2) first, followed after an infinitesimal time interval by a projective measurement on the right one (i = L/2 + 1). In a sequence of repeated measurements, a “double salvo" of such two measurements is separated from the next one by the time interval Δ T. As in the domain-wall setup, the bipartite entanglement entropy S is computed with respect to partition in the middle of the main chain. Considering the quench from the ground state, we analyze statistics over a larger number of individual trajectories, namely about 300, compared to 40 in Sec. <ref>. The numerical results for M = 1, 5, 50 in the double-ancilla setup of Fig. <ref> are shown in Fig. <ref>. For M=1, the entanglement dynamics for the quench from the ground state with two ancillas is very much similar to that for a single ancilla. Specifically, the entanglement entropy grows with time in a smooth manner, gradually evolving to the thermal value, with a rather narrow distribution of the entropy over individual quantum trajectories. The shape of the typical entropy curves here is similar to the curve in Figs. <ref>(c) and <ref> for M=1. The trajectory-averaged ancilla density n_a [Fig. <ref>(b)] quickly loses memory of the initial condition for M=1, fluctuating around the average n_a = 1/2, also similar to the single-ancilla case [Fig. <ref>(b)]. When coupling M becomes stronger, M=5, the distribution of entropy broadens in analogy with the case of a single ancilla. At the same time, an essential difference is observed. Specifically, the entropy growth for individual quantum trajectories does not show sharp branching between the two well-defined types of behavior described by the low- and high-entanglement curves (compare the panels for M=5 in Fig. <ref> and Fig. <ref>). Instead, the growth follows individual curves that are characterized by a rather broad distribution without apparent bimodal features. This can be attributed to the mutual “interference” effect of the two ancillas, tending to blur the quantum-Zeno-valve phenomenon. Another striking difference is the emergence of up-and-down jumps, about 0.5–1 in magnitude, visible in the quantum trajectories of the entropy for M=5 and M=50. They are the signature of measurement-induced fluctuations of the particle number in halves of the chain, as is evidenced by a comparison between the timing of the jumps in the entropy and the timing and position of the density fluctuations. Note that the size of the jumps is close to the contribution to the entanglement entropy associated with a single Bell pair: ln 2. For M=50, the entropy behaves similarly to the single-ancilla case in that there are two classes of trajectories (with low and high entanglement). However, there are also clear differences. First, a smaller fraction of trajectories features high entanglement. This is because the blocking by two ancillas is more efficient, since it is operative when at least one particle occupies sites L/2 and L/2+1. Thus, the averaged entropy is strongly suppressed for the case of two ancillas. This provides indication for a formation of a disentangling phase for many ancillas, as discuss below in Sec. <ref>. To complete the statistical description of quantum trajectories, we consider the distribution of the entanglement entropy as a function of time. We characterize the distribution by the variance: var_S(t) = 1/R∑_j=1^R [S_j(t) - S_avg(t)]^2, where R is the number of realizations and S_avg is the mean value of the entropy. The result is shown in Fig. <ref>, where a strong dependence on the coupling strength M is evident, with a crossover around M=4 from a narrow distribution to a wide one. The fluctuations of S are nonmonotonic in time, with an initial increase of the distribution width followed by a decrease, except for the largest values of M. The narrowing of the distribution can be attributed to more trajectories converging to the thermal state for large t. For M=25 and 50, this narrowing is not seen because only a few rare trajectories are strongly entangled. §.§.§ Density distribution in the main chain for two ancillas In Fig. <ref>(a), we show the averaged density in the chain as a function of time. For M=1, we observe an initial suppression of the density at the sites coupled to the ancillas. This inhomogeneity is faded away with time. For strong coupling, M=50, a very different behavior is observed: the density is enhanced and this enhancement stays approximately constant within the observation time. This is a manifestation of a formation of a repulsive bound state, discussed above (for a single ancilla). To further illustrate this, we present the density dynamics on the least and the most entangled trajectories for various M in Fig. <ref>. The panels for the least entangled states at M=5 and M=50 in Fig. <ref>(a) clearly demonstrate the emergence, as M is increased, of a stable blocking region, with at least one of the two sites connected to the ancilla staying occupied. By contrast, the most entangled trajectories feature chaotic dynamics with frequent occurrence of holes on these sites. Recall that holes do not interact with the ancilla in Eq. (<ref>), whereas a particle on the site coupled to the ancilla is pinned by strong interaction with the ancilla particle. This is much the same physics as pertinent to freezing-out of the system with the domain-wall initial state in the limit of large M. We also observe that the dynamics of the density is correlated with the entanglement dynamics. As an example, the one-to-one correspondence between the entropy jumps and the fluctuations of the particle number imbalance can be clearly seen by comparing the evolution of S on the least entangled trajectory for M=5 in Fig. <ref>(c) and the density fluctuation for M=5 on this trajectory in Fig. <ref>(a). The presence of spikes in the entropy curves for the double-ancilla setup, as opposed to the case of a single ancilla, again indicates the importance of correlation in the dynamics of multiple ancillas caused by their interaction with the same main chain in the course of joint unitary evolution. § MEASUREMENTS ON EVERY SITE With the above detailed analysis of the previous cases with one and two ancillas, we are now in a position to proceed with the setup where each site of the main chain is coupled to an ancilla pair. §.§ Measurement protocol The consideration of a finite density of ancillas allows us to study the ancilla-measurement approach in the context of entanglement transitions. As in Sec. <ref>, we initialize the main chain in the ground state with U = 1/4. We address system sizes up to L = 24, so that the total number of sites in the system is L_tot = 24 × 3 = 72. This means that we can no longer use an unrestricted bond dimension (i.e., exact numerical simulations) as before. We furthermore restrict our time window to t ∈ [0, 50] as the simulations become more computationally demanding. The measurement protocol is now as follows. At each step Δ T, we perform projective measurements on all ancilla sites connected to the main chain, using the Born rule as before. These measurements are performed consecutively from left to right, so that each ancilla is projected at every measurement cycle, with an infinitesimal delay between the measurements on neighboring sites. §.§ Overview of entanglement dynamics In the double-ancilla case, we observed that having two ancilla pairs in the system is not sufficient to reach a disentangled state for every trajectory, even for very large measurement strength M = 50. It is therefore of interest to see whether a disentangling phase forms in the limit of a large number of ancilla pairs. Furthermore, it is even more intriguing questions whether our setup yields an entanglement transition and what are then the properties of the entangling phase. In particular, while we observed a thermalizing (reaching nearly maximum entanglement) behavior for one and two ancillas with M=1, it is not obvious whether this behavior survives in the case of all sites coupled to ancillas. In Fig. <ref>, we show the resulting dynamics of the entanglement entropy for the chain length L=20 and various choices of M = 1, 3, 4, 5, 10. For all values of M we see an initial growth of S at t≲ 10, induced by the quench. For longer times the dynamics for different couplings is very different. For large measurement strength, M = 10, we observe that entanglement is suppressed on average, as shown by the thick black line, compared to entanglement in the initial state. The most entangled trajectory as quantified by Ξ (thin purple line), unlike the double-ancilla case, also exhibits suppressed entanglement. We have verified (result not shown) by comparing the cases of bond dimensions χ=128 and χ = 256 that the value of the average entropy in the large-M limit (M ≥ 5) is not dependent on the cutoff determined by the bond dimension. For sufficiently weak measurements, M = 1, we observe a fast growth of entanglement with a quick saturation at the cutoff value. The entropy follows the curve that is almost identical in shape to those for the single- and double-ancilla setups at M=1, which approached the Page value proportional to the system size (see Sec. <ref>). Therefore, we expect that in this case the exact dynamics (which is numerically inaccessible for long times t≳ 15 due to strong entanglement growth) should closely approach the Page value indicated by the dotted black line. The clear difference in the entanglement dynamics for M=1 and M=10 strongly suggests an existence of a measurement-induced entanglement transition in between. Let us now inspect the entanglement dynamics for the intermediate values of M. The dynamics for M = 5 is very similar to the case M = 10, indicating that M=5 also belongs to the disentangling phase. For M=4, we observe a very weak increase of the entropy, which is followed by a saturation at a value that is only slightly larger than the initial one. For M = 3, there is a clear trend of increasing entropy with time, although the increase is much slower than for M=1. At this value of M, we observe a dependence on the bond dimension χ at later times. This consideration suggests that the entanglement transition takes place in the vicinity of M=4. To shed more light on this, we will consider the L-dependence of the entropy in Sec. <ref>. For all values of M>1 in Fig. <ref>, we see a broad distribution of the entanglement entropy over quantum trajectories. At the same time, no bimodality features are observed in these distributions. We argue that the bimodality that we encountered in the single- and two-ancilla cases (Sec. <ref>) is smeared by the interference of multiple ancillas in our implementation of measurements. §.§ System-size dependence of entropy and measurement-induced transition In Fig. <ref>, we investigate the disentangling-entangling transition more closely by showing the behavior of the entropy as a function of time for various system sizes L = 8, 12, 16, 20, 24 for M = 1, 3, 4, 5. For M = 1 (entangling phase) we observe a clear dependence on the system size. Furthermore, we observe a fast increase, approximately linear in time for t ≲ L, of the entropy towards its maximum value. For small systems (L=8, 12), this maximum is determined by the Page entropy and for larger system sizes (L ≥ 16) it is determined by the cutoff established by the bond dimension χ. This implies that in this case the exact dynamics is reproduced up to times t≈ 15 for L ≥ 16. In view of the bond-dimension cutoff, the curves for L =16, 20, and 24 remain nearly identical in the whole considered time range. The saturation value increases approximately proportional to L for L= 8, 12, and 16. This indicates a volume-law dependence of the entanglement entropy. Weak fluctuations around the average, as seen in Fig. <ref>, like in the single- and double-ancilla setups for M=1 (where the entropy reached the Page value), are also suggestive of volume-law behavior. A volume law for not too large M is consistent with earlier results on interacting Hamiltonian systems obtained with different implementations of measurements <cit.>. On the other hand, for M = 5, the curves collapse within error bars for different L, a feature of area-law scaling. In the intermediate case M = 4, we observe a very weak growth with system size, just about distinguishable within error bars. This behavior can be potentially described as a logarithmic growth with a relatively small prefactor (which is still larger than that in the Luttinger-liquid ground state <cit.>). Alternatively, the weak L dependence here can, in principle, be attributed to finite-size corrections convergent to an area-law curve. More work is needed to distinguish between the two possibilities. Note that this value of M corresponds to the crossover scale identified in the previous section, associated with the width of the distribution of the entropy for different trajectories in the two-ancilla case (see Fig. <ref>). For M=3, we observe a clear increase of the long-time value of entropy (note that the largest-L curves start to be slightly affected by the bond-dimension cutoff, as mentioned above). This supports the interpretation of M=3 belonging to the entangling phase. This analysis confirms our above expectation that the transition is located near M=4. A more precise determination of the critical measurement strength would necessitate larger system sizes, as well as a longer time window for the simulations. Likewise, unambiguously distinguishing a volume-law asymptotic behavior from a possible critical phase with logarithmically growing entanglement with a large prefactor is difficult in numerical simulations in view of system-size and finite-time limitations. §.§ Density fluctuations To make a connection between the entropy and density dynamics, we also analyze the fluctuations in the particle density by considering the particle imbalance ℐ. As before, this quantity measures the number of particles in the main chain that have moved from the right side of the system to the left side. We depict the dynamics in the bottom panels of Fig. <ref>. As might be expected, these fluctuations generally increase as a function of M, with the average state nonetheless being close to homogeneous density. For the results in the disentangling phase, with M = 5 and M = 10, we find occasional “plateaus” in the imbalance, associated, in particular, with weakly entangled trajectories (thin blue lines in both upper and lower panels of Fig. <ref>). Such a plateau can be associated with weakly entangled trajectories corresponding to blocking regions in the two-ancilla case (see Fig. <ref>a). In the many-ancilla case, such a blocking region can occur at any point in the chain. This is illustrated by considering the trajectories corresponding to the weakest bipartite entanglement, as shown in Fig. <ref>. Here we again see a clear qualitative difference between the cases M = 3 (entangling phase) and M = 5 (disentangling phase). In the former case, the density in the main chain rarely approaches the fully polarized values corresponding to a particle or hole, but shows only moderate fluctuations around the half-filling density. For M = 5, however, one can observe polarized regions with strongly peaked density close to 1 and 0 (dark red and dark blue, respectively). The case M = 10 looks qualitatively similar to M = 5, supporting the observation that both cases correspond to the disentangling phase. We note that in these cases, the minimum entropy trajectories feature the blocking regions at the center of the chain, where the bipartition cut is located (as in the two-ancilla setup). The above analysis demonstrates the capability of the method to detect the measurement-induced transition in a correlated many-body system. Our results provide further indications that measurement-induced transitions are a universal feature across a variety of measurement techniques, through employing a measurement protocol different from those conventionally used in this context. § DISCUSSION AND CONCLUSIONS In this paper, we have developed an implementation of a realistic measurement apparatus for generalized measurements on quantum many-body systems, based on the coupling of the system sites to ancillary degrees of freedom that are stroboscopically projected. Our protocol can be implemented in experimental setups with, e.g., ultracold atoms, trapped ions or superconducting qubits, and can be readily adapted to generalized measurements of observables other than density by adjusting the measurement part of the Hamiltonian (<ref>). With this tool, we have studied the effect of projective measurements of the ancillary degrees of freedom on the dynamics of hard-core bosons in the “main” system. Our work complements previous numerical studies in the field of measurement-induced transitions by implementing the measurements in such a way that the effect on the system of interest is indirect, through the ancilla sites that are entangled with the system. In this sense, it is a more realistic description than the approaches that involve a direct measurement of observables of the system, or the approaches that mimic continuous monitoring through an ad hoc stochastic formalism. We have started with the setups where the system is coupled to a single or a double ancilla pair, and then explored the setup with all sites of the main chain coupled to ancillas. Analysis of the density and entanglement dynamics in the single- and double-ancilla setups (Sec. <ref>) reveals a “quantum-Zeno-valve effect” for sufficiently strong coupling M between the ancilla and a site of the chain. This effects leads to long-lived quantum trajectories with very low entanglement corresponding to the blocking of the dynamics by a formation of a nearly bound state. This phenomenon is a manifestation of the interplay between the measurement backaction and the accumulated feedback of the measured system exerted on detectors in the absence of their reinitialization. As discussed in Secs. <ref> and <ref>, the single- and double-ancilla setups already feature certain signatures of the many-ancilla entanglement transition. Specifically, we observe a crossover with increasing M between two distinct types of behavior. For relatively small M, represented by M=1, all the quantum trajectories correspond to fast thermalization, with entanglement entropy approaching its maximum value and weak fluctuations. On the other hand, for large M, we have observed a strong variance of entanglement entropy, with some trajectories exhibiting strong entanglement across the system and others having strongly suppressed entanglement in comparison to thermal, chaotic states. This feature persists even for strong coupling M = 50, implying that typical trajectories are not representative for the dynamics at large M. The crossover between these types of dynamics occurs around M=4. Applying only a single (or double) measurement is not sufficient to disentangle all trajectories; we always observe at least a few trajectories that are strongly entangled, even for a very large measurement strength. Therefore, we conclude that a genuine entanglement transition requires a finite density of ancilla pairs. We have confirmed the existence of the transition explicitly in a setup where every site of the chain is measured, Sec. <ref>. We have observed entangling behavior for weak measurement strength and disentangling behavior for strong measurement strength. Using matrix product states, we have studied the system sizes of up to L=24 sites in the main chain (ergo, 24 × 3 = 72 sites in the system as a whole). Analyzing the dependence of the entanglement entropy on time t and the system size L, we have found evidence of a transition for a measurement strength M ≈ 4. Remarkably, this is essentially the same value as was identified for the crossover in the setups with one or two ancillas. Our numerics provides indications of a volume-law entropy scaling in the entangling phase, which is in line with previous results on correlated chains, where different implementations of measurements were employed <cit.>. Note that such a behavior is also typical for the entanglement transition in random quantum circuits <cit.>. Our work demonstrates the capability of the developed approach to capture the physics of measurement-induced entanglement transition in a realistic setup. It sheds more light on the nature of the entanglement transition, by disclosing the tomographic signatures of the transition in terms of quantum trajectories. Moreover, it reveals a high degree of universality of the measurement-induced entanglement transition in correlated many-body systems with respect to implementation of measurements. To conclude, let us list possible future directions opened by this work. By considering an intermediate probability 0 < P < 1 of measuring the ancilla within a given measurement interval, one can enrich the phase diagram of a measurement-induced transition. Furthermore, a quantitative analysis of the measurement-induced transition and associated critical behavior in our framework remains to be performed. In addition, it would be interesting to explore how the resetting of ancillas affects the entanglement transition. Another possible avenue for future work is that in the noninteracting case U = 0, the volume-law behavior was argued to be unstable in the archetype measurement protocols. It is interesting to study whether this statement holds for our measurement setup, where the interaction between ancillas and the chain site is inevitably present even for U=0. Finally, the developed framework is expected to be valuable for applications in the context of quantum engineering and information processing, such as steering of quantum states. The present analysis may be adapted to many realistic interacting systems of interest, ranging from cold-atom or trapped-ion quantum simulators to superconducting qubit arrays. § ACKNOWLEDGMENTS We thank J. Behrends, C. Castelnovo, H. Perrin, I. Poboiko, P. Pöpperl, and A. Štrkalj for useful discussions. EVHD and IVG gratefully acknowledge collaboration with Oleg Yevtushenko at early stages of this work. The work was supported by the Deutsche Forschungsgemeinschaft (DFG): Project No. 277101999 – TRR 183 (Project C01) and Grants No. EG 96/13-1 and No. GO 1405/6-1, by the Helmholtz International Fellow Award, and by the Israel Binational Science Foundation – National Science Foundation through award DMR-2037654. The simulations in this work have been performed using the library, versions and <cit.>. We acknowledge support by the State of Baden-Württemberg through bwHPC. Data availability.— Numerical data and source codes are available from the corresponding author upon reasonable request. § NUMERICAL METHOD For our simulations, we apply the procedure similar to the one outlined in Ref. <cit.>. That is, we use the time-dependent variational principle (TDVP) as applied to matrix product states (MPS) <cit.> to exactly compute the unitary evolution. The difference between the procedure outlined in Ref. <cit.> and the method used in this work, is that projections of the ancilla site are instantaneous. This is effected through the application of a strong imaginary on-site potential on the first ancilla site: ℋ_proj = ± iA a_1^† a_1. The dynamics is then computed over the time interval [0,η], resulting in imaginary-time propagation. In the limit A,η→∞, this results in a projection onto either the |01⟩ or |10⟩ state for the ancilla, depending on the sign ± in Eq. (<ref>). As explained in the main text, this sign is determined stochastically, by applying the Born rule to the ancilla density n_a. During this process, the original Hamiltonian is “switched off.” In the case where there is more than one ancilla pair, each projection is performed consecutively, starting from the left side of the main chain (i = 1). We choose the numerical parameters A = 3 · 10^8 and η = 10^-6. Numerically, the procedure is subdivided into 10 smaller steps to improve stability. A larger value of η would provide a closer approximation of the projection (typically, we find a deviation of around 10^-4 in the ancilla density after projection); we choose a smaller value to reduce computational effort without significantly compromising accuracy. For implementing the model using MPS, we map the system of the main chain plus the ancilla pairs onto a next-nearest-neighbor Hamiltonian, by “folding” the ancillary sites into the main chain. One can avoid introducing next-next-nearest neighbour terms by alternating the ancilla and main sites in the resulting one-dimensional chain—this structure facilitates the calculation. Note, however, that, because of this choice, it is not straightforward to find, e.g., multipartite entanglement involving the ancilla pair as one of the partitions. In the case of many ancillas (Sec. <ref>) there is not sufficient “space” to fold both ancilla sites into the main chain, and we do use a next-next-nearest neighbor Hamiltonian. We observe that, for the cases of one and two ancillas studied in Sec. <ref>, there is no parameter range for which entanglement is strongly suppressed in all trajectories. Hence, there will be some trajectories poorly approximated in case the MPS is truncated. For that reason, we do not restrict the bond dimension of the MPS when dealing with one or two ancilla pairs; our approach then becomes equivalent to exact diagonalization and we are restricted to modest system sizes of ≈ 20 sites. Alternatively, one could post-select only those trajectories that have limited entanglement, keeping the maximum bond dimension, at the cost of losing information about the more entangled trajectories. This might be an interesting avenue for future work. For the case of ancillas coupled to every site (Sec. <ref>), the total number of sites in the whole system is up to 24 × 3 = 72. Obviously, an exact diagonalization is not possible any more. Our analysis of the entanglement transition proceeds similarly to Ref. <cit.>. We perform the calculations with the bond dimensions χ=128 and 256. For sufficiently strong coupling to ancillas, M ≥ 4, all quantum trajectories exhibit suppressed entanglement, so that the truncation of the bond dimension to the lower-entanglement subspace does not influence the result. On the other hand, for weaker couplings, M ≤ 3, the bond-dimension truncation starts limiting the growth of the entanglement at long times, as pointed out in the main text. § PROTOCOL FOR THE RECONSTRUCTION OF THE DENSITY IN THE MAIN CHAIN In the main text, Sec. <ref>, we have discussed the effect of the main chain on the ancilla dynamics. As a “proof of concept," let us provide a demonstration of how one could infer the main chain density n_i from the ancilla dynamics coupled to the site i. For the same parameters as used in Fig. <ref>, we consider 100 measurement cycles, i.e., continue the runs up to time t = 200. For each cycle, we infer the density on the chain site connected to the ancilla by fitting Eqs. (<ref>) and (<ref>) to the ancilla density. Here, we only consider those cases where a particle was measured in the ancilla, separately for each measurement step. For sufficiently coupling strength M, this procedure provides remarkably good agreement to the true density (as numerically obtained from the full many-body wave function). For M = 3 and M = 5, the two curves match very closely, while for weaker measurement strength M = 1, in which case the two-level approximation does not work as well, the agreement is qualitative. Thus, the employed ancilla-based setup is indeed capable of measuring the true density in the chain. Of course, in a realistic experimental setting, we do not have direct access to the ancilla dynamics n_a(t) while measurements are not performed, as such measurements would necessarily be at least partially destructive and affect the ancilla dynamics itself. However, Fig. <ref> does clearly demonstrate that the information about the main chain density n_L/2 is strongly correlated to the ancilla density. Thus, it is in principle possible to obtain this information from projective measurements. One possible approach is to restrict the measurement interval Δ T in such a way that it is smaller than half a period of the ancilla Rabi oscillations for the non-measured case , leading to the simple criterion JΔ T < π. Then, a one-to-one mapping of n_a to n_L/2 can be constructed by cross-correlating the results of projective measurements for many different choices of M over several trajectories. § DYNAMICS IN THE ABSENCE OF MEASUREMENTS In the main text, we have discussed the case where the main chain of the system is coupled to one or more ancilla pairs with a nearest-neighbour interaction M, and projective measurements are “stroboscopically” performed at regular intervals Δ T. For the sake of comparison, it is useful to consider the cases where there is no interaction, and where there is coupling to ancilla but no measurement is performed on the ancilla. We will consider both cases in this Appendix. §.§ Uncoupled case Let us first compare the dynamics of the measured system to that in the “uncoupled” case without any interaction between the main chain and the ancilla, M = 0. We show the result in the left panel of Fig. <ref>, starting from the domain-wall initial state. In this case, the density evolution is believed to feature integrable Kardar-Parisi-Zhang dynamics <cit.>. Interestingly, in the finite-size setup, we observe multi-scale oscillations of the density, apparently related to the interference effects resulting from multiple bounces of the excitations from the boundaries in this integrable model. The analytical description of this phenomenon is, however, beyond the scope of this work. Comparing to the case of repeated measurements as shown in the main text (Sec. <ref>), we see that a finite ancilla-chain coupling M in a single-ancilla setup destroys the peculiar features of Kardar-Parisi-Zhang domain-wall melting by breaking the integrability of the system. To provide a further benchmark, we have compared the integrable dynamics computed using our TDVP implementation at M = 0 (decoupled ancilla) to the dynamics obtained from the exact simulation of the main-chain dynamics only (i.e., no ancilla present) as obtained from a fully independent calculation using the library <cit.>, finding an excellent agreement (result not shown). §.§ No ancilla projections It is also useful to investigate the case where the system is quenched at t = 0, but no measurement is performed. Let us consider the case of a single ancilla with the domain-wall initial condition. The results for M = 1 and M = 5 are shown in the middle and right panels of Fig. <ref>, the corresponding entropy dynamics is shown in Fig. <ref>, and the dynamics of the ancilla is shown in Fig. <ref>. Comparing the case of only this single quench to the case of repeated measurements (Fig. <ref> of the main text), we see dramatic differences. Indeed, in the absence of measurements, the state remains close its initial form: the density distribution remains close to the domain-wall initial condition, the entropy stays well below the Page value, and the ancilla density n_a stays close to unity (well above 1/2). On the other hand, repeated measurements lead to thermalization that manifests itself in the entropy approaching the Page value, domain wall smearing, and the ancilla density relaxing towards 1/2 (cf. Fig. <ref>). This demonstrates that the physics studied in this paper is not just governed by the coupling to an ancilla but is rather induced by projective measurements of the ancilla degrees of freedom. § ADDITIONAL DATA TO THE SINGLA-ANCILLA SETUP WITH A DOMAIN-WALL INITIAL STATE In Sec. <ref>, we have studied the dynamics of the entanglement entropy and its fluctuations for the setup with one ancilla and a domain-wall initial state. In this Appendix, we provide additional numerical data for this setup. In Fig. <ref>, the five least entangled trajectories, as quantified by the integrated entropy Ξ [Eq. (<ref>)], are displayed. For the case of relatively weak measurement strength, M = 1, all curves are similar, with only minor fluctuations, and the least entangled trajectories are close to the average, as is also expected from the corresponding data in Fig. <ref>(c). For the resonant case M = 3, we see that all curves fall on top of each other, implying that the measurement results are identical for each of these trajectories. For a still stronger (and non-resonant) coupling, M = 5, all five trajectories are identical (and are characterized by a very small entropy) up to a time t ≈ 120, after which they start exhibiting branching off towards high entropy state. These three panels provide an additional manifestation of a distinct physics in the case of relatively weak (M=1), resonant (M=3), and strong off-resonant (M=5) couplings. We also show in Fig. <ref> the entropy dynamics for the case M = 50 in the domain-wall setup. The data complement those shown in Fig. <ref>c for the same setup and M= 1, 3, and 5. As pointed out in the main text, the dynamics here almost totally freezes for all trajectories since the Rabi oscillation amplitude approaches zero as M →∞ and a “nearly bound state” is formed. The domain wall is therefore robust up to very long times. This manifests itself in very low values of the entanglement entropy for all trajectories.

http://arxiv.org/abs/2303.06650v1

20230312124422

Signatures of winds and jets in the environment of supermassive black holes

astro-ph.HE

firstpage–lastpage Interpreting Hidden Semantics in the Intermediate Layers of 3D Point Cloud Classification Neural Network Cheng Wang* =========================================================================================================== The Event Horizon Telescope Collaboration (EHTC) has presented first - dynamic-range limited - images of the black hole shadows in M87 and Sgr A*. The next generation Event Horizon Telescope (ngEHT) will provide higher sensitivity and higher dynamic range images (and movies) of these two sources plus image at least a dozen others at ≤100 gravitational radii resolution. We here perform an exploratory study of the appearance of winds and jets in such future observations. To do this we use M87 and Sgr A* as reference systems: we do not aim to exactly reproduce them, but rather to determine how their observed images will depend on specific physical assumptions. Even in the case of similar or the same dynamics, the images depend significantly on global parameters such as the black hole mass and the mass accretion rate. Our results provide guidance in the interpretation of future high-resolution images, particularly if a wind or jet is detected. black hole physics – radiative transfer – methods: numerical – magnetohydrodynamics – accretion, accretion discs § INTRODUCTION Astrophysical black holes (BHs) are found with masses ranging from a few solar masses to supermassive BHs of masses greater than 10^6 M_. BHs initially were a basic outcome of Einstein's General Theory of Relativity (GR) defined by a one-way causal boundary in space time also known as the event horizon from which even light cannot escape, making the direct detection of BHs impossible. BHs can be detected only indirectly by their gravitational influence on the surrounding matter and/or space-time, by their accretion which results in often luminous emission from radio to γ-rays, or by gravitational wave emission. The Event Horizon Telescope (EHT) has presented first images of the gravitionally-lensed ring, and BH shadow, of the supermassive BH of M87 <cit.> in the heart of Virgo cluster, and of SgrA* at the center of the Galaxy <cit.>. Active galactic nuclei (AGN), accretion powered supermassive BHs in the nuclei of some galaxies, have intrigued astronomers for ages as being the most powerful sources in the sky. They can outshine even the entire stellar population of the galaxy which hosts them such as in quasars which are among the most luminous sources in the sky <cit.>. These are believed to be powered by high accretion rates on to the supermassive BH through a geometrically thin but optically thick disk <cit.>. On the other hand, AGN in the local universe, including M87 and SgrA*, dominantly show lower accretion rates onto the central BH <cit.>. Many AGN display signatures of outflows and large-scale jets which are observed across the spectrum. Jets appear as collimated structures launched at relativisitic velocities from very close to the BH. The EHT has recently imaged the innermost jets of 3C 279, CenA, and the blazar J1924-2914 <cit.> with 20 μ as resolution paving a way to observe and understand the physical processes involved in launching jets around supermassive BHs. Jets may either be powered by magnetic fields threading the event horizon, extracting the rotational energy of the BH, <cit.> or by the magneto-rotational acceleration of matter from the accretion flow <cit.>. The presence of strong magnetic fields close to the event horizon thus leads to synchrotron radiation which peaks between the radio and the far-infrared <cit.> and could originate from the accretion flow, jet base or disk wind. The first ever images of the BH shadows of M87 and SgrA* have paved the way for understanding the physical processes around supermassive BHs. While both sources have roughly the same flux and angular size in their gravitationally lensed ring, they are individually interesting as their masses, and potentially spins, are on the two extreme ends of the mass and spin spectrum of supermassive BHs. Although the BH mass of M87 (M_BH=6.5 × 10^9 M_⊙) is three orders of magnitude larger than that of SgrA* (M_BH=4.16 × 10^6 M_⊙), it is also at a distance approximately 3 orders of magnitude farther (D_M87=16.8 Mpc) than SgrA* (D_SgrA*=8.127 kpc) [see for e.g. <cit.>]. This gives rise to interesting observational similarities such as their lensed photon rings project similar angular sizes on earth (39 μarcsec for M87 and 52 μarcsec for SgrA*), which make them the most interesting targets for the EHT. Also, they both display similarities in the flux received through our telescopes. The total accretion power of M87 is estimated to be typically ∼ 10-100 times larger than SgrA* <cit.>. The higher power of M87 could mean that it is being fed directly from a larger reservoir of gas than the stellar winds which are expected to feed SgrA*. This may be the cause for another significant difference between the sources i.e. the prominent, powerful jet launched from M87, which can be observed at multiple wavelengths and across nearly eight orders of magnitude in size <cit.>. The extended one-sided M87 jet, and the relative faintness of the nuclear counter-jet, provides strong constraints on the orientation of the jet, and thus the spin axis of the black hole: an inclination of ∼ 20 to the line of sight was used for numerical simulations of M87 <cit.>. For SgrA* there are no strong constraints on the spin axis yet, although the model fits from the EHTC suggest an inclination angle of 30 to the line of sight <cit.>). These interesting characteristics of M87 and SgrA* make them the best candidates for testing different theoretical models. The Event Horizon Telescope, in science operation since 2017, combines a network of 8 to 10 globally distributed mm-wave telescopes using the technique of very large baseline interferometry (VLBI). The next generation EHT (ngEHT), currently in the design stage, aims to double this number of telescopes, double bandwidth (thus sensitivity), and implement dual 230/345 GHz observing. This will allow significant improvement in image quality, thus observation of finer structures and time-lapse movies of horizon scale structure in M87 and SgrA*, and the imaging of additional sources at better than 100 gravitational radius resolution. On the longer term the spatial resolution can be bettered either using higher frequency (e.g., 700 GHz) on the ground, or longer baselines via ground-satellite interferometry. Resolutions can be significantly improved with earth-Geostationary (GEO-VLBI) or earth-L2 (L2-VLBI) baselines. In addition to improving the image quality with extended baselines, ngEHT also plans to improve on image construction algorithms from the data to be able to obtain the the strongest possible constraints on the complex structures around the BH event horizon. Given the limits of the 2017 EHT intrinsic resolution and dynamic range, it has been difficult to address some questions such as the whether the emission signatures are from the accretion disk or disk wind or jet base, whether the ring features are in complete agreement with standard GR or if there is a possibility to constrain the spin and accretion rate accurately. The ngEHT will already be an improvent with better image fidelity (more u-v coverage) and sensitivity with a resolution gain from moving to 345 GHz as well as super-resolution techniques. However,we expect that the longer base-lines of future earth-space based VLBI (a possibility through space based ngEHT) will result in an enhancement in the resolution which will allow us to distinguish between various physical scenarios. Through this work we intend to demonstrate how to distinguish certain theoretical models with current and future VLBI probes, given the mass and flux constraints for two systems with contrasting mass and spin. While here our aim is not to reproduce specific images, we will take some guidance from the well-studied systems M87 and Sgr A*, with the aim of motivating the properties of a typical source. For this purpose, we will consider the black hole masses and the reported spectral energy distribution (SED) information at low frequencies of these two sources reported in the literature [for reference see<cit.>] to derive some reference systems that have similar emission properties. Based on these reference systems (M87 or SgrA*), we will study the possibility to identify signatures of winds and jets in future observations. The paper is organised as follows: In section [<ref>] we describe briefly the dynamical set up of our simulations while in section [<ref>] the post-processing and ray-tracing methods are discussed especially in the context of the ones chosen for this work. In section [<ref>], we discuss how the accretion rates are determined from the synthetic SED which fit to the observed data for our simulations for M87 and SgrA* like systems while in section [<ref>], we discuss the features obtained in the intensity maps and profiles with the parameters obtained from the spectral fitting. A comparision of the deduced results for a M87 and SgrA* like system is discussed in section [<ref>]. Finally in section [<ref>] we provide the primary insights obtained from this investigation. § METHODOLOGY Our approach follows a two-step procedure. First, we carry out resistive general relativistic magneto hydro dynamic (GR-MHD) simulations of the disk-jet evolution for a range of parameters. Secondly, we post-process these dynamical data applying ray-tracing procedures. Note that the dynamical data are scale-free by definition as these are pure magneto hydro dynamic (MHD) simulations. We thus need to re-scale our physical parameters, such as mass density, gas pressure, and magnetic field strength to astrophysical units. In fact, this approach allows us to apply our dynamical models to accreting BH systems with various physical conditions such as BHs of different mass, accretion rates and magnetic flux. In turn, we may derive these parameters of the BHs by comparing our results to observed systems. §.§ Dynamical Modeling The GR-MHD models we use here are the same as in <cit.>. Here, we summarize their main features and the numerical setup of the simulations for convenience. We have applied the resistive GR-MHD code rHARM3D <cit.> that is based on the well-known HARM code <cit.>. The application of a physical resistivity is essential for such simulations as it allows both for a long-term mass loading of the disk wind and jet, and for a smooth accretion process. For such simulations, it is believed that physical resistivity is essential as it allows both for a long-term mass loading of the disc wind and jet, and for a smooth accretion process. This is a consequence of the involved magnetic diffusivity. In addition, this allows to treat reconnection in a way that is physically well defined. The GR-MHD simulations considered here are axisymmetric on a spherical grid, implying that the vector components of all 3 dimensions are considered, however, any derivatives in the ϕ direction are neglected partly due to the exceptionally high CPU demand in case of resistive MHD. However, we believe that our approach is fully sufficient in order to investigate the radiative signatures of the black hole-disc-wind-jet system. Of course, any 3D features like instabilities in the disc or in the jets cannot be treated, nor any orbiting substructures can be found in these components. For our qualitative investigation, these assumptions are sufficient. It is worth summarizing the differences of our dynamical model setup in comparison to other attempts to model intensity maps and spectra from GR-MHD simulations such as e.g. published by <cit.>. We work with resistive GR-MHD, which allows for smooth disc accretion and launching of a disc wind. Our disc is thin, in Keplerian rotation, and is threaded by a large-scale magnetic flux. This allows to drive strong disc winds (that, on larger spatial scales, are supposed to evolve into a jet). In these cases, it is the magnetized disc wind that is more efficient in angular momentum removal due to its large lever arm. A common feature of all GR-MHD simulations is the application of a so-called floor density set as a lower density threshold in order to keep the MHD simulation going. This numerical necessity strongly affects the resulting dynamics of the axial Blandford-Znajek (BZ) jet – the mass load carried away with the jet and the jet velocity <cit.>. On the other hand, the disk wind dynamics is governed by the mass loading from the disk, which is self-consistently derived from MHD principles. Our simulations follow the setup of <cit.>. The output of our GR-MHD simulations comes in normalized units where G=c=M=1. In particular, all length scales used in these simulations are in gravitational radii R_ g, while velocities are normalized to the speed of light. All physical variables need to be properly re-scaled for post-processing of radiation. A further discussion of our model approach including the parameter space involved, as well as maps of the essential dynamical variables, can be found in <cit.>. In brief, we investigate (i) a reference simulation labeled as 20EF which applies similar simulation parameters as <cit.>, (ii) a simulation 21EF with a (ten times) higher floor density, thus a higher mass loading of the BZ jet, (iii) a simulation 22EF with a Kerr parameter a=0, (iv) a simulation 23EF with an even higher floor density (another factor ten) resulting in the absence of BZ jet, (v) a simulation 24EF with a lower plasma-β (factor ten), resulting in a stronger magnetic flux, and (vi) a simulation 26EF with an even stronger magnetic flux (another factor ten). The differences in the model setups are mainly in terms of the magnetic field strength, BH rotation and floor density. However, these parameter have an immediate impact on the jet dynamics, namely the jet speed (Doppler boosting), and mass flux. In combination with the magnetic field, the synchrotron emission is also affected. Note, however, that the latter also depends on the scaling of the simulation variables in astrophysical units. The accretion rate given in the GR-MHD simulations is in code units, Ṁ≃ 0.1, which can be re-scaled to astrophysical units. As an example typical AGN accretion rates of Ṁ≡ρ_0 R_g^2 cṀ≃ 10^-3 M_⊙yr^-1 constrain the (maximum) disk density to ρ_0 = 4× 10^-15g/cm^3[ Ṁ/10^-3 M_⊙/yr] [ M_BH/ 5× 10^9 M_⊙]^-2[ Ṁ/ 0.1 ]^-1 <cit.>. The re-scaled astrophysical density is ρ = ρ_0 ρ̅, where ρ̅ in code units follows from our simulations. For illustration, for the example of M87 one can use the values provided by the EHT collaboration <cit.> with an accretion rate of Ṁ∼ 2.7× 10^-3 M_ yr^-1 and a BH mass of M = 6.2×10^9 M_. §.§ Parameter fitting and ray-tracing GR-MHD codes like the one we have used in this work typically apply normalized units where G = c = M = 1. These simulations need to be post processed with physical parameters such as black hole mass, accretion rates etc., to obtain a physical realization of an AGN. Here we post process our simulations for the BH masses of M87 and SgrA* and vary the accretion rate (fixing other parameters) to obtain synthetic SED comparable to the observed data in the radio bands for these sources. The post processing of these GR-MHD simulations are done using ray tracing codes. The image of a BH is characterized by a "shadow" region towards which no photons arrive due to the curvature of space-time around the BH; the unlensed "photon ring" (3 R_g), corresponding to bound photon orbits around the BH; and as a result of lensing, the photons from the inner part of the disk appear somewhat further outside that region leading to lensed ring marked by inner (5.2 R_g) and outer (6.2 R_g) boundaries <cit.>. These features are a result of strong GR effects around the vicinity of the BH. The use of a ray-tracing code is to post process the simulated dynamical model to calculate the observed intensity on locations (pixels) of an observer’s camera for a given model of emission and absorption. Here we use the publicly available general relativistic radiative transport code GRTRANS <cit.>. Within GRTRANS, the Boyer–Lindquist coordinates of the photon trajectories are calculated from the observer towards the BH (i.e. the rays are traced) for each pixel in the camera. The observed polarization basis is then parallel-transported into the fluid frame. The local emission and absorption properties at each point are calculated. The radiative transfer equations are solved along those rays. We refer to the original paper by <cit.> for details on the working of the code. The jet and wind dominated regions are highly magnetized and thus it can be assumed that synchrotron emission is the most dominant emission mechanism in these regions <cit.>. In this work we consider synchrotron emission from a population of thermal electrons as the primary emission process. This emission, often referred to as thermal synchrotron, depends on the thermal state of the electrons where the electron temperature should ideally be determined from the plasma physics. However, currently setting up simulations of a fully functional plasma dynamics in a GR set up is computationally expensive. In the absence of such a physically complete model, an alternate prescription is applied which replicates the plasma temperature model. In this scenario, the gas temperature at each point in the grid is determined from the internal energy at those grid points assuming an ideal gas prescription. The electron temperature is deduced from the gas temperature assuming that the plasma consists only of ionised hydrogen. As prescribed by <cit.>, the electron temperature is approximated from the following relation as T_e = T_ gas/(1+R), R = T_ p/T_ e = R_ highb^2/1+b^2+R_ low1/1+b^2, with b=β/β_ crit, β=P_ gas/P_ mag and P_ mag=B^2/2. The value of β_ crit is assumed to be unity, and R_ high and R_ low are the temperature ratios that describe the electron-to-proton coupling in the weakly magnetized (e.g. discs) and strongly magnetized regions (e.g. jets), respectively. The emissivity for a given distribution function of electrons is given as j=∫_1^∞dγ N(γ)η, where η=√(3)e^2/8π cν_Bsin θ_BH(ν,θ_B) is the vacuum emissivity <cit.> with e- the electron charge, c- the speed of light, ν_B=eB/2π mc, H=F(ν/ν_c), ν- the emitted frequency, γ- the electron Lorentz factor, ν_c=(3/2)ν_B sin(θ_B)γ^2 and F(x)=x∫_x^∞dyK_5/3(y) is an integral of the modified Bessel function. For a thermal distribution the distribution function in eqn.[<ref>] is N(γ)=nγ^2βexp(-γ/θ_e)/θ_e K_2(1/θ_e), where n is the number density of electrons and θ_e=kT_ e/mc^2 is the dimensionless electron temperature. As shown by <cit.>, the post-processing parameters which significantly affect the synthetic SED and emission maps are the BH mass and the accretion rate. The accretion rate and black hole mass scale the density, pressure and the true magnetic field from the simulations. The temperature of the gas is obtained from the density and pressure assuming an ideal gas equation as already mentioned. The temperature and the magnetic field are the important physical parameters which affect the synchrotrom emission through the distribution function (eqns. [<ref>] and [<ref>]). In this work, we have used the high resolution spectral data from <cit.> for M87 and that from <cit.> for SgrA*. Modeled spectra from the simulations are generated by fixing all the other parameters (except for the black hole mass and accretion rate) to constant values for the two systems for simplicity. As this is a qualitative study about the emission features around supermassive BHs, the small changes in the SED due to the inclination angle or the choice of R_high and R_low in eqn.[<ref>] are almost insignificant compared to the variation introduced due to the underlying dynamics from the simulations. Once the best fit SED are obtained for each of the simulations, the synthetic images are generated at 230 GHz using the best fit value of the Eddington ratio or the accretion rate for each given BH mass. We use these synthetic images to obtain the intensity profile for several cross-sections of the image. These synthetic images are further convolved with a Gaussian beam-width to account for the telescope resolution of the EHT and the future ngEHT to obtain secondary synthetic images which are expected to be similar to that we observe with the telescopes. At 230 GHz, an intrinsic (FWHM) resolution of ∼20 μ as is achieved for the longest baseline (D_1∼ 11,000 Km). Thus using these values, we can obtain approximate values of the resolution for earth space baselines. A Geo-VLBI baseline (D_2∼47,000Km) and a L2-VLBI baseline (D_2∼1.5 million Km) would give us approximate resolutions of θ_2∼ 5 μ as and θ_2∼0.16 μ as respectively. We assume these resolutions in the sections below (see also <cit.>). § SPECTRAL PROPERTIES The primary aim of this work is to investigate the origin of emission signatures from disk winds, jets or the accretion flow for a set of GR-MHD simulations when subject to the known extreme ends of the mass and spin spectrum for supermassive BHs in the local universe. The origin of emission signatures from winds, jets and disk from the different simulations can be distinguished when we can obtain a similar spectral fit from each of them. We thus post processed our simulations with the masses of M87 and SgrA* and varied the accretion rates (by varying the Eddington ratio ṁ=Ṁ_ acc/Ṁ_ Edd) for each of our simulations to generate synthetic SED which fit best to the high resolution multi-wavelength data <cit.> of M87 and SgrA*. Generally in literature (for reference, see <cit.>, etc.) the SED for the sources of interest are obtained considering the synchrotron emission and its Compton scattering using the GRMONTY code <cit.>. However, in this work, the spectrum obtained with GRTRANS <cit.> using synchrotron emission is sufficient as we are mainly interested in a qualitatively fit of the synthetic SED to the high resolution data in the radio band where we know that synchrotron emission is the most dominant emission process. M87 has a mass (6.5 × 10^9 M_⊙) which is 3 orders of magnitude higher than that of SgrA* (4.12 × 10^6 M_⊙) but is also at a distance (16.9 Mpc) 3 orders of magnitude farther than SgrA* (8.127 kpc) which is in the center of our Galaxy <cit.>. Table [<ref>] shows that the Eddington ratios (7th and 8th columns) required to match the SED for each of the simulations are higher for M87 compared to SgrA*. This result is similar to previous literature conclusions and implies that M87, the CD galaxy of the Virgo cluster, is fed by a more massive reservoir of gas compared to the trickle from disk winds as speculated for SgrA*. The higher accretion rates in M87 lead to a higher luminosity but since it is also at a distance 3 orders of magnitude farther than SgrA*, both result in similar fluxes. As can be seen from the spectral fits in Fig.[<ref>], 26EF (the simulation with spin parameter 1 and the lowest value of β parameter) fits best to the M87 high resolution data at radio frequencies, while 22EF (the simulation with spin 0) fits best to the high resolution data for SgrA*. It can be noted here that the accretion rate that fits the M87 data for 26EF is 2 orders of magnitude smaller than what has been inferred by the EHTC <cit.>, which could be due to the an intrinsic higher accretion rates (Ṁ in Eq. [<ref>]) in the dynamical simulations used in this work. It is expected that the SED generated with a power-law, hybrid or Kappa distribution would be the result of accretion rates different from the values deduced in this work. However, this work focuses on the SED fits resulting from a thermal synchrotron emission to deduce the underlying physical parameters from expected observations. § INTENSITY MAPS & PROFILES Once the accretion rates are determined by fitting the modeled spectra to the observed data, the intensity maps for the simulations are generated. The existence of jets and winds if at all present are expected to be detected with a edge-on view as the direction of these flows in general are perpendicular to the accretion flow. However, in reality not all AGN have an orientation with edge-on views. As an example both M87 and SgrA* have very low inclination angles (for M87 this is inferred from observations of the large scale jet while for SgrA*, this is deduced through model fitting <cit.>). Hence we first post process our simulations assuming an edge-on view to identify the presence of disk wind and jet signatures and then look for these signatures for lower inclination angles. Fig.[<ref>] shows the edge-on views of the post processed images of our simulation for M87-"like" (upper 6 rows of each figure) and SgrA*-"like" (lower 6 rows of each figure) systems at 86 GHz, 230 GHz, 345 GHz and 700GHz respectively. Since low frequency photons cannot escape the inner regions due to synchrotron self-absorption, only the outer regions are visibile at lower frequencies and hence the likelihood of the visibility of any signatures of winds and jets at lower frequencies. At higher frequencies such as at 345 GHz, the emission from regions in the vicinity of the black holes event horizon are visible from which the photons with the highest energies are emitted. As can be seen in Fig. [<ref>] the emission signatures of jet and disk wind are more prominent at 86 GHz (see for example the emission maps for 21EF and 23EF for SgrA* like system) while those same features are not prominent at 345 GHz where the emission is expected from the inner most regions. However, 230 GHz is an intermediate frequency where both these effects are visible. These images in general are brighter on the left including the ones with jet and disk features due to Doppler beaming. At present it is difficult to obtain good resolution images at 86 GHz even with an earth-size telescope. As can be observed from Fig.[<ref>], 230 GHz is thus the ideal frequency to trace the jet and wind signatures back to horizon scales with the resolutions available in the present. As we move towards higher frequencies, it is difficult to distinguish the traces of any jet or outflow signature distinctly and this is independent of any telescope resolution. Although at higher frequencies, we may able to probe regions in the proximity of the BH event horizon, it may still be difficult to comment on the origin of jets and outflows at these high energies. This may however change if the accretion rate is increased in which case there will be an overall increase in flux throughout the spectrum (the images for low inclination angle at 345 and 700 GHz are in the Appendix A). Here we, thus, concentrate on investigating emission features of our interest at 230 GHz. The simulations are thus further processed at 230 GHz for a face-on view and low inclination angle (17) as shown in fig.[<ref>]. Although from a glance distinguishing emission features from winds and jets is not obvious for low inclination angles or face-on views, however observing the emission features from different simulations especially from their emission profiles hint at the existence of disk winds or jets as shown in fig.[<ref>]. As an example one can deduce from the figure that the emission profile for 21EF is broader than 20EF for SgrA* like system for both face-on and a small inclination angle of 17 for a similar Eddington ratio. This comparison although not quite evident from the edge-on view at 230 GHz but the edge-on view at 86 GHz clearly shows a jet signature for 21EF. Similarly a large scale wind signature is visible for 23EF at 86 GHz in the edge on view for SgrA*-like system for an accretion rate which is slightly lower than those for 20EF and 21EF. for the same system at lower inclination angles and face-on views, the emission profile is broader and extends to much larger scales. For M87-like system too, with an accretion rate lower by an order than 20EF and 21EF, the emission profile is similar or slightly broader for 23EF at 230 GHz. Such a broader emission profiles for a similar physical set-up could indicate the presence of a disk wind which is due to velocity dispersion within the wind or jet. In addition to tracing jets and wind signatures deduced from the width of the emission profiles, there are several interesting observations from these emission maps. As an example, the upper left panel of Fig.[<ref>] shows the face-on views of the emission where the unlensed photon ring and the outer and inner boundaries (corresponding to the prograde and retrograde spins) of the lensed photon ring are marked by green contours of radii 3 R_g, 5.2 R_g and 6.2 R_g, respectively. The lensed photon ring visible in face-on views (0) coincides with the 5.2 R_g radius for almost all the cases (as expected for prograde spins). The presence of a disk wind or jet base smear out the photon ring features as in 21EF for SgrA*-like system. As expected for face-on views, the flux profiles along different slices of the images coincide as can be seen in the lower left panel of Fig.[<ref>], while for small inclination angle such as 17^ the profiles separate and one of the peaks is higher than the other due to the additional effect of Doppler beaming. We also observe that for an inclination angle of 17, the emission peaks shifts from the expected inner or outer boundries of the lensed photon ring in all models except for 22EF which is an accreting Schwarzschild BH. The asymmetry is an intrinsic property of BHs with non-zero spins. Simulations 20, 21 and 23 have very similar properties except for their floor density chosen (which are not expected to influence the emission signatures due to the assumed σ_cut). Simulations 24EF and 26EF are however set-up considering a larger mean magnetic field strength. Thus in general the accretion rates deduced for these simulation is much lower than those for the ones mentioned above as synchrotron emissivity is also dependent on the magnetic field strength. It is also observed that the accretion rates obtained from fitting the simulated spectrum to the data for M87 and SgrA* for 24EF and 26EF are of the same order for both cases but the emission region for 26EF is quite compact compared to 24EF. The magnetic field strength in case of 26EF is an order of magnitude higher, thus the emission from a smaller region is enough to obtain a similar amount of flux strength as obtained that from 24EF. All our simulations are set-up for Kerr BHs except for 22EF which is a spin 0 Schwarzchild BH. All other parameters for this simulation are same as those set up for simulations 20, 21 and 23. Being a Schwarzschild BH, the ISCO radius is much larger implying that for similar accretion rates, the temperature of the in-falling material will be lesser than its Kerr counterpart. Thus to compensate for this in order to obtain similar fluxes, a higher accretion rate is required for both M87-like and SgrA* like systems. Simulations 21EF and 24EF display very similar emission features at all frequencies, both in the edge-on views and low inclination angle views for both M87 and SgrA* like systems. 24EF although has a lower accretion rate compared to 21EF by more than an order of magnitude in both the sources, but 24EF also has an intrinsic magnetic field strength which is an order of magnitude higher. It seems that these two effects almost balance each other. The images obtained at lower inclination angles are further processed to account for the resolution of the current EHT(20 μ as) and for the future earth-space based VLBI resolutions (5 μ as for Geo-VLBI and 0.16 μ as for L1-VLBI) as shown in Fig.[<ref>], to predict if the origin of emission will be resolved today or in future for a better understanding of the physical processes. Given the current EHT resolution, 20EF, 21EF, 23EF and 24EF appear qualitatively indistinguishable (Fig.[<ref>]). 22EF shows large scale wings showing the signature from the large scale outflow while a much more compact emission profile is seen in 26EF for both M87-like and SgrA*-like systems. With the current EHT resolution, the shadow region is not well resolved for a M87 like system for 23EF, 24EF and especially 26EF whose emission region is quite compact. For the expected Geo-VLBI and L1-VLBI like resolutions, these similarities in the emission profiles seem to break down and and the intrinsic features start to appear (Fig.[<ref>]). We would like to emphasize here that in this work we have considered only a thermal distribution for the electron distribution function for a simplistic approach. In reality we expect the jet emissions to be non-thermal in nature due to the presence of non-thermal processes such as magnetic field driven turbulence, magnetic reconnection, etc. We would however like to note that for simulations 20EF, 21EF, 22EF and 23EF this might be a good approximation which have a higher values of plasma β (β_0=P_gas/P_mag=10) However, simulations 24EF and 26EF have β=1 and β=0.1, respectively, implying that the magnetic pressure is equal to or ten times greater than the gas pressure. In these it might not be sufficient to just consider a thermal distribution function for the electrons. For a true understanding of the origin of disk and wind signatures of a realistic system, it might not be sufficient to just consider the emissions originating from a population of thermal electrons. However, considering other distribution function is beyond the scope of this investigation. § COMPARISON STUDY The emission region for a M87-like system in general is more compact compared to SgrA* like systems in R_g scales. As can be seen in fig.[<ref>], 22EF shows a large scale outflow both for M87 and SgrA like systems visible prominently at 86 GHz. Similarly, 23EF shows a disk wind resulting in a large scale outflow visible distinctly at 86 GHz for both systems. 21EF displays a strong disk wind signature and a trace of the BZ jet along the upper half of the equatorial plane at 86 GHz (left panel of Fig.[<ref>]) in an edge-on view for SgrA* like system. The trace of the BZ jet is also visible at 230 GHz in the emission profile with the face-on view (0) (bottom left panel of Fig.[<ref>]) where a non-zero flux is visible in the expected shadow region. 20EF and 21EF both have similar accretion rates but the emission profiles at 230 GHz display a broader emission profile of 21EF for SgrA like system and a greater Doppler beaming effect for M87 which leads to the conclusion that such a feature can arise if there is emission from region with additional velocities such as disk wind or outflows. Similarly 23EF has a lower accretion rate than 20EF but broader emission profiles at 86 and 230 GHz for SgrA like systems and a comparable width in the emission profile for M87 at 86 GHz and 230 GHz. This is again possible if the emission has a contribution from the disk wind. At higher frequencies such as 700 GHz, however, we see contrary effects in the case of 23EF both for M87 and Sgr A like systems where the emission feature is peaked near the photon ring and is dimmer than 20EF or 21EF which is possible when the emission is from the accreting disk because at higher frequencies it is the energetics of the accreting disk which becomes relevant. § DISCUSSION AND CONCLUSIONS Most AGN are found to exist in nature with jets and outflows which in some cases are detected easily. It is thus important to understand the physical processes which launch different types of jets or outflows for which it is also important to trace their origin through observations. Through this work we have tried to demonstrate how a set of accreting systems with different physical conditions, results in varied observational features for systems with similar BH mass and spectral behaviours as M87 and SgrA*. BH shadows for both SgrA* and M87 have already been observed with the EHT <cit.>. In this work our intention is not to make a quantitative comparison with the observations. However, this work intends to motivate how the same set of simulations, with similar spectral behaviour under different physical conditions, result in generating completely different observational signatures. In addition this work also intends to describe qualitative methods to extracting jet and wind signatures from observations. In our study the BH masses and SED are the primary constraining parameters. The accretion rates are varied in each simulation until the modeled SED is similar to the observed data. This results in generating intrinsic features in the emission maps, thus enabling to compare the different physical processes which lead to such emission signatures. The emission maps are further processed to make predictions for current and future VLBI observations. In general one of the criteria for accepting theoretical models is in comparing the jet power <cit.> to the real values; however this investigation is more qualitative in its approach of demonstrating observational signatures of winds and jets, rather than testing alternate models for M87 and SgrA*. Thus in this work we do not calculate the jet power from these models as it is not relevant for the current investigation. The emission maps here are shown in linear scale as generally done in the literature with the normalization being done with the maximum intensity in each frame. However we would like to mention that the mean value of intensity per pixel (which is quite low10^-7 Jy) is similar for all the frequencies. A normalization with the mean value generates similar emission maps displaying a thick emission ring at all frequencies but the fine structures arising due to the presence of winds and jets are not visible at all frequencies. However, normalizing the images with the mean saturates the images and is not practical because the value is much less than the flux sensitivity of any telescope. On this note we would like to emphasize that even though we have taken into account the smudging of images due to the limited beam-width introduced by the resolution of telescopes, in reality the images obtained are further limited by thermal noise, limited u-v coverage and calibration errors. We would like to emphasize here that while post-processing, in addition to the accretion rate, it is also important to consider a proper electron distribution function depending on the underlying physical parameters of the GR-MHD which may affect the resulting spectrum significantly. Although considering a thermal electron distribution function may be sufficient for a weaker magnetic field limit but may not hold true in the strong field limit. In the strong field limit, we expect non-thermal processes to dominate arising due to effects such as magnetic re-connection and also holds true for the emission from jets where the magnetic fields dominate. In such cases considering a pure thermal electron distribution function will be an underestimation. The outflow features are easily distinguishable at lower frequencies but tracing them to their origin from regions close to the BH would mean observing them with the highest resolutions which is possible only at higher frequencies and using telescopes with larger baselines. Currently highest resolution observations are possible with the technique of radio interferometery with different VLBI arrays such as the EHT which has an Earth sized baselines. Through this investigation we deduce that with fluxes like that we observe for M87 or SgrA*, observing the signatures of jets or outflows closest to horizon scale will be possible at 230 GHz. It might be possible to observe some of the features even at 345 GHz but currently not all EHT telescopes are designed to function at 345 GHz, which thus will introduce additional noise during image reconstruction. Although making strong deductions about different physical scenarios will not be possible with the current EHT resolution, it may however be possible with future Earth-Space baseline probes with ngEHT. Apart from improving telescope resolution to obtain finer details of emissions from horizon scale around SMBHs with longer baselines and tuning the telescopes to observe at higher frequencies, it is also important to improve on the image reduction algorithms which generates the observed images with better resolution. In addition it is also important to consider the telescope sensitivity to obtain the threshold to distinguish the finer structures. As a final note we would like to mention that through this work we primarily intended to demonstrate whether with current and future VLBI arrays it would be possible to detect signatures of disk winds or jets at horizon scales for two different physical systems. In the process we thus used a set of disk-wind-jet simulated models for AGNs to show how each model would differ from the other observationally. The SED was the primary constraint that was used and only the accretion rate was varied to fit the modeled SED to the observed spectrum. A more quantitative complete approach would include considering the variation in many other physical parameters (such as a considering different temperature models or distribution function for electrons etc.) to obtain precise fit to the SEDs. However for this investigation we think a qualitative approach is good enough to demonstrate the tools to extract certain observational features from simulated emission maps and making predictions if such an analysis would be possible for the current and future VLBI facilities. § ACKNOWLEDGEMENTS We acknowledge funding from ANID Chile via Nucleo Milenio TITANs (NCN19-058), Fondecyt 1221421, and Basal projects AFB-170002 and FB210003. All GR-MHD simulations were performed on the VERA clusters of the Max Planck Institute for Astronomy. C.F. thanks Qian Qian and Christos Vourellis for developing the resistive version rHARM-3D of the original GR-MHD code HARM-3D, kindly provided by Scott Noble. § DATA AVAILABILITY Data available on request. The data underlying this article will be shared on reasonable request to the corresponding author. mnras § EMISSION MAPS AND PROFILES FOR 345 GHZ AND 700 GHZ In Fig.[<ref>], we show the modeled emission maps and intensity profiles for M87-like and SgrA*-like systems from our simulations for 345 GHz and 700 GHz. The figure is discussed in Section [<ref>].

http://arxiv.org/abs/2303.07105v1

20230313134255

Measuring Multi-Source Redundancy in Factor Graphs

cs.IT

B-.05emi-.025em b-.08em T-.1667em.7exE-.125emX thmTheorem corollaryCorollary lemmaLemma formulaFormula propositionProposition definitionDefinition computationComputation *remarkRemark shapes.geometric, arrows,decorations.pathreplacing pose = [circle, text centered, draw=black, fill=orange!30, minimum size = 1.25cm] spur = [circle, text centered, draw=black, fill=violet!30, minimum size = 1.25cm] factor = [rectangle, text centered, draw=black, fill=blue!30] rendezFactor = [rectangle, text centered, draw=black, fill=violet!30] rendezFactorEliminated = [rectangle, text centered, draw=black!40, text = black!40, fill=violet!10] factorNew = [rectangle, text centered, draw=black, fill=green!30] factorNewEliminated = [rectangle, text centered, draw=black!40, text=black!40, fill=green!10] poseEliminated = [circle, text centered, draw=black!40, text=black!40, fill=orange!10] spurEliminated = [circle, text centered, draw=black!40, text=black!40, fill=violet!10, minimum size = 1.25cm] factorEliminated = [rectangle, text centered, draw=black!40, text=black!40, fill=blue!10] intra = = [thick,->,>=stealth] inter = = [dashed,-, very thick,draw=blue] factorLine = = [thick,-] factorLineEliminated = = [thick,-,draw=black!40] factorLineNew == [thick, green, -] BNarrow = = [dashed, blue, ->] interrobotComms = [dashed,->,draw=violet,thick] warning [colback=yellow!10!white,colframe=orange!70!yellow,title =Warning] assumption[section] assumption[2][] assumption #1 frametitle= [baseline=(current bounding box.east),outer sep=0pt] [anchor=east,rectangle,fill=blue!20] Assumption ; frametitle= [baseline=(current bounding box.east),outer sep=0pt] [anchor=east,rectangle,fill=blue!20] Assumption : #1; innertopmargin=10pt,linecolor=blue!20, linewidth=2pt,topline=true, frametitleaboveskip=- [] metric[section] metric[2][] metric #1 frametitle= [baseline=(current bounding box.east),outer sep=0pt] [anchor=east,rectangle,fill=violet!20] Metric ; frametitle= [baseline=(current bounding box.east),outer sep=0pt] [anchor=east,rectangle,fill=blue!20] Metric : #1; innertopmargin=10pt,linecolor=blue!20, linewidth=2pt,topline=true, frametitleaboveskip=- [] Measuring Multi-Source Redundancy in Factor Graphs Jesse Milzman, Andre Harrison, Carlos Nieto-Granda, John Rogers DEVCOM Army Research Laboratory Adelpi, MD USA {jesse.m.milzman.civ, andre.v.harrison2.civ, carlos.p.nieto2.civ, john.g.rogers59.civ}@army.mil March 30, 2023 =========================================================================================================================================================================================================================================================================== Factor graphs are a ubiquitous tool for multi-source inference in robotics and multi-sensor networks. They allow for heterogeneous measurements from many sources to be concurrently represented as factors in the state posterior distribution, so that inference can be conducted via sparse graphical methods. Adding measurements from many sources can supply robustness to state estimation, as seen in distributed pose graph optimization. However, adding excessive measurements to a factor graph can also quickly degrade their performance as more cycles are added to the graph. In both situations, the relevant quality is the redundancy of information. Drawing on recent work in information theory on partial information decomposition (PID), we articulate two potential definitions of redundancy in factor graphs, both within a common axiomatic framework for redundancy in factor graphs. This is the first application of PID to factor graphs, and only one of a few presenting quantitative measures of redundancy for them. factor graphs, redundancy, SLAM, partial information decomposition, information theory § INTRODUCTION Autonomous agents operating within an unknown environment must often simultaneously solve several related yet distinct optimization problems. Concurrently, they model their surrounding environment, estimate their own pose trajectory, the poses of other agents with respect to themselves and the environment, and classify, localize, and track objects within that environment. These optimization tasks are often modeled in robotics research as a factor graph consisting of variables which represent the unknown quantities and factors that define functions that act on a subset of the variables. Typically factors operate only on a limited set of variables, corresponding to measurements on those variables. Moreover, when several of these agents operate together information is shared amongst them to generate faster and more accurate estimates of the given target variable(s) that they need to estimate. In some of the most austere environments these autonomous systems will form a distributed mobile ad-hoc network (MANET) as there may be no supporting communication infrastructure or equipment. A recurrent research question in these scenarios is how to maintain or even improve estimation accuracy (and speed) using the least amount of information, in order to both minimize communication between agents as well as reduce the number of variables incorporated into the optimization problem. This can be done by estimating the individual or sets of information sources that provide the most unique or synergistic information for estimating the target variable while simultaneously minimizing information from redundant sources of information <cit.>. While minimizing the communication of redundant information can also reduce computational complexity, it can make the network brittle as the corruption or invalidation of any information source will reduce the estimation accuracy of the target variable(s). Incorporating a certain amount of redundancy for an information source or set of sources may lead to a more robust distributed estimation network. Redundancy also plays an important role in the implementation of resilient estimation. Alternative communication and information-gathering pathways can facilitate adaptation to unexpected failure modes <cit.>. Cummingham et al. proposed one the first distributed implementations of Smoothing and Mapping (DDF-SAM) resulting in systems with increased resilience to node failures and network topology changes <cit.>. Khedekar et al. <cit.> proposed a multi-modal approach to enable resilient SLAM through the use of redundant sets of heterogeneous sensors. To summarize, in many autonomy scenarios, adding heterogeneous sensors or additional agents can be used to increase the overall robustness and resilience of factor graph-based SLAM estimates. However, precise measures of the amount of redundancy from difference sources of information present within a factor graph hasn't received as much attention. Wang et al. proposed the use of maximal information content (MIC) to estimate the similarity between different features within a factor graph as part of an unsupervised feature selection approach (FGUFS) <cit.>. The partitioning of information sources at the level of granularity — to fully capture the uniqueness, synergy, and redundancy among information sources — has not historically been possible with standard information theoretic concepts of mutual information or conditional mutual information  <cit.>. Recent developments in information theory have seen the advancement of the partial information decomposition (PID) of multivariate information to address such questions. At its core, PID untangles these notions by hierarchically decomposing the information content among many variables with a redundancy function. However, this approach, while it has been applied to problems in feature selection and information fusion  <cit.> has not been applied to factor graph analysis. In this paper, we present the initial development of a framework for estimating the redundancy within a factor, building upon PID. We present two measures of redundancy: the direction application of PID as an information-theoretic measure, and a statistical distance-based approach with an analogous form. The contributions made in this paper are the following: * Building upon the work in PID, we outline a formal axiomatic framework for defining multi-source redundancy in factor graphs. * We define two potential redundancies within this framework: an information-theoretic redundancy and a statistical distance-based redundancy. * We investigate these definitions analytically for supplemental factors in linear factor graphs, rewriting them as closed-form integrals on the base graph. * We apply these to simulations of a 2D landmark SLAM task, in order to compute the redundancy of the information contributed by two landmarks. § PRELIMINARIES In this section, we briefly review the key definitions from the PID framework, as well as those for linear factor graphs. For factor graphs in particular, we will largely use our own notation in order to simplify the later presentation of our PID-based redundancy framework in Sec. <ref>. This work is primarily focused on linear factor graphs. §.§ Partial Information Decomposition The partial information decomposition was originally put forward in <cit.> as a one-sided multivariate extension of mutual information, and as a resolution to the signed nature of interaction information. In its most common form, PID decomposes the information I(T; X_1, ..., X_n) into D(n)-2 components, where D(n) is the n-th Dedekind number. For two sources X_1 and X_2, this resolves into the so-called bivariate PID, i.e. the sum I(T;X_1,X_2) = S + R + U_1 + U_2, with each summand (atom) respectively referred to as the synergistic, redundant, and unique information(s). Crucially, every PID is uniquely defined by a redundancy function , such as the function defined in <cit.>. In the bivariate case, the redundancy function determines R. More generally, it determines the redundant information shared between multiple sources, formalized as collections of random variables. The PID has primarily been applied to theoretical and experimental neuroscience <cit.>, although there have been a few works applying its tools to other areas <cit.>. Although the framework was received with great interest from interdisciplinary information theorists and complex systems scientists, PID remains ill-posed due to the multiplicity of proposed redundancy functions. The original redundancy function remains controversial, since it does not distinguish between sources providing the same information regarding a target outcome from the same amount of information <cit.>. At present, we do not find this critique relevant to the application at hand, as we are more concerned with choosing a posterior that sufficiently constrains the target variables for accurate state inference. The redundancy function takes the PID lattice of source antichains as its domain, so we must first define that set. Let Z_1, ..., Z_n ∈ℝ^N be a collection of random variables in our N-dimensional configuration space, referred to as predictor variables, and Z = (Z_i)_i=1^N be the full n × N vector. In keeping with many previous PID works including <cit.>, when it will not cause confusion, we will identify sets of indexed variables with vectors ordered along increasing index, i.e. Z = (Z_i)_i = { Z_i }_i. Let 𝒫(A) to denote the powerset of any set A, and [z] = { 1, ..., z} for any integer z. Then 𝒫(Z) denotes the collection of sources, i.e. subsets of predictor variables, which we denote Z_J for some index vector J ∈𝒫([n]). The PID lattice, denoted 𝒜 (Z), is the collection of antichains of the poset (𝒫(Z), ⊆). Put differently, the PID lattice is comprised of all possible collections of sources in 𝒫(Z) such that no two sources in the same collection are comparable by inclusion. Put formally, for any nonempty α∈𝒫(𝒫(Z)), α∈𝒜(Z) ⇔∀Z_J,Z_J'∈α, Z_J ⊄Z_J'. The antichain lattice 𝒜(Z) comes equipped with the following partial ordering: α≼β⇔∀Z_J∈β, ∃Z_J'∈α, J' ⊆ J Thus, as one ascends the lattice, individual sources within the antichain will encompass more variables, non-monotonically decreasing the number of distinct sources in the antichain by the top. In particularly, if α_min = {{Z_i }}_i (i.e. the antichain of singletons) and α_max = {Z} (i.e. the source containing all predictors), we have α_min≼β≼α_max for all β∈𝒜(Z). When convenient, we may use an abuse of notation whereby the antichain α contains the index sets for its sources, e.g. J ∈α. In this work, every source variable Z_i has a smooth density p_Z_i supported everywhere on the configuration space. Moreover, every joint density p_Z_J, including p_Z, is smooth and supported everywhere on the product space. Let the target variable T ∈ℝ^M be given, and assume that it has a smooth joint density p_Z,T with the predictors, supported everywhere.[We assume full support to simplify our presentation.] The specific information of a predictor source Z_J toward the target variable is the integral function defined pointwise on target realizations T=t. Let the predictors and target variable and (Z, T) have a smooth joint density p_Z,T. Then for any t such that p_T(t) > 0, we define the specific information of Z_J about T=t to be I_Z_j(t) = D ( p_Z_J|T(Z_J|t) || p_Z_J(Z_J) ). Moreover, we let I_Z(t) = 0 whenever p_T(t) = 0. We emphasize that we use the subscript Z_J to label our specific informations, but they are not functions of the variable Z_J, only T. The redundancy function is defined on 𝒜(Z) as follows: Let α∈𝒜(Z) be the antichain composed of sources α = (Z_J_k)_k. Then (α;T) = 𝔼min_k I_Z_J_k(T) The redundancy function satisfies several desirable properties: * Nonnegativity (P) (α;T) ≥ 0 ∀α∈𝒜(Z) * Monotonicity (M) (α;T) ≤(β;T) if β⊆α * Self-redundancy (SR) ({Z_J}, T) = I(Z_J, T) The latter two of these are typically enumerated among the Williams-Beer (WB) axioms, along with a symmetry property that follows from being well-defined on 𝒜(Z). This structure allows the function to be used to decompose the full information I(T; Z) into nonnegative atoms in the discrete case, using the Möbius inversion of , typically denoted Π: Π(α) = (α) - ∑_β≺α(β) In the bivariate case, the redundancy, synergy, and unique information are given by R= Π({Z_1}{Z_2}), S = Π({Z}), and U_i = Π({Z_i}). Although these other atoms may be of future interest for the analysis of factor graphs, in this work we restrict our attention to redundancies. §.§ Linear Factor Graphs Factor graphs <cit.> are graphical models that are well suited to modeling complex perception estimation problems in robotics, such as Simultaneous Localization and Mapping (SLAM), or Structure from Motion (SFM) <cit.> <cit.>. Similar to other probabilistic graphical models (PGMs) such as Bayesian networks and Markov random fields, factor graphs allow high-dimensional state variables to be decomposed into components with sparse dependencies. A common application of factor graphs in robotics is in pose graph optimization (PGO), in which there have been many advancements in recent years. For centralized PGO, there has been a development of certifiably correct estimation methods that are capable of efficiently recovering globally optimal solutions of generally intractable estimation problems <cit.> <cit.>. In distributed PGO, Tian et al. <cit.> presented the first Distributed Certifiably Correct Pose- Graph Optimization (DC2-PGO). Given all measurements and priors on the state z, factor graphs formally factor the posterior density p(x|z) into the product of functions specific to each measurement, typically unnormalized densities of the relationship between a state observable h_i(x) and the corresponding measurement z_i. For instance, in SLAM and PGO, measurements of one or two keyframe(s) from the trajectory are pulled from odometry, IMU, LiDAR scans, and/or visual perception. Taken together, all these measurements allow for the inference of the full trajectory from its posterior. However, by representing this distribution in a factor graph, the maximal a posteriori (MAP) trajectory estimate can be computed piecemeal by exploiting the sparsity of the graph <cit.>. For the theoretical elements of this paper, we will use the closed-form expression for the MAP of a linear factor graph. Consider a system comprised of n variables in N-dimen- sional configuration space, i.e. X_1, ..., X_n ∈ℝ^N. A factor graph is comprised of these variables and an associated collection of factors ℱ = {ϕ_j}_j=1^m, which is a collection of real-valued functions, with each ϕ_j taking a subset of variables X_i_j as its argument. Most often, factors take one or two variables as their argument. Formally, a factor graph is the bipartite graph 𝒢 = (𝒱,ℰ), where 𝒱 = ℱ∪X and { i,j }∈ℰ iff i ∈i_j, i.e. iff X_i is an argument for ϕ_j. In robotic state estimation, univariate factors can encode priors (including starting position and exogenous map information) and GPS measurements, while bivariate factors might encode wheel odometry, visual perception, or LiDAR point-cloud comparisons <cit.>. In this work, we model every factor as a multivariate Gaussian distribution, produced by a perfectly calibrated Gaussian measurement. We use 𝒩_z(m, Σ) to represent Gaussian densities at Z=z for mean m and covariance matrix Σ. For any vector z, we denote the Mahalanobis norm with its precision matrix, i.e. z ^2_M = z^⊤ M z, since we will more often work with precision matrices instead of noise matrices. Every factor ϕ_j is given by: ϕ_j (x_i_j) = exp -1/2A_j x - z_j _Γ_j^2 for some N × n N measurement matrix A_j. This is analogous to the Jacobian matrix of the measurement operator in nonlinear factor graphs <cit.>. We will often merely refer to the collection of factors ℱ as the factor graph, and subsets of it as subgraphs. Such subsets are denoted ℱ_J⊆ℱ for some vector of factor indices J ⊂ [m]. Using powerset notation, the collection of all subgraphs forms a partially ordered set (poset) under inclusion: (𝒫(ℱ), ⊆). Note that ℱ_J'⊂ℱ_J iff J' ⊂ J, and we will often prefer the latter notation for convenience. Any factor subgraph ℱ_J is associated to the product factor Φ_ℱ_J = ∏_j ∈ Jϕ_j(x_i_j). By concatenating the measurements and measurement matrices, we may rewrite these Φ_ℱ_J (x) = exp -1/2A_Jx - z_J_Γ_J^2 where A_J = [ A_j_1; ⋮; A_j_|J| ], z_J = [ z_j_1; ⋮; z_j_|J| ], Γ_J = [ Γ_j_1 0 ; ⋱ ; 0 Γ_j_|J| ] For the full graph ℱ, we omit subscripts on A, z, and ℱ. In the event that (A_J)=n, the product factor Φ_ℱ_J presents the unnormalized density for a joint distribution on X. In this situation, we say that ℱ_J is full-rank. By solving the quadratic form in the exponential, i.e. refactoring it into the standard Gaussian form, we have Φ_ℱ_J ( ) ∝𝒩_ (μ_J, Λ_J^-1) Λ_J = A_J^⊤Γ_J A_J μ_J = Λ_J^-1 A_J^⊤Γ_J where, as before, we omit subscripts when referencing the full graph ℱ. In fact, for any full-rank ℱ_J, this distribution Φ_ℱ_J corresponds to posterior state distribution after accounting for the included measurements, i.e. p(x | z_J) ∝Φ_ℱ_J(x) Viewed from this perspective, we see that μ_J is the maximum a posteriori (MAP) state estimate <cit.>. In practice, the dimensionality of Λ will be quite large, and it is often computationally difficult if not intractable to naively solve for μ. Factor graphs are employed in contexts in which we expect them to be sparse, i.e. we expect A to be sparse, or (as the Jacobian) sparse at every point for nonlinear factor graphs. This allows one to exploit this sparsity to solve for μ iteratively <cit.>. We assume that every measurement z_j is independently sampled, each from the conditional distribution p(z_j | x) ∼𝒩(A_J x, Γ_z_j^-1). Note that the density of this distribution is merely a normalization of ϕ_j. This assumption will allow us to estimate the information content of individual factors, given some prior on X. However, in practice, all of our information regarding X is encoded in ℱ. Rather than assuming a prior distribution separate from ℱ, we instead partition our factors into two sets: base factors (forming the base graph) and supplementary factors. ℱ = ℱ_B∪ℱ_B̅ where B̅ = [m] ∖ B. Taking the density induced by the base graph as our prior distribution, we denote our prior and posteriors as: p_X(x) := p(x | z_B) ∝Φ_ℱ_B p_X|Z_J(x | z_J) := p(x | z_B ∪ J) ∝Φ_ℱ_B ∪ J where we omit subscripts p(x), p(x|z_J) where they are redundant. To simplify notation, we denote J̃ = B ∪ J for all J ⊂B̅, so that we have p(x|z_J) ∼𝒩_x(μ_J̃, Λ_J̃) as per (<ref>). Moreover, from here on, we will employ Δ_J in place of Λ_J for the matrix expression in (<ref>) when it is ambiguous as to whether ℱ_J is full-rank, i.e. whether Δ_J is nonsingular. If Δ_J is singular, then μ_J is not defined. Even if Δ_J is singular, it will still be symmetric and positive semi-definite, and thus ℱ_J̃ will be full-rank, as we will see below (<ref>). Hence, (<ref>) and μ_J̃ are always well-defined. We summarize this exposition by introducing the shorthand of a supplemented factor graph, i.e. factor graphs with such a partitioning of their factors as described above. A supplemented factor graph is represented by the triplet (X, ℱ, B), where X are the target variables, ℱ is a collection of m linear Gaussian factors of the form (<ref>), and B ⊂ [m] is a subset of indices such that ℱ_B is full-rank for X. The elements of ℱ_B are referred to as base factors, and determine the prior distribution p_X as per (<ref>). We refer to ℱ_B̅ as the supplemental factors, Z_B̅ as the supplemental measurements, and any ℱ_J̃ as a supplemented subgraph of ℱ. A supplemented subgraph ℱ_J̃ determines the Z_J̃-posterior distribution p_X|Z_J as per (<ref>). Now that we have meaningful prior and posterior distributions, we will be able to quantify and decompose the information provided by the supplementary factors in this model. To conclude this section, we provide some useful properties that we will need in the next section. Let J,K ⊂ [m] be mutually exclusive sets of factors. Then we have that: Δ_J ∪ K = Δ_J + Δ_K If ℱ_J ∪ K is full-rank,μ_J ∪ K = Λ_J ∪ K^-1( A_J^⊤Γ_Jz_J + A_K^⊤Γ_Kz_K) In particular, for a given base graph B and supplementary set J ⊂B̅, we have that Λ_J̃ = Λ_B + Δ_J μ_J̃ = Λ_J̃^-1( Λ_B μ_B + A_J^⊤Γ_Jz_J ) By expanding A_J∪ K, z_J ∪ K, and Γ_J ∪ K into their J and K block components, we arrive at (<ref>) by direct computation of (<ref>). Similarly, (<ref>) follows from (<ref>). Moreover, the poset of subgraphs is compatible with the Loewner ordering on their information matrices. For any J,J' ⊆ [m], J' ⊆ J ⇒Δ_J'≤Δ_J Follows readily from (<ref>). Using the prior/posterior from (<ref>-<ref>), we may compute the mutual information for any supplemented factor graph ℱ_J̃⊂ℱ. Let ℱ_J̃⊂ℱ be a supplemented subgraph for some J ⊂B̅. Then the mutual information between factor measurements Z_J and the state X are given by I(Z_J ; X) = 1/2log|Λ_J̃|/|Λ_B| = 1/2log |I + Δ_J Λ_B^-1| We recall that for a Gaussian vector U ∼𝒩(μ, Σ), the differential entropy is given by h(U) = 1/2ln |2 π e Σ| <cit.>. Thus, using (<ref>-<ref>) and (<ref>-<ref>), we compute I(Z_J ; X) = h(X) - h(X | Z_J) = 1/2log|2 π e Λ_B^-1|/|2 π e Λ_J̃^-1| = 1/2log|Λ_J̃|/|Λ_B| This computation gives us an information-theoretic monotonicity for the poset of subgraphs. For any pair of full-rank subgraphs ℱ_J, ℱ_J'⊆ℱ, J' ⊆ J ⇒ I(Z_J' ; X) ≤ I(Z_J ; X). There are a few conditional moments of μ_J̃ that we will need in the next section. We enumerate those here: Let supplemented subgraph ℱ_J̃⊂ℱ be given, with prior and posterior distributions as in (<ref>-<ref>). Then we have the following conditional moments for all matrices T and vectors m: 𝔼( μ_J̃ | x) = Λ_J̃^-1( Λ_B μ_B + Δ_J x) 𝔼( μ_J̃ + m_T^2 | x) = ( TΛ_J̃^-1Δ_J Λ_J̃^-1) + Λ_J̃^-1( Λ_B μ_B + Δ_J x) + m_T^2 The first moment (<ref>) follows readily from (<ref>) and 𝔼( Z_J |x)= A_J x. Then let m' = 𝔼 ( μ_J̃|x ) be this conditional mean. We rewrite the quadratic form: μ_J̃ - m' _T^2 = Λ_J̃^-1A_J^⊤Γ_J( z_J - A_J x) _T^2 = z_J - A_J x_H^⊤TH^2 letting H = Λ_J̃^-1A_J^⊤Γ_J. Since p(z_J | x) ∼𝒩(A_J x, Γ_J^-1), in expectation we have that: 𝔼_p(z_J | x) μ_J̃ - m' _T^2 = (Γ_J^-1H^⊤TH) = ( TΛ_J̃^-1Δ_J Λ_J̃^-1) taking advantage of the cyclic property to rearrange the trace. The extension to Eq. (<ref>) is an application of a well-known recentering formula <cit.>, whereby for U∼𝒩(v, Σ): 𝔼U - v' ^2_M = (MΣ) + v - v' _M^2. § MEASURING REDUNDANCY IN LINEAR FACTOR GRAPHS In this section, we introduce two metrics for measuring the redundancy of information provided by supplemental factors in a factor graph. One of these metrics is a straight-forward application of the redundancy function from PID, while the latter will consider redundancy pointwise on the target variable(s) in a manner analogous to but from the perspective of statistical distance. In order to maintain notational similarity between our metrics without causing confusion with PID functions from the literature (which, ultimately, are all extensions of mutual information and denoted with an annotated I), we denote the -based metric as . We then define another redundancy metric based upon pointwise statistical distance, denoted in reference to the Wasserstein-2 distance. These two functions and are by no means the only possibilities, but demonstrate distinct ways to quantify redundancy “pointwise” with respect to target variables. Before we define each metric in turn, we will first introduce a minimal formalism for redundancy that both will follow, generalizing the PID lattice from  <cit.> to more arbitrary quality metrics of posterior distributions. The motivation here is to provide a unified language for redundancy, independent of the quality function that is being redundantly guaranteed. Let the supplemented factor graph (X, ℱ, B) be given as in Def. <ref>. Further, consider the measurements Z as the collection of predictor variables for target X as in Sec. <ref>, with the collection of antichains 𝒜(Z) defined as in Eqs. (<ref>-<ref>). A PID-like redundancy measure is the pair (𝚁, 𝚀), where 𝚀: 𝒫(Z_B̅) → [Q_min, Q_max] is the quality function and 𝚁: 𝒜(Z_B̃) → [Q_min, Q_max] is the redundancy function, when they fulfill the following axioms: (MR) Monotonicity of 𝚁 For all α, β∈𝒜(Z), we have that 𝚁(α) ≤𝚁(β) whenever β⊂α. (MQ) Monotonicity of 𝚀 For all Z_J, Z_J'∈𝒫(Z), we have that 𝚀(Z_J) ≤𝚀(Z_J') whenever Z_J⊆Z_J' (SR) Self-Redundancy 𝚁({Z_J }) = 𝚀(Z_J) for all J ⊂B̅. A notable feature of the PID is that it quantifies redundancy by looking at the minimal specific information provided by any source, in expectation, about the target variable pointwise. That is, for any realization X=x of the target, accounts for the least specific information that any source might provide under conditional expectation. Although this definition of redundancy has generated controversy in the PID literature, we consider it intuitively aligned with what we would want from a redundancy metric in factor graphs. Namely, redundancy ought to (a) provide robustness guarantees on the minimal information known if one or more sources fail, and (b) identify superfluous information sources in resource-constrained scenarios. We want this same target-pointwise specificity for any redundancy function. Thus, we offer the following definition as a further specification of Def. <ref>: A PID-like redundancy measure (𝚁,𝚀) is said to be a target integral if there exists a mapping 𝚂: 𝒫(Z_B̅) → L^1(ℝ^nN, p(x) dx), denoted Z_J ↦𝚂_Z_J, such that 𝚁(α) = ∫ p( x) min_J ∈α𝚂_Z_J ( x) d x 𝚀(Z_J) = ∫ p( x) 𝚂_Z_J ( x) d x We refer to this mapping as the specific quality function, with 𝚂_Z_J(x) as the specific quality about x provided by Z_J. This notion of specific quality is motivated by specific information from <cit.>, which was used to define PID in <cit.>. Both and are target integrals, and with Defs. <ref>-<ref> can now be entirely defined by their specific quality functions. Moreover, it is evident that any target integral redundancy measure will satisfy axioms (MR) and (SR). Note that computing (<ref>) explicitly requires one to partition the target state space into regions where a given specific quality is minimal. In practice, it will be easier to numerically approximate this integral with Monte Carlo sampling from p(x). §.§ An Information-Theoretic Redundancy Measure The first redundancy metric we consider is the application of the PID from <cit.>, (Sec. <ref>) to the supplementary measurements in factor graphs (Sec. <ref>). Here, the posterior quality metric will be standard mutual information, i.e. ≡ I(·; X), decomposed by the ≡( · ; X) redundancy function from Def. <ref>. It is already well-known that , as the PID, satisfies (MR) and (SR) axioms in Def. <ref>, while (MQ) follows from Cor. <ref>. Let the supplemented factor graph (X, ℱ, B) be given. For every supplemental source Z_J ⊂Z_B̅, the specific information provided about X=x is given by, _Z_J(x) = 𝔼( logp(x | Z_J)/p(x) | x) For any (X,ℱ, B), we may compute the closed-form of this function. _Z_J(x) = I(Z_J ; X) - 1/2[ ( M_J') - x - μ_B_ M_J^2 ] where M_J = Λ_B - Λ_BΛ_J̃^-1Λ_B M'_J = Δ_JΛ_J̃^-1 Λ_J̃ = Λ_B + Δ_J Since Λ_J̃ > Λ_B, we have that M_J > 0, i.e. is positive definite. Thus, the term x - μ_B_ M_J^2 ≥ 0 for all x, and 𝔼x - μ_B_ M_J^2 = (M'_J). Therefore _Z_J(x) () reduces to mutual information when there is only one source in α, i.e. the self-redundancy axiom (SR). What distinguishes and PID from previous information-theoretic approaches to redundancy (e.g. interaction information <cit.>) is the pointwise decomposition of information. As we mentioned above, for the PID, this pointwise decomposition is only with respect to the target variable, not the predictors. We see this pointwise nature explicitly in the quadratic form in (<ref>). For any realized error x - μ_B between the true state and the prior mean, the pointwise minimal source Z_J^⋆(x), where J^⋆=_J ∈α_Z_J(x), will be that source which least corrects this error — assuming all predictor sources J ∈α contribute information matrices Δ_J with similar profiles. Proposition <ref> will follow readily from the following computation. 𝔼( x - μ_J̃_Λ_J̃^2 - x - μ_B_Λ_B^2 | x) = (M_J') - x - μ_B_M_J^2 From (<ref>) from Prop. <ref>, we compute the moment 𝔼( x - μ_J̃_Λ_J̃^2 | x) = (Λ_J̃Λ_J̃^-1Δ_J Λ_J̃^-1) + Λ_J̃^-1(Λ_B μ_B + Δ_J x) - x^2_Λ_J̃ = (M'_J) + x - μ_B ^2_Λ_B Λ_J̃^-1Λ_B By recombining with the conditional constant x - μ_B_Λ_B^2, we arrive at (<ref>). We may now prove Proposition <ref>. Using the expressions from Sec. <ref>, we may write the pointwise mutual information term in (<ref>): logp(x | z_J)/p(x) = logΦ_ℱ_B ∪ J(x)/Φ_ℱ_B(x) = 1/2log|Λ_J̃|/|Λ_B| - 1/2[ - μ_J̃_Λ_J̃^2 - - μ_B _Λ_B^2 ] We take this in expectation with respect to p(z_J | x). The first term is constant, and we use Lemma <ref> to integrate the second, arriving at (<ref>). §.§ A Statistical Distance Redundancy Measure For our second redundancy metric, we consider a different approach. We instead define redundancy using the statistical distances between the source-defined posterior distributions p(X'| Z_J) and ground-truth target δ_x. Our redundancy metric will look at the pointwise reduction in this statistical distance relative to that of the prior distribution p(X') from the base graph. Let the supplemented factor graph (X, ℱ, B) be given. We define the specific Wasserstein error reduction (WER) function to be _Z_J( x) = 𝔼( d_𝒲^2^2(p(X'), δ_x) - d_𝒲^2^2(p(X' | Z_J), δ_x) | x) where d^2_𝒲^2 is the square Wasserstein-2 distance. For arbitrary Gaussian density f(u) ∼𝒩(m,Σ) and point mass δ_u', this takes the form: d_𝒲^2^2( f(u), δ_u' ) = (Σ) + m - u' ^2 Equivalently, one could also define _Z_J( x) as a reduction in the target-centered L^2-norm of the posterior pose-graph distribution, relative to the prior: _Z_J( x) = 𝔼( X' - x^2 - X” - x^2 | x) where X' ∼ p_X and X”∼ p_X | Z_J (<ref>-<ref>). Before we proceed to computing a closed-form for , we will first compute the closed form of the quality function induced by Def. <ref> through Def. <ref>. This quality function will serve a role analogous to single-source mutual information in PID. The quality function for the redundancy measure induced by Def. <ref> has the closed form: (Z_J) = 2 (Λ_B^-1 - Λ_J̃^-1) Moreover, this function satisfies monotonicity (MQ). Combining Defs. <ref>&<ref>, we have that (Z_J) = 𝔼 d_𝒲^2^2(p(X'), δ_X) - 𝔼 d_𝒲^2^2(p(X' | Z_J), δ_X) Since 𝔼| μ_J' - x ^2 = (Λ_J'^-1) for all Z_J'⊂Z, we use (<ref>) to find that find that 𝔼 d_𝒲^2^2(p_X, δ_X) = 2 (Λ_B^-1) and 𝔼 d_𝒲^2^2(p_X|Z_J, δ_X) = 2 (Λ_J̃^-1). The monotonicity of 𝚀 follows from Prop. <ref> and the monotonicity of the trace. We now compute the closed-form solution of our specific quality function. _Z_J( x) = ( N'_J ) + μ_B - x^2_N_J whereN'_J = Λ_B^-1 - Λ_J̃^-1 - Λ_J̃^-1Δ_J̃Λ_J̃^-1 N_J = I - Λ_B Λ_J̃^-2Λ_B From (<ref>) and (<ref>), we have that _Z_J( x) = (Λ_B^-1 - Λ_J̃^-1) + 𝔼( μ_B - x^2 - μ_J̃ - x^2 | x) As a special case of the moment (<ref>) from Prop. <ref>, we have 𝔼( μ_J̃ - x^2 | x) = (Λ_J̃^-1Δ_J Λ_J̃^-1) + μ_B - x^2_Λ_B Λ^-2_J̃Λ_B Substituting this into (<ref>) and recombining terms, we arrive at (<ref>). § AN APPLICATION TO 2D LANDMARK SLAM In order to investigate the plausibility and behavior of our proposed and redundancy metrics, we now apply them to simulations of 2D landmark-based SLAM task. In our simulation, a single robot randomly traverses a 2D environment populated by two landmarks. By taking measurements of both landmarks, it will supplement its pose graph with additional, potentially redundant information. We will use our redundancy metrics to estimate the redundancy that comes from using both landmarks, rather than one. In these simulations, we dispense with the linearity restriction we have assumed up this point. Our robot's configuration space will be the 2D rigid body transformations (2), representing its pose (translation and rotation) relative to the origin in ℝ^2. For all factor graphs operations, we use the GTSAM <cit.> toolbox for nonlinear factor graphs. Allow y^t∈ℝ^2 to represent the translation of any y ∈(2), and y^t = (y_i^t)_i=1^n for any vector y∈( (2) )^n. We denote the robot's pose trajectory as X = (X_0, ..., X_n), taking values in SE(2) over n timesteps, and let L_0 and L_1 denote the positions of the landmarks in ℝ^2. Collectively, these n+2 variables are represented in the robot's factor graph, yet we will compute the redundancy metrics only with respect to trajectory variables X. Both the robot's initial position and those of the landmarks are independently sampled from uniform distributions at initialization, i.e L_0, L_1 ∼𝒰([-C,C]^2) and X_0^t∼𝒰([-C/2,C/2]^2) for some C>0. We generate the robot's trajectory with a stationary random walk, i.e. η_i = X_i-1^-1 X_i is sampled with the same mean and covariance for each i. At each step i=1,...,n, the robot takes an unbiased odometry measurement Z_i of η_i with fixed noise matrix Σ_odom in (x,y,θ)-coordinates. These odometry measurements compose the base graph. Moreover, at every timestep i, the robot takes a range-bearing measurement of each landmark, indexed Z_(s+1)n+i for landmark s ∈{0,1}, and adds it to its factor graph. The bearing measurement has fixed variance, while the variance of the range measurement scales quadratically with distance. We will be estimating the redundancy of the 2-antichain α = {Z_J_0, Z_J_1} where J_s = { i | (s+1)n < i ≤ (s+2) n} — that is, the redundancy between the cumulative measurements of L_0 and those of L_1. In order to compute and , we compute the integral in (<ref>) by sampling from the base graph X' ∼ p_X. We linearize each subgraph ℱ_J̃_s about its solution μ_J̃. We then estimate each of the specific quality functions 𝚂_Z_J̃_s at x using the linear formulae from Props. <ref> and <ref> as a first approximation. Given ground-truth, one can evaluate the localization component of a SLAM pipeline using absolute or relative trajectory error (ATE and RTE, respectively) <cit.>. It is less immediately obvious what ground-truth redundancy estimates should be compared against. We will define a variant of ATE that aims at ground-truth error from a redundancy perspective. Let the nonlinear supplemented factor graph (X, ℱ, B) be given for ground-truth trajectory x∈2^n. Moreover, let an antichain of supplemental factors α = (Z_J_1, ... Z_J_k) be given as well, with μ_J̃_k as the MAP estimate of X from each ℱ_J̃_k. The worst-case absolute trajectory error (WC-ATE) is given by: WC-ATE(x, z) = ∑max_J_kx^t_i - R_J_kμ_J̃, i^t - t_J_k^2 where R_J_k, t_J_k = (R,t)(x^t,μ_J̃_k^t) are the transformation and translation from the Umeyama method for aligning point-sets <cit.>. In this case, we compute the WC-ATE for the antichain α = {Z_J_0, Z_J_1} to compare to the redundancy scores. We present exploratory results from our experiments in Figs. <ref>-<ref>. In Fig. <ref>, we plot the WC-ATE of 500 simulated trajectories of n=10 poses against the and redundancies estimated from the robot's factor graph. For the , we see that higher redundancies suggest, with high probability, a lower WC-ATE — i.e, the error in any pose is likely to be small regardless of which landmark is used for SLAM. However, the converse is not true: lower estimated does not necessarily imply that both landmarks are needed in every realization. This is expected: WC-ATE is based upon the fully realized trajectory x and measurements z, whereas the redundancies are computed in expectation. Put differently, it will often be the case that μ_J̃_s for both s=0,1 will be closer to x in realization than expectation. In Fig. <ref>, we explore a possible source of heterogeneity between simulations that we would expect to affect redundancy: the distance of the landmarks to the robot. Since the simulated range-bearing measurements degrade with distance, we expect the information quality contributed by each landmark to scale down with distance. Moreover, we expect maximal redundancy in the contributed quality when both landmarks are close to the robot. We see that this is generally the case: the highest redundancy scores for both and were computed from trajectories that remained close to both landmarks. § CONCLUSION AND FUTURE WORK In this paper, we have begun the exploration in the formal and theoretical direction for developing a theory of multi-source redundancy for factor graphs, drawing upon ideas from the partial information literature. We adapt the PID axiomatic framework to the factor graph setting, generalizing the language of redundancy for quality metrics beyond mutual information (as in PID). Within this framework, we contribute two redundancy metrics on linear factor graphs. These metrics were applied to simulations of a 2D SLAM scenario. We are currently working toward creating a rigorous toolbox for estimating and employing redundancy in linear factor graphs. In future work, we will generalize the redundancy metrics proposed to nonlinear factor graphs. ieeetr

http://arxiv.org/abs/2303.07368v1

20230313180003

$C_2$ generalization of the van Diejen model from the minimal $(D_5,D_5)$ conformal matter

hep-th

=1 toc lof lot

http://arxiv.org/abs/2303.07071v1

20230313124652

Correlates of Programmer Efficacy and Their Link to Experience: A Combined EEG and Eye-Tracking Study

cs.SE

Correlates of Programmer Efficacy and Their Link to Experience]Correlates of Programmer Efficacy and Their Link to Experience: A Combined EEG and Eye-Tracking Study Saarland University, Saarland Informatics Campus Saarbrücken Germany Saarland University, Saarland Informatics Campus, Graduate School of Computer Science Saarbrücken Germany German Research Center for Artificial Intelligence, Saarland Informatics Campus, Saarbrücken, Germany Chemnitz University of Technology Chemnitz Germany German Research Center for Artificial Intelligence, Saarland Informatics Campus, Saarbrücken, Germany NC State University Raleigh North Carolina USA Chemnitz University of Technology Chemnitz Germany Saarland University, Saarland Informatics Campus Saarbrücken Germany Background: Despite similar education and background, programmers can exhibit vast differences in efficacy. While research has identified some potential factors, such as programming experience and domain knowledge, the effect of these factors on programmers' efficacy is not well understood. Aims: We aim at unraveling the relationship between efficacy (speed and correctness) and measures of programming experience. We further investigate the correlates of programmer efficacy in terms of reading behavior and cognitive load. Method: For this purpose, we conducted a controlled experiment with 37 participants using electroencephalography (EEG) and eye tracking. We asked participants to comprehend up to 32 Java source-code snippets and observed their eye gaze and neural correlates of cognitive load. We analyzed the correlation of participants' efficacy with popular programming experience measures. Results: We found that programmers with high efficacy read source code more targeted and with lower cognitive load. Commonly used experience levels do not predict programmer efficacy well, but self-estimation and indicators of learning eagerness are fairly accurate. Implications: The identified correlates of programmer efficacy can be used for future research and practice (e.g., hiring). Future research should also consider efficacy as a group sampling method, rather than using simple experience measures. <ccs2012> <concept> <concept_id>10003120.10003121.10011748</concept_id> <concept_desc>Human-centered computing Empirical studies in HCI</concept_desc> <concept_significance>300</concept_significance> </concept> <concept> <concept_id>10003120.10003121.10003122</concept_id> <concept_desc>Human-centered computing HCI design and evaluation methods</concept_desc> <concept_significance>300</concept_significance> </concept> </ccs2012> [300]Human-centered computing Empirical studies in HCI [300]Human-centered computing HCI design and evaluation methods [ Sven Apel March 30, 2023 ================== § INTRODUCTION Proficient programmers are essential for providing the critical infrastructure and functioning applications for our modern society <cit.>. Although the learning strategies, education, and pathways to becoming a programmer may differ, the general expectation is that, with more time and experience, a programmer should generally become more proficient. For example, many research studies of programmers use years of experience <cit.>, education level <cit.>, or employment status <cit.> as foundational measures for proficiency. While all these choices are sensible, they all implicitly encode the expectation that higher proficiency should correlate with more experience. However, in practice, reported observations violate this expectation. Hiring managers and technical founders <cit.> report that they routinely encounter “engineers with years of experience who couldn't competently program” <cit.>, who “struggle with tiny problems” <cit.>, and “Senior Engineers who can't write basic code” <cit.>. The original creator of the infamous FizzBuzz interview question <cit.> only did so after seeing that the “majority of comp sci graduates... and self-proclaimed senior programmers” had difficulty solving simple problems in a timely manner. Research has also found that, in companies, the seniority level showed little correlation to actual programming skill <cit.>, and programmers with similar education and background can exhibit vast differences in productivity, up to factors of ten <cit.>. Research is still at a loss when it comes to explaining the cause of these differences, accounting for different trajectories of learning that underlie programming education and training <cit.>, or identifying proficient programmers during a hiring process <cit.>. In this paper, we explore the idea of unraveling programmer efficacy[Efficacy specifically refers to the ability to quickly produce the intended result. Notably, this differs from expertise, which involves additional facets, such as deep knowledge, effort, and mastery of skills.]—based on speed and correctness—with the help of programmer experience—amount of learning or practice. To improve our understanding of the relationship between efficacy and experience, we conducted a combined electroencephalography (EEG) and eye-tracking study, allowing us to take a close look at how differences in efficacy and experience are related to cognitive differences among programmers. In our study, 37 participants with varying levels of experience performed program-comprehension tasks. We found that programmers with higher efficacy read code more targeted, with shorter fixations, fewer (re)fixations, and skipping more code elements. They also complete their tasks with lower cognitive load, in less time, and make fewer errors than programmers with lower efficacy. Interestingly, we found that commonly used experience measures do not correlate with the observed efficacy, but instead self-estimation and learning indicators have considerable predictive power. To this end, we have identified correlates of high programmer efficacy as well as experience measures that provide a strong link to efficacy. In summary, we make the following contributions: * A combined EEG and eye-tracking experiment to investigate programmer efficacy with a diverse participant pool. * Confirmation of prior results that programmers with high efficacy read source code more efficiently and with lower cognitive load. * Empirical evidence that conventional experience measures have only poor predictive power for programmer efficacy. Self-estimation and indicators of learning eagerness are better suited. * An online replication package[<https://github.com/brains-on-code/NoviceVsExpert>], including experiment design, raw data, and executable analysis scripts. § RESEARCH QUESTIONS AND VARIABLES Our study on programmer efficacy builds on the methodology of previous experiments investigating program comprehension and programming experience. Our aim is to incorporate measures of program comprehension from these experiments into a single coherent study as shown in <Ref>. We specifically designed our study to better understand programmer efficacy across a wide range of experience levels in the context of program-comprehension tasks. Programmer efficacy is therefore the independent variable for our experiment. We operationalize programmer efficacy as follows: 𝑝𝑟𝑜𝑔𝑟𝑎𝑚𝑚𝑒𝑟 𝑒𝑓𝑓𝑖𝑐𝑎𝑐𝑦 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑐𝑜𝑑𝑒 𝑐𝑜𝑚𝑝𝑟𝑒ℎ𝑒𝑛𝑠𝑖𝑜𝑛 𝑡𝑎𝑠𝑘𝑠/𝑐𝑜𝑚𝑝𝑙𝑒𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒 𝑖𝑛 𝑚𝑖𝑛𝑢𝑡𝑒𝑠 Note that programmer efficacy captures both the speed and correctness of a participant's behavior. Due to the strict one-hour time limit of the experiment (see <Ref>), only fast participants were able to see all snippets before the experiment ended. Our definition of programmer efficacy eliminates effects arising from a difference in the number of attempted tasks and is in line with prior work <cit.>. To guide our experiment design regarding dependent variables, we defined several research questions, which we introduce next. §.§ Reading Behavior (RQ1) Several studies have shown that experienced programmers show a different reading behavior than novices based on eye-gaze measures. The reading behavior describes a programmer's eye movements while they are comprehending a source-code snippet. Therefore, we state the following research question: RQ1Do different levels of programmer efficacy exhibit differences in reading behavior (in terms of navigation strategy and code element coverage)? Operationalization For RQ1, we adopt 7 measures that have been identified in the literature: In a longitudinal study, Al Madi et al. observed differences in the navigation strategy at the token level <cit.>. Specifically, they analyzed fixation duration and how likely it was that a participant (re)fixated on a token. Experts showed significantly shorter fixation durations, a lower chance of refixating on a token, and a higher chance to skip tokens. Likewise, Aljehane et al. found differences between novices and experts in terms of code element coverage <cit.>. It refers to the number of elements on which a participant fixates during a task in contrast to the total number of code elements. They found that novices read, in particular, more method signatures, variable declarations, identifiers, and keywords. Finally, Busjahn et al. have also identified a difference in code element coverage, in that novices fixate more code elements than experts <cit.>. They also identified a difference in the reading order. However, our replication study revealed that this effect is principally driven by the execution order of the snippets <cit.>. As our snippets are not balanced for this aspect, we will not consider reading order as a measure; these studies used different measures to pinpoint the participants' reading behavior. In our study, we aim at answering how programmer efficacy affects reading behavior across all described measures to increase comparability. §.§ Cognitive Load (RQ2) Similar to reading behavior, previous studies found that different levels of programming experience can be distinguished by the observed neural correlates of cognitive load, which leads us to our next research question: RQ2Do different levels of programmer efficacy exhibit differences in neural correlates of cognitive load? Operationalization Related to RQ2, Crk et al. used event-related desynchronization along alpha and theta powerbands measured by an EEG device to classify participants into two programming experience levels <cit.>. Similarly, Lee et al. found a similar effect also using EEG, but using a more comprehensive analysis across more EEG bands <cit.>. Ishida and Uwano found an increase in the alpha frequency band for programmers who successfully finished tasks <cit.>. However, Madeiros et al. compared several EEG measures to distinguish different experience levels and suggested the ratio between theta and alpha power as a measure of cognitive load <cit.>. Notably, this ratio is commonly used as a cognitive load measure in other fields <cit.>. We analyze both powerbands and the ratio measure in our study to better understand the link between programmer efficacy and cognitive load. §.§ Experience Measures (RQ3) Finally, we focus on programmer efficacy as a distinguishing factor between our participants. Prior research most commonly uses rather simple experience measures to separate participants (e.g., years of programming). In this vein, we pose our last research question: RQ3Do different levels of programmer efficacy correlate with common measures of programming experience? Operationalization Many commonly used experience measures can be inconsistent in their predictive power, which led us to use programmer efficacy as the distinguishing factor. In RQ3, we analyze widely used experience measures[We share the full list of experience measures on the project's Web site.] and distill the relevant measures that correlate best with the observed programmer efficacy. § STUDY DESIGN AND CONDUCT To answer our research questions, we designed and conducted a study, which we describe in this section. All materials including snippets, tasks, and the experiment script, are available on the project's Web site.^2 §.§ Experiment Plan We opted for a within-subject experiment design <cit.>, which we illustrate in <Ref>. We presented up to 32 source-code snippets with a program-comprehension task (cf. <Ref>), which were pseudo-randomized to avoid learning and fatigue effects. The experiment ended after the 32 program-comprehension tasks were completed or after 60 minutes, whichever happened first. The tasks were presented in three runs of 20 minutes and a voluntary break of 5 minutes between runs. In addition to the program-comprehension tasks, we included a search task, in which participants had to count brackets. This serves as a baseline for neural activation and is common in neuroscience studies of program comprehension <cit.>. We chose a 4:1 design, so a participant completes four program-comprehension tasks and one search task. Between tasks, we included a 30 second rest condition, in which participants were instructed to focus their eyes on a fixation cross and relax. For the comprehension tasks, the participants could choose among four answer options (cf. <Ref>) as well as the option “next” to skip a task. Participants used the left hand to press space as “submit” and “continue” buttons, and the right hand to navigate with the arrow keys between answer options. After a short training session on the experiment flow (i.e., presentation of an example snippet, example input/output task, answering possible clarification questions), participants could use the keyboard without constantly looking at their hands, which minimizes motion artifacts. Our study was approved by the ethical review board of the faculty of Mathematics and Computer Science at Saarland University. §.§ Snippet Selection Example source-code snippet with intermediate complexity that checks a string for the existence of a substring. An example input provided to the participant would be with answer options “”, “” (correct), “”, and “”. [ fontsize=7pt7pt, linenos, xleftmargin=10pt, numbersep=5pt, frame=lines]Java public boolean containsSubstring(String word, String substring) boolean containsSubstring = false; for (int i = 0; i < word.length(); i++) for (int j = 0; j < substring.length(); j++) if (i + j > word.length()) break; if (word.charAt(i + j) != substring.charAt(j)) break; else if (j == substring.length() - 1) containsSubstring = true; break; return containsSubstring; A crucial element of our experiment are the source-code snippets. We aimed at selecting snippets covering a variety of complexities. This ranges from simple snippets, with only a few lines that can be understood within seconds, to complex source-code snippets that require substantial mental effort to comprehend. Thus, our snippets require different levels of cognitive effort to comprehend them. This helps us to comprehensively capture program comprehension in relation to programmer efficacy. We started our search by selecting Java snippets from a variety of previous studies of program comprehension <cit.>. We then searched for further snippets implementing algorithms of comparable complexity (in terms of size, nesting depth, and execution length), which lead to a pool of 38 snippets. Three of the authors independently assessed each snippet's complexity and suitability for the study (e.g., whether prior knowledge is necessary to comprehend it). While the snippets are in Java, we aimed at selecting snippets that did not require deep knowledge of the language. This resulted in 32 snippets. We show a sample snippet that checks for the existence of a substring in <Ref>. We list all included snippets in <Ref> and provide them in the replication package. §.§ Program-Comprehension Task We subdivided the program-comprehension task into two distinct steps to isolate the underlying cognitive processes of program comprehension and mental calculation of the result. To this end, we first presented the source-code snippet and instructed participants to comprehend it. Once a participant confirmed they comprehended its functionality (by pressing a button), we presented a sample input. Then, the participant had to mentally calculate the resulting output. The sample input was not shown until after the participant fully comprehended the code snippet, to prevent premature mental calculation without fully understanding the snippet. We informed participants of this multiple-step process beforehand and ensured their understanding with two training snippets. §.§ Experiment Execution and Data Collection All participants provided their informed consent and completed our experience questionnaire (see <Ref>). We put the EEG cap on the participant's head, calibrated the eye-tracker, and started the experiment. After the experiment, we conducted a semi-structured interview, including questions on their subjective views about the experiment as well as each snippet. The EEG laboratory is in a dim-light room with minimized distractions, such as external sounds or mobile devices. Participants sat in a comfortable chair to prevent unnecessary muscle movements to reduce noise and artifacts in the EEG signal. EEG signals were recorded using LiveAmp 64[Brain Products GmbH, <https://brainvision.com/products/liveamp-64/>], which is a 64-channel EEG device. The sampling rate was set to 500 Hz, and the international 10–20 system of electrode placement <cit.> was used to cover the entire scalp and obtain spatial information from the brain recordings. For simultaneous eye tracking, we used the Tobii Pro Fusion eye-tracker[Tobii AB, <https://www.tobiipro.com/product-listing/fusion/>] attached to the screen. The eye gaze was recorded with a sampling rate of 250 Hz. The experiment was run with a PsychoPy <cit.> script (available in the replication package). §.§ Participants We recruited participants at Saarland University via e-mail lists and online bulletin boards. Participants received 10 Euro compensation each. The prerequisite for participation was, at least, one year of experience with Java or, at least, three years of experience with a related programming language, such as . It was important to recruit a diverse set of programmers with a wide range of experience levels to explore the differences and commonalities across different efficacy levels. In <Ref>, we provide an overview of our participants' demographics. Based on the conventionally used experience measures, our participant sample exhibits a wide range of programming experience, from programmers in their first year of programming at the university to 30 years of experience in industry. We invited 39 participants to start the experiment. 38 out of the 39 measured participants completed the experiment.[One participant aborted the experiment early due to personal time constraints.] For one participant, the eye-tracker could not be calibrated, but all other modalities are available and the data are included in the analysis. We later excluded one complete participant data set due to their behavioral data consisting of a majority of outliers (cf. <Ref>), leaving us with 37 participants included in the data set. We based our programming experience questions on a validated questionnaire <cit.>, but extended it to cover more topics for further investigation (e.g., programming-content consumption and production and daily work; the questionnaire is available in the replication package). § DATA ANALYSIS In this section, we describe our data-analysis procedure for eye-tracking data, EEG data, and experience measures. §.§ Outlier Removal We started with removing outliers in response time for comprehending a source-code snippet. Specifically, we discarded the slowest 5% (i.e., a response time of over 2 min 32 s) as well as the fastest 5% (under 11 s), but only if a participant chose the option “Next”, since some tasks can be rapidly answered by proficient programmers (e.g., the snippet). This way, data points were only removed when participants did not thoroughly attempt to understand the snippet and program comprehension may not have occurred. With these rules, we removed 9 data points[One data point here is the response time for one participant to one snippet] in the lower 5% and 55 in the upper 5%, leaving 1072 data points for further analysis. This led to the removal of more than half of the data points for one (slow) participant, so we excluded the participant's entire data set from the study. §.§ Eye-Tracking Data We defined areas-of-interests (AOIs) for each source-code snippet to relate eye-gaze behavior with particular regions and elements of code. For this purpose, we first obtained the abstract syntax tree of each source-code snippet to identify each token and code element. Then, we manually identified higher-order syntactical structures such as the head and body of loops or if-else-statements for each snippet and thereupon defined AOIs based on categories defined in previous work <cit.>. We preprocessed the eye-tracking data to separate fixations and saccades from the raw stream of coordinates. For this purpose, we used , which is a noise-robust algorithm <cit.>. Then, we computed the token-level and element-coverage measures described in <Ref>. §.§ EEG Data Our cognitive load measures are based on a spectral analysis of EEG data. To improve the robustness of our analysis, we calculate cognitive load using two different approaches: theta/alpha power ratio and relative power analysis. §.§.§ Data Cleaning First, we removed noise from the EEG data, especially movement artifacts, using established preprocessing methods: We filtered the data using Hamming-windowed finite impulse response (FIR) filters. Power line noise was removed by a notch filter with a lower cutoff frequency of 49 and an upper cutoff frequency of 51. Second, to obtain a more robust performance in the subsequent analysis, we also removed baseline drifts and high-frequent noise with a bandpass filter ranging from 4 to 200 <cit.>. Due to the nature of the experiment, where participants have to look at different points of the screen, eye movements and other muscle activity cannot be fully avoided. Therefore, we performed blind source separation as a second step to remove corresponding artifacts. The used toolbox <cit.>[Version 2021.1, <https://sccn.ucsd.edu/eeglab/>] provides an automated implementation of the independent component analysis (ICA). For each component, it yields a classification with a confidence level, which we can use to automatically reject noisy components. Rejection thresholds for eye and muscle artifacts were set to the default values of 0.9, while components without a clear assignment to a group were rejected at 0.95. §.§.§ Theta/Alpha Power Ratio As a first technique, we computed the cognitive load based on the work of Holm et al. <cit.> and Kartali et al. <cit.> as the ratio of the relative power of the theta and alpha band of the EEG. We calculated the power spectral density on a sliding window obtaining the mean power within each frequency band of interest. The sliding window had a size of 3 and was shifted by 0.1 over each task. The cognitive load measure is obtained by dividing the mean power within a window. By using the sliding window, we can observe the time course of the cognitive load during the tasks. §.§.§ Relative Power Analysis As a second technique, we also computed a per-band and per-channel relative power analysis along alpha, beta, gamma, and theta band based on the technique established by Lee et al. <cit.>, which is sensitive to the spatial location of brain activation by considering the channel of the electrode.[To facilitate comparison with Lee et al.'s work <cit.>, which used two groups, we formed two groups post-hoc based on experimental scores of efficacy. We separated our participants by performance into thirds. For increased potential, we contrasted the high-performing group (i.e., efficacy >= 0.35, n=12) and the low-performing group (i.e., efficacy <= 0.29, n=13), leaving out the middle group.] §.§ Efficacy and Experience For RQ3, we correlated the observed efficacy with all collected numerical and categorical experience measures. To this end, we used Spearman's ρ correlation due to the mix of continuous and ordinal nature of our data. § RESULTS AND DISCUSSION In this section, we present the results of our EEG and eye-tracking data analysis, along with their subsequent interpretation. Applying our programmer efficacy definition to the data, we find that, on average, 1 task was correctly solved around every 3 minutes (i.e., mean programmer efficacy of 0.33 ± 0.10). We visualize the distribution of participant efficacy in <Ref>, which is a non-normal distribution according to a Shapiro-Wilk test (W = 0.94, p = 0.046). It could be normalized by removing the upper tail (2 participants)—neither which were the most experienced programmers—but since we already removed outliers, this appears to be the real distribution in our sample. §.§ Reading Behavior (RQ1) Our eye-tracking data show a change in reading behavior with increasing efficacy levels. Based on the token-level measures from Al Madi et al., we found that increased efficacy leads to: * shorter (first) fixations (ρ=-0.14), * shorter gaze duration (ρ=-0.19), * a much lower chance that a token is revisited (ρ=-0.37), and * a slightly higher probability that a token is skipped (ρ=0.08). We visualize these relationships in <Ref>. Regarding code element coverage, based on Aljehane et al. and Busjahn et al., we find a similar result, such that higher efficacy leads to a lower code element coverage (ρ=-0.35). Notably, this difference increases throughout the task, as we illustrate in <Ref>. To underline the difference between programmer with low and high efficacy, we show an example scanpath in <Ref> (Page fig:EyetrackingScanpathSample). Clearly, the programmer with high efficacy requires fewer fixations and refixations to comprehend the source-code snippet. Overall, this leads to the following answer to our research question: RQ1We can confirm prior results that programmers with higher efficacy read code more efficiently in terms of shorter fixations on fewer code elements. However, some measures show only weak correlations. Discussion Our results for RQ1 confirm prior results and corroborate the theory that proficient programmers read source code more efficiently <cit.> and are actively looking for an efficient way to solve a task <cit.>. This supports the view that their knowledge guides their eyes <cit.>. However, the established measures that we use capture this effect to different degrees. Fixation duration measures show only weak correlations (ρ=-0.14), while fixation probabilities show, at best, medium correlations with programmer efficacy (ρ=-0.37). These results support that proficient programmers read individual elements faster, and, in particular, focus on the important elements—skipping several parts that they deem unimportant. This narrow focus on important elements is further highlighted in the code element coverage measure. Programmers with high efficacy read fewer code elements throughout the task, increasing over time as shown in <Ref>. Even at the end of a task, they still only fixated on 69 % of the tokens, on average—skipping tokens they rate as irrelevant. While a narrow focus enables high performers to quickly solve a task, they might overlook relevant elements. For example, research in psychology has shown that experts' internal filtering heuristics can lead them to miss relevant information and make mistakes <cit.>. In programming, this effect was observed with experts missing obvious syntax errors that novices consistently notice <cit.>. §.§ Cognitive Load (RQ2) The observed ratio of alpha and theta band as an indicator of cognitive load can be analyzed regarding different features, which we summarize in <Ref>. While generally the cognitive load appears to be lower for programmers with higher efficacy (ρ = -0.09), the correlations are weak. Still, it has to be taken into account that programmers with higher efficacy tend to complete their tasks faster. In <Ref>, we visualize the cognitive load over time depending on efficacy, which shows that the level of cognitive load generally stays lower, and with fewer spikes, for programmers with higher efficacy. Results from our powerband analysis confirm results from Lee et al. <cit.>. Like Lee et al., we separated our participants into a high-performing and low-performing group (cf. <Ref>). We found a lower beta power for programmers with higher efficacy, and higher alpha and gamma power, which is illustrated in <Ref>. RQ2The cognitive load is slightly lower for programmers with higher efficacy, despite faster completion times. Programmers with higher efficacy further experience fewer high spikes of cognitive load. This is in line with previous findings of lower beta and higher alpha and gamma power for programmers with higher efficacy. Discussion The results of RQ2 indicate that programmers with higher efficacy can not only comprehend source code faster, but also with less mental effort. This could be explained by an increased neural efficiency, which has been shown in other fields <cit.> or as an effect of source code structures serving as beacons for program comprehension <cit.>, but not as a factor of proficiency in programming. One explanation would be a difference in underlying cognitive processing: Experts may see the presented snippets as a task, that is, they have a strategy to solve the problem, and simply need to implement the solution. This is unlike novice programmers, who also need to find a strategy first, then implement the solution, which leads to higher cognitive load and longer response times <cit.>. Due to the lower cognitive-load levels, programmers with high efficacy likely can sustain longer periods of work. By contrast, programmers with lower efficacy are more likely to be overwhelmed by constant spikes of cognitive load (<Ref>). Synthesis of RQs1–2 Our study was inspired by several prior experiments investigating program comprehension with various operationalizations and different definitions of experience. We aimed at finding a common ground across their measures, specifically when focusing on programmer efficacy. Overall, our study confirmed the accuracy of some of these measures, but to different degrees. Programmers with higher efficacy can be particularly identified due to their efficient reading strategy and lower spikes in cognitive load. This is not only important for future research, but also has practical implications. For example, a lot of hiring processes use technical interviews in front of a whiteboard, which artificially introduce stress and high cognitive load <cit.>. An alternative solution to evaluate potential talent with measures that allow for accurate and quick responses, without inducing unnecessary stress and cognitive load, would be private interviews, comprehension tasks, or other alternative interview methods, for example, as proposed by Behroozi et al. <cit.>. Regarding programming language design, the combination of eye tracking and EEG may have potential. Future work shall explore whether we can use eye-tracking data to pinpoint spikes in cognitive load to specific code elements, similar to Fakhoury et al.'s work with functional near-infrared spectroscopy (fNIRS) <cit.> or our prior work on combining eye tracking and fMRI <cit.>. Such combination may reveal code structures that are particularly difficult for programmers to comprehend. §.§ Experience Measures (RQ3) In <Ref>, we show the strength of correlations between programmer efficacy and a subset of popular measures of programming experience. Notably, some commonly used experience measures (e.g., years of (professional) programming <cit.>) show little predictive power to our participants' efficacy. But, several other dimensions of experience show, at least, medium-strength correlations. Specifically, self-estimation in comparison to peers (ρ=0.59) and a 10-year professional programmer (ρ=0.38) show that our participants seem to be keenly aware of their proficiency level. The number of known programming languages also shows a medium-strength correlation (ρ=0.42). For the subset of professional programmers (n=26), several questions that capture learning eagerness (e.g., learning, mentoring, code review) show medium correlations. However, strikingly, the time professionals spend on pure programming does not correlate to their performance in our experiment (ρ=0.06). Overall, these results allow us to answer this research question: RQ3Programmer efficacy does not correlate with commonly used experience measures, such as years of programming. Self-estimation and indicators of learning eagerness show, at least, medium correlations with observed programmer efficacy. Discussion Our experiment highlights two fundamental problems for studying programmers. One issue is that programming is such a diverse field with different technologies that require different skill sets, so it is incredibly difficult to accurately measure a programmer's experience. Therefore, many researchers rely on simple measures, such as years of programming. However, this can become problematic if the difference between the selected measure and the actual proficiency level becomes a significant confounding factor <cit.>. Our experiment, in line with prior research <cit.>, underlines how limited common experience measures are in predicting programmers' efficacy. Thus, future research must carefully consider collecting more comprehensive experience data, in particular, when using it to separate the participants into groups. While this may not be completely new insights, our experiment further corroborates this critical point. The use of too simplistic measures can potentially weaken empirical studies of programmers and their conclusions. § THREATS TO VALIDITY In this section, we discuss threats to construct, internal, and external validity of our study. §.§ Construct Validity Our programmer efficacy measure may deviate from a participant's true efficacy due to the nature of our selected snippets and programming language. We mitigated this threat by selecting a variety of source-code snippets such that prior knowledge plays a minuscule role, and we ensured sufficient Java knowledge for each participant. §.§ Internal Validity We have operationalized program comprehension in a multiple-step design, in which participants first had to comprehend a source-code snippet, then compute an input/output task, and finally select the output from four answers. Clearly, program comprehension is a multi-faceted phenomenon for which a variety of operationalizations are possible <cit.>, but our approach ensures that participants genuinely comprehended each snippet. Regarding data collection, we limited the experiment to a maximum of 1 hour (with two breaks) to avoid fatigue effects. Depending on each participant's speed, this led to an unequal number of collected data points (17 of 37 participants completed all 32 snippets). While we could have ended the experiment after, for example, 25 instead of 32 snippets for everyone, we would have lost around 10% of data. Therefore, we chose a small potential bias as a trade-off for a notably larger data set (to gain more statistical power and external validity). We mitigated the threat of an unequal number of snippets by randomizing the presentation order of snippets. Our results indicate a strong influence regarding measures of several programming activities, including the amount of code review and testing, on observed efficacy. This relationship between observed efficacy and time spent with these identified programming activities may be overly emphasized by our experiment design since they are close in nature to our program-comprehension task. §.§ External Validity Due to our focus on internal validity, our presented study is limited regarding its generalizability. While our participant sample covers a wide range of experience and proficiency levels, the presented tasks were limited to algorithms in an object-oriented programming language. Thus, our study on program comprehension may not generalize to all software-engineering activities, including comprehension of large code bases and other programming paradigms. Still, our study pinned down an effect with high internal validity; future work shall replicate, vary, and extend the setup toward external validity. § RELATED WORK Our study is at the intersection between the fields of program comprehension, eye tracking, neuroimaging, and programming expertise, which we each discuss below. Program Comprehension There is a multitude of program-com- prehension studies <cit.>, on which we build by re-using source-code snippets as a foundation for our experiment. In the early years, behavioral or reflective studies were popular <cit.>. Our study shares a similar design with them in terms of snippets and tasks, but our observed measures are more in line with the recent movement toward more objective measurement methods, such as eye tracking or EEG <cit.>, which we discuss next. Eye Tracking Eye tracking enables researchers to objectively observe attention during complex tasks and has increased in popularity, including program comprehension <cit.>. Again, there is a multitude of studies that use various eye-gaze measures <cit.>. We have used recently established eye-gaze measures for differentiation of programmer expertise <cit.>, but we applied these measures to a larger pool of diverse programmers than before and included insights from neural correlates via EEG data. With efficacy, we also arguably used a better way to separate participants (cf. <Ref>). Neural Correlates of Program Comprehension In addition to eye tracking, some studies have employed methods to observe neural correlates of program comprehension. Some have used fMRI <cit.> or fNIRS <cit.>, others have used EEG <cit.>. Our study adopted the measures of cognitive load of Lee et al. (i.e., alpha and beta power) <cit.> and Medeiros et al. (i.e., ratio of theta and alpha) <cit.>, but it differs in terms of the goal and the participant pool. Lee et al. focused on classification between two distinct groups, while Medeiros et al. focused on software metrics <cit.>. By contrast, we focused on the programmer and investigated their efficacy on a continuous distribution, without the need to create two groups. Programming Experience and Expertise There have been multiple waves of program-comprehension research regarding expertise. In the early years, several studies have devised theories of expertise, in particular related to plans <cit.>, mental models <cit.>, strategy <cit.> and knowledge <cit.>. More recently, researchers have focused on visual attention <cit.> and implementation style <cit.>. Our study differs in that we operationalize on a more narrow scope of expertise based on programmer efficacy, rather than measures of knowledge, mental representations, or other intangible factors. We argue our operationalization offers a higher relevance to practical and educational problems as well as a clearer definition over an obfuscated experience measure, because of its better link to task performance. In summary, our experiment is a fusion of established research designs from different fields and is novel by integrating several measures in one experiment. Furthermore, we focused on inviting a diverse participant pool as well as on efficacy as a separating factor (while still collecting a wide array of experience measures). § PERSPECTIVES AND CONCLUSION Measuring Programming Experience For future studies, one major obstacle is to select the correct measure of programming experience. Cognitive psychology has investigated the relationship between experience and expertise in many fields beyond programming <cit.>. One consistent finding is that the length of experience can be part of expertise, as people continue to acquire skill <cit.>, but is not everything <cit.>. In programming, studies from the 1990s identified a similar effect, in that there are two elements at play: the time and learning <cit.>. In our questionnaire, we aimed to capture many aspects of programming experience, including measures of learning, but also regarding content creation and work-time distribution. One theme that consistently showed a strong connection with observed efficacy is the eagerness of learning and inquisitiveness. The number of known programming languages, how much time is spent on mentoring, and on code review are all highly relevant factors. Psychology describes this conscious effort to keep learning as deliberate practice. It also highlights that not necessarily the length of practice, but the intensity and goal-orientation is critical to achieve expertise in a topic <cit.>. The effect of deliberate practice has been observed in programming education and practice, as well. For example, students show higher ability when participating in programming competitions <cit.>. In an industry survey, software engineers shared the notion that a great software engineer is shown by their open-mindedness and eagerness to learn. They considered the ability to learn as more important than technical skill <cit.>. This was later confirmed in an experimental setting, in which company seniority level showed little correlation to actual programming skill <cit.>. Overall, the results of our experiment underlined that the learning component is a highly relevant measure to capture efficacy, which leads us to suggest it as important measure(s) for future experiments. Implications for Research Beyond learning measures, our results indicate that future research could maximize its potential when collecting self-estimation measures. This result is in line with a similar finding by Siegmund et al., which also highlighted self-estimated comparison as a differentiating factor among students <cit.>. In the same vein, Kleinschmager and Hanenberg found self-estimated comparison, at least, as effective as pretests or university grades in its predictive power of efficacy <cit.>. Our shared experience questionnaire provides a proven basis for others to use to capture several dimensions of programming experience. Implications for Industry From a practical perspective, our empirical results provide scientific evidence that technical interviews may be necessary for establishing efficacy, since common measures selected for decisions (e.g., years of experience), have limited to no predictive power. Our questionnaire provides a few promising directions for operationalizing a more effective method for assessing efficacy in industry, such as peer review and mentoring experience, peer comparison, and learning behaviors. However, in high-stake decisions, such as promotion or hiring, the degree to which self-estimation can be reliable remains an open question. Conclusion Experience, expertise, and efficacy are three dimensions of characteristics of programmers that are not well understood. In particular, the relationship between them is unclear. In this paper, we have presented correlates of efficacy in terms of reading order and cognitive load across a wide range of programmers. Commonly used experience measures do not correlate well to observed efficacy. Instead, we underline to use self-estimation and learning eagerness as more accurate measures for programming experience. Despite encouraging results, future work shall explore programmer efficacy in more detail. For example, the link between reading behavior and cognitive load could be explored to understand causalities. Is the more efficient reading strategy of programmers with high efficacy the cause for lower cognitive load, or vice versa? We thank all participants of our study. Furthermore, we thank Julia Hess, Tobias Jungbluth, and Johannes Ihl for their support during data acquisition. Rekrut's work is supported by the German Federal Ministry of Education and Research (01IS12050). Parnin's work is supported by the National Science Foundation under grant number 2045272. Siegmund's work is supported by DFG grant SI 2045/2/̄2. Apel's work is supported by ERC Advanced Grant 101052182 as well as DFG Grant 389792660 as part of TRR 248 – CPEC. ACM-Reference-Format

http://arxiv.org/abs/2303.06851v1

20230313044608

On the Regret of Online Edge Service Hosting

cs.DC

prakash.14191@gmail.com;sriprakash@ee.iitb.ac.in nikhilk@ee.iitb.ac.in sharayum@ee.iitb.ac.in organization=Indian Institute of Technology Bombay, addressline=Powai, postcode=400076, city=Mumbai, country=India We consider the problem of service hosting where a service provider can dynamically rent edge resources via short term contracts to ensure better quality of service to its customers. The service can also be partially hosted at the edge, in which case, customers' requests can be partially served at the edge. The total cost incurred by the system is modeled as a combination of the rent cost, the service cost incurred due to latency in serving customers, and the fetch cost incurred as a result of the bandwidth used to fetch the code/databases of the service from the cloud servers to host the service at the edge. In this paper, we compare multiple hosting policies with regret as a metric, defined as the difference in the cost incurred by the policy and the optimal policy over some time horizon T. In particular we consider the Retro Renting (RR) and Follow The Perturbed Leader (FTPL) policies proposed in the literature and provide performance guarantees on the regret of these policies. We show that under i.i.d stochastic arrivals, RR policy has linear regret while FTPL policy has constant regret. Next, we propose a variant of FTPL, namely Wait then FTPL (W-FTPL), which also has constant regret while demonstrating much better dependence on the fetch cost. We also show that under adversarial arrivals, RR policy has linear regret while both FTPL and W-FTPL have regret O(√(T)) which is order-optimal. § INTRODUCTION Software as a Service (SaaS) instances like online navigation platforms, Video-on-Demand services, etc., have stringent latency constraints in order to provide a good quality of experience to their customers. While most SaaSs use cloud resources, low latency necessitates the use of storage/computational resources at the edge, i.e., close to the end-user. A service is said to be hosted at the edge if the code and databases needed to serve user queries are stored on the edge servers and requests can be served at the edge. In this work, the service can also be partially hosted at the edge, in which case, customers' requests can be partially served at the edge <cit.>. For instance, consider a SaaS application which provides access to news articles on demand. A typical news article includes some text, some images and possibly some embedded videos. One way to partially host such a service is to store the text corresponding to the news articles at the edge and store the images/videos in the cloud. Each user request will then be served partially by the edge. We consider the setting where third-party resources can be rented via short-term contracts to host the service, and the edge hosting status of the SaaS can be dynamically changed over time. If the service is not hosted at the edge, it can be fetched from the cloud servers by incurring a fetch cost. The performance of a hosting policy is a function of the rent cost, the fetch cost, and the quality of experience of the users. We refer to the algorithmic challenge of determining what fraction of the service to host at the edge over time as the service hosting problem. Note that we study the service hosting problem from the perspective of a specific SaaS provider which can rent third-party edge resources to improve their customers' quality of experience by reducing the latency of their service. Novel online service hosting policies with provable performance guarantees have been proposed in <cit.>. The metric of interest in <cit.> is the competitive ratio, defined as the ratio of the cost incurred by an online policy to the cost incurred by an optimal offline policy for the same request arrival sequence. Since the competitive ratio is multiplicative by definition, the absolute value of the difference in the cost incurred by an online policy and the optimal policy can be large even though the competitive ratio of the online policy is close to one. This motivates studying the performance of candidate policies in terms of regret, defined as the difference in the cost incurred by the policy and the optimal policy. Regret is a widely used metric in online learning <cit.>, including recently for the caching problem <cit.>, which is closely related to the service hosting problem. In this work, one of our goals is to design hosting policies with provable guarantees in terms of the regret. Another key dimension in the performance guarantees of online policies is the assumption made on the request arrival process. Commonly studied settings include the stochastic setting and the adversarial setting, and we provide performance guarantees for both settings. §.§ Our Contributions We propose a variant of the FTPL policy called Wait then Follow the Perturbed Leader (W-FTPL). W-FTPL is a randomized policy that takes into account the fetch cost in its decision-making. More specifically, W-FTPL does not fetch the service for an initial wait period which depends on the request arrivals and is an increasing function of the fetch cost. Following the wait period, W-FTPL mimics the FTPL policy. For i.i.d. stochastic request arrivals and the regret metric, we show that RR is sub-optimal while FTPL and W-FTPL are both order-optimal with respect to the time horizon. While the regret of FTPL can increase linearly with the fetch cost, the regret of W-FTPL increases at most polylogarithmically. The improved performance of W-FTPL over FTPL is a consequence of the fact that W-FTPL avoids most of the fetches made by FTPL in the initial time-slots and by the end of the wait period, its estimate of the arrival rate is accurate enough to avoid multiple fetches. Further, we characterize the competitive ratio of FTPL and W-FTPL for the setting where a finite number of partial hosting levels are allowed. We compare these results with the competitive ratio of RR when a single partial hosting level is allowed <cit.> to conclude that FTPL and W-FTPL outperform RR with respect to the competitive ratio. For the adversarial setting, we first characterize a fundamental lower bound on the regret of any online hosting policy. In terms of regret, we then show that RR is strictly suboptimal, while FTPL and W-FTPL have order-optimal performance with respect to time. We show that the competitive ratios of FTPL and W-FTPL are upper bounded by a constant (with respect to time) for any finite number of partial hosting levels. This is a more general setting than the one considered in <cit.>, where the competitive ratio of a variant of RR is characterized when only one partial hosting level is permitted. §.§ Related work Several emerging applications such as Augmented / Virtual Reality (AR/VR), online gaming, the Internet of Things (IoT), and autonomous driving have very stringent latency requirements. In addition, many of these applications are also compute-and-storage heavy. This necessitates the migration of some of the storage and computation requirements of these services from remote cloud servers to the edge. To enable such offloading, several edge computing architectures have been proposed in the literature and their performance has been analyzed extensively, see for example <cit.> and references therein. The service hosting problem has received a lot of attention recently, and various settings have been studied. One approach has been to pose the design problem of where to host which service as a one-shot large-scale optimization problem, see for example <cit.>. Our work differs from this line of work in that we consider an online setting where we adaptively decide when (and what fraction of) service to host at any time, depending on the varying number of requests seen thus far. There have been recent works <cit.> which consider the online service hosting problem, and design adaptive schemes whose performance is characterized in terms of the competitive ratio with respect to an oracle which known the entire request sequence apriori. Our work differs from these in a couple of ways. Firstly, the above mentioned works study the problem from the perspective of an edge resource provider who has to decide which amongst a collection of services to host at each time on a storage-constrained edge server. On the other hand, similar to some other works <cit.>, we study the problem from the perspective of a service provider which needs to decide when to host its application at the edge so as to minimize the overall cost. Secondly, most of the works mentioned above do not consider partial hosting of the service, which is allowed in our framework. While <cit.> does study partial hosting and proposed the Retro-Renting (RR) policy which is shown to have a constant competitive ratio, it only allows one level of partial hosting of the service. On the other hand, we consider a much more general setting where multiple partial service hosting levels are allowed and are able to show that both the FTPL and its variant W-FTPL are able to achieve a constant competitive ratio. While the works mentioned above study competitive ratio as a performance metric, another popular measure of performance for online algorithms is the regret which considers the difference in the cost incurred by an online policy and the optimal (static) policy. In particular, for the closely related caching problem, several policies inspired by the recent advances in online convex optimization have been proposed, including Online Gradient Ascent <cit.>, Online Mirror Descent <cit.>, and Follow the Perturbed Leader (FTPL) <cit.>. In particular, FTPL has been shown to have order-optimal regret for the caching problem in the adversarial setting <cit.>. In this work, we study regret for the RR and FTPL policies for the service hosting problem under both the stochastic and adversarial settings. Since RR is a deterministic policy, its regret performance in the adversarial setting is poor. While FTPL does achieve order-optimal regret (with respect to time), one limitation is that it makes hosting decisions agnostic of the fetch cost. As a result, in some cases, FTPL is prone to fetching and evicting the service multiple times in the initial time-slots when its estimate of the request arrival rate is noisy, thus leading to poor performance. This motivated our design and regret analysis of W-FTPL, a variant of FTPL which takes into account the fetch cost. One recent work which also considers regret as a performance metric is <cit.>, which studied the problem of joint online service hosting and routing and proposed a two time-scale algorithm based on online gradient descent with provably order-optimal regret under adversarial requests. Finally, there are several other approaches which have been used to address the online service hosting problem, including Markov Decision Processes <cit.>, Reinforcement Learning <cit.>, and Multi-Armed Bandits <cit.>. § SYSTEM SETUP We consider a system consisting of a back-end server and an edge-server to serve customers of a service. The back-end server always hosts the service and can serve any requests that are routed to it. In this work, we allow the service to be partially hosted at the edge-server. When the service is partially hosted at the edge, requests are partially served at the edge and partially at the back-end server. Further, the fraction of service hosted at the edge can be changed over time. If the service is not hosted or partially hosted at the edge, parts of the service can be fetched from the back-end server to host at the edge. Sequence of events in each time-slot: We consider a time-slotted system. In each time-slot, we first make the service hosting decision for that time-slot. Following this, requests may arrive and are served at the edge and/or the back-end server. Request arrivals: We consider two types of arrival processes in this paper. - Stochastic arrivals: In this case, request arrivals are i.i.d. across time-slots with mean μ. - Adversarial arrivals: In this case, the request arrival sequence is generated by an oblivious adversary[ The entire request sequence is assumed to be fixed apriori.]. Let r_t denote the number of request arrivals in time-slot t. We consider bounded requests, specifically, 0≤ r_t≤κ, and let R_t=∑_i=1^tr_i denote the cumulative requests observed by the end of time-slot t. In both cases, we assume that there are at most κ arrivals per slot. Costs: We model three types of costs. - Rent cost (𝒞^𝒫_R,t): The system incurs a cost of c units per time-slot to host the entire service at the edge. For hosting f fraction of service, the system incurs a cost of cf units per time-slot. - Service cost (𝒞^𝒫_S,t): This is the cost incurred to use the back-end server to serve (parts of) requests. If f fraction of service is hosted at the edge, the system incurs a service cost of g(f) units per request, where g(f) is a decreasing function of f. We fix g(0)=1, i.e., service cost is one unit when the service is not hosted at the edge. - Fetch cost (𝒞^𝒫_F,t): The system incurs a cost of M> 1 units for each fetch of the entire service from the back-end server to host on the edge-server. For fetching f fraction of service, the system incurs a cost of Mf units. Let ρ_t^𝒫 denote the edge hosting status of the service in time slot t under policy 𝒫. We consider the setting where ρ_t^𝒫∈{α_1,α_2, ⋯ , α_K-1,α_K} and ρ_t^𝒫=α_i implies that α_i fraction of the service is hosted at the edge-server in time-slot t. Note that α_i∈ [0,1] and in particular we consider α_1=0, α_K=1 and α_i<α_j for i<j. It follows that ρ_t^𝒫=1 implies that the entire service is hosted at the edge in time-slot t, and ρ_t^𝒫=0 implies that the service is not hosted at the edge in time-slot t. We make some assumptions about the various system parameters. c≤κ. This assumption is motivated by the fact that if κ < c, it is strictly sub-optimal to host the entire service at the edge, irrespective of the number of arrivals in a time-slot. For a fraction of service α_i, 1≤ i≤ K, α_i+g(α_i)≤ 1. In <cit.>, it is shown that the benefits of partial hosting are limited to the setting where α_i+g(α_i)≤ 1. Motivated by this, we consider the case where α_i+g(α_i)≤ 1 for all candidate values of α_i. For any fraction of service α_i, 1≤ i≤ K, cα_i+g(α_i)r_t≤κ Assumption <ref> follows from Assumptions <ref> and <ref> For stochastic arrivals, μ>0 and for adversarial arrivals, R_T≥ 1. This assumption eliminates degenerate cases where there are no request arrivals in the time-horizon of interest. The total cost incurred in time-slot t by policy 𝒫, denoted by 𝒞^𝒫_t(r_t), is the sum of the rent, service, and fetch costs. It follows that, 𝒞^𝒫_t(r_t) =cρ_t^𝒫+ g(ρ_t^𝒫) r_t+ M (ρ_t^𝒫-ρ_t-1^𝒫)^+. Let r={r_t}_t≥1 denote the request arrival sequence and 𝒞^𝒫(T,r) denote the cumulative cost incurred by policy 𝒫 in time-slots 1 to T. It follows that, 𝒞^𝒫(T,r) = ∑_t=1^T 𝒞^𝒫_t(r_t). Performance metrics: We consider two performance metrics, namely regret and competitive ratio. For i.i.d. stochastic arrivals, the regret of a policy 𝒫, denoted by ℛ^𝒫_S(T), is defined as the difference in the total expected cost incurred by the policy and the oracle static hosting policy. The oracle static hosting policy makes a hosting decision at t=1 using the knowledge of the statistics of the request arrival process but not the entire sample-path. For ease of notation, we define μ=[r_t], μ_i=cα_i+g(α_i)μ. It follows that the expected cost incurred by the optimal static hosting policy in time-slots 1 to T is min_i{cα_iT+g(α_i)μ T+Mα_i}, and therefore, ℛ^𝒫_S(T) = _𝒫,r[𝒞^𝒫(T,r)] -min_i{cα_iT+g(α_i)μ T+Mα_i} =_𝒫,r[𝒞^𝒫(T,r)] -min_i{μ_i T+Mα_i}, ≤_𝒫,r[𝒞^𝒫(T,r)] -min_i{μ_i T}, where _𝒫,r represents the expectation over the randomization in policy 𝒫 and the request arrival sequences r. For i.i.d. stochastic arrivals, the competitive ratio of a policy 𝒫, denoted by σ^𝒫_S(T), is defined as the ratio of the expected total cost incurred by the policy and expected total cost incurred by the optimal static hosting policy as discussed above. It follows that, σ_S^𝒫(T)= _𝒫,r[𝒞^𝒫(T,r)]/min_i{cα_iT+g(α_i)μ T+Mα_i}. For adversarial arrivals, the regret of a policy 𝒫, denoted by ℛ^𝒫_A(T), is defined as the worst case difference in the total expected cost incurred by the policy and the optimal static hosting policy. The expectation is taken over the randomization in policy 𝒫. It follows that for any request sequence r, the total cost incurred by the optimal static hosting policy in time-slots 1 to T is min_i{cα_iT+g(α_i)R_T+Mα_i}, and therefore, ℛ^𝒫_A(T,r) =_𝒫[𝒞^𝒫(T,r)] - min_i{cα_iT+g(α_i)R_T+Mα_i}, ℛ^𝒫_A(T) =sup_r(ℛ^𝒫_A(T,r) ). We have the following relation between the regret induced by a policy under i.i.d stochastic arrivals and adversarial arrivals. For any online policy 𝒫 and stochastic arrivals we have, ℛ_S^𝒫(T)≤ℛ_A^𝒫(T). The proof can be found in Section <ref>. For adversarial arrivals, the competitive ratio of a policy 𝒫, denoted by σ^𝒫_A(T), is defined as the worst case ratio of the expected total cost incurred by the policy and the cost incurred by the offline optimal policy. The optimal offline policy can change the hosting status over time using the knowledge of the entire request process a priori. Let the cost incurred by the optimal offline hosting policy in time-slots 1 to T be denoted by 𝒞^OFF-OPT(T,r). It follows that, σ^𝒫_A(T)=sup_r_𝒫[𝒞^𝒫(T,r)]/𝒞^OFF-OPT(T,r). Goal: The goal is to design online hosting policies with provable performance guarantees with respect to regret and competitive ratio. Notations: We define some notations which will be used in the rest of this paper. Let s=[0 , α_2, α_3, ⋯, 1]', f=[1,g(α_2), g(α_3), ⋯, 0]', and ρ_t^𝒫∈𝒳, where 𝒳 is the collection of all possible one hot vectors in {0,1}^K and thus |𝒳|=K. The position of 1 in the one hot vector ρ_t^𝒫 represents the level of service hosted at the edge in slot t, that is ρ_t^𝒫=⟨ρ_t^𝒫,𝐬⟩. The total cost in slot t as given in (<ref>) can be rewritten using the above notations as follows, 𝒞_t^𝒫(r_t) = c⟨ρ_t^𝒫,s⟩ + r_t ⟨ρ_t^𝒫,f⟩ + M (ρ_t^𝒫-ρ_t-1^𝒫)^+, =⟨ρ_t^𝒫,θ_t⟩ + M (ρ_t^𝒫-ρ_t-1^𝒫)^+, where θ_t= cs+r_tf. For ease of notation, for stochastic arrivals we define μ = [r_t], μ_i=cα_i+g(α_i)μ, Δ_ij=μ_i-μ_j, i^⋆=_i μ_i , Δ_i=μ_i-μ_i^⋆, Δ_min=min_i Δ_i, Δ_max=max_i Δ_i. We summarize important notations used in this paper in  <ref>. § POLICIES Our Policy: Our policy called Wait then Follow the Perturbed Leader (W-FTPL), is a variant of the popular FTPL policy. A formal definition of FTPL is given in Algorithm <ref>. FTPL is a randomized policy and is known to perform well for the traditional caching problem, in fact achieving order-wise optimal regret for adversarial arrivals <cit.>. Under FTPL, in any time slot t, for each i, FTPL considers the cumulative cost that would have been incurred had the system used α_i hosting fraction in the entire duration [1, t -1], and then perturbs this by an appropriately scaled version of a sample of a standard Gaussian random variable. FTPL then hosts the fraction of service for which the perturbed cost is minimum in that slot. We propose a policy called Wait then Follow The Perturbed Leader (W-FTPL). The key idea behind the W-FTPL policy is to not host the service for an initial wait period. This is to reduce the number of fetches made initially when the estimate of the arrivals is noisy. The duration of this wait period is not fixed apriori and is a function of the arrival pattern seen till that time. Following the wait period, W-FTPL mimics the FTPL policy. Refer to Algorithm <ref> for a formal definition of W-FTPL. Let T_s be the time slot, after which W-FTPL starts acting as FTPL. Formally we define T_s=min{t:t<(max_i j(Θ_t+1,i-Θ_t+1,j))^2/κ^2β (log M)^1+δ}, where β>1, δ>0 are constants. Retro-Renting (RR) <cit.>: The RR policy is a deterministic hosting policy, and it either hosts the complete service or hosts nothing in a slot. The key idea behind this policy is to use recent arrival patterns to make hosting decisions. We refer the reader to Algorithm 1 in <cit.> for a formal definition of the RR policy. The performance of RR with respect to the competitive ratio was analyzed in <cit.>. α-RR <cit.> is a variant of RR where one partial level of hosting the service is allowed. The formal definition of α-RR, borrowed from <cit.> is given in Algorithm <ref>. Similar to RR, the key idea behind α-RR policy is to check whether the current hosting status is optimal in the hindsight using recent arrival patterns to make hosting decisions. α-RR checks the optimal offline policy cost for the recent request arrivals and chooses the fraction of service with the minimum cost. For ρ_t^𝒫∈{0,1}, α-RR behaves as RR <cit.>. Therefore we refer α-RR as RR in this paper. In this section we consider different polices we want to analyze with the metrics defined in Section <ref>. §.§ Retro Renting (RR) This policy was proposed in <cit.>. It is a deterministic policy and tries to mimic the optimal off-line policy and it is considered as optimal off-line policy with some delay. The Algorithm <ref> represents the RR policy. The competitive ratio analysis for RR is also there in <cit.>. We analyze the regret of RR in this paper. §.§ Follow The Perturbed Leader (FTPL) This policy already exists in the content caching, on-line learning literature <cit.>, <cit.> where as we adapted it to the service caching setting and analyzed its regret. The Algorithm <ref> with β=0 represents the FTPL policy. Note that it is a randomized algorithm and does not consider M while making decisions and leads to large fetch cost if M is large. §.§ Our policy (W-FTPL) In our policy there will be an initial wait period in which service is not fetched or hosted at edge and all the requests are forwarded to edge. After the wait period we follow the FTPL policy. So we named our policy as Wait then FTPL (W-FTPL) which is given in Algorithm <ref>. Note that the initial wait period depends on M, which helps in reducing the fetch cost when M is large compared to FTPL.

http://arxiv.org/abs/2303.07230v1

20230313160414

Systematic Evaluation of Deep Learning Models for Failure Prediction

cs.SE

Systematic Evaluation of Deep Learning Models for Failure Prediction This work was supported by the Canada Research Chair and Discovery Grant programs of the Natural Sciences and Engineering Research Council of Canada (NSERC), by a University of Luxembourg’s joint research program grant, and by European Union's Horizon 2020 Research and Innovation Programme under grant agreement No. 957254 (COSMOS). The experiments conducted in this work were enabled in part by Digital Alliance of Canada (alliancecan.ca). Fatemeh Hadadi Joshua H. Dawes Donghwan Shin Domenico Bianculli Lionel Briand ================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================== With the increasing complexity and scope of software systems, their dependability is crucial. The analysis of log data recorded during system execution can enable engineers to automatically predict failures at run time. Several Machine Learning (ML) techniques, including traditional ML and Deep Learning (DL), have been proposed to automate such tasks. However, current empirical studies are limited in terms of covering all main DL types—Recurrent Neural Network (RNN), Convolutional Neural network (CNN), and transformer—as well as examining them on a wide range of diverse datasets. In this paper, we aim to address these issues by systematically investigating the combination of log data embedding strategies and DL types for failure prediction. To that end, we propose a modular architecture to accommodate various configurations of embedding strategies and DL-based encoders. To further investigate how dataset characteristics such as dataset size and failure percentage affect model accuracy, we synthesised 360 datasets, with varying characteristics, for three distinct system behavioral models, based on a systematic and automated generation approach. Using the F1 score metric, our results show that the best overall performing configuration is a CNN-based encoder with Logkey2vec. Additionally, we provide specific dataset conditions, namely a dataset size >350 or a failure percentage >7.5%, under which this configuration demonstrates high accuracy for failure prediction. Keywords: Logs, Failure Prediction, Deep Learning, Embedding Strategy, Synthesised Data Generation, Systematic Evaluation § INTRODUCTION As software systems continue to increase in complexity and scope, reliability and availability play a critical role in quality assurance and software maintenance <cit.>. During runtime, software systems often record log data about their execution, designed to help engineers monitor the system’s behavior <cit.>. One important quality assurance activity is to predict failures at run time based on log analysis, as early as possible before they occur, to enable corrective actions and minimise the risk of system disruptions <cit.>. However, software systems typically generate a vast quantity of log data which makes manual analysis error-prone and extremely time-consuming. Therefore, a number of automatic log analysis methods, particularly for failure prediction <cit.> and anomaly detection <cit.>, have been proposed over the past few years. Machine Learning (ML) has played a key role in automatic log analysis, from Traditional ML methods (e.g., Random Forest (RF) <cit.>, Support Vector Machine (SVM) <cit.>, Gradient Boosting (GB) <cit.>) to Deep Learning (DL) methods (e.g., DeepLog <cit.>, LogRobust <cit.>, LogBERT <cit.>) relying on various DL network architectures, including Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and transformers <cit.>. Although several studies have explored the use of DL models with various log sequence embedding strategies <cit.>, they have been limited in terms of evaluating the three main types of DL networks—RNN, CNN, and transformer—combined with different embedding strategies; for instance, two studies by <cit.> and <cit.> included CNN-based models but did not cover transformer-based models. Moreover, previously studied models were often applied to a limited number of available datasets, which severely limited the generalizability of results <cit.>. Indeed, because these few datasets exhibit a limited variety of characteristics, studying the robustness and generalizability of DL models, along with their embedding strategies, is unlikely to yield practical guidelines. In this paper, we aim to systematically investigate the combination of the main DL architectures and embedding strategies, based on datasets whose main characteristics (e.g., dataset size and failure percentage) are controlled. To achieve this, we first introduce a modular architecture for failure prediction, where alternative log embedding strategies and DL models can be easily applied. The architecture consists of two major steps: an embedding step that converts input logs into log embedding vectors followed by a classification step that predicts failures by processing the embedding vectors using encoders that are configured by different DL models, called DL encoders. In the embedding step, two alternative strategies, i.e., the pretrained BERT <cit.> and trainable Logkey2vec <cit.>, are considered. In the classification step, four types of DL models, including LSTM <cit.>, BiLSTM<cit.>, CNN<cit.>, and transformer <cit.>, with an attention mechanism <cit.> for BiLSTM and transformer, are used. Also, to address the issue of the limited availability of adequate datasets, we designed a rigorous approach for generating synthesised data relying on behavioral models built by applying model inference algorithms <cit.> to available system logs. When synthesizing data, we control key dataset characteristics such as the size of the dataset and the percentage of failures. Additionally, we define patterns that are associated with system failures and are used to classify logs for the failure prediction task. The goal is to associate failures with complex patterns that are challenging for failure prediction models. Further, based on our study, we investigated how the dataset characteristics determine the accuracy of model predictions and then derive practical guidelines. Our empirical results conclude that the best model includes the improved CNN-based encoder with Logkey2vec as an embedding strategy. Using a wide variety of datasets showed that this combination is also very accurate when certain conditions are met in terms of dataset size and failure percentage. To summarise, the main contributions of this paper are: * A large-scale, systematic investigation of the application of various DL encoders–LSTM-, BiLSTM-, CNN-, and transformer-based– and embedding strategies–BERT <cit.> and Logkey2vec <cit.>– for failure prediction modeling * A systematic and automated approach to synthesise log data, with a focus on experimentation in the area of failure prediction, to enable the control of key data set characteristics while avoiding any other form of bias. * Practical guidelines for using DL-based failure prediction models depending on dataset characteristics such as dataset size and failure rates. * A publicly available replication package, containing the implementation, generated datasets with behavioural models, and results. The rest of the paper is organised as follows. Section <ref> presents the basic definitions and concepts that will be used throughout the paper. Section <ref> illustrates related work. Section <ref> describes the architecture of our failure predictor with its different configuration options. Section <ref> describes our research questions, empirical methodology, and synthetic log data generation. Section <ref> reports empirical results. Section <ref> discusses the implications of the results. Section <ref> concludes the paper and suggests future directions for research and improvements. § BACKGROUND In this section, we provide background information on the main concepts and techniques that will be used throughout the paper. First, we briefly introduce the concepts related to finite state automata (FSA) and regular expressions in  <ref> and execution logs in  <ref>. We then describe two important log analysis tasks (anomaly detection and failure prediction) in  <ref> and further review machine-learning (ML)-based approaches for performing such tasks in  <ref>. We conclude by providing an overview of embedding strategies for log-based analyses in  <ref>. §.§ Finite State Automata and Regular Expressions A deterministic FSA is a tuple ℳ = ⟨ Q, A, q_0, Σ, δ⟩, where Q is a finite set of states, A ⊆ Q is the set of accepting states, q_0 ∈ Q is the starting state, Σ is the alphabet of the automaton, and δ Q ×Σ→ Q is the transition function. The extended transition function δ^* : Q ×Σ^* → Q, where Σ^* is the set of strings over Σ, is defined as follows: * For every q ∈ Q, δ^*(q,ϵ) = q, where ϵ represents the empty string; * For every q ∈ Q, every y ∈Σ^*, and every σ∈Σ, δ^*(q, yσ) = δ(δ^*(q, y), σ). Let x ∈Σ^*; the string x is accepted by ℳ if δ^*(q_0, x) ∈ A and is rejected by ℳ, otherwise. The language accepted by an FSA ℳ is denoted by ℒ(ℳ) and is defined as the set of strings that are accepted by ℳ; more formally, ℒ(ℳ) = {w |δ^*(q_0, w) ∈ A }. A language accepted by an FSA is called a regular language. Regular languages can also be defined using regular expressions; given a regular expression r we denote by ℒ(r) the language it represents. A regular expression r over an alphabet Σ is a string containing symbols from Σ and special meta-symbols like “|” (union or alternation), “.” (concatenation), and “*” (Kleene closure or star), defined recursively using the following rules: * ∅ is a regular expression denoting the empty language ℒ(∅)=∅; * For every a ∈Σ, a is a regular expression corresponding to the language ℒ(a)={a}; * If s and t are regular expressions, then r= s | t and r=s.t (or r=st) are regular expressions denoting, respectively, the union and the concatenation of ℒ(s) and ℒ(t); * If s is a regular expression, then r=s^* is a regular expression denoting the Kleene closure of ℒ(s). §.§ Logs In general, a log is a sequence of log messages generated by logging statements (e.g., , ) in the source code <cit.>. A log message is textual data composed of a header and content <cit.>. In practice, the logging framework determines the header (e.g., ) while the content is designed by developers and is composed of static and dynamic parts. The static parts are the fixed text written by the developers in the logging statement (e.g., to describe a system event), while the dynamic parts are determined by expressions (involving program variables) evaluated at runtime. For instance, let us consider the execution of the log printing statement ; during the execution, assuming variable is equal to , the log message is printed. In this case, is the static part while is the dynamic part, which changes depending on the value of at run time. A log template (also called event template or log key) is an abstraction of the log message content, in which dynamic parts are masked with a special symbol (e.g., ); for example, the log template corresponding to the above log message is . Often, each unique log template is identified by an ID number for faster analysis and efficient data storage. A log sequence is a fragment of a log, i.e., a sequence of log messages contained in a log; in some cases, it is convenient to abstract log sequences by replacing the log messages with their log templates. Log sequences are obtained by partitioning logs based on either log message identifiers (e.g., session IDs) or log timestamps (e.g., by extracting consecutive log messages using a fixed/sliding window). For a log sequence l, |l| indicates the length of the log sequence, i.e., the number of elements (either log templates or log messages), not necessarily unique, inside the sequence. Figure <ref> shows an example summarizing the aforementioned concepts. On the left side, the first three log messages are partitioned (using a fixed window of size three) to create a log sequence. The first message in the log sequence (LogMessage1) is . It is decomposed into the header and the content . The log template for the content is ; the dynamic parts are and . §.§ Log Analysis Tasks In the area of log analysis, several major tasks for reliability engineering such as anomaly detection and failure prediction have been automated <cit.>. §.§.§ Anomaly Detection Anomaly detection is the task of identifying anomalous patterns in log data that do not conform to expected system behaviors <cit.>, indicating possible errors, faults, or failures in software systems. To automate the task of anomaly detection, log data is often partitioned into smaller log sequences. This partitioning is typically based on log identifiers (e.g., session_ID or block_ID), which correlate log messages within a series of operations; alternatively, when log identifiers are not available, timestamp-based fixed/sliding windows are also used. Labeling of partitions is then required, each partition usually being labeled as an anomaly either when an error, unknown, or failure message appears in it or when the corresponding log identifier is marked as anomalous. Otherwise, it is labeled as normal. §.§.§ Failure Prediction Failure prediction attempts to proactively generate early alerts to prevent failures, which often lead to unrecoverable outages <cit.>. Similar to anomaly detection, log data is often partitioned into sequences using log identifiers. Partitioned log sequences are labeled as failures according to mechanisms that are specific to the application being monitored. Like for anomalies, failures can be associated with complex patterns in log sequences. §.§ DL Techniques in Log Analysis In recent years, a variety of deep learning (DL) techniques have been applied to log analysis, and more specifically to failure prediction and anomaly detection. Compared to traditional ML techniques such as Random Forests (RF) and K-nearest Neighbours (KNN), DL techniques incrementally learn high-level features from data, removing complex feature extraction activities based on domain expertise. According to <cit.>, there are three main categories of DL approaches in log analysis: (1) Recurrent Neural network (RNN), (2) Convolutional Neural Network (CNN), and (3) transformer. In each category, different variations can be adopted; for instance, Long Short-Term Memory networks (LSTM) and Bidirectional Long Short-Term Memory networks (BiLSTM), which fall into the RNN category, have been repeatedly used for anomaly detection and failure prediction <cit.>. We now explain the major features of each category as well as their variations. §.§.§ RNN LSTM <cit.> is an RNN-based model commonly used in both anomaly detection and failure prediction <cit.>. An LSTM network consists of multiple units, each of which is composed of a cell, an input gate, an output gate <cit.>, and a forget gate <cit.>. An LSTM-based model reads an input sequence (x_1, …, x_n) and produces a corresponding sequence (y_1, …, y_n) with the same length. At each time step t>1, an LSTM unit reads the input x_t as well as the previous hidden state h_t-1 and the previous memory c_t-1 to compute the hidden state h_t. The hidden state is employed to produce an output at each step. The memory cell c_t is updated at each time step t by partially forgetting old, irrelevant information and accepting new input information. The forget gate f_t is employed to control the amount of information to be removed from the previous context (i.e., c_t-1) in the memory cell c_t. As a recurrent network, an LSTM shares the same parameters across all steps, which reduces the total number of parameters to learn. Learning is achieved by minimizing the error between the actual output and the predicted output. Moreover, to improve the regularization of an LSTM-based model, a dropout layer is applied between LSTM layers. It randomly drops some connections between memory cells by masking their value. LSTM-based models have shown significant performance in several studies in log-based failure prediction and anomaly detection  <cit.>. BiLSTM is an extension of the traditional LSTM <cit.>. However, BiLSTM reads the sequence in both directions, enabling it to comprehend the relationships between the previous and the upcoming inputs. To make this possible, a BiLSTM network is composed of two layers of LSTM nodes, whereby each of these layers learns from the input sequence in the opposite direction. At time step t, the output h_t is calculated by concatenating h_t^f (the hidden states in a forward pass) and h_t^b (the hidden states in a backward pass). By allowing this bi-directional computation, BiLSTM is able to capture complex dependencies and produce more accurate predictions. The BiLSTM-based model has achieved accurate results for anomaly detection <cit.>. §.§.§ CNN CNN is a neural network primarily employed for image recognition <cit.>. It has a unique architecture designed to handle 2D and 3D input data such as images and matrices. A CNN leverages convolutional layers to perform feature extraction and pooling layers to downsample the input. The 1D convolutional layer uses a set of filters to perform convolution operation with the 2D input data to produce a set of feature maps (CNN layer output). According to <cit.>, let w ∈𝑅^k × d be a filter which is applied to a window of k elements in a d-dimension input log sequence, and let x_i represent the i-th elements in the sequence. A feature c_i ∈𝑅 is calculated as c_i = σ(w . x_i:i+k-1+b), where σ is the activation function (i.e., ReLu), x_i:i+k-1 represents the concatenation of elements {x_i, x_i+1, ..., x_i+k-1}, and b ∈𝑅 denotes a bias term. After this filter is applied to each window in the sequence ({x_1:k, x_2:k, ..., x_n-k+1:n}), a feature map c = [c_1, c_3, ..., c_n-k+1] is produced, where c ∈𝑅^n-k+1. Parameter k represents the kernel size; it is as an important parameter of the operation. Note that there is no padding added to the input sequence, leading to feature maps smaller than the input sequence. Padding is a technique employed to add zeros to the beginning and/or end of the sequence; it allows for more space for the filter to cover, controlling the size of output feature maps. Padding is commonly used so that the output feature map has the same length as the input sequence <cit.>. The pooling layer reduces the spatial dimensions of the feature maps extracted by the convolutional layer and simplifies the computational complexity of the network. Recently, CNNs have shown high-accuracy performance in anomaly detection <cit.>. §.§.§ Transformer The transformer is a type of neural network architecture designed for natural language processing tasks, introduced by <cit.>. The main innovation of transformers is the self-attention mechanism. More important parts of the input receive higher attention, which facilitates learning the contextual relationships from input data. This is implemented by calculating a weight for each input element, which represents the importance of that element with respect to the adjacent elements. Hence, a model with self-attention (not necessarily a transformer) can capture long-range dependencies in the input. Since the transformers do not process inputs sequentially like LSTM, positional encoding is needed. Positional encoding vectors are fixed-size, added to the input to provide information about the position of each element in the input sequence. Further, a transformer involves a stack of multiple transformer blocks. Each block contains a self-attention layer and a feed-forward neural network layer. In the self-attention layer, the model computes attention scores (weights) for each element, allowing it to capture the relationship between all input elements. The feed-forward layer is used to transform the representation learned by the self-attention layer into a new representation entering the next transformer block. In the area of log analysis, transformers have been recently applied in a few studies on anomaly detection <cit.>, showing outstanding performance. §.§ Log Sequence Embedding Strategies When analyzing log sequences, the textual data of log sequences' elements must be converted into a vector representation that is understandable by a machine; such a conversion is called the log sequence embedding. Generally, there are two main approaches for doing this: 1) based on count vectors <cit.> or ID numbers of the sequence elements, or 2) based on the contextual information of sequence elements. Here, we cover one widely used example for each case in the following sections. §.§.§ Logkey2vec There are many studies that have achieved high accuracy results by using log embedding strategies that rely on the ID numbers or count vectors of log sequence elements <cit.>. Advantages include the speed of processing and model simplicity since text preprocessing (e.g., tokenization) is not required. Similar to methods like word2vec <cit.>, which assigns a unique vector to each different word based on a context mapping table, Logkey2vec maps each unique log template ID to a vector representation. While Word2vec is a pre-trained tool, Logkey2vec is a trainable layer implemented inside a neural network. It relies on a matrix called “codebook”, where the number of rows is the vocabulary size and the number of columns is the embedding vector size of each log template ID. The embedding vectors are first initialised by random numbers and are improved through backpropagation during training. For a log sequence, Logkey2vec computes the embedding vector of each log template based on its log template ID; each row of the matrix represents the whole log sequence. §.§.§ BERT In the past few years, Bidirectional Encoder Representations from Transformers (BERT) has provided significant improvements in the semantic embedding of textual information by taking the contextual information of text into account. It has been used in a few studies in log-based anomaly detection <cit.>. This model fares better than the other pretrained transformer-based models: GPT2 <cit.> and RoBERTa <cit.> in log sequence embedding <cit.>. The pre-trained BERT base model <cit.> provides the embedding matrix of log sequences where each row is the representation vector of its corresponding log template inside the sequence. The BERT model is applied to each log template separately and then the representation is aggregated inside a matrix. To embed the information of a log template into a 768-sized vector, the BERT model first tokenizes the log template text. BERT tokenizer uses WordPiece <cit.>, which is able to handle out-of-vocabulary (OOV) words to reduce the vocabulary size. Further, the tokens are fed to the 12 layers of BERT's transformer encoder. After obtaining the output vectors of a log template's tokens, the log template embedding is calculated by getting the average of output vectors. This process is repeated for all the log templates inside the log sequence to create an n × 768 matrix representation where n is the size of the log sequence. § RELATED WORK There are several papers reporting empirical studies of different DL-based methods for log-based anomaly detection and failure prediction. In our review, we include studies that covered more than one DL model, possibly based on the same DL type; given our focus, non-DL models such as RF, SVM, and clustering are not included. The main studies are summarised in Table <ref>. Column “DL Type(s)” indicates the type of DL network covered in each paper. We indicate the Log Sequence Embedding (LSE) strategies, introduced in  <ref>, in the next column; notice there are a few models not using LSE, such as DeepLog <cit.>. Column “Dataset(s)” indicates which datasets (whether existing datasets or synthesised ones) were used in the studies. Column “DataSet Char Control” indicates whether the dataset characteristics were controlled during the experiment and lists such characteristics. In the last column, the labeling scheme indicates the applied method(s) for log partitioning, as mentioned in  <ref>, based on either a log identifier or timestamp (represented by L and T, respectively). We now briefly explain the papers with the aim to motivate our study and highlight the differences. We note that, unless we mention it, LSE strategies are implemented specifically for one DL model (combinations are not explored). Indeed, many of the reported techniques tend to investigate one such combination or simply do not rely on any embeddign strategy. The studies are listed in chronological order. <cit.> (2018) introduced CNN for anomaly detection as well as the Logkey2vec embedding strategy (see  <ref>). They compared it to LSTM and MLP networks, also relying on the Logkey2vec embedding strategy. <cit.> (2019) developed LogAnomaly, an LSTM-based model, using their proposed embedding strategy, Template2Vec (a log-specific variant of Word2Vec). <cit.> (2020) introduced a state-of-the-art LSTM-based model, Aarohi, for predicting failures using log data and compared it to two LSTM-based models Dash <cit.> and DeepLog <cit.> with the aim of online prediction efficiency. None of these technqiues rely on a log embeddign strategy. The first study considering transformers in their DL comparison is by <cit.> (2020), featuring three DL models: HitAnomaly (transformer-based), LogRobust <cit.> (BiLSTM-based), and DeepLog (LSTM-based). HitAnomaly utilises transformer blocks (see  <ref>) as part of its LSE strategy, called Log Encoder. LogRobust employed TF-IDF technique <cit.> while DeepLog did not utilise any LSE strategy. The authors also controlled dataset characteristics by manipulating the unstable log ratios. <cit.> (2021) proposed the GRU-based <cit.> PLELog and compared it to LogRobust and DeepLog. PLELog used the TF-IDF technique, similar to LogRobust. <cit.> (2021) proposed a transformer-based model, LogBERT, and compared its performance with two LSTM-based models, LogAnomaly and DeepLog. LogBERT uses an Embedding Matrix for its embedding strategy, which is similar to Logkey2vec. <cit.> (2022) evaluated their proposed transformer-based model, Neurallog, against LogRobust (BiLSTM-base) and DeepLog (LSTM-based). The LSE strategies for the models were a pre-trained BERT (see  <ref>) for Neurallog and Log2Vec <cit.> for DeepLog. Finally, <cit.> (2022) conducted a comprehensive evaluation of several DL models including LSTM-based models such as DeepLog and LogAnomaly, GRU-based model PLELog, BiLSTM-based model LogRobust, and CNN. The study focused on various aspects including data selection, data partitioning, class distribution, data noise, and early detection ability. Datasets. Due to security concerns, in many of the works in the literature, the data sources are unavailable such as Clay-HPC (Clay high-performance computing (HPC) systems). Studies that used available datasets are limited to the following: Hadoop Distributed File System (HDFS) collected in 2009, and three HPC datasets, BGL, Spirit, and Thunderbird, collected between 2004 and 2006. Besides, there is the OpenStack dataset (2017) created by injecting a limited number of anomalies at different execution points which not only does limit the diversity of anomalies, it may not accurately reflect real-world scenarios. Motivating this work. Although the above studies used various DL models (13 models), none of them covers all the main DL types, RNN, CNN, and transformer. Moreover, the LSE strategies are rarely evaluated across all the DL networks, keeping them inside which original model they were proposed. Furthermore, the investigation of various combinations of LSE strategies and DL types has been limited in existing studies; many popular models tend to adopt a fixed combination of these two elements. From table <ref> we can also see that recent studies are incomplete evaluations of DL models with a limited number of datasets that are not recent. These remarks highlight the need for a systematic and comprehensive study that 1) explores various DL types as well as their combinations with different LSE strategies; and 2) examines how various dataset characteristics affect performance. § FAILURE PREDICTION ARCHITECTURE To systematically evaluate various deep learning models for failure prediction, we rely on a generic architecture that can use different modules with respect to the embedding strategy and the DL encoder. This modular architecture will allow us to easily change individual modules corresponding to the various DL techniques and log sequence embedding strategies. Figure <ref> depicts the modular architecture. The architecture consists of two main steps, embedding and classification, designed to adopt different embedding as well as DL techniques, respectively. We note that preprocessing is not required in this architecture since log sequences are based on log templates which are already preprocessed from log messages. In the embedding step, log sequences are given as input, and each log sequence is in the form (x_1,x_2,...,x_i,..,x_n), where x_i is a log template ID and n is the length of the log sequence. An embedding technique (e.g., BERT) converts each x_i to a θ-dimensional vector representing the semantics of x_i, where θ is the size of log sequence embedding. Then each log sequence forms a matrix X ∈𝑅^n×θ. Different log sequence embedding strategies can be applied; more information is provided in  <ref>. In the classification step, the embedding matrix is processed to predict whether the given log sequence is failure or not. A DL model, as an encoder Φ encodes the matrix X into a feature vector z = Φ(X) ∈𝑅^m, where m is the number of features, which is a variable depending on the architecture of Φ. Different DL encoders can be applied; more information is provided in  <ref>. Similar to related studies <cit.>, the output feature vector z is then fed to a feed-forward network (FFN) and softmax classifier to create a vector of size d (d=2), indicating the prediction of the input unit label. More specifically, the FFN activation function is rectified linear unit (ReLu), and the output vector of the FNN r is defined as r = max(0, z W_1+b_1) where W_1 ∈𝑅^m× d_f and b_1 ∈𝑅^d_f are a trainable parameter, and d_f is the dimensionality of FNN. Further, the calculation of the softmax classifier is as follows. o = rW_2+b_2 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(o_p) = exp(o_p)/∑_jexp(o_j) where W_2∈𝑅^d_f × d and b_2 ∈𝑅^d are trainable parameters to convert r to t ∈𝑅^d before applying softmax; o_p represents the p-th component in the o vector, and exp is the exponential function. After obtaining the softmax values, the position with the highest value determines the label of the input log sequence. Overall, the combination of an embedding strategy and a DL encoder forms a language model that takes textual data as input and transforms it into a probability distribution <cit.>. This language model handles the log templates as well as learning the language of failure patterns to predict the label of sequences. To train the above architecture, a number of hyper-parameters should be set such as the choice of the optimizer, loss function, learning rate, input size (for some deep learning models), batch size, and the number of epochs. Tuning these hyper-parameters is highly suggested as it increases the chances of achieving the best failure prediction accuracy. Section <ref> will detail the training and hyper-parameter tuning in our experiments. After the model is trained, it is evaluated with a test log split from the dataset with stratified sampling. We used stratified sampling to keep the same distribution of failure log sequences as the original dataset. Similar to training data, the embedding step transforms the test log sequences into embedding matrices. The matrices are then fed to the trained DL encoder to predict whether log sequences are normal or not. §.§ Embedding Strategies. While the modular architecture can accommodate various log sequence embedding options, we only selected two in order to limit the number of possible models resulting from the architecture, given our experimental constraints[More details are provided in  <ref>]. Our selection criteria aimed to include both trainable and pretrained options, as well as an advanced strategy based on transformers. Hence, given the description provided in  <ref>, the “Embedding step” of our base architecture is instantiated using two strategies: BERT or Logkey2vec. The BERT sequence embedding technique is pretrained and was recently adopted for log sequences <cit.> while Logkey2vec is trainable and was proposed earlier <cit.>. Note that these two techniques were not compared in the same study before, according to Table <ref>. BERT. The maximum number of input tokens for BERT (see  <ref>) is 512 tokens. This limit does not constitute a problem in this work since the log templates in our datasets are relatively short and the total number of tokens in each log template is always less than 512. Even if log templates were longer than 512, there are related studies suggesting approaches to use BERT accordingly <cit.>. Each layer of the transformer encoder contains multi-head attention sub-layers and FFNs to compute a context-aware embedding vector (θ=768) for each token. This process is repeated for all the log templates inside the log sequence to create a matrix representation of size n × 768, where n is the length of the input log sequence. Logkey2vec. For Logkey2vec (see  <ref>), we set the embedding size to 768, similar to BERT for better comparison. The vocabulary size is a parameter of Logkey2vec that is going to be set during the experiments. §.§ Deep Learning Encoder In this section, we illustrate the main features of the four DL encoders that can be used in the “Classification step” when instantiating our base architecture. We selected four encoders (LSTM-, BiLSTM-, CNN-, and transformer-based) because they cover the main DL types in addition to the recently emerged mechanism of attention, as mentioned in  <ref>. LSTM-based. This DL model is inspired by the LSTM architecture suggested by related works, including DeepLog <cit.>, Aarohi <cit.>, and Dash <cit.>. The model has one hidden layer of LSTM with 128 nodes and ReLu activation. A Dropout with a rate of 0.1 is applied to make the model generalise better. The output of the model is a feature vector of size 128. BiLSTM-based. The model has an architecture similar to LogRobust, which was proposed for anomaly detection. Due to its RNN-based architecture, its output is a feature vector with the same size as the input log sequence length <cit.>. CNN-based. The CNN architecture is a variation of the convolutional design for the CNN-based anomaly detection mode <cit.>. Based on our preliminary experimental results, 20 filters, instead of one, for each of the three 1D convolutions (see  <ref>) are used in parallel to capture relationships between log templates at different distances. To ensure that feature maps of each convolution have the same dimension as the input, the padding technique is used. Hence, the length of the output feature vector is the product of the number of filters (20), the number of convolutions (3), and the input size of the log sequence. Transformer-based. Our architecture of the transformer model is inspired by recent work in anomaly detection <cit.>. The model is composed of two main parts: positional embedding and transformer blocks. One transformer block is adopted after positional embedding, set similarly to a recent study <cit.>. After global average pooling, the output matrix is mapped into one feature vector the same size as the log template embedding θ = 768, previously explained in  <ref>. § EMPIRICAL STUDY DESIGN §.§ Research Questions The goal of this study is to systematically evaluate the performance of failure predictors, by instantiating our base architecture with different combinations of DL encoders and log sequence embedding strategies, for various datasets with different characteristics. The ultimate goal is to rely on such analyses to introduce practical guidelines to select the right failure prediction model based on the characteristics of a given dataset. To achieve this, we investigate the following research questions: * What is the impact of different DL encoders on failure prediction accuracy? * What is the impact of different log sequence embedding strategies on failure prediction accuracy? * What is the impact of different dataset characteristics on failure prediction accuracy? RQ1 and RQ2 investigate how failure prediction accuracy varies across DL encoders and embedding strategies reported in the literature. Most of them have been evaluated in isolation or with respect to a few alternatives, often using ad-hoc benchmarks (see  <ref> for a detailed comparison). To address this, we comprehensively consider all variations of our base architecture, obtained by combining four DN encoders (LSTM, CNN, transformer, and BiLSTM) with two log sequence embedding strategies (Logkey2vec and BERT) that have been widely used in failure prediction and anomaly detection. Furthermore, we systematically vary the characteristics of the input datasets in terms of the number of log sequences, the length of log sequences, and the proportion of normal log sequences. The answer to these questions is expected to lead to practical guidelines for choosing the best failure prediction model given a dataset with certain characteristics. RQ3 additionally investigates the impact of the input dataset characteristics on failure prediction accuracy with a focus on the best DL encoder and log sequence embedding strategy found in RQ1 and RQ2. The answer to this question will help us better understand under which conditions the combination of the best DL encoder and log sequence embedding strategy works sufficiently well for practical use, possibly leading to practical guidelines to best prepare input datasets for increasing failure prediction accuracy. §.§ Methodology As discussed in  <ref>, we can instantiate the base architecture for failure prediction with different DL encoders and log sequence embedding strategies. To answer RQ1 and RQ2, we train different configurations of the base architecture while systematically varying training datasets' characteristics (e.g., size and failure types). Then, we evaluate the relative performance of the configurations in terms of failure prediction accuracy, using test datasets having the same characteristics but not used during training. To answer RQ3, we analyze the results of the best configuration (i.e., the best DL encoder and log sequence embedding strategy) identified in RQ1 and RQ2. Specifically, we build regression trees <cit.> to automatically infer conditions describing how the failure prediction accuracy of the best configuration varies according to the dataset characteristics. §.§.§ Log Sequence Embedding Strategies and DL Encoders As for different log sequence embedding strategies, we consider Logkey2vec and BERT, which have shown to be accurate in the literature as discussed in  <ref>. For BERT, we use the original pretrained model with 512 input tokens, where each token is represented as a 768-dimensional embedding vector. For Logkey2vec, we set the size of an embedding vector to be the same as BERT for a fair comparison. Also, Logkey2vec has an additional parameter: vocabulary size. We set it to 200, which is large enough for all datasets used in our evaluation. As for different DL encoders in RQ1 and RQ2, we consider four encoders that have been previously used in failure prediction and anomaly detection. We configured the encoders based on the recommendations reported in the literature (see  <ref> for further details). §.§.§ Datasets with Different Characteristics As for the characteristics of datasets, we consider four factors that are expected to affect failure prediction performance: (1) dataset size (i.e., the number of logs in the dataset), (2) log sequence length (LSL) (i.e., the length of a log sequence in the dataset), (3) failure percentage (i.e., the percentage of log sequences with failure patterns in the dataset), and (4) failure pattern type (i.e., types of failures). The dataset size is important to investigate to assess the training efficiency of different DL models. To consider a wide range of dataset sizes while keeping the number of all combinations of the four factors tractable, we consider six levels that cover the range of real-world dataset sizes reported in a recent study <cit.>: 200, 500, 1000, 5000, 10000, and 50000. The LSL could affect failure prediction since a failure pattern that spans a longer log might be more difficult to predict correctly. Similar to observed lengths in real-world log sequences across publicly available datasets <cit.>, we vary the maximum[We set the maximum LSL for to simplify control.] LSL across five levels: 20, 50, 100, 500, and 1000. The failure percentage determines the balance of classes in a dataset, which may affect the performance of DL models <cit.>. The training dataset is perfectly balanced at 50%. However, the failure percentage can be much less than 50% in practice, as observed in real-world datasets <cit.>. Therefore, we vary the failure percentage across six levels: 5%, 10%, 20%, 30%, 40%, and 50%. Regarding failure patterns, we aim to consider patterns with potential differences in terms of learning effectiveness. However, failure patterns defined in previous studies are too simple; for example, <cit.> consider a specific, consecutive sequence of problematic log templates, called a “failure chain”. But in practice, not all problematic log templates appear consecutively in a log. To address this, we use regular expressions to define failure patterns, allowing non-consecutive occurrences of problematic log templates. For example, a failure pattern “x(y|z)” indicates a pattern composed of two consecutive templates that starts with template x and ends with either template y or template z. In addition, we consider two types of failure patterns (in the form of regular expressions), Type-F and Type-I, depending on the cardinality of languages accepted by the regular expressions (finite and infinite, respectively). This is because, if the cardinality of the language is finite, DL models might memorise (almost) all the finite instances (i.e., sequences of log templates) instead of learning the failure pattern. For example, the language defined by the regular expression “x(y|z)” is finite since there are only two template sequences (i.e., xy and xz) matching the expression “x(y|z). In this case, the two template sequences might appear in the training set, making it straightforward for DL models to simply memorise them. On the contrary, the language defined by the regular expression “x^*(y|z)” is infinite due to infinite template sequences that can match the sub-expression `x^*'; therefore simply memorising some of the infinitely many sequences matching “x^*(y|z)” would not be enough to achieve high failure prediction accuracy. To sum up, we consider 360 combinations (six dataset sizes, five maximum LSLs, six failure percentages, and two failure pattern types) in our evaluation. However, we could not use publicly available datasets for our experiments due to the following reasons. First, although <cit.> reported several datasets in their survey paper, they are mostly labeled based on the occurrence of error messages (e.g., log messages with the level of ERROR) instead of considering failure patterns (e.g., sequences of certain messages). Furthermore, there are no publicly available datasets covering all the combinations of the four factors defined above, making it impossible to thoroughly investigate their impact on failure prediction. To address this issue, we present a novel approach for synthetic log data generation in  <ref>. §.§.§ Failure Predictor Training and Testing We split each artificially-generated dataset into two disjoint sets, a training set and a test set, with a ratio of 80:20. Further, 20% of the training set is separated as a validation set, which is used for early stopping <cit.> during training to avoid over-fitting. For training failure predictors, to control the effect of highly imbalanced datasets, oversampling <cit.> is performed on the minority class (i.e., failure logs) to achieve a 50:50 ratio of normal to failure logs in the training dataset. For all the training datasets, we use the Adam optimizer <cit.> with a learning rate of 0.001 and the sparse categorical cross-entropy loss function <cit.> considering the Boolean output (i.e., failure or not) of the models. However, we use different batch sizes and numbers of epochs for datasets with different characteristics since they affect the convergence speed of the training error (particularly the dataset size, the maximum LSL, and the failure percentage). It would however be impractical to fine-tune the batch size and the number of epochs for 360 individual combinations. Therefore, based on our preliminary evaluation results, we use larger batch sizes with fewer epochs for larger datasets to keep the training time reasonable without significantly affecting training effectiveness. Specifically, we set the two hyperparameters as follows: * Batch size: By default, we set it to 10, 15, 20, 30, 150, and 300 for dataset sizes of 200, 500, 1000, 5000, 10000, and 50000, respectively. If the failure percentage is less than or equal to 30 (meaning more oversampling will happen to balance between normal and failure logs, increasing the training data size), then we increase the batch size to 10, 15, 30, 60, 300, and 600, respectively, to reduce training time. Furthermore, regardless of the failure percentage, we set the batch size to 5 if the maximum LSL is greater than or equal to 500 to prevent memory issues during training. * Number of epochs: By default, we set it to 20. If the maximum LSL is greater than or equal to 500, we reduce the number of epochs to 10, 10, 5, and 5 for dataset sizes of 1000, 5000, 10000, and 50000, respectively, to reduce training time. Table <ref> summarises the above conditions, where FP is the failure percentage and MLSL refers to the maximum LSL. Once failure predictors are trained, we measure their accuracy on the corresponding test set in terms of precision, recall, and F1 score. We conducted all experiments with cloud computing environments provided by the Digital Research Alliance of Canada <cit.>, on the Cedar cluster with a total of 94528 CPU cores for computation and 1352 GPU devices. §.§ Synthetic Data Generation In defining a set of factors, the methodology described in  <ref> makes it clear that there is a need for a mechanism that can generate datasets in a controlled, unbiased manner. For example, let us consider the factor of failure percentage ( <ref>). Such a factor requires that one be able to control whether the log sequence being generated does indeed correspond to a failure; this would ultimately allow one to control the percentage of failure log sequences in a generated dataset. While, for smaller datasets, one could imagine manually choosing log sequences that represent both failures and normal behaviour, for larger datasets this is not feasible. When considering the other factors defined in  <ref>, such as LSL, the case for a mechanism for automated, controlled generation of datasets becomes yet stronger. §.§.§ Key Requirements We now describe a set of requirements that must be met by whatever approach we opt to take for generating datasets. In particular, our approach should: req req:controlledAllow datasets' characteristics to be controlled. This requirement has already been described, but we summarise it here for completeness. We must be able to generate datasets for each combination of levels (of the factors defined in  <ref>). Hence, our approach must allow us to choose a combination of levels, and generate a dataset accordingly. req:realisticBe able to generate realistic datasets. A goal of this work is to present results that are applicable to real-world systems. Hence, we must require that the datasets with which we perform any evaluations reflect real-world system behaviours. req:diverseBe able to generate datasets corresponding to a diverse set of systems. While we require that the datasets that we use be realistic, we must also ensure that the data generator can be applied to any system for which we have this automaton, rather than being limited to a single system. req:unbiasedAvoid bias in the log sequences that make up the generated datasets. The previous requirement ensures that we do not introduce an approach that only works with one system. If our approach were to work solely with one system, we might say that it was biased at the use case level. The second kind of bias is at the log sequence-level. For example, for a given system, we wish to generate datasets containing log sequences that explore as much of the system's behaviour as possible (rather than being biased to a particular part of the system). §.§.§ Automata for System Behaviour Our approach is based on finite-state automata. In particular, we use automata as approximate models of the behaviour of real-world systems. We refer to such automata as behaviour models, since they represent the computation performed by (i.e., behaviour of) some real-world system. We chose automata, or behaviour models, because some of our requirements are met immediately: R<ref>. Existing tools <cit.> allow one to infer behaviour models of real-world systems from collections of these systems' logs (in a process called model inference). Such models attach log messages to transitions, which is precisely what we need. Importantly, collections of logs used are unlabelled, meaning that the models that we get from these tools have no existing notion of normal behaviour or failures. R<ref>. A result of meeting R<ref> is that one can easily infer behaviour models for multiple systems, provided the logs of those systems are accessible. The remaining sections will give the complete details of our automata-based data generation approach. In presenting these details, we will show how R<ref> and R<ref> are met. §.§.§ Behaviour Models We take a behaviour model ℳ to be a deterministic finite-state automaton ⟨ Q, A, q_0, Σ, δ⟩, with symbols as defined in  <ref>. A behaviour model has the particular characteristic that its alphabet Σ consists of log template IDs (see  <ref>). A direct consequence of this is that one can extract log sequences from behaviour models. In particular, if one considers a sequence of states (i.e., a path) q_0, q_i, q_i+1, …, q_n through the model, one can extract a sequence of log template IDs using the transition function δ. For example, if the first two states of the sequence are q_0 and q_i, then one need only find s ∈Σ such that δ(q_0, s) = q_i, i.e., it is possible to transition to q_i from q_0 by observing s. Finally, by replacing each log template ID in the resulting sequence with its corresponding log template, one obtains a log sequence (see  <ref>). These sequences can be divided into two categories: failure log sequences and normal log sequences. We describe failures using regular expressions. This is natural since behaviour models are finite state automata, and sets of paths through such automata can be described by regular expressions. Hence, we refer to such a regular expression as a failure pattern, and denote it by 𝑓𝑝. By extension, for a given behaviour model ℳ, we then denote by 𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ) the set {𝑓𝑝_1, 𝑓𝑝_2, …, 𝑓𝑝_𝑛} of failure patterns paired with the model ℳ. Based on this, we characterise failure log sequences as such: Failure log sequence. For a system whose behaviour is represented by a behaviour model ℳ, we say that a log sequence represents a failure of the system whenever its sequence of log template IDs matches some failure pattern 𝑓𝑝∈𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ). Since this definition of failure log sequences essentially captures a subset of the possible paths through ℳ, we define normal log sequences as those log sequences that are not failures: Normal log sequence. For a system whose behaviour is represented by a behaviour model ℳ, we say that a log sequence l is normal, i.e., it represents normal behaviour, whenever l ∈ℒ(ℳ) and l ∉⋃_𝑓𝑝∈𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ)ℒ(𝑓𝑝) (we take ℒ(ℳ) and ℒ(𝑓𝑝) to be as defined in  <ref>). Hence, defining a normal log sequence requires that we refer to both the language of the model ℳ, and the languages of all failure patterns associated with the model ℳ. §.§.§ Generating Log Sequences for Failures Let us suppose that we have inferred a model ℳ from the execution logs of some real-world system, and that we have defined the set 𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ). Then we generate a failure log sequence that matches some 𝑓𝑝∈𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ) by: * Computing a subset of ℒ(𝑓𝑝). We do this by repeatedly generating single members of ℒ(𝑓𝑝). Ultimately, this leads to the construction of a subset of ℒ(𝑓𝑝). In practice, the Python package exrex<cit.> can be used to generate random words from the language ℒ(𝑓𝑝), so we invoke this library repeatedly. If the language of the regular expression is infinite, we can run exrex multiple times, each time generating a random string from the language. The number of runs is set based on our preliminary results with respect to the range of dataset size (2500 times for each failure pattern). Doing this, we generate a subset of ℒ(𝑓𝑝). * Choosing at random a log sequence l from the random subset ℒ(𝑓𝑝) computed in the previous step, with |l| ≤𝑚𝑙𝑠𝑙 where mlsl refers to the value of maximum log sequence manipulated by LSL factor. (see  <ref>) (maximum LSL, see  <ref>). The Python package random<cit.> was employed for this. We highlight that failure patterns are designed so that there is always at least one failure pattern that can generate log sequences whose length falls within this bound. More details on this are provided in  <ref>. For requirement R<ref>, since our approach relies on random selection of log sequences from languages generated by the exrex tool, we highlight that the bias in our approach is subject to the implementation of both exrex, and the Python package, random. Exrex is a popular package for RegEx that has more than 100k monthly downloads. Its method for generating a random matching sequence is implemented by a random selection of choices on the RegEx's parse tree nodes. Random package uses the Mersenne twister algorithm <cit.> to generate a uniform pseudo-random number used for random selection tasks. §.§.§ Generating Log Sequences for Normal Behaviour While our approach defines how failures should look using a set of failure patterns 𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ) defined over a model ℳ, we have no such definition of how normal behaviour should look. Instead, this is left implicit in our behaviour model. However, based on the definition of normal log sequences given in  <ref>, such log sequences can be randomly generated by performing random walks on behaviour models. This fact forms the basis of our approach to generating log sequences for normal behaviour. However, we must also address key issues: 1) the log sequences generated by our random walk must be of bounded length, and 2) the log sequences must also lack bias. There are two reasons for enforcing a bound on the length of log sequences: * Deep learning models (such as CNN) often accept inputs of limited size, so we have to ensure that the data we generate is compatible with the models we use. * One of the factors introduced in  <ref> is LSL, so we need to be able to control the length of log sequences that we generate. For bias, we have two sources: 1) bias to specific regions of the behavioral model, 2) bias to limited variation of LSL. We must minimise bias in both cases. Algorithm <ref> gives our procedure for randomly generating a log sequence representing normal behaviour of a system. Algorithm <ref> itself makes use of Algorithm <ref>. In particular, Algorithm <ref> generates a normal log sequence by: * Generating a random log sequence by random walk (invoking Algorithm <ref>); * Looking for a failure pattern 𝑓𝑝∈𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ) that matches the generated log sequence; * Repeating until a log sequence is generated that matches no failure pattern. Ultimately, Algorithm <ref> is relatively lightweight; the weight of the work is performed by Algorithm <ref>, which we now describe in detail. The input arguments of Algorithm <ref>, which defines the procedure filteredRandomWalk, are a behaviour model ℳ and the maximum LSL, 𝑚𝑙𝑠𝑙. The algorithm proceeds as follows. First, on line 1, we invoke the calculateSValues function to compute a map that sends each state s ∈ Q of ℳ to the length of the shortest path from that state to an accepting state in A. Next, on line 2, the sequence variable is initialised to an empty sequence. As the algorithm progresses, this variable stores the generated sequence of log template IDs. To help with this, the variable currentState is initialised to keep track of the state that the algorithm is currently in during the walk of the behavioral model. Hence, this variable is initialised on line 4 as the initial state. The final step in the setup stage of our algorithm is to initialise the maximumWalk variable, which serves as a counter to ensure the limit on the length of the generated log sequence (defined by 𝑚𝑙𝑠𝑙) is respected. In the while loop (line 5), as long as the current state, 𝑐𝑢𝑟𝑟𝑒𝑛𝑡𝑆𝑡𝑎𝑡𝑒, is not yet an accepting state, the random walk transits from the current state to a new state. The set of possible transitions to take is computed by line 7, and stored in the variable 𝑡𝑟𝑎𝑛𝑠𝑖𝑡𝑖𝑜𝑛𝑠. Each transition is represented by a triple containing the starting state, the symbol to be observed, and the state resulting from observation of that symbol. Once this set has been computed, the algorithm performs a filtering step. In particular, in order to ensure that we respect the limit imposed on the length of the generated path by 𝑚𝑙𝑠𝑙, we only consider transitions that lead to a state q' such that 𝑠𝑉𝑎𝑙𝑢𝑒(q') < 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘. The resulting list of valid options is then held in the variable 𝑜𝑝𝑡𝑖𝑜𝑛𝑠. Once the set 𝑜𝑝𝑡𝑖𝑜𝑛𝑠 has been computed, one transition ⟨ q, s, q' ⟩ will be chosen randomly from the set (line 13). This random choice eliminates bias because, each time we choose the next state to transition to, we do not favour any particular state (there is no weighting involved). This, extended over an entire path, means that we do not favour any particular region of a behaviour model. Now, from the randomly chosen transition ⟨ q, s, q' ⟩, the log template ID, s, is added to sequence (via sequence concatenation); currentState is updated to the next state, q'; and maximumWalk is decreased by one. Based on the condition of the while loop (line 5), when 𝑐𝑢𝑟𝑟𝑒𝑛𝑡𝑆𝑡𝑎𝑡𝑒∈ A (i.e., the algorithm has reached an accepting state), the generated sequence 𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒 is returned. While Algorithm <ref> generates an unlabelled log sequence, Algorithm <ref> generates a normal log sequence. To do this, it starts by generating a log sequence, by invoking the filteredRandomWalk procedure (Algorithm <ref>). Since the sequence generated by Algorithm <ref> is unlabelled, we must ensure that we do not generate a failure log sequence. We do this by checking whether the generated log sequence, 𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒, belongs to the language of any failure pattern in 𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ). If this is indeed the case, another sequence must be generated. This process is repeated (line 2) until the log sequence generated by the call of 𝑓𝑖𝑙𝑡𝑒𝑟𝑒𝑑𝑅𝑎𝑛𝑑𝑜𝑚𝑊𝑎𝑙𝑘 does not match any failure pattern in 𝖿𝖺𝗂𝗅𝗎𝗋𝖾𝖯𝖺𝗍𝗍𝖾𝗋𝗇𝗌(ℳ). Once a failure log sequence has been generated, it is returned. We acknowledge that this process could be inefficient (since we are repeatedly generating log sequences until we get one with the characteristics that we need). However, we highlight that failure patterns describe only a small part of a behaviour model (this is essentially the assumption that failure is a relatively uncommon event in a real system). Hence, normal log sequences generated by random walks can be generated without too many repetitions. Correctness and lack of bias. We now provide a sketch proof of the correctness of Algorithm <ref>, along with an argument that the algorithm eliminates bias. To prove correctness, we show that, for a behaviour model ℳ, the algorithm always generates a sequence of log template IDs that correspond to the transitions along a path through ℳ. The algorithm begins at q_0, by setting 𝑐𝑢𝑟𝑟𝑒𝑛𝑡𝑆𝑡𝑎𝑡𝑒 to q_0 (line 4). From q_0, and each successive state in the path, the possible next states must be adjacent to 𝑐𝑢𝑟𝑟𝑒𝑛𝑡𝑆𝑡𝑎𝑡𝑒 (line 7). Hence, the final value of 𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒 after the while-loop at line 5 must be a sequence of log template IDs that correspond to the transitions along a path through ℳ. Further, we must show that the sequence of log template IDs generated does not only correspond to a path through the behaviour model, but is of length at most 𝑚𝑙𝑠𝑙 (one of the inputs of Algorithms <ref> and <ref>). This is ensured by three factors: * The initialisation of the variable 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘 on line 3. * The subsequent decrease by one of that variable each time a new log template ID is added to 𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒. * Filtering of the possible next states in the random walk on line 9. In particular, on line 9 we ensure that, no matter which state we transition to, there will be a path that 1) leads to an accepting state; and 2) has length less than 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘. Finally, bias is minimised by two factors: * On line 13, we choose a random next state. Of course, here we rely on the implementation of random choice that we use. * On line 9, while we respect the maximum length of the sequence of log template IDs, we do not enforce that we reach this maximum. Hence, we can generate paths of various lengths. Example. To demonstrate Algorithm <ref>, we now perform a random walk over the behaviour model shown in Figure <ref>. We start with the behaviour model's starting state, q_0, with 𝑚𝑙𝑠𝑙 set to 5. Since q_0 ∉A, we can execute the body of the while-loop at line 5. Hence, we can determine the set 𝑡𝑟𝑎𝑛𝑠𝑖𝑡𝑖𝑜𝑛𝑠 of transitions leading out of q_0: {⟨ q_0,a,q_2⟩, ⟨ q_0,b,q_2⟩, ⟨ q_0,c, q_1⟩, ⟨ q_0,d,q_1⟩}. Our next step is to filter these transitions to ensure that the state that we move to allows us to reach an accepting state within 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘 states. To do this, we filter the set 𝑡𝑟𝑎𝑛𝑠𝑖𝑡𝑖𝑜𝑛𝑠 with respect to the values in Table <ref>. After this filtering step, the resulting set, 𝑜𝑝𝑡𝑖𝑜𝑛𝑠, is {⟨ q_0,a,q_2⟩, ⟨ q_0,b,q_2⟩, ⟨ q_0,c, q_1⟩, ⟨ q_0,d,q_1⟩}. All states in 𝑡𝑟𝑎𝑛𝑠𝑖𝑡𝑖𝑜𝑛𝑠 are safe to transition to. To take one transition as an example, ⟨ q_0, a, q_2⟩ has 𝑠𝑉𝑎𝑙𝑢𝑒(q_2) = 1 < 5, so is kept. Once we have computed 𝑜𝑝𝑡𝑖𝑜𝑛𝑠, we choose a transition at random. In this case, we arrive at ⟨ q_0,c,q_1⟩, meaning that we set 𝑐𝑢𝑟𝑟𝑒𝑛𝑡𝑆𝑡𝑎𝑡𝑒 to q_1. Before we progress to the next iteration of the main loop of the algorithm, we also decrease 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘. This means that, during the next iteration of the while loop, we will be able to choose transitions leading to states from which an accepting state must be reachable within less than 4 states. Indeed, from q_1, there are four transitions, for which we compute the set {⟨ q_1,a,q_0⟩, ⟨ q_1,b,q_1⟩, ⟨ q_1,c, q_3⟩, ⟨ q_1,d,q_3⟩}. From this set, each possible next state has 𝑠𝑉𝑎𝑙𝑢𝑒 greater than 𝑚𝑎𝑥𝑖𝑚𝑢𝑚𝑊𝑎𝑙𝑘 (equal to 4), so all of them would be possible options for the next step. Suppose that we choose ⟨ q_1,a,q_0⟩ at random. Hence, q_0 is the next state and a is added to the 𝑠𝑒𝑞𝑢𝑒𝑛𝑐𝑒. For the remaining steps, a possible run of the procedure could yield the sequence of transitions ⟨ q_0,b,q_2⟩, ⟨ q_2,d,q_1⟩, ⟨ q_1,d,q_3⟩, in which case the final sequence of log template IDs would be c, a, b, d, d. §.§.§ A summary of requirements met We now describe how the approach that we have described meets the remaining requirements set out in  <ref>. R<ref> is met because we have two procedures for generating failure log sequences ( <ref>) and normal log sequences ( <ref>). By having these procedures, we can precisely control the number of each type of log sequence in our dataset. R<ref> is met because of the randomisation used in our data generation algorithm, described in Sections <ref> and <ref>. §.§ Experimental Setting for Syntactic Data Generation To generate diverse log datasets with the characteristics described in  <ref>, using the syntactic data generation approach described in  <ref>, we need two main artifacts: behavior models and failure patterns. §.§.§ Behavior Models Regarding behavior models, as discussed in  <ref>, we can infer accurate models of real-world systems from their execution logs using state-of-the-art model inference tools, i.e., MINT <cit.> and PRINS <cit.>. Among the potential models we could generate using the replication package of these tools, we choose models that satisfy the following criteria based on the model size and inference time reported by <cit.>: * The model should be able to generate (accept) a log with a maximum length of 20 (i.e., the shortest maximum LSL defined in  <ref>; * Since there is no straightforward way of automatically generating failure patterns for individual behavior models considering the two failure pattern types, we had to manually generate failure patterns (detailed in  <ref>). Therefore, the size of the model should be amenable to manually generating failure patterns by taking into account the model structure (i.e., the number of all states and transitions is less than 1000); * The model inference time should be less than 1 hour; and * If we can use both PRINS and MINT to infer a model that satisfies the above criteria for the same logs, then we use PRINS, which is much faster than MINT in general, to infer the model. As a result, we use the following three models as our behavior models: ℳ_1 (generated from NGLClient logs using PRINS), ℳ_2 (generated from HDFS logs using MINT), and ℳ_3 (generated from Linux logs using MINT). Table <ref> reports about the size of the three behavior models in terms of the number of templates (#Templates), states (#States), and transitions (#Transitions). It additionally shows the number of states in the largest strongly connected component (#States-NSCC) <cit.>, which indicates the complexity of a behavior model (the higher, the more complex). §.§.§ Failure Patterns Regarding failure patterns, recall a failure pattern 𝑓𝑝 of a behavior model ℳ is a regular expression where ℒ(𝑓𝑝) ⊂ℒ(ℳ), as described in  <ref>. Also, note that we need two types of failure patterns (Type-F and Type-I), and the failure log sequences generated from the failure patterns must satisfy the dataset characteristics (especially the maximum LSL) defined in  <ref>. To manually create such failure patterns (regular expressions) in an unbiased way, we used the following steps for each behavior model and failure pattern type: * We randomly choose the alphabet size of a regular expression and the number of operators (i.e., alternations and Kleene stars; the latter is not used for Type-F). * Using the chosen random values, for a given behavior model ℳ, we manually create a failure pattern (regular expression) 𝑓𝑝 to satisfy ℒ(𝑓𝑝) ⊂ℒ(ℳ) and the maximum LSL within the time limit of 1 hour; if we fail (e.g., if the shortest log in ℒ(𝑓𝑝) is longer than the maximum LSL of 20), we go back and restart Step 1. * We repeat Steps 1 and 2 ten times to generate ten failure patterns and then randomly select three out of them. As a result, we use 18 failure patterns (i.e., 3 failure patterns × 3 behavior models × 2 failure pattern types) for synthetic data generation. § RESULTS This section presents the results of RQ1 (DL encoders), RQ2 (log sequence embedding strategies), and RQ3 (dataset characteristics), respectively. §.§ RQ1: DL Encoders  <ref> shows boxplots of the failure prediction accuracy (F1 score) for different DL encoders (i.e., transformer-based, LSTM-based, CNN-based, and BiLSTM-based models) on the datasets generated by different behavior models (i.e., M1, M2, and M3). Each box is generated based on 360× 2 data points since we have 360 combinations of dataset characteristics and two log sequence embedding strategies. In each box, a triangle indicates the mean value. Overall, the CNN-based model achieves the best performance in terms of F1 score for all behavior models. It has the highest mean values with the smallest interquartile ranges (IQRs), meaning that the CNN-based model consistently works very well regardless of dataset characteristics and log sequence embedding strategies. The BiLSTM-based model also shows promising results. However, the CNN-based model's results are significantly higher for all the behavioral models (paired t-test p-values << 0.001). In contrast, the LSTM-based and transformer-based models show poor results (low F1 score on average with very large IQRs). These patterns are independent from both the embedding strategy and the model. Further, the large variance for LSTM-based and transformer-based models suggests that these models are very sensitive to the dataset characteristics. The poor performance of the transformer-based encoder can be explained by the fact that the transformer blocks in the encoder are data-demanding (i.e., requiring much training data). When the dataset size is small (below 1000), the data-demanding transformer blocks are not well-trained, leading to poor performance. This limitation is thoroughly discussed in the literature <cit.>. The LSTM-based encoder, on the other hand, has two simple layers of LSTM units. Recall that an LSTM model sequentially processes a given log sequence (i.e., a sequence of templates), template by template. Although LSTM attempts to address the long-term dependency problem of RNN by having a forget gate (see  <ref>), it is still a recurrent network that has difficulties to remember a long input sequence <cit.>. For this reason, since our log datasets contain long log sequences (up to a length of 1000), the LSTM-based encoder did not work well. The BiLSTM-based encoder involves LSTM units and therefore has the weakness mentioned above. However, for BiLSTM, the input sequence flows in both directions in the network, utilizing information from both sides. Furthermore, it is enhanced by the attention mechanism that assigns more weight to parts of the input which are associated with the failure pattern <cit.>. Thus, the BiLSTM-based encoder can more easily learn the impact of different log templates on the classification results. However, the attention layer is more data-demanding than the convolution layers (see  <ref>) in the CNN-based encoder, and this explains why the BiLSTM-based encoder does not outperform the CNN-based encoder. The high performance of the BiLSTM-based and CNN-based encoders can be explained by the number of trainable parameters; for these two encoders, unlike the transformer-based and LSTM-based ones, the number of trainable parameters increases as the input sequences get longer. The larger number of parameters makes the encoders more robust to longer input log sequences. Furthermore, CNN additionally processes spatial information (i.e., conjunctive relationships among templates) using multiple filters with different kernel sizes <cit.>, which makes failure prediction more accurate even when the input size (sequence length) is large. These characteristics make the CNN-based encoder the best choice in our application context. The answer to RQ1 is that the CNN-based encoder significantly outperforms the other encoders regardless of dataset characteristics and log sequence embedding strategies. §.§ RQ2: Log Sequence Embedding Strategies  <ref> shows the boxplots of the failure prediction accuracy (F1 score) for the different log sequence embedding strategies considered in this study (i.e., BERT and Logkey2vec) on the datasets generated by the three behaviour models (M1, M2, and M3). Each box is generated based on 360× 4 data points since we have 360 combinations of dataset characteristics and four DL encoders. Similar to  <ref>, the triangle in each box indicates the mean value. We now inspect the plots shown inside with the aim of answering our research questions. The plots based on precision and recall are excluded since they draw similar conclusions.  <ref> shows that the BERT embedding strategy performs better than Logkey2vec for all behaviour models in terms of mean values and smaller IQRs. This means that, on average, for all DL encoders, the semantic-aware log sequence embedding using BERT fares better than the embedding that solely relies on log template IDs using Logkey2vec, which does not account for the semantic information of templates. To better understand the impact of log sequence embedding strategies on the performance of different DL encoders, we additionally performed paired t-tests to compare the F1 score distributions of BERT and Logkey2vec for each of the four DL encoders. Table <ref> reports the statistical test results. For example, the low p-value in column M2 and row CNN indicates that Logkey2vec is significantly better than BERT when the CNN-based encoder is used on datasets generated by the M2 behaviour model. Interestingly, BERT is statistically better than or equal to Logkey2vec for all DL encoders except the CNN-based encoder (i.e., the best-performing DL encoder as investigated in  <ref>), a trend that is clearly observable in  <ref>, depicting the F1 score distributions of BERT and Logkey2vec for the CNN encoder. In other words, the combination of the CNN-based encoder and the Logkey2vec embedding strategy is the best combination of DL encoders and log sequence embedding strategies. Although, in contrast to BERT, Logkey2vec does not consider the semantic information of log templates, it accounts for the order of template IDs in each log sequence. Furthermore, Logkey2vec is trained together with the DL encoder, while BERT is pre-trained independently from the DL encoder. We suspect that such characteristics of Logkey2vec play a positive role in combination with the CNN-based encoder. We note that BERT is still an attractive strategy for log sequence embedding when any other encoder than CNN is used. Although BERT is considerably larger than Logkey2vec in terms of parameters, using BERT does not require significantly more time and resources than Logkey2vec since BERT minimises repeated calculations by mapping each log template ID to its corresponding BERT embedding vector. The answer to RQ2 is that the performance of the log sequence embedding strategies varies depending on the DL encoders used. Although BERT outperforms Logkey2vec overall across all encoders, Logkey2vec outperforms BERT when the CNN-based encoder is used. §.§ RQ3: Dataset Characteristics  <ref> shows the distributions of F1 scores according to different dataset characteristic values. We only use the data generated by the best combination of the DL encoder and the log sequence embedding strategy, i.e., the CNN-based encoder and Logkey2vec, based on the results from RQ1 and RQ2. Below, we discuss how the failure prediction accuracy of the best-performing combination varies with each of the dataset characteristics. In  <ref>(a), we can see the impact of dataset size on the failure prediction accuracy; it is clear that accuracy decreases with smaller datasets, regardless of the behaviour models used to generate the log datasets. For example, when dataset size is 200, accuracy goes down below 0.7 in the worst case, whereas it always stays very close to 1.0 when dataset size is above or equal to 5000. Since larger datasets imply more training data, this result is intuitive.  <ref>(b) depicts the impact of maximum LSL values (mlsl) on the failure prediction accuracy. Compared to the impact of data set size, we can see that the impact of log sequence length is relatively small. This implies that the CNN-based encoder with the Logkey2vec embedding strategy works fairly well for long log sequence lengths of up to 1000. We suspect that the impact of log sequence length could be significant for much longer log sequences. However, log sequences longer than 1000 are not common in the publicly available, real-world log datasets <cit.> as explained in Section <ref>. Nevertheless, the investigation of much longer log sequences remains for future work. The relationship between the failure percentage and the failure prediction accuracy (F1 score) is depicted in  <ref>(c). It is clear that, overall, the F1 score increases as the failure percentage increases. This is intuitive since a larger failure percentage means more instances of failure patterns in the training data, making it easier to learn such patterns. An interesting observation is that the average failure prediction accuracy is above 0.9 even when the failure percentage is 10%. This implies that the failure predictor (i.e., DL encoder and log sequence embedding strategy) can cope well with unbalanced data.  <ref>(d) shows the failure prediction accuracy for different failure pattern types. There is no consistent trend across models M1, M2, and M3; Type-F (the corresponding language is finite) is better detected than Type-I (the corresponding language is infinite) in M2 and M3, whereas the opposite happens in M1. Considering the complexity of failure patterns, it is unclear why, in M1, detecting less complex failure patterns (Type-F) is more difficult than detecting more complex patterns (Type-I). We may not have defined failure pattern types in a way that is conducive to explaining variations in accuracy and different hypotheses will have to be tested in future work with respect to which pattern characteristics matter. We focused above on the impact of individual characteristics on the accuracy of failure prediction, not their combinations. To identify conditions describing how the failure prediction accuracy of the best-performing failure predictor (i.e., the CNN-based encoder with the Logkey2vec embedding strategy) varies according to combinations of dataset characteristics, we built a regression tree <cit.> where we explain variations in F1 scores according to dataset characteristics. Out of 1080 data points (360 × 3 since we have 360 data points for each of the three behavior models), we used 720 (66.7%) and 360 (33.3%) of them, respectively, for training and testing the tree.  <ref> shows the resulting regression tree; each non-leaf node presents a partial condition and each leaf node presents the predicted accuracy of the condition corresponding to the path from the root to the leaf. For example, the left-most leaf node means that the average failure prediction accuracy is predicted to be 0.516 if the dataset size is less than 350 and the failure percentage is less than 7.5. Otherwise, the failure average prediction accuracy will be at least 0.9. Further, we calculated the IQR of the F1 scores when the conditions for high accuracy are met, IQR reaches below 0.01. From the tree, it is clear that dataset size and failure percentage are the two main factors that explain failure prediction accuracy. Furthermore, we can see that the CNN-based encoder with the Logkey2vec embedding works very well, except when both the dataset size and failure percentage are small. The practical implication of this finding will be further discussed in Section <ref>. The answer to RQ3 is that, for the CNN-based encoder with the Logkey2vec embedding strategy, increasing the dataset size and failure percentage significantly increases the failure prediction accuracy, whereas the other factors (i.e., LSL and failure pattern type) do not have a clear relationship with failure prediction accuracy. Especially, the failure predictor is very accurate (F1 score above 0.95) and robust (IQR is below 0.01) when the dataset size is above 350 or the failure percentage is above 7.5%. §.§ Replication package We plan to make the replication package, including implementation, generated datasets with behavioral models, and results, publicly available upon acceptance. § DISCUSSION §.§ Findings and Implications Our study leverages the main DL types (LSTM, CNN, and transformer), along with different LSE strategies (BERT and Logkey2vec). In contrast to other studies mentioned in  <ref>, the full combination of DL types and LSE strategies are evaluated. Moreover, instead of using a limited number of datasets, using generated data enables us to control dataset characteristics to identify necessary conditions for achieving high-accuracy models. Several major findings are reported in  <ref>. First, the CNN-based DL encoder fares the best among different DL encoders, including the ones based on LSTMs, transformers, and BiLSTMs. Second, the CNN-based DL encoder works best with the Logkey2vec embedding strategy, although BERT fares better than Logkey2vec overall for all DL encoders. Third, for the best combination, i.e., the CNN-based encoder with Logkey2vec, both the size and the failure percentage of input log datasets significantly drive the failure prediction accuracy, whereas the log sequence length and the failure pattern type do not. Last but not least, we found that the best combination works well if either the dataset size is above 350 or the failure percentage is above 7.5%. These findings carry practical implications for both researchers and engineers. Although the CNN-based DL encoder and the Logkey2vec embedding strategy are not the most recent techniques in their respective fields, interestingly, their combination works best for failure prediction. Based on this finding, we can recommend using the CNN-based encoder with Logkey2vec for accurate failure prediction among various combinations of DL encoders and log sequence embedding strategies. However, we have explored the ranges of dataset characteristics that have been observed in the literature. Different results may be observed beyond these ranges and for different failure patterns. Furthermore, the conditions driving failure prediction accuracy suggest practical guidelines. For example, for a log dataset size below 350 and a failure percentage below 7.5%, failure prediction will be inaccurate. In that case, one can increase either the log dataset size or the failure percentage to improve the situation. Although the failure percentage is inherent to the system under analysis and might not be easy to control in practice, collecting more log sequences during the operation of the system to increase the dataset size is usually feasible. §.§ Threats to Validity There are a number of potential threats to the validity of our experimental results. Hyper-parameter tuning of models. The hyper-parameters of failure predictors, such as optimizers, loss functions, and learning rates, can affect the results. To mitigate this, we followed recommendations from the literature. For the batch size and the number of epochs, as mentioned in  <ref>, we chose values for different combinations of dataset characteristics based on preliminary evaluation results. Better results could be obtained with different choices. Synthetic data generation process. Due to the lack of a method to generate the datasets satisfying different dataset characteristics mentioned in  <ref>, we proposed a new approach, with precise algorithms, that can generate datasets in a controlled, unbiased manner as discussed in  <ref>. To mitigate any risks related to synthetic generation, we provided proof of the correctness of the algorithms and explained why it is unbiased during the generation process. To further support the validity of the generation process, we compared the results on actual datasets reported in the literature with those of the synthesised datasets for corresponding key parameters (e.g., dataset sizes and failure percentage). Results show to be remarkably consistent, thus backing up the validity of our experiments. For example, a BiLSTM-based model in the study of <cit.>, achieves results comparable to the synthesised data with similar characteristics [Using real-world datasets of BGL and Spirit, the failure predictor achieved 1.0 and 0.95 as F1 score, respectively, while the F1 score results of similar synthesised datasets using the same DL type as encoder are 0.99 and 0.99]. Behavioral models and failure patterns. The behavioral models and failure patterns used for the generation of synthetic datasets may have a significant impact on the experimental results. We want to remark that this is the first attempt to characterise failure patterns for investigating failure prediction performance. To mitigate this issue, we carefully chose them based on pre-defined criteria described in  <ref>. Nevertheless, more case studies, especially considering finer-grained failure patterns, are required to increase the generalizability of our findings and implications and, for that purpose, we provide in our replication package all the artifacts required. Possible bugs in the implementation. The implementation of the DL encoders, the log sequence embedding strategies, the dataset generation algorithms, and the scripts used in our experiments might include unexpected bugs. To mitigate this risk, we used the replication packages of existing studies <cit.> as much as possible. Also, we carefully performed code reviews. § CONCLUSION In this paper, we presented a comprehensive and systematic evaluation of alternative failure prediction strategies relying on DL encoders and log sequence embedding strategies. We presented a generic, modular architecture for failure prediction which can be configured with specific DL encoders and embedding strategies, resulting in different failure predictors. We considered BERT and Logkey2vec as embedding strategies. We also covered the main DL categories, plus the attention mechanism, resulting in four DL encoders (LSTM-, BiLSTM-, CNN-, and transformer-based). Our selection was inspired by the previously used DL models in the literature. We evaluated the failure prediction models on diverse synthetic datasets using three behavioural models inferred from available system logs. Four dataset characteristics were controlled when generating datasets: dataset size, failure percentage, Log Sequence Length (LSL), and failure pattern type. Using these characteristics, 360 datasets were generated for each of three behavioral models. Evaluation results show that the accuracy of the CNN-based encoder is significantly higher than the other encoders regardless of dataset characteristics and embedding strategies. Between the two embedding strategies, pretrained BERT outperformed the trainable Logkey2vec overall, although Logkey2vec fared better for the CNN-based encoder. The analysis of dataset characteristics confirms that increasing the dataset size and failure percentage increases the failure prediction accuracy, while the other factors (i.e., LSL and failure pattern type) did not show a clear relationship with failure prediction accuracy. Furthermore, the accuracy of the best configuration (i.e., CNN-based with Logkey2vec) consistently yielded high accuracy when the dataset size was above 350 or the failure percentage was above 7.5%, which makes it widely usable in practice. As part of future work, we plan to further evaluate the best-performing configuration of the failure prediction architecture on real-world log data to further investigate the effect of other factors, such as log parsing techniques, on model accuracy. Additionally, by using real-world data, we also plan to include time-aware evaluation metrics, such as lead time <cit.>, to assess the accuracy of these models at predicting failures early on. spbasic 60 urlstyle [Bauer and Adams(2012)]bauer2012reliability Bauer E, Adams R (2012) Reliability and availability of cloud computing. John Wiley & Sons [Black(2020)]SCC Black PE (2020) Strongly connected component. Dictionary of Algorithms and Data Structures <https://www.nist.gov/dads/HTML/stronglyConnectedCompo.html> [Breiman(2001)]breiman2001random Breiman L (2001) Random forests. Machine learning 45(1):5–32, 10.1023/A:1010933404324 [Breiman et al.(1984)Breiman, Friedman, Olshen, and Stone]breiman1984classification Breiman L, Friedman JH, Olshen RA, Stone CJ (1984) Classification and Regression Trees. Wadsworth [Carvalho et al.(2019)Carvalho, Soares, Vita, da P. Francisco, Basto, and Alcalá]CARVALHO2019106024 Carvalho TP, Soares FAAMN, Vita R, da P Francisco R, Basto JP, Alcalá SGS (2019) A systematic literature review of machine learning methods applied to predictive maintenance. Computers & Industrial Engineering 137:106024, https://doi.org/10.1016/j.cie.2019.106024, <https://www.sciencedirect.com/science/article/pii/S0360835219304838> [Chen et al.(2019)Chen, Yang, Lin, Zhang, Dong, Xu, Li, Kang, Zhang, Gao, Xu, and Dang]Chen2019 Chen Y, Yang X, Lin Q, Zhang D, Dong H, Xu Y, Li H, Kang Y, Zhang H, Gao F, Xu Z, Dang Y (2019) Outage prediction and diagnosis for cloud service systems. The Web Conference 2019 - Proceedings of the World Wide Web Conference, WWW 2019 pp 2659–2665, 10.1145/3308558.3313501 [Chen et al.(2022)Chen, Li, Li, Guo, Du, and Xu]chen2022ai Chen Y, Li L, Li W, Guo Q, Du Z, Xu Z (2022) AI Computing Systems: An Application Driven Perspective. Elsevier Science, <https://books.google.ca/books?id=RSWJEAAAQBAJ> [Chen et al.(2021)Chen, Liu, Gu, Su, and Lyu]experimentreport Chen Z, Liu J, Gu W, Su Y, Lyu MR (2021) Experience report: Deep learning-based system log analysis for anomaly detection. 10.48550/ARXIV.2107.05908, <https://arxiv.org/abs/2107.05908> [Cho et al.(2014)Cho, Van Merriënboer, Bahdanau, and Bengio]cho2014learning Cho K, Van Merriënboer B, Bahdanau D, Bengio Y (2014) Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:14061078 [Chollet(2017)]padding Chollet F (2017) Xception: Deep learning with depthwise separable convolutions. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [Cortes and Vapnik(1995)]cortes1995support Cortes C, Vapnik V (1995) Support-vector networks. Machine learning 20(3):273–297 [Das et al.(2018)Das, Mueller, Siegel, and Vishnu]Das2018 Das A, Mueller F, Siegel C, Vishnu A (2018) Desh: Deep learning for system health prediction of lead times to failure in HPC. HPDC 2018 - Proceedings of the 2018 International Symposium on High-Performance Parallel and Distributed Computing pp 40–51, 10.1145/3208040.3208051 [Das et al.(2020)Das, Mueller, and Rountree]aarohi Das A, Mueller F, Rountree B (2020) Aarohi: Making Real-Time Node Failure Prediction Feasible. Proceedings - 2020 IEEE 34th International Parallel and Distributed Processing Symposium, IPDPS 2020 pp 1092–1101, 10.1109/IPDPS47924.2020.00115 [Deerwester et al.(1990)Deerwester, Dumais, Furnas, Landauer, and Harshman]countvector Deerwester S, Dumais ST, Furnas GW, Landauer TK, Harshman R (1990) Indexing by latent semantic analysis. Journal of the American society for information science 41(6):391–407 [Devlin et al.(2018)Devlin, Chang, Lee, and Toutanova]BERTbasemodel Devlin J, Chang MW, Lee K, Toutanova K (2018) Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:181004805 [Digital Research Alliance of Canada(2016)]computecanada Digital Research Alliance of Canada (2016) <https://alliancecan.ca/>, accessed: March 2, 2023 [Ding et al.(2020)Ding, Zhou, Yang, and Tang]Ding2020CogLTXAB Ding M, Zhou C, Yang H, Tang J (2020) Cogltx: Applying bert to long texts. In: Neural Information Processing Systems [Du et al.(2017)Du, Li, Zheng, and Srikumar]DeepLog Du M, Li F, Zheng G, Srikumar V (2017) Deeplog: Anomaly detection and diagnosis from system logs through deep learning. In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, Association for Computing Machinery, New York, NY, USA, CCS '17, p 1285–1298, 10.1145/3133956.3134015, <https://doi.org/10.1145/3133956.3134015> [Gers et al.(2000)Gers, Schmidhuber, and Cummins]GersLSTM Gers FA, Schmidhuber JA, Cummins FA (2000) Learning to forget: Continual prediction with lstm. Neural Comput 12(10):2451–2471, 10.1162/089976600300015015, <https://doi.org/10.1162/089976600300015015> [Gu et al.(2018)Gu, Wang, Kuen, Ma, Shahroudy, Shuai, Liu, Wang, Wang, Cai, and Chen]CNNadvances Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, Liu T, Wang X, Wang G, Cai J, Chen T (2018) Recent advances in convolutional neural networks. Pattern Recognition 77:354–377, https://doi.org/10.1016/j.patcog.2017.10.013, <https://www.sciencedirect.com/science/article/pii/S0031320317304120> [Guo et al.(2021)Guo, Yuan, and Wu]logBERT Guo H, Yuan S, Wu X (2021) Logbert: Log anomaly detection via bert. In: 2021 International Joint Conference on Neural Networks (IJCNN), pp 1–8, 10.1109/IJCNN52387.2021.9534113 [He et al.(2021)He, He, Chen, Yang, Su, and Lyu]He2021 He S, He P, Chen Z, Yang T, Su Y, Lyu MR (2021) A Survey on Automated Log Analysis for Reliability Engineering. ACM Computing Surveys 54(6), 10.1145/3460345, 2009.07237 [Hochreiter and Schmidhuber(1997a)]LSTM Hochreiter S, Schmidhuber J (1997a) Long short-term memory. Neural Comput 9(8):1735–1780, 10.1162/neco.1997.9.8.1735, <https://doi.org/10.1162/neco.1997.9.8.1735> [Hochreiter and Schmidhuber(1997b)]HochSchm97 Hochreiter S, Schmidhuber J (1997b) Long short-term memory. Neural Computation 9(8):1735–1780 [Huang et al.(2020)Huang, Liu, Fung, He, Zhao, Yang, and Luan]HitAnomaly Huang S, Liu Y, Fung C, He R, Zhao Y, Yang H, Luan Z (2020) HitAnomaly: Hierarchical Transformers for Anomaly Detection in System Log. IEEE Transactions on Network and Service Management 17(4):2064–2076, 10.1109/TNSM.2020.3034647 [Huang et al.(2015)Huang, Xu, and Yu]BiLSTMfirstpaper Huang Z, Xu W, Yu K (2015) Bidirectional lstm-crf models for sequence tagging. 10.48550/ARXIV.1508.01991, <https://arxiv.org/abs/1508.01991> [Johnson and Khoshgoftaar(2019)]ImbalanceLearning Johnson JM, Khoshgoftaar TM (2019) Survey on deep learning with class imbalance. Journal of Big Data 6(1), 10.1186/s40537-019-0192-5, <https://doi.org/10.1186/s40537-019-0192-5> [Jones(1979)]jones1979study Jones KS (1979) A study of the tf-idf weight scheme for text classification. In: Proceedings of the 3rd International Conference on Research in Information Retrieval, ACM, pp 100–107 [Kim(2014)]CNNdefinition Kim Y (2014) Convolutional neural networks for sentence classification. arXiv preprint arXiv:14085882 [Kingma and Ba(2015)]adam Kingma DP, Ba J (2015) Adam: A method for stochastic optimization. In: Bengio Y, LeCun Y (eds) 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, <http://arxiv.org/abs/1412.6980> [Le and Zhang(2021)]Neurallog Le VH, Zhang H (2021) Log-based anomaly detection without log parsing. In: 2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE), pp 492–504, 10.1109/ASE51524.2021.9678773 [Le and Zhang(2022)]10.1145/3510003.3510155 Le VH, Zhang H (2022) Log-based anomaly detection with deep learning: How far are we? In: Proceedings of the 44th International Conference on Software Engineering, Association for Computing Machinery, New York, NY, USA, ICSE '22, p 1356–1367, 10.1145/3510003.3510155, <https://doi.org/10.1145/3510003.3510155> [Lin et al.(2018)Lin, Hsieh, Dang, Zhang, Sui, Xu, Lou, Li, Wu, Yao, Chintalapati, and Zhang]unbalance_failure Lin Q, Hsieh K, Dang Y, Zhang H, Sui K, Xu Y, Lou JG, Li C, Wu Y, Yao R, Chintalapati M, Zhang D (2018) Predicting node failure in cloud service systems. In: Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, Association for Computing Machinery, New York, NY, USA, ESEC/FSE 2018, p 480–490, 10.1145/3236024.3236060, <https://doi.org/10.1145/3236024.3236060> [Lipton(2015)]Lipton2015ACR Lipton ZC (2015) A critical review of recurrent neural networks for sequence learning. ArXiv abs/1506.00019 [Liu et al.(2019)Liu, Ott, Goyal, Du, Joshi, Chen, Levy, Lewis, Zettlemoyer, and Stoyanov]Roberta Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, Levy O, Lewis M, Zettlemoyer L, Stoyanov V (2019) Roberta: A robustly optimized bert pretraining approach. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) [Lu et al.(2018)Lu, Wei, Li, and Wang]CNN Lu S, Wei X, Li Y, Wang L (2018) Detecting anomaly in big data system logs using convolutional neural network. IEEE Access 6:21929–21940, 10.1109/ACCESS.2018.2811530 [Matsumoto and Nishimura(1998)]Matsumoto1998 Matsumoto M, Nishimura T (1998) Mersenne Twister: A 623-Dimensionally Equidistributed Uniform Pseudo-Random Number Generator. ACM Transactions on Modeling and Computer Simulation 8(1):3–30, 10.1145/272991.272995 [Meng et al.(2019)Meng, Liu, Zhu, Zhang, Pei, Liu, Chen, Zhang, Tao, Sun, and Zhou]LogAnomaly Meng W, Liu Y, Zhu Y, Zhang S, Pei D, Liu Y, Chen Y, Zhang R, Tao S, Sun P, Zhou R (2019) Loganomaly: Unsupervised detection of sequential and quantitative anomalies in unstructured logs. In: International Joint Conference on Artificial Intelligence [Meng et al.(2020)Meng, Liu, Huang, Zhang, Zaiter, Chen, and Pei]Log2Vec Meng W, Liu Y, Huang Y, Zhang S, Zaiter F, Chen B, Pei D (2020) A semantic-aware representation framework for online log analysis. In: 2020 29th International Conference on Computer Communications and Networks (ICCCN), pp 1–7, 10.1109/ICCCN49398.2020.9209707 [Mikolov et al.(2013)Mikolov, Chen, Corrado, and Dean]word2vec Mikolov T, Chen K, Corrado GS, Dean J (2013) Efficient estimation of word representations in vector space. In: International Conference on Learning Representations [Nedelkoski et al.(2020)Nedelkoski, Bogatinovski, Acker, Cardoso, and Kao]LogSy Nedelkoski S, Bogatinovski J, Acker A, Cardoso J, Kao O (2020) Self-attentive classification-based anomaly detection in unstructured logs. Proceedings - IEEE International Conference on Data Mining, ICDM 2020-Novem(Icdm):1196–1201, 10.1109/ICDM50108.2020.00148, 2008.09340 [O'Shea and Nash(2015)]CNNfirstpaper O'Shea K, Nash R (2015) An introduction to convolutional neural networks. 10.48550/ARXIV.1511.08458, <https://arxiv.org/abs/1511.08458> [Package(2019)]random Package RP (2019) <https://docs.python.org/3/library/random.html>, accessed 2022-11-14 [Prechelt(1998)]earlystopping Prechelt L (1998) Early stopping-but when? In: Neural Networks: Tricks of the Trade, Springer, pp 55–69 [Radford et al.(2019)Radford, Wu, Child, Luan, Amodei, and Sutskever]GPT2 Radford A, Wu J, Child R, Luan D, Amodei D, Sutskever I (2019) Language models are unsupervised multitask learners. In: Neural Information Processing Systems [Sahoo et al.(2003)Sahoo, Oliner, Rish, Gupta, Moreira, Ma, Vilalta, and Sivasubramaniam]Sahoo2003 Sahoo RK, Oliner AJ, Rish I, Gupta M, Moreira JE, Ma S, Vilalta R, Sivasubramaniam A (2003) Critical event prediction for proactive management in large-scale computer clusters. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining pp 426–435, 10.1145/956750.956799 [Salfner et al.(2010)Salfner, Lenk, and Malek]Salfner2010 Salfner F, Lenk M, Malek M (2010) A survey of online failure prediction methods. ACM Computing Surveys 42(3), 10.1145/1670679.1670680 [Schuster and Paliwal(1997)]BiLSTM Schuster M, Paliwal K (1997) Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11):2673–2681, 10.1109/78.650093 [Shin et al.(2022)Shin, Bianculli, and Briand]PRINS Shin D, Bianculli D, Briand L (2022) Prins: Scalable model inference for component-based system logs. Empirical Softw Engg 27(4), 10.1007/s10664-021-10111-4, <https://doi.org/10.1007/s10664-021-10111-4> [Sun et al.(2019)Sun, Qiu, Xu, and Huang]10.1007/978-3-030-32381-3_16 Sun C, Qiu X, Xu Y, Huang X (2019) How to fine-tune bert for text classification? In: Sun M, Huang X, Ji H, Liu Z, Liu Y (eds) Chinese Computational Linguistics, Springer International Publishing, Cham, pp 194–206 [Tauber(2018)]exrex Tauber A (2018) exrex: Irregular methods for regular expressions. <https://github.com/asciimoo/exrex>, accessed 2022-11-14 [Upton and Cook(2008)]upton2008dictionary Upton G, Cook I (2008) A Dictionary of Statistics. Oxford Paperback Reference, OUP Oxford, <https://books.google.ca/books?id=u97pzxRjaCQC> [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin]AttentionIA Vaswani A, Shazeer NM, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention is all you need. ArXiv abs/1706.03762 [Walkinshaw et al.(2013)Walkinshaw, Taylor, and Derrick]MINT Walkinshaw N, Taylor R, Derrick J (2013) Inferring extended finite state machine models from software executions. In: 2013 20th Working Conference on Reverse Engineering (WCRE), pp 301–310, 10.1109/WCRE.2013.6671305 [Weijie et al.(2021)Weijie, Yunyi, Jing, and Xuchen]9587007 Weijie D, Yunyi L, Jing Z, Xuchen S (2021) Long text classification based on bert. In: 2021 IEEE 5th Information Technology,Networking,Electronic and Automation Control Conference (ITNEC), vol 5, pp 1147–1151, 10.1109/ITNEC52019.2021.9587007 [Wu et al.(2016)Wu, Schuster, Chen, Le, Norouzi, Macherey, Krikun, Cao, Gao, Macherey, Klingner, Shah, Johnson, Liu, Kaiser, Gouws, Kato, Kudo, Kazawa, Stevens, Kurian, Patil, Wang, Young, Smith, Riesa, Rudnick, Vinyals, Corrado, Hughes, and Dean]wordpiece Wu Y, Schuster M, Chen Z, Le QV, Norouzi M, Macherey W, Krikun M, Cao Y, Gao Q, Macherey K, Klingner J, Shah A, Johnson M, Liu X, Kaiser L, Gouws S, Kato Y, Kudo T, Kazawa H, Stevens K, Kurian G, Patil N, Wang W, Young C, Smith J, Riesa J, Rudnick A, Vinyals O, Corrado G, Hughes M, Dean J (2016) Google's neural machine translation system: Bridging the gap between human and machine translation. CoRR abs/1609.08144, <http://arxiv.org/abs/1609.08144>, 1609.08144 [Xu et al.(2020)Xu, Kumar, Yang, Zi, Tang, Huang, Cheung, Prince, and Cao]trans-largedata Xu P, Kumar D, Yang W, Zi W, Tang K, Huang C, Cheung JCK, Prince S, Cao Y (2020) Optimizing deeper transformers on small datasets. In: Annual Meeting of the Association for Computational Linguistics [Xu et al.(2019)Xu, Li, Li, Li, and Li]logembedding Xu Y, Li Z, Li Z, Li X, Li H (2019) Log embedding strategies for anomaly detection in log data. IEEE Access 7:58264–58277 [Yang et al.(2021)Yang, Chen, Wang, Wang, Jiang, Dong, and Zhang]PLELog Yang L, Chen J, Wang Z, Wang W, Jiang J, Dong X, Zhang W (2021) Semi-supervised log-based anomaly detection via probabilistic label estimation. In: 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pp 1448–1460, 10.1109/ICSE43902.2021.00130 [Zhang et al.(2019)Zhang, Xu, Lin, Qiao, Zhang, Dang, Xie, Yang, Cheng, Li, Chen, He, Yao, Lou, Chintalapati, Shen, and Zhang]LogRobust Zhang X, Xu Y, Lin Q, Qiao B, Zhang H, Dang Y, Xie C, Yang X, Cheng Q, Li Z, Chen J, He X, Yao R, Lou JG, Chintalapati M, Shen F, Zhang D (2019) Robust log-based anomaly detection on unstable log data. ESEC/FSE 2019 - Proceedings of the 2019 27th ACM Joint Meeting European Software Engineering Conference and Symposium on the Foundations of Software Engineering pp 807–817, 10.1145/3338906.3338931

http://arxiv.org/abs/2303.07265v1

20230313164428

Multimodal Reinforcement Learning for Robots Collaborating with Humans

cs.RO

CrysFieldExplorer: a software for rapid optimization of crystal field Hamiltonian Huibo Cao* March 30, 2023 ================================================================================= Robot assistants for older adults and people with disabilities need to interact with their users in collaborative tasks. The core component of these systems is an interaction manager whose job is to observe and assess the task, and infer the state of the human and their intent to choose the best course of action for the robot. Due to the sparseness of the data in this domain, the policy for such multi-modal systems is often crafted by hand; as the complexity of interactions grows this process is not scalable. In this paper, we propose a reinforcement learning (RL) approach to learn the robot policy. In contrast to the dialog systems, our agent is trained with a simulator developed by using human data and can deal with multiple modalities such as language and physical actions. We conducted a human study to evaluate the performance of the system in the interaction with a user. Our designed system shows promising preliminary results when it is used by a real user. § INTRODUCTION Assistive robots built with the intention of helping older people and people with disabilities with activities of daily living (ADLs) such as cooking or cleaning must be able to work in tandem with their users to complete their tasks. These interactions are typically multimodal (involving speech, gestures and physical actions) and collaborative (requiring some level of engagement from both participants). As with other autonomous robots, these assistive robots can be decomposed into three main components that perform the typical sense-plan-act cycle throughout the task: a perception module that collects and processes sensory data (potentially from multiple modalities) to understand the environment and the user's actions; an interaction manager that then determines an appropriate response for the robot; and an execution module that enables the robot to perform this action (Fig. <ref>). One way of designing an effective interaction manager for such a robot is to study how two humans interact with each other in collaborative tasks. We previously proposed a novel architecture called a Hierarchical Bipartite Action-Transition Networks (HBATNs) for multimodal human-robot interaction management <cit.>, which was developed largely based on our corpus of interactions between elderly individuals (labeled ELD) and nursing students (labeled HEL) assisting them in completing ADLs <cit.>. In particular, we targeted what we call the Find task, in which two participants collaborate to locate in the environment an object that is not visible. We observed the participants in this task interacting via speech, pointing gestures, and haptic-ostensive (H-O) actions that bring physical objects into conversational focus through touch. We showed that HBATNs can effectively provide a policy for the robot, but while the construction of the HBATN for the Find task was completely grounded in our human-human interaction data, it was ultimately crafted by hand and it would need to be re-designed for new tasks. In this paper, we propose a more scalable interaction manager that employs reinforcement learning (RL) to automatically extract an optimal policy for a robot participating in a collaborative task. With RL, the robot learns to differentiate desirable actions from undesirable actions through trial and error with feedback from its interactive environment in the form of rewards and penalties <cit.>. We demonstrate the effectiveness of this RL-based interaction manager by training the agent for the robot assuming the HEL role that collaborates with a human assuming the ELD role in the Find task. A major challenge in training this agent is in building an interactive environment that can provide the agent with rewards. Here, the environment includes not only the physical space in which the agent acts but also the human user acting as ELD. To be able to perform RL at scale, we developed a simulated environment with a neural network-based user simulator that was inspired by Behavioral Cloning <cit.> and was based on Find task data from our ELD-HEL interaction corpus. Additionally, because our corpus is relatively small for RL-based training and because interactions involving a robot are more prone to error than between two humans due to imperfect sensors and algorithms, we augmented our data with synthetic, data-driven misunderstandings and equipped our environment with an error-injecting module. Our contributions in this paper are twofold. Firstly, we propose an interpretable RL-based interaction manager for multimodal collaborative robots. The reward function is simple, and some preconditions are enforced to prevent unwanted policy searches and speed up the training process. Secondly, the proposed neural network-based, end-to-end user simulator provides a training environment in RL training. On the one hand, it handles "erupt" interactions, like suddenly changing the goal object. On the other hand, it can provide realistic actions when the dataset or the HBATN framework is not considered. Moreover, our preliminary user study showed that the robot agent meets high accuracy and satisfaction. The rest of the paper is organized as follows. Section <ref> reviews related literature, Section <ref> presents our user simulator, Section <ref> outlines our RL framework, and Section <ref> describes a user study we carried out to evaluate the performance of our learned interaction manager. In our previous study, motivated by our corpus of human-human interactions between elderly individuals (elder role: ELD) and nursing students (helper role: HEL) assisting in ADLs <cit.>, we developed a Hierarchical Bipartite Action-Transition Network (HBATN) as the main core of our multimodal Human-Robot Interaction Manager (MIM) for assistive robots <cit.>. The task that was studied in detail in our work was the Find task, an interaction scenario in which a human and a robot work together to find an object in the environment. The Find task is decomposed into a set of subtasks to identify two main unknowns: the target object (O) and its location (L). The four main subtasks are determining the desired object type (Det(O_T)), determining a potential location to check (Det(L)), opening the location (Open(L)), and determining the actual object (Det(O)). These are modeled as ActNets. Later on, we generalized our model to enable the robot to be either the ELD or HEL by decomposing the subtasks into what we call primitive subtasks <cit.>. The core of our framework is Hierarchical Bipartite Action-Transition Networks (HBATNs) that model both agents simultaneously to maintain the state of a task-driven multimodal interaction and plan subsequent robot moves. Althogh the HBATN ultimately provides the robot agent policy, it is a rule-based handcrafted network. We would like to make our agent intelligent enough to extract the optimal policy automatically. In this study, we propose our Multimodal Reinforcement Learning Framework which enables our robot agent to extract and perform the optimal policy automatically. However, there are some critical challenges we need to tackle in order to make our robot agent intelligent. The first and foremost challenge would be the small size of our ELDERLY-AT-HOME corpus <cit.> on which the HBATN is developed. HBATN does a great job replicating the interactions of the Find-task but when it comes to developing an intelligent agent based on the same data, it will get a lot more complex. For compensating the small amount of data, we could perform a Data Augmentation on our current data. When two humans collaborate on a multimodal task, errors such as mishearing the utterances and misinterpreting the gestures are inevitable. Hence, another concern arises about applying the augmentation on all modalities in such a way that it models those errors the same way a human naturally acts. The next challenge would be the need to have another agent acting as a human while training our robot agent. Inspired by Behavioral Cloning (BC) concept <cit.>, we tackle this problem by proposing our Black-Box Human Agent. However, we don't exactly perform BC for developing and training our human agent since we need to map the current state of human agent to it next (state, action) pair. Instead of ActNet, use HBATN. HBATN is much more than ActNet. HBATN ultimately provides robot agent policy. The premise is that HBATN is handcrafted, now we can extract/synthesize automatically. RL needs data and we don't have lots of it. User-agent simulator+reward function (high-level) compensate for the lack of data, but they are grounded in the data we have. How can we use our (small) ELDERLY-AT-HOME corpus to the maximum extent? Highlight the challenges of data augmentation given our interaction is multimodal. § RELATED WORK Machine learning has become an indispensable tool in robotics. The critical aspect of interaction with the environment makes RL a good candidate for learning a policy from sequential human-robot interaction data. After theoretical guarantees for Q-learning were published in the '80s <cit.>, RL algorithms have developed greatly, including Deep-Q-Learning, Actor-Critic, and Trust Region Policy Optimization. <cit.>. For example, in <cit.>, the authors demonstrate how RL agents can outperform humans in the Atari game environment. Inverse-RL is also popular since there is limited size data, and human guidance can significantly reduce the set of states needed to be explored. As an example, in <cit.> the authors introduce a preference-inference Inverse-RL model for assistive robots that learns different preferences by observing users performing tasks. In this work, we are interested in learning an agent for an intelligent assistive robot and to that end, we propose a Deep-Q-Learning (DQL) framework with a DAGGER warm-up<cit.> which helps us tune the RL reward function more efficiently. In RL, a mandatory component to complete the feedback loop for training is a simulated user or an interactive environment. In human-robot interaction, the simulated user (which mimics humans) should be reliable but also generate a variety of actions to ensure the RL exploration. The authors of <cit.> examine whether the human-given reward is compatible with the traditional RL reward signal. In <cit.>, authors assume the human agent is an expert in interacting with the robot, i.e., humans will always give actions designed to be easy-understood by the robot. In this case, the human agent's policy is a fixed and deterministic function of the robot's movements. In our case, this expert assumption for the user simulator is inaccurate since the user, who might have a disability and need help, can not be seen as an expert. Instead of giving the best decision, our proposed user simulator can capture the user's decision preference. One can personalize the user simulator by training it based on their own collected data. Our RL agent can always successfully assist these personalized users by acquiring information or taking the initiative. Several works exist that propose RL frameworks for training interactive dialogue systems to assist users <cit.>. For example, in <cit.>, a Deep Q-learning network architecture is proposed for a dialogue system that can help humans in accessing information or completing tasks. The agent is trained with a language generation machine as the user simulator. One of the main differences between this work and ours is that we work with actual human data to develop our user simulator. Besides, in dialogue systems, the agent is limited to a single modality (language), while in this paper, we deal with a multi-modal system which makes it more complex. Another related work is <cit.> in which Q-learning is employed to learn a personalized policy based on children's facial reactions and pose data in the class. Although they work with actual human data, they are only concerned with visual data from poses and faces captured by a camera. Throughout the recent literature, there is not much work that considers a multi-modal setting with gestures, physical actions, and dialogue while training the intelligent agent based on actual human data. In this paper, we build upon the HBATN framework proposed in <cit.> which allowed the agent to manipulate the objects in the environment and interact with users through pointing gestures, haptic-ostensive (H-O) actions <cit.>, and speech. This data-driven model was developed based on our previously collected corpus of human-human <cit.>. The task studied in detail was the Find task, an interaction scenario in which a human and a robot work together to find an object of interest in the environment. HBATNs model both agents simultaneously to maintain the state of a task-driven multi-modal interaction and plan the subsequent robot moves. Although this model provides an interpretable decision-making process (it is rule-based and constructed by hand), it lacks scalability. Also, extracting the policy manually is a daunting task. One of our main contributions in this paper is to automate this policy extraction process using RL. § USER SIMULATOR As the assistive robot can be explained in the See-Think-Act cycle in <ref>, our user simulator will take the place of the human agent on the left side, and the perception module and RL robot agent will be the interaction manager module. Especially, the user simulator takes a vector of Helper's dialogue, actions, and environment information as input and outputs the elder's dialogue and action to the RL robot agent. A critical component of the RL training cycle is the environment in which the agent to be trained can perform actions and receive feedback. In a collaborative setting, the environment also includes the participant that the agent interacts with. As illustrated in Fig. <ref>, our HEL agent decides on its next action based on the Perceived Human Intent. As a result, we need to provide the HEL agent with this input in the RL training cycle as well. Because the number of trials required to effectively perform RL grows significantly with the size of the agent's search space and putting a human user in the training cycle to provide the Perceived Human Intent input to the RL agent is neither realistic nor feasible we developed a user simulator. We focus on the Find task, for which we have annotated human-human interaction data. To train an agent to perform the role of HEL in the Find task using RL, we thus require a user simulator that can act as ELD and provide the same kind of response we observed in our data. We previously developed an HBATN to represent the states and actions of both participants in the Find (and more generally, a collaborative) task <cit.>. The task was decomposed into subtasks with the goal of identifying a target object type O_T, identifying locations L potentially containing the object, and, ultimately, locating the target object O. In <cit.>, we showed that our HBATN, equipped with a trained classifier that determines which subtask the participants are currently in, can model and perform both HEL and ELD behavior; however, training an agent using RL requires the agent to be able to explore its space of states and actions, some of which are never seen in our data, which can lead to erratic behavior that the subtask classifier was not trained to handle. We, therefore, propose an end-to-end neural network-based user simulator that can reasonably engage with the agent as it undergoes RL. As depicted in Fig <ref>, for training our Reinforcement Learning (RL) agent in an RL cycle, we need a User Simulator acting as the ELD agent. To build the user simulator, ELD Agent, we use the same ELDERLY-AT-HOME data on which we previously developed our HBATN <cit.>. We develop a Black-Box User Simulator which can take care of all sub-tasks at once and predict the state of the world, i.e. the state in which the ELD is as well as the ELD’s next move. The ELD agent is supposed to predict if and what move the ELD is going to take, the dialogue act (DA) and action, as well as the state of the world. The state of the world is determined by the ELD’s belief of the HEL’s object type (O_T), location (L), and object (O). §.§ Feature Extraction We require a state representation that can be the input to the user simulator (ELD) to reasonably determine an output as an appropriate response to the agent (HEL) action. We consider three main parameters: (1) ELD's belief of HEL's knowledge of O_T, which can be one of three values (ELD believes HEL does not know the target O_T, ELD believes HEL does know the target O_T, or ELD believes HEL is thinking of a different O_T), (2) ELD's belief of HEL's knowledge of L, and (3) ELD's belief of HEL's knowledge of O. We supplement these with additional features representing HEL's action that ELD responds to: (4) what or where HEL pointed to, if a pointing gesture was performed, (5) what H-O action HEL performed and on what, if an H-O action was performed, and (6) the Dialogue Act (DA) – roughly, intent indicated through speech – of HEL's utterance, which we have extensively studied and built classifiers for <cit.>. We assume that both ELD and HEL have a finite set of actions. We developed a list of actions for each role based on the Find task corpus so that HEL's action label could be input to the user simulator and so that the user simulator can generate its action labels as output. ELD actions include providing the O_T or the L and giving an affirmative or negative response, while HEL actions include requesting the O_T or the L and verifying a potential O_T, L, or O. We designed our user simulator to output ELD's next action label, the DA tag associated with ELD's action (to simulate speech recognition and classification when the agent interacts with a human user), and the updated state representation of ELD's beliefs of HEL's current knowledge of the task. The action label and DA tag can then be passed to the agent during training, while the updated state representation of ELD's beliefs get passed into the user simulator at the subsequent step. When interacting with a human user, the agent will need to be equipped with classifiers to determine the user's action label and DA given sensory input. This paper addresses the RL cycle, and we plan to implement and test a full human-robot interaction in our future work. We first determine the information our model needs to know in order to decide which state the ELD is in and what action the ELD is going to take. Then, we’ll extract the features from the data. The features extracted and used to train the ELD Agent are as follows: (1) whether the HEL is going to take consecutive moves; (2) whether or not an object type has been uttered; (3) whether or not a location is uttered; (4) the ELD’s belief of the HEL’s object type, could be 0, 1, 2. 0 is when it has not been determined yet, 1 is when it matches the ELD’s object type, 2 is when it mismatches; (4) the ELD’s belief of the HEL’s location, the same as previous feature could be 0, 1, 2; (5) the ELD’s belief of the HEL’s object, the same as previous feature could be 0, 1, 2; (6) whether or not a pointing gesture has been performed, If so, it is targeting an object or a location; (7) whether or not an H-O action has been performed, if so, it is targeting an object or a location and what kind of H-O it is; (8) the current HEL’s action tag; (9) the current HEL’s utterance, the DA tag. §.§ Data Annotation We turn to the Find task data available in the ELDERLY-AT-HOME corpus <cit.>. This data was previously transcribed and annotated for DAs, pointing gestures, and H-O actions. To train our user simulator with this data, we supplied additional annotations around ELD beliefs of HEL's knowledge of O_T, L, and O, and ELD and HEL actions using our list of actions. ELD beliefs were annotated based on the heuristic that ELD updates their belief whenever HEL gives some indication of knowing or not knowing O_T, L, or O. For example, ELD believes HEL does not know the L until HEL has acknowledged ELD giving the L or performs an action on the intended L, but ELD believes HEL is thinking of a different L if HEL asks for verification of the wrong L (e.g., ELD says, "Check down there," while pointing to a cabinet, but HEL asks, "Here?" while pointing to a different cabinet). ELD and HEL actions were labeled by two annotators. Because these action labels did not follow a strict heuristic and were, therefore, more subjective, we measured inter-annotator agreement on both types of actions using Cohen's kappa. We chose 40 random ELD actions and 40 random HEL actions from the corpus and had both annotators independently label them prior to annotating the remaining actions. We found a high level of agreement between the annotators for both ELD actions (κ=1.0) and HEL actions (κ=0.81), indicating that we have reliable action labels. §.§ Data Augmentation The Find task data provided a strong foundation for training our user simulator, but there were no instances in which ELD believed HEL had the wrong O_T or L in mind. The data also contained very few instances in which ELD believed HEL did not know O_T. Only in the beginning of each trial did this occur; after ELD provided the O_T, HEL never needed to ask for O_T again. The lack of these states is unsurprising given that the interactions were between humans who could hear and see each other well. However, when we train our HEL agent with RL, we expect that the agent will make mistakes, misinterpreting or not even interpreting O_T or L, so the user simulator needs to be able to respond appropriately. We, therefore, augment our data with synthetic but grounded examples that cover these missing or infrequent states. In the real world, the cases where the human misinterprets another human's actions would happen frequently. As a result, we expect that to be the case for ELD’s belief of the HEL’s object type (O_T), location (L), and object (O) in our data too. However, due to the small amount of data, there are not enough samples of these error cases in our data. To compensate for that and balance the data, we executed a data augmentation procedure before developing and training the model. To increase the number of states in which ELD believes HEL does not know O_T, we sample instances in which ELD believes HEL knows the O_T (regardless of whether ELD believes HEL knows the L that ELD wants to search) and replace HEL's utterance with an utterance from the data that corresponds with the action of requesting the O_T (see Fig. <ref>). The subsequent ELD action would then be to give the O_T again. To increase the number of states in which ELD believes HEL has the wrong O_T in mind, we sample instances in which HEL mentions the target O_T in their utterance and replace the O_T in HEL's utterance with an incorrect O_T (see Fig. <ref>). The subsequent ELD action would then be to give the O_T again. Analogously, we increase the number of states in which ELD believes HEL is not thinking of the same L or O by sampling instances in which HEL performs an action on an object or location and replacing the object or location with an incorrect one, for which the subsequent ELD action would be to give the L or the O_T again (see Fig. <ref>). * (O_T, L, O) = (0, 0, 0) and (0, 1, 0); we take random HEL's actions and replace utterance with a random (0, 0, 0) /(0, 1, 0) utterance. * (O_T, L, O) = (2, 0, 0) and (2, 1, 0); we take random HEL's actions containing the target O_T and replace O_T in utterance with a random different O_T. * (O_T, L, O) = (1, 2, 0) and (1, 1, 2); we randomly replace L and O in pointing and H-O actions with a different L or O. With this data augmentation scheme, we increase the number of data points we have from 693 to 1101. §.§ Model Architecture and Training Our designed network includes two fully connected (dense) layers with ReLU activation function, and a dropout layer (ratio=0.5) for better generalization and to avoid over-fitting. Cross-Entropy loss function and Adam optimizer are used for optimization. The model is trained for at most 100 epochs. However, evaluating the model on the validation set is performed alongside the training so training can be stopped early. The inputs to the network are the features described in the previous section and the outputs are the ELD's next state, DA, and action which are manually annotated in the data. While training, we adopt sample re-weighting to resolve the class imbalance issue. Pythorch library is used for implementation <cit.>. In addition to overall accuracy, we evaluated the model on the classification accuracy of each individual output of the model; i.e. the classification accuracy for: (1) the predicted belief ELD holds of HEL's knowledge of O_T, L, and O; (2) the predicted DA; (3) the predicted action. The accuracy results are provided in tables <ref> and <ref>. Pythorch library is used for the model implementation. The design of the network is as follows: a two-layer feed-forward Pythorch network; dropout=0.5 for better generalization and to avoid over-fitting; and Relu activation function. The relatively low overall accuracy results from the low action classification accuracy and DA classification accuracy. The main reason is still the small amount of our data. We do not want to overdo the data augmentation since having a large proportion of data generated this way becomes counterproductive. Moreover, our action accuracy and DA accuracy are overly strict since the classifier output is considered correct only when the predicted tag matches the ground-truth label. However, the predicted DA could be acceptable if it still conveys the right intent. For instance, consider the case where the HEL utters: “Where should I look for the cup?” and the simulator predicts “Instruct” and “Give O_T, L” for the DA and action, respectively, while the labels are “Reply-w” (meaning that the user replies to a Where or What question) and “Give L”. The simulator output would be labeled as wrong, but it does convey the right intent. We thus consider our current user simulator to be adequate because what is important is the interaction between the user simulator and the RL agent. As long as the simulator conveys the user's intent correctly to the RL agent, the essential prerequisites for training the RL agent are taken care of. Improving the accuracy of the user simulator is an important research problem, but it is beyond our current study and will be considered in the future. § REINFORCEMENT LEARNING FRAMEWORK In this RL problem, we employ a Deep-Q-Network <cit.> model to act as our assistive robot (HEL) agent by having it interact with the interaction environment containing the user simulator proposed in the previous section. To enhance the training, we warm up the agent before starting the training loop using DAGGER, an Imitation Learning (IL) algorithm <cit.>. The warm-up phase happens before initiating the RL training cycles. This step is crucial due to the fact that learning the appropriate agent behaviors with a simple reward function is extremely difficult and training such an agent from scratch is tedious. The flow of the interactions between the HEL agent and the user simulator while training the HEL agent is illustrated in Fig. <ref>. The state of the HEL agent consists of three variables: the state of HEL's O_T, HEL's L, and HEL's O. The possible values for each variable are 0, 1, or 2. The variable is encoded with 0 when it has not been determined yet, with 1 when it matches the user simulator’s belief about it, and with 2 when there is a mismatch between the HEL value and user simulator’s belief. As depicted in Fig. <ref>, the HEL agent takes the user simulator's action and the previous HEL's state as input and outputs a DA tag and an action vector encoding the HEL's utterance and physical action, respectively. We develop and train a Deep-Q-Network <cit.> acting as our HEL agent by having it interact with the ELD agent we proposed in the previous section. We also need to warm up the RL agent before starting the RL training using an Imitation Learning (IL) algorithm. We do the warm-up phase before RL training because modeling human behavior using a simple reward function and training from scratch is very hard and not efficient at all. As a result, we take advantage of DAGGER IL algorithm <cit.> in this regard. The flow of the interactions between HEL and ELD agents while training the HEL agent is illustrated in Fig. <ref>. The state of the HEL agent consists of three variables: (1) the state of HEL's O_T could be 0 when it has not been determined yet, 1 when it matches the ELD’s object type, 2 is when it mismatches the ELD’s object type; (2) the state of HEL's L could be 0, 1, 2, same conditions as HEL's O_T apply; (3) the state of HEL's O could be 0, 1, 2 the same as HEL's O_T, L. §.§ Model Architecture The HEL agent network includes two fully connected layers followed by a dropout layer (ratio=0.1). Then the output of the dropout layer is fed to the output layer followed by ReLU activation and gives a vector encoding HEL's (DA, action) output pair. For implementation, we used the Pythorch library <cit.>. Pythorch library is used for implementing this framework. The design of the HEL agent's network is as follows: a two-layer feed-forward Pythorch network; dropout=0.1 for better generalization and avoiding over-fitting; and Relu activation function. §.§ DAGGER Warm-up First, we have our HEL agent interact with the user simulator and run the DAGGER algorithm as a warm-up stage. We don't let the agent get fully trained, we only use it to obtain a good initial guess for the subsequent RL training. Using DAGGER only, without subsequent RL training, results in poorer performance of the HEL agent. The human simulator initiates the interaction, the state of the HEL agent is updated according to the received input and it picks an action according to its current state and the user simulator's action. An error module is deployed before passing the HEL's state to the HEL agent. The error module is necessary due to the fact that in the data when O_T/L/O information is given to the HEL by the ELD, most of the time the state corresponding to that changes to 1, i.e. the HEL's understanding of O_T/L/O is the same as what human has uttered. However, we would like to generalize the framework better so that it also covers the states when the HEL's understanding of O_T/L/O does not match the human's utterances. Thus, the error module in 25% of the cases where HEL's understanding of O_T/L/O is 1, changes that to 2. First, we have our HEL agent interact with the human simulator and run the DAGGER algorithm as a warm-up stage. At this stage we don't let the agent get fully trained, we only push the policy network weights toward the expert's policy extracted from the ELDERLY-AT-HOME data. The previously trained ELD agent initiates the interaction by taking a random action, based on the action ELD takes, the state of the HEL agent is updated. Subsequent to that, the HEL agent takes an action getting its current state and the ELD's action as the inputs. An error module is deployed before inputting the HEL's state to the HEL agent to add errors to HEL's (O_T, L, O) state. The error module is devised here due to the fact that in the data when O_T, L, O are uttered to the HEL, most of the time the state corresponding to the changes to 1, i.e. the HEL's understanding of O_T, L, O is the same as what ELD has uttered. However, we would like to generalize the framework better so that it also covers the states when the HEL's understanding of O_T, L, O doesn't match the ELD's utterances. Thus, the error module in 25% of the cases where HEL's understanding of O_T or L or O is 1, changes that to 2. For training the HEL agent using DAGGER algorithm, we need to run the algorithm for N iterations. Here we call each iteration one Episode since the interactions between the ELD and the HEL are defined to be episodic. During each episode, one entire interaction consisting of at most M turns between the HEL agent and the user simulator takes place. That interaction is successful if the HEL agent finds the object the user simulator requested before reaching turn M. Otherwise, the interaction is unsuccessful. During each episode, the HEL agent executes the current learned policy. Throughout execution, at each turn, the expert's action, which we get from the roll-outs extracted from our ELDERLY-AT-HOME data, is also recorded but not executed. After sufficient data is collected, it is aggregated together with all of the data that was previously collected. Eventually, the cross-entropy algorithm generates a new policy by attempting to optimize performance on the aggregated data. This process of execution of the current policy, correction by the expert, and data aggregation and training is repeated. Fig. <ref> shows training loss, success rate, and average turns during each episode for DAGGER training on the agent for 50 episodes. However, as mentioned before, we do not want a fully DAGGER trained HEL agent. Therefore, the HEL agent's network weights are saved at episode 8 (a good mid-point where the loss is not minimized and the success rate is not maximized yet) to be used later on for training the HEL agent using the Deep-Q-Learning (DQL) algorithm. §.§ Deep-Q-Learning For training our HEL agent in the RL cycle, we need to initialize a Policy Network and a Target Network. The former network's weights will be optimized for obtaining the optimal policy, and the latter network will be used to track the target Q-values associated with each individual action for input states <cit.>. The policy we extracted at episode 8 of DAGGER training is used to initialize the weights of the Policy Network and the Target Network for DQL. The model architectures for the Policy Network and the Target Network are exactly the same as the model used previously for DAGGER training due to the fact that we simply copy the weights extracted from DAGGER network into these two networks. This makes running the DQL algorithm on our RL framework much more efficient in terms of time and space. Another advantage of warming up the HEL agent by DAGGER training is that by pushing the weights of the policy toward the expert's policy, we don't need to hard code complex human behavior in the Reward Function. A simple Reward Function combined with DAGGER-half-trained HEL agent makes the DQL algorithm on our HEL agent run much faster. For our reward function, we considered a small negative reward, -r, for each HEL agent's move. This is to motivate the agent to finish the task sooner than later. If the interaction is unsuccessful and ends by getting to turn M, the transition is penalized by -2 r. If the interaction successfully ends before reaching turn M, that transition is rewarded as 2 r. We also set some ground rules as Preconditions. The Preconditions are as follows: (1) the HEL agent must not take the action of verifying O_T before O_T is uttered by the user simulator; (2) the HEL agent must not take the action of verifying L before L is uttered by the user simulator; (3) the HEL agent must not take the action of verifying O before both O_T and L are uttered by the user simulator. If any of the Preconditions are violated, that action is penalized with a very large negative number -Z. Every action the HEL agent takes is rewarded with a small negative number, -r. If the interaction is unsuccessful and ends by getting to turn M, that transition is rewarded as -2*r. If the interaction is successfully ended before reaching to turn M, that transition is rewarded as 2*r. We also set some ground rules as Preconditions. The Preconditions are as follows: (1) the HEL agent must not take the action of verifying O_T before O_T is uttered by the ELD agent; (2) the HEL agent must not take the action of verifying L before L is uttered by the ELD agent; (3) the HEL agent must not take the action of verifying O before O_T or L are uttered by the ELD agent. If any of the aforementioned Preconditions are violated, that action is rewarded a very large negative number, -inf. The interactions between the HEL agent and the user simulator while running the DQL algorithm also follow the flow in Fig. <ref>. A tuple of (state, reward, action, next state) is stored in the memory at each turn. Every C episodes, the weights of the Policy Network are optimized using the Mean-Squared-Error cost, and every m C episodes, the weights of the Policy Network are copied into the Target Network. Fig. <ref> shows training loss, average reward, success rate, and average turns at each episode for the full DQL algorithm training of the HEL agent. The reason that average turns start to increase and the success rate gradually decreases after about episode 1250 is that the memory of the agent has been filled up, so part of the memory is emptied and starts to get filled with new samples. The agent has been over-fitted at this point. We thus use the policy obtained just before this phenomenon occurs. § EXPERIMENTAL EVALUATIONS To evaluate our framework in the real world, we designed a user interface to have humans interact with the HEL agent obtained through RL. In the following section, we explain the details of our interface as well as the evaluation procedure. §.§ User Interface Implementation Since our framework is trained on multi-modal human data, it is capable of interpreting and executing multi-modal actions, and in particular language and physical actions. However, in this work, we focus on language so the Perceived Human Intent vector that results from human subjects is limited to their utterances. The HEL agent is also capable of performing pointing and H-O actions in addition to communicating through language. However, in our evaluation, the agent generates the speech based on its predicted DA, and then also informs the human about the other modalities through speech. For instance, when the agent is verifying if the human uttered “drawer" with a pointing gesture it announces: “The agent points to the drawer." For the human subjects to be able to have a smooth interaction with our agent, we developed and implemented a friendly user interface following the architecture for the Perception Module and the Execution Module in Fig. <ref>. Since the HEL agent is trained based on DA tags and the action vector generated by the user simulator, it expects to get DA tags and action vectors of the human subject in the test phase as well. As a result, we need to extract the DA tags and the action vectors from the human utterances. Therefore, we have Speech-to-Text, Action Extractor, and DA Classifier components which are explained in detail later on. And for the agent to respond back to the human we first need to transform its DA tag and action vector into a sentence and a textual description of the action, respectively. To this end, we developed a rule-based text generator inspired by the ELDERLY-AT-HOME corpus. Finally, the Pyttsx3 Python library <cit.> is used to transform the agent's sentences into speech. Human speech is passed to the Google Cloud Speech-to-Text API <cit.> for speech recognition. Then, the transcribed text is passed to a DA classifier as well as an action extractor. We developed our DA classifier by deploying Sentence-BERT <cit.> and extracting embeddings from our data as the input to the classifier. We then trained a two-layer (one hidden layer followed by the output layer) feed-forward Pythorch neural network, with a dropout layer (ratio=0.1), and ReLU activation function after output layer. We used the gold-standard DA tags as the ground-truth labels and optimized the weights by employing the cross-entropy loss function and Adam optimizer. The accuracy of our Sentence-BERT-based DA classifier tested on 10% of ELDERLY-AT-HOME data is 67.23%. Our action extractor module works based on the objects and locations it extracts from human utterances. For example, if it detects one of the objects in the utterance, that's labeled as “Give O_T" action; if it detects one of the locations in the utterance, that's labeled as “Give L" action; if it detects one of the objects as well as one of the locations in the utterance, that's labeled as “Give O_T, L" action. For extracting these words of interest, we built a dictionary based on NLTK dictionary <cit.> consisting of all the words that could be pronounced similarly or close to our words of interest (to compensate for the speech recognition returning similarly sounding words that do not make sense in the context). §.§ User Study We performed a human user study where 9 healthy adults were recruited to interact with our HEL agent. Each subject performed 4 to 5 trials (entire interactions) with the HEL agent adding up to a total of 42 trials. The hypothetical experiment environment would be a room with a drawer, a shelf, and a cabinet. The user can choose between red, green, and yellow cups and red, green, yellow, and white balls. At the beginning of each trial, objects are randomly scattered in different locations. The user only knows there are aforementioned locations and objects in the room but doesn't know which item is located where. The subjects were instructed to choose the object of interest at the beginning of the trial, and guide the agent through different locations to find the object of interest. The data collected in this user study will be added to our publicly available ELDERLY-AT-HOME corpus. §.§ Results Inspired by our previous study <cit.>, we measured the performance of the system by calculating the accuracy of each component independently as well as the overall quality of the system. For evaluating speech recognition we measured Speech-to-Text (STT) accuracy which is the percentage of words transcribed correctly. We report the average Speech-to-Text confidence score as well. To evaluate the DA classification, we calculated the accuracy based on the percentage of the DAs classified correctly. For obtaining the ground-truth gold-standard DA labels, two annotators annotated the texts that were previously extracted from the Speech-to-Text component. We chose 40 random sentences and had both annotators independently label them prior to annotating the remaining sentences. We calculated Cohen's kappa and found a moderate level of agreement between the two annotators (κ=0.6). Although it's not a high level of agreement, the DA accuracy still is consistent with the test accuracy (67.23%) we got from ELDERLY-AT-HOME data before. For evaluating the HEL agent itself, i.e. whether or not the pair of (DA, action) outputs of the HEL agent makes sense according to the DA tag and the action vector it gets as inputs, we compare its actions to those of HBATN's (as our ground truth labels) and report the action accuracy. Moreover, we calculate and report the percentage of the HEL's non-eligible actions with respect to the human's original speech. We had two human transcribers listen to the audio data recorded from human subjects. Their transcriptions from audio data matched about 99.2%. A non-eligible action is one where the agent's response does not correspond to the human's previous action. For example, if the human says: “Please get a cup." and the agent responds with: “Did you say inside the cabinet?", this is a non-eligible action. In other words, non-eligible actions are those actions that do not make sense to the human participants in the study based on the speech they uttered. Non-eligible actions could result from speech recognition errors, DA classification errors, and/or wrong actions by the agent itself. We also report the average length of interactions as well as the success rate over all trials. A successful trial is a trial in which the agent is able to find the human subject's object of interest in less than 15 turns. If the agent reaches the 15th turn without finding the object, that trial is unsuccessful. Eventually, we asked the participants to rate their experience on a Likert scale of 1-10. A score of 1 meant “significantly worse than expected" and a score of 10 meant “significantly better than expected". The results of our preliminary user study in Table <ref> show that our robot agent does very well in extracting and performing the optimal policy. Two very important metrics to evaluate how well our overall framework performs are the average success rate of 92.86% and the number of average turns of 8.38. There are a couple of reasons why the success rate doesn't reach 100%. The main reason is that our framework is not solely the robot agent itself. The user interface is implemented on top of the agent network and each component of this user interface introduces errors to the overall framework. These errors make our robot agent take some wrong actions occasionally which leads to taking more turns to complete the task. The average non-eligible action rate shows that the HEL agent in 30.57% of the cases takes an action that does not make sense to the human at that point. The action accuracy of 90.2% indicates that 9.8% of the actions the robot agent takes are wrong actions. This small rate of wrong actions by the robot agent itself is due to the fact that compared to HBATN, our agent is trained through machine learning, not rule-based. It covers a broader set of states while getting trained and as a result when it comes to testing it, the process of making decisions is more complicated. About 20% of all actions are non-eligible actions resulting from errors in other components. To investigate this more, we calculate the contribution of each component to those wrong actions. Table <ref> summarizes the contribution of each component to those wrong actions taken by the robot agent. 46.73% of non-eligible actions (corresponding to about 14% of all actions) resulted from errors made by the speech recognition component. The DA classifier also causes 21.49% of non-eligible actions (corresponding to about 6% of all actions). All in all, we argue that the success rate of 92.86% is remarkable, and the average number of turns of 8.38, and the average Likert score of 7.9 lead us to conclude that our RL framework for multimodal human-robot interaction performs exceptionally well. Things to discuss that are not covered in the evaluation/results section. Action accuracy: the match is 90%, what about the remaining 10%. It comes from the robot. Some thoughts on how to make it better. Thought: we are trying to make the agent better, but we lack data. That is not quite enough, HBATN helps with that, but RL needs a bit more help. One thing: hard to express human behavior with a simple reward, we don't want to hand-craft it. But inverse reinforcement learning may be useful. In that case, how to get around not having enough data? Is human simulator similar to imitation learning/behavior cloning? § CONCLUSION In this paper, we propose a Reinforcement Learning (RL) approach to automate the policy synthesis for a multimodal interactive service robot. The agent trained by this algorithm acts as the interaction manager of our assistive robot in collaborative tasks. The main advantage of this approach over the hand-crafted rules is that it can be applied to other service robots only by adapting the features according to the available data and the required task. Moreover, this approach is more efficient in terms of time and space. We employ the Deep Q-Learning network with a DAGGER warm-up to train the model. During the RL training process, the agent interacts with a human simulator through multiple modalities. In contrast to most of the RL-based models in dialog systems, which mainly work with a single modality (speech), we developed our own human simulator using our previously collected data, the ELDERLY-AT-HOME corpus. Our human simulator is an end-to-end neural network predicting the human's state and action according to the human's previous state, HEL's Dialogue Act, and HEL's action. To evaluate our trained model, we have conducted a human study and assessed user satisfaction and system performance. IEEEtran

http://arxiv.org/abs/2303.07066v1

20230313124140

The magnetic sensitivity of the (250-278nm) Fe II polarization spectrum

astro-ph.SR

dafonso@iac.es Instituto de Astrofísica de Canarias, E-38205 La Laguna, Tenerife, Spain Departamento de Astrofísica, Universidad de La Laguna, E-38206, La Laguna, Tenerife, Spain Instituto de Astrofísica de Canarias, E-38205 La Laguna, Tenerife, Spain Departamento de Astrofísica, Universidad de La Laguna, E-38206, La Laguna, Tenerife, Spain Instituto de Astrofísica de Canarias, E-38205 La Laguna, Tenerife, Spain Departamento de Astrofísica, Universidad de La Laguna, E-38206, La Laguna, Tenerife, Spain Consejo Superior de Investigaciones Científicas, Spain This paper presents a theoretical investigation on the polarization and magnetic sensitivity of the near-ultraviolet (near-UV) solar spectral lines of Fe2 between 250 and 278 nm. In recent years, UV spectropolarimetry has become key to uncover the magnetism of the upper layers of the solar chromosphere. The unprecedented data obtained by the CLASP2 suborbital space experiment across the Mg2 h and k lines around 280 nm are a clear example of the capabilities of near-UV spectropolarimetry for the magnetic field diagnostics throughout the whole solar chromosphere. Recent works have pointed out the possible complementary diagnostic potential of the many Fe2 lines in the unexplored spectral region between 250 and 278 nm, but no quantitative analysis of the polarization and magnetic sensitivity of those spectral lines has been carried out yet. To study the polarization signals in these spectral lines, we create a comprehensive atomic model including all the atomic transitions resulting in strong spectral lines. We then study the magnetic sensitivity of the linear and circular polarization profiles in a semi-empirical model representative of the quiet sun. We present a selection of Fe2 spectral lines with significant linear and circular polarization signals and evaluate their diagnostic capabilities by studying their formation heights and magnetic sensitivity through the action of the Hanle and Zeeman effects. We conclude that when combined with the CLASP2 spectral region these Fe2 lines are of interest for the inference of magnetic fields throughout the solar chromosphere, up to near the base of the corona. § INTRODUCTION Spectral lines encode information about the physical conditions of the emitting plasma. Therefore, the study of their formation in given models of the solar atmosphere is very important for reaching the goal of inferring its physical properties from the observed spectral line radiation. In particular, the linear and circular polarization profiles carry information about the magnetic field strength and geometry in the their region of formation. Even though the solar spectrum is populated by plenty of spectral lines forming everywhere in the solar atmosphere, it is in the ultraviolet (UV) region of the spectrum where we find abundant spectral lines forming in the solar chromosphere and transition region. Determining the magnetic field in these regions of the upper solar atmosphere is of fundamental importance for understanding how the energy is transported from the underlying layers and dissipated in the chromosphere and the corona, where the explosive events that can impact the heliosphere take place. For these reasons, ultraviolet spectropolarimetry is increasingly recognized as a key diagnostic tool for facilitating new breakthroughs in our empirical understanding of solar and stellar magnetic fields<cit.>. The Interface Region Imaging Spectrograph <cit.> mission is one of the main examples of the relevance of UV spectroscopy for investigating the upper layers of the solar atmosphere. IRIS has been in operation since 2013 measuring the intensity profiles in several windows of the UV solar spectrum, including spectral lines such as Mg2 h and k (279.6 and 280.3 nm), C2 (133.4/133.5 nm), or Si4 (139.4/140.3 nm). The research exploiting the data provided by the IRIS mission has significantly improved our knowledge about the solar chromosphere <cit.>. Concerning UV spectropolarimetry, the most recent relevant advances are the Chromospheric Lyman-Alpha SpectroPolarimeter <cit.>, which in 2015 observed the intensity and linear polarization of the H1 Ly α line <cit.>, and the Chromospheric LAyer SpectroPolarimeter <cit.>, which in 2019 acquired unprecedented spectropolarimetric data of the 279.2 - 280.7 nm near-UV spectral region that contains the Mg2 h and k lines and several other lines of great diagnostic interest <cit.>. These suborbital space missions were motivated by a series of theoretical investigations <cit.>, which predicted that the polarization signals are measurable and magnetically sensitive and showed the importance of the effects of partial frequency redistribution (PRD), J-state interference, and the joint action of the Hanle, Zeeman, and magneto-optical effects on the linear polarization signals produced by scattering processes. Moreover, the analysis of the circular polarization profiles measured by CLASP2 in the resonance lines of Mg2 and Mn1 allowed for the first mapping of the longitudinal component of the magnetic field from the photosphere to the upper chromosphere, just below the transition region <cit.>. Despite all these novel advances, there are still many unexplored regions in the UV spectrum of the Sun. The polarization signals of the near-UV Fe2 spectral lines that are located blueward of the CLASP2 spectral region have been proposed as potentially useful for facilitating the exploration of chromospheric magnetism <cit.>. In particular, in the 250 - 278 nm spectral window we can find many lines belonging to this ion, including a strong resonant multiplet at 260 nm. Due to the lack of spectroscopic or spectropolarimetric solar observations in this spectral region, previous works found in the literature are mainly focused on the study of the intensity spectrum of these spectral lines in other stars <cit.> or of other Fe2 spectral lines located in observed spectral regions, such as the small Fe2 emission line in the red wing of the Ca2 line at 369.94 nm <cit.>. Moreover, as far as we know, there are not theoretical investigations about the polarization signals of these spectral lines or about their magnetic sensitivity. In this work we present the first detailed theoretical study on the polarization and magnetic sensitivity of the Fe2 lines that are located in the 250 - 278 nm near-UV spectral region, as well as their suitability to infer magnetic fields in the solar atmosphere. In Sec. <ref> we introduce the methods and tools used in our radiative transfer modeling. The atomic model used in the calculation is described in Sec. <ref>. In Sec. <ref> we show the results of our modeling, including a selection of the most interesting spectral lines, following certain criteria, as well as an analysis of their magnetic sensitivity and suitability to infer magnetic fields in the solar atmosphere. Finally, present our conclusions in Sec. <ref>. § FORMULATION OF THE PROBLEM To study the formation and magnetic sensitivity of the Fe2 lines in the spectral window between 250 and 278 nm we solve the problem of generation and transfer of polarized radiation in optically-thick magnetized plasmas out of local thermodynamic equilibrium (NLTE). This problem is both non-local and non-linear, requiring the simultaneous solution of the atomic excitation (populations of the atomic substates and quantum coherence between them) described by the statistical equilibrium equations (SEE) and of the radiation propagation throughout the atmosphere governed by the radiative transfer (RT) equations. We solve this problem by applying the HanleRT code <cit.> to a 1D plane-parallel model of the solar atmosphere. We have chosen the semi-empirical C model of the quiet Sun described in <cit.>. Even though PRD effects and quantum interference can have an impact on the linear scattering polarization profiles <cit.>, we neglect both for the following reasons: first, these effects are particularly relevant in the far wings of these spectral lines and, as we discuss in more detail in Sec. <ref>, in the near-UV region of the solar spectrum the effect of spectral line blanketing is notably strong, masking the impact of these effects and making the use of the line wings for atmospheric diagnostic extremely difficult. Second, including these physical ingredients has a significant computational cost, especially when considering the size of the atomic model needed to study all the spectral lines considered in this work. For these reasons, the calculations shown in this paper assume complete frequency redistribution (CRD) and neglect quantum interference between the substates pertaining to different atomic levels, which are often good approximations to model the spectral line core. § ATOMIC MODEL As far as we know, there are not previous studies on the formation of the Fe2 polarization profiles in the spectral window between 250 and 278 nm and, therefore, on the atomic model necessary to reliably model the polarization of all these spectral lines. Consequently, the first step is to build a suitable atomic model including all the atomic transitions that could produce (or impact) any interesting polarization signal in the spectral region of interest. In the solar chromosphere the iron atoms are mostly singly ionized, and, consequently, Fe2 is the dominant ion <cit.>. Our atomic model thus includes just the Fe2 levels and the ground level of Fe3. We take the atomic level energies and the Einstein coefficients for spontaneous emission from the NIST database <cit.>. We approximate the photoionization cross sections as hydrogenic <cit.>. The rates of ionizing inelastic collisions with electrons are calculated using the approximation of <cit.>. The rate of excitation due to inelastic collisions with electrons are calculated following <cit.> for pairs of levels fulfilling the electric dipole selection rules, and following <cit.> for pairs of levels that do not fulfill those selection rules. The rate of depolarizing collisions with neutral hydrogen atoms are calculated assuming the Van der Waals potential <cit.>. We neglect inelastic collisions with hydrogen atoms. In our atomic model we only include those transitions with a spontaneous emission probability above a certain threshold (A_ul≥ 10^6 s^-1), which can produce a significant (in terms of its depths in the emergent spectrum) spectral line. Moreover, once we have solved the self-consistent RT problem with the whole atomic model we then remove one third of the levels, the least populated ones, as well as those not radiatively connected with any other atomic level. Using the reduced atomic model results in relative differences of no more than 5% at the center of the emergent intensity profiles for the spectral lines in the 250 – 280 nm range. Our atomic model to study the formation and polarization of the Fe2 spectral lines between 250 and 278 nm has 453 Fe2 atomic levels (and the ground level of Fe3) and 2225 radiative transitions. The model is thus complete enough to ensure a reliable calculation of the population balance of the levels involved in the formation of any of the spectral lines in the range between 250 and 280 nm. In Fig. <ref> we show the Grotrian diagram for the described atomic model, including all the radiative transitions. The red line in the diagram indicates the resonant multiplet that produces a series of intense spectral lines around 260 nm. A detailed Grotrian diagram of this multiplet is shown in the inset at the top-left corner of Fig. <ref>. The blue and green lines highlight some spectral lines analyzed in detail in Secs. <ref> and <ref>. The rates of inelastic collisions are critical for the population balance of the Fe2 atomic levels <cit.>. To build our atomic model we use very rough approximations to these rates, what can have a significant effect on the emergent Stokes profiles. In order to study how the inelastic collisions affect the emergent profiles, we create a reduced Fe2 atomic model with 90 levels (including the Fe3 ground level) and 113 radiative transitions, which can correctly model the emergent intensity profiles of the resonant multiplet around 260 nm; that is, it is able to reproduce the same emergent intensity profiles for these strong resonant lines (Fig. <ref>). Using the reduced model, we perform different numerical experiments in which we multiply the inelastic collisional rates between pairs of atomic levels fulfilling (not fulfilling) the electric dipole selection rules with a factor Ω_A (Ω_F). Due to the lack of solar spectroscopic observations, it is not possible to compare our calculations with the solar spectrum at these wavelengths. We have thus opted for a comparison with the flux of a solar analog star, α Cen A, a very close G2V star <cit.>, the same spectral type than the Sun. We compare the intensity flux profile of α Cen A[ This spectrum is available in The Advanced Spectral Library Project (ASTRAL, <https://archive.stsci.edu/prepds/astral/>). For further details see <cit.>] observed with the Space Telescope Imaging Spectrograph <cit.> on board of the Hubble Space Telescope (HST) with calculations in the FAL-C model for different values of Ω_A and Ω_F (Fig. <ref>). The first evident difference between the theoretical and observed fluxes is the significantly diminished continuum intensity in the observation, showing considerably wider lines. The large amount of spectral lines in the UV spectrum, contributing to the opacity, produce the so called UV blanketing, which consistently reduces the intensity flux across the spectral region, giving the impression of a much lower continuum and of wider spectral lines. Consequently, we focus our analysis to the core and near wings of the Fe2 spectral lines. With the inelastic collisions given by the above-mentioned approximations (Ω_A=Ω_F=1.0; blue curve) the flux in the line core is significantly larger than the observed flux, with significant emission features in the near wings which are absent in the observation. We find the best agreement when all the inelastic collisional rates are reduced by a factor of ten (Ω_A=Ω_F=0.1; red curve); the intensity flux at the line cores becomes remarkably similar to the observed flux and the near wing emission features completely vanish (see bottom row of Fig. <ref>). Our atomic model is thus able to adequately reproduce the intensity flux of the most intense Fe2 atomic lines in the near-UV spectrum of a solar-like star. § RESULTS We use the atomic model described in Sec. <ref> to model the Stokes profiles of the Fe2 atomic lines in the UV spectral window between 250 and 278 nm. In particular, we study their linear and circular polarization profiles and their capability to uncover the magnetic fields of the solar atmosphere. For this purpose, we perform RT calculations including scattering polarization and the Hanle and Zeeman effects assuming CRD. We study the fractional linear polarization signals for a line-of-sight (LOS) near the solar limb (μ=cosθ=0.1, with θ the heliocentric angle). We then impose different magnetic fields to investigate the magnetic sensitivity of their linear polarization. Finally, we study the formation of Zeeman-induced circular polarization signals for the disk center LOS. We then study the suitability of the weak-field approximation (WFA) to infer the longitudinal component of the magnetic field (B_L) from the circular polarization. §.§ Linear Polarization Signals. Fig. <ref> shows the fractional linear polarization Q/I profile in the unmagnetized FAL-C model atmosphere for the spectral region of the Fe2 resonant multiplet around 260 nm for the LOS at μ=0.1 (see blue curve in Fig. <ref>). While some of the strongest lines, in terms of their absorption features in the intensity profile, hardly show any significant linear polarization signal (see lines at 259.914, 261.265, or 261.840 nm) others exhibit relatively large signals shown in detail in the bottom panels of Fig. <ref>. The colored background in Fig. <ref> shows the logarithm of the normalized contribution function for Stokes Q, which gives a measure of how much each height layer in the model atmosphere contributes, at each wavelength, to the linear polarization of the emergent radiation. The red and blue colors indicate positive or negative contributions, respectively. We can conclude, from the contribution function, that the linear polarization at the line core of the lines shown in the figure is formed between the middle and upper chromosphere (above ≈ 1 Mm in the FAL-C model). We have selected the spectral lines with zero-field scattering polarization signals Q/I>0.2 % at the line core, and we have analyzed their magnetic sensitivity in order to determine which ones are suitable for the diagnostics of chromospheric magnetic fields. In the left panel of Fig. <ref> we show the height where the optical depth is unity at the line center (which gives an idea of the region of formation) for the selected lines, as well as their Hanle critical field (B_H), their line center zero-field Q/I signal, and information about the blends with neighbor lines. The Hanle critical field is the magnetic field strength for which the Zeeman splitting equals the natural width of the line <cit.>: B_H = 1.137·10^-7/t_ lifeg , with t_ life and g the radiative lifetime and Landé factor of the line level under consideration, respectively. Typically, a spectral line is sensitive, via the Hanle effect, to magnetic fields with strengths between approximately 0.2B_H and 5B_H. To characterize the blends with other spectral lines we compare the theoretical and observed fluxes from Fig. <ref> with a synthetic profile including all the transitions from the Kurucz database in LTE <cit.>, similarly to the approach of <cit.>. Through visual inspection, we look for any spectral line close enough to the selected Fe2 line to affect their profiles, and we classify them in three categories: no nearby blends (circles in Fig. <ref>), a moderate blend in the wings, far enough from the line core (squares), or a significant blend close to the line core (diamonds). Many of the selected lines show a nearby blend but, in these cases, the line core (line region sensitive to the Hanle effect) should not be affected significantly. Most of the spectral lines in our selection form in the middle chromosphere (around 1.25 Mm in the FAL-C semi-empirical model). A relatively small number of lines form in the lower chromosphere and near the FAL-C temperature minimum (around 0.5 Mm). Moreover, several stronger lines form in the upper chromosphere (above around 1.8 Mm). These latter transitions belong to a resonant multiplet (260.017 and 262.907 nm) and the metastable multiplets a^4D - z^4P^∘ (258.336 and 259.231 nm), a^4D - z^4D^∘ (274.035 nm), and a^4D - z^4F^∘ (274.730, 275.013 and 275.658 nm). Most of the lines in our selection have critical Hanle fields below 50 G and, for those lines forming in the upper chromosphere, we always find B_H<30 G. Consequently, these spectral lines are thus expected to be sensitive to magnetic fields with strengths considerably smaller than 100 G and, for those lines forming in the upper chromosphere, to magnetic fields around 20 G or less. These magnetic field strengths are of the order we expect for the chromospheric magnetic field in the quiet Sun and, therefore, these spectral lines seem to be good candidates to study the quiet Sun magnetism in the upper chromosphere. To actually study the magnetic sensitivity of these spectral lines, we have selected five specific cases, choosing lines forming in different layers of the solar atmosphere and with different Hanle critical fields. Table <ref> summarizes the atomic data of the five chosen transitions, also highlighted in blue in the Grotrian diagram of Fig. <ref>. In the FAL-C model we impose a horizontal magnetic field (θ_ B=90^∘, ϕ_ B=0, with θ_ B and ϕ_ B the polar and azimuthal angles of the magnetic field vector with respect to the local vertical) with different strengths, and we compare the calculated Q/I profiles of the radiation emerging at the μ=0.1 LOS (see Fig. <ref>): 0 (black), 10 (blue), 20 (green) and 50 G (red). ccccc Wavelength, level configuration and energy, Einstein coefficient for spontaneous emission, and critical Hanle magnetic field of the upper level for the spectral lines studied in detail in Sec. <ref>. λ [nm] Transition Energies [cm^-1] A_ul [10^8 s^-1] B_H^u [G] 260.609 b^2D_3/2 - x^2D^ o_3/2 36 126 - 74 498 2.00 ∼40 252.257 c^2G_7/2 - w^2G^ o_7/2 33 501 - 73 143 2.60 ∼57 258.336 a^4D_3/2 - z^4P^ o_3/2 8 680 - 47 389 0.88 ∼20 262.908 a^6D_1/2 - z^6D^ o_3/2 977 - 39 013 0.87 ∼16 275.655 a^4D_7/2 - z^4F^ o_9/2 7 955 - 44 232 2.20 ∼22 The background color in Fig. <ref> shows the logarithm of the normalized contribution function as in Fig. <ref>, what confirms that the linear polarization in this selection of lines form at different heights ranging from a few hundred kilometers above the temperature minimum (>0.5 Mm) and the upper chromosphere (⪆1.7 Mm). For the two metastable transitions at 260.608 and 252.257 nm (first and second panels in Fig. <ref>) most of the contribution is localized in a relatively extense region in the middle chromosphere of the model. These two transition have B_H=40 and 50 G, respectively and, consequently, a magnetic field of just 10 G already induces a significant depolarization of the line core linear polarization. A magnetic field of 50 G already reduces the fractional linear polarization by more than a factor ten. Contrarily to the rest of the selected transitions, the spectral line at 258.336 nm shows a strong negative fractional linear polarization signal (radial polarization; see third panel of Fig.<ref>) of about -0.8 %. This line forms in a extense region of the upper chromosphere of the model and it has B_H≈20 G. Consequently, a 10 G magnetic field already more than halved the amplitude of the linear polarization and it is dramatically depolarized by a 50 G magnetic field. Finally, regarding the spectral lines at 262.908 and 275.655 nm, their line core linear polarization forms in a narrow area in the top of the FAL-C model's chromosphere (between 1.9 and 2.2 Mm and between 2.0 and 2.2 Mm, respectively). Their critical Hanle fields are also relatively small (16 and 22 G, respectively), as evident from the strong depolarization induced by a magnetic field of 20 G (see fourth and fifth panels of Fig. <ref>). Nevertheless, these lines are sensitive to magnetic fields with strengths of a few gauss, of the order of the typical magnetic field strengths expected in the upper chromosphere of the quiet Sun. Moreover, note that the chosen magnetic field geometry, namely horizontal with respect to the local vertical, is such that it maximizes the depolarization of the fractional linear polarization Q/I profiles. §.§ Circular Polarization Signals. We now study the circular polarization profiles of the Fe2 UV lines and their suitability to infer the longitudinal component of the magnetic field via the WFA. For this purpose, we calculate the fractional circular polarization profiles V/I for a LOS at the disk center (μ=1), imposing in the FAL-C model a uniform vertical magnetic field of 50 G. We show in the right panel of Fig. <ref> those lines with a circular polarization amplitude above 0.5%, with the color of the marker indicating such amplitude and its size indicating the effective Landé factor (g_ eff) calculated assuming L-S coupling. Although the lobes of the circular polarization profiles typically sample the atmosphere at regions significantly below those sampled by the line cores, we will use the height at which the line-center optical depth is unity to indicate the approximate height at which these profiles are formed. A total of 103 spectral lines show circular polarization signals larger than 0.5 %. Most of these lines form between the upper photosphere and lower chromosphere. A considerable amount of Fe2 lines in the selection form in the middle and upper chromosphere, e.g., the transitions of the resonant multiplet a^6D - z^6D^∘. Given the computational demands for solving the full NLTE RT problem, it is ideal to have inference methods that allow for a quick determination of the magnetic field without the need of carrying out the NLTE modeling of the observations. The WFA, if suitable, allows for the very quick determination of the longitudinal component of the magnetic field <cit.>. For this approximation to be suitable, the line's Doppler width (Δλ_D) must be much larger than the magnetic Zeeman splitting between the atomic magnetic sublevels (Δλ_B), g_ effΔλ_B/Δλ_D << 1, where Δλ_B = 4.6686·10^-13Bλ_0^2 (with B in Gauss and λ_0 in Å) and Δλ_D = λ_0/c√(2kT/m + v_m^2) (with T and v_m the temperature and turbulent velocity in the formation region of the line, with c the speed of light, k the Boltzmann constant, and m the mass of the atom). In this way, we can roughly expect the WFA to be valid when the magnetic field strength is below the value which makes g_ effΔλ_B/Δλ_D = 0.1. Fig. <ref> shows the maximum magnetic field strength for which this condition is fulfilled. The included lines are those selected in the right panel of Fig. <ref>. For those spectral lines whose line core forms between the middle and upper chromosphere the WFA is valid, through the chosen criteria, for any magnetic field strength below 500 G (except for the one spectral line at 262.245 nm). We can thus be confident that the WFA necessary condition is met both in quiet sun regions and in the weaker manifestations of active regions, such as plages. For spectral lines originating between the upper photosphere and lower chromosphere we can distinguish two cases: spectral lines in which the WFA applicability condition is met for B_L<=200 G, and thus only applicable in quiet Sun regions, and spectral lines for which the applicability condition is B_L<=600 G, and thus it can also be applied in relatively weak active regions such as plages. ccccc Wavelength, level configuration and energy, Einstein coefficient for spontaneous emission, and effective Landé factor for the spectral lines studied in detail in Sec. <ref> λ [nm] Transition Energies [cm^-1] A_ul [10^7 s^-1] g_ eff 259.608 b^4F_9/2 - z^4H^ o_4 22 637 - 61 156 0.12 2.5 262.148 b^4F_4 - z^4G^ o_4 22 810 - 60 956 3.40 1.1 262.119 a^6D_3/2 - z^6D^ o_3/2 862 - 39 013 0.43 1.9 262.908 a^6D_1/2 - z^6D^ o_3/2 977 - 39 013 8.70 1.5 To study in detail the actual applicability of the WFA to infer the longitudinal magnetic field, we select four spectral lines, relatively close in wavelength, and infer the magnetic field from a theoretical circular polarization profile calculated in the FAL-C model imposing a certain magnetic field stratification. Table <ref> summarizes the atomic data of the four chosen transitions, also highlighted in green in the Grotrian diagram of Fig. <ref>. Among these spectral lines we find two resonant transitions (262.119 and 262.908 nm) and two transitions with metastable lower levels in the b^4F term (259.608 and 262.148 nm). The dotted curves in the first and second column in Fig. <ref> show the fractional circular polarization profile of the radiation emerging from the FAL-C model in which we have added a longitudinal magnetic field decreasing exponentially with height, from 200 G at the base of the photosphere to 40 G at the top of the chromosphere. The colored background in the same panels show the logarithm of the normalized contribution function for Stokes V, which gives an idea of the regions in the model atmosphere which contribute the most to the emergent circular polarization profile at each wavelength. The main contribution to the circular polarization of the spectral line at 259.608 nm comes from the photosphere, below the temperature minimum. The line at 262.147 nm forms at slightly larger heights, with significant contribution from the regions just above the temperature minimum. Finally, the contribution function of the resonance lines at 262.119 and 262.908 nm have their peak in the upper layers of the chromosphere, mainly around 1 Mm for the former, and even higher for the latter. The colored curves show the WFA fit to the theoretical circular polarization profiles. Because the contribution function of the 259.608 and 262.148 nm lines is more or less concentrated and coherent in sign at every wavelength, it is possible to find a good fit to the whole spectral range. The profiles are fitted with a magnetic field of 178 and 150 G, respectively. In the model atmosphere, these fields are found at heights ∼0.2 and ∼0.4 Mm, which correspond to approximately the middle of the region in the model with significant contribution function values (see colored background in the corresponding panels). For the resonant lines at 262.119 and 262.908 nm, the contribution function is significantly more extended and it changes sign with height at some wavelengths. For this reason, it is not possible to find a unique satisfactory fit to the whole spectral range (different regions of the profile are being affected by different magnetic fields, because they do not form in the same atmospheric region). We can fit the central part of the profiles with a magnetic field of 93 and 68 G for the two lines, respectively. These magnetic fields are found in the model atmosphere at heights ∼1 and ∼1.4 Mm, respectively, which also correspond to approximately those heights showing the largest value in the contribution function at the fitted wavelengths (see colored background in the corresponding panels). The WFA thus seem suitable for the quick inference of the longitudinal magnetic field at different layers of the solar atmosphere by combining the diagnostic of multiple Fe2 with different formation regions. § SUMMARY AND CONCLUSIONS We have theoretically investigated the formation of the intensity and polarization of the Fe2 spectral lines in the relatively unexplored 250 - 278 nm window of the near-UV solar spectrum. We identified those lines for which we predict significant linear or circular polarization signals and investigated their magnetic sensitivity, as well as their suitability to infer the magnetic field in the solar atmosphere. By means of NLTE RT calculations in a semi-empirical model of the solar atmosphere, we have obtained the Stokes profiles of the emergent spectral line radiation. To this end, we built a model atom with all possible transitions of interest in the above mentioned spectral region. Due to the lack of solar observations in this region of the solar spectrum, we use the observed intensity flux of α Cen A (a solar-like star) in order to adjust one of the most important and approximated parameters in our atomic model, namely the rate of inelastic collisions with electrons. We find that by reducing all these rates by a factor ten we obtain a theoretical flux very similar to the α Cen A observation at the core of the Fe2 resonant lines. It is important to emphasize that this region of the spectrum is severely affected by the so-called UV blanketing, which we cannot accurately include in our modeling, and that is evident when comparing the continuum level between the theoretical and observed flux profiles. While the impact of the UV blanketing, a priori, limits the diagnostic capabilities of the weaker lines and the wings of the strongest lines, the metallicity of α Cen A is considerably larger than that of the Sun. Therefore the impact of UV blanketing is expected to be less significant in the solar spectrum than in α Cen A. Future solar observations in this spectral window will make possible to study in detail how the UV blanketing affects the conclusions of this work and to further validate and improve the atomic data for the modeling of the Fe2 spectral lines. Most of the stronger (in intensity) Fe2 lines do not show remarkable linear polarization signals for a close to the limb LOS (μ=0.1). Nevertheless, there are numerous lines with predicted signals with amplitudes larger than 0.2 %, and therefore we expect that they can be observed with spectropolarimetric instrumentation similar to that of CLASP2. We show the region of formation, critical Hanle field, and information about blends for all these lines. Most of them form around the middle chromosphere, with some forming in the upper chromosphere. We study in detail the magnetic sensitivity of the linear polarization of five of these lines, showing that those forming in the middle (upper) chromosphere are typically sensitive to magnetic fields with strength up to ∼50 G (∼20 G). These lines thus encode information about the expected weak magnetic fields in the chromosphere of the quiet Sun. We find that, for a magnetic field strength of 50 G, more than a hundred Fe2 transitions are predicted to show circular polarization signals above 0.5% strong enough to be able of use the WFA to reliably infer the longitudinal component of the magnetic field. We show the region of formation, effective Landé factor, and information about blends for all these lines. Most of them form in the upper photosphere and lower chromosphere, with some of them forming in the upper chromosphere. We estimate the order of magnitude of the magnetic field for which the WFA may not hold. We find that, for the lines forming in the photosphere, these values are mostly between 200 and 400 G, while in the middle and upper chromosphere the typical values are between 700 and 900 G. By applying the WFA to the theoretical circular polarization profiles of four lines with different heights of formation we are able to recover the magnetic field values of the imposed exponential stratification. When checking the height in the atmospheric model at which the inferred magnetic field is located, we can see that they coincide with the information provided by the circular polarization contribution function regarding the region of formation of each of the selected spectral lines. The narrower 279.2 - 280.7 nm spectral window of the CLASP2 suborbital space experiment includes several spectral lines of already demonstrated utility for diagnosing the magnetic field across the whole solar chromosphere, namely the Mg2 h and k lines and the resonance lines of Mn1. In addition, as we shall show in a forthcoming publication, it includes several other weaker spectral lines with measurable circular polarization signals in active region plages, such as two hitherto unexplored Fe2 lines. The circular polarization of these Fe2 lines provides information about the magnetic field which is complementary to the Mg2 and Mn1 lines in terms of the region of the atmosphere that they sample. All these near-UV lines of the CLASP2 spectral region are very suitable to map the magnetic field from the lower to the upper chromosphere. Nevertheless, the results of the present investigation on the magnetic sensitivity of the Fe2 lines in the 250 - 278 nm spectral region, especially the prediction that they should show measurable circular polarization signals and sample different layers of the solar chromosphere when suitably combined, lead us to conclude that including these lines would further help determine the magnetic field throughout the whole solar chromosphere. We acknowledge the funding received from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (ERC Advanced Grant agreement No 742265). aasjournal

http://arxiv.org/abs/2303.06834v1

20230313033129

DarkVisionNet: Low-Light Imaging via RGB-NIR Fusion with Deep Inconsistency Prior

cs.CV

Label Information Bottleneck for Label Enhancement Qinghai Zheng^1,  Jihua Zhu^2[1]Corresponding author, E-mail: zhujh@xjtu.edu.cn,  Haoyu Tang^3 ^1College of Computer and Data Science, Fuzhou University, China ^2School of Software Engineering, Xi'an Jiaotong University, Xi'an, China ^3School of Software, Shandong University, Jinan, China Received: date / Accepted: date ======================================================================================================================================================================================================================================================================================================================== RGB-NIR fusion is a promising method for low-light imaging. However, high-intensity noise in low-light images amplifies the effect of structure inconsistency between RGB-NIR images, which fails existing algorithms. To handle this, we propose a new RGB-NIR fusion algorithm called Dark Vision Net (DVN) with two technical novelties: Deep Structure and Deep Inconsistency Prior (DIP). The Deep Structure extracts clear structure details in deep multiscale feature space rather than raw input space, which is more robust to noisy inputs. Based on the deep structures from both RGB and NIR domains, we introduce the DIP to leverage the structure inconsistency to guide the fusion of RGB-NIR. Benefiting from this, the proposed DVN obtains high-quality low-light images without the visual artifacts. We also propose a new dataset called Dark Vision Dataset (DVD), consisting of aligned RGB-NIR image pairs, as the first public RGB-NIR fusion benchmark. Quantitative and qualitative results on the proposed benchmark show that DVN significantly outperforms other comparison algorithms in PSNR and SSIM, especially in extremely low light conditions. § INTRODUCTION High-quality low-light imaging is a challenging but significant task. On the one hand, it is the cornerstone of many important applications such as 24-hour surveillance, smartphone photography, etc. On the other hand, though, massive noise of images in extremely dark environments hinders algorithms from the satisfactory restoration of low-light images. RGB-NIR fusion techniques provide a new perspective for the challenge: it enhances the low-light noisy color (RGB) image through rich, detailed information in the corresponding near-infrared (NIR) image (The high quality of NIR images in dark environments comes from invisible near-infrared flash), which greatly improves the signal-to-noise ratio (SNR) of the restored RGB image. Under the constraint of cost, size and other factors, RGB-NIR fusion becomes the most promising technique to restore the vanished textual and structural details from noisy RGB images taken in an extremely low-light environments, as shown in Figure <ref>(a). However, the existing RGB-NIR fusion algorithms suffers from the problem of structure inconsistency between RGB and NIR images, resulting in unnatural appearance and loss of key information, which limits the application of RGB-NIR fusion algorithm in low-light imaging. Figure <ref> illustrates two typical examples of structure inconsistency between RGB and NIR images: Figure <ref>(b) shows the absence of NIR shadows in the RGB image (grass shadows only appear on the book edge in the NIR image), and the nonexistence of RGB color structure in the NIR image (text ‘complete’ almost disappears on the book cover in the NIR image). Fusion algorithms need to tackle these structure inconsistency to avoid visual artifacts in output images. There are two categories of RGB-NIR fusion methods currently, i.e. traditional methods and neural-network-based methods, and modeling the structure of the paired RGB-NIR images plays an important role in both of them. Traditional methods, such as Scale Map <cit.>, tackle the structure inconsistency problem by manually designed functions. Some neural-network-based methods <cit.>, on the other hand, utilize deep learning techniques to automatically learn the structure inconsistency by a large amount of data. Both of them perform well under certain circumstances. However, when confronted with extreme low-light environments, existing methods fail to maintain satisfactory performance, since the structure inconsistency is dramatically exacerbated by massive noise in the RGB image. As shown in Figure <ref>(b), the dense noise in the RGB image makes it difficult for Scale Map to extract structural information, causing the failure of distinguishing which structures in the NIR image should be eliminated, and result in the unnatural ghost images on the book edge. Deformable Kernel Networks (DKN) <cit.> falsely weakens gradients of input RGB image that do not exist in the corresponding NIR image, which leads to the blurriness of letters on the book cover. Even though these structural inconsistency of corresponding RGB and NIR images can be captured by human eyes effortlessly, they still confuse most of the existing fusion algorithms. In this paper, we focus on improving the RGB-NIR fusion algorithm for extremely low SNR images by tackling the structure inconsistency problem. Based on the above analysis, we argue that the structure inconsistency under extremely low light can be handled well by introducing prior knowledge into deep features. To achieve this, we propose a deep RGB-NIR fusion algorithm called Dark Vision Net (DVN), which explicitly leverages the prior knowledge of structure inconsistency to guide the fusion of RGB-NIR deep features, as shown in Figure <ref>. With DVN, two technical novelties are introduced: (1) We find a new way, referred to as deep structures, to represent clear structure information encoded in deep features extracted by the proposed Deep Structure Extraction Module (DSEM). Even facing images with low SNR, the deep structures can still be effectively extracted and represent reliable structural information, which is critical to the introduction of prior knowledge. (2) We propose Deep Inconsistency Prior (DIP), which indicates the differences between RGB-NIR deep structures. Integrated into the fusion of RGB-NIR deep features, the DIP empowers the network to handle the structure inconsistency. Benefiting from this, the proposed DVN can obtain high-quality low-light images. In addition, to the best of our knowledge, there is no available benchmark dedicated for the RGB-NIR fusion task so far. The lack of benchmarks to evaluate and train fusion algorithms greatly limits the development of this field. To fill this gap, we propose a dataset named Dark Vision Dataset (DVD) as the first RGB-NIR fusion benchmark. Based on this dataset, we give qualitative and quantitative evaluations to prove the effectiveness of our method. In summary, the main contributions of this paper are as follows: * We propose a novel RGB-NIR fusion algorithm called Dark Vision Net (DVN) With Deep Inconsistency Prior (DIP). The DIP explicitly integrates the prior of structure inconsistency into the deep features, avoiding over-relying on NIR features in the feature fusion. Benefits from this, DVN can obtain high-quality low-light images without visual artifacts. * We propose a new dataset Dark Vision Dataset (DVD) as the first public dataset for training and evaluating RGB-NIR fusion algorithms. * The quantitative and qualitative results indicate that DVN is significantly better than other compared methods. § RELATED WORK §.§.§ Image Denoising. In recent years, denoising algorithms based on deep neural networks have continually emerged and overcome the drawbacks of analytical methods <cit.>. The image noise model is gradually improved simultaneously <cit.>. <cit.> applied an encoder-decoder network to suppress the noise and recover high-quality images. <cit.> presented a denoising network to process blind noise denoising. <cit.> attempted to remove the noise from real noisy images. There are also deep denoising algorithms trained without clean data supervision <cit.>. However, in extremely dark environments, fine texture details damaged by the high-intensity noise are very difficult to restore. In that case, denoising algorithms tend to generate over-smoothed outputs. By the way, low-light image enhancement algorithms <cit.> try to directly restore high-quality images in terms of brightness, color, etc. However, these algorithms cannot deal with such high-intensity noise as well. §.§.§ RGB-NIR Fusion. To obtain high-quality low-light images, researchers <cit.> try to fuse NIR images with RGB images. Recently, <cit.> pointed out the gradient inconsistency between RGB-NIR image pairs, and proposed Scale Map to try to solve it. Among the methods based on deep neural network, Joint Image Filtering with Deep Convolutional Networks (DJFR) <cit.> constructs a unified two-stream network model for image fusion, CU-Net <cit.> combines sparse encoding with Convolutional Neural Networks (CNNs), DKN <cit.> explicitly learns sparse and spatially-variant kernels for image filtering. <cit.> innovatively constructs a network that directly decouples RGB and NIR signals for 24-hour imaging. In general, current RGB-NIR fusion algorithm has two main problems: insufficient ability to deal with RGB-NIR texture inconsistency, leading to heavy artifacts on the final fusion images, inadequate noise suppression capability especially when dealing with high-intensity noise in extremely low-light environments. §.§.§ Datasets. There is only a small amount of data that can be used for RGB-NIR fusion studies because of the difficulty to obtain aligned RGB-NIR image pairs. Some studies <cit.> focus on obtaining hyperspectral datasets, and strictly aligned RGB-NIR image pairs can be obtained by integrating hyperspectral images on the corresponding band. <cit.> present a prototype camera to collect RGB-NIR image pairs. However, these datasets are too small to be used to comprehensively measure the performance of fusion algorithms. More importantly, due to the lack of data on actual scenarios, they cannot encourage follow-up researchers to focus on the valuable problems that RGB-NIR will encounter in applications. § APPROACH §.§ Prior Knowledge of Structure Inconsistency As previously described, the network needs to be aware of the inconsistent regions on the two inputs. We design an intuitive function to measure the inconsistency from image features. Firstly binary edge maps are extracted from each feature channel. Then the inconsistency is defined as ℱ (edge^C_:, edge^N) = λ (1 - edge^C_:)(1 - edge^N) + edge^C_:· edge^N where C_:∈ℝ^H × W and N ∈ℝ^H × W denote R/G/B channel of the clean RGB image and NIR image, edge^C_: and edge^N respectively represent the binarized edge maps of C_: and N, which is obtained by binarizing its mean value as a threshold after Sobel filtering. As shown in Figure <ref>, ℱ(·,·) equals to 0 in the regions where edge^C_: and edge^N shows severe inconsistency. On the contrary, ℱ(·,·) equals to 1 in the regions where the structures of RGB and NIR are consistent. And in other regions, ℱ(·,·) is set to a hyperparameter λ (0 < λ < 1), indicating that there is no significant inconsistency. Utilising the output inconsistency map of ℱ, the inconsistent NIR structures can be easily suppressed by a direct multiplication. §.§ Extraction of Deep Structures Even though function ℱ subtly describes the inconsistency between RGB and NIR images, it cannot be applied directly in extremely low light cases. As shown in Figure <ref>, the calculated inconsistency map contains nothing but non-informative noise when facing extremely noisy RGB image. To avoid the influence of noise in the structure inconsistency extraction, we propose the Deep Structure Extraction Module (DSEM) and Deep Inconsistency Prior (DIP), where we compute the structure inconsistency in feature space. Considering the processing flow of RGB and NIR are basically the same, we give a unified description here to keep the symbols concise. The detailed architecture of DSEM is shown in Figure <ref>(a). DSEM takes multi-scale features feat_i (i represents scale) from restoration network R and outputs multi-scale deep structure maps struct_i. In order for DSEM to predict high-quality deep structure maps, we introduce a clear supervision signal struct_i^gt (addressed later) for DSEM and the training loss is calculated as: ℒ_stru = ∑_i = 1^3∑_c = 1^Ch_i Dist(struct_i, c, struct_i, c^gt), where Ch_i is the channel number of the deep structures in the ith scale, Dist is Dice Loss <cit.>, struct_i, c is the cth channel of the predicted deep structures in the ith scale and struct_i, c^gt is the corresponding ground-truth. The Dice loss is given by Dist(P, G) = ( ∑_j^N p_j^2+∑_j^N g_j^2 ) / (2 ∑_j^N p_j g_j), where p_j, g_j is the value of the jth pixel on the predicted structure map P and ground-truth G. §.§.§ Supervision of DSEM Considering that it is almost impossible for DSEM to naturally output feature maps that only contain structural information, we have to introduce a clear supervision for the output of DSEM to predict high-quality deep structure maps. The supervision signal is set up following the idea of Deep Image Prior <cit.> and struct_i, c^gt is acquired from a pretrained AutoEncoder <cit.> network AE [See the training details in supplementary material.]. The base architecture of AE is exactly the same as R with skip connections removed, as Figure <ref>(b) shows. Multi-scale decoder features dec_i, c are extracted from the pretrained AutoEncoder network AE and the supervision signal is calculated by: struct_i, c, j^gt = 0 if (∇ dec_i, c, j - m_∇ dec_i, c) <= 0, 1 if (∇ dec_i, c, j - m_∇ dec_i, c) > 0. where struct_i, c, j^gt is the jth pixel of struct_i, c^gt, ∇ represents the Sobel operator, ∇ dec_i, c, j is the jth pixel in ∇ dec_i, c and m_∇ dec_i, c is the global average pooling result of ∇ dec_i, c. The supervision signal obtained by this design effectively trains the DSEM and clear deep structure maps are predicted as shown in Figure <ref>. §.§ Calculation of DIP and Image Fusion The extracted deep structures contain rich structure information and are robust to noise. With struct_i of the noisy RGB and NIR images, we can introduce inconsistency function ℱ to obtain high-quality knowledge of structure inconsistency: M^DIP_i, c = ℱ(struct_i, c^C, struct_i, c^N) where C ∈ℝ^H × W × 3 and N ∈ℝ^H × W denote the noisy RGB image and NIR image, struct_i, c is the cth channel of the features from the ith scale and M^DIP_i, c is the corresponding inconsistency measurement. Since M^DIP_i, c represents the structure inconsistency instead of intensity inconsistency, we apply M^DIP_i, c directly to struct_i, c^N instead of feat^N_i, c in the form of: ŝt̂r̂ûĉt̂_̂î,̂ ̂ĉ^̂N̂ = M^DIP_i, c· struct_i, c^N. Under the guidance of the DIP, ŝt̂r̂ûĉt̂_̂î,̂ ̂ĉ^̂N̂ discards the structures that are inconsistent with RGB, thus empowering the deep features with prior knowledge to tackle structure inconsistency. As we will show in the experiments later, inconsistent structures in the NIR structure maps can be significantly suppressed after multiplying with M^DIP_i, c. To further fuse the rich details contained in ŝt̂r̂ûĉt̂_̂î,̂ ̂ĉ^̂N̂ into RGB features, we designed a multi-scale fusion module as shown in Figure <ref>(c). As pointed out in <cit.>, denoising first and fusion later can improve the fusion quality. So we follow <cit.> to reuse the denoised output of the restoration network R set up for noisy RGB as the input of the multi-scale fusion module. §.§ Loss Function The total loss function we used is formulated as: ℒ = ℒ_rec^C + ℒ_rec^Ĉ + ℒ_rec^N + λ_1 ·ℒ_stru^C + λ_2 ·ℒ_stru^N where ℒ_stru^C and ℒ_stru^N are the loss function for RGB/NIR deep structures prediction, which is described above. λ_1 and λ_2 are the corresponding coefficients and set to 1/1000 and 1/3000. ℒ_rec^C, ℒ_rec^Ĉ and ℒ_rec^N represent the reconstruction loss of fused-RGB/coarse-RGB/NIR image respectively. All of them are Charbonnier loss <cit.> in the form of: ℒ_rec = √(𝐗-𝐗_gt^2+ε^2) where 𝐗 and 𝐗_gt represent the network output and the corresponding supervision. The constant ε is set to 10^-3 empirically. § EXPERIMENT §.§ Datasets §.§.§ Data Collection. In order to obtain the aligned RGB-NIR image pairs in the easiest and direct way, we collect all RGB-NIR image pairs by switching an optical filter placed directly in front of the camera without an IR-Cut, and we divide them into two types of image pairs for different usages. We collect reference image pairs in normal-light environments. In order to obtain high-quality references, multiple still captures are averaged to remove noise and a matching algorithm <cit.> with manual double-check is applied to ensure the alignment of image pairs. In the following experiments, we add synthetic noise to these references to quantitatively evaluate the performance of fusion algorithms. To facilitate training, the collected images are cropped into 256*256 image patches. We collect real noisy image pairs of 1920*1080 pixels in low-light environments. The post-processing steps are the same as those used in collecting references image pairs, except that multi-frame average is not performed to noisy RGB images. In the following experiments, we use these noisy image pairs to qualitatively evaluate the performances of fusion algorithms in handling real low-light images. §.§.§ Dataset for Experiment. To make the synthetic data closer to the real images, we follow <cit.> to add synthetic noise to reference image pairs for training and testing. Considering that all the comparison methods use sRGB (standard Red Green Blue) images as input, we convert the collected raw data into sRGB through a simple isp-pipeline (Gray World for white balance, Gamma correction, Demosaicing) <cit.>, to make a fair comparison. In the following experiments, we use 5k reference image pairs (256*256) as the training set. Another 1k reference image pairs (256*256) along with 10 additional real noisy image pairs (1920*1080) are used for testing. §.§ Implementation Details All experiments are conducted on a device equipped with two 2080-Ti GPUs. We train the proposed DVN from scratch in an end-to-end fashion. Batchsize is set to 16. Training images are randomly cropped in the size of 128*128, and the value range is [0, 1]. We augment the training data following MPRNet <cit.>, including random flipping and rotating. Adam optimizer with momentum terms (0.9, 0.999) is applied for optimization. The whole network is trained for 80 epochs, and the learning rate is gradually reduced from 2e-4 to 1e-6. λ in function ℱ is set to 0.5 for all configurations. The AutoEncoder network used to provide supervision signals for DSEM is pretrained in the same way, except that it only trained for 5 epoches and the input is clean RGB and NIR images separately. See supplementary material for more training details. The synthesis of low-light data for training includes two steps. The first step is to reduce the average value of raw images taken under normal light to 5 (10-bit raw data). The second step is to add noise to the pseudo-dark raw images, including Gaussian noise with variance equals to σ, and Poisson noise with a level proportional to σ. §.§ Performance Comparison §.§.§ Results on DVD Benchmark. We evaluate and compare DVN with representative methods in related fields, including single-frame noise reduction algorithms NBNet <cit.> and MPRNet <cit.>, joint image filtering algorithms GIF <cit.>, DJFR <cit.>, DKN <cit.> and CUNet <cit.>, as well as Scale Map <cit.> which specially designed for RGB-NIR fusion. All methods are trained or finetuned on DVD from scratch. We use PSNR and SSIM <cit.> for quantitative measurement. Qualitative comparison is shown in Figure <ref>, and quantitative comparison under different noise intensity settings (σ = 2, 4, 6, 8, the larger the σ, the heavier the noise) on DVD benchmark is shown in Table <ref>. The qualitative comparison in Figure <ref> clearly illustrates the superiority of the proposed DVN on noise removal, detail restoration and visual artifacts suppression. In contrast, image denoising algorithms (i.e. NBNet and MPRNet) cannot restore texture details when the noise intensity becomes high, and the output turns into pieces of smear even though the noise is effectively suppressed. GIF and DJFR output images with heavy noise as the 3rd and 4th column in Figure <ref> shows, which greatly affects the fusion quality. The fusion effect of DKN and CUNet (5th and 6th column in Figure <ref>) under mild noise (e.g. σ=2) is acceptable. But under heavy noise, obvious color deviation appears in the DKN output, and neither of them can deal with structure inconsistency (see the 4th row in Figure <ref>), resulting in severe artifacts in the fusion images. Scale Map outputs images with rich details. However, it cannot reduce the noise in the areas where texture is lacking in the NIR image. In addition, it is hard to achieve a balance between noise suppression and texture migration when applying Scale Map. §.§.§ Generalization on Real Noisy RGB-NIR. To evaluate the performance of algorithms when facing real low-light images, we conduct a qualitative experiment on several pairs of RGB-NIR images captured in real low-light environments. As shown in Figure <ref>, outputs of DVN have obviously low noise, rich details, and are visually more natural when handling RGB-NIR pairs with real noise, even if the network is trained on a synthetic noisy dataset. §.§.§ Comparison on Public Dataset. So far, there is no high-quality public RGB-NIR dataset like DVD yet. For example, RGB-NIR pairs in IVRG <cit.> are not well aligned. Even so, we retrained DVN and other methods on IVRG and give quantitative comparison in Table <ref>. It is clear that DVN still performs well. §.§.§ Comparison with Low-Light Enhancement Methods. We also compare our method with the low-light enhancement methods. We retrained SID <cit.> and SGN <cit.>, the comparison can be seen in Table <ref>. It is clear that our proposed DVN still shows great superiority. §.§ Effectiveness of DIP In this section, we verify that the proposed DIP is effective in handling the mentioned structure inconsistency. For comparison, we retrain a baseline, which is the same as the proposed DVN only without the DIP module. As Figure <ref>(a) shows, the NIR shadow of the grass still remains in the fusion result without DIP, but not in the fusion result with DIP. This directly proves that DIP can handle the structure inconsistency. Figure <ref>(b) shows that DIP can also deal with serious structure inconsistency caused by the misalignment between RGB-NIR images to a certain extent (this example pair cannot be aligned even after image registration). This has practical value because the problem of misalignment frequently occurs in applications. Taking into account the nature of DIP, the remaining artifacts are in line with expectations, since they are concentrated near the pixels with gradients in the RGB image. In addition, Figure <ref> also visualizes the deep structure of RGB, NIR, consistent NIR (DIP-weighted) as well as DIP Maps. It is obvious that even facing noisy input, the RGB deep structure still contains clear structures. The visual comparison between the NIR deep structure and the consistent NIR deep structure proves that the introduction of DIP can handle structure inconsistency in deep feature space. §.§ Ablation Study We evaluate the effectiveness of each component in the proposed algorithm on the DVD benchmark quantitatively in this section. PSNR and SSIM are reported in Table <ref>. The baseline network directly fuse NIR features with RGB features (row 1 in Table <ref>). Intermediate supervision ℒ_rec^Ĉ and ℒ_rec^N effectively improve the performance as Table <ref> (row 1 and 2) shows. This indicates the necessity of enhancing the noise suppression capability of the network for clean structure extraction. Applying DSEM to learn deep structures without DIP can improve performance as well as Table <ref> (row 1 and 3) shows. However, since the inconsistent structures are not removed, the benefits are not obvious, even if we use intermediate supervision and DSEM simultaneously as row 4 shows. As Table <ref> (row 5) shows, after introducing DIP to deal with the structure inconsistency, the network performance can be further improved by a large margin. This demonstrates the effectiveness of our proposed algorithm and the necessity to focus on the structure inconsistency problem on RGB-NIR fusion problem. § CONCLUSION In this paper, we propose a novel RGB-NIR fusion algorithm called Dark Vision Net (DVN). DVN introduces Deep inconsistency prior (DIP) to integrate the structure inconsistency into the deep convolution features, so that DVN can obtain a high-quality fusion result without visual artifacts. In addition, we also proposed the first available benchmark, which is called Dark Vision Dataset (DVD), for RGB-NIR fusion algorithms training and evaluation. Quantitative and qualitative results prove that the DVN is significantly better than other algorithms. § ACKNOWLEDGEMENTS This paper is supported by the National Key R&D Plan of the Ministry of Science and Technology (Project No. 2020AAA0104400) and Beijing Academy of Artificial Intelligence (BAAI).

http://arxiv.org/abs/2303.07165v1

20230313150850

Almost sharp lower bound for the nodal volume of harmonic functions

math.AP

Effective lower bound]Almost sharp lower bound for the nodal volume of harmonic functions. [Andrea Sartori]Departement of Mathematics, Tel Aviv University, Tel Aviv, Israel, IL [Andrea Sartori]sartori.andrea.math@gmail.com [Lakshmi Priya]Departement of Mathematics, Tel Aviv University, Tel Aviv, Israel, IL [Lakshmi Priya]lpriyame@gmail.com [Alexander Logunov]Université de Genève, Section de mathématiques, rue du Conseil-Général 7-9, 1205 Genève, Switzerland and Department of Mathematics, Princeton University, Princeton, NJ, USA [Alexander Logunov]log239@yandex.ru This paper focuses on a relation between the growth of harmonic functions and the Hausdorff measure of their zero sets. Let u be a real-valued harmonic function in ℝ^n with u(0)=0 and n≥ 3. We prove ℋ^n-1({u=0}∩ B(0,2)) ≳_ N^1-, where the doubling index N is a notion of growth defined by sup_B(0, 1)|u| = 2^N sup_B(0,1/2)|u|. This gives an almost sharp lower bound for the Hausdorff measure of the zero set of u, which is conjectured to be linear in N. The new ingredients of the article are the notion of stable growth, and a multi-scale induction technique for a lower bound for the distribution of the doubling index of harmonic functions. It gives a significant improvement over the previous best-known bound ℋ^n-1({u=0}∩ 2B)≥exp (c log N/loglog N ), which implied Nadirashvili's conjecture. [ Alexander Logunov, Lakshmi Priya M.E., Andrea Sartori March 30, 2023 ========================================================== § INTRODUCTION Given a non-constant harmonic function u in ℝ^n, its zero set {u=0} is non-empty and has Hausdorff dimension (n-1). Locally the geometry and complexity of the zero set of u is controlled in terms of its growth. In particular, one can bound the (n-1) dimensional Hausdorff measure of the set {u=0} in every Euclidean ball B in terms of growth of u around B. One of the multiple ways to quantify the growth of a function is given by the notion of the doubling index. Given a positive number k and an Euclidean ball B=B(x,r) with center at x and radius r>0, we denote by kB the scaled ball B(x,kr). For any function h in a ball 2B⊂ℝ^n, the doubling index of h in B is defined by 𝒩_h(B):= log_2 sup_2B|h|/sup_B|h|. We often write 𝒩(x,r) instead of 𝒩_h(B(x,r)) and often omit the dependence on h in the notation and simply write 𝒩(B). Let B⊂ℝ^n be a unit ball. There exists a constant C=C(n)>1 such that ℋ^n-1({u=0}∩ B)≤ C 𝒩(B), for all harmonic functions u: 2B →ℝ. The main result of this paper states that there is also a lower bound in terms of growth. Let B⊂ℝ^n be a unit ball and n≥ 3. For every ε>0, there exists a constant c=c(n,ε)>0 such that for every harmonic function u: 4B →ℝ with u(0)=0, we have ℋ^n-1({u=0}∩ 2B)≥ c 𝒩(1/2B)^1-ε. Let us make a few remarks. Given a unit ball B⊂ℝ^n, Harnack's inequality for harmonic functions implies that if 𝒩_u(1/2B) is sufficiently large, then 2B must contain a zero point of u. If the harmonic function u is zero at the center of the ball B, then 𝒩(B)≥ c_n for some numerical constant c_n > 0 (see Claim <ref> in Appendix <ref> for the proof). Hence, in Theorem <ref> there is a uniform lower bound, which was conjectured by Nadirashvili. Nadirashvili's conjecture. If the harmonic function u is zero at the center of the unit ball B⊂ℝ^3, then the area of the zero set of u in B is bounded from below by a positive numerical constant. Nadirashvili's conjecture was recently proved in <cit.>. The fact sounds elementary, but there is no simple proof known, and there are applications including the lower bound in Yau's conjecture for nodal sets of Laplace eigenfunctions. We refer to <cit.> for the statement of Yau's conjecture and related results. This article does not pursue the purpose of generality. We restrict our attention only to the Euclidean case. There isn't a methodological obstacle preventing to extend the main result to the case of smooth Riemannian manifolds, though we don't do this job. We hope that this decision makes the article more accessible for the reader. One can compare this paper to the approach for Nadirashvili's conjecture in <cit.>, which gave a slowly growing bound ℋ^n-1({u=0}∩ 2B)≥exp (c log N/loglog N ), where N= 𝒩(1/2B) is assumed to be bigger than 10 just to make the right hand side well-defined and u vanishes at the center of B. Compared to <cit.>, this article introduces the new notion of stable growth and a new multi-scale induction technique, which gives significantly improved lower bounds and allows to come closer to the following folklore conjecture. Given a unit ball B⊂ℝ^n, n≥ 3, and a harmonic function u in 4B, which is zero at the center of B, the following lower bound holds: ℋ^n-1({u=0}∩ 2B) ≥ c 𝒩(1/2B), for some numerical constant c=c(n)>0. Before discussing the conjecture let us mention a few properties of the doubling index. The doubling index is scale invariant. A non-trival fact states that the doubling index for a harmonic function u is almost monotone function of a set in the sense that 𝒩_u(b) ≤ C 𝒩_u(B) if a ball B contains the ball 2b. Here C>1 is a dimensional constant, but one can find a sharper statement in <ref> about the doubling index. The monotonicity property is a powerful tool often used in geometrical analysis. However, besides almost monotonicity it is unclear how the doubling index behaves as a function of a set. Rescaling Conjecture <ref> and Theorem <ref>, one can see that scaled doubling index 𝒮𝒩(B):=𝒩(B) (B)^n-1 is comparable to the Hausdorff measure of the zero set of u in the following sense. Rescaled Theorem <ref>. ℋ^n-1({u=0}∩ B)≤ C ·𝒮𝒩(B). Rescaled Conjecture <ref>. 𝒮𝒩(1/2B)≤ C ·ℋ^n-1({u=0}∩ 2B) + C ·(B)^n-1. The extra additive term C ·(B)^n-1 is required (u can be positive and have no zeroes in 4B), but as a rule it plays no role. A challenge in nodal geometry is to understand the distribution of the doubling index as a function of a set. If Conjecture <ref> is true and the scaled doubling index for harmonic functions is comparable to the Hausdorff measure of the zero set, then one can conclude that the scaled doubling index is almost additive as a function of a set. The main Theorem <ref> is likely to have applications in nodal geometry as it gives a weak version of almost additivity for scaled doubling index with subpolynomial error term. In this paper we introduce a new notion of stable growth. At some moment in this paper we will choose a dimensional constant C=C(n)>1 (its choice will be presented in Remark <ref>). A harmonic function u:ℝ^n is said to have stable growth in a ball B⊆^n if C^-1(12B) ≤(2B) ≤ C (12B). We note that the first inequality holds due to almost monotonicity of the doubling index (see Lemma <ref> below) and can be omitted in the definition of stable growth. The assumption of stable growth is usually earned for free or for a small price of a subpolynomial error term, but it simplifies many of the proofs. Conjecture (<ref>) together with Theorem <ref> imply that if a harmonic function u has stable growth in a unit ball B and A is a constant such that 𝒩(B) ≫ A>1, then on the scale 1/A the following property of the doubling index holds: ∫_B 𝒩(x, A^-1)dx ≍𝒩(B)/A. In other words, it is anticipated that the average doubling index drops proportionally to the scale factor. § ACKNOWLEDGMENT. The authors are indebted to Mikhail Sodin for his enormous patience and numerous fruitful discussions that motivated this work. The authors are grateful to Tel Aviv University, Princeton University and University of Geneva, where this work was conducted. A.L. was supported in part by the Packard Fellowship, Swiss NSF grant 200020-200400 and by NCCR SwissMAP (Swiss NSF grant 205607). L.P.M.E. was supported by ERC Advanced Grant 692616, ERC consolidator grant 101001124 (UniversalMap) and ISF grant 1294/19. A.S. was supported by the ISF Grant 1903/18 and the BSF Start up Grant no. 20183. §.§ Notation As mentioned above, given a ball B=B(x,r)⊂ℝ^n with center at x∈ℝ^n and radius r>0, we denote by ℓ B the ball ℓ B= B(x,ℓ· r) for a given scaling parameter ℓ>0. Sometimes we will not need to explicate the center of the ball and we will simply write B(x,r)= B(r). We also write B to denote a closed ball. Similarly, given a cube Q⊂ℝ^n we write ℓ Q for the cube which is the homothetic copy of Q with same center and with homothety coefficient ℓ>0. We define the maximal doubling index of a cube Q by 𝒩^*(Q) := sup_x ∈ Q 0<ρ≤(Q) 𝒩(B(x,ρ)). We denote by c,c_1,c_2,… (small) positive constants and by C,C_1,C_2,… (large) positive constants, which may change from line to line. Given two quantities A,B, we write A≲ B and A≳ B to designate the existence of two constants c,C>0 such that A≤ CB and A≥ cB, respectively. Here the constants c, C are allowed to depend on the dimension n. If these constants depend on some auxiliary parameter γ, we write A≲_γ B and A≳_γ B. If A≲ B and B≲ A, we write A≍ B and A≍_γ B if one (or both) constants implied in the notation depend on the auxiliary parameter γ. Finally, we write A≪ B to mean that, for a sufficiently small constant c>0, we have A≤ cB and A≫ B if, for a sufficiently large constant C≥1, we have A≥ CB. Again, we write A≪_γ B and A≫_γ B if the constants implied in the notation depend on the parameter γ. Sometimes we will just write ≍, ≲ and ≪ instead of ≍_n, ≲_n and ≪_n. Given some x∈ℝ, we write ⌊ x⌋ for the largest integer smaller than x and ⌈ x ⌉ for the smallest integer larger than x. Given a finite set F, we denote by # F the number of elements in F. For a set S ⊂^n and ℓ >0, define the ℓ-neighbourhood of S by S_+ℓ := {x ∈^n : x= s+b, where s∈ S and b∈ B(0,ℓ) }. The zero sets of harmonic functions are often called nodal sets. We will sometimes refer to the Hausdorff measure of the nodal set as nodal volume. When we say that a harmonic function u has bounded doubling index in the ball B, we mean that there exists a numerical constant C>0 such that 𝒩_u(B)≤ C. § FIRST THOUGHTS AND OBSERVATIONS This section discusses elementary ideas and some examples for building intuition about the zero sets and the doubling index of harmonic functions. §.§ Bounded doubling index The following simple observation gives a lower bound for the nodal volume of a harmonic function in terms of its doubling index. We note that this bound is useful only when the doubling index is not large. Let B=B(0,1)⊂ℝ^n be the unit ball and let u be a harmonic function in 2B. Suppose that u(0)=0, then there exists a constant c=c(n)>0 such that ℋ^n-1({u=0}∩ 2B )≥ c 𝒩(B)^1-n. Before starting the proof, we formulate a consequence of Harnack's inequality. Let B=B(x,r) ⊂^n be any ball. Then for any non-constant harmonic function u defined in 2B satisfying u(x) ≥ 0, we have sup_43B u ≳_nsup_B |u|. Let M and m denote the supremum and the infimum of u in B. If m≥ 0, then the conclusion (<ref>) holds trivially. Let M' denote the supremum of u in 43B. It suffices to consider the case m <0, and show that -m ≲_n M'. Let x_m, x_M ∈∂ B be such that u(x_m) = m and u(x_M) = M. Consider the positive harmonic function h := M' -u in 43B. By Harnack's inequality, there is a constant C=C(n) >1 such that (M' - m) = h(x_m) ≤ C h(x_M) = C(M' - M) ≤ CM'. Hence, -m ≲_n M', and the claim follows. We will estimate the Hausdorff measure of the zero set of u in the spherical layer {4/3≤ |x|≤5/3}. Let N:= 20⌈𝒩(B)⌉. Consider the behavior of u in the concentric spheres S_i= ∂ ( r_i B )= {x∈ 2B : |x|=r_i}, where r_i= 4/3 + i/3N for i=0,…,N and define m_i^+ := max_S_iu,  and  m_i^- := min_S_iu. We may assume that u is not identically zero. Since u(0)=0, the maximum principle implies m_i^+>0, m_i^-<0, m_i^+ < m_i+1^+, and |m_i^-| < |m_i+1^-|. We aim to show that there are two balls of radius c_0 N^-1 in B, for some c_0=c_0(n)>0, such that u is positive in one ball and negative in the other ball. The following geometrical fact shows that existence of two such balls guarantees a lower bound for the nodal volume. Let f:D be a continuous function in a convex set D⊂ℝ^n. Assume that there are two balls B^+, B^-⊂ D both of radius r>0 such that f>0 in B^+ and f<0 in B^-. Then ℋ^n-1 ({f=0}) ≳_n r^n-1. Consider line segments starting in B^+ and ending in B^-, which are parallel to the line connecting the centers of the balls B^+ and B^-. Since f is positive at one end and negative at the other end, every such line segment contains a zero of f. Consider the orthogonal projection of the zero set {f=0} onto a hyperplane orthogonal to these line segments. The zeros of f in the line segments project onto a (n-1) dimensional ball of radius r in this hyperplane. Since the orthogonal projection does not increase the distances, the Hausdorff measure of {f=0} is at least the Hausdorff measure of its projection, which contains a (n-1) dimensional ball of radius r. This completes the proof of Claim <ref>. We are ready to proceed with the proof of Lemma <ref>. We will show that there are at least N/10 disjoint spherical layers {x∈ 2B: r_i-1≤ |x|≤ r_i} such that in each of them, we can find two balls of radius c_0 N^-1, for some c_0=c_0(n)>0, so that u is positive in one of the balls and negative in the other. The increment of u from 43B to 53B can be written as a telescopic product m_1^+/m_0^+·m_2^+/m_1^+⋯m_N^+/m_N-1^+ = sup_5/3Bu/sup_4/3 Bu. By Claim <ref> and recalling that N= 20⌈𝒩(B)⌉, we have sup_5/3Bu/sup_4/3 Bu≲_n sup_2B|u|/sup_B|u| = 2^𝒩(B)≤ e^N. Similarly, m_1^-/m_0^-·m_2^-/m_1^-⋯m_N^-/m_N-1^-≲_n e^ N. Hence we conclude that there exists C_1=C_1(n) >0 such that |m_i+1^-|≤ C_1 |m_i^-| holds for at least 3/4N values of i (and similarly for m_i^+). Thus, we have m_i+1^+≤ C_1 m_i^+ and |m_i+1^-|≤ C_1 |m_i^-|, for at least N/2 values of i. From now on, we will consider only i satisfying (<ref>). Let x_i^+ and x_i^-∈ S_i be points where u attains its maximum and minimum respectively, i.e., u(x_i^+) = m_i^+ and u(x_i^-) = m_i^-. Define B_i^+ := B(x_i^+, 110N) and B_i^- := B(x_i^-, 110N). We will now show that there is a small constant c_0>0 such that u >0 in c_0 B_i^+, and u <0 in c_0 B_i^-, by estimating the gradient of u as follows: sup_B_i^+ |∇ u| ≲_n N sup_2B_i^+ |u| ≲_n Nsup_83B_i^+u ≤Nsup_S_i+1 u = N m_i+1^+≤_(<ref>) C_2 N m_i^+, where the first inequality follows from the standard gradient estimates (Claim <ref>), the second inequality follows from Claim <ref>, and the third inequality follows from the inclusion 83B_i^+⊂ r_i+1B. All in all, we have u(x_i^+) = m_i^+ and sup_B_i^+|∇ u| ≲_n Nm_i^+, where B_i^+ := B(x_i^+, 110N). We may conclude that there is a small constant c_0 ∈ (0,1) such that u >0 in c_0 B_i^+. A similar argument implies that (for a small c_0∈ (0,1)) we have u <0 in c_0B_i^-. Finally, the Euclidean Claim <ref> implies the lower bound for the nodal volume. One can modify the argument in Lemma 2.1 to prove a slightly better estimate for the nodal volume: ℋ^n-1({u=0}∩ 2B) ≳(B)^2-n, by showing that ℋ^n-1( {u=0}∩{x∈ 2B: r_i-1≤ |x|≤ r_i+1})≳1/N^n-1 for at least N/2 indices i. We recall from <ref> the known upper bound and the conjectured lower bound for the nodal volume of a harmonic u vanishing at the center of a ball B ⊂^n (12B)r^n-1Conj.<ref>≲_nℋ^n-1({u=0}∩ B) Thm.<ref>≲_n(B) r^n-1. In the case when u has bounded doubling index in B 𝒩(1/2B)≍𝒩(B)≍ 1, Theorem <ref> and Lemma <ref> imply that the nodal volume of u is comparable to the surface area of B. Let B ⊂^n be a ball and let C>1 be some constant. Then for all harmonic functions u in 2B, vanishing at the center of B, such that _u(B) ≤ C, we have ℋ^n-1({u=0}∩ B) ≍_C,n r^n-1. The nodal volume is also comparable to the scaled doubling index: ℋ^n-1({u=0}∩ B) ≍_C,n𝒮𝒩(B). §.§ Chopping Given a harmonic function u vanishing at the center of a ball B⊂^n, how can we estimate the volume of the zero set when the doubling index is large? When we start looking at the zero set at small scales, the doubling index in small balls around the zero drops down asymptotically in the following way: lim_ 0(B(x,)) = order of vanishing of u at x. The latter limit is 1 for all x in the zero set except the singular set, that is the set where u and |∇ u| vanish simultaneously. The singular set does not contribute to the nodal volume as it has Hausdorff dimension (n-2), see <cit.>, and also <cit.> for a quantitative estimate. The non-singular zero set can be covered by countably many balls with bounded doubling index. In order to prove a lower bound for the nodal volume, a naive idea is to find small disjoint balls B_k = B(x_k,r_k) in B with doubling index (B_k) ≤ C such that the harmonic function has a zero at the center of each B_k. Then in each of the B_k we have a lower bound for the nodal volume by Claim <ref>, and therefore ℋ^n-1({u=0}∩ B) ≥∑_B_kℋ^n-1({u=0}∩ B_k) ≍∑_B_k r_k^n-1. To find balls with bounded doubling index, one can attempt to use the following idea of chopping. Take a cube Q (which is easier to chop than balls) and repeatedly divide it into smaller subcubes {Q_k}. At any point in this process if we have a subcube with bounded doubling index, we stop chopping, and if a subcube has large doubling index we continue chopping. Finally, in subcubes with small doubling index and a zero of u, one can use a version of Claim <ref> for cubes. There are two problems to implement this idea. The first one is to show that for most of the subcubes the doubling index drops down. The second problem is to get some lower bound for the sum of scaled doubling indices ∑(Q_k) (diam Q_k)^n-1, where the sum is over subcubes Q_k containing a zero of u. We choose to work with (<ref>) as it imitates the nodal volume. In order to obtain quantitative bounds for the nodal volume, one has to understand the distribution of the doubling index on many scales. More precisely, given a partition of cube Q ⊂^n into equal subcubes and a harmonic function in 2Q, what can we say about the doubling indices in these subcubes compared to the doubling indices of Q and 1/4Q? In the next section we discuss two simple examples of the distribution of doubling index. Old and new results about the distribution of doubling index of harmonic functions will be presented in <ref> §.§ Examples We consider the harmonic function u(x,y)= exp (Nx)sin (Ny), for some large positive integer N. Partition Q=[0,1]^2 into A^2 many equal subcubes {Q_i}_i of side length A^-1. If 1 ≪ A ≪ N, then 𝒩^*(Q)≍ N and 𝒩^*(Q_i)≍N/A. In this case, the doubling index is a linearly decreasing function of the size of the subcubes. The zero set of u consists of parallel lines, see Figure <ref> below. Consider u(x,y)= ((x+iy)^N) for some large integer N>1. The zero set of the function u is the union of 2N rays emanating from the origin, see Figure <ref> below. We consider Q=[-2,2]^2 and partition it into A^2 equal subcubes {Q_i}_i. We have 𝒩^*(Q)≍ N. Unlike the previous example, the doubling index of the subcubes depend on the proximity of the subcube to the origin, which is a singular zero of u. Consider z∈ Q with |z| = r>0, let us now get an estimate for (z,1/A). In polar coordinates, u can be written as u(r,θ) = r^N cos Nθ. Hence, for 1 ≪ A≪ N, we have logsup_B(z,2/A)|u|/sup_B(z,1/A)|u|≍log(r+2A)^N/(r+1A)^N = log(1+2rA)^N/(1+1rA)^N. If (1/rA) ≪ 1, or equivalently r≫ 1/A, then we have (z,1/A) ≍N/rA. Thus, for subcubes Q_i close to the origin, we have ^*(Q_i) ≍ N. And for subcubes away from the origin, e.g. the ones in B(0,1)\ B(0,1/2), we have ^*(Q_i) ≍ N/A. § DOUBLING INDEX AND THE FREQUENCY FUNCTION The frequency of a harmonic function is a well-studied close cousin of the doubling index. Due to its analytic nature, it is sometimes easier to study the frequency than the doubling index. In this section, we first define the frequency of a harmonic function and then present some of its well-known properties. Later, we discuss the relation between the frequency and the doubling index. All the results presented in this section are well-known; for completeness, we give their proofs in Appendix <ref>. §.§ Frequency of a harmonic function Given a harmonic function u in ℝ^n, x∈ℝ^n and r>0, we define H(x,r):= ∫_∂ B(x,r) u^2 dS, where dS is the surface measure on ∂ B(x,r). Moreover, we also let G(x,r):= ∫_B(x,r) |∇ u|^2 d. The frequency function is defined by β(x,r):= r G(x,r)/H(x,r). Agmon <cit.> and Almgren <cit.> proved that β(·) is a monotonic function. For any harmonic function in ℝ^n, its frequency function β(r):=β(0,r) is a non-decreasing function of r. Garofalo and Lin <cit.> generalized it to elliptic operators with variable coefficients, see also the work of Kukavica <cit.> and Mangoubi <cit.>. We would like to mention an elegant work of Colding and Minicozzi <cit.>, which proves monotonicity of the frequency function for parabolic operators in a suprisingly elementary way and in great generality. An equivalent way to define the frequency function is β(x,r):= rH'(x,r)/2H(x,r) - n-1/2, where the derivative is taken with respect to r, see Appendix <ref>. That is the frequency is the logarithmic derivative of H, up to the factor of r and constants. The following is an immediate corollary of this definition and the monotonicity of β. For any harmonic function in ℝ^n, let β(r)= β(0,r) be its frequency function. For any 0<r_1<r_2, we have H(r_2)/r_2^n-1= H(r_1)/r_1^n-1exp( 2 ∫_r_1^r_2β(r)/rdr). Moreover, we also have (r_2/r_1)^2β(r_1) + n-1≤H(r_2)/H(r_1)≤(r_2/r_1)^2β(r_2) + n-1. To define the frequency, it is not necessary for the harmonic function to be defined in the whole of ^n. For any R>1 and any harmonic function defined in a neighbourhood of the closed ball B(x,R)⊂^n, the frequency is well defined by (<ref>) for r∈ [0,R]. All the results about the frequency (namely Theorem <ref>, Corollary <ref>, and Lemmas <ref> & <ref>) hold in any ball where the harmonic function is defined. §.§ Relation between the frequency and the doubling index Thanks to the standard elliptic estimates (see the proofs in Appendix <ref>), we have the following comparison between the frequency in a ball and the doubling index in a slightly larger ball. Let δ>0 be a sufficiently small parameter and let 𝒩(r):= 𝒩(B(0,r)), for r>0. There exists a constant C=C(δ,n)≥ 1 such that β (r(1+δ))(1-100δ) -C≤𝒩(r) ≤β (2r(1+δ))(1+100δ) +C uniformly for all 0<r≤ 4^-1 and all harmonic functions u defined in B(0,2). The above lemma together with the monotonicity of the frequency (Theorem <ref>) immediately imply the following almost monotonicity property of the doubling index. Let δ>0 be a sufficiently small parameter and let u be a harmonic function in B(0,2). Let 𝒩(r):= 𝒩(B(0,r)), for r>0. There exists a constant C=C(δ,n)>0, independent of u, such that (r_2/r_1)^𝒩(r_1)(1-δ)-C≤sup_B(0,r_2)|u|/sup_B(0,r_1)|u|≤(r_2/r_1)^𝒩(r_2)(1+δ)+C, for all r_1≤ r_2/2≤ 4^-1. In particular, we have 𝒩(r_1)(1-δ) - C≤𝒩(r_2)(1+δ) + C. Let B⊂ℝ^n be any ball and let u be a harmonic function in 2B. There exists some constant C=C(n)>1 such that 𝒩_u(b) ≤ C 𝒩_u(B) for all balls b with 2b⊂ B. § DISTRIBUTION OF THE DOUBLING INDEX AND HOW TO USE IT In this section we review the known results from <cit.>, <cit.> about the distribution of the doubling index and formulate a new bound (Proposition <ref>) , which is a crucial ingredient of the paper for proving a lower bound for the nodal volume. §.§ Upper bounds for the distribution of doubling index The following result is borrowed from <cit.>. A sketch of the proof is provided in Appendix <ref>. The proposition below states that if we chop a big cube into smaller subcubes, then most of the subcubes have smaller doubling index than that of the big cube. Let a cube Q in ℝ^n be partitioned into A^n equal subcubes Q_i. Let u be a harmonic function in 4√(n)· Q. There exist numerical (depending only on the dimension n) constants A_0,C > 1 and c>0 such that if A>A_0, then the number of Q_i with ^*(Q_i)>max(^*(Q)exp(- c log A/ loglog A ), C) is smaller than A^n-1-c. The work of Donnelly–Fefferman <cit.> gives a different kind of information about the distribution of the doubling index. Let Q be a unit cube in ^n and let u be a harmonic function in 4√(n)· Q. Let ε>0 and let Q be partitioned into equal subcubes {Q_i}_i of side length ≍_nc_u^*(Q) for some sufficiently small c=c(n,ε)>0. There exists a constant C = C(n,ε) > 1 such that ^*(Q_i)≤ C for at least (1-ε) portion of the Q_i. Although Proposition <ref> is known to experts and is attributed to Donnelly and Fefferman, we could not find a precise reference. A version of it appeared in the study of the doubling index for Laplace eigenfunctions <cit.>. Roughly speaking, if we consider a Laplace eigenfunction on a compact Riemannian manifold with eigenvalue λ, and cover the manifold with balls of radius ≍λ^-1/2, the doubling index for most of these balls is bounded. We will not use Proposition <ref> by Donnelly and Fefferman in our present work. Nevertheless, it follows in a straightforward way from their results in <cit.> about holomorphic functions and the method of holomorphic extension. A formal deduction of Proposition <ref> is contained in Appendix <ref>. The result is formulated here for the sake of completeness and in connection to Conjecture <ref> (almost additivity). To explain the connection we have to formulate another conjecture. Assume that a harmonic function u has stable growth in a unit ball B and 𝒩_u(B)=N ≫ 1. Let us cover 2B by a lattice of cubes Q_i of size 1/N. Then the number of Q_i in 2B, which contain a zero of u, is at least cN^n for some c>0 depending on the dimension n. Let us make a few remarks and speculations. The constant c in Conjecture <ref> might become worse as we change the constant in the definition of stable growth. Conjecture <ref> about many separated zeroes and Proposition <ref> imply that Conjecture <ref> is true under the assumption of stable growth. Roughly speaking they imply that one can find ≍ N^n cubes Q_i of size 1/N (N=𝒩_u(B)), which contain zeroes of u and have a bounded doubling index. Claim <ref> implies that 2Q_i must have nodal volume ≳ 1/N^n-1. So ℋ^n-1({u=0}∩ 2B) ≳ N in the case of stable growth (this additional assumption can be earned for free). §.§ New result on the distribution of doubling index Let us remind the definition of scaled doubling index. For a harmonic function u in a ball 2B ⊂^n, we define the scaled doubling index (s.d.i.) of B, denoted by 𝒮𝒩(B), to be 𝒮𝒩(B) := (B) r^n-1, where r is the radius of B. For a harmonic function vanishing at the center of a ball B, 𝒮𝒩( B) is conjectured (Conjecture <ref>) to give a lower bound for its nodal volume in 2B. Complementary to the known results (<ref>), our main new result also gives a lower bound for the distribution of the doubling index. Given a ball B and a harmonic function on 4B, we show that there is a collection of disjoint balls with doubling index smaller than (2B), but whose total s.d.i. is almost as large as ( B). Let B⊂ℝ^n be a unit ball, u: 4B a harmonic function, and A>1 a sufficiently large parameter. Let N_1:= (1/2B) and N_2:= 𝒩(2B). Then there exists a constant C_A>1 such that whenever N_1≥ C_A, there exists a collection of disjoint balls {B_i = B(x_i,r_i)}_i ∈ℐ such that * We have an upper bound on the doubling index: max(2B_i) ≲_n N_2/A. * We have a lower bound on the sum of the s.d.i.'s: ∑𝒮𝒩(12B_i)=∑(12B_i) r_i^n-1≳_n N_1/(log Aloglog A)^n-1. * Vanishing and non-intersection properties: u(x_i) = 0, 4B_i ∩ 4 B_j=∅, and 4B_i⊂ 2B, for all i≠ j. We now briefly sketch how one can use Proposition <ref> to implement the naive idea discussed in <ref> in order to prove a lower bound for the zero set. One can recursively apply Proposition <ref> in order to obtain a collection of balls with bounded doubling indices. Roughly speaking, in each step we apply Proposition <ref> to each of the balls with large doubling index and replace it by a collection of sub-balls with smaller doubling index. Conditions (<ref>) and (<ref>) guarantee that in a finite number of steps all balls will have bounded doubling index and u must vanish at the center of each ball. The number of steps can be estimated in terms of N_2 and A. Condition (<ref>) should be interpreted as a condition ensuring that after recursive applications of Proposition <ref>, we don't lose too much of the nodal set. Here is an informal explanation. The sum ∑𝒮𝒩( B_i) imitates the nodal volume. If we believe in Conjecture <ref>, then ∑ℋ^n-1({u=0}∩ 2B_i) ≳∑𝒮𝒩( B_i) ≳N_1/(log Aloglog A)^n-1, and the loss of the nodal volume in each step is subpolynomial in A. In the last step all balls have bounded doubling index. Recall that if u has a bounded doubling index in a ball B and vanishes at its center, then Claim <ref> gives the following estimate for the nodal volume 𝒮𝒩( B) ≍ℋ^n-1({u=0}∩ B) ≍ (radius(B))^n-1. We can apply Claim <ref> to each of the balls in the last step and the control (<ref>) of the sum of s.d.i. provides a lower bound for the nodal volume. In the next section we formalize the idea and quantify what kind of a lower bound we get for the nodal volume. §.§ How to use Proposition <ref> We will start proving Proposition <ref> only in <ref>. Now, we show how Proposition <ref> implies the following lower bound for the nodal volume. Let B⊂ℝ^n be a unit ball, and u: 4B a harmonic function. Assume that ε>0, N_1 := 𝒩(1/2B) ≫_n,ε 1, and N_2 := 𝒩(2B). Then we have ℋ^n-1( {u=0}∩ 2B)≳_n,ε N_1^1-ε(N_1/N_2)^ε. We observe that if u has stable growth in a unit ball B, i.e. the ratio N_2/N_1 is bounded by some numerical constant (possibly depending on the dimension n), then Lemma <ref> immediately implies Theorem <ref> under the assumption of stable growth. Unfortunately, the bound in Lemma <ref> gets worse as the ratio N_2/N_1 gets larger. However, in <ref> we will show that Lemma <ref> implies Theorem <ref> by reducing the general case to the stable growth case. Given ε>0, we will choose a large constant A=A(ε)≥ 2 later. Let C_A≥ 1 be given by Proposition <ref>, and assume that C_A≫_n 1. Starting with S_0 := {B}, for every k ∈, we will recursively use Proposition <ref> to construct a collection of balls S_k = {B_α}. This sequence will eventually stabilize, that is, there exists K≥ 0 such that S_k = S_K for every k ≥ K. Moreover, the final collection S_K will be such that for every ball B_α∈ S_K, u vanishes at its center and (12B_α) ≤ C_A. Let S_0 := {B}. Given S_k = {B_α}, we will define S_k+1. Suppose B_α∈ S_k is such that (1/2 B_α) ≤ C_A, then we retain B_α in S_k+1. If (1/2 B_α) > C_A, then we apply Proposition <ref> to B_α to obtain a collection of balls {B_α,i}_i satisfying the three conditions of Proposition <ref>, and this collection will replace B_α in S_k+1, i.e., S_k+1 := {B_α: B_α∈ S_k and (12 B_α) ≤ C_A}∪{B_α,i: B_α∈ S_k and (12 B_α) > C_A}. For B_α∈ S_k, let x_α and r_α be its center and radius respectively. We claim that the collection S_k satisfies the following properties: * For some (large) C=C(n)≥ 1, we have for every B_α∈ S_k: (2B_α) ≤ C^k N_2/A^k or (12 B_α) ≤ C_A. * For some (small) c=c(n)>0, we have ∑_B_α∈ S_k(12B_α) r_α^n-1≥ c^k N_1/(log Aloglog A)^k(n-1). * Vanishing and non-intersection properties: u(x_α) = 0 , 4B_α∩ 4B_γ=∅, and 4B_α⊂ 2B, for all B_α,B_γ∈ S_k with B_α≠ B_γ. We prove by induction that the above properties <ref>-<ref> hold for every S_k. Properties <ref>-<ref> hold for S_1 by the statement of Proposition <ref>. Assuming that all the above properties hold for S_k, let us show that they hold for S_k+1 as well. It is an immediate consequence of Proposition <ref> and the way S_k is constructed that <ref> and <ref> hold. We now show that <ref> also holds. We partition S_k into S_k^(1) and S_k^(2) where S_k^(1) = {B_α∈ S_k: (12 B_α) ≤ C_A} and S_k^(2) = S_k ∖ S_k^(1). For B_α∈ S_k^(2), we denote by {B_α,i}_i the collection of balls obtained by applying Proposition <ref> to B_α. Let r_α,i be the radius of B_α,i, then we have ∑_B_j ∈ S_k+1(12 B_j) r_j^n-1 = ∑_B_α∈ S_k^(1)(12 B_α) r_α^n-1 + ∑_B_α∈ S_k^(2)∑_i(12 B_α,i) r_α,i^n-1. It follows from the second condition of Proposition <ref> that ∑_i(12 B_α,i) r_α,i^n-1≥c/(log A loglog A)^n-1·(12 B_α) r_α^n-1. We thus have ∑_B_j ∈ S_k+1(12 B_j) r_j^n-1 ≥∑_B_α∈ S_k^(1)(12 B_α) r_α^n-1 + c/(log A loglog A)^n-1∑_B_α∈ S_k^(2)(12 B_α) r_α^n-1, ≥c/(log A loglog A)^n-1∑_B_α∈ S_k(12 B_α) r_α^n-1, ≥c^k+1/(log A loglog A)^(k+1)(n-1)· N_1, where the last inequality follows by our induction hypothesis that <ref> holds for S_k. We now show that S_k stabilizes in ≍log N_2/log A steps. Because of <ref>, at every step and for every ball B_α∈ S_k either (12 B_α) ≤ C_A (and then B_α is frozen forever) or the doubling index of 2B_α drops by a factor of A/C, which can be crudely estimated by √(A) (assuming that A is sufficiently large). The latter scenario can happen only ≍log N_2/log A times as we start with N(2B)=N_2. Indeed, otherwise doing k ≥ 2 log N_2/log A steps, we have a ball B_α∈ S_k with (2B_α) ≤ N_2/ A^k/2≤ 1 and by almost monotonicity of the doubling index it implies ( B_α) ≤ C (2B_α) ≤ C ≤ C_A. So after K := ⌈ 2 log N_2/log A ⌉ steps all balls have bounded doubling index. By Lemma <ref> we have ℋ^n-1({u=0}∩ 2B) ≥∑_B_α∈ S_Kℋ^n-1({u=0}∩ B_α) ≥ c ∑_B_α∈ S_K(12 B_α)^1-n r_α^n-1 ≥ c · C_A^1-n∑_B_α∈ S_k r_α^n-1. It is time to use <ref> to get a lower bound (<ref>): C_A ∑_B_α∈ S_k r_α^n-1≥∑_B_α∈ S_k(12 B_α) r_α^n-1≥c^K/(log A loglog A)^K(n-1) N_1, thus, ℋ^n-1({u=0}∩ 2B) ≳ C_A^-nc^K/(log A loglog A)^K(n-1) N_1. We may assume that N_2≫ 1 because N_1 ≫ 1. Since K = ⌈ 2 log N_2/log A ⌉, we can always choose A=A(ε)≥ 2 so that c^K≥ N_2^-ε/2 and (log Aloglog A)^K(n-1)≤ N_2^ε/2. Hence, absorbing the factor C_A^-n, which now depends only on ε, in the ≲_ε notation, we have C_A^-n· c^K N_1/(log Aloglog A)^K(n-1)≳_ε N_1 · N_2^-ε= N_1^1-ε(N_1/N_2)^ε, concluding the proof of Lemma <ref>. § STRUCTURE OF THE PROOFS OF THEOREM <REF> AND PROPOSITION <REF> block/.style=rectangle split, draw, rectangle split parts=2, text width=14em, text centered, rounded corners, minimum height=4em, grnblock/.style=rectangle, draw, fill=white, text width=14em, text centered, rounded corners, minimum height=4em, sgrnblock/.style=rectangle, draw, fill=white, text width=7em, text centered, rounded corners, minimum height=2em, newsgrnblock/.style=rectangle, draw, fill=white, text width=9em, text centered, rounded corners, minimum height=2em, slgrnblock/.style=rectangle, draw, fill=white, text width=13em, text centered, rounded corners, minimum height=2em, whtblock/.style=rectangle, draw, fill=white!20, text width=14em, text centered, minimum height=4em, line/.style=draw, -Latex[length=2mm,width=1mm], cloud/.style=draw, ellipse,fill=white!20, node distance=3cm, minimum height=4em, container/.style=draw, rectangle,dashed,inner sep=0.28cm, rounded corners,fill=white,minimum height=4cm In Figure <ref>, we present a logical progression of ideas/results leading to our main result Theorem <ref>. We now indicate where to find these results, and also what we do in the remaining sections. We advise the reader to keep Figure <ref> and the following as a guide while navigating through the rest of the paper. <ref> Distribution of the doubling index and how to use it: In this section, we present old and new results on the distribution of the doubling index of a harmonic function. Given a cube Q ⊂^n which is partitioned into equal subcubes Q_i, Proposition <ref>, which was established in <cit.>, gives an upper bound on the number of cubes with large doubling index. Proposition <ref> is a new result about the distribution of doubling index. For a harmonic function u defined in a ball B, Proposition <ref> gives a collection of disjoint balls in B satisfying the following properties: u vanishes at their centers, their total s.d.i. is large, and their doubling index is much smaller than (B). We explain how to use Proposition <ref> in Lemma <ref>, which gives a lower bound for the nodal volume. <ref> Stable growth: In this section, we study the behaviour of a harmonic function in a ball with stable growth (Definition <ref>). In <ref>, we prove Lemma <ref>, which gives estimates for the growth of a harmonic function in a spherical layer where the frequency is almost constant. In some instances, we will prove results about the nodal set of a harmonic function in a ball B, under the simplified assumption of stable growth. Claim <ref>, proved in <ref>, serves as a bridge between the stable growth case and the general case by implying that for any harmonic function defined in ball B, there are several balls B_k ⊂ B of significant size with stable growth. Lemma <ref> implies Theorem <ref> under the additional assumption of stable growth. We show that this along with Claim <ref> implies the general case of Theorem <ref>. <ref> Reducing Proposition <ref> to the stable growth case: We state Lemma <ref>, which is a version of Proposition <ref> with the additional assumption of stable growth. Using Claim <ref> again, we reduce Proposition <ref> to Lemma <ref>. <ref> Zeros of harmonic functions in a tunnel: In this section, we study harmonic functions defined in a tunnel, that is a hyper-rectangle in ^n with all sides but one of equal length, while the length of the remaining side is much larger than the others. In Lemma <ref>, we show that for a harmonic function u defined in a good tunnel (which is a tunnel where u has significant growth along the longer side of the tunnel, and there is some control on the doubling index at all points in the tunnel), there are many disjoint balls where u vanishes and the sum of their s.d.i.'s is large. <ref> Connecting the dots: Proof of Lemma <ref>: In this final section, equipped with Proposition <ref>, Lemma <ref> and Lemma <ref>, we finish the proof of Lemma <ref>. Recall that Lemma <ref> is Proposition <ref> with the additional assumption of stable growth. We first use Lemma <ref> to get several tunnels where the harmonic function has significant growth along their longer sides. We then use Proposition <ref> to control the doubling index at all points in at least half of these tunnels, thus giving us several good tunnels. Finally in these good tunnels, we apply Lemma <ref> to conclude that there are several disjoint balls satisfying the conclusions of Lemma <ref>. § STABLE GROWTH The notion of stable growth has already appeared a few times in the earlier sections; we are finally ready to formally introduce it. §.§ Definition of stable growth and some consequences Let C>1 be a constant depending only on the dimension n (which will be chosen in Remark <ref>). A harmonic function u:ℝ^n is said to have stable growth in a ball B⊆^n if C^-1(12B) ≤(2B) ≤ C (12B). In (<ref>), the first inequality C^-1( B) ≤(2B) automatically holds as a consequence of the almost monotonicity of the doubling index (Lemma <ref>). The main point of the stable growth condition (<ref>) is the other inequality, namely, (2B) ≤ C( B). Any homogeneous harmonic polynomial p of degree N ≥ 1 has stable growth in every ball B(0,r) ⊆^n. Such p can be written as p(r,θ) = r^N q(θ), for r≥ 0, θ∈𝕊^n-1, where q: 𝕊^n-1 is a spherical harmonic of degree N. Then sup_B(0,r)|p| = r^Nsup_𝕊^n-1|q|, and the doubling index (0,r) ≡ N, for every r>0. Any harmonic polynomial P has the following representation: P(x) = ∑_k=m^M p_k(x), where p_k is a homogeneous harmonic polynomial in ^n, of degree k. For 0<r≪ 1 and R ≫ 1, we have P(r,θ) = p_m(r,θ) + O(r^m +1), P(R,θ) = p_M(R,θ) + O(R^M-1). Hence for small r, we have (0,r) ≍ m; for large R, we have (0,R) ≍ M. Hence P has stable growth in balls centered at the origin, whose radius is either small or large. In the case M=2m, for the polynomial P(x) = ∑_k=m^2m p_k(x) we have (0,r) ≍ m for small r, and (0,R) ≍ M ≍ m for large R. So by the monotonicity of the doubling index, (0,t) ≍ m for all t>0. Hence P has stable growth in every ball B(0,r). We observe that if a harmonic function u has stable growth in a ball B=B(x,r) ⊂^n with 𝒩(1/2B)=: N≥ 1, then because of the close relation between the doubling index and the frequency (Lemma <ref>), the frequency β satisfies β_u(x,t)∈[c_1 N, C_1 N], ∀ t∈ [1.1r,1.9r] where constants c_1,C_1>0 depend only on n. §.§ Choosing the stability constant Let B=B(0,1) ⊂^n be the unit ball and let u be a harmonic function in 4B. Consider the spherical layer 𝒮_ρ,w:={x∈ℝ^n : ρ-10w≤ |x|≤ρ + 10 w}⊂ B(2)\ B(1), with width 20w ∈ (0,ρ). Suppose that the frequency of u in this layer satisfies β_u(0,t) ∈ [N,10N] , for t∈ [ρ- 10w, ρ + 10w]. Also assume that the layer is not too narrow in terms of N, in the sense that N w(log1w)^-1n≫ 1. Then there exists a constant C >0 and a ball D of radius w such that 4D ⊂𝒮_ρ,w satisfying C^-1( D) ≤(2D) ≤C( D), and (2D) ≍_n wN. From now on, we fix the stability constant in Definition <ref> to be C, as given in the statement of Lemma <ref>. The following result captures the growth of a harmonic function with stable growth. Let B=B(0,1) ⊂^n be the unit ball and let u be a harmonic function in 4B. Consider the spherical layer 𝒮_ρ,w:={x∈ℝ^n : ρ-10w≤ |x|≤ρ + 10 w}⊂ B(2)\ B(1), with width 20w ∈ (0,ρ). Suppose that the frequency of u is comparable to N>1 in this layer, that is, β_u(0,t) ≍_n N for t∈ [ρ- 10w, ρ + 10w]. Also assume that the layer is not too narrow in terms of N, in the sense that N w(log1w)^-1n≫ 1 (the larger N, the smaller w can be). Define the function M(t):= sup_tB|u|. Then, for all t_1<t_2∈ [ρ- 3w,ρ+ 3w] and |t_2-t_1|≥ w/2, we have exp(λ(t_2-t_1)N)<M(t_2)/M(t_1)<exp (Λ (t_2-t_1)N), for some constants λ=λ(n)>0 and Λ=Λ(n)>1. Lemma <ref> will be used to prove Lemma <ref> and also Lemma <ref>. Under the assumptions of Lemma <ref> and maintaining the same notation, we have the following. Let x_0 ∈∂(ρ B) be a point where |u| attains its maximum on ρ B, that is, |u(x_0)| = M(ρ). If D is any ball of radius w such that x_0 ∈∂D, then (D) ≲_n wN. * If D ⊂ρ B is a ball of radius w such that x_0∈∂ D, then 4D ⊂𝒮_ρ,w and (12D)≍_n 𝒩(2D)≍_n wN. * If D̃ is any ball of radius w such that x_0 ∈∂D̃, then we have the following upper bound for (D̃): (D̃) ≲_n wN. We are now ready to prove Lemma <ref>. To simplify the notation, we write H(r) = ∫_∂(rB)u^2 dS, and β(r) instead of the frequency β(0,r). We know that H is a non-decreasing function, and thanks to Corollary <ref>, we also know a great deal about the growth of H in an interval where β is nearly constant. For harmonic functions, we can compare the L^2 norm √(H(·)) with the L^∞ norm M(·). Hence we should also be able to get information about the growth of M(·) in an interval where β is nearly constant. We now establish the upper bound in (<ref>), the lower bound can be shown using a similar argument. Let x_2 ∈∂ (t_2 B) be such that M(t_2)= |u(x_2)|, and let us write s= (t_2-t_1). We have u(x_2)^2 ≲ s^-2n(∫_B(x_2,s)|u| )^2 ≲ s^-n∫_B(x_2,s)u^2≤ s^-n∫_B(0,t_2+s)u^2. Since H(·) is non-decreasing, we have ∫_rB u^2= ∫_0^r∫_∂(tB) u^2(x) dS(x) dt = ∫_0^r H(t)dt≤ r H(r). Since (t_2 + s)<4, we have the following comparison between H and M: M(t_2)^2 ≲_(<ref>) s^-n∫_B(0,t_2 +s)u^2 ≲_(<ref>) s^-n H(t_2 +s). Since (t_2+s) ∈ (ρ-10w, ρ + 10w), we have by our assumption on β that β(t_2+s) ≍ N. Corollary <ref> implies that for some C=C(n)>0, we have H(t_2+s)/H(t_1)≤(t_2+s/t_1)^C N= t_1 +2s/t_1^C N≤ (1+2s)^C N≤exp(2C sN), since t_1 >1. Hence H(t_1) ≥ H(t_2 +s) exp(-2Cs N) ≳_(<ref>) s^n exp(-2Cs N) M(t_2)^2. From the trivial comparison H(t_1) ≲ M(t_1)^2 and (<ref>), we have M(t_2)^2 ≲ M(t_1)^2 exp(2Cs N + n log1s). Since s ≍ w, and Nw ≫log1w, we also have Ns ≫log1s. Hence there is a constant Λ = Λ(n) such that 2Cs N + n log1s≤ 2Λ s N = 2Λ (t_2 - t_1)N, and this establishes the upper bound in (<ref>). We now prove Lemma <ref> and Lemma <ref>. Let x_0 be a point where |u| attains its maximum in ρ B. Let D ⊂ρ B be a ball of radius w such that x_0∈∂ D. Then we have 4D ⊂ (ρ+3w)B and by Lemma <ref> sup_4D|u|≤ M(ρ+3w) ≤ M(ρ) exp(3Λ w N). Since D ⊂ (ρ- w)B, by Lemma <ref> we have sup_1/2D |u| ≤ M(ρ - w) ≤ M(ρ) exp(-λ w N/2). Using sup_D|u|=M(ρ) and sup_2D|u|≥ M(ρ), we have λ w N/2≤logsup_D |u|/sup_1/2D|u| = (12D) and (2D) = logsup_4D|u|/sup_2D|u|≤ 3Λ w N, which concludes the proof of Lemma <ref>. The proof is similar to the proof of Lemma <ref> and follows by observing that sup_D|u| ≥ M(ρ) and sup_2D|u| ≤ M(ρ) exp(3Λ wN), where the second inequality holds because 2D⊂ (ρ+3w)B. This completes the proof of Lemma <ref>. §.§ Proof of Theorem <ref> assuming Proposition <ref>. We now state a useful claim about the frequency. Although we state it specifically for the frequency, an analogous statement is true for more general monotone functions φ:[a,b]. It essentially says that for any monotone function, it is possible to find a subinterval of significant length where the function is almost constant. Let u be a harmonic function defined in 4B, with B =B(0,1) ⊂ℝ^n. Suppose that 𝒩(1/2B)=N_1≫_n 1 and let 𝒩(2B)=N_2. Then there exists some N satisfying N_1 ≲_n N ≲_n N_2 such that β(0,t)∈ [N, eN], ∀ t ∈ℐ where ℐ⊂ [1.1,1.9] is some interval of length |ℐ|≳_n (log N)^-2. In this proof, we write β(r) instead of β(0,r). We first observe that as a consequence of Lemma <ref> and assumption N_1 ≫_n 1, we have 1 ≪_n N_1 ≲_n β(1.1) ≤β(1.9) ≲_n N_2. Depending on the growth of the frequency β, we partition the interval [1.1,1.9] as follows. Let k ∈ be such that k:=⌊logβ(1.9)β(1.1)⌋. For 0 ≤ j ≤ k, let a_j ∈ [1.1,1.9] be defined by a_j := inf{t∈ [1.1,1.9]: β(t)≥ 10^jβ(1.1)}. If there is some j such that 0≤ j ≤ k-1 and a_j+1 - a_j≥1/100 (logβ(a_j))^-2, then we can take ℐ = [a_j,a_j+1]. If (<ref>) does not hold for any j, then we have ∑_j=0^k-1 a_j+1 - a_j ≤1/100∑_j=0^k-1(logβ(a_j))^-2≤1/100·∑_j=0^k-11/(logβ(1.1) + j)^2, a_k - 1.1 ≤1/100·∑_j=0^∞1/j^2≤ 0.02, and hence a_k ≤ 1.12. By our definition of k, we also have 1≤β(1.9)/β(a_k) ≤ e. Hence β(t) ∈ [β(a_k),eβ(a_k)],∀ t∈ [a_k,1.9]. Also notice that 1.9 - a_k ≥ 0.7 ≫ (logβ(1.1))^-2≥ (logβ(a_k))^-2. This shows that at least one of the intervals [a_j,a_j+1] or [a_k,1.9] satisfies the required condition, and this completes the proof of Claim <ref>. First let us define the points r_0= 1.1 and r_k+1=r_k + 1/100 logβ(r_k), for all k≥ 0 subject to r_k+1< 1.8. Suppose, by contradiction, that β(r_k+1)≥ 10 β(r_k), for all k. Then, since β(1.1)≥ N_1/2 by Lemma <ref> (with δ=10^-10) and the fact that N_1 is large, we have β(r_k)≥ 10^k and r_k+1- r_k = 1/100 logβ(r_k)≤1/100 k^2. Since 100^-1∑_k k^-2< 0.5, the monotonicity of β(·), that is Theorem <ref>, implies β(r_k)≤β(1.6), a contradiction to β(r_k)≥ 10^k. All in all, we have shown that there exists some r_k_0∈ [1.1,1.8] such that N_1/2≤β(r_k_0)≤ 2N_2 and β(r_k_0+1)≤ 10 β(r_k_0), where we have used again Lemma <ref> (with δ=10^-10). Hence, we may take β(r_k_0)= N and I_n=[r_k_0+1, r_k_0], as required. Lemma <ref> implies Theorem <ref>, under the stable growth assumption. We now show that Theorem <ref> also follows in the general case. Let us recall the setting of Theorem <ref>. We have the unit ball B=B(0,1), and a harmonic function u in 4B, which vanishes at the center of B. We wish to get a lower bound for the nodal volume of u in 2B, in terms of ( B). Define N_1:= ( B) and fix ε >0. We may assume that N_1 ≫_ε, n 1, otherwise we may employ the elementary Lemma <ref>. We first use Claim <ref> to conclude that there is N ≳_n N_1, and an interval ℐ of length |ℐ| ≍ (log N)^-2 on which β∈ [N,eN]. Then we apply Lemma <ref> to ℐ and conclude that there is a ball D of radius ≍ (log N)^-2 with stable growth satisfying: (D) ≍ N(log N)^-2 and 4D ⊂ 2B. Finally, we use Lemma <ref> for this ball to get the following lower bound: ℋ^n-1({u=0}∩ 2D) ≳_n, (N(log N)^-2)^1-· (log N)^-2(n-1)≥ N^1-(log N)^-2n≥ N^1-2, where the last inequality follows because N ≫ 1. Now since 4D ⊂ 2B, we can also conculde that ℋ^n-1({u=0}∩ 2B) ≳_n, N^1-2, and this completes the reduction of Theorem <ref> to Lemma <ref>. Lemma <ref> was already reduced to Proposition <ref>, which is the only thing left to prove. Let N_1=𝒩(1/2B) and let N_0 be as in Claim <ref>. If N_1≤ N_0, then we may apply Lemma <ref> and Theorem <ref> follows by taking the constant c sufficiently small. Therefore, we may assume that t N_1=𝒩(1/2B)≥ N_0. Thanks to Claim <ref>, there exists some N_1/2≤ N and some interval I=I_N⊂ [1/2,2] of length |I|≍_n (log N)^-2 such that β_u(t)∈ [N, 10N] for all t∈ I. Therefore, by Lemma <ref>, there exists a ball B̃ of radius r≍_n (log N)^-2 such that 4B̃⊂ B and (1/2B̃)≍_n 𝒩(2B̃)≍_nN/(log N)^2. Finally, applying a rescaled version of Lemma <ref>, we find ℋ^n-1( {u=0}∩ 2B)≥ℋ^n-1( {u=0}∩ 2B̃) ≥ r^n-1(1/2B̃)^1-ε/2≥ N^1-ε, where we have absorbed the logarithmic factors into the term N^-ε. This concludes the proof of Theorem <ref>. § REDUCING PROPOSITION <REF> TO THE STABLE GROWTH CASE The aim of this section is to show that it suffices to prove Proposition <ref> under the additional assumption that u has stable growth in B. The following result is exactly Proposition <ref> with the extra assumption of stable growth. Let B be a unit ball, u: 4B⊂ℝ^n be a harmonic function and A>1 be a sufficiently large parameter. There exists a constant C_A>1 such that the following holds. Assume that u has stable growth in B and (1/2B)=N≥ C_A. Then there exists a collection of disjoint balls B_i = B(x_i,r_i) such that: * We have an upper bound on the doubling index: max_i (2B_i) ≲_n N/A. * We have a lower bound on the total s.d.i.'s: ∑ (12B_i) r_i^n-1≳_n N_1/(log Aloglog A)^n-1. * Vanishing at the center and non-intersection properties: u(x_i) = 0, 4B_i ∩ 4 B_j=∅, and 4B_i⊂ 2B,∀ i≠ j. Let x be the center of B; we will write β(·) for β(x,·). Depending on the growth of the frequency β, we divide the interval [1.1,1.9] as follows. Let k ∈ be such that k:=⌊logβ(1.9)β(1.1)⌋. For 0 ≤ j ≤ k, let a_j ∈ [1.1,1.9] be such that β(a_j) = 10^jβ(1.1). Define the interval I_j:= [a_j, a_j+1] of length 20w_j := a_j+1 - a_j. Observe that on each of the intervals I_j, the frequency β does not grow more than by a factor of 10. If w_j satisfies 10^j β(1.1) ≥ C_n ·1w_jlog1w_j, for some sufficiently large constant C_n>0, then we can use Lemma <ref> to get a ball B_j of stable growth in B(a_j+1)∖ B(a_j). But some of these intervals I_j can be too short and (<ref>) may not hold for these intervals, and we will discard them. We will show that the sum of the w_j satisfying (<ref>) is at least 1/2 if β(1.1) is large enough (which holds because β(1.1)≳𝒩(1/2B) ≥ C_A ≫ 1). Since w_j ∈ (0,1), and x≥log x for x>1, we have 1/w_j^2≥1/w_jlog1/w_j. Hence if w_j^2· 10^j β(1.1) ≥ C_n, then (<ref>) holds. Thus, letting 𝒥 be the collection of indices j for which w_j^2· 10^j β(1.1) ≤ C_n, we have ∑_j ∈𝒥 w_j ≤√(C_n/β(1.1)) ∑_j ∈1/10^j/2≤√(C_n/β(1.1)). Hence the sum ∑_j ∈𝒥 w_j < 1/3 for β(1.1)≫ 1. Define G = {0,1,…,k}∖𝒥. We have ∑_j ∈G w_j ≥ 1/2, and for each I_j, j ∈ G, we can apply Lemma <ref> to find a ball B_j such that: * u has stable growth in B_j, and (B_j) ≍ w_j 10^j β(1.1). * The radius of B_j is w_j, and 4B_j ⊂ B(0,a_j+1)∖ B(0,a_j). The latter property implies that 4B_j are disjoint. In order to simplify the notation, let us write Φ(A)= (log A loglog A)^n-1 until the end of the proof. For each B_j, j ∈G, we use Lemma <ref> to get a collection of disjoint sub-balls {B_i,j}_i, with radii r_i,j and centers x_i,j, satisfying the following conditions: * Upper bound for the doubling index: max_i (2B_i,j) ≲_n (B_j)/A ≍_n w_j 10^j β(1.1)/A. * Lower bound for the total s.d.i.: ∑_i (12B_i,j) r_i,j^n-1≳_n (B_j)/Φ(A)· w_j^n-1≍β(1.1)/Φ(A)· 10^j w_j^n. * Vanishing at the centers and disjointness: u(x_i,j) = 0, 4B_i,j∩ 4 B_k,j=∅, and 4B_i,j⊂ 2B_j,∀ i≠ k. Since w_j≤ 1, <ref> and Lemma <ref> imply that max_i (2B_i,j) ≲_n w_j 10^j β(1.1)/A ≲_n β(1.9)/A≲_n 𝒩(2B)/A. We observe that <ref> implies (3) in Proposition <ref>. Therefore, we are left to show (2). It suffices to show that, for j ∈ G, we have ∑_i,j(12B_i,j)  r_i,j^n-1≳_n ( B)/Φ(A). Indeed, summing (<ref>) over all j ∈ G and using the bound β(1.1)≳𝒩(1/2B), which follows from Lemma <ref>, we get ∑_j ∈ G∑_i(12B_i,j)  r_i,j^n-1≳_n ( B)/Φ(A)∑_j ∈ G 10^j w_j^n. We will get a lower bound for ∑_j ∈ G 10^j w_j^n by using Holder's inequality which says that, for α_m, γ_m >0, and p,q ≥ 1 such that p^-1 + q^-1 = 1, we have ∑_mα_m γ_m ≤∑_m α_m^p^1/p∑_m γ_m^q^1/q. Choosing p=n, q=nn-1, α_j = w_j 10^j/n, γ_j = 10^-j/n in (<ref>), we get ∑_j ∈ G w_j ≤(∑_j ∈ G 10^j w_j^n)^1/n·( ∑_j ∈ 10^-j/n-1)^n-1/n. By (<ref>), ∑_j ∈ Gw_j ≥ 1/2. We deduce that there exists constant c_n>0 such that ∑_j ∈ G 10^j w_j^n≥ c_n. Using this in (<ref>), we have shown that (<ref>) holds. This concludes the proof of Proposition <ref> (assuming Lemma <ref>). § ZEROS OF A HARMONIC FUNCTION IN A TUNNEL In this section we introduce the notion of a tunnel. A tunnel 𝒯⊂^n is a closed hyper-rectangle with all sides but one of length h>0, while the length ℓ of the remaining side is much larger than h. That is, ℓ≫_n h. We call h and ℓ the width and the length of the tunnel respectively. We will always tacitly assume that m := ℓ/ h ∈. A tunnel has two faces whose sides are all of length h. We designate one of these to be the beginning face and the other one to be the end face, and we will denote them by F_beg and F_end respectively. We can find a unit vector v ∈^n (which is parallel to the longer side of ) and a∈ such that F_beg = ∩{x ∈^n: v · x = a} and F_end = ∩{x∈^n: v · x = a+ℓ}, where · denotes the usual inner product in ^n. We partition into equal hyper-cubes of side length h, and denote this collection by {q_i}_i=1^m. More precisely, for 1≤ k ≤ m: q_k := ∩{x∈^n : a+ (k-1)h ≤ v· x ≤ a + kh}. Note that F_beg⊂ q_1 and F_end⊂ q_m. We will call subcubes q_1 and q_m the beginning and the end of respectively. Henceforth, whenever we talk about a tunnel, we implicitly assume that it comes with a choice of the beginning and the end. With as above, a harmonic function u defined in a neighbourhood of is said to have logarithmic multiplicative increment (l.m.i.) Z>0 in if logsup_q_m|u|/sup_q_1|u|= Z. We now recall the notion of the doubling index of a cube introduced in (<ref>): given a cube Q with diameter r>0, and a harmonic function on √(n)Q, we write 𝒩^*(Q)= sup_x ∈ Q, .05cm ρ∈ (0,r) (x,ρ). We recall the notion of ℓ-neighbourhood of a set S ⊂^n: S_+ℓ := {x ∈^n : x= s+b, where s∈ S and b∈ B(0,ℓ) }. In the following lemma, we show that if a harmonic function u has a large enough multiplicative increment Z in a tunnel , and the doubling indicies in all of the subcubes q_k are much smaller than Z, then the zero set of u in a neighbourhood of is non-empty and is well spread out. That is, there are many disjoint balls (of radius ≍ h) where u vanishes. Let B ⊂^n be a unit ball and let u be a harmonic function in 4B. Let ⊂1.6B be a tunnel of length ℓ and width h. Suppose that u has l.m.i. Z>1 in , and that m= ℓ/ h≪_n Z. Moreover, suppose that max_k 𝒩^*(100√(n)q_k)≤ Z/V, for some positive number V ≪_n Z. Then there exist ≳_n V many disjoint balls B_k = B(y_k,r_k) such that B_k⊂_+50√(n)h and the following holds. * Vanishing and non-intersection properties: u(y_i) = 0, 4B_i ∩ 4 B_j=∅, and 4B_i⊂ 2B, for all i ≠ j. * A lower bound for the sum of the doubling indices: Z ≲_n ∑_k 𝒩(12B_k). Let z_k be the center of q_k. We start with the following claim. Let S ⊂{1,2,…,m-1} be the collection of of k's such that (2q_k ∪ 2q_k+1) contains a zero of u. Then we have |S| ≳_n V and ∑_k ∈ S(z_k, 4√(n)h) ≳_n Z. We first observe that Z = logsup_q_m|u|/sup_q_1|u|= ∑_k=1^m-1logsup_q_k+1|u|/sup_q_k|u|. If u does not vanish on (2q_k∪ 2q_k+1), then by Harnack's inequality, there exists some constant C_0=C_0(n)>0 such that logsup_q_k+1|u|/sup_q_k|u|≤ C_0. Therefore, writing S^c={1,2,…,m-1}\ S, we have Z = ∑_k ∈ S^clogsup_q_k+1|u|/sup_q_k|u| + ∑_k ∈ Slogsup_q_k+1|u|/sup_q_k|u|, ≤ C_0 m + ∑_k ∈ Slogsup_q_k+1|u|/sup_q_k|u|. Since m≪ Z, we have C_0 m ≤ Z/2 and hence Z/2 ≤∑_k ∈ Slogsup_q_k+1|u|/sup_q_k|u|. Since B(z_k, h2) ⊂ q_k, and q_k+1⊂ B(z_k, 2√(n)h), we conclude from Lemma <ref> that logsup_q_k+1|u|/sup_q_k|u| ≤logsup_B(z_k, 2√(n)h)|u|/sup_B(z_k, h2)|u|, ≤log [(4√(n))^2(z_k, 4√(n)h) + C_1], ≤ C_2 (z_k, 4√(n)h) +C_2, where C_1, C_2>0 are dimensional constants. From (<ref>) and (<ref>), we have Z/2≤∑_k ∈ S[C_2 (z_k, 4√(n)h) + C_2 ] ≤ C_2 ∑_k ∈ S(z_k, 4√(n)h) + C_2 m. Since m≪ Z, we have Z≲∑_k ∈ S(z_k, 4√(n)h). By our assumption, each ^*(4√(n) q_k) ≤ Z/V, and hence (z_k,4 h) ≤ Z/V. Hence we have |S| ≳_n V. This completes the proof of Claim <ref>. The collection of balls obtained in Claim <ref> satisfy (<ref>), but the balls might intersect each other and the zeros are not necessarily at their centers. It is easy to modify this collection to also satisfy (<ref>). For k ∈ S, let y_k ∈ (2q_k ∪ 2q_k+1) be a point where u vanishes. The distance between y_k and z_k is at most 3√(n)h. So B(z_k, 8√(n)h) ⊂ B(y_k, 16√(n)h). By almost monotonicity of the doubling index for non-concentric balls (Lemma <ref>) (z_k, 4√(n)h) ≲_n (y_k, 16√(n)h). Thus by Claim <ref> ∑_k ∈ S(y_k, 16√(n)h) ≳_n∑_k ∈ S(z_k, 4√(n)h) ≳_n Z. Define B_k := B(y_k, 32√(n)h). Then ∑_k ∈ S(12 B_k) ≳_n Z . Since each of the balls 4B_k intersects at most C=C(n) of other balls 4B_l, we can split the set { 4B_k}_k ∈ S into at most C collections of balls such the balls in each collection do not intersect. Let {4B_k}_k ∈S̃ be the collection of disjoint balls with the maximal sum ∑_k ∈S̃(12 B_k). Then ∑_k ∈S̃(12 B_k) ≥1/C∑_k ∈ S(12 B_k) ≳_n Z. The collection of balls {B_k}_k ∈S̃ satisfies (<ref>) and (<ref>). This completes the proof of Lemma <ref>. The collection of balls obtained in Claim <ref> satisfy (<ref>), but the balls might intersect and the zeros are not necessarily at their centers. It is easy to modify this collection to also satisfy (<ref>). For k ∈ S, let y_k ∈ (2q_k ∪ 2q_k+1) be a point where u vanishes. The following inclusion is easy to see: B(z_k, 8√(n)h) ⊂ B(y_k, 16√(n)h). From this and Lemma <ref>, we can conclude that (z_k, 4√(n)h) ≲_n (y_k, 16√(n)h). Thus from Claim <ref>, and the fact that m ≪ Z, we conclude that ∑_k ∈ S(y_k, 16√(n)h) ≳_n Z. For k ∈ S, define B_k := B(y_k, 32√(n)h). Except for the fact that B_k's can have non-trivial intersection, this collection of balls has all the other desired properties, i.e., u vanishes at the center of each B_k and (<ref>) holds. Note that there is a constant K_0 ∈ such that if |k-k'|≥ K_0, then 4B_k∩ 4B_k' = ∅. Hence we can find a subcollection of balls, denoted {B_k}_k ∈ℐ where ℐ⊂ S, such that |ℐ| ≥ |S|/K_0 ≳_n V, and ∑_k ∈ℐ(y_k, 22√(n)h) = ∑_k ∈ℐ(12 B_k) ≥1/K_0∑_k ∈ S(y_k, 22√(n)h) ≳_n Z. The collection of balls {B_k}_k ∈ℐ has all the desired properties and this completes the proof of Lemma <ref>. § CONNECTING THE DOTS: PROOF OF LEMMA <REF> Throughout this section, we will use the notation for tunnels introduced in <ref>. The proof of Lemma <ref> is split into three parts. In <ref> we show how to construct tunnels in 2B= B(0,2), which have l.m.i. at least N=𝒩(1/2B). In <ref> we use information about the distribution of doubling index from Proposition <ref> to prove that many of these tunnels are good. Good tunnels are those for which the l.m.i. is at least N, and the doubling index in each of its subcubes is much smaller than N. Finally in <ref> we use Lemma <ref> in good tunnels to conclude the proof of Lemma <ref>. §.§ Step 1: Constructing tunnels with large multiplicative increment We first start with a lemma about the growth of harmonic functions. Let B=B(0,1) ⊂^n and let u be a harmonic function in B. There exists C=C(n)>0 such that the following holds. For every r_1, r_2 satisfying 0< 2r_1 < r_2 < 4^-1, and y ∈∂(r_2B), we have sup_B(0,r_1)|u| ≥ |u(y)|(r_2/r_1)^-2𝒩(0,r_2) - C. We use monotonicity of doubling index (Lemma <ref>) to see that |u(y)|/sup_B(0,r_1)|u|≤sup_B(0,r_2)|u|/sup_B(0,r_1)|u|≤(r_2/r_1)^2𝒩(0, r_2) +C, as required. Consider a harmonic function u with stable growth in the unit ball B=B(0,1), with (1/2B) =N. In the following claim, we consider a tunnel whose end is near a point where |u| attains its maximum on ∂(1.5 B), and whose beginning is contained well within 1.5B (for example in 1.3B). We show that if 𝒯 is not too narrow, then u has l.m.i. Z ≳ N in . Let B=B(0,1) ⊂^n. Assume that a harmonic function u in 4B has stable growth in B. Let x_0 ∈∂ (1.5B) be a point where |u| attains its maximum in 1.5B. Let 𝒯 be any tunnel of length 1/4 and width h such that the beginning q_1 satisfies q_1 ⊂ 1.3B. Denote by z_m the center of the ending q_m of and put r=|z_m -x_0|. Assume that h ≪_n 1 and h/2≤ r ≪_n (log1h)^-1 and ( B) = N ≫_n 1hlog1h. Then u has multiplicative increment Z ≳_n N in 𝒯. Let M=|u(x_0)|. We can apply Lemma <ref> to u in the spherical layer {x∈ℝ^n: 1.1≤ |x|≤ 1.9} to conclude that sup_q_1 |u|≤sup_B(0,1.3) |u| ≤sup_B(0,1.5) |u|exp(-c_1 N) = M exp(-c_1 N) for some constant c_1=c_1(n)>0. We conclude from Claim <ref> that sup_q_m |u| ≥sup_B(z_m, h/4) |u| ≥ |u(x_0)| 4r/h^-2(z_m,r)-C, for some C=C(n)>0. Since Nr(log1r)^-1≥ Nh ( log1h)^-1≫_n 1 we can use Lemma <ref> to get (z_m, r) ≲_n rN. Let's plug the latter inequality in (<ref>) and use rN ≫ 1 : sup_q_m |u| ≥ M (2r/h)^-C_1rN = Mexp(-C_1rN log(4r/h)) ≥ Mexp(-C_1rN log (1/h)) for some C_1=C_1(n)>1. Hence, we have the following bound for the multiplicative increment of u in : Z = logsup_q_m|u|/sup_q_1|u|≳_(<ref>),(<ref>) N(c_1 - C_1rlog (1/h)) ≳ N. In the last inequality we use the assumption r ≪_n (log1h)^-1. The proof of Lemma <ref> is finished. §.§ Step 2: Seeking for many good tunnels A randomly chosen tunnel from Lemma <ref> may have a subcube with doubling index comparable to N. In order to apply Lemma <ref>, we need to ensure that all the subcubes in the tunnel have doubling index ≪_n N. Such a tunnel will be called a good tunnel. The goal of this section is to construct many good tunnels in Lemma <ref>. We are proving Lemma <ref> and we keep the notation introduced in its statement. Namely, u is harmonic on 4B and has stable growth in B=B(0,1), with (12B) =N. Let x_0 ∈∂ (1.5B) be such that |u(x_0)| = sup_1.5B|u|. Denote by ℛ a tunnel of length 1/4 and width w ≪ 1, whose longer side is parallel to the vector x_0, and x_0 is the center of its end face. We will assume that (4w)^-1∈ so that we could chop ℛ into equal subcubes. Then ℛ⊂ 1.6B and its beginning is contained in 1.3B. See Figure <ref> We will start chopping ℛ into subcubes and we will need an integer (chopping) parameter K ≫ 1. K and w will be chosen later, depending on A in the statement of Lemma <ref>. Partition ℛ into equal subcubes {Q_j: 1≤ j ≤ (4w)^-1} of side length w. Also partition ℛ into equal subtunnels {_i: 1 ≤ i ≤ K^n-1} of length 1/4 and width w/K. The tunnels _i are such that all their beginning faces and end faces are contained in the beginning face and the end face of ℛ respectively. We denote by {q_i,t: 1≤ t ≤ K(4w)^-1} the subcubes of tunnel _i with side length w/K. See Figure <ref> for a pictorial explanation for this partitioning. It is easy to see that Q_j consists of K^n equal subcubes q_i,t: Q_j = ∪{q_i,t: 1+ (j-1)K ≤ t ≤ jK and 1 ≤ i ≤ K^n-1}. We will show that in at least half of the tunnels _i, the doubling index of all the scaled subcubes 100√(n)q_i,t is much smaller than N, by using a modified version of Proposition <ref>, which is proved in Appendix <ref>. Let a unit cube Q ⊂^n be partitioned into K^n equal subcubes {q_α: 1≤α≤ K^n}. Let u be a harmonic function in 4√(n) Q. Let ^*(2Q) = N. If K ≫_n 1, then we have #{α :^*(100√(n)q_α) ≥max(N· 2^-clog K/loglog K,C)}≤ K^n-1-c, where c,C >0 are dimensional constants. The next claim is a straightforward consequence of the almost monotonicity of doubling index (Lemma <ref>). Let B=B(0,1) ⊂^n. Then for every harmonic function u defined in 4B, and every cube Q⊂ B(0, 1.6) of side length at most (10√(n))^-1, we have 𝒩^*(Q)≲_n 𝒩(2B). We are now ready to show the existence of many good tunnels in ℛ. With the notation as above, we define N^*(𝒯_i) = max_t=1,2,…, K(4w)^-1𝒩^*(100 q_i,t). If K ≫_n 1 and w≍_n (log K)^-1, then #{i: N^*(𝒯_i) ≤ N/ C_K }≥K^n-1/2, where C_K = exp(c' log K/loglog K) for some constant c' = c'(n)>0. In total there are K^n-1 tunnels _i. The above lemma implies that in a neighborhood of at least half of these tunnels the doubling index significantly drops down. If w ≪ 1, we can apply Claim <ref> to get ^*(2Q_j) ≤ C N, for some constant C=C(n)>1. By Corollary <ref> applied to Q_j we have #{q_i,t⊂ Q_j : ^*(100√(n)q_i,t) ≥ NC2^-clog K/loglog K}≤ K^n-1-c, Put c':=c/2 and C_K := exp(c' log K/loglog K). If K≫ 1,then N C 2^-clog K/loglog K≤ N/C_K. There are (4w)^-1 subcubes Q_j. Using the assumptions w^-1≍_n log K and K ≫ 1 we have #{q_i,t⊂ℛ : ^*(100√(n)q_i,t) ≥ N/C_K}≲ w^-1 K^n-1-c≪K^n-1/2, Since the total number of tunnels _i is K^n-1, we conclude that at least half of them satisfy N^*(𝒯_i) ≤ N/C_K. §.§ Step 3: Concluding the proof of Lemma <ref> In order to finish the proof we need to carefully choose the parameters K,w in the construction and chopping of the tunnel ℛ. Recall that ℛ has length 1/4 and width w. Its subtunnels _i have length 1/4 and width h=w/K. We can choose a dimensional constant c_0=c_0(n)>0 and put w=c_0/ log K so that the the following holds. If K≫_n 1 and N ≫_n K(log K)^2, then every tunnel _i ⊂ℛ has a multiplicative increment Z ≳_n N. Formally, we just need to check that the assumptions of Lemma <ref> hold to conclude that the multiplicative increment Z ≳_n N. For that we need to consider the end cube q of _i and its center z. Recall that x_0 is the maximum point of u on ∂(1.5 B) and r:=|z - x_0|. The assumptions needed for Lemma <ref> are the following: h/2 ≤ r, r ≪_n (log1/h )^-1 and 1/hlog1/h≪_n N . The first inequality holds because x_0 is on the end face of ℛ and z is the center of the end cube q, which is of side-length h. Moreover, the diameter of the end face of ℛ is ≍_n w. See figure <ref>. So r ≲_n w and for the second inequality in (<ref>) it is sufficeint to have w ≪_n ( log1/h )^-1. In view of w=c_0 / log K and h=K/w, it is equivalent to c_0/log K·logK log K/c_0≪_n 1. This is the place where we choose the dimensional constant c_0>0, keeping in mind that K≫_n 1. The third inequality in (<ref>) holds because we assume N ≫_n K(log K)^2 and 1/hlog1/h = K log K/c_0logK log K/c_0≲_n K (log K)^2 ≪_n N. Connecting the dots and choosing the parameters. We have defined w=c_0 / log K. The constant K≫_n 1 is not chosen yet and we assume N ≫_n K(log K)^2 . By the previous Claim <ref> every tunnel _i ⊂ℛ has a multiplicative increment Z ≳_n N. Lemma <ref> ensures that half of {_i}_i satisfy N^*(_i) ≤ N/ C_K. We can find a positive portion of the good tunnels {_i}_i, which are well-separated. That is, there is S ⊂{1,…, K^n-1} such that |S|≳_n K^n-1, N^*(_i) ≤ N/ C_K for each i ∈ S, and for distinct i,j ∈ S: (_i)_+50 w/K∩ (_j)_+50 w/K = ∅. Finally for every i ∈ S, we can use Lemma <ref> in _i to obtain a collection of disjoint balls {B_i,k}_k with center y_i,k and radius r_i,k≍ w/K satisfying the following properties. * The doubling index drops down: max_k𝒩 (2B_i,k)≤ N/C_K. * From condition (<ref>) of Lemma <ref>, we have the following lower bound on the sum of the doubling indices: N ≲_n ∑_k 𝒩( 12B_i,k). * From condition (<ref>) of Lemma <ref>, we have u(x_i,j) = 0, 4B_i,j∩ 4 B_i,k=∅, and 4B_i,k⊂ (_i)_+50 w/K⊂ 2B, for all i∈ S and all j≠ k. Finally, we will show that the collection ⋃_i∈ S{B_i,k} satisfies conditions (1), (2) and (3) of Lemma <ref>. Condition (3) of Lemma <ref>. almost follows from <ref>. We need also to show that for two different tunnels 𝒯_i and 𝒯_j with i,j ∈ S any balls B_i,k and B_j,l are disjoint. This is why we constructed the tunnels to be well-separated. Namely, (_i)_+50 w/K, i ∈ S are disjoint while the balls 4B_i,k and 4B_j,l must lie in 50 w/K neighbourhood of the corresponding tunnels. To get (1) of Lemma <ref> we will choose K depending on A and n. Namely, we choose K so that C_K = e^c'log K/loglog K > A Then condition <ref> translates to condition (1) in Lemma <ref>. We also assumed many times that K≫_n 1 which follows from the assumption A ≫_n 1, in Lemma <ref>. Finally we make our choice of K so that log A ≍_n log K/loglog K and C_K = e^c'log K/loglog K > A . Now the constants w and K are chosen, and we can specify the constant C(A)>1 in Lemma <ref>. The condition N≫ K (log K)^2 translates to N≫ C(A) withlog C(A) ≍log A loglog A. We now need to check (2) of Lemma <ref> and we do this below. Recall that for every i∈ S, each r_i,k≍ w/K with w≍ (log K)^-1. Hence from (<ref>) and the fact that |S| ≳_n K^n-1, we have ∑_i∈ S∑_k ( 12B_i,k )  r_i,k^n-1≳_n |S| w/K^n-1 N ≳_n w^n-1 N ≍_n N/(log K)^n-1. Since log A ≍_n log K/loglog K, we have log K ≲_n log A ·loglog A, for all large enough K. Hence we get the following lower bound from (<ref>): ∑_i∈ S∑_k (12B_i,k)  r_i,k^n-1≳_n N/(log A ·loglog A)^n-1. as required. This completes the proof of Lemma <ref>. § PROOF OF CLASSICAL RESULTS In this first part of this section, we prove several known results about the frequency function which was presented in <ref>. Recall that the frequency of a harmonic function u in ℝ^n is defined by β(x,r)= r G(x,r)/H(x,r), where x∈ℝ^n, r>0 and H(x,r)= ∫_∂ B(x,r) u^2 dS, G(x,r)= ∫_B(x,r) |∇ u|^2 d. The following well-known identity will be useful in proving that β(r):=β(0,r) is a non-decreasing function of r, see <cit.>. With the above notation, we have ∫_∂ (rB) |∇ u|^2 = n-2/r∫_rB |∇ u|^2d + 2∫_∂ (rB)(∇ u·n̂)^2 dS. where n̂ is then unit normal vector to the sphere ∂ (rB) and B=B(0,1). Writing 1= r̂·n̂ / r, where n̂= r̂r^-1 is the unit normal to the sphere ∂(rB), we have ∫_∂ (rB) |∇ u|^2= r^-1∫_∂ (rB) |∇ u|^2 r̂·n̂ dS= r^-1∫_rB( |∇ u|^2 ·r̂) d, where the second equality follows by the divergence theorem. Expanding the divergence operator and using the fact that (r̂)=n, we obtain ∫_rB( |∇ u|^2 ·r̂) d = n∫_rB |∇ u|^2 d + 2∫_rB (H (u) r̂)·∇ u d = n G(r) + 2∫_rB (H (u) r̂)·∇ u d, where H(u) is the Hessian of u. Using the identity ∇(∇ u ·r̂) = H(u) r̂ + ∇ u, the second term in (<ref>) can be re-written as ∫_rB (H (u) r̂)·∇ u d = ∫_rB∇ (∇ u·r̂) ·∇ u d - ∫_rB |∇ u|^2 d = ∫_∂ (rB) (∇ u ·r̂) (∇ u ·n̂) dS - ∫_rB |∇ u|^2 d. Therefore, inserting (<ref>) into (<ref>) and using the fact that n̂= r̂ r^-1, we conclude ∫_∂ (rB) |∇ u|^2=n-2/r G(r) + 2∫_∂ (rB)(∇ u·n̂)^2 dS, as required. thm: monotonicity frequency [Monotonicity of the frequency] For any harmonic function in ℝ^n, its frequency function β(r):=β(0,r) is a non-decreasing function of r. Taking the logarithmic derivative in the definition of the frequency function (<ref>) we have β'(r)= β(r)(1/r + G'(r)/G(r) - H'(r)/H(r)). We first consider the term H'/H on the r.h.s. of (<ref>). Observe that H(r)= ∫_∂ (rB) |u|^2 dS= r^n-1∫_∂ B |u(r·)|^2 dS. Thus, we have H'(r) = (n-1)r^n-2∫_∂ B |u(r·)|^2 dS + r^n-1∫_∂ B∂_r |u(r·)|^2 dS, = n-1/rH(r) + 2 ∫_∂ (rB) u ( ∇ u ·n̂ )dS, which gives H'(r)/H(r)= n-1/r + 2 ∫_∂ (rB) u ( ∇ u ·n̂ )dS/∫_∂ (rB) u^2dS. We now compute the term G'/G on the r.h.s. of (<ref>). Differentiating G(r), we find G'(r)=∫_∂ (rB) |∇ u|^2 dS. Therefore, Lemma <ref> gives G'(r)= n-2/rG(r) + ∫_∂ (rB)( ∇ u ·n̂ )^2dS. Since, for a non-constant harmonic function u, G(r)= ∫_B(x,r) |∇ u|^2 d = ∫_∂ (rB) u( ∇ u ·n̂ )dS>0, we deduce G'(r)/G(r)= n-2/r + 2 ∫_∂ (rB) ( ∇ u ·n̂ )^2dS/∫_∂ (rB) u( ∇ u ·n̂ )dS. Finally, inserting (<ref>) and (<ref>) into (<ref>), we obtain β'(r)= 2β(r)(∫_∂ (rB) ( u ·n̂ )^2dS/∫_∂ (rB) u( ∇ u ·n̂ )dS- ∫_∂ (rB) u ( ∇ u ·n̂ )dS/∫_∂ (rB) u^2dS), which is non-negative by the Cauchy-Schwartz inequality. cor: formulas for the growth For any harmonic function in ℝ^n, let β(r)= β(0,r) be its frequency function. For any 0<r_1<r_2, we have H(r_2)/r_2^n-1= H(r_1)/r_1^n-1exp( 2 ∫_r_1^r_2β(r)/rdr). Moreover, we also have (r_2/r_1)^2β(r_1) + n-1≤H(r_2)/H(r_1)≤(r_2/r_1)^2β(r_2) + n-1. Since we may write G(r)= ∫_∂ (rB) u (∇ u·n̂)dS, formula (<ref>) implies the following expression for the frequency function 2β(r)/r = 2G(r)/H(r)= H'(r)/H(r)- n-1/r Integrating both sides, we obtain the first claim of Corollary <ref>. Since, by Theorem <ref>, we have β(r_1)≤β(r)≤β(r_2) for all r≤ r_2, the second claim in Theorem <ref> follows from the first. Thanks to the Poisson formula for harmonic functions in ℝ^n, it is possible to bound the L^∞ norm of a harmonic function in a ball by its L^2 norm on the boundary of a slightly larger ball. Let δ>0 and let B⊂ℝ^n be a unit ball. There exists some constant C=C(n,δ)>1 such that sup_B|u|^2 ≤ C∫_∂((1+δ)B) |u|^2 dS, for any harmonic function u defined in a neighbourhood of (1+δ)B. lem: comparability di and frequency Let δ>0 be a sufficiently small parameter and let 𝒩(r):= 𝒩(B(0,r)), for r>0. There exists a constant C=C(δ,n)≥ 1 such that β (r(1+δ))(1-100δ) -C≤𝒩(r) ≤β (2r(1+δ))(1+100δ) +C uniformly for all 0<r≤ 4^-1 and all harmonic functions u defined in B(0,2). We first establish the lower bound in Lemma <ref>. Let us write B=B(0,1)⊂ℝ^n and observe that, by elliptic estimates, we have sup_rB|u|^2 ≲_δ1/r^n-1∫_∂ (r(1+δ)B) |u|^2 dS ≲_δH((1+δ)r)/r^n-1. We also have sup_2rB|u|^2 ≳1/r^n-1∫_∂ (2rB) |u|^2 dS≳H(2r)/r^n-1. Therefore, 2𝒩(rB) = logsup_2rB|u|^2/sup_rB|u|^2≥logH(2r)/H((1+δ)r) + C_0, for some C_0=C_0(n,δ)≥ 1. Using the second part of Corollary <ref> and taking δ>0 sufficiently small (so that log_2(2/(1+δ))≥ 1-100δ), we obtain logH(2r)/H((1+δ)r)≥ 2β (r(1+δ))(1-100δ) + C_1, for some C_1=C_1(n,δ)≥ 1, and the lower bound in Lemma <ref> follows. We are now going to establish the upper bound in Lemma <ref>. By elliptic estimates, we have sup_2rB|u|^2 ≲_δ1/r^n-1∫_∂ (2r(1+δ)B) |u|^2 dS ≲_δH(2(1+δ)r)/r^n-1, and we also have sup_rB|u|^2 ≥1/r^n-1∫_∂ (rB) |u|^2 dS= H(r)/r^n-1. Therefore, 2𝒩(rB) = logsup_2rB|u|^2/sup_rB|u|^2≤logH(2r(1+δ))/H(r) + C_2, for some C_2=C_2(n,δ)>0. By Corollary <ref> and taking δ>0 sufficiently small (so that log_2(2/(1+δ))≤ 1+100δ), we have logH(2r)/H((1+δ)r)≤ 2β (2r(1+δ))(1+100δ) + C_3, for some C_3=C_3(n,δ)>0, and the upper bound in Lemma <ref> follows. lem: monotonicity doubling index[Almost monotonicity of the doubling index] Let δ>0 be a sufficiently small parameter and let u be a harmonic function in B(0,2). Let 𝒩(r):= 𝒩(B(0,r)), for r>0. There exists a constant C=C(δ,n)>0 such that (r_2/r_1)^𝒩(r_1)(1-δ)-C≤sup_B(0,r_2)|u|/sup_B(0,r_1)|u|≤(r_2/r_1)^𝒩(r_2)(1+δ)+C, for all r_1≤ r_2/2≤ 4^-1. In particular, we have 𝒩(r_1)(1-δ) - C≤𝒩(r_2)(1+δ) + C. Let B=B(0,1)⊂ℝ^n. We will first prove the upper bound in Lemma <ref>. Let δ_1>0 be some, sufficiently small, parameter to be chosen in terms of δ>0. By elliptic estimates, we have sup_r_2B|u|^2≲_δ_1H(r_2(1+δ_1))/r_2^n-1. By Corollary <ref>, we have H(r_2(1+δ_1))/r_2^n-1≲_δ_1H(r_1)/r_1^n-1(r_2(1+δ_1)/r_1)^2β(r_2(1+δ_1)) +n-1. By Lemma <ref>, we have β(r_2(1+δ_1)) ≤𝒩(r_2)/(1+100δ_1) +C_1≤𝒩(r_2)(1+200δ_1) + C_1 , for some C_1=C(δ_1,n)≥ 1. Therefore, have shown that sup_r_2B|u|^2≲_δ_1H(r_1)/r_1^n-1(r_2(1+δ_1)/r_1)^2𝒩(r_2)(1+200δ_1) + C_1. Choosing δ_1=δ/1000 (so that (1+200δ_1)log(1+δ_1)≤δ/10, say), we deduce sup_r_2B|u|^2≤H(r_1)/r_1^n-1(r_2/r_1)^2𝒩(r_2)(1+δ) + C_2, where we have absorbed the constant in the ≲ notation into C_2. Hence, the upper bound in Lemma <ref> follows upon noticing that H(r_1)/r_1^n-1≤sup_r_1B|u|^2, and taking the square-root. We are now going to prove the lower bound in Lemma <ref>. First, we observe that we may assume r_2≥ 2(1+δ)r_1. Indeed, if 2r_1≤ r_2≤ 2(1+δ)r_1, then (r_2/r_1)^𝒩(r_1)(1-δ)≤ (2(1+δ))^𝒩(r_1)(1-δ)≤ 2^𝒩(r_1), for all δ>0 sufficiently small so that (2(1+δ))^1-δ≤ 2. Therefore, in this case, we have sup_r_2B|u|≥sup_2r_1B|u|= 2^𝒩(r_1)sup_r_1B|u|≥(r_2/r_1)^𝒩(r_1)(1-δ)sup_r_1B|u|. From now we will assume that r_2≥ 2(1+δ)r_1. Let δ_1 ∈ (0,δ/2) be a parameter to be chosen later in terms of δ>0. By Corollary <ref>, we have sup_r_2B|u|^2≥H(r_2)/r_2^n-1≥H(2r_1(1+δ_1))/(2r_1(1+δ_1))^n-1(r_2/2r_1(1+δ_1))^2β(2r_1(1+δ_1)), and, by Lemma <ref>, we also have β(2r_1(1+δ_1))≥𝒩(r_1)/1+100δ_1 - C_3, for some C_3=C_3(δ_1,n)≥ 1. Moreover, by elliptic estimates, we have H(2r_1(1+δ_1))/(2r_1(1+δ_1))^n-1≳_δ_1sup_2r_1B|u|^2≳_δ_1 2^2𝒩(r_1)sup_r_1B|u|^2. All in all, we have shown that sup_r_2B|u|^2 ≳_δ_1sup_r_1B |u|^2 (r_2/2r_1(1+δ_1))^2𝒩(r_1)(1+100δ_1)^-1 -C_32^2𝒩(r_1) . Taking the square root and choosing δ_1=δ/10000>0 (so that (1+100δ_1)^-1≥ 1-δ/10), we conclude that sup_r_2B|u| ≥sup_r_1B |u| (r_2/r_1)^𝒩(r_1)(1-δ) -C_4, where we have absorbed the constant implied in the ≲ notation into C_4, as required. v First, we observe that, by the monotonicity of the doubling index Lemma <ref> and the definition of 𝒩^*(·), it is enough to prove that 𝒩(x,r)≤ C_1 𝒩(0,2) + C_2, for all x∈ B(0,1.6) and r=10^-1 (say). In order to see (<ref>), we will first show that there exists constants C_3=C_3(n)≥ 1 and C_4=C_4(n)≥ 1 such that 𝒩(x,1)≤ C_3𝒩(0,2) + C_4, for all |x|≤ 99/100. Indeed, given the inclusion of balls B(x,2) ⊂ B(0,4) and B(0,10^-4)⊂ B(x,1), the definition of 𝒩(·) and Lemma <ref> (with δ=10^-10, say) give sup_B(x,2)|u|≤sup_B(0,4) |u|= 2^𝒩(0,2)sup_B(0,2)|u| ≤exp (C_3 𝒩(0,2) + C_4)sup_B(0,10^-4)|u|≤exp (C_3 𝒩(0,2) + C_4) sup_B(x,1)|u|, and (<ref>) follows. Now, given x∈ B with |x|≤ 99/100, we may apply (rescaled) (<ref>) to u(· -x) to see that 𝒩(y,1/2)≤ C_3𝒩(x,1) + C_4,≤ C_5𝒩(0,2) + C_6, for some C_5=C_5(n)≥ 1 and C_6=C_6(n)≥ 1, for all |y-x|≤ 2^-1(99/100)^2 and all |x|≤ 99/100. Repeating (<ref>) a third time, we conclude that 𝒩(x,1/4)≤ C_7𝒩(0,2) + C_8, for some C_7,C_8=C_7(n),C_8(n)≥ 1, for all x∈ 2B with |x|≤ 99/100+ 1/2(99/100)^2 + 1/4(99/100)^3 =1.7226..., as required. Partition Q' into (A')^n many equal subcubes {q'_j: 1≤ j ≤ (A')^n} with side length 100√(n)ℓ A'^-1 each, where A' ≍ A. Define and ' to be the collections := {100√(n)q_α}_α and ' := {q'_j}_j. Proposition <ref> implies that ♯{1≤ j ≤ (A')^n-1: ^*(q'_j) ≥max(N· 2^-clog A/loglog A,C)}≤ (A')^n-1-c_0≲ A^n-1-c_0. It is easy to see that every 100√(n)q_α∈ can be covered by cubes from '; and for every q'_j ∈', there are at most L = L(n) many cubes in which have non-trivial intersection with q'_j. Also the side length of every cube 100√(n)q_α∈𝒞 is smaller than the side length of every cube q'_j∈𝒞'. Because of this, we can conclude ♯{α : ^*(100√(n)q_α) ≥max(N· 2^-clog A/loglog A,C)} ≲_L♯{j: ^*(q'_j) ≥max(N· 2^-clog A/loglog A,C)}≲ A^n-1-c_0. as claimed, and this finishes the proof of Corollary <ref>. [Gradient estimates for harmonic functions] Let B(p,r)⊂^n be any ball, and let u be a harmonic function defined in a neighbourhood of its closure. Then | ∇ u (p)| ≲_n 1/rsup_B(p,r)|u|. Since the partial derivatives of u are also harmonic functions, the mean value theorem and the divergence theorem imply that ∇ u (p) = 1/ B(p, r)∫_B(p,r)∇ u d = 1/ B(p, r)∫_∂ B(p,r) u(x)n_x dS(x), where n_x is the exterior normal vector to ∂ B(p,r) at x ∈∂ B(p,r); and dS denotes the surface measure on ∂ B(p,r). Hence | ∇ u (p)| ≲_n 1/rsup_B(p,r)|u|, as claimed. Let B⊂ℝ^n be the unit ball, and let u be a non-zero harmonic function on 2B such that u(0)=0. Then, there exists some constant c=c(n)>0, independent of u, such that 𝒩(B)≥ c. Let m_1^+:=sup_B u, m_2^+:= sup_2B u and m_1^-:=-inf_B u, m_2^-:= -inf_2B u and observe that, since u(0)=0, the Mean Value Theorem implies m_1^+,m_2^+,m_1^-,m_2^->0. Now, assume that m_1^+=sup_B|u| and consider the harmonic function h:= m_2^+-u. Then h≥ 0 in 2B and Harnack's inequality implies sup_B h≤ C inf_B h. Let x∈ B be the point such that u(x)= m_1^+=sup_B|u|. We obtain 1/Ch(0)≤ h(x) ≤ C h(0), that is C^-1 m_2^+≤ m_2^+- m_1^+≤ C m_2^+, which, upon rearranging, gives m_1^+≤ m_2^+ (1- 1/C). All in all, we have shown that (1- 1/C)^-1≤m_2^+/m_1^+≤sup_2B|u|/sup_B|u|, as required. If -m_1^-=sup_B|u|, we apply the above argument to the harmonic function -u. This concludes the proof of Claim <ref>. Let B=B(0,1) be the unit ball and let u be a non-constant harmonic function in 4B. There exists a numerical constant C=C(n)>1, independent of u, such that C^-1𝒩(1/2B)≤𝒩(2B). By Lemma <ref>, we have 1/2𝒩(1/2B)-C_1 ≤ 2𝒩(2B), for some constant C_1= C_1>1. Thus, Claim <ref> is true in the case 𝒩(1/2B)≥ C_2 for some constant C_2=C_2(n)>0. Let us now suppose that 𝒩(1/2B)≤ C_2. By multiplying u by a constant, we may assume that sup_1/2B|u|=1 sup_B|u|=sup_B u =: M. Let us also write p∈∂ B for the point such that u(p)=M and let M'= sup_2B|u|. Finally, let us write D= B(p,1), D_1= D∩1/2B, D_2= D∩ (B\1/2B) and D_3= D∩ (2B\ B). By the mean value theorem for harmonic functions, we have M=u(p)= 1/(D)∫_D u d≤1/(D)((D_1) + M (D_2) + M'(D_3) ). Since (D) = (D_1) + (D_2)+ (D_3), we obtain (M-1) (D_1)≤ (M'-M)(D_3), that is M +(M-1)(D_1)/(D_3)≤ M'. Letting c= (D_1)/(D_3), we have 1 + cM-1/M≤M'/M. To conclude the proof of Claim <ref>, let us first consider the case, M= 1+ε for some small ε≤ε_0. Then, we have 𝒩(1/2B)=log (1+ε)≍ε, and 𝒩(2B)≥log(1+cε/1+ε)≍ cε. Thus, Claim <ref> follows. For M≥ 1+ε_0, (<ref>) implies that there exists some constant c_1=c_1(n,ε_0)>0 such that 𝒩(2B)≥ c_1, which implies Claim <ref>, in light of the fact that 𝒩(1/2B)≤ C_2, as required. Let Q⊂ℝ^n be a unit cube. There exists some Γ=Γ(n)≥ 1 such that for any harmonic function u in Γ Q, we have 𝒩^*(Q)≲_nlog∫_Γ Q|u|^2 d/∫_Q|u|^2 d . Up to a translation, we may assume that Q is centered at zero. By the monotonicity of the doubling index for non-concentric balls, Lemma <ref>, it is enough to prove that 𝒩(4√(n)B)≲_n log∫_Γ Q|u|^2 d/∫_Q|u|^2 d , where B=B(0,1). To see this, the elliptic estimates imply sup_8√(n)B|u| ≲1/Γ^n∫_Γ Q|u|^2 d, for all sufficiently large Γ=Γ(n). On the other hand, we have sup_4√(n)B|u|≳_n ∫_Q|u|^2 d , and Claim <ref> follows. Most of the remaining section is devoted to proving the following lemma about almost monotonicity of the doubling index for non-concentric balls. lem: almost monotonicity [Almost monotonicity for non-concentric balls] Let B⊂ℝ^n be any ball and let u be a harmonic function in a neighbourhood of 2B. There exists a constant C=C(n)>1 such that 𝒩(b) ≤ C 𝒩(B) for all balls b with 2b⊂ B. We will prove Lemma <ref> in several steps. The main step is to prove the following special case of the lemma when b is concentric to B. Let B=B(0,1) ⊂^n be the unit ball and let u be a harmonic function in a neighbourhood of 2B. There exists a constant C=C(n)>1 such that the following holds for every r ∈ (0,]: 𝒩(rB)≤ C𝒩(B). To establish Claim <ref>, we need several auxiliary results and we prove them now. The following claim proves Claim <ref> in two cases: when either (B) or r is bounded away from 0. Let B=B(0,1) ⊂^n be the unit ball and let u be a harmonic function in a neighbourhood of 2B. Then the following hold. * If (B) ≥ c >0, then 𝒩(rB) ≲_n,c𝒩(B),  ∀ r ∈ (0,]. * For every t ∈ (0,), there is a dimensional constant C_t >0 such that (rB) ≤ C_t·(B),  ∀ r ∈ [t, ]. We first show the first part of the claim. Recall from Lemma <ref> that 𝒩(rB) ≤ 2𝒩(B) + C, ∀ r ∈ (0,] for some constant C= C(n)>1. If (B) ≥ c >0, then it follows from (<ref>) that 𝒩(rB) ≤ (2 + Cc) ·𝒩(B), ∀ r ∈ (0,], and we get the desired conclusion. 2mm We now prove the second part of the claim. If (B) ≥ C, then (<ref>) implies (rB) ≤ 3 (B), ∀ r ∈ (0,]. Hence it suffices to prove (<ref>) in the case when ( B) <C, and we do this below. We fix t ∈ (0,), and let r ∈ [t, ]. Again it follows from (<ref>) that if ( B) <C, then (rB) ≤ 3C. Multiplying u by a constant, we may assume that sup_rB|u|=1,   and  sup_B|u|=sup_B u =: M. Let p∈∂ B be a point such that u(p)=M, and put M'= sup_2B|u|. With the above notation, since 2r≤ 1, one has (rB) ≤log M ,  and  (B) = logM'/M. Finally, define D:= B(p,1),  D_1:= D∩ rB,  D_2:= D∩ (B\ rB), and D_3:= D∩ (2B\ B). By the mean value theorem for harmonic functions, we have M=u(p)= 1/ (D)∫_D u d≤1/(D)((D_1) + M (D_2) + M'(D_3) ). Since (D) = (D_1) + (D_2)+ (D_3), we obtain (M-1) (D_1)≤ (M'-M)(D_3), that is M +(M-1)(D_1)/(D_3)≤ M'. Let c= c(t)= (D ∩ t B)/(D_3), since (D ∩ t B) ⊆ D_1, we have 1 + cM-1/M≤M'/M. To conclude (<ref>), we consider two cases: the first being when M is very close to 1, and when M is not very close to 1. In the first case M= 1+ε, for some small ε≤ε_0. Then 𝒩(rB) ≤log M = log (1+ε)≍ε. And from (<ref>) we get 𝒩(B)≥log(1+cε/1+ε)≍ cε. Thus (<ref>) follows. Let us consider the second case: M≥ 1+ε_0. Then (<ref>) implies that there exists some constant c_1=c_1(n,ε_0)>0 such that 𝒩(B)≥ c_1, which implies (<ref>), in light of the fact that 𝒩(rB)≤ 3C. This completes the proof of Claim <ref>. The following claim gives a comparison of the doubling indices in two balls when their radii are much larger than the distance between their centers. Let B=B(0,1) ⊂^n be the unit ball, and let u be a harmonic function in a neighbourhood of 2B. Let r >0, and p,q ∈ B such that |p-q|< r. Also assume that 8r <1, then (p,2r) ≲_n (q,4r). Observe that B(q,r) ⊂ B(p,2r) ⊂ B(p,4r) ⊂ B(q,8r). Thus one has sup_B(p,4r)|u|/sup_B(p,2r)|u|≤sup_B(q,8r)|u|/sup_B(q,r)|u| = sup_B(q,8r)|u|/sup_B(q,4r)|u|·sup_B(q,4r)|u|/sup_B(q,2r)|u|·sup_B(q,2r)|u|/sup_B(q,r)|u|, from which one gets (p,2r) ≤(q,4r) + (q,2r) + (q,r). It follows from part two of Claim <ref> that both (q,r), (q,2r) ≲_n (q,4r). So we get the desired conclusion (p,2r) ≲_n (q,4r). With Claim <ref> and Claim <ref> in hand, we conclude in the following claim that if the doubling index of a harmonic function in a ball is small enough, then it is non-vanishing in a smaller concentric ball. Let B=B(0,1) ⊂^n be the unit ball, and let u be a harmonic function in a neighbourhood of 2B. If (14B) ≪_n 1, then u is non-vanishing in 116B. Assume that there is a point p ∈116B where u vanishes. Then it follows from Claim <ref> that (p,18) ≥ c, where c>0 is as in Claim <ref>. Using Claim <ref> for u with q=0, p as above, and r= 1/16, one can conclude that c ≤(p,18) ≲_n (0,14)=(14B). Hence if u vanishes somewhere in 116B, then (14B) ≳_n 1. This proves Claim <ref>. We now present one final auxiliary result before proving Claim <ref>. The following claim will be helpful for comparing doubling indices in two sub-balls when the harmonic function has very small doubling index in a ball. Let B=B(0,1) ⊂^n be the unit ball, and let u be a positive harmonic function in a neighbourhood of 2B with u(0)=1. There exist small dimensional constants ℓ, c_0 >0 such that if (B) ≤ c_0, then (B) ≳_n sup_(ℓ/2) B |u-1|. We start by choosing ℓ. Claim <ref> and Claim <ref> guarantee the existence of constants C>1 and c>0 respectively such that the following holds. For any non-constant harmonic function f in a neighbourhood of 2B vanishing at its center, one has _f(B) ≥log C,   and  sup_2Bf ≥ c sup_B|f|. Fix a large k ∈ such that c· C^k≥ 4. For this k, let ℓ∈ (0,1) be such that 18≤ 2^k ℓ≤14. The choice of c_0 ∈ (0,1) will be made later. For u as in the statement of the claim, define v:= u-u(0) = u-1. Then v is harmonic in 2B and vanishes at its center. Using the first conclusion of (<ref>) repeatedly for v, one gets sup_2^kℓ B |v| ≥ C^ksup_ℓ B |v|. This together with the second conclusion of (<ref>) gives sup_2^k+1ℓ Bv ≥ c ·sup_2^kℓ B |v| ≥ c· C^ksup_ℓ B |v| ≥ c· C^ksup_ℓ B v. By the choice of k, we have c · C^k≥ 4, and hence sup_2^k+1ℓ Bv ≥ 4 sup_ℓ B v. Hence sup_2^k+1ℓ Bu ≥ 1+ 4 sup_ℓ B v. We now get a lower bound for the doubling index of u in B. We have 1+ 4 sup_ℓ B v /1 + sup_ℓ B v≤_(<ref>)sup_2^k+1ℓ B u /sup_ℓ B u = sup_2^k+1ℓ B u /sup_2^kℓ B u·sup_2^kℓ B u /sup_2^k-1ℓ B u⋯sup_2ℓ B u /sup_ℓ B u. Thus log1+ 4 sup_ℓ B v /1 + sup_ℓ B v≤∑_j=0^k(2^j ℓ B) ≲_n(B), where the last inequality holds because each term in the sum is ≲(B) by the second part of Claim <ref>. Since u is positive in 2B and u(0) = 1, by Harnack's inequality one has u ≍_n 1 in B. Hence there is a dimensional constant c_1 >0 such that 1+ 4 sup_ℓ B v /1 + sup_ℓ B v = 1+ 3sup_ℓ B v /sup_ℓ B u≥ 1 + c_1 sup_ℓ B v. Thus from (<ref>) and (<ref>), one has log (1 + c_1 sup_ℓ B v) ≲_n (B). Note that if (B) ≪ 1, then 1≫(B) ≳_n log (1 + c_1 sup_ℓ B v) ≍sup_ℓ B v ≥_(<ref>) c sup_(ℓ/2) B|v|. Hence if c_0 >0 is small enough, then (B) ≤ c_0 implies (B) ≳_nsup_(ℓ/2) B |u-1|. We are finally ready to prove Claim <ref>. Recall that in Claim <ref>, it was shown that (rB) ≲(B) holds whenever either r or (B) is separated from 0. Hence it suffices to prove that there exist dimensional constants r_0 ∈ (0,1) and C_0 >1 such that (rB) ≤ C_0 ·(B), ∀ r ∈ (0,r_0). Let c_0,ℓ >0 be as in Claim <ref>. Let c_1 >0 be a small dimensional constant, to be chosen later. We now prove (<ref>) for r_0 := ℓ/2^10. If (14B) > c_1, then it follows from part two of Claim <ref> that (B) ≳(14B) ≥ c_1, and from part one of Claim <ref> we get that (rB) ≲_n(B), ∀ r∈ (0,12]. Hence it suffices to consider the case when (14B)≤ c_1. Claim <ref> guarantees that if c_1 is small enough, then u is non-vanishing in 116B. We will choose c_1 satisfying this and an additional condition to be specified soon. 2mm Let >0 and 2r∈ (0,r_0) be such that (rB) =. We will now show that with the above assumption of u being non-vanishing in 116B, one has (116B) ≳_n/r and this will prove (<ref>). Let M = sup_rB u, then sup_2rB u = 2^M. Hence there is a point p ∈ 2rB ⊂ℓ2^10B such that |∇ u(p)| ≳M(2^-1)/r≳M/r. Multiplying u by a constant, we may assume that u(p) = 1, in which case by Harnack's inequality M ≍_n 1. Hence we have the following bound for the gradient at p: |∇ u(p)| ≳/r. From the standard gradient estimates (Claim <ref>), we get sup_B(p,ℓ/2^6) |u-1| ≳ |∇ (u-1)(p)| = |∇ u(p)| ≳/r. Recalling that ℓ <1, we have the following comparison for the doubling indices from Claim <ref> (applied twice): (p,132) ≲(0,14)=(14B) ≤ c_1. We will choose c_1 small enough so that the above inequality implies (p,132) ≤ c_0. This together with (<ref>) implies that we can use Claim <ref> to conclude that (p,132) ≳r. Once again using Claim <ref>, one can conclude that (116B) = (0,116) ≳(p,132) ≳/r≥ = (rB). This completes the proof of Claim <ref>. We prove this claim in several steps. 2mm Step 1: We first show that if (B) is bounded below by a positive constant, then the conclusion of the claim holds. Recall from Lemma <ref> that 𝒩(rB) ≤ 2𝒩(B) + C_1, ∀ r ∈ (0,] for some constant C_1= C_1(n)>1. Suppose (B) ≥ c_1 >0, then it follows from (<ref>) that 𝒩(rB) ≤ (2 + C_1c_1) ·𝒩(B), ∀ r ∈ (0,] and hence the claim follows. 2mm Step 2: We next show that the claim is true when r is away from 0. We will now show that for every ℓ∈ (0,), there is C_ℓ >0 such that (rB) ≤ C_ℓ·(B), ∀ r ∈ [ℓ, ]. Observe that if (B) ≥ C_1, then (<ref>) implies (rB) ≤ 3 (B), ∀ r ∈ (0,]. This shows that it suffices to prove (<ref>) in the case when ( B) <C_1, and we do this below. Again it follows from (<ref>) that if ( B) <C_1, then (rB) ≤ 3C_1. We fix ℓ∈ (0,), and let r ∈ [ℓ, ]. Multiplying u by a constant, we may assume that sup_rB|u|=1,   and  sup_B|u|=sup_B u =: M. Let p∈∂ B be a point such that u(p)=M, and put M'= sup_2B|u|. With the above notation, since 2r≤ 1, one has (rB) ≤log M ,  and  (B) = logM'/M. Finally, define D:= B(p,1),  D_1:= D∩ rB,  D_2:= D∩ (B\ rB), and D_3:= D∩ (2B\ B). By the mean value theorem for harmonic functions, we have M=u(p)= 1/(D)∫_D u d≤1/(D)((D_1) + M (D_2) + M'(D_3) ). Since (D) = (D_1) + (D_2)+ (D_3), we obtain (M-1) (D_1)≤ (M'-M)(D_3), that is M +(M-1)(D_1)/(D_3)≤ M'. Let c= c(ℓ)= (D ∩ℓ B)/(D_3), since (D ∩ℓ B) ⊆ D_1, we have 1 + cM-1/M≤M'/M. To conclude (<ref>), we consider two cases: the first being when M is very close to 1, and the other is when M is not very close to 1. In the first case M= 1+ε, for some small ε≤ε_0. Then 𝒩(rB) ≤log M = log (1+ε)≍ε. And from (<ref>) we get 𝒩(B)≥log(1+cε/1+ε)≍ cε. Thus (<ref>) follows. Let us consider the second case: M≥ 1+ε_0. Then (<ref>) implies that there exists some constant c_1=c_1(n,ε_0)>0 such that 𝒩(B)≥ c_1, which implies (<ref>), in light of the fact that 𝒩(rB)≤ 3C_1. This completes Step 2. 2mm Step 3: We now show that there is r_0 >0 small enough and C_0 >1 such that (rB) ≤ C_0 ·(B), ∀ r ∈ (0,r_0). In this step, we consider two cases: the first being when 172B contains a zero of u, and the other when u is non-vanishing in 172B. Let us consider the first case now. Let q ∈172B be such that u(q)=0. Then 14 B ⊂ B(q,2072) ⊂ B(q,3072) ⊂ B. From this and Claim <ref>, it follows that (14B) ≥ c, for some c=c(n)>0. This together with the conclusion of Step 2 implies that (B) ≳_n 1, and hence the claim follows from Step 1. We now consider the case when u is non-vanishing on 172B. In light of Step 2, it suffices to prove that (rB) ≤(14· 72B), ∀ r ∈ (0,116· 72). Because of this, we may assume w.l.o.g. that u>0 in 2B, u(0)=1, and show that (rB) ≤( B), ∀ r ∈ (0,18). Consider the harmonic function v := u - u(0). Then from the standard gradient estimates (Claim <ref>), one has The following claim gives a comparison between the doubling indices of a sub-ball and a ball, when the center of the sub-ball is close to the center of the ball. Let B=B(0,1) ⊂^n be the unit ball and let u be a harmonic function in 2B. Then there is a constant C=C(n)>1 such that for every z ∈12B, the following holds: (z,1-|z|2) ≤ C (B). Also, for any ball b which is centered at z and satisfies 2b ⊂ B, we have (b) ≤ C (B). We first observe that D = B(z,1-|z|2) is the largest ball centered at z such that 2D ⊂ 2B. Using |z|<12, one has B(0,14) ⊆ B(z,1-|z|2) ⊆ B(z,2-|z|) ⊆ B(0,2). Hence sup_B(z,2-|z|)|u|/sup_B (z,1-(|z|/2)) |u|≤sup_B(0,2)|u|/sup_B(0,14)|u| = sup_B(0,2)|u|/sup_B(0,1)|u|·sup_B(0,1)|u|/sup_B(0,12)|u|·sup_B(0,12)|u|/sup_B(0,14)|u|. Therefore (z,1-|z|2) ≤(0,1) + (0,12) + (0,14) ≲(0,1), where the last inequality follows from Claim <ref>, and this proves the first part (<ref>) of the claim. Let b=B(z,ℓ). Since 2b ⊂ B, one has 2ℓ≤ 1-|z| ≤ 1-|z|2, and hence 2b ⊂ D. Now using Claim <ref> and (<ref>), one can conclude (z,ℓ) ≲(z,1-|z|2) ≲(0,1), and this proves the last part (<ref>) of the claim. OLD VERSION OF THE ABOVE CLAIM Let B=B(y,R) ⊂^n be a ball and let u be a harmonic function in 2B. Then there is a constant C=C(n)>1 such that for every z ∈12B, the following holds: (z,R-|z-y|2) ≤ C (B). Also, for any ball b = B(z,ℓ) with 2b ⊂ B, we have 2b ⊂ B(z, R - |z-y|2) and the following holds: (b) ≤ C (B). Because of the scaling and translation invariance of the doubling index, we may assume w.l.o.g. that B=B(0,1) and |z| ≤. In this setting, (<ref>) translates to (z,1-|z|2) ≲(B). We first observe that D = B(z,1-|z|2) is the largest ball centered at z such that 2D ⊂ 2B. One has B(0,14) ⊆ B(z,1-|z|2) ⊆ B(z,2-|z|) ⊆ B(0,2). Hence sup_B(z,2-|z|)|u|/sup_B (z,1-(|z|/2)) |u|≤sup_B(0,2)|u|/sup_B(0,14)|u| = sup_B(0,2)|u|/sup_B(0,1)|u|·sup_B(0,1)|u|/sup_B(0,12)|u|·sup_B(0,12)|u|/sup_B(0,14)|u|. Therefore (z,1-|z|2) ≤(0,1) + (0,12) + (0,14) ≲(0,1), where the last inequality follows from Claim <ref>, and this proves the first part (<ref>) of the claim. Since 2b ⊂ B, one has ℓ≤1- |z|2. Since 2ℓ≤ 1-|z| ≤ 1-|z|2, (<ref>) follows. Now using Claim <ref> and (<ref>), one can conclude (z,ℓ) ≲(z,1-|z|2) ≲(0,1), and this proves the last part (<ref>) of the claim. We are finally ready to prove Lemma <ref>. Once again, we may assume w.l.o.g. that B=B(0,1). Let b = B(x,r) be such that 2b ⊂ B. If |x| ≤, then (<ref>) in Claim <ref> immediately implies the conclusion of the lemma. We now prove the lemma when |x| ∈ (,1). Let x = |x| θ, where θ∈∂ B(0,1). Let C be as in Claim <ref>. By (<ref>) of Claim <ref> one has (12θ, 34) ≤ C (B). and 𝒩(78θ, 916)≤ C(12θ, 34) ≤ C^2 (B). Recall that 2b⊂ B(x,1-|x|). We observe that, if |x|∈ [12,78], then 2b ⊂ B(x,1-|x|)⊂ B(12θ, 12)⊂ B(12θ, 34), and x∈12B(12θ, 34). Using (<ref>) of Claim <ref> for B(12θ, 34) gives 𝒩(b)≤ C (12θ, 34) ≤ C^2 (B). If x∈ [78,1], as illustrated in Figure <ref>, we have 2b ⊂ B(x,1-|x|)⊂ B(78θ, 18)⊂ B(78θ, 916), and x∈12B(78θ, 916). Using (<ref>) of Claim <ref> for the ball B(78θ, 916) gives 𝒩(b)≤ C (78θ, 916)≤ C^3 𝒩(B), as required. We end this section by showing how Proposition <ref> implies the following modified version of it. cor_of_prop4.1 [A modified version of Proposition <ref>] Let a unit cube Q ⊂^n be partitioned into K^n equal subcubes {q_α: 1≤α≤ K^n}. Let u be a harmonic function in 4√(n) Q. Let ^*(2Q) = N. If K ≫_n 1, then we have #{α :^*(100√(n)q_α) ≥max(N· 2^-clog K/loglog K,C)}≤ K^n-1-c, where c,C >0 are dimensional constants. Partition 2Q into L^n many equal subcubes {s_j : 1≤ j ≤ L^n}, where L := ⌊K/100√(n)⌋. Hence the side length of each s_j is at least 200√(n)/K. Let c_0 and C_0 denote the constants given by Proposition <ref>. Using Proposition <ref> for the cube 2Q with the above partition, one can conclude that if L ≫_n 1 (which is equivalent to K ≫_n 1, since K ≍_n L), then #{1 ≤ j ≤ L^n: ^*(s_j) > max(Nexp(-c_0 log L/loglog L), C_0)}≤ L^n-1-c_0. For cubes 100√(n)q_α and s_j, one has ^*(100√(n)q_α) = sup{(x,r): x∈ 100√(n)q_α and r≤ 100n/K}, ^*(s_j) ≥sup{(x,r): x∈ s_j and r≤ 200n/K}. It is easy to see that every cube 100√(n)q_α can be covered by cubes from {s_j}_j; and there is M ∈ such that for every s_j, there are at most M many cubes in {100√(n)q_α}_α which have non-trivial intersection with s_j. From the above discussion, it follows that ^*(100√(n)q_α) can be large only when 100√(n)q_α non-trivially intersects s_j for which ^*(s_j) is large. Hence we have #{α: ^*(100√(n) q_α) ≥max(Nexp(-c_0 log L/loglog L), C_0)}, ≤ M ·#{j : ^*(s_j) ≥max(Nexp(-c_0 log L/loglog L), C_0)}. The desired conclusion now follows from (<ref>) since K ≍_n L. § FEW WORDS ABOUT THE PROOF OF PROPOSITION <REF> Maximal Doubling index for cubes. For a given cube Q in ℝ^n, we denote by 2Q the cube, which is the homothetic copy of Q with same center and with homothety coefficient 2. In some of the multiscale arguments it is convenient to work with a maximal version N^*_u(Q) of the doubling index: N^*_u(Q) = sup_x ∈ Q, r ≤(Q) logmax_B(x, 2 r)|u|/max_B(x,r)|u| because N^*_u(Q) is a monotone function of a set. §.§ Informal guide to the proof of Proposition <ref> In the proof of the lower effective bound in Nadriashvili's conjecture we used once Proposition <ref> as a black box. Let us remind the formulation of the Proposition. prop: lemma 1. Let a cube Q in ℝ^n be partitioned into A^n equal subcubes Q_i. Let u be a harmonic function in 2Q. There exist numerical (depending only on the dimension n) constants A_0,C > 1 and c>0 such that if A>A_0, then the number of Q_i with N^*_u(Q_i)>max(N^*_u(Q)/exp(- c log A/ loglog A ), C) is smaller than A^n-1-c The goal of this section is to provide an informal guide for those who would like to understand the proof of the Proposition <ref>. In order to understand the full details the reader may read pages 221-231 in <cit.> and pages 249-254 in <cit.>, but there are only two facts about harmonic functions, which are used in the proof. The first fact is the monotonicity property of the frequency, which is a powerful tool. The second fact is quantitative Cauchy uniqueness property. The monotonicity property of the frequency and some considerations from Euclidean geometry imply the following quasi-additive property of the doubling index. There exist a small number c>0 and and a large number C>1, depending on the dimension n, such that the following holds for any harmonic function u in ℝ^n. Suppose that n+1 points x_1,x_2, ..., x_n+1 in ℝ^n form an equilateral simplex S with side length r. Assume that max_1≤ i ≤ n+1 N_u(B(x_i,ρ)) ≥ A, where A is any positive number and ρ≤ r/2. Let x be the center of S. Then N_u(B(x,C r)) ≥ A(1+c) - C. In the statement above one can replace n+1 balls with centers at the vertices of the equilateral simplex by any other n+1 balls, whose centers form a non-degenerate simplex and the radii are bounded by the diameter of the simplex. But the constants c>0 and C>1 depend on how degenerate the simplex is. We refer to the formal statement of simplex lemma in <cit.>, where the constants depend on the ratio of the width of the simplex and its diameter. The second fact about harmonic functions, which is used in the proof of Proposition <ref>, is quantitative Cauchy uniqueness property. Consider a half-ball B_+ = {(x,y)∈ℝ^n-1×ℝ: |x|^2+y^2< 1, y> 0 } . There exist α∈(0,1) and C>0 such that if h is harmonic in B_+, h∈ C^1(B_+) and satisfies the inequalities |h|≤ 1, |∇ h|≤ 1 in B_+ and |h|≤ε, |∂_y h|≤ε on {(x,y)∈B_+, y=0}, ε≤ 1, then |h(x)|≤ Cε^α when x∈1/3B_+. We refer to Appendix of <cit.> for elementary proof. Quantitative Cauchy uniqueness property is used to prove another quasi-additive property of the frequency. There exists a large integer K>100 and C>1 such that the following holds. Let u be a harmonic function in ℝ^3. Consider a finite lattice of points L_K:={ (i,j,0): i=- K, …, K, j=-K, …, K }. If N(B(x,1)) ≥ A for each x∈ L_K and some number A>C, then N(B(0,K))≥ 2A. In the statement of the hyperplane lemma one can replace the finite lattice of balls B(x,1), x ∈ L_K by any collection of balls B(p_x,r),x ∈ L_K, with the centers p_x ∈ B(x,1) and any radius r≤ 1, and the statement will remain true. In the n-dimensional version of hyperplane lemma one should take (n-1)-dimensional lattice of balls. Hyperplane lemma and simplex lemma can be combined together to get the following useful fact about distribution of doubling indices. Let u be a harmonic function in ℝ^n. There exist numerical (depending only on the dimension n) constants A_0 ∈ℕ, C > 1 and c_1>0 such that if we partition a cube Q in ℝ^n into A_0^n equal subcubes Q_i, then the number of Q_i with N^*_u(Q_i)>max(N^*_u(Q)/(1+c_1), C) is smaller than 1/2 A_0^n-1. In the proof Theorem 5.1 from <cit.> the following intuitive principle is used. The simplex lemma implies that “bad" Q_i with N^*_u(Q_i) > N^*_u(Q)/(1+c_1) should be close to some (n-1)-dimensional hyperplane, and then one can find a finite “lattice" of “bad" cubes and arrive to contradiction with the hyperplane lemma. We refer to <cit.> for details. Proposition <ref> follows from Theorem 5.1 from <cit.> by iterative subdivision of Q into smaller and smaller cubes. One can think that A=A_0^k, and we subsequently chop cubes into 1/A_0 smaller subcubes and each time we apply Theorem 5.1 to get more information on the distribution of doubling indices. The formal iterative argument is given on pages 249-254 in <cit.>. § DONNELLY AND FEFFERMAN'S BOUND FOR THE DOUBLING INDEX The aim of this section is to show how Proposition <ref> follows from the work of Donnelly and Fefferman on complex extensions of real-analytic functions, more precisely from <cit.>. We now recall the statement of Proposition <ref>: lem: half of the doubling index Let Q be a unit cube in ^n and let u be a harmonic function in 4√(n)· Q. Let ε>0 and let Q be partitioned into equal subcubes {Q_i}_i of side length ≍_nc_u^*(Q) for some sufficiently small c=c(n,ε)>0. There exists a constant C = C(n,ε) > 1 such that ^*(Q_i)≤ C for at least (1-ε) portion of the Q_i. Proposition <ref> follows with the help of the following fact about holomorphic function <cit.>: Let k be a sufficiently large integer depending on n. Let G(z) be holomorphic in |z|≤ 3^k, z∈ℂ^n, and satisfying max_|z|≤ 2^k |G(z)|≤ |G(0)| exp (D), for some D≥ 1 sufficiently large depending on n. Assume that G(z) is real and non-negative for real x∈ Q_0:={(x_1,...,x_n)∈ℝ^n: |x_i|≤ 1, 1≤ i≤ n }. Suppose that R⊂12Q_0 is a sub-cube and subdivide R into equal sub-cubes R_i of side length ≍_n 1/D. Let ε>0 be given. Then there exists a subset E⊂ Q_0 of measure (E)≤ε and some constant C=C(n,ε)≥ 1 such that |log G(x)- log(1/ (R_i)∫_R_i G d)|≤ C for all x∈ R_i\ E. Let k be a sufficiently large integer depending on n. Let G(z) be holomorphic in |z|≤ 3^k, z∈ℂ^n, and satisfying max_|z|≤ 2^k |G(z)|≤ |G(0)| exp (D), for some D≥ 1 sufficiently large depending on n. Assume that G(z) is real and non-negative for real x∈ Q_0:={(x_1,...,x_n)∈ℝ^n: |x_i|≤ 1, 1≤ i≤ n }. Let δ>0, suppose that R⊂12Q_0 is a sub-cube and subdivide R into equal sub-cubes R_i of side length ≲_n c/D for some sufficiently small c=c(n,δ)>0. Then we have ∫_R_i G d≍_δ∫_2R_i G d for (1-δ) portion of the R_i. We apply Proposition <ref> twice, for D and D'≫ D instead of D, to obtain two sets E and E' of volume ≤ε (to be chosen later) and two collections of sub-rectangles R_i and R_i' of side length ≍1D and ≍aD, correspondingly, for some a to be chosen later subject to 0<a≪ 1. For all x ∈ R \ (E ∪ E'), x∈ R_i ∩ R_j' for some i,j, and ∫_R_i G d≍ G(x)(R_i)≍ G(x)(R_j')≍∫_R_j' G d, where the constant implied in the ≍ notation depend on n,ε,a. We will call R_j' good if 2R_j'⊂ R_i for some R_i and R_j' has at least one point in R\ (E∪ E'). The portion of good R_j' is at least (1-δ) if ε and a are small enough (depending on δ). For good R_j' we have ∫_2R_j' G d≤∫_R_i G d≍∫_R_j' G d, as required. One may change conclusion (<ref>) to ∫_R_i G d≍∫_Γ R_i G d, fo any parameter Γ≥ 2, and the statement of Corollary <ref> will remain true with c and the constant implied in the ≍ notation also depending on Γ (and δ>0). In order to use Proposition <ref> in the proof of Lemma <ref>, we will need a claim about the complex extension of a harmonic function u, denote by u^ℂ, to a subset of ℂ^n. There exists some constant C=C(n)≥ 1 such that for any harmonic function u on a ball B⊂ℝ^n centered at 0 the following holds. There exists a holomorphic function u^ℂ defined in Ω:={x+ iy ∈ℂ^n: x∈15B , y∈15B} such that u^ℂ(x)= u(x) for all x∈15B and sup_Ω |u^ℂ| ≤ C sup_B |u|. By considering a slightly smaller ball and rescaling, we may assume that B=B(0,1)⊂ℝ^n and u is continuous on B. Then u has a Poisson representation in terms of its boundary values given by u(x) = ∫_∂ B P(x,ζ) u(ζ)dS(ζ), where P(x,ζ)= c_n 1 - |x|^2 /|x-ζ|^n. The holomorphic extension of the Poisson kernel is P(x+iy,ζ)= c_n 1 - ∑_k (x_k +iy_k)^2 /(√(∑_k (x_k +iy_k-ζ_k)^2))^n, where x=(x_1,...,x_n), y=(y_1,...,y_n) and ζ=(ζ_1,...,ζ_n). Indeed, since √(z) is a well-defined holomorphic function for |1-z|<1, it is enough to check | 1-∑_k (x_k +iy_k-ζ_k)^2|<1. For all x+iy ∈Ω, we have |∑_k (x_k+iy_k)^2|≤ 2 (∑_k x_k^2 +y_k^2)≤4/25, and, since ζ∈∂ B, |1-∑_k (x_k +iy_k-ζ_k)^2| = | ∑_k(x_k +i y_k)^2 + 2ζ_k(x_k +iy_k)| ≤| ∑_k(x_k +i y_k)^2| +2 √(∑_k ζ_k^2)·√(∑_k (x_k+iy_k)^2)≤4/25 + 4/5=1-1/25. Therefore, P(x+iy,ζ) is a well-defined holomorphic function of x+iy in Ω and ζ∈∂ B. Moreover, |∑_k (x_k +iy_k-ζ_k)^2| ≥1/25. Thus, the absolute value of the denominator in (<ref>) is uniformly bounded from below, and sup_x+iy ∈Ω|P(x+iy,ζ)|≤ C_n. Hence, the Poisson representation gives a holomorphic extension of u onto Ω u^ℂ(x+iy) = ∫_∂ B P(x+iy ,ζ) u(ζ)dS(ζ), satisfying sup_Ω |u^ℂ| ≤ C sup_B |u|. Let k be a sufficiently large integer depending on n. Let u(x) be a harmonic function in |x|≤ 5· 3^k, x∈ℝ^n, and satisfying max_|x|≤ 5· 2^k |u(x)|≤ |u(0)| exp (D), for some D≫_n 1. Let Q_0:={(x_1,...,x_n)∈ℝ^n: |x_i|≤ 1, 1≤ i≤ n } and suppose that R⊂12Q_0 is a sub-cube. Let ε>0. Subdivide R into equal sub-cubes R_i of side length ≲_n c/D for some sufficiently small c=c(n,ε)>0. Then there exists some constant C=C(n,ε)>1 such that 𝒩^*(R_i)≤ C, for at least (1-ε) portion of the R_i. Thanks to Claim <ref>, we can extend u holomorphically onto |z|≤ 3^k. Therefore, the function G(z)=(u^ℂ(z))^2 is holomorphic for |z|≤ 3^k, non-negative and real on Q_0. Moreover, G satisfies max_|z|≤ 2^k |G|≤ |G(0)| exp (2D). Therefore, we may apply Corollary <ref> and Remark <ref> to see that ∫_Γ R_i|u|^2 d/∫_R_i|u|^2 d ≤ C_1, for at least (1-ε) portion of the R_i. Hence, Claim <ref> finishes the proof of Lemma  <ref>. We are finally ready to present the proof of Proposition <ref>. Let N=𝒩^*(Q) and partition Q into equal sub-cubes Q_j of side length ≍_ncN for some c=c(n,ε)>0 to be chosen later. We wish to cover the unit cube Q with at most C_n balls B_i=B(x_i,r) of small radius 0<r=r(n)≪_n 1 such that sup_2^10k B_i|u|≤ u(x_i)exp(C N), where k=k(n) is as in Lemma <ref>. We start with any covering of the unit cube by C_n balls B_i' of radius r/2. Put B_i=B(x_i,r), and let x_i be the point in ∂ B_i such that sup_B_i'|u| = |u(x_i)|. For each B_i', we have 𝒩(2^10k+2 B_i')≤ N (by the definition of the doubling index of the cube and assuming that r is sufficiently small). Therefore, the almost monotonicity of the doubling index, Lemma <ref>, implies sup_2^10k B_i |u|≤sup_2^10k+2 B_i'|u|≤sup_B_i'|u|exp(C k N)= |u(x_i)| exp(2k N). We may assume that k is sufficiently large. Consider a cube R_i with the following properties. The faces of R_i are parallel to the faces of Q, R_i is a union of a finite number of sub-cubes Q_j, and B_i⊂ R_i ⊂ 2^k B_i . Since the B_i cover Q, every cube Q_j is contained in at least one of the R_i. We apply rescaled Lemma <ref> to the cube R_i partitioned into the Q_j which lie inside R_i. We conclude that, if c=c(n,ε)>0 (the parameter in the side length of the Q_j) is sufficiently small, there exists C=C(n,ε)>1 such that the following holds. For at least (1-ε) portion of the Q_j in R_i, we have 𝒩^*(Q_j) ≤ C. Finally, considering all the C_n choices of R_i, we find that, for at least (1-C_nε) portion of all the Q_j, we have 𝒩^*(Q_j) ≤ C. siam

http://arxiv.org/abs/2303.07298v1

20230313172307

Wild solutions to scalar Euler-Lagrange equations

math.AP

Improvement of Neutron-HP package: Doppler broadening of the neutron elastic scattering kernel and cross sections [ March 30, 2023 ================================================================================================================= We study very weak solutions to scalar Euler-Lagrange equations associated with quadratic convex functionals. We investigate whether W^1,1 solutions are necessarily W^1,2_, which would make the theories by De Giorgi-Nash and Schauder applicable. We answer this question positively for a suitable class of functionals. This is an extension of Weyl's classical lemma for the Laplace equation to a wider class of equations under stronger regularity assumptions. Conversely, using convex integration, we show that outside this class of functionals, there exist W^1,1 solutions of locally infinite energy to scalar Euler-Lagrange equations. In addition, we prove an intermediate result which permits the regularity of a W^1,1 solution to be improved to W^1,2_ under suitable assumptions on the functional and solution. § INTRODUCTION Let f ∈ C^2(^n) be uniformly convex and Ω⊂^n any bounded Lipschitz domain. The equation EL (Df(Du)) = 0 in Ω is the Euler-Lagrange equation of the energy functional ℰ_f W^1,2(Ω) → defined by ℰ_f(u) ∫_Ω f(Du) dx. This is a standard model in the Calculus of Variations. The classical regularity theory was obtained in the fundamental works by E. De Giorgi in <cit.> and J. Nash in <cit.>. In this paper, we devote our attention to distributional solutions of (<ref>) belonging to W^1,1(Ω). We pursue the question whether such solutions are of locally finite energy which would make the theories of De Giorgi-Nash <cit.> and Schauder <cit.> applicable. In other words, we study functions u ∈ W^1,1(Ω) for which we have ∫_Ω Df(Du) · Dφ dx = 0, ∀φ∈ C_0^∞(Ω). Notice that if Df is of linear growth, the expression above is well-defined. Our study is focused on existence and non-existence of distributional solutions u ∈ W^1,1(Ω) ∖ W^1,2_(Ω). Since the energy ℰ_f of such solutions is infinite, we sometimes call such solutions non-energetic. For the Laplace equation it is well-known from Weyl's lemma that any distributional solution to Δ u = 0 in Ω is smooth, proving that no non-energetic solutions exist for the Laplace equation. The following result due to V. Šverák and X. Yan <cit.> treats the vectorial version of (<ref>): [<cit.>] There exists f ^3 × 3→ smooth, uniformly convex such that |D^2 f| ≤ c in ^3 × 3 and the equation (Df(Du)) = 0 admits a solution in W^1,p∖ W^1,2 (1 < p < 2). This in particular proves non-uniqueness of solutions in W^1,p. The existence and non-existence of non-energetic solutions has already been extensively studied for the p-Laplace equation as well as the linear equation (A(x) · Du) = 0 in Ω where λ I ≤ A(x) ≤Λ I and A is measurable. We give a swift outline of the situation for (<ref>). For the standard regularity theory we refer to the fundamental works by E. De Giorgi and J. Nash <cit.> (see also <cit.>). Subsequently, J. Serrin <cit.> provided an example of a distributional solution to (<ref>) not belonging to L^∞_. The solutions constructed by J. Serrin belong to W^1,p for all 1 < p < 2 showing that u ∈ W^1,2 is a crucial assumption in the theory by De Giorgi-Nash. Following this striking achievement, J. Serrin <cit.> conjectured that if A is Hölder continuous, then any W^1,1 solution to (<ref>) belongs to W^1,2_loc. The conjecture was confirmed by R. Hager and J. Ross <cit.> for W^1,p solutions with 1 < p < 2 and by H. Brezis <cit.> for W^1,1 solutions. The results by H. Brezis are more general than what the conjecture requires, see <cit.>. For further results regarding very weak solutions to (<ref>), we refer to <cit.>. We also give a brief description of the situation for the p-Laplace equation. For the standard regularity theory we refer to <cit.>. In <cit.>, T. Iwaniec and C. Sbordone conjectured that any W^1,r solution with r > max{1, p-1 } to the p-Laplace equation is W^1,p_ and hence classical in the interior of the domain. The conjecture was confirmed in the case when p - δ < r < p for some δ > 0 depending on n and p by T. Iwaniec and C. Sbordone <cit.>. Recently, in <cit.>, the conjecture was proved to be false in full generality. The main goal of this paper is to establish an optimal condition (see (<ref>) below) under which no non-energetic solutions exist. Our first main result is the following, which treats more general problems than (<ref>). It states that if the Euler-Lagrange equation resembles that of a convex quadratic form, then a Weyl-type result holds. Let n ≥ 1 and Ω⊂^n be a Lipschitz domain. Let f ∈ C^1(^n) be a function so that there exist 0 ≤γ < 1, A ∈^n × n an elliptic matrix with ellipticity constants m and M as well as a constant K > 0 for which we have N |Df(x) - Ax| ≤ K(1 + |x|^γ), ∀ x ∈^n. Let h ∈ L^p(Ω) for some 1 < p < n and u ∈ W^1,1(Ω) be a distributional solution of (Df(Du)) = h. Then u ∈ W^1,p^⋆_(Ω) (where p^⋆np/n-p) with u _W^1, p^⋆(Ω^')≤ C ( 1 + u _W^1, 1(Ω) + h _L^p(Ω)) for all Ω^'⋐Ω for some constant C depending only on Ω, Ω^', n, p, γ, m, M and K. Note that, although we require A to be elliptic, we do not require f to be convex. Theorem <ref> implies that if u ∈ W^1,1(Ω) is a distributional solution of (<ref>) and f satisfies (<ref>), then u ∈ W^1,p_(Ω) for all 1 < p < ∞. Our second main result demonstrates that if (<ref>) is not satisfied, e.g. if (<ref>) is satisfied with A elliptic but γ = 1, then, one can in general not improve the regularity from W^1,1 to W^1,2_ for solutions to (<ref>). Let Ω⊂^2 be a Lipschitz domain and 0 < λ < Λ < ∞ be arbitrary. Then there is a smooth and uniformly convex function f ∈ C^∞(^2) with λ I ≤ D^2f(x) ≤Λ I, ∀ x ∈^2 such that for any p < p_λ, Λ2 √(Λ)/√(λ) + √(Λ) and any α∈ (0,1), (<ref>) admits infinitely many solutions u ∈ W^1,p(Ω) ∩ C^α(Ω) with identical affine boundary datum and ∫_B |Du|^2 dx = + ∞ for any ball B ⊂Ω. We observe that p_λ, Λ→ 2 as Λ→∞ for fixed λ. Moreover, we observe that the exponent p_λ, Λ coincides with the lower critical exponent for 2D linear elliptic equations, see the works by F. Leonetti and V. Nesi <cit.>, K. Astala <cit.> and S. Petermichl and A. Volberg <cit.> (see also <cit.> and <cit.>). As a byproduct, the proof of Theorem <ref> shows that Theorem <ref> is optimal. Indeed, consider the case n = 2 and assume that f ^2 → satisfies (<ref>). Then lim_a → - ∞∂_1 f(a,b)/a = lim_a → + ∞∂_1 f(a,b)/a ∀ b ∈ and lim_b → - ∞∂_2 f(a,b)/b = lim_b → + ∞∂_2 f(a,b)/b ∀ a ∈. The functions considered in the proof of Theorem <ref> are such that lim_a → + ∞∂_1 f(a,b)/a = λ and lim_a → - ∞∂_1 f(a,b)/a = Λ ∀ b ∈ and satisfy (<ref>) with γ = 1 and some elliptic matrix A. This demonstrates that the assumptions of Theorem <ref> cannot be relaxed. Our proof of Theorem <ref> uses the convex integration method invented by S. Müller and V. Šverák <cit.> and ever since used to attack a wide range of problems related to wild solutions of PDE's, compare for instance <cit.>. Staircase laminates, which were introduced by D. Faraco in <cit.> (and subsequently further developed and used in e.g. <cit.>) are a central part of the proof. Such techniques were recently used in <cit.> to build very weak solutions to the p-Laplace equation. Finally, our last main result reveals that under a pinching condition on f and one-sided bounds on solutions u to (<ref>), regularity can be improved from W^1,1 to W^1,2_. Let n ≥ 1 and Ω⊂^n be a Lipschitz domain. Let f ∈ C^2(^n) be a convex function so that there exist 0 < λ < Λ < ∞ for which we have λ I ≤ D^2f(x) ≤Λ I for all x ∈^n. Assume that the following pinching condition Λ^2(1 - n^-1) < λ^2 holds. Let u ∈ W^1,1(Ω) be a distributional solution of (Df(Du)) = h where h ∈ L^1(Ω). Assume that there exist σ_i ∈{ -1 , 1} and L_i ∈ such that σ_i ∂_i u ≥ L_i a.e. in Ω for all i = 1, …, n. Then u ∈ W^1,2_(Ω) with u _W^1,2(Ω^')≤ C ( Du _L^1(Ω; ^n) + h _L^1(Ω)) ( u _L^1(Ω) + (L_1, …, L_n) _ℓ^2) for all Ω^'⋐Ω for some constant C depending only on Ω, Ω^', n, Λ and λ. A comparable result for the p-Laplace equation is already available. We refer the reader to <cit.>, the proof of which is close to that of Theorem <ref>. Theorem <ref> leaves open whether the pinching condition (<ref>) is necessary. §.§ Plan of the paper In Section <ref>, we introduce some notation and preliminary results. We prove Theorem <ref> in Section <ref>. In the next section we prove Theorem <ref>. The remainder of the paper is dedicated to proving Theorem <ref>. In Section <ref> we do the necessary geometric construction. In Section <ref>, we give a detailed description of our convex integration scheme. The two last sections prove Theorem <ref>. § ACKNOWLEDGEMENTS The author was supported by the SNSF Grant 182565 and by the Swiss State Secretariat for Education, Research and lnnovation (SERI) under contract number MB22.00034. The author thanks Riccardo Tione and Maria Colombo for suggesting to investigate the topic of very weak solutions to Euler-Lagrange equations. Their suggestions have been of great value. This paper is an outcome of improvements made to results proved in the master thesis of the author carried out at EPFL in fall 2021 <cit.>. The advice from Xavier Fernández-Real regarding Theorem <ref> is gratefully acknowledged. § NOTATION AND PRELIMINAIRES In this section, we introduce some notation and results from classical convex integration, see <cit.>. §.§ Notation Throughout this paper, Ω⊂^n denotes any bounded Lipschitz domain in ^n. We will use | · | to denote the n-dimensional Lebesgue measure in ^n. We also use this notation to denote the Euclidian norm i.e. |x| = |(x_1, …, x_n)| = ( ∑_i = 1^n x_i^2 )^1/2, x ∈^n. For any map u Ω→^m the j-th component is denoted by u^j. A continuous Lipschitz map u Ω→^m will be called piecewise affine if there is a collection ℱ of open convex mutually disjoint sets covering Ω up to a set of Lebesgue measure 0 such that for all U ∈ℱ, u |_U is affine. Let n,m ≥ 2. The set of real-valued matrices with n rows and m columns is denoted by ^n × m. A matrix A ∈^n × n is said to be elliptic if there exists constants 0 < m ≤ M such that mI ≤ A ≤ MI i.e. m |x|^2 ≤ Ax · x ≤ M |x|^2 ∀ x ∈^n. The constants m and M are called the ellipticity constants of A. The set of real-valued symmetric matrices with n rows and n columns is denoted by ^n × n_. We say that two matrices A, B ∈^n × m are rank-one connected if A - B≤ 1. If A and B are rank-one connected, the segment [A,B] { tA + (1 - t)B : t ∈ [0,1] } is called a rank-one segment. A function is called rank-one convex if it is convex along any rank-one segment. We let ρ∈ C^∞_0(^n) be a function such that supp(ρ) ⊂ B_1 and ∫_B_1ρ dx = 1. We define ρ_∈ C^∞_0(^n) as ρ_(x) 1/^nρ( x/). For any u ∈ W^1,1(Ω), we denote its convolution with ρ_ as u_ u ∗ρ_. Note that u_ is only well-defined in { x ∈Ω: xΩ^c > }. To overcome this restriction in the case where u has affine boundary datum, we extend u in an affine manner to ^n. Then u_ is well-defined on the whole Ω. This is also true for (Du)_. Finally, given x_1, …, x_N, we define ∑_k = i^j x_k 0 and ∏_k = i^j x_k 1 whenever j < i. §.§ Results from Lipschitz convex integration In this subsection we introduce the building blocks of the convex integration scheme invented by Müller and Šverák and some extensions due to Astala, Faraco and Székelyhidi. The content of this subsection comes from <cit.>. The barycenter of a probability measure ν on ^n × m is ν = ∫_^n × m X dν(X) ∈^n × m. A probability measure on ^n × m ν is a laminate if f(ν) ≤∫_^n × m f(X) dν(X) for all f ^n × m→ rank-one convex. Assume ν = ∑_j = 1^r λ_jδ_A_j with A_j ≠ A_k whenever j ≠ k. The measure ν^' is said to be obtained from ν by an elementary splitting if for some j and some λ∈ [0,1], there exists a rank-one segment [B_1, B_2] containing A_j with (1 - s)B_1 + sB_2 = A_j such that ν^' = ν + λλ_j ( (1 - s)δ_B_1 + sδ_B_2 - δ_A_j). It follows from the definition of laminates that any probability measure that is obtained by an elementary splitting of a lamiate is also a laminate. This justifies the notion of laminates of finite order: We say that a laminate ν is a laminate of finite order if there exists probability measures ν_1, …, ν_m such that ν_1 = δ_A for some A, ν_m = ν and for each j = 1, …, m-1, ν_j+1 can be obtained from ν_j by an elementary splitting. Let ν = ∑_j = 1^r λ_j δ_A_j (with A_j ∈^n × m for all j and A_j ≠ A_k whenever j ≠ k) be a laminate of finite order and let A = ν be the barycenter of ν. Then for any α∈ (0,1), any > 0, any b ∈^m and any 0 < δ < 1/2min{ |A_i - A_j| : i ≠ j } there is a piecewise affine mapping u Ω→^m such that * u - (Ax + b) _C^α(Ω; ^m) <; * u = Ax + b on ∂Ω; * |{ x ∈Ω : |Du(x) - A_j| < δ}| = λ_j |Ω| for all j = 1, …, r. If m = n and all the matrices A_j ∈^n × n_, then the piecewise affine mapping u can be selected so that there exists g ∈ W^2, ∞(Ω) for which we have u = Dg. The last sentence in the previous lemma is particularly important in our construction. Indeed, we will use the following which is a consequence of the lemma above: let ν = ∑_j = 1^r λ_j δ_A_j be a laminate of finite order with A_j ∈^n × n_ for all j = 1, …, r, A_i ≠ A_j whenever i ≠ j and A = ν. In addition, for all j = 1, …, r, let U_j ⊂^n × n_ be an open neighbourhood of A_j in ^n × n_. Then for any α∈ (0,1), any > 0 and any b ∈^n, there exists a piecewise affine mapping u Ω→^n such that * u - (Ax + b) _C^α(Ω; ^n) <; * u = Ax + b on ∂Ω; * |{ x ∈Ω : Du(x) ∈ U_j }| = λ_j |Ω| for all j = 1, …, r. § PROOF OF THEOREM <REF> In this section, we prove Theorem <ref>. We first need a preliminary lemma. We recall that for any p ∈ [1,n), we define p^⋆ as p^⋆ = np/n - p. Let 1 < p < n, 1 < q < ∞ and A ∈^n × n be an elliptic matrix with ellipticity constants m and M. If u ∈ W^1,1(Ω) is a distributional solution of (A Du) = h + G for some h ∈ L^p(Ω) and G ∈ L^q(Ω; ^n). Then u ∈ W^1, min(p^⋆, q)_(Ω) with u _W^1, p^⋆(Ω^')≤ C ( u _W^1, 1(Ω) + h _L^p(Ω) + G _L^q(Ω; ^n) ) for all Ω^'⋐Ω for some constant C depending only on Ω, Ω^', n, p, q, m and M. Moreover, letting m and M be real numbers such that m I ≤ A ≤ M I, then for any Ω^'⋐Ω there exists C depending only on Ω, Ω^', n, p, m and M such that u _W^1, p^⋆(Ω^')≤ C ( u _W^1, 1(Ω) + h _L^p(Ω) + G _L^q(Ω; ^n) ) Step 1: Assume that A is the identity matrix. Then u solves Δ u = h + G. Let B ⊂Ω be an arbitrary ball. By <cit.> and <cit.>, there exists u_B ∈ W^1, min(p^⋆, q)(B) solving Δ u_B = h + G on B with u_B _W^1, min(p^⋆, q)(B)≤ c ( h _L^p(B) + G _L^q(B; ^n)) for some c depending only on B, n, p and q. Hence u - u_B is harmonic on B. Thus u ∈ W^1, min(p^⋆, q)_(B). By Caldéron-Zygmund estimates, for any B^'⋐ B ⊂Ω u - u_B _W^1, min(p^⋆, q)(B^')≤ c^' u - u_B _W^1, 1(B). for some c^' depending only on B, B^', n, p and q. Therefore, u _W^1, min(p^⋆, q)(B^') ≤ u - u_B _W^1, min(p^⋆, q)(B^') + u_B _W^1, min(p^⋆, q)(B^') (<ref>) ≤ c^' u - u_B _W^1, 1(B) + u_B _W^1, min(p^⋆, q)(B^') ≤ 2 c^' ( u _W^1, 1(B) + u_B _W^1, min(p^⋆, q)(B) ) (<ref>) ≤ 2 c c^' ( u _W^1, 1(B) + h _L^p(B) + G _L^q(B; ^n) ). Since u ∈ W^1, min(p^⋆, q)_(B) for any ball B ⊂Ω, we conclude that u ∈ W^1, min(p^⋆, q)_(Ω). In virtue of (<ref>), for any Ω^'⋐Ω, u _W^1, min(p^⋆, q)(Ω^')≤ C ( u _W^1, 1(Ω) + h _L^p(Ω) + G _L^q(Ω; ^n) ) for some constant C depending only on Ω, Ω^', n, p and q. Step 2: Let us now assume that A is an arbitrary symmetric positive definite matrix. Then there exists an invertible matrix R ∈^n × n such that A = R^T I R. Let R^-TΩ{ R^-T x : x ∈Ω} and define v ∈ W^1,1(R^-TΩ) as v(x) u(R^T x). Then v solves Δ v = h̃ + G where h̃∈ L^p(R^-TΩ) and G∈ L^q(R^-TΩ; ^n) are defined as h̃(x) h(R^T x) and G(x) R^-T G(R^T x). By Step 1, v ∈ W^1, min(p^⋆, q)_(R^-TΩ). We conclude that u ∈ W^1, min(p^⋆, q)_(Ω). By (<ref>), for any Ω^'⋐Ω, u _W^1, min(p^⋆, q)(Ω^')≤ C ( u _W^1, 1(Ω) + h _L^p(Ω) + G _L^q(Ω; ^n) ) for some constant C depending only on Ω, Ω^', n, p, q, m and M. Step 3: Assume that A is only known to be elliptic. Let A_ be the symmetric part of A is defined as A_ = 1/2(A + A^T). Observe that (A_ Du) = (A Du) = h + G. By Step 2, u ∈ W^1, min(p^⋆, q)_(Ω) and (<ref>) holds. This ends the proof of Lemma <ref>. We can now prove Theorem <ref>. We begin with the case when 0 < γ < 1. We find that (A Du) = (A Du - Df(Du)_=: G ) + (Df(Du)) = G + h. Since | G | ≤ |A Du - Df(Du)| ≤ K(1 + |Du|^γ), we find that G ∈ L^1γ(Ω; ^n) with G _L^1γ(Ω; ^n)≤ c_K,γ (1 + Du _L^1(Ω; ^n)) for some constant c_K,γ depending only on K and γ. Therefore by Lemma <ref>, u ∈ W^1, min(1γ, p^⋆)_(Ω). Due to (<ref>), for any Ω^'⋐Ω, there exists a constant c depending only Ω, Ω^', n, p, m, M and γ such that u _W^1, min(1γ, p^⋆)(Ω^') ≤ c ( u _W^1, 1(Ω) + h _L^p(Ω) + G _L^1γ(Ω; ^n) ) ≤ c ( u _W^1, 1(Ω) + h _L^p(Ω) + c_K,γ (1 + Du _L^1(Ω; ^n)) ) ≤ 2 c c_K,γ (1 + u _W^1, 1(Ω) + h _L^p(Ω)). Thus, by (<ref>), G ∈ L^1γmin(1γ, p^⋆)_(Ω; ^n). By applying the same argument as above, u ∈ W^1, min(1γ^2, p^⋆)_(Ω) with u _W^1, min(1γ^2, p^⋆)(Ω^')≤ c^' (1 + u _W^1, 1(Ω) + h _L^p(Ω)) for all Ω^'⋐Ω for some constant c^' depending only on Ω, Ω^', n, p, γ, m, M and K. By a bootstrapping argument on a finite increasing sequence of compacts between Ω^' and Ω, u ∈ W^1, p^⋆_(Ω) with u _W^1, p^⋆(Ω^')≤ C (1 + u _W^1, 1(Ω) + h _L^p(Ω)) for all Ω^'⋐Ω for some constant C depending only on Ω, Ω^', n, p, γ, m, M and K. If γ = 0, then the argument above gives that G ∈ L^∞(Ω; ^n). Thus u ∈ W^1, p^⋆_(Ω; ^n) and (<ref>) holds. This finishes the proof. § PROOF OF THEOREM <REF> In this section, we shall prove Theorem <ref>. First, however, we need a preliminary lemma. We recall the notation w_ w ∗ρ_ from Subsection <ref>. Let n ≥ 1, Ω⊂^n and u ∈ W^1,1(Ω). Assume that for all i = 1, …, n, there exist σ_i ∈{ -1, 1} and L_i ∈ such that σ_i ∂_i u ≥ L_i a.e. on Ω. Then for any Ω^'⋐Ω and < 1/2Ω^'Ω^c (|Du|)_≤ n^12 |Du_| + 2 n^12 (L_1, …, L_n) _ℓ^2 a.e. on Ω^'. Step 1: Fix some arbitrary Ω^'⋐Ω and < 1/2Ω^'Ω^c. We start by assuming that all derivatives ∂_i u preserve their sign over Ω. By this, we mean that there exists σ_i ∈{ -1, 1} for all i= 1, …, n so that ∂_i u = σ_i |∂_i u|. Then, for all x ∈Ω^' (|Du|)_(x) = ∫_Ω( ∑_i = 1^n |∂_i u|^2 )^12 (x-y) ρ_(y) dy ≤∫_Ω∑_i = 1^n |∂_i u (x-y)| ρ_(y) dy = ∑_i = 1^n | ∫_Ω∂_i u(x-y) ρ_(y) dy | ≤ n^12( ∑_i = 1^n| ∫_Ω∂_i u(x-y) ρ_(y) dy |^2 )^12 = n^12 |Du_|(x). Step 2: Once again, fix some arbitrary Ω^'⋐Ω and < 1/2Ω^'Ω^c. Now we assume that there are σ_i ∈{ -1, 1 } and L_i ∈ such that σ_i ∂_i u ≥ L_i a.e. on Ω. Define a new function as ũ(x) u(x) - ∑_i = 1^nσ_i L_i x_i. The function ũ satisfies the conditions of Step 1 and therefore (|Dũ|)_≤ n^12 |Dũ_| a.e. on Ω^'. We conclude that (|Du|)_(x) ≤ (|Dũ|)_(x) + (L_1, …, L_n) _ℓ^2 ≤ n^12 |Dũ_|(x) + (L_1, …, L_n) _ℓ^2 ≤ n^12 |Du_|(x) + 2 n^12 (L_1, …, L_n) _ℓ^2 for all x ∈Ω^'. We can now prove Theorem <ref>. The proof is close to the proof of <cit.>. Step 1: In this first step, we assume that n(Λ^2 - 2λ + 1) < 1. The reason for this assumption is not clear at this point but will become clear in the proof. Moreover, without loss of generality, we can assume that λ < Λ. Indeed, if λ = Λ, then there is a λ^' < Λ such that n(Λ^2 - 2λ^' + 1) < 1 and λ^' I ≤ D^2 f(x) ≤Λ I still holds. Notice that, since we assume λ < Λ, we have Λ^2 - 2 λ + 1 > Λ^2 - 2 Λ + 1 = (Λ - 1)^2 ≥ 0. In addition, we may, without loss of generality, assume that Df(0) = 0. We have ∫_Ω Df(Du) · D φ dx = ∫_Ω h φ dx ∀φ∈ C^∞_0(Ω). Let Ω^'⋐Ω^''⋐Ω^'''⋐Ω with Ω^'Ω^c≤ 2 Ω^'''Ω^c and let < Ω^'''Ω^c. We find that ∫_Ω^'' (Df(Du))_· D φ dx = ∫_Ω^'' h_φ dx ∀φ∈ C^∞_0(Ω^''). Let Ψ∈ C^∞_0(Ω^'') such that Ψ = 1 on Ω^' and D Ψ_L^∞(Ω^''; ^n)≤ 5 Ω^'Ω^''^c^-1. Then ∫_Ω^'' (Df(Du))_· D(Ψ u_) dx = ∫_Ω^'' h_Ψ u_ dx. which leads to ∫_Ω^''((Df(Du))_· Du_) Ψ dx = - ∫_Ω^''((Df(Du))_· D Ψ) u_ dx + ∫_Ω^'' h_Ψ u_ dx. The left-hand side in (<ref>) equals ∫_Ω^'' |Du_|^2 Ψ dx + ∫_Ω^'' ((Df(Du))_ - Du_) · Du_Ψ dx. Observe that |Df(Du) - Du|^2 = |Df(Du)|^2 - 2 Df(Du) · Du + |Du|^2 = |Df(Du)|^2_≤Λ^2 |Du|^2 - 2 (Df(Du) - Df(0)) · (Du - 0)_≥λ |Du|^2 + |Du|^2 ≤ (Λ^2 - 2 λ + 1) |Du|^2 and hence, by Lemma <ref>, we have |(Df(Du))_ - Du_| ≤ |Df(Du) - Du|_≤ (Λ^2 - 2 λ + 1)^12(|Du|)_ (<ref>) ≤ n^12(Λ^2 - 2 λ + 1)^12|Du_| + 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2. Therefore | ∫_Ω^'' ((Df(Du))_ - Du_) · Du_Ψ dx | ≤ n^12(Λ^2 - 2 λ + 1)^12∫_Ω^'' |Du_|^2 Ψ dx + 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2∫_Ω^'' |Du_| Ψ dx. We observe that due to <cit.> and (<ref>) there exists a constant C = C(n, Ω, Ω^'Ω^c) such that u_^'_L^∞(Ω^'')≤ C( u_^'_L^1(Ω^''') + (L_1, …, L_n) _ℓ^2) ∀^' < Ω^'''Ω^c. Then, by combining (<ref>) with (<ref>) and (<ref>), we find ( 1 - n^12(Λ^2 - 2 λ + 1)^12) ∫_Ω^'' |Du_|^2 Ψ dx ≤ 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2∫_Ω^'' |Du_| Ψ dx + | ∫_Ω^''((Df(Du))_· D Ψ) u_ dx | + | ∫_Ω^'' h_Ψ u_ dx | ≤ 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2 Du__L^1(Ω^''; ^2) + (Df(Du))__L^1(Ω^''; ^n) D Ψ_L^∞(Ω^''; ^n) u__L^∞(Ω^'') + h__L^1(Ω^'')Ψ_L^∞(Ω^'') u__L^∞(Ω^'') (<ref>) ≤ 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2 Du__L^1(Ω^''; ^2) + C Λ Du _L^1(Ω; ^2) D Ψ_L^∞(Ω^''; ^n) ( u__L^1(Ω^''') + (L_1, …, L_n) _ℓ^2 ) + C h _L^1(Ω) ( u__L^1(Ω^''') + (L_1, …, L_n) _ℓ^2 ) ≤ 2 n^12 (Λ^2 - 2 λ + 1)^12 (L_1, …, L_n) _ℓ^2 Du _L^1(Ω; ^2) + C ( Λ Du _L^1(Ω; ^2) D Ψ_L^∞(Ω^''; ^n) + h _L^1(Ω))( u _L^1(Ω) + (L_1, …, L_n) _ℓ^2 ) =: C^' By lower semi-continuity, ∫_Ω^' |Du|^2 dx ≤∫_Ω^'' |Du|^2 Ψ dx ≤lim inf_→ 0∫_Ω^'' |Du_|^2 Ψ dx ≤C^'1 - n^12(Λ^2 - 2 λ + 1)^12 < ∞. Thus, u ∈ W^1,2(Ω^'). Therefore, we conclude u ∈ W^1,2_(Ω) and (<ref>) holds. Step 2: In this step, we drop the assumption that n (Λ^2 - 2 λ + 1) < 1. Since u solves the equation (Df(Du)) = h, u also solves the equation ( 1/βDf(Du) ) = 1/β h for all β > 0. By direct computations, we find that the assumption that Λ^2(1 - n^-1) < λ^2 is sufficient to deduce that there exists a β > 0 such that n ( Λ^2/β^2 - 2 λ/β + 1 ) < 1. We can then apply the arguments from Step 1 to the equation (<ref>) to conclude the proof. Since the choice of β depends only on n, Λ and λ, (<ref>) still holds. § GEOMETRIC CONSTRUCTIONS (PART 1 OF THE PROOF OF THEOREM <REF>) In this section, we do the necessary geometric constructions to be able to perform a convex integration scheme in the next section. In Subsection <ref>, we show that (<ref>) is equivalent to a differential inclusion Dw ∈ K_f. In this first subsection, we also select f depending on the values of λ and Λ. The set K_f is central in the geometric constructions. In Subsection <ref>, we construct laminates of finite order partly supported on K_f. Although partly supported in the set K_f, it is not clear how to use these laminates together with Lemma <ref> (see also Remark <ref>) in order to construct a functioning convex integration scheme. Therefore, in Subsection <ref>, we define interpolation maps. In Subsection <ref>, we use these maps to modify the already constructed laminates so that they can be used to construct a functioning convex integration scheme. §.§ Equation (<ref>) as a differential inclusion and choice of f The aim of this subsection is to show that (<ref>) is equivalent to a differential inclusion. Define K_f {[ a b; ∂_2 f(a,b) - ∂_1 f(a,b); ] : a, b ∈}. Assume u ∈ W^1,1(Ω) solves (<ref>). Define J [ 0 1; -1 0; ]. Since u solves (<ref>), JDf(Du) = 0. Due to this, there is ũ∈ W^1,1(Ω) such that D ũ = JDf(Du). Define w (u, ũ) ∈ W^1,1(Ω; ^2). Then Dw(x) ∈ K_f for a.e. x∈Ω. Now instead assume that for some w ∈ W^1,1(Ω; ^2), we have Dw(x) ∈ K_f for a.e. x ∈Ω. Then (Df(Dw^1)) = Dw^2 = 0. Therefore, equation (<ref>) is equivalent to the differential inclusion Dw(x) ∈ K_f for a.e. x ∈Ω. Given 0 < λ < Λ as in Theorem <ref>, let φ∈ C^∞() by any function such that 0 < λ < inf_a ∈φ^''(a) ≤sup_a ∈φ^''(a) < Λ, lim_a → - ∞φ^'(a)/a = Λ and lim_a → + ∞φ^'(a)/a = λ. We assume, without loss of generality, that λ < 1 < Λ. Then, choose f ^2 → to be f(a,b) = φ(a) + b^22. The set K_f becomes K_f = {[ a b; b - φ^'(a); ] : a, b ∈}⊂^2 × 2_. §.§ Construction and parametrisation of laminates of finite order In this subsection, we build laminates of finite order depending on a set of parameters. The space of parameters is defined as 𝒫 (1,2) × (1,2) × (-1,1) ⊂^3. We define the maps a_0^+, a_0^-, b 𝒫→ as follows. Let P = (a_0^+, a_0^-, b) ∈𝒫 and define a_0^+(P) a_0^+, a_0^-(P) a_0^-, and b(P) b. In other words, a_0^+ is the projection from 𝒫 into the first coordinate, a_0^- is the projection from 𝒫 into the second coordinate and so on. Let 1 < r < 2 be a constant to be determined later (see Lemma <ref>). For each P ∈𝒫, we define the sequences { a_k^+(P) }_k = 1^∞ and { a_k^-(P) }_k = 1^∞ as a_k^+(P) a_0^+(P) + r^k and a_k^-(P) a_0^-(P) + √(λ/r Λ) r^k for all k ≥ 1. The justification for this notation is that a_k^-(P) will always be used with a minus sign in front of it and a_k^+(P) without a minus sign. For each P ∈𝒫 and each i ≥ 1, we define A_i(P) [ a_i^+(P) b(P); b(P) - φ^'(- a_i^-(P)) ]. We split the matrix A_i(P) along the rank-one direction [ 0 0; 0 x ], x ∈ with endpoints B_i(P) and C_i(P) where B_i(P) [ a_i^+(P) b(P); b(P) - φ^'(a_i^+(P)) ] and C_i(P) [ a_i^+(P) b(P); b(P) - φ^'(- a_i+1^-(P)) ]. We have A_i(P) = λ_i^1(P) B_i(P) + (1 - λ_i^1(P)) C_i(P) with λ_i^1(P) φ^'(- a_i^-(P)) - φ^'(- a_i+1^-(P))/φ^'(a_i^+(P)) - φ^'(- a_i+1^-(P)). Then we split the matrix C_i(P) along the rank-one direction [ x^' 0; 0 0; ], x^'∈ with endpoints D_i(P) and E_i(P) where D_i(P) [ - a_i+1^-(P) b(P); b(P) - φ^'(- a_i+1^-(P)) ] and E_i(P) [ a_i+1^+(P) b(P); b(P) - φ^'(- a_i+1^-(P)) ]. We have C_i(P) = λ_i^2(P)/1 - λ_i^1(P) D_i(P) + λ_i^3(P)/1 - λ_i^1(P) E_i(P) where λ_i^2(P) φ^'(a_i^+(P)) - φ^'(- a_i^-(P))/φ^'(a_i^+(P)) - φ^'(- a_i+1^-(P))a_i+1^+(P) - a_i^+(P)/a_i+1^-(P) + a_i+1^+(P) and λ_i^3(P) φ^'(a_i^+(P)) - φ^'(- a_i^-(P))/φ^'(a_i^+(P)) - φ^'(- a_i+1^-(P))a_i+1^-(P) + a_i^+(P)/a_i+1^-(P) + a_i+1^+(P). Notice that E_i(P) = A_i+1(P). By the construction above, we have A_i(P) = λ_i^1(P) B_i(P) + λ_i^2(P) D_i(P) + λ_i^3(P) E_i(P) = λ_i^1(P) B_i(P) + λ_i^2(P) D_i(P) + λ_i^3(P) A_i+1(P) and μ^(i)(P) λ_i^1(P) δ_B_i(P) + λ_i^2(P) δ_D_i(P) + λ_i^3(P) δ_A_i+1(P) is a laminate of finite order with barycenter A_i(P). For an illustration of the matrices A_i(P), B_i(P), C_i(P), D_i(P) and E_i(P), we refer to Figure <ref>. We end this subsection by a result about the behaviour of λ_i^3 as i is large. We recall that p is a fixed arbitrary quantity in Theorem <ref>, which satisfies p < p_λ, Λ = 2 √(Λ)/√(λ) + √(Λ). For any 0 < δ < 1/10(p_λ, Λ - p), there exists 1 < r < 2 and an integer I_0 such that for all i ≥ I_0 1/r^2 < inf_P ∈𝒫λ_i^3(P) ≤sup_P ∈𝒫λ_i^3(P) < 1/r^p + 2 δ. By standard calculations, lim sup_i →∞sup_P ∈𝒫λ_i^3(P) ≤( λ√(r) + √(Λλ)/λ√(r) + √(Λλ) r)^2 ≤lim inf_i →∞inf_P ∈𝒫λ_i^3(P). Then we observe that lim_r → 1log_r [ ( λ√(r) + √(Λλ) r/λ√(r) + √(Λλ))^2 ] = 2 √(Λ)/√(λ) + √(Λ) = p_λ, Λ. Thus, there exists an 1 < r < 2 such that 2 > log_r [ ( λ√(r) + √(Λλ) r/λ√(r) + √(Λλ))^2 ] > p + 2 δ which implies r^2 > ( λ√(r) + √(Λλ) r/λ√(r) + √(Λλ))^2 > r^p + 2 δ. We conclude (<ref>). From now on, we let δ be a fixed quantity satisfying 0 < δ < 1/10(p_λ, Λ - p). The quantities r and I_0 denote those given by the above lemma. In addition, we let C_r,p≥ 0 denote a constant for which we have j r^-p(j - 1)≤ C_r,p ∀ j ≥ 1. §.§ Definition and properties of interpolation maps The goal of this subsection is to define interpolation maps and prove useful properties of these maps. For any i ≥ 1, t ∈ (0,1), we define Φ_i,t^1, Φ_i,t^2 𝒫→^2 × 2_sym as Φ_i,t^1(P) = tB_i(P) + (1 - t)A_i(P) and Φ_i,t^2(P) = tD_i(P) + (1 - t)A_i+1(P). For an illustration of the role of the matrices Φ_i,t^1(P) and Φ_i,t^2(P) in the construction, we refer the reader to Figure <ref>. Having defined these maps, the main goal of the remainder of this subsection is to prove the following two lemmas. For all i ≥ 1, A_i is an open map. For all t ∈ (0,1) and all i ≥ 1, Φ_i,t^1 and Φ_i,t^2 are open maps. Since the dimension of the domain and the codomain agree, it suffices to show that A_i is continuous injection (see for instance <cit.>). The continuity of the map A_i follows from the fact that a_i^+, a_i^-, b, φ^' are continuous. Now, we only need to show that A_i is injective. Assume that for two P, P̃∈𝒫, we have A_i(P) = A_i(P̃). Write P = (a_0^+, a_0^-, b) and P̃ = (ã_0^+, ã_0^-, b̃). From (<ref>), b = b̃ and a_i^+(P)= a_i^+(P̃) which leads to a_0^+ = ã_0^+. In addition, (<ref>) also implies that φ^'(- a_i+1^-(P)) = φ^'(- a_i+1^-(P̃)) and since φ^'' > 0, a_i+1^-(P) = a_i+1^-(P̃) which leads to a_0^- = ã_0^-. This proves the injectivity of A_i as wished. The proof is divided into 2 similar steps. In the first, we treat Φ_i,t^1 and the second treats Φ_i,t^2. Step 1: We first prove that Φ_i,t^1 is an open map for all t ∈ (0,1). Define the map π_1 ^2 × 2_→^2 × 2_ as π_1 ( [ a_11 a_12; a_12 a_22; ]) = [ a_11 a_12; a_12 - φ^'(a_11) ] and note that B_i = π_1 ∘ A_i for all i ≥ 1. Define ψ_t^1 ^2 × 2_→^2 × 2_ as ψ_t^1(X) = t π_1(X) + (1 - t)X and note that Φ_i,t^1 = ψ_t^1 ∘ A_i. Since A_i is an open map by Lemma <ref>, it suffices to show that ψ_t^1 is an open map. To show that ψ_t^1 is an open map, we will show that the differential has full rank (see for instance <cit.>). Instead of considering ψ_t^1 as a mapping from ^2 × 2_ to ^2 × 2_, we choose (without changing the notation) to view ψ_t^1 as a mapping from ^3 to ^3. Then ψ_t^1(a_11, a_12, a_22) = (a_11, a_12, -t φ^'(a_11) + (1 - t) a_22 ) and the differential is D ψ_t^1(a_11, a_12, a_22) = [ 1 0 0; 0 1 0; - t φ^''(a_11) 0 1-t; ]. Then D ψ_t^1(a_11, a_12, a_22) = 1 - t > 0 for all t ∈ (0,1), which means that ψ_t^1 is an open map. Hence Φ_i,t^1 is an open map. Step 2: We now prove that Φ_i,t^2 is an open map for all t ∈ (0,1). By standard arguments, φ^' admits an inverse maps (φ^')^-1. Define π_2 ^2 × 2_→^2 × 2_ as π_2 ( [ a_11 a_12; a_12 a_22; ]) = [ (φ^')^-1(- a_22) a_12; a_12 a_22; ]. Note that D_i = π_2 ∘ E_i = π_2 ∘ A_i+1 for all i ≥ 1. Define ψ_t^2 ^2 × 2_→^2 × 2_ as ψ_t^2(X) = t π_2(X) + (1-t)X and note that Φ_i,t^2 = ψ_t^2 ∘ A_i+1. Since A_i+1 is an open map, it suffices to show that ψ_t^2 is an open map to deduce that Φ_i,t^2 is an open map. We can proceed in the same way as in Step 1 to show that ψ_t^2 is an open map. This ends the proof. §.§ Construction and parametrisation of interpolation laminates In this subsection, the aim is to define new laminates with help of the interpolation maps from Subsection <ref>. As explained in the beginning of this section, these laminates are needed to overcome the difficulty caused by the errors in Lemma <ref>. Recall the definition of μ^(i) in (<ref>). We consider this laminate but replace B_i(P) by Φ_i,t^1(P) and D_i(P) by Φ_i,t^2(P) (see Figure <ref>). The coefficients corresponding to λ_i^1(P), λ_i^2(P) and λ_i^3(P) in (<ref>) also have to be adapted. For all i and t ∈ (0,1), we define μ^(i, i+1)_t(P) λ_i,t^1(P) δ_Φ_i,t^1(P) + λ_i,t^2(P) δ_Φ_i,t^2(P) + λ_i,t^3(P) δ_A_i+1(P) where λ_i,t^1(P) λ_i^1(P)/t + (1-t)λ_i^1(P), λ_i,t^2(P) λ_i^2(P)/t + (1-t)λ_i^1(P), and λ_i,t^3(P) t λ_i^3(P) - (1-t)λ_i^2(P) /t + (1-t)λ_i^1(P). By Lemma <ref>, there is T_0 ∈ (0,1) and an integer I_1 ≥ 3 such that for all i ≥ I_1 and all t ∈ (T_0, 1) we have 1/r^2 < inf_t ∈ (T_0, 1)inf_P ∈𝒫λ_i,t^3(P) ≤sup_t ∈ (T_0, 1)sup_P ∈𝒫λ_i,t^3(P) < 1/r^p + 2 δ. By standard computations, the barycenter of μ^(i, i+1)_t(P) is μ^(i, i+1)_t(P) = A_i(P). We define a new laminate of finite order μ_t^(i,i+2)(P) by replacing the delta measure δ_A_i+1(P) in (<ref>) by μ_t^(i+1,i+2)(P). In other words, μ^(i, i+2)_t(P) λ_i,t^1(P) δ_Φ_i,t^1(P) + λ_i,t^2(P) δ_Φ_i,t^2(P) + λ_i,t^3(P) λ_i+1,t^1(P) δ_Φ_i+1,t^1(P) + λ_i,t^3(P) λ_i+1,t^2(P) δ_Φ_i+1,t^2(P) + λ_i,t^3(P) λ_i+1,t^3(P) δ_A_i+2(P). We continue in this way which results in the following definition μ^(i,j)_t(P) ∑_k = i^j-1( ∏_m = i^k-1λ_m,t^3(P) ) λ_k,t^1(P) δ_Φ_k,t^1(P) + ∑_k = i^j-1( ∏_m = i^k-1λ_m,t^3(P) ) λ_k,t^2(P) δ_Φ_k,t^2(P) + ( ∏_m = i^j-1λ_m,t^3(P) ) δ_A_j(P) ∀ i < j, t ∈ (0,1). Note that the barycenter of μ_t^(i,j)(P) is μ_t^(i,j)(P) = A_i(P). In order to correct errors (see Lemma <ref>) and make sure that the gradients of our sequence of functions progressively approaches the set K_f, we define the following laminates of finite order μ^1, (i,j)_t,t^'(P) tt^'δ_Φ_i,t^'^1(P) + ( 1 - tt^') μ^(i,j)_t^'(P), μ^2, (i,j)_t,t^'(P) tt^'δ_Φ_i,t^'^2(P) + ( 1 - tt^') μ^(i+1,j)_t^'(P). Notice that the barycenter of these two laminates are μ^1, (i,j)_t,t^'(P) = Φ_i,t^1(P) and μ^2, (i,j)_t,t^'(P) = Φ_i,t^2(P). Now, we define collections of sets that will be useful in the proof in the next section. For each i ≥ 1 and t ∈ (0,1), we define the sets 𝒱_i A_i(𝒫), 𝒰_i^1 B_i(𝒫), 𝒰_i^2 D_i(𝒫), 𝒲_i,t^1 Φ_i,t^1(𝒫), 𝒲_i,t^2 Φ_i,t^2(𝒫). By Lemma <ref> and <ref>, the sets 𝒱_i, 𝒲_i,t^1 and 𝒲_i,t^2 are open for all i ≥ 1 and t ∈ (0,1). The laminates (<ref>), (<ref>), (<ref>), the openness of the sets 𝒱_i, 𝒲_i,t^1 and 𝒲_i,t^2 combined with Lemma <ref> (see also Remark <ref>) immediately leads to the three following propositions (for which we shall provide an explicit proof only for Proposition <ref> because the proof of Proposition <ref> is essentially the same and the proof of Proposition <ref> is in the same spirit but slightly less involved. Recall the constant I_1 and T_0 defined in (<ref>)) Let j > i ≥ I_1, P ∈𝒫, α∈ (0,1), t ∈ (T_0, 1), b ∈^2 and U convex and open. For any > 0, there is a piecewise affine Lipschitz mapping w U →^2 such that * w - (A_i(P)x + b) _C^α(U, ^2) <; * w = A_i(P)x + b on ∂ U; * Dw(x) ∈⋃_k = i^j-1𝒲_k,t^1 ∪⋃_k = i^j-1𝒲_k,t^2 ∪𝒱_j for a.e. x ∈ U; * (1 - 1/r^p) ( 1/r^2 )^k-i |U| ≤ |{ Dw ∈𝒲_k, t^1 ∪𝒲_k, t^2 }| ≤ ( 1 - 1/r^2) ( 1/r^p + 2 δ )^k-i |U| for all i ≤ k < j; * (1/r^2)^j-i |U| ≤ |{ Dw ∈𝒱_j}| ≤ (1/r^p + 2 δ)^j-i |U|. Let j > i ≥ I_1, P ∈𝒫, α∈ (0,1), T_0 < t < t^' < 1, b ∈^2 and U convex and open. For any > 0, there is a piecewise affine Lipschitz mapping w U →^2 such that * w - (Φ_i,t^1(P) x + b) _C^α(U, ^2) <; * w = Φ_i,t^1(P) x + b on ∂ U; * Dw(x) ∈⋃_k = i^j-1𝒲_k,t^'^1 ∪⋃_k = i^j-1𝒲_k,t^'^2 ∪𝒱_j for a.e. x ∈ U; * t/t^' |U| ≤ |{ Dw ∈𝒲_i, t^'^1 ∪𝒲_i, t^'^2 }| ≤ |U|; * (1 - 1/r^p) (1 - t/t^') ( 1/r^2 )^k-i |U| ≤ |{ Dw ∈𝒲_k, t^'^1 ∪𝒲_k, t^'^2 }| ≤ ( 1 - 1/r^2) (1 - t/t^') ( 1/r^p + 2 δ )^k-i |U| for all i < k < j; * (1 - t/t^') (1/r^2)^j-i |U| ≤ |{ Dw ∈𝒱_j}| ≤ (1 - t/t^') (1/r^p + 2 δ)^j-i |U|. Let j > i ≥ I_1, P ∈𝒫, α∈ (0,1), T_0 < t < t^' < 1, b ∈^2 and U convex and open. For any > 0, there is a piecewise affine Lipschitz mapping w U →^2 such that * w - (Φ_i,t^2(P) x + b) _C^α(U, ^2) <; * w = Φ_i,t^2(P) x + b on ∂ U; * Dw(x) ∈⋃_k = i^j-1𝒲_k,t^'^1 ∪⋃_k = i^j-1𝒲_k,t^'^2 ∪𝒱_j for a.e. x ∈ U; * t/t^' |U| ≤ |{ Dw ∈𝒲_i, t^'^1 ∪𝒲_i, t^'^2 }| ≤ |U|; * (1 - 1/r^p) (1 - t/t^') ( 1/r^2 )^k-i-1 |U| ≤ |{ Dw ∈𝒲_k, t^'^1 ∪𝒲_k, t^'^2 }| ≤ ( 1 - 1/r^2) (1 - t/t^') ( 1/r^p + 2 δ )^k-i-1 |U| for all i < k < j; * (1 - t/t^') (1/r^2)^j-i-1 |U| ≤ |{ Dw ∈𝒱_j}| ≤ (1 - t/t^') (1/r^p + 2 δ)^j-i-1 |U|. The laminate of finite order μ^1, (i,j)_t,t^'(P) from (<ref>) satisfies the property that μ^1, (i,j)_t,t^'(P) = Φ_i,t^1(P). Moreover, the following estimates follow from the definition of μ^1, (i,j)_t,t^'(P): μ^1, (i,j)_t,t^'(P)( ⋃_k = i^j-1𝒲_k,t^'^1 ∪⋃_k = i^j-1𝒲_k,t^'^2 ∪𝒱_j) = 1; t/t^'≤μ^1, (i,j)_t,t^'(P)( 𝒲_i,t^'^1 ∪𝒲_i,t^'^2 ) ≤ 1; ( 1 - 1/r^p) ( 1 - t/t^') ( 1/r^2)^k-i≤μ^1, (i,j)_t,t^'(P)( 𝒲_k,t^'^1 ∪𝒲_k,t^'^2 ) ≤( 1 - 1/r^2) ( 1 - t/t^') ( 1/r^p + 2 δ)^k-i ∀ i < k < j; ( 1 - t/t^') ( 1/r^2)^j-i≤μ^1, (i,j)_t,t^'(P) ( 𝒱_j) ≤( 1 - t/t^') ( 1/r^p + 2 δ)^j-i. Therefore, since the sets 𝒲_k,t^'^1, 𝒲_k,t^'^2 (i < k < j) and 𝒱_j are open, by applying Lemma <ref> (see also Remark <ref>), we find a piecewise affine Lipschitz mapping w U →^2 satisfying all the desired properties. This concludes the proof of Proposition <ref>. Finally, we end this subsection and section by the following observation: There is a finite constant C > 1 such that |X_11|^p + |X_22|^p < C r^p ℓ ∀ X ∈𝒰_ℓ^1 ∪𝒰_ℓ^2 ∪𝒱_ℓ∪⋃_t ∈ (T_0,1), r = 1,2𝒲_ℓ,t^r. This follows from the definition of the sets 𝒰_ℓ^1, 𝒰_ℓ^2, 𝒱_ℓ, 𝒲_ℓ, t^1 and 𝒲_ℓ, t^2 combined with direct computations. § CONVEX INTEGRATION SCHEME (PART 2 OF THE PROOF OF THEOREM <REF>) In this section, we describe the convex integration scheme and finish the proof of Theorem <ref>. More precisely, we describe the scheme in Subsection <ref>. At the end of this first subsection, we will have finished the proof of Theorem <ref>. In order to do this, we refer to several properties that we prove in the remaining subsections. §.§ Description of the Convex Integration Scheme Recall I_1 and T_0 defined in (<ref>). Select[To select such I and T_0^', it suffices to choose T_0^' and Ĩ≥ I_1 such that for all i ≥Ĩ, 𝒲_i, T_0^'^1, 𝒲_i, T_0^'^2 ⊂{ X ∈^2 × 2_ : X_22 < 0 }. This is possible because 0 < λ, Λ. Hence (<ref>) holds for any I ≥Ĩ and T_0^'. Then it only remains to select I ≥Ĩ so that (<ref>) holds.] I ≥ I_1 and 0 < T_0^' < 1 such that r^I ≥1/r^2 δ - 1; 1 < dist(⋃_j = 0^∞𝒱_I + j,⋃_j ≥ 0, t > T^'_0𝒲_I + j, t^1 ∪⋃_j ≥ 0, t > T^'_0𝒲_I + j, t^2 ). Take M ∈𝒱_I. Define v_I Ω→^2 as v_I(x) Mx. Let {δ_ℓ}_ℓ = 1^∞ be a decreasing sequence such that δ_ℓ≤ 2 δ for all ℓ and δ_ℓ→δ. Let { t_ℓ}_ℓ = 1^∞ be an increasing sequence such that (recall C_r,p defined in (<ref>)): 1 - 1/r^I+2 ℓ ≤ t_ℓ≤ 1; 2 C_r,p (t_ℓ + 1 - t_ℓ) ≤( 1/r^p + δ_ℓ + 1)^ℓ+1 - ( 1/r^p + δ_ℓ)^ℓ+1; t_1 ≥max(9/10, T_0, T_0^'). The existence of the sequences {δ_ℓ}_ℓ = 1^∞ and { t_ℓ}_ℓ = 1^∞ will be proved in Subsection <ref>. Let > 0 be arbitrary. Step 1: We describe the first step of the scheme. Let 0 < _I < 2. Since M ∈𝒱_I there is P ∈𝒫 such that M = A_I(P). By Proposition <ref>, there is a piecewise affine mapping v_I+1Ω→^2 such that * v_I+1 - v_I _C^α(Ω; ^2) < _I/2^I; * Dv_I+1(x) ∈𝒲_I,t^1 ∪𝒲_I,t^2 ∪𝒱_I+1 for a.e. x ∈Ω; * (1 - 1/r^p) |Ω| < |{ Dv_I+1∈𝒲_I, t_1^1 ∪𝒲_I, t_1^2 }| < ( 1 - 1/r^2) |Ω|; * 1/r^2 |Ω| ≤ |{ Dv_I+1∈𝒱_I+1}| ≤1/r^p + 2 δ |Ω| ≤1/r^p + δ_1 |Ω|. Let ℱ_I+1 be a collection of open convex mutually disjoint sets covering Ω up to a set of measure 0 so that v_I+1 |_U is affine for each U ∈ℱ_I+1. Without loss of generality, sup_U ∈ℱ_I+1U < 2^I+1. Select 0 < _I+1 < min(_I, 2^-(I+1)) such that Dv_I+1∗ρ__I+1 - Dv_I+1_L^1(Ω, ^2 × 2) < 2^-(I+1). Inductive step, (ℓ+1)-th step: We assume that we are given v_I + ℓΩ→^2, _I + ℓ - 1 and _I + ℓ such that * v_I+ℓ - v_I+ℓ-1_C^α(Ω; ^2) < _I+ℓ-1/2^I + ℓ-1; * Dv_I + ℓ(x) ∈⋃_j = 0^ℓ - 1𝒲_I+j,t_ℓ^1 ∪⋃_j = 0^ℓ - 1𝒲_I+j,t_ℓ^2 ∪𝒱_I+ℓ for a.e. x ∈Ω; * (1 - 1/r^p) ( 1/r^2 )^j t_j+1/t_ℓ |Ω| ≤ |{ Dv_I + ℓ∈𝒲_I+j, t_ℓ^1 ∪𝒲_I+j, t_ℓ^2 }| ≤ ( 1 - 1/r^2) ( 1/r^p + δ_ℓ )^j |Ω| for all 0 ≤ j < ℓ; * (1/r^2)^ℓ |Ω| ≤ |{ Dv_I + ℓ∈𝒱_I+ℓ}| ≤ (1/r^p + δ_ℓ)^ℓ |Ω|; * Dv_I+ℓ∗ρ__I+ℓ - Dv_I+ℓ_L^1(Ω, ^2 × 2) < 2^-(I+ℓ); * there is a collection ℱ_I+ℓ of open convex mutually disjoint sets covering Ω up to set of measure 0 such that v_I+ℓ |_U is affine for all U ∈ℱ_I+ℓ and sup_U ∈ℱ_I+ℓU < 2^-(I+ℓ). We will build v_I + ℓ + 1 by modifying v_I+ℓ on each U ∈ℱ_ℐ + ℓ. For each U ∈ℱ_I+ℓ, the construction is as follows: If Dv_I+ℓ |_U ∈𝒲_I+j,t_ℓ^1 for some 0 ≤ j < ℓ: There is P ∈𝒫 such that Dv_I+ℓ |_U = Φ_I + j, t_ℓ^1(P). By Proposition <ref>, there is a piecewise affine v_I + ℓ +1, U U →^2 such that * v_I+ℓ+1, U - v_I+ℓ_C^α(U; ^2) < _I+ℓ/2^I + ℓ + 1; * v_I + ℓ + 1, U = v_I + ℓ on ∂ U; * Dv_I+ℓ+1, U(x) ∈⋃_k = j^ℓ𝒲_I+k,t_ℓ+1^1 ∪⋃_k = j^ℓ𝒲_I+k,t_ℓ+1^2 ∪𝒱_I+ℓ+1 for a.e. x ∈Ω; * t_ℓ/t_ℓ + 1 |U| ≤ |{ Dv_I+ℓ+1, U∈𝒲_I+j, t_ℓ + 1^1 ∪𝒲_I+j, t_ℓ + 1^2 }| ≤ |U|; * ( 1 - t_ℓ/t_ℓ + 1) (1/r^2)^m (1 - 1/r^p) |U| ≤ |{ Dv_I+ℓ, U∈𝒲_I+j+m, t_ℓ + 1^1 ∪𝒲_I+j+m, t_ℓ + 1^2 }| ≤ ( 1 - t_ℓ/t_ℓ + 1) (1/r^p + 2 δ)^m ( 1 - 1/r^2) |U| for all 1 ≤ m ≤ℓ - j; * ( 1 - t_ℓ/t_ℓ + 1) (1/r^2)^ℓ + 1 - j |U| ≤ |{ Dv_I+ℓ+1, U∈𝒱_I+ℓ+1}| ≤ ( 1 - t_ℓ/t_ℓ + 1) (1/r^p + 2 δ)^ℓ + 1 - j |U|. If Dv_I+ℓ |_U ∈𝒲_I+j,t_ℓ^2 for some 0 ≤ j < ℓ: There is P ∈𝒫 such that Dv_I+ℓ |_U = Φ_I + j, t_ℓ^2(P). By Proposition <ref>, there is a piecewise affine v_I + ℓ +1, U U →^2 such that * v_I+ℓ+1, U - v_I+ℓ_C^α(U; ^2) < _I+ℓ/2^I + ℓ + 1; * v_I + ℓ + 1, U = v_I + ℓ on ∂ U; * Dv_I+ℓ+1, U(x) ∈⋃_k = j+1^ℓ𝒲_I+k,t_ℓ+1^1 ∪⋃_k = j^ℓ𝒲_I+k,t_ℓ+1^2 ∪𝒱_I+ℓ+1 for a.e. x ∈Ω; * t_ℓ/t_ℓ + 1 |U| ≤ |{ Dv_I+ℓ+1, U∈𝒲_I+j, t_ℓ + 1^1 ∪𝒲_I+j, t_ℓ + 1^2 }| ≤ |U|; * ( 1 - t_ℓ/t_ℓ + 1) (1/r^2)^m-1 (1 - 1/r^p) |U| ≤ |{ Dv_I+ℓ, U∈𝒲_I+j+m, t_ℓ +1^1 ∪𝒲_I+j+m, t_ℓ + 1^2 }| ≤ ( 1 - t_ℓ/t_ℓ + 1) (1/r^p + 2 δ)^m-1 ( 1 - 1/r^2) |U| for all 1 ≤ m ≤ℓ - j; * ( 1 - t_ℓ/t_ℓ + 1) (1/r^2)^ℓ - j |U| ≤ |{ Dv_I+ℓ+1, U∈𝒱_I+ℓ+1}| ≤ ( 1 - t_ℓ/t_ℓ + 1) (1/r^p + 2 δ)^ℓ - j |U|. If Dv_I+ℓ |_U ∈𝒱_I+ℓ: There is P ∈𝒫 such that Dv_I+ℓ |_U = A_I+ℓ(P). By Proposition <ref>, there is a piecewise affine v_I + ℓ +1, U U →^2 such that * v_I+ℓ+1, U - v_I+ℓ_C^α(U; ^2) < _I+ℓ/2^I + ℓ + 1; * v_I + ℓ + 1, U = v_I + ℓ on ∂ U; * Dv_I+ℓ+1, U(x) ∈𝒲_I+ℓ, t_ℓ+1^1 ∪𝒲_I+ℓ,t_ℓ+1^2 ∪𝒱_I+ℓ+1 for a.e. x ∈Ω; * (1 - 1/r^p) |U| ≤ |{ Dv_I+ℓ+1, U∈𝒲_I+ℓ, t_ℓ + 1^1 ∪𝒲_I+ℓ, t_ℓ + 1^2 }| ≤ ( 1 - 1/r^2) |U|; * 1/r^2 |U| ≤ |{ Dv_I+ℓ+1, U∈𝒱_I+ℓ+1}| ≤1/r^p + 2 δ |U|. Now, we can define v_I + ℓ + 1. We define the function v_I + ℓ + 1Ω→^2 as v_I + ℓ + 1(x) = v_I + ℓ + 1, U(x) if x ∈ U for all U ∈ℱ_I + ℓ. Now we estimate the following quantities (for which we also introduce new notation): * 𝒬^Step (ℓ+1)_I+j |{ Dv_I+ℓ+1∈𝒲_I+j, t_ℓ+1^1 ∪𝒲_I+j, t_ℓ+1^2 }| for all 0 ≤ j ≤ℓ; * 𝒬^Step (ℓ+1) |{ Dv_I+ℓ+1∈𝒱_I+ℓ+1}|. For 0 ≤ j < ℓ, we find 𝒬^Step (ℓ+1)_I+j<ref>, <ref>, <ref>, <ref> ≤∑_k = 0^j-1( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + 2 δ)^j-k( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 }| + ∑_k = 0^j-1( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + 2 δ)^j-k-1( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^2 }| + |{ Dv_I+ℓ∈𝒲_I+j, t_ℓ^1 }| + |{ Dv_I+ℓ∈𝒲_I+j, t_ℓ^2 }| ≤( 1 - 1/r^2) ( 1 - t_ℓ/t_ℓ + 1) ∑_k = 0^j-1( 1/r^p + δ_1)^j-k-1 |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 ∪𝒲_I+k, t_ℓ^2 }| + |{ Dv_I+ℓ∈𝒲_I+j, t_ℓ^1 ∪𝒲_I+j, t_ℓ^2 }| <ref> ≤( 1 - 1/r^2)^2 ( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + δ_1)^j-1 |Ω| ∑_k = 0^j-1( r^p + δ_1/r^p + δ_ℓ)^k + ( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^j |Ω| ≤( 1 - 1/r^2)^2 (1 - t_ℓ/t_ℓ + 1) j (r^(δ_1 - δ_ℓ)/r^p + δ_1)^j-1 |Ω| + ( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^j |Ω| (<ref>), (<ref>) ≤( 1 - 1/r^2) ( 2 C_r,p( 1 - 1/r^2) (t_ℓ+1 - t_ℓ) + ( 1/r^p + δ_ℓ)^j ) |Ω| ≤( 1 - 1/r^2) ( 2 C_r,p (t_ℓ+1 - t_ℓ) + ( 1/r^p + δ_ℓ)^j ) |Ω| (<ref>)≤( 1 - 1/r^2) ( 1/r^p + δ_ℓ + 1)^j |Ω|; 𝒬^Step (ℓ+1)_I+j<ref>, <ref> ≥t_ℓ/t_ℓ + 1 |{ Dv_I+ℓ∈𝒲_I+j, t_ℓ^1 ∪𝒲_I+j, t_ℓ^2 }| <ref> ≥t_ℓ/t_ℓ + 1( 1 - 1/r^p) ( 1/r^2)^j t_j+1/t_ℓ |Ω| ≥( 1 - 1/r^p) ( 1/r^2)^j t_j+1/t_ℓ+1 |Ω| and 𝒬^Step (ℓ + 1)_I+ℓ<ref>, <ref>, <ref> ≤∑_k = 0^ℓ-1( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + 2 δ)^ℓ-k( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 }| + ∑_k = 0^ℓ-1(1 - t_ℓ/t_ℓ + 1) ( 1/r^p + 2 δ)^ℓ-k-1( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^2 }| + ( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒱_I+ℓ}| ≤( 1 - 1/r^2) ( 1 - t_ℓ/t_ℓ + 1) ∑_k = 0^ℓ-1( 1/r^p + δ_1)^ℓ-k-1 |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 ∪𝒲_I+k, t_ℓ^2 }| + ( 1 - 1/r^2) |{ Dv_I+ℓ∈𝒱_I+ℓ}| <ref>, <ref> ≤( 1 - 1/r^2)^2 ( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + δ_1)^ℓ-1 |Ω| ∑_k = 0^ℓ-1( r^p + δ_1/r^p + δ_ℓ)^k + ( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^ℓ |Ω| ≤( 1 - 1/r^2)^2 ( 1 - t_ℓ/t_ℓ + 1) ℓ( r^(δ_1 - δ_ℓ)/r^p + δ_1)^ℓ - 1 |Ω| + ( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^ℓ |Ω| (<ref>)(<ref>) ≤( 1 - 1/r^2) ( 2 C_r,p( 1 - 1/r^2) (t_ℓ + 1 - t_ℓ) + ( 1/r^p + δ_ℓ)^ℓ) |Ω| ≤( 1 - 1/r^2) ( 2 C_r,p (t_ℓ + 1 - t_ℓ) + ( 1/r^p + δ_ℓ)^ℓ) |Ω| (<ref>)≤( 1 - 1/r^2) (1/r^p + δ_ℓ + 1)^ℓ |Ω|; 𝒬^Step (ℓ + 1)_I+ℓ<ref> ≥( 1 - 1/r^p) |{ Dv_I+ℓ∈𝒱_I+ℓ}| <ref>≥( 1 - 1/r^p) ( 1/r^2)^ℓ |Ω|; 𝒬^Step (ℓ+1)<ref>, <ref>, <ref> ≤∑_k = 0^ℓ-1( 1 - t_ℓ/t_ℓ + 1) ( 1/r^p + 2 δ)^ℓ + 1 - k |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 }| + ∑_k = 0^ℓ-1( 1 - t_ℓ/t_ℓ + 1) (1/r^p + 2 δ)^ℓ - k |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^2 }| + 1/r^p + 2 δ |{ Dv_I+ℓ∈𝒱_I+ℓ}| ≤( 1 - t_ℓ/t_ℓ + 1) ∑_k = 0^ℓ-1( 1/r^p + δ_1)^ℓ - k |{ Dv_I+ℓ∈𝒲_I+k, t_ℓ^1 ∪𝒲_I+k, t_ℓ^2 }| + 1/r^p + δ_1 |{ Dv_I+ℓ∈𝒱_I+ℓ}| <ref>, <ref> ≤( 1 - 1/r^2) (1 - t_ℓ/t_ℓ+1) (1/r^p + δ_1)^ℓ∑_k = 0^ℓ-1(r^p + δ_1/r^p + δ_ℓ)^k |Ω| + (1/r^p + δ_ℓ)^ℓ+1 |Ω| ≤( 1 - 1/r^2) (1 - t_ℓ/t_ℓ+1) ℓ(r^δ_1 - δ_ℓ/r^p + δ_1)^ℓ-1 |Ω| + (1/r^p + δ_ℓ)^ℓ+1 |Ω| (<ref>), (<ref>) ≤( 2 C_r,p (t_ℓ + 1 - t_ℓ) + (1/r^p + δ_ℓ)^ℓ+1) |Ω|(<ref>)≤( 1/r^p + δ_ℓ+1)^ℓ+1 |Ω|; 𝒬^Step (ℓ+1)<ref>, <ref>, <ref> ≥1/r^2 |{ Dv_I+ℓ∈𝒱_I+ℓ}| <ref>≥( 1/r^2)^ℓ+1 |Ω|. Select 0 < _I + ℓ + 1 < min(_I + ℓ, 2^-(I + ℓ + 1)) so that D v_I + ℓ + 1∗ρ__I + ℓ + 1 - D v_I + ℓ + 1_L^1(Ω; ^2 × 2) < 2^-(I + ℓ + 1). Let ℱ_I + ℓ + 1 be a collection of open convex mutually disjoint sets covering Ω up to a set of measure 0 such that v_I + ℓ + 1 |_U is affine for each U ∈ℱ_I + ℓ + 1. This ends the (ℓ + 1)-th step. Therefore, by induction, we have built a sequence { v_I + ℓ}_ℓ = 1^∞ such that for all ℓ≥ 1: v_I+ℓ - v_I + ℓ -1 _C^α(Ω; ^2) < _I + ℓ/2^I + ℓ; Dv_I + ℓ(x) ∈⋃_j = 0^ℓ - 1𝒲_I+j,t_ℓ^1 ∪⋃_j = 0^ℓ - 1𝒲_I+j,t_ℓ^2 ∪𝒱_I+ℓ for a.e. x ∈Ω; ( 1 - 1/r^p) ( 1/r^2)^j t_j+1/t_ℓ |Ω| ≤ |{ Dv_I + ℓ∈𝒲_I+j, t_ℓ^1 ∪𝒲_I+j, t_ℓ^2 }| ≤( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^j |Ω|, ∀ 0 ≤ j < ℓ; ( 1/r^2)^ℓ |Ω| ≤ |{ Dv_I + ℓ∈𝒱_I+ℓ}| ≤( 1/r^p + δ_ℓ)^ℓ |Ω| ; D v_I + ℓ∗ρ__I + ℓ - D v_I + ℓ_L^1(Ω; ^2 × 2) < 2^-(I + ℓ); and there is a collection ℱ_I+ℓ of open convex mutually disjoint sets covering Ω up to set of measure 0 such that v_I+ℓ |_U is affine for all U ∈ℱ_I+ℓ and sup_U ∈ℱ_I+ℓU < 2^-(I+ℓ). Due to computations that we will carry out in Subsection <ref>, we find that there is some p > 1 such that sup_ℓ≥ 1 v_I + ℓ_W^1,p(Ω; ^2) < ∞. Therefore there exists a function v ∈ W^1,p(Ω; ^2) such that v_I + ℓ converges weakly to v in W^1,p(Ω; ^2) (no need to pass to subsequences since v ∈ C^α(Ω; ^2) and v_I + ℓ→ v in C^α(Ω; ^2)). In Subsection <ref>, we prove that Dv_I + ℓ→ Dv in L^1(Ω; ^2 × 2) as ℓ→∞. Then we show that Dv(x) ∈ K_f for a.e. x ∈Ω in Subsection <ref>. This proves, in virtue of the arguments from Subsection <ref>, that the first component of v, denoted by v^1, solves equation (<ref>). In Subsection <ref>, we show that for any ball B ⊂Ω we have ∫_B |∂_1 v^1|^2 dx = + ∞. In Subsection <ref>, we explain why our scheme yields infinitely many such solutions to (<ref>) with the same boundary datum. This ends the proof of Theorem <ref>. §.§ Proof of existence of the sequences {δ_ℓ}_ℓ = 1^∞ and { t_ℓ}_ℓ = 1^∞ Define the sequence {δ_ℓ}_ℓ = 1^∞ by δ_ℓδ + 1/ℓ^2δ = ( 1 + 1/ℓ^2) δ By standard computations, δ_ℓ - δ_ℓ + 1≥2/(ℓ+1)^3δ. This implies that (δ_ℓ - δ_ℓ + 1) (ℓ + 1)^3 ≥ 2 δ. Define the sequence { t_ℓ}_ℓ = 1^∞ by t_ℓ 1 - 1/r^N + I + 2p(ℓ + 1)^3. where N ≥log_r (2 C_r,p) is selected sufficiently large so that (<ref>) holds. The fact that (<ref>) hold is immediate. It remains to prove (<ref>). We have t_ℓ + 1 - t_ℓ≤1/r^N + I + 2p(ℓ + 1)^3 and ( 1/r^p + δ_ℓ + 1)^(ℓ + 1)^3 - ( 1/r^p + δ_ℓ)^(ℓ + 1)^3 = (r^p + δ_ℓ)^(ℓ + 1)^3 - (r^p + δ_ℓ + 1)^(ℓ + 1)^3/(r^2p + δ_ℓ + δ_ℓ + 1)^(ℓ + 1)^3 = 1/r^p(ℓ + 1)^3r^(δ_ℓ - δ_ℓ + 1)(ℓ + 1)^3 - 1/r^δ_ℓ (ℓ + 1)^3 ≥1/r^2 p (ℓ + 1)^3 (r^(δ_ℓ - δ_ℓ + 1)(ℓ + 1)^3 - 1) (<ref>)≥1/r^2 p (ℓ + 1)^3 (r^2 δ - 1) (<ref>) ≥1/r^I + 2 p (ℓ + 1)^3 = r^N 1/r^N + I + 2 p (ℓ + 1)^3(<ref>)≥ 2 C_r,p (t_ℓ + 1 - t_ℓ). Therefore, 2 C_r,p (t_ℓ + 1 - t_ℓ) ≤( 1/r^p + δ_ℓ + 1)^(ℓ + 1)^3 - ( 1/r^p + δ_ℓ)^(ℓ + 1)^3≤( 1/r^p + δ_ℓ + 1)^(ℓ + 1) - ( 1/r^p + δ_ℓ)^(ℓ + 1) for all ℓ≥ 1. This proves (<ref>). §.§ Proof that sup_ℓ≥ 1 v_I + ℓ_W^1,p(Ω; ^2) < ∞ First we notice that sup_ℓ≥ 1∂_2 v_I + ℓ^1 _L^∞(Ω; ^2 × 2) , sup_ℓ≥ 1∂_1 v_I + ℓ^2 _L^∞(Ω; ^2 × 2)< 1. In addition, we have (with C given by (<ref>)) ∫_Ω |∂_1 v_I + ℓ^1|^p + |∂_2 v_I + ℓ^2|^p dx = ∑_j = 0^ℓ - 1∫_{ Dv_I + ℓ∈𝒲_I+j, t_ℓ^1 }∪{ Dv_I + ℓ∈𝒲_I+j, t_ℓ^2 } |∂_1 v_I + ℓ^1|^p + |∂_2 v_I + ℓ^2|^p dx + ∫_{ Dv_I + ℓ∈𝒱_I + ℓ} |∂_1 v_I + ℓ^1|^p + |∂_2 v_I + ℓ^2|^p dx (<ref>) ≤ C r^p I∑_j = 0^ℓ - 1 r^pj |{ Dv_I + ℓ∈𝒲_I+j, t_ℓ^1 ∪𝒲_I+j, t_ℓ^2 }| + Cr^p (I + ℓ) |{ Dv_I + ℓ∈𝒱_I+ℓ}| (<ref>), (<ref>) ≤ C r^p I |Ω| ∑_j = 0^ℓ - 1 r^pj( 1 - 1/r^2) ( 1/r^p + δ_ℓ)^j + Cr^p ( I + ℓ) |Ω| ( 1/r^p + δ_ℓ)^ℓ≤ C r^p I |Ω| ∑_j = 0^∞( r^p/r^p + δ)^j < ∞. Therefore, since the last term does not depend on ℓ, we conclude that sup_ℓ≥ 1 v_I + ℓ_W^1,p(Ω; ^2) < ∞. §.§ Proof that Dv_I + ℓ→ Dv in L^1(Ω; ^2 × 2) as ℓ→∞ We have, by (<ref>), Dv_I + ℓ - Dv _L^1(Ω; ^2 × 2) ≤ Dv_I + ℓ - Dv_I + ℓ∗ρ__I + ℓ_L^1(Ω; ^2 × 2)_≤ 2^-(I + ℓ) + Dv_I + ℓ∗ρ__I + ℓ - Dv ∗ρ__I + ℓ_L^1(Ω; ^2 × 2) + Dv ∗ρ__I + ℓ - Dv _L^1(Ω; ^2 × 2)._→ 0 as ℓ→∞ The first and third term converge to 0 as ℓ→∞. Therefore, it suffices to show that the second term converges to 0 as ℓ→∞. Inside Ω (recall the conventions fixed in Subsection <ref> and that v_I + ℓ has affine boundary datum), we have (Dv_I + ℓ - Dv)∗ρ__I + ℓ = (v_I + ℓ - v) ∗ Dρ__I + ℓ. We have (v_I + ℓ - v) ∗ Dρ__I + ℓ_L^1(Ω; ^2 × 2)≤ D ρ_L^1(B_1; ^2)/_I + ℓ v_I + ℓ - v _L^1(Ω; ^2)≤ D ρ_L^1(B_1; ^2) |Ω|/_I + ℓ v_I + ℓ - v _L^∞(Ω; ^2). In addition, v_I + ℓ - v _L^∞(Ω; ^2)≤∑_j = I + ℓ^∞ v_j+1 - v_j_L^∞(Ω; ^2)(<ref>)≤∑_j = I + ℓ^∞_I+j/2^I+j≤_I+ℓ/2^I+ℓ-1. Thus the second term in (<ref>) converges to 0 as ℓ→∞. From this, we deduce that Dv_I + ℓ→ Dv in L^1(Ω; ^2 × 2) as ℓ→∞. §.§ Proof that Dv(x) ∈ K_f for a.e. x ∈Ω We begin by proving that Dv(x) ∈⋃_j = 0^∞𝒰_I+j^1 ∪⋃_j = 0^∞𝒰_I+j^2 ∪⋃_j = 0^∞𝒱_I+j_ =: K̃_f for a.e. x ∈Ω. From (<ref>), we find that XK̃_f≤1/2^ℓ ∀ X ∈⋃_j = 0^ℓ - 1𝒲_I+j, t_ℓ^1 ∪⋃_j = 0^ℓ - 1𝒲_I+j, t_ℓ^2 ∪𝒱_I+ℓ. The fact that Dv_I + ℓ(x) ∈⋃_j = 0^ℓ - 1𝒲_I+j, t_ℓ^1 ∪⋃_j = 0^ℓ - 1𝒲_I+j, t_ℓ^2 ∪𝒱_I+ℓ for a.e. x ∈Ω, combined with (<ref>) and the fact that Dv_I+ℓ→ Dv in L^1(Ω; ^2 × 2) allows us to conclude that Dv(x) ∈K̃_f for a.e. x ∈Ω. It remains to prove that | { Dv(x) ∈⋃_j = 0^∞𝒱_I + j}| = 0. First we point out that since Dv_I+ℓ→ Dv in L^1(Ω; ^2 × 2), Dv_I+ℓ→ Dv in measure. Notice that { Dv(x) ∈⋃_j = 0^∞𝒱_I + j} ⊂{ Dv(x) ∈⋃_j = 0^∞𝒱_I + j, |Dv_I+ℓ - Dv| > 1 } ∪{ Dv(x) ∈⋃_j = 0^∞𝒱_I + j, |Dv_I+ℓ - Dv| ≤ 1 } (<ref>) ⊂{ |Dv_I+ℓ - Dv| > 1 }∪{ Dv_I + ℓ∈𝒱_I + ℓ, }. Therefore | { Dv(x) ∈⋃_j = 0^∞𝒱_I + j}| ≤| { |Dv_I+ℓ - Dv| > 1 }| + | { Dv_I + ℓ∈𝒱_I + ℓ, }| (<ref>) ≤| { |Dv_I+ℓ - Dv| > 1 }| + ( 1/r^p + δ)^ℓ |Ω|. Since Dv_I+ℓ→ Dv in measure, the first term converges to 0 as ℓ→∞. Since the second term converges to 0 as well, we find (<ref>), as wished. From (<ref>) we conclude that Dv(x) ∈⋃_j = 0^∞𝒰_I+j^1 ∪⋃_j = 0^∞𝒰_I+j^2 ⊂ K_f for a.e. x ∈Ω. §.§ Proof that ∫_B |∂_1 v^1|^2 dx = + ∞ for any ball B ⊂Ω Let B be an open ball of arbitrary radius. Let B^'⊂ B be an arbitrary ball which is a subset of B. Since sup_U ∈ℱ_I+ℓdiam(U) < 2^- (I + ℓ) for all ℓ≥ 1, for ℓ large enough there is an open set U ∈ℱ_I+ℓ so that U ⊂ B^'. By construction, there is a subset Ũ∈ℱ_I + ℓ + 1, Ũ⊂ U so that v_I + ℓ + 1|_Ũ is affine and Dv_I + ℓ + 1|_Ũ∈𝒱_I + ℓ + 1. By construction, there exist _Ũ > 0 and ^'_Ũ > 0 such that |{ Dv_I + ℓ + 3 |_Ũ∈𝒲_I + ℓ + 1, t_ℓ + 3^1 ∪𝒲_I + ℓ + 1, t_ℓ + 3^2 }| > _Ũ |Ũ|; |{ Dv_I + ℓ + 3 |_Ũ∈𝒲_I + ℓ + 2, t_ℓ + 3^1 ∪𝒲_I + ℓ + 2, t_ℓ + 3^2 }| > _Ũ^' |Ũ|. By <ref> and <ref>, we obtain |{ Dv_I + ℓ + 4 |_Ũ∈𝒲_I + ℓ + 1, t_ℓ + 4^1 ∪𝒲_I + ℓ + 1, t_ℓ + 4^2 }| > t_ℓ + 3/t_ℓ + 4_Ũ |Ũ|; |{ Dv_I + ℓ + 4 |_Ũ∈𝒲_I + ℓ + 2, t_ℓ + 4^1 ∪𝒲_I + ℓ + 2, t_ℓ + 4^2 }| > t_ℓ + 3/t_ℓ + 4_Ũ^' |Ũ|. Letting r ≥ 5 be an arbitrary integer and applying <ref> and <ref> (r - 4) times more yield, thanks to (<ref>) |{ Dv_I + ℓ + r |_Ũ∈𝒲_I + ℓ + 1, t_ℓ + r^1 ∪𝒲_I + ℓ + 1, t_ℓ + r^2 }| > t_ℓ + 3/t_ℓ + r_Ũ |Ũ| > 9/10_Ũ |Ũ|; |{ Dv_I + ℓ + r |_Ũ∈𝒲_I + ℓ + 2, t_ℓ + r^1 ∪𝒲_I + ℓ + 2, t_ℓ + r^2 }| > t_ℓ + 3/t_ℓ + r_Ũ^' |Ũ| > 9/10_Ũ^' |Ũ|. Due to the fact that Dv_I + ℓ + r→ Dv in L^1(Ω; ^2 × 2) as r →∞, this implies that |{ Dv |_Ũ∈𝒰_I + ℓ + 1^1 ∪𝒰_I + ℓ + 1^2 }| ≥9/10_Ũ |Ũ| and |{ Dv |_Ũ∈𝒰_I + ℓ + 2^1 ∪𝒰_I + ℓ + 2^2 }| ≥9/10_Ũ^' |Ũ|. Thus we obtain, by assuming ℓ sufficiently large sup_x,y ∈ B^' |∂_1 v^1(x) - ∂_1 v^1(y)| > 1. Since this last inequality holds for all B^'⊂ B, v^1 is not C^1 on B. Since f is smooth, due to the results by De Giorgi and Nash <cit.> v^1 |_B ∉W^1,2(B). Indeed, otherwise, if v^1 |_B ∈ W^1,2(B), then by these results we deduce that v^1 ∈ C^∞(B). This is a contradiction. §.§ Why does this convex integration scheme yield infinitely many solutions with identical boundary datum? To see that the convex integration scheme yields infinitely many solutions, all having the same boundary datum, we first observe that with v given by the scheme above v - v_I _C^α(Ω; ^2)≤∑_j > I v_j - v_j-1_C^α(Ω; ^2)(<ref>)≤∑_j > I_j/2^j < _I ∑_j > I1/2^j < _I < . Keep in mind that > 0 can be chosen arbitrarily at the beginning of the convex integration scheme. Assume for a contradiction that it is only possible to find finitely many v^(i) (i = 1 , …, M) such that Dv^(i)∈ K_f a.e. Then select (recall that v_I(x) = Mx) min_i = 1, …, M v^(i) - v_I _C^α(Ω; ^2)/2. By the convex integration scheme, we can construct a function ṽ∈ W^1,p(Ω; ^2) ∩ C^α(Ω; ^2) such that ṽ - v_I _C^α(Ω; ^2) < and D ṽ∈ K_f a.e.. This is a contradiction. This proves that there are infinitely many solutions to (<ref>), all with identical boundary datum. plain

http://arxiv.org/abs/2303.07016v1

20230313112532

HOOV: Hand Out-Of-View Tracking for Proprioceptive Interaction using Inertial Sensing

cs.HC

HOOV: Hand Out-Of-View Tracking for Proprioceptive Interaction using Inertial Sensing]HOOV: Hand Out-Of-View Tracking for Proprioceptive Interaction using Inertial Sensing Department of Computer Science ETH Zürich Switzerland Department of Computer Science ETH Zürich Switzerland Department of Computer Science ETH Zürich Switzerland 0000-0001-9655-9519 Department of Computer Science ETH Zürich Switzerland Current Virtual Reality systems are designed for interaction under visual control. Using built-in cameras, headsets track the user's hands or hand-held controllers while they are inside the field of view. Current systems thus ignore the user's interaction with off-screen content—virtual objects that the user could quickly access through proprioception without requiring laborious head motions to bring them into focus. In this paper, we present HOOV, a wrist-worn sensing method that allows VR users to interact with objects outside their field of view. Based on the signals of a single wrist-worn inertial sensor, HOOV continuously estimates the user's hand position in 3-space to complement the headset's tracking as the hands leave the tracking range. Our novel data-driven method predicts hand positions and trajectories from just the continuous estimation of hand orientation, which by itself is stable based solely on inertial observations. Our inertial sensing simultaneously detects finger pinching to register off-screen selection events, confirms them using a haptic actuator inside our wrist device, and thus allows users to select, grab, and drop virtual content. We compared HOOV's performance with a camera-based optical motion capture system in two folds. In the first evaluation, participants interacted based on tracking information from the motion capture system to assess the accuracy of their proprioceptive input, whereas in the second, they interacted based on HOOV's real-time estimations. We found that HOOV's target-agnostic estimations had a mean tracking error of 7.7 cm, which allowed participants to reliably access virtual objects around their body without first bringing them into focus. We demonstrate several applications that leverage the larger input space HOOV opens up for quick proprioceptive interaction, and conclude by discussing the potential of our technique. <ccs2012> <concept> <concept_id>10003120.10003121.10003124.10010866</concept_id> <concept_desc>Human-centered computing Virtual reality</concept_desc> <concept_significance>500</concept_significance> </concept> <concept> <concept_id>10003120.10003121.10003125</concept_id> <concept_desc>Human-centered computing Interaction devices</concept_desc> <concept_significance>300</concept_significance> </concept> <concept> <concept_id>10010147.10010257.10010293.10010294</concept_id> <concept_desc>Computing methodologies Neural networks</concept_desc> <concept_significance>100</concept_significance> </concept> <concept> <concept_id>10010147.10010178.10010224.10010245.10010253</concept_id> <concept_desc>Computing methodologies Tracking</concept_desc> <concept_significance>500</concept_significance> </concept> <concept> <concept_id>10010147.10010178.10010224.10010226.10010238</concept_id> <concept_desc>Computing methodologies Motion capture</concept_desc> <concept_significance>100</concept_significance> </concept> </ccs2012> [500]Human-centered computing Virtual reality [300]Human-centered computing Interaction devices [100]Computing methodologies Neural networks [500]Computing methodologies Tracking [100]Computing methodologies Motion capture < g r a p h i c s > is a wireless sensing method that complements existing virtual and augmented reality headsets to support hand tracking outside the field of view of the headset's cameras. In this example app of our system, a participant is building a structure composed of blocks with different sizes. When the user places an object onto the structure, his hand is tracked by the headset's cameras (left). As the user reaches for a block just to his right (right), his hand leaves the field of view of the cameras. Now, continuously estimates the 6D position and orientation of the wrist using the 6-axis inertial measuring unit attached to the wrist during the time the hand stays outside the field of view. The figure consists of two pictures of a user wearing an Oculus Quest 2 headset with HOOV attached to his right hand's wrist. In the left picture, the user interacts with building blocks in a tower structure within the field of view of the headset's cameras. In the right picture, the user grabs for another building block outside the cameras' field of view. [ Christian Holz March 30, 2023 ================== § INTRODUCTION Mixed Reality systems are increasingly designed to be standalone, both for Augmented Reality (AR) as well as Virtual Reality (VR) scenarios. They typically integrate all necessary components for an immersive virtual experience in a user-worn headset. These include a stereoscopic display, stereo sound, processors for rendering, as well as sensors such as inertial measurement units (IMUs) and cameras for inside-out head-pose tracking in world space. Standalone systems are highly portable, allowing them to operate in mobile scenarios, but requiring them to make several trade-offs in their design. Because their provided comfort strongly affects the user experience, size, form factor, and weight are the limiting factors in the implementation. These considerations ultimately limit the available computing power, battery capacity, and display size, which confines the visible field of view (FOV) for the wearer. They also affect the number, type, and placement of integrated sensors, in particular the cameras that track the user's surroundings and, thus, the effective tracking FOV. In addition to scene perception, these embedded cameras increasingly deliver the signal for detecting user input through hand gestures and interaction with virtual and physical objects. Therefore, the effective tracking FOV also determines the operational range for hand input on these devices. In practice, today's headsets typically cover an operational range that slightly exceeds the user's visible FOV, thereby requiring all hand-object interaction to happen under visual control. While the operational range could be increased with additional headset cameras, which would incur additional needs for compute, power, and physical space inside the headset, certain areas around the body might still be outside the line of sight due to occlusion caused by clothing, hair, or other parts of the user's body. As a result, today's systems implicitly require all interaction to occur in the visual FOV, which neglects most of the available space around the user for input. We argue that this is a missed opportunity, as humans perceive more than 210 <cit.>, allowing us to perform hand-object interactions even at the edge of our periphery. Even without visual control, we routinely rely on proprioception to quickly place and retrieve physical objects in our vicinity <cit.>—without requiring a turn of the head, which would slow down such motions. Since such short-term use cases do not justify the cost of additional cameras, off-screen interaction is not part of the interaction vocabulary on today's headsets. And with mixed reality devices' ever-decreasing form factor, we do not expect future devices to substantially widen the tracking FOV. In this paper, we introduce , a method to track interaction outside a headset's tracking field-of-view in immersive environments. Using our novel data-driven inertial estimation pipeline, complements current headsets by leveraging the continuous signals from a 6-axis IMU inside a wrist-worn band to estimate the user's current hand position outside the headset's tracking FOV. §.§ Outside field-of-view interaction <ref> shows an example application of our method. Here, a user is building a structure composed of blocks with different sizes, placed by his side for easy access. The user can pick up a building block by pinching it, dragging it to the tower, and dropping it at the desired location and with the hand-held orientation. The building blocks by his side replenish, such that grabbing one immediately produces another one at the same location. Having built up the tower so far, the user has internalized the location of building blocks by his side by now. Naturally, as time progresses, he has to rely less and less on visual operation and, instead, simply reaches for one of the three locations to grab the corresponding piece. This aptly supports his construction process as the complexity of the structure advances, allowing him to keep his visual attention fixated on the tower and the spot where he plans to place the next block, while reaching out to grab it. The underlying implementation of this is powered by our method . When the user's hand leaves the headset's tracking FOV, takes over the tracking and continuously provides the hand's current 3D position to the application. For this, processes the signals from the 3-axis accelerometer and 3-axis gyroscope sensor integrated inside the wrist-worn band as input. From these, estimates the wrist's current 3D orientation and feeds it into our novel temporal machine learning model, which estimates the current hand position. To register input commands from the user, leverages the same IMU to detect hand gestures. In <ref>, the user pinches to grab the intended building block, which recognizes as a performed gesture in the signal stream. 's pinch detection complements the interaction paradigm on today's VR platforms, in which pinching is commonly used either for direct or remote selection through a ray <cit.>. §.§.§ Performance Evaluation We conducted two studies to evaluate 's performance during outside field-of-view interaction in stationary scenarios. Both studies consisted of three tasks each. In the first task, participants grabbed objects outside their FOV while facing forward, resembling our examples in <ref> but scaled up to 17 potential targets. In the second task, participants placed virtual spheres into 1 of 17 discrete drop zones by their side. In a final compound task, participants repeatedly switched between retrieving and placing objects outside their FOV, which constituted a compound and, thus, longer task. In , participants' wrists were tracked by an 8-camera OptiTrack Prime 13 system, permitting an analysis of the human ability to interact based on proprioception only, under nearly optimal tracking conditions. Participants achieved a success rate of ∼91% when grabbing and dropping objects off-screen. In , participants were operating exclusively using 's tracking when their hand left the headset's view. They were successful in 86.18%, 87.65% and 86.18% of cases for the grab, placement, and compound task respectively. When considered as a target-agnostic 3D tracking system, 's simulated position estimates form a 3D trajectory that has a mean absolute error of 7.77 cm from the OptiTrack's reference path in . We observed that 's error was lowest within the first three seconds, covering the duration of most interactions outside the user's field of view. In our simulations of lower tracking FOV in the headset (i.e., earlier moments at which the headset hands off tracking to ), we found that 's tracking error slightly increased, with a mean absolute error of 9.28 cm in a simulated 120 tracking FOV and a 10.16 cm error in a simulated 90 tracking FOV. In , the average error remained below 16 cm during real-time interactive use over longer off-screen trajectories. In its current implementation, 's inference pipeline runs in real-time on a desktop machine with an NVIDIA GeForce RTX 3090 GPU, requiring 4 ms to predict a 3D hand pose, which amounts to a maximum update frequency of 250 Hz. In both studies, participants also completed a separate condition of the same tasks under visual control. In this condition, the hand was only tracked by the headset, in our case an Oculus Quest 2, requiring participants to turn their head to keep the hand inside the headset's field of view. While the average success rate rose to 95% across all tasks in , the average task completion time also increased, most notably by the compound task. Due to the monotony and repetitiveness of the task, most participants pointed out the unfavorable need for turning their heads in this condition and the strain this puts on their neck over time, preferring interaction through over interaction under visual control. §.§.§ Applications enables several scenarios that are directly applicable and useful for immersive systems. We show this at the example of three demo applications that benefit from the knowledge of the current hand position beyond the trackable FOV of the headset's cameras. In the first application, users can interact with building blocks outside their FOV as illustrated in <ref> and <ref>. In the second scenario, improves a Beat Saber gaming experience by tracking quick arm movements that cross hardly tracked areas around the user's body. Finally, facilitates tracked arm movements of extended range as demonstrated in our third application, an archery game. We conclude this paper with a discussion of our applications and the implications of our method for future immersive systems. §.§ Contributions In this paper, we contribute * a tracking method that enables short-term interaction beyond the FOV of the cameras inside AR/VR headsets. Our method merely relies on the signals from a 6-axis IMU, integrated into our custom inertial sensing device placed on the user's wrist, which captures the information for outside-field-of-view 6D hand tracking and pinch detection and provides haptic feedback during interaction events. * a novel temporal deep learning architecture that comprises a Transformer for the estimation of the current hand position from the IMU signals and the latest available output of the headset's visual hand tracking pipeline. * a user study with 12 participants to investigate human proprioceptive abilities and 's tracking performance, where participants retrieved and placed spheres at 17 different target locations outside the visual and the camera's field of view. Using an 8-camera OptiTrack system, participants grabbed and placed the correct object ∼91% of the time. In an offline simulation, they would have grabbed and placed the correct object in ∼85% of all cases using . Compared to a baseline with a commodity headset where participants grabbed and placed objects under visual control, their success rate increased to ∼94% at the expense of 20% slower task completion and repeated neck movement. * a user study with 10 participants where participants completed the selection and placement tasks using in real-time, achieving success rates of more than 86%. § RELATED WORK is related to body capture using worn sensors, inertial tracking, and interactions driven by spatial memory, proprioception, and kinesthesia. §.§ Body capture using worn sensors Information about the user's body pose, especially the posture of the arm, enables VR and AR systems to correctly embody the user and to detect input through gestures. In contrast to external body pose capture systems that include high-end commercial marker-based camera systems <cit.>, depth and RGB <cit.> cameras as well as non-optical systems <cit.>, body-worn systems enable tracking in mobile settings, and do not require the user to remain within a constraint area. To receive a wider view of the user's body for tracking, cameras with wide angle-lenses were attached to various parts of the body including the wrist <cit.>, the feet <cit.>, and the chest <cit.> or alternatively integrated into controllers <cit.> or using a suspension on the headset <cit.> further away from the user's body. However, these solutions tend to suffer from occlusion and are often obtrusive to wear. Alternative approaches directly estimate the full-body pose based solely on the available temporal motion information of the user's head and hands <cit.>. Moreover, various specialized mobile hand-held or body-worn systems have been built for tracking that make use of sensing modalities such as magnetic <cit.>, mechanical <cit.>, or acoustic <cit.> sensing. Prior research has also shown that body-attached IMUs can be used in a standalone system for tracking human pose. Depending on the number of body locations that are instrumented, the problem varies in difficulty. Sparse Inertial Poser <cit.> estimates the 3D human pose based on 6 IMU sensors attached to the wrists, lower legs, the back and the head using a joint optimization framework incorporating anthropometric constraints. Deep Inertial poser <cit.> trains a recurrent neural network (RNN) for this task. Transpose <cit.> and PIP <cit.> further improve on the predicted output by incorporating joint-specific loss terms and physic-aware motion optimizers. §.§ IMU-based arm and wrist tracking Instead of tracking the full body, attached three IMU sensors to the hand, forearm and upperarm to track a user's arm pose through solving inverse kinematics <cit.>. Closely related to our work, presented an online and an offline tracking method to estimate the posture of the arm based on a single wrist-worn 9-axis IMU integrated within a smartwatch <cit.>. The method finds an optimal path through a set of potential candidate arm postures that are retrieved from a dictionary based on the IMU-estimated wrist orientation at each time step. The offline version uses Viterbi decoding to estimate an optimal tracking path, and achieves a medium tracking error of 9.2 cm for the wrist. However, since it occurs a complexity of 𝒪(N^3T), it is not suitable for real time-tracking. A simpler online version that estimates the arm position from the candidate postures at a single step using a frequency-weighted average achieves a median tracking error of 13.3 cm.  <cit.> propose to improve the computation based on Hidden Markov Model state reorganization, achieving an accuracy of 12.94 cm. With an updated algorithm including a Particle Filter, a median error of 8.8 cm was reported in a static setting <cit.>. Similarly, propose an RNN architecture to track the arm posture from a 6-axis IMU assuming a fixed shoulder position <cit.>. They report a median error of 15.4 cm and a MAE of 16.4 cm for the wrist in a leave-one-subject-out evaluation scheme. LimbMotion uses an additional edge device for acoustic ranging to estimate the arm posture with a median wrist error of 8.9 cm <cit.>. Compared to the previous approaches, our method does not aim to support stand-alone arm tracking based on a 6-axis IMU but to support a visual tracking system for short periods of time where the hand moves outside the cameras field of view. This allows us to correct for drift, especially in the yaw direction which is more pronounced due to the missing magnetometer, as soon as the hand is visible by the headset's cameras. Moreover, our method does not assume a fixed shoulder position. §.§ Spatial Memory, Proprioception, and Kinesthesia The human ability to operate without visual control is mainly supported by the following factors: spatial memory <cit.>, proprioception <cit.>, and kinesthesia <cit.>. Spatial memory refers to the part of memory responsible for recording the position and spatial relations between objects <cit.>. Proprioception refers to the sense of position and orientation of one's body parts with respect to each other <cit.>. Kinesthesia, which is often used interchangeably with proprioception, refers to the perception of one's body movements and motions <cit.>. In prior research, there is substantial work focused on characterizing the aforementioned cognitive and perceptual abilities of people <cit.>. Gutwin et al. <cit.> and Cockburn and McKenzie <cit.>, for instance, studied people's capabilities of building spatial memory in 2D and 3D spaces. Hocherman <cit.>, Soechting and Flanders <cit.>, and Medendrop et al. <cit.> examined the extent to which people can rely on their proprioception to perform target selections. Andrade and Meudell <cit.> and Postma and De Haan <cit.> showed that people are capable of learning spatial locations without paying particular attention to them. Within the human-computer interaction community, it is generally acknowledged that spatial memory, proprioception, and kinesthesia can be exploited to support rich and efficient interactions <cit.>. Li et al.'s Virtual Shelves technique <cit.> leveraged users' kinesthetic memory for triggering programmable shortcuts. Using Virtual Shelves, users can select shortcuts by pointing a spatially-aware device at 28 different locations in front of themselves. Yan et al. <cit.> build on top of this work and experimentally studied eye-free target acquisition in VR including the additional aspect of user comfort. Cockburn et al. <cit.> presented a design space exploration for the interaction termed "air pointing". Gustafson et al.'s Imaginary Interfaces <cit.> sought to enable bi-manual empty-handed interactions without visual feedback by relying on users' short-term memory. Imaginary phone <cit.> demonstrated the potential of transferring users' spatial memory from a familiar physical device to operating an "imaginary" equivalence on their palm. Additional works explored proprioception-driven interactions for menu usage <cit.>, allowing information access via a hip-attached touch device <cit.>, supporting mobile phone access for visually impaired users <cit.>, and enabling interaction with the back of a headset <cit.>. Also closely related to , many works support users in proprioceptively performing interactions through haptics.  <cit.> guided visually-impaired users in target selection on a large wall-mounted display with vibrotactile feedback. Barham et al.'s CAVIAR device <cit.> similarly used vibrotactile actuators to guide users' hands with continuous stimuli. Vo and Brewster  <cit.> studied ultrasonic haptic feedback to support spatial localization. HOOV ultimately aims to enable the aforementioned spatial memory- and proprioception-driven interactions by expanding the tracking space of current VR devices for hand-object interactions. We focus on addressing the technical challenge of performing the sort of eye-free interactions explored by Yan et al. <cit.> when the available headset tracking field-of-view is limited. We further support users in performing eye-free interactions with haptic feedback via an actuator integrated into our wrist-worn device. § METHOD: ESTIMATING PINCH POSITIONS FROM OBSERVATIONS OF WRIST ORIENTATIONS We now introduce our method that enables the tracking of the wrist position, 𝐩^w∈ℝ^3, and orientation, 𝐑^w∈ SO(3), based on the captured acceleration, 𝐚^w∈ℝ^3, and angular velocity signals, ω^w∈ℝ^3, of a 6-axis IMU placed at the user's wrist. Besides the wrist-worn band's motion signals, also receives as input the headset's current position, 𝐩^h∈ℝ^3, and orientation, 𝐑^h∈ SO(3), as well as the sequence of τ_t last available head and hand poses {𝐩^w, 𝐑^w 𝐩^h, 𝐑^h}_-(τ_t-1):0 before the hand has left the tracking FOV of the headset. We describe this mapping f by {𝐩^w, 𝐑^w}_1:τ = f({𝐚^w,ω^w,𝐩^h, 𝐑^h}_1:τ_i, {𝐩^w, 𝐑^w, 𝐩^h, 𝐑^h}_-(τ_t-1):0), where τ matches the number of considered time steps predicted since the hand has left the tracking FOV, and τ_i is the number of inputs sampled in the same period. This is a challenging problem since we only receive the noisy observations of the relative motions of the wrist whose position lies in a 5-DoF space given a static shoulder position (3 DoF through shoulder joint and 2 DoF through lower arm) <cit.> and a 7-DoF space given a constant neck position (additional 2 DoF from clavicle) as input. However, the set of possible wrist poses is severely constrained given the knowledge of the forearm's orientation and by incorporating the anatomical constraints of the individual human joints <cit.>. While orientation estimation from a 6-axis IMU suffers from yaw drift due to the lack of a magnetometer, we demonstrate that a learning-based method can correct for this drift for short-term out-of-view interactions in static settings. Our method does this by estimating the most likely arm trajectory from an implicitly acquired approximated distribution of previously seen trajectory observations. In addition to tracking the 3D positions of the wrist, detects input commands from the wearer. We continuously process the signals of the accelerometer for characteristic patterns to detect gestures, such as pinching or fist clenching. Using the build-in haptic actuator, our prototype is capable of rendering haptic feedback to the user in response via a knocking sensation. §.§ Inertial 6D hand pose estimation <ref> shows an overview of our inertial 6D hand pose estimation pipeline. §.§.§ Input and output representation The input consists of the accelerometer and gyroscope signals from the IMU within the wrist-worn band as well as the current head position and orientation estimated by the inside-out tracking of the VR headset. The spacing between the samples of tracked head positions is matched to the sampling interval of the IMU through cubic interpolation. We further estimate an initial orientation of the wrist by directly applying an extended Kalman Filter <cit.> to the gyroscope and accelerometer signals <cit.>. We convert all orientations to their 6D representation to ensure continuity <cit.>. For each time step, we concatenate the acceleration, angular velocity, head position and orientation, and the output of the extended Kalman Filter 𝐑^k. This input sequence, 𝐗∈ℝ^τ_i× 21, starts from the moment when the hand leaves the trackable FOV of the headset. We obtain another input sequence 𝐒∈ℝ^τ_t× 18 of the last τ_t head and hand poses that were tracked by the headset before the hand left the tracking FOV. Based on this input, our estimator directly predicts wrist position 𝐩^w∈ℝ^3 and rotation 𝐑^w∈ℝ^6 within the world. Using the orientation of the wrist, we produce a more refined estimate of the pinch position, 𝐩^p∈ℝ^3, by applying an offset of 15 cm to the wrist position, 𝐩^p = 𝐑^w(0, -0.15, 0)^T + 𝐩^w, where 𝐑^w is the rotation matrix corresponding to 𝐑^w and the y-axis of the right-handed local coordinate system centered at the wrist points along the forearm to the shoulder with the z-axis pointing downwards through the palm. We took this option because we expected position estimates to be better than representations that encode the rotations for each joint along the kinematic tree from the head to the wrist. Since the interaction happens outside the user's visible FOV, visualization artifacts due to varying bone lengths are of limited concern. §.§.§ Network architecture relies on a neural network to approximate the mapping f. The architecture of the network consists of a downsampling module that receives the input sequence 𝐗 containing the inertial motion information as input and reduces the sequence along the temporal axis by a factor of 8. The module consists of three blocks that each reduce the signal's temporal resolution by a factor of 2 and consist of two convolutional layers with a kernel of size 3 followed by a max pooling layer. The samples of the remaining features are converted to linear embeddings that are fed to a Transformer consisting of two encoder layers. At the start of the sequence, we add an initial token that embeds the information from the headset-tracked hand and head poses 𝐒, extracted through a 2-layer Elman RNN. We apply a sinusoidal positional encoding to the input of the Transformer <cit.>, and avoid that future samples are attended to for the estimation of any output samples by applying a corresponding mask to the sequence. For each output sample of the Transformer corresponding to the down-sampled features from the inertial input, we predict the position and orientation of the wrist using a series of fully-connected layers. This design supports sequences of variable length. §.§.§ Loss function Our loss function, ℒ = ℒ_p + ℒ_R = ∑_t=1^τ (|𝐩̂^w_t-𝐩^w_t| + |𝐑̂^w_t-𝐑^w_t|), consists of two terms, where ℒ_p penalizes the positional and ℒ_R the rotational offset using the L1 loss function, and 𝐩^w_t and 𝐑^w_t are the ground-truth and 𝐩̂^w_t and 𝐑̂^w_t the predicted position and orientation of the wrist at time t respectively. §.§ Pinch detection To detect pinch events, we threshold the running rate-of-change score c, which accumulates the absolute change in the acceleration signals 𝐚^w captured by the IMU at the wrist across time <cit.>, c_t = 1Dc_t-1 + |‖𝐚^w_t‖_2-‖𝐚_t-1^w‖_2|. Pinch events cause sudden changes in a^w_t that lead to a strong increase c_t, and thus, can be detected through a threshold. The exponential reduction factor D attenuates past accumulations. § IMPLEMENTATION is implemented to run as a real-time interactive tracking system on an 8-core Intel Core i7-9700K CPU at 3.60 GHz with an NVIDIA GeForce RTX 3090 GPU. consists of three components: 1) a wrist-worn sensing platform, 2) a virtual reality interface, and 3) a central control and sensor fusion unit that handles the communication to the wristband hardware and the virtual reality headset as well as the estimation of the hand pose outside the FOV. §.§ Hardware We custom-designed an embedded platform for for sensing and actuation during interaction in mid-air. As shown in <ref>, our electronics platform centers around a System-on-a-Chip (NRF52840, Nordic Semiconductors) that samples a 6-axis IMU (LSM6DSOX, STMircroelectronics) at 427 Hz. Our prototype streams the inertial data to a PC, either wirelessly over BLE or through a wired serial interface. The prototype integrates a separate board that embeds an audio amplifier (MAX98357A, Maxim Integrated) to drive a Lofelt L5 actuator to produce haptic feedback. All components of our prototype, including a battery, are housed in a 3D-printed case, which attaches to the wrist using a strap. §.§ Virtual Reality We implemented our VR application in Unity 2021 for the Oculus Quest 2 VR headset. Our application communicates with our outside-field-of-view tracking pipeline via a web socket. As tracking source, our VR environment receives the hand pose computed by the headset, sampled at ∼70 Hz when the user's hand is tracked, and the estimates produced by otherwise. Input events are triggered through 's pinch detection algorithm in both states. §.§ Sensor fusion and outside FOV tracking The main control unit of our implementation is a state machine that moves between states depending on the current tracking status of the VR headset. During the outside FOV state, the headset forwards the received head poses and the inertial input data together with the last 5 available hand and head poses from the VR headset to the deep learning-based hand pose estimation pipeline. We implemented the main control unit in Python 3.8. The program runs across multiple cores to handle the communication with the sensing hardware and the processing of the corresponding signals, and to set the current hand pose for the VR interface. Upon detecting a pinch gesture using an exponential reduction factor D of 1.07 and a threshold of 4.9 m/s^2, our system triggers an event inside the VR environment and sends a command to the motor driver to activate the haptic feedback once. 's network is implemented in PyTorch and has 4,408,199 trainable parameters. We use the Kaiming initialization <cit.>, the Adam optimizer <cit.> with a learning rate of 10^-4, and a batch size of 16 where we randomly group sequences of equal length. § : EVALUATION WHERE PARTICIPANTS INTERACTED BASED ON TRACKING To evaluate 's potential, we conducted a user study in which participants placed and retrieved objects outside their FOV. served two purposes: (1) It allowed us to quantify 's accuracy of tracking wrist positions outside the headset's tracking FOV. (2) We could evaluate how quickly and accurately participants selected and placed objects outside their field of view compared to performing the same task under visual control by turning their heads. This quantified the effect of operating under proprioceptive control. To evaluate the upper bound of our approach, we conducted this first evaluation while participants interacted based on OptiTrack tracking to exclude the impact of tracking errors on participants' behavior. (See Section <ref> for our real-time evaluation of for the same tasks ().) §.§ Study design §.§.§ Apparatus Our experiment apparatus consisted of an Oculus Quest 2 VR headset, a wristband as described above, a computer for logging, and an optical motion capture system. VR headset Throughout the study, participants wore a Quest 2 VR headset with integrated visual hand tracking. For the study setup, we implemented a VR environment featuring objects, placement zones, and task protocols in Unity 2021. Participants saw a visualization of their right hand when it was within their FOV. 's IMU streams and haptic feedback Participants wore a wristband on their right arm, which continuously streamed the data from the IMU to a PC for processing. From the continuous stream of accelerations, our apparatus detected pinch events as described in Section <ref>. To eliminate the impact of wireless connectivity onto participants' performance, we connected the wristband through thin and flexible magnet wires to the PC for power and reliable serial communication in this study. Motion capture for 3D wrist and head trajectories To compare the accuracy of 's estimated wrist position to a ground-truth baseline, we tracked rigid-body markers attached to the participant's headset, shoulder, elbow, and wristband using an 8-camera OptiTrack system with sub-mm accuracy to obtain positions and rotations at a sampling frequency of 240 Hz. The rigid bodies are made out of cardboard that we attached to worn elbow and shoulder pads, and the HOOV wristband (<ref>). To track the headset, we directly attached four 9-mm tracking markers to the device. An experimenter ensured a consistent placement of the rigid bodies across participants ahead of the experiment and verified throughout that the markers did not slip. We calibrated the OptiTrack to the coordinate system of the VR headset through two pre-defined calibration points in the physical space that we marked with a controller in the virtual environment. §.§.§ Participants We recruited 12 participants from our local university (3 female, 9 male, ages 22–35, mean=25.7 years). Using the OptiTrack, we measured the distance between the worn headset and the floor as well as the distance between the neck and the wrist while participants stood in a T-pose. The average headset distance from the ground was 173 cm (SD=12 cm, min=149 cm, max=188 cm), and arm lengths ranged between 62 cm and 80 cm (mean=72 cm, SD=6 cm). Participants self-reported their prior experience with VR technology on a 5-point Likert scale (from 1–never to 5–more than 20 times). Participants' ratings ranged between 1 and 5 and their median prior experience was 3. Each participant received a small gratuity for their time. §.§.§ Task The study consisted of three grab-and-place tasks. Throughout the experiment, participants stood at a fixed point that was highlighted on the floor using tape, such that the experimenter could verify their position. In the first task (<ref>), participants stood next to a set of 17 spherical target objects outside their FOV in the virtual environment, grabbed an intended sphere, and placed it into the blue dropzone in front of them. Participants saw a down-sized illustration of the task environment in front of them, which highlighted the sphere to pick next. Participants grabbed a sphere using a pinch gesture, selecting the sphere that was closest to their right hand when in range. Participants then moved it towards the dropzone and released it using a second pinch. In the second task , participants picked up a spherical target in front of them, and placed it within one of 17 drop zones to their right outside their FOV. Again, a set of three matching illustrations highlighted the task's target drop zone next to them. Participants first grabbed the sphere using a pinch gesture, before moving it to the target drop zone. After pinching again, the apparatus placed the sphere in the drop zone that was closest to the hand's position at the time of the second pinch. The third task combined Tasks 2 and 1. First, participants grabbed the sphere within their field of view in front of them and dropped it into one of the 17 drop zones outside their field of view as instructed (<ref>b). Then, they were instructed to grab one of the 17 spheres from outside their field of view and release it within the drop zone in front of them (<ref>a). As shown in <ref>, the spheres and drop zones were arranged in a grid of 3 rows, 3 columns, and 2 layers in the lateral direction of the participant. We excluded the closer center location behind the participant, because reaching it would require contorting one's arm in an uncomfortable pose without turning one's upper body. All other targets outside the FOV were convenient to reach. The spacing of the targets and drop zones remained static throughout the study, and was adjusted at the beginning of the experiment to each participant's eye level and arm length to ensure all targets could be reached. The spacing in the lateral direction between the two layers was equal to half the length of the participant's arm. The spacing between objects and drop zones in the sagittal and ventral plane was equal to half the distance between the neck and the wrist of the participant when performing a T-pose, corresponding to an angular spacing of around 30 between the drop zones and targets further away from the participant. The target and drop zone grid was centered at the height of the participant's shoulder, and placed at a distance so that the participant's hand would reach all elements of the second layer when elongated. §.§.§ Conditions Participants performed the three tasks in two conditions. In the condition, the participant's hand was solely tracked by the Quest 2 headset. Thus, participants needed to turn their heads during the task and follow their hands to ensure that the headset tracked the hand's position when performing the task. In the condition, participants were instructed not to turn their head and to look straight forward. An experimenter ensured this throughout by monitoring the head-mounted display through a secondary screen. During this condition, the wrist position and orientation was provided to the VR environment by the OptiTrack motion capture system. In both conditions, the apparatus detected pinch gestures for grabbing and releasing the spheres using the IMU inside the wristband as in Section <ref>. Participants received haptic feedback for detected pinch events and were notified of erroneous selections and placements with audio feedback. §.§.§ Procedure The study started with a brief introduction of the tasks. The experimenter then noted down the participant's age and gender. Participants performed a T-pose to calibrate the system to their body size. The study started with a training session of all conditions and tasks, followed by the experiment. During training, participants performed 17 trials, one for each target location, for each task. When using the condition, participants received visual guidance for the and task through the down-sized visualization to allow them to refine their internal model of their hand position in 3-space. The visualization highlighted the corresponding sphere or drop zone as soon as the participant's hand came into contact with it. This online highlighting was exclusive to the training session and was not available during the actual experiment. After training, participants performed all tasks in each of the conditions without guidance. We counterbalanced the order of the task and conditions across participants. Participants performed 34 trials for and , and 68 trials for the task, where a single trial consisted of placing or retrieving a single target object. For and , each drop zone and sphere was used as target twice with the order across targets randomized. The same drop zone or target sphere was not used in direct succession. For the task, we fixed the order of locations between directly successive and subtasks but randomized the order across combinations. We ensured that each valid position was used twice as drop zone and target object location. In total, participants completed 2 conditions × (34  trials + 34  + 68  trials) = 272 trials in under 45 minutes. Participants were instructed to complete the tasks as fast as possible while avoiding mistakes. §.§.§ Measures As dependent variables, our apparatus measured the time it took participants to complete each trial and the percentage of successfully completed trials. A trial was successful when the participant placed the correct sphere in the correct drop zone. §.§ Comparing with high-accuracy tracking from a motion capture system Our study apparatus logged all signals and, thus, recorded a labeled dataset of IMU signals and ground-truth wrist and head positions from the motion capture system. We used the recordings to train and test 's estimation network for an offline evaluation. This allowed us to not only quantify 's performance given the specific configuration of the headset, but we could also simulate a variety of FOV configurations and, thus, moments at which took over tracking from the headset due to our knowledge of the hand positions relative to the head pose at every time step. The condition thereby acts as the upper bound for participants' overall performance given the tracking system's high accuracy. Using the OptiTrack data, we could evaluate participants' ability to interact with objects outside their FOV under near-perfect tracking conditions. This enables us to decompose 's error into contributions from participants' inherent limitation to perform proprioceptive interactions and limitations associated with our method itself. Training and evaluation We evaluated our network in a 6-fold cross-validation where each fold used the data of two randomly selected participants for testing and the other 10 participants for training. We further augmented our training data with the recordings from three additional participants that we acquired while piloting our study setup. For training, we use all data from the condition, including the trajectories recorded during participants' initial training sessions. To further augment our dataset, we trained our network on sequences that simulate a horizontal FOV between 40 and 120 in steps of 5. We also generated training input based on the logged tracking information from the Quest 2. We trained our network until convergence on validation data from a participant that was extracted and excluded from each training dataset. We trained for at least 250,0000 iterations on an NVIDIA GeForce RTX 3090, which takes approximately 12 hours. For the test set, we only use the trials of the condition from the actual experiment. We simulated and evaluated three headset tracking FOVs, including 90 (i.e., the visible FOV of the Quest 2), 120 (i.e., the approximate FOV of human binocular vision), and the actual tracking FOV of the Quest 2. The latter exceeds 120 as we empirically determined through the headset's reported tracking state. We evaluated 's output in terms of mean absolute distance error (MAE) and mean angle of the difference rotation (MAD) <cit.>, θ=2arccos(𝐪̂^w·𝐪^w), to the measurements reported by the motion capture system, averaged over the whole out-of-field-of-view tracking path. Here, 𝐪̂^w and 𝐪^w are the unit quaternions corresponding to the ground-truth orientation 𝐑̂^w and predicted orientation 𝐑^w respectively, and · denotes the inner (or dot) product of vectors. We also compare 's success rates for the simulated FOV ranges to and . §.§ Results §.§.§ Success rate <ref> (left) shows an overview of participants' performance success during this study. Using , participants were careful not to make mistakes when performing the tasks under visual control. Participants reached an average success rate of 94.12% for (max=100.00, min=88.24, SE=1.30), 94.61% for (max=100.00, min=79.41, SE=1.69), and 94.12% for (max=100.00, min=88.24, SE=0.93). Using , participants' mean success rate was 90.93% for (max=100.00, min=76.47, SE=1.91), 90.69% for (max=97.06, min=82.35, SE=1.50), and 90.56% for (max=100.00, min=83.82, SE=1.26). When simulating the tracking of participants' wrist motions outside the FOV with , participants' overall success rate was 87.50% for (SE=1.63, max=97.06, min=76.47), 84.56% for (SE=2.02, max=97.06, min=67.65) and 83.82% for the task (SE=1.83, max=97.06, min=72.06). §.§.§ Trial Duration As shown in <ref> (left), participants' mean completion time was shorter for the three tasks when not looking at the outside FOV regions, i.e., during , , and in the condition. On average, participants took 2.37 s to complete a trial (max=3.27, min=1.66, SE=0.14), 2.65 s for a trial (max=3.50, min=1.96, SE=0.15), and 2.79 s for a trial (max=3.59, min=2.07, SE=0.17). As shown in <ref>, participants' mean completion time increased across all tasks in the condition. Here, the average mean trial completion time was 2.93 s for (max=5.36, min=2.29, SE=0.22), 3.12 s for (max=4.58, min=2.33, SE=0.17), and 3.29 s for the task (max=4.36, min=2.57, SE=0.16). §.§.§ Absolute tracking accuracy of Comparing the wrist poses estimated by to the poses captured by the OptiTrack system when participants' hands were outside the tracking FOV of the Quest 2, we found an MAE of 7.77 cm across participants. The median position offset averaged across participants was 7.00 cm. <ref> illustrates the development of the average tracking error over time from the point when a participant's hand left the FOV of the headset. When dropping an object outside the FOV, 's position estimates had an average MAE to the actual hand position of 6.95 cm. The average MAE when selecting objects outside the FOV was 7.72 cm. The orientation error as measured by the mean angle of the difference rotation is 6.50 (see <ref>). As mentioned above, we simulated a series of tracking FOV to assess how 's tracking capabilities may deteriorate for larger areas outside the FOV and, thus, increased of -based tracking. For a simulated horizontal FOV of 120, the average positional MAE across participants increased to 9.28 cm, and the average MAD increased to 7.67. The offset to the actual release position and grab position while performing the and stayed within 10 cm, amounting to 8.92 cm and 9.86 cm, respectively, on average. For a simulated horizontal FOV of only 90, estimated positions with an average MAE of 10.16 cm and estimated orientations with an MAD of 8.12. §.§ Discussion The results of underline the potential of our method to complement camera-based headset tracking for motions and interactions outside their tracking FOV. Most obviously was participants' speed increase when they could directly interact with targets outside their FOV without turning their heads to first bring them into the view. On average, mean completion times were 19% lower for the and 15% lower for the and task when participants were able to interact outside their FOV. Because participants were able to directly compare the conditions during their experience in the study, several mentioned, unprompted, that turning their heads felt like a burden during this task. The results are also promising, as our study task simulated a common scenario in real life—performing an action for hand-object interaction while focusing on another object or action somewhere else. Real-world examples include grabbing a water bottle or shifting gears while driving, or switching brushes while painting. However, the gain in speed came at the cost of accuracy as expected. Under visual control, participants achieved higher success rates for grabbing and dropping targets than operating outside their FOV without vision, which has been quantified in previous experiments <cit.>. Note that our targets were world-anchored and static; while the experimenter ensured that participants did not move from the position on the floor, shifts in upper-body pose may already result in a considerable effect. Upon closer inspection of where the outside FOV input events were least correct, we saw a disproportional amount of errors for the spheres and drop zones behind the user. In these cases, participants found it difficult to locate their hand outside their field of view correctly, possibly because these locations were outside participants' proprioceptive range and thus the area they would naturally interact with outside their FOV in other use cases. Related, in erroneous cases, failed to detect the correct object especially when participants placed their hand just between two targets. Then, even small deviations in estimated positions led to binary changes of target outcomes. We note that future systems may take this into account and potentially increase the separation between targets to correct for the tracking inaccuracies, introducing a dead band between targets where input events have no effect. In terms of 's potential to complement the tracking capabilities of headsets with limited FOVs, we first highlight its small median error of < 8 for all three simulated FOV conditions in <ref>. Several applications in VR benefit from stable orientation estimation, such as games (e.g., Beat Saber) or 3D interactive environments (e.g., for ray-casting). Next, we underline 's capability to estimate absolute positions outside the headset's tracking FOV. Given a suitable interface design for virtual content outside the user's FOV, the median error of 7 cm—about the length of an index finger—offers sufficient tracking accuracy to navigate through a range of targets in the hemisphere defined by the arm's length. For this purpose, elements would need sufficient size and spacing that would allow users to traverse them in conjunction with the haptic feedback rendered by our wrist device. While 's tracking estimates are closest to the ground-truth within the first three seconds after the hand leaves the headset's FOV, showed promise in that its tracking error deteriorated only minimally under our simulated 90 tracking FOV condition. The mean position error of ∼10 cm and the mean angle of the difference rotation of ∼8offer encouraging benefits for to work in conjunction with future headsets optimizing on power and form factor. Regarding participants' qualitative feedback (<ref>), their overall reception of performing the tasks in the condition was positive. 8 out of 12 participants reported that they preferred the condition over , as their necks felt strained from repetitively turning around. Several participants also mentioned that this condition felt more efficient, even if we saw during the analysis that they tended to make more mistakes. In contrast, 4 of the 12 participants said that they would rather look at the object while performing the task, because it felt `safer.' Comparison to related approaches is not directly comparable to other methods that estimate the wrist position from a single IMU, because we evaluated our system on a different dataset and receives an initial position of the hand as input. However, 's relatively small median position error of 7 cm demonstrates the strength of our estimation pipeline to handle the accurate estimation of the wrist pose over short periods. Comparing the results to the values reported in the literature, performs 15.5% better than ArmTrak's offline algorithm <cit.>, 11.7% better than the Particle Filter-based approach by  <cit.>, 12.7% better than LimbMotion <cit.>, which uses an additional remote edge device, and 49.5% better than the RNN proposed by  <cit.> (see <ref>). In terms of human performance, the results from our study generally align with the conclusions drawn by  <cit.>. However, we also showed that participants were able to differentiate between objects in the lateral direction outside their field of view. In order to understand 's estimation performance during real-time use, we conducted another evaluation where participants' interaction was based solely on 's tracking. § : EVALUATION WHERE PARTICIPANTS INTERACTED BASED ON 'S REAL-TIME ESTIMATES In , we conducted another evaluation where participants interacted based on 's real-time predictions, such that any potential tracking error of our method affected their behavior and thus performance. §.§ Study design §.§.§ Apparatus We built on the experiment apparatus from our previous study (Section <ref>). To additionally evaluate the accuracy of our pinch detection and position, we attached two motion capture markers to participants' thumb and index fingernails before the study. Tracking them, we simultaneously assessed the detection accuracy of pinch gestures as well as the accurate position of the fingers relative to the tracked wristband. Our apparatus detected pinch events when the two markers converged to a distance below 3 cm in a sudden motion. In contrast to , the OptiTrack system was merely present in our apparatus to record ground-truth wrist poses for later analysis and had no influence on participants' behavior, the processed input events, or the virtual scene during the study. Likewise, the apparatus continued to detect pinch events based on the IMU signals (and not based on the attached OptiTrack markers). The VR front-end of our study apparatus received the wrist position and orientation from 's real-time estimations. For this, a PC (32 GB RAM, 3.60 GHz CPU) with an NVIDIA GeForce RTX 3090 received the headset's pose, hand positions from the VR headset when visible, and the stream of IMU signals from 's wristband and processed them in real-time using the model trained on all data from (<ref>). §.§.§ Participants We recruited a fresh set of 10 participants (4 female, 6 male, ages 22–28, mean=25.3 years). Headset distances from the ground were 157—190 cm (mean=170 cm, SD=9 cm), and arm lengths ranged between 57—76 cm (mean=69 cm, SD=6 cm). Participants' prior experience with VR technology was lower than in our first study, ranging between 1–4 on a 5-point Likert scale (median=2). Participants received a small gratuity for their time. §.§.§ Task Participants completed the same three grab-and-place tasks (, & ) as in while standing at an indicated stationary position. The placement of the 17 virtual spheres and drop zones was also identical to our first study. §.§.§ Conditions Participants again completed each task in two different conditions. The first condition was identical to the condition in (Section <ref>), where participants visually followed their hands by turning their heads to interact with out-of-view targets. In the second condition , participants faced straightforward and completed the tasks without turning their heads. Unlike in , our apparatus tracked participants' interaction, rendered positions, and executed input events based on 's predictions. Participants again received haptic feedback at grab and release events and audio feedback depending on the outcome of a trial. §.§.§ Procedure At the start of the study, participants received an introduction to the tasks, provided information about their age and gender, and performed the T-pose calibration procedure to adjust the virtual environment to their body size. Participants received training for each condition and task that included visual feedback for the condition. Overall, participants completed 272 trials (= 2 conditions × (34 trials + 34  trials + 68  trials)) as fast as they could while avoiding errors. Each drop zone and sphere appeared as target twice for each combination of task and condition. §.§.§ Measures Using our study apparatus, we measured the duration of each trial as well as participants' success rates. We evaluated the tracking performance of compared to the ground-truth tracking from the OptiTrack's marker-based tracking system. Using the additional optical markers, we also analyzed the accuracy of our pinch gesture detection. §.§ Results §.§.§ Success rate Using the , participants correctly performed the task in 97.65% (max=100.00, min=91.18, SE=0.83) of all trials for , 95.88% (max=100.00, min=85.29, SE=1.43) for , and 97.94% (max=100.00, min=92.65, SE=0.72) for . Using 's real-time tracking, participants correctly completed the task in 87.65% (max=97.06, min=70.59, SE=2.04) of trials for , 86.18% (max=100.00, min=73.53, SE=2.28) for , and 86.18% (max=91.18, min=76.47, SE=1.23) for . Analyzing the sub-mm accuracy measurements from the motion capture system for comparison, participants' hands were in the correct position 92.35% (max=97.06, min=76.47, SE=1.83) of the trials for , 95.59% (max=100.00, min=79.41, SE=1.75) for and 92.21% (max=100.00, min=72.06, SE=2.09) during the task. Note that this analysis of human performance during interaction based on proprioception is hypothetical, as any real-time tracking error incurred by had an impact on participants' behavior during this study. §.§.§ Trial duration Using , participants on average completed trials in 3.17 s (max=4.06, min=2.10, SE=0.20) for , 3.74 s (max=5.26, min=2.48, SE=0.22) for , and 3.90 s (max=6.41, min=2.49, SE=0.32) for . Using 's real-time tracking and without participants turning their heads, mean trial duration was 2.99 s (max=3.63, min=2.44, SE=0.11) for , 3.54 s (max=4.07, min=2.93, SE=0.10) for , and 3.38 s (max=4.11, min=2.87, SE=0.13) for . §.§.§ Pinch accuracy Compared to the detected pinches using the two additional optical markers, 's pinch detection achieved a recall of 94.85% and a precision of 98.05% across all three tasks. Any false-negative detection in during the study led participants to pinch again, since tracking had no impact on the behavior of the visual environment and study procedure. §.§.§ Absolute tracking accuracy of HOOV When considering as a continuous tracker of wrist position, 's median position error was 15.11 cm (MAE=16.97 cm) averaged across all trials and participants. In terms of accuracy during input events outside the FOV, the position error of 's predicted release and grab events was 14.43 (±4.16) cm for , 16.41 (±4.21) cm for , and 14.85 (±3.19) cm for . Analyzing 's refined estimates of fingertip positions (15 cm offset to the wrist) using the markers attached to participants' fingernails, the average error was 4.9 cm when using the 's wrist position as the basis and 14.0 cm when using 's estimated wrist orientation and position as the basis for refinement (which encompasses 's error of predicting wrist positions). §.§.§ Qualitative feedback As shown in <ref> (bottom), participants again preferred performing the study using overall, that is without requiring them to turn their heads to bring off-screen targets into view first. 7 out of the 10 participants reported a preference for the condition over . Participants again mentioned the efficiency of not having to look for close-by off-screen targets before selecting them, although our analysis again showed that they had a slightly decreased success rate. §.§ Discussion Overall and compared to , participants were more careful in performing tasks in , both in the and the condition. This is evident as the success rates for and both exceeded participants' results in the first study using the OptiTrack system (see <ref>). This also explains the longer average trial completion time in as shown in <ref>. Specific to the condition, participants achieved a similar performance in terms of success rate. This demonstrates the capability of our learning-based method to generalize to and support interaction for users unseen during training without the need for calibration or fine-tuning. It also shows our method's suitability for proprioceptive interaction when possibly imperfect tracking influences user behavior and thus alters their interaction. Further supporting our method was participants' qualitative feedback, which highlighted that 's tracking capabilities that extend and complement those of a regular headset were well-received. In terms of tracking accuracy, produced a higher position error than in . Again, this likely resulted from the longer time that participants spent completing trials and, thus, the increased amount of time they interacted outside the field of view. This result is also commensurate with our assessment of tracking error over time as shown in <ref>. Nevertheless, the absolute error in position remained relatively stable over time and still enabled participants to complete the intended tasks successfully. The results from also highlight the suitability of our pinch detection. Our system is tuned towards a more conservative prediction of pinch events, resulting in a low number of wrongly detected peaks. This rarely required participants to pinch more than once when our online detection missed an event. Finally, this evaluation also validated our previous assumption of a constant offset between 's predicted wrist position and the participant's fingertip. The offset we found from the pinch position reported by the optical tracker was moderate and well below the error introduced by interacting based on proprioception outside the FOV in the first place. Comparison to related approaches. The wrist position error for the interactive tracking in is comparable to the results reported for 's RNN-based method <cit.>. However, their results were obtained through an offline evaluation (similar to ) using a Kinect camera with limited accuracy, and the displayed user motions had a limited range in the yaw direction. While our performance falls short of the reported median tracking error of 8.8 cm for MUSE <cit.> in a static setting with unsupervised user motions, the error becomes comparable in a medical rehabilitation (MR) application involving reach-and-grasp tasks (see <ref>). In addition, MUSE relies on a 9-axis IMU with a magnetometer prone to interference and performs the tracking in a 5-DoF space with respect to the torso. § APPLICATIONS Being complementary to the optical tracking abilities of commercial headsets, a multitude of applications can benefit from during operation to support outside FOV interaction. A large number of game titles and immersive experiences rely on external tracking to obtain the position of the user's hand-held controllers around the body. Such external tracking also supports comfortable interaction, where users can let their arms hang loosely by their sides, but still engage in an interactive experience thanks to controller input. With , we aim to enable similar levels of interactivity while supporting the portability-optimized form factor of current VR headsets, which therefore track all input inside-out as opposed to relying on additional sensing infrastructure. To that end, we prototyped three demonstrations that illustrate the benefits of in various application areas. §.§ Beat Saber To first illustrate the use of our system's extended tracking capabilities, we extended the Beat Saber gameplay to include omnidirectional targets. Rather than presenting targets to users from a single direction, our game scene contains four tracks. Since users do not have a full visual understanding of the arriving targets, they have to rely more greatly on cues from the music to execute movements. For instance, they may be encouraged to swing backward on a regular beat. With inside-out tracking capabilities of HMDs, these movements would be difficult to track, precluding such interactions from being implemented in current applications. resolves this issue and extends opportunities for gameplay to take advantage of a more dynamic range of movements. §.§ Archery Similar to Beat Saber, firing a bow requires movements that sometimes involve moving one's hand out of the tracking field-of-view of the head-mounted display. Unlike Beat Saber, archery further requires the user to steadily maintain an out-of-view hand position. nonetheless supports this and further enables the release interaction with the on-wrist IMU. §.§ Block Builder Lastly, we implemented a block builder application to demonstrate the applicability of in areas beyond games and play. The Block Builder application consists of a workbench that is instrumented with containers for blocks around the edges. Users can grab specific blocks from each container and piece them together on the workbench. As in a variety of other everyday activities, such as typing, we do not necessarily interact with the world with visual control, but rather sometimes rely on our sense of proprioception for control. As users gradually gain familiarity with the environment, we can expect users to rely more so on their spatial memory to grab items as opposed to visual control. enables this interaction. § LIMITATIONS AND FUTURE WORK While the results of the method we introduced in this paper are promising, several limitations exist that deserve further investigation in future work. as a method to compensate for IMU drift Combating the drift arising from estimating positions based on observations from inertial motions is a challenging problem. With , we so far address this for the limited space of a person's reach, only for a short period of time, and only for stationary and standing scenarios. <ref> shows the development of tracking error over time, highlighting the challenge and remaining work needed to improve dead reckoning based on IMU signals. can therefore not be considered a general-purpose hand position estimator using inertial sensing. The results we presented in this paper also assume reaching and grabbing tasks and may therefore not translate to other tasks users may perform in their proprioceptive reach, such as gesturing, drawing, or other kinds of spatial navigation. The assumptions about the task, space, and scenarios we made to develop specifically for stationary VR scenarios considerably set apart the problem we addressed from more general inertial-based odometry, which attempts dead reckoning over much larger areas in mobile and moving settings. Non-stationary environments During our evaluation, participants stood still at a fixed position in the experimental space while performing all tasks as instructed. Though calibrated to each participant's body dimensions once at the beginning of the experiment, all targets were static and remained world-anchored. Although the experimenter verified throughout that participants did not step away from this position, the nature of the task did not allow us to control for a completely steady posture. Therefore, shifts in upper-body posture or slight shoulder rotations may have contributed to the errors we observed, as participants obtained no feedback about their position in the virtual space relative to the targets. Future evaluations could extend our evaluation to include non-stationary environments where participants can move around. This would additionally evaluate the robustness of the method to motion artifacts. Because 's algorithm currently needs a PC with a good GPU, such an evaluation would require a mobile reimplementation of our method, using model-size reduction techniques and deployment on the embedded platform to support such operation. Computational complexity Due to the self-attention layers, 's neural network incurs a complexity that grows quadratically with the length of the input sequence, limiting the maximum outside-the-field-of-view tracking duration. Future work should consider architectures with recurrent elements and memory <cit.> to avoid the repeated processing of the whole input sequence. Tracking accuracy While the tracking accuracy of is state-of-the-art for the given input modality, running from IMU-based orientation estimates only given the last hand observation from the headset, the existing error leaves room for improvement, especially compared to the tracking capabilities of an optical tracking system or a head-mounted display. We believe that a part of this challenge can be addressed by acquiring a much larger data set, which would allow our method to be refined for personalization and to adapt to varying body sizes. Instrumentation of the wrist requires the user to wear a wristband with an embedded IMU to be tracked outside the FOV and to submit input commands. While this is a common sensor inside every smartwatch today, it still requires a separate device to operate (parts of) VR experiences, which is counter to the intuition of manufacturers to integrate all VR components into just the headset. Given the existing ecosystem of watches and their various programming interfaces, we are excited about the possibility of transferring and adapting our method to run on commodity devices and to support out-of-the-box use. This will necessarily entail optimization strategies on the performance level as mentioned above in order to reduce the computational cost of our model or outsource its estimations to the neural computing unit of a headset. § CONCLUSION We have presented , a wrist-worn sensing and position estimation method that complements existing AR and VR headsets in tracking the user's hand interactions outside the tracking field of view of their built-in cameras. obtains its input from a 6-axis IMU that is embedded in a wrist-worn band, which captures 3-axis acceleration and 3-axis rotational velocity. From these observations, 's learning-based inference pipeline estimates the user's current wrist orientation and position as soon as the hand leaves the tracking range of the headset. Our user studies showed the promise of our method, allowing participants to leverage their proprioception and interact with one out of 17 targets outside their FOV with an overall mean success rate of around 85% () and 87% (). The outside FOV interaction in our study supported speed improvements over what would be required on today's standalone headsets—turning one's head to bring the virtual content into view first and exclusively interact with it under visual control. While under visual control, participants successfully interacted with the correct targets in 19 of 20 cases, we quantified the drop in success rates under outside FOV operation: Monitored with an external motion capture system, participants' success rate dropped to 18 of 20 trials when interacting purely based on their proprioception (and with targets that remained static and world-anchored). Of these 18 successful trials, ∼17 interactions were correctly detected using —without any external infrastructure or cameras. Our study also uncovered 's potential as a tracking complement for raw 3D positions outside the headset's tracking FOV. 's median tracking error of 7 cm for short-term outside-field-of-view interactions outperforms related techniques while requiring a commodity sensor that is commonly found in smartwatches and current hand-held controllers (<ref>). opens up an exciting opportunity to broaden the interactive space for AR and VR headsets that may now be able to better leverage users' proprioception within the large outside FOV space around the user that can accommodate convenient reach. Collectively, we believe that will support the development of a future generation of AR and VR headsets that optimize for form factor and power consumption by outsourcing some of the computation to other devices. We see an opportunity for adaptive interaction techniques—as commonly used for selection tasks where input is ambiguous (e.g., touchscreen typing)—to operate in conjunction with 's estimates to support users' proprioceptive interactions. Even more, we see the future potential for to support hand tracking and interaction inside the headset's FOV, thereby alleviating the computational cost of performing computer vision-based camera processing in real-time with a hand position estimation that may run on an embedded wrist-worn device. We sincerely thank Manuel Meier for helpful discussions and comments. We are grateful to NVIDIA for the provision of computing resources through the NVIDIA Academic Grant. We thank the anonymous reviewers and all participants of our user studies. ACM-Reference-Format

http://arxiv.org/abs/2303.06919v2

20230313083630

NeRFLiX: High-Quality Neural View Synthesis by Learning a Degradation-Driven Inter-viewpoint MiXer

cs.CV

NeRFLiX: High-Quality Neural View Synthesis by Learning a Degradation-Driven Inter-viewpoint MiXer Kun Zhou1,21 Wenbo Li3Equal contribution Yi Wang4 Tao Hu3 Nianjuan Jiang2 Xiaoguang Han1 Jiangbo Lu2Corresponding author ^1SSE, CUHK-Shenzhen, ^2SmartMore Corporation ^3CUHK ^4Shanghai AI Laboratory kunzhou@link.cuhk.edu.cn, {wenboli,taohu}@cse.cuhk.edu.hk hanxiaoguang@cuhk.edu.cn,{jiangbo.lu,wygamle}@gmail.com March 30, 2023 ===================================================================================================================================================================================================================================================================================================================================================================== < g r a p h i c s > figureWe propose NeRFLiX, a general NeRF-agnostic restorer that is capable of improving neural view synthesis quality. The first example is from Tanks and Temples <cit.>, the second/third examples are from LLFF <cit.>, and the last one is a user scene captured by a mobile phone. RegNeRF-V3 <cit.> means the model trained with three input views. Neural radiance fields (NeRF) show great success in novel view synthesis. However, in real-world scenes, recovering high-quality details from the source images is still challenging for the existing NeRF-based approaches, due to the potential imperfect calibration information and scene representation inaccuracy. Even with high-quality training frames, the synthetic novel views produced by NeRF models still suffer from notable rendering artifacts, such as noise, blur, etc. Towards to improve the synthesis quality of NeRF-based approaches, we propose NeRFLiX, a general NeRF-agnostic restorer paradigm by learning a degradation-driven inter-viewpoint mixer. Specially, we design a NeRF-style degradation modeling approach and construct large-scale training data, enabling the possibility of effectively removing NeRF-native rendering artifacts for existing deep neural networks. Moreover, beyond the degradation removal, we propose an inter-viewpoint aggregation framework that is able to fuse highly related high-quality training images, pushing the performance of cutting-edge NeRF models to entirely new levels and producing highly photo-realistic synthetic views. Our project page is available at <https://redrock303.github.io/nerflix/>. § INTRODUCTION Neural radiance fields (NeRF) can generate photo-realistic images from new viewpoints, playing a heated role in novel view synthesis. In light of NeRF's <cit.> success, numerous approaches <cit.> along these lines have been proposed, continually raising the performance to greater levels. In fact, one prerequisite of NeRF is the precise camera settings of the taken photos for training  <cit.>. However, accurately calibrating camera poses is exceedingly difficult in practice. Contrarily, the shape-radiance co-adaption issue <cit.> reveals that while the learned radiance fields can perfectly explain the training views with inaccurate geometry, they poorly generalize to unseen views. On the other hand, the capacity to represent sophisticated geometry, lighting, object materials, and other factors is constrained by the simplified scene representation of NeRF <cit.>. On the basis of such restrictions, advanced NeRF models may nonetheless result in notable artifacts (such as blur, noise, detail missing, and more), which we refer to as NeRF-style degradations in this article and are shown in Fig. <ref>. To address the aforementioned limitations, numerous works have been proposed. For example, some studies, including <cit.>, jointly optimize camera parameters and neural radiance fields to refine camera poses as precisely as possible in order to address the camera calibration issue. Another line of works <cit.> presents physical-aware models that simultaneously take into account the object materials and environment lighting, as opposed to using MLPs or neural voxels to implicitly encode both the geometry and appearance. To meet the demands for high-quality neural view synthesis, one has to carefully examine all of the elements when building complex inverse rendering systems. In addition to being challenging to optimize, they are also not scalable for rapid deployment with hard re-configurations in new environments. Regardless of the intricate physical-aware rendering models, is it possible to design a practical NeRF-agnostic restorer to directly enhance synthesized views from NeRFs? In the low-level vision, it is critical to construct large-scale paired data to train a deep restorer for eliminating real-world artifacts <cit.>. When it comes to NeRF-style degradations, there are two challenges: (1) sizable paired training data; (2) NeRF degradation analysis. First, it is unpractical to gather large-scale training pairs (more specifically, raw outputs from well-trained NeRFs and corresponding ground truths). Second, the modeling of NeRF-style degradation has received little attention. Unlike real-world images that generally suffer from JPEG compression, sensor noise, and motion blur, the NeRF-style artifacts are complex and differ from the existing ones. As far as we know, no previous studies have ever investigated NeRF-style degradation removal which effectively leverages the ample research on image and video restoration. In this work, we are motivated to have the first study on the feasibility of simulating large-scale NeRF-style paired data, opening the possibility of training a NeRF-agnostic restorer for improving the NeRF rendering frames. To this end, we present a novel degradation simulator for typical NeRF-style artifacts (e.g., rendering noise and blur) considering the NeRF mechanism. We review the overall NeRF rendering pipeline and discuss the typical NeRF-style degradation cases. Accordingly, we present three basic degradation types to simulate the real rendered artifacts of NeRF synthetic views and empirically evaluate the distribution similarity between real rendered photos and our simulated ones. The feasibility of developing NeRF-agnostic restoration models has been made possible by constructing a sizable dataset that covers a variety of NeRF-style degradations, over different scenes. Next, we show the necessity of our simulated dataset and demonstrate that existing state-of-the-art image restoration frameworks can be used to eliminate NeRF visual artifacts. Furthermore, we notice, in a typical NeRF setup, neighboring high-quality views come for free, and they serve as potential reference bases for video-based restoration with a multi-frame aggregation and fusion module. However, this is not straightforward because NeRF input views are taken from a variety of very different angles and locations, making the estimation of correspondence quite challenging. To tackle this problem, we propose a degradation-driven inter-viewpoint “mixer" that progressively aligns image contents at the pixel and patch levels. In order to maximize efficiency and improve performance, we also propose a fast view selection technique to only choose the most pertinent reference training views for aggregation, as opposed to using the entire NeRF input views. In a nutshell, we present a NeRF-agnostic restorer (termed NeRFLiX) which learns a degradation-driven inter-viewpoint mixer. As illustrated in Fig. <ref>, given NeRF synthetic frames with various rendering degradations, NeRFLiX successfully restores high-quality results. Our contributions are summarized as * Universal enhancer for NeRF models. NeRFLiX is powerful and adaptable, removing NeRF artifacts and restoring clearly details, pushing the performance of cutting-edge NeRF models to entirely new levels. * NeRF rendering degradation simulator. We develop a NeRF-style degradation simulator (NDS), constructing massive amounts of paired data and aiding the training of deep neural networks to improve the quality of NeRF-rendered images. * Inter-viewpoint mixer. Based on our constructed NDS, we further propose an inter-viewpoint baseline that is able to mix high-quality neighboring views for more effective restorations. * Training time acceleration. We show how NeRFLiX makes it possible for NeRF models to produce even better results with a 50% reduction in training time. § RELATED WORKS NeRF-based novel view synthesis. NeRF-based novel view synthesis has received a lot of attention recently and has been thoroughly investigated. For the first time, Mildenhall  <cit.> propose the neural radiance field to implicitly represent static 3D scenes and synthesize novel views from multiple posed images. Inspired by their successes, a lot of NeRF-based models <cit.> have been proposed. For example, point-NeRF <cit.> and DS-NeRF <cit.> incorporate sparse 3D point cloud and depth information for eliminating the geometry ambiguity for NeRFs, achieving more accurate/efficient 3D point sampling and better rendering quality. Plenoxels <cit.>, TensoRF <cit.>, DirectVoxGo <cit.>, FastNeRF <cit.>, Plenoctrees <cit.>, KiloNeRF <cit.>, and Mobilenerf <cit.>, aim to use various advanced technologies to speed up the training or inference phases. Though these methods have achieved great progress, due to the potential issue of inaccurate camera poses, simplified pinhole camera models as well as scene representation inaccuracy, they still suffer from rendering artifacts of the predicted novel views. Degradation simulation. Since no existing works have explored the NeRF-style degradation cases, we will overview the real-world image restoration works that are most related to ours. The previous image/video super-resolution approaches <cit.> typically follow a fix image degradation type (e.g., blur, bicubic/bilinear down-sampling). Due to the large domain shift between the real-world and simulated degradations, the earlier image restoration methods <cit.> generally fail to remove complex artifacts of the real-world images. In contrast, BSRGAN <cit.> design a practical degradation approach for real-world image super-resolution. In their degradation process, multiple degradations are considered and applied in random orders, largely covering the diversity of real-world degradations. Compared with the previous works, BSRGAN achieves much better results quantitatively and qualitatively. Real-ESRGAN <cit.> develops a second-order degradation process for real-world image super-resolution. In this work, we propose a NeRF-style degradation simulator and construct a large-scale training dataset for modeling the NeRF rendering artifacts. Correspondence estimation. In the existing literature, the video restoration methods <cit.> aim to restore a high-quality frame from multiple low-quality frames. To achieve this goal, cross-frame correspondence estimation is essential to effectively aggregate the informative temporal contents. Some works <cit.> explore building pixel-level correspondences through optical-flow estimation and perform frame-warping for multi-frame compensation. Another line of works <cit.> tries to use deformable convolution networks (DCNs <cit.>) for adaptive correspondence estimation and aggregation. More recently, transformer-based video restoration models <cit.> implement spatial-temporal aggregation through an attention mechanism and achieve promising performance. However, it is still challenging to perform accurate correspondence estimation between frames captured with very distinctive viewpoints. § PRELIMINARIES In this section, we review the general pipeline of NeRF-based novel view synthesis and discuss potential rendering artifacts. As shown in Fig. <ref>, three main steps are involved in the rendering: (1) Ray Shooting. To render the color of a target pixel in a particular view, NeRF utilizes the camera's calibrated parameters π to generate a ray 𝐫(𝐨,𝐝) through this pixel, where 𝐨,𝐝 are the camera center and the ray direction. (2) Ray Marching. A set of 3D points are sampled along the chosen ray as it moves across the 3D scene represented by neural radiance fields. The NeRF models encode a 3D scene and predict the colors and densities of these points. (3) Radiance Accumulation. The pixel color is extracted by integrating the predicted radiance features of the sampled 3D points. Discussion. We can see that establishing a relationship between 2D photos and the 3D scene requires camera calibration. Unfortunately, it is very challenging to precisely calibrate the camera poses, leading to noisy 3D sampling. Meanwhile, some previous works <cit.> also raise other concerns, including the non-linear pinhole camera model <cit.> and shape-radiance ambiguity <cit.>. Because of these inherent limitations, as discussed in Section <ref>, NeRF models still synthesize unsatisfied novel test views. § METHODOLOGY Overview. In this work, we present NeRFLiX, a general NeRF-agnostic restorer which employs a degradation-driven inter-viewpoint mixer to enhance novel view images rendered by NeRF models. It is made up of two essential components: a NeRF-style degradation simulator (NDS) and an inter-viewpoint mixer (IVM). As seen in Fig. <ref>(a), during the training phase, we employ the proposed NDS to create large-scale paired training data, which are subsequently used to train an IVM for improving a NeRF-rendered view using two corresponding reference pictures (reference views). In the inference stage, as illustrated in Fig. <ref>(b), IVM is adopted to enhance a rendered view by fusing useful information from the selected most relevant reference views. §.§ NeRF-Style Degradation Simulator (NDS) Due to the difficulties in gathering well-posed scenes under various environments and training NeRF models for each scene, it is infeasible to directly collect large amounts of paired NeRF data for artifact removal. To address this challenge, motivated by BSRGAN <cit.>, we design a general NeRF degradation simulator to produce a sizable training dataset that is visually and statistically comparable to NeRF-rendered images (views). To begin with, we collect raw data from LLFF-T[the training parts of LLFF <cit.>.] and Vimeo90K<cit.> where the adjacent frames are treated as raw sequences. Each raw sequence consists of three images {I^gt,I_1^r,I_2^r}: a target view I^gt and its two reference views {I_1^r,I_2^r}. To construct the paired data from a raw sequence, we use the proposed NDS to degrade I^gt and obtain a simulated degraded view I, as shown in Fig. <ref>(a). The degradation pipeline is illustrated in Fig <ref>. We design three types of degradation for a target view I^gt: splatted Gaussian noise (SGN), re-positioning (Re-Pos.), and anisotropic blur (A-Blur). It should be noted that there may be other models for such a simulation, and we only utilize this route to evaluate and justify the feasibility of our idea. Splatted Gaussian noise. Although additive Gaussian noise is frequently employed in image/video denoising, NeRF rendering noise clearly differs. Rays that hit a 3D point will be re-projected within a nearby 2D area because of noisy camera parameters. As a result, the NeRF-style noise is dispersed over a 2D space. This observation led us to present a splatted Gaussian noise, which is defined as I^D1 = (I^gt + n) ⊛ g , where n is a 2D Gaussian noise map with the same resolution as I^gt and g is an isotropic Gaussian blur kernel. Re-positioning. We design a re-positioning degradation to simulate ray jittering. We add a random 2D offset δ_i,δ_j ∈ [-2,2] with probability 0.1 for a pixel at location (i,j) I^D2(i,j) = I^D1(i,j) if   p>0.1 I^D1(i+ δ_i,j+δ_j) else   p≤0.1 where p is uniformly distributed in [0, 1]. Anisotropic blur. Additionally, from our observation, NeRF synthetic frames also contain blurry contents. To simulate blur patterns, we use anisotropic Gaussian kernels to blur the target frame. Region adaptive strategy. Neural radiance fields are often supervised with unbalanced training views. As a result, given a novel view, the projected 2D areas have varying degradation levels. Thus, we carry out each of the employed degradations in a spatially variant manner. More specifically, we define a mask M as a two-dimensional oriented anisotropic Gaussian <cit.> M(i,j)= G(i-c_i, j-c_j; σ_i, σ_j, A), where (c_i,c_j),(σ_i, σ_j) are the means and standard deviations and A is an orientation angle. After that, we use the mask M to linearly blend the input and output of each degradation, finally achieving region-adaptive degradations. As shown in Fig. <ref>, our simulated rendered views visually match the real NeRF-rendered ones. All the detailed settings of NDS are elaborated in our supplementary materials. At last, with our NDS, we can obtain a great number of training pairs, and each paired data consists of two high-quality reference views {I_1^r,I_2^r}, a simulated degraded view I, and the corresponding target view I^gt. Next, we show how the constructed paired data {I,I_1^r,I_2^r|I^gt} can be used to train our IVM. §.§ Inter-viewpoint Mixer (IVM) Problem formulation. Given a degraded view I produced by our NDS or NeRF models, we aim to extract useful information from its two high-quality reference views {I_1^r,I_2^r} and restore an enhanced version Î. IVM architecture. For multi-frame processing, existing techniques either use optical flow <cit.> or deformable convolutions <cit.> to realize the correspondence estimation and aggregation for consistent displacements. In contrast, NeRF rendered and input views come from very different angles and locations, making it challenging to perform precise inter-viewpoint aggregation. To address this problem, we propose IVM, a hybrid recurrent inter-viewpoint “mixer" that progressively fuses pixel-wise and patch-wise contents from two high-quality reference views, achieving more effective inter-viewpoint aggregation. There are three modules i.e., feature extraction, hybrid inter-viewpoint aggregation and reconstruction, as shown in Fig. <ref>. Two convolutional encoders are used in the feature extraction stage to process the degraded view I and two high-quality reference views {I_1^r,I_2^r}, respectively. We then use inter-viewpoint window-based attention modules and deformable convolutions to achieve recurrent patch-wise and pixel-wise aggregation. Finally, the enhanced view Î is generated using the reconstruction module under the supervision Loss = |Î - I^gt|, where Î = f(I,I_1^r,I_2^r;θ), where θ is the learnable parameters of IVM. The framework architecture is given in our supplementary materials. §.§ View Selection In the inference stage, for a NeRF-rendered view I, our IVM produces an enhanced version by aggregating contents from two neighboring high-quality views. But, multiple input views are available and only a part of them are largely overlapped with I. In general, only the most pertinent input views are useful for the inter-viewpoint aggregation. To this end, we develop a view selection strategy to choose two reference views {I_1^r,I_2^r} from the input views that are most overlapped with the rendered view I. Specifically, we formulate the view selection problem based on the pinhole camera model. An arbitrary 3D scene can be roughly approximated as a bounding sphere in Fig. <ref>, and cameras are placed around it to take pictures. When camera-emitted rays hit the sphere, there are a set of intersections. We refer to the 3D point sets as Φ_i={p_0^i,p_1^i,⋯,p_M_i^i} and Φ_j={p_0^j,p_1^j,⋯,p_M_j^j} for the i-th and j-th cameras. For m_i-th intersection p_m_i^i ∈Φ_i of view i, we search its nearest point in view j with the L2 distance p_m_i^i → j = min_p ∈Φ_j ( || p - p_m_i^i||_2^2 ). Then the matching cost from the i-th view to the j-th view is calculated by C_i → j = ∑_m_i=0^M_i|| p_m_i^i - p_m_i^i → j||_2^2 . We finally obtain the mutual matching cost between views i and j as C_i ↔ j = C_i → j + C_j → i. In this regard, two reference views {I_1^r,I_2^r} are selected at the least mutual matching costs for enhancing the NeRF-rendered view I. Note that we also adopt this strategy to decide the two reference views for the LLFF-T <cit.> data during the training phase. § EXPERIMENTS §.§ Implementation Details We train the IVM for 300K iterations. The batch size is 16 and the patch size is 128. We adopt random cropping, vertical or horizontal flipping, and rotation augmentations. Apart from the inherent viewpoint changes over {I,I_1^r,I_2^r}, random offsets (±5 pixels) are globally applied to the two reference views (I_1^r,I_2^r) to model more complex motion. We adopt an Adam <cit.> optimizer and a Cosine annealing strategy to decay the learning rate from 5 × 10^-4 to 0. We train a single IVM on the LLFF-T and Vimeo datasets and test it on all benchmarks (including user-captured scenes). §.§ Datasets and Metrics We conduct the experiments on three widely used datasets, including LLFF <cit.>, Tanks and Temples <cit.>, and Noisy LLFF Synthetic. LLFF <cit.>. LLFF is a real-world dataset, where 8 different scenes have 20 to 62 images. Following the commonly used protocols <cit.>, we adopt 1008× 756 resolution for LLFF-P1 and 504× 376 resolution for LLFF-P2. Tanks and Temples <cit.>. It contains 5 scenes captured by inward-facing cameras. There are 152-384 images in the 1920× 1080 resolution. It should be noted that the viewpoints of different frames are significantly larger than LLFF. Noisy LLFF Synthetic <cit.>. There are 8 virtual scenes, each of which has 400 images with a size of 800 × 800. To simulate noisy in-the-wild calibration, we ad hoc apply camera jittering (random rotation and translation are employed) to the precise camera poses. Metrics. Following previous NeRF methods, we adopt PSNR (↑)/SSIM <cit.> (↑)/LPIPS <cit.>(↓) for evaluation. §.§ Improvement over SOTA NeRF Models We demonstrate the effectiveness of our approach by showing that it consistently improves the performance of cutting-edge NeRF approaches across various datasets. LLFF. In order to fully verify the generalization ability of our NeRFLiX, we investigate six representative models, including NeRF <cit.>, TensoRF <cit.>, Plenoxels <cit.>, NeRF-mm <cit.>, NLF <cit.>, and RegNeRF <cit.>. Using rendered images of NeRF models as inputs to our model, we aim to further improve the synthesis quality. The quantitative results are provided in Table <ref>. We find that under both of the two protocols, our method raises NeRF model performance entirely to new levels. For example, NeRFLiX improves Plenoxels <cit.> by 0.61dB/0.025/0.054 in terms of PSNR/SSIM/LPIPS. Tanks and Temples. Due to large variations of camera viewpoints, even advanced NeRF models, e.g., TensoRF <cit.> and DIVeR <cit.>, show obviously inferior rendering quality on this dataset. As illustrated in Table <ref>, we show that our NeRFLiX can still boost the performance of these models by a large margin, especially TensoRF <cit.> achieves 0.51dB/0.01/0.022 improvements on PSNR/SSIM/LPIPS. Noisy LLFF Synthetic. Apart from in-the-wild benchmarks above, we also demonstrate the enhancement capability of our model on noisy LLFF Synthetic. From the results shown in Table <ref>, we see that our NeRFLiX yields substantial improvements for two SOTA NeRF models. Qualitative results. In Fig. <ref>, we provide some visual examples for qualitative assessment. It is obvious that our NeRFLiX restores clearer image details while removing the majority of NeRF-style artifacts in the rendered images, clearly manifesting the effectiveness of our method. More results are provided in our supplementary materials. §.§ Training Acceleration for NeRF Models In this section, we show how our approach makes it possible for NeRF models to produce better results even with a 50% reduction in training time. To be more precise, we use NeRFLiX to improve the rendered images of two SOTA NeRF models after training them with half the training period specified in the publications. The enhanced results outperform the counterparts with full-time training, as shown in Table <ref>. Notably, NeRFLiX has reduced the training period for Plenoxels <cit.> from 24 minutes to 10 minutes while also improving the quality of the rendered images. §.§ Ablation Study In this section, we conduct comprehensive experiments on LLFF <cit.> under the LLFF-P1 protocol to analyze each of our designs. We use TensoRF <cit.> as our baseline[The TensoRF results (26.70dB/0.838/0.204) we examined differ slightly from the published results (26.73dB/0.839/0.204).]. §.§.§ NeRF-Style Degradation Simulator Simulation quality. We first examine the simulation quality of the proposed NeRF-style degradation simulator. To this end, we analyze the distribution of our degraded images, BSR <cit.> degraded images and NeRF rendered images on LLFF <cit.>. We use t-SNE <cit.> to visualize deep image features (by Inception-v3 <cit.>) and results are shown in Fig. <ref>. Our simulated data is statistically much closer to the real rendered images than BSR. This conclusion is also supported by Table <ref>, which demonstrates that our NDS significantly surpasses BSR and yields 0.6-1.0dB improvements when used for learning NeRF degradations. Degradation type. We also evaluate the detailed contribution of each data degradation. We use simulated data to train our models by gradually including four types of degradation, as illustrated in Table <ref>. From quantitative comparisons on LLFF <cit.>, we observe that all employed degradations are beneficial to our system. §.§.§ Inter-viewpoint Mixer View selection strategy. We develop a view selection strategy to make full use of high-quality reference views. As shown in Fig. <ref>, our system can identify the most relevant views for quality enhancement when compared to random selection. Also, the quantitative results in Table  <ref> suggest that our view selection achieves significantly improved results, illustrating the usefulness of our method. Hybrid recurrent multi-view aggregation. To handle large viewpoint differences in reference and rendered views, we develop a hybrid recurrent inter-viewpoint aggregation network. We train models using either pixel-wise or patch-wise aggregation and test different iterations to assess the proposed IVM. Models using a single aggregation approach, as illustrated in Table <ref>, perform worse than our full configuration. Additionally, by gradually increasing the iteration numbers from 1 to 3, we achieve improvements of 0.12dB and 0.06dB, albeit at an additional cost of 66 ms and 46 ms for aggregation. Last, compared with the existing state-of-the-art models in Table <ref>, thanks to the recurrent hybrid aggregation strategy, our IVM outperforms all of these models in terms of quantitative results, demonstrating the strength of our aggregation design. Limitation. Though NeRFLiX achieves promising progress for its universal improvements over existing NeRF models. There are still some future directions that deserve further exploration. (1) Our NDS is one of many possible solutions for NeRF degradation simulation. (2) Exploring real-time inter-viewpoint mixers is interesting and useful. § CONCLUSION We presented NeRFLiX, a general NeRF-agnostic restoration paradigm for high-quality neural view synthesis. We systematically analyzed the NeRF rendering pipeline and introduced the concept of NeRF-style degradations. Towards to eliminate NeRF-style artifacts, we presented a novel NeRF-style degradation simulator and constructed a large-scale simulated dataset. Benefiting from our simulated dataset, we demonstrated how SOTA deep neural networks could be trained for NeRF artifact removal. To further restore missing details of NeRF rendered frames, we proposed an inter-viewpoint mixer that is capable of aggregating multi-view frames captured from free viewpoints. Additionally, we developed a view selection scheme for choosing the most pertinent reference frames, largely alleviating the computing burden while achieving superior results. Extensive experiments have verified the effectiveness of our NeRFLiX. Code will be made publicly available. ieee_fullname § NERF DEGRADATION SIMULATOR Raw data collection. We collect raw sequences from Vimeo90K <cit.> and LLFF-T <cit.>. In total, Vimeo90K contains 64612 7-frame training clips with a 448×256 resolution. Three frames (two reference views and one target view) are selected from a raw sequence of Vimeo90K in a random order. As described in Sec. 5.1, apart from the inherent displacements within the selected views, we add random global offsets to the two reference views, largely enriching the variety of inter-viewpoint changes. On the other hand, we also use the training split of the LLFF dataset, which consists of 8 different forward-facing scenes with 20-62 high-quality input views. Following previous work, we drop the eighth view and use it for evaluation. To construct a training pair from LLFF-T, we randomly select a frame as the target view and then use the proposed view selection algorithm (Sec. 4.3) to choose two reference views that are most overlapped with the target view. Hyper-parameter setup. In Eq. (1), the 2D Gaussian noise map n is generated with a zero mean and a standard deviation ranging from 0.01 to 0.05. The isotropic blur kernel g has a size of 5 × 5. We employ a Gaussian blur kernel to produce blurry contents by randomly selecting kernel sizes (3-7), angles (0-180), and standard deviations (0.2-1.2). Last, in order to obtain a region-adaptive blending map M in Eq. (3), we use random means (c_i,c_j ∈(-16,144)), standard deviations (σ_i ∈ (13,25),σ_j ∈(0,24)), and orientation angles (A ∈(0,180)). Additionally, we visualize some generated masks using different hyper-parameter combinations ([c_i, c_j; σ_i, σ_j, A]) in Fig. <ref>. Training data size. We investigate the influence of training data size. Under the same training and testing setups, we train several models using different training data sizes. As illustrated in Table <ref>, we can observe that the final performance is positively correlated with the number of training pairs. Also, we notice the IVM trained with only LLFF-T data or additional few simulated pairs (10% of the Vimeo90K) fails to enhance the TensoRF-rendered results, , there is no obvious improvement compared to TensoRF <cit.>. This experiment demonstrates the importance of sizable training pairs for training a NeRF restorer. § INTER-VIEWPOINT MIXER In Sec. 4.2, we briefly describe the framework architecture of our inter-viewpoint mixer (IVM). Here we provide more details. As illustrated in Fig. <ref>(a), there are two convolutional modules (“Encoder 1/2") to extract features of the degraded view I and its two reference views {I_1^r,I_2^r}, respectively. Then, we develop a hybrid recurrent aggregation module that iteratively performs pixel-wise and patch-wise fusion. At last, a reconstruction module is implemented by a sequence of residual blocks (40 blocks) to output the enhanced view Î. The default channel size is 128. Feature extraction. Given a rendered view I and its two reference views I_1,2^r, we aim to utilize the two encoders to extract deep image features 𝐟 and 𝐟_1,2^r, respectively. As detailed in Fig. <ref>(a), the two encoders share an identical structure. A convolutional layer is first adopted to convert an RGB frame to a high-dimensional feature. Then we further extract the deep image feature using 5 stacked residual blocks followed by another convolution layer. Hybrid recurrent aggregation. As depicted in Fig. <ref>(a), we employ three hybrid recurrent aggregation blocks (termed “Hybrid-R1(2,3)") to progressively fuse the inter-viewpoint information from the image features (𝐟 and 𝐟_{1,2}^r). Next, we take the first iteration as an example to illustrate our aggregation scheme. Pixel-wise aggregation. As shown in Fig. <ref>(b), we first merge the target view feature 𝐟 and one of the reference features 𝐟_{1,2}^r by channel concatenation. Then we use a convolutional layer to reduce the channel dimension and five residual blocks followed by another convolutional layer to obtain a fused deep feature. Later on, the fused feature and the reference feature are further aggregated through a deformable convolution. And the other reference image follows the same processing pipeline. In this case, we finally obtain two features after the pixel-wise aggregation. Patch-wise aggregation. We adopt a window-based attention mechanism <cit.> to accomplish patch-wise aggregation. In detail, the pixel-wisely fused features are first divided into several 3D slices through a 3D patch partition layer. Then, we obtain 3D tokens via a linear embedding operation and aggregate patch-wise information using a video Swin transformer block. Finally, 3D patches are regrouped into a 3D feature map. In the next iteration, we split the 3D feature map into two “reference" features ḟ_{1,2}^r and repeat the pixel-wise and patch-wise aggregation. Note that, the weights of pixel-wise and patch-wise modules are shared across all iterations to reduce the model complexity. Comparisons with SOTA image restorers. Thanks for this good suggestion. We have tested five off-the-shelf SOTA image/video restoration models (“M1-5" refer to BSRGAN in ICCV 2021, MPRNet in CVPR 2021, Real-ESRGAN in ICCVW 2021, Restormer and RealBasicVSR in CVPR 2022, respectively) to remove NeRF-rendered artifacts. These methods, though good for real-world image restoration, are not designed nor trained specifically to tackle NeRF-style artifacts. In fact, as shown in the Table <ref> below, the quality of the synthesized images is basically not improved or becomes even worse. Conversely, if using our simulated NDS dataset to re-train the BSRGAN model, we find it obviously outperforms the original BSRGAN by 0.62dB (Table 3b of the main paper). This experiment indicates the existing image/video restoration model cannot enhance NeRF-rendered frames, confirming the necessity of NeRFLiX. § ADDITIONAL RESULTS Number of reference views. By default, we perform inter-viewpoint aggregation using two reference views (termed IVM-2V). We train another three models (IVM-0V, IVM-1V, and IVM-3V) adopting different numbers of reference views. The results are shown in Table <ref>. The model without using reference views (IVM-0V) achieves the lowest PSNR and SSIM values compared with other models. Meanwhile, it is observed that the more reference views used, the higher IVM performance, indicating the importance of utilizing high-quality reference views. View selection. Fig. <ref> exhibits the selected views by our algorithm in different NeRF scenes. We see that the proposed view selection strategy is able to choose the most relevant ones from freely captured views. Qualitative results. Here, we provide more visual examples to adequately validate the effectiveness of our approach. As shown in Fig. <ref>, Fig. <ref>, Fig. <ref>, Fig. <ref>, NeRFLiX consistently improves NeRF-rendered images with clearer details and fewer artifacts for all NeRF models. For example, NeRFLiX successfully recovers recognizable characters, object textures, and more realistic reflectance effects, while effectively eliminating the rendering artifacts. Video demo. We also provide a video demo[available at our project website <https://redrock303.github.io/nerflix/>] for a clear visual comparison. First, we show some NeRF-rendered views and the restored counterparts of NeRFLiX. Then, we provide two video cases (one is from LLFF and the other is an in-the-wild scene) to compare the rendered views of TensoRF <cit.> and enhanced results of our NeRFLiX. It is observed that NeRFLiX is capable of producing clearer image details and removing the majority of the NeRF rendering artifacts.

http://arxiv.org/abs/2303.12736v2

20230313134039

DPPMask: Masked Image Modeling with Determinantal Point Processes

cs.CV

DPPMask: Masked Image Modeling with Determinantal Point Processes Junde Xu 1, 2 Equal Contribution. Zikai Lin 1, 2 * Donghao Zhou 1, 2 Yaodong Yang 4 Xiangyun Liao 1 Bian Wu 3 Guangyong Chen 3Corresponding Author (mailto:gychen@zhejianglab.comgychen@zhejianglab.com). Pheng-Ann Heng 1, 4 1 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 2 University of Chinese Academy of Sciences 3 Zhejiang Lab 4 The Chinese University of Hong Kong March 30, 2023 ========================================================================================================================================================================================================================================================================================================================================================================================================================== empty Masked Image Modeling (MIM) has achieved impressive representative performance with the aim of reconstructing randomly masked images. Despite the empirical success, most previous works have neglected the important fact that it is unreasonable to force the model to reconstruct something beyond recovery, such as those masked objects. In this work, we show that uniformly random masking widely used in previous works unavoidably loses some key objects and changes original semantic information, resulting in a misalignment problem and hurting the representative learning eventually. To address this issue, we augment MIM with a new masking strategy namely the DPPMask by substituting the random process with Determinantal Point Process (DPPs) to reduce the semantic change of the image after masking. Our method is simple yet effective and requires no extra learnable parameters when implemented within various frameworks. In particular, we evaluate our method on two representative MIM frameworks, MAE and iBOT. We show that DPPMask surpassed random sampling under both lower and higher masking ratios, indicating that DPPMask makes the reconstruction task more reasonable. We further test our method on the background challenge and multi-class classification tasks, showing that our method is more robust at various tasks. § INTRODUCTION Self-supervised learning aims to extract semantic features by solving auxiliary prediction tasks (or pretext tasks) with pseudo labels generated solely based on input features. While various tasks have been proposed for self-supervised learning, one intuitive idea is learning representations by recovering the original data from the corrupted structure. The philosophy behind such methods is simple: what the model generates can examine whether the model understands. This principle was first introduced by the denoising autoencoder <cit.> which has supported significant advances in NLP <cit.>. Methods that follow this idea such as BERT <cit.> now have become a dominant routine. In the field of image processing tasks, although reconstruction-based pre-training was first put forth by <cit.>, it wasn't until recently that methods based on this concept were brought back to state-of-the-art performance. Benefiting from the new network architectures like ViT <cit.>, Masked Image Modeling (MIM) has become highly popular, and there is a series of more aggressive masking strategies like MAE <cit.>, simMIM <cit.>. However, simply random masking can be problematic in practice. An important fact is that it is unreasonable to force the model to reconstruct something beyond recovery. Consider a simple case as Fig. <ref>. The top row of Fig. <ref> shows the underlying logic of MIM: successful reconstruction implies the network captures the correct semantic features. While the bottom row of Fig. <ref> shows a failure case: if the masking process happens to drop an important semantic of the original image, the book in Fig. <ref>, then will change the semantics of the original image and make the network hardly recover it from the rest. In this case, if the model continues to be forced to reconstruct the original image, the model might fill in the obscured image with whatever is feasible, which will interfere with the process of learning the original features. Furthermore, as the masking rate increases, the original semantic information is distorted with a higher probability. We refer to such a situation as a misalignment problem, i.e. the semantics of masked image and the original image are miss-aligned. Consequently, the misalignment problem will cause the alignment of improper sample pairs, which will eventually harm the performance of the downstream task. Some studies also proposed constructing better sample pairs for MIM. In MaskFeat <cit.>, they change the pixel reconstruction task to HOG reconstruction, to reduce the impact of some ambiguous situations for the network to prediction, such as colors, and textures. However, MaskFeat can only reduce the impact of hardly recoverable high-frequency signals. If the masked part contains the whole object instance, there is still not enough information for the network to rebuild. In ADIOS <cit.> and SemMAE <cit.>, they train an extra segmentation network to partition the image into different semantics. However, the number of semantics varies significantly in different images, thus, it is hard to find an optimal semantic partition network. In AttMask <cit.>, they mask the most attended patches according to their attention score to construct more challenging MIM tasks. However, AttMask needs an attention map to perform sampling, which can not fit into reconstruction-based MIM methods (e.g., MAE) seamlessly. Recently, AMT <cit.> adding attention map guided masking to MAE. Unfortunately, these algorithms do not take into account the misalignment problem, which leads to inferior performance. In contrastive learning, an InfoMin principle suggests that two augmented views of an image should retain task-relevant information while minimizing irrelevant nuisances <cit.>. Analog to MIM, we can summarize the following two conditions: First, the selected patches should be representative enough to cover the whole semantic information of the original image. Second, the masking ratio should be set to a high level to minimize the irrelevant information shares between different masks of the same image. While the second constraint is easy to satisfy, the problem is how to retain the task-relevant original semantics under the limited input ratio. To address this, we propose a novel masking strategy based on Determinantal Point Process (DPPs). DPPs are elegant probabilistic models on sets that can capture both quality and diversity when a subset is sampled from a ground set <cit.>, making them ideal for modeling the set that contains more information of original images as possible. During the sampling process, DPPs will compute the distance of each patch, and select patches that are dis-similar from the selected subset. This process makes the network focus on the patches with more representative information. For example, the unique color, texture, etc. We show that our new sampling strategy can obtain more representative patches to keep the semantics unchanged and alleviate the impact of the misalignment problem. More importantly, We show that DPPMask surpassed random sampling under both lower and higher masking ratios, indicating that DPPMask makes the reconstruction task more reasonable. Furthermore, our method needs no extra training process and achieves minimal computational resource consumption. Our contribution can be summarized as follows: * We analyze the training behavior of reconstruction-based MIM and discuss the impact of the misalignment problem. * To alleviate the impact of misalignment in MIM, we proposed a novel plug-and-play sampling strategy called DPPMask based on DPPs. Our method can generate more reasonable training pairs, is simple yet effective, and requires no extra learning parameters. * We verify our method on two representative MIM frameworks, our experiments evidence that features learned by fewer misalignment problems achieve better performance in downstream tasks. § RELATED WORK Self-supervised learning. Classic deep learning trains the parameters of the model by utilizing labeled data. Instead, self-supervised learning(SSL) expects to acquire representations with unlabeled data by a pre-text task. Among them, Masked language modeling (MLM) has taken the lead to be a highly influential self-supervised learning model before. e.g.,, BERT <cit.> and GPT <cit.> are such successful methods that the academia has focused on these two models for pre-training in NLP. These models leverage visible tokens in a sequence and predict invisible tokens to gain appropriate representations, which have been proved to successfully repaint the field <cit.>. In other fields of SSL, there have been numerous methods that focus on different pretext tasks like reconstructing original tokens from image/patch operations <cit.>and Spatio-temporal operation <cit.>. A well-known method is contrastive learning that capitalizes on augmentation invariance in the feature space and could be evaluated by linear probing <cit.>, which was the previous mainstream based on SSL. Masked Image Modeling(MIM) Masked Image Modeling recently has shown capability to reconstruct pixels <cit.> from corrupted images. MIM can be seen as a generalized Denoising AutoEncoders (DAE) <cit.>, which aims to reconstruct the original tokens from corrupt input. e.g., inputing missing color channels <cit.> or missing pixels <cit.>. Context Encoder <cit.> reconstructs a rectangle area of the original images using convolutional networks. Then ViT <cit.> and iGPT <cit.> recall the learning approach of predicting patches with a contrastive predictive coding loss on the modern vision Transformers, and show strong potential in representation learning. BEiT <cit.> proposes to use a pre-trained discrete VAE <cit.>as the tokenizer, and improves MIM's performance further. However, the tokenizer needs to be offline pre-trained with matched model and dataset which limits its adaptivity. To end this, iBOT <cit.> presents a new framework that performs masked prediction with an online tokenizer and gains prominence achievement. Recently, equipped with a more aggressive masking strategy, SimMIM <cit.> and MAE <cit.> further demonstrate that simple pixel reconstruction can achieve competitive results from previous pre-training methods. Determinantal Point Process. Determinantal point processes (DPPs) are probabilistic models of configurations that favor diversity <cit.>. Its repulsion brings new potential to enhance diversity in multiple machine learning problems, such as feature extract from high dimensional data <cit.>, texture synthesis in image processing <cit.>, building informative summaries by selecting diverse sentences <cit.>. In addition to DPPs, there are some previous methods like Markov random fields(MRFs). However, MRF assumes that repulsion does not depend on context too much, so it cannot express that, say, there can be only a certain number of selected items overall <cit.>. The DPPs can naturally implement this kind of restriction through the rank of the kernel. § METHODS In this section, we first discuss the impact of the misalignment problem on the downstream tasks, then, we give a brief introduction to DPPs. Finally, we propose our method of applying DPPs in the patch masking process. §.§ Misalignment in MIM Let x_1, x_2, z be the masked image, reconstruction target, and hidden vector. In the ideal situation, the input x_1 and reconstruction target x_2 all followed an identical distribution x_1, x_2 ∼ P(z). However, consider the input only captures the partial information of the original image, in this case, x_1 ∼ P(z') where z' defines a different distribution to z. For pixel reconstruction tasks, we denote the encoder and the decoder parameterized by ϕ and θ respectively. Following <cit.>, MIM training can be viewed as variational autoencoder training <cit.>, which can be described as a Maximum A Posteriori (MAP) Estimation: max_ϕ, θ𝔼_z'∼ q_ϕ(z'|x_1)log p_θ (x_2|z') Suppose the encoder is capable of capturing the semantic information of x_1 and the decoder is capable of recovering the image described by z and z', by unfolding the decoder part, we get: p_θ(x_2|z') = ∫ p_θ(x_2|z)P(z|z')dz The Eq.<ref> indicates that the pixel reconstruction task is minimizing the distance between representations of masked image z' and original image z. Now, let's zoom the lens to multiple steps, MIM can receive multiple different masked images of the original image. By training the network to reconstruct original images from different masked ones, MIM minimizes the distance between those different masks. Fig. <ref> shows a diagram of such training behavior, where different masking result lies at a different location of a hyperplane and construct the semantic space. During the reconstruction training process, data points in the semantic space are pushed to align with the location of the original image. Consequently, as they are aligned with the same data point, the distance between each data point in semantic space is also minimized. While misalignment is a false aggregation of data points that have skewed semantics (orange dot in Fig. <ref>). For other MIM reconstruction targets, such as visual tokens in BEiT <cit.> and iBOT <cit.>, it is easy to verify that they also share similar training behavior. A similar conclusion has also been reported in <cit.>, however, they focus on the dimensional collapse issue in MAE and neglect the misalignment problem of MIM. In practice, semantics are not evenly distributed in images. These semantics are likely to be ignored by random masking strategies. As the MIM pulls masked samples together, two images with different semantics are miss-aligned. If the changed semantics is an important clue for image understanding, such a problem can seriously affect downstream performance. To this end, we propose a new sampling strategy to select as representative patches as possible. §.§ Determinantal Point Process Our core technical innovation is modeling the patch masking process with DPPs. To this end, we start with a high-level overview of DPPs. Brief intro. A determinantal point process (DPPs) is a distribution over configurations of points. The defining characteristic of the DPP is that it is repulsive, which makes it useful for modeling diversity <cit.>. Formally, a point process 𝒫 on a discrete set S = {1, 2, ..., N} is a probability measure on 2^S, the set of all subsets of S. 𝒫 is called a determinantal point process if, when A is a random subset drawn according to 𝒫, we have, 𝒫(Y=A) ∝det(L_A), where L ∈ R^N× N is a real, symmetric, positive semi-definite kernel, and L_A ∈ R^|A| × |A| is a submatrix of L indexed by elements of A. Note this is an unnormalized probability of sampling a set of A. The normalization constant is defined as the sum of the unnormalized probabilities over all subsets of the S, i.e. ∑_A ⊆ Sdet( L_A ). We can compute the normalized constant by the following theorem <cit.>: For any A ⊆ S: ∑_A ⊆ Y ⊆ Sdet(L_Y)=det(L+I_A̅), where I_A̅ is a diagonal matrix such that I_ii = 0 for indices i ∈ A and I_ii = 1 for i ∈A̅. Setting A = ∅, we obtain the following corollary: ∑_A ⊆ Sdet(L_A)=det(L+I_S). Therefore, for any A ⊆ S, we can compute its probability by: 𝒫(Y=A)=det(L_A)/det(L+I ), where I is the identity matrix. We give a simple example to illustrate how DPPs model diversity. Suppose we have two patches i, j to select, i.e. A = {i, j}, we denote the vector of each patch as S_i, S_j ∈ R^1× n, where n is the dimension of elements in S. We can compute the L-ensemble L_ij = S_i^T S_j and their co-occurrence probability by Eq.<ref>. The numerator of Eq.<ref> can be written as: det(L_A) = L_ii× L_jj - L_ij× L_ji. Note that L_ij and L_ji measure the similarity between elements i and j, being more similar lowers the probability of co-occurrence. On the other hand, when the subset is very diverse, i.e. L_ij× L_ji becomes small, the determinant is bigger and correspondingly its co-occurrence is more likely. The DPP thus naturally diversifies the selection of subsets. Unfortunately, to our best knowledge, the implementation of exact DPP sampling needs matrix decomposition <cit.>, which is an unacceptable computation cost during the training iteration. Thus, to apply DPPs in MIM, an approximation is needed. Greedy approximation. Considering we only select the subset Y with the highest probability under a cardinality constraint n, then such problem can be defined as: Y_MAP = max_Y ⊆ Sdet(L_Y). This problem is known as maximum a posteriori (MAP) inference and has been proved as an NP-hard problem in DPPs <cit.>. Instead, the greedy algorithm is widely used for approximation <cit.> for MAP inference, justified by the fact that the log-probability of set in DPPs f(Y) = log det(L_Y) is sub-modular <cit.>. Thus, the selection process of our method can be described as follows: j = max_i∈ S∖ Y_g f(Y_g ∪{i}) - f(Y_g), where Y_g is the subset of Y. In each iteration, we add an item that maximizes the marginal gain to Y_g, until the maximal marginal gain emerges negative or goes against the cardinality constraint. We adopt a fast implementation of the greedy MAP inference algorithm for DPPs following <cit.>. Formally, since L is a PSD matrix, the Cholesky decomposition of L_Y_g is available as L_Y_g = V V^T, where V is an invertible lower triangular matrix. For any i ∈ Z ∖ Y_g, the Cholesky decomposition of L_Y_g ∪{i}can be derived as: L_Y_g∪{i}=[[ L_Y_g L_Y_g, i; L_i, Y_g L_i i ]]=[[ V 0; c_i d_i ]][[ V 0; c_i d_i ]]^⊤, where row vector c_i and scalar d_i ≥ 0 satisfies: V c_i^⊤ =L_Y_g, i, d_i^2 =L_i i-c_i_2^2. Then the determinate of L_Y_g∪{i} can be written as det(L_Y_g∪{i})=det(V V^⊤) · d_i^2=det(L_Y_g) · d_i^2. Therefore, Eq.<ref> is equivalent to select the element i with maximum d_i^2. After solving the equation, the Cholesky factor of L_Y_gcan therefore be efficiently updated after a new item is added to Y_g. With these approximations, the selecting process can be fit in the GPU training loops. In our experiments, the acceleration ratio is up to 10 times faster with respect to exact DPPs sampling and brings the time cost of DPPs to the same level of random, more details can be found in the supplementary. §.§ Purge misalignment with DPPs In this section, we introduce two key factors of DPPMask: kernel and purge ratio. Kernels. A common-used type of kernel is the class of Gaussian kernels <cit.>. Defined by ∀ S_i, S_j ∈ S, L_ij = exp(- ||S_i - S_j||^2/ϵ). Where ϵ is called the bandwidth or scale parameter. This kernel depends on the squared Euclidean distance between the intensity values of pairs of patches. It is often used as a similarity measure on patches. The value of the parameter ϵ has an impact on how repulsive the DPPs are. However, note we only choose patches that maximize Eq.<ref>, and the value of ϵ influences a little to model performance. This is because ϵ will not change the order of distances between patches. Thus, for better numerical stability, we normalize each patch before computing the distances and set ϵ to 1 empirically. Purge ratio. As shown in Fig. <ref>, DPPMask aims to purge those cases in that semantic information has been changed by masking. However, DPPMask can get over-purged in some cases. As Fig. <ref> shows, due to patches of the sky being too similar to each other, then the greedy selection will only focus on the foreground, as it is more diverse than the background. This situation makes the MIM task too easy and purges most of useful augmented inputs, which is not helpful in feature learning. A simple modification can tackle this problem. Instead of letting the selection process hit the cardinality constraint, we set a parameter called purge ratio τ∈ (0, 1) as the threshold of maximal marginal gain. Concretely, in each iteration, we monitor the distance of the next patch to the selected subsets, if the distance is below the purge ratio τ, abort the greedy selection process and fill the subset will random patches. The purge ratio plays a role of adjust how "severe" the DPPMask is. In particular, τ = 0 indicates the selection process becomes fully greedy and τ = 1 indicates fully random sampling. Fig. <ref> shows greedy selection under three different purge ratios, a higher purge ratio can prevent DPPMask get over-purged and maintain the input diversity for training the network. In this section, we first analyze the training behavior of MIM, then we give our method namely DPPMask which uses DPPs to model the repulsion of image patches, in order to sample the most representative patches and preserve the original semantic information of images. We summarize our algorithm in Alg. <ref>. § EXPERIMENT §.§ Implementation details To examine the effectiveness of our method, we perform DPPMask on two representative MIM methods: MAE <cit.>, iBOT <cit.>, which represent two different MIM frameworks: pixel reconstruction and feature contrast. For MAE, images are patched by convolutional kernels and added with a position embedding, after that, we compute the distances of patches. For iBOT, images are fed into a teacher model to get semantic tokens, which we compute distances based on. Compare to direct computing with pixel intensities, semantic tokens may contain more useful information for partitioning images. For example, to identify instances that share similar appearances. We adopt two different scales of backbones, ViT-Base and ViT-Small for MAE and iBOT respectively. We mainly evaluate our algorithms on the ImageNet-100 dataset, which is derived from ImageNet-1K <cit.>. We train MAE and iBOT for 400 and 100 epochs respectively. Unfortunately, our computational resources can not support us to make out larger-scale experiments, such as more training epochs and heavier backbones. We leave this for future work. We set the masking ratio is set to 0.75 for MAE and 0.7 with 0.05 variance for iBOT by default. §.§ A detailed study of misalignment We give our main result in Tab. <ref> and Tab. <ref>. The final feature vector for classification is obtained by global pooling. For fine-tuning tasks, we run each setting with three random seeds and report their average performance. We train the iBOT model under the fine-tuning and linear probing parameter setting of MAE for reducing the experiment's complexity. The performance gain brings by DPPMask. We first observed a steady performance gain on fine-tuning tasks in both MAE and iBOT frameworks. In MAE, we make 0.2% accuracy gains. In iBOT, we make 0.4% accuracy gains. This shows our method can improve the representation power by purging the misalignment samples. As the threshold τ increases, the pre-train loss becomes smaller in response. However, when τ is too low, greedy sampling goes over-purged and makes the task too simple for the network to learn useful features. Besides, iBOT and MAE show different preferences of τ, this is because we apply DPPs on the output of the teacher model, which is a different distribution from MAE. For linear probing, we notice a significant performance drop of MAE. This does not surprise us as we analyze in Sec. <ref>. DPPs make the sample space shrunk in order to purge the misalignment samples. Aggregating fewer samples together makes the feature space more continuous and less linear separable. Notably, the features also become more precise to describe an image, which can reflect the fine-tuned performance. This behavior is not observed on iBOT, which uses extensive augmentation to further expand the scale of positive samples. Instead, iBOT got 0.3% performance gain on linear probing as well as cluster performance (NMI and ACC), which proves our method indeed purged improper samples that hurt feature learning. Tab. <ref> shows our method alongside other advanced sampling methods on ImageNet-1K, our methods surpassed other sampling methods. DPPMask makes the MIM task more reasonable. Another key factor of successfully applying MIM is the masking ratio of input images <cit.>. It should be high enough to construct a meaningful reconstruction target while preventing the task from degenerating to simply copying-pasting from neighboring patches. However, the root cause of the misalignment problem is also the aggressive masking strategy. To better understand the relationship between the masking ratio and the misalignment problem, we study the fine-tuned performance of MAE under different masking ratios. We perform each experiment three times and report their mean accuracy of ImageNet-100. For DPPs sampling, we run two values of τ, 0.90 and 0.85, we report the higher performance. As Fig. <ref> shows, we find that the original 0.75 masking ratio is not the optimal setting for fine-tuned performance. Instead, lowering the masking ratio significantly improves the accuracy, this is further evidence of the impact of the misalignment problem on feature learning, as a higher masking ratio raises the probability of misalignment. With DPPs sampling equipped, our method has achieved higher performance in all masking ratio settings. In the masking ratio 0.7, the maximal performance boost reached 0.4%. When the masking ratio gets lower, the pretext task actually becomes more simple, which leads to a performance drop. Notably, we make a better result than the best in MAE (masking ratio at 0.6) with both higher and lower masking ratios (0.5, 0.7). This is meaningful, as the reconstruction problem becomes more simple while our method still performs better than MAE, which shows our sampling method makes the pre-train task more reasonable rather more simple. §.§ Robustness The misalignment problem makes the network align images with different semantics. In some severe cases, the network may are required to align the original image with the background. Thus, the misalignment problem can interfere with the network decision by letting the network more focusing the background of images. To verify this, we evaluate the quality of the learned feature on the background challenge <cit.>. We run fine-tuned models of each method on 4 different variations from the original image. Each variation replaces the original background with empty (O.F.), with another image in the same class (M.S.), with a random image in any class (M.R.), or with an image from the next class (M.N.), examples are shown in Fig. <ref>. Tab. <ref> shows the performance of our method on the background challenge. Both iBOT and MAE witnessed a steady performance gain in four different variations of original images. Notably, our method on MAE achieves 1.08% and 1.13% improvements on O.F. and M.N images respectively, which is higher than the original images (0.86%). Such results were also observed in the iBOT framework. Our method achieves 0.42% improvements on original images, while other variations both improved by a large margin. In particular, we achieve 1.58% improvements on M.S. which largely surpasses the original improvements. This shows our methods are more robust to background changes, as we do not impose the network to align the background to the original image as the random strategy does. §.§ Multi-label classification To further examine the influence of the misalignment problem, we also test our method on a multi-label classification task. An intuitive understanding is that: the network can not reflect the semantic changes, which are trained to reconstruct objects whether are been masked or not. Where in multi-label classification, every semantics is important, which makes them suitable to test whether the network is capable to extract whole information of images. We test our method on CLEVR <cit.> and MS-COCO <cit.> datasets, which are widely used in multi-label classification. The CLEVR dataset contains 24 binary labels, each indicating the presence of a particular color and shape (8 × 3) combination in the image <cit.>. For the MS-COCO dataset, we report the following fine-tuned performance on the validation set: mean average precision (mAP), average per-class F1 score (F1_class), and the average overall F1 score (F1_all) <cit.>. For the CLEVR dataset, we train MAE with and without DPPs for 200 epochs, we report the linear probing and fine-tuned performance of F1_micro, F1_macro and F1_weighted, where ‘micro’ evaluate F1 across the entire dataset, ‘macro’ evaluates an unweighted average of per-label F1 score, and ‘weighted’ scale the per-label F1 score by a number of examples when taking the average. We show our result in Tab. <ref> and Tab. <ref>. Our method shows better performance on both COCO and CLEVR datasets. This shows our sampling method can select more informative patches for the network to reconstruct or align, which reduces the impact of the misalignment problem. §.§ Qualitative analysis of DPPs sampling To better understand the sampling behavior of DPPs, we compare the DPPs sampling result with random sampling. Fig. <ref> shows the MAE reconstruction result from the ImageNet validation set, each triplet from left to right indicates the original image, reconstruction result with random sampling and DPPs sampling. For each image, we fix the random seed in order to find the difference between DPPs and random. We show coincide patches with white boxes, patches in random sampling while not in DPPs are shown in red boxes, and patches in DPPs while not in random are shown in blue boxes. Our experiment shows the reconstruction result of DPP sampling is better than random sampling, which proves that DPPs can represent more complete semantics than random. In particular, the sampling result shows two important properties of DPPs. First, DPPs can catch the appearance of each object more precisely, which is an important clue for image understanding. For example, the slot of the mailbox is crucial evidence to classify with cabin. Another important property is DPPs can retain more small foreground information, which is highly likely omitted in random sampling. Such properties show our method successfully alleviates the impact of misalignment problem, and achieve better performance in feature learning. In our experiments, the MAE does not reconstruct the unobserved semantics, indicating that false positive samples are not perfectly aligned. A proper guess of such a phenomenon can be the diversity of ImageNet or the representative capabilities of networks. However, despite the network does not fall into over-fit, the incorrect gradient of misalignment will still interfere with the learning process. Our experiment also shows the potential of MIM with fewer misalignment problems. § CONCLUSION In this paper, we show that uniformly random masking widely used in previous works unavoidably loses some key objects and changes original semantic information, resulting in a misalignment problem and hurting the representative learning eventually. To this end, we propose a new masking strategy namely the DPPMask to reduce the semantic change of the image after masking. We show that DPPMask can make the MIM task more reasonable by purging the misalignment of training pairs. We hope our work can provide insights to help design a better MIM algorithm. ieee_fullname § IMPLEMENTATION DETAILS OF DPPMASK Suppose we add i into the subset Y_g ∪{j}. From Eq. <ref>, we have [ [ V 0; c_j d_j ]] c_i^'⊤ = L_Y_g∪{j}, i= [ [ L_Y_g, i; L_j i ]], where c_i^'=[ [ c_i (L_j i-⟨ c_j, c_i⟩) / d_j ]] ≐[[ c_i e_i ]]. For updating d_i, we have d_i^' 2 =L_i i-c_i^'_2^2 =L_i i-c_i_2^2-e_i^2 =d_i^2-e_i^2. With Eq. <ref> and Eq. <ref>, we can update d incrementally. § IMAGENET-100 SETTING We follow the original MAE <cit.> experiment setting, except for the learning rate of fine-tuning task. We list our fine-tuning parameters in Tab. <ref>. For iBOT <cit.>, we train a ViT-small backbone with 100 epochs. We change the block mask strategy with random, and set masking ratio to 70% with 5% variation. We list our fine-tuning parameters in Tab. <ref> § GREEDY APPROXIMATION PERFORMANCE We examine the approximation performance with respect to original DPPs. We run each setting with one entire epoch and report the mean time cost of each iteration. As Tab. <ref> shows, the greedy approximation achieved 10x faster than original DPPs, which makes it possible to fit into a GPU training loop. § MORE DISCUSSION ABOUT RELATIVE WORKS Our method is different from recent masking strategies. Recent strategies can roughly divide into two lines of work, learning-based and attention-based. Learning-based strategies include ADIOS<cit.> and SemMAE<cit.>. They both need extra learning parameters. ADIOS trains a network to propose masks adversarially, in order to find out more meaningful masks for MIM tasks. However, the semantic meaningful masks proposed by the network are hard to predict, which is less explainable. SemMAE separates the mask learning process from pre-training and makes the training process into two stages. The type of masks they learned are more like semantic parts, such as heads, arms, etc. However, as the semantics varies in images, the number of classes is hard to define, therefore weakening the application of such methods. Attention-based strategies include AttMask<cit.> and AMT<cit.>. This line of work selects patches according to the attention map. Despite different policies to manipulate the attention map, both intend to retain some patches with high attention scores to give a "hint" to the model as such patches are more likely to have more semantics. Note this policy aligns with our ideas. However, they do not associate it with misalignment problems in MIM, thus leading to an inferior policy. Furthermore, attention maps require an extra forward pass to compute, which brings more computation. § ALIGNMENT VERSUS DIVERSITY As we discussed in <ref>, DPPMask aims to purge the training pairs that are polluted by misalignment problems. However, the network also needs irrelevant information between different masks to perform feature learning which can be measured as the variance of sampled masks. Here, we show that DPPMask achieves the adjustment between alignment and diversity. We first obtain the original semantic representation by feeding the network with unmasked images and saving the cls token. Then, we obtain the masked semantic representation by saving the cls tokens of masked images under different masking strategies. We compute the L2 distance between masked semantics and original semantics to illustrate the alignment of different masking strategies. For masked semantic representation, we run 5 independent trails and compute the L2 distance between each trail to illustrate the diversity of different masking strategies. As shown in <ref>, random masking has the most diversity, but it also suffers from the misalignment problem, i.e. the farthest distance between masked semantics and original semantics <ref>. As the τ decrease, the distance between masked semantics and original semantics has been reduced, indicating more alignment to the original semantics. However, a lower τ cause less diversity of different masks, which can purge some useful training pairs and is not helpful for feature learning <ref>. § BROADER IMPACT Despite the eye-catching performance of MIM algorithms, what makes a good mask for MIM tasks still remains unclear. DPPMask provides a possible answer to this question. By analog to the InfoMin principle of contrastive learning. We conclude two properties of MIM. Masked images should retain the original semantics while minimizing shared information from different masks. Minimizing shared information can be achieved by setting a high mask ratio, while how to retain the original semantics is a non-trivial problem. Furthermore, DPPMask also models the probability-of-co-occurrence of each patch and thus can serve as a potential tool to study the relationship between such two properties.

http://arxiv.org/abs/2303.07334v1

20230313175656

Assessing the performance of spatial cross-validation approaches for models of spatially structured data

stat.CO

Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need Da-Wei Zhou^1, Han-Jia Ye^1[2], De-Chuan Zhan^1, Ziwei Liu^2[2] ^1 State Key Laboratory for Novel Software Technology, Nanjing University ^2 S-Lab, Nanyang Technological University {zhoudw, yehj, zhandc}@lamda.nju.edu.cn, ziwei.liu@ntu.edu.sg March 30, 2023 =============================================================================================================================================================================================================================================================== Evaluating models fit to data with internal spatial structure requires specific cross-validation (CV) approaches, because randomly selecting assessment data may produce assessment sets that are not truly independent of data used to train the model. Many spatial CV methodologies have been proposed to address this by forcing models to extrapolate spatially when predicting the assessment set. However, to date there exists little guidance on which methods yield the most accurate estimates of model performance. We conducted simulations to compare model performance estimates produced by five common CV methods fit to spatially structured data. We found spatial CV approaches generally improved upon resubstitution and V-fold CV estimates, particularly when approaches which combined assessment sets of spatially conjunct observations with spatial exclusion buffers. To facilitate use of these techniques, we introduce the package which provides tooling for performing spatial CV as part of the broader tidymodels modeling framework. Keywords • cross-validation spatial data machine learning random forests simulation Shaded[breakable, borderline west=3pt0ptshadecolor, sharp corners, enhanced, interior hidden, frame hidden, boxrule=0pt] sec-introduction § INTRODUCTION Evaluating predictive models fit using data with internal spatial dependence structures, as is common for Earth science and environmental data (Legendre and Fortin 1989), is a difficult task. Low-bias models, such as the machine learning techniques gaining traction across the literature, may overfit on the data used to train them. As a result, model performance metrics may be nearly perfect when models are used to predict the data used to train the model (the “resubstitution performance” of the model) (Kuhn and Johnson 2013), but deteriorate when presented with new data. In order to detect a model's failure to generalize to new data, standard cross-validation (CV) evaluation approaches assign each observation to one or more “assessment” sets, then average performance metrics calculated against each assessment set using predictions from models fit using “analysis” sets containing all non-assessment observations. We refer to the assessment set, which is “held out” from the model fitting process, as D_out, while we refer to the analysis set as D_in. Splitting data between D_out and D_in is typically done randomly, which is sufficient to determine model performance on new, unrelated observations when working with data whose variables are independent and identically distributed. However, when models are fit using variables with internal spatial dependence (often referred to as spatial autocorrelation), random assignment will likely assign neighboring observations to both D_out and D_in. Given that neighboring observations are often more closely related (Legendre and Fortin 1989), this random assignment yields similar results as the resubstitution performance, providing over-optimistic validation results that over state the ability of the model to generalize to new observations or to regions not well-represented in the training data (Roberts et al. 2017; Bahn and McGill 2012). One potential solution for this problem is to not assign data to D_out purely at random, but to instead section the data into “blocks” based upon its dependence structure and assign entire blocks to D_out as a unit (Roberts et al. 2017). For data with spatial structure, this means assigning observations to D_out and D_in based upon their spatial location, in order to increase the average distance between observations in D_out and those used to train the model. The amount of distance required to ensure accurate estimates of model performance is a matter of some debate, with suggestions to use variously the variogram ranges for model predictors, model outcomes, or model residuals (Le Rest et al. 2014; Roberts et al. 2017; Telford and Birks 2009; Valavi et al. 2018; Karasiak et al. 2021). However, it is broadly agreed that increasing the average distance between observations in D_out and those used to train the model may produce more accurate estimates of model performance. This objective – evaluating the performance of a predictive model – is subtly but importantly distinct from evaluating the accuracy of a map of predictions. Map accuracy assessments assume a representative probability sample in order to produce unbiased accuracy estimates (Stehman and Foody 2019), while assessments of models fit using spatial data typically assume a need to estimate model performance without representative and independent assessment data. Such situations emerge frequently across model-based studies (Gruijter and Braak 1990; Brus 2020), such as during hyperparameter tuning (Schratz et al. 2019); when extrapolating spatially to predict into “unknown space” without representative assessment data (Meyer and Pebesma 2021, 2022); or when working with data collected via non-representative convenience samples (any non-probability sample where observations are collected on the easiest to access members of a population) as commonly occurs in ecology and environmental science (Martin, Blossey, and Ellis 2012; Yates et al. 2018). In these situations, understanding the predictive accuracy of the model on independent data is essential. Although recent research has argued against spatial CV for map accuracy assessments (Wadoux et al. 2021), spatial CV remains an essential tool for evaluating the performance of predictive models. For many situations spatial CV has been shown to provide empirically better estimates of model performance than non-spatial methods (Bahn and McGill 2012; Schratz et al. 2019; Meyer et al. 2018; Le Rest et al. 2014; Ploton et al. 2020), and as such is popularly used to evaluate applied modeling projects across domains (Townsend, Papeş, and Eaton 2007; Meyer et al. 2019; Adams et al. 2020). As such, a variety of methods for spatially assigning data to D_out have been proposed, many of which have been shown to improve performance estimates over randomized CV approaches. However, the lack of direct comparisons between alternative spatial CV approaches makes it difficult to know which methods may be most effective for estimating model performance. Understanding the different properties of various CV methods is particularly important given that many spatial modeling projects rely on CV for their primary accuracy assessments (Bastin et al. 2019; Fick and Hijmans 2017; Hoogen et al. 2019; Hengl et al. 2017). Here, we evaluated the leading spatial CV approaches found in the literature, using random forest models fit on spatially structured data following the simulation approach of Roberts et al. (2017). To facilitate comparison among methods, we offer a useful taxonomy and definition of these CV approaches, attempting to unify disparate terminology used throughout the literature. We then provide comprehensive overview of the performance of these spatial CV methods across a wide array of parameterizations, providing the first comparative performance evaluation for many of these techniques. We found that spatial CV methods yielded overall better estimates of model performance than random assignment. Approaches that incorporated both D_out of spatially conjunct observations and buffers yielded the best results. Lastly, to facilitate further evaluation and use of spatial CV methods, we introduce the R package that implements each of the approaches evaluated here, and avoids many of the pitfalls encountered by prior implementations. spatialsample-and-the-tidymodels-framework § SPATIALSAMPLE AND THE TIDYMODELS FRAMEWORK The tidymodels framework is a set of open-source packages for the R programming language that provide a consistent interface for common modeling tasks across a variety of model families and objectives following the same fundamental design principles as the tidyverse (R Core Team 2022; Kuhn and Silge 2022; Wickham et al. 2019). These packages help users follow best practices while fitting and evaluating models, with functions and outputs that integrate well with the other packages in the tidymodels and tidyverse ecosystems as well as with the rest of the R modeling ecosystem. Historically, data splitting and CV in the tidymodels framework has been handled by the package (Frick et al. 2022), with additional components of the tidymodels ecosystem providing functionality for hyperparameter tuning and model evaluation (Kuhn 2022; Kuhn and Frick 2022). However, primarily focuses on randomized CV approaches, and as such it has historically been difficult to implement spatial CV approaches within the tidymodels framework. A new tidymodels package, , addresses this gap by providing a suite of functions to implement the most popular spatial CV approaches. The resamples created by rely on the same infrastructure as those created by , and as such can make use of the same tidymodels packages for hyperparameter tuning and model evaluation. As an implementation of spatial CV methods, improves on alternate implementations in R by relying on the and packages for calculating distance in geographic coordinate reference systems, improving the accuracy of distance calculations and CV fold assignments when working with geographic coordinates (Pebesma 2018; Dunnington, Pebesma, and Rubak 2021). The package is also able to appropriately handle data with coordinate reference systems that use linear units other than meters and to process user-provided parameters using any units understood by the package (Pebesma, Mailund, and Hiebert 2016). The package also allows users to perform spatial CV procedures while modeling data with polygon geometries, such as those provided by the US Census Bureau. Distances between polygons are calculated between polygon edges, rather than centroids or other internal points, to ensure that adjacent polygons are handled appropriately by CV functions. Finally, provides a standard interface to apply inclusion radii and exclusion buffers to all CV methodologies (Section <ref>), providing a high degree of flexibility for specifying spatial CV patterns. sec-overview § RESAMPLING METHODS This paper evaluates a number of the most popular spatial CV approaches, based upon their prevalence across the literature and software implementations. We focus on CV methods which automatically split data either randomly or spatially but did not address any assessment methods that divide data based upon pre-specified, user-defined boundaries or predictor space. As such, we use “distance” and related terms to refer to spatial distances between observations, unless we specifically refer to “distance in predictor space”. resubstitution §.§ Resubstitution Evaluating a model's performance when predicting the same data used to fit the model yields what is commonly known as the “apparent performance” or “resubstitution performance” of the model (Kuhn and Johnson 2013). This procedure typically produces an overly-optimistic estimate of model performance, and as such is a poor method for estimating how well a model will generalize to new data (Efron 1986; Gong 1986; Efron and Gong 1983). We include an assessment of resubstitution error in this study for comparison, but do not recommend it as an evaluation procedure in practice. randomized-v-fold-cv §.§ Randomized V-fold CV Perhaps the most common approach to CV is V-fold CV, also known as k-fold CV. In this method, each observation is randomly assigned to one of v folds. Models are then fit to each unique combination of v - 1 folds and evaluated against the remaining fold, with performance metrics estimated by averaging across the v iterations (Stone 1974). V-fold CV is generally believed to be the least biased of the dominant randomized CV approaches (Kuhn and Johnson 2019), though research has suggested it overestimates model performance (Varma and Simon 2006; Bates, Hastie, and Tibshirani 2021). This optimistic bias is even more notable when training models using data with internal dependency structures, such as spatial structure (Roberts et al. 2017). For this study, V-fold CV was performed using the function in . blocked-cv §.§ Blocked CV One of the most straightforward forms of spatial CV is to divide the study area into a set of polygons using a regular grid, with all observations inside a given polygon assigned to D_out as a group (Valavi et al. 2018; Brenning 2012; Wenger and Olden 2012). This technique is known as “spatial blocking” (Roberts et al. 2017) or “spatial tiling” (Brenning 2012). Frequently, each block is used as an independent D_out (leave-one-block-out CV, Wenger and Olden 2012), though many implementations allow users to use fewer D_out, combining multiple blocks into sets either at random or via a systematic assignment approach (Valavi et al. 2018). Spatial blocking attempts to address the limitations of randomized V-fold CV by introducing distance between D_in and D_out, though the distance from observations on the perimeter of a block to D_in will be much less than that from observations near the center of the block (O'Sullivan and Unwin 2010). This disparity can be addressed through the use of an exclusion buffer around D_out, wherein points within a certain distance of D_out (depending on the implementation, calculated alternatively as distance from the polygon defining each block, the convex hull of points in D_out, or from each point in D_out independently) are excluded from both D_out and D_in (Valavi et al. 2018). A challenge with spatial blocking is that dividing the study area using a standard grid often results in observations in unrelated areas being grouped together in a single fold, as regular gridlines likely will not align with relevant environmental features (for instance, blocks may span significant altitudinal gradients, or reach across a river to combine disjunct populations). Although this can be mitigated through careful parameterization of the grid, it is difficult to create meaningful D_out while still enforcing the required spatial separation between D_in and D_out. For this study, spatial blocking was performed using the function in . Each block was treated as a unique fold (leave-one-block-out CV). clustered-cv §.§ Clustered CV Another form of spatial CV involves grouping observations into a pre-specified number of clusters based on their spatial arrangement, and then treating each cluster as a D_out in V-fold CV (Brenning 2012; Walvoort, Brus, and Gruijter 2010). This approach allows for a great degree of flexibility, as alternative distance calculations and clustering algorithms may produce substantially different clusters. Similarly to spatially-blocked CV this approach may produce unbalanced D_out and folds that combine unrelated areas, though in practice most clustering algorithms typically produce more sensible fold boundaries than spatial blocking. Clustering is also similar to spatial blocking in that observations closer to the center of a cluster will be much further spatially separated from D_in than those near the perimeter, although clustering algorithms typically produce more circular tiles than blocking methods, and as such generally have fewer points along their perimeter overall. As with spatial blocking, exclusion buffers can be used to ensure a minimum distance between D_in and D_out. A notable difference between spatial clustering and spatial blocking is that, depending upon the algorithm used to assign data to clusters, cluster boundaries may be non-deterministic. This stochasticity means that repeated CV may be more meaningful with clustered CV than with spatial blocking, but also makes it difficult to ensure that cluster boundaries align with meaningful boundaries. For this study, spatial clustering was performed using the function in . Each cluster was treated as a unique fold (leave-one-cluster-out CV). buffered-leave-one-observation-out-cv-blo3-cv §.§ Buffered Leave-One-Observation-Out CV (BLO3 CV) An alternative approach to spatial CV involves performing leave-one-out CV, a form of V-fold cross validation where v is set to the number of observations such that each observation forms a separate D_out, with all points within a buffer distance of D_out omitted from D_in (Telford and Birks 2009; Pohjankukka et al. 2017). As the other methods investigated here are all examples of leave-one-group-out CV, with groups defined by spatial positions, we refer to this procedure as buffered leave-one-observation-out CV (BLO3 CV). This approach may be more robust to different parameterizations than spatial clustering or blocking, as the contents of a given D_out are not as dependent upon the precise locations of blocking polygons or cluster boundaries. Many studies have recommended buffered leave-one-out cross validation for models fit using spatial data, with the size of the exclusion buffer variously determined by variogram ranges for model predictors, model outcomes, or model residuals (Le Rest et al. 2014; Roberts et al. 2017; Telford and Birks 2009; Valavi et al. 2018; Karasiak et al. 2021). For this study, BLO3 CV was performed using the function in . leave-one-disc-out-cv-lodo-cv §.§ Leave-One-Disc-Out CV (LODO CV) The final spatial CV method investigated here is leave-one-disc-out CV, following Brenning (2012). This method extends BLO3 by adding all points within a certain radius of each observation to D_out, increasing the size of the final D_out. Data points falling within the exclusion buffer of any observation in D_out, including those added by the inclusion radius, is then removed from D_out (Figure <ref>). Similarly to blocked and clustered CV, LODO CV evaluates models against multiple D_out of spatially conjunct observations. Similar to BLO3 CV, LODO CV approach may be more robust to different parameter values than methods assigning D_out based upon blocking polygons or cluster boundaries. Unlike any of the other approaches investigated, observations may appear in multiple D_out, with observations in more intensively sampled regions being selected more often. In this study, spatial leave-one-disc-out CV was performed using the function in . An important feature of this implementation of leave-disc-out CV is that the exclusion buffer is calculated separately for each point in D_out. Where other implementations remove all observations within the buffer distance of the inclusion radius to create a uniform “doughnut” shaped buffer, only removes observations that are within the buffer distance of data in D_out, potentially retaining more data in D_in by creating an irregular buffer polygon. sec-methods § METHODS sec-simulation §.§ Landscape Simulation To compare the validation techniques described above, we extended the simulation approach used by Roberts et al. (2017, Box 1). We simulated 100 landscapes, representing independent realizations of the same data-generating process, generating a set of 13 variables calculated using the same stochastic formulations across a regularly spaced 50 x 50 cell grid, for a total of 2,500 cells per landscape (Table <ref>). Simulated predictors included eight random Gaussian fields, generated using the R package (Schlather et al. 2015), which uses stationary isotropic covariance models to generate spatially structured variables. Five additional variables were calculated as combinations of the randomly generated variables to imitate interactions between environmental variables. This simulation approach was originally designed to resemble the environmental data that might be used to model species abundance and distribution; further interpretation of what each predictor represents is provided in Appendix 2 of Roberts et al. (2017). tbl-simulations These predictors were then used to generate a target variable y using Equation <ref>. eq-y y = {[ min(y), if X4≠ 0; X11, if y≥X11; X1 + X5 + X6 + X12 + X13, otherwise ]} One instance of the spatially clustered y values produced by this process is visualized in Figure <ref>. Models were then fit using variables X2, X3, and X6 - X10. Of these seven variables, three were involved in calculating the target variable y (X2 and X3, as components of X4 and X5; and X6, used directly) and therefore provide useful information for models, while the remaining four (X7 - X10) were included to allow overfitting. sec-resampling §.§ Resampling Methodology We divided each simulated landscape into folds using each of the data splitting approaches (Section <ref>) across a wide range of parameter sets (Table <ref>; Table <ref>) in order to evaluate the usefulness of spatial CV approaches. Spatial blocking, spatial clustering, and leave-one-disc-out used a “leave-one-group-out” approach, where each D_out was made up of a single block or cluster of observations, with all other data (excluding any within the exclusion buffer) used as D_in. BLO3 used a leave-one-observation-out approach. We additionally evaluated spatial blocking with fewer D_out than blocks, resulting in multiple blocks being used in each D_out. Each simulated landscape was resampled independently, meaning that stochastic methods (such as V-fold CV and spatial clustering) produced different CV folds across each simulation. All resampling used functions implemented in the and packages (Frick et al. 2022; Mahoney and Silge 2022). Examples of spatial clustering CV, spatially blocked CV, and leave-one-disc-out CV are visualized in Figure <ref>. tbl-whichparams tbl-paramdefs sec-models §.§ Model Fitting and Evaluation For each iteration, we modeled the target variable y using random forests as implemented in the R package (Breiman 2001; Wright and Ziegler 2017), fit using variables X2, X3, and X6 - X10. Random forests generally provide high predictive accuracy even without hyperparameter tuning (Probst, Bischl, and Boulesteix 2018), and as such all random forests were fit using the default hyperparameter settings of the package, namely 500 decision trees, a minimum of 5 observations per leaf node, and two variables to split on per node. Model accuracy was measured using root-mean-squared error (RMSE, Equation <ref>). To find the “ideal” error rate that we would expect CV approaches to estimate, we fit 100 separate random forest models, each trained using all values within one of the 100 simulated landscapes. We then calculated the RMSE for each of these models when used to predict each of the 99 other landscapes. As each landscape is an independent realization of the same data-generation process, the relationships between predictors and y is identical across landscapes, although the spatial relationships between y and variables not used to generate y are likely different across iterations. As such, RMSE values from a model trained on one landscape and used to predict the others represent the ability of the model to predict y based upon the predictors and without relying upon spatial structure. These RMSE estimates therefore represent the “true” range of RMSE values when using these models for spatial extrapolation to areas with the same relationship between predictors and the target feature, but without any spatial correlation to the training data itself. We defined the success of model evaluation methods as the proportion of iterations which returned RMSE estimates between the 5th and 9th percentile RMSEs of this “ideal” estimation procedure. To find the error rate of the resubstitution approach, we fit 100 random forests, one to each landscape, and then calculated the RMSE for each model when used to predict its own training data. To find the error of each CV approach, we first used each CV approach to separate each landscape into n folds (Section <ref>). We then fit models to each combination of n - 1 of these folds, and calculated RMSE when using the model to predict the remaining D_out (Equation <ref>). eq-rmseRMSE = √((1/n)∑_i=1^n(y_i - ŷ_̂î)^2) We then calculated the variance of the RMSE estimates of each method across the 100 simulated landscapes, as well as the proportion of runs for each method which fell between the 5th and 95th percentiles of the “true” RMSE range. Based upon prior research, we expected the optimal spacing between D_in and D_out to be related to the range of spatial dependence either in the outcome variable or in model residuals (Le Rest et al. 2014; Roberts et al. 2017; Telford and Birks 2009). As such, we quantified the range of spatial autocorrelation in both the target variable y and in resubstitution residuals from random forest models using the automated variogram fitting approach implemented in the R package (Hiemstra et al. 2008). sec-results § RESULTS AND DISCUSSION spatial-cv-improves-model-performance-estimates §.§ Spatial CV Improves Model Performance Estimates Spatial cross-validation methods consistently produced more accurate estimates of model performance than non-spatial methods, which were optimistically biased (producing too-low estimates of RMSE) (Table <ref>; Figure <ref>). CV produced the best estimates when D_out of spatially conjunct observations were combined with exclusion buffers (Table <ref>). Spatially clustered CV and LODO, both of which enforce D_out of spatially conjunct observations, were among the most consistently effective CV methods (Figure <ref>). Removing too much data from D_in, such as by clustering with only two folds or blocking with only two blocks resulted in pessimistic over-estimates of RMSE (Figure <ref>). tbl-overall tbl-winners The best parameter sets for CV methods consistently separated the center of D_out from D_in by 25%-41% of the grid length (Clustering 25%-29%; LODO 36%-39%; Blocked 32%-41%; BLO3 24%-30%). Given that the target variable had a mean autocorrelation range of 24.61% of the grid length (Figure <ref>), this suggests that spatial cross-validation approaches produce the best estimates of model performance when D_out is sufficiently separated from D_in such that there is no spatial dependency in the outcome variable between the two sets. Clustering appeared to be the spatial CV method most robust to different parameterizations (Figure <ref>; Figure <ref>), with the highest proportion of all iterations within the target RMSE range (Clustered 36.97% of all iterations; LODO 31.70%; Blocked 27.90%; BLO3 ) (Figure <ref>). This may, however, simply reflect the relatively narrow range of parameters evaluated with clustering, as both blocking and LODO had a wide range of parameters which returned estimates within the target range at least half the time (Figure <ref>). While BLO3 exhibited increasing RMSE with increasing buffer radii (Figure <ref>), as frequently reported in the literature, we found it only rarely produced RMSE estimates within the target range (Figure <ref>). RMSE estimates from spatial blocking were inversely related to the number of D_out used, likely due to the distance between D_out and D_in increasing if adjacent blocks were assigned to the same D_out (Figure <ref>). RMSE estimates generally increased gradually as the number of D_out decreased, though notable increases were observed when blocks were assigned via the continuous systematic method and the number of cells in each grid row were evenly divisible by the number of D_out (e.g., when 1/16th cell sizes produced by a 4-by-4 grid were divided into 4 folds). In these situations, each column of the grid will be entirely assigned to the same D_out, somewhat resembling the CV method of Wenger and Olden (2012), producing D_out which have no neighboring D_in observations in the y direction and therefore a greater average distance between D_out and D_in. Overall, these results suggest that using fewer D_out than blocks may be appropriate when the range of autocorrelation in the outcome variable is relatively small and there is concern about large blocks restricting predictor space (Roberts et al. 2017); however, with longer autocorrelation ranges it is likely best to use a leave-one-block-out approach with fewer, larger blocks. As such, the recommendations from this study are clear: CV-based performance assessments of models fit using spatial data benefit from spatial CV approaches. Those spatial CV approaches are most likely to return good estimates of true model accuracy if they combine D_out of spatially conjunct observations with exclusion buffers, such that the average observation is separated from D_in by enough distance that there is no spatial dependency in the outcome variable between D_in and D_out. limitations §.§ Limitations This simulation study assumed that spatial CV could take advantage of regularly distributed observations, such that all locations had a similar density of measurement points. This assumption is often violated, as it is often impractical to obtain a uniform sample across large areas, and as such observations are often clustered in more convenient locations and relatively sparse in less accessible areas (Meyer and Pebesma 2022; Martin, Blossey, and Ellis 2012). Alternative approaches not investigated in this study may be more effective in these situations; for instance, when the expected distance between training data and model predictions is known, Milà et al. (2022) proposes an alternative nearest neighbor distance matching CV approach which may equal or improve upon buffered leave-one-out CV. An alternative approach put forward by Meyer et al. (2018) uses meaningful, human-defined locations as groups for CV, which may produce better results than the automated partitioning methods investigated in this study. While we believe our results clearly demonstrate the benefits of spatial CV for sampling designs resembling our simulation, we do not pretend to present the one CV approach to rule them all. We additionally do not expect these results, focused upon using CV to evaluate the accuracy of predictive models, will necessarily transfer to map accuracy assessments. Stehman and Foody (2019) explained that design-based sampling approaches provide unbiased assessments of map accuracy, while Wadoux et al. (2021) demonstrated that spatial CV methods may be overly pessimistic when assessing map accuracy, and Bruin et al. (2022) suggested sampling-intensity weighted CV approaches for map accuracy assessments in situations where the study area has been unevenly sampled. However, we expect these results will be informative in the many situations requiring estimates of model accuracy, particularly given that traditional held-out test sets are somewhat rare in the spatial modeling literature. Lastly, we did not investigate any CV approaches which aim to preserve outcome or predictor distributions across D_out. When working with imbalanced outcomes, random sampling may produce D_out with notably different outcome distributions than the overall training data, which may bias performance estimates. Assigning stratified samples of observations to D_out can address this, but it is not obvious how to use stratified sampling when assigning groups of observations (such as a spatial cluster or block), with one outcome value per observation, to D_out as a unit. The package allows stratified CV when all observations within a given group have identical outcome values (that is, when groups are strictly nested within the stratification variable), but this condition is rare and difficult to enforce when using unsupervised group assignment based on spatial location, as with all the spatial CV methods we investigated. Creating D_out based on predictor space, rather than outcome distributions, has also been proposed as a solution to spatial CV procedures restricting the predictor ranges present in D_in. This is a particular challenge if the predictors themselves are spatially structured, and may unintentionally force models to extrapolate further in predictor space than would be expected when predicting new data (Roberts et al. 2017). As increasing distance in predictor space often correlates with increasing error (e.g. Thuiller et al. 2004; Sheridan et al. 2004; Meyer and Pebesma 2021), Roberts et al. (2017) suggest blocking approaches to minimize distance in predictor space between folds, although to the best of our knowledge these approaches are not yet in widespread use. A related field of research suggests methods for calculating the applicability domain of a model (Netzeva et al. 2005; Meyer and Pebesma 2021), which can help to identify when predicting new observations will require extrapolation in predictor space, and will likely produce predictions with higher than expected errors. Such methods are particularly well-equipped to supplement spatial CV procedures, as it adjusts the permissible distance in predictor space based upon the distance between D_in and D_out. sec-conclusion § CONCLUSION These results reinforce that spatial CV is essential for evaluating the performance of predictive models fit to data with internal spatial structure, particularly in situations where design-based map accuracy assessments are not practical or germane. Techniques that apply exclusion buffers around assessment sets of spatially conjunct observations, such as spatial clustering and LODO, are likely to produce the best estimates of model performance. The most accurate estimates of model performance are produced when the assessment and analysis data are sufficiently separated so that there is no spatial dependence in the outcome variable between the two sets. acknowledgements § ACKNOWLEDGEMENTS We would like to thank Posit, PBC, for support in the development of the and packages. software-data-and-code-availability § SOFTWARE, DATA, AND CODE AVAILABILITY The package is available online at https://github.com/tidymodels/spatialsample . All data and code used in this paper are available online at https://github.com/cafri-labs/assessing-spatial-cv . references § REFERENCES tocsectionReferences refs 10 preref-adams2020 Adams, Matthew D., Felix Massey, Karl Chastko, and Calvin Cupini. 2020. “Spatial Modelling of Particulate Matter Air Pollution Sensor Measurements Collected by Community Scientists While Cycling, Land Use Regression with Spatial Cross-Validation, and Applications of Machine Learning for Data Correction.” Atmospheric Environment 230 (June): 117479. <https://doi.org/10.1016/j.atmosenv.2020.117479>. preref-bahn2012 Bahn, Volker, and Brian J. McGill. 2012. “Testing the Predictive Performance of Distribution Models.” Oikos 122 (3): 321–31. <https://doi.org/10.1111/j.1600-0706.2012.00299.x>. preref-bastin2019 Bastin, Jean-Francois, Yelena Finegold, Claude Garcia, Danilo Mollicone, Marcelo Rezende, Devin Routh, Constantin M. Zohner, and Thomas W. Crowther. 2019. “The Global Tree Restoration Potential.” Science 365 (6448): 76–79. <https://doi.org/10.1126/science.aax0848>. preref-Bates2021 Bates, Stephen, Trevor Hastie, and Robert Tibshirani. 2021. “Cross-Validation: What Does It Estimate and How Well Does It Do It? arXiv:2104.00673v2 [Stat.ME].” <https://doi.org/10.48550/arXiv.2104.00673>. preref-Breiman2001 Breiman, Leo. 2001. “Random Forests.” Machine Learning 45: 5–32. <https://doi.org/10.1023/A:1010933404324>. preref-brenning2012 Brenning, Alexander. 2012. “Spatial Cross-Validation and Bootstrap for the Assessment of Prediction Rules in Remote Sensing: The R Package sperrorest.” 2012 IEEE International Geoscience and Remote Sensing Symposium, July. <https://doi.org/10.1109/igarss.2012.6352393>. preref-debruin2022 Bruin, Sytze de, Dick J. Brus, Gerard B. M. Heuvelink, Tom van Ebbenhorst Tengbergen, and Alexandre M.J-C. Wadoux. 2022. “Dealing with Clustered Samples for Assessing Map Accuracy by Cross-Validation.” Ecological Informatics 69 (July): 101665. <https://doi.org/10.1016/j.ecoinf.2022.101665>. preref-brus2020 Brus, Dick J. 2020. “Statistical Approaches for Spatial Sample Survey: Persistent Misconceptions and New Developments.” European Journal of Soil Science 72 (2): 686–703. <https://doi.org/10.1111/ejss.12988>. preref-s2 Dunnington, Dewey, Edzer Pebesma, and Ege Rubak. 2021. s2: Spherical Geometry Operators Using the S2 Geometry Library. <https://CRAN.R-project.org/package=s2>. preref-efron1986 Efron, Bradley. 1986. “How Biased Is the Apparent Error Rate of a Prediction Rule?” Journal of the American Statistical Association 81 (394): 461–70. <https://doi.org/10.1080/01621459.1986.10478291>. preref-efron1983 Efron, Bradley, and Gail Gong. 1983. “A Leisurely Look at the Bootstrap, the Jackknife, and Cross-Validation.” The American Statistician 37 (1): 36–48. <https://doi.org/10.1080/00031305.1983.10483087>. preref-fick2017 Fick, Stephen E., and Robert J. Hijmans. 2017. “WorldClim 2: New 1-Km Spatial Resolution Climate Surfaces for Global Land Areas.” International Journal of Climatology 37 (12): 4302–15. <https://doi.org/10.1002/joc.5086>. preref-rsample Frick, Hannah, Fanny Chow, Max Kuhn, Michael Mahoney, Julia Silge, and Hadley Wickham. 2022. rsample: General Resampling Infrastructure. <https://CRAN.R-project.org/package=rsample>. preref-gong1986 Gong, Gail. 1986. “Cross-Validation, the Jackknife, and the Bootstrap: Excess Error Estimation in Forward Logistic Regression.” Journal of the American Statistical Association 81 (393): 108–13. <https://doi.org/10.1080/01621459.1986.10478245>. preref-degruijter1990 Gruijter, J. J. de, and C. J. F. ter Braak. 1990. “Model-Free Estimation from Spatial Samples: A Reappraisal of Classical Sampling Theory.” Mathematical Geology 22 (4): 407–15. <https://doi.org/10.1007/bf00890327>. preref-hengl2017 Hengl, Tomislav, Jorge Mendes de Jesus, Gerard B. M. Heuvelink, Maria Ruiperez Gonzalez, Milan Kilibarda, Aleksandar Blagotić, Wei Shangguan, et al. 2017. “SoilGrids250m: Global Gridded Soil Information Based on Machine Learning.” Edited by Ben Bond-Lamberty. PLOS ONE 12 (2): e0169748. <https://doi.org/10.1371/journal.pone.0169748>. preref-automap Hiemstra, P. H., E. J. Pebesma, C. J. W. Twenhöfel, and G. B. M. Heuvelink. 2008. “Real-Time Automatic Interpolation of Ambient Gamma Dose Rates from the Dutch Radioactivity Monitoring Network.” Computers & Geosciences. <https://doi.org/10.1016/j.cageo.2008.10.011>. preref-vandenhoogen2019 Hoogen, Johan van den, Stefan Geisen, Devin Routh, Howard Ferris, Walter Traunspurger, David A. Wardle, Ron G. M. de Goede, et al. 2019. “Soil Nematode Abundance and Functional Group Composition at a Global Scale.” Nature 572 (7768): 194–98. <https://doi.org/10.1038/s41586-019-1418-6>. preref-karasiak2021 Karasiak, N., J.-F. Dejoux, C. Monteil, and D. Sheeren. 2021. “Spatial Dependence Between Training and Test Sets: Another Pitfall of Classification Accuracy Assessment in Remote Sensing.” Machine Learning 111 (7): 2715–40. <https://doi.org/10.1007/s10994-021-05972-1>. preref-tune Kuhn, Max. 2022. tune: Tidy Tuning Tools. <https://CRAN.R-project.org/package=tune>. preref-dials Kuhn, Max, and Hannah Frick. 2022. dials: Tools for Creating Tuning Parameter Values. <https://CRAN.R-project.org/package=dials>. preref-Kuhn2013 Kuhn, Max, and Kjell Johnson. 2013. Applied Predictive Modeling. Vol. 26. Springer. <https://doi.org/10.1007/978-1-4614-6849-3>. preref-kuhn2019 ———. 2019. Feature Engineering and Selection. Chapman; Hall/CRC. <https://doi.org/10.1201/9781315108230>. preref-tmwr Kuhn, Max, and Julia Silge. 2022. Tidy Modeling with R. O'Reilly. preref-lerest2014 Le Rest, Kévin, David Pinaud, Pascal Monestiez, Joël Chadoeuf, and Vincent Bretagnolle. 2014. “Spatial Leave-One-Out Cross-Validation for Variable Selection in the Presence of Spatial Autocorrelation.” Global Ecology and Biogeography 23 (7): 811–20. <https://doi.org/10.1111/geb.12161>. preref-legendre1989 Legendre, Pierre, and Marie Josée Fortin. 1989. “Spatial Pattern and Ecological Analysis.” Vegetatio 80 (2): 107–38. <https://doi.org/10.1007/bf00048036>. preref-spatialsample Mahoney, Michael, and Julia Silge. 2022. spatialsample: Spatial Resampling Infrastructure. <https://CRAN.R-project.org/package=spatialsample>. preref-martin2012 Martin, Laura J, Bernd Blossey, and Erle Ellis. 2012. “Mapping Where Ecologists Work: Biases in the Global Distribution of Terrestrial Ecological Observations.” Frontiers in Ecology and the Environment 10 (4): 195–201. <https://doi.org/10.1890/110154>. preref-meyer2021 Meyer, Hanna, and Edzer Pebesma. 2021. “Predicting into Unknown Space? Estimating the Area of Applicability of Spatial Prediction Models.” Methods in Ecology and Evolution 12 (9): 1620–33. <https://doi.org/10.1111/2041-210x.13650>. preref-meyer2022 ———. 2022. “Machine Learning-Based Global Maps of Ecological Variables and the Challenge of Assessing Them.” Nature Communications 13 (1). <https://doi.org/10.1038/s41467-022-29838-9>. preref-meyer2018 Meyer, Hanna, Christoph Reudenbach, Tomislav Hengl, Marwan Katurji, and Thomas Nauss. 2018. “Improving Performance of Spatio-Temporal Machine Learning Models Using Forward Feature Selection and Target-Oriented Validation.” Environmental Modelling & Software 101 (March): 1–9. <https://doi.org/10.1016/j.envsoft.2017.12.001>. preref-meyer2019 Meyer, Hanna, Christoph Reudenbach, Stephan Wöllauer, and Thomas Nauss. 2019. “Importance of Spatial Predictor Variable Selection in Machine Learning Applications – Moving from Data Reproduction to Spatial Prediction.” Ecological Modelling 411 (November): 108815. <https://doi.org/10.1016/j.ecolmodel.2019.108815>. preref-miluxe02022 Milà, Carles, Jorge Mateu, Edzer Pebesma, and Hanna Meyer. 2022. “Nearest Neighbour Distance Matching Leave-One-Out Cross-Validation for Map Validation.” Methods in Ecology and Evolution 13 (6): 1304–16. <https://doi.org/10.1111/2041-210x.13851>. preref-netzeva2005 Netzeva, Tatiana I., Andrew P. Worth, Tom Aldenberg, Romualdo Benigni, Mark T. D. Cronin, Paola Gramatica, Joanna S. Jaworska, et al. 2005. “Current Status of Methods for Defining the Applicability Domain of (Quantitative) Structure-Activity Relationships.” Alternatives to Laboratory Animals 33 (2): 155–73. <https://doi.org/10.1177/026119290503300209>. preref-osullivan2010 O'Sullivan, David, and David J. Unwin. 2010. Geographic Information Analysis. John Wiley & Sons, Inc. <https://doi.org/10.1002/9780470549094>. preref-sf Pebesma, Edzer. 2018. “Simple Features for R: Standardized Support for Spatial Vector Data.” The R Journal 10 (1): 439–46. <https://doi.org/10.32614/RJ-2018-009>. preref-units Pebesma, Edzer, Thomas Mailund, and James Hiebert. 2016. “Measurement Units in R.” R Journal 8 (2): 486–94. <https://doi.org/10.32614/RJ-2016-061>. preref-ploton2020 Ploton, Pierre, Frédéric Mortier, Maxime Réjou-Méchain, Nicolas Barbier, Nicolas Picard, Vivien Rossi, Carsten Dormann, et al. 2020. “Spatial Validation Reveals Poor Predictive Performance of Large-Scale Ecological Mapping Models.” Nature Communications 11 (1). <https://doi.org/10.1038/s41467-020-18321-y>. preref-pohjankukka2017 Pohjankukka, Jonne, Tapio Pahikkala, Paavo Nevalainen, and Jukka Heikkonen. 2017. “Estimating the Prediction Performance of Spatial Models via Spatial k-Fold Cross Validation.” International Journal of Geographical Information Science 31 (10): 2001–19. <https://doi.org/10.1080/13658816.2017.1346255>. preref-Probst2018 Probst, Philipp, Bernd Bischl, and Anne-Laure Boulesteix. 2018. “Tunability: Importance of Hyperparameters of Machine Learning Algorithms.” arXiv. <https://doi.org/10.48550/ARXIV.1802.09596>. preref-R R Core Team. 2022. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. <https://www.R-project.org/>. preref-Roberts2017 Roberts, David R., Volker Bahn, Simone Ciuti, Mark S. Boyce, Jane Elith, Gurutzeta Guillera-Arroita, Severin Hauenstein, et al. 2017. “Cross-Validation Strategies for Data with Temporal, Spatial, Hierarchical, or Phylogenetic Structure.” Ecography 40 (8): 913–29. https://doi.org/<https://doi.org/10.1111/ecog.02881>. preref-RandomFields Schlather, Martin, Alexander Malinowski, Peter J. Menck, Marco Oesting, and Kirstin Strokorb. 2015. “Analysis, Simulation and Prediction of Multivariate Random Fields with Package RandomFields.” Journal of Statistical Software 63 (8): 1–25. <https://doi.org/10.18637/jss.v063.i08>. preref-schratz2019 Schratz, Patrick, Jannes Muenchow, Eugenia Iturritxa, Jakob Richter, and Alexander Brenning. 2019. “Hyperparameter Tuning and Performance Assessment of Statistical and Machine-Learning Algorithms Using Spatial Data.” Ecological Modelling 406 (August): 109–20. <https://doi.org/10.1016/j.ecolmodel.2019.06.002>. preref-sheridan2004 Sheridan, Robert P., Bradley P. Feuston, Vladimir N. Maiorov, and Simon K. Kearsley. 2004. “Similarity to Molecules in the Training Set Is a Good Discriminator for Prediction Accuracy in QSAR.” Journal of Chemical Information and Computer Sciences 44 (6): 1912–28. <https://doi.org/10.1021/ci049782w>. preref-stehman2019 Stehman, Stephen V., and Giles M. Foody. 2019. “Key Issues in Rigorous Accuracy Assessment of Land Cover Products.” Remote Sensing of Environment 231 (September): 111199. <https://doi.org/10.1016/j.rse.2019.05.018>. preref-Stone1974 Stone, M. 1974. “Cross-Validatory Choice and Assessment of Statistical Predictions.” Journal of the Royal Statistical Society. Series B (Methodological) 36 (2): 111–47. <http://www.jstor.org/stable/2984809>. preref-telford2009 Telford, R. J., and H. J. B. Birks. 2009. “Evaluation of Transfer Functions in Spatially Structured Environments.” Quaternary Science Reviews 28 (13-14): 1309–16. <https://doi.org/10.1016/j.quascirev.2008.12.020>. preref-thuiller2004 Thuiller, Wilfried, Lluis Brotons, Miguel B. Araújo, and Sandra Lavorel. 2004. “Effects of Restricting Environmental Range of Data to Project Current and Future Species Distributions.” Ecography 27 (2): 165–72. <https://doi.org/10.1111/j.0906-7590.2004.03673.x>. preref-townsend2007 Townsend, Peterson A., Monica Papeş, and Muir Eaton. 2007. “Transferability and Model Evaluation in Ecological Niche Modeling: A Comparison of GARP and Maxent.” Ecography 30 (4): 550–60. <https://doi.org/10.1111/j.0906-7590.2007.05102.x>. preref-valavi2018 Valavi, Roozbeh, Jane Elith, José J. Lahoz-Monfort, and Gurutzeta Guillera-Arroita. 2018. “blockCV: An R Package for Generating Spatially or Environmentally Separated Folds for k-Fold Cross-Validation of Species Distribution Models.” Edited by David Warton. Methods in Ecology and Evolution 10 (2): 225–32. <https://doi.org/10.1111/2041-210x.13107>. preref-varma2006 Varma, Sudhir, and Richard Simon. 2006. “Bias in Error Estimation When Using Cross-Validation for Model Selection.” BMC Bioinformatics 7 (1). <https://doi.org/10.1186/1471-2105-7-91>. preref-wadoux2021 Wadoux, Alexandre M. J.-C., Gerard B. M. Heuvelink, Sytze de Bruin, and Dick J. Brus. 2021. “Spatial Cross-Validation Is Not the Right Way to Evaluate Map Accuracy.” Ecological Modelling 457 (October): 109692. <https://doi.org/10.1016/j.ecolmodel.2021.109692>. preref-walvoort2010 Walvoort, D. J. J., D. J. Brus, and J. J. de Gruijter. 2010. “An R Package for Spatial Coverage Sampling and Random Sampling from Compact Geographical Strata by k-Means.” Computers & Geosciences 36 (10): 1261–67. <https://doi.org/10.1016/j.cageo.2010.04.005>. preref-wenger2012 Wenger, Seth J., and Julian D. Olden. 2012. “Assessing Transferability of Ecological Models: An Underappreciated Aspect of Statistical Validation.” Methods in Ecology and Evolution 3 (2): 260–67. <https://doi.org/10.1111/j.2041-210x.2011.00170.x>. preref-wickham2019 Wickham, Hadley, Mara Averick, Jennifer Bryan, Winston Chang, Lucy McGowan, Romain François, Garrett Grolemund, et al. 2019. “Welcome to the Tidyverse.” Journal of Open Source Software 4 (43): 1686. <https://doi.org/10.21105/joss.01686>. preref-ranger Wright, Marvin N., and Andreas Ziegler. 2017. “ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R.” Journal of Statistical Software 77 (1): 1–17. <https://doi.org/10.18637/jss.v077.i01>. preref-yates2018 Yates, Katherine L., Phil J. Bouchet, M. Julian Caley, Kerrie Mengersen, Christophe F. Randin, Stephen Parnell, Alan H. Fielding, et al. 2018. “Outstanding Challenges in the Transferability of Ecological Models.” Trends in Ecology & Evolution 33 (10): 790–802. <https://doi.org/10.1016/j.tree.2018.08.001>.

http://arxiv.org/abs/2303.06633v1

20230312110503

The case of Quantum Gravity with Spontaneous Collapse of the Wave Function

gr-qc

Wigner Research Centre for Physics, H-1525 Budapest 114, P.O.Box 49, Hungary Eötvös Loránd University, H-1117 Budapest, Pázmány Péter stny. 1/A diosi.lajos@wigner.hu When about half a century ago the concept of universal spontaneous collapse of the wave function was conceived it was an attempt to alter standard non-relativistic quantum physics. As such, it was largely ignored by relativistic field theory and quantum gravity communities. A central motivation of spontaneous collapse community has been to replace the standard collapse-by-measurement that annoyed many. With few exceptions, it did not annoy the field theory and quantum gravity communities. Concept of certain general-relativity-related universal irreversibility in quantum field theory had been initiated very long ago by Wheeler, Hawking and a few others independently from the concept of spontaneous collapse. Lately the two concepts started to converge and support each other. § INTRODUCTION Allow the author to introduce the topics in a self-quotation from 2015 <cit.>. `The inception of a universal gravity-related irreversibility took place originally in quantum cosmology but it turned out soon that a universal non-unitary dynamics is problematic itself. Independent investigations of the quantum measurement postulate clarified that a non-unitary dynamics is of interest already in the non-relativistic context. An intricate relationship between Newton gravity and quantized bulk matter might result in universal non-relativistic violation of unitarity - also called spontaneous decoherence. The corresponding gravity-related spontaneous decoherence model is now on the verge of detectability in optomechanical experiments. It is also a toy-model of cosmic quantum gravitational non-unitarity, illuminating that the bottle-neck of quantum gravity is the quantum measurement postulate instead of quantum cosmology.' We want to build our message around the appearance of the concept of fundamental irreversibility in two separate fields of foundational theoretical physics. One is quantum cosmology. It has no consistent theory valid down towards the Planck scale, despite various efforts using various advanced mathematical models. At the same time, semiclassical and heuristic approaches imply that fundamental irreversibilities may be part of quantum cosmology. But quantum irreversibility itself, especially when universal, happens to be non-trivial both technically and conceptually. And here another research field gains importance. Its tasks seem opaque compared to quantum cosmology's. They include, e.g., the measurement problem, the dynamics of wave function collapse, the quantum-classical transition and hybrid dynamics, but the relevant task is the quantum theory of macroscopia, coined as the  cat problem. You should discuss the options of macroobjects' quantum behavior, and you can do it non-relativistically, before you go for quantum cosmology! The hypothesis of spontaneous universal collapse in massive degrees of freedom, assuming a fundamental gravity-related irreverisibility, may have perspectives for quantum cosmology as well, especially if fundamental irreversibility remains part of it. § FUNDAMENTAL IRREVERSIBILITY? — TWO COMMUNITIES The two fundamental issues in question are the following: unified theory of space-time with quantized matter (cf. quantum gravity, quantum cosmology) and the physics of quantum measurement (cf. wave function collapse, quantum-classical transition). These problems were considered unrelated for long time, studied by two separate research communities. We refer to the first one as quantum cosmologists and to other as measurement problem solvers or, with a spectacular touch, as  cat killers. Quantum cosmology has always been part of main stream physics, using heavy artillery of mathematics, while  cat killers used light weapons and sometimes whimsical identification of their problems. §.§ Irreversible Quantum Gravity/Cosmology at the Planck Scale Emergence of irreversibility within standard physics was not derived by exact derivations but by heuristic arguments. A selection of earliest theoretical signatures could be the following. In 1936, Bronstein <cit.> was the first to point out that space-time metric must have an ambiguity δ g_ab of different nature from the dictum of Heisenberg uncertainty relationships. Much later in 1962, Wheeler <cit.> proposed that space-time has a certain foamy structure at the Planckian scale. After another decade, Bekenstein <cit.> proved that black-holes behave thermodynamically, they have entropy: S_BH=A/4, where A is the surface aera of the black-hole in Planck units (Boltzmann's constant k_B=1). Then Hawking <cit.> showed that, yes, in accordance with the above entropy, black-holes emit thermal radiation of temperature T_BH=1/8π M, where M is the black-hole mass. Later, Hawking <cit.> went deeper and conjectured that unitarity of standard quantum theory is lost due to the instanton mechanism. The scattering processes must be non-unitary, the non-unitary super-scattering operator $ acts on the density operator : →$ where the r.h.s. does not factorize into SS^† with any unitary scattering operator S. Instead of Hawking's scattering process, Banks, Susskind and Peskin <cit.> considered a detailed dynamical equation d/dt=-i[,]-∫∫ [(x),[(y), ]]h(x-y)d^3xd^3y, where is the Hamiltonian, is a relativistic quantum field and h is a positive kernel. The authors show that the irreversible term on the r.h.s. violates the energy-momentum conservation which cannot be restored in local field theories. §.§ Irreversible Quantum Mechanics for Massive Objects Among diverse incentives, including a metaphysical discontent about von Neumann's quantum measurement theory, the only relevant motivation is the problematic extendibility of quantum theory for massive degrees of freedom. This led to heuristic modifications of standard quantum physics. According to that, the unitarity is violated in massive degrees of freedom so that macroscopically different superpositions, aka  cat states like |Ψ⟩=|f_1⟩+|f_2⟩/√(2), collapse spontaneously into one of the components. Milestones of the consept could be the following. In 1966, Károlyházy <cit.> proposed that, due to a conjectured spectrum of space-time metric fluctuations δ g_ab at the Planck scale,  cat states collapse before they get too large. He outlined a naive qualitative model of when and how the collapse happens. Independently and much later in 1986, Ghirardi, Rimini and Weber <cit.> constructed a simple model of universal spontaneous collapse in exact mathematical terms. They postulated extreme rare and weak spontaneous collapses for the elementary constituents' wave functions which in turn yield robust localization of the macrosopic center-of-mass, i.e., the collapse of the  cat state. The GRW model has no reference to gravity. However, the present author <cit.> proposed gravity-related universal spontaneous collapse of massive superpositions, like  cats, based on a master equation: d/dt=-i/ħ[,]-G/2ħ∫∫ [(x),[(y), ]]1/| x-y|d^3xd^3y, where is the non-relativistic quantized mass density field and G is Newton's constant. The master equation corresponds to a certain ambiguity δΦ=12 c^2 δ g_00 of the Newton potential. §.§ Fundamental irreversibility? — Parallel pursuits Thus, the desired quantum gravity and the unwanted  cats led both the heavy-armed relativistic and light-horse non-relativistic studies to the same structure of heuristic master equations. And they suffer of the same problem of spontaneous creation of energy and momentum. This coincidence is a spectacular instance to illustrate that the two, apparently distant theoretical tasks, share similar proposals and problems. Another instance has been the surprising coincidence between the  cat's collapse rate predicted by the non-relativistic master equation <cit.> and by the general relativistic arguments of Penrose <cit.>. For possible reasons of coincidence, see ref. <cit.>. Mention should be made of Gell-Mann and Hartle <cit.> who were perhaps the only ones at the time to recognize that quantum cosmology needed a more general measurement theory than von Neumann's. For the relationship of their theory of decoherent histories to spontaneous collapse theories, see <cit.>. Most recently, hybrid quantum-classical master equations have been tested in semiclassical cosmology by Oppenheim et al. <cit.>, with numerous references to spontaneous collapse models which are, in fact, the alternative formalism, see <cit.>. § SUMMARY The failures of quantum gravity models may not be due to the inadequacy of the quantization methods tried so far, but the inadequacy of standard quantum theory for macroobjects, e.g., the spontaneous decay of massive superpositions, aka -cats. We recalled that the idea of fundamental irreversibility has also been raised in quantum cosmology. The roots of irreversibility may be common in quantum cosmology and in speculative models of spontaneous collapse of the wave function. This may be so despite the fact that one is predicted at the Planck scale while the other comes from non-relativistic discussion of massive superpositions <cit.>. Considerations of quantum cosmologists and advocates of spontaneous collapse (referred here extravagantly as  cat killers) will have to come closer to each other. To support this suggestion, here are two quotes. The author (cat killer) claimed `the measurement problem culminates in quantum cosmology" <cit.>. The cosmologist claims `the principles of general relativity must influence, and actually change, the very formalism of quantum mechanics' <cit.>. This research was funded by the Foundational Questions Institute and Fetzer Franklin Fund, a donor-advised fund of the Silicon Valley Community Foundation (Grant No's. FQXi-RFPCPW-2008, FQXi-MGA-2103), the National Research, Development and Innovation Office (Hungary) “Frontline” Research Excellence Program (Grant No. KKP133827), and the John Templeton Foundation (Grant 62099). § REFERENCES 99 Erice2015 Diósi L 2015 Gravity related spontaneous decoherence: from Wheeler-Bekenstein-Hawking to optomechanics (talk given in Erice) www.wigner.hu/∼diosi/slides/erice2015.pdf Bro36 Bronstein M 1936 Phys. Z. Sowjetunion 9 140 Whe62 Wheeler J A 1962 Geometrodynamics (New York: Academic Press) Bek73 Bekenstein J D 1973 Phys. Rev. D 7 2333 Haw75 Hawking S W 1975 Commun. Math. Phys. 43 199 Haw83 Hawking S W 1982 Commun. Math. Phys. 87 395 BPS84 Banks T, Susskind L and Peskin M E 1984 Nucl. Phys. B 244 125 EllMohNan89 Ellis J, Mohanty S and Nanopoulos D V 1989 Phys. Lett. B 221 113 GelHar90 Gell-Mann M and Hartle J B 1993 Phys. Rev. D 47 3345 Pen96 Penrose R 1996 Gen. Relativ. Gravit. 28 581 KwoHog16 Kwon O and Hogan C J 2016 Class. Quantum Grav. 33 105004 Sud21 Sudarsky D 2021 Do Wave Functions Jump? Perspectives of the Work of GianCarlo Ghirardi, ed V Allori, A Bassi, D Dürr, N Zanghi (Cham: Springer Nature Switzerland) p 291 AnaHu22 Anastopoulos C and Hu B L 2021 AVS Quantum Science 4 015602 Opp22 Oppenheim J and Weller-Davies 2022 J. High Energ. Phys. 2022 80; Oppenheim J, Sparaciari C, Šoda B and Weller-Davies Z 2022 Preprint arXiv:2203.01982 Kar66 Karolyhazy F 1966 Nuovo Cim. 42 390 Pea76 Pearl P 1976 Phys. Rev. D 13 857 Gis84 Gisin N 1984 Phys. Rev. Lett. 52 1657 GRW86 Ghirardi G C, Rimini A and Weber T 1986 Phys. Rev. D 34 470 Dio87 Diósi L 1987 Phys. Lett. A 120 377 Dio89 Diósi L 1989 Phys. Rev. A 40 1165 GRP90 Ghirardi G C, Rimini A and Pearl P 1990 Phys. Rev. A 42 78 Bedetal14 Bedingham D, Dürr D, Ghirardi G, Goldstein S, Tumulka R and Zanghi N 2014 J. Stat. Phys. 154 623 TilDio16 Tilloy A and Diósi L 2016 Phys. Rev. D 93 024026 Basetal17 Bassi A, Grossardt A and Ulbricht H 2017 Class. Quantum Grav. 34 193002 Dio22 Diósi L 2022 AVS Quantum Sci. 4 015605 Dioetal95 Diósi L, Gisin N, Halliwell J and Percival I C 1995 Phys. Rev. Lett. 74 203 Dio23 Diósi L 2023 Preprint arXiv::2302.13418 Dio18 Diósi L 2018 Entropy 20 496; 2019 J. Phys. Conf. Ser. 1275 Dio92 Diósi L 1992 Quantum Chaos—Quantum Measurement (Springer) p 299 Pen14 Penrose R 2014 Found. Phys. 44 557

http://arxiv.org/abs/2303.07190v1

20230313153151

Transverse single-spin asymmetry of midrapidity $π^{0}$ and $η$ mesons in $p$+Au and $p$+Al collisions at $\sqrt{s_{_{NN}}}=$ 200 GeV

nucl-ex

[PHENIX Spokesperson: ]akiba@rcf.rhic.bnl.gov Deceased PHENIX Collaboration Presented are the first measurements of the transverse single-spin asymmetries (A_N) for neutral pions and eta mesons in p+Au and p+Al collisions at √(s__NN)=200 GeV in the pseudorapidity range |η|<0.35 with the PHENIX detector at the Relativistic Heavy Ion Collider. The asymmetries are consistent with zero, similar to those for midrapidity neutral pions and eta mesons produced in p+p collisions. These measurements show no evidence of additional effects that could potentially arise from the more complex partonic environment present in proton-nucleus collisions. Transverse single-spin asymmetry of midrapidity π^0 and η mesons in p+Au and p+Al collisions at √(s__NN)= 200 GeV L. Zou March 30, 2023 ================================================================================================================= § INTRODUCTION Transverse single-spin asymmetries (TSSAs) in particle production for hadronic collisions involving a transversely polarized proton result from nonperturbative spin-momentum correlations in the proton and/or the process of hadronization <cit.>. For recent discussions of TSSAs measured in polarized collisions at the Relativistic Heavy Ion Collider (RHIC) and the possible mechanisms contributing to them, see Refs. <cit.>. In hadronic collisions involving a nucleus, the underlying partonic origins of the asymmetries could be affected by the presence of more complex quantum-chromodynamics environments. For example, relations between TSSAs and the physics of small parton momentum fractions have been proposed, in particular how comparisons of asymmetries measured in and collisions for forward hadron production could be used to probe gluon saturation effects in the nucleus <cit.>. Further theoretical works have explored these ideas <cit.>. At RHIC, TSSA measurements for proton-nucleus collisions have been performed for forward charged hadrons <cit.> and forward J/ψ mesons <cit.> by PHENIX, and for forward π^0 production by STAR <cit.>, revealing some nuclear dependencies that remain to be understood in detail. PHENIX has additionally measured the TSSAs for far forward neutron production, with the observed nuclear dependence of the asymmetries understood to be due to the interplay of hadronic and electromagnetic interactions in ultra-peripheral collisions <cit.>. No experimental measurements exist and very little theoretical work has been done to explore possible nuclear effects for midrapidity TSSA observables, which can only be studied at RHIC. § ANALYSIS This brief article reports the first measurement of the TSSAs of neutral pions and eta mesons in proton-Gold () and proton-Aluminum () collisions at =200 GeV at midrapidity (|η|<0.35). The data were taken in 2015 at RHIC and total integrated luminosities of approximately 202 and 690 nb^-1, respectively, were collected. Measurements were performed with collisions of a vertically polarized proton beam on an ion beam (Au or Al). The proton or ion bunches are separated by 106 ns in the RHIC rings. Each polarized proton bunch is assigned a polarization direction, either up or down, so that measurements with both spin directions can be performed nearly simultaneously. This significantly reduces any possible systematic uncertainties related to the detector performance with time. The average proton beam polarization was 0.60 and 0.57 in and collisions respectively <cit.>, with a relative uncertainty of 3% due to uncertainty in polarization normalization. The data analysis procedure follows almost exactly from the recent TSSA measurement for π^0 and η mesons in √(s) = 200 GeV polarized p+p collisions <cit.>, with the distinction that only the proton beam is polarized in collisions. Only events with a collision z-vertex within ± 30  cm from the nominal collision point were selected. The collision or minimum-bias trigger, as well as vertex position, were determined by two beam-beam counters (BBC) located at ± 144  cm from the nominal collision point along the beam line, and covering the pseudorapidity range 3.1<|η|<3.9 with full azimuthal coverage. Neutral pions and eta mesons were reconstructed through their two-photon decay in the electromagnetic calorimeters (EMCal) of PHENIX. The EMCal is located in two nearly back-to-back central arm spectrometers (west and east), each covering Δϕ=π/2 in azimuth and ±0.35 in pseudorapidity. The EMCal comprises two types of calorimeters, six sectors of sampling lead-scintillator (PbSc) calorimeter and 2 sectors of Čerenkov lead-glass (PbGl) calorimeter <cit.>. The two calorimeter systems have different granularity (Δϕ×Δη =0.011 × 0.011 in PbSc and 0.008 × 0.008 in PbGl) and also have a different response to charged hadrons, which provides important systematic cross-checks for the measurement. The PHENIX EMCal was also used to generate a high-photon trigger to tag events with a high-energy cluster in the EMCal. The high-photon trigger (with an energy threshold of 1.5 GeV) in coincidence with a minimum-bias trigger was used to collect the and η statistics in this analysis. The efficiency of such a trigger for 's increased from 20% at =3 to 90% at >6, with the plateaued efficiency level defined by the acceptance of the live trigger tiles. Photons were identified in the EMCal by placing selection criteria on the shower profile and time-of-flight (TOF) with | TOF| < 5 ns, and with a minimum energy selection of 0.5 GeV to reduce the contribution from electronic noise in the EMCal, and combinatorial background in π^0 and η reconstruction. A charged track veto was also implemented to eliminate clusters that are geometrically associated with a track, in order to suppress the background from electrons and charged hadrons. Photon pairs were reconstructed by finding a high-p_T trigger photon, and pairing it with another photon from the same event and spectrometer arm. Photon pairs passing an energy-asymmetry requirement: α = |E_1 - E_2|/(E_1+E_2)<0.8 were selected for further analysis. fig:invmass shows the two-photon invariant mass distributions around the and η peaks for photon pairs within 4<p_T<5 GeV/c in the west central arm spectrometer for and collisions. The and η meson yields were defined to be within the signal-invariant mass window ([blue] leftward-hatched regions in fig:invmass) of ± 25 and ± 70 MeV/c^2 from the and η mass peaks, respectively, in the two-photon invariant mass distribution for each bin. The side-band regions used to approximate the combinatorial background under the signal peak ([red] rightward-hatched regions in fig:invmass) are defined as 47–97 and 177–227 MeV/c^2 for the and 300–400 and 700–800 MeV/c^2 for the η mesons. The same signal and side-band regions were used in a previous PHENIX analysis <cit.>. The combinatorial background for the two-photon invariant mass spectrum (described by a third-order polynomial and shown as the [green] solid lines in fig:invmass) was used to quantify the fraction of background existing under the signal peaks. For the , this ranged from 14% (13%) to 6% (6%) from the lowest to the highest bins in () collisions, while for the η, the combinatorial background under the signal peak ranged from 79% (77%) to 48% (43%) in () collisions. Similar to the recent p+p π^0 and η TSSA analysis <cit.>, the transverse single-spin asymmetry A_N is determined with the “relative-luminosity” formula, which is calculated separately for the two detector arms. This yields measurements from two independent data sets that are verified for consistency and then averaged to obtain the final result. The equation for the relative-luminosity TSSA is A_N = 1/P<cos(ϕ)>N^↑-ℛN^↓/N^↑+ℛN^↓, where P is the beam polarization, and ℛ is the relative luminosity, defined as the ratio of integrated luminosities between the bunches with ↑ and ↓ spin states. <cos(ϕ)> is the acceptance factor which reflects the detector azimuthal coverage, calculated separately for each p_T bin and spectrometer arm, and N refers to the yields, with the arrows referring to the up (↑) or down (↓) polarization of the proton beam. Another method to calculate the asymmetry is the “square-root” formula, which is used as a cross-check. The square-root formula is defined as A_N = 1/P<cos(ϕ)>√(N_L^↑N_R^↓)-√(N_L^↓N_R^↑)/√(N_L^↑N_R^↓)+√(N_L^↓N_R^↑), and it is used to calculate the asymmetry for both spectrometer arms simultaneously, where the L and R subscripts of N correspond to yields measured in the left and right detector arms respectively, with respect to the polarized-proton-going direction. This results in only one measurement of A_N that can be compared with the weighted average of the left and right relative-luminosity asymmetry calculation. The comparison showed that results using eq:relan and eq:sqrtan were statistically consistent, with any deviations assigned as a conservative systematic uncertainty. As an additional cross-check, the TSSA is calculated as a function of ϕ, in which case a cosine modulation is fit to extract the asymmetry. This was found to be statistically consistent with the main asymmetry results. The measured asymmetries were corrected for the background as follows, A_N^ sig = A_N-r · A_N^ BG/1-r, where r is the background fraction under the or η peaks, calculated from the background fits ([green] solid lines) shown in fig:invmass. A_N^ BG is the background asymmetry, which was evaluated from the side bands on both sides of the and η peaks, also shown in fig:invmass. The background asymmetry was consistent with zero in all bins and all collision systems for both the and η mesons. Asymmetries were calculated separately for each accelerator fill, during which the detector performance and beam conditions are considered to be relatively stable. The final asymmetry was obtained from the weighted average over the accelerator fills. The possible sources of systematic uncertainties considered are 1) the background contribution r in eq:sb, and 2) possible variation in detector performance and beam conditions. A systematic uncertainty on the background fraction is quantified by varying the fit ranges used to calculate r, and computing how much the background corrected asymmetry changes. The variations in detector performance and beam conditions were tested with a technique known as “bunch shuffling,” and from the comparison of the results calculated with the “relative-luminosity formula" eq:relan and the “square-root formula" eq:sqrtan. The asymmetries calculated with the different formulas were found to be statistically consistent, but a conservative systematic uncertainty is still applied based on the difference. In the bunch shuffling procedure, the polarization of each bunch is randomly assigned to be up or down, and a distribution of A_N in each p_T bin is obtained by repeating the procedure 10,000 times <cit.>. While most p_T bins were found to be consistent with the statistical variation, some of the lower p_T bins for both the π^0 and the η in both collision systems included up to 15% variation beyond what was expected from statistical fluctuations. To account for this effect, an additional systematic uncertainty was assigned in any p_T bin showing variations significantly beyond expected statistical fluctuations. § RESULTS AND DISCUSSION fig:ANpT shows the measurement of A_N for and η mesons in and collisions at = 200 GeV. The measured asymmetries in and are consistent with zero across the entire range for both the and η mesons. Table <ref> lists the asymmetries, statistical uncertainties, and total systematic uncertainties for the π^0 and η mesons in and collisions. The TSSA measurements presented here probe the complex dynamics of partons within a nucleus. Measurements of asymmetries with heavy nuclei have not been performed at collider energies before 2015. Therefore, it is unclear to what extent the nuclear environment affects TSSAs. Collisions with a nucleus explore spin-momentum correlations in an environment where larger multiplicities and stronger color fields could play an additional role. In a factorized picture, initial-state spin-momentum correlations in the polarized proton cannot be affected by the presence of a nucleus; however, it is possible for final-state spin-momentum correlations in the process of hadronization to be affected as the scattered parton passes through the nuclear matter. Allowing for factorization-breaking effects, the larger color field of the nuclear remnant in collisions as compared to the proton remnant in collisions could potentially modify the observed asymmetries <cit.>. Neutral-pion measurements in the forward region <cit.> and charged-hadron measurements in the intermediate rapidity region <cit.> show sizable TSSAs in collisions, with moderate nuclear modifications in for the former and strong nuclear modifications in for the latter. In contrast, the π^0 and η meson asymmetries at midrapidity are consistent with zero in all collision systems, showing no difference between and collisions. § SUMMARY The data presented here were motivated by the outstanding questions regarding the physical origin of transverse single-spin asymmetries. The TSSAs of midrapidity and η mesons were measured in , and collisions at =200 GeV by the PHENIX experiment at RHIC. The measured asymmetries are consistent with zero up to very high precision in both collision systems for both meson species. The data presented here will contribute to the understanding of transverse spin phenomena in the more complex environment present in proton-nucleus collisions. In particular, we find that at midrapidity the presence of a heavy nucleus in the collision does not significantly modify the magnitude of the measured TSSAs. We thank the staff of the Collider-Accelerator and Physics Departments at Brookhaven National Laboratory and the staff of the other PHENIX participating institutions for their vital contributions. We acknowledge support from the Office of Nuclear Physics in the Office of Science of the Department of Energy, the National Science Foundation, Abilene Christian University Research Council, Research Foundation of SUNY, and Dean of the College of Arts and Sciences, Vanderbilt University (U.S.A), Ministry of Education, Culture, Sports, Science, and Technology and the Japan Society for the Promotion of Science (Japan), Natural Science Foundation of China (People's Republic of China), Croatian Science Foundation and Ministry of Science and Education (Croatia), Ministry of Education, Youth and Sports (Czech Republic), Centre National de la Recherche Scientifique, Commissariat à l'Énergie Atomique, and Institut National de Physique Nucléaire et de Physique des Particules (France), J. Bolyai Research Scholarship, EFOP, the New National Excellence Program (ÚNKP), NKFIH, and OTKA (Hungary), Department of Atomic Energy and Department of Science and Technology (India), Israel Science Foundation (Israel), Basic Science Research and SRC(CENuM) Programs through NRF funded by the Ministry of Education and the Ministry of Science and ICT (Korea). Ministry of Education and Science, Russian Academy of Sciences, Federal Agency of Atomic Energy (Russia), VR and Wallenberg Foundation (Sweden), University of Zambia, the Government of the Republic of Zambia (Zambia), the U.S. Civilian Research and Development Foundation for the Independent States of the Former Soviet Union, the Hungarian American Enterprise Scholarship Fund, the US-Hungarian Fulbright Foundation, and the US-Israel Binational Science Foundation. 33 fxundefined [1] ifx#1 fnum [1] #1firstoftwo secondoftwo fx [1] #1firstoftwo secondoftwo noop [0]secondoftwo ref[1]@startlink#1@href href[1]#1@endlink anitize@url [0]` 12`$12`&12`#12`1̂2`_12`%12 startlink[1] endlink[0] rl [1]href #1 @bib@innerbibempty [Anselmino et al.(2020)Anselmino, Mukherjee, and Vossen]TSSA_review author author M. Anselmino, author A. Mukherjee, and author A. Vossen, title title Transverse spin effects in hard semi-inclusive collisions, https://doi.org/10.1016/j.ppnp.2020.103806 journal journal Prog. Part. Nucl. Phys. volume 114, pages 103806 (year 2020)NoStop [Acharya et al.()Acharya et al.]PPG246 author author U. Acharya et al. (collaboration PHENIX Collaboration), @noop title Improving constraints on gluon spin-momentum correlations in transversely polarized protons via midrapidity open-heavy-flavor electrons in p^↑+p collisions at √(s)=200 GeV, note arXiv:2204.12899NoStop [Acharya et al.(2022a)Acharya et al.]PPG247 author author U. Acharya et al. (collaboration PHENIX Collaboration), title title Transverse-single-spin asymmetries of charged pions at midrapidity in transversely polarized p+p collisions at √(s)=200 GeV, https://doi.org/10.1103/PhysRevD.105.032003 journal journal Phys. Rev. D volume 105, pages 032003 (year 2022a)NoStop [Acharya et al.(2021a)Acharya et al.]PPG235 author author U. Acharya et al. (collaboration PHENIX Collaboration), title title Probing Gluon Spin-Momentum Correlations in Transversely Polarized Protons through Midrapidity Isolated Direct Photons in p^↑+p Collisions at √(s)=200 GeV, https://doi.org/10.1103/PhysRevLett.127.162001 journal journal Phys. Rev. Lett. volume 127, pages 162001 (year 2021a)NoStop [Acharya et al.(2021b)Acharya et al.]PPG234 author author U. Acharya et al. (collaboration PHENIX Collaboration), title title Transverse single-spin asymmetries of midrapidity π^0 and η mesons in polarized p+p collisions at √(s)=200 GeV, https://doi.org/10.1103/PhysRevD.103.052009 journal journal Phys. Rev. D volume 103, pages 052009 (year 2021b)NoStop [Acharya et al.(2021c)Acharya et al.]PPG238 author author U. Acharya et al. (collaboration PHENIX Collaboration), title title Transverse momentum dependent forward neutron single spin asymmetries in transversely polarized p+p collisions at √(s) = 200 GeV, https://doi.org/10.1103/PhysRevD.103.032007 journal journal Phys. Rev. D volume 103, pages 032007 (year 2021c)NoStop [Abdallah et al.(2022)Abdallah et al.]STAR:2022hqg author author M. Abdallah et al. (collaboration STAR Collaboration), title title Azimuthal transverse single-spin asymmetries of inclusive jets and identified hadrons within jets from polarized pp collisions at √(s) = 200 GeV, https://doi.org/10.1103/PhysRevD.106.072010 journal journal Phys. Rev. D volume 106, pages 072010 (year 2022)NoStop [Adam et al.(2021a)Adam et al.]STAR:2020nnl author author J. Adam et al. (collaboration STAR Collaboration), title title Measurement of transverse single-spin asymmetries of π^0 and electromagnetic jets at forward rapidity in 200 and 500 GeV transversely polarized proton-proton collisions, https://doi.org/10.1103/PhysRevD.103.092009 journal journal Phys. Rev. D volume 103, pages 092009 (year 2021a)NoStop [Adamczyk et al.(2018a)Adamczyk et al.]STAR:2017wsi author author L. Adamczyk et al. (collaboration STAR Collaboration), title title Transverse spin-dependent azimuthal correlations of charged pion pairs measured in p^↑+p collisions at √(s) = 500 GeV, https://doi.org/10.1016/j.physletb.2018.02.069 journal journal Phys. Lett. B volume 780, pages 332 (year 2018a)NoStop [Adamczyk et al.(2018b)Adamczyk et al.]STAR:2017akg author author L. Adamczyk et al. (collaboration STAR Collaboration), title title Azimuthal transverse single-spin asymmetries of inclusive jets and charged pions within jets from polarized-proton collisions at √(s) = 500 GeV, https://doi.org/10.1103/PhysRevD.97.032004 journal journal Phys. Rev. D volume 97, pages 032004 (year 2018b)NoStop [Kang and Yuan(2011)]Kang:2011ni author author Z. B. Kang and author F. Yuan, title title Single Spin Asymmetry Scaling in the Forward Rapidity Region at RHIC, https://doi.org/10.1103/PhysRevD.84.034019 journal journal Phys. Rev. D volume 84, pages 034019 (year 2011)NoStop [Kovchegov and Sievert(2012)]Kovchegov:2012ga author author Y. V. Kovchegov and author M. D. Sievert, title title A New Mechanism for Generating a Single Transverse Spin Asymmetry, https://doi.org/10.1103/PhysRevD.86.034028 journal journal Phys. Rev. D volume 86, pages 034028 (year 2012), note [Phys. Rev. D 86, 079906(E) (2012)]NoStop [Schäfer and Zhou(2014a)]Schafer:2014xpa author author A. Schäfer and author J. Zhou, title title Color entanglement for γ-jet production in polarized pA collisions, https://doi.org/10.1103/PhysRevD.90.094012 journal journal Phys. Rev. D volume 90, pages 094012 (year 2014a)NoStop [Schäfer and Zhou(2014b)]Schafer:2014zea author author A. Schäfer and author J. Zhou, title title Transverse single spin asymmetry in direct photon production in polarized pA collisions, https://doi.org/10.1103/PhysRevD.90.034016 journal journal Phys. Rev. D volume 90, pages 034016 (year 2014b)NoStop [Kovchegov and Sievert(2016)]Kovchegov:2015zha author author Y. V. Kovchegov and author M. D. Sievert, title title Calculating TMDs of a Large Nucleus: Quasi-Classical Approximation and Quantum Evolution, https://doi.org/10.1016/j.nuclphysb.2015.12.008 journal journal Nucl. Phys. B volume 903, pages 164 (year 2016)NoStop [Zhou(2015)]Zhou:2015ima author author J. Zhou, title title Transverse single spin asymmetry in Drell-Yan production in polarized pA collisions, https://doi.org/10.1103/PhysRevD.92.014034 journal journal Phys. Rev. D volume 92, pages 014034 (year 2015)NoStop [Hatta et al.(2016)Hatta, Xiao, Yoshida, and Yuan]Hatta:2016wjz author author Y. Hatta, author B. W. Xiao, author S. Yoshida, and author F. Yuan, title title Single Spin Asymmetry in Forward pA Collisions, https://doi.org/10.1103/PhysRevD.94.054013 journal journal Phys. Rev. D volume 94, pages 054013 (year 2016)NoStop [Hatta et al.(2017)Hatta, Xiao, Yoshida, and Yuan]Hatta:2016khv author author Y. Hatta, author B. W. Xiao, author S. Yoshida, and author F. Yuan, title title Single spin asymmetry in forward pA collisions II: Fragmentation contribution, https://doi.org/10.1103/PhysRevD.95.014008 journal journal Phys. Rev. D volume 95, pages 014008 (year 2017)NoStop [Zhou(2017)]Zhou:2017sdx author author J. Zhou, title title Single spin asymmetries in forward p-p/A collisions revisited: the role of color entanglement, https://doi.org/10.1103/PhysRevD.96.034027 journal journal Phys. Rev. D volume 96, pages 034027 (year 2017)NoStop [Benić and Hatta(2019)]Benic:2018amn author author S. Benić and author Y. Hatta, title title Single spin asymmetry in forward pA collisions: Phenomenology at RHIC, https://doi.org/10.1103/PhysRevD.99.094012 journal journal Phys. Rev. D volume 99, pages 094012 (year 2019)NoStop [Kovchegov and Santiago(2020)]Kovchegov:2020kxg author author Y. V. Kovchegov and author M. G. Santiago, title title Lensing mechanism meets small- x physics: Single transverse spin asymmetry in p^↑+p and p^↑+A collisions, https://doi.org/10.1103/PhysRevD.102.014022 journal journal Phys. Rev. D volume 102, pages 014022 (year 2020)NoStop [Benić et al.(2022)Benić, Horvatić, Kaushik, and Vivoda]Benic:2022qzv author author S. Benić, author D. Horvatić, author A. Kaushik, and author E. A. Vivoda, title title Odderon mechanism for transverse single spin asymmetry in the Wandzura-Wilczek approximation, https://doi.org/10.1103/PhysRevD.106.114025 journal journal Phys. Rev. D volume 106, pages 114025 (year 2022)NoStop [Aidala et al.(2019)Aidala et al.]PPG215 author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Nuclear Dependence of the Transverse Single-Spin Asymmetry in the Production of Charged Hadrons at Forward Rapidity in Polarized p+p, p+Al, and p+Au Collisions at √(s__NN)=200 GeV, https://doi.org/10.1103/PhysRevLett.123.122001 journal journal Phys. Rev. Lett. volume 123, pages 122001 (year 2019)NoStop [Abdulameer et al.()Abdulameer et al.]PPG226 author author N. J. Abdulameer et al. (collaboration PHENIX Collaboration), @noop title Transverse single-spin asymmetry of midrapidity charged hadrons in p+Au and p+Al collisions at √(s__NN)=200 GeV, note (to be published)NoStop [Aidala et al.(2018a)Aidala et al.]PPG211 author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Single-spin asymmetry of J/ψ production in p+p, p+Al, and p+Au collisions with transversely polarized proton beams at √(s__NN)=200 GeV, https://doi.org/10.1103/PhysRevD.98.012006 journal journal Phys. Rev. D volume 98, pages 012006 (year 2018a)NoStop [Adam et al.(2021b)Adam et al.]STAR:2020grs author author J. Adam et al. (collaboration STAR Collaboration), title title Comparison of transverse single-spin asymmetries for forward π^0 production in polarized pp, pAl and pAu collisions at nucleon pair c.m. energy √(s_NN)= 200 GeV, https://doi.org/10.1103/PhysRevD.103.072005 journal journal Phys. Rev. D volume 103, pages 072005 (year 2021b)NoStop [Aidala et al.(2018b)Aidala et al.]PPG203 author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Nuclear Dependence of the Transverse-Single-Spin Asymmetry for Forward Neutron Production in Polarized p+A Collisions at √(s_NN)=200GeV, https://doi.org/10.1103/PhysRevLett.120.022001 journal journal Phys. Rev. Lett. volume 120, pages 022001 (year 2018b)NoStop [Acharya et al.(2022b)Acharya et al.]PPG244 author author U. Acharya et al. (collaboration PHENIX Collaboration), title title Transverse single spin asymmetries of forward neutrons in p+p, p+Al and p+Au collisions at √(s__NN)=200 GeV as a function of transverse and longitudinal momenta, https://doi.org/10.1103/PhysRevD.105.032004 journal journal Phys. Rev. D volume 105, pages 032004 (year 2022b)NoStop [Schmidke et al.(2018)Schmidke et al.]RHIC_polarimetry author author W. B. Schmidke et al. (collaboration The RHIC Polarimetry Group), https://doi.org/10.2172/1473643 title RHIC polarization for Runs 9-17 (year 2018)NoStop [Aphecetche et al.(2003)Aphecetche et al.]Aphecetche:2003zr author author L. Aphecetche et al. (collaboration PHENIX Collaboration), title title PHENIX calorimeter, https://doi.org/10.1016/S0168-9002(02)01954-X journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 499, pages 521 (year 2003)NoStop [Adare et al.(2014)Adare et al.]bunchShuffling author author A. Adare et al. (collaboration PHENIX Collaboration), title title Inclusive double-helicity asymmetries in neutral-pion and eta-meson production in p+p collisions at √(s)=200GeV, https://doi.org/10.1103/PhysRevD.90.012007 journal journal Phys. Rev. D volume 90, pages 012007 (year 2014)NoStop [Rogers and Mulders(2010)]fact_break_2010 author author T. C. Rogers and author P. J. Mulders, title title No Generalized TMD-Factorization in Hadro-Production of High Transverse Momentum Hadrons, https://doi.org/10.1103/PhysRevD.81.094006 journal journal Phys. Rev. D volume 81, pages 094006 (year 2010)NoStop [Rogers(2013)]extra_asymmetries author author T. C. Rogers, title title Extra spin asymmetries from the breakdown of transverse-momentum-dependent factorization in hadron-hadron collisions, https://doi.org/10.1103/PhysRevD.88.014002 journal journal Phys. Rev. D volume 88, pages 014002 (year 2013)NoStop

http://arxiv.org/abs/2303.07329v1

20230313175003

Application of targeted maximum likelihood estimation in public health and epidemiological studies: a systematic review

stat.AP

1]Matthew J. Smith* 2]Rachael V. Phillips 1,3]Miguel Angel Luque-Fernandez^± 1]Camille Maringe^± Smith et al. [1]Inequalities in Cancer Outcomes Network, London School of Hygiene and Tropical Medicine, London, England, United Kingdom [2]Division of Biostatistics, School of Public Health, University of California at Berkeley, California, United States of America [3]Department of Statistics and Operations Research, University of Granada, Granada, Spain *Matthew J. Smith. matt.smith@lshtm.ac.uk London School of Hygiene and Tropical Medicine, Keppel Street, London, England, United Kingdom, WC1E 7HT ^± Both authors contributed equally to this work. [Abstract] Background The Targeted Maximum Likelihood Estimation (TMLE) statistical data analysis framework integrates machine learning, statistical theory, and statistical inference to provide a least biased, efficient and robust strategy for estimation and inference of a variety of statistical and causal parameters. We describe and evaluate the epidemiological applications that have benefited from recent methodological developments. Methods We conducted a systematic literature review in PubMed for articles that applied any form of TMLE in observational studies. We summarised the epidemiological discipline, geographical location, expertise of the authors, and TMLE methods over time. We used the Roadmap of Targeted Learning and Causal Inference to extract key methodological aspects of the publications. We showcase the contributions to the literature of these TMLE results. Results Of the 81 publications included, 25% originated from the University of California at Berkeley, where the framework was first developed by Professor Mark van der Laan. By the first half of 2022, 70% of the publications originated from outside the United States and explored up to 7 different epidemiological disciplines in 2021-22. Double-robustness, bias reduction and model misspecification were the main motivations that drew researchers towards the TMLE framework. Through time, a wide variety of methodological, tutorial and software-specific articles were cited, owing to the constant growth of methodological developments around TMLE. Conclusions There is a clear dissemination trend of the TMLE framework to various epidemiological disciplines and to increasing numbers of geographical areas. The availability of R packages, publication of tutorial papers, and involvement of methodological experts in applied publications have contributed to an exponential increase in the number of studies that understood the benefits, and adoption, of TMLE. Application of targeted maximum likelihood estimation in public health and epidemiological studies: a systematic review [ March 30, 2023 ======================================================================================================================= § BACKGROUND Public health decisions across many clinical specialities are often informed by research exploring the relationship between exposures and patient health outcomes. To ascertain a causal relationship, randomised control trials (RCT) are considered the gold standard because, through randomisation of subjects to a treatment, they reduce the possibility of bias. Observational data offer invaluable opportunities to study causal relationships in contexts where clinical trials might prove infeasible or unethical, as well as for studying groups of the population typically excluded from trials or beyond the initial target population. Under correct adjustment for selection bias, missingness, interference, and confounding, observational data complements the evidence coming from RCTs. In both RCT and observational studies, a causal exposure-outcome relationship is presented with a causal estimand, such as the average treatment effect (ATE). Methodological statistical developments for causal inference attempt to produce the least biased estimate of the estimand and accurate inference. G-computation, Propensity Score (PS), and Inverse Probability of Treatment Weighting (IPTW) estimators rely on parametric modelling assumptions, which are susceptible to model misspecification (i.e., exclusion of the underlying and unknown data-generating distribution from the model, which can lead to biased estimates with small standard errors and thus misleading results). Double-robust methods, like Augmented Inverse Probability of Treatment Weighting (AIPTW) and Targeted Maximum Likelihood Estimation (TMLE), aim to minimise model misspecification by requiring estimation of both the outcome and exposure mechanisms. They provide a consistent estimator as long as either the outcome or exposure model is correctly specified. Double-robust methods often outperform single-robust methods in point and interval estimation.<cit.> TMLE, also known as targeted minimum loss-based estimation, was introduced by van der Laan and Rubin in 2006.<cit.> In general, TMLE is a two-step process that involves (1) initial estimation of the outcome and intervention models, and then (2) in a “targeting” step, uses information from them to optimise the bias-variance trade-off for the target estimand (e.g., ATE), rather than the whole outcome probability distribution. Furthermore, to avoid model misspecification, ensemble machine learning algorithms are used to estimate the initial models. In particular, the Super Learner (SL) algorithm for stacked ensemble machine learning is most commonly used as it is theoretically grounded and proven to perform optimally in large samples.<cit.> We lightly detail the technical steps involved in the TMLE of the ATE, i.e., the effect of a binary exposure A on a post-exposure outcome Y, adjusted by baseline covariates W.<cit.> The prediction function for the mean outcome Y, given exposure A and covariates W is estimated, most commonly, using SL. We could use this estimated prediction function, Ê[Y|A,W], to arrive at an estimate of the ATE. Specifically, we would obtain predicted outcomes under a counterfactual scenario where all subjects receive the exposure/treatment versus another scenario where no one receives it. The average difference between these predicted counterfactual outcomes is an estimate of the ATE. However, formal statistical inference (i.e., confidence intervals and p-values) cannot be obtained for this estimate and it is susceptible to residual confounding; the latter can be reduced by using the information on how each individual was assigned or allocated to each level of the exposure. We, therefore, estimate the function for predicting the probability of being exposed, given the covariates W, using SL (exposure model, i.e. propensity score). These first steps are common to other double-robust estimators of the ATE, such as AIPTW. We then calculate the so-called “clever covariate” for the ATE, which is the individual values of the binary exposure weighted by the predicted probabilities of the exposure, given W. This is similar to IPTW, except here we weight the predicted probability of each exposure level instead of the outcome. The fluctuation parameter (ϵ) describes the difference between the observed outcome Y and the initial predictions of the outcome from the outcome model. It is calculated through maximum likelihood estimation (MLE) by regressing the clever covariate on the observed outcome and incorporating as offset the predicted outcome values. When the fluctuation parameter is estimated as close to 0 there is little difference between the observed and predicted outcomes; thus, the propensity score does not provide additional information for the initial estimate of the outcome model because it was correctly specified. If the fluctuation parameter is not close to 0, then this indicates the presence of residual confounding in the initial estimate. The initial outcome model's predictions for each level of the binary exposure are updated using the fluctuation parameter ϵ as a weight, and the final ATE estimate is calculated from these updated estimates. The functional delta method based on the influence function can be used to derive the standard error of the ATE and construct Wald-type confidence intervals. Since 2006, the TMLE framework has experienced a growing number of theoretical and applied developments, and it has expanded further after a book that shared the TMLE framework to the international community of applied researchers was published in 2011.<cit.> Targeting specifically applied researchers, efforts were made to provide lay-language descriptions of the framework and exemplify its applications.<cit.> Furthermore, in 2018, a second book was published disseminating more advanced applications of the TMLE framework to data scientists with a particular focus on longitudinal settings.<cit.> TMLE is a robust framework for statistical analysis in clinical, observational and randomised studies. Since 2016, the American Causal Inference Conference has hosted a data challenge in which teams compete to estimate a causal effect in simulated data sets based on real-world data, such as from healthcare or education.<cit.> The competition is a proving ground for cutting-edge causal inference methods that have the potential to transform program evaluation. TMLE has consistently been a top-performing method.<cit.> The use of robust statistical methods is key to obtaining reliable results for public health and epidemiological research and maximising their benefit to society. Evidence shows that TMLE, by blending flexible machine learning methods and causal inference principles, is one such step towards robust causal claims that bear significant and practical effects. We reviewed the literature around public health and epidemiological applications of TMLE to date, alongside key TMLE developments over the last 20 years. We highlight the speed at which the field developed and spread through the scientific community, and identify areas for further development to increase the utility of the TMLE framework in epidemiological and applied research. § METHODS Protocol registration and reporting standards This study is reported using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guideline. We registered this systematic review with PROSPERO (ID: CRD42022328482). Information sources We searched the PubMed medical literature database for published epidemiological studies using TMLE in any epidemiological field (i.e., observational settings in biomedical sciences, including clinical epidemiology and public health). We searched for publications from any time up to 20th May 2022, the date the search was executed. The search strategy comprised of two specific groups of search terms focusing on TMLE and epidemiology. Relevant Mesh headings were included as were free-text terms, searched in the title, abstract or keyword fields. We used general and specific TMLE search terms, such as “targeted maximum likelihood estimation”, “targeted minimum loss-based estimation”, and “targeted machine learning”. Epidemiological search terms included “epidemiology”, “public health”, “population”, or “treatment”. The two specific groups of terms were combined with ‘AND’ to retrieve the final set of results. Search strategies were developed with an information specialist (MALF). The full search strategy is shown in Table <ref>. Eligibility criteria We excluded studies that did not report the use of TMLE as a tool to explore their (possibly causal) estimand of interest. We also excluded experimental studies, such as RCTs (n=15). We did not consider manuscripts that compared the performance of TMLE to other estimators, when there was no new development proposed and even if there was an applied question of interest, such as in Luque-Fernandez et al.<cit.> Studies were restricted to the English language and primary research studies. Secondary research studies, such as reviews and comments of TMLE, conference abstracts and brief reports, and preprints were not searched. We classified the retained manuscripts into observational, methodological and tutorial articles. TMLE methodological development articles and tutorials were considered separately, even if they contained a methodological development specifically designed to investigate an epidemiological question within the same article. We make reference to these methodological articles throughout this review, as they underpin the applied publications. Study selection, data extraction and management All retrieved publications were imported into the Endnote reference software where they were checked for duplication. Two of the three lead researchers (authors MJS, MALF and CM) were randomly allocated two-thirds of the 226 articles, to screen titles and abstracts of each publication independently and classify these into (1) observational, (2) methodological developments, (3) tutorial, (4) systematic review, (5) RCT, or (6) not relevant (Figure <ref>). Disagreements were discussed and adjudicated by the third independent reviewer, where necessary. Two researchers (authors MJS and CM) independently reviewed the full text of all eligible observational publications for data extraction. § RESULTS We found 226 unique publications published prior to 20th May 2022 in PubMed (Figure <ref>). Of these, 95 articles were a methodological development (including theoretical - or software-based), eight were tutorials, five were systematic reviews, and fifteen were RCTs. Of the 22 articles that were not relevant, three mentioned “TMLE” only in the author fields, one was a discussion of currently existing methods, twelve were assessments of learning (educational) programs that are targeted towards clinical environments, and six were comparisons of machine learning algorithms to other prediction models (without mention of TMLE). Overall, we focused on 81 observational studies in this systematic review for which full texts were available; six publications were not open-access and full texts were obtained from the corresponding authors. §.§ Dissemination and uptake of the TMLE framework There has been a growing uptake of the TMLE framework over time, with five or fewer applied publications per year until 2017, and up to 21 in 2021. The majority (58, 71.6%) of publications using TMLE were published in the last three and a half years, starting in 2019. Most studies (77, 95.1%) cited the authors of particular TMLE methods they apply, whereas four (4.9%) did not cite any TMLE references. The large majority of these first epidemiological studies benefit from the expert knowledge of an author who is (or was) part of Professor Mark van der Laan's lab. (Table <ref>) Of the 81 studies included, two-thirds were conducted in the United States of America (US) (53, 65.4%, Figure <ref>),<cit.> with 100% of articles before 2017 being published in the US and down to 50% of all 2021 articles. Publications from Europe or Scandinavia (10, 12.3%),<cit.> Africa (4, 4.9%),<cit.> Middle East (5, 6.2%),<cit.> and Oceania or Asia (8, 9.9%)<cit.> represent between 20% (in 2017) and 70% (in 2022, up to May 2022) of all applied studies published in the last 6 years. (Figure <ref>, Table <ref>) In the US, the majority of publications (28) were from California, including 20 from the University of California at Berkeley, where TMLE was first described. In the early years, the first authors tended to be qualitative academic experts, but we saw more variety in expertise and a larger number of practising clinicians leading observational studies in many epidemiological and public health fields in recent publications. The most common epidemiological sub-discipline was non-communicable disease (23, 28.4%),<cit.> followed by behavioural epidemiology (17, 21.0%),<cit.> and then infectious disease epidemiology (12, 14.8%).<cit.> Through time we see an uptake of TMLE in many more disciplines, such as pharmaco-epidemiology,<cit.> policy,<cit.> biomarker epidemiology,<cit.> environmental epidemiology,<cit.> occupational epidemiology,<cit.> and health economy.<cit.> We also studied the evolution of citations. When only methodological overviews of the TMLE framework were available, these were cited despite their heavy statistical requisite. Since 2016, tutorials were published and started to be cited alongside references for statistical packages. <cit.>(Table <ref>) Cohort study design<cit.> was the most prevalent design (42, 51.9%), followed by cross-sectional (32, 39.5%) (Appendix Table <ref>).<cit.> Other types of epidemiological design included case-control<cit.> and ecological.<cit.> Many articles reported results from other statistical methods, in addition to reporting those obtained from TMLE. Over one-quarter of the studies used adjusted parametric regression (24, 29.6%),<cit.> one sixth (12, 14.8%) used IPTW,<cit.> one (1.2%) used AIPTW,<cit.> three (3.7%) used non-parametric methods (e.g. Kaplan Meier),<cit.> and seven (8.6%) used unadjusted regression.<cit.> Some studies included more than one comparative method. SuperLearner (SL) provides a flexible machine learning approach to the estimation of the initial outcome and intervention models. Of the 81 articles, more than half (47, 58.0%) used the SL algorithm, <cit.> 18 (22.2%) used logistic regression,<cit.> and 16 (19.8%) did not specify the approach for the estimation of either the outcome or intervention model.<cit.> The average number of machine-learning algorithms considered by the SL was 6.3 (range 1 - 16), 19 different machine-learning algorithms were used across the articles (a machine-learning algorithm is a wrapper included within the SuperLearner<cit.> library in R software). Variance of point estimates obtained from TMLE were estimated using the influence function (n=19, 23.5%),<cit.> bootstrap (n=6, 7.4%),<cit.> and Wald tests (n=2, 2.5%),<cit.> while 54 (66.7%) studies did not specify how standard errors were obtained.<cit.> The Causal Inference Roadmap<cit.> contains seven recommended criteria to define a causal effect: (i) specify the scientific question, (ii) specify the causal model, (iii) define the target causal quantity, (iv) link the observed data to the causal model, (v) assess identifiability assumptions, (vi) estimate the target statistical parameters, and (vii) interpretation of the results. On average, 5.3 (SD 1.0) criteria were complete per article. We considered a version of the Targeted Learning Roadmap<cit.> that contains five criteria: (i) specify the observed data and describe the data-generating experiment, (ii) specify a statistical model representing a set of realistic assumptions about the underlying true probability distribution of the data, (iii) define a target estimand of the data distribution that “best" approximates the answer to the scientific question of interest, (iv) given statistical model and target estimand, construct an optimal plug-in estimator of the target estimand of the observed data distribution, while respecting the model, and (v) construct a confidence interval by estimating the sampling distribution of the estimator. When the scientific question of interest is causal, step (iii) of the Targeted Learning Roadmap incorporates steps (ii)–(v) of the Causal Inference Roadmap.<cit.> On average, 3.3 (SD 1.0) criteria were complete per article. Most studies have room to state the necessary content for at least one more criteria. Most publications (77, 95.1%) used R software to perform TMLE,<cit.> except four that used STATA.<cit.> Nonetheless, nine articles reported using another software tool (i.e., Stata/SAS/SPSS) alongside R for TMLE.<cit.> The most commonly used R software packages were tmle<cit.> (40, 49.4%) and ltmle<cit.> (16, 19.8%). §.§ Showcase of the TMLE framework in different disciplines In all disciplines and applications, applying the TMLE framework to their specific research question encouraged authors to review the strengths and limitations of their data and carefully consider how their data and setting might violate identifiability assumptions, which are assumptions necessary for causal inference but not TMLE. If, and only if, the identifiability assumptions are assumed to hold, the estimated effect is a causal effect. However, for observational studies, it cannot be known whether identifiability assumptions hold. Therefore, if an estimate is interpreted as a causal effect, then this interpretation should be accompanied by a discussion of the plausibility of identifiability assumptions. All disciplines and disease areas highlight issues with missing data and measurement errors and incorporate subject-matter knowledge. (Appendix Table <ref>) Model misspecification, which might result from imposing constraints that are unrealistic or not informed by subject-matter knowledge, was the first and foremost driver for using the TMLE framework. Three-quarters of the studies (n=68, 84.0%) provided at least one justification for using TMLE compared to another method. (Table <ref>) Standard regression techniques in settings with low incidence,<cit.> rare outcomes<cit.> or low sample size<cit.> may over-fit the data or not converge: careful SL specifications overcome these limitations.<cit.> Least biased and double-robust estimation were also widely cited advantages of the framework. These mean (i) the estimated parameter will be closest to the true value of our quantity of interest, even when our sample size is low, and (ii) only one of our initial models needs to be correctly specified. There was a range of disease areas covered in the 23 noncommunicable disease epidemiology studies. The appealing property of TMLE was that it is a semiparametric estimator, allowing the use of machine learning algorithms to minimise model misspecification.<cit.> Additionally, extensions of TMLE have developed ways to appropriately handle the dual nature of time-varying confounding, which have been utilised in longitudinal studies analysing data on depression,<cit.> survival from acute respiratory distress syndrome,<cit.> caries arising from intake of soda,<cit.> effects of smoking on rheumatoid arthritis,<cit.> and effects of asthma medication on asthma symptoms.<cit.> Improved predictive performance<cit.> and adjusting for informative censoring<cit.> were additional reasons for using TMLE. Furthermore, the extension of TMLE to case-control studies, in which sampling is biased with respect to the disease status, provided a platform for analysing the causal effect of reproductive factors on breast cancer by using case-control weighted TMLE.<cit.> In infectious disease epidemiology (IDE) articles, most were concerned with having a flexible modelling approach that does not impose assumptions on the functional form of the exposure-outcome relationship.<cit.> A key feature of the IDE subdiscipline is that baseline confounders and exposures may change over time and can obscure the causal effect of interest.<cit.> Standard survival modelling assumes that censoring and survival processes are independent, which is likely violated in this setting, and it assumes there is no time-dependent confounding.<cit.> TMLEs with a working marginal structural model and for time-to-event outcomes permit evaluation of the effect of an exposure at multiple time points, which is beneficial when the interpretation of causal effects from hazard models is often difficult.<cit.> Other studies have overcome this issue by using TMLE in a target trial framework or case-cohort studies.<cit.> In behavioural epidemiology manuscripts, the behavioural nature of the topics covered implied that RCTs are mostly unethical, bear prohibitive costs or have very small sample sizes. There are several key challenges for using observational data to study the causal effects of childhood adversities,<cit.> physical activity,<cit.> alcohol consumption<cit.> or supply <cit.> on various outcomes, including fractures,<cit.> mental health,<cit.> asthma,<cit.> and pregnancy outcomes.<cit.> They include a risk for reverse causation;<cit.> high dimensional data and in particular, multidimensional exposures;<cit.> and measurement error resulting from self-reported exposures or outcomes.<cit.> Longitudinal relationships and time-varying confounding, where confounders of the effect of an exposure on an outcome can themselves be affected by prior exposures, as well as sample attrition,<cit.> are particular challenges faced by survey data that are collected in consecutive waves.<cit.> TMLE adjusts for time-varying confounders affected by prior exposure and employs a doubly robust estimation approach that allows for flexible model fitting. Additionally, as pointed out in 2016 by Ahern et al.,<cit.> “TMLE with machine learning addresses the challenge of a multidimensional exposure because it facilitates `learning' from the data the strength of the relations between each adversity [dimensions of the exposure] and outcome, incorporating any interactions or nonlinearity, specific to each [sub-group]". The field of biomarker epidemiology is driven by the search for sets of candidate biomarkers that are important in determining given outcomes. Ranking the contributions of these candidate biomarkers is also of interest. Some studies used TMLE to measure variable importance in biomarker research<cit.> and in other fields.<cit.> Dimension reduction for the estimation of causal effects is an aim in some biomarker examples.<cit.> In the models presented in the publications, there are complex joint effects to consider in large correlated data, as well as longitudinal patterns and time-dependent confounding.<cit.> Furthermore, two manuscripts present covariate selection algorithms ahead of causal effect estimation.<cit.> Research published in environmental epidemiology highlights challenges around the clear definitions of exposure and outcomes,<cit.> as there are likely many proxy and surrogate measures of exposure,<cit.> coupled with potential exposure misclassification and measurement errors.<cit.> Nonetheless, TMLE was successfully applied to determine either causal attributable risk,<cit.> or risk differences.<cit.> The only observational study of TMLE in health economics explored the relationship between financial resources leading to food insecurity and healthcare expenditure in a pay-to-access healthcare system. It uses ecological measures of exposure and outcome and leads to evidence for policy.<cit.> Two publications focused on occupational epidemiology.<cit.> A key aspect of occupational epidemiology is accounting for the healthy worker survivor effect: a bias arising due to healthier workers accruing more exposure over time. These studies looked at exposure to particulate matter from aluminium or metalworking fluids in metal factory workers, which varied depending on the length of employment. Both studies benefited from TMLE's flexibility to allow for time-varying confounding of the exposure. The field of pharmacoepidemiology is concerned with assessing treatment's efficacy in real-world settings and monitoring long-term side effects of treatments. Both objectives would be either impractical or too costly to study in RCTs, given the limited follow-up time available in clinical trials. TMLE has been used in this setting, as it provides a robust procedure for estimation.<cit.> In particular, the flexibility of TMLE, provided through the specification of a diverse and rich set of machine learning algorithms in the SL, is crucial for appropriately adjusting for confounding in observational studies.<cit.> Policy epidemiology assesses the effects of population programs or mandates. Lack of randomisation, such as in studies examining the association between speciality probation and public safety outcomes,<cit.> leads to an imbalance in the covariate distribution by exposure levels. Studies of cost-effectiveness may involve dealing with outliers which can be addressed with TMLE.<cit.> Other challenges include zero-inflation, such as the assessment of the effect of primary care physician density on arthroplasty outcomes, in which some areas had zero density.<cit.> This is dealt with by using a mixture of models to assess the probability of non-exposure (i.e., very low density).<cit.> Other policy studies presented challenges around missing data,<cit.> reliance on epidemic modelling assumptions,<cit.> target trial emulation,<cit.> or infeasible randomisation process.<cit.> 23cm! [ht] Distribution of observational papers by year of publication and selected characteristics 1l 28cYear of publication 1l 2-29 2c2009 2c2010 2c2011 2c2012 2c2013 2c2014 2c2015 2c2016 2c2017 2c2018 2c2019 2c2020 2c2021 2c2022 ^⋆ Total 2-30 N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) N 2-30 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1lPublications 2 1 1 1 2 1 5 4 6 12 16 21 9 81 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l gray1lTMLE expert (author)^± 2 (100) 1 (100) 1 (100) 1 (100) 1 (50) 1 (100) 3 (60) 1 (25) 4 (67) 6 (50) 6 (38) 5 (24) 1 (11) 33 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l 1l gray1lUSA-based publication 2 (100) 1 (100) 1 (100) 2 (100) 1 (100) 5 (100) 3 (75) 6 (100) 9 (75) 10 (63) 11 (52) 2 (22) 52 1l gray1lDiscipline Behavioural epi 1 (100) 1 (20) 1 (17) 2 (17) 6 (38) 4 (19) 2 (22) 17 Biomarker 2 (100) 3 (60) 1 (6) 2 (10) 8 Environmental epi 1 (100) 1 (25) 1 (17) 3 (25) 1 (5) 7 Health economy 1 (8) 1 Infectious disease 1 (50) 1 (25) 1 (8) 4 (25) 2 (10) 3 (33) 12 Non-Communicable Disease 1 (100) 1 (50) 3 (50) 3 (25) 4 (25) 9 (43) 2 (22) 23 Occupational epi 1 (100) 1 (8) 2 Pharmaco-epi 1 (25) 1 (6) 2 (10) 4 Policy 1 (20) 1 (25) 1 (17) 1 (8) 1 (5) 2 (22) 7 1l gray1lMotivations ^† Bias 1 (50) 2 (100) 1 (100) 1 (20) 1 (25) 2 (33) 5 (42) 5 (31) 12 (57) 6 (67) 36 Double-robust 1 (50) 1 (20) 1 (25) 2 (33) 2 (17) 5 (31) 6 (29) 4 (44) 22 Efficient 2 (17) 5 (31) 2 (10) 3 (33) 12 Finite sample 1 (100) 2 (10) 3 Model misspecification 1 (50) 1 (100) 1 (100) 3 (60) 1 (25) 2 (33) 1 (8) 3 (17) 1 (57) 2 (22) 16 Positivity assumption 1 (17) 1 (8) 1 (6) 3 Power 1 (6) 1 None specified 1 (100) 1 (20) 2 (50) 5 (42) 4 (25) 2 (10) 1 (11) 16 1l gray1lExpertise first author Biostatistician 1 (50) 1 (50) 1 (100) 1 (25) 2 (33) 2 (17) 2 (13) 4 (19) 2 (22) 16 Epidemiologist 1 (100) 1 (100) 1 (50) 3 (60) 1 (17) 4 (33) 6 (38) 7 (33) 1 (11) 25 MD 1 (20) 2 (50) 1 (8) 1 (6) 5 (24) 2 (22) 12 MD, MPH 1 (8) 1 (6) 2 MD, PhD 1 (100) 3 (25) 2 (13) 3 (14) 3 (33) 12 Other 2 (13) 2 (10) 4 PhD 1 (50) 1 (25) 2 (33) 1 (6) 5 Not known 1 (20) 1 (17) 1 (8) 1 (6) 1 (11) 5 1l gray1lCitations ^⊡ Overall TMLE method. 2 (100) 1 (50) 1 (100) 5 (56) 1 (33) 4 (33) 4 (50) 3 (33) 6 (19) 13 (42) 20 (38) 4 (16) 64 Specific TMLE method. 1 (50) 1 (100) 4 (44) 1 (33) 6 (50) 1 (13) 3 (33) 15 (47) 7 (23) 13 (25) 8 (32) 60 Tutorial 1 (11) 7 (22) 7 (23) 14 (27) 7 (28) 36 R software 1 (33) 2 (17) 3 (38) 2 (22) 4 (13) 4 (13) 5 (10) 6 (24) 27 30l2*c]@l@ ^⋆ Up to 20th May 2022 ^± TMLE expert is a current or past member of M.J. van der Laan's Lab. ^† Proportions calculated over the number of publications within that year. ^⊡ Proportions calculated over the total number of citations within that year. §.§ Methodological developments and their implementation Over the years since the TMLE framework was first laid out,<cit.> many contributions have been made to expand the settings in which TMLE is used, provide tools for implementation in standard software, and describe the TMLE framework and application in lay language. Thanks to this, the community of public health researchers and epidemiologists have started implementing the TMLE framework and its latest developments to obtain double robust, least biased and efficient estimates and statistical inference from studies. The properties of TMLE, in contrast to other estimators commonly used for causal inference, include that it is loss-based, well-defined, unbiased, efficient and can be used as a substitution estimator. Figure <ref> shows schematically when and why extensions of TMLE have happened in the last 15 years, as well as extensions and uptake. The 81 applied epidemiological studies are classified by methodological development used during the study. Appendix Table <ref> provides further details regarding the methodological development references between 2006 and mid-2022 that are highlighted in Figure <ref>. TMLE’s superior efficiency and power are evidenced in small sample size settings where marginal effects from logistic regression models adjusted for (possibly many) covariates would not be recommended.<cit.> The implementation of TMLE in complex causal effect estimation problems is discussed in many publications, such as in settings with multiple time point interventions,<cit.> longitudinal data,<cit.> post-intervention effect modifiers,<cit.> dependence of the treatment assignment between units<cit.> or censoring,<cit.> causally connected units,<cit.> hierarchical data structures,<cit.> randomisation at the cluster level,<cit.> large electronic health record data, <cit.> and in meta-analyses.<cit.> The TMLE framework is extended and discussed in the setting of case-control studies. One study matched cases to controls,<cit.> another used two-stage sampling and nested case-control design.<cit.> Other studies required the design to be adaptive to possibly invalid assumptions of independent units<cit.> or if the sample population differs from the (possibly ill-defined) target population.<cit.> The collaborative TMLE (C-TMLE), introduced in 2010,<cit.> is an extension of TMLE, in which information on the causal parameter of interest is used when estimating and selecting initial model(s). C-TMLE aims to improve the robustness and efficiency of the TMLE. Schnitzer et al.<cit.> highlight the pitfalls and their consequences of automated variable selection in causal inference, such as in the propensity score model, and how C-TMLE corrects for this. C-TMLE was later extended to measure variable importance<cit.> and to longitudinal data settings.<cit.> Proposals to enhance the C-TMLE algorithm include ordering covariates to decrease C-TMLE time complexity,<cit.> using LASSO with C-TMLE for the estimation of the propensity scores,<cit.> and adaptive truncation of the propensity scores with C-TMLE to ensure positivity.<cit.> The pooled TMLE<cit.> was developed for the context of longitudinal data structures with baseline covariates, time-dependent intervention nodes, intermediate time-dependent covariates, and a possibly time-dependent outcome. Extensions include advice for the optimal discretisation of time<cit.> and to the hazard function.<cit.> The one-step TMLE aims to preserve the performance of the original two-step TMLE, and achieves bias reduction in one step (i.e., without additional iterations of the TMLE update step and possible over-fitting in finite samples).<cit.> This one-step TMLE was later extended to counterfactual average survival curves<cit.> and heterogeneous treatment effects.<cit.> The robust TMLE was proposed in 2017 for transporting intervention effects from one population to another.<cit.> The cross-validated TMLE (CV-TMLE) provides asymptotic inference under minimal conditions (i.e., non-parametric smoothness<cit.>) keeping the bounds of the parameter estimates. It is also used in the estimation of data-adaptive target parameters, like optimal treatment regimes. Recently, TMLE was shown to be useful in defining thresholds and marking specified levels of risks.<cit.> The set of observational articles that use TMLE in their main or sensitivity analyses shows that TMLE has successfully been used to examine associations,<cit.> causation,<cit.> and variable importance. <cit.> It has been used to analyse data with varying numbers of observations, from less than 100 to over hundreds of thousands, from clinical trials, cohort studies,<cit.> and observational studies. § DISCUSSION We aimed to investigate the use of the TMLE framework in epidemiology and public health research and to describe the uptake of its methodological developments since its inception in 2006. We focused on TMLEs for point treatment, time-to-event/survival, and longitudinal exposure-outcome relationships. We did not discuss TMLE methodological developments and software for mediation analyses.<cit.> We found that the TMLE framework and its different estimators were implemented in at least 81 epidemiological observational studies. The majority of these studies have come from the US, many of which are from the University of California, Berkeley. Recently, the use of TMLE has spread across the world. Until 2016, TMLE in observational studies was used by select groups of expertise and disciplines, such as biostatisticians or epidemiologists in academia exploring noncommunicable and infectious diseases, or behavioural epidemiology. From 2016 onward, there was a faster uptake amongst a wider range of expertise and disciplines. There is potential for even wider dissemination and acceptance, both geographically and in some specific disease areas or epidemiological disciplines. We hope this review of explicit and applied examples will contribute to enhancing the relevance of the TMLE framework and increasing its uptake and acceptance in settings where challenges with regard to data, unrealistic assumptions, or subject-matter knowledge lend themselves to the framework. Initially, causal inference methods and estimators relied on parametric modelling assumptions but, to quote Box (1976), “all models are wrong but some are useful".<cit.> It highlights that model misspecification was and remains a challenge, even with ever-growing datasets and computing power: exclusion of the unknown data-generating distribution leads to biased results. Semi-parametric and non-parametric causal inference estimators, such as AIPTW, double-debiased,<cit.> and TMLE<cit.> aim to provide the least biased estimate of the effect of an exposure on an outcome.<cit.> Maximum Likelihood Estimation (MLE) based methods (stratification, propensity score and parametric regression adjustment) and other estimating equations (AIPTW) do not have all of the properties of TMLE, and evidence shows that they under-perform in comparison to TMLE in specific settings.<cit.> For more detailed comparisons between the different methods, the interested reader is referred to Chapter 6 of van der Laan and Rose’s TMLE textbook.<cit.> It is important to highlight that in contrast to the AIPTW estimator, TMLE respects the global constraints of the statistical model (i.e., P(0 < Y < 1) = 1) and solves the score estimation equations being equal to zero using substitution and fluctuation approaches.<cit.> TMLE augments the initial estimates to obtain an optimal bias-variance trade-off for the target estimand of interest and produces a well-defined, unbiased, efficient substitution estimator. Furthermore, the targeting step (i.e., update of the initial estimate) may remove finite sample bias. Lastly, the TMLE framework can be tailored to specific research questions that are difficult to answer using other causal inference methods, such as rare diseases,<cit.> ordinal<cit.> or continuous exposures,<cit.> dynamic treatment regimes,<cit.> and missing outcome data.<cit.> We argue that dissemination of any new statistical methodology relies on five key factors: (i) software availability, (ii) accessibility of available material (e.g., quality of software help files, language used in publications, etc.), (iii) number of experts in the area, (iv) teaching, and (v) collaborations. In the following, we discuss the dissemination of TMLE with regard to each of them. (i) Software availability: Various TMLEs have been developed for complex study designs, such as those with time-to-event outcomes, case-control studies, hierarchical data structures (including cluster randomised trials), longitudinal data, and time-dependent confounding. These methodological developments were accompanied by the release of R software packages, increasing the usability of TMLE. Such software developments include the SuperLearner<cit.> R package in 2007 and the tmle R package in 2012.<cit.> TMLE software for survival analysis (survtmle),<cit.> longitudinal data (ltmle),<cit.> doubly-robust confidence intervals (drtmle),<cit.> and estimation of the survival curve under static, dynamic and stochastic interventions (stremr)<cit.> were implemented in 2017. To match the expanding framework, further software developments occurred in the following years, such as the tlverse suite of software packages for Targeted Learning (<https://tlverse.org/tlverse-handbook/>), which includes R packages for cross-validation (origami),<cit.> highly adaptive lasso (HAL, hal9001),<cit.> super learning (sl3),<cit.> and TMLEs for a range of target estimands, such as effects under static interventions on an exposure (tmle3),<cit.> optimal dynamic treatment regimes for binary and categorical exposures (tmle3mopttx),<cit.> and stochastic treatment regimes that shift the treatment mechanism of a continuous exposure (tmle3shift).<cit.> Additional recently developed packages in R include ctmle for collaborative TMLE,<cit.> haldensify for conditional density estimation with HAL, <cit.> txshift for estimating causal effects of stochastic interventions,<cit.> and lmtp for longitudinal modified treatment policies.<cit.> Although the TMLE framework is well developed in the R software, applied epidemiological research is performed in several other software languages, such as Stata, Python, and SAS. TMLE implementations for binary point exposure and outcome studies are available in all of these languages. A SAS macro for the general implementation of TMLE was programmed in 2016.<cit.> TMLE has been developed for the Python software language in the library zEpid.<cit.> The number of applied researchers in epidemiological studies using Python is relatively low but is increasing; thus, this tool is not currently widely used among applied health sciences researchers. Further development could improve on software packages in the widely used statistical software in health sciences and econometrics, such as Stata.<cit.> Nonetheless, the development version of the user-written Stata command eltmle is currently available to Stata users.<cit.> Not all features of TMLE are available in this Stata command, such as longitudinal analysis and cross-validated TMLE. Additionally, eltmle provides ensemble learning capabilities by accessing the SuperLearner R package. Lastly, any new software development needs to have a friendly user interface, together with standard programming features to be easily disseminated and quickly adopted. (ii) Accessibility of available material: The TMLE framework is a series of potentially statistically-complex modelling approaches and computational algorithms, grounded in statistical theory that requires a solid understanding of highly advanced statistics (i.e., theory for semi-parametric estimation, asymptotics, efficiency, empirical processes, functional analyses, and statistical inference). Tutorials in a more lay language targeting applied researchers and epidemiologists have become more common over the past five years and the uptake of TMLE is expected to increase in the future because of them.<cit.> Their beneficial impact is evident from this review, as these articles are highly referenced in applied work, from the year of their publication, alongside more methodologically heavy contributions to start with, and as sole references in later years. This shows evidence of the importance of speaking the language of the target audience and disseminating advanced mathematical statistics and algorithms from an applied perspective. Additionally, the gradual dissemination of the TMLE framework was evident from our systematic review of the methods sections of the 81 selected manuscripts. We observed that papers published in the early years lay out their TMLE strategy and carefully describe each step in the methods section; whereas, more recently, publications of applied research have placed details of the methods in appendices (or supplementary material) and only cite tutorials and software packages. This shows that the community (e.g., authors, editors, reviewers, readers, etc.) is now aware of the TMLE framework, its utility, and its advantages. A wide range of journals have published the applied research articles studied here, from non-specific public health journals to statistical or disease-specific journals. (iii) Experts: Dissemination outside the US needs further work, as evidenced in our systematic review. We have shown that the TMLE framework appears to be well consolidated in the US, and adoption from Europe and other regions are lower in comparison. This may be related to the delayed introduction of causal inference education outside the US. Fostering targeted local seminars and dedicated short courses for the interested applied audience could be a useful strategy to disseminate the framework. Disease- or discipline-specific experts would be useful for the wider distribution of the methods in specific areas that would benefit from improved methodology. TMLE remains dominant in non-communicable or infectious disease epidemiology compared to other disciplines, but it has high applicability in many disciplines and its use has increased in several of them. The slower uptake of the TMLE framework in other disciplines might be due to a lack of empirical examples of how one performed and interpreted statistical analyses using TMLE in a specific disease area. We aimed to provide such a showcase of the application of the methods in specific settings, based on the available literature, and we demonstrated how the framework was successfully used to advance research by providing robust results. We believe interested readers will find it useful to refer to the studies that faced similar challenges, or were based in settings comparable to theirs. (iv) Teaching: There have been tremendous efforts of dissemination of causal inference methods across disciplines, with a particular emphasis on epidemiology and econometrics sciences in the US during the last 20 years. Most graduate programs in epidemiology have included the teaching of causal inference as a leading topic in the field. In Europe, the trends have not been as fast-paced and for a long time, introductions to causal inference methods have mainly been provided through week-long intensive short courses and at international conferences. These different approaches have major impacts on how quickly the methods are adopted by the community of researchers, journal editors, public health groups, and regulatory agencies. In recent years, there has been a development and acceptance of real-world evidence in various public-health fields, such as the Food and Drug Administration's 21st Century Cures Act of 2016 in the US, which specifically promotes the use of causal inference methodology and designs, such as the emulated trial and TMLE frameworks.<cit.> (v) Collaborations: The Center for Targeted Machine Learning and Causal Inference (CTML) is an interdisciplinary research centre at the University of California at Berkeley that is focused on applications of causal inference and targeted learning. The CTML mission is to advance, implement and disseminate methodology to address problems arising in public health and clinical medicine (<https://ctml.berkeley.edu/home>). CTML provides a great resource for courses, ongoing research, partners, collaborators, and Berkeley faculty members involved in TMLE. CTML sponsors include the Danish multinational pharmaceutical company, Novo Nordisk A/S, the Patient-Centered Outcomes Research Institute (pcori), Kaiser Permanente, the US National Institutes of Health, and the Bill and Melinda Gates Foundation. Academic partners include the University of Washington, University of Copenhagen, UCLA David Geffen School of Medicine, University of California at San Francisco, and Monash University. Conclusions Evidence shows that cross-validated, double-robust, efficient and unbiased estimators are at the forefront of causal inference and statistics, as they aim to avoid model misspecification, bias and invalid inference. The TMLE framework for causal and statistical inference was first developed in 2006 and its expansion in applied studies arose in 2018 via applied epidemiological work, tutorials and user-friendly software. The theoretical properties and practical benefits of the TMLE framework have been highlighted across different fields of applied research (such as various epidemiological, public health and clinical disciplines). More can be done to reach a wider audience across varied disease areas and scientific fields (e.g., genomics, econometrics, political and sociological sciences), including the development of software packages outside the R software, tutorial articles as well as seminars and courses targeted to audiences in specific disciplines, lay-language demonstration, such as by example, of the benefits of TMLE in improving epidemiological output, to name only a few ideas. Many recent TMLE developments answer a variety of methodological problems that expand across scientific disciplines and further efforts can be made to disseminate the framework. This would facilitate the conscientious application of TMLE for causal inference and statistical data analyses, so more researchers could use it in their applied work to minimise the risk of reporting misleading results that are biased due to misspecification. § FUNDING This work was supported by the Medical Research Council [grant number MR/W021021/1]. A CC BY or equivalent licence is applied to the Author Accepted Manuscript (AAM) arising from this submission, in accordance with the grant’s open access conditions. Camille Maringe is supported by a Cancer Research UK Population Research Committee Programme Award (C7923/A29018). § AUTHORS CONTRIBUTIONS The article arose from the motivation to disseminate the principles of modern epidemiology among clinicians and applied researchers. All authors developed the concept and wrote the first draft of the article. MJS and CM reviewed the literature. MJS, RVP, MALF and CM drafted and revised the manuscript. RVP, SG and MJL provided comments on the draft manuscript. RVP contributed to drafting some sections. All authors read and approved the final version of the manuscript. CM is the guarantor of the article. § ACKNOWLEDGMENTS The motivation and some parts of the manuscript come from MALF's work in a visiting academic position in the Division of Biostatistics at the Berkeley School of Public Health in 2019. We acknowledge Dr Susan Gruber for her careful read over one draft manuscript and Professor Mark J. van der Laan for his additional guidance. For the purpose of open access, the authors have applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising. [pages=1, scale=0.95, angle=90, pagecommand=tableArticles by discipline]Table_of_disciplines.pdf [pages=2-,scale=0.95, angle=90, pagecommand=]Table_of_disciplines.pdf

http://arxiv.org/abs/2303.06812v1

20230313020155

Weighted Euclidean balancing for a matrix exposure in estimating causal effect

stat.ME

1]Juan Chen 1]Yingchun Zhou Corresponding author: yczhou@stat.ecnu.edu.cn [1]Key Laboratory of Advanced Theory and Application in Statistics and Data Science-MOE, School of Statistics, East China Normal University. Weighted Euclidean balancing for a matrix exposure in estimating causal effect [ ============================================================================== In many scientific fields such as biology, psychology and sociology, there is an increasing interest in estimating the causal effect of a matrix exposure on an outcome. Covariate balancing is crucial in causal inference and both exact balancing and approximate balancing methods have been proposed in the past decades. However, due to the large number of constraints, it is difficult to achieve exact balance or to select the threshold parameters for approximate balancing methods when the treatment is a matrix. To meet these challenges, we propose the weighted Euclidean balancing method, which approximately balance covariates from an overall perspective. This method is also applicable to high-dimensional covariates scenario. Both parametric and nonparametric methods are proposed to estimate the causal effect of matrix treatment and theoretical properties of the two estimations are provided. Furthermore, the simulation results show that the proposed method outperforms other methods in various cases. Finally, the method is applied to investigating the causal relationship between children's participation in various training courses and their IQ. The results show that the duration of attending hands-on practice courses for children at 6-9 years old has a siginificantly positive impact on children's IQ. Keywords: causal inference, matrix treatment, weighting methods, overall imbalance, observational study. § INTRODUCTION For decades, causal inference has been widely used in many fields, such as biology, psychology and economics, etc. Most of the current research is based on univariate treatment (binary, multivalued or continuous treatment) (<cit.>; <cit.>; <cit.>; <cit.>; <cit.>; <cit.>; <cit.>; <cit.>; <cit.>). However, one may be interested in the causal effect of a matrix treatment. For example, in studying the impact of children's participation in training courses on children's intelligence (measured by IQ), the exposure is a matrix, whose rows represent different age groups, columns represent the types of trainging courses and each element represents the number of hours of class per week. The existing methods are not suitable for matrix exposure and there have been few research on this. Therefore, the goal of this paper is to develop a new method to estimate the causal effect function for matrix exposure. To estimate causal effects in observational studies, it is common to use propensity scores (<cit.>; <cit.>; <cit.>). There are several classes of propensity score-based methods, such as matching, weighting and subclassification, that have become part of applied researchers' standard toolkit across many scientific displines (<cit.>; <cit.>). In this article we focus on the weighting method. In the past decade, various weighting methods have been proposed to balance covariates in the estimation procedure (<cit.>; <cit.>; <cit.>). The key idea of these methods is to estimate propensity score ( <cit.>; <cit.>; <cit.>; <cit.>). When using the parametric method to model the propensity score, the estimation bias of the causal effect will be large if the model is mis-specified. Therefore, some nonparametric methods for estimating the propensity score have been proposed, such as the kernel density estimation (<cit.>). In addition, in recent years, some studies have used optimized weighting methods to directly optimize the balance of covariates (<cit.>; <cit.>; <cit.>). These methods avoid the direct construction of the propensity scores, therefore the obtained estimate achieves higher robustness. One of the methods, the entropy balancing method, has been established as being doubly robust, in that a consistent estimate can still be obtained when one of the two models, either the treatment assignment model or the outcome model, is correctly specified (<cit.>). Furthermore, this method can be easily implemented by solving a convex optimization problem. Here we extend the entropy balancing method to the matrix treatment scenario to balance the covariates. The methods mentioned above assume that all balancing conditions hold exactly, that is, they are exact balancing methods. However, the balancing conditions cannot hold exactly when the dimension of covariate or treatment is high as there will be too many equations to hold simultaneously. For matrix treatment, it is even more difficult for the balancing conditions to hold exactly. To meet this challenge, literatures have shown that approximate balance can trade bias for variance whereas exact balance cannot and the former works well in practice in both low- and high-dimensional settings (<cit.>; <cit.>). The potential limitations of the existing approximate balancing methods are that they directly control univariate imbalance, which cannot guarantee the overall balance especially in the high-dimensional scenario. Besides, there is no principled way to select the threshold parameters simultaneously in practice. Another potential issue of the univariate approximate balancing methods is that it is difficult to handle high-dimensional constraints since the theoretical results require that the number of the balancing constraints should be much smaller than the sample size(<cit.>). To alleviate the limitations of univariate balancing methods, we propose an overall balancing approach, which is called Weighted Euclidean balancing method. The weight is obtained by optimizing the entropy function subject to a single inequality constraint, hence the issue of tuning multiple threshold parameters in univariate balancing methods is solved. The Weighted Euclidean distance is defined to measure the overall imbalance and a sufficient small value of the distance suggests that the covariates are approximately balanced from the overall perspective. Moreover, we propose an algorithm to deal with high-dimensional constraints, so that the proposed method is not restrictive to the low-dimensional setting. The main contributions of the paper are summarized as follows. First, an overall balancing method (Weighted Euclidean balancing method) is proposed, which extends the binary univariate entropy approximate balancing method to the matrix treatment scenario. Unlike univariate approximate balancing method, the Weighted Euclidean balancing method controls the imbalance from the overall perspective. Moreover, to the best of our knowledge, it is the first time that matrix treatment is studied by weighting method in causal inference literature. Second, both parametric and nonparametric causal effect estimation methods for matrix treatment are proposed. Under the parametric framework, a weighted optimization estimation is defined and its theoretical properties are provided. Under the nonparametric framework, B-splines are used to approximate the causal effect function and the convergence rate of the estimation is provided. Third, the proposed method is applied to explore the impact of children's participation in training courses on their IQ and meaningful results are obtained. The remainder of this article is organized as follows: In Section 2, the preliminaries are introduced. In Section 3, the Weighted Euclidean balancing method (WEBM) is proposed. In Section 4, the theoretical properties of the WEBM method are shown. In section 5, a numerical simulation is performed to evaluate the performance of the WEBM method under finite samples. In Section 6, the WEBM method is applied to analyze a real problem. The conclusions and discussions are summarized in Section 7. § PRELIMINARY §.§ Notation and assumptions Suppose an independent and identically distributed sample (𝐙_1,…,𝐙_n) is observed, where the support of 𝐙 = (𝐓,𝐗,Y) is 𝒵=(𝒯×𝒳×𝒴). Here 𝐓∈ R^p× q denotes a matrix exposure, 𝐗∈ R^J denotes a vector of covariates, and Y ∈ R denotes the observed outcome. Since the causal effect is characterized by potential outcome notion, let Y(t) for all t∈𝒯 denotes the potential outcome that would be observed under treatment level 𝐭, i.e. Y = Y(t) if T= t. In this paper, our goal is to estimate the causal effect function 𝔼(Y(𝐭)), which is defined in terms of potential outcomes that are not directly observed. Therefore, three assumptions that commonly employed for indentification are made (<cit.>; <cit.>). Assumption 1 (Ignorability): T_i ⊥ Y_i(t) |X_i, which implies that the set of observed pre-treatment covariates 𝐗_i, is sufficiently rich such that it includes all confounders , i.e. there is no unmeasured confounding. Assumption 2 (Positivity): f_T|X(T_i = t|X_i ) > 0 for all t∈𝒯, which means that treatment is not assigned deterministically. Assumption 3 (SUTVA): Assume that there is no interference among the units, which means that each individual's outcome depends only on their own level of treatment intensity. Under the above assumptions, we first define the stabilized weight as w_i = f(T_i)/f(T_i |X_i), then one can estimate the causal effect function based on the stabilized weight with observational data. §.§ Exact entropy balancing and approximate entropy balancing for matrix exposure The entropy balancing method (<cit.>) is used to determine the optimal weight for inferring causal effects. It has been used for univariate treatment and here this method is extended to matrix exposure and to balance covariates approximately. Note that the stabilized weight w_i = f(T_i)/f(T_i |X_i) satisfies the following conditions for any suitable functions u(𝐓) and v(𝐗): 𝔼(w_iu(𝐓_i)v(𝐗_i)) = ∬f(T_i)/f(T_i |X_i)u(𝐓_i)v(𝐗_i)f(𝐓_i, X_i)dT_idX_i =∫{f(T_i)/f(T_i |X_i)u(T_i)f(T_i |X_i)dT_i } v(X_i)f(X_i)dX_i =𝔼(u(𝐓_i))𝔼(v(𝐗_i)) Besides, it also satisfies that 𝔼(w_i) = ∬f(T_i)/f(T_i |X_i)f(T_i, X_i)dT_idX_i = 1. However, Equation (2) implies an infinite number of moment conditions, which is impossible to solve with a finite sample of observations. Hence, the finite dimensional sieve space is considered to approximate the infinite dimensional function space. Specifically, let u_K1(𝐓) = (u_K1,1(𝐓), u_K1,2(𝐓),…, u_K1,K1(𝐓))^', v_K2(𝐗) = (v_K2,1(𝐗), v_K2,2(𝐗), …, v_K2,K2(𝐗))^' denote the known basis functions, then 𝔼(w_i u_K1(𝐓_i) v_K2(𝐗_i)^') = 𝔼(u_K1(𝐓_i))𝔼(v_K2(𝐗_i)^'). In practice, the covariate balancing conditions given in Equation (4) cannot hold exactly with high dimensional covariates or treatments. It is even more difficult to hold exactly for matrix exposure. To overcome this difficulty, approximate balance is considered rather than exact balance, which has been demonstrated to work well in practice in both low- and high-dimensional settings (<cit.>; <cit.>; <cit.>). Specifically, the balancing weights that approximately satisfy the conditions in Equation (4) are the global minimum of the following optimization problem: min_𝐰∑_i=1^nw_ilog(w_i) s.t. |1/n∑_i=1^nw_i u_K1,l(𝐓_i)v_K2,l̃(𝐗_i) - (1/n∑_i=1^n u_K1,l(𝐓_i)) (1/n∑_i=1^nv_K2,l̃(𝐗_i)) |≤δ_l,l̃, where u_K1,l(𝐓_i) and v_K2,l̃(𝐗_i) denote the lth and l̃th components of u_K1(𝐓_i) and v_K2(𝐗_i), respectively. Let m_K(𝐓_i, 𝐗_i) = vec(1/n u_K1(𝐓_i)v_K2(𝐗_i)^') and m̅_K = vec (1/nu̅_K1v̅_K2^') denote two column vectors with dimension K, where K= K1 K2, the lth and l̃th components of u̅_K1 and v̅_K2 are defined as u̅_K1,l = 1/n∑_i=1^n u_K1,l(𝐓_i) and v̅_K2,l̃ = 1/n∑_i=1^n v_K2,l̃(𝐗_i), then condition (5) is equivalent to min_w ∑_i=1^nw_ilog(w_i) s.t. |∑_i=1^nw_i m_K,k(𝐓_i, 𝐗_i) - nm̅_K,k|≤δ_k, k= 1,…,K. However, there is a large number of tuning parameters (δ_1,…,δ_K) which is very time-consuming to determine and there is lack of guideline on tuning these parameters simultaneously in practice. § METHODOLOGY Due to the potential issues of univariate approximate balancing methods, the weighted Euclidean balancing method is proposed in this section, whose key idea is to control the overall imbalance in the optimization problem (7). §.§ Weighted Euclidean balancing method Define the following weighted Euclidean imbalance measure (WEIM) as a weighted version of the squared Euclidean distance: WEIM = ∑_k=1^K{λ_k^2 [∑_i=1^n w_i (m_K,k(𝐓_i,𝐗_i)-m̅_K,k)]^2 }. The weighted Euclidean balancing obtains the balancing weights that approximately satisfy the condition (4) by solving the following convex optimization problem: min_w ∑_i=1^nw_ilog(w_i) s.t. ∑_k=1^K{λ_k^2 [∑_i=1^n w_i (m_K,k(𝐓_i,𝐗_i)-m̅_K,k)]^2 }≤δ, where (λ_1,…, λ_K) is a pre-sepecified weight vector and δ≥ 0 is a threshold parameter. Assume that condition (3) holds exactly, whose sample condition is 1/n∑_i=1^n w_i = 1. Note that univariate exact balance is equivalent to the overall exact balance, in the sense that ∑_i=1^n w_i (m_K,k(𝐓_i,𝐗_i)-m̅_K,k)=0, k= 1,…,K is equivalent to WEIM=0. However, the univariate approximate balance does not imply the overall approximate balance since it is possible that ∑_i=1^n w_i (m_K,k(𝐓_i,𝐗_i)-m̅_K,k), k= 1,…,K is small while the WEIM is large. The pre-specified vector (λ_1,…, λ_K) reflects the importance of each univariate constraint. In this paper, we set λ_k= σ_k^-1, where σ_k^2 is the variance of m_K,k(𝐓,𝐗). Since problem (9) is difficult to solve numerically, its dual problem is considered here, which can be solved by numerically efficient algorithms. Theorem 1 provides the dual formulation of problem (9) as an unconstrained problem. Theorem 1. Assume that max_i (max_k |λ_km_K,k(𝐓_i,𝐗_i) |) < ∞, the dual of problem (9) is equivalent to the following unconstrained problem min_θ∈ R^K∑_i=1^nexp( ∑_k=1^Kθ_j λ_j (m_K,k(𝐓_i,𝐗_i)-m̅_K,k)) + √(δ)||θ||_2, and the primal solution ŵ_i is given by ŵ_i = exp{∑_k=1^Kθ̂_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k) -1} , i=1,…,n, where θ̂ is the solution to the dual optimization problem (10). The proof of Theorem 1 is in Appendix A.1. §.§.§ Selection of tuning parameter Another practical issue that arises with weighted Euclidean weights is how to choose the degree of approximate balance δ. A tuning algorithm is proposed as follows. First, determine a range of positive values 𝒟 for δ, then the optimal value of δ is selected by the following algorithm. Algorithm 1. Selection of δ. For each δ∈𝒟, 1. Compute the dual parameters θ̂ by solving the dual problem (10); 2. Compute the estimated weights ŵ_i using equation (11); 3. Calculate WEIM in (8) using ŵ_i; Output δ^* that minimizes WEIM. §.§.§ Weighted Euclidean balancing with high-dimensional covariates In the high-dimensional or ultra high-dimensional covariate setting with K relatively large compared to n or K>>n, it becomes difficult to control the overall imbalance using the Weighted Euclidean balancing method. To meet this challenge, we propose an algorithm to select a small subset of the covariates in the sparse setting. Specifically, consider v_K2(𝐗) = (1, 𝐗) in the high-dimensional setting. Let Bcor_j be the ball correlation (<cit.>) between X_j and 𝐓, j=1,…,L. Rank X_1, …,X_L as X_(1),…,X_(L) such that X_(1) has the largest Bcor value, X_(2) has the second largest Bcor value, and so forth. The covariates are added successively according to the rank of ball correlation until there is a break point of WEIM's, and the set added before the break point appears is the target set. The key idea hings on WEIM, which represents the contribution of the j most imbalanced covariates to the overall imbalance after WEBM weighting. If WEIM remains stable as j inceases, it indicates that the overall imbalance can be controlled. However, if there is a break point at Step j, it implies that adding the jth covariates greatly inceases WEIM, which is harmful to the control of the overall imbalance. Therefore, the algorithm should be stopped and print the outputs at Step j-1. Specifically, procedures to select the subset of covariates are given by the following algorithm. Algorithm 2. Subset selection of covariates in the high dimensional case. For each j ∈{ 1,…, L}, compute the estimated weights ŵ_i^(j) using v_K2(𝐗) = (1,X_(1),…, X_(j)); calculate WEIM^(j) in (8) using ŵ_i^(j); add (j,WEIM^(j)) to the x-y plot and observe whether there is a break point at (j,WEIM^(j)): If no, let j=j+1; If yes, stop and output L_0 = j-1. The selected subset of the covariates is (X_(1),…, X_(L_0)). §.§ Causal effect estimation In this subsection, both parametric and nonparametric approaches are developed to estimate the causal effect function. A weighted optimization estimation is defined under the parametric framework and broadcasted nonparametric tensor regression method (<cit.>) is used to estimate the causal effect function under the nonparametric framework. §.§.§ Parametric approach The causal effect function is parametrized as s(𝐭;β), assume that it has a unique solution β^* defined as β^* = agrmin_β∫_𝒯𝔼[Y(t)- s(𝐭;β) ]^2f_T(t)dt. The difficulty of solving Equation (12) is that the potential outcome Y(t) is not observed for all t. Hence, Proposition 1 is proposed to connect the potential outcome with the observed outcome. Proposition 1 Under Assumption 1, it can be shown that 𝔼[w(Y- s(𝐭;β) )^2] = ∫_𝒯𝔼[Y(t)- s(𝐭;β) ]^2f_T(t)dt. The proof of Proposition 1 can be found in Appendix A.2. Note that Y(t) on the right hand side of Equation (13) represents the potential outcome and Y on the left hand side represents the observed outcome. Proposition 1 indicates that by having w on the left hand side of Equation (13), one can represent the objective function with the potential outcome (right side) by that with the observed outcome (left side). Therefore, the true value β^* is also a solution of the weighted optimization problem: β^* = argmin_β𝔼[w(Y- s(𝐭;β))^2]. This result implies that the true value β^* can be identified from the observational data. One can obtain the estimator based on the sample, which is β̂ = argmin_β∑_i=1^nŵ_̂î(Y_i- s(𝐓_i;β) )^2. §.§.§ Nonparametric approach Suppose 𝔼(Y(𝐭)) = s(𝐭). In a similar manner to the proof of Proposition 1, it can be shown that 𝔼(wY |𝐓=t) = 𝔼(Y(𝐭)). Existing work of nonparametric tensor regression suffers from a slow rate of convergence due to the curse of dimensionality. Even if one flattens the tensor covariate into a vector and applies common nonparametric regression models such as additive models or single-index models to it, this issue still exists. Besides, when dealing with a vectorized tensor covariate, one would ignore the latent tensor structure and this might result in large bias. To meet these challenges, we adopt the broadcasted nonparametric tensor regression method (<cit.>) to estimate the causal effect function s(𝐭). The main idea of the broadcasted nonparametric tensor regression is to use the (low-rank) tensor structure to discover important regions of the tensor so as to broadcast a nonparametric modeling on such regions. Specifically, assume that s(𝐓) = c+1/pq∑_r=1^R<β_1^(r)∘β_2^(r), F_r(𝐓)>, where c∈ R, 𝐓∈ R^p × q, F_r(𝐓) = ℬ(f_r, 𝐗), ℬ is a broadcasting operator, which is defined as (ℬ(f,𝐓))_i_1,i_2 = f(T_i_1,i_2), for all i_1,i_2. The broadcasted functions f_r, r=1,…,R, will be approximated by B-spline functions, i.e., f_r(x) ≈∑_d=1^Dα_r,db_d(x), where 𝐛(x) = (b_1(x),…,b_D(x))^' is a vector of B-spline basis functions and α_r,d's are the corresponding spline coefficients. Define α_r = (α_r,1,…,α_r,D)^' and (Φ(𝐓))_i_1,i_2,d = b_d(T_i_1,i_2), the regression function (16) can be approximated by s(𝐓) ≈ c+1/pq∑_i=1^R<β_1^(r)∘β_2^(r)∘α_r, Φ(𝐓)>. To separate out the constant effect from f_r's, the condition ∫_0^1 f_r(x) dx=0 is imposed, which leads to ∫_0^1∑_d=1^Dα_r,db_d(x)dx=0, r=1,…, R. Then the following optimization problem is considered: argmin_c,𝐆 ∑_i=1^n (ŵ_iY_i- c-1/pq<𝐆,Φ(𝐓_i)>)^2 s.t. 𝐆 = ∑_r=1^Rβ_1^(r)∘β_2^(r)∘α_r ∑_d=1^Dα_r,d∫_0^1 b_d(x)dx =0, r=1,…,R, and the estimated regression function is ŝ(𝐓) = ĉ+1/pq <𝐆̂,Φ(𝐓)>, where (ĉ,𝐆̂) is a solution of (20). Since optimization problem (20) contains too many constraints, it is not computationally efficient to solve it directly. To further simplify the optimization problem, an equivalent truncated power basis (<cit.>) is used to reduce the constraints. Specifically, let b̃_d(x), d=1,…,D denote the truncated basis: b̃_1(x)=1,b̃_2(x)=x,…, b̃_ς(x) = x^ς-1, b̃_ς+1(x) = (x-ξ_2)_+^ς-1,…, b̃_D(x)=(x-ξ_D-ς+1)_+^ς-1, where ς and (ξ_2,…,ξ_D-ς+1) are the order and the interior knots of the aforementioned B-spline, respectively. Based on these basis functions, consider the following optimization argmin_c̃,𝐆̃ ∑_i=1^n (ŵ_iY_i- c̃-1/pq<𝐆̃,Φ̃(𝐓_i)>)^2 s.t. 𝐆̃ = ∑_r=1^Rβ_1^(r)∘β_2^(r)∘α̃_r, where Φ̃(𝐓)_i_1,i_2,d = b̃_d+1(𝐓_i_1,i_2), d=1,…,D and α̃_r ∈ R^D-1 is the vector of coefficients. Compared with (20), the mean zero constraints are removed by reducing one degree of freedom in the basis functions. According to <cit.>, Lemma B.1, one can show that ŝ(𝐓) = ĉ̃̂+1/pq <𝐆̂̃̂,Φ̃(𝐓)>, where (ĉ̃̂,𝐆̂̃̂) is the solution of (22). The optimization problem (22) can be solved by the scaled-adjusted block-wise descent algorithm (<cit.>). § THEORETICAL PROPERTIES In this section, the large sample properties of the proposed estimators in section 3 are established. First the consistency of the estimated weight in section 3.1 is shown, then the consistency of the parametric estimator in section 3.2.1 and the convergence rate of the nonparametric estimator in section 3.2.2 are shown. The following assumptions are made. Assumption 4 * There exists a constant c_0 such that 0 < c_0 < 1, and c_0 ≤exp (z-1) ≤ 1-c_0 for any z= M̃_K(𝐭,𝐱)^'θ with θ∈int(Θ). Besides, exp (z-1) = O(1) in some neighborhood of z^* = M̃_K(𝐭,𝐱)^'θ^*, where M̃_K(𝐭,𝐱)= Λ (m_K(𝐭,𝐱)-m̅_K) and Λ= diag(λ_1,…,λ_K). * There exists a constant C such that E {M̃_K(𝐓_i, 𝐗_i)M̃_K(𝐓_i, 𝐗_i)^'}≤ C. * δ = o(n). * sup_(𝐓,𝐗) exp{∑_j=1^Kθ_j^* λ_j [m_K,j(𝐓,𝐗)-Em_K,j(𝐓,𝐗)] } = O(1). Assumption 5 * The parameter space Θ_1 is a compact set and the true parameter β_0 is in the interior of Θ_1. * (Y-s(T;β))^2 is continuous in β, 𝔼[sup_β(Y-s(T;β))^2] < ∞ and sup_β𝔼[(Y-s(T;β))^4] < ∞. Assumption 6 * s(𝐭;β) is twice continuously differentiable in β∈Θ_1 and let h(𝐭;β) ≡▽_β s(𝐭;β). * 𝔼{ w(Y-s(𝐓;β))h(𝐓;β) } is differentiable with respect to β and U ≡ - ▽_β𝔼{ w(Y-s(𝐓;β))h(𝐓;β) }|_β=β^* is nonsingular. * 𝔼[sup_β| Y-s(T;β) |^2+δ] < ∞ for some δ >0 and there extists some finite positive constants a and b such that 𝔼[sup_β_1: ||β_1-β|| < δ_1| s(T;β_1) - s(T;β) |^2]^1/2 < a·δ_1^b for any β∈Θ_1 and any small δ_1 >0. Assumption 7 * The treatment 𝐓∈ [0,1]^p× q has a continuous probability density function f, which is bounded away from zero and infinity. * The vector of random errors, ϵ = (ϵ_1,…,ϵ_n)^', has independent and identically distributed entries. Each ϵ_i is sub-Gaussian with mean 0 and sub-Gaussian norm σ < ∞. * The true broadcasted functions f_0r∈ℋ, r= 1,…,R_0. Here ℋ is the space of functions from [0,1] to R satisfying the Hölder condition of order ω, i.e., ℋ = { g: | g^(l) (x_1)-g^(l) (x_2) |≤ S_1| x_1-x_2 |^ω, ∀ x_1,x_2 ∈ [0,1] }, for some constant S_1>0, where g^(l) is the l-th derivative of g, such that ω∈ (0,1] and τ = l+ω >1/2. * The order of the B-spline used in (16) satisfies ς≥τ+1/2. Let 0= ξ_1 < ξ_2 <… < ξ_D-ς+2=1 denote the knots of B-spline basis and assume that h_n = max_d=1,…, D-ς+1|ξ_d+1 -ξ_d |≍ D^-1 and h_n/min_d=1,…, D-ς+1|ξ_d+1 -ξ_d |≤ S_2 for some constant S_2>0. Assumption 4(1) enables consistency of θ̂ to translate into consistency of the weights. Assumption 4(2) is a standard technical condition that restricts the magnitude of the basis functions; see also Assumption 4.1.6 of <cit.> and Assumption 2(2) of <cit.>. Assumption 4(3) requires that the threshold parameter δ should be much smaller than the sample size. Assumption 4 (4) is needed for consistency of the estimated weight. Assumption 5(1) is a commonly used assumption in nonparametric regression. Assumption 5(2) is an envelope condition applicable to the uniform law of large numbers. Assumption 6(1) and (2) impose sufficient regularity conditions on the causal effect function and its derivative function. Assumption 6(3) is a stochastic equicontinuity condition, which is needed for establishing weak convergence (<cit.>). Assumption 7(1), (3) and (4) are common in nonparametric regression models. In particular, Assumption 7(3) and (4) regularize the space where the true broadcasted functions lie in and guarantee that they can be approximated welll by B-spline functions. Similar assumptions can be found in <cit.> and <cit.>. Assumption 7(2) is a standard tail condition of the error. Based on these assumptions, the following theorems are established. Theorem 2. Let θ̂ denote the solution to Problem (10) and ŵ_i = exp{∑_k=1^Kθ̂_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k) -1} , i=1,…,n, then under Assumptions 1-4, * ∫|𝐰̂-𝐰^* |^2 dF(𝐭,𝐱) = O_p(n^-1). * 1/n∑_i=1^n|ŵ_i- w_i^* |^2 = O_p(n^-1). Based on Theorem 2, we can establish the consistency of the parametric estimator and the convergence rate of the nonparametric estimator. Theorem 3 * Under Assumptions 1-5, ||β̂-β^* || →_p 0. * Under Assumptions 1-6, √(n)(β̂-β^*) →_d N(0,V), where V = 4U^-1·𝔼{ w^2(Y-s(𝐓;β^*))^2h(𝐓;β^*)h(𝐓;β^*)^'}· U^-1 Theorem 4 If Assumptions 1-4 and 7 hold, R≥ R_0, and n > S_1 h_n^-2-2/log(h_n)(log^-2(h_n))(R^3+R(p+q)+RD) for some large enough constant S_1>0, then ||ŝ(𝐓)-s_0(𝐓) ||^2 = O_p ( R^3+R(p+q)+RD/n )+O_p ( {∑_r=1^R_0||vec(𝐁_0r)||_1/pq}^2 1/D^2τ ), where s_0(𝐓) = c_0+1/pq∑_r=1^R_0<β_1^(0r)∘β_2^(0r), F_0r(𝐓)> represents the true regression function. The proofs of Theorem 2, 3 and 4 can be found in Appendix A.3, A.4 and A.5, respectively. § SIMULATION To evaluate the finite sample performance of the proposed method, simulation studies are carried out under different data settings. The main motivation of the simulation is to compare the proposed method with three other methods when the outcome model are linear and nonlinear in various ways. §.§ The low-dimensional covariate setting In this subsection, we compare the performance of the proposed method (WEBM) with the unweighted method (Unweighted), entropy balancing method (EB) and univariate approximate balancing method (MDABW) in the low-dimensional covariate setting, where EB refers to the method proposed by <cit.> that balances covariates exactly and MDABW refers to the method proposed by <cit.> that balances covariates approximately. Since the covariates are shared across all scenarios, their data generating process is first described. Specifically, we independently draw 5 covariates from a multivariate normal distribution with mean 0, variance 1 and covariance 0.2, that is, 𝐗 = (X_1,....,X_5)^'∼ N_5(μ,Σ) with μ= [ 0; ⋮; 0 ]and Σ=[ 1 0.2 … 0.2; 0.2 1 0.2 … 0.2; 4; 0.2 0.2 … 1 ]_5×5. Consider a linear treatment assignment model, which is defined as 𝐓_i= X_i1𝐁_1+X_i2𝐁_2+X_i3𝐁_3+𝐄_i, where 𝐁_j = [ 1 0; 0 1; 1 1 ]_3×2, j=1,2,3 denotes the jth coefficient matrix, and 𝐄_i ∈ R^3× 2 denotes the error matrix, whose element follows a standard normal distribution. For the outcome model, we consider four scenarios and conduct 100 Monte Carlo simulations for each scenario. The first two scenarios assume an outcome model that is linear in treatment and the others assume a nonlinear relationship. In scenario 1, the linear outcome model is defined as Y_i = 1+ <𝐁, 𝐓_i>+X_i1+(X_i2+1)^2+X_i4^2+ϵ_i, where 𝐁 = [ 1 0; 0 1; 1 1 ]_3×2 and ϵ_i ∼ N(0,2^2). In this scenario, v_k2(𝐗) = (1, 𝐗, 𝐗*𝐗)^', where * represents the Hadamard product of two matrices, and the corresponding element of 𝐗*𝐗 is (X*X)_ij = (x_ijx_ij). In scenario 2, the linear outcome model is defined as Y_i = 1+ <𝐁, 𝐓_i>+X_i2+X_i3+X_i1X_i2+…+X_i4X_i5+ϵ_i, where 𝐁 and ϵ_i are the same as in Equation (21). In this scenario, the interaction terms are strong confounders, hence set v_k2(𝐗) = (1, 𝐗, X_jX_k)^', 1≤ j < k ≤ 5. In scenarios 3 and 4, the nonlinear outcome models are considered and are defined as Y_i = 1+ <𝐁, F_1(𝐓_i)>+X_i1+(X_i2+1)^2+X_i4^2+ϵ_i, and Y_i = 1+ <𝐁, F_1(𝐓_i)>+X_i2+X_i3+X_i1X_i2+…+X_i4X_i5+ϵ_i, respectively. Here 𝐁 and ϵ_i are the same as in Equation (24), (F_1(𝐓))_k_1,k_2= f_1(T_k_1,k_2)= T_k_1,k_2+0.6sin{2π(T_k_1,k_2-0.5)^2 }. For all four scenarios, consider u_K1(𝐓) =(1, vec(𝐓)^')^' for simplicity. For each method, the mean RMSE and its standard deviation of the coefficient estimates for the linear outcome model, and those of the fitted values for the nonlinear outcome model are reported based on 100 data replications. Table 1 shows the mean RMSE and its standard deviation of the coefficient matrix for the linear outcome model. Observe that the mean RMSE of WEBM is the smallest among the four methods, and the standard deviation of WEBM is the second smallest while that of Unweighted is the smallest in both scenario 1 and scenario 2. Besides, the results of scenario 2 indicate that MDABW and EB have poor performance when the basis function of covariates includes interaction entries, and their mean RMSEs are even larger than those of Unweighted. The mean RMSE and standard deviation of all methods decreases as the sample size increases. Table 2 shows the mean RMSE and its standard deviation of the fitted values of the nonlinear outcome model. As can be seen, the results for both scenario 3 and 4 are similar to those in Table 1, that is, WEBM performs the best with both the smallest mean RMSE in all cases. Similarly, MDABW and EB methods have poor performance when the basis function of covariates includes interaction entries and the mean RMSE and standard deviation of all methods decreases as the sample size increases. §.§ The high-dimensional covariate setting In this subsection, since the MDABW and EB methods can only deal with the low-dimensional covariate case, we compare the performance of the proposed method (WEBM) with the Unweighted method and the method (Mapping) proposed by <cit.>, which selects a small important subset of covariates by a joint screening procedure, in the high-dimensional covariate setting. Since the Mapping method can only deal with linear outcome model, these two methods are compared in the linear outcome model setting. For both methods, set the sample size n=500 and the dimension of 𝐓 to be 3 × 2. Consider two scenarios (scenario 5-6) with the dimension of covariates L=49 and L=99, respectively. The motivation for such settings is to consider the number of constraints K < n (scenario 5) and K>n (scenario 6), respectively. For the basis functions, u_K1(𝐓)= (1,vec(𝐓)^')^' and v_K2(𝐗) = (1,𝐗^')^' are considered. Additionally, covariates are drawn from a multivariate normal distribution with mean 0, variance 1 and covariance 0.2. For each scenario, consider a linear treatment assignment model, which is defined as 𝐓_i= X_i1𝐁_1+X_i2𝐁_2+…+X_i5𝐁_5+𝐄_i, where 𝐁_j = [ 1 0; 0 1; 1 1 ]_3×2, j=1,2,3,4,5 denotes the jth coefficient matrix, and 𝐄_i ∈ R^3× 2 denotes the error matrix, whose element follows a standard normal distribution. Moreover, the linear outcome model is defined as Y_i = 1+ <𝐁, 𝐓_i>+X_i1+X_i2+…+X_i5+ϵ_i, where 𝐁=[ 1 0; 0 1; 1 1 ]_3×2, ϵ_i ∼ N(0,2^2). For each method, the mean RMSE and its standard deviation of the tensor regression estimation for the linear outcome model are reported based on 100 data replications. Table 3 shows the mean RMSE and its standard deviation of coefficent matrix for the linear outcome model in the high-dimensional covariate setting. The results indicate that WEBM performs the best with the smallest mean RMSE in all cases. § APPLICATION Intelligence Quotient (IQ) is based on biological attributes, most of which are inherited from parents. It mainly refers to a person's cognitive abilities, such as logical reasoning, pattern recognition, and short-term memory. Of course, due to genetic mutations, parents with average IQ may have offspring with superior intelligence, and vice versa. But IQ also has social attributes. Studies have found that IQ has obvious plasticity, and environmental factors affect IQ levels. From 1947 to 2002, the IQ level of developed countries rose steadily at a rate of 3% every 10 years, which is called the "Flynn effect." This effect has been repeatedly observed in various countries, various age groups, and a large number of different environments, and has become a reliable evidence that "environment affects IQ" (<cit.>). In particular, literatures have shown that taking training lessons, such as music, sports, chess and so on, can significantly enhance children's IQ (<cit.>; <cit.>; <cit.>; <cit.>). Unfortunately, there are some limitations of existing studies. First, existing studies have only analyzed the relevant effects of attending single training course at a fixed age on children's IQ. In practice, however, a child may attend different training courses at the same age and participation may also vary by age. Second, existing studies mainly investigated the impact of whether or not to attend training classes, ignoring the effect of class duration. Third, existing studies only analyzed the correlation between children's participation in training courses on children's IQ, while their causal relationship is much more of a concern. To overcome the aforementioned limitations, the proposed method is applied to a matrix treatment, which contains the structural information of children's participation in training courses, to investigate their causal impact on children's IQ. The data are obtained from the Brain Science Innovation Institute of East China Normal University's Child Brain Intelligence Enhancement Project, whose goal is to explore brain development and help improve brain power. The treatment we are interested in is a 2 × 5 matrix about children's participation in training courses, whose rows represent age groups, including 3-6 years old and 6-9 years old, columns represent the types of training courses, including knowledge education (Chinese, mathematics and English, etc.), art (music, art, calligraphy, etc.), sports (swimming, ball games, etc.), hands-on practice (STEM, Lego, etc.) and thinking training (logical thinking, EQ education, attention, etc.), and each element of the matrix represents the number of hours of class per week. The outcome is children's IQ and the pre-treatment covariates include children's gender as well as parental education, which have been shown to be associated with both the treatment and outcome variables (<cit.>; <cit.>; <cit.>; <cit.>). A complete-case analysis is conducted with a sample of 103 participants. Before estimating the causal effect, we first examine the covariate balancing of WEBM, MDABW and EB methods based on the WEIM statistics. The statistic WEIM defined in equation (8) is 0.1050 for the WEBM method, 0.3761 for the MDABW method and 0.9881 for the EB method, which implies that WEBM balances covariates well while EB does not. Assume a linear tensor outcome model and the bootstrap method with 200 replicates is used to obtain confidence intervals for the parameter estimates. The results are shown in Table 4. Table 4 shows the estimated causal effects of the duration of attending different classes at different ages on children's IQ. It can be seen that most methods (except EB) suggest that the duration of attending hands-on practice courses at 6-9 years old has a siginificantly positive impact on children's IQ. This finding is not only consistent with previous findings that participation in hands-on practice classes can improve children's IQ (<cit.>; <cit.>; <cit.>), but further suggests that longer participation in hands-on practice classes is more beneficial to children's IQ. The results imply that future work of an intervention study about attending hands-on practice training courses, which in turn may improve children's IQ, is suggested. Besides, the width of confidence interval of the estimated causal effect based on WEBM is the smallest, which implies that the estimation accuracy of WEBM is the highest. § CONCLUSION AND DISCUSSION In this study, the weighted Euclidean balancing method is proposed, which obtains stabilized weights by adopting a single measure that represents the overall imbalance. An algorithm for the high-dimensional covariate setting is also proposed. Furthermore, parametric and nonparametric methods are developed to estimate the causal effect and their theoretical properties are provided. The simulation results show that the proposed method balances covariates well and produces a smaller mean RMSE compared to other methods under variaous scenarios. In the real data analysis, the WEBM method is applied to investigate causal effect of children's participation in training courses on their IQ. The results show that the duration of attending hands-on practice at 6-9 years old has a siginificantly positive impact on children's IQ. Since the causal effect function 𝔼(Y(t)) is more general, we mainly consider it as the estimand for matrix treatment in this paper. Actually, one can also consider the average treatment effect (𝔼(Y(t+ t)-Y(t))) or average partial effect (𝔼Y(t+ t)-𝔼Y(t)/ t) , which can be easily estimated based on the estimates of causal effect function (<cit.>). Indeed, the causal effect function provides a complete description of the causal effect, rather than a summary measure. Moreover, parametric and nonparametric methods are developed to estimate the causal effect function. Parametric method is recommended when reasonable assumptions can be made about the true model since it is easier to implement and requires less sample size. Despite nonparametric method has higher requirements of the sample size, one can choose to use it according to the real situations due to its higher flexibility. Besides, this paper mainly focuses on the small-scale matrix treatment. Large-scale matrix treatment with low-rank structure can also be considered. In such case, one may control the overall imbalance by only balancing their non-zero elements based on some decomposition technology, and this will be investigated in future work. * apalike § APPENDIX §.§ A.1.Proof of Theorem 1 The primal problem is min_𝐰∑_i=1^nw_ilog(w_i) s.t. ∑_j=1^K{λ_j^2 [∑_i=1^n w_i (m_K,j(𝐓_i,𝐗_i)-m̅_K,j)]^2 }≤δ. Let ||θ||_2 = √(θ_1^2+…+θ_K^2) be the l_2 norm for an arbitrary K-dimensional vector θ =(θ_1,…,θ_K)^' and Λ = diag(λ_1,…,λ_K), then the inequality constraint in the primal problem can be rewritten as ||∑_i=1^n w_i Λ (m_K(𝐓_i,𝐗_i)-m̅_K) ||_2 ≤√(δ). Let 𝒜⊆ R^K be a convex set such that 𝒜 = { a ∈ R^K: || a||_2 ≤√(δ)}. Define I_𝒜(a) = 0 if a ∈𝒜 and I_𝒜(a) = ∞ otherwise. Then, the primal problem (15) is equivalent to the following optimaization problem: min_𝐰 ∑_i=1^nw_ilog(w_i)+I_𝒜( ∑_i=1^n w_i Λ (m_K(𝐓_i,𝐗_i)-m̅_K)). Let h(w) = ∑_i=1^n w_i log(w_i), the conjugate function of h is h^*(w) = sup_t(∑_i=1^n w_i t_i-∑_i=1^n w_i log(w_i)) = sup_t ∑_i=1^n (w_i t_i-w_i log(w_i)) = ∑_i=1^n sup_t_i(w_i t_i-w_i log(w_i)) = ∑_i=1^n f^*(w_i), where f^*(w_i) = sup_t_i(w_i t_i-w_i log(w_i)) is the conjugate function of f(w_i) = w_i log(w_i). Let g(θ) = I_𝒜(θ) for any θ∈ R^K, then the conjugate function of g is g^*(θ) = sup_a (∑_k=1^Kθ_ka_k -T_𝒜(a) ) = sup_|| a ||_2 ≤√(δ) (∑_k=1^Kθ_ka_k) = sup_|| a ||_2 ≤√(δ) (||θ||_2 || a ||_2) = √(δ)||θ||_2. Define the mapping H: R^n → R^K such that Hw = ∑_i=1^n w_i Λ (m_K(𝐓_i,𝐗_i)-m̅_K), then H is a bounded linear map. Let H^* be the adjoint operator of H, then for all θ = (θ_1,…,θ_K)^'∈ R^K, H^*θ = (∑_k=1^Kθ_k λ_k (m_K,k(𝐓_1,𝐗_1)-m̅_K,k), …, ∑_k=1^Kθ_k λ_k (m_K,k(𝐓_n,𝐗_n)-m̅_K,k) )^'. Define θ̃ = Hw̃ = 1/n^r∑_i=1^nΛ (m_K(𝐓_i,𝐗_i)-m̅_K), where w̃ = (1/n^r,…,1/n^r)^'∈ dom(F). Here, we choose b to be sufficiently large such that ||θ̃||_2 ≤√(δ), then we obtain that g(θ̃) = 0 and g is continuous at θ̃. Therefore, θ̃∈ H(dom(F) ∩ cont(g)), which implies that H(dom(F) ∩ cont(g)) ≠∅. Here, dom(F) and cont(g) denotes the domain of F and the continuous set of g, respectively. Therefore, the strong duality condition of the Fenchel duality theorem is verified. Moreover, F(H^*θ)+g^*(-θ) = ∑_i=1^n f^*(∑_k=1^Kθ_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k))+√(δ)||θ||_2. According to the Fenchel duality theorem (Mohri et al. (2018), Theorem B.39), we have min_w ∑_i=1^nw_ilog(w_i)+I_𝒜( ∑_i=1^n w_i Λ (m_K(𝐓_i,𝐗_i)-m̅_K)) = min_w ∑_i=1^n f^*(∑_k=1^Kθ_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k))+√(δ)||θ||_2. Furthermore, since the strong duality condition holds, we can conclude that H^*θ̂ is a subgradient of F at ŵ. That is, ∑_k=1^Kθ̂_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k) = log(ŵ_i)+1. Therefore, ŵ_i = exp(∑_k=1^Kθ̂_k λ_k (m_K,k(𝐓_i,𝐗_i)-m̅_K,k)-1). The proof of theorem 1 is completed. §.§ A.2.Proof of Proposition 1 Using the law of total expectation and Assumption 1, we can deduce that 𝔼[w(Y- s(𝐓;β) )^2] =E[f(T)/f(T|X)(Y-s(𝐓;β) )^2] = 𝔼(𝔼[f(T)/f(T|X)(Y- s(𝐓;β) )^2] |T=t,X=x ) = 𝔼(f(t)/f(t|x)𝔼([(Y- s(𝐓;β) )^2] |T=t,X=x) ) = ∫_𝒯×𝒳f(t)/f(t|x)𝔼[(Y(T)- s(𝐓;β) )^2 |T = t, X= x]f(t|x)dtdx =∫_𝒯×𝒳𝔼[(Y(T)- s(𝐓;β) )^2 |T = t, X= x]f(t)f(x)dtdx = ∫_𝒯×𝒳𝔼[(Y(t)- s(𝐓;β) )^2 |X= x]f(t)f(x)dtdx (using Assumption 1) = ∫_𝒯𝔼[(Y(t)- s(𝐓;β) )^2] f(t)dt . Hence, we complete the proof of Proposition 1. §.§ A.3.Proof of Theorem 2 The first order optimality condition for the dual probblem (10) is ∑_i=1^nexp{∑_j=1^Kθ̂_jλ_jM_K,j(𝐓_i,𝐗_i) }·λ_jM_K,j(𝐓_i,𝐗_i) +√(δ)θ̂_j/||θ̂||_2 =0, j=1,…,K, where M_K,j(𝐓_i,𝐗_i) = m_K,j(𝐓_i,𝐗_i)- m̅_K,j, M_K(𝐓_i,𝐗_i)= (M_K,1(𝐓_i,𝐗_i), …, M_K,K(𝐓_i,𝐗_i))^'. Let Λ = diag(λ_1,…,λ_K) and 1/n∑_i=1^nΦ(𝐓_i,𝐗_i;θ) = 1/n∑_i=1^nexp{∑_j=1^Kθ_jλ_j[m_K,j(𝐓_i,𝐗_i)-𝔼(m_K,j)] }Λ [m_K(𝐓_i,𝐗_i) -𝔼(m_K)], which is a set of K estimating functions. Note that |𝔼( Φ(𝐓_i,𝐗_i;θ^*)) | = |𝔼{exp{∑_j=1^Kθ_j^*λ_j[m_K,j(𝐓_i,𝐗_i)-Em_K,j] }Λ [m_K(𝐓_i,𝐗_i) -Em_K] }| ≤sup_(𝐓_i,𝐗_i) exp{∑_j=1^Kθ_j^*λ_j[m_K,j(𝐓_i,𝐗_i)-Em_K,j] }·|𝔼Λ [m_K(𝐓_i,𝐗_i) -Em_K] | ≤ O(1) ·0 = 0, hence we have 𝔼( Φ(𝐓_i,𝐗_i;θ^*))=0, which implies that θ^* is the unique solution of 𝔼( Φ(𝐓_i,𝐗_i;θ))=0. Therefore, by the estimating equation theory (Van der Vaart (2000)), the solution of the estimating equations 1/n∑_i=1^nΦ(𝐓_i,𝐗_i;θ) = 0, denoted by θ̃, is asymptotically consistent for θ^*. Furthermore, by the Taylor expansion, we have √(n)(θ̃ - θ^*) →_d N(0, Σ), where Σ = { E(∂Φ/∂θ^') }^-1 E(ΦΦ^'){ E(∂Φ/∂θ) }^-1. Moreover, by the assumption that δ = o(n), we have 1/n√(δ)θ_j/||θ||_2 = o_p(n^-1/2) for any θ∈ int(Θ). Therefore, by the Slutsky's theorem, we obtain that √(n)(θ̂ - θ^*) →_d N(0, Σ). Let M̃_K,j(𝐓_i,𝐗_i) = λ_j(m_K,j(𝐓_i,𝐗_i)- m̅_K,j), then ŵ_i = exp{∑_j=1^Kθ̂_j λ_j (m_K,j(𝐓_i,𝐗_i)-m̅_K,j) -1} = exp{M̃_K(𝐓_i,𝐗_i)^'θ̂ -1} By Mean Value Theorem, we can deduce that ∫|ŵ-w^*|^2dF(𝐭,𝐱) ≤sup_(𝐭,𝐱) |exp{M̃_K(𝐭,𝐱)^'θ_1 -1}|^2 ×∫|M̃_K(𝐭,𝐱)^'(θ̂-θ^*) |^2 dF(𝐭,𝐱) ≤ O_p(1) ·∫|M̃_K(𝐭,𝐱)^'(θ̂-θ^*) |^2 dF(𝐭,𝐱), where θ_1 lies between θ̂ and θ^*. Since ∫|M̃_K(𝐭,𝐱)^'(θ̂-θ^*) |^2 dF(𝐭,𝐱) = ∫M̃_K(𝐭,𝐱)^'(θ̂-θ^*)(θ̂-θ^*)^'M̃_K(𝐭,𝐱) dF(𝐭,𝐱) = tr{ (θ̂-θ^*)(θ̂-θ^*)^'∫M̃_K(𝐭,𝐱)M̃_K(𝐭,𝐱)^'dF(𝐭,𝐱) } ≤ C tr{ (θ̂-θ^*)(θ̂-θ^*)^'} = C||θ̂-θ^*||^2 = O_p(n^-1). Then we have ∫|ŵ-w^*|^2dF(𝐭,𝐱) = O_p(n^-1). Furthermore, one can show that 1/n∑_i=1^n|M̃_K(𝐭,𝐱)^'(δ̂-δ^*) |^2 - ∫|M̃_K(𝐭,𝐱)^'(δ̂-δ^*) |^2 dF(𝐭,𝐱)=o_p(1). Hence, 1/n∑_i=1^n|ŵ_i- w_i^*|^2 ≤sup_(𝐭,𝐱) |exp{M̃_K(𝐭,𝐱)^'θ_1 -1}|^2 ·1/n∑_i=1^n|M̃_K(𝐭,𝐱)^'(θ̂-θ^*) |^2 ≤ O_p(1) ∫|M̃_K(𝐭,𝐱)^'(θ̂-θ^*) |^2 dF(𝐭,𝐱) +o_p(1) = O_p(n^-1) Therefore, the proof of Theorem 2 is completed. §.§.§ A.4.Proof of Theorem 3 We first show that the conclusion of Theorem 3(1). Since β̂ (as a estimator of β^*) is a unique minimizer of 1/n∑_i=1^nŵ_i(Y_i-s(T_i;β))^2(regarding 𝔼[w(Y-s(T;β))^2], according to the theory of M-estimation (van der Vaart, 2000, Theorem 5.7), if sup_β∈Θ_1|1/n∑_i=1^nŵ_̂î(Y_i-s(T_i;β))^2-𝔼[w(Y-s(T;β))^2]) |→_p 0, then β̂→_p β^*. Note that sup_β∈Θ_1|1/n∑_i=1^nŵ_̂î(Y_i-s(T_i;β))^2-𝔼[w(Y-s(T;β))^2]) | ≤sup_β∈Θ_1|1/n∑_i=1^n(ŵ_̂î-w_i)(Y_i-s(T_i;β))^2 | +sup_β∈Θ_1|1/n∑_i=1^nw_i(Y_i-s(T_i;β))^2-𝔼[w(Y-s(T;β))^2]) |. We first show that sup_β∈Θ_1|1/n∑_i=1^n(ŵ_̂î-w_i)(Y_i-s(T_i;β))^2 | is o_p(1). Using the Causchy-Schwarz inequality and the fact that ŵ→^L^2 w, we have sup_β∈Θ_1|1/n∑_i=1^n(ŵ_̂î-w_i)(Y_i-s(T_i;β))^2 | ≤{1/n∑_i=1^n(ŵ_̂î-w_i)^2 }^1/2sup_β∈Θ_1{1/n∑_i=1^n(Y_i-s(T_i;β))^2 }^1/2 ≤ o_p(1){sup_β∈Θ_1𝔼[w(Y-s(T;β))^2]+o_p(1) }^1/2 =o_p(1). Thereafter, under Assumption 5, we can conclude that sup_β∈Θ_1|1/n∑_i=1^nw_i(Y_i-s(T_i;β))^2-𝔼[w(Y-s(T;β))^2]) | is also o_p(1) (Newey and McFadden (1994), Lemma 2.4). Hence, we complete the proof for Theorem 3(1). Next, we give the proof of Theorem 3(2). Define β̂^* = argmin_β∑_i=1^n w_i(Y_i-s(𝐓_i;β))^2. Assume that 1/n∑_i=1^n w_i(Y_i - s(𝐓_i;β̂^*))h(𝐓_i;β̂^*)) = o_p(n^-1/2) holds with probablility to one as n →∞ By Assumption 5 and the uniform law of large number, one can get that 1/n∑_i=1^n w_i(Y_i-s(T_i;β))^2 →𝔼{ w(Y-s(T;β))^2 } in probability uniformly over β, which implies ||β̂^* -β^* ||→_p 0. Let r(β) = 2𝔼{ w(Y-s(T;β))h(T;β) }, which is a differentiable function in β and r(β^*) = 0. By mean value theorem, we have √(n)r(β̂^*)- ▽_β r(ζ) ·√(()n)(β̂^* - β^*) =√(n)r(β^*) =0 where ζ lies on the line joining β̂^* and β^*. Since ▽_β r(β) is continuous at β^* and ||β̂^* -β^* ||→_p 0, then √(n)(β̂^* - β^*) = ▽_β r(β^*)^-1·√(n) r(β̂^*) +o_p(1) Define the empirical process G_n(β)= 2/√(n)∑_i=1^n{ w_i(Y_i-s(T_i;β))h(T_i;β) - 𝔼{ w(Y-s(T;β))h(T;β) }}. Then we have √(n)(β̂^* - β^*) = ▽_β r(β^*)^-1·{√(n) r(β̂^*) - 2/√(n)∑_i=1^n{ w_i(Y_i-s(T_i;β̂^*))h(T_i;β̂^*) + 2/√(n)∑_i=1^n{ w_i(Y_i-s(T_i;β̂^*))h(T_i;β̂^*) } = -▽_β r(β^*)^-1· G_n(β̂^*)+o_p(1) = U^-1·{ G_n(β̂^*)-G_n(β^*) +G_n(β^*) } +o_p(1). By Assumption 5, 6, Theorem 4 and 5 of Andrews(1994), we have G_n(β̂^*)-G_n(β^*) →_p 0. Thus, √(n)(β̂^* - β^*) = U^-12/√(n)∑_i=1^n{ w_i(Y_i-s(T_i;β^*))h(T_i;β^*) } +o_p(1), then we can get that the asymptotic variance of √(n)(β̂^* - β^*) is V. Therefore, √(n)(β̂^* - β^*) →_d N(0,V). Next, we will prove β̂→_p β̂^*. Since sup_β∈Θ_1|1/n∑_i=1^nŵ_̂î(Y_i-s(T_i;β))^2-1/n∑_i=1^nw_i(Y_i-s(T_i;β))^2) | ≤sup_β∈Θ_1|1/n∑_i=1^n(ŵ_̂î-w_i)(Y_i-s(T_i;β))^2 | ≤{1/n∑_i=1^n(ŵ_̂î-w_i)^2 }^1/2sup_β∈Θ_1{1/n∑_i=1^n(Y_i-s(T_i;β))^2 }^1/2 ≤ o_p(1){sup_β∈Θ_1𝔼[w(Y-s(T;β))^2]+o_p(1) }^1/2 =o_p(1), which implies β̂^* →_p β̂. Then by Slutskey's Theorem, we can draw the conclusion that √(n)(β̂ - β^*) →_d N(0,V). Therefore, we have completed the proof of Theorem 3. §.§.§ A.5.Proof of Theorem 4 For convenience, we use a mapping Ω: R^p× q × D× R → R^p× q × D to represent the operator of absorbing the constant into the coefficients of B-spline basis for the first predictor. More precisely, Ω is defined by 𝐆^b= Ω(𝐆,c), where 𝐆_i_1,i_2,d= 𝐆_i_1,i_2,d for (i_1,i_2) ≠ (1,1) and 𝐆_1,1,d=𝐆_1,1,d+pqc, d=1,…,D. It then follows from the property of B-spline functions that c+1/pq<𝐆,Φ(𝐓)> = 1/pq <𝐆^b,Φ(𝐓)>. We also write 𝐆_0= ∑_r=1^R_0𝐁_0r∘α_0r, r= 1,…, R_0. Suppose 𝐆̂,ĉ) is a solution to (19) and 𝐆̂ = ∑_r=1^Rβ̂_1^(r)∘β̂_2^(r)∘α̂_r, then by <cit.>, Lemma B.1, there exists č∈ R and 𝐆̌ = ∑_r=1^Rβ̂_1^(r)∘β̂_2^(r)∘α̌_r, such that č+1/pq<𝐆̌ ,Φ(𝐓)> = ĉ+1/pq<𝐆̂ ,Φ̃(𝐓)>, where α̌_r= (α̌_r,1,…,α̌_r,D )^' satisfying ∑_d=1^Dα̌_r,du_d=0 with u_d= ∫_0^1 b_d(x)dx. Using (27), we have ∑_i=1^n (ŵ_iy_i-č-1/pq<𝐆̌,Φ(𝐓_i)>)^2 ≤∑_i=1^n (ŵ_iy_i-c_0-1/pq<𝐆_0,Φ(𝐓_i)>)^2. Let 𝐆̌^b= Ω(α̌,č) and 𝐆_0^b = Ω(𝐆_0,c_0), then ∑_i=1^n (ŵ_iy_i-1/pq<𝐆̌^b,Φ(𝐓_i)>)^2 ≤∑_i=1^n (ŵ_iy_i-1/pq<𝐆_0^b,Φ(𝐓_i)>)^2. Therefore, we have ∑_i=1^n ((ŵ_i-w_i+w_i)y_i-1/pq<𝐆̌^b,Φ(𝐓_i)>)^2 ≤∑_i=1^n ((ŵ_i-w_i+w_i)y_i-1/pq<𝐆_0^b,Φ(𝐓_i)>)^2, which leads to ∑_i=1^n (w_iy_i-1/pq<𝐆̌^b,Φ(𝐓_i)>)^2 ≤∑_i=1^n (w_iy_i-1/pq<𝐆_0^b,Φ(𝐓_i)>)^2 +2 ∑_i=1^n (ŵ_i-w_i)y_i(1/pq<𝐆̌^b-𝐆_0,Φ(𝐓_i)). Let 𝐆^# =𝐆̌^b-𝐆_0^b, 𝐚^# =vec(𝐆^# ), 𝐚_0^b =vec(𝐆_0^b), 𝐚̌^b= vec(𝐆̌^b) and 𝐙 = (𝐳_1,…,𝐳_n)^'∈ R^n× pqD, where 𝐳_i = vec(Φ(𝐓_i)),i=1,…,n. Let y_ŵ= (ŵ_1y_1,…, ŵ_ny_n)^' and y_w = (w_1y_1,…, w_ny_n)^', then using (31) and working out the squares, we obtain 1/p^2q^2||𝐙𝐚^#||^2 ≤ 2<1/pq𝐙𝐚̌^b,y_w>-2< 1/pq𝐙𝐚_0^b,y_w> -21/p^2q^2<𝐙𝐚^#,𝐙𝐚_0^b>+2<1/pq𝐙𝐚^#, y_ŵ-y_w> = 2<1/pq𝐙𝐚^#,y_ŵ-y_w>+2<1/pq𝐙𝐚^#,ϵ>+2< 1/pq𝐙𝐚^#,y_w-ϵ-1/pq𝐙𝐚_0^b> First, we show the upper bound of <1/pq𝐙𝐚^#,y_ŵ-y_w>. Using the Cauchy-Schwarz inequality, we have <1/pq𝐙𝐚^#,y_ŵ-y_w> ≤||ŵ-w ||_2 ·||1/pq𝐙𝐚^#||_2 ≤C_1√(nh_n)/pq·||ŵ-w ||_2 ·||𝐚^#||_2 By the conclusion of Theorem 2(2), we have ||ŵ-w ||_2 = O_p(1). Applying (37) to (36), we can obtain that <1/pq𝐙𝐚^#,y_ŵ-y_w> ≤C_2√(nh_n)/pq||𝐚^#||_2. Second, by the conclusion of <cit.>, (A.18) and (A.20), we can obtain the upper bound of <1/pq𝐙𝐚^#,ϵ> and < 1/pq𝐙𝐚^#,y_w-ϵ-1/pq𝐙𝐚_0^b>, which are <1/pq𝐙𝐚^#,ϵ> ≤C_3/pq||𝐚^#||_2 { nh_n(R^3+R(p+q)+RD)}^1/2. and <1/pq𝐙𝐚^#,y_w-ϵ-1/pq𝐙𝐚_0^b> ≤C_4/pq||𝐚^#||_2 {∑_r=1^R_0||vec(𝐁_0r)||_1/pq}n√(h_n)/D^τ Therefore, applying (38), (39) and (40) to (35), we have C_5/pq||𝐚^#||_2^2 ≤ R_1 ||𝐚^#||_2, where R_1 = C_6√(D/n)+C_7 {D(R^3+R(p+q)+RD)/n}^1/2+C_8{∑_r=1^R_0||vec(𝐁_0r)||_1/pq}1/D^τ-1/2. By solving the second order inequality (41), we have C_5/pq||𝐚^#||_2 ≤ R_1 Further, by Assumption 6 and <cit.>, (A.38) of Lemma A.2, we have ||ŝ(𝐓) -s(𝐓) ||^2 ≤ C_9 h_n 1/p^2q^2||𝐚^#||^2 = C_10R_1^2/D = O_P(1/n)+O_p(R^3+R(p+q)+RD/n)+O_p({∑_r=1^R_0||vec(𝐁_0r)||_1/pq}^2 1/D^2τ) =O_p(R^3+R(p+q)+RD/n)+O_p({∑_r=1^R_0||vec(𝐁_0r)||_1/pq}^2 1/D^2τ). Hence, the proof of Theorem 4 is completed. § APPENDIX REFERENCE Mohri, M., Rostamizadeh, A., and Talwalkar, A. (2018). Foundations of ma- chine learning. MIT press. 0.2cm Newey, W. K. and McFadden, D. (1994). Large sample estimation and hypoth- esis testing. Handbook of econometrics, 4:2111–2245. 0.2cm Van der Vaart, A. W. (2000). Asymptotic statistics, volume 3. Cambridge university press. 0.2cm Zhou, Y., Wong, R., and He, K. (2020). Broadcasted nonparametric tensor regression.

http://arxiv.org/abs/2303.07039v1

20230313120122

Time-Optimal Path Tracking for Cooperative Manipulators: A Convex Optimization Approach

cs.RO

[t1]This work was supported in part by the Swedish Department of Education within the strategic research environment ELLIIT and in part by the Vinnova Competence Center LINK-SIC at Linköping University. fn1]Hamed Haghshenascor1 hamed.haghshenas@liu.se fn1]Anders Hansson anders.g.hansson@liu.se fn2]Mikael Norrlöf mikael.norrlof@se.abb.com [cor1]Corresponding author [fn1]Hamed Haghshenas and Anders Hansson are with the Division of Automatic Control, Department of Electrical Engineering, Linköping University, 581-83 Linköping, Sweden. [fn2]Mikael Norrlöf is with ABB Robotics Division, 721-36 Västerås, Sweden. This paper studies the time-optimal path tracking problem for a team of cooperating robotic manipulators carrying an object. Considering the problem for rigidly grasped objects, we show that it can be cast as a convex optimization problem and solved efficiently with a guarantee of optimality. When formulating the problem, we avoid using a particular wrench distribution and exploit the full actuation available to the system. Then, we consider the problem for grasps using frictional forces and show that this problem also, under a force-closure grasp assumption, can be formulated as a convex optimization problem and solved efficiently and to optimality. To ensure a firm grasp, internal forces have been taken into account in this approach. Path tracking, time-optimal, cooperative manipulation, convex optimization, contact with friction. § INTRODUCTION Robotic systems are nowadays the key technology in a wide range of application domains, from construction, manufacturing and agriculture to search and rescue, and service robotics. The increasing demand for performance of robotic systems is often met by using multiple robots for a specific task. A team of cooperating robots outperforms the functionality of a single robot; like a human using two arms has an advantage over one using only one arm. When a task conducted by multiple robots involves manipulation of an object, the multi-robot system is said to perform a cooperative manipulation task. Cooperative manipulation is an important capability for extending the domain of robotic applications. Typical examples include industrial manipulators manipulating large or heavy objects as well as transportation tasks conducted by multiple robots. One particular subject that most applications have in common is path tracking. Path tracking is the second stage of the so-called decoupled approach <cit.>; an approach for solving motion planning problems. The first stage of the decoupled approach, known as path planning, determines a path while taking geometric aspects of the task and the environment into account, whereas the path tracking stage is concerned with the dynamic aspects of the robot and the task. In many robotic applications, the path is determined by the task and its specifications. Hence, assuming that a desired geometric path is given, the path tracking problem can be studied independently and is of value on its own. For a wide range of robotic applications, it is desired to minimize the execution time of a task or some other criteria such as the energy consumed during the motion. For example, robotic manipulators performing a variety of tasks are included in almost every production line, and there is an obvious relation between the execution time of the tasks and productivity. Hence, time-optimal and in general optimal motion planning is of great significance for robotic systems. By optimizing the robots' motions while taking their dynamics into account, robots can fully exploit their capabilities. Motivated by these observations, this paper studies the time-optimal path tracking problem for a team of cooperating robotic manipulators. We consider a scenario where an object is grasped by multiple manipulators. Time-optimal cooperative path tracking addresses the problem of moving the object along a predefined geometric path in minimum time, which requires realizing a velocity as high as possible along the path while satisfying the imposed constraints on the motion. Note that solving this problem gives the nominal trajectories. For online path tracking, the nominal, feedforward trajectories are combined with feedback in a control architecture such that robustness to uncertainties in model or environment is achieved (see e.g., <cit.>). Combining the feedforward trajectories with a feedback is beyond the scope of this work and our focus is on solving the time-optimal path tracking problem. Most of the works on the area of time-optimal path tracking are focused on a single manipulator. Contrary to the case of a single manipulator, a given time parameterization of the path in the case of cooperative manipulators may correspond to infinitely many possible torque profiles. One possible way to deal with this redundancy is to consider a predefined load distribution. It is however clear that this would give rise to slower trajectories than what can be obtained using the full actuation available to the system. Our goal in this paper is to extend the time-optimal path tracking to the case of cooperative manipulators so that it can optimally handle the actuation redundancy. There are different methods for solving the time-optimal path tracking problem, most of which exploit the fact that the motion along a predetermined path can be described by a single parameter, called path coordinate and denoted by s, and its time derivatives <cit.>. These approaches can basically be divided into three groups: numerical integration, dynamic programming, and convex optimization. The numerical integration approach <cit.> is based on the Pontryagin maximum principle: the time-optimal velocity profile in the (s,ṡ) plane is known to be bang–bang and can be obtained by integrating successively the maximum and minimum accelerations s̈. This requires finding the switch points between accelerating and decelerating segments, which constitutes a major implementation difficulty as well as the main cause of failure <cit.>. This is because of the robustness issues associated with the so-called dynamic singularities which occur when the maximum velocity curve is nondifferentiable (see <cit.>). Furthermore, finding the switch points and computing the maximum velocity curve become computationally expensive for complex constraints with many inequalities <cit.>. This approach is theoretically faster than the other two approaches, since it exploits the bang-bang structure of the problem. A limitation of this approach is that it is restricted to the time-optimal case. An algorithm for handling of dynamic singularities for robots subject to joint velocity and acceleration constraints is presented in <cit.>. In <cit.>, a more general approach that can account for dynamic singularities resulting from second-order constraints together with an open-source implementation of the algorithm in C/Python is presented. In <cit.>, a complete and time-optimal algorithm based on numerical integration is proposed. In <cit.>, an algorithm that solely treats jerk limits is proposed. The numerical integration approach is extended to the case of redundantly actuated systems in <cit.>. An object manipulated by several manipulators can be considered as a redundantly actuated system. In their approach, they use the equations of motion of an open-chain system obtained by cutting the closed-chain system at some joints. This requires obtaining the dynamics of the open-chain system where the chain is cut. If the dynamic models of the manipulators are already available, obtaining the dynamics of another open-chain system can be an extra task. Also, to compute the optimal accelerations at the singular points, one cannot use the analytical value suggested in <cit.> and has to search numerically for an approximate optimal acceleration. Furthermore, after obtaining the optimal parameterization, to obtain the torques at each actuated joint another optimization problem must be solved whose objective is specified by the user. Thus, the approach proposed in <cit.> does not use the full actuation available to the system. Compared to that work, our approach exploits the full actuation available to the manipulators. Also, it is straightforward to impose constraints on the wrenches acting on the object in our approach, whereas in <cit.> it is not clear how to do this. Time-optimal path tracking for cooperative multi-manipulator systems was independently studied in <cit.> and <cit.> based on numerical integration, where linear programming was used to find the maximum acceleration and deceleration at a given point in the (s,ṡ) plane. However, since the algebraic form of extreme accelerations is not available from linear programming, they are not able to develop a systematic search scheme to find the switching points, and their proposed procedures give an approximation to the time-optimal trajectory. In the dynamic programming approach, the problem is solved by dividing the (s,ṡ) plane into a grid and subsequently using dynamic programming (see <cit.>). One advantage of this method is the ability to account for a general form of the objective functions and constraints. The main disadvantage of this approach is the high computational cost due to the need for solving a problem with a large number of variables. In <cit.>, an algorithm which is capable of taking limits on joint jerks and torque rates into account is presented. Additionally, viscous friction is included in the dynamic model of the manipulator, which has been made possible by utilizing the idea of dynamic programming. The convex optimization approach converts the original problem, which turns out to be an infinite-dimensional optimization problem, into a finite-dimensional one through discretization of the s-axis. One of the early works using this approach is <cit.>, where it is shown that by exploiting a nonlinear change of variables, the time-optimal path tracking problem can be cast as a convex optimization problem and subsequently as a second-order cone program (SOCP). This allows for efficiently solving the problem by utilizing the wide variety of algorithms and software developed for convex optimization <cit.>. Furthermore, the convexity of the problem guarantees that any locally optimal solution is also a globally optimal one. This approach has also the advantage that other objective functions such as energy or torque rate, can be incorporated. It must be noted that the change of variables used in <cit.> has been known since at least 1985 (see e.g., <cit.>), although the resulting convexity of the problem was not noted back then. The main disadvantage of the convex optimization approach is computational cost, which can be an order of magnitude greater than numerical integration-based approaches. To reduce computational cost, <cit.> proposes an approach based on sequential linear programming. Further, <cit.> shows that using a special discretization scheme, the time-optimal velocity profile can be obtained by linear programming with the benefit of lower computation cost with respect to convex solvers. Because of the advantages that the convex formulation offers, much work has been conducted to extend the range of applicability of these ideas to other problems such as path tracking for different types of vehicles as in <cit.>, or to include new types of constraints in the robot path tracking problem <cit.>. Similarly as in <cit.>, viscous friction which renders the problem nonconvex is not considered in the robot model in <cit.>. In this context, some works such as <cit.> and <cit.> have addressed the incorporation of the constraints that destroy the convexity of the problem. The work <cit.> incorporates speed dependent constraints into the problem and replaces the nonconvex constraints with their convex approximations to preserve the convexity of the problem, whereas the work <cit.> considers several robotic applications that result in nonconvex problems and employs sequential convex programming to solve the corresponding problem by writing the nonconvex constraints as differences of convex functions. In <cit.>, change rate of torque and voltage of the DC motor are incorporated. By introducing some new variables which are tight approximation of change rate of torque and motor voltage, the original nonconvex problem is formulated as an approximate convex optimization problem. The work by <cit.> formulates the time-optimal path tracking problem for a dual-robot system as a nonconvex problem and solves an approximate problem using sequential convex programming. The change of variables introduced in <cit.> is also used in other studies. In <cit.>, a two-step iterative learning algorithm for path tracking problem of a robotic manipulator is proposed which compensates for a possible model-plant mismatch and improves the tracking performance. From the mentioned approaches for solving the time-optimal path tracking problem, only the numerical integration approach has been applied to cooperative manipulators. As mentioned earlier, the works that use this method for cooperative manipulators either do not use the full actuation available to the system <cit.> or are not able to develop a systematic search for finding the switch points <cit.>. Additionally, this approach suffers from robustness issues associated with the dynamic singularities and is limited to the time-optimal case. Because of the advantages that the convex optimization approach offers, in this paper we use this method to solve the time-optimal path tracking problem for a team of cooperating manipulators. With the availability and maturity of SOCP solvers, this approach reliably and robustly solve the problem of interest. With this approach, it is also possible to incorporate objective functions other than time into the problem. First, we consider a scenario where an object is rigidly grasped by multiple manipulators. The rigid grasping assumption implies that the manipulators can apply forces and moments along all directions to the object. By exploiting the mentioned change of variables we show that this problem can be formulated as a convex optimization problem and subsequently as an SOCP. Due to multiple manipulators being involved in the manipulation task, there exists an infinite number of forces/moments exerted by the end-effectors that result in the same desired force/moment on the centre of mass of the object. This redundancy is fully exploited in our formulation. Others (e.g., <cit.>) use specific force/moment distributions, also called wrench distribution, which result in suboptimal solutions. We then relax the assumption on the rigid grasping and consider contacts with friction. These types of contacts complicate the problem as each contact may only apply forces/moments that respect friction cone constraints to prevent slip, instead of arbitrary forces/moments associated with rigid contacts. We show that this problem also, under a force-closure grasp assumption, can be formulated as a convex optimization problem. In summary, the following are the contributions of this paper: * The time-optimal path tracking problem for a team of cooperative manipulators rigidly grasping an object is formulated as a convex optimization problem. * The results from the first contribution are extended to consider contacts with friction. * Our approach optimally handles the actuation redundancy and uses the full actuation available to the system. In <cit.>, we have presented a preliminary study of the time-optimal path tracking problem for a particular setup comprised of two two-link planar manipulators with non-actuated end-effectors rigidly grasping a bar. Compared to that work where only two manipulators are involved and the object is a bar grasped at the two end points, in this paper we consider a general scenario with N generic manipulators and a generic object. Furthermore, a particular wrench distribution was assumed in <cit.>, whereas here we avoid using any wrench distribution. The rest of this paper is organized as follows. Section <ref> presents some known results on time-optimal path tracking problem for a single manipulator. Section <ref> states the problem of interest. Section <ref> presents the modeling of the coupled kinematics and the dynamics of the manipulators and the object. Section <ref> studies the time-optimal cooperative path tracking problem with rigid contacts. Section <ref> extends the results in Section <ref> to consider contacts with friction. Section <ref> contains some simulation results. Section <ref> summarizes our findings and presents ideas for future work. § PRELIMINARIES Here, we collect some known results about the time-optimal path tracking problem for a single manipulator that will be used in the rest of the paper. Manipulator motion along a prescribed trajectory can be written as a function of a single parameter, s, either in task space <cit.> or in joint space <cit.>. Given a prescribed geometric path q(s) in joint space, the joint velocities and accelerations can, using the chain rule, be written as q̇(s) = q'(s) ṡ, q̈(s) = q'(s) s̈ + q”(s) ṡ^2, where q'(s) = ∂ q(s)/∂ s, q”(s) = ∂^2q(s)/∂ s^2, ṡ = ds/dt and s̈ = d^2s/dt^2. It is shown in <cit.> that the time-optimal path tracking problem for a single robotic manipulator subject to lower and upper bounds on the torques, can be cast as the convex optimization problem \beginmini![2] ∫_0^11/√(b(s))ds τ(s)=m(s)a(s)+c(s)b(s)+g(s), b(0)=ṡ_0^2, b(1)=ṡ_T^2, b(s)≥0, b'(s)=2a(s), τ(s) ≤τ(s) ≤τ(s), ∀s ∈[0,1], \endmini! where the optimization variables in addition to joint torques, τ(s), are the acceleration and square of the speed along the path coordinate, i.e., a(s) = s̈(t), b(s) = ṡ(t)^2. The key behind this result is to use the acceleration as a free variable, which can be accomplished by representing the joint velocities and accelerations as (<ref>), and substituting them in the equations of motion. The constraint in (<ref>) shows the equations of motion of a single manipulator after this substitution and using (<ref>) as optimization variables. The change of integration variable from time t to s allows to write the objective function, i.e. the duration of the motion, as (<ref>). In the above optimization problem, τ and τ are the lower and upper bounds on the joint torques, respectively, and can be functions of s, and ṡ_0 and ṡ_T are, respectively, the initial and final velocities along the path, usually chosen to be 0. In this problem, the inequalities are interpreted as component-wise inequalities. The optimization problem in (<ref>) is convex since the objective function is convex in b(s), which follows from the fact that integration preserves convexity and that 1/√(b(s)) is convex in b(s), and since all the constraints are linear in the optimization variables. Extensions of the problem to incorporate other objective functions and constraints that preserve convexity can be found in <cit.>. In particular, it is shown that symmetric lower and upper bounds on the joint velocities can be translated into upper bounds on b(s). We denote this upper bound by b(s), and the corresponding constraint will be incorporated into the forthcoming optimization problems. The convex optimization problem in (<ref>) is an infinite dimensional problem with infinitely many optimization variables and constraints. Therefore, the direct transcription method is employed and the problem in (<ref>) is formulated as a large sparse optimization problem in <cit.>. The resulting problem is finally transformed into an SOCP, which can be solved efficiently using dedicated solvers, such as MOSEK <cit.>. The reader is referred to <cit.> for more details. § PROBLEM STATEMENT This section addresses the problem of time-optimal path tracking for multiple cooperative manipulators rigidly grasping an object. Consider N fully actuated robotic manipulators, indexed by the set 𝒩 = {1, …, N}, rigidly grasping an object as in  <ref>. The rigidity assumption implies that the manipulators can exert forces and moments along all directions of the object. The objective is to minimize the traversal time required to move the object with a desired orientation along a prescribed geometric path subject to constraints on the joint torques and velocities. We assume that the prescribed geometric path is given for the centre of mass of the object, and that both the path and the object's orientation are given as functions of a scalar path coordinate s. Without loss of generality, assume that the trajectory starts at t=0, ends at t=T, and that s(0)=0 ≤ s(t) ≤ s(T)=1. Furthermore, it is assumed that ṡ(t) ≥ 0, ∀ t ∈ [0,T]. In other words, we always move forward along the path. Throughout the paper, all quantities are expressed with respect to an inertial frame of reference {I} whenever the reference frame is not explicitly indicated by a superscript. Let {O} and {E_i} denote the reference frames attached to the object at the centre of mass and the ith end-effector, respectively. Let ϕ_ O(s) = [φ_ O(s), ϑ_ O(s), ψ_ O(s)]^T be the object's orientation in terms of a triplet of Euler angles. In the following section, we present the modeling of the coupled kinematics and the dynamics of the manipulators and the object. § KINEMATIC AND DYNAMIC MODELS §.§ Kinematics Let p_ O∈ℝ^3 be the position of the object's centre of mass. Let p_i ∈ℝ^3 and ϕ_i be the position and orientation of the ith end-effector, respectively. In view of  <ref>, one can see that the pose of the ith end-effector and the object's centre of mass are related by p_i(s) = p_ O(s) + p_i O(s) = p_ O(s) + R_ E_i(s) p_i O^ E_i, ϕ_i(s) = ϕ_ O(s) + ϕ_i O, where p_i O^ E_i and ϕ_i O are the constant distance and orientation offsets between the reference frames {O} and {E_i}, respectively, and are considered to be known. The 3× 3 matrix R_ E_i denotes the rotation matrix of the reference frame {E_i} with respect to {I}. §.§ Manipulator Dynamics The equations of motion for the ith manipulator with n_i DOF and the joint variables q_i∈ℝ^n_i can be described by <cit.> M_i(q_i)q̈_i+C_i(q_i,q̇_i)q̇_i+g_i(q_i)=τ_i - J_i(q_i)^T h_i, where τ_i∈ℝ^n_i are the joint torques, M_i(q_i) ∈ℝ^n_i × n_i is the positive definite inertia matrix, C_i(q_i,q̇_i) q̇_i∈ℝ^n_i × 1 is the vector of Coriolis and centrifugal forces, g_i(q_i) ∈ℝ^n_i is the vector of gravitational terms, J_i∈ℝ^6 × n_i is the ith manipulator's geometric Jacobian, and h_i∈ℝ^6 is the vector of generalized forces exerted by the ith end-effector on the object. A vector of generalized forces acting on a rigid body consists of a linear component (pure force) and an angular component (pure moment). We will refer to a force/moment pair as a wrench. In (<ref>), the torques J_i^T(q_i) h_i are a portion of the actuation torques that is needed to balance the torques induced at the joints by the contact forces. §.§ Object Dynamics Regarding the object, the pose and the velocity of its centre of mass are denoted by x_ O = [p_ O^T, ϕ_ O^T]^T and v_ O = [ṗ_ O^T, ω_ O^T]^T, respectively, where ω_ O∈ℝ^3 is the object's angular velocity. The following second-order dynamics, which is based on the Newton-Euler formulation can be considered for the object: M_ O(x_ O)v̇_ O + C_ O(x_ O,ẋ_ O)v_ O + g_ O = h_ O, where M_ O = [ m I_3 0_3 × 3; 0_3 × 3 I_ O ], C_ O = [ 0_3 × 3 0_3 × 3; 0_3 × 3 S(ω_ O) I_ O ], g_ O = [ - m g; 0_3 ], and where m is the mass of the object, g ∈ℝ^3 is the gravitational acceleration vector, I_ O∈ℝ^3 × 3 is the object's inertia tensor relative to the centre of mass when expressed in a frame parallel to {I} with origin at the centre of mass, and h_ O∈ℝ^6 is the wrench acting on the object's centre of mass. Here, S(·) is the skew-symmetric matrix operator performing the cross product, i.e., S(u)w = u × w for u,w ∈ℝ^3. § TIME-OPTIMAL COOPERATIVE PATH TACKING WITH RIGID CONTACTS In this section, we show that the time-optimal path tracking problem for cooperative manipulators rigidly grasping an object can be cast as a convex optimization problem. The first step is to rewrite the manipulator dynamics in terms of s̈, ṡ^2 and s. Given p_ O(s) and ϕ_ O(s), the pose of each end-effector can be computed by (<ref>). Then, inverse kinematics can be employed to compute the corresponding joint variables as functions of the path coordinate s. Using the chain rule, the joint velocities and accelerations of the ith manipulator can be written as q̇_i(s) = q'_i(s) ṡ, q̈_i(s) = q'_i(s) s̈ + q”_i(s) ṡ^2, where q'_i(s) = ∂ q_i(s)/∂ s and q”_i(s) = ∂^2q_i(s)/∂ s^2. Substituting these expressions in (<ref>) yields the following expression for the equations of motion: τ_i(s) = m_i(s) s̈ + c_i(s) ṡ^2 + g_i(s) + J_i(s)^T h_i(s), where m_i(s) = M_i(q_i(s)) q'_i(s), c_i(s) = M_i(q_i(s)) q”_i(s) + C_i(q_i(s),q'_i(s)) q'_i(s), for each i ∈𝒩. The second step is to show that the object dynamics in (<ref>), similar to Equation <ref> can be written as a function of s̈, ṡ^2 and s. This allows us to use s̈ and ṡ^2 as optimization variables later. To this end, we start by writing the first- and the second-order time derivatives of the pose of the object's centre of mass as ẋ_ O(s) = x_ O'(s) ṡ, ẍ_ O(s) = x_ O'(s) s̈ + x_ O”(s) ṡ^2, respectively, where x_ O'(s) = ∂ x_ O(s)/∂ s and x_ O”(s) = ∂^2x_ O(s)/∂ s^2. Next, we find the relationship between the velocity and the time derivative of the pose of the object's centre of mass. For this, we know that it is possible to find the relationship between the object's angular velocity ω_ O and its rotational velocity ϕ̇_ O for a given set of orientation angles <cit.>. Let T(ϕ_ O) be the transformation between these two velocities, i.e., ω_ O = T(ϕ_ O) ϕ̇_ O. Using this equation, the velocity of the object's centre of mass can be written as v_ O(s) = T_ O(s) ẋ_ O(s), where T_ O(s) = [ I_3 0_3 × 3; 0_3 × 3 T(ϕ_ O(s)) ]. Substituting (<ref>) into (<ref>) and its derivative, yields the following expressions for v_ O and v̇_ O, respectively: v_ O(s) = T_ O(s) x_ O'(s) ṡ, v̇_ O(s) = (∂ T_ O(s)/∂ s x_ O'(s)+T_ O(s) x_ O”(s))ṡ^2 + T_ O(s) x_ O'(s) s̈. Finally, substituting (<ref>) and (<ref>) into (<ref>), an expression for the object dynamics can be obtained as follows: m_ O(s)s̈ + c_ O(s)ṡ^2 + g_ O = h_ O(s), where m_ O(s) = M_ O(x_ O(s)) T_ O(s) x_ O'(s), c_ O(s) = M_ O(x_ O(s)) ∂ T_ O(s)/∂ s x_ O'(s) + M_ O(x_ O(s)) T_ O(s) x_ O”(s) + C_ O(x_ O(s),x_ O'(s)) T_ O(s) x_ O'(s). The final step concerns the relationship between the wrench h_ O acting on the object's centre of mass and the wrenches h_i, i ∈𝒩 exerted by the end-effectors to the object at the grasping points. Let v_i = [ṗ_i^T, ω_i^T]^T be the velocity of the ith end-effector, where ω_i∈ℝ^3 is the angular velocity. Differentiation of (<ref>) together with the fact that ω_i = ω_ O due to the grasping rigidity, leads to v_i(s) = J_ O_i(s) v_ O(s), ∀ i ∈𝒩, where J_ O_i is called the object-to-agent Jacobian matrix, with J_ O_i(s) = [ I_3 -S(p_i O(s)); 0_3 × 3 I_3 ], ∀ i ∈𝒩, which is full-rank due to the grasp rigidity. The grasp matrix G ∈ℝ^6 × 6N is formed by stacking J_ O_i(s)^T, i ∈𝒩 as G(s) = [J_ O_1(s)^T ⋯ J_ O_N(s)^T], and has full row rank. Let v(s) = [v_1(s)^T ⋯ v_N(s)^T]^T. Equation (<ref>) can now be written in the following form: v(s) = G(s)^T v_ O(s). Let h = [h_1^T ⋯ h_N^T]^T. The kineto-statics duality <cit.> together with the grasp rigidity suggest that the wrenches h_ O and h_i, i ∈𝒩 are related by h_ O(s) = G(s) h(s) = ∑_i ∈𝒩^ G_i(s) h_i(s), where G_i is the ith block of the grasp matrix, i.e., G_i = J_ O_i^T. Finally, having the dynamic equations of the manipulators and the object written in terms of s̈, ṡ^2 and s, and having the relationship between the wrenches h_ O and h_i, i ∈𝒩, we are ready to present the convex formulation of the problem of interest. From the equations in (<ref>), (<ref>) and (<ref>) it can be seen that choosing a(s) = s̈ and b(s) = ṡ^2 together with h_ O(s) and h_i(s), τ_i(s), i ∈𝒩 as optimization variables, makes these equations, which enter as constraints in the optimization problem, linear in the optimization variables. The objective function for our problem of interest is the same as (<ref>), which is convex in b(s). With the introduced optimization variables, the problem of interest can be cast as the following convex optimization problem: [2] ∫_0^11/√(b(s))ds τ_i(s) = m_i(s) a(s) + c_i(s) b(s) + g_i(s) + J_i(s)^T h_i(s), i ∈𝒩, h_O(s) = m_O(s)a(s) + c_O(s)b(s) + g_O, h_O(s) = ∑_i ∈𝒩^ G_i(s) h_i(s), b(0)=ṡ_0^2, b(1)=ṡ_T^2, op:general:convex 0 ≤b(s) ≤b(s), b'(s)=2a(s), τ_i(s) ≤τ_i(s) ≤τ_i(s), i ∈𝒩, ∀s ∈[0,1], with variables a, b, h_ O and τ_i, h_i, i ∈𝒩. In (<ref>), τ_i and τ_i are the lower and upper bounds on the ith manipulator's joint torques, respectively. This problem can finally be reformulated as an SOCP using the direct transcription method and a procedure similar to the one used in <cit.>. Note that to consider all possible combinations of wrenches when finding the optimal solution, the relationship between the wrenches h_ O and h_i, i ∈𝒩 is intentionally kept as (<ref>), instead of solving (<ref>) for h_i, i ∈𝒩 and representing h_i, i ∈𝒩 based on a particular wrench distribution. For a cooperative manipulation task, in general, there exists an infinite number of wrenches exerted by the end-effectors that result in the same desired wrench on the centre of mass of the object. This is due to multiple manipulators being involved in the manipulation task. When using a predefined wrench distribution, this redundancy is lost. Our approach uses the full actuation available to the system and optimally handles the actuation redundancy. Compared to a case where a particular wrench distribution is used, a smaller minimal traversal time is expected from our approach. A numerical comparison using some common wrench distributions is carried out and reported in Section <ref> which confirms this point. § EXTENSION TO GRASPING WITH FRICTION In this section we relax the assumption on the rigid grasping points. We consider the problem stated in Section <ref>, but with a different type of grasp. In particular, we assume that contacts between the manipulators and the object are of point contact with friction or soft-finger types. Also, we assume a force-closure grasp, which will be defined later. For grasps using frictional forces, a model for friction must be provided. For this purpose, we will use a simple model which is referred to as the Coulomb friction model. This model is an empirical model which states that the allowed force in the tangential directions to a surface is proportional to the applied force at the normal direction. The constant of proportionality which is referred to as the coefficient of friction, is determined by the materials that are in contact. A point contact with friction model is used when friction exists between the manipulator tip and the object, in which case forces can be exerted in any direction that is within the friction cone for the contact. For a soft-finger contact, not only forces are allowed to be applied in a cone about the surface normal, but also torques about that normal <cit.>. These types of contacts complicate the problem since each contact may only apply a wrench that respects friction cone constraints to prevent slip, instead of an arbitrary wrench associated with rigid contacts. We start by describing the grasp. The ith contact can be modeled using a wrench basis, B_ C_i∈ℝ^6× m_i, and a friction cone, FC_ C_i <cit.>. The dimension of the wrench basis, m_i, indicates the number of independent forces/moments that can be applied by the contact. Let us consider the planar grasp shown in  <ref>, with point contacts with friction. For convenience, the contact coordinate frame 𝒞_i, is chosen in such a way that its x-axis points in the direction of the inward surface normal at the point of contact. Let h_ C_i be the wrench applied at the ith contact point. We represent this wrench with respect to a basis of directions which are consistent with the friction model: h_ C_i = B_ C_i f_ C_i = [ 1 0 0 0 0 0; 0 1 0 0 0 0; ]^T f_ C_i, f_ C_i∈ FC_ C_i, where the friction cone for the ith contact is FC_ C_i = { f= [ f_x f_y ]^T ∈ℝ^2 : | f_y | ≤μ f_x, f_x ≥ 0 }, In (<ref>), f_x and f_y are the normal and tangential force components, respectively, and μ > 0 is the coefficient of friction. For a soft-finger contact with a coordinate frame where its x-axis points in the direction of the inward surface normal, the wrench basis and the friction cone are as follows: B_ C_i = [ 1 0 0 0 0 0; 0 1 0 0 0 0; 0 0 1 0 0 0; 0 0 0 1 0 0; ]^T, FC_ C_i = { f ∈ℝ^4 : √(f_y^2+f_z^2)≤μ f_x, f_x ≥ 0, | t_x | ≤γ f_x }, where f= [ f_x , f_y , f_z , t_x ]^T, and where γ > 0 is the coefficient of torsional friction. The contact map, G̅_i ∈ℝ^6× m_i, is defined to be the linear map between the ith contact wrench represented with respect to B_ C_i, and the object wrench at the centre of mass: G̅_i = G_i B_ C_i. With this definition, the object wrench can be written as h_ O = G̅ f_ C, f_ C∈ FC, where G̅ = [ G_1 B_ C_1⋯ G_N B_ C_N; ], is called grasp map, and where f_ C = [ f_ C_1^T ⋯ f_ C_N^T; ]^T ∈ℝ^m, FC = FC_ C_1×⋯× FC_ C_N⊂ℝ^m, m = m_1 + ⋯ + m_N. The full friction cone, FC, is the Cartesian product of all the friction cones. A grasp is completely described by the grasp map G̅ and the friction cone FC. Wrenches in the null space of G̅ correspond to those wrenches that can be exerted at the contact points without resulting in a net wrench on the object. These are referred to as internal forces. For collaborative manipulation with the mentioned types of contacts, it is critical to ensure that the object does not slip. To ensure this, each contact wrench must remain in the respective friction cone. Internal forces can be used to insure that contact wrenches satisfy friction cone constraints. Any given vector of contact wrenches can be brought into the friction cone by adding a sufficiently large wrench in the null space of G̅, if the condition N(G̅) ∩int(FC) ≠∅, where N(·) denotes the null space of a matrix and int(FC) is the interior of the friction cone, is satisfied <cit.>. In order to be able to firmly grasp an object, it is desirable that internal forces exist and lie in the interior of the friction cone. It is also desirable that any given wrench on the object can be achieved by an appropriate choice of contact wrenches lying in the friction cone. The mathematical characterization of this ability is G̅(FC)=ℝ^6, i.e., the grasp map G̅ should map the friction cone FC onto ℝ^6. This ability is linked to the ability of a grasp to resist any applied wrench, which is called force-closure. A grasp is a force-closure grasp if given any external wrench h_e ∈ℝ^6 applied to the object, there exist contact wrenches f_ C∈ FC such that G̅ f_ C = - h_e. It follows directly from this definition that a grasp is force-closure if and only if G̅(FC)=ℝ^6 <cit.>. It is also shown in <cit.> that the existence of an internal force belonging to the interior of the friction cone is a necessary condition for a grasp to be force-closure. In order to have a grasp with the two desirable properties mentioned above, i.e., existence of an internal force which belongs to the interior of the friction cone and G̅(FC)=ℝ^6, we assume that the grasp in our problem is a force-closure grasp. Constructing force-closure grasps is not the focus of this paper and the reader is referred to the related work on this subject, see e.g., <cit.>. One particularly simple method for the case of two contact points is proposed in <cit.>. A planar grasp with two point contacts with friction is force-closure if and only if the line connecting the contact points lies inside both friction cones. This result can be extended to the case of a spatial grasp with two soft-finger contacts. Now, we are ready to present the optimization problem associated with the problem of interest in this section. In order to ensure that slipping does not occur, the friction cone constraints will enter as constraints in the optimization problem, and it can be seen from (<ref>) and (<ref>) that these constraints are convex. This follows from the fact that every norm on ℝ^n is convex <cit.>. In order to consider internal forces, we divide the wrench applied at the ith contact point into two parts: a motion-inducing part, denoted by h_ M_i, and a part associated with internal forces, denoted by h_ I_i. In other words, h_ C_i = h_ M_i + h_ I_i, i ∈𝒩, where h_ I = [h_ I_1^T ⋯ h_ I_N^T]^T ∈ N(G). The constraint on h_ I being in the null space of G can be written as ∑_i ∈𝒩^ G_i h_ I_i = 0. Let f_ M_i and f_ I_i be the contact wrenches corresponding to h_ M_i and h_ I_i, represented with respect to B_ C_i, respectively, i.e., h_ M_i = B_ C_i f_ M_i and h_ I_i = B_ C_i f_ I_i. As mentioned above, in order to have a firm grasp, it is desirable that the internal forces lie in the interior of the friction cone. The constraints f_ I_i∈int(FC_ C_i), i ∈𝒩 will be added to the optimization problem for this purpose. Note that to prevent slip, f_ C_i = f_ M_i + f_ I_i must also respect the friction cone constraints. By introducing h_ M_i, h_ I_i, f_ M_i and f_ I_i as additional optimization variables and with the discussed changes, the problem of interest in this section can be cast as the following convex optimization problem: [2]<b> ∫_0^11/√(b(s))ds τ_i(s) = m_i(s) a(s) + c_i(s) b(s) + g_i(s) + J_i(s)^T (h_M_i(s) + h_I_i(s)), i ∈𝒩, h_O(s) = m_O(s)a(s) + c_O(s)b(s) + g_O, h_O(s) = ∑_i ∈𝒩^ G_i(s) (h_M_i(s) + h_I_i(s)), ∑_i ∈𝒩^ G_i(s) h_I_i(s)= 0, op:FC:convex h_M_i = B_C_i f_M_i, i ∈𝒩, h_I_i = B_C_i f_I_i i ∈𝒩, (f_M_i(s) + f_I_i(s)) ∈FC_C_i, i ∈𝒩, f_I_i(s) ∈int(FC_C_i), i ∈𝒩, b(0)=ṡ_0^2, b(1)=ṡ_T^2, 0 ≤b(s) ≤b(s), b'(s)=2a(s), τ_i(s) ≤τ_i(s) ≤τ_i(s), i ∈𝒩, ∀s ∈[0,1], with variables a, b, h_ O and τ_i, h_ M_i, h_ I_i, f_ M_i, f_ I_i, i ∈𝒩. This optimization problem is convex since the objective function and all the constraints are either linear or convex in the optimization variables. Again, this problem can be reformulated as an SOCP after employing the direct transcription method. § NUMERICAL SIMULATION In this section, we provide some numerical simulations to illustrate the results. We start by presenting the simulation results for the case of rigid contacts. §.§ Rigid Contacts Consider the problem setup described in Section <ref> with two 6 DOF Stanford manipulators (see <cit.>,  2.25). Suppose the manipulators are located at [0 -1.4 0]^T and [0 1.4 0]^T in {I} and are placed in such a way that the rotation matrices of the base frames {B_i}, i=1,2, with respect to {I} are R_ B_1 = I_3 and R_ B_2= diag(-1,-1,1), where diag is the diagonal matrix with the vector elements as diagonal elements. Also, suppose that the manipulators are rigidly grasping a cuboid object of dimensions 0.07 × 0.07 × 0.4 m^3 and with mass 10 kg. We assume that the centre of mass of the object is located at its centroid. The dynamic parameters used for the simulation are given in Tables <ref>-<ref>, where m_ij and l_ij denote the mass and the length of the jth link of the ith manipulator, respectively. We assume that all links except the third one are cylindrical with radius of 0.12, 0.1, 0.06, 0.05, 0.05 (all in meter), respectively, and compute the inertia tensors accordingly. The third link is assumed to be a cuboid with equal length and width of size 0.12 m. The desired path for the centre of mass of the object is given by p_ O(s) = [ -0.3+1.2 s; sin(s-0.4); 0.4+0.6 s ]. For the orientation of the object, we have used ZYZ angles ϕ_ O(s) =[-0.5 s+0.2 0.5 sin(s) -0.4]^T. We assume that gravity acts along the negative z-axis. For the simulation, the path coordinate s is discretized on 80 grid points. The optimization problem is implemented and solved in SOCP form in MATLAB using the optimization modeling toolbox YALMIP <cit.>, and the solver MOSEK <cit.>. The simulation is performed on a laptop with a 2.6 GHz Intel Core i7-5600U processor. In the approach presented in this paper we need to compute the partial derivatives q'_i(s) and q”_i(s) at the descretization points. Any method that can provide these partial derivatives can be used for this purpose. For a manipulator with 6 DOF, computing these derivatives analytically can be computationally expensive. In this simulation, we are using a numerical approach to compute q'_i(s) and q”_i(s) at the discretization points. The approach presented in this paper results in a minimal traversal time of 0.422 sec for this simulation.  <ref> shows the resulting joint torques or forces and the corresponding bounds for both manipulators as functions of the path coordinate s. The bounds are depicted in the same color as the joint torque or force, but with a dashed line. Once the optimization variables a and b are obtained, Equations (<ref>) and (<ref>) can be used to compute the joint velocities and accelerations, respectively.  <ref> shows the evolution of the joint velocities with respect to the path coordinate s for both manipulators, where the initial and the final velocities have been set to zero. The considered bounds on the joint velocities can be seen in the same figure and are depicted with dashed lines in the same color as each joint velocity. From the plots it can be seen that for every point along the path, either one of the joint torques/forces of the first or second manipulator is in saturation, or one of the joint velocities has reached its bound. This is in line with the findings in the work by <cit.>, where it is shown that the minimum-time control for multiple manipulators requires that at least one of the actuators is always saturated on any finite time subinterval, while the rest of them adjust their torques so that other constraints on the motion are not violated. We have also performed the simulation for different grid sizes to get an idea of computation time. Solver time and YALMIP time are reported in Table <ref> for grid sizes of K = 30, 100, 300, 1000. It can be seen that the increase in YALMIP time is greater than the increase in solver time. Since all possible combinations of wrenches are considered in our approach, a smaller minimal traversal time from our approach is expected compared to a case where a particular wrench distribution is used. For the purpose of comparison, we have carried out some simulations with the same setup described above but different path and orientation for the object and with wrench distributions borrowed from other works. The resulting minimal traversal times using the wrench distribution from <cit.>, <cit.> and <cit.>, denoted by Ref. 1, Ref. 2 and Ref. 3, respectively, together with the results for the rigid and the frictional (see next subsection) contacts are reported in Table <ref>. In the simulations, the path coordinate s is discretized on 50 grid points. The paths and the orientations that are used for this comparison can be seen in Tables <ref>-<ref>, where p_ O^x(s), p_ O^y(s) and p_ O^z(s) are the x, y and z components of the position of the object, respectively, and [φ_ O(s), ϑ_ O(s), ψ_ O(s)]^T is the object's orientation in terms of ZYZ Euler angles. The considered bounds on joint torques and velocities are same as the ones in  <ref> and  <ref>. In the tables, P.i refers to the ith path. As expected, the obtained minimal traversal time using our approach in the case of rigid contacts is significantly smaller than other values where a particular wrench distributions is used. This is because our approach exploits the full actuation available to the system. §.§ Contacts with Friction Here, we present the simulation results for the problem considered in Section <ref>. We consider the same simulation setup described in the previous subsection. The considered grasp can be seen in  <ref>. We assume that the contacts are soft-finger contacts and that the z-axes of the contact coordinate frames point in the direction of the inward surface normals. The line connecting the two contact points lies inside the friction cones, and hence according to <cit.>, the grasp is force-closure. We have considered the coefficients of friction to be μ = 1 and λ = 1, where μ = 1 corresponds to a friction cone angle of 45°, i.e., the angle of the cone with respect to the surface normal. The interior of the friction cone for the ith contact has been considered as int (FC_ C_i) = { f ∈ℝ^4 : √(f_x^2+f_y^2)≤μ f_z - δ_1, f_z ≥ 0, | t_z | ≤γ f_z -δ_2 }, where δ_1 and δ_2 are small positive numbers and f= [ f_x , f_y , f_z , t_z ]^T. For this simulation, we have chosen δ_1 = δ_2 = 0.5.  <ref> and <ref> show the resulting joint torques/forces and joint velocities as functions of the path coordinate s, together with the considered bounds on them for both manipulators, respectively. The bounds are depicted in dashed lines. Similar to the results from the case with rigid contacts, it can be seen that for every point along the path, either one of the joint torques or forces of the first or second manipulator is in saturation, or one of the joint velocities has reached its bound. The simulation resulted in a minimal traversal time of 0.567 sec, which is greater that the minimal traversal time of the rigid contacts case. This is because of the additional constraints associated with friction cones that the contact forces must respect. Comparison between the minimal traversal times of rigid and frictional contact cases for other paths and orientations of the object can be found in Table <ref>. It can be seen that in all cases, the result obtained for rigid contacts is smaller than the one for frictional contacts. As mentioned earlier, in order to be able to firmly grasp an object, it is desirable that internal forces are present during the motion. For the spatial grasp of  <ref>, internal forces are in the form of f_ I_1 = [ 0 , 0 , α , β ]^T and f_ I_2 = [ 0 , 0 , α , β ]^T, i.e., the force or the torque components of contact wrenches in the normal direction with equal magnitudes. For these wrenches to be in the interior of the friction cone, α must be positive, meaning that the forces in the normal direction are applied in the positive direction of z-axis at each contact point. In order to investigate the presence of internal forces in the obtained solution from the simulation, the components of the wrenches f_ C_1 and f_ C_2 in the normal direction are shown in  <ref>. The force components for both contacts are shown in the upper plot and the torque components can be seen in the lower plot. Internal forces are the force or the torque components in the normal directions with equal magnitudes. That is, the least of the two values of normal components for the forces. This is shown by the black dashed line in the upper plot in  <ref>. Note that the obtained solution for the normal force components must be positive at every point along the path, since the wrenches must be in the interior of the friction cones. This is the case for the obtained solution and can be seen from the upper plot in  <ref>. For the torque components, if at a point along the path both torques have the same sign, then the least of the magnitudes of the two torques, will be part of the internal force. This can be seen by the black dashed line in the lower plot. The tangential components of the contact wrenches are plotted in  <ref>. This simulation result shows that the solution of the proposed optimization problem involves internal forces belonging to the interior of the friction cone during the whole motion, which is a desirable property since it ensures a firm grasp. § CONCLUSION We have formulated the time-optimal path tracking problem for a multi-manipulator system carrying an object as a convex optimization problem. Two types of contact were considered: rigid contacts and contacts with friction. Numerical simulations were carried out for both cases. From the simulation results in both cases, it was seen that for every point along the path, either one of the joint torques or forces of the first or second manipulator was in saturation, or one of the joint velocities had reached its bound. This is in line with the results that can be found in the literature about the structure of the minimum-time control for multiple cooperating manipulators. The resulting minimal traversal time from our approach was compared to the simulation results from using a particular wrench distribution for different paths, where the wrench distributions were chosen from other studies. For each path, our approach resulted in a smaller minimal traversal time, since it considers all possible combinations of wrenches and exploits the full actuation available to the system. The case with contacts with friction resulted in a greater minimal traversal time compared to the case with rigid contacts. This is because of the additional constraints associated with friction cones that the contact wrenches must respect in the case of contacts with friction. Furthermore, in the same case it was seen that the solution of the proposed optimization problem involves internal forces that lie in the interior of the friction cone during the whole motion, which is a desirable property for the purpose of having a firm grasp. In future work we plan to study the extension to mobile manipulators, where manipulators are mounted on a mobile platform. Extension to consider rolling contacts where grasping involves moving rather than fixed contact points is another possibility. § DECLARATION OF COMPETING INTEREST The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. elsarticle-harv

http://arxiv.org/abs/2303.07370v1

20230313180004

Exotic energy injection in the early universe II: CMB spectral distortions and constraints on light dark matter

astro-ph.CO

L[1]> m#1 C[1]> m#1 R[1]> m#1 e-.05em V e-.05em V pc mystyle language = Python, commentstyle=, keywordstyle=, keywordstyle=[1], keywordstyle=[2], keywordstyle=[3], keywordstyle=[4], commentstyle=, numberstyle=, stringstyle=, basicstyle=, breakatwhitespace=false, breaklines=true, captionpos=b, keepspaces=false, numbers=none, numbersep=5pt, showspaces=false, showstringspaces=false, showtabs=false, tabsize=2, morekeywords=True, False, len, columns=flexible literate= *001 111 221 331 441 551 661 771 881 991 e-e-2 e0e02 e1e12 e2e22 e3e32 e4e42 e5e52 e6e62 e7e72 e8e82 e9e92 e-e-2 -0-02 -1-12 -2-22 -3-32 -4-42 -5-52 -6-62 -7-72 -8-82 -9-92 style=mystyle

http://arxiv.org/abs/2303.06944v1

20230313092248

A Human Subject Study of Named Entity Recognition (NER) in Conversational Music Recommendation Queries

cs.CL

Context-Aware Selective Label Smoothing for Calibrating Sequence Recognition Model Shuangping Huang^1,2, Yu Luo^1, Zhenzhou Zhuang^1, Jin-Gang Yu^1,2*, Mengchao He^3, Yongpan Wang^4 =========================================================================================================== We conducted a human subject study of named entity recognition on a noisy corpus of conversational music recommendation queries, with many irregular and novel named entities. We evaluated the human NER linguistic behaviour in these challenging conditions and compared it with the most common NER systems nowadays, fine-tuned transformers. Our goal was to learn about the task to guide the design of better evaluation methods and NER algorithms. The results showed that NER in our context was quite hard for both human and algorithms under a strict evaluation schema; humans had higher precision, while the model higher recall because of entity exposure especially during pre-training; and entity types had different error patterns (e.g. frequent typing errors for artists). The released corpus goes beyond predefined frames of interaction and can support future work in conversational music recommendation. § INTRODUCTION Music recommendation systems (RSs), fundamental to streaming services nowadays, learn from user listening history or music content which artists or tracks to suggest next <cit.>. Most of these algorithms provide personalized music content to the users when logging in the streaming apps or websites, or when triggered with pre-defined utterances via voice assistants <cit.>. More recent conversational RSs aim to help users to express their recommendation needs by supporting interactions via queries in natural language <cit.>. However, despite existing in the scientific literature, such conversational RSs are not widely deployed because of multiple issues, one being NER. The processing of recommendation queries entails the extraction of named entity mentions <cit.>. This sub-task faces multiple challenges, even when queries are framed as pre-defined utterances. The transcriptions of the voice queries results in lower-case noisy text, often with misspellings <cit.>. The lack of capitalisation in entities and misspelled words are often present in text-based queries too <cit.>. Music entities, or those coming from the creative content domains, are highly irregular: they do not follow inherent patterns as it is the case with people's names, and there is little to no separation between the vocabularies of entity and context words, especially for creative works <cit.> (e.g. common words like "I" or "love" in track titles). Also, new music entities appear all the time. Major music streaming services ingest one new track almost every second <cit.>. Previous works have already shown that NER systems struggle with the aforementioned challenges <cit.>. Thus, multiple approaches have been proposed to address them, either focused 1) on collecting more and relevant data for training / fine-tuning standard NER sequential models <cit.>; or 2) on model's design choices that favour generalisation <cit.>. Most solutions focused on the latter objective have been motivated by the human NER linguistic behaviour, e.g. make the model rely more on context cues than on named entity mentions or learn from a few examples only, as humans do. However, apart from some scarce, partially related works <cit.>, there is no systematic investigation of how humans actually perform NER on noisy text with many new and irregular named entities. Moreover, in the case of music recommendation, we are not aware of any existing dataset of queries in natural language, annotated with named entities. Thus, our goal is to investigate the human NER linguistic behavior when confronted with these challenging conditions. For that, we create MusicRecoNER, a new corpus of noisy natural language queries for music recommendation in English that simulates human-music assistant interactions. We then conduct a human subject research study to establish a human baseline and learn from it. Finally, we perform a detailed comparison of humans and the most popular NER systems nowadays, fine-tuned transformers, that covers multiple evaluation schemes (strict named entity segmentation and typing, exact segmentation only, or partial segmentation with strict named entity typing) and scenarios including entities previously seen or unseen by the model or humans. The results showed that the task was challenging for humans. Given an aggregated metric such as F1 score, human and algorithmic performances were on par. However, the detailed evaluation revealed that humans struggled more with recall while the best model with precision. The high recall obtained by the model was partially a result of entity exposure during pre-training or fine-tuning. Also, music entities had different error patterns and, in some queries, had ambiguous context that made their segmentation and typing quite hard. To sum up, our research contribution[Code and data are available at <https://github.com/deezer/music-ner-eacl2023>.] are: * MusicRecoNER, a corpus of noisy complex natural language queries for music recommendation collected from human-human conversations in English, but which simulates human-music assistant interactions, annotated with Artist and WoA (work of art) entities. This dataset is not limited to pre-defined utterances as it would be the case if collected from interactions with conversational or voice assistants. Thus, it contains entities in diverse context, being also a useful resource for future work on conversational music recommendation. * A human subject study design for NER in noisy text with many new and irregular named entities. The proposed method is transferable to other creative content domains that face similar challenges to music such as books, movies, videos, but also to any other domain with scarce data, which wants to learn more about the NER task before building a system. * An extensive music NER benchmark on noisy text which compares the performance of human versus automatic baselines under multiple evaluation schemes, scenarios and by controlling for the novelty of named entities. § RELATED WORK Analysing human and algorithmic performance was done for multiple NLP tasks in the past. <cit.> ran an annotation campaign on the GLUE benchmark with the goal to estimate the effort needed by existing models to catch up with the humans under limited-data regimes. <cit.> conducted a large-scale human study on topic shift identification in order to discover patterns of disagreements and consolidate the evaluation metrics. <cit.> analysed the human behaviour for understanding ambiguous text-based or spoken sentences to guide the development of a machine learning system. Multiple machine translation works challenged the human parity claim <cit.> and proposed a secondary evaluation method to reveal detailed differences between humans or algorithms <cit.>. Compared to these, we benchmark humans and models on a different task—named entity recognition, but we share similar goals—to estimate the human-algorithmic performance gap and to identify patterns that could support the design of better evaluation methods or automatic solutions. Human annotation is frequent in NER especially when targeting a new domain such as archaeology <cit.>, or a new language such as Indonesian <cit.>. However, we are not aware of any annotated corpus of conversational queries for recommendation in the music domain. Some other related works propose corpora of noisy social media text containing new entities including irregular ones <cit.>, a noisy dataset of movie-related queries <cit.>, a dataset of music artist biographies annotated for entity linking <cit.>, or a corpus of tweets associated with a classical music radio channel <cit.>. Previous works have showed that transformers fine-tuned for NER are strong baselines, especially when training data is scarce <cit.>. A more recent line of research employs these pre-trained models as few-short learners <cit.>. However, the results are still below those obtained with a fine-tuning approach. In order to improve the bare-bone fine-tuned transformers, other works adopted distant supervision <cit.>, and the inclusion of gazetteers <cit.> or contextual triggers <cit.>. Though these solutions are interesting and relevant to our problem and context, in the current research, we want to rely on the results of this study before making any design choices for an advanced NER system in the music domain. When conducting human subject studies, the quality of annotations (inter-rater agreement or reliability) is often assessed with Kappa statistic or its variations <cit.>. Yet, for NER, or more generally for labelling phrases, this statistic is less applicable as the number of negative cases on which it relies is ill-defined <cit.>. To address this issue, multiple imperfect solutions have been proposed such as to compute the Kappa statistic at the token level <cit.>—however, this does not reflect the task well as each token is not tagged individually; or to estimate the negative cases by enumerating all n-grams or noun phrases from a text—however, this lacks accuracy <cit.>. <cit.> show that when the number of negative cases gets very large, the Kappa statistic approaches the F1 score. Thus, F1 is considered a better metric, which we also adopt to measure the performance of humans and compare the NER human and algorithmic baselines. § HUMAN SUBJECT NER STUDY §.§ Data Collection For data collection we have chosen the music suggestions subreddit[<www.reddit.com/r/musicsuggestions/>] as a relevant data source. Reddit is a discussion website where members can submit questions, share content and interact with other members. It is organised in subreddits built around dedicated topics. Each discussion starts with an initial post that has a title and description. From this post, threads of conversations develop. We were interested only in posts triggered by a music information seeking or recommendation need. We crawled the full subreddit with 8615 initial posts. This number corresponds to the posts in the beginning of 2020. We did not consider posts' comments. These humans-to-humans posts asking for music recommendations are particularly relevant to study as they go beyond pre-defined frames of interaction with a text or voice-based assistant. Hence, they exhibit a realistic human use of language, which although more challenging, could help with the development of the next generation of music assistants. For NER, the existence of queries in natural language translates in a more diverse context surrounding named entities, thus in a higher query generalisation for music recommendation. By manually checking this data, we noticed that many mentioned artists or music titles were not popular. Thus, we expected most named entities to be new to the annotators, an aspect we wanted to control for, as mentioned in Section <ref>. §.§ Data Cleaning and Pre-processing As we aimed at creating a corpus of music recommendation queries simulating human-assistant interactions, we made multiple decisions to pre-process the collected posts. We performed a manual cleaning of this data by removing those posts which directly shared music with the community; were aimed at promoting music or other music-related entities; contained explicit words; or contained only links to external music resources. Then, we focused on titles only as the post content was rather long, specific to asynchronous communication; as human-assistant interactions happen synchronously, the written or spoken queries are expected to be short, composed of a few short sentences at most <cit.>. We removed all references to specific music-related services in order to obtain generic queries (e.g. we removed "Youtube" from the request "music similar to my Youtube playlist"). We also removed words which were explicit markers of human-human interaction in order to ensure compatibility with human-assistant interaction. For instance, we removed phrases such as "hello guys" or "could anybody". We performed the rest of the pre-processing steps to ensure that the queries contained, to some extent, the kind of noise that could be found in transcribed voice queries too, such as those obtained when interacting with a voice assistant. For this, we transformed the text in lowercase and removed punctuation marks and emoticons (with some exceptions when the symbol was part of the named entity's pronunciation such as "&"). We kept content from parentheses when found at the end of a post title, otherwise we removed it. Although very common in automatic transcriptions, we did not introduce any artificial noise regarding the spelling of named entities. Still some noise was present as Reddit authors sometimes made misspelling errors. These steps were done automatically. We release both the original and pre-processed data. All keywords used in the described steps are in Appendix <ref>. We show multiple query examples in Table <ref>. §.§ Annotation Guidelines and Procedure We sampled multiple subsets of 600 queries each from the cleaned and pre-processed corpus. This number was established by estimating the required time for the experiment to be maximum 2 hours per annotator, based on an initial trial on 751 queries. The annotation guidelines were also tested in the trial experiment and refined after. The subjects were informed that the goal was to identify names of artists (e.g. bands, singers, composers) and titles of works of art (e.g. albums, tracks, playlists, soundtracks) in unformatted music-related queries. We requested the annotators not to consult the Internet as we wanted them to rely on the query content only and on their own previous knowledge. We then introduced the labels: Artist_known, Artist_deduced, WoA_known, WoA_deduced, and Artist_or_WoA_deduced with examples. The last one was for ambiguous cases of named entity typing, but allowed the annotators to segment. Segmentation is still very relevant when parsing natural language queries for music recommendation as the type could be eventually disambiguated with the help of a search engine, for instance. The other labels corresponded to Artist and WoA types, completed by whether the annotator knew the entity from before or deduced it from query's content, as we wanted to keep track of entity's novelty. Then, we introduced challenging annotation cases with guidelines on how to proceed. We instructed the annotators to include Artist and WoA named entities from other domains too such as movies or video games, but to ignore all the other entity types such as countries or music genres; to consider the innermost entities in case of nested entities; to ignore implicit entities such as "this singer"; to always include the "'s" from the possessive case as part of the named entity; and to consider a named entity with misspelled, translated and transliterated words as correct. The final form of the guidelines is shown in Appendix <ref>. Ten annotators (1 for the trial, and 9 for the main study) were recruited from our organisation with the condition to be fluent in English. Each set of 600 queries (DS1, DS2, and DS3) was given to three annotators. The annotation campaign was performed using Doccano <cit.>. The guidelines and the annotation tool were presented in a 30-minute workshop where annotators could ask questions. They could consult the guidelines and contact the researchers if they needed any clarification during the experiment too. After, one week was set aside for each annotator to complete the annotations individually. §.§ Ground-truth MusicRecoNER Corpus Often in related works, a ground-truth corpus is obtained by using full agreement or majority voting <cit.> (e.g. tag named entities on which at least two out of three human annotators agreed). However, here, because we wanted to establish a human baseline and have a corpus exhaustively annotated, we labelled the ground-truth corpus ourselves from scratch. Compared to the settings of the human subject study, we had access to the original Reddit post titles including capitalised text and punctuation. During the annotation, we used web and music streaming search engines to check if certain entities were Artist or WoA. The full Reddit post was also used to disambiguate cases when a name could be both an Artist or a WoA. The most challenging examples were discussed among us. The ground-truth preparation together with the adjudication discussions happened over several weeks, as the process to disambiguate entities was more complex. Statistics about each dataset are presented in Part I of Table <ref>. Artist mentions are more common than WoA mentions. Regardless of the type, we could notice that a large majority of entity mentions are unique in each dataset. The mean number of entity mentions per query is around 2, with the maximum varying between 6 and 10. From these, the proportion of queries with no entity is on average 56%. § EVALUATION PROTOCOL §.§ Fine-tuned Transformer Baselines The goals of the human subject NER study are to establish a human baseline on this challenging dataset of noisy queries for music recommendation and to learn from the human linguistic behavior in comparison to the most common NER systems nowadays, the fine-tuned transformers. We consider three language models proven to have good results in various natural language tasks including language understanding, sequence labeling or text classification: BERT <cit.>, RoBERTa <cit.> and MPNet <cit.>. BERT <cit.> is a multi-layer bidirectional encoder based on the original Transformer architecture <cit.>. It is pre-trained on: 1) the cloze task, i.e. to predict a masked token from the left and right context; and 2) next sentence prediction, i.e. to predict the next sentence from a given one. RoBERTa <cit.> has the same architecture as BERT, but incorporates multiple training steps proven to lead to an increased performance than the original model: the training of the model using more data, with larger batches, on longer sequences and for a longer time; and keeping only the cloze task as a pre-training objective while applying a dynamic masking schema to the input training data. MPNet <cit.> proposes a new pre-training objective by integrating the masked language modeling objective of BERT and the permuted language modeling objective introduced in XLNet <cit.>. That is, it models the dependency among the masked tokens at prediction (i.e. takes into account the already predicted masked tokens to generate the current one), while providing visibility on the position information of the full sentence (i.e. the positions of the masked token and the next ones to be predicted). We fine-tune the pre-trained versions of these models released in the huggingface transformers library <cit.> for token classification / sequence labeling. We took the largest available version for each of them: bert-large-uncased, roberta-large, and mpnet-base. From all, only BERT is pre-trained on uncased text. During experiments, we noticed that the model initialisation had a large impact on the results. This instability is well-documented in the past work, especially when the corpus for fine-tuning was small <cit.>. Thus, to overcome bad initialisation and have more coherent results over different runs, we re-initialized the last layer of each pre-trained model. This also led to faster convergence and more efficient fine-tuning. We also tried to increase the number of the re-initialized layers to 2, but the results were similar or sometimes worse. §.§ Evaluation Metrics and Schemes Precision (P), recall (R) and F1 are commonly used to evaluate automatic NER systems <cit.>. In our evaluation, we extend these metrics to support a more detailed benchmark and understanding of the kind of errors a NER system makes. Namely, we also allow for a relaxed system's evaluation, when either segmentation or typing is correct, but not necessarily both. A NER system can produce various types of outcomes <cit.>. Inspired by this and <cit.>, all NER outcomes, which we denote O, can be: * Correct outcomes (O_c): predicted and ground-truth entities match. * Missing outcomes (O_m): system entirely fails to spot a ground-truth entity. * Spurious outcomes (O_s): false entities are produced by the system. * Incorrect outcomes (O_i): predicted and ground-truth entities do not match because of either typing or segmentation errors. To classify the predictions of a NER system in these categories, we first need to fix an evaluation schema. The most common one in the literature is the Strict match <cit.> when both segmentation and typing are correct. Under the Strict schema, a prediction is incorrect when its boundaries were correct but not its type, or when its type was correct but not its boundaries. All other cases (e.g. partial segmentation with incorrect type) are classified as spurious. The Exact schema classifies a prediction as correct when its boundaries match those of the ground-truth, regardless of its type. In contrast, the Entity schema classifies a prediction as correct when its type matches that of the ground-truth, regardless of its boundaries. For these latter schemes, incorrect is adapted from its definition in Strict; missed and spurious are the same too. We use another class of outcomes, partial (O_p), only when computing the human performance. As described in Section <ref>, humans could annotate a text as Artist_or_WoA_deduced. Thus, whenever a human prediction had this label and matched exactly the boundaries of the ground-truth entity, partial was incremented and contributed to the final scores with a factor of 0.5 <cit.>, as follows: R = (|O_c| + 0.5 * |O_p|) / (|O| - |O_s|) P = (|O_c| + 0.5 * |O_p|) / (|O| - |O_m|) We exemplify the different outcomes under the mentioned schemes in Table <ref>. One practical detail regarding the calculation of the evaluation metrics is that we had to apply some segmentation corrections before, to cover the situations when human annotations started or finished in the middle of a word. This could appear because Doccano did not force automatically an alignment to a desired tokenization (entire words). Thus, we corrected the start or end index of the concerned span by moving them to the left or right, based on a simple heuristic with regard to the closest found separating character (space or newline) to the concerned word. We did not intervene when an entity was composed of multiple words and only a part of them were annotated, but we captured this type of errors with the used evaluation schemes. No correction was needed in the case of model annotations as, during fine-tuning, we propagated the label of the first word token to the rest; hence, the labels were always consistent for all word tokens. §.§ Evaluation Scenarios We explicitly consider the novelty of entities. In the case of humans, this was encoded in the annotation process as we introduced the labels suffixed with _known. Fine-tuned models could have seen music entities from the test set during pre-training, when they were exposed to a large amount of unlabelled data or during fine-tuning, if the train and test sets had common entities. While this latter exposure could be easily checked, the pre-training exposure is more challenging to assess as it requires access to the pre-training data or to find other ways to test exposure based on the model only <cit.>. The solution we adopted targeted BERT, which performed on par with the other models as revealed in Section <ref>. BERT is pre-trained on Wikipedia and BookCorpus <cit.>. Thus, music entities could be found more likely in the Wikipedia content. However, some music entities could be quite rarely mentioned in Wikipedia compared to others. To quantify BERT's exposure to an entity e we used the following method. First, we tried to link each entity to Wikipedia by querying the Wikidata knowledge base <cit.>. We re-ranked the returned results to give priority to music entities and returned the first entity whose type was in a pre-defined type list (see Appendix <ref>). Second, we computed exposure by adapting the metric proposed by <cit.>: expo(e) = {[ log|𝒮|-logrank(e) e ∈ Wiki.; 0 e ∉ Wiki.; ]. where 𝒮 represents all Wikipedia named entities and the function rank considers entity popularity (higher the popularity, lower the rank). We retrieve 𝒮 and entity counts from Wikipedia2Vec <cit.>. We manually checked the linking for 300 random entities. 82% were correct, either linked or not found on Wikipedia correctly. 14% were linked to music-related entities but not the right ones and the rest were errors or missed entities. Examples of entities with high exposure values are: the beatles, elvis, pink floyd, metallica, drake, johnny cash, eminem, nirvana, and coldplay. We could notice that all are of type Artist. § RESULTS AND DISCUSSION We report scores using 4-fold cross-validation on the datasets presented in Table <ref>. Means and standard deviations (std.) are computed over different folds, different initialisation seeds for the model, and different human annotators. In most cases, this was over 12 data points as, for each model, the results were aggregated over each dataset as a test and 3 different initialisation seeds[All the models were trained and tested on the ground-truth datasets, and did not consider annotator-specific sets.] and for the human evaluation, over each dataset as a test and 3 human predictors per dataset. When comparing BERT and the other models in Table <ref>, BERT and human baselines in Tables <ref> and <ref>, and results on Seen versus Unseen entities obtained either by humans or BERT in Table <ref>, scores in bold are statistically larger (p-value=0.05). We test statistical significance with the Mann-Whitney U Test (Wilcoxon Rank Sum Test, ), which assesses under the null hypothesis that two randomly selected observations X and Y come from the same distribution. §.§ Fine-tuned Transformer Baselines Table <ref> shows that the fine-tuned BERT, pre-trained on uncased text, and MPNet yield the largest F1 scores for each entity type or overall. RoBERTa is statistically comparable and only marginally lower than the other models. Although MPNet and RoBERTa share the same pre-training corpus and the Transformer architecture, the addition of the permuting language objective to the cloze task gives a slight advantage to MPNet. We use BERT for the rest of the experiments. §.§ Humans vs. Fine-tuned BERT Table <ref> shows that the performance of BERT is comparable to that of the human baseline in terms of F1 score. However, Table <ref> shows that humans and BERT perform differently in terms of precision and recall. Humans have a higher precision, for both Artist and WoA, whilst BERT has a marginal or significantly larger recall than humans, especially for Artist. We confirmed that this phenomenon was not due to a particular precision / recall compromise by testing various precision / recall value and optimizing on F1. Also, BERT has a lower precision than the recall, but we see the opposite for humans. Considering Equations 1 and 2, the model appears to hypothesize spurious entities more often, while humans tend to miss entities more often. Table <ref> also shows that the F1 scores under Exact and Entity schemes are larger than under Strict as some of the errors produced are because of segmentation or typing. However, we can notice a different behaviour for the two entity types for both BERT and humans. In the case of WoA, the Entity F1 scores are slightly larger than those obtained under the Exact schema, showing that boundary errors happen more frequently. On the contrary, for Artist entities, the segmentation is more often correct, but the typing is wrong. §.§ Error Analysis Figure <ref>, showing a detailed error analysis, confirms that indeed BERT has more often spurious outcomes than humans, for both entity types. Also, humans miss to annotate ground-truth entities more often than BERT. We can equally observe that BERT is highly superior in identifying correct named entities. Previous works on NER <cit.> have discussed that a system should learn to exploit the context (i.e. the non-entity words) rather than entity memorisation to generalise. However, the high number of correctly recognised entities as well as the frequent spurious entities suggest that this may not be the case here; and BERT's behaviour may be linked to entity exposure. As shown at the end of Section <ref>, the entities with the highest exposure score were of type Artist. We could see in Figure <ref> that there are a lot more correct Artist entities, and the number of missed and spurious outcomes for Artist is lower, which seems to be aligned with our hypothesis related to entity exposure. §.§ Impact of Entity Exposure In Table <ref>, Part II, we show the percentage of entities known by at least one annotator among the three in each dataset. This varies between 24% and 30%. In practice, each annotator has known at most this number of entities, which confirms that most entities from the collected corpus were new to our subjects. The entity exposure is much larger for BERT. While the train and test sets share only maximum 15% of the entities, BERT has seen up to a half of corpus' entities during pre-training. To check the model's performance on seen versus unseen entities, we show Recall scores for these groups in Table <ref>. Seen entities are those present in the train set or with expo(e)>1. Unseen entities have expo(e)=0 and are not known to humans. The rest of the entities are discarded from the evaluation. BERT's recall on Seen is much larger than on Unseen, which confirms our hypothesis that memorisation plays a role. However, the model seems to rely significantly on context too given that the results on Unseen are still quite high. We also report the results of humans in Table <ref> and see a similar pattern. Although the split is made considering the model's exposure, humans are also very likely to know entities from Seen. The lower humans' scores on Unseen show that the recognition of these entities is quite challenging, possibly because of insufficient context. For example, "songs bands similar to sales getting it on off and on and porches mood" contains an enumeration that is difficult to segment and type (Artist: "sales", "porches"; WoA: "getting it on", "off and on", "mood"). Also, entity typing is ambiguous in "anything similar to some people say" (WoA). For these imperfectly recognised entities, including external resources such as gazetteers or search engines might be an option to explore. § CONCLUSION In this work, we investigated the human linguistic behavior when performing NER in the music domain. We created MusicRecoNER, a new corpus of complex noisy queries for recommendations annotated with Artist and WoA entities. We then designed and conducted a human subject research study to establish a human baseline and learn from its comparison with the most popular systems nowadays, fine-tuned transformers. We performed a thorough evaluation covering multiple metrics, schemes and scenarios, including a careful analysis of the impact of entity exposure on results. The results obtained by the algorithmic baselines were comparable to the human ones. Yet, the detailed evaluation showed that humans yielded a better precision while the model had a better recall, linked also to entity exposure during pre-training and fine-tuning. Thus, when evaluating fine-tuned pre-trained models, checking their performance on new entities shows their real generalisation ability. Regarding the NER evaluation protocol, human performances were much better under a more relaxed schema focused on segmentation or typing only. Such a schema could prove a more realistic setup to aim to when training models too. Also, we noticed that the relevant schema depended on the entity type as Artist was better segmented, while WoA better typed. Contrary to previous claims, we show that, in our domain, NER in challenging conditions such as noisy text, and irregular or novel entities is rather hard for humans even when provided with complex instructions and multiple examples. Thus, although we could learn from the human linguistic behaviour, we should not, by default, assume their results to be a target for any NLP problem. For some tasks, it is common when establishing a human baseline to consider it as an upper bound for the model. This is not necessarily a desirable outcome in our case as it would imply mislabelling 1/3 WoA entities. More generally, as we also showed by studying the impact of entity exposure, algorithms can store a lot more knowledge than humans and one may want to leverage this as much as possible. As for proposing a better system to perform music NER, one next step would be to continue the model's pre-training on more related data, in our case music, to get even more exposure, or to integrate gazetteers. Still, given the rate of new entities in our domain, forcing the model to rely more on context, when context is not confusing, is another desirable future direction. In case of context ambiguity, asking questions to clarify the request and supporting user interaction in natural language could be ultimately the answer towards a more suitable, but still very challenging solution. We plan to explore these ideas as future work. § LIMITATIONS We further discuss the limitations of our work. The corpus of noisy complex queries in natural language we use in the human subject study and we release is built based on a single source, Reddit. The demographics of the users using Reddit are relatively narrow, with a majority being male, young, and educated[<https://foundationinc.co/lab/reddit-statistics/>]. Moreover, users seeking music recommendations on this type of forums may be rather "music enthusiasts" and may not represent regular music listeners. The implications are that the language employed in these queries could be specific to this category of population. Also, the mentioned entities could reflect the music taste of this type of profiles only. This latter implication turned to be an advantage for us as we ended up with many novel entities, unknown by the annotators who participated in the study. As for the first implication, we manually checked the queries, and found them quite diverse, not necessarily using a specific vocabulary but more general language expressions. An alternative to creating such a corpus could have been a Wizard of Oz experimental setup <cit.>. However, this would require significantly more costs and would highly depend on the type of profiles interested in participating in such a music discovery experiment. Second, we pre-processed the corpus in order to simulate written or transcribed speech-based human-computer interactions. However, the steps we took may be largely insufficient to simulate the kind of noise found in transcriptions. As we also discussed in Section <ref>, we did not inject any artificial noise for named entities, while spelling errors when automatically transcribing them are a common problem. Another limitation regarding named entities is the computation of the model's exposure by leveraging Wikipedia. Our linking was quite rudimentary and imperfect, as we reported in Section <ref>. Moreover, for retrieving entity ranks, we used Wikipedia2Vec <cit.>, which is built on a slightly older Wikipedia version than the one BERT was trained on. Therefore, the results obtained by the model on the Unseen dataset may be slightly larger, as the model could have been exposed to some of these entities. However, our goal was to prove a phenomenon—the impact of named entity exposure on the results, even if this impact may be marginally underestimated. Finally, the annotators recruited from our organisation have similar age and demographics. Also they likely have a richer musical background compared to regular human subjects. This signifies that, in reality, the number of novel entities could be higher, which could also impact the overall results obtained with the human baseline. Nevertheless, this hypothesis could be tested only by running subsequent studies including more subjects. § ETHICAL CONSIDERATIONS We have provided most of the details about data collection, data cleaning and pre-processing, and the annotation procedure and guidelines in Section <ref> and Appendices. We discuss further various ethics-related aspects not covered yet in the paper. The dataset was gathered from the music suggestion subreddit via the Reddit API. According to the privacy policy of Reddit[<https://www.reddit.com/policies/privacy-policy>], third parties can freely access public content via the API. We have not gathered any other information besides the public posts—their titles and descriptions. As previously mentioned, the annotators were recruited from our organisation. They performed the annotation tasks during their regular paid hours. Moreover, the participation was fully on voluntarily basis, following a public call for participation by the authors within the organisation. acl_natbib § DATA FILTERING KEYWORDS The list of keywords used in the data cleaning and pre-processing steps are presented in Table <ref>. These keywords were used to filter out posts, which were manually verified after. The outcome of the verification was either to exclude these posts from the data, or to keep the posts as they were or after having removed specific words (as described in Section <ref>). We have considered both lower and upper case variations of each keyword. § ANNOTATION GUIDELINES §.§ Introduction The goal of this annotation experiment is to identify names of artists (e.g. bands, singers, composers) and names of works of art (e.g. albums, tracks, playlists, soundtracks, movies, video games) in music-related requests. The requests could be single- or multi-line. Also, they are unformatted, meaning that they contain no capitalized letters or punctuation marks. Also contractions such as "Artist’s first album", "don’t" are written as if pronounced, specifically "artists first album" and "dont". Through this experiment, we study how well humans can identify named entities (artists and works of art) in unformatted text by relying on the request content only, and on one's own knowledge. For this reason, it is important that during annotation you do not consult the Internet to verify if some parts of text are named entities, but rely on your intuition after reading the text. §.§ Named Entity Categories There are two categories referring to the entity type Artist; two categories referring to the entity type Work of Art (WoA); and one category for dealing with ambiguous cases as follows: 1. Artist_known. This category should be used for sequences of words denoting an artist that is previously known to the annotator. In the next request I recognize "queen" and "the clash" to be Artist entities because I knew them from the past: i like queen and the clash what else should i listen to. Note that when "the" is part of the name (e.g. "the clash"), it should be annotated likewise. 2. Artist_deduced This category should be used for sequences of words denoting an artist that is not known to the annotator, but deduced from the text. In the next request I recognize "stephan forté" to be an Artist: looking for something like the first album of stephan forté. I have never heard of this artist before, but I deduced it from the request’s content. 3. WoA_known This category should be used for sequences of words denoting a work of art that is previously known to the annotator. In the next request I recognize "karma police" to be a WoA because I knew it before: im a very picky music listener but i love karma police any other suggestions. 4. WoA_deduced This category should be used for sequences of words denoting a work of art that is not known to the annotator, but deduced from text. In the next request I recognize "special affair" to be a WoA: songs like special affair. I have never heard of this work of art before, but I deduced it from the request’s content. 5. Artist_or_WoA_deduced This category is used for sequences of words recognised to denote an artist or a work of art, but choosing between the two entity types is challenging. In the next request I recognize "superunloader" to be either an Artist or a WoA: music like superunloader I have never heard of this before and it is difficult for me to distinguish between the two options. §.§ (Challenging) Examples Please read the following examples carefully and re-consult them during the experiment whenever needed. Relevant named entities not related to music. A text could contain other types of works of art such as movies or video games. Annotate these names using the category WoA. Similarly, annotate with the Artist category movie directors, filmmakers, music composers and so on. All the other types of named entities not related to Artist and WoA must be ignored (e.g. names of countries, music genres etc.). In the example below, "gemini" is annotated as WoA and "ang lee" as Artist: i recently watched this film gemini made by ang lee and liked the soundtrack any similar recommendations of this. Multiple named entities clearly delimited. A text could contain multiple entities which are clearly delimited by other words such as "by", "from", "and" etc. Please annotate all of them individually. In the example below, "heartbeat" is a WoA and "annie" is an Artist: songs with similar vibe and structure as heartbeat by annie. In the example below, "hallelujah" is a WoA and "jeff buckley" is Artist: other beautiful songs by jeff buckley apart from hallelujah. Multiple named entities with no delimitation. A text could contain multiple entities which are not clearly delimited. Try to annotate each segment of text individually with its corresponding named entity category. In the example below, if the annotator recognizes the entities, then 3 separate Artist entities should be selected, namely "imagine dragons", "bastille", and "daya": singers bands like imagine dragons bastille and daya. However, if not all entities are known from the past, then the unknown span of text could be annotated either as Artist_or_WoA_deduced, Artist_deduced or WoA_deduced depending on the content and the annotator's intuition. For instance, if the annotator recognizes only "imagine dragons" but not the rest, then "bastille and daya" could be considered either 1 entity ("bastille and daya") or 2 entities ("bastille", "daya") and further annotated with any of the 3 categories mentioned above. Named entities collated with 's from the possessive case. In this case, include the "s" as part of the named entity. In the example below, "toni braxton" is the real name of the artist, but "toni braxtons" (coming from "toni braxton's") is actually annotated as an Artist entity: newer 2005+ ballads in the style of toni braxtons un break my heart and stevie wonders all in love is fair. Similarly for "stevie wonders" (coming from "stevie wonder's"). Nested named entities. A text could contain nested entities. This means that there is a larger text segment that could be considered as an entity and a smaller text segment inside the larger one that could be also considered as an entity. In this case, always favor the innermost text segment with an exception which is described below. Multiple examples are given further. In the example below, "treasure planet" is annotated as WoA and not "treasure planet soundtrack" (which is also a WoA, but the innermost one is chosen): looking for calm violin music very similar to the first 34 seconds of 12 years later from the treasure planet soundtrack. In the example below, "ezra collective" and "ty" are annotated as 2 separate Artist entities and not as one: "ezra collective feat ty": recommend me some good jazz hip hop songs with rap like chapter 7 by ezra collective feat ty. There is also a third entity, "chapter 7", annotated as WoA): In the example below, although "i took a pill in ibiza seeb remix" could be considered as a WoA, the innermost entities are annotated instead, namely "i took a pill in ibiza" as WoA and "seeb" as Artist: songs similar to i took a pill in ibiza seeb remix. Exception: if the name of a well-known band that you recognize is composed of 2 or more individual artist names, then annotate the band name using the category Artist_known. In the example below, I recognized that "emerson lake and palmer" is the name of a band despite the fact that it refers to three individual artists ("emerson", "lake", "palmer"): other artists similar to emerson lake and palmer. Explicit versus implicit named entities. There are cases when an Artist or a WoA are mentioned in text, but these entities are not explicitly named. For instance, neither "the last album", nor "this singer" are explicit named entities in the request below; hence they must not be annotated: show me something similar to the last album of this singer. (Incorrect) Variations of the original named entities. The text may contain variations of the original names of the entities (including misspelled, missing, translated or transliterated words). Normally, in order to recognize an incorrectly written named entity, the named entity must be already known to the annotator. In these cases, even if the named entity does not match exactly the real name, the annotator is required to annotate the corresponding span of text using the named entity categories ending with the "_known" suffix. In the example below, the annotator recognizes "hey ponchuto" to be mistakenly written: fast dancey blues or songs like hey ponchuto from the mask. The original named entity which the annotator knows from the past is "hey pachuco". Thus, "hey ponchuto" is annotated as WoA_known. Note that "the mask" is a WoA too (a movie). § PRE-DEFINED ENTITY TYPES FOR WIKIDATA LINKING In order to ensure that the entity linking gives priority to music-related entities, we re-rank the returned results. Specifically, we return the first entity whose type matches any of the criteria below: * Artist: type matches exactly one of the following types—musical group, rock group, supergroup, musical ensemble, girl group, or it contains one of the following strings—band, duo, musician, singer. WoA: type matches exactly one of the following types—album, musical work/composition, song, single, extended play, or it contains one of the following strings—album, song. If the previous matching fails, the fallback is the first entity of type human for an Artist entity, or of type video or film for a WoA entity. If none of these type criteria is met, then an empty string, corresponding to failed linking is returned. § COMPUTATIONAL INFORMATION For training and evaluation, we had a 32-core Intel Xeon Gold 6134 CPU @ 3.20GHz CPU with 128GB RAM, equipped with 4 GTX 1080 GPUs, each with 11GB RAM. Fine-tuning a single model on three datasets from the four we annotated during 3 epochs and testing it on the hold-out dataset on a single GPU took about 2 minutes.

http://arxiv.org/abs/2303.06889v1

20230313062758

Finding the minimum distance and decoding linear codes with Gaussian Elimination Method

cs.IT

[Tianshuo Yang: Department of Robotics Engineering, Widener University, Chester, PA 19013, USA, Email: tyang3@widener.edu.] We propose an algorithm using the Gaussian elimination method to find the minimal Hamming distance and decode received messages of linear codes. This algorithm is easy to implement as it requires no Gröbner bases to compute solutions for systems of polynomial equations. Finding the minimum distance and decoding linear codes with the Gaussian elimination method Tianshuo Yang March 30, 2023 =========================================================================================== § INTRODUCTION Let 𝕂 be a field and 𝕂^n the n-dimensional 𝕂-vector space. A linear code C of length n over 𝕂 is a subspace of 𝕂^n. Let v=[[ v_1 … v_n ]] and w=[[ w_1 … w_n ]] be two vectors in 𝕂^n. Recall the (Hamming) distance from v to w, denoted by d(v, w), is defined to be the number of positions at which v and w differ. The (Hamming) weight of v, denoted by wt(v), is defined as d(v, 0), i.e., the number of entries in v that are different from zero. Let C be a code containing at least two codewords. The minimum (Hamming) distance of C is defined as d(C)= min{d(v, w) | v, w∈ C, v≠w} and the minimum (Hamming) weight of C is defined as wt(C)= min{ wt (v) | v∈ C, v≠0}. In the case that C is a linear code, one has d(C)= wt(C). By an [n, k, d] linear code C we mean a k-dimensional subspace of 𝕂^n such that the minimum distance d(C)=d. Under the standard coding theory, the code C can detect up to d-1 errors and correct up to ⌊d-1/2⌋ errors. Thus, determining the value of d is critical for understanding the error detection/correction capability of C. However, Vardy <cit.> showed that for general linear codes, computing the minimum distance is an NP-hard problem and the corresponding decision problem is NP-complete. Hence any general algorithm for computing the minimum distance will run in super polynomial time unless P=NP. Historical techniques of decoding and computing the minimal distance reduce the problems to systems of polynomial equations in several variables over finite fields and then use Gröbner bases to solve them (see for example <cit.>, <cit.>, and <cit.>). Gröbner bases can be computed via computer algebra packages such as Axiom, CoCoA, Gap, Macaulay, Singular, etc. The complexity of computing them is exponential in the case of a finite set of solutions. Motivated by the work of M. De Boer and R. Pellikaan <cit.>, B. Anzis and S. Tohăneanu <cit.>, and R. Burity, S. Tohăneanu and Y. Xie <cit.>, we reduce the problem of decoding and computing the minimal distance to systems of linear equations in several variables and use Gaussian elimination in linear algebra to compute their solutions. This paper is structured as follows: In next section, we propose an algorithm using Gaussian elimination for computing the minimal distance and codewords with minimal distance for a linear code. Then we provide a similar algorithm to decode received messages of a linear code in Section 3. After that, we show examples to illustrate our algorithms. In the last section, we close the paper with concluding remarks. § THE MINIMAL HAMMING DISTANCE Let C be an [n, k, d] linear code over the field 𝕂. Since C is a k-dimensional subspace of the vector space 𝕂^n, one can use a basis of C to form a generating matrix of rank k G=[[ a_11 a_12 ⋯ a_1n; a_21 a_22 ⋯ a_2n; ⋮ ⋮ ⋮; a_k1 a_k2 ⋯ a_kn ]], where a_ij∈𝕂. Observe C is the image of the injective linear map ϕ: 𝕂^k 𝕂^n via ϕ(x)=xG for x∈𝕂^k. Assume G is nondegenerate, i.e., none of the columns of G is the zero column vector in 𝕂^k. Let R=𝕂[x_1, …, x_k] be a polynomial ring of k variables over the field 𝕂. Observe the n columns in the generating matrix G define n nonzero homogeneous linear forms ℓ_j=∑_i=1^ka_ijx_i 1≤ j≤ n, in R. These linear forms ℓ_1, …, ℓ_n are called the defining linear forms for the linear code C. For 1≤ a≤ n, the ideal generated by a-fold products of linear forms of C is defined as I_a(C)=⟨{ℓ_i_1⋯ℓ_i_a | 1≤ i_1<⋯<i_a≤ n}⟩. Let ℙ_𝕂^k-1 be the projective (k-1)-space. Recall the projective variety defined by a homogeneous ideal J⊂ R is V(J)={P∈ℙ_𝕂^k-1 | f(P)=0 f∈ J}. Let v be a nonzero codeword in C. Then v=xG, where 0≠x∈𝕂^k. Observe wt(v)≤ e if and only if all products of e+1 distinct entries of v are zero. This means x is a nonzero solution for all of the equations ℓ_i_1⋯ℓ_i_e+1=0, where 1≤ i_1<⋯<i_e+1≤ n, i.e., x∈ V(I_e+1(C)). One has that (see <cit.>) V(I_a(C))={x∈ℙ_𝕂^k-1 | wt(v)<a with v=xG}, and d= min{a | V(I_a+1(C))≠∅}. Let Γ(C) be the set of all linear prime ideals generated by linear forms in {ℓ_1, …, ℓ_n}. Let 𝔭∈Γ(C) and ν_C(𝔭) be the number of linear forms in {ℓ_1, …, ℓ_n} that belong to 𝔭. The irrelevant maximal ideal 𝔪=⟨ x_1, …, x_k⟩=⟨ℓ_1, …, ℓ_n⟩ and ν_C(𝔪)=n. By <cit.>, for 1≤ a≤ n, the ideal I_a(C) has the primary decomposition I_a(C)=∩_𝔭∈Γ(C)𝔭^a-n+ν_C(𝔭), where if a-n+ν_C(p)≤ 0, then the corresponding component is replaced with the ring R. Consider the case a=1 and I_1(C)=⟨ℓ_1, …, ℓ_n⟩=⟨ x_1, …, x_k⟩. One has V(I_1(C))=∅ as ℓ_1, …, ℓ_n span a k-dimensional vector space so that the homogeneous equations ℓ_1=0, …, ℓ_n=0 have only the trivial solution 0 in 𝕂^k. By the above results, to compute the minimal distance d, one needs to find the maximum number of linear forms in {ℓ_1, …, ℓ_n} that span a k-1-dimensional vector space. In other words, d is the minimal number of linear forms we have to delete from {ℓ_1, …, ℓ_n} so that the remaining linear forms span a k-1-dimensional vector space. The question of whether a set of linear forms span a (k-1)-dimensional vector space can be answered by reducing the matrix formed by the coefficients of these linear forms to a row echelon matrix and looking at the rank. Since we need at least k-1 linear forms to span a k-1-dimensional vector space, one has that 1≤ d≤ n-k+1. We state the algorithm using Gaussian elimination for computing the minimal distance and codewords with minimal distance in the following: Algorithm 2.1 * Input a generating matrix G of size k× n * For 1≤ j≤ n-k+1 * Compute S_j={{i_1, …, i_j} | 1≤ i_1<i_2<⋯<i_j≤ n} * For each {i_1, …, i_j}∈ S_j * Let G_i_1⋯ i_j be the matrix obtained from G by deleting j columns i_1, …, i_j * Reduce G_i_1⋯ i_j^T to a row echelon matrix H. Compute rank(H). * If rank(H)<k Solve the system of linear equations Hx=0 in the projective space ℙ_𝕂^k-1 Store the solutions in the set X * If X≠∅, let d=j and Y=X^TG. Return d, X, and Y. In the above algorithm, observe |S_j|=nj for each 1≤ j≤ n-k+1. Thus the complexity of this algorithm is large as the number of all possible subsets of {1, …, n} grows exponentially. However, the algorithm is easy to implement and requires no symbolic computations in computer algebra systems. § DECODING LINEAR CODES Let C be an [n, k, d] linear code with a generating matrix G as in Section 2. Suppose that a codeword w=[[ w_1 … w_n ]]∈𝕂^n is received. The most commonly used rule for decoding w is to find the codeword v∈ C which minimizes wt(w-v) (i.e., v is the nearest neighbor of w in C), and decode w to v. Of course, a codeword w∉C might have more than one nearest neighbors. In this case the nearest neighbor decoding rule fails. As we have mentioned above, the minimal distance d determines the error detection/correction capability of C as it can detect up to d-1 errors and correct up to ⌊d-1/2⌋ errors. Traditionally one translates the syndrome decoding algorithm into the language of varieties (called syndrome varieties) and use computational algebraic techniques (such as Gröbner bases) to find the error and the nearest neighbor of a received word (see for example <cit.>, <cit.>, and <cit.>). In 2015, it was shown in <cit.> that any error with weight up to ⌊d-1/2⌋ in data transmission can be computed as the codeword of minimum weight of a new linear code C(w) with a generating matrix G(w)=[[ a_11 a_12 ⋯ a_1n; ⋮ ⋮ ⋮; a_k1 a_k2 ⋯ a_kn; w_1 w_2 ⋯ w_n ]], which is created from the generator matrix G of C by augmenting the received word w as a new row (a code with such a generating matrix is called an augmented code). Let d(w)= min{ wt(e) | e∈𝕂^n with w-e∈ C}= min{ wt(e) | e∈ C(w)}. Then one can compute the codeword of minimum weight by solving the ideal I_d(w)+1(C(w)) using Gröbner bases or by finding a primary decomposition of this ideal. Both methods require symbolic computations. Applying similar idea to the new linear code C(w) as in Section 2, we can use Gaussian elimination to compute the codeword of minimal weight in C(w). Hence we obtain the following algorithm for decoding a linear code C with minimal distance d: Algorithm 3.1 * Input a generating matrix G of size k× n and a received message w * Let G(w) be the matrix obtained by appending w to G in the last row * For 1≤ j≤⌊d-1/2⌋ * Compute S_j={{i_1, …, i_j} | 1≤ i_1<i_2<⋯<i_j≤ n} * For each {i_1, …, i_j}∈ S_j * Let G(w)_i_1⋯ i_j be the matrix obtained from G(w) by deleting j columns i_1, …, i_j * Reduce G(w)_i_1⋯ i_j^T to a row echelon matrix H. Compute rank(H). * If rank(H)<k+1 Solve the system of equations Hx=0 in the projective space ℙ_𝕂^k Let x be the nonzero solution (with the last entry =1) Let e=x^TG(w) and v=w-e Return e and v * Return w is non-decodable § EXAMPLES In this section, we provide examples to illustrate the above algorithms for computing the minimal distance and decoding linear codes. Example 4.1 Let 𝕂=𝔽_2. Consider the linear code C with a generating matrix G=[[ 1 0 0 1 1 0; 0 1 0 1 0 1; 0 0 1 0 1 1 ]]. This code has 6 homogeneous linear forms x_1, x_2, x_3, x_1+x_2, x_1+x_3, x_2+x_3 generated by the columns of G. Applying Algorithm 2.1, we found that d=d(C)=3 and there are 4 linear prime ideals of height 2; each is generated by n-d=6-3=3 linear forms: 𝔭_1=⟨ x_1+x_2, x_1+x_3, x_2+x_3⟩, 𝔭_2=⟨ x_2, x_3, x_2+x_3⟩, 𝔭_3=⟨ x_1, x_3, x_1+x_3⟩, and 𝔭_4=⟨ x_1, x_2, x_1+x_2⟩. To find ∪_i=1^4V(𝔭_i), we solve the 4 homogeneous linear systems with augmented matrices (formed by the coefficients of linear forms in 𝔭_i, 1≤ i≤ 4) [[ 1 1 0 0; 1 0 1 0; 0 1 1 0 ]], [[ 0 1 0 0; 0 0 1 0; 0 1 1 0 ]], [[ 1 0 0 0; 0 0 1 0; 1 0 1 0 ]], [[ 1 0 0 0; 0 1 0 0; 1 1 0 0 ]], and obtain 4 nonzero solutions x_1=[[ 1 1 1 ]]^T, x_2=[[ 1 0 0 ]]^T, x_3=[[ 0 1 0 ]]^T, and x_4=[[ 0 0 1 ]]^T. This yields 4 codewords in C of minimal weight 3: y_1 = x_1^TG=[[ 1 1 1 0 0 0 ]], y_2 = x_2^TG=[[ 1 0 0 1 1 0 ]], y_3 = x_3^TG=[[ 0 1 0 1 0 1 ]], y_4 = x_4^TG=[[ 0 0 1 0 1 1 ]]. Hence the outputs of Algorithm 2.1 are d=3, X={[[ 1 1 1 ]]^T, [[ 1 0 0 ]]^T, [[ 0 1 0 ]]^T, [[ 0 0 1 ]]^T}, and Y={[ [[ 1 1 1 0 0 0 ]] [[ 1 0 0 1 1 0 ]]; ; [[ 0 1 0 1 0 1 ]] [[ 0 0 1 0 1 1 ]] ]}. Indeed, there are 8 codewords in C, i.e., C={[ [[ 0 0 0 0 0 0 ]] [[ 1 0 0 1 1 0 ]] [[ 0 1 0 1 0 1 ]] [[ 0 0 1 0 1 1 ]]; ; [[ 1 1 0 0 1 1 ]] [[ 1 0 1 1 0 1 ]] [[ 0 1 1 1 1 0 ]] [[ 1 1 1 0 0 0 ]] ]}. One can see that d(C)=3 and there are 4 codewords in C of minimal weight 3. Since d(C)=3, the linear code C can fix one error. Suppose the codeword w=[[ 0 1 1 1 0 0 ]] is received. Then we have the augumented code C(w) with the augumented matrix G(w)=[[ 1 0 0 1 1 0; 0 1 0 1 0 1; 0 0 1 0 1 1; 0 1 1 1 0 0 ]]. The augumented code C(w) has 6 homogeneous linear forms x_1, x_2+x_4, x_3+x_4, x_1+x_2+x_4, x_1+x_3, x_2+x_3 generated by the columns of G(w). Applying Algorithm 3.1, we have d(C(w))=1 and there is a unique linear prime ideal of height 3 generated by n-d=6-1=5 linear forms: 𝔭=⟨ x_1, x_2+x_4, x_3+x_4, x_1+x_2+x_4, x_2+x_3⟩. Again to find V(𝔭), we solve the homogeneous linear system with augmented matrix (formed by the coefficients of linear forms in 𝔭) [[ 1 0 0 0 0; 0 1 0 1 0; 0 0 1 1 0; 1 1 0 1 0; 0 1 1 0 0 ]] and obtain the unique nonzero solution x=[[ 0 1 1 1 ]]^T. The error codeword e=x^TG(w)=[[ 0 0 0 0 1 0 ]]. Hence the codeword in C with minimal distance to w is v=w-e=[[ 0 1 1 1 1 0 ]]. The outputs of Algoritm 3,1 are e=[[ 0 0 0 0 1 0 ]] and v=[[ 0 1 1 1 1 0 ]], and we decode w to v. Example 4.2 Let 𝕂=𝔽_2. Consider the [7, 4] cyclic code C generated by the polynomial g(x)=1+x^2+x^3. Then this code has a generating matrix (note that we identify a vector with a polynomial) G=[[ g(x); xg(x); x^2g(x); x^3g(x) ]]= [[ 1 0 1 1 0 0 0; 0 1 0 1 1 0 0; 0 0 1 0 1 1 0; 0 0 0 1 0 1 1 ]]. This code has 7 homogeneous linear forms x_1, x_2, x_1+x_3, x_1+x_2+x_4, x_2+x_3, x_3+x_4, x_4 generated by the columns of G. Applying Algorithm 2.1, we found that d=d(C)=3 and there are 7 linear prime ideals of height 3 generated by n-d=7-3=4 linear forms: 𝔭_1=⟨ x_1, x_2, x_1+x_3, x_2+x_3⟩, 𝔭_2=⟨ x_1, x_2, x_1+x_2+x_4, x_4⟩, 𝔭_3=⟨ x_1, x_1+x_3, x_3+x_4, x_4⟩, 𝔭_4=⟨ x_1, x_1+x_2+x_4, x_2+x_3, x_3+x_4⟩, 𝔭_5=⟨ x_2, x_1+x_3, x_1+x_2+x_4, x_3+x_4⟩, 𝔭_6=⟨ x_2, x_2+x_3, x_3+x_4, x_4⟩, and 𝔭_7=⟨ x_1+x_3, x_1+x_2+x_4, x_2+x_3, x_4⟩. To find ∪_i=1^7V(𝔭_i), we solve the 7 homogeneous linear systems with augmented matrices (formed by the coefficients of linear forms in 𝔭_i, 1≤ i≤ 7) [[ 1 0 0 0 0; 0 1 0 0 0; 1 0 1 0 0; 0 1 1 0 0 ]], [[ 1 0 0 0 0; 0 1 0 0 0; 1 1 0 1 0; 0 0 0 1 0 ]], [[ 1 0 0 0 0; 1 0 1 0 0; 0 0 1 1 0; 0 0 0 1 0 ]], [[ 1 0 0 0 0; 1 1 0 1 0; 0 1 1 0 0; 0 0 1 1 0 ]], [[ 0 1 0 0 0; 1 0 1 0 0; 1 1 0 1 0; 0 0 1 1 0 ]], [[ 0 1 0 0 0; 0 1 1 0 0; 0 0 1 1 0; 0 0 0 1 0 ]], [[ 1 0 1 0 0; 1 1 0 1 0; 0 1 1 0 0; 0 0 0 1 0 ]], and obtain 7 nonzero solutions, i.e., X={x_i | 1≤ i≤ 7}, where x_1=[[ 0 0 0 1 ]]^T, x_2=[[ 0 0 1 0 ]]^T, x_3=[[ 0 1 0 0 ]]^T, x_4=[[ 0 1 1 1 ]]^T, x_5=[[ 1 0 1 1 ]]^T, x_6=[[ 1 0 0 0 ]]^T, and x_7=[[ 1 1 1 0 ]]^T. This yields 7 codewords in C of minimal weight 3, i.e., Y={y_i | 1≤ i≤ 7}, where y_1 = x_1^TG=[[ 0 0 0 1 0 1 1 ]], y_2 = x_2^TG=[[ 0 0 1 0 1 1 0 ]], y_3 = x_3^TG=[[ 0 1 0 1 1 0 0 ]], y_4 = x_4^TG=[[ 0 1 1 0 0 0 1 ]], y_5 = x_4^TG=[[ 1 0 0 0 1 0 1 ]], y_6 = x_4^TG=[[ 1 0 1 1 0 0 0 ]], y_7 = x_4^TG=[[ 1 1 0 0 0 1 0 ]]. By computing the 16 codewords in C, one can verify that d(C)=3 and there are 7 codewords of minimal weight 3. Again since d(C)=3, the cyclic code C can only fix one error. Suppose w=[[ 1 1 0 1 0 1 1 ]] is the received codeword. Then we have the augumented code C(w) with the augumented matrix G(w)=[[ 1 0 1 1 0 0 0; 0 1 0 1 1 0 0; 0 0 1 0 1 1 0; 0 0 0 1 0 1 1; 1 1 0 1 0 1 1 ]]. This augumented code C(w) has 7 homogeneous linear forms x_1+x_5, x_2+x_5, x_1+x_3, x_1+x_2+x_4+x_5, x_2+x_3, x_3+x_4+x_5, x_4+x_5 generated by the columns of G(w). Applying Algorithm 3.1, we have d(C(w))=1 and there is a unique linear prime ideal of height 4 generated by n-d=7-1=6 linear forms: 𝔭=⟨ x_1+x_5, x_2+x_5, x_1+x_3, x_1+x_2+x_4+x_5, x_2+x_3, x_4+x_5⟩. To find V(𝔭), we solve the homogeneous linear system with augmented matrix (formed by the coefficients of linear forms in 𝔭) [[ 1 0 0 0 1 0; 0 1 0 0 1 0; 1 0 1 0 0 0; 1 1 0 1 1 0; 0 1 1 0 0 0; 0 0 0 1 1 0 ]] and obtain the unique nonzero solution x=[[ 1 1 1 1 1 ]]^T. The error codeword e=x^TG(w)=[[ 0 0 0 0 0 1 0 ]]. Hence the codeword in C with minimal distance to w is v=w-e=[[ 1 1 0 1 0 0 1 ]]. The outputs of Algoritm 3,1 are e=[[ 0 0 0 0 0 1 0 ]] and v=[[ 1 1 0 1 0 0 1 ]], and we decode w to v. § CONCLUDING REMARKS The purpose of this paper is to propose a practical algorithm for computing the minimal distance and decoding general linear codes without using symbolic computations in computer algebra systems. The computational complexity of this algorithm is large as one cannot hope for a polynomial algorithm to decode and compute the minimal distance for general linear codes. However since the implementation is not hard, one can use it to compute certain examples to verify your intuition. This can help design new linear codes. One can modify the above algorithm to compute other things for an [n, k] linear code C such as the primary decomposition of I_a(C) for 1≤ i≤ n, the weight distribution {(i, α_i) | i=0, 1, …, n}, where α_i denotes the number of codewords in C of weight i, the polynomial W_C(X, Y)=∑_i=0^nα_iX^n-iY^i, and the MacWilliams identity W_C^⊥(X, Y)=q^-kW_C(X+(q-1)Y, X-Y), where 𝕂=𝔽_q is a finite field of q elements and C^⊥ is the dual code of C. This algorithm may be improved to reduce the computational complexity in special classes of linear codes. For example, if the generator matrix G is of the form [[ G_1 ⋯ G_s ]], where G_1, …, G_s are matrices such that the first k columns of G_j form the k× k identity matrix for all j=1, …, s, we know that any linear combination of r rows with non-zero coefficients gives a codeword of weight at least rs. Hence s≤ d≤ n-k+1 and we can start our loop from j=s. In the future, one may implement this algorithm to study special classes of linear codes such as cyclic codes, maximum distance separable (MDS) codes, BCH codes, Golay codes, etc. 0.5in amsalpha 10 AT B. Anzis and S. Tohăneanu, Error-correction of linear codes via colon ideals, J. Algebra 443 (2015), 479-493. BP S. Bulygin and R. Pellikaan, Decoding and finding the minimum distance with Gröbner bases: history and new insights, pp. 585-622 in: I. Woungang, S. Misra, S. C. Misra (editors). Series on Coding Theory and Cryptology vol. 7, Selected Topics in Information and Coding Theory, World Scientific 2010. BTX R. Burity, S. Tohăneanu and Y. Xie, Ideals generated by a-fold products of linear forms have linear graded free resolution, to appear in Michigan Math J., arXiv: 2004.07430. DP1 M. De Boer and R. Pellikaan, Gröbner bases for codes, In: Some Tapas of Computer Algebra, 237-259, Springer, Berlin 1999. DP2 M. De Boer and R. Pellikaan, Gröbner bases for decoding, In: Some Tapas of Computer Algebra, 260-275, Springer, Berlin 1999. V A. Vardy, The intractability of computing the minimum distance of a code, IEEE Trans. Inf. Theory, 43 (1997), 1757-1766.

http://arxiv.org/abs/2303.06983v2

20230313103333

Measuring the environment of a Cs qubit with dynamical decoupling sequences

quant-ph

APS/123-QED widera@rptu.de Department of Physics and Research Center OPTIMAS, University of Kaiserslautern-Landau, Erwin-Schrödinger-Straße 46, D-67663 Kaiserslautern, Germany We report the experimental implementation of dynamical decoupling on a small, non-interacting ensemble of up to 25 optically trapped, neutral Cs atoms. The qubit consists of the two magnetic-insensitive Cs clock states |F=3, m_F=0⟩ and |F=4, m_F=0⟩, which are coupled by microwave radiation. We observe a significant enhancement of the coherence time when employing Carr-Purcell-Meiboom-Gill (CPMG) dynamical decoupling. A CPMG sequence with ten refocusing pulses increases the coherence time of 16.2(9) by more than one order of magnitude to 178(2). In addition, we make use of the filter function formalism and utilize the CPMG sequence to measure the background noise floor affecting the qubit coherence, finding a power-law noise spectrum 1/ω^α with α = 0.89(2). This finding is in very good agreement with an independent measurement of the noise in the intensity of the trapping laser. Moreover, the measured coherence evolutions also exhibit signatures of low-frequency noise originating at distinct frequencies. Our findings point toward noise spectroscopy of engineered atomic baths through single-atom dynamical decoupling in a system of individual Cs impurities immersed in an ultracold ^87Rb bath. Measuring the environment of a Cs qubit with dynamical decoupling sequences Artur Widera March 30, 2023 =========================================================================== § INTRODUCTION The tremendous progress in manipulating and measuring ultracold quantum gases makes them a versatile platform for quantum technologies. Electric and magnetic field control enable the precise creation of quantum superpositions for targeted applications in quantum computing <cit.>, quantum simulation <cit.>, and quantum sensing <cit.>. In all these applications, information is stored in the relative phase of such a quantum superposition in so-called quantum bits (qubits). These coherent superpositions, however, are usually very fragile because they are sensitive to external perturbations. In particular, even in highly specialized labs, quantum systems cannot be perfectly isolated from their environment, which will lead to dissipation and eventually to decoherence, i.e., the decay of a pure quantum state into a statistical mixture. In this regard, the reduction of decoherence and the extension of coherence times is a prime challenge in quantum science which will enable the broad applications of quantum technologies. One of the main sources of decoherence is dephasing caused by time-dependent (classical) fluctuations. It is known that such dephasing can, in principle, be reversed by so-called dynamical decoupling, a coherent control-pulse method that was originally developed in the field of solid-state physics. The primary purpose of dynamical decoupling is the effective reduction of the coupling between the qubit and a mostly generic environment by a sequence of pulses. These single <cit.> and multi-pulse sequences <cit.> have already been used to extend coherence in various physical systems, ranging from spin ensembles <cit.>, semiconductor quantum dots <cit.>, nitrogen-vacancy centers in diamond <cit.>, superconducting qubits <cit.>, ensembles of ultracold atoms and ions <cit.> to the limit of single atoms and ions <cit.>. Consequently, many key characteristics and possible causes of qubit decoherence in ultracold atomic systems have already been discussed and explained in various experimental and theoretical works <cit.>. Remarkably, such sequences cannot only be utilized to partially reverse the decoherence dynamics but also to perform noise spectroscopy, i.e., to measure details of the coupled, unknown degrees of freedom of the environment <cit.>. This first appears to be counter-intuitive since sensing is usually more effective if the coupling between the probe and the measurement object is strong. The idea of noise spectroscopy through dynamical decoupling, by contrast, is the modification of the spectral properties of the qubit which probes the environmental noise spectrum in different frequency ranges. While one could expect that the measurement of all environmental effects is a hopeless task, the most relevant effects can be directly extracted from the detailed knowledge of the qubit's dynamics. This creates the possibility to recover details of the environment by controlling and measuring the qubit's time evolution. From a fundamental point of view, the dynamical driving of the qubit can be used for the creation of correlated states between the environmental and the qubit's degrees of freedom, closely related to spin-boson <cit.>, polaron <cit.>, and Kondo physics <cit.>. From a more practical point of view, the gained knowledge about the environment can be utilized to improve the experiment by applying optimized sequences tailored to the situation in the lab. In this paper, we exploit the dynamical decoupling method to determine essential properties of the environmental noise spectrum affecting the qubit coherence in a small ensemble of optically trapped, neutral ^133Cs atoms. We use the two magnetic-insensitive Zeeman levels |F=3, m_F=0⟩ and |F=4, m_F=0⟩ of the 6 S_1/2 ground state as qubit states and couple them by microwave radiation. Such atoms have been immersed into an ultracold atomic gas recently <cit.> for probing coherence and dephasing dynamics of individual Cs impurities beyond the effect of electromagnetic fields. In this work, we demonstrate how dynamical qubit control can give access to essential properties of a coupled environment. Our results represent a subsequent stepping stone towards utilizing the Cs qubit as a quantum probe for correlated environments consisting of interacting quantum particles. The paper is structured as follows. Sec. <ref> comprises a description of the experimental setup and a detailed characterization of the qubit. In Sec. <ref>, we introduce the theoretical description of the dynamically driven qubit in a decohering environment. Then, in Sec. <ref>, we present the experimental realization of the dynamical decoupling sequence and a comparison to theory. After that, we conclude our results in Sec. <ref> and provide a brief outlook. § EXPERIMENTAL SYSTEM §.§ Experimental setup A schematic overview of the trapping geometry in our experimental setup is given in Fig. <ref>. The preparation of the Cs atoms is a three-stage process. First, about 50 Cs atoms are captured in a high-gradient magneto-optical trap and loaded into a crossed optical dipole trap formed by the intersection of two 1064 laser beams. Subsequently, the Cs atoms are further cooled to a temperature of T_Cs = 5.9(3) using degenerate Raman sideband cooling (see Refs. <cit.> for additional information on the temperature measurement). In addition to the cooling effect, Raman cooling optically pumps the Cs atoms to the absolute ground state |F=3, m_F=3⟩. The Cs atoms are then transferred to the magnetic-insensitive state |F=3, m_F=0⟩ by four successive microwave Landau-Zener sweeps <cit.>. We remove residual Cs atoms that are not in |F=3, m_F=0⟩ from the trap by a subsequent cleaning scheme, which is based on the spin selective readout in Ref. <cit.> and consists of a combination of microwave Landau-Zener sweeps and resonant laser pulses. This preparation process eventually leaves about 15 to 25 Cs atoms in the desired state |F=3, m_F=0⟩ and, at most, two Cs atoms in states |F=3, m_F ≠ 0⟩ <cit.>. In the last step, a one-dimensional optical lattice formed by two counter-propagating 790 laser beams is used as a conveyor belt to transport the Cs atoms to a second crossed optical dipole trap located at an axial distance of about 208. Here, a coherent superposition of the Cs clock states is prepared, as explained in the subsequent section below, and the dynamical decoupling sequence is applied. Importantly, the Cs atoms are always trapped in the static optical lattice during the sequence. The lattice creates a repulsive potential for the Cs atoms along the horizontal direction and thereby freezes the atoms' position in this dimension. This results in trapping frequencies of ω_ax = 2 π×60.7 and ω_rad = 2 π×694 in the axial and the radial direction, respectively. Moreover, it ensures that Cs-Cs interactions are negligible since, at most, one Cs atom can be trapped in each lattice site. At the end of each dynamical decoupling sequence, a resonant laser pulse removes all Cs atoms in |F=4⟩ from the trap, and the optical lattice is used for spatially-resolved fluorescence imaging of Cs atoms in |F=3, m_F=0⟩. §.§ Qubit characterization We start with a detailed characterization of the qubit transition before providing information about the dynamical decoupling sequence in Sec. <ref>. For all measurements shown in the following, the external magnetic field at the atoms' position is calibrated to a value of 198.5 (see Appendix <ref> for further details on the magnetic field calibration). The qubit's transition frequency is measured using microwave spectroscopy, taking care to avoid power broadening. The corresponding resonant Rabi frequency Ω_R = 2 π×619(2) is independently measured, and Fig. <ref>(a) shows the spectrum of a microwave square pulse with τ = 725 duration. The signal follows the typical shape of the spectrum of a Fourier-limited square pulse N_Cs(δν_0) = A Ω_R^2/Ω^2sin^2 (Ωτ/2) + C, where Ω^2 = 4 π^2 (δν_0 - δν_0, res)^2 + Ω_R^2, and A and C are the amplitude and the offset of the fit, respectively. The peak frequency is shifted from the exact atomic resonance frequency ν_0 = 9.192 631 77 by δν_0, res= -133(12). The two main contributions to this shift are the differential light shift δ_DT of the trapping field and the quadratic Zeeman shift δ_B. We find that the measured shift is in good agreement with the expected shift δν_0, theo= -123 calculated from experimental trapping parameters and the Breit-Rabi formula. For dynamical decoupling sequences with multiple pulses, the pulse durations are important timescales determining the duration of the sequences and must be short compared to all other experimental timescales. We achieve for all dynamical decoupling measurements a resonant Rabi frequency of Ω_R = 2 π×21.14(4) [Fig. <ref>(b)], which corresponds to a π-pulse duration of τ_π = 23.65. The Cs clock transition chosen is electric-dipole forbidden so that the excited state is long-lived, and T_1 decay occurs on the timescale of seconds and will therefore be neglected. Another important timescale is the lifetime of the Cs atoms in the optical lattice, which is limited by heating due to the relative phase noise between the two laser beams to about 685(81). Moreover, a sufficient timescale separation between coherence time and π-pulse duration is required for realizing multi-pulse sequences. To measure the coherence time in the absence of dynamical decoupling, we apply a Ramsey sequence consisting of two microwave π/2-pulses with variable sequence length T [Fig. <ref>(c)]. The microwave frequency is set to the unperturbed transition frequency ν_0, i.e., detuned by approximately 133 from the actual qubit resonance so that several fringes can be observed within the decay time. However, the pulses can still be approximated as near-resonant, as the detuning is much smaller than the bare Rabi frequency. For such a sequence, the expected population dynamics can be described by N_Cs(T) = A/2[1 - cos(2π |δν_0, res| T) e^ -(T/τ̃_c)^2] + C, where A and C are the amplitude and offset of the fit. Previously, this functional behavior was a heuristic assumption that describes accurately the dynamics of the experiment. Later in this paper, in Sec. <ref>, we will provide a justification of this formula which is a consequence of the measured noise spectrum. We fit Eq. (<ref>) to our data in Fig. <ref>(c) and find a frequency shift of |δν_0, res|= 136.8(8), close to the frequency shift measured through microwave spectroscopy, and a coherence time of τ̃_c = 15.7(7). A detailed study on inhomogeneous and homogeneous dephasing effects causing decoherence in an ensemble of Cs atoms confined by a far-off-resonant standing-wave optical dipole trap can be found in Ref. <cit.>. Inhomogeneous dephasing is mainly caused by inhomogeneous differential light shifts originating from the thermal motion of the atoms in the trapping potential, which lead to slightly different resonance frequencies among the ensemble. The effect of this dephasing mechanism can be reversed by applying a spin-echo sequence comprising an additional π-pulse between the two Ramsey π/2-pulses. Homogeneous dephasing effects, instead, lead to homogeneous, time-dependent fluctuations of the transition frequency among the ensemble. The main causes of homogeneous dephasing in our experimental setup are intensity fluctuations of the dipole trap laser and heating of the atoms (e.g., due to relative phase noise between the two lattice laser beams). Fluctuating magnetic fields are another source of broadening which is, however, far less pronounced because of the use of magnetic-field-insensitive states. Additional homogeneous dephasing sources, that play a minor role in our experimental setup, are the pointing instability of the dipole trap laser and imperfections of microwave pulses. Dephasing caused by homogeneous broadening cannot be simply reversed by the spin-echo sequence. Here, we need to apply dynamical decoupling sequences, which extend the spin-echo scheme by using multiple π-pulses, to compensate for noise originating at different frequencies. The theoretical background for this is presented in the following section. § QUBIT COHERENCE IN A NOISY ENVIRONMENT We use a semiclassical model to describe the dynamics of the driven qubit coupled to environmental noise <cit.>. Hereby, we study the dynamics of the density matrix ρ̂ of the qubit, which is governed by the von Neumann equation ∂ρ̂/∂ t=1/iħ[Ĥ,ρ̂]. The corresponding Hamiltonian is Ĥ=ħΩ(t)/2σ̂_x+ħβ(t)/2σ̂_z, and reported in the frame rotating with the qubit eigenfrequency. Here, we have introduced the Pauli matrices σ̂_i with i=x,y,z of the two qubit states |F=3, m_F=0⟩ and |F=4, m_F=0⟩. The driving of this qubit is described by the coherent, time-dependent Rabi frequency Ω(t). The influence of the environment is encoded in fluctuations described by the stochastic process β(t). We assume that this noise has a vanishing mean, ⟨β(t)⟩=0, and is determined by the time-correlation function C(t-t^')=⟨β(t)β(t^')⟩. The expectation values are taken over noise realizations. The stochastic fluctuations are characterized by the noise spectrum S(ω)=∫_-∞^∞ dt^' e^-iω t^' C(t^'). In our description, we have neglected noise-driven spin flips, which are expected to be relevant only on timescales much longer than we describe in the following. Moreover, in this work, we will consider a spectral density of the form S(ω)=S_0/|ω|^α, with a positive real number S_0>0 and a positive real exponent α>0, describing an algebraic decay of the spectral density. This spectral density describes a variety of noise processes in various physical scenarios including 1/f-noise. Similar to Ref. <cit.>, we derive a master equation which is valid in the weak coupling approximation. For this, we transform into a frame which is rotating with the driving field Ω(t), with ϱ̂=Ûρ̂Û^† and Û(t)=exp(iΦ(t)σ̂_x/2), Φ(t)=∫_0^tdt^' Ω(t^'). The dynamics of ϱ̂ is then governed by the master equation ∂ϱ̂(t)/∂ t=-∫_0^tdt^' C(t-t^')/4{[Ĵ(t),Ĵ(t^')ϱ̂(t)]+h.c.}, where we introduced Ĵ(t)=Û(t)σ̂_zÛ^†(t). This master equation can be mapped to the Bloch equations for the polarization p=(p_x,p_y,p_z)^T=Tr[σϱ̂], with σ=(σ̂_x,σ̂_y,σ̂_z)^T. The Bloch equations are given by d p/dt=-∫_0^t dt^' C(t-t^') M(t, t^') p(t) with the matrix elements of M(t, t^') M_xx= cos[Φ(t)]cos[Φ(t^')]+sin[Φ(t)]sin[Φ(t^')], M_yy= cos[Φ(t)]cos[Φ(t^')], M_yz= -cos[Φ(t)]sin[Φ(t^')], M_zy= -sin[Φ(t)]cos[Φ(t^')], M_zz= sin[Φ(t)]sin[Φ(t^')]. All other matrix elements are equal to zero. We will consider the case where Ω(t) is a series of δ-shaped π-pulses applied at times t_j, described by Ω(t)=∑_j=1^Nπδ(t-t_j). In this case, we find that sin(Φ)=0 for all times. The initial state in this protocol is obtained after applying a π/2-pulse to the ground state. Therefore, the initial state is polarized in the y direction such that our initial conditions are p_x(0)=0, p_y(0)=1, p_z(0)=0. With these initial conditions, we can solve Eq. (<ref>) and find p_x(T)=0, p_y(T)=V(T), p_z(T)=0, where T is the total integration time and where we have introduced the visibility V(T)=e^-χ(T). The visibility is determined by the coherence integral χ(T) = T^2/2π∫_0^∞ dω S(ω) g(ω, T) and the filter function g(ω,T)= 1/T^2|∫_0^T dt^' e^iω t^'cos[Φ(t^')]|^2 . For the sequence of π-pulses, we obtain g_N(ω, T) = | 1 + (-1)^1+N e^i ω T + 2 ∑_j=1^N (-1)^j e^i ω t_j|^2/(ω T)^2. Until now, we have neither specified the instances t_j for the π-pulses nor the total integration time T. In fact, the result given in Eq. (<ref>) is general and can be used for an arbitrary sequence of π-pulses. In this paper, we employ the N-Carr-Purcell-Meiboom-Gill (N-CPMG) sequence, whose π-pulses are applied at times t_j = (2j-1)T/(2N) <cit.>, and the 1-CPMG sequence corresponds to the spin-echo sequence. This pulse sequence consists of temporally equidistant π-pulses where the first pulse is applied at time t_1=T/(2N) and the last pulse at t_N=T-T/(2N) [see Fig. <ref>(a)]. The choice of the CPMG sequence enables a very efficient decoupling from noise with a spectrum S(ω) given by Eq. (<ref>), especially if we consider no high-frequency cut-off <cit.>. This results from the filter function g_N(ω, T), which is shown in Fig. <ref>(b) for different N-CPMG sequences and a fixed sequence length T. The filter functions' peaks shift to higher frequencies as the number of π-pulses is increased, resulting in different frequency ranges of the noise spectrum contributing to decoherence [Eq. (<ref>)]. Consequently, since the spectral density S(ω) [Eq. (<ref>)] is decreasing with ω, we obtain a lower total decoherence for an increasing number N of π-pulses. We mention, however, that we could also have chosen another pulse sequence such as the Uhrig sequence <cit.> which would show a similar performance in reducing the decoherence. Besides the decoupling from noise, this band-pass filter property of g_N(ω, T) can also be exploited to perform noise spectroscopy. Here, we vary the number N of π-pulses and the sequence length T such that the filter function samples the environmental noise spectrum S(ω). From the measured decoherence signal, we can then reconstruct the spectrum. To demonstrate this, we first use that the sum in Eq. (<ref>) for the N-CPMG sequences can be written with the help of a geometric series g_N(u)=|1+(-1)^1+Ne^iu+2e^iu/(2N)(-1)^Ne^iu-1e^iu/N+1|^2/u^2. Here, we have used that the expression of g_N(ω,T) for the N-CPMG sequence only depends on u=ω T. The expression (<ref>) exhibits for large N a sharp peak whenever u=(2k+1)Nπ, where the denominator in the absolute value vanishes. For large N, we can approximate the filter function as an infinite series g_N(u) ≈∑_k=0^∞4/[(2k+1)π]^2sinc^2(u-(2k+1)Nπ/2), with the introduction of sinc(x)=sin(x)/x. Since the slope of the spectral density Eq. (<ref>) is decreasing, and the appearing sinc functions in Eq. (<ref>) evaluate the noise spectrum at larger and larger frequencies, we can exchange in Eq. (<ref>) sinc^2(x)≈πδ(x) in the large N limit. With this approximation, we can derive from Eq. (<ref>) the following formula χ_N(T) ≈(T/τ_c)^1+α, where we have defined the coherence time τ_c= N^α/(1+α)τ_1, using τ_1={4S_0π^-2-α[1-2^-2-α]ζ(2+α)}^-1/(1+α), and the Riemann ζ function ζ(s)=∑_n=1^∞ 1/n^s. This finding is in agreement with Ref. <cit.>. Equation (<ref>) shows that the exponent α can directly be inferred from the visibility V(T) <cit.>. The latter exhibits an exponential behavior where the exponent of the argument is 1+α, determined by the algebraic behavior of the spectral density. Moreover, the coherence time τ_c [Eq. (<ref>)] is algebraically growing with the number N of pulses. This increasing coherence time can be understood by the shift of the filter function's peak towards higher frequencies for an increasing number N of π-pulses [see Eq. (<ref>) and Fig. <ref>(b)]. The latter leads to a reduction in the net integrated noise and thereby to an enhancement of the coherence time. The exponent of the growth is α/(1+α) which is monotonically increasing with increasing α. Therefore, the noise reduction is more pronounced if the noise spectrum decays faster. For α→∞, the coherence time is increasing linearly with N. § N-CPMG MEASUREMENTS In the experiment, the effect of a N-CPMG sequence on the coherence time is investigated by tracking the visibility of Ramsey fringes for different sequence lengths T. To measure a Ramsey fringe for fixed T, we scan the phase φ of the second π/2-pulse from -185 to 185 in 10 steps and measure the Cs population in |F=3, m_F=0⟩ (Figs. <ref>(a) and <ref>(c), see Appendix <ref> for details on the normalization). We fit a sinusoidal function <cit.> P(T, φ) = a sin^2 (- π/360φ + π/360Φ) + c to the data, where a, c and Φ are the amplitude, the offset, and the free phase of the fit, respectively. The fringe visibility is defined as V(T) = P_max-P_min/P_max + P_min = a/a + 2c. Each data point in Fig. <ref>(c) is an average of typically 12 to 15 experimental runs resulting in a total measurement time of about 18 to 24 for each N-CPMG sequence. Figure <ref> shows the measured visibility evolution for different numbers of π-pulses ranging from N=0 to N=10. The data points predict, in general, a decaying dynamics of the visibility, which is slower for a larger number N of π-pulses. To get a more quantitative access to this data, we fit V(T) = V_0 e^-χ_N (T) + b. to the measured visibility. For the calculation of χ_N, we calculate the coherence integral in Eq. (<ref>) with the expected form of the spectral density in Eq. (<ref>) and the filter function given by Eq. (<ref>) for every number N of π-pulses. This assures that we also take into account small-N effects which are not present in the analytical result shown in Eq. (<ref>). Importantly, each N-CPMG data set owns its specific amplitude V_0 and offset b. In contrast, the parameters S_0 and α are global fitting parameters shared among all N-CPMG data sets. The fits are visible as the solid lines in Fig. <ref> and agree well with the data points. From our fit, we find an exponent α = 0.89(2) and the value S_0=1288(122)^1-α. In addition, we can also extract the coherence time τ_c, which is defined as exp[-χ_N(τ_c)]=1/e, from these fits. The values τ_c are shown in Fig. <ref> as data points in a color scheme that matches the one used for the visibility data shown in Fig. <ref> (N=0 black, N=1 green, etc.). In the absence of dynamical decoupling (N=0), we find a coherence time of τ_c = 16.2(9). This value is in agreement with the coherence time of τ̃_c = 15.7(7) obtained by fitting the population dynamics in Fig. <ref>(c) with Eq. (<ref>). We want to underline that we have assumed a heuristic decay with exp[-(T/τ̃_c)^2] in Eq. (<ref>). To validate this assumption, we can now compare this decay to the analytical result in Eq. (<ref>) which predicts a decay with exp[-(T/τ_c)^1+α] ≈exp[-(T/τ_c)^1.89]. Consequently, the accurate result obtained from Eq. (<ref>) can be explained by the exponent α=0.89(2) which is close to one. For a growing number N of π-pulses, we obtain an increasing coherence time. In particular, we find for the maximum value of N=10 a coherence time of τ_c=178(2), which corresponds to an enhancement of more than one order of magnitude when compared to the N=0 case. To compare these results with analytical estimates, we have calculated Eq. (<ref>) from the fitted value of α and S_0. The result is shown as solid gray line in Fig. <ref> and agrees better for larger numbers of π-pulses. The analytical estimate predicts an algebraic growth with N^α/(1+α)≈ N^0.47. Compared to the theory, the experimental results also suggest that the observed coherence time can be further extended by adding more π-pulses to the CPMG sequence since the Ramsey fringe visibility doesn't drop significantly. This also implies that imperfections in the π-pulses play a minor role in our system. The maximum number of π-pulses is eventually limited by the minimum pulse duration and the maximum sequence length. The former is determined by the available Rabi frequencies and the latter is constrained by the lifetime of the Cs atoms in the optical lattice. In contrast to the strictly monotonous decaying fits visible in Fig. <ref>, our experimental data also show narrow dips in the visibility for N ≥ 6. These dips are highlighted in the insets of Fig. <ref> for N=6,8,10. We expect that these dips originate from several resonances in the noise spectrum. In fact, already a single resonance gives rise to the appearance of multiple dips in the visibility. This can be seen by considering a spectral density S(ω)=S_1e^-(ω-ω_0)^2/(2Δω^2) which describes phenomenologically a resonance at ω_0 with a finite width of Δω. Using Eq. (<ref>) and Eq. (<ref>) in Eq. (<ref>), we expect multiple dips at times T_k(N)=(2k+1)Nπ/ω_0 for k=0,1,2,…. This makes it very difficult to determine from the experimental data whether the dips result from different resonances or from a single resonance. In particular, we could assume that the two dips in Fig. <ref> for N=8 found at T_a and T_b with T_a≈120 and T_b≈200 originate from the same resonance ω_0. In addition, if they correspond to neighboring dips, we can assume T_b/T_a=(2k+3)/(2k+1) which can be compared to the data, T_b/T_a≈1.67. This suggests that k=1 is a good fit with T_b=T_2(N=8) and T_a=T_1(N=8). From this, we can estimate the frequency which is approximately ω_0≈ 2π×100. In order to identify dominant contributions to the experimentally determined noise spectrum S(ω), we investigate common effects, such as intensity fluctuations of the dipole trap laser and magnetic field fluctuations, that have already been seen to cause homogeneous dephasing in ultracold atomic systems <cit.>. The intensity fluctuations are measured by shining the trapping light of the horizontal dipole trap onto a fast photodiode. We record the photodiode voltage as a function of time and calculate the power spectral density (PSD) S_DT of our discrete-time signal. In Fig. <ref>(a), S_DT is visible as the red curve. The resulting PSD S_DT comprises two noise components S_DT (ω) = S_PLNω^α̃+ S_WN, where the first component describes power-law noise with a characteristic, positive exponent α̃ and the second component describes white noise with a constant power spectral density S_WN. We use Eq. (<ref>) as a fit model for our data in Fig. <ref>(a) and find an exponent α̃ = 0.8904(2). This fit is visible as black solid line in Fig. <ref>(a) and agrees very well with S_DT. Importantly, the fitted exponent coincides with α = 0.89(2) obtained from the measured visibility evolutions. This suggests that the dominant contribution to the dephasing of the qubit originates from noise in the laser intensity which generates the dipole trap. Moreover, previously, we have estimated that some of the dips in the visibility data originate from a resonance of S(ω) at ω_0≈ 2π×100. Remarkably, a resonance at ω_0 is found in S_DT and can therefore explain such features in the visibility. To quantify the magnetic field fluctuations, we cannot directly measure the fluctuations outside of the vacuum system because they differ from the fluctuations at the atoms' position. Instead, we record the current through the magnetic field coils as a function of time and calculate the PSD of the signal. The resulting PSD, S_B, is shown as brown curve in Fig. <ref>(b). Fundamentally different from S_DT, the PSD S_B is rather flat with no clear decay. Such a noise spectrum would result in a very different evolution of the visibility. Therefore, we conclude that the background noise floor inferred from the measured coherence evolution is mainly caused by intensity fluctuations of the trapping light and not by magnetic field fluctuations. This finding highlights the magnetic-insensitivity of the two Cs clock states which have been used as stable qubit states. § CONCLUSION AND OUTLOOK In conclusion, we have studied the dynamics of a noise-coupled qubit realized in ^133Cs under the effect of dynamical driving. Hereby, we have focused on a dynamical decoupling protocol, in particular the CPMG pulse sequence, that has increased the coherence time of the qubit by an order of magnitude. Moreover, from the visibility data, we were able to analyze properties of the environmental noise spectrum including its algebraic decay. From this, we have calculated analytical estimates for the coherence time which are in good agreement with the measured data. To find the origin of the noise affecting the qubit dynamics, we have measured the PSDs determining the noise spectra of the dipole trap laser intensity and the magnetic field. With these, we could show that the main noise source originates from the trapping laser intensity while magnetic field fluctuations play only a minor role. The latter originates from the magnetic-insensitivity of the used qubit states. Remarkably, the algebraic decay obtained from the visibility data is in excellent agreement with the algebraic decay found in the PSD of the dipole trap laser intensity. This highlights the ability to measure spectral properties of the environment with a dynamical probe. In future work, we will investigate the dynamics of the ^133Cs qubit when it is coupled to a bosonic environment of ^87Rb atoms. The qubit dynamics could then be analyzed to potentially measure properties of an interacting environment. This is the next step towards the implementation of a single ^133Cs qubit as a versatile, nondestructive quantum probe for noise spectroscopy of engineered atomic baths. An important prerequisite for this is that the lifetime of the qubit without the bosonic bath exceeds the lifetime in the presence of the ^87Rb atoms. By comparing typical lifetimes obtained for the mixture <cit.>, τ_c,Rb≈1, which are one order of magnitude shorter than the shortest lifetime obtained in this paper, τ_c = 16.2(9), we conclude that this requirement is fulfilled. This will enable the distinction of dynamical features emerging from, e.g., noise in the dipole trap laser intensity and features emerging from the bosonic ^87Rb bath. Besides its potential as a quantum sensor, this setup can also be used to explore the entanglement dynamics of the Cs qubit with the surrounding ^87Rb bath under dynamical driving. This creates the possibility to study non-equilibrium phenomena in quantum many-body systems closely connected to spin-boson and polaron physics. Additionally, it will be interesting to investigate the possibility of tailoring the qubit-bath coupling with periodic driving which opens the door to quantum-state engineering in open quantum systems. This work was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) via the Collaborative Research Centers SFB/TR185 (Project No. 277625399). S.B. acknowledges funding by Studienstiftung des deutschen Volkes. apsrev4-2 § MAGNETIC FIELD CALIBRATION We employ microwave spectroscopy on the |F=1,m_F=0⟩→|F=2,m_F=1⟩ transition of the ^87Rb ground state to precisely calibrate the magnetic field B_exp at the atoms' positions by compensating for the contributions B_bg of ambient fields. The Rb atoms are trapped in the second crossed optical dipole trap and are initially prepared in the |F=1, m_F=0⟩ state. The microwave frequency is fixed to the expected transition frequency given by the Breit-Rabi formula for the desired field magnitude B_exp=198.5, neglecting the small differential light shift δ_DT = 39 of the trapping field. We vary the externally applied magnetic field B_coil, and thereby also the transition frequency of the Rb atoms, and probe the excited state population by standard absorption imaging using a time-of-flight measurement [Fig. <ref>(a)]. The microwave pulse duration τ = 243.9 is chosen such that full population transfer occurs as soon as the resonance condition B_coil+B_bg=B_exp is fulfilled. The signal follows the typical shape of the spectrum of a Fourier-limited square pulse N_Rb(B_coil) = A Ω_R^2/Ω^2sin^2 (Ωτ/2) + C, where Ω^2 = Δ^2 + Ω_R^2, Δ = 2 π× (B_coil - B_coil,res) ×0.7 in the limit of small magnetic fields and A and C are the amplitude and the offset of the fit, respectively. Notably, the resonant Rabi frequency Ω_R = 2 π×2.180(5) is measured independently [Fig. <ref>(b)] so that A, C, and B_coil,res are the only free parameters of the fit. § NORMALIZATION OF THE RAMSEY FRINGES The Ramsey fringes with total Cs atom number [Fig. <ref>(a)] need to be normalized to calculate the fringe visibility according to Eq. (<ref>). We rescale the data for a given number N of π-pulses and pulse sequence length T according to P(T, φ) = N_Cs(T, φ)/N_Cs, tot(T), where the total number of Cs atoms contributing to the signal N_Cs, tot(T) depends on T due to the finite lifetime of the atoms in the optical lattice. For each data set with a fixed number N of π-pulses, the total atom number N_Cs, tot(T) is deduced from two additional measurements. First, we extract the maximal number of Cs atoms contributing to the signal N_Cs,0 from a Ramsey fringe with N=0 π-pulses and shortest possible pulse sequence length T=27 [Fig. <ref>(b)]. Second, we measure the lifetime τ_LT of the Cs atoms in the optical lattice by monitoring the Cs atom number as a function of the pulse sequence length T without applying microwave radiation [Fig. <ref>(c)]. We find typical lifetimes between τ_LT = 539(37) and τ_LT = 831(72). The total atom number, which sets the scaling factor for the normalization, is then given by N_Cs, tot(T) = N_Cs,0exp(-T / τ_LT).

http://arxiv.org/abs/2303.07136v1

20230313140507

Comparison of confidence regions for quantum state tomography

quant-ph

ICFO - Institut de Ciències Fotòniques, The Barcelona Institute of Science and Technology, 08860 Castelldefels (Barcelona), Spain. Naturwissenschaftlich-Technische Fakultät, Universität Siegen, Walter-Flex-Straße 3, 57068 Siegen, Germany Física Teórica: Informació i Fenòmens Quántics, Departament de Fisica, Universitat Autònoma de Barcelona, E-08193, Bellaterra (Barcelona), Spain The quantum state associated to an unknown experimental preparation procedure can be determined by performing quantum state tomography. If the statistical uncertainty in the data dominates over other experimental errors, then a tomographic reconstruction procedure must produce a statistical confidence region in state space rather than a single state. Naturally, the size of this region decreases when increasing the number of samples, but it also depends critically on the construction method of the region. We compare recent methods for constructing confidence regions as well as a reference method based on a Gaussian approximation. For the comparison, we propose an operational measure with the finding, that there is a significant difference between methods, but which method is preferable can depend on the details of the state preparation scenario. Comparison of confidence regions for quantum state tomography Gael Sentís March 30, 2023 ============================================================= § INTRODUCTION Quantum state tomography is a complete and agnostic procedure to characterize an unknown quantum state from experimental data <cit.>. As such, it plays a central role as a quality control subroutine in quantum experiments or quantum technological applications, allowing us to assess the outcome of a state-preparation procedure and verify that a targeted quantum state is indeed being produced. The outcome of a tomographic procedure is usually a point estimate ρ̂ of the true density matrix ρ, along with some sort of error bar. A common approach is to use maximum likelihood estimation <cit.> to construct ρ̂, and report the variance of the point estimate obtained by bootstrapping <cit.>. However, this approach accounts only for the statistical fluctuations of the maximum likelihood point estimate, which can turn out to be smaller than its actual bias <cit.>. More importantly, such error bars tell us nothing about how far off is ρ̂ from the true state ρ. To answer this question rigorously in statistical terms, confidence regions (CRs) are constructed: these are regions in the space of Hermitian operators that intersect the state space and contain the true state with high probability. Generally speaking, we can regard a confidence region C as a function from the experimental data to the space of Hermitian operators fulfilling [ ρ∈ C(data)|data∼ρ ] ≥ 1-δ for every state ρ, where 1-δ is the confidence level. There exist diverse methods to construct CRs from tomographic data in the literature <cit.>. Methods may differ on the theoretical grounds they are built on, in turn leading to differences in their tightness, speed of convergence, and computational cost. However, these features are not always conveyed in a way that allows for a direct comparison between different methods, particularly in the practically relevant regime of finite data. As a result, the natural question of which CR should one choose to report the results of a tomographic experiment has, to this day, no straightforward answer. The main objective of this work is to close this gap by developing an operational benchmark under which methods to build CRs can be fairly contrasted. Generically, for the same dataset and confidence level, the smaller a CR is, the better. To illustrate why this simple criterion is problematic to evaluate in practice, let us have a first look at two recent results in Refs. <cit.>, which we will further analyze in this work. A usual way of expressing Eq. (<ref>) (see, for example, Refs. <cit.>) is [ ⋆ρ̂-ρ≤ϵ ] ≥ f(ϵ,d,N), where d is the dimension of the quantum system, N is the number of tomographic experiments, and the form of f varies with the norm ⋆ (usual examples are trace, Hilbert-Schmidt, or operator norm) and with the estimator used for ρ̂. In terms of speed of convergence with the number of samples taken, Guţă et al. <cit.> report an almost optimal scaling of f with respect to N, d, but fixed ϵ (showing an improvement over Ref. <cit.> for the same norms). In contrast, Wang et al. <cit.> provide an explicit construction of the function C as a polytope, and the convergence of the region with N is implicit, preventing a direct comparison to Eq. (<ref>) based only on the region size. For finite N, neither of these works consider the tightness of the CRs quantitatively. Then the question arises, how can we judge whether one of these CRs is smaller than the other in a given experiment? First, the regions have incomparable shapes (a sphere in a norm and a polytope), hence a parametric comparison is not possible. One could look at volumes to disregard shape, but the relevant volume to consider would be that of the intersection of the CRs with the state space, hence we would require numerical estimates (for example, by Monte Carlo sampling) that are not scalable beyond relatively small systems and will be in any case dependent on the chosen integration measure. Finally, we would encounter the same problems when comparing CRs of the form in Eq. (<ref>) if these are expressed in different norms. In this case, we could use inequalities between norms to transform one into another, but we would always be underestimating or overestimating one of the regions. In this work we propose to assess the power of a CR by how well can it distinguish a pair of states. More precisely, we simulate a tomography experiment on two given target states, and find, for a given confidence level, the minimum number of repetitions necessary for their corresponding CRs (built with the same method) to stop intersecting, which ensures that the two states can be distinguished with the chosen confidence. While, clearly, this is not the most efficient test to discriminate the two states, as a test it has several desirable features: It has an operational meaning, it is agnostic to the specifics of a CR, and it is computationally efficient. We use this test to carry out a comparative analysis of recent CRs in paradigmatic scenarios involving pairs of states of up to 4 qubits, and we focus on performing local Pauli-basis measurements as our tomographically complete set of measurements for its practical relevance and widespread use. Our results show that there is no single CR that outperforms the others in all cases. The performance of the CRs varies not only with the dimension of the system, but also with the pair of states selected for the distinguishability test. An example of this is the finding that while a CR may allow to distinguish a pure state from the completely mixed state with the fewest number of samples, it underperforms when distinguishing two pure states that are very close to each other. This notwithstanding, our analysis allows us to make some general recommendations. We apply our comparative analysis to the recent results in Ref. <cit.> and Ref. <cit.>. In addition, we provide a method to build CRs based on a Gaussian approximation, which will serve us as a reference. While there are other works on CRs in the literature, the rationale behind focusing on these ones is the following. First, we concentrate on unconditional tomography, leaving aside Bayesian methods that require assuming a prior distribution over the possible states (and result in credibility regions instead of CRs) <cit.>, methods that assume rank-deficient states like compressed sensing <cit.>, permutationally invariant states <cit.>, or matrix product states <cit.>, and heuristic methods like self-guided tomography <cit.> or CRs based on likelihood ratios <cit.>. Second, we select those works which have not yet been clearly superseded (for example, Ref. <cit.> improves on Ref. <cit.>, and Ref. <cit.> on Ref. <cit.>). Finally, it is worth mentioning with Ref. <cit.> another recent comparative study on quantum state tomography, which evaluates a different aspect in tomography schemes, namely the typical error (distance between the estimate and the true state) of various point estimators. The paper is structured as follows. In Section <ref> we summarize the CRs considered in our study. Section <ref> details the methodology used for comparing between CRs in a series of simulated tomography experiments, and includes the results of our testing. We end the paper with a short discussion in Section <ref>. In the Appendix, we include details on the derivation of the Gaussian CRs as well as technical considerations about the simulations. § METHODS FOR BUILDING CONFIDENCE REGIONS Here we summarize the results that we use for our comparative analysis. A common feature of all selected methods is that the CRs are built for the free least-squares estimator for ρ̂, see Appendix <ref>. Importantly, this estimator is unbiased, although due to finite statistics in experiments it generally produces a non-physical density matrix. This is not necessarily problematic, in the sense that the associated CR will almost certainly have some overlap with the state space. CRs for physical (hence biased) estimators exist, for example, for projected least-squares <cit.> or other variants <cit.>, but they are derived from a CR for the free least squares estimator and thus do not improve over them. Hence, to treat all methods on equal footing, in this paper we restrict ourselves to free least squares. For easy referencing, we label the CRs used in our analysis as A, B, C_1, and C_2. The first result we review is a non-asymptotic concentration bound for the operator-norm error of the free least-squares estimator. In Ref. <cit.>, Guţǎ et al. show that the free least-squares estimator obeys Confidence region A [ ∞ρ̂-ρ≥ϵ ] ≤ R_d(ϵ√(N)) where ∞X is the operator norm of X, N is the total number of state preparations, and d is the dimension of the underlying Hilbert space. Hence, the CR is of the form in Eq. (<ref>). One has log R_d(ξ)=-38ξ^2 g(d)+log d, where the factor g(d) takes different forms depending on the chosen measurement scheme. For multiqubit states and local Pauli-basis measurements, we have g(d)= 3^-log_2 d. Another method by Wang et al. <cit.> constructs a polytope-shaped region Γ(f⃗) as a function of the empirical frequencies f⃗=(f_1,…,f_M) of a single positive operator-valued measure (POVM), that is, positive semidefinite operators {E_i}_i=1^M obeying ∑_i E_i =, applied on the state ρ, where f_i=n_i/N, n_i is the number of occurrences of outcome i, and N the total number of state preparations. The region Γ(f⃗) obeys Confidence region B [ ρ∉Γ(f⃗) ]≤δ. Each facet of the polytope is determined by the half-space Γ_i(f_i)=ρ | (E_i ρ)≤ f_i+ϵ_i , where ϵ_i is the positive root of D(f_i‖ f_i+ϵ_i)=-log(δ_i)/N, with ∑_i=1^M δ_i = δ, and D(x‖ y) = x log (x/y) + (1-x)log[(1-x)/(1-y)] is the relative binary entropy. Then, the CR is given by the intersection of all half-spaces, that is, Γ(f⃗)=⋂_i Γ(f_i). Because each half-space corresponds to an independent confidence interval for the frequency f_i, we are free to choose each individual confidence level by fixing δ_i as long as they add up to the total confidence parameter δ. This method thus allows us to shrink the CR along a preferred measurement direction at the expense of enlarging it in other ones, which adds versatility. For our comparative analysis, we make the uniform choice δ_i=δ/M. We also provide bounds based on a Gaussian approximation. Here we make the approximation that the empirical frequencies f⃗ of the measurement outcomes are Gaussian distributed with mean p⃗(ρ) and a nonsingular covariance matrix Σ(ρ). Then we obtain the Gaussian CR defined by Confidence region C_1 [ 2Σ(ρ)^-1/2(f⃗-p⃗(ρ)) > ϵ ]=P_k(ϵ^2), where P_k(x) is the survival function of the central χ^2-distribution with k=(Σ) degrees of freedom, see Appendix <ref> for details. Note that Σ(ρ) here scales with the number of samples. This CR is exact, but comes with the disadvantage that the region depends nonlinearly on ρ through Σ. We can improve on this at the cost of getting a more relaxed bound, see Appendix <ref>, Confidence region C_2 [ 2p⃗(ρ̂)-p⃗(ρ) >ϵ ]≤ P_d^2-1(σ^-2_maxϵ^2), with σ^2_max the largest eigenvalue of Σ(ρ) across all states. Note that the resulting CR forms an ellipsoid in the state space. For multinomial distributions, we have σ^2_max=1/(2n), where n is the number of samples per measurement setting. For example, for local Pauli-basis measurements on a q-qubit state, we have 3^q distinct measurement settings, N=3^q n (see Appendix <ref>). § METHODOLOGY FOR COMPARING CRS AND RESULTS We propose two universal tests to benchmark CRs irrespective of their specifications. The first test consists in computing the (1-δ)-quantile of the distribution of the ratio r=⋆…/ϵ(δ) over several repetitions of the same tomography experiment with confidence level (1-δ). The terms in r depend on the chosen CR, as they appear in [ ⋆… > ϵ(δ) ] ≤δ and note that the norm can be defined either in probability space or in the state space. This test quantifies the tightness of a CR in a scale-independent way. On the one hand, a value 100% for the (1-δ)-quantile of r indicates that the region is exact, inasmuch as the probability of obtaining ρ̂ at a distance no more than ϵ from ρ coincides with the confidence level, that is, the inequality in Eq. (<ref>) is saturated. On the other hand, a small value of r tells us that the CR is unnecessarily large, since in (1-δ)× 100% of the experiments ρ̂ is actually much closer to ρ than the size of the region ϵ. We run simulations of tomography experiments on Haar-random pure states of up to 4 qubits. We simulate using N=60 000 samples and local Pauli-basis measurements, see Appendix <ref>, and evaluate r for a confidence level of 1-δ=99%. For Confidence Region A, Confidence Region C_1 and Confidence Region C_2, and for each state considered, we determine the empirical (1-δ)-quantile from 10 000 samples of the ratio r; the results are summarized in Table <ref>. Note that, since Confidence Region B is not norm-based, we do not run this test on it. We observe that Confidence Region C_1 is an exact region, which supports the validity of the Gaussian sampling assumption in our numerical experiments. For Confidence Region A and C_2 we see that C_2 is tighter for the 1-qubit, 2-qubit and 3-qubit cases and then becomes looser for states of 4 qubits, whereas the tightness of Confidence Region A is more stable with an increasing number of qubits. As we see next, these results correlate well with those of our second test, even though they quantify different features of the CRs. The second test we devise is aimed at quantifying the distinguishability power of CRs. Choosing two distinct target states over which we perform tomography, we ask the question of how many samples N are required so that their respective CRs do not contain a common state. At this point we can distinguish the two states with confidence 1-δ based on the CRs, see Figure <ref>. In this way we can attach a number to a CR and a given pair of states which is indirectly related to the size of the region. This test thus allows us to rank CRs according to an operationally well defined task and independently of their specifications. We run this test with Confidence Region A, B, and C_2 for two types of pairs of states of up to 4 qubits: pairs comprised by a pure state and the completely mixed state, and pairs of two pure states relatively close to each other, with fidelity ⟨ϕ_1|ϕ_2|⟩^2= 0.63. The specific states are detailed in Appendix <ref>, and we fix the confidence level to 1-δ=90%. Importantly, the value of N at which the CRs stop intersecting is a stochastic variable, since it directly depends on the estimate ρ̂ where the regions are centered and thus on the data obtained in each tomography experiment. For increasing values of N, we look at the frequency of finding an intersection over many repetitions (3000 for q=1,2; 1000 for q=3 and 500 for q=4). We also note that testing whether two regions share a common state can be computed efficiently via a convex optimization program, see Appendix <ref>. This is the case with the exception of Confidence Region C_1 since this region depends nonlinearly on the true state through Σ(ρ)^-1/2, and for this reason we do not include it in this test. In Figure <ref> we show the frequency that the CRs of the two target states contain a common state, for each of the three CRs considered. Each plot corresponds to a different pair of target states of one and two qubits. For the case of one-qubit states, Confidence Region C_2 performs best in this test, closely followed by Confidence Region B, regardless of whether we try to distinguish a pure state versus the completely mixed state or two pure states close to each other, whereas Confidence Region A requires roughly a factor of two more state preparations. For two-qubit states we observe an interesting phenomenon: Confidence Region A distinguishes a Bell state from the identity with the least number of state preparations and Confidence Region C_2 has almost the same performance, but it becomes notably more inefficient when distinguishing between a Bell state and the same state with a slight local rotation applied (with fidelity 0.63), for which case Confidence Region C_2 is again the best one. An important remark is that the states selected for the test are not aligned with any measurement direction, in order to represent a generic case of state tomography. Particularly for one-qubit states, if the state would point in the x,y or z direction, we find that Confidence Region B slightly outperforms Confidence Region C_2. This is due to the facets of the polytope Γ(f⃗) being perpendicular to the measurement directions by design. We show the results of the intersection test for pairs of 3-qubit and 4-qubit states in Figure <ref>. Here we observe a clear advantage of Confidence Region A in all cases. This is consistent with the results of Ref. <cit.>. Indeed, the derivation of Confidence Region A is based on the observation that ρ̂ can be seen as a sum of N random matrices and the subsequent application of a matrix concentration inequality, which becomes sharper as N increases (which is necessary for a larger number of qubits). Also, the ranking between the analyzed CRs is maintained throughout the studied 3-qubit and 4-qubit cases, with Confidence Region C_2 in a second position and Confidence Region B requiring the largest number of state preparations. § DISCUSSION We have proposed and tested two methods to benchmark confidence regions for quantum state tomography. One of them, based on the distinguishability power of a given confidence region construction, is operationally defined and applicable to any such construction regardless of its shape, specifications, and whether the region is defined in probability space or in state space. In addition, we proposed a second method which determines how tight a given region is. These new metrics allow us to compare existing (and future) methods to build confidence regions that were until now not directly comparable. Thus, we provide a robust tool to help in deciding which confidence region to report in any state tomography experiment. Our numerical experiments show that this choice should depend on the dimension of the system at hand and the goal for which tomography is used. We see this variability as a reflection of the diverse techniques used to derive confidence regions. For general purposes, the method by Guţǎ et al. <cit.> appears to be a safe choice, particularly for large systems, while in special cases involving 1-qubit and 2-qubit states, one may consider using the method of Wang et al. <cit.> or a Gaussian approximation as in Confidence Region C_2. In addition, Confidence Region C_2 has also been found to be useful in deriving optimal tests for detecting properties of the measured state directly from the tomographic data, without going through the intermediate step of computing a point estimate of the density matrix. An example of this is a proposal for the experimental detection of bound entanglement <cit.>. With this work, we argue that unconditional full quantum state tomography is almost settled as a research topic. A point estimate of the unknown state obtained by linear inversion equipped with a confidence region is computationally inexpensive and should suffice for any purpose where tomography is needed. As for which confidence region is best to use, as we have seen, it depends. There may still be room for improvement in light of our (1-δ)-quantile tests, since, for example, for the method of Guţǎ et al. we observe it is only around 40%. This means there could be an even tighter construction, although this seemingly suboptimal slackness does not have a great effect on our intersection tests. We thank Richard Kueng, Maciej Lewenstein, and Michael Skotiniotis for the helpful discussions. This work was supported by ERC (Advanced Grant NOQIA and Consolidator Grant 683107/TempoQ); Ministerio de Ciencia y Innovation Agencia Estatal de Investigaciones (PGC2018-097027-B-I00/10.13039/501100011033, CEX2019-000910-S/10.13039/ 501100011033, Plan National FIDEUA PID2019-106901GB-I00, FPI, QUANTERA MAQS PCI2019-111828-2, QUANTERA DYNAMITE PCI2022-132919, Proyectos de I+D+I “Retos Colaboración” QUSPIN RTC2019-007196-7, PID2019-107609GB-I00); MICIIN with funding from European Union NextGenerationEU(PRTR-C17.I1) and by Generalitat de Catalunya; Fundació Cellex; Fundació Mir-Puig; Generalitat de Catalunya (European Social Fund FEDER and CERCA program, AGAUR Grant No. 2021 SGR 01452, QuantumCAT U16-011424, QuantumCAT 001-P-001644 co-funded by ERDF Operational Program of Catalonia 2014-2020); Barcelona Supercomputing Center MareNostrum (FI-2022-1-0042); EU Horizon 2020 FET-OPEN OPTOlogic (Grant No 899794); EU Horizon Europe Program (Grant Agreement 101080086 – NeQST), National Science Centre, Poland (Symfonia Grant No. 2016/20/W/ST4/00314); ICFO Internal “QuantumGaudi” project; European Union's Horizon 2020 research and innovation program under the Marie-Skłodowska-Curie grant agreement No 101029393 (STREDCH) and No 847648 (“La Caixa” Junior Leaders fellowships ID100010434: LCF/BQ/PI19/11690013, LCF/BQ/PI20/11760031, LCF/BQ/PR20/11770012, LCF/BQ/PR21/11840013); Deutsche Forschungsgemeinschaft (DFG, German Research Foundation, project numbers 447948357 and 440958198); Sino-German Center for Research Promotion (Project M-0294); German Ministry of Education and Research (Project QuKuK, BMBF Grant No. 16KIS1618K). Views and opinions expressed in this work are, however, those of the author(s) only and do not necessarily reflect those of the European Union, European Climate, Infrastructure and Environment Executive Agency (CINEA), nor any other granting authority. Neither the European Union nor any granting authority can be held responsible for them. § CONFIDENCE REGION FOR NORMAL-DISTRIBUTED TOMOGRAPHIC DATA Here we provide the proofs for Confidence Region C_1 and Confidence Region C_2. For that, we write f⃗ for the collection of the empirical frequencies and assume that f⃗ is normal distributed with mean p⃗(ρ) and covariance matrix Σ(ρ), where ρ is the state on which the tomography is performed. Statistically dependent entries are eliminated from p⃗ and f⃗—for example for a multinomial distribution, ∑_i p_i=1 and hence we remove one entry from p⃗. Then Σ(ρ) is nondegenerate, except for a subset of states which has measure zero, and hence is irrelevant. We observe that Δ⃗_ρ = Σ(ρ)^-1/2 [ p⃗(ρ)-f⃗ ] is normal distributed with mean 0⃗ and covariance matrix . Consequently 2Δ⃗_ρ^2 is distributed according to the central χ^2(k)-distribution with k=(Σ) degrees of freedom and survival function [ 2Δ⃗_ρ^2 > y ]=P_k(y). This already yields Confidence Region C_1 via [ 2Δ⃗_ρ > ϵ ] = P_k(ϵ^2). We mention that for a multinomial distribution and n samples one has (Δ⃗_ρ)^2= n ∑_i [ p_i(ρ)-f_i ]^2/p_i(ρ), where the sum includes the element that had been omitted from p⃗ and f⃗ in order to make Σ(ρ) nondegenerate. For Confidence Region C_2 we consider M(ρ-ρ')=p⃗(ρ)-p⃗(ρ') as linear map M on the space of zero-trace Hermitian matrices, which we consider as a real vector space of dimension d^2-1. We write M^+ for the pseudo-inverse of M and P is a map onto ℝ^d^2-1 with P=PMM^+ and M=P^TPM. Note that P can can be constructed by orthogonalizing the row vectors of M. We use now the free least-squares estimate ρ̂=_X2p⃗(X)-f⃗^2 | X=X^† and (X)=1 where the result is not required to be positive semidefinite. We have ρ̂=M^+f⃗ and ρ=M^+p⃗(ρ). It follows that PM(ρ̂-ρ)=PMM^+[f⃗-p⃗(ρ)]=P[f⃗-p⃗(ρ)] has mean 0⃗ and covariance Σ'(ρ)=PΣ(ρ)P^T. Note that Σ'(ρ) is nonsingular (even if Σ(ρ) is degenerate) because M is injective and has dimensions (d^2-1)×(d^2-1). Therefore we have [ 2Σ'(ρ)^-1/2P(p⃗(ρ̂)-p⃗(ρ)) >ϵ ] = P_d^2-1(ϵ^2) The main advantage of both confidence regions derived so far is that they do not contain any approximations up to the assumption of a normal distribution. Hence they are optimal in this respect. However, they do not have a simple shape because they depend in a nonlinear way on ρ via Σ(ρ). For this reason we replace Σ(ρ) by σ_max^2 with σ^2_max the largest eigenvalue of the covariance matrix over all states ρ. Using P^TP[p⃗(ρ̂)-p⃗(ρ)]= p⃗(ρ̂)-p⃗(ρ), this gives us Confidence Region C_2 via [ 2p⃗(ρ̂)-p⃗(ρ) >ϵ ]≤ P_d^2-1(σ^-2_maxϵ^2). We mention that the spectral decomposition of ϵ (M^TM)^-1/2 yields the semiaxes of the confidence ellipsoid in the state space. We can relax Confidence Region C_2 even further. For that we let μ be the smallest (nonzero) singular value of M. Since M is injective, it follows that 2M(ρ-ρ̂)≥μ2ρ̂-ρ. This gives us an additional CR defined by [ 2ρ-ρ̂> ϵ ] ≤ P_d^2-1(μ^2 σ_max^-2ϵ^2). Although this bound has a very similar structure to Confidence Region A, it cannot be expected to yield a good CR. This can be understood geometrically, because we approximated the previous CR, which gave us an ellipsoid, by simply fitting a sphere around it with the radius of the largest half-axis. § LOCAL PAULI-BASIS MEASUREMENTS FOR QUANTUM STATE TOMOGRAPHY In this appendix we provide specifications for computing the CRs analyzed in the main text when choosing local Pauli-basis measurements as a tomographically complete set of measurements. Generically, a measurement on a d-dimensional quantum state ρ can be described by a POVM, that is, as M positive semi-definite matrices E_1,…,E_M, each with dimension d× d, such that ∑_i E_i=. Each measurement of ρ results in a corresponding outcome and the probability of observing outcome i=1,…,M is given by p_i=( E_i ρ ). This probability can be estimated based on the outcome frequency when measuring many copies of ρ. The frequency corresponding to outcome i after n repetitions of the measurement E_1,…,E_M is f_i= n_i/n for i=1,…,M , and we have f_i→ p_i in the limit n→∞. §.§ Local Pauli-basis measurements We now particularize this generic POVM for state tomography with local Pauli-basis measurements. For one qubit, measuring in the three Pauli bases can be viewed as a single six-outcome POVM, whose elements correspond to the eigenstates of each Pauli matrix σ_j, j=x,y,z. Each POVM element has two labels, the direction j and the two possible outcomes along that direction, +1 and -1. Therefore, we write E_j,±=(±σ_j)/2=w_j,±. When each measurement σ_j is performed n times, the frequency vector corresponding to the six-outcome POVM has components f_j,±=n_j,±/(3n). Generalizing the tomography scheme for q qubits, the dimension of the system is d=2^q, and there are 3^q Pauli measurement settings m⃗=(m_1,…,m_q)∈x,y,z^q, each with d possible outcomes o⃗ = (o_1,…,o_q), where o_j∈± 1. This tomography scheme can also be written as a single POVM, whose elements are now specified by the measurement setting m⃗ and the outcome o⃗, as E_m⃗,o⃗ = 3^-q⊗_j=1^q w_m_j,o_j. If each measurement setting m⃗ is performed n times, we obtain for the single POVM the frequencies f_m⃗,o⃗ = n_m⃗,o⃗/(3^q n) where the total number of samples is N=3^q n. At this point we note that, formally, there is no difference between measuring each combination of Pauli matrices m⃗ separately as 2^q-outcome projective measurement and considering the single POVM E_m⃗,o⃗ with 6^q outcomes. However, in the finite data regime there is a potential difference when it comes to the covariance matrix of the data. Indeed, when considering a single POVM there is only one frequency that is dependent of the others, by the normalization ∑_m⃗,o⃗ f_m⃗,o⃗=1. If, instead, we consider applying the Pauli measurements specified by each m⃗ independently and aggregating the data at a later stage, we have now 3^q dependent frequencies, since now ∑_o⃗ f_m⃗,o⃗ = 1. This is relevant when computing the covariance matrix Σ(ρ) that appears in Confidence Region C_1: if we do not eliminate frequencies corresponding to the trivial outcomes from f⃗, Σ(ρ) will not have a well defined inverse. Moreover, as we saw in Appendix <ref>, discarding these outcomes results in a tighter CR. When detailing how to compute Confidence Region A and Confidence Region B for local Pauli-basis measurements, we use the single-POVM formulation to follow the original references <cit.> and <cit.> closer. We switch to the individual measurement settings m⃗ when discussing our bound Confidence Region C_2 (<ref>). §.§ Adaptations for the Confidence Regions Generalized Bloch representation. It is convenient for our computations to write the density matrix in the generalized Bloch representation. For the tuple of 4^q matrices μ⃗=2^-q/2(,σ_x,σ_y,σ_z)^⊗ q, any Hermitian matrix can be written as X = x⃗·μ⃗ with x⃗∈ℝ^4^q . Due to the choice of the normalization, we have (XY)=x⃗·y⃗ for X=x⃗·μ⃗ and Y=y⃗·μ⃗. Free last-squares estimator. The free least-squares estimator in state tomography is the solution to the least-squares problem that seeks to minimize the distance between the frequencies and the probabilities, see also Eq. (<ref>), ρ̂=_X∑_i ( f_i- ( E_i X ))^2 | X=X^† and (X)=1 . Note that the optimization is performed over all Hermitian matrices with unit trace and hence the resulting estimate ρ̂, in general, will fail to be positive semidefinite. For local Pauli-basis measurements, where f⃗ has 6^q entries, this problem has an analytical solution <cit.>. In our definitions, the frequencies have the form f_m⃗,o⃗= n_m⃗, o⃗/(3^q n), and the free least-squares estimate is given by ρ̂=∑_m⃗,o⃗f_m⃗,o⃗⊗_j=1^q (3w_m_j,o_j -). We use this explicit solution when evaluating Confidence Region A. Confidence Region C_1 and C_2. As indicated in Appendix <ref>, when considering Confidence Region C_1 and Confidence Region C_2, we assume that each measurement setting m⃗ is measured independently n times, yielding frequencies f_m⃗,o⃗=n_m⃗,o⃗/n with ∑_o⃗ f_m⃗,o⃗ = 1. Therefore, there is one statistically dependent frequency per measurement setting m⃗, which we can compute from the remaining frequencies of the same setting. We thus omit the 3^q dependent frequencies to avoid having a singular covariance matrix, which reduces the dimension of the data f⃗ to 3^q(2^q-1) components. § INTERSECTION OF CONFIDENCE REGIONS VIA CONVEX OPTIMIZATION Given two regions of the types A, B, or C_2, we can determine whether they have a non-empty intersection via convex optimization, see, for example, Ref. <cit.>. In our numerical experiments we use the Splitting Conic Solver SCS <cit.> for convex-cone optimization with the Python programming language. Intersection for Confidence Region A. We have two CRs, R_1 and R_2, for two tomography experiments on the states ρ_1 and ρ_2, and corresponding free least-squares estimates ρ̂_1 and ρ̂_2. Confidence Region A for confidence level 1-δ is given by all states ρ obeying ∞ρ-ρ̂_i≤ϵ(δ), with ϵ(δ)=N^-1/2R_2^q^-1(δ)=√(3^q log(2^q/δ) 8/(3N)) and N the total number of samples. We can run the following semidefinite program to test for the existence of a state ρ∈ R_1 ∩ R_2: find ρ such that ρ≥ 0, (ρ)=1, -ϵ(δ) ≤ρ-ρ̂_i ≤ϵ(δ) for i=1,2. If the above program does not find a feasible solution, we conclude that the intersection R_1∩ R_2 is either empty or does not contain physical states. Intersection for Confidence Region C_2. For Confidence Region C_2 we proceed analogously, where here the region with confidence level 1-δ is given by all states ρ satisfying 2p⃗(ρ)-p⃗(ρ̂)^2 ≤ϵ^2(δ), with ϵ(δ)^2 = P_d^2-1^-1(1-δ)/(2n) and n the number of samples per measurement setting, hence N=3^qn. A state ρ in the intersection can be found by the quadratic convex program find ρ such that ρ≥ 0, (ρ)=1, ∑_j=1^3^q(2^q-1)((E_jρ)-(E_jρ̂_i))^2≤ϵ_i^2(δ) for i=1,2. Intersection of Confidence Region B. For regions of type B, we need to find a state in the intersection Γ^(1)∩Γ^(2) of the two confidence polytopes. The facets of either polytope Γ^(i), i=1,2, are given by the linear inequalities (E_jρ)≤ f_j^(i)+ϵ^(i)_j(δ), for j=1,… 6^q, where ϵ_j^(i) is determined by the root of the equation D(f_j^(i)‖ f_j^(i)+ϵ_j^(i) )=-1/Nlog (δ_j^(i)), N is the total number of samples, and we make the uniform choice δ_j^(i)=6^-qδ. Finding a state ρ∈Γ^(1)∩Γ^(2) hence corresponds to the semidefinite program find ρ such that ρ≥ 0, (ρ)=1, (E_jρ) ≤ f_j^(i)+ϵ_j^(i)(δ) for i=1,2 and j=1,…,6^q. § TARGET QUANTUM STATES The target quantum states used for evaluating the empirical (1-δ)-quantiles of different CRs in Table <ref> are uniformly random pure states sampled according to the Haar measure. For the tests of intersecting CRs in Figs. <ref> and <ref>, we choose pairs of target states ρ^(1), ρ^(2). The first state is always chosen as ρ^(1) = Ψ_ target, where |Ψ_ target⟩ is obtained by applying uniformly random rigid local unitaries on exemplary pure states of up to 4 qubits, i.e., |Ψ_ target⟩ = U^⊗ q|ψ_q⟩ , where q=1,2,3,4 is the number of qubits of |ψ_q⟩, U∼μ_ Haar(SU(2)), and |ψ_1⟩ = 1/√(2)(|0⟩+|1⟩) , |ψ_2⟩ = |Φ^+⟩ = 1/√(2)(|00⟩+|11⟩) , |ψ_3⟩ = | GHZ⟩ = 1/√(2)(|000⟩+|111⟩) , |ψ_4⟩ = 1/2( |0000⟩+|0011⟩+|1100⟩-|1111⟩) . Then, we distinguish two cases: the second state is either the maximally mixed state ρ^(2) = 2^-q, or another pure state defined as ρ^(2) = U_z(θ)^⊗ qρ^(1) (U_z(θ)^⊗ q)^†, where U_z(θ) = exp(i θσ_z), and θ is chosen such that (ρ^(1)ρ^(2)) = 0.63.

http://arxiv.org/abs/2303.06863v1

20230313051531

Semantically Secure Private Set Intersection over Outsourced Multi-Owner Secret-Shared Databases

cs.CR

University of Nevada, Reno United States dzhao@unr.edu Private set intersection (PSI) aims to allow users to find out the commonly shared items among the users without revealing other membership information. The most recently proposed approach to PSI in the database community was Prism, which is built upon secret sharing and the assumption that multiple non-colluding servers are available. One limitation of Prism lies in its semantic security: the encoding on the servers is deterministic, implying that the scheme cannot be indistinguishable under a chosen-plaintext attack (IND-CPA). This paper extends the original PSI scheme of Prism by two orthogonal primitives, namely Kaleido-RND and Kaleido-AES: the former exhibits highly efficient performance with randomized encoding and the latter is provably secure under CPA attacks with more computational overhead. A system prototype is implemented and deployed on a 34-node cluster of SQLite instances. Extensive experiments on the TPC-H benchmark and three real-world applications confirm the effectiveness of the proposed Kaleido primitives. printfolios=true Semantically Secure Private Set Intersection over Outsourced Multi-Owner Secret-Shared Databases Dongfang Zhao Received: date / Accepted: date ================================================================================================ § INTRODUCTION Private set intersection (PSI) aims to allow users to find out the commonly shared items among the users without revealing other membership information. As can be used as a versatile cryptographic primitive, PSI has been actively studied in the literature of applied cryptography <cit.> and recently in privacy-preserving outsourced databases <cit.>. As a concrete example, PSI has been recognized as a crucial building block to support various operations in outsourced databases, such as semi-joins <cit.>. The most recently proposed approach to PSI in the database community was Prism <cit.>, which is built upon secret sharing and the assumption that multiple non-colluding servers are available. Secret sharing <cit.>, as the name suggests, allows users to split a plaintext into multiple portions such that a single portion (stored on a distinct server) does not reveal any information about the plaintext. and is usually considered as a specific form of secure multiparty computation (MPC) <cit.>. Prism demonstrates a high efficiency that is based on group-homomorphic exponentiation. One limitation of Prism <cit.> lies in its semantic security: Although the exact number of shared entities is masked through number-theoretical modular exponentiation, such encoding is deterministic, implying that the scheme cannot be indistinguishable under a chosen-plaintext attack (IND-CPA). The first goal of this paper is, therefore, to describe an attack that can reveal information more than the intersection among parties. To guarantee the semantic security of Prism, the second goal of this paper is to extend the original PSI scheme of Prism by two orthogonal primitives: (i) Kaleido-RND, an efficient module to randomize the group generator based on which the exponentiation is computed in the codomain of a multiplicative group, and (ii) Kaleido-AES, a pseudorandom function to ensure the IND-CPA security of the entire PSI scheme. As the name suggests, our proposed schemes aim to garble the original Prism scheme, ending up a scattered prism—looking like a kaleidoscope. While Kaleido-RND is highly efficient and does offer randomization on the ciphertext, Kaleido-RND is not proven IND-CPA secure. However, we will demonstrate that Kaleido-AES is provably semantically secure with the cost of computational overhead. To demonstrate the effectiveness of the proposed Kaleido schemes, we implement a system prototype with about 1,000 lines of Python and Shell script and deploy the system on a 34-node cluster of SQLite instances on CloudLab <cit.>. We evaluate Prism, Kaleido-RND, and Kaleido-AES with the TPC-H benchmark and three real-world applications, where we scale the number of database owners (i.e., clients) between 2 and 32. Our results show that Kaleido-RND only incurs insignificant overhead on the server encoding and is generally negligible in the entire system. Moreover, although Kaleido-AES seems to introduce much computational overhead compared to the original Prism, the overhead is orders of magnitude lower than the communication cost, demonstrating the practicality of Kaleido-AES in addition to its provable security. In summary, this paper makes the following technical contributions. * We demonstrate a security vulnerability of a state-of-the-art PSI scheme, Prism, in outsourced databases. We provide both the intuition behind an effective attack and detail such an attack with concrete examples. (<ref>) * We propose a series of new primitives, namely Kaleido, to extend Prism such that the new PSI scheme is semantically secure. We prove the IND-CPA security of Kaleido under the widely-accepted computational assumption and exemplify its correctness with both theoretical analysis and case studies. (<ref>) * We implement the Kaleido scheme with about 1,000 lines of Python and Shell scripts. We deploy the system on two servers and 32 database owners hosted in the public cloud, CloudLab <cit.>. Extensive experiments on the TPC-H benchmark and three real-world applications confirm the effectiveness of the proposed Kaleido schemes. (<ref>) § PRELIMINARY §.§ Private Set Intersection and Secure Multi-Party Computation Private set intersection (PSI) <cit.> has been extensively studied in outsourced databases, such as being used as a building block for privacy-preserving joins between outsourced databases <cit.>. Indeed, PSI can be considered a special form of set operations among multiple parties, and therefore can be implemented through a general-purpose secure multi-party computation scheme, which mostly reply on arithmetic or boolean gates that incur significant computation overhead. Secure multiparty computation (MPC) <cit.> has a long history <cit.>. The goal of MPC is more ambitious: in addition to keeping the plaintext confidential, we want to calculate an arbitrary function of the original plaintexts by touching on only the encoded data on multiple parties. The original problem was solved by the so-called garbled circuits <cit.>, whose idea was pretty simple: we can ask each party to encode the input with its private key, shuffle the encrypted ciphertexts, and then enumerate all the keys to decrypt the result. Since we assume the encryption scheme is secure, the only way that the result can be revealed is that the correct combination of private keys is applied to one of the garbled outputs. This is indeed a feasible solution, at least theoretically; in practice, the circuits may grow exponentially and result in efficiency issues. There are many more efficient MPC solutions, such as <cit.>. One of the main limitations of PSI schemes lies in their scalability. In fact, many existing PSI schemes support only two parties, such as <cit.>. More recent works <cit.> emerged to focus efficient PSI schemes on more than two parties. This work falls into the category of the latter: The proposed Kaleido scheme supports an arbitrary of parties for the PSI operation. §.§ Secret Sharing The idea of a secret sharing scheme (SSS) is straightforward: a given plaintext pt is converted into a set of encoded bytes ct's such that only a specific subset of ct's can reconstruct the original pt. The goal of SSS is to reduce the risk of disclosing the plaintext; instead of compromising the holder of the plaintext, the malicious adversary needs to subvert multiple entities before any of the shareholders detect the attack. Even for weaker attacks where only semi-honest adversaries are assumed, dispersing the secret shares to more parties raises the bar of a successful eavesdropping attack. In practice, a SSS can be tuned by the subset size. Formally, a (t,n)-threshold SSS (TSSS) is defined as follows. A (t,n)-TSSS is comprised of two algorithms: * Share: a randomized algorithm that takes as input a plaintext pt and returns a sequence S = (s_1, …, s_n) of shares. * Reconstruct: a deterministic algorithm that takes a set of at least t shares and returns the plaintext. The number t is called the threshold of the TSSS. Let U of size t be a subset of n shares, |U| ≥ t and U ⊆{s_1, …, s_n}, we require that a TSSS holds the following property: Reconstruct(U) = pt. As we will see in the next section <ref>, the definition of TSSS leads to a slightly different security definition compared with the conventional encryption schemes. The canonical example of (t,n)-TSSS is due to Shamir <cit.>, in which the secrets were revealed through a (t-1)-degree polynomial. In essence, each share can reconstruct the coefficient of a specific degree of unknowns through the LaGrange polynomials. In addition to Shamir's construction, other schemes exist. Ito et al. <cit.> proposed the replicated secret-sharing scheme, which was based on finite fields where each share is a vector. One nice property of replicated secret-sharing is its linearity: the addition and subtraction of local shares can be linearly transformed into the addition and subtraction of the plaintexts. A simpler variant of replicated secret-sharing is additive secret sharing, where each share is a scalar value and the threshold t is set to n. §.§ Probably Security When employing an encryption scheme in an application, it is highly desirable to demonstrate its security provably. Formally, we need to identify the following three important pieces for the provable security of a given encryption scheme: security goal, threat model, and assumption. The security goal spells out the desired effect when the application is under attack; the threat model articulates what an adversary can do with the attack, such as what information of the plaintext/ciphertext can be collected and the resource/time limitation of the attack; the assumption lists the presumed conditions of the cryptographic scheme (e.g., factoring a product of two big primes, finding the discrete logarithmic root). The security goal and threat model are usually called security definition collectively. One well-accepted security definition with a good balance between efficiency and security is that the adversary can launch a chosen-plaintext attack (CPA), defined as follows. Given a security parameter n, i.e., the bitstring length of the key, an adversary can obtain up to poly(n) of plaintext-ciphertext pairs (m, c), where m is arbitrarily chosen by the adversary and poly(·) is a polynomial function in n. With such information, the adversary tries to decrypt a c' that is not included in the polynomial number of known ciphertexts. The polynomial requirement mandates that the adversary should only be able to run a polynomial algorithm without unlimited resources. Accordingly, we want to design encryption schemes that are CPA secure: even if the adversary 𝒜 can obtain those extra pieces of information, 𝒜 should not be able to decode the ciphertext better than a random guess up to a very small probability. To quantify the degree of this small probability, negligible function is defined as below. A function μ(·) is called negligible if for all polynomials poly(n) the inequality μ(n) < 1/poly(n) holds for sufficiently large n's. For completeness, we list the following lemmas for negligible functions that will be used in later sections. We state them without proof, which can be found in introductory cryptography or complexity theory texts. Let μ_1(n) and μ_2(n) be both negligible functions. Then the following functions are all negligible: μ_1(n) + μ_2(n), μ_1(n) - μ_2(n), μ_1(n) ×μ_2(n), and μ_1(n)/μ_2(n). The canonical method to prove the security of a proposed encryption scheme, such as IND-CPA, is through reduction <cit.>. Usually, breaking the scheme is reduced to solving a hard mathematical problem, which means that if an attack is possible for the scheme then the mathematical problem would be efficiently solved. That is, the encryption scheme is at least as hard as the mathematical problem. The scheme is modeled as a subroutine, whose inputs are simulated such that the adversary cannot tell whether it is being involved in an attack or in a subroutine to help solve the hard problem. Although forward proof is possible, the more commonly used technique is a contradiction: by assuming that the adversary could distinguish two ciphertexts in an experiment with a non-negligible advantage, the reduction would lead to a non-negligible probably to efficiently solve the hard mathematical problem that is believed to be intractable, thus leading to a contradiction. § PSI This section discusses the potential security issues of Prism <cit.>. We start by describing the Prism PSI scheme (<ref>), then provide a security analysis (<ref>), and finally, exemplify the security issues using a real-world example (<ref>). §.§ Scheme Definition of Prism PSI This subsection provides a brief introduction to Prism <cit.>. §.§.§ Assumptions Non-colluding Servers Prism assumes that a cluster of non-colluding servers is available for storing secret shares generated from the original plaintext. In practice, this assumption can be realized by, for example, acquiring virtual machines from distinct cloud computing vendors such as Google Cloud Platform and Amazon Web Services. Hardness of Discrete Logarithmic Problem Let 𝔾 denote a (multiplicative) cyclic group and g as a generator. That is, ∀ y ∈𝔾, ∃ x ∈ [0, |𝔾| - 1] such that y = g^x. Informally, the discrete logarithmic (DL) problem states that if the carnality of 𝔾 is sufficiently large, then it is infeasible to find x in polynomial time even if both y and g are given. This is a well-accepted assumption in most modern cryptographic schemes, such as ECDSA <cit.> and ElGamal <cit.>. Knowledge of Clients We assume that the clients are aware of the domain of the attribute on which the PSI is carried on. In the context of PSI, we are mostly interested in the categorical or discrete data <cit.>, implying that each client can represent its local items with a boolean vector, where each element is a boolean value indicating whether a specific item exists on the client. §.§.§ Architecture While the full-fledged Prism deployment comprises four subsystems: clients (i.e., database owners), non-colluding servers, the oracle (initiator), and the announcer, the PSI functionality can be completed through the first three subsystems, i.e., without the announcer. Clients (Database Owners) The clients are the owners of the sensitive data that should be kept confidential from other clients. We expect the clients are honest during the PSI procedure: (i) clients do not have the incentive to lie about their local data items for PSI and are not interested in probing or analyzing any intermediate results; (ii) clients are not compromised by adversaries without being detected, which can be enforced by digital signatures <cit.> and Byzantine fault-tolerant protocols <cit.>. Servers In the PSI scheme of Prism, two servers are available and they do not collude. This implies that the servers cannot be active adversaries; that is, both servers are not malicious. However, servers are not trusted in the sense that they may behave as passive adversaries, meaning that they could be interested in learning about the plaintexts from its involvement in the outsourced databases. Because the data touched on by each of the two servers are random due to the protocol of generating secret shares, each server has no way to probe the plaintext and is thus information-theoretically secure. Oracle The oracle is also called an initiator, whose job is to set up the initial parameters for both the clients and the servers. For example, two groups are required for Prism and it is the oracle that specifies the group order and generator. The oracle is useful only at the beginning of the PSI scheme; if the context is clear, we will skip the oracle's initialization step and assume it has been done in the offline stage. It should be noted that the oracle in Prism is not a conventional oracle that has been widely used in cryptography: in the latter case, an oracle usually refers to an entity whose functionality is “ideal” and does not reveal any internal machinery, i.e., as a black box. Figure <ref> illustrate the system model of PSI over multi-owner outsourced databases. We assume there are four clients {C_0, C_1, C_2, C_3} and two non-colluding servers {S_0, S_1}. Before the clients start outsourcing the local data, the oracle sends parameters to clients and servers. The clients then apply additive secret-sharing primitives to the local data and send them to both servers through a secure communication channel, such as OpenSSL <cit.>. Both servers execute a protocol to encode the secret shares and broadcast the encoded values back to all clients. §.§.§ Protocols Initialization The oracle picks the parameters of two cyclic groups and broadcasts them to servers and/or clients. Some of those parameters (e.g., generator of the multiplicative group) should be kept confidential to the servers only such that the clients cannot learn about any information beyond the PSI. In addition to the parameters, the oracle is also responsible for configuring the platform such as the host names and other metadata information. Vectorization of Attribute Domain Each client C_i generates a local vector, V_i, such that ∀ v ∈ V_i v ∈{0, 1}. A value v = 1 in V_i if and only if the corresponding item exists on C_i; otherwise v = 0. Indeed, here we assume that all clients agree upon the same order of items in the domain of the attribute, denoted by A_c. On straightforward approach to agreeing on the order of domain values is to take the alphanumeric order. Secret Sharing Each client C_i splits its local vector V_i into two secret vectors, denoted by V_i^0 and V_i^1 that are to be sent to S_0 and S_1, respectively. One requirement for the splitting is that the secrets must be additive, meaning that V_i^b, b ∈{0, 1}, must be able to reconstruct V_i by the addition operation on the secrets. The additive property exploited by Prism is the modular addition defined in an additive group ℤ_p = {0, 1, …, p-1}, where p denotes a large prime. That is, given a plaintext v ∈{0, 1}[Recall that the elements of a vector are all boolean values, 0 or 1.], we randomly pick the first share v^0 ∈ [0, m) and calculate the second share v^1 as follows: v^1 ≡ v - v^0 (mod p). If we apply the above splitting to all elements of a local vector V_i, we can generalize the above equation from scalars to vectors, in an element-wise fashion, as the following: V_i^1 ≡ V_i - V_i^0 (mod p). Aggregating Shares After server S_b receives all the secret shares V_i^b, b ∈{0, 1} and 0 ≤ i < m, S_b carries out a local aggregation over the secret shares from all clients C_i. That is, server S_b computes a new vector V^b as follows: V^b ≡ -[m^b] + ∑_i = 0^m-1 V_i^b (mod p), where [m^b] denotes the extended vector (with the same length of V^b) of a secret share generated from the number of clients m, b ∈{0, 1}, in the same way we previously compute for v^b. The operand of -m^b is added because we want to in later steps cancel out the value “1” in a vector that represents the existence of a shared attribute value. Encoding Shares The server should not simply send the locally aggregated shares back to the client because doing so would reveal the exact number of shared items among clients. Instead, the servers must somehow encode the partial shares such that the client cannot learn about the shared items except for the intersection. The encoding scheme used by Prism unsurprisingly leverages the algebraic property of a homomorphic function Hom between two cyclic groups 𝔾 = (G, +) and ℍ = (H, ×) such that: Hom: G → H x ↦ g^x (mod q), where g is a generator and q is the order of the multiplicative group ℍ. It is obvious that the function Hom() is homomorphic, ∀ x_0, x_1 ∈ G: Hom(x_0) × Hom(x_1) = g^x_0× g^x_1 = g^x_0 + x_1 = Hom(x_0 + x_1). We can then apply Hom() to the elements of V^b on server S_b. Let U^b denote the encoded vector in which each element is applied with the Hom() function. Server S_b broadcasts U^b to all clients. Calculating PSI Each client receives two vectors U^0 and U^1 from two servers S_0 and S_1, respectively. The client calculates the element-wise product of U^0 and U^1 without knowing the value of g—the generator of the multiplicative group. Let n denote the cardinality of A_c's domain; if U^0 = (u^0_0, u^0_1, …, u^0_n-1) and U^1 = (u^1_0, u^1_1, …, u^1_n-1), then we define the element-wise product as follows: U = U^0 ⊙ U^1 (u^0_0 · u^1_0, u^0_1 · u^1_1, …, u^0_n-1· u^1_n-1). Evidently, ∀ i ∈ [0, n), u_i u^0_i · u^1_i = 1 if and only if the summation of secret shares of all clients equals m. Indeed, this is the criterion for detecting the intersection of the attribute values: only those values whose encoding is 1 from all m clients should be included in the PSI. On the other hand, if not all boolean ciphertexts are 1, then u_i cannot be 1 implying that the i-th value of the attribute domain is not in PSI. Furthermore, because g is visible only to the servers, clients cannot learn about the exact number of clients who share the i-th value of the attribute domain. §.§ Security Analysis of Prism §.§.§ High-Level Intuition While Prism hides the exact number of clients who share the i-th value of the attribute domain unless every client holds it, this section will show that there is other information leaked from the Prism scheme. The key insight is that the function of Eq. (<ref>) is deterministic, implying that repeated inputs lead to the same output. As a result, if the same number of clients share a specific attribute value, then Prism would leak such information to an adversary. §.§.§ Chosen-Plaintext Attacks Security Notion Informally, a Chosen-Plaintext Attack (CPA) refers to an attack where the adversary can distinguish the ciphertexts of two arbitrarily chosen plaintexts with a non-negligible probability, even after running the attack algorithms polynomial times in the security parameter, usually the bitstring length of the problem size, e.g., the cardinality of the attribute domain. A negligible function of argument n is defined as negl(n) o(n^-c), where c denotes any constant. That is equivalent to: ∀ c > 0 lim_n →∞negl(n)/n^-c = 0. The formal proof of indistinguishability under CPA (IND-CPA) for a scheme Π usually takes a form of a simulation, where we assume that Π is not IND-CPA and use Π to construct a protocol that would solve an intractable problem, thus leading to a contradiction. In the context of PSI, IND-CPA implies that all elements in U should look random except for 1, which indicates a shared attribute value. An adversary can trivially break the IND-CPA of Prism in two rounds: * The adversary inserts a new attribute value x_n to all but the first client and retrieves the corresponding element in U, say u_n, from any client. * The adversary inserts a new attribute value x_n+1 to all but the first client and another new attribute value x_n+2 to all but the first two clients. Similarly to the first round, the adversary retrieves two ciphertexts in U: u_n+1 and u_n+2. The adversary can win the IND-CPA game with a probability of one by comparing u_n with u_n+1 and u_n+2. While it can be argued that the practicality of allowing the adversary to insert new attribute values into database owners is debatable, the CPA security does consider the “worst-case” scenario. Even if this IND-CPA may be thought of as theoretical interest only, the following section constructs a more practical attack. §.§.§ Inference beyond PSI Suppose that each client computes the encoded vector U = (u_0, u_1, …, u_n-1) from C_i, i ∈ [0, m), where n denotes the cardinality of the attribute domain. Without loss of generality, let 0 ≤ i ≠ j < n and u_i = u_j ≠ 1. This means the i-th and the j-th values of attribute A_c are shared by up to m-1 clients. Let [·] denote the positional element of an array, a specific client C_k can infer the following information: * If V_k[i] = V_k[j], C_k knows that A_c[i] and A_c[j] are equally “popular” on all database owners. * If V_k[i] < V_k[j], i.e., V_k[i] = 0 and V_k[j] = 1, then C_k knows that there is exactly one additional client who holds A_c[i] rather than A_c[j]. Depending on the applications, such information might introduce bias or unfairness, e.g., in a market of competitors. * If V_k[i] > V_k[j], client C_k can infer similar information that is symmetric to the second case above. Evidently, any of the above three scenarios does reveal some information beyond the intersection of items among the clients. §.§ CPA Attack on Prism This section demonstrates a running example of Prism's PSI protocol and attacks on it. §.§.§ System Configuration We assume m = 4, i.e., there are four clients, C_i, 0 ≤ i < 4. If not otherwise stated, let b {0, 1}, i.e., b is assigned 0 or 1 with 50% probability[In the literature of cryptography, it is also common to use $, _R, or other varieties to denote the same thing.]. There are a fixed number of non-colluding servers; for simplicity, two servers are available: S_0 and S_1. Let the additive group 𝔾 = (ℤ_5, +) and the multiplicative group ℍ = (ℤ_11^*, ×) and g=3 as a generator of a subgroup of ℍ. Note that the aforementioned parameters are taken for the sake of explanation; they are usually much larger in practice to be resilient to brute-force attacks. §.§.§ Schema and Data Schema and Metadata For simplicity, we assume that all of the four clients hold their local relations of a single-attribute schema. Since there is only a single attribute, we can safely represent the column of the relationship as a vector. Let's also assume the data type of the single attribute is an integer in the semi-open interval [0, 5). We use n to denote the cardinality of the attribute domain, i.e., n = 5 in our example. This means that although each client could store an arbitrary number of values, the length of each plaintext vector is 5, i.e., |V_i| = 5, 0 ≤ i < 4. Data Let R_i denote the single-attribute relation on client C_i as a list: R_0 = (0, 1, 3) R_1 = (1, 3, 4) R_2 = (3, 4, 4) R_3 = (1, 2, 3, 4) The goal of PSI is for every C_i to find out the commonly shared item, in this case integer 3. §.§.§ Vectorization The plaintext data on four clients in Eq. (<ref>) can be vectorized to the following: V_0 = (1, 1, 0, 1, 0) V_1 = (0, 1, 0, 1, 1) V_2 = (0, 0, 0, 1, 1) V_3 = (0, 1, 1, 1, 1) The elements that should be included in the PSI are those values whose positional vector elements are one. In our example, the only element satisfying this is integer 3. If we denote the PSI as a function over the set of relations, we have PSI(ℛ) = {3}, where ℛ{R_i}, 0 ≤ i < m. We also introduce another parameterized function Card(·, k), which returns the set of attribute values that are shared by k clients: Card: R^m ×ℤ_m → 2^A_c (ℛ, k) ↦{y | y ⊆ A_c}, where R denotes an arbitrary vector of integers between 0 and n and R^m denotes the cross-product m times over R. For simplicity, we sometimes place the second parameter in the subscript, i.e., Card_k(·) Card(·, k). By definition, it is evident that PSI(ℛ) is a special case of Card_k(ℛ), i.e., PSI(ℛ) ≡ Card_|ℛ|(ℛ). With this notation, we can write the following for our example: Card_0(ℛ) = ∅ Card_1(ℛ) = {0, 2} Card_2(ℛ) = ∅ Card_3(ℛ) = {1, 4} Card_4(ℛ) = {3} We notice that the cardinality of some Card_k() is equal, such as |Card_1(ℛ)| = |Card_3(ℛ)| = 2. As we will see later, this will reveal information that can be exploited by the adversary. §.§.§ Secret Sharing Recall that the client relies on the additive group 𝔾 = (ℤ_5, +) for additive secret sharing, as discussed in <ref>. Therefore, the V's in Eq. (<ref>) can be split into the following: V_0^0 = (3, 4, 1, 2, 0); V_0^1 = (3, 2, 4, 4, 0) V_1^0 = (1, 2, 4, 3, 4); V_1^1 = (4, 4, 1, 3, 2) V_2^0 = (0, 4, 2, 3, 1); V_2^1 = (0, 1, 3, 3, 0) V_3^0 = (2, 3, 4, 1, 0); V_3^1 = (3, 3, 2, 0, 1) It can verified that V_i = V_i^0 + V_i^1 in 𝔾, 0 ≤ i < 4. Client C_i then sends V_i^0 to server S_0 and V_i^1 to server S_1. §.§.§ Server Encoding Each of the two servers receives m = 4 vectors of shares. That is, server S_b receives V_i^b, b ∈{0, 1}, 0 ≤ i < m. As specified in Eq. (<ref>), m should be split as well on 𝔾. In this example, let m^0 = 1 and m^1 = 3. Therefore, the first element of the aggregated vector V^0 on S_0 can be calculated as V^0[0] = -m^0 + ∑_i=0^3 V_i^0 = -1 + (3 + 1 + 0 + 2) ≡ 0 (mod 5). Similarly, we can compute both aggregated vectors V^0 = (0, 2, 0, 3, 4) V^1 = (2, 2, 2, 2, 0) The two servers then apply Eq. (<ref>) to V^b and generated two new vectors (note that q = 11 in our example): U^0 = (1, 9, 1, 5, 4) U^1 = (9, 9, 9, 9, 1) Both servers then broadcast U^b to all clients. §.§.§ Client Decryption All clients receive two vectors U^b from both servers. Each client computes the element-wise multiplication of U^b under group ℍ, i.e., modular 11. For example, U^0[1] × U^1[1] = 9 × 9 = 81 ≡ 4 (mod 11), and U^0[3] × U^1[3] = 5 × 9 = 45 ≡ 1 (mod 11). The final vector on each client is therefore U = (9, 4, 9, 1, 4). All clients at this point know that the fourth value of A_c is shared by all clients because U[3] = 1 and have no idea about how many other clients hold what other values because those encrypted values are meaningless due to Eq. (<ref>). §.§.§ Information Leakage All clients learn that U[0] = U[2] = 9. Let's see what client C_0 can learn from this. Recall that client C_0 knows its local vector, e.g., V_0[0] = 1 and V_0[2] = 0. This means that C_0 can infer the following fact: there must be at least one client C_j, j ≠ 0, such that V_j[2] = 1, because otherwise U[0] cannot equal U[2]. That is, by following the Prism PSI scheme C_0 learns that the integer “2” exists on at least one of the other three clients, although “2” is not shared by all of the four clients (because U[2] = 9 ≠ 1). § PSI To fix the information leakage of Prism PSI, we design a new PSI scheme called Kaleido for multi-owner databases. As the name suggests, Kaleido introduces disorder to the PSI calculation among multi-owner databases and therefore is considered “garbled” Prism. The key idea of Kaleido is fairly simple: instead of using a fixed generator, let's randomize it such that the encoded shares on the server cannot be distinguished even for the same inputs. The technical challenge of applying this idea is to ensure both the correctness and the security of PSI. §.§ Server Protocol of Kaleido Description We elaborate on the randomness introduced by Kaleido servers in Algorithm <ref>. Lines 1 – 7 simply aggregate the secret shares sent by all clients and subtract the secret share of the number of clients m. Line 10 ensures that (i) the output of the parameterized function is indistinguishable from a random string, and (ii) the output is identical on both S_0 and S_1 because the secret key k is agreed upon between both servers. In practice, PRF_k(·) can be implemented by a block cipher, such as AES <cit.>. Kaleido then maps the randomized output to an element in the group ℍ, which is done in Line 11. As in Line 10, both S_0 and S_1 generate the same g because ℍ is revealed to both servers. Lines 12–14 ensure that g is a generator of a (sub)group of order p. Note that we use |g| (Line 12) to denote the order of g in the cyclic group rather than the absolute value of scalar g in arithmetic. On Line 15, the elevated value in group ℍ replaces the element in the group 𝔾. Line 16 ensures that even the same elements in 𝔾 map to distinct elements in ℍ due to different values of ct and g. Server S_b finally broadcasts its local encoding V^b to all clients, as shown in Line 18. Figure <ref> illustrates a simple example where both servers take an initial vector (IV) to randomly encode the secret shares submitted by clients. Intuitively, because both servers parameterize the pseudorandom function PRF(·) with the same key k initialized by the oracle at the beginning of the execution, the base number g (cf. Line 11, Alg. <ref>) is identical on both servers and yet look random in different rounds. Complexity Suppose the cost of computing PRF(·) is c. Lines 2 – 7 take 𝒪(nm). Lines 9 – 17 take 𝒪(nqc). Therefore, the overall computational complexity of Alg. <ref> is 𝒪(nm+nqc). The overall number of messages is straightforward: Line 18 incurs m messages. Since there are two servers, the overall number of network messages is 2m. The overall number of communication rounds is 1, also incurred by Line 18 (assuming broadcasting is implemented as an asynchronous primitive). §.§ Correctness of Kaleido We need to show that U[i] = 1, 0 ≤ i < n, if and only if V_j[i] = 1, 0 ≤ j < m. §.§.§ V_j[i] = 1 U[i] = 1 If V_j[i] = 1, recall that this means the j-th client holds the i-th positional value of attribute A_c. Let g_i denote the i-th base of elevation, as shown in Line 12 of Alg. <ref>. As discussed in <ref>, g_i is identical for both S_0 and S_1. According to Eq. (<ref>), the following holds U[i] ≡ U^0[i] × U^1[i] (mod q) = g_i^V^0[i]× g_i^V^1[i] ≡ g_i^V^0[i] + V^1[i] (mod p) = g_i^ - m^0 - m^1 + ∑_j=0^m-1 V_j[i] = g_i^ - m + ∑_j=0^m-1 1 = g_i^0 = 1. §.§.§ U[i] = 1 V_j[i] = 1 If U[i] = 1, then according to Eq. (<ref>), the following equality must hold: - m^0 - m^1 + ∑_j=0^m-1 V_j[i] = 0, or equivalently: ∑_j=0^m-1 V_j[i] = m. Recall that V_j[i] ∈{0, 1}, cf. Eq. (<ref>). Therefore, the only way to satisfy the above equation is V_j[i] = 1, 0 ≤ j <m, which is desired. §.§ Provable Security of Kaleido We claim that the PSI scheme backed by Alg. <ref> is IND-CPA except for the attribute values shared by all clients. This means that the adversary is not allowed to query the servers to return the ciphertext of an attribute value that is shared by all clients. We formulate the above in the following proposition. If the PRF(·) function in Alg. <ref> is indistinguishable from an ideal random number generator, the elements of U generated by V^b in Alg. <ref>, b ∈{0, 1}, are provably secure in the IND-CPA security model. Let 𝒜 denote the adversary who can break the security of Alg. <ref>. We write MOD^CPA_𝒜, Kaleido to denote the private-set-intersection (PSI) experiment in Multi-Owner Databases (MOD): * 𝒜 outputs two attribute values x_0 and x_1 in 𝔾 such that there exists at least one client that does not hold x_0 or x_1, respectively; * Kaleido servers agree upon a value b ∈{0, 1}, run Alg. <ref>, and send Enc(x_b) to 𝒜; * 𝒜 outputs b' ∈{0, 1}; * The output of the experiment is defined to be 1 if b' = b, and 0 otherwise. We write MOD^CPA_𝒜, Kaleido = 1 if the output of the experiment is 1 and 𝒜 is said to be successful. Let g_0 and g_1 denote the bases of x_0 and x_1, respectively, as specified in Line 12 of Alg. <ref>. We consider two cases: (i) g_0 ≠ g_1 and (ii) g_0 = g_1. * If g_0 ≠ g_1, we have Pr[MOD^CPA_𝒜, Kaleido = 1 | g_0 ≠ g_1] = 1/2, because the best 𝒜 can do is a random guess of b. According to elementary probability theory, we also know that the following holds: Pr[MOD^CPA_𝒜, Kaleido = 1 ∧ g_0 ≠ g_1] ≤1/2. * If g_0 = g_1, since PRF(·) is indistinguishable from a random generator, the following equality holds: Pr[g_0 = g_1] = 1/2^λ. It follows that Pr[MOD^CPA_𝒜, Kaleido = 1 ∧ g_0 = g_1] ≤1/2^λ. Combining Eqs. (<ref>) and (<ref>), we have Pr[MOD^CPA_𝒜, Kaleido = 1] ≤1/2 + 1/2^λ = 1/2 + negl(λ), implying that 𝒜 cannot succeed significantly better than a random guess, which completes the proof. §.§ Case Study of Kaleido We revisit the example discussed before in <ref> to illustrate the effectiveness of the proposed Kaleido PSI. Recall that the probability of having g_i = g_j, i ≠ j, is negligible. Let's assume that g_i equals i+2 % 11. For example, U^0[3] is calculated as U^0[3] = (3+2 % 11) ^ V^0[3] % 11 = 5^3 % 11 = 4, and U^1[3] is calculated as U^1[3] = (3+2 % 11) ^ V^1[3] % 11 = 5^2 % 11 = 3. Following the above algorithm, U^b can be calculated as: U^0 = (1, 9, 1, 4, 9) U^1 = (4, 9, 5, 3, 1) Consequently, the Kaleido PSI U is: U = (4, 4, 5, 1, 9). We see that U[3] = 1 still implies that the fourth positional value of A_c is shared by all clients, which is desired. However, the equality or inequality between other values does not reveal the number of clients sharing the values. For example, although U[0] = U[1] = 4, it does not indicate that integers “0” and “1” are shared by the same number of clients; in fact, “0” is hold only by C_0 and “1” is held by all clients but C_2. Similarly, although U[0] ≠ U[2] = 5, it does not mean that integers “0” and “2” are not shared by the same number of clients—both are indeed shared by the same number of clients (a single client C_0 for “0” and a single client C_3 for “2”). § EVALUATION §.§ System Implementation We have implemented the proposed Kaleido scheme with about 1,000 lines of Python code and Bash script, which will be released at <https://github.com/>. We choose the lightweight SQLite as the local database instance. Note that SQLite is a file-based database and does not support network access. We thus implement a communication layer among remote SQLite instances through the paramiko library for secure data transfer and remote query invocation. Some of the most important libraries and dependencies include: python 3.8.0, sqlite 3.31.1, numpy 2.21.0, paramiko 2.12.0, scp 0.14.4, and cryptography 39.9.0. §.§ Experimental Setup §.§.§ Test Bed We deploy Kaleido and other baseline schemes alone with SQLite <cit.> on a 34-node cluster hosted at CloudLab <cit.>. Each node is equipped with two 32-core Intel Xeon Gold 6142 CPUs, 384 GB ECC DDR4-2666 memory, and two 1 TB SSDs. The operating system image is Ubuntu 20.04.3 LTS, and the page size is 4 KB. All servers are connected via a 1 Gbps control link (Dell D3048 switches) and a 10 Gbps experimental link (Dell S5048 switches). We only use the experimental links for our evaluation. Specifically, we name the 34 nodes in the cluster as node0–node33. The two servers run on node0 and node1, and the database owners (i.e., clients) are deployed on node2–node33. All 34 nodes are enabled with password-less SSH connection for convenient communication since our evaluation focuses on performance metrics rather than security measurement. §.§.§ Systems under Comparison For all the systems under comparison, the orders of the additive and multiplicative groups are 113 and 227, respectively. Prism Prism <cit.> assigns a static generator during the server encoding. As discussed in prior sections, Prism can leak information other than the PSI, such as which items are shared by the same number of clients. However, because the generator is fixed, Prism incurs no extra cost for randomizing the ciphertext, which leads to the highest performance compared with the schemes proposed in this paper. Kaleido-RND (Naive Randomness) One straightforward way to resolve static generators in Prism is to randomly pick a generator, which we refer to as Kaleido-RND. This requires that both servers synchronize the random seeds and therefore incur server-to-server communication, which is against the assumption that the PSI servers should be non-colluding. Moreover, we are unaware of an effective model to formally prove the security of this scheme. Kaleido-AES (Provably Secure) We chose AES-128 <cit.> to implement the pseudorandom function in our Kaleido prototype system. PKCS7 is used for padding the plaintext data and the cipher block chaining (CBC) mode is adopted for the AES cipher. The initial vector (IV) is set to 1234. §.§.§ Workloads TPC-H Benchmark The synthetic data set is TPC-H version 3.0.0 <cit.>, a standard database benchmark. There are overall eight tables in TPC-H; in our evaluation, we select the LineItem table—the same one used in Prism <cit.>, which comprises 6,001,215 tuples. The attribute that we will focus on is L_ORDERKEY, which consists of 1,500,000 distinct values. Each database owner, i.e., the client, randomly select 5,000,000 values from the L_ORDERKEY attribute, and the PSI goal is to find out those L_ORDERKEY values shared by the set of clients. COVID-19 Patient Record The first application is the U.S. national COVID-19 statistics from April 2020 to March 2021 <cit.>. The data set has 341 days of 16 metrics, such as death increase, positive increase, and hospitalized increase. Each client randomly selects 300 days of data as their local data. Bitcoin Trade History The second application is the history of Bitcoin trade volume <cit.> since it was first exchanged in the public in February 2013. The data consists of the accumulated Bitcoin exchange on a 3-day basis from February 2013 to January 2022, totaling 1,086 large numbers. Each client randomly selects 1,000 trade data points in their local database. Human Gene #38 Sequence The third application is the human genome reference 38 <cit.>, commonly known as hg38, which includes 34,424 rows of singular attributes, e.g., transcription positions, coding regions, and number of exons, last updated in March 2020. Each client randomly selects 30,000 rows in its local data set. §.§ Performance §.§.§ TPC-H Benchmark Server Encoding Figure <ref> reports the encoding performance of various schemes. We compare the server processing time of all three schemes when serving different numbers of clients ranging from 2 to 32. Unsurprisingly, Prism achieves the highest performance because of the fixed generator g. Kaleido-RND is slightly slower than Prism because of the communication between the two servers and the randomization cost on both servers. Kaleido-AES introduces about 500 seconds overhead for encoding five million tuples; this is the cost for ensuring the semantic security of PSI over multi-owner databases. Please note that the overhead is a computational cost, which can be reduced by parallel processing that has not been explored in this work. Client Computation In our implementation, the work on the client can be broken into four stages: (i) Load the data from the local SQLite database, (ii) Hash the raw value and map it to a binary flag, (iii) Split the binary flag into two additive secret shares, and (iv) Recover the PSI result from two encoded vectors from servers. Table <ref> reports the time consumption for each of the above four stages. We observe that the most costly stage is Split, which takes almost 10 seconds. Again, this can be parallelized with multiple CPU cores because the elements in the vector are independent, which we will explore in our future work. I/O Cost Table <ref> reports the local I/O cost on both clients and servers. Note that the numbers are for individual clients or servers, e.g., each client persists a vector of 15 MB in size—the total I/O overhead of the entire system is thus 15 × m MB, where m denotes the number of database owners (clients). Communication Cost Figure <ref> reports the communication time spent on different numbers of clients. We observe that the upstream communication is almost constant when scaling out the number of clients; however, the downstream traffic takes proportionally more time when more clients are involved. When the number of clients is sufficiently large, e.g., 32, server-to-client communication takes more time than the other direction. §.§.§ Applications Due to limited space, we only report the performance of these three applications on client scales: two, four, and eight. We observe that in these real-world applications, the performance overhead lies in communication: both the client and the server spend most of their time sending the vectors, orders of magnitude more than other stages. COVID-19 Figure <ref> reports the stage-wise performance of Kaleido and Prism on the COVID-19 data. For clients, we observe that the communication from clients to serves dominates the cost. For servers, again, the communication from servers to clients takes the majority of the overall cost, and the portion increases when more clients are involved. As a result, although the cryptographic cost introduced by Kaleido is significant from the computational perspective, the real bottleneck of the entire system lies in communication. In fact, the computational cost of the cryptographic component can be further reduced (not explored in this work). Bitcoin Figure <ref> reports the performance of the Bitcoin trade history. We observe a similar pattern to the COVID-19 data. Even more evidently, the numbers show that server communication dominates the overall cost on the server side: When working with only two clients, the communication time of the servers takes 10x more time than the cryptographic component (i.e., AES <cit.>). These results reaffirm the practicality of the proposed Kaleido scheme: Although the cost is significantly more than the naive random method, the cost is an order of magnitude lower than the communication time. Human Gene #38 Sequence Figure <ref> reports the performance on a larger data set—over 30,000 human gene #38 sequence. We observe that the server exhibits roughly the same performance as the previous two data sets; however, the clients are more occupied this time. This is understandable as the clients in this case must work on more data items regarding their local databases. § RELATED WORK This section reviews three other important techniques for ensuring the confidentiality of outsourced databases. All of these techniques are orthogonal to the secret-share-based schemes (e.g., Prism, Kaleido). §.§ Encrypted Storage The database instance from the cloud vendor is considered as storage of encrypted data and the client is responsible for nontrivial queries. This solution is viable only if (i) the relations touched on by the query are small enough that the network overhead of transmitting those relations is acceptable, and (ii) the user has the capability (both computation and storage) to execute the query locally. We stress that this solution might defeat the purpose of outsourcing the database service to the cloud. §.§ Encrypted Tuples Every tuple of the original relation R is encrypted into a ciphertext that is stored in column T of a new relation R^s. For each attribute A_i in R, there is a corresponding attribute A_i^s in R^s, whose value is the index of R.A_i. The index is usually assigned by a random integer based on some partitioning criteria and can be retrieved with the metadata stored on the client, i.e., the user's local node. As a result, the schema stored at the cloud provider is R^s(T, A_1^s, …, A_i^s, …). When the user submits a query Q, the client splits Q into two subqueries Q_s and Q_c. Q_s serves as a filter to eliminate those unqualified tuples based on the indices in R^s and transmits the qualified tuples (in ciphertexts) to the client. Q_c then ensures that those false-positive tuples are eliminated after the encrypted tuples are decrypted using the secret key presumably stored on the client. This approach involves both the client (i.e., the user) and the server (i.e., the cloud provider) when completing a query, often referred to as information hiding approaches <cit.>. §.§ Encrypted Fields The third approach aims to minimize the involvement of clients when processing the query over the encrypted data stored at the cloud vendor. The idea is to encrypt the relations at a finer granularity—each attribute of a relation is separately encrypted. The key challenge of this approach lies in its expressiveness, e.g., how to apply arithmetic or string operations over the encrypted fields. While fully homomorphic encryption (FHE) <cit.> can support a large set of computing problems, the performance of current FHE implementations cannot meet the requirements of practical database systems <cit.>. An alternative solution is a partially homomorphic encryption (PHE) schemes <cit.>, which are orders of magnitude faster than FHE but only support a single algebraic operation. Traditional PHE schemes are designed for public-key (asymmetric) encryption, which is desirable for straightforward key distribution over insecure channels but significantly more expensive than secret-key (symmetric) encryption. However, in the context of outsourced databases, the user usually serves as both the sender and the receiver and there is no need to distribute the key. To this end, symmetric (partially) homomorphic encryption, was proposed <cit.>. § CONCLUSION AND FUTURE WORK This paper first demonstrates a security vulnerability of a state-of-the-art PSI scheme, Prism, in outsourced databases. The paper then proposes a series of new primitives, namely Kaleido, to extend Prism such that the new PSI scheme is semantically secure. Both the correctness and security are formally proven, and the intuition is explained with a running example. A system prototype of Kaleido is implemented and deployed to two servers and 32 database owners hosted at CloudLab <cit.>. Extensive experiments on the TPC-H benchmark and three real-world applications confirm the effectiveness of the proposed Kaleido schemes. Our future work is two-fold. First, we will leverage the underlying multiple/many CPU cores to parallelize the costly cryptographic operations in Kaleido. The challenge lies in how to ensure the multiple (non-colluding) servers do not communicate and yet stay synchronized regarding the randomized group generators. Second, we will adopt some compression schemes to reduce the vectors, generated by both the clients and the servers, such that the communication cost can be reduced. The challenge of the second research direction lies in the balance between the computational cost for compressing and decompressing the vectors and the communication saving out of the compressed vectors. plain

http://arxiv.org/abs/2303.06977v1

20230313102422

Topological dynamical quantum phase transition in a quantum skyrmion phase

cond-mat.str-el

Department of Physics, Indian Institute of Technology (Banaras Hindu University) Department of Physics and Medical Engineering, Rzeszów University of Technology, 35-959 Rzeszów, Poland Max Planck Institute of Microstructure Physics, Weinberg 2, D-06120 Halle, Germany Institute for Theoretical Physics, Johannes Kepler University, Altenberger Straße 69, 4040 Linz, Austria Max Planck Institute of Microstructure Physics, Weinberg 2, D-06120 Halle, Germany Radboud University, Institute for Molecules and Materials, Heyendaalseweg 135, 6525AJ Nijmegen, Netherlands Department of Physics, Indian Institute of Technology (Banaras Hindu University) Quantum skyrmionic phase is modelled in a 2D helical spin lattice. This topological skyrmionic phase retains its nature in a large parameter space before moving to a ferromagnetic phase. Next nearest-neighbour interaction improves the stability and it also causes a shift of the topological phase in the parameter space. Nonanalytic behaviour of the rate function observed, when the system which is initially in a quantum skyrmion phase is quenched to a trivial quantum ferromagnetic phase, indicates a dynamical quantum phase transition. Dynamical quantum phase transition is absent when the system initially in a skyrmion phase is quenched to a helical phase. Topological dynamical quantum phase transition in a quantum skyrmion phase S. K. Mishra March 30, 2023 ========================================================================== Progresses made in the last decade in controlling matter at quantum levels allow access to real-time dynamics of closed quantum many-body systems realisable <cit.>. This progress lifted the border between experimentally feasible physical reality and model systems. Ultra-cold atoms in optical lattices and trapped ions are examples in which such dynamical phenomena were observed in real-time. Nowadays, we have full access to the real-time dynamics of quantum many-body and finite systems, either isolated or coupled to the Markovian or non-Markovian environment. The experiments with THz pulses in solids <cit.>, high magnetic field pulse experiments <cit.>etc. are also developments in recent past which can be aided by theoretical understanding of dynamical properties of the corresponding quantum systems especially the study of the evolution of the system after a sudden change in its parameter or a quantum quench. When we talk about changing the parameter of a system, the first thing that pops up in our mind is the term phase transition. Phase transitions are the points in the parameter space of a system around which a small change in the control parameter manifests a drastic change to its characteristics. In a classical/thermal phase transition, the thermal fluctuations cause the destruction of long-range ordering and facilitate the phase transition. But when we study the changes of the parameters at zero temperature or ground states the characteristics of the phase transition become purely quantum, because here the phase transition is facilitated by quantum fluctuations instead of thermal fluctuations. Such phase transitions are known as quantum phase transitions. Equilibrium quantum phase transitions(EQPT) are studied extensively but we have a lesser understanding of quantum systems out of equilibrium. In order to theoretically aid the experimental developments mentioned earlier we need to understand the dynamical quantum phase transitions(DQPT). Recent experimental developments identified a signature of dynamical behaviour after a quench in a Haldane-like system <cit.>. Experiments with trapped ions were able to directly observe DQPTs <cit.>. The theoretical inspiration for the DQPT can be extracted from Lee-Yang theorem, Fisher zero's and accompanying analysis. (For more details refer to <cit.>) From Lee-Yang analysis <cit.> one can arrive at the conclusion that for a partition function of external fields (like magnetic field), which also depended on the temperature T as Z(T), when zeros exist and have a positive real value then each of those roots corresponds to nonanalyticity in the free energy, i.e., phase transition. Fisher extended this analysis considering partition functions with complex temperature z instead of T, Z(z). When Fisher's zeros of Z(z) overlap with the real axis it produces nonanalyticities or phase transitions, however, no such overlaps is observed in the course of an EQPT. The real values of Fisher's zeros correspond to a different kind of phase transition, namely DQPT <cit.>. Using these analyses essence of DQPTs can be explained briefly as follows <cit.>: Short-lived non-equilibrium phase transitions accompanied with a nonanalytic behaviour of physical quantities as a function of time is a characteristic feature of DQPTs <cit.>. To study this we utilise a quantity 𝒢 the Loschmidt amplitude as a function of time t, given as 𝒢(t)=⟨ψ_0| e^-iĤ t|ψ_0⟩. In a sense, it plays the same role in the study of states out-of-equilibrium as the partition function Z=Tr(e^-βĤ) in thermodynamic equilibrium case. Here Ĥ is the Hamiltonian of the system and β is the inverse temperature and ψ_0 is the initial state of the system. Here Z can be seen as the kind of partition function considered by Lee and Yang(with inverse temperature in place of temperature, still the conclusions of the analysis holds). And in the exponent term of 𝒢(t) if we take (it) as the complex temperature, 𝒢 behaves like partition function of complex temperature considered by Fisher. Another quantity of interest is the rate function of the return probability, Eq. <ref> (hereinafter referred to as rate function) which is analogous to the thermodynamic free energy. As discussed earlier during a phase transition the thermodynamic free energy, F=-ln(Z)/β, turns out to be a nonanalytic function of a control parameter. Based on the analogy we established so far, we expect to see a nonanalytic behaviour on ℒ when there is a DQPT, since it is a dynamical analogue of thermodynamic free energy. The rate function is given as: ℒ(t)=-lim_N→∞1/Nln|𝒢(t)|^2, where N is the number of degrees of freedom of the system. DQPT is intensively studied during the last decade <cit.>. In the present work we are interested in the interplay between topology and DQPTs considering the quantum skyrmion <cit.>. The scientific community is still fascinated about topological states of matter even though it has been over four decades since the discovery of the Quantum Hall state, the first discovered topological state <cit.>. Topologically distinct states or topological states are those states, which are classified based on a certain invariant <cit.>. Such states are said to be identical when we can move from one state to another by applying continuous smooth deformations(deformations which do not close the bulk energy gap) without changing the value of the invariant <cit.>. When the system shows this kind of resistance to deformation we say that it is topologically protected. Conventional states, which are earlier believed to be the same may become topologically distinct. We discover this only when the accompanying physical behaviour is detected, like in the case of the Quantum Hall effect. In the past few decades, scientists discovered many topological materials and states like topological insulators <cit.>, topological crystaline insulators <cit.>, topological semi-metals <cit.> etc. This classification is based on the behaviour of the corresponding band Hamiltonians in the reciprocal state. There are also nontrivial topological objects determined by their characteristics in real space such as various types of topological defects in condensed matter <cit.>. Magnetic skyrmions are among the most popular types of topological defects studied now. Skyrmions are particular examples of solitons that can be informally defined as localized waves with a stable shape (for more accurate definition and detailed discussion see e.g. <cit.>). They are related to peculiar localized non-colinear magnetic textures within magnetic systems <cit.>. Skyrmions are promised to be potential information carriers for the next generation of spintronic devices <cit.>. New studies suggest macroscopic skyrmion qubits design suitable for quantum computing technology controlling the helicity and dynamics of the skyrmions through electric fields <cit.>. Helicity arises due to the peculiar non-collinear orientation of spins in helical spin systems(magnets) like skyrmions. In conventional magnets spins align either parallel or anti-parallel, i.e. spins are collinear. Helicity is a measure of how much the spins are rotated about a reference direction and its orientation. These developments make the study of the dynamics of quantum skyrmions demanding. The formation of magnetic order in spin systems depends on different factors and competing interactions. A ferromagnetic or an antiferromagnetic spin order is a relatively simpler case. Formation of non-colinear magnetic textures is mainly fueled either by competing nearest-neighbor ferromagnetic and next nearest-neighbor antiferromagnetic or asymmetric exchange interactions termed as Dzyaloshinskii–Moriya interaction (DMI). In most cases, the DMI is the dominant mechanism forming non-conventional magnetic textures <cit.>. Recently, the quantum analog of magnetic skyrmions has been suggested and studied <cit.>. However, contrary to the classical skyrmion the quantum skyrmion is not topologically stable in a rigorous sense. Qualitatively, it is not protected with respect to the quantum tunneling to the topologically trivial vacuum state. At the same time, it presents a quantum spin state with quite a special character reminiscent of its topologically protected classical analog. Here we use the words "topological phase" for the case of quantum skyrmions in this, not completely rigorous but intuitively clear, sense. To better understand the nature of this state we studied its real-time dynamical properties. Unlike skyrmions, <cit.> helical magnetic textures do not possess a topological invariant and a topological protection. Any classification of 2D helical phases is difficult, even properties of quantum skyrmions are not well explored. Recent attempts to discover quantum skyrmions show a certain degree of success. Lohani et.al. <cit.> and Gauyacq et.al. <cit.> could identify magnetization patterns of a quantum skyrmion but topological protection of skyrmion phase was not explored. Sotnikov et.al. used a quantum scalar chirality to identify a topological protection<cit.>. Siegl et.al. using topological index and winding parameters, were able to identify a skyrmion phase and quantify its stability <cit.>. All these works could only find a skyrmionic phase in the ground state. In the present work, we investigate the stability of quantum skyrmions in higher excited levels as well. An important question is the strength of the DM interaction. DMI can be determined in experiments or calculated accurately from first-principles <cit.>. Our interest is focused on materials with large DMI: <cit.>. In recent experimental works H. Yang et.al <cit.> has shown that magnetic films sandwiched between nonmagnetic layers exhibit DMI with an electrically tunable strength, i.e., Co films sandwiched between nonmagnetic layers or MgO/Fe/Pt. The value of DMI in such materials linearly increases with the applied external electric field D=D_0+g_MEE, where D_0 is the intrinsic DMI part, g_ME is the magnetoelectric coupling and E is the external electric field. Enhanced DMI can formally reach the order of exchange interaction for a large electric field. The dynamic control of intrinsic magnetic interactions by varying the strength of a high-frequency laser field allows further enhancing of the ratio between DMI and exchange interaction constants <cit.>. The idea relies on the fact that DMI and exchange interaction are both based on hopping processes and that time-periodic fields renormalize the electronic tunneling, leading to the effective rescaled DMI and exchange constants D≈4tΔ/U_0-U and J≈4t^2/Ω^2(U_0-U). Here Δ describes the Rashba spin-orbit coupling, t denotes the hopping amplitude, U_0-U is the gain in the Coulomb energy due to the electron displacement, and Ω is the frequency of the laser field. Because of the factor 1/Ω^2, the high-frequency laser field can substantially reduce the rescaled exchange constant to achieve the condition D>J. In order to study Helical-Skyrmionic-Ferromagnetic phases we consider an array of spins formed in a triangular lattice. The spins are arranged in such a way that it has a six-fold rotation symmetry (see Fig. <ref>), also with periodic boundary conditions (PBC) it possesses translation symmetry. The Hamiltonian of the system has the form: Ĥ=B∑_i 𝐒̂_̂î^̂ẑ+J_1∑_⟨ i,j⟩𝐒̂_̂î𝐒̂_̂ĵ+ J_2∑_⟨⟨ i,j⟩⟩𝐒̂_̂î𝐒̂_̂ĵ +∑_i<jD_i,j[𝐒̂_̂î×𝐒̂_̂ĵ], where summation in single brackets is taken over the nearest neighbors and in double brackets over the next-nearest-neighbors, B is an external magnetic field, and the DMI vector D is aligned perpendicular to the bond between lattice sites i and j, see inset of Fig. (<ref>). The direction of DMI vectors is chosen to ensure that the six-fold rotation symmetry holds. We note that in helical multiferroic insulators, the parameter D is an effect of the magnetoelectric coupling g_ME with an external electric field E, i.e., D=g_MEE. Thus, the strength of the DMI term can be controlled externally <cit.>. Here we consider both J_1(a ferromagnetic nearest neighbor exchange) and J_2 (an anti-ferromagnetic next nearest neighbor exchange) interactions along with the DMI. An array of spins read along with the above Hamiltonian with a specific parameter set forms a quantum skyrmion. The nearest-neighbor ferromagnetic J_1 term encourages colinear spin orientation while DMI term compels non-colinear spin texture. These two competing interactions form classical skyrmions stabilized by the applied magnetic field. Below we show that adding even small next nearest neighbor antiferromagnetic Heisenberg term J_2 improves the stability of quantum skyrmion structures. In general, it is well-known that the J_2 term leads to the spin frustration and formation of antiferromagnetic classical skyrmions <cit.> However, quantum skyrmion structures are quite specific as compared to classical skyrmions. In particular, quantum skyrmions do not possess continuous magnetic texture and topological charge. Therefore to infer the quantum skyrmion state, we exploit another tool, such as scalar chirality. The energy levels of the system Eq.(<ref>) show a six-fold degeneracy even at very high magnetic field values. We calculated the expectation values considering maximally mixed state of these degenerate states with equal probability. We used spin correlation functions to characterise quantum non-trivial magnetic structures. In particular, we explore the Fourier transform of the longitudinal spin correlation function G_(q)=1N∑_ijG_(r_ij)exp(-iqr_ij), where N is the number of spins and the correlation function is given by G_(r_ij)=1/Z∑_n⟨ n|Ŝ_i^zŜ_j^z| n⟩exp(-β E_n). where E_ns are the eigenvalues of the Hamiltonian of the system and r_ij is the distance between the lattice sites i and j. Topologically protected systems like skyrmions tend to show resistance to a deformation in its configuration<cit.>. Throughout this work an applied magnetic field is considered as a cause of deformation. In Fig. <ref>(A) and Fig. <ref>(B) panels are arranged in the increasing order of applied magnetic field from left to right. Fig. <ref>(A) shows that the local magnetisation is uniform and increasing magnetic field causes the local magnetisation to increase in the direction of the field. This observation fails to identify any magnetic textures in the system. Fig. <ref>(B), multiple intensity spots (Bragg peaks) observed for nonzero q(Fourier conjugate of r or wave vector in reciprocal space) confirms formation of the quantum non-trivial magnetic texture <cit.>. This quantity could distinguish between helical and ferromagnetic phases but it fails to identify skyrmionic phase within the helical phase. So we require another quantifier that can trace out the skyrmion phase from helical and ferromagnetic phases. Scalar chirality <cit.> C_Ψ =𝒩/π⟨Ŝ_1[ Ŝ_2×Ŝ_3] ⟩. is considered as a distinguishing property of a helical spin system. When the chirality has a non-zero value we say that the system is in a helical phase. It is proposed that chirality can distinguish quantum skyrmion phase from other phases of the system <cit.>. In this equation 𝒩 is the number of non-overlapping elementary triangular patches covering the lattice. Three adjacent spins form a patch. The scalar chirality for any of these three adjacent spin combinations is the same, because of the translational and rotational symmetries of the lattice. In Fig. <ref>(C) from B=0D to B=0.46D chirality increases almost steadily. In this region no two deformed states have same chirality value, therefore, all those states are topologically non-identical. From B=0.46D till B=0.86D the plateau of scalar chirality implies that all the states in this region have a common topological invariant and we call them topologically identical phase or simply topological phase. After crossing B=0.86D the system briefly falls back to a helical phase. It is represented by a dip in chirality. Then around B=1.0D the system goes to the trivial ferromagnetic phase indicated by zeros of scalar chirality. Here we notice that the scalar chirality is again invariant for the ferromagnetic phase. So we may say that ferromagnetic phases are also topologically identical. But that is too trivial to be worthy of any mention. In Fig. <ref>(C) The magnetisation graph is telling us that for the region where we have plateau in chirality the magnetisation is not same for any two states i.e., the system is in fact undergoing deformation during the constant chirality plateau also. The phase diagram (Fig. <ref>(D)) shows that for non zero J_2 the invariance of scalar chirality extends for larger values of the applied magnetic field compared to J_2=0. The skyrmion phase is embedded into the helical phase. For a larger value of the interaction parameter J_2, the system retains the skyrmion phase for a longer range of applied magnetic field. We studied the scalar chirality not only in the ground state but also in several excited states. We see that the skyrmion state survives in first and second excited states as well as a certain degree of topological invariance can be seen in higher excited states, see Fig. <ref>. When J_2=0D we arrive at the results from O. M. Sotnikov et. al. <cit.>. For higher values of J_2(J_2>0.35D) the plateau gets distorted. Here we show that an important information about helical and quantum skyrmion phases can be obtained from the analyses of topological DQPTs. Namely, when quenching the system from a quantum skyrmion to a topologically trivial phase, we observe a characteristic signature of a DQPT as a nonanalyticity in the rate function. On the other hand, a quench between a skyrmion and a helix does not lead to a DQPT. Let the system be prepared in the ground state |ψ_0⟩ of the Hamiltonian Ĥ_0=Ĥ(λ_0). At t=0 the parameter λ is quenched to a new value Ĥ(λ_f) and the initial wave function is evolved to a new state |ψ(t)⟩=e^-iĤ t|ψ_0⟩ under this new Hamiltonian. In order to describe the DQPT we study the Loschmidt amplitude (return amplitude) and the rate function, in the thermodynamic limit <cit.>. Our primary interest concerns DQPTs between a quantum skyrmion phase and a trivial phase. The quench protocol applied on a topological phase to a trivial (skyrmion to ferromagnet) phase produced a nonanalytic behaviour of the rate function with respect to time. The result is plotted in Fig. <ref>(A). On the other hand, when switching between a skyrmionic phase to a helical phase, we do not observe any DQPT. We can confirm this from the lack of nonanalytic behaviour of the rate function when plotted against time in Fig. <ref>(B). Here we looked for nonanalyticity with different values of the parameter B but we could not find any. The presence of nonanalyticity in the former case can be explained by a sharp and discontinuous transition of C_ψ from skyrmionic to trivial phase in Fig. <ref>(C). Similarly the lack of nonanalyticity in the later case is due to the smooth and continuous transition of C_ψ from helical to skyrmionic phase. From Fig. <ref>(A) the critical exponent α = 0.7020 ± 0.0233 around t_c = 9.66. The value turned out to be very close to this for other critical points as well (see supplementary methods for details <cit.>, see, also, references <cit.> therein). Further study is required with larger system sizes(L=3i(i+1)+1, i=3, 4, 5, ...) in order to comment on the universality of the critical exponent that we calculated. We have computational limitations to analyzing finite-size effects and artifacts of a particular quantum skyrmion state in our system. However, we performed the study of the finite-size effects using another quantum skyrmion state obtained for a slightly different model in <cit.>. The results of the calculations are shown in <cit.>. We see the same trend for this case also, i. e., non-analytic singularities in the rate function during the dynamical quantum phase transition between the skyrmion and FM phases. Thus obtained results are pretty universal and apply to any quantum skyrmion. The reason for the universal effect is the orthogonality of quantum skyrmion and FM states. This argument is valid for any quantum skyrmion independently of its size. Apart from this, we achieved a significant improvement of the topological phase stability with our model compared to the previous works <cit.>. We note that within a skyrmionic phase a larger J_2 value gives a topological phase protection against a larger range of applied magnetic fields. At high values of J_2 the topological invariance is destroyed. This result tells us that the key for tunability and improved stability of quantum skyrmions can be the interaction parameter, this may be useful when choosing the material to realise skyrmions for experiments. This kind of model is realized in Pd/Fe/Ir(111) system with Co surrounded edges <cit.>. Also nanoscale skyrmions are reported at room temperature with large DMI interaction in Ir/CoFeB/MgO systems <cit.>. In conclusion, the quantum skyrmion model proposed above shows significant improvement in topological protection. In the said model we identified a robust DQPT when quench protocol is applied from a skyrmionic state to a ferromagnetic state. Robust DQPTs accompany many interesting properties. An interesting direction is to look for the connection between entanglement dynamics and DQPTs in a skyrmionic phase. Certain systems have reported to show an enhanced entanglement entropy around critical points of DQPTs <cit.>. § ACKNOWLEDGEMENT We would like to sincerely thank O. M. Sotnikov for discussion related to the initial stage of this work. We acknowledge National Supercomputing Mission (NSM) for providing computing resources of ‘PARAM Shivay’ at Indian Institute of Technology (BHU), Varanasi, which is implemented by C-DAC and supported by the Ministry of Electronics and Information Technology (MeitY) and Department of Science and Technology (DST), Government of India. A.E. acknowledges funding by Fonds zur Förderung der wissenschaftlichen Forschung (FWF) grant I 5384. The work of MIK is supported by European Research Council via Synergy Grant 854843 - FASTCORR. SKM acknowledges Science and Engineering Research Board, Department of Science and Technology, India for support under Core Research Grant CRG/2021/007095. apsrev4-1

http://arxiv.org/abs/2303.07302v1

20230313172748

Shallower CNOT circuits on realistic quantum hardware

quant-ph

The category of extensions and idempotent completion [ March 30, 2023 ==================================================== We focus on the depth optimization of CNOT circuits on hardwares with limited connectivity. We adapt the algorithm from Kutin et al. <cit.> that implements any n-qubit CNOT circuit in depth at most 5n on a Linear Nearest Neighbour (LNN) architecture. Our proposal is a block version of Kutin et al.'s algorithm that is scalable with the number of interactions available in the hardware: the more interactions we have the less the depth. We derive better theoretical upper bounds and we provide a simple implementation of the algorithm. Overall, we achieve better depth complexity for CNOT circuits on some realistic quantum hardware like a grid or a ladder. For instance the execution of a n-qubit CNOT circuit on a grid can be done in depth 4n. § INTRODUCTION Quantum decoherence is a major obstacle to the scaling of quantum computing. It is very hard to maintain the qubits isolated from the environment during a calculation and once the qubits interact with external elements the result of the current calculation is lost. The decoherence time is used to designate this limited amount of available computing time. This time is given by the hardware and adds an additional constraint at the software level and more precisely during compilation: the sequence of instructions generated by the compiler for the machine must be able to be executed in a sufficiently short time. In the quantum circuit model, this is equivalent to say that the depth of the circuit must be as low as possible. In this article we tackle the depth optimization of a specific class of quantum operators, namely the linear reversible operators or equivalently the CNOT circuits. CNOT circuits are themselves a subclass of the so-called Clifford circuits which play a major role in many different area of quantum computation such as quantum error correction <cit.>, randomized benchmarking protocols <cit.>, quantum state distillation <cit.>. CNOT circuits are also used in other classes of quantum circuits, for instance in the synthesis of phase polynomials <cit.>. More directly, the optimization of CNOT circuits has also been useful for the optimization of more general quantum circuits like arithmetic circuits <cit.>. We take into account some architectural constraint between the qubits. The CNOT gate is a two-qubit gate, if two qubits are not close enough in the hardware their interactions is not physically achievable and a CNOT gate between these two qubits cannot be executed. This limits the pair of qubits on which one can apply a CNOT gate and adds even more constraints to the compiler. In an ideal case, all the qubits are connected: then we talk about full connectivity or all-to-all connectivity, otherwise we talk about partial or constrained connectivity between the qubits. The optimization of CNOT circuits have attracted a lot of attention in recent years. Two metrics are generally used to evaluate the cost of running a CNOT circuit: its size or its depth. There are also two types of connectivity: full connectivity and partial connectivity. In total, this gives four cases to be treated. If we can find recent works that optimize the size in full connectivity <cit.> and in partial connectivity <cit.>, and works that optimize the depth in full connectivity <cit.>, nothing to our knowledge has been recently proposed to optimize the depth of CNOT circuits with architectural constraints. As far as we know, only two 15-year old works proposed implementations of CNOT circuits in the specific case of a Linear Nearest Neighbour (LNN) architecture <cit.>. The main result is that n-qubit CNOT circuits can be implemented in depth at most 5n for the LNN architecture. No improvement of this result nor extensions to other architectures were proposed since. We propose a block generalization of the algorithm proposed in <cit.>. Our algorithm works for any architecture where the qubits can be packed into groups of equal size such that the groups are arranged as a line. So for instance with groups of size 2 our algorithm can treat the case of a ladder like the architecture IBM QX5. With groups of size 4 we can deal with the grid. While not being universal for any qubit connectivity, our algorithm is versatile enough to work for realistic quantum architectures. We show that the algorithm skeleton does not depend on a specific architecture or the group size. Our algorithm consists in a series of small problems involving boolean matrices to zero: these problems are the atoms of our algorithm and this is where the block size and the actual architecture specify the constraints with which we have to solve them. We propose several strategies to solve these problems for different problem sizes and different architectures. Overall, we show that any n-qubit CNOT circuit can be executed in depth at most 4n in the case of the ladder. For the grid, a depth of 4n is sufficient and a depth of 15n/4 is enough if we add the diagonals. The structure of the article is a follows: in Section <ref> we give a brief background about CNOT circuits synthesis and we give a tuned formulation of Kutin et al.'s algorithm proposed in <cit.> for the LNN architecture. Then in Section <ref> we propose our block extension of Kutin et al.'s algorithm. We detail the general structure of the algorithm, independent of the architecture and the block size, and the subproblems we have to solve. Then in Section <ref> we give several ways to solve these subproblems for different cases of block sizes and architectures. We conclude in Section <ref>. § BACKGROUND AND KUTIN ET AL.'S ALGORITHM §.§ Background The CNOT gate is a classical reversible operator. It applies a NOT gate on a target qubit if the value of a control qubit is True. This is equivalent to writing CNOT(x_1, x_2) = (x_1, x_1 ⊕ x_2) where x_1, resp. x_2, is the logical input value of the control, resp. target, qubit and ⊕ is the XOR operator. By extension any CNOT circuit on n qubits applied to a bitstring x = (x_1, , x_n) outputs a bitstring y = (y_1, , y_n) where each y_i is a linear combination of the x_i's. In other words, y = Ax = [ A_11 x_1 ⊕ A_12 x_2 ⊕⊕ A_1n x_n; ⋮; A_n1 x_1 ⊕ A_n2 x_2 ⊕⊕ A_nn x_n ]. A ∈𝔽_2^n × n completely characterizes the functionality of the CNOT circuit. By reversibility of the operator, A is necessarily invertible. Given a CNOT circuit implementing an operator A, outputting a bitstring y=Ax, executing an additional CNOT gate with control i and target j will perform the operation CNOT(y_i, y_j) = (y_i, y_i ⊕ y_j) = A'x where A' is given from A by adding the row i to the row j. We write A' = E_ijA where E_ij = I ⊕ e_ji and e_ji is zero everywhere expect in the entry (j,i), and we note that E_ij^-1 = E_ij. The simulation of CNOT circuits is therefore efficient and its optimization affordable for a compiler. We just showed an equivalence between applying a CNOT gate and performing an elementary row operation on the matrix operator A. If one finds a suitable sequence of N row operations (E_i_k, j_k)_k such that ( E_i_N,j_N× E_i_N-1, j_N-1×× E_i_1, j_1) A = I then A = ∏_k=1^N E_i_k, j_k and one gets a direct implementation of A as a CNOT circuit by reading the different elementary row operations. Such decomposition is always possible, for instance any Gaussian elimination algorithm works. To summarize the synthesis of a CNOT circuit is equivalent to reducing an invertible boolean matrix to the identity with the use of elementary row operations. One is interested in an efficient algorithm to do such reduction. To evaluate the efficiency of a synthesis algorithm, two metrics are used: * the number of CNOT gates in the circuit, in other words the number of row operations used, * the depth of the circuit. This corresponds to the number of time steps needed to execute the circuit given that two gates that act on distinct qubits can be executed simultaneously. Moreover, the CNOT is a two-qubit gate, and therefore requires the interactions between two qubits to be executed. This might not be always physically possible to perform such interactions due to hardware constraints. We talk about graph connectivity to encode the available interactions: the nodes are the qubits and an edge indicates that the two qubits can interact and that a CNOT gate can be executed. When the graph is complete, all qubits are connected and we have a full qubit connectivity. Otherwise, the connectivity is said to be partial. We give some examples of realistic and existing qubit connectivities in Figure <ref>. The first algorithm improving the Gaussian elimination algorithm was the Patel-Markov-Hayes (PMH) algorithm <cit.>. It works in the case of a full qubit connectivity and generates circuits of size O(n^2/log_2(n)) where n is the number of qubits. Then several algorithms were proposed in recent years to improve the PMH algorithm <cit.> and extensions to partial connectivities were also proposed <cit.>. For the depth optimization, surprisingly, the first works treated directly the case of a LNN architecture <cit.>. The main result is that n-qubit CNOT circuit can be executed in depth at most 5n <cit.>. More surprisingly, to our knowledge, no other work was proposed to either improve the complexity in the LNN case or to extend it to other connectivities. For a full qubit connectivity, several algorithms were proposed in recent years <cit.>, notably the asymptotic optimum of O(n/log_2(n)) is achievable <cit.>. For completeness and for clarity, we now detail Kutin et al.'s algorithm achieving a depth of 5n on a line. Our block extension is natural with a suitable formulation of this algorithm. §.§ Kutin et al.'s algorithm In <cit.>, an algorithm for synthesizing any n-qubit CNOT circuit in depth at most 5n for the LNN architecture has been proposed. We propose a slightly reformulated version of this algorithm. The algorithm consists in two parts: * first, reduce the operator A to a north-west triangular operator B, i.e, such that B[i,j] = 0 if i+j > n-1 (the indices start at 0). * secondly, reduce B to the identity operator. In both steps, each qubit/row of the matrix will be given a label and the algorithm will consist in sorting the labels while maintaining some invariants. When the labels are sorted, the invariants impose the expected result: in the first step A will be north-west triangular, in the second step B will be the identity operator. First, write A = UPL where U is upper triangular, L is lower triangular and P is a permutation matrix. Such decomposition is always possible with a variant of the Gaussian elimination algorithm whose pseudo-code is given in Algorithm <ref>. Let J_n be the exchange matrix of size n, write A = UPJ_nJ_nL = U P' W where P' = PJ_n and W is a north-west triangular matrix. One can check that north-west triangular matrices are stable by left-multiplication with upper triangular matrices. Therefore, reducing A to a north-west triangular matrix is equivalent to reducing an upper triangular matrix with columns permuted into an upper triangular matrix. In other words we want to do U P' → U' for some upper-triangular matrix U', during the first step of the algorithm and U'W → I during the second step (UW being north-west triangular). §.§.§ From upper with columns permuted to upper triangular Given U and P, we label the rows of UP as the following: row i has label j if P[i,j]=1. Then one can check that matrix A=UP satisfies the following property that will be our invariant throughout the synthesis: 35em Invariant n°1: for each row i of A with label k, for each row j>i, A[j,k]=0. Also, we always have A[i,k]=1. Notice that if the labels are sorted, i.e row i has label i, and the invariant holds, then the matrix is upper-triangular (i.e. P is trivial). Figure <ref> depicts a matrix A=UP and its labeling. Let A be the operator we manipulate during the synthesis with elementary row operations. Initially A = UP. We now show that given two adjacent qubits i, i+1, one can always apply a linear reversible operator between the two qubits such that the two labels can be swapped while maintaining the invariant on A. First, note that both rows verify A[i,k] = A[i+1,k] = 0 for any label k of the rows 1 i-1. It is clear that this property will remain true after any linear reversible operation between those two rows. The values A[i,k], A[i+1,k] for any label k of the rows k > i+1 are arbitrary and do not participate in the truth value of the invariant. So now it should be clear that to maintain the invariant we only need to focus on the 2 × 2 submatrix B = A[[i,i+1], [k,k']] where k is the label of row i and k' the label of row i+1. Because of the invariant B can only have two values: B = [ 1 1; 0 1 ] or B = [ 1 0; 0 1 ]. To swap labels k and k', we have to apply a 2-qubit linear reversible operator C such that CB = [ ⋆ 1; 1 0 ] where ⋆ can be either 0 or 1. With CB of this shape, we can assign label k to row i+1 and label k' to row i while maintaining A[i,k'] = 1, A[i+1,k]=1 and A[i+1,k'] = 0. The choice of C is simple: * if B = [ 1 1; 0 1 ] then C = E_i,i+1 works, one CNOT is sufficient, * if B = [ 1 0; 0 1 ] then C = E_i,i+1E_i+1,i works, and two CNOT are sufficient. To summarize, given a matrix A with labels on the rows and that satisfies invariant n°1, we have a procedure that modifies A such that we can swap the labels of two adjacent rows while maintaining the invariant. This procedure, that we call a box (to follow the terminology of <cit.>) requires at most 2 CNOT gates. Once the labels are sorted, the invariant guarantees us that A will be upper triangular. What remains to do is to propose a quantum circuit made of boxes that guarantees that the labels are sorted after its execution. We want this circuit to be the shallowest possible. What is proposed in <cit.> is a LNN sorting network, an example on 7 bits is given in Figure <ref>. The circuit, as a box-based circuit, is of depth at most n. Therefore, given that each box is of depth at most 2, the CNOT-based circuit is of depth at most 2n. §.§.§ From northwest triangular to the identity The principle is exactly the same than in the first step, except this time we have to reduce a north-west triangular matrix A. For this step, we use the following invariant: 35em Invariant n°2: for each row i of A with label k, for each row j, if j>i then A[j,k]=0 ; if j < i and row j has label k' < k then A[j,k]=0. Also, we always have U[i,k]=1. Initially, A is northwest triangular and we assign to row i the label n-i+1. One can check that the property is satisfied. Figure <ref> depicts such a matrix A and its labeling. Given a pair of qubits i, i+1, we can again swap the labels with some modifications on A. Again, for the same reasons that in the first step, note that we only have to focus on the submatrix B = A[[i,i+1], [k,k']]. If k < k' then necessarily B = [ 1 0; 0 1 ] and there is nothing to be done. If k' < k then B can still have only two values: B = [ 1 1; 0 1 ] or B = [ 1 0; 0 1 ]. The difference this time compared to the first step is that we impose to apply an operator C such that CB = [ 0 1; 1 0 ] otherwise the invariant would not be preserved. So we have two possibilities for C: * if B = [ 1 1; 0 1 ] then C = E_i+1,iE_i,i+1 works, two CNOT are sufficient, * if B = [ 1 0; 0 1 ] then C = E_i,i+1E_i+1,iE_i,i+1 works, and three CNOT are sufficient. With the same sorting network, we can sort the labels and reduce A to the identity operator. As a box-based operator, the sorting network is of depth n. But now each box can be of depth at most 3 and therefore the second step can be executed in depth at most 3n. This gives one of the main result of <cit.>: Any n-qubit linear reversible operator can be executed in depth at most 5n on a LNN architecture. Simply concatenate the two steps, the first step can be executed in depth at most 2n, the second step can be executed in depth at most 3n, hence the result. § A BLOCK VERSION OF KUTIN ET AL.'S ALGORITHM The main contribution of our article is a block version of Kutin et al.'s algorithm. For our block version to work, we will need our target qubit topology to verify a simple structure. Let n be the number of qubits, we assume n is a multiple of some integer p which will be the block size. We pose m=n/p the number of blocks. We write b_1 = [1,2, , p], b_2 = [p+1, p+2, , p+p], , b_m the different blocks. We further assume that each block b_i induces a connected subgraph of the topology. Furthermore, the blocks are connected in line, meaning that in each block b_i, i < m, there exists a qubit that is connected to a qubit in b_i+1. §.§ Description of the algorithm Given the description we made of Kutin et al.'s algorithm, the extension to a block version is straightforward. The idea is to perform Kutin et al.'s algorithm on the line of blocks. We use the notation A[b_i,:] for the natural indexing of the rows (i-1)*p+1, i*p of A. Similar notations are used to index the columns as well. The general procedure is the same as in the case of the path graph: we start from A = UPW, for an upper-triangular matrix U and a north-west triangular matrix W. We label the rows of UP similarly according to P and we reduce UP to a block upper triangular matrix U'. Then we reduce U'W, a block north-west triangular matrix, to the identity. During the process we will sort the labels by block while maintaining two invariants. §.§.§ From UP to block upper triangular During this first step, the invariant we maintain is the following: 35em Invariant n°3: for each block b_i of A with labels k=[k_1, , k_p], for each block b_j, j>i, U[b_j,k]=0. Also, we have A[b_i,k] invertible. This new invariant is illustrated in Figure <ref>. Similarly to invariant n°1, this invariant holds for any UP with U upper triangular and P a permutation matrix by giving label j to row i if P[i,j]=1. Moreover, if the labels are block-sorted, i.e each row of b_i has a label in b_i, then this specifies a block upper triangular matrix. Given two adjacent blocks b_i, b_i+1 with labels k, k', we show how to assign any p-sized subset k” of k∪ k' to block b_i while maintaining the invariant. Let k”' = k∪ k' ∖ k”. Similarly to the LNN case, we only need to focus on a particular submatrix, namely a 2p × 2p matrix B = A[[b_i, b_i+1], [k, k']]. Let's write B = [ A_1 A_3; 0 A_2 ] where A_1, A_2, A_3 are p × p, by invariant n°2 A_1 and A_2 are invertible, B is full rank, and the lower-left p× p block is 0. If we want to assign labels k” to block b_i we are more interested in a column-permuted version of B, namely B[:,[k”,k”']] = [ B_1 B_2; B_3 B_4 ], where B_1, B_2, B_3, B_4 are arbitrary as long as B is full rank. If one finds a suitable linear reversible operator C such that CB[:,[k”,k”']] = [ B_5 B_6; 0 B_7 ] then one can assign labels k” to block b_i. B_6 is arbitrary. CB is full rank so we have B_5 and B_7 invertible and the invariant is fully maintained. Any choice for k” works but in our case we want k” to be the p smallest labels among k,k' to perform a sorting. Then using an LNN sorting network on the blocks one can finally block sort the labels and transform A into a block northwest triangular matrix. Note that we do not need in fact to focus on the columns labeled by k”' of B. Our goal was to show that the suitable subblocks of B are still invertible, but in practice we only need to focus on A[[b_i, b_i+1], k”] = [ A_1; A_2 ] Because the labels k” can somehow be arbitrary, we cannot conclude anything on A_1 and A_2 except that they form a matrix of rank p. Our goal is to find a C such that CA[[b_i, b_i+1], k”] = [ A_3; 0 ] with A_3 necessarily invertible. §.§.§ From block northwest triangular to the identity For this second step, each block b_n-i+1 has initially the labels k_i = (i-1)· p+1, i· p . During the second step, the labels will be block sorted but labels within each block label will not be modified. The invariant we maintain is the following: 35em Invariant n°4: for each block b_i of U with labels k_l, for each block b_j, j>i, A[b_j,k_l]=0. For each block b_j, j < i, with label k_h, if h < l, then A[b_j, k_l]=0. Also, we have A[b_i,k_l] invertible. Again, one can check that this property is satisfied for a block north-west triangular matrix with the chosen labeling. This is illustrated in Figure <ref>. Once the labels are block sorted, then the invariant specifies a block-diagonal matrix. Given two adjacent blocks b_i, b_i+1 with labels k_j, k_j', j > j', we show how to swap the blocks of labels while maintaining the invariant. We focus on the 2p × 2p submatrix B = A[[b_i, b_i+1], [k_j', k_j]]. We write B = [ A_1 A_3; A_2 0 ] where A_2, A_3 are invertible. We want to find a linear reversible operator C such that CB = [ A_4 0; 0 A_5 ]. After applying C the two blocks of labels can be swapped, obviously A_4 and A_5 are invertible and the invariant is preserved. Again we use an LNN sorting network on the blocks such that we eventually reduce the northwest triangular matrix into a block diagonal matrix. §.§.§ From block diagonal to the identity This final step is straightforward: we assume we can perform a direct synthesis on each block in parallel that reduces each block to the identity gate. This can be done in depth O(p), so provided the block size is a O(1) this step can be done in depth O(1). §.§.§ Summary Given an hardware composed of m blocks of p qubits laid out as a line, we have shown that we can perform Kutin et al.'s algorithm directly on the blocks. The resulting quantum circuit consists of two sorting networks made of "block-boxes", i.e, linear reversible operators that act on 2 adjacent blocks (2p adjacent qubits). During the first step, those block-boxes transform binary matrices of the form [ A_1; A_2 ] to [ A_3; 0 ] where A_3 is invertible, A_1 and A_2 are arbitrary. This is our Problem 1: Input ∘ an integer p > 0, ∘ A full rank 2p × p boolean matrix B = [ B_1; B_2 ], ∘ a connectivity graph G of size 2p giving the available row operations on B, Problem 1 Find a sequence of row operations C G-compliant such that: ∘ CB = [ B_3; 0 ]. During the second step, the block-boxes transform binary matrices of the form [ A_1 A_3; A_2 0 ] to [ A_4 0; 0 A_5 ] where A_2, A_3, A_4 and A_5 are invertible. A_1 is arbitrary. This is our Problem 2: Input ∘ an integer p > 0, ∘ A 2p × 2p boolean matrix B = [ B_1 B_3; B_2 0 ] with B_2 and B_3 invertible, ∘ a connectivity graph G of size 2p giving the available row operations on B, Problem 2 Find a sequence of row operations C G-compliant such that: ∘ CB = [ B_4 0; 0 B_5 ]. A third step simply consists on the synthesis of linear reversible operators on p qubits. This is our Problem 3. A pseudo-code is given in Algorithm <ref>. A detailed example on 8 qubits with blocks of size 2 and a topology of a ladder with diagonals is given in Figure <ref>. For the moment our framework is very generic because we have not proposed algorithms to synthesize shallow block-boxes. This corresponds to the function "ZeroBlock" in our pseudo-code. This will be the subject of the next section. Let's now derive a general formula for the total depth. If we write d_1(p), resp. d_2(p), the maximum depth needed to perform the transformations necessary during the first step, resp. the second step, and if we note d^*(p) the maximum depth required to do the synthesis of a linear reversible operator on p qubits, then d(n) ≤n/p×(d_1(p) + d_2(p) ) + d^*(p). The main question now is to propose algorithms for different architectures that will give interesting values for d_1, d_2, d^*. § PRACTICAL IMPLEMENTATIONS FOR DIFFERENT QUBIT CONNECTIVITIES Before proposing strategies, note that our two problems are unchanged after column operations. Either during Problem 1 when working on [ A_1; A_2 ] or during Problem 2 with [ A_1 A_3; A_2 0 ] any column operation on A_1, A_2 during Problems 1 and 2 gives another problem whose solutions are exactly the same. One can check that undoing the column operations on [ A_3; 0 ] resp., [ A_4 0; 0 A_5 ] will not modify the desired structure. §.§ Exhaustive search for small blocks For small p, it might be interesting to try an exhaustive search in order to have the best possible bounds on the depth. The search will consist in a breadth-first search. To perform such search, we need three elements: * starting roots, these elements need depth 0 to solve Problem 1 or 2, * a set of all available operations of depth 1, * a characterization of the set of all elements we need to find. The matrix at a maximum distance from the roots will give the maximum depth required to solve Problem 1 or 2. The roots for Problem 1 are all matrices of the form [ A_3; 0 ] with A_3 invertible. Similarly for Problem 2 the roots are all matrices of the form [ A_4 0; 0 A_5 ] with A_4, A_5 invertible. For Problem 1, we need to cover all full rank matrices of size 2p × p. For Problem 2, we need to cover all matrices of the form [ A_1 A_3; A_2 0 ] with A_2, A_3 invertible and A_1 is arbitrary. Fortunately, we can rely on the invariance of the problems by column operations to reduce the search space. With column operations, we can put our matrices in reduced column-echelon form. Such form is unique for one given matrix. This has two consequences: * first, both Problem 1 and 2 have only one root, namely [ I_p; 0 ]   for Problem 1 and [ I_p 0; 0 I_p ] for Problem 2, * secondly, for Problem 1 we only need to cover the set of matrices of size 2p × p in reduced column-echelon form. For Problem 2 we will have to search through the set of matrices of the form [ R_1 R_2 ] where R_1, R_2 are 2p × p and are in reduced column-echelon form. Note that not all matrices of this form are of interest (the targeted matrices have more structure) but this is the smallest stable set in which we can do the search. The set of available operations can be computed in the following way: * enumerate all possible matchings of the connectivity graph G, * each edge of a matching is a pair of qubits on which two different CNOT gates can be applied, depending on which qubit is the control and the target. For each matching do all possible combinations, i.e, if for each edge we assign a 0 or a 1 characterizing which CNOT is applied. The results of this exhaustive search are given in Table <ref>. First, note that for Problem 1 we recover the number of full rank matrices under reduced row echelon form. This is equivalent to counting the number of different p dimensional subspaces of the 2p dimensional vector space over GF(2). This is known to be the Gaussian binomial coefficient 2pp_2 (see, e.g, <cit.>). For Problem 2, all target matrices can be put in the canonical form [ A I; I 0 ] where A is an arbitrary p × p boolean matrix. We recover the fact that there can be 2^p^2 of them. From those maximal depth values, we deduce the complexities summarized in Table <ref> where we explicit the depth to solve each Problem for each architecture and the total depth required for full operator synthesis. Table <ref> gives an overview of the achieved asymptotic depth complexities for the various explored architectures. The main results, in our opinion, are the ones for the grid: any CNOT circuit on n qubits can be executed in depth 4n + O(1) on a grid and 15n/4 + O(1) when we add diagonal interactions. We also want to emphasize that when there is a full connectivity between blocks of size 4, the total depth is 7n/4 + O(1), which is better than the 2n result given by the extension of Kutin at al.'s to the full qubit connectivity <cit.>. The best practical method so far synthesizes CNOT circuits in depth n + o(n) <cit.>. This shows that with slightly more connectivity between the qubits than the LNN architecture we are able to significantly reduce the computational depth of CNOT circuits and have close to similar results to the full connectivity case. §.§ Combining block layouts Equipped with the asymptotic bounds for step 1 and 2 of our algorithm for different architectures, one can try to combine them in order to improve the overall depth bound in some case. We propose two possible improvements formalized in the following propositions. In a 2× 2L grid layout (n=4L), any n-qubits linear reversible operator can be implemented in depth at most 15n/4 + O(1). In this hardware setting, one can either divide the grid in 2L blocks of size p=2 and use the “Width-2 ladder” bounds, or divide the grid into L blocks of size 4 and use the “Grid” bounds. One can combine step 1 of the grid layout (7n/4) with step 2 of the Width-2 ladder layout (2n). In order to achieve this, we need to be able to turn a block north-west triangular matrix with block size 4 into a block north-west triangular matrix with block size 2. This step is at most as costly as performing a full block synthesis, and can thus be bounded by d^*(4), which is constant. Hence, the overall depth bound is of 7n/4 + 2n + O(1) = 15n/4 + O(1). In a 2L× 2K grid layout (n=4LK) with 2L-2 additional connections as in Figure <ref>, any n-qubits linear reversible operator can be implemented in depth at most 15n/4 + O(1). In a 2L× 2K grid layout (n=4LK) with diagonals (every other row) with 2L-2 additional connections as in Figure <ref>, any n-qubits linear reversible operator can be implemented in depth at most 7n/2 + O(1). The proofs of those propositions are similar to the one of Proposition <ref>. We use the additional connections in order to allow for a Width-2 ladder block layout to perform the second step. Consider for instance a square grid of even dimensions n=4L^2. In that setting, by adding 2L-2 = o(n) connections, we can obtain an asymptotic improvements of n/4 with respect to the complexity obtained for a standard Grid and Grid + diagonals layouts. §.§ All-to-all connectivity between larger blocks We now explore the more general case where we have a full connectivity between any pair of qubits in neighboring blocks and where the block size is arbitrary. In other words, in that setting, all the qubits in one block are fully connected, and all the qubits in block b_i are connected to all the qubits in block b_i+1. This is an hypothesis that can become true in some hardware models where each qubit has an interaction radius: each qubit can interact with any qubit within a certain distance <cit.>. In the case of a grid, with suitable interaction radius, our hypothesis can be true and the larger the radius the larger the block size can be. We propose two ways to solve the two problems, our method essentially relies on previous works about the synthesis of linear reversible circuits for unconstrained architecture <cit.>. §.§.§ Problem 1 By assumption we are given a full rank 2p × p binary matrix of the form [ A_1; A_2 ]. With a CNOT circuit of depth 1 we can add suitable rows of A_2 to A_1 to make A_1 invertible. Then, we consider the matrix B = A_2A_1^-1 that we zero with the following available operations: * elementary row operations on A_2 → available row operations on B, * elementary row operations on A_1 → available column operations on B, * elementary row operations from A_1 to A_1 → flip one entry of B. It is well-known that we can encode B in a bipartite graph such that any matching on B corresponds to a set of parallel row operations between A_1 and A_2. If B has at most k entries equal to 1 on each row and column, then the corresponding bipartite graph has degree k and can be decomposed as a sum of k matchings. Therefore, B can be zeroed in depth at most k. B is p × p so we are ensured that B can be zeroed in depth at most p. Problem 1 can be solved in depth at most 1+p. We can improve this result by using a technique used in <cit.>, they show that we can write B = B' ⊕1· v_1^T ⊕ v_2 ·1^T where v_1, v_2 are two arbitrary vectors and B' has at most p/2 entries on each row and column. Therefore step 1 can be realized first by performing B ⊕ B' in depth p/2. Then B ⊕ B' = [ 1 v_2 ]·[ v_1 1 ]^T = w_1 · w_2^T where w_1, w_2 are p × 2. We can reduce in parallel both w_1 and w_2 with elementary row and column operations on B ⊕ B'. Note that only 2 different nonzero rows can be found in w_1: [1,0] and [1,1], similarly in w_2 we can only find [0,1] and [1,1]. In both cases, we can zero any duplicate in depth at most log(p): given m occurences of [1,0] for instance, we can zero m/2 of them in depth 1, and we repeat the process. We eventually have the top 2 × 2 entries of w_1 and w_2 that are nonzero: this corresponds to a 2 × 2 block in B that can be zeroed in depth at most 2. Problem 1 can be solved in depth at most 3+p/2+log(p). §.§.§ Problem 2 In the all-to-all case, Problem 2 is very similar to Problem 1. By assumption we are given a 2p × 2p binary matrix of the form [ A_1 A_3; A_2 0 ] where A_2 and A_3 are invertible. In depth 2 we can obtain the matrix [ A_2 0; A_1 ⊕ A_2 A_3 ] and we have to zero A_1 ⊕ A_2 using A_2 in a similar way that we did during first step. Problem 2 can be solved in depth at most 2+p. Problem 2 can be solved in depth at most 4+p/2+log(p). §.§.§ Third step For this step we can also rely on the work already done for linear reversible circuits synthesis on unconstrained architectures <cit.>. The best method so far is a divide-and-conquer algorithm but for simplicity we use the adaptation of Kutin et al.'s algorithm to the all-to-all connectivity: any p-qubit linear reversible operator can be synthesized in depth at most 2p+6. §.§.§ Results In a quantum hardware with blocks of p qubits arranged on a line and with full connectivity between any pair of consecutive blocks, any n-qubit linear reversible operator can be synthesized in depth at most (2 + 3/p)n+ 2p + 6. We simply add the depths: d(n) = n/p × (1+p+2+p) + 2p+6 = 3n/p + 2n + 2p + 6. In a quantum hardware with blocks of p qubits arranged on a line and with full connectivity between any pair of blocks, any n-qubit linear reversible operator can be synthesized in depth at most (1+7/p+2log(p)/p)n + 2p + 6. We simply add the depths: d(n) = n/p × (3+p/2+log(p) + 4+p/2+log(p)) + 2p+6 < n + 7n/p + 2nlog(p)/p + 2p + 6. § CONCLUSION We proposed a block generalization of Kutin et al.'s algorithm that synthesizes CNOT circuits for an LNN architecture. Our generalization needs the blocks of qubits to be arranged as a line in the hardware. Despite this prerequisite, some realistic quantum hardware can fit into our framework. We showed that the depth complexity of our algorithm essentially depends on the solving of two elementary problems involving O(p)-sized boolean matrices where p is the size of the block. We brute-forced the solution for some small blocks (p ≤ 4) and we gave an algorithm for general p when the blocks are fully connected. As a result we improved the depth complexity for useful quantum hardware such as the grid. As a future work, it would be interesting to extend this framework to other classes of circuits: CZ circuits, Clifford circuits, phase polynomials. We could have used normal form for Clifford circuits <cit.> to propose preliminary results but we believe deeper analysis can be done. § ACKNOWLEDGMENTS This work has been supported by the French state through the ANR as a part of Plan France 2030, projects NISQ2LSQ (ANR-22-PETQ-0006) and EPiQ (ANR-22-PETQ-0007), as well as the ANR project SoftQPro (ANR-17-CE25-0009). The authors would like to thank Jérôme Pioux for his patience: running the enumerations took a bit longer than anticipated :) abbrv

http://arxiv.org/abs/2303.06827v1

20230313030003

Kernel Density Bayesian Inverse Reinforcement Learning

cs.LG

Large statistical learning models effectively forecast diverse chaotic systems William Gilpin March 30, 2023 ============================================================================== Inverse reinforcement learning (IRL) is a powerful framework to infer an agent's reward function by observing its behavior, but IRL algorithms that learn point estimates of the reward function can be misleading because there may be several functions that describe an agent's behavior equally well. A Bayesian approach to IRL models a distribution over candidate reward functions, alleviating the shortcomings of learning a point estimate. However, several Bayesian IRL algorithms use a Q-value function in place of the likelihood function. The resulting posterior is computationally intensive to calculate, has few theoretical guarantees, and the Q-value function is often a poor approximation for the likelihood. We introduce kernel density Bayesian IRL (KD-BIRL), which uses conditional kernel density estimation to directly approximate the likelihood, providing an efficient framework that, with a modified reward function parameterization, is applicable to environments with complex and infinite state spaces. We demonstrate KD-BIRL's benefits through a series of experiments in Gridworld environments and a simulated sepsis treatment task. § INTRODUCTION Reinforcement learning (RL) methods find policies that maximize an agent's long-term expected reward within a Markov decision process (MDP). In many observational data settings, we observe a sequence of states and actions for an agent carrying out a policy driven by an unknown reward function. It can be useful to learn this reward function to identify the factors driving the agent's behavior. For example, in a hospital setting, we see a patient's treatment schedule and measurements of physiological state. To make sense of the underlying factors influencing treatment decisions, we can identify the clinician's reward function (i.e., objectives), and examine how this function drives treatment decisions given a patient's state. It is particularly difficult to infer the reward function in this setting because the vector of observed covariates for a patient at a given time is often noisy and partially missing, and there may be several candidate reward functions that can explain the doctor's behavior. Inverse reinforcement learning (IRL) methods infer an agent's reward function given observations of the agent's behavior. Early IRL algorithms were used to identify point estimates of the reward function that best explained an agent’s behavior <cit.>, and were applied to problems in path planning <cit.>, urban navigation <cit.> and robotics <cit.>. A point estimate can also aid in imitation learning, where the inferred reward function is used to fit RL policies that replicate desired behavior. Despite the success of early IRL approaches, there are limitations to inferring a point estimate. First, the IRL problem is often non-identifiable <cit.>, meaning there may be multiple (and possibly infinite) reward functions that explain a set of behaviors equally well. Second, for finite demonstration data, point estimates fail to capture the uncertainty and noise in the data-generating process. Thus, it is advantageous to take a Bayesian approach, which treats the reward function as inherently random and communicates a degree of uncertainty that relies on the dataset distribution. A Bayesian approach to IRL computes a posterior distribution that places mass on reward functions proportional to how well they explain the observed behavior <cit.>. However, existing Bayesian IRL methods can be computationally demanding. In Bayesian modeling, likelihood specification has a large impact on the resulting posterior distribution. The formulation of the likelihood function (i.e., the function describing the probability of observing a state-action pair given a reward function) is unknown in the IRL setting. Existing approaches replace it with an optimal Q-value function, denoted by Q^⋆, which best approximates the long-term expected reward for a given state-action tuple(<cit.>). The Q-value function must be learned using Q-learning <cit.>, a “forward RL” algorithm, or one that solves an environment's MDP. The original algorithm pioneered by  <cit.> and the majority of its successors use Markov chain Monte Carlo (MCMC) sampling to compute a posterior over the reward function, and every iteration of MCMC requires forward RL for each sampled reward function. This is computationally expensive, especially with infinite or high-dimensional state spaces. Additionally, a posterior that uses Q^⋆ as a likelihood is equivalent to a Gibbs posterior <cit.> and lacks desirable theoretical properties <cit.>. We address these challenges with kernel density Bayesian inverse reinforcement learning (KD-BIRL), a method that (1) estimates the likelihood function directly, leading to theoretical guarantees for the consistency of the resulting posterior distribution, and (2) disassociates the number of times forward RL is required from the number of iterations of MCMC sampling, thus reducing computational complexity. The contributions of our work are as follows: * We propose KD-BIRL, a Bayesian IRL method that uses conditional kernel density estimation to directly approximate the likelihood function (Section <ref>). * We justify our method theoretically by proving posterior consistency, and demonstrating that the posterior contracts to the equivalence class of the expert reward function (Section <ref>). * We show that KD-BIRL's posterior estimates efficiently and accurately capture agent priorities in Gridworld environments (Section <ref>). * We demonstrate that, with a feature-based reward function, KD-BIRL can successfully infer rewards in complex state spaces such as a sepsis management task (Section <ref>). § PRELIMINARIES §.§ Background: Inverse reinforcement learning (IRL) The goal of IRL methods is to infer the reward function of an agent, given its behavior. An RL agent interacts with and responds to an environment that can be defined using an MDP. An MDP is represented by (𝒮, 𝒜, P, R), where 𝒮 is the state space; 𝒜 is the set of actions; P(𝐬^e_t+1 | 𝐬^e_t, 𝐚_t) defines state-transition probabilities from time t to t+1; and R : 𝒮→ℝ is a reward function, where R ∈ℛ and ℛ denotes the space of reward functions. The input to an IRL algorithm is a set of expert demonstrations, {(s^e_t, a^e_t)}_t=1^n, where each demonstration is a 2-tuple (s^e_t, a^e_t) representing an agent's state and chosen action at time t. These demonstrations are assumed to arise from an agent acting according to policy, π^⋆:𝒮→𝒜, that is optimal for a fixed but unknown reward function R^⋆. Given these demonstrations, IRL algorithms seek a reward function R^⋆ such that π^⋆ is optimal with respect to R^⋆. Bayesian approaches to IRL treat the reward function R as inherently random. By specifying a prior distribution over R and a likelihood function for the observed data, these methods then infer a posterior distribution over R given n expert demonstrations of an agent {(s_i^e,a_i^e)}_i=1^n. Using Bayes rule, the posterior density is equivalent to the product of the prior distribution on the reward, p(R), and the likelihood of the expert demonstrations given the reward function, with a normalizing constant that corresponds to the probability of the expert demonstrations: p(R | {(s_i^e,a_i^e)}_i=1^n) = p(R)∏_i=1^n p(s_i^e,a_i^e|R)/p({(s_i^e,a_i^e)}_i=1^n). In the initial formulation of Bayesian IRL (BIRL) <cit.>, the authors propose using a Q-value function to calculate the likelihood in <Ref>. The Q-value function for a given policy π at time t is Q^π(s_t, a_t) = r_t + γ𝔼_s^'∼ P[V^π(s^')], where s_t, a_t, and r_t are the state, action, and reward at time t, and γ∈ [0, 1] is a discount factor. This approach uses an optimal Q-value function, Q^⋆, as a component of the likelihood within a Gibbs posterior framework. The “likelihood” takes the form p(s, a | R) ∝ e^α Q^⋆(s, a, R), where Q^⋆(s, a, R) is the optimal Q-value function for reward function R, and α > 0 is an inverse temperature parameter that represents confidence in the agent's ability to select optimal actions. There are several potential challenges with learning the aforementioned BIRL posterior. First, the optimal Q^⋆ is found using Q-learning, a forward RL algorithm, and Q^⋆ is typically expensive to estimate for a new R. BIRL uses MCMC, which requires learning Q^⋆ on every iteration of sampling for a new R. In addition, because Equation <ref> is a loss-based function (rather than a true likelihood), the resulting function is not a classical Bayesian posterior <cit.> and does not have theoretical guarantees regarding posterior contraction (more details in Appendix Section 2,3). Additionally, Q-value estimates can be incorrect for states that are very rarely visited, as often happens in infinite state spaces, leading to incorrect likelihood estimates that can affect the accuracy of the BIRL posterior. §.§ Related Work Several extensions to the original BIRL algorithm <cit.> have been proposed. The first set of methods identifies nonparametric reward functions <cit.>. These algorithms use a variety of strategies such as Gaussian processes <cit.>, Indian buffet process (IBP) priors <cit.>, and Dirichlet process mixture models <cit.> to learn reward functions for MDPs with large state spaces that may include sub-goals. Other methods reduce computational complexity by using either more informative priors <cit.>, different sampling procedures (e.g., Metropolis-Hastings <cit.> or expectation-maximization <cit.>), variational inference to approximate the posterior <cit.>, or by learning several reward functions that each describe a subset of the state space <cit.>. However, all of these approaches use a Q-value function in place of the likelihood and hence still suffer from consistency and computational issues; this construction is both inefficient and limits desirable Bayesian behaviors with regard to posterior sampling and uncertainty quantification. To address these computational and consistency issues, it is necessary to either directly estimate the likelihood, reduce the number of times forward RL is performed, or modify the reward function parameterization. Recent work proposes a variational Bayes framework, Approximate Variational Reward Imitation Learning (AVRIL) <cit.>, to approximate the full posterior. This method improves upon existing work by avoiding re-estimating Q^⋆ for every sampled reward function, allowing it to bypass some of the computational inefficiencies of <cit.>'s initial formulation. However, AVRIL still requires the use of a Q-value function, resulting in a misspecified optimization objective. One method avoids using a Q-value function entirely for the likelihood <cit.> and instead approximates it using real-time dynamic programming or action comparison. Other work proposes a feature-based reward function, which parameterizes the reward as a linear combination of a set of weights and a low-dimensional feature encoding of the state <cit.>. This approach can be beneficial because the posterior inference is over a lower dimensional reward vector. More recent work builds on this approach and proposes a method that enables imitation learning in complex control problems <cit.>. All of these techniques are best suited for environments with a closed-loop controller that provides instant feedback. § METHODS §.§ Conditional kernel density estimation As discussed earlier, directly estimating the likelihood function can lead to theoretical guarantees of posterior consistency. To estimate the likelihood p(s,a | R), we first observe that it can be viewed as the conditional density of the state-action pair given the reward function. Thus, any appropriate conditional density estimator could be applied; examples include the conditional kernel density estimator (CKDE, <cit.>) and Gaussian processes (GPs). We adopt the CKDE because it is nonparametric, has a closed form, and is straightforward to implement <cit.>. Motivated by the conditional probability equation p(y|x)=p(x,y)/p(x) (where x and y are two generic random variables), the CKDE estimates the conditional density p(y|x) by approximating the joint distribution p(x,y) and marginal distribution p(x) separately via kernel density estimation (KDE). Given pairs of observations {(x_j,y_j)}_j=1^m, the KDE approximations for the joint and marginal distributions are p(x,y)=1/m∑_j=1^m K(x-x_j/h) K'(y-y_j/h'), p(x)=1/m∑_j=1^m K(x-x_j/h), where K and K' are kernel functions with bandwidths h, h' > 0, respectively. To approximate the conditional density, the CKDE simply takes the ratio of these two KDE approximations: p(y|x) =p(x,y)/p(x) =∑_j=1^mK(x-x_j/h) K'(y-y_j/h')/∑_ℓ=1^m K(x-x_ℓ/h). §.§ Kernel Density Bayesian IRL We propose kernel density Bayesian inverse reinforcement learning (KD-BIRL), which uses a CKDE approximation p_m(s,a | R) to estimate the likelihood p(s,a | R). While the standard form of the CKDE (<Ref>) uses the difference between two samples (e.g., x - x_j) as input to the kernel functions, this difference can be replaced by any suitable distance metric <cit.>. To estimate the joint and marginal distributions, p(s, a, R) and p(R), we must specify two distance functions: one for comparing state-action tuples and one for comparing reward functions. We denote these as d_s: (𝒮×𝒜) × (𝒮×𝒜) →ℝ_+ and d_r: ℛ×ℛ→ℝ_+, respectively, and we discuss specific choices for them later. The CKDE approximation is then p_m(s,a | R) = p_m(s,a, R)/p_m(R) = ∑_j=1^m K(d_s((s,a), (s_j,a_j))/h) K'(d_r(R, R_j)/h') /∑_ℓ=1^m K(d_r(R, R_l)/h') , where h, h' > 0 are the bandwidth hyperparameters. Note that fitting a CKDE for the likelihood requires estimating the density across a range of reward functions and state-action pairs. To better enable this, we construct an additional set of demonstrations and reward functions – which we call the training dataset {(s_j,a_j, R_j)}_j=1^m – to augment the observed expert demonstrations {(s_i^e, a_i^e)}_i=1^n (from an agent acting according to the data generating reward R^⋆). The training dataset contains demonstrations from agents whose policies optimize for reward functions that are likely distinct from those of the expert. Each sample in the training dataset is a state-action pair associated with a reward function. There will be many state-action pairs that correspond to the same reward function; therefore, R_j is not unique. In a simulated setting where the training demonstrations are not available already, we choose k training set reward functions, learn k optimal policies that optimize for each of these functions, and generate ⌊ m/k ⌋ demonstrations from each policy. We sample the reward functions R_1, …, R_k ∼ u, where u is a distribution on the space of reward functions ℛ. The resulting density estimate is more accurate when R_1, …, R_k are uniformly distributed across ℛ. As such, in our experiments with simulated environments, we use a uniform distribution for u. Using the CKDE in <Ref>, we can now estimate the posterior density function of R given n expert demonstrations, m training demonstrations, and prior p(R): p_m^n(R|{s_i^e,a_i^e}_i=1^n) ∝ p(R) ∏_i=1^n p_m(s_i^e,a_i^e | R) = p(R) ∏_i=1^n ∑_j=1^m K(d_s((s_i^e,a_i^e), (s_j,a_j))/h) K'(d_r(R, R_j)/h') /∑_ℓ=1^m K(d_r(R, R_l)/h') . The choice of the prior, p(R), and the distance metrics d_s, d_r can be altered depending on information known about the reward function or state space in advance <cit.>. For example, if the reward function is assumed to be a linear function of the state, the cosine distance is more appropriate for d_r. Several non-uniform priors may be appropriate for p(R) depending on the characteristics of the MDP, including Gaussian <cit.>, Beta <cit.>, and Chinese restaurant process (CRP) <cit.>. In KD-BIRL, the kernel K(x)=exp(-x^2) is chosen to be Gaussian. We choose a Gaussian kernel because it can approximate bounded and continuous functions well. The bandwidth hyperparameters can be chosen using rule-of-thumb procedures <cit.>. To infer the posterior estimate in <Ref>, we sample rewards using a Hamiltonian Monte Carlo algorithm <cit.> (additional details in Appendix Section 9). Note a key computational gain of our approach over BIRL, which is also a sampling-based algorithm: we only use forward RL to generate the training dataset (in the case of simulated environments or when it is not already present), and avoid it in each iteration of MCMC. This is possible because <Ref> does not depend on Q^⋆. §.§.§ Feature-based reward function While KD-BIRL can be applied as-is in many environments, the CKDE is known to scale poorly to high-dimensional functions. Thus, before KD-BIRL can work in environments with large state spaces where the corresponding reward function has a large number of parameters, it is necessary to re-parameterize the reward function. In our existing formulation, the reward function is parameterized as a vector where each index corresponds to the reward in one of the states. Under this formulation, in a 10×10 Gridworld environment, the reward function would be represented as a vector of length 100. In practice, the CKDE increases in computational cost with respect to both the length of the vector and the number of samples in the expert and training datasets <cit.>, and it would not be suited to learn 100 parameters. We propose a formulation of KD-BIRL that uses a feature-based reward function <cit.>. This method of parameterizing a reward is one of three broad categories of IRL formulations <cit.>. The feature-based reward function R(s,a) = w^⊤ϕ(s,a) where w ∈ℛ^q and ϕ: S × A →ℛ^q is advantageous because it does not scale with the dimensionality of the state s and does not rely on the state space being discrete like our earlier approach. Here, ϕ is a known function that maps a state-action tuple to a feature vector of length q. Intuitively, this feature vector is a low-dimensional representation of the original state that facilities reward inference. In this setup, the goal is to find w^⋆ such that: w^⋆⊤ E[∑_t=0^∞γ^t ϕ(s_t, a_t) | π^⋆] ≥ w^⋆⊤ E[∑_t=0^∞γ^t ϕ(s_t, a_t) | π] where γ is a discount factor, π is a policy, and s_t, a_t is the state and action at time t. In a Bayesian setting, the resulting posterior is over w rather than R. Now we generate n expert and m training demonstrations. Recall that the CKDE requires as input the reward function parameters corresponding to each training dataset sample. A given training dataset sample here is {(s_j,a_j,w_j)} where w_j is the weight vector of length q associated with the reward function that was used by the agent to generate the sample s_j, a_j. w_j can be repeated and is not unique. A sample from the expert demonstration dataset is still {(s_i^e, a_i^e)}_i=1^n, where the data-generating weights w^⋆ are used to generate these demonstrations. The procedure for learning the CKDE and the resulting posterior inference then stays the same. The CKDE formulation is now: p_m(s,a | w) = p_m(s,a, w)/p_m(w) = ∑_j=1^m K(d_s((s,a), (s_j,a_j))/h) K'(d_r(w,w_j)/h') /∑_ℓ=1^m K(d_r(w,w_l)/h') , where d_r measures the similarity between weight vectors, and d_s is the distance between state-action tuples. The posterior is then: p_m^n(w|{s_i^e,a_i^e}_i=1^n) ∝ p(w) ∏_i=1^n p_m(s_i^e,a_i^e | w) = p(w) ∏_i=1^n ∑_j=1^m K(d_s((s_i^e,a_i^e), (s_j,a_j))/h) K'(d_r(w,w_j)/h') /∑_ℓ=1^m K(d_r(w,w_l)/h') . § THEORETICAL GUARANTEES OF KD-BIRL KD-BIRL estimates the density function of a true Bayesian posterior distribution (<Ref>), so we can reason about the posterior's asymptotic behavior. In particular, we want to ascertain that this posterior estimate contracts as it receives more samples. Because the IRL problem is non-identifiable, the “correct” reward function as defined by existing methods <cit.>, may not be unique. In this work, we assume that any two reward functions that lead an agent to behave in the same way are equivalent. Said another way, if a set of observations is equally likely under two reward functions, the functions are considered equal: R_1≃ R_2 if p(·|R_1)-p(·|R_2)_L_1 = 0. We can then define the equivalence class [R^⋆] for R^⋆ as [R^⋆] = {R ∈ℛ : R ≃ R^⋆}. An ideal posterior distribution places higher mass on reward functions in the equivalence class [R^⋆]. We first focus on the likelihood estimation step of our approach and show that, when the size of the training dataset m approaches ∞ and the m samples arise from sufficiently different reward functions that cover the space of ℛ, the likelihood estimated using a CKDE (Equation (<ref>)) converges to the true likelihood p(s,a | R). Let h_m, h_m'>0 be the bandwidths chosen for the CKDE. Assume that both p(s,a | R) and p(R) are square-integrable and twice differentiable with a square-integrable and continuous second order derivative, and that mh_m^p/2→∞ and mh_m^'^p/2→∞ as m→∞. Then, p_m(s,a|R)p(s,a | R), ∀(s,a,R)∈𝒮×𝒜×ℛ. <Ref> verifies that we can estimate the likelihood using a CKDE, opening the door to Bayesian inference. We now show that as n, the size of expert demonstrations, and m, the size of the training dataset, approach ∞, the posterior distribution generated using KD-BIRL contracts to the equivalence class of the expert demonstration generating reward [R^⋆]. Assume the prior for R, denoted by Π, satisfies Π({R:KL(R^⋆,R)<ϵ})>0 for any ϵ>0, where KL is the Kullback–Leibler divergence. Assume ℛ⊆ℝ^d is a compact set. Then, the posterior measure corresponding to the posterior density function p_m^n defined in <Ref>, denoted by Π_m^n, is consistent w.r.t. the L_1 distance; that is, Π^n_m({R:p(·|R)-p(·|R^⋆)_L_1<ϵ)}) 1. <Ref> implies that the posterior Π_m^n assigns almost all mass to the neighborhood of [R^⋆]. This means that the reward function the KD-BIRL posterior contracts to with a large enough sample size is practically equivalent to the data-generating reward function R^⋆. Note that this not a statement regarding the posterior contraction rate, just a certification of contraction. Proofs for both <Ref> and <Ref> are in Appendix Section 4. § EXPERIMENTS Here, we evaluate the accuracy and computational efficiency of KD-BIRL. We compare KD-BIRL to AVRIL <cit.>, a recent method that that simultaneously learns an imitator policy and performs reward inference on the expert demonstrations, and the original Bayesian IRL algorithm (BIRL) <cit.>. We demonstrate results using a Gridworld environment <cit.> and a sepsis management clinical environment <cit.>. To quantitatively evaluate the reward functions learned by IRL methods, previous studies have used Expected Value Difference (EVD) <cit.>. EVD is defined as |V^⋆(r^A) - V^π^⋆(r^L)(r^A)| where V^π = ∑_s p_0(s) V^π is the value of policy π with initial state distribution p_0, r^A is the true data-generating reward, r^L is the learned reward, and V^⋆ is the value function associated with the optimal policy π^⋆. Intuitively, the EVD measures the difference in reward obtained by an agent whose policy is optimal for the true reward and the reward obtained by an agent whose policy is optimal for the learned reward. We use EVD because it allows us to compare KD-BIRL to related methods without needing to directly compare two reward function samples based on their functional form. The lower the EVD, the better our learned reward recapitulates the expert reward (see Appendix Section 10 for more details). §.§ Gridworld environment We begin our analysis in a Gridworld environment. The MDP here is defined by the grid's g× g discrete state space 𝒮 where a given state is represented as a one-hot encoded vector in ℛ^g× g, e_i, where the i'th index is 1 and corresponds to the state in which the agent is in, and g is the size of the grid; the action space contains 5 possible actions {, , , , }, each represented as a one-hot encoded vector; and the true reward function R^⋆, which is unobserved by the IRL algorithms, is a vector of length g × g. We structure the reward function such that each state has an independent scalar reward parameter. We specify the domain of each of these parameters to be the unit interval; thus, each feasible reward function can be represented by a vector R ∈ [0, 1]^g× g. To fit KD-BIRL we use Stan <cit.>, which uses a Hamiltonian Monte Carlo algorithm. To fit the BIRL and AVRIL posteriors, we first generate the same number of expert demonstration trajectories as used for KD-BIRL. BIRL and AVRIL use an inverse temperature hyperparameter, α; we set α=1 for all methods. AVRIL uses two additional hyperparameters γ, δ, which we set to 1. Unless otherwise specified, KD-BIRL uses a uniform prior for the reward r_s ∼Unif(0, 1) for s=1,…,g × g and Euclidean distance for d_s, d_r. §.§.§ Visualizing KD-BIRL's posterior distribution First, we visualize KD-BIRL's posterior distribution in comparison to those recovered by AVRIL and BIRL. To do this, we show density plots of samples from all three distributions marginalized at each state in the 2×2 Gridworld environment. All methods use a data-generating reward function R^⋆ = [0, 0, 0, 1]. We find that the posterior samples from KD-BIRL and BIRL are more concentrated around R^⋆ than those from AVRIL (<Ref>). §.§.§ KD-BIRL requires fewer instances of Q-learning Next, we quantify the computational complexity associated with performing reward inference in a 4×4 Gridworld in which only the state [3,3] contains a reward of 1. As discussed earlier, much of the computational cost associated with learning a posterior distribution in existing methods arises from repeated instances of forward RL. BIRL requires forward RL during every iteration of MCMC sampling; several thousand iterations are required for the sampler to converge. AVRIL uses one instance of forward RL to learn an approximate posterior. KD-BIRL also minimizes the use of forward RL, only using it during dataset generation, in the case that these observations are not already available. Here, we vary the number of iterations of forward RL and plot the EVDs for reward samples from the resulting posterior distributions for the three methods. Our results indicate that with fewer instances of forward RL, KD-BIRL reward samples better replicate the behavior of the expert demonstrations than those of BIRL (partly because the x-axis implies too few iterations of MCMC sampling); consequently, even though AVRIL requires fewer instances of forward RL, it is at the expense of accuracy in the posterior distribution, as highlighted by the stagnant EVD (<Ref>). §.§ Limitations of the CKDE It is well known that CKDE has difficulty scaling to high-dimensional probability density functions <cit.>. Regardless, we want to identify the limits of the CKDE used in the original KD-BIRL setup without a feature-based reward function. To do so, we use a 5×5 Gridworld environment. In <Ref>, despite the fact that the number of reward parameters is larger than what we expect the CKDE to successfully model, KD-BIRL is able to estimate a posterior whose mean is in the equivalence class of R^⋆. That is, the posterior mean and R^⋆ encourage the same behavior in an agent, which implies that the expert demonstrations are equally likely under both. However, there are states ([2,3], [4, 3] in <Ref>, Panel 2) in the 5×5 Gridworld in which the mean estimated reward is notably incorrect, which indicates that the CKDE struggles to learn 25 independent reward parameters successfully. §.§ Feature-based rewards We now study three methods of reward function featurization that can enable KD-BIRL to perform reward inference in environments with large state spaces. §.§.§ Using known features in a 10x10 Gridworld As discussed in <Ref>, without using feature-based rewards, the original KD-BIRL algorithm would not be able to perform inference in the 10×10 Gridworld because the reward vector length (100) is too high. In the 10×10 Gridworld, the MDP is identical to the earlier Gridworld settings, except the state space is the series of one-hot encoded vectors of length 100. In this setting, we select ϕ(s) = [x, y] to be a simple function that ignores the action and maps the state vector of length 100 to the spatial coordinates of the agent. In this way, we treat the coordinates of the agent as a “feature vector”. Then, we choose weights w^⋆ such that R^⋆ is a linear combination of the features and w^⋆. <Ref> visualizes the resulting posterior distributions for two choices of w^⋆. We use a Normal prior for p(w) with mean 0 and variance 1 for w^⋆=[-1, 1], and a Normal prior with mean 0.5 and variance 0.5 for w^⋆=[1, 1]. We find that KD-BIRL accurately recovers the relative magnitude and sign of the individual components of w^⋆ for both chosen reward functions. §.§.§ Manually curated features in a sepsis treatment environment Now we perform inference in a sepsis treatment simulator based on de-identified data from MIMIC-III <cit.>, a database of electronic health records (EHR). Sepsis arises when the body responds to infection in a way that is harmful to its own tissues and organs <cit.>. In the original simulator <cit.>, the state is a vector of length 46, where each element contains information about a given physiological covariate. There are 25 possible actions, each corresponding to a different combination of treatments. The transition function was learned using deep RL models <cit.> (see Appendix Section 12 for details). Sepsis treatment depends heavily on organ failure metrics, and fast increases in these metrics warrant urgent care <cit.>. Since we have observed that in the Gridworld environment, KD-BIRL can successfully model a reward that is a function of a small state space, we choose ϕ to be a function that ignores the action and extracts three Sequential Organ Failure Assessment (SOFA) <cit.> covariates present in the state: sofa, quick sofa, and quick sofa systolic blood pressure score. The result is a feature-based reward function with manually selected features based on prior knowledge. Our reward is now a linear combination of the difference between the state features at time t and t+1, R(s_t) = [ a; b; c ]^⊤[ s(cov_1)_t-s(cov_1)_t+1; s(cov_2)_t-s(cov_2)_t+1; s(cov_3)_t-s(cov_3)_t+1, ] where [a, b, c] are the weights, and s(cov_1), s(cov_2), s(cov_3) are the three organ failure features in state s. We choose the true (unobserved) weights to be [a=0.8, b=0.6, c=0.4]. We compare our method to AVRIL and avoid fitting BIRL due to computational constraints. Our results indicate that KD-BIRL generates reward samples with lower and more concentrated EVDs than the AVRIL trajectories (<Ref>). This indicates that KD-BIRL estimates a posterior distribution that is concentrated around the equivalence class of R^⋆. §.§.§ Using a VAE to identify features in the sepsis environment Finally, we explore the use of a variational auto-encoder (VAE) to learn ϕ in the sepsis environment. More specifically, we use a VAE to learn a low-dimensional representation of state-action tuples, and aim to learn the set of weights that modifies this representation to form the reward function. To do this, we first learn ϕ on a set of state-action tuples independent of the training or expert demonstrations. The input dimension to the VAE is 47 (46 state features + 1 action), and the low dimensional representation has 3 features. The VAE uses 4 linear layers for the encoder and decoder, and optimizes for a downsampled representation with low reconstruction error using Adam. Once ϕ is known, it can be used to generate the required datasets. To do this, we first select a set of weights w^⋆ for the expert demonstrations, and generate state-action tuples that optimize R(s,a) where R(s,a) = ϕ(s,a) × w^⋆. We repeat this procedure for several sets of uniformly selected weights w_0, ⋯, w_c to generate the training dataset. Finally, we fit KD-BIRL and evaluate the learned weights using EVD as before. We report results in <Ref>, and find that across a variety of w^⋆ values, KD-BIRL's posterior samples generate comparable, if not much lower values than AVRIL. This, coupled with the additional theoretical guarantees, makes KD-BIRL a good choice for performing IRL in complex environments. § DISCUSSION AND CONCLUSION In this work, we present kernel density Bayesian inverse reinforcement learning (KD-BIRL), an efficient IRL algorithm that improves upon existing methods by estimating a posterior distribution on the reward function while avoiding Q-learning for every iteration of MCMC sampling, and by providing theoretical guarantees of posterior consistency. We show that KD-BIRL generates concentrated posteriors and is more computationally efficient than existing methods in a Gridworld environment. Additionally, we demonstrate that with a feature-based reward function, KD-BIRL can perform inference in a complex healthcare environment, and the resulting posterior outperforms a leading method. Taken together, our results suggest that, in complex environments, KD-BIRL can enable an accurate probabilistic description of clinician objectives that is not possible with current methods. Several future directions remain. This work is best-suited for on-policy (i.e., simulation) environments, and additional work is necessary to apply it directly to off-policy environments such as retrospective clinical decision-making settings. In particular, we would need behavior demonstrations from multiple agents in order to define a training dataset. Thus, it will be necessary to be able to associate actions with an agent (i.e., clinician). Additionally, the particular choices of distance metrics and hyperparameters used in the CKDE depend on the environment and reward function parameterization; additional experimentation is required to adapt this to different environments. Furthermore, a limitation of the conventional CKDE is that it performs poorly in high dimensions. One solution is to consider a modified version of the CKDE to speed it up <cit.>. Another solution is to replace the CKDE with another nonparametric conditional density estimator <cit.>. Because KD-BIRL is a framework that estimates the likelihood as a conditional density, it can be easily modified to accommodate other choices for the CKDE. Finally, in this work, we discuss efforts to re-parameterize the reward function, and it is of interest to apply this work in additional environments with continuous or infinite state spaces, such as real-world EHR. We thank Alex Chan for providing code associated with the AVRIL method. This work was funded by the Helmsley Trust grant AWD1006624, NIH NCI 5U2CCA233195, NIH NHLBI R01 HL133218, and NSF CAREER AWD1005627. BEE is on the SAB of Creyon Bio, Arrepath, and Freenome. A. Mandyam was supported in part by a Stanford Engineering Fellowship. D. Cai was supported in part by a Google Ph.D. Fellowship in Machine Learning. § CODE Our experiments were run on an internally-hosted cluster using a 320 NVIDIA P100 GPU whose processor core has 16 GB of memory hosted. Our experiments used a total of approximately 200 hours of compute time. Our code uses the MIT License and is available at https://github.com/bee-hive/kdbirlhttps://github.com/bee-hive/kdbirl. § RATIONALITY OF BAYESIAN IRL The original Bayesian IRL algorithm uses a Q-value function as a component of the likelihood calculation. While this can be sufficient for imitation learning, it does not allow for posterior predictive sampling, and updates to the Gibbs posterior are not rational unless the Q-value function satisfies certain conditions <cit.> (see <Ref>). Intuitively, rational updates imply that new evidence is appropriately incorporated to modify the posterior from the prior. The Q-value function can be approximated using any real-value prediction method from neural networks to linear regression, and the rationality of the resulting posterior updates is not discussed in prior work. A rational posterior update implies that if the updating function, in this case, Q^⋆, has a lower value, the posterior probabilities should be lower, and vice versa. However, there is no way to confirm that Q^⋆ satisfies Assumption 3 of the guidelines for updating belief distributions in <cit.>. Furthermore, a Q-value function is a poor approximation of density or likelihood. A higher Q-value associated with a state means that the Q-learning algorithm finds that this state is more likely to generate a reward. However, this does not imply that an agent is more likely to visit that state. Bayesian IRL formulations that use a Q-value function equate these two phenomena, and this can be incorrect depending on (1) the optimality of such an agent, (2) the quality of the Q-learning estimator, and (3) whether the agent has even visited the state. § RATIONALITY CONDITIONS FOR POSTERIOR DISTRIBUTION UPDATES  <cit.> make several assumptions before guaranteeing rational updates to a posterior distribution. These are: ψ[l(θ, x_2), ψ{l(θ, x_1), π(θ)}] ≡ψ{l(θ, x_1) + l(θ, x_2), π(θ)} where θ is the parameter of interest, π(θ) is the prior distribution on θ, l is a loss function, ψ is the update function, and x_1, x_2 are data points. For any set A ⊂Θ, ψ{l(θ, x), π(θ)}/∫_A ψ{l(θ, x), π(θ)} dθ = ψ{l(θ, x), π_A(θ)} where π_A is π normalized to A. Lower evidence for a state should yield smaller posterior probabilities under the same prior. So, if for some A⊂Θ, l(θ, x) > l(θ, y) for θ∈ A ⊂Θ and l(θ, x) = l(θ, y) for θ∈ A^c, then ∫_A ψ{l(θ, x), π(θ)} dθ < ∫_A ψ{l(θ, y), π(θ)}dθ. If l(θ, x) ≡ constant, then ψ{l(θ, x), π(θ)}≡π(θ). If l̃(θ, x) = l(θ, x) + c for some constant c, then ψ{l̃(θ, x), π(θ)} = ψ{l(θ, x), π(θ)}. If ψ is a Q-value function, which can be parameterized by anything from a linear model to a deep neural network, the rationality of the subsequent posterior updates is not discussed in previous work. In particular, it is not possible to verify these assumptions for a Q-value function. § PROOFS FOR LEMMA 4.1, THEOREM 4.2 The proof associated with Lemma 4.1 follows. We now use the continuity of the likelihood, the finite sample analysis of multivariate kernel density estimators in <cit.>[Section 4.4, Equation 4.16](<Ref>), which defines the Mean Integrated Square Error (MISE) of the density function, and Theorem 1(<Ref>) of <cit.>[Section 2.6-2.9], which asserts that as the sample size increases, the mean of the density estimator converges and variance prevents the mean from exploding. We can use Theorem 1 because we assume that the density function is square-integrable and twice differentiable and that the bandwidth approaches 0 as the dataset size increases. Then, up to a constant, for a given state-action pair (s, a), 1/m∑_j=1^m e^-d_s((s,a),(s_j,a_j))^2/(2h)e^-d_r(R,R_j)^2/(2h')p(s,a,R). The same holds true for d_r, 1/m∑_ℓ=1^m e^-d_r(R,R_ℓ)^2/(2h')p(R). By the Continuous Mapping Theorem <cit.>, we conclude that p_m(s,a|R)p(s,a,R)/p(R)=p(s,a|R). The proof associated with Theorem 4.2 follows. By Lemma 4.1, as m→∞, p_m converges to the true likelihood, so we can adopt existing tools from Bayesian asymptotic theory. We first define an equivalence relation on ℛ, denoted by ≃: R_1≃ R_2 iff p(·|R_1)=p(·|R_2), a.e. Note that ≃ satisfies reflexivity, symmetry, and transitivity and is, therefore an equivalence relation. We denote the equivalence class by [·], that is, [R]={R':R'≃ R}, and the quotient space is defined as ℛℛ / ≃ = {[R]:R∈ℛ}. The corresponding canonical projection is denoted by : ℛ→ℛ,  R↦[R]. Then, the projection induces a prior distribution on ℛ denoted by Π: Π(A)Π(^-1(A)). Moreover, ℛ admits a metric d: d([R_1],[R_2])p(·|R_1)-p(·|R_2)_L^1. Because this metric uses the L^1 norm, it satisfies symmetry and triangular inequality. Additionally, it is true that d([R_1],[R_2])=0⟺ p(·|R_1)=p(·|R_2), a.e. ⟸ R_1≃ R_2⟺ [R_1]=[R_2], so d fulfills the identity of indiscernibles principle. As a result, d is a valid distance metric on ℛ. Then consider the following Bayesian model: (s,a)|[R]≃ p(s,a|[R]), [R]∈ℛ, [R]≃Π. This model is well-defined since p(s,a|[R]) is independent of the representative of [R] by the definition of the equivalence class. Observe that KL(R,R^*)=KL([R],[R^*]) by the definition of the equivalence class. Then, let A={[R]:KL([R],[R^*])<ϵ}⊂ℛ. We can define ^-1(A) = {R∈ℛ:KL(R,R^*)<ϵ}⊂ℛ. As a result, Π({[R]:KL([R],[R^*])<ϵ})=Π({R:KL(R,R^*)<ϵ})>0 for any ϵ>0, that is, the KL support condition is satisfied . Moreover, the mapping [R]→ p(·|R) is one-to-one. Because the Bayesian model is parameterized by [R] and we assume that ℛ is a compact set, by  <cit.>[Lemma 10.6](<Ref>) there exist consistent tests as required in Schwartz's Theorem . Then, by <cit.>(<Ref>), the posterior Π_n on ℛ is consistent. That is, for any ϵ>0, Π^n_m({[R]:d([R],[R^*])<ϵ}) 1. Put in terms of the original parameter space, Π_m^n({R:p(·|R)-p(·|R^⋆)_L_1<ϵ)})=Π^n_m({[R]:d([R],[R^*])<ϵ}), ∀ϵ>0. § ASYMPTOTIC EXPANSION OF THE MEAN INTEGRATED SQUARED ERROR THEOREM 1 (i) The integrated squared bias of the kernel density estimator can be expanded as ISB{f̂(·;H)} = 1/4 c_2(K)^2vec^⊤R(D^⊗ 2f)(vec H)^⊗ 2 + o(||vec H||^2). (ii) The integrated variance of the kernel density estimator can be expanded as IV{f̂(·;H)} = m^-1 |H|^-1/2 R(K) + o(m^-1|H|^-1/2). Here, f̂ is the estimated density function, H is a matrix of bandwidth values, c_2(K) = ∫_ℛ^d z_i^2 K(z) dz for all i=1, …, d is the variance of the kernel function K, and m is the size of the training dataset. In our work, H is a diagonal matrix where every element on the diagonal is the same bandwidth h_m. In this work, we assume that f is square-integrable and twice differentiable and that the bandwidth matrix H→ 0 is m →∞. Because we use a Gaussian kernel for K, we know that it is square integrable, spherically symmetric, and has a finite second-order moment. § WAND AND JONES, EQUATION 4.16 The mean integrated squared error (MISE) of a multivariate kernel density estimator is defined as: MISE{f̂(·;H)} = n^-1 (4π)^-d/2 |H|^-1/2 + w^⊤{(1-n^-1)Ω_2 - 2Ω_1 + Ω_0}w where H is a matrix of bandwidth values, n is the size of the dataset, Ω_a denotes the k× k matrix with (l, l') entry equal to ϕ_a H + Σ_l + Σ_l' (μ_l - μ_l'), ϕ_d is a d-variate Normal kernel, w=(w_1, …, w_k)^⊤ is a vector of positive numbers summing to 1, and for each l=1, …, k, μ_l is a d × 1 vector and Σ_l is a d× d covariance matrix. § SCHWARTZ'S THEOREM If p_0 ∈ KL(Π) and for every neighborhood 𝒰 of p_0 there exist tests ϕ_n such that P_0^nϕ_n → 0 and sup_p ∈𝒰^c P^n(1 - ϕ_n) → 0, then the posterior distribution Π_n(·| X_1, …, X_n) in the model X_1, …, X_n | p ∼^iid p and p ∼Π is strongly consistent at p_0. § ASYMPTOTIC STATISTICS, LEMMA 10.6 Suppose that Θ is σ-compact, P_θ≠ P_θ' for every pair θ≠θ', and the maps θ→ P_θ are continuous for the total variation norm. Then there exists a sequence of estimators that is uniformly consistent on every compact subset of Θ. Here, Θ is the space of parameters, P is the probability density function, and θ∈Θ is a parameter. § KD-BIRL ALGORITHM <Ref> is the general version of the KD-BIRL algorithm where the expert demonstrations and training dataset are already available. A version of the algorithm that is suited for simulated datasets which require dataset generation can be seen in <Ref>. § CALCULATING EXPECTED VALUE DIFFERENCE (EVD) The procedure to calculate EVD varies depending on the method. For all methods, this process requires a set of reward samples. Because KD-BIRL and BIRL both use MCMC sampling, we can use the reward samples generated from each iteration. AVRIL does not use MCMC, so we have to modify the approach to generating samples depending on the structure of the reward function. When the reward function is a vector with length equal to the cardinality of the state space, we use the AVRIL agent to estimate the variational mean and standard deviation of reward at each state in the environment. Using these statistics, we then assume that the reward samples arise from a multivariate normal distribution and use numpy to generate samples according to the mean and standard deviation information collected. Once we have samples, we can then calculate EVD and 95% confidence intervals. Recall that EVD is defined as |V^*(r^A) - V^π^*(r^L)(r^A)| where r^A is the ground truth known reward and r^L is the learned reward. For a given method, we calculate 95% confidence intervals across the EVD for each of the reward samples. Given a reward sample, we can calculate EVD, where r^L is the sample, and r^A is the known reward. To determine the value of the policy optimizing for a particular reward function, we train an optimal agent for that reward function, and generate demonstrations characterizing its behavior; the value is then the average reward received across those demonstrations. Finally, we calculate the difference between the value of the policy for r^A and r^L, and report confidence intervals across these values for all samples for a given method. In the Sepsis environment, we use two methods of reward function featurization. In both, because the reward function does not consider individual discretized states, we cannot use the earlier approach to calculate EVD for AVRIL. To generate EVDs, we used the trained AVRIL agent to generate trajectories using agent-recommended actions starting from an initial state; the EVD here is the difference between the total value of these trajectories and the total value of trajectories generated using the correct reward function (independent of AVRIL). § CHOOSING BANDWIDTH HYPERPARAMETERS FOR KD-BIRL Now, we investigate the effect of the bandwidth hyperparameters h and h' on KD-BIRL's posterior distribution for the reward function R^⋆= [0, 0, 0, 1], marginalized at state [1, 1] (<Ref>). Recall that h and h' correspond to the bandwidth for the state-action pairs and reward functions, respectively. Note that the density of reward samples varies more with h' than h, but it is important to tune both these hyperparameters because the resulting posterior distributions can change substantially. § SEPSIS ENVIRONMENT The Sepsis environment is a simulation setting that models Sepsis treatment. There are 48 state features(<Ref>) in the original environment, comprised of 46 physiological covariates and an action and state index. Feature Description Albumin Measured value of Albumin, a protein made by the liver Anion Gap Measured difference between the negatively and positively charged electrolytes in blood Bands Measuring band neutrophil concentration Bicarbonate Measured arterial blood gas Bilirubin Measured bilirubin BUN Measured Blood Urea Nitrogen Chloride Measured chloride Creatinine Measured Creatinine DiasBP Diastolic blood pressure Glucose Administered glucose Glucose Measured glucose Heart Rate Measured Heart Rate Hematocrit Measure of the proportion of red blood cells Hemoglobin Measured hemoglobin INR International normalized ratio Lactate Measured lactate MeanBP Mean Blood Pressure PaCO2 Partial pressure of Carbon Dioxide Platelet Measured platelet count Potassium Measured potassium PT Prothrombin time RespRate Respiratory rate Sodium Measured sodium SpO2 Measured oxygen saturation SysBP Measured systolic blood pressure TempC Temperature in degrees Celsius WBC White blood cell count age Age in years is male Gender, true or false race Ethnicity (white, black, hispanic or other) height Height in inches Weight Weight in kgs Vent Patient is on ventilator SOFA Sepsis related organ failure score LODS Logistic organ disfunction score SIRS Systemic inflammatory response syndrome qSOFA Quick SOFA score qSOFA Sysbp Score Quick SOFA that incorporates systolic blood pressure measurement qSOFA GCS Score Quick SOFA incorporating Glasgow Coma Scale qSofa Respirate Score Quick SOFA incorporating respiratory rate Elixhauser hospital Hospital uses Elixhauser comorbidity software Blood culture positive Bacteria is present in the blood The original Sepsis environment uses the features described in <Ref> to represent the patient state. In the first method of featurization we consider in our experiments, we consider only three features that directly affect the reward received in a given state. These are SOFA, qSOFA, and qSOFA Sysbp Score. Intuitively, these can be thought of as features of the state that are most relevant to the reward function. In the second method of featurization, we consider all of the state features as well as the action chosen, and use a variational auto-encoder to generate a low dimensional representation.

http://arxiv.org/abs/2303.07047v1

20230313121326

Optimization of Velocity Ramps with Survival Analysis for Intersection Merge-Ins

cs.AI

Optimization of Velocity Ramps with Survival Analysis for Intersection Merge-Ins Tim Puphal^1†, Malte Probst^1†, Yiyang Li^2, Yosuke Sakamoto^2 and Julian Eggert^1† ^1 Honda Research Institute (HRI) Europe, Carl-Legien-Str. 30, 63073 Offenbach, Germany Email: {tim.puphal, malte.probst, julian.eggert}@honda-ri.de ^2 Honda Innovation Lab (HIL) Tokyo, Honda R&D Co., Ltd. 5-3-1 Akasaka, 107-6327 Tokyo, Japan Email: {Yiyang_Li, Yosuke_Sakamoto}@n.t.rd.honda.co.jp † The authors contributed equally to this work Received; accepted ============================================================================================================================================================================================================================================================================================================================================================================================================================================================== We consider the problem of correct motion planning for T-intersection merge-ins of arbitrary geometry and vehicle density. A merge-in support system has to estimate the chances that a gap between two consecutive vehicles can be taken successfully. In contrast to previous models based on heuristic gap size rules, we present an approach which optimizes the integral risk of the situation using parametrized velocity ramps. It accounts for the risks from curves and all involved vehicles (front and rear on all paths) with a so-called survival analysis. For comparison, we also introduce a specially designed extension of the Intelligent Driver Model (IDM) for entering intersections. We show in a quantitative statistical evaluation that the survival method provides advantages in terms of lower absolute risk (i.e., no crash happens) and better risk-utility tradeoff (i.e., making better use of appearing gaps). Furthermore, our approach generalizes to more complex situations with additional risk sources. § INTRODUCTION Most Advanced Driver Assistance Systems (ADAS) enable semi-autonomous driving at low velocities in parking areas or at high velocities on highways, but not for moderate velocities in complex inner-city scenarios <cit.>. They are separably developed, only reactive and not predictive and do not adapt to the driver's needs. Especially at intersections, the interplay of vehicles plays a dominant role and ADAS face a wide range of risk types (e.g. collision, curve, occlusion and traffic rules). This requires an ADAS that employs behavior prediction for risk estimation and subsequently plans safe behaviors. It has to generalize over scenarios and holistically incorporate different risk sources. By optimizing integral risk and benefit factors over the prediction horizon employing the survival analysis <cit.>, we compute trajectories consisting of velocity ramps on the ego path. The performance of ROPT is evaluated for merge-ins at unsignalized T-intersections. Merge-in support systems have sequential planning requirements due to curve taking and multiple other vehicles. We compare ROPT with an extended intersection version of the Intelligent Driver Model (IDM) <cit.> using a statistical analysis. Particularly, we look at mean trends of gap number and gap size as well as of minimum back and front distance for varying parameter values of the models. The next Section <ref> gives an overview of related work focusing on the research community. The introduction of the Intersection IDM (IIDM) is divided into Section <ref> on longitudinal dynamics and Section <ref> on the lateral risk extension. We then explain more in detail ROPT's trajectory generation and optimization in Section <ref> and <ref>. Finally, Section <ref> shows the experiment setup together with results analysis and Section <ref> a summary and discussion for future developments. §.§ Related Work Besides the automotive domain, numerous approaches for motion planning exist in robotics and physics. A survey of state of the art was conducted in <cit.>. Traffic simulators are categorized into microscopic and macroscopic models. While microscopic models, such as IDM, treat each car's dynamics seperately, macroscopic models look at traffic density, flow and average velocity with fluid dynamic equations <cit.>. A subclass of microscopic models are cellular models, which discretize the time and space and thus work also well for larger road networks <cit.>. The remaining planning methods are either using trajectory optimization or search algorithms. In <cit.>, velocity profiles of trapezoidal shapes are optimized with Model Predictive Control (MPC) and then evaluated against their proximity to dynamic obstacles, their smoothness and speed. Supplementary, <cit.> evaluates besides the costs of the ego car, costs from the perspective of other cars to find a cooperative trajectory, which is applied to traffic scenes with priority orders. As a risk indicator, the Time-To-Collision (TTC) zone is used. Crossing of intersections can also be learned based on potential fields and the Levenberg-Marquardt method <cit.>. Especially for T-intersections, <cit.> obtain personal critical gaps with Maximum Likelihood estimation and give out recommendations of safe gaps ahead. Standard search procedures find a path through static obstacle maps, but do not consider dynamic entities over time. For example <cit.> combine task planning with spatial exploration to detect free areas with circle-shaped spatial probability propagation. For this reason, Optimal Reciprocal Collision Avoidance (ORCA) is applied in <cit.> to search the velocity instead of position space in all directions and retrieve a collision-free velocity vector. Similarly, the authors of <cit.> connect velocities along obstacle tangent points in the path-time space under acceleration bounds. § INTERSECTION IDM §.§ Basics and Curve Driving The IDM <cit.> is a popular traffic model describing the dynamics of a leading and following vehicle pair driving along the same straight path. Its differential equation outputs a safe and efficient acceleration profile v̇_f for the follower v̇_f = a(1-(v_f/v_c)^δ)_ - a(d_0+v_fT/d+v_f(v_f-v_l)/2d√(ab))^2_. In general, the follower approaches the cruising velocity v_c with the maximal acceleration a and acceleration exponent δ in the free term. However, once another vehicle is in front at a distance d with velocity v_l, the interaction term reduces v̇_f to reach a balance point defined by the minimal distance d_0 and time headway T. The contained reaction time can be set with the desired deceleration b. Similar to <cit.>, we additionally account for curves in the path by searching for the next maximal curvature κ_ in a curve segment [κ_, κ_]. The resulting maximal lateral acceleration a_y leads to an altered desired velocity v_c = √(a_y/κ_), if κ_ > κ_ and κ_ < κ_. Hence, on sharp curve segments exceeding the threshold κ_, the follower will converge to lower v_f due to the free term. §.§ Lane Projection Since the IDM is suitable for the simulation of longitudinal scenarios, it is often used in freeways. To improve the applicability, lane-change decisions have been modeled in <cit.>. Here, the accelerations from the IDM are compared between the driver to the leader and the follower on the current as well as the adjacent lane. If the acceleration gain of the driver is higher than the loss of his followers, a lane change is executed. We extend this idea for the crossing of intersections. First, we find the intersecting point between the current and other path and obtain the distance d_I of the driver to the point. At segments close to the intersection, the position of the driver is projected and shifted along d_I onto the other path. Then, we locate leading and following cars around this “hypothetical” driver. The IDM equations are taken for the driver relative to the leader and the follower relative to the driver to retrieve a_d and a_f, respectively. For the case of the current path, we assume constant acceleration for the follower a_f and a stop maneuver for the driver a_d. If the incentive criterion a_d - a_d_ + p(a_f - a_f)_ > Δ a_ is not fulfilled, the driver performs a_d. Otherwise the driver crosses the intersection by using a_d. The politeness factor p and threshold Δ a_ modify the needed advantage at which the the driver will pass in front of the follower. With the incentive criterion alone, it is possible that the follower crashes into the driver for small p. We therefore check simultaneously the safety criterion a_f ≥ b_s with a safe deceleration b_s. The resulting method is called the IIDM. § RISK OPTIMIZATION METHOD In each time step of the simulation, ROPT receives information about the current state of the environment. This includes the latest measured position and velocity of all traffic participants as well as their associated future paths extracted from map data. The goal of ROPT is to predict an optimal velocity profile for the ego car along its path. The computation consists of three basic steps. Initially, a trajectory for each other car is extrapolated over the prediction horizon.[ROPT uses a constant velocity assumption. However, it is also possible to use other methods, e.g. constant acceleration, or variants.] Second, a set of potential ego trajectories are created. Third, one of the created ego trajectories is selected based on its integral risk (caused by curve and collision), utility (distance travelled) and comfort (strength and frequency of velocity change). Since we use an optimization algorithm for the trajectory generation, steps two and three are heavily intertwined. §.§ Trajectory Optimization For scenarios with only one risk source (i.e., car following, curve driving and driving straight on an intersection), basic trajectory sampling methods are usually sufficient. Accordingly in previous research, the Foresighted Driver Model (FDM) <cit.> was developed. It samples via gradient descent an acceleration and deceleration profile and balances risk with utility to find correct ego behaviors. Merge-ins at intersections are however more complex scenarios in which the planned velocity profile has to obey maximum curve speed and match the speed of the traffic flow. The FDM would only converge to the local minimum from the curve risk. Here, it is necessary to generate a velocity profile that considers multiple spatio-temporal risk sources. In a different approach, we incrementally constructed velocity paths through predictive risk maps with Rapidly-exploring Random Trees (RRT) <cit.>. It was proven to work in intersection scenarios with multiple other cars, but the step-wise evaluation makes it harder to constrain the convergence of the solution to specific simple trajectories. ROPT therefore optimizes parametrized velocity profiles consisting of two consecutive acceleration and deceleration ramps (see left-hand side of Figure <ref>). A double-ramp profile is described by the end velocities of the first and second ramp v_r,1 and v_r,2 plus the start time of the second ramp s_r,2. Each ramp has a fixed duration s_d=[2.5]s, whereas the first ramp starts with the current velocity v_0. We use the Nelder-Mead opimization algorithm <cit.>, a downhill simplex method that does not require gradient information. In each optimization iteration, the simple double-ramp velocity profile is converted into a trajectory, which in turn is evaluated for risk, utility and comfort (see Section <ref>). Trajectories that violate constraints on velocity or accleration are penalized. Since the Nelder-Mead optimization is a local search, it depends on the initial value of the parameters. Hence, ROPT simultaneously optimizes k trajectories, starting from different initial values v_r,1^j = v_r,2^j = v_r^j with j∈1,…,k and v_r^j evenly spaced in [[2]m/s,v_max]. If trajectory j was chosen in the previous time step, the optimization continues with its previous parameters. Additionally, it shifts the beginnings and ends of all ramps by an offset o that corresponds to the total time that the trajectory has been selected.[If the trajectory has been active for a duration equal to the ramp length (o=s_r), we insert a new ramp after ramp two if s_2<s_r or before ramp two otherwise.] In other words, if a planned trajectory is executed for multiple time steps, it is time-shifted and fine tuned in each step. In addition to the optimized trajectries, ROPT samples three simple trajectories (see right-hand side of Figure <ref>): A constant velocity trajectory v_c = v_0, a stopping trajectory with the end point (s_b, v_b = 0) and an acceleration trajectory to the set speed v_a at the predicted time s_a. §.§ Trajectory Evaluation §.§.§ Risk Prediction An indicator for risk is the probability function P_crit(s;t,Δ t) that a critical event will happen from the current time t around a future time t+s during an interval of size Δ t. A compact risk measure R(t) then comprises the entire accumulated future risk contained in P_crit(s;t,Δ t), s∈[0,∞]. By also weighting P_crit(s;t,Δ t) with the predicted damage of the event D_crit(s;t,Δ t), we obtain risk as the expected future severity. ROPT combines the Gaussian method <cit.> for the estimation of critical event probabilities with the survival analysis <cit.> to gain the overall risk. In the following, we consider the situation evolution of an ego car (green) encountering another car (red) indexed with i=1,2 as depicted in Figure <ref>. The car's future trajectories are predicted as in Section <ref>. It cannot be assumed that the cars follow exactly the predicted trajectory. In reality, they undergo variations in speed and heading. Consequently, we model their respective spatial position with a normal distribution f_i. Since we predict the vehicles to drive along predefined paths, we furthermore define that the longitudinal uncertainty along the path is higher than the lateral uncertainty. In this way, we obtain 2D ellipses that are specified by an uncertainty matrix Σ^'_i around the mean position vector μ_i with μ_i = [ μ_i,x; μ_i,y; ], Σ^'_i = [ σ^2_i,lon 0; 0 σ^2_i,lat; ] . A collision occurs if both cars coincide at the same position, which is analog to f_coll=f_1f_2. To retrieve the product of two Gaussian functions, the uncertainties Σ^'_i have to be transformed into the same global coordinate system x,y according to Σ_i = 𝐑Σ^'_i𝐑^T and𝐑 = [ cosα_i -sinα_i; sinα_i cosα_i; ]. The collision probability is eventually given by spatially integrating f_coll over all positions P_coll(s;t,Δ t) = |2π (Σ_1+Σ_2)|^-1/2 * exp{-1/2 (μ_2-μ_1)^T(Σ_1 +Σ_2)^-1 (μ_2-μ_1)}. Due to temporal uncertainty, the deviation of the real from the predicted trajectory differs with increasing prediction times. We extrapolate the kinematics of the current state to achieve trajectories. After a prediction step of size Δ s, their longitudinal position on the path l_i is shifted by Δ l_i and we get l_i(s+1) = l_i(s) + Δ l_i = l_i(s) + v_i(s) Δ s with velocities v_i. Knowing this, the growth of spatial uncertainty can be derived from a velocity uncertainty factor c_i using σ_i,v(s+1) := σ_i,v(s) + c_i v_i(s) Δ s. The probability for the ego car to drive off at sharp curves is formulated similarly. Here, we assume 1D circles with uncertainty σ_1 and look at the lateral acceleration a_y(s) = √(0pt0.3cmκ v_1), which is influenced by the curvature of the road κ. We then compare a_y with its maximal possible value a_y,max from vehicle dynamics constraints to retrieve P_curv(s;t,Δ t) = 1/√(2πσ^2_1) exp{-max(a_y,max-|a_y|,0)^2/2σ_1^2}. If |a_y| approaches a_y,max, the probability P_curv(s;t,Δ t) will thus increase. Next in the survival analysis, accident occurrences are modeled as a thresholding process based on Poisson-like event probabilities. A Poisson process is defined by a situation state-dependent total event rate τ^-1(z_t:t+s), which characterizes the mean time between events and consists of a critical event rate τ^-1_crit and a constant escape rate τ^-1_0 (comprising behavioral options that mitigate critical events). For τ^-1_crit, we consider collision risks represented by the single event rate τ_coll^-1 and the risk of losing control in curves τ^-1_curv which leads to τ^-1(z_t:t+s)=τ^-1_0 + τ^-1_crit=τ^-1_0+τ^-1_coll+τ^-1_curv with τ_crit^-1(z_t:t+s) = P_crit(s;t,Δ t) / Δ t, P_crit(s;t,Δ t) = P_coll(s;t,Δ t) + P_curv(s;t,Δ t). The survival function indicates the probability that the vehicle will not be engaged in an event like an accident from t until t+s in compliance with S(s;t,z_t:t+s)=exp{-∫_0^s τ^-1(z_t:t+s') ds'}. It has been empirically shown that human injury compared to the kinetic energy of the accident (proportional to velocity vectors 𝐯_i) has the behavior of a logistic function <cit.>. We postulate that the damage of the involved cars has the same qualitative relationship. Car-to-car collision and curve control loss damage is given by D_coll(s;t,Δ t) = D_/1+{ k_ (𝐯_2 - 𝐯_1 - β_) }, D_curv(s;t,Δ t) = D_/1+{ k_(𝐯_1 - β_) }, where the parameter k is the damage increase factor, β the damage midpoint and D_ the maximal damage. As a result, we acquire the overall risk engaging in a future critical event by temporally integrating the term of probabilities, damages and survival function R(t) = ∫_0^∞ (τ_coll^-1D_coll+τ_curv^-1D_curv)S ds. §.§.§ Utility and Comfort Prediction A driver tries to minimize the risk, but maximize his benefit as well. In ROPT, the considered benefit consists of the needed time to arrive at the goal and the comfort of the travel. The former is defined by the velocity course of the ego car v_1 and the latter takes the acceleration and jerk profile a_1 and j_1 into account.[The double ramp parameters have limited influence in j_1. Here, polynoms as velocity profiles can be beneficial due to their property of continuity.] We weight the components with driver-specific constants b^t, b^c and b^j and compute the integral future benefit with B(t)=∫_0^∞ (b^t |v_1|-b^c |a_1|- b^j |j_1|) S ds. For higher s, we also consider the survival function S in the evaluation. In this way, high-risk situations result into lower B(t). ROPT finally evaluates for each generated trajectory the cost function C(t) = R(t) - B(t) and executes the trajectory with the lowest C(t). For comparability, R(t) and B(t) have to be transformed into the same unit. Since the severity factors in R(t) require a monetarization, we use for C(t). § EXPERIMENTS §.§ Setup In our simulation, IIDM and ROPT face the complex task of planning succesful merge-ins at T-intersections. Figure <ref> illustrates the ego car waiting at the stop line, while other cars pass from left to right. We model the traffic as a Poisson distribution P(λ) with different intervals λ and count the number of missed gaps n_gap.[For continuous λ, we add a uniform random noise offset in the range of [[-0.5]s,[0.5]s].] If the gap size t_gap is large enough, the ego car needs to increase v until the maximum curve velocity and subsequently accelerate even further to match the velocity of the traffic flow. The angle of the right turn with [90]^∘ and the constant traffic speed with v_f=[10]m/s are chosen in a way that makes single-ramp trajectories only feasible for very large gaps of more than [100]m. Besides systematically changing λ which results in different traffic densities, we modify the politeness factor p of IIDM and the travel benefit b^t for ROPT to control the driving behavior. For every model and traffic setting, we then run several simulations (ca. 200) to ensure statistical significance. Once the ego car advances past the stop line, the merge-in procedure has started and the encountered distances to the front d_front and rear vehicle d_back are recorded. In the evaluation, we look at the mean minimal distances d̅_back,min and d̅_front,min, which indicate risk, as well as the mean utility indicators n̅_gap and t̅_gap.[Ideal environment conditions are assumed, in which position and velocity of the involved cars are synchronized and known without sensor errors.] §.§ Results Figure <ref> plots the distances d̅_back,min and d̅_front,min for IIDM with p = [0.5,4] and ROPT with b^t=[[0.1]/km,[10]/km] for different λ=[[2]s,[5]s]. As expected, d̅_back,min decreases with lower λ for both models. Likewise, it declines with lower p or higher b^t. The ego car enters smaller gaps, where the rear car approaches more closely due to its excess speed. More importantly, the lower bound for d̅_back,min over all runs (purple line) is always 0 in the case of the IIDM. Merge-ins cannot be performed properly with accidents happening in [37]% of the runs.[A crash or near-crash case is assumed, if the distance between two cars is less than [1]m.] The reason can be identified in the simulation: At time step t_start, the IIDM decides to merge in based on the incentive criterion of Equation (<ref>). However due to the restricted curve speed, it cannot accelerate quickly enough in front of the follower. At some later time step t_end, the criterion becomes invalid and a brake is executed. If the resulting stopping position is in the way of the crossing traffic, an accident occurs. In contrast, the lower bound of ROPT stays above [15]m. ROPT does not expect any deceleration, but assumes constant velocity for the follower and yields consistently safe behavior. Regarding d̅_front,min, only runs with safe mergers are listed for IIDM (otherwise there is no notion of a back car). While d̅_front,min is approximately the same for IIDM and ROPT with different λ, it also descends with smaller p or bigger b^t. In ROPT, bigger b^t lead to less weight on risk and earlier merge-in begins. With IIDM, d_front is usually constant. Equation (<ref>) does not consider the front car. Nevertheless, due to unsuccessful merge-in attempts at previous gaps, it is possible that the ego car has already advanced into the intersection. Figure <ref> pictures n̅_gap and t̅_gap with the same parameter variations. Both models choose earlier gaps (less n̅_gap) based on willingness of driving into lower t̅_gap with decreasing p or increasing b^t. In general, IIDM has sharper slopes, because of its risk-proneness. For dense traffic settings (small λ), the models need to wait longer with higher n̅_gap and the taken gap has lower t̅_gap. In total, t̅_gap lies in between [4 - 7]s which are realistic critical gaps <cit.>. §.§.§ Predictive IIDM We saw that IIDM is unable to cope with merge-in scenarios. It implicitly predicts the behavior of the driver and leading vehicle, but lacks a proper prediction of the follower. For this reason, we extend the IIDM further. In the spirit of ROPT, we explicitly extrapolate trajectories of other cars with constant velocity and the ego trajectory using Equations (<ref>) and (<ref>). At every predicted step, the criterions (<ref>) and (<ref>) are evaluated. Only if both hold true for the complete horizon, a merge-in is started. The outputs of the predictive IIDM are shown in Figure <ref>. The lower bound for d̅_back,min approaches [14]m for p→0.5 , because b_safe is set to [-2]m/s^2 in the safety criterion. The courses of d̅_front,min are constant [7]m for different λ. Independent of p and λ, a safe behavior is achieved. Additionally, n̅_gap and t̅_gap are shifted to higher values and similar to ROPT. §.§.§ Discussion ROPT distributes the available t_gap more to d_front than d_back, while keeping the absolute risk low (i.e . no crash happens) for varying b^t and λ. As opposed to this, IIDM is not able to cap the risk with its heuristics. The predictive IIDM generates safe behavior, but different p can only change its influence to d_back. This is undesirable, because in merge-ins the behavior of front cars is better manageable than back cars. Moreover, neither versions of IIDM can derive a continous acceleration course. They only switch between braking or following the front car. ROPT inherently possesses an explicit risk-utility tradeoff (i.e., making better use of t_gap). For intersections of arbitrary geometry and traffic constellations, ROPT works out of the box due to its instantaneous adaptation using a full holistic predictive risk model. For instance, a crossing of X-intersections could be also handled in the predictive IIDM with additional go/no-go decisions. However, lateral risks are only considered shortly before the intersection and not afterwards. In complex interactions between the ego and multiple other cars, the combination of heuristics will potentially lead to unsafe behavior. § CONCLUSION AND OUTLOOK In this work, we outlined a longitudinal velocity planner along map geometries. ROPT alternates between trajectory generation and evaluation in an optimization cycle. Multiple double-ramp velocity profiles as well as fixed constant velocity, acceleration and deceleration trajectories are constructed. For the arising dynamic traffic scenario predictions, the integral risk, utility and comfort are calculated. ROPT infers curve and collision risks based on Gaussian probability estimation and a future risk accumulation with the survival analysis. The cycle is repeated until the cost threshold is satisfied for each sample in the set and the velocity course with the lowest cost is chosen for execution. As a comparison, we extended the IDM for curve driving and intersection scenarios to obtain the IIDM. It uses a modified free term for maximal curvature adherence and lane change criterions altering the interaction term. In simulations of right turning at T-intersections with several cross traffic densities, ROPT considers the risks properly in all parameter settings and distributes it to the two risk sources evenly. IIDM has qualitatively similar results but can yield crash cases in suboptimal configurations, because of if-then type behavior and improper prediction. We thus created the predictive IIDM, which applies the heuristics for all time steps and has no accidents but weak risk-utility tradeoff. Overall, ROPT is generalizable to more complex situations due to its holistic risk prediction. In <cit.>, we modeled occlusion risks from buildings around intersections with virtual cars. In future work, traffic elements at intersections (e.g. traffic lights or right-before-left) could be incorporated in a similar way. Otherwise, the ego car will not comply with special rule-based behaviors. The survival method has been shown to detect collision risks early with few false positives <cit.>. Nevertheless for high prediction horizons, the 2D Gaussians might cover opposite lanes when turning and project errors. It remains to be investigated, if the road curvatures are usable to create Gaussian Mixture Models (GMM) with bent shapes. Vehicle dynamics are currently only modeled with a double integrator system in ROPT. For online application, physical steering models with delay of gas pedal to acceleration for the ego car and sensor models with noisy and late detected positions of other cars may be included as a pre-processing step. At last, the weighting of risk against benefit depends on the driver type. We expect that self-adjustment of the corresponding parameters would yield a personalization of planned velocites. § ACKNOWLEDGMENT This work has been partially supported by the European Unions Horizon 2020 project VI-DAS, under the grant agreement number 690772. The authors would like to thank Fabian Müller for his support. IEEEtran

http://arxiv.org/abs/2303.06673v1

20230312142656

SSGD: A smartphone screen glass dataset for defect detection

cs.CV

Universal Instance Perception as Object Discovery and Retrieval Bin Yan^1This work was performed while Bin Yan worked as an intern at ByteDance. Email: mailto:yan_bin@mail.dlut.edu.cnyan_bin@mail.dlut.edu.cn. ^† Corresponding authors: mailto:jiangyi.enjoy@bytedance.comjiangyi.enjoy@bytedance.com, mailto:wdice@dlut.edu.cnwdice@dlut.edu.cn., Yi Jiang^2, †, Jiannan Wu^3, Dong Wang^1, †, Ping Luo^3, Zehuan Yuan^2, Huchuan Lu^1,4 ^1 School of Information and Communication Engineering, Dalian University of Technology, China ^2 ByteDance ^3 The University of Hong Kong ^4 Peng Cheng Laboratory March 30, 2023 =========================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================== Interactive devices with touch screen have become commonly used in various aspects of daily life, which raises the demand for high production quality of touch screen glass. While it is desirable to develop effective defect detection technologies to optimize the automatic touch screen production lines, the development of these technologies suffers from the lack of publicly available datasets. To address this issue, we in this paper propose a dedicated touch screen glass defect dataset which includes seven types of defects and consists of 2504 images captured in various scenarios. All data are captured with professional acquisition equipment on the fixed workstation. Additionally, we benchmark the CNN- and Transformer-based object detection frameworks on the proposed dataset to demonstrate the challenges of defect detection on high-resolution images. Dataset and related code will be available at <https://github.com/VincentHancoder/SSGD>. touch-screen-glass, dataset, defect detection § INTRODUCTION Nowadays, smart terminals are increasingly crucial to the intelligence process of contemporary society. Almost everyone has at least one smartphone in their hands. As the essential accessories, the production quality of the smartphone screen directly determines the display effect, service life, and user evaluation of smartphones. To this end, almost all screens must pass a production quality inspection before they leave the factory. However, defect detection via human beings only is labor-intensive and inefficient, which is insufficient for the vast market of smartphones. In this case, the use of data-driven computer vision inspection methods can significantly improve efficiency and reduce judgment errors brought on by human factors. There have been several attempts to apply deep learning techniques to the task of detecting defects in industry. In the metal generic surface defect detection area, Bao et al. <cit.> presented a dedicated dataset called NEU-Dataset, and Song et al. <cit.> raised a matching method with a set of data used for defect detection of silicon steel strip micro surface. Similarly, a railway surface defect dataset <cit.> was made for detection in 2017. Besides metal surface detection tasks, increasing number of industrial products dataset <cit.> have been widely collected for specific detection scenarios. Some works focused on fabric regions <cit.> have been produced and extended to a challenging competition to encourage contestants to approach higher detection accuracy. Recently more in-depth, increasing dimensions of the electronics industry production have been using neural network technologies to detect the quality of products and defect distribution. Pramerdorfer et al. <cit.> announced a dataset of Printed circuit board (PCB) which was made to facilitate the computer vision tasks in the challenge of PCB deficiency in producing process. In 2019, Deitsch et al. <cit.> presented a dataset presented a dataset about the solar panels damage situation. However, to the best of our knowledge, there is no publicly available dataset for defect detection of smartphone screens. This seriously hinders the application of computer vision technology in the defects inspection of screens. To solve this problem, we firstly propose an open-source Smartphone Screen Glass Dataset, dubbed as SSGD, which contains basically common types of defect occurring on the glass panels. Specifically, the proposed SSGD is made up of 2504 images and contains seven types of defects commonly existing in the production process. All images are in an uniformed resolution of 1500× 1000 pixels. Figure <ref> shows some samples of SSGD. After the procedure of data collecting and annotating, we conduct extensive experiments based on the general platform to evaluate the performance of popular object detection on SSGD (Sec. <ref>). Our main contributions can be summarized as follows: * We collect and annotate a dataset for defect detection of smartphone screens, which possesses various annotations categories, relatively high image information quality, and public availability. * We benchmark many popular object detectors on the proposed dataset, including CNN- and Transformer-based frameworks. § DATASET CREATION PROCEDURE In this section, we will introduce SSGD from three aspects: (1) the process of image capturing, (2) the overall properties of the dataset, and (3) the composition of the dataset. §.§ Data Collection Before information acquisition, numerous samples with defects have been collected and selected as origin materials for the purpose of capturing. The background scenario is decorated with the color black to weaken the influence of other visible light in the environment on acquired image information. During the image collection process, smartphone touch screens are placed on a specific capture platform which has been calibrated by a leveler to ensure the correctness of the shooting angle. On the platform, the surrounding region of the screen is the ink area, and the middle is the visible area protected by specific film. The device used to collect images is a line-scan camera which is designed for industry-level information acquisition. After the unified collection and initial data cleaning process, we uniformly change the image resolution into 1500× 1000 pixels in order to ensure every training object provides the same size background information. The entire dataset is grouped by different capturing platforms and given specific file numbers according to groups. After collection, we employ the labelme[<https://github.com/wkentaro/labelme>] to annotate the seven kinds of defects and the corresponding locations (bounding box format) on the screens and get an XML file for each image. §.§ Dataset Properties SSGD consists of 2504 images and includes seven common types of defects in actual smartphone screen process. Aiming to show the properties of SSGD clearly, we illustrate the causality in following perspectives: Categories: Seven types of defects including crack, broken, spot, scratch, light-leakage, blot, broken-membrane. Some samples are provided in Fig. <ref>. Workstation: Two workstations are used to capture images. SSGD is therefore divided into two parts, called Part I and Part II, respectively. Workstation Content: Two workstations captured 1258 and 1246 images (corresponding to Part I and Part II). Part I Content: the possession situation of each type of defects is that crack: 988, broken: 304, spot: 175, scratch: 99, light-leakage: 63, blot: 18, broken-membrane: 10. Part II Content: the possession situation of each type of defects is that crack: 787, broken: 756, spot: 467, scratch: 163, light-leakage: 60, blot: 13, broken-membrane: 11. §.§ Dataset Distribution As shown in Fig. <ref> (a), there are two data bounding box characteristic distribution maps whose horizontal axis and vertical axis represent for the height and the width of each bounding box which are distinguished by different workstations. The position on the map of points in different colors show how their bounding boxes look like. There are two curves in blue and red symbolize thresholds to define small, middle and large target detection object. If the point is lower than red curve, it is a small target covering an area less than 32× 32 pixels. Similarly, it is thought as a large target when more distant from origin point than blue curve. And the target which is between blue and red curve will be thought as a middle size detection object. For the purpose of showing quantity distribution of large, medium and small detection targets, we summarize the following related information that can be corroborated by the relevant information in the Fig. <ref> (a) as well: Part I: the amount of different types detection object is small: 241, middle: 378, large: 1038. Part II: the amount of different types detection object is small: 783, middle: 441, large: 1033. However, there are some extreme points existing on the map such as the points gathering at the upper left area of the map, which means that bounding boxes represented by those points are at a nearly 7:1 width and length ratio. As the input of network, bounding box in a such extreme shape may influence final convergence direction. The color intensity shows how gathering a type of defect points are. It is shown that on the both maps (Part I and Part II) category crack basically distributed above the blue line while category scratch mainly gather at the zone in middle. In other words, we get a pattern that most samples of above two categories are large size and middle size detection object. Obviously, there are still some categories, such as blot, that are unable to find the pattern of bounding box size. Such a phenomenon is acceptable as well. § EXPERIMENT §.§ Settings We benchmark most basic object detection models on the proposed SSGD, including Faster R-CNN <cit.>, FCOS <cit.>, and YOLO series <cit.>. To reduce random bias, we use 5-fold cross-validation to measure all models, and results are averaged over the five folds. In the implementation, most experiments are based on MMDetection <cit.>. YOLOv5 and YOLOX <cit.> follow their official repositories[<https://github.com/ultralytics/yolov5>][<https://github.com/Megvii-BaseDetection/YOLOX>]. Before training, we initialize the model with the weight pre-trained on COCO <cit.> dataset and new layers in the classification head with the Normal scheme. During training, for ResNet-50-based <cit.> models, we utilize the SGD <cit.> optimizer and the 2× (24 epochs) schedules with a global batch size of 16 on 4 GPUs. We also adopt the multi-scale training, where the short side of input images is randomly resized to [800, 1500] pixels, and the long side is at most 2250 pixels. For Transformer-based models <cit.>, we utilize the AdamW <cit.> optimizer and the 2× schedules with a global batch size of 8 on 8 GPUs. The resolution of input images is 1500× 1000 pixels. Other settings remain the same as MMDetection. For YOLO series, the input with a resolution of 1500× 1000 pixels is padded to 1500× 1500 pixels. We train the model 100 epochs with a global batch size of 16 on 4 GPUs. Other settings remain the same as the original repositories. During testing, input images maintain their original resolution without any augmentation. §.§ Results Results on CNN-based methods. We evaluate mainstream anchor-based <cit.> and anchor-free <cit.> object detectors on SSGD. Besides YOLOv5 and YOLOX, all other models take ResNet-50 as the backbone. Due to an adaptive training sample selection strategy, ATSS obtains the best performance in all ResNet-50-based models. ATSS outperforms the two-stage Cascade R-CNN by 1.4 AP and 0.3 AP on SSGD Part I and Part II, respectively, while only possessing half the parameters. However, Cascade R-CNN achieves better performance on large objects. Results on Transformer-based methods. We evaluate Swin <cit.>, PVT <cit.>, ScalableViT <cit.>, and UniFormer <cit.> on the proposed SSGD using the Faster R-CNN framework. As reported in Table <ref>, ScalableViT-S achieves 21.1 AP on SSGD Part I under a single-scale training strategy, which surpasses most ResNet-50-based methods using the multi-scale training strategy. However, on SSGD Part II, Swin-T obtains better performance (27.0 AP) than other Vision Transformer counterparts. When compared to ResNet-50, Swin-T gets 3.2 AP gains. Nevertheless, Transformer-based models have an obvious disadvantage in the speed that is required in industrial scenarios. Specifically, when input resolution is 1500 × 1000 pixels, Swin-T-based Faster R-CNN can only process 18 images per second, but ResNet-50-based ones can process 26 images. Moreover, PVT and ScalableViT are slower than Swin because the method that shrinks spatial tokens of Keys and Values may no longer be applicable in high-resolution images. Therefore, a Vision Transformer, friendly to high-resolution images and industrial scenes, needs to be developed with both higher accuracy and lower latency. § DISCUSSION AND CONCLUSION In this paper, we present the first publicly available Smartphone Screen Glass Dataset for defect detection. We adapted professional capturing device and non-single workstations to collect images which involves various categories of defects commonly existing in actual producing procedure. Then, we elaborately analysed the object distribution of this dataset Next, abundant experiments are conducted to show the performance of popular methods on proposed dataset. Based on the comparison of experimental results, we find that the Vision Transformers perform worse and are much slower than their CNN counterparts at high resolution input. At the same time, a dynamic assignment strategy during training is very important to this dataset. In the future, we will continue the investigation to develop more promising and approachable methods to improve the detection effect in the testing process. We hope this paper paves the way for the application of computer vision in the defect detection of screens. § ACKNOWLEDGEMENT This work was supported by the National Key R&D Program of China 505 (Grant No.2020AAA0108303), the Shenzhen Science and Technology Project (Grant No.JCYJ20200109143 041798) and Shenzhen Stable Supporting Program (WDZC202 00820200655001). Partial samples and analytical methods are provided by Shenzhen Zhihan Equipment Ltd., Li Xinghui and Wang Xiaohao. IEEEbib

http://arxiv.org/abs/2303.07110v1

20230313134404

Upcycling Models under Domain and Category Shift

cs.CV

Upcycling Models under Domain and Category Shift Sanqing Qu^1Equal Contribution , Tianpei Zou^1, Florian Röhrbein^2, Cewu Lu^3, Guang Chen^1Corresponding author: guangchen@tongji.edu.cn , Dacheng Tao^4,5, Changjun Jiang^1 ^1Tongji University, ^2Chemnitz University of Technology, ^3Shanghai Jiao Tong University, ^4JD Explore Academy, ^5The University of Sydney March 30, 2023 ================================================================================================================================================================================================================================================================================================================================= Deep neural networks (DNNs) often perform poorly in the presence of domain shift and category shift. How to upcycle DNNs and adapt them to the target task remains an important open problem. Unsupervised Domain Adaptation (UDA), especially recently proposed Source-free Domain Adaptation (SFDA), has become a promising technology to address this issue. Nevertheless, existing SFDA methods require that the source domain and target domain share the same label space, consequently being only applicable to the vanilla closed-set setting. In this paper, we take one step further and explore the Source-free Universal Domain Adaptation (SF-UniDA). The goal is to identify “known" data samples under both domain and category shift, and reject those “unknown" data samples (not present in source classes), with only the knowledge from standard pre-trained source model. To this end, we introduce an innovative global and local clustering learning technique (GLC). Specifically, we design a novel, adaptive one-vs-all global clustering algorithm to achieve the distinction across different target classes and introduce a local k-NN clustering strategy to alleviate negative transfer. We examine the superiority of our GLC on multiple benchmarks with different category shift scenarios, including partial-set, open-set, and open-partial-set DA. Remarkably, in the most challenging open-partial-set DA scenario, GLC outperforms UMAD by 14.8% on the VisDA benchmark. The code is available at <https://github.com/ispc-lab/GLC>. § INTRODUCTION At the expensive cost of given large-scale labeled data and huge computation resources, deep neural networks (DNNs) have made remarkable progress in various tasks. However, DNNs often generalize poorly to the unseen new domain under domain shift and category shift. How to upcycle DNNs and adapt them to target tasks is still a long-standing open problem. In the last decade, many efforts have been devoted to unsupervised domain adaptation (UDA) <cit.>, which capitalizes on labeled source data and unlabeled target data in a transduction manner, and has achieved significant success. Despite this, the access to source raw data is inefficient and may violate the increasingly stringent data privacy policies <cit.>. Recently, Source-free Domain Adaptation (SFDA) <cit.> has become a promising technology to alleviate this issue, where only a pre-trained source model is provided as supervision rather than raw data. However, to avoid model collapse, most existing methods <cit.> assume that the label space is identical across the source and target domain, thus being only applicable to vanilla closed-set scenarios. In reality, target data may come from a variety of scenarios. Therefore, it is too difficult to hold such a strict assumption. For a better illustration, we suppose 𝒴_s and 𝒴_t as the label space of source domain and target domain, respectively. In addition to the well-studied vanilla closed-set (𝒴_s = 𝒴_t.), we often encounter several other situations, e.g., the partial-set (𝒴_s ⊃𝒴_t), the open-set (𝒴_s ⊂𝒴_t), and the open-partial-set (𝒴_s ∩𝒴_t ∅, 𝒴_s ⊈𝒴_t, 𝒴_s ⊉𝒴_t). Currently, there have been several source data-dependent works <cit.> developed to target category shift. However, methods devised for one situation are commonly infeasible for others. In practice, the target domain is unlabeled and we cannot know which of these category shifts will occur in advance. Not to mention that the requirement to source raw data makes it inefficient and potentially violates data protection policies. To tackle these limitations, and handle those category shifts in a unified manner, in this paper, we take one step further and delve into the Source-free Universal Domain Adaptation (SF-UniDA). The goal is to upcycle the standard pre-trained source models identifying “known" data samples and rejecting those “unknown" data samples (not present in source classes) under domain and category shift. We conceptually present the SF-UniDA in Fig. <ref>. Note that, very few works <cit.> have studied the source-free model adaptation in open-partial-set scenarios. Nevertheless, their approaches demand dedicated model architectures, greatly limiting their practical applications. SF-UniDA is appealing in view that model adaptation can be resolved only on the basis of a standard pre-trained closed-set model, i.e., without specified model architectures. To approach such a challenging DA setting, we propose a simple yet generic technique, Global and Local Clustering (GLC). Different from existing pseudo-labeling strategies that focus on closed-set scenarios, we develop a novel one-vs-all global clustering based pseudo-labeling algorithm to achieve “known" data identification and “unknown" data rejection. As we have no prior about the category shift, we utilize the Silhouettes <cit.> metric to help us realize adaptive global clustering. To avoid source private categories misleading, we design a global confidence statistics based suppression strategy. Although the global clustering algorithm encourages the separation of “known" and “unknown" data samples, we find that some semantically incorrect pseudo-label assignments may still occur, leading to negative knowledge transfer. To mitigate this, we further introduce a local k-NN clustering strategy by exploiting the intrinsic consensus structure of the target domain. We validate the superiority of our GLC via extensive experiments on four benchmarks (Office-31 <cit.>, Office-Home <cit.>, VisDA <cit.>, and Domain-Net <cit.>) under various category shift situations, including partial-set, open-set and open-partial-set. Empirical results show that GLC yields state-of-the-art performance across multiple benchmarks, even with stricter constraints. Our contributions can be summarized as follows: * To the best of our knowledge, we are the first to exploit and achieve the Source-free Universal Domain Adaptation (SF-UniDA) with only a standard pre-trained closed-set model. * We propose a generic global and local clustering technique (GLC) to address the SF-UniDA. GLC equips with an innovative global one-vs-all clustering algorithm to realize “known" and “unknown" data samples separation under various category-shift. * Extensive experiments on four benchmarks under various category-shift situations demonstrate the superiority of our GLC technique. Remarkably, in the open-partial-set DA situation, GLC attains an H-score of 73.1% on the VisDA benchmark, which is 14.8% and 16.7% higher than UMAD and GATE, respectively. § RELATED WORK Unsupervised Domain Adaptation: To alleviate performance degeneration caused by domain shift, unsupervised domain adaptation (UDA) has received considerable interest in recent years. Existing methods can be broadly classified into three categories: discrepancy based, reconstruction based, and adversarial based. Discrepancy based methods <cit.> usually introduce a divergence criterion to measure the distance between the source and target data distributions, and then achieve model adaptation by minimizing the corresponding criterion. Reconstruction based methods <cit.> typically introduce an auxiliary image reconstruction task that guides the network to extract domain-invariant features for model adaptation. Inspired by GAN <cit.>, adversarial based approaches <cit.> leverage domain discriminators to learn domain-invariant features. Despite of effectiveness, these methods typically focus only on the vanilla closed-set domain adaptation. Universal Domain Adaptation: To handle category-shift, there have been some methods proposed for partial-set <cit.>, open-set <cit.>, and open-partial-set domain adaptation <cit.>. However, most of these methods are designed for a specified situation, and are typically not applicable to other category-shift situations. As an example, an open-partial-set method <cit.> even underperforms the source model in the partial-set scenario. Recently, <cit.> propose a truly universal UDA method, which is applicable to all three category-shift situations. Nevertheless, most existing methods need access to source data during adaptation, which is inefficient and may violate the increasing data protection policies <cit.>. Source-free Domain Adaptation: Recently, several works <cit.> have attempted to achieve domain adaptation with knowledge from only the pre-trained source model rather than raw data. However, to avoid model collapse, these methods commonly focus on the vanilla closed-set domain adaptation, significantly limiting their usability. Very recently, few works <cit.> have studied the source-free domain adaptation in open-partial-set scenarios. Nevertheless, the requirement of dedicated source model architectures, e.g., specialized two-branch open-set recognition frameworks, greatly limits their deployment in reality. In this paper, we target for achieving truly universal model adaptation, including partial-set, open-set, and open-partial-set scenarios, with the knowledge from vanilla source closed-set model. § METHODOLOGY §.§ Preliminary In this paper, we aim to achieve model upcycling under both domain shift and category shift, i.e., the source-free universal domain adaptation (SF-UniDA). In particular, we consider the K-way classification. In this setting, there is a well-designed source domain 𝒟_s = {(x^i_s, y^i_s)}^N_s_i=1 where x^i_s ∈𝒳_s, y_s^i ∈𝒴_s, and an unlabeled target domain 𝒟_t = {(x_t^i, ?)}^N_t_i=1 where x_t^i ∈𝒳_t. For a better illustration, we denote 𝒴 = 𝒴_s ∩𝒴_t as the common label space, 𝒴̅_s = 𝒴∖𝒴_t as the source private label space, and 𝒴̅_t = 𝒴∖𝒴_s as the target private label space, respectively. As aforementioned, there are three possible category shifts, i.e., the partial-set DA, PDA, (𝒴_s ⊃𝒴_t); the open-set DA, OSDA, (𝒴_s ⊂𝒴_t); and the open-partial-set DA, OPDA, (𝒴∅, 𝒴̅_s ∅, 𝒴̅_t ∅). The final goal is to identify “known" samples (belonging to 𝒴) and reject “unknown" samples (belonging to 𝒴̅_t) of 𝒟_t, with the knowledge only from source pre-trained model f_s. 𝒟_s is not available, and we do not have prior knowledge of what kind of category shift we are facing. There have been few works <cit.> explored source-free domain adaptation under category shift. However, these methods are limited to specific category shift, and require dedicated source model architectures. To address these limitations, we propose to achieve SF-UniDA on the basis of only the vanilla closed-set model. Following existing closed-set source-free domain adaptation methods <cit.>, given a source model f_s = h_s ∘ g_s, consisting of a feature module g_s and a classifier module h_s, we capitalize on the source hypothesis to achieve source and target domain alignment. That is, we only learn a target-specific feature module g_t and keep the classifier h_t = h_s. To realize “known" data identification and “unknown" data rejection under both domain shift and category shift, we devise a novel, adaptive, global one-vs-all clustering algorithm. Besides, we further employ a local k-NN clustering strategy to alleviate negative transfer. The pipeline is presented in Fig. <ref>. More details will be described in the following. §.§ One-vs-all Global Clustering Pseudo-labeling is a promising technique in unsupervised learning. Traditional pseudo-labeling strategies <cit.> assign pseudo labels directly based on sample-level predictions, which are often noisy, especially in the presence of domain shifts. To mitigate this, there are some pseudo-labeling strategies <cit.> exploit the data structure of the target domain, i.e., the target-specific prototypes. However, these strategies assume that the source and target domain share identical label space, making it infeasible under category shift. Therefore, a question naturally arises: How to achieve pseudo-labeling with inconsistent label space? Especially, for universal domain adaptation, we have no prior about the category shift between 𝒴_s and 𝒴_t. To tackle this, we first view this problem from a simplified perspective: If 𝒴_s ⊂𝒴_t (i.e., the OSDA setting), and we were to know the number of categories in the target domain is C_t, what kind of pseudo-labeling strategy should we apply? Intuitively, target domain, in this case, should be grouped into C_t clusters, each corresponding to a specific category. We can then assign pseudo labels via the nearest cluster centroid classifier. However, even though we apply existing clustering algorithms, such as K-means <cit.>, to divide the target domain into C_t clusters. It is still challenging to associate the corresponding semantic category for each cluster, in particular for the SF-UniDA, as we have no access to the source raw data. In view of this, to ease the challenging semantic association, we devise a novel one-vs-all global clustering pseudo-labeling algorithm. The main idea is that For a particular “known" category c ∈ C_s, in order to decide whether a data sample belongs to the c-th category, we need to figure out what is and what is not the c-th category. The detailed procedure is presented as follows: * For a particular c-th category, we first aggregate the top-K δ_c(f_t(x_t)) scores represented instances along all target domain 𝒟_t as positive 𝒫_c , and the rest as negative 𝒩_c. Here, δ_c(f_t(x_t)) denotes the soft-max probability of target instance x_t belonging to the c-th class. We empirically set K = N_t / C_t. * Then, we obtain the positive prototype representation p_c (i.e., what is the c-th category), and negative prototypes {n_c^i}^M_i=1 (i.e., what are not the c-th category) via K-means. Noting that we have employed multiple prototypes to represent the negatives since the negatives contain distinct classes. We set M to C_t instead of C_t - 1, considering that the “known" category of the target domain typically involves some hard samples that are difficult to be selected by top-K sampling. p_c = 1/K∑_x_t∈𝒫_c g_t(x_t), {n_c^i}_i=1^M = Kmeans_x_t ∈𝒩_c (g_t(x_t)). * Thereafter, we decide whether data sample x_t belongs to the c-th category via the nearest centroid classifier: q̂_c = { 1,if S(g_t(x_t), p_c) ≥max{S(g_t(x_t), n_c^i)}_i=1^M 0,if S(g_t(x_t), p_c) < max{S(g_t(x_t), n_c^i)}_i=1^M. where S(a, b) measures the similarity between a and b. We apply the cosine similarity function by default. * Finally, we iterate the above process to obtain the pseudo labels ŷ_t for all “known" category c ∈ C_s. Since each data sample either belongs to the unknown or to one of the categories in the source domain, it is not possible to belong to multiple categories at the same time. Thereby, we introduce a filtering strategy to avoid semantic ambiguity. Here, we just set the category with maximum similarity as the target. It is worth noting that our algorithm does not require the above pseudo-label ŷ_t to be one-hot encoded. Those pseudo labels with all-zero encoding mean that these data samples belong to the “unknown" target-private categories 𝒴̅_t. To realize “known" and “unknown" separation, we then manually set those all-zero encoding pseudo labels to a uniform encoding, i.e., q̂_c = 1/C_s. §.§ Confidence based Source-private Suppression In the above section, we developed the one-vs-all global clustering algorithm to assign pseudo labels for OSDA, i.e., 𝒴_s ⊂𝒴_t, when the number of categories in the target domain C_t is available. However, in addition to OSDA, we may also encounter PDA and OPDA, where the source domain contains categories absent in the target domain. To make the above algorithm applicable to both OSDA, PDA and OPDA, it is necessary to tailor the proposed algorithm to prevent those source-private categories from misleading pseudo-label assignments. We empirically found that on positive data group 𝒫 sampled with top-K on the target domain, those source-private categories generally yield lower mean prediction confidence than those source-target shared categories. In light of this observation, we design a source-private category suppression strategy based on the mean prediction confidence of the positive data group 𝒫. Specifically, for a particular category c∈ C_s, we tailored the Eq. <ref> to: ϵ_c = ρ + 1-ρ/K∑_x_t∈𝒫_cδ_c(f_t(x_t)), q̂_c = { 1,if ϵ_c· S(g_t(x_t), p_c) ≥max{S(g_t(x_t), n_c^i)}_i=1^M 0,if ϵ_c· S(g_t(x_t), p_c) < max{S(g_t(x_t), n_c^i)}_i=1^M. where ϵ_c is the designed source-private suppression weight for the c-th category, and ρ is a hyper-parameter to control this weight. We empirically set ρ to 0.75 for all datasets. Its sensitivity analysis can be found in the experiment. §.§ Silhouette Based Target Domain C_t Estimation Based on the previous sections, we now have achieved the pseudo-labeling algorithm for SF-UniDA. However, it is still not applicable yet, due to the requirement of prior information, i.e., the number of categories C_t in the target domain, which is commonly unavailable in reality. Therefore, the last obstacle for us is: How to determine the number of categories C_t in the target domain? To address this, a feasible solution is to first enumerate the possible values of the number of categories C_t in the target domain and divide the target domain into the corresponding clusters by applying a clustering algorithm like K-means <cit.>. Then the clustering evaluation criteria <cit.> can be employed to determine the appropriate number of target domain categories C̃_t. In this paper, we employ the Silhouette criterion <cit.> to facilitate estimating C̃_t. Technically, for a data sample x_t ∈𝒞_I, the Silhouette value s(x_t) is defined as: a(x_t) = 1/|𝒞_I| - 1∑_x ∈𝒞_I, x x_t d(x_t, x), b(x_t) = min_J≠ I1/|𝒞_J|∑_x ∈𝒞_J d(x_t, x), s(x_t) = b(x_t) - a(x_t)/max{a(x_t), b(x_t)}. where a(x_t) and b(x_t) measure the similarity of x_t to its own cluster 𝒞_I (cohesion) and other clusters 𝒞_J, J≠ I (separation), respectively. d(x_i, x_j) measures the distance between data points x_i and x_j, and |𝒞_I| denotes the size of cluster 𝒞_I. The Silhouette value s(x_t) ranges from -1 to +1, where a high value indicates that the data sample x_t has a high match with its own cluster and a low match with neighboring clusters. Therefore, if most of the data samples have high Silhouette values, then the clustering configuration is appropriate; otherwise, the clustering configuration may have too many or too few clusters. Since it is challenging to obtain the exact number of target domain categories C_t, in our implementation, we empirically enumerate the possible values of C̃_t as [1/3C_s, 1/2C_s, C_s, 2C_s, 3C_s], taking into account the scenarios may encounter. Note that we only estimate the value of C̃_t at the beginning, and subsequently, we do not change the value of C̃_t considering the overall efficiency. §.§ Local Consensus Clustering Although the global one-vs-all clustering pseudo-labeling algorithm encourages the separation between “known" and “unknown" data samples, semantically incorrect pseudo-label assignments still occur due to domain shift and category shift, resulting in negative transfer. To mitigate this, we further introduce a local k-NN consensus clustering strategy that exploits the intrinsic consensus structure of the target domain 𝒟_t. Specifically, during model adaptation, we maintain a memory bank 𝒢_t = {g_t(x_t), δ(f_t(x_t))}_x_t ∈𝒟_t, which contains the target features and corresponding prediction scores. The local k-NN consensus clustering is then realized by: l_c^i = 1/|L^i|∑_x_t∈L^iδ_c(f_t(x_t)), ℒ_tar^loc = -1/N∑_i=1^N∑_c=1^C_s l_c^ilogδ_c(f_t(x_t^i)). where δ_c(f_t(x_t)) denotes the soft-max probability of data instance x_t belonging to the c-th class, L^i refers to the set of nearest neighbors of data x_t^i in the embedding feature space. Here, we apply the cosine similarity function to find the nearest neighbors L^i of x_t^i in the memory bank 𝒢_t. We then encourage minimizing the cross entropy loss between x_t^i and the nearest neighbors L^i to achieve the local semantic consensus clustering. §.§ Optimization Objective The overall training loss of GLC can be written as: ℒ_tar^glb = -1/N∑_i=1^N∑_c=1^C_sq̂_c^ilogδ_c(h_t(g_t(x_t^i)))), ℒ_tar = ηℒ_tar^glb + ℒ_tar^loc. where q̂_c^i denotes the global clustering pseudo label for data sample x_t^i, and ℒ_tar^glb is the corresponding global cross-entropy loss. η > 0 is a trade-off hyper-parameter. §.§ Inference Details As there is only one standard classification model, we apply the normalized Shannon Entropy <cit.> as the uncertainty metric to separate known and unknown data samples: I(x_t) = - 1/log C_s∑_c=1^C_sδ_c(f_t(x_t)) logδ_c(f_t(x_t)) where C_s is the class number of source domain 𝒟_s, and δ_c(f_t(x_t)) denotes the soft-max probability of data sample x_t belonging to the c-th class. The higher the uncertainty, the more the model f_t tends to assign an unknown label to the data sample. During inference stage, given an input sample x_t, we first compute I(x_t) and then predict the class of y(x_t) with a pre-defined threshold ω as: y(x_t) = { unknown, if I(x_t) ≥ω argmax(f_t(x_t)), if I(x_t) < ω. which either rejects the input sample x_t as unknown or classifies it into a known class. In our implementation, we set ω = 0.55 for all standard benchmark datasets. Its sensitivity analysis can be found in the experiments. § EXPERIMENTS §.§ Setup Dataset: We utilize the following standard datasets in DA to evaluate the effectiveness and versatility of our method. Office-31 <cit.> is a widely-used small-sized domain adaptation benchmark, consisting of 31 object classes (4,652 images) under office environment from three domains (DSLR (D), Amazon (A), and Webcam (W)). Office-Home <cit.> is another popular medium-sized benchmark, consisting of 65 categories (15,500 images) from four domains (Artistic images (Ar), Clip-Art images (Cl), Product images (Pr), and Real-World images (Rw)). VisDA-C <cit.> is a more challenging benchmark with 12 object classes, where the source domain contains 152,397 synthetic images generated by rendering 3D models and the target domain consists of 55,388 images from Microsoft COCO. Domain-Net <cit.>, is the largest domain adaptation benchmark with about 0.6 million images, which contains 345 classes. Similar to previous works <cit.>, we conduct experiments on three subsets from it (Painting (P), Real (R), and Sketch (S)). We evaluate our GLC on partial-set DA (PDA), open-set DA (OSDA), and open-partial-set DA (OPDA) scenarios. Detailed classes split are summarized in Table <ref>. Evaluation protocols: For a fair comparison, we utilize the same evaluation metric as previous works <cit.>. Specifically, in PDA scenario, we report the classification accuracy over all target samples. In OSDA and OPDA scenarios, considering the trade-off between “known" and “unknown" categories, we report the H-score, i.e., the harmonic mean of the accuracy of “known" and “unknown" samples. Implementation details: We adopt the same network architecture with existing baseline methods. Specifically, we adopt the ResNet-50 <cit.> pre-trained on ImageNet <cit.> as the backbone for all datasets. For preparing the source model, here, we utilize the same network structure and training recipe as SHOT <cit.>. We present more details about source model training in the supplementary. During target model adaptation, we apply the SGD optimizer with momentum 0.9. The batch size is set to 64 for all benchmark datasets. We set the learning rate to 1e-3 for Office-31 and Office-Home, and 1e-4 for VisDA and DomainNet. For hyper-parameter, as we described in previous sections, we set ρ to 0.75 for all datasets. For local k-NN consensus clustering, |L| is set to 4 for all benchmarks. As for η, we set it to 0.3 for Office-31, VisDA, and 1.5 for Office-Home and DomainNet. All experiments are conducted on an RTX-3090 GPU with PyTorch-1.10. §.§ Experiment Results To verify the effectiveness of our GLC, we conduct extensive experiments on three possible category-shift scenarios, i.e., open-partial-set DA (OPDA), open-set DA (OSDA), and partial-set DA (PDA). We compare GLC with data-dependent and more recent data-free methods to empirically demonstrate the merit of GLC. In adaptation, data-dependent methods typically require access to source raw-data, while data-free methods require source pre-trained models. In particular, GLC requires only a standard pre-trained source model, i.e., without any dedicated model architectures as <cit.>. For a fair comparison, all methods are performed without the prior knowledge of category-shift, except those designed only for specific scenarios. Results on OPDA: We first conduct experiments on the most challenging setting, i.e., OPDA, in which both source and target domains involve private categories. Results on Office-Home are summarized in Table <ref>, and results on Office-31, VisDA and DomainNet are summarized in Table <ref>. As shown in Table <ref> and Table <ref>, our GLC achieves new state-of-the-arts, even compared to previous data-dependent methods. Especially, on VisDA, GLC achieves the H-score of 73.1%, which surpasses GATE <cit.> and UMAD <cit.> by a wide margin (16.7% and 14.8%). On the largest benchmark, i.e., DomainNet, GLC still achieves consistent performance improvements compared to UMAD and GATE, with gains of approximately 8.0% and 3.0%. Results on OSDA: We then conduct experiments on OSDA, where only the target domain involves categories not presented in the source domain. Results on Office-Home, Office-31, and VisDA are summarized in Table <ref>. As shown in Table <ref>, GLC still achieves state-of-the-art performance. Specifically, GLC obtains 69.8% H-score on Office-Home and 72.5% H-score on VisDA, with an improvement of 3.4% and 5.7% compared to UMAD. Results on PDA: We last verify the effectiveness of GLC on PDA, where the label space of the target domain is a subset of the source domain. Results summarized in Table <ref> show that GLC still achieves comparable performance compared to methods tailored for PDA. In a fairer comparison, GLC clearly outperforms UMAD, specifically achieving performance gains of 6.2%, 4.6%, and 7.7% on Office-31, Office-Home, and VisDA, respectively. §.§ Experiment Analysis Ablation Study: To verify the effectiveness of different components within GLC, we conduct extensive ablation studies on Office-31, Office-Home, and VisDA in OPDA scenarios. The results are summarized in Table <ref>. Here GLC w/o ℒ_tar^glb refers to that we only employ the local k-NN consensus clustering loss to regulate model adaptation, while GLC w/o ℒ_tar^loc denotes that we only employ the global one-vs-all clustering based pseudo-labeling algorithm to achieve model adaptation. From these results, we can conclude that our local and global clustering strategies are complementary to each other. And global clustering is of vital importance to help us distinguish “known" and “unknown" categories. For example, on VisDA, with only ℒ_tar^glb, we can advance the source model from the H-score of 25.7% to 66.0%, and outperform GATE by 9.6%. Hyper-parameter Sensitivity: We first study the parameter sensitivity of η and ρ on Office-31 under OPDA setting in Fig. <ref> (a-b), where η is in the range of [0.1, 0.2, 0.3, 0.4, 0.5], and ρ is in the range of [1/2, 2/3, 3/4, 4/5, 1.0]. Note that ρ = 1.0 denotes that we do not introduce the confidence based source-private suppression mechanism. It is easy to find that results around the selected parameters η = 0.3 and ρ = 0.75 are stable, and much better than the source model. By oracle validation, we may find better hyper-parameter settings, e.g., η = 0.1 and ρ = 0.50. In Fig. <ref> (c), we present the H-score with respect to ω on Office-Home in OPDA. For all benchmarks, we pre-define ω = 0.55 to separate “known" and “unknown" samples. The results show that H-score is relatively stable around our selection, and we could achieve better performance when setting ω to 0.65 via oracle validation. Besides, in Fig. <ref> (d), we illustrate the H-score convergence curves on VisDA. Varying Unknown Classes: As increasing “unknown" classes, it becomes more difficult to correctly identify the “unknown" and “known" objects. To examine the robustness of GLC, we compare GLC with other methods when varying unknown classes on Office-Home under OPDA setting. Fig. <ref> shows that GLC achieves more stable and much better performance against existing methods. §.§ Discussion So far, most existing domain adaptation methods designed for category shift are not applicable to the vanilla closed-set DA (CLDA). To verify the effectiveness of GLC in CLDA, we have conducted experiments on Office-31 and Office-Home in the Appendix. Moreover, existing methods usually perform experiments only on standard computer science benchmarks. Here, we have validated the effectiveness of GLC in more realistic applications, including remote sensing in PDA, wildlife classification in OSDA, and single-cell RNA sequence identification in OPDA. These results are also presented in the Appendix. § CONCLUSION In this paper, we have presented Global and Local Clustering (CLC) for upcycling models under domain shift and category shift. Technically, we have devised an innovative one-vs-all global clustering strategy to realize “unknown" and “known" data separation, and introduced a local k-NN clustering strategy to alleviate negative transfer. Compared to existing approaches that require source data or are only applicable to specific category shifts, GLC is appealing by enabling universal model adaptation on the basis of only standard pre-trained source models. Extensive experiments in partial-set, open-set, and open-partial-set DA scenarios across several benchmarks have verified the effectiveness and superiority of GLC. Remarkably, GLC significantly outperforms existing methods by almost 15% on the VisDA benchmark in open-partial-set DA scenario. Acknowledgment: This work was supported by Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01), ZJ Lab, and Shanghai Center for Brain Science and Brain-Inspired Technology, the Shanghai Rising Star Program (No.21QC1400900), and the Tongji-Westwell Autonomous Vehicle Joint Lab Project. § SOURCE MODEL PREPARING As aforementioned, in this paper, we focus on the K-way classification. For a given source domain 𝒟_s = { (x_i^s, y_i^s) }_i=1^N_s where x^i_s ∈𝒳_s and y^i_s ∈𝒴_s ⊂ℝ^K, we adopt the same recipe as SHOT <cit.> and BMD <cit.> to prepare the source model. Specifically, the source model f_s parameterized by a deep neural network consists of two modules: the feature encoding module g_s: 𝒳_s →ℝ^d and the classifier module h_s: ℝ^d →ℝ^K, i.e., f_s = h_s ∘ g_s. We optimize f_s with the following loss: ℒ_src = -1/N∑_i=1^N∑_k=1^K q_k logδ_k(f_s(x^i_s)) where δ_k (f_s(x_s)) denotes the softmax probability of source sample x_s belonging to the k-th category, q_k is the smoothed one-hot encoding of y_s, i.e., q_k = (1 -α)*1_[k=y_s] + α / K, and α is the smoothing parameter which is set to 0.1 for all benchmarks. § EXPERIMENTS ON CLOSED-SET ADAPTATION Existing methods designed for category shift, typically do not perform well for the vanilla closed-set domain adaptation scenario (CLDA). To examine the effectiveness and robustness of GLC, we further conduct experiments on Office-31, and Office-Home. All implementation details are the same as before, e.g., we adopt the ResNet-50 <cit.> as the backbone, the learning rate is set to 1e-3, and ρ is set to 0.75. The results are listed in Table <ref> of this supplementary material. As shown in this Table, despite the fact that the GLC is not tailored for CLDA, we still attain comparable or even better performance compared to existing methods designated for CLDA, e.g., MDD <cit.>. Specifically, GLC obtains the overall accuracy of 88.1% and 70.4% on Office-31 and Office-Home, respectively. While MDD attains 88.9% and 68.1% on Office-31 and Office-Home, respectively. In a fairer comparison, GLC significantly outperforms UMAD <cit.>, which is also model adaptation method designed for category-shift. Specifically, GLC outperforms UMAD by 6.4% and 7.3% on Office-31 and Office-Home. § EXPERIMENTS ON REALISTIC APPLICATIONS So far, most existing methods usually perform experiments only on standard computer science benchmarks. Here, we have further validated the effectiveness and superiority of GLC in realistic applications, including remote-sensing recognition, wild-animal classification, and single-cell RNA sequence identification. We present more details in the following. §.§ Partial-set Model Adaptation on Remote Sensing Recognition Remote sensing has great potential to manage global climate change, population movements, ecosystem transformations, and economic development. However, due to data protection regulations <cit.>, it is difficult for researchers to obtain multi-scene, high-resolution satellite imagery. For example, there are strict data regulation policies in China for high-resolution remote sensing images in meteorological, oceanic, and environmental scenarios <cit.>. To validate the effectiveness of GLC on remote sensing, we conduct experiments on two existing large-scale datasets, the PatternNet <cit.> and the NWPU45 <cit.> dataset. PatternNet is one of the largest satellite image datasets collected from Google Earth imagery in the US. It contains 38 scene classes and 30,400 high-resolution (0.2∼6m per pixel) satellite images, such as airport, beach, dense residential, forest, etc. NWPU45 dataset consists of 45 scene classes and 31,500 satellite images covering more than 100 countries and regions around the world. Its spatial resolution varies from about 30 to 0.2 m per pixel. The heterogeneity of spatial resolution and geographic location poses a significant challenge to model adaptation. In this paper, we set the PatternNet as source dataset and the NWPU45 as target dataset to investigate the model adaptation from high-resolution satellite images to low-resolution satellite images. There are 21 overlapping scenes classes between PatternNet and NWPU45. Thus, we transfer the scene classes from the PatternNet to the 21 overlapping scene classes in the NWPU45 and compare the results with the original labels from the NWPU45 for performance evaluation. An illustration of boxplot in Fig. <ref>b basically demonstrates that GLC effectively realizes model adaptation and achieves more accurate performance with less variance than existing methods. Quantitatively, GLC achieves 64.6 ± 0.22 % overall accuracy with 5 different random seeds. In contrast to GLC, the baseline methods DCC, ETN, DANCE, and BA3US obtain 55.2 ± 0.80 %, 54.9 ± 0.77 %, 62.0 ± 1.02 % and 59.6 ± 1.25 % overall accuracy, respectively. To verify the robustness, we further conduct an ablation experiment on decreasing the number of overlapping scenes between source and target domains. The number gets smaller, there is more probability of overlapping samples being categorized into other scenes. Despite this, the results in Fig. <ref>c show that GLC is still capable of addressing this challenge and even achieving better performance. Quantitatively, GLC obtains 64.5% average accuracy in four different target situations. In contrast, the baseline methods DCC, ETN, DANCE, and BA3US attain 55.4%, 55.0 %, 52.4%, and 49.6% overall accuracy, respectively. We attribute this to our global one-vs-all clustering algorithm, which is able to discover non-existent scene categories and suppress model adaptation over these categories. §.§ Open-set Model Adaptation on Wild Animal Classification We next study a more challenging setting, the open-set model adaptation on wild animal classification. Having the ability to accurately classify wild animals is important for studying and protecting ecosystems <cit.>, especially the ability to identify novel species <cit.>. However, it is almost impossible for a database to cover all animal species, and it is also typically difficult to collect and annotate a large number of wild animal images in practice. Thereby, it would be ideal if we develop an animal classification system based on the existing large number of virtual animal images on the Internet. In this article, we execute experiments on the I2AWA <cit.> benchmark to investigate open-set model adaptation from virtual to real-world. I2AWA consists of a virtual source domain dataset and a real-world target domain dataset with a total of 50 animal categories. We divide the first 30 into known categories in alphabetical order and the remaining 20 into unknown categories. The source domain dataset consists of 2,970 virtual animal images collected through the Google-Image search engine, while the target domain dataset comes from the AWA2 <cit.> dataset with a total of 37,322 images from the real world. Due to differences in image styles between virtual and real-world datasets, directly deploying a DNN model trained on the virtual images can lead to severe performance degradation. For a qualitative demonstration, we enumerate some animal images on target domain in Fig. <ref>a and apply the Grad-CAM heatmap <cit.> technique to compare the source model with the adapted target model by our GLC technique. From this, we can conclude that the source model typically fails to locate and extract key information for animal identification, while the upcycled model overcomes these failures well. For quantitative performance evaluation, we compare GLC with the methods dedicated to open-set domain adaptation (DANN <cit.>, OSBP <cit.>), and the methods designed for universal domain adaptation (DANCE <cit.>, OVANet <cit.>). An inspection of the tSNE plots (Fig. <ref>b) indicates that our GLC algorithm effectively realizes known animal classification and unknown animal separation. This observation is further demonstrated by the quantitative metric in Fig. <ref>c. Specifically, GLC achieves 79.1 ± 0.28 % overall H-Score. In contrast, the baseline methods DANN, OVANet, OSBP and DANCE obtains 70.1 ± 0.85%, 70.8 ± 0.58%, 72.2 ± 1.61%, 74.5 ± 0.32% overall H-Score. As presented in Fig. <ref>d, compared to existing methods, GLC further provides significant savings in target-domain side computational resource overhead (about 48.4% training time reduction in this case). This is due to the fact that our GLC merely fine-tunes the source model to realize adaptation, while existing source data-dependent methods need to train the target models from scratch. To visually assess the separation between the known and unknown classes, we present the uncertainty density distribution in Fig. <ref>e. The higher the uncertainty, the more the model treated the input animal image as an unknown species. The results show that while OVANet and DANCE are able to achieve promising classification of known classes of animals, they have trouble in unknown animal separation. In contrast, our GLC draws a better trade-off between known animal classification and unknown animal identification. We further examine the robustness of GLC in different open-set situations, e.g., varying target domain unknown categories and source domain known categories. As illustrated in Fig. <ref>f, we can find that GLC maintains a promising H-score compared to existing methods. Specifically, GLC achieves 78.3% overall H-Score in four different target domain unknown categories situations, while DANN, OSBP, OVANet, and DANCE obtain 69.3%, 72.9%, 70.8%, 75.4% average H-Score, respectively. Similarly, when source domain known categories varies, GLC arrives 83.4% overall H-Score in four different situations, still significantly outperforming existing methods. §.§ Open-partial-set Model Adaptation on Single-cell Identification We finally consider the most challenging scenario, open-partial-set adaptation, where both source and target domains contain private categories. Here, we implement experiments on single-cell identification. It has great potential in the studies of cell heterogeneity, developmental dynamics, and cell communications <cit.>. Currently, there are two main types of single-cell sequencing technologies, namely scRNA-seq and scATAC-seq. However, it has been noted that the extreme scarcity of scATAC-seq data tends to limit its ability for cell type identification. In contrast, large amounts of well-annotated scRNA-seq datasets have been curated as cell atlases. It motivates us to upcycle models trained on the scRNA-seq datasets and adapt them to the scATAC-seq datasets. Nevertheless, the cell types contained in different atlas data are generally inconsistent, which poses substantial challenges for model adaptation across atlases. In this article, we apply our GLC to two mouse cell atlases, the Tabula Muris atlas <cit.> for scRNA-seq data and the Cusanovich atlas <cit.> for scATAC-seq data. The Tabula Muris atlas consists of 73 cell types totaling 96,404 cells from 20 organs with two protocols profiling transcriptomics, while the Cusanovich atlas consists of 29 cell types totaling 81,173 cells from 13 tissues. There are 19 cell types common between the Tabula Muris atlas and the Cusanovich atlas. For performance evaluation, as in the wild animal experiments above, we utilize the harmonic mean accuracy H-Score of the known cell types and the unknown cell types as the quantitative metric. We compare our GLC with recently developed and applied methods for scRNA-seq and scATAC-seq integration, including the scJoint <cit.> and the Seurat v.3 <cit.>. We illustrate the tSNE plots in Fig. <ref>a to compare with the ground truth labels annotated in the Cusanovich atlas <cit.>. The tSNE plots are generated by applying the singular value of the term frequency-inverse document frequency (TD-IDF) transformation of scATAC-seq peak matrix as in the Cusanovich atlas <cit.>. It observes that GLC achieves a better trade-off between known cell types identification and unknown cell types separation than the other methods. This observation is further quantitatively demonstrated by the H-Score metric. Specifically, GLC obtains 62% overall H-Score compared with 58% for Seurat and 56% for scJoint. As presented in Fig. <ref>c, not only is there a significant performance improvement, but our GLC also brings significant savings in target domain computational resource overhead (about 75.9% training time reduction). In addition to tSNE plots, we also present the uncertainty density distribution in Fig. <ref>b, where the higher the uncertainty, the more the model tends to group the cell into the unknown cell types group. To find the best trade-off point, a global decision boundary search was performed for all methods. The decision boundary for GLC is 0.5 compared with 0.25 for Seurat and 0.05 for scJoint. It further indicates that our GLC attains an optimal trade-off in cell types identification to the other methods. § DISCUSSION During the past decades, deep neural networks (DNNs) have achieved remarkable success in various applications and fields. However, DNNs are typically restricted to the training data domain. If the test data is collected in another modality or from other types of instruments, we will typically suffer from a significant performance degradation <cit.>. This phenomenon is likely to worsen when training and testing data do not share the same ground-truth class space. Although DNNs can be adapted to different application scenarios with additional supervised learning, this paradigm asks for annotation of large-scale target domain data. It would require significant resources and experts in real-world applications, such as clinical staff for medical imaging diagnosis and genetic scientists for single-cell sequence analysis, making it extremely expensive and impossible for most scenarios. In this paper, we find that it is possible to productively upcycle existing pre-trained models and adapt them to new scenarios. Numerous empirical evidences on standard computer science benchmarks and simulated realistic applications basically demonstrate that GLC is a promising, simple, and general solution for a variety of real-world application tasks, including single-cell sequence analysis, remote sensing recognition, and other such domain-dependent problems. ieee_fullname

http://arxiv.org/abs/2303.07354v1

20230313063938

MetaTroll: Few-shot Detection of State-Sponsored Trolls with Transformer Adapters

cs.CL

State-sponsored trolls are the main actors of influence campaigns on social media and automatic troll detection is important to combat misinformation at scale. Existing troll detection models are developed based on training data for known campaigns (e.g. the influence campaign by Russia's Internet Research Agency on the 2016 US Election), and they fall short when dealing with novel campaigns with new targets. We propose MetaTroll, a text-based troll detection model based on the meta-learning framework that enables high portability and parameter-efficient adaptation to new campaigns using only a handful of labelled samples for few-shot transfer. We introduce campaign-specific transformer adapters to MetaTroll to “memorise” campaign-specific knowledge so as to tackle catastrophic forgetting, where a model “forgets” how to detect trolls from older campaigns due to continual adaptation. Our experiments demonstrate that MetaTroll substantially outperforms baselines and state-of-the-art few-shot text classification models. Lastly, we explore simple approaches to extend MetaTroll to multilingual and multimodal detection. Source code for MetaTroll is available at: https://github.com/ltian678/metatroll-code.git <ccs2012> <concept> <concept_id>10010147.10010178.10010179</concept_id> <concept_desc>Computing methodologies Natural language processing</concept_desc> <concept_significance>500</concept_significance> </concept> <concept> <concept_id>10010147.10010178.10010179.10003352</concept_id> <concept_desc>Computing methodologies Information extraction</concept_desc> <concept_significance>300</concept_significance> </concept> </ccs2012> [500]Computing methodologies Natural language processing [300]Computing methodologies Information extraction Direct tomography of quantum states and processes via weak measurements of Pauli spin operators on an NMR quantum processor Arvind March 30, 2023 =========================================================================================================================== § INTRODUCTION State-sponsored trolls, along with bots, are the main actors in influence operations and misinformation campaigns on social media. The 2016 US Presidential election for example, highlighted how foreign states can carry out mass influence campaign, and in this case by Russia's Internet Research Agency <cit.>.[<https://www.adelaide.edu.au/newsroom/news/list/2021/12/09/understanding-mass-influence-activities-is-critical>.] In another context, <cit.> studied how influence operations undermined the vaccine debate in public health. To help combat misinformation on social media, Twitter began releasing user accounts associated with state-sponsored influence activities. Our work uses this dataset for troll detection, and as such a troll in our context is a state-sponsored agent with links to an influence campaign. Existing troll detection models focus on extracting signals from user online posts and user activities <cit.>. Recent neural approaches explore fusing different feature representations for troll detection, e.g. social structure, source posts and propagated posts such as retweets on Twitter <cit.>. These studies generally formulate the task as a standard supervised learning problem which requires a substantial amount of labelled data for known campaigns. As such, they are ill-equipped to detect novel campaigns sponsored by another state for a different target. Meta-learning is a well-established framework for few-shot learning <cit.>. The idea of meta-learning is to leverage the shared knowledge from previous tasks to facilitate the learning of new tasks. In our context, meta-learning has the potential to quickly adapt a troll detection model to a new campaign via few-shot transfer, once it's meta-trained on a set of known campaigns. However, in a continual learning setting where the troll detection model needs to be constantly updated to work with new campaigns over time, the standard meta-learning framework suffers from catastrophic forgetting — the problem where the model “forgets” how to detect trolls from the older campaigns as they are updated <cit.>. The main contributions of this work are as follows: * We introduce a troll detection problem under a realistic setting where novel campaigns continually emerge, mimicking real-world events on social media platforms. * We propose MetaTroll, a text-based meta-learning framework for troll detection, which includes a three-stage meta-training process that allows it to learn knowledge across different campaigns for fast adaptation to new campaigns. * MetaTroll tackles catastrophic forgetting by introducing campaign-specific transformer adapters <cit.> (i.e. each campaign has its own set of adapter parameters), which can “memorise” campaign-specific knowledge. * MetaTroll has the ability to to work with multiple languages (multilingual) by using a pretrained multilingual model and incorporate images as an additional input (multimodal) by either encoding them using pretrained image classifiers or converting them into text via optical character recognition. * Large-scale experiments on a real-world dataset of 14 Twitter campaigns showed the superior performance of MetaTroll. § RELATED WORK Our related work comes from three areas, troll detection, meta-learning and few-shot text classification. §.§ Troll detection Early studies of troll detection focus on extracting hand-engineered features from the textual contents of user posts for troll detection <cit.>. Signals such as writing style, sentiment as well as emotions have been explored <cit.>. User online activities have also been used to detect trolls <cit.>. <cit.> presents a study on anti-social behavior in online discussion communities, focusing on users that will eventually be banned. More recently, approaches that combine user posts and online activities for troll detection have emerged <cit.>. <cit.> identify 49 linguistic markers of deception and measure their use by troll accounts. They show that such deceptive language cues can help to accurately identify trolls. <cit.> propose a detection approach that relies on users' metadata, activity (e.g. number of shared links, retweets, mentions, etc.), and linguistic features to identify active trolls on Twitter. <cit.> show that Russian trolls aim to hijack the political conversation to create distrust among different groups in the community. <cit.> similarly demonstrate how trolls act to accentuate disagreement and sow division along divergent frames, and this is further validated by <cit.> in relation to Russian ads on Facebook. In <cit.>, the authors study the effects of manipulation campaigns by analysing the accounts that endorsed trolls' activity on Twitter. They find that conservative-leaning users re-shared troll content 30 times more than liberal ones. <cit.> compare troll behavior with other (random) Twitter accounts by recognising the differences in the content they spread, the evolution of their accounts, and the strategies they adopted to increase their impact. <cit.> leverage both social structure and textual contents to learn user representations via graph embedding. We focus on how to detect trolls in emergent social activities with limited labelled data. §.§ Meta-learning Meta-learning aims to extract some transferable knowledge from a set of tasks to quickly adapt to a new task. These approaches can be divided into three categories: metric-based, optimisation-based, and model-based. Metric-based meta-learning approaches <cit.> aim to learn an embedding function to encode the input and a metric function to learn the distance (e.g. cosine distance and euclidean distance) between the query data and support data. Optimisation-based approaches <cit.> are designed to learn good parameter initialisation that can quickly adapt to new tasks within a few gradient descent steps. Model-based approaches <cit.> use neural networks to embed task information and predict test examples conditioned on the task information. Our approach combines both optimisation and model-based ideas in that we adopt <cit.> to update model parameters and a novel architecture that involves adapters and adaptive classification layers to learn task information. §.§ Few-shot text classification Text classification has shifted from task-specific training to pre-trained language models (such as BERT) followed by task-specific fine-tuning <cit.>. Recently, the language model GPT-3 <cit.> has shown strong few-shot performance for many natural language processing tasks. Few-shot text classification refer to a text classification setting where there are novel unseen tasks (domains) with only a few labelled examples for training a classification model, and meta-learning have been shown to produce strong performance  <cit.>. <cit.> introduce distributional signatures, such as word frequency and information entropy, into a meta-learning framework. <cit.> combine attention mechanism with prototypical network to solve noisy few-shot relation classification task. <cit.> leverage the dynamic routing algorithm in meta-learning to solve sentiment and intent text classification tasks for English and Chinese datasets. <cit.> propose an adaptive metric-based method that can determine the best weighted combination automatically. In our problem setting, our focus is on adapting the classifier to a new task (i.e. campaign) under the meta-learning framework. § PROBLEM STATEMENT We first explain some preliminaries of meta-learning, and contextualise it under our problem setting, before describing our model in approach. Let ℰ = {e_0, e_1, e_2,...,e_l} be a set of campaigns. These campaigns are split into two partitions: ℰ_train for meta-training and ℰ_test for meta-testing (in our case, we have 6 campaigns for ℰ_train, and 4 for ℰ_test; see Table <ref>). For each campaign e, we have labeled users 𝒰, where each user 𝐮 consists of a list of text posts and images, defined as 𝐮 = {(c_0, m_0, p_0), ..., (c_n, m_n, p_n)}, where c refers to the textual content of the post, m a set of images, and p the timestamp of the post. Each user 𝐮 is associated with a ground-truth label y ∈ Y, where Y represents the label set (troll or non-troll; binary). Our goal is to leverage the knowledge learned from past (meta-train) campaigns (ℰ_train) to adapt quickly to new (meta-test) campaigns (ℰ_test) via few-shot learning (e.g. 5 or 10 users for the new campaign). In our problem setting, the notion of a task (from meta-learning) is the binary troll detection task for a particular campaign. Each task 𝒯_i = ( 𝒟^s, 𝒟^q ) includes a support set 𝒟^s with S data points and a query set 𝒟^q with Q data points, where 𝒟^s = {𝐮_i^s, y_i^s}_i=1^S , 𝒟^q = {𝐮_i^q, y_i^q}_i=1^Q, and 𝐮_i is a user with a list of posts and images and y_i∈{ 0,1 } corresponding to troll/non-troll labels (as defined above). During meta-training, we iterate through different campaign in ℰ_train and sample support and query instances to train the model. Once that's done, to adapt to a new campaigns in ℰ_test, we follow a similar training process but we update the model parameters using only the support instances from the new campaign, and reserve the query set for evaluation (i.e. the query set is not used for model update). MetaTroll Input/hyper-parameters: meta-train campaigns ℰ_train, meta-test campaigns ℰ_test, learning rate β, γ and δ, model M Parameters: campaign-general adapter Ψ, campaign-specific adapter Ψ_e, BERT Φ, linear classifier Ω, learning rate α § APPROACH We present the overall architecture of MetaTroll in Figure <ref>. MetaTroll has two modules, a BERT-based  <cit.> feature extractor and an adaptive linear classifier. We first explain at a high level how we train MetaTroll in three stages for the feature extractor and linear classifier (Algorithm <ref>). In the first stage, we fine-tune an off-the-shelf BERT and update all its parameters (Φ) using all training samples from the meta-train campaigns ℰ_train (line 1 in Algorithm <ref>). At this stage we do not introduce adapters <cit.> to BERT and it is optimised to do binary classification of trolls and non-trolls. The idea of this step is to train a feature extractor for the troll detection task (in other words, this step is standard fine-tuning to adapt BERT for binary troll detection). In the second stage, we introduce adapters <cit.> to BERT and train the adapter parameters (Ψ) using model agnostic meta-learning (MAML: <cit.>; line 3–16 in Algorithm <ref>). Note that the adapter is shared across all campaigns (i.e. the adapter is not campaign-specific), and the idea of this stage is to learn a good initialisation for the adapter to do general troll detection; in the next stage, the learned adapter parameters will be used to initalise campaign-specific adapters. During this stage, only the adapter parameters (Ψ) are updated while the BERT parameters (Φ) are frozen. In the third and final stage, we introduce campaign-specific adapters and adaptive linear classifiers to MetaTroll, creating the full model. The campaign-specific adapters are initialised using the campaign-general adapter from stage 2. The idea of using campaign-specific adapters is to address catastrophic forgetting when MetaTroll is continuously updated for new campaigns that emerge over time in an application setting: the campaign-specific adapter solves the `forgetting' problem because the knowledge of detecting older/past campaigns are stored in their adapters which will not be overwritten as MetaTroll is continually updated. As for the adaptive linear classifiers, they are also campaign-specific and designed to encourage MetaTroll to learn campaign representations that are distinct for different campaigns.[The adaptive linear classifier parameters are initialised randomly for a new campaign.] We update the campaign-specific adapter (Ψ_e) and classifier parameters (Ω_e) via meta-learning, similar to stage 2 training (line 17–32 in Algorithm <ref>), noting that the BERT parameters (Φ) are frozen in this stage. After MetaTroll is trained (over the 3 stages), to adapt it to trolls of a new campaign at test time, we follow the third stage training to learn campaign-specific adapter (Ψ_e) and classifier (Ω_e) parameters for the new campaign. Once adapted, MetaTroll can be used to classify users from this new campaign. We next describe the training stages and test inference in detail. §.§ Stage One Training In the first stage, MetaTroll is a standard BERT fine-tuned with standard cross entropy loss to do binary classification of trolls vs. non-trolls (ignoring the campaigns). Given a user 𝐮 with posts {c_0,...,c_n}: v = BERT([CLS] ⊕ c_0 ⊕ ... ⊕ c_n) ŷ = softmax (W v + b) where v is the contextual embedding of [CLS] and ⊕ is the concatenation operation. §.§ Stage Two Training In the second stage, we add adapters to BERT (see Figure <ref>). Following <cit.>, we insert two adapter modules containing bottleneck layers into each transformer layer of BERT. Note that the adapter is shared across all campaigns here, as the goal in this stage is to learn good set of initial adapter parameter values that can be used to initialise campaign-specific adapters in the next stage. Correspondingly, Equation <ref> is now modified to: v = AdapterBERT([CLS] ⊕ c_0 ⊕ ... ⊕ c_n) We meta-train the adapter parameters using MAML (line 3–16 in Algorithm <ref>). We first sample a task 𝒯_i for campaign e from ℰ_train to create the support D^s and query set D^q. Denoting M as the model, Φ the BERT parameters and Ψ the adapter parameters, we next compute the inner loop to update Ψ as follows <cit.>: ℒ_𝒯_i(M_Φ,Ψ, 𝒟^s) = 1/|𝒟^s|∑_𝐮_i ∈𝒟^s-log p(y_i|𝐮_i ; Φ, Ψ) Ψ^' = Ψ - α∇_Ψℒ(M_Φ,Ψ, 𝒟^s) where α is the learning rate (computed next in the outer loop) and ℒ_𝒯_i(M_Φ,Ψ, 𝒟^s) is the cross entropy loss over the support set. Next we compute the cross-entropy loss based on the query set, using the updated parameters: ℒ_𝒯_i(M_Φ,Ψ^', 𝒟^q) = 1/|𝒟^q|∑_𝐮_i ∈𝒟^q-log p(y_i|𝐮_i ; Φ, Ψ^') This inner loop is carried out for multiple steps of gradient descent (using several tasks from the same campaign). Note that in this stage Φ (BERT parameters) is frozen. Once that's done, we update the adapter parameters Ψ and inner loop learning rate α:[In our implementation, α is layer-specific, i.e. we have a separate learning rate for each adapter in different layers.] Ψ←Ψ-β∇_Ψ𝔼_𝒯_i ∼𝒯[ℒ_𝒯_i(M_Φ,Ψ^', 𝒟^q)] α←Ψ-β∇_α𝔼_𝒯_i ∼𝒯[ℒ_𝒯_i(M_Φ,Ψ^', 𝒟^q)] where β is the learning rate for the outer loop (set to 1e^-5 in our experiments). §.§ Stage Three Training In this stage, we introduce campaign-specific adapters for each campaign, and they are all initialised using the campaign-general adapter learned from the previous stage. Formally, Equation <ref> is now updated to: v = AdapterBERT_e([CLS] ⊕ c_0 ⊕ ... ⊕ c_n) where e is the campaign. We introduce campaign-specific adapters to our model for two reasons: (1) they are more efficient to train and less vulnerable to overfitting (which is important in a few-shot learning setting), since they contain only a small number of parameters compared to the alternative where we have one BERT model for every campaign); and (2) they alleviate catastrophic forgetting in a continual learning setting, as each campaign has its own adapter. Inspired by <cit.>, we next introduce an adaptive linear classifier network that replaces W and b in the linear layer used for classification (Equation <ref>). Intuitively, this adaptive classifier works by first computing the aggregate troll and non-troll representation for each campaign, and then learn campaign-specific projections to classify between trolls vs. non-trolls. Let 𝒟_0^s and 𝒟_1^s denote the support set where the labels are trolls (y=0) and non-trolls (y=1) respectively, we compute W_e, b_e and ŷ for troll campaign e as follows: W_e^0 = 1/|𝒟_0^s|∑_v_ ∈𝒟_0^s( v ) b_e^0 = 1/|𝒟_0^s |∑_v_ ∈𝒟_0^s( v ) W_e^1 = 1/|𝒟_1^s|∑_v_ ∈𝒟_1^s( v ) b_e^1 = 1/|𝒟_1^s |∑_v_ ∈𝒟_1^s( v ) ŷ = softmax (W_e v + b_e) where W^i_e denotes the (i-1)^th column of W_e. In other words, MetaTroll classifies a user based on whether its representation (v) is closer to the (average) troll or non-troll representation. The campaign-specific adaptive linear classifier parameters Ω_e and the adapters parameters Ψ_e are trained using MAML, just like stage 2 training (line 17–31 in Algorithm <ref>). §.§ Meta-testing After MetaTroll is trained, to adapt it to a new campaign e , we follow the process of the third stage training. To simulate few-shot learning, we sample only a small number of instances for the support and query set (e.g. 5 each), and use only the support set for updating the adapter (Ψ_e) and classifier parameters (Ω_e) (line 25–26 in Algorithm <ref>) and do not run the outer loop ((line 30–31). Here the query set is used only for computing performance (i.e. in line 27 we compute accuracy instead of loss over the query set). § EXPERIMENTS AND RESULTS §.§ Datasets and models We use the information operations dataset published by Twitter for our experiments.[<https://transparency.twitter.com/en/reports/information-operations.html>] This dataset contains different groups of users banned by Twitter since October 2018 for engaging in state-sponsored information operations, and each group represents a campaign in our work. For example, the “Iran-2018-Palestine” campaign refers to trolls sponsored by Iran for an information campaign targeting Palestine in 2018.[<https://blog.twitter.com/en_us/topics/company/2018/enabling-further-research-of-information-operations-on-twitter>] To clarify, these campaigns are defined by Twitter when they release the data, and each campaign is associated with a blogpost that explains the information operation. For each campaign, we filter users and keep only those who have posted a tweet within the 6-month event period (“Event Time” in data) to remove users who are inactive during the event.[The event period is determined as follows: (1) the end date is the last post in the campaign; and the (2) start date is 6 months from the end date.] For each user, we also filter their tweets to keep only their most recent 20 posts that have a timestamp within the event period. To create the non-troll users, we combine two sources: (1) “Random”, random accounts that are sampled by generating random numeric user IDs and validating their existence following <cit.>; and (2) “Hashtag”, users whose posts contain popular hashtags used by a campaign, where popular hashtags are defined as hashtags that collectively cover 75% of trolls' posts. The reason why we have two types of non-troll users is that if we only sample random users as non-trolls, the post content of non-trolls would be topically very different to that of the trolls, and the detection task would degenerate into a topic detection task. The non-troll users sampled using the “Hashtag” approach is designed to circumvent this and makes the detection task more challenging. Table <ref> present some statistics for trolls and non-trolls in different campaigns, where 6 are used for meta-training and 4 for meta-testing. Note that the non-trolls of a particular campaign are users sampled with the “Hashtag” approach, and the last row corresponds to non-troll users sampled using the “Random” approach. For these campaigns, at least 80% of the trolls' posts are in English, and so they are used for the monolingual (English) experiments in the paper.[For non-troll users, we only keep English tweets (based on the predicted language given in the metadata).] The trolls and non-trolls in Table <ref> represent the pool of users which we draw from to construct the final training/testing data. In all of our experiments, we keep the ratio of trolls to non-trolls to 50/50 through sampling, and when sampling for non-troll users, the ratio from the two sources (“Random” and “Hashtag”) is also 50/50.[We also only sample “Random” users whose most recent post is in the event time.] As an example, “Uganda-2021-NRM” has 334 troll users. We therefore sample 334 non-troll users, where 167 are from “Random” and another 167 from “Hashtag”. We compare our MetaTroll model against the following baselines including state-of-the-art meta-learning methods and few-shot text classification models: * BERT <cit.>[<https://huggingface.co/docs/transformers/model_doc/bert>]: BERT fine-tuned using the support set of meta-test data. * KNN <cit.>: K-nearest neighbour classifier with off-the-shelf BERT as the feature extractor.[K is selected from [5,10].] * AdBERT <cit.>[<https://adapterhub.ml/>]: BERT that fine-tunes an adapter for each campaign. * GPT3 <cit.>[<https://gpt3demo.com/apps/openai-gpt-3-playground>]: a very large pretrained model adapted to our tasks using prompt-based learning <cit.>.[We use “text-davinci-002” in our experiments. The prompts are a small set of instances (e.g. 5) with their labels added to the beginning of the input.] * MAML <cit.>[<https://github.com/tristandeleu/pytorch-meta/tree/master/examples/maml>]: BERT trained with the MAML algorithm for few-shot learning. * CNP <cit.>[<https://github.com/deepmind/neural-processes>]: a model-based meta-learning framework that consists of a shared encoder, aggregator and decoder. * ProtoNet <cit.>[<https://github.com/orobix/Prototypical-Networks-for-Few-shot-Learning-PyTorch>]: a deep metric-based approach using sample average as class prototypes and the distance is calculated based on euclidean-distance. * Induct <cit.>: a few-shot classification model that uses a dynamic routing algorithm to learn a class-wise representation. * HATT <cit.>[<https://github.com/thunlp/HATT-Proto>]: A classification model of metric-based meta-learning framework together with attention mechanism. * DS <cit.>[<https://github.com/YujiaBao/Distributional-Signatures>]: A few-shot text classification model that uses distributional signatures such as word frequency and information entropy for training. §.§ Results We first present English troll detection performance. In Sections <ref> and <ref> we extend MetaTroll to work with non-English data and images. As we focus on English here, non-English posts are discarded (although this only reduces the data size by a negligible amount, as most data is in English, as explained in datasets). All reported figures are an average accuracy performance over 5 runs with different random seeds.[Noting that the performance for one run is an average performance over the query sets from multiple tasks. ] Table <ref> presents 5-shot and 10-shot results for the 4 meta-test campaigns.[5-shot means only 5 labelled instances for each class are given in the support set.] MetaTroll is the best model, achieving an average of 76.35% accuracy over all campaigns. That said, some of the few-shot text classifiers (Induct e.g.73.57%) are not far behind. Most models only benefit marginally by seeing 5 more examples going from 5- to 10-shot, with the exception of GPT3 where we see a substantial performance boost (average 14.64% gain). §.§ Continual learning performance In continual learning <cit.>, new tasks appear over time, where the goal of learning is to adapt the model accordingly to the new tasks without forgetting the previous tasks. This is a more realistic setting for a troll detection system, as it should continually adapt to new campaigns that appear over time. But in this setting it will suffer from catastrophic forgetting <cit.>, where after adapting to newer campaigns its performance to classify older campaigns will degrade. To simulate this continual learning setting, we next evaluate the troll detection models on a past campaign after it has been adapted for a number of campaigns in sequence. For example, a system is first adapted to GRU-2020-NATO (G), and then to IRA-2020-Russia (I), Uganda-2021-NRM (U) and China-2021-Xinjiang (C) in sequence (denoted as G→I→U→C). We then test the system using trolls using the past campaigns, i.e. G, I and U. One challenge with MetaTroll under this continual learning evaluation is that at test time it needs to know which adapter to use — information that most other systems do not require as they don't have campaign-specific parameters (exception: AdBERT and GPT-3[GPT3 technically does not have campaign-specific parameters, but it needs to be given campaign-specific prompts, and so requires the campaign label.]). We experiment a simple approach to solve this: have MetaTroll classify a user using all its adapters, and select the outcome that has the highest probability.[We do the same for AdBERT and GPT3.] We present (5-shot) troll classification results under this continual learning setting in Table <ref>. AdBERT and GPT-3 suffers little catastrophic forgetting, as they have campaign-specific parameters or prompts (they are unaffected by continual learning as their base model is unchanged), although their performance is only marginally above random chance in the first place. MetaTroll is the clear winner here, with only <5% accuracy degradation over time.[Note that performance of these systems on older campaigns will still degrade slightly over more adaptations, as there are more campaign-specific adapters or prompts to select from.] In contrast, the meta-learning methods and few-shot classifiers suffer from catastrophic forgetting and their performance on older campaigns drops substantially (e.g. MAML's performance for GRU-2020-NATO drops from 68.32 to 65.15 and 57.62 after several updates). An unintended benefit of that MetaTroll classifies trolls using multiple adapters is that it is effectively doing multi-class rather than binary classification (with over 85% accuracy for campaign classification; Table <ref>). What this means is that not only MetaTroll is able to retain its performance for classifying trolls vs. non-trolls from different campaigns over time, it can also predict which campaign an instance belongs to – an arguably more difficult task. §.§ Multilingual performance We extend MetaTroll to multilingual. To encode multilingual input text, we replace BERT with XLM-R <cit.>.[Specifically, we use the following library that implements adapters into XLM-R: <https://docs.adapterhub.ml/classes/models/xlmroberta.html>] For comparison, we include baselines BERT, KNN, and AdBERT, and meta-learning methods ProtoNet and MAML as their base model can be replaced with XLM-R. We exclude the few-shot classifiers as they are designed for English and cannot be trivially adapted to other languages. In terms of data, we expand the meta-test campaigns by adding four new campaigns (Thailand-2020-RTA, Mexico-2021-Election, Venezuela-2021-Gov, and China-2021-Changyu) where the predominant language is not in English and present their statistics in Table <ref>. For the 10 English campaigns (data), we restore the previously discarded non-English posts and include them for meta-training and meta-testing. Results are presented in Table <ref>. Generally we see that all models' performance has somewhat degraded when their feature extractor is replaced with a multilingual model, although MetaTroll manages to keep its accuracy around 60%. Interestingly, China-2021-Changyu appears to be the most difficult campaign, and we suspect it may be due to its diverse set of languages (38% French, 12.3% English, 12.1% Simplified Chinese). §.§ Multimodal performance Next we consider incorporating images posted by users, as image is an effective communication device (e.g. memes). Note that we only present results for different variants of MetaTroll here, as we have demonstrated that it is the most competitive detection system. To process images, we use pre-trained ResNet18 <cit.> as an off-the-shelf tool to extract image embeddings. We also explore using a multilingual OCR model to extract text information from images (which will be useful for processing memes).[https://github.com/JaidedAI/EasyOCR] As we have multiple images for each user, we aggregate the ResNet18 image embeddings via max-pooling, and concatenate the max-pooled vector with the text representation (v in Equation <ref>). For texts that are extracted by OCR, we concatenate them into a long string and process them with another AdapterBERT (Equation <ref>; their parameters are not shared), and similarly concatenate the final CLS representation to the text representation. Results are in Table <ref>. “+image#” is a baseline where we concatenate a numeric feature that denotes the number of images posted by the user to the text representation. Interestingly, even with the baseline approach we see a small improvement, indicating that trolls use more images (e.g. average number of images used by trolls is 20 vs. 5 for non-trolls in GRU-2020-NATO). Incorporating either ResNet or OCR encodings boosts performance further (with OCR being marginally more beneficial), and that adding them both produces the best performance. § CONCLUSION We propose MetaTroll, a few-shot troll detection model with campaign-specific adapters that tackles catastrophic forgetting in a continual learning setting. Experimental results show that MetaTroll outperforms existing state-of-the-art meta-learning and few-shot text classification models, and it can be extended to handle multilingual and multimodal input. § ACKNOWLEDGEMENTS This research is supported in part by the Australian Research Council Discovery Project DP200101441. Lin Tian is supported by the RMIT University Vice-Chancellor PhD Scholarship (VCPS). ACM-Reference-Format § IMPLEMENTATION DETAILS We implement our models in PyTorch using the HuggingFace library[<https://github.com/huggingface>] and their pretrained BERT[<https://huggingface.co/bert-base-cased>] and XML-R[<https://huggingface.co/xlm-roberta-base>]. The adapter-based models are from AdapterHub[urlhttps://github.com/adapter-hub/adapter-transformers]. To handle images, we use ResNet18 [<https://pytorch.org/vision/main/models/generated/torchvision.models.resnet18.html>] and EasyOCR [<https://github.com/JaidedAI/EasyOCR>]. We set maximum token length = 320 and dropout rate = 0.2 for BERT embedding and dropout rate = 0.1 for meta-learning process. Learning rate is tuned in the range between [1e^-5, 5e^-5] for BERT in the first stage. Learning rate warmup is set up 10% of steps. The learning rate β for MAML outer loop is 1e^-5. In the first stage, BERT uses the Adam optimiser <cit.>. Search space for learning rate γ and δ is from [1e^-4, 1e^-5, 2e^-5,3e^-5,4e^-5,5e^-5]. The outer loop learning rate for the third stage δ is set as 1e^-5. The inner loop learning rate γ for BERT-based models is set as 2e^-5 and for XLM-R-based models is set as 5e^-5. We experimented with the number of training tasks in the range of 60,000 to 100,000, with 80,000 tasks generally yielding the best results. Our experiments are running using A100 GPU with 40GB Memory. § ABLATION STUDY We conduct ablation study to justify the importance of our three stages training processes. By removing the adaptive classifier, we attach a standard linear layer for the final classification. For “s1+s2”, we attach a standard linear layer for the final classification with removing the adaptive classifier. “s2+s3” is initialise the BERT text encoder without pre-training on any troll data. “s1+s3” drops the meta-trained task-specific adapter. We still allocate one adapter to each task with randomly initialised adaptor. We report the average accuracy on four English meta-testing campaigns of different model variants in Table <ref>. Results show that our model, the combination of all three training stages achieves the best average accuracy of 76.35%. Without pretraining the base text encoder, the framework performs the worst, resulting in 3.04% drops on average accuracy. § DATASET DETAILS We further include a detailed statistics of our Twitter troll data with detailed topic for each campaign and online post link attached.

http://arxiv.org/abs/2303.07237v1

20230313161012

The atomic-to-molecular hydrogen transition in the TNG50 simulation: Using realistic UV fields to create spatially resolved HI maps

astro-ph.GA

firstpage–lastpage Sim-to-Real Deep Reinforcement Learning based Obstacle Avoidance for UAVs under Measurement Uncertainty Bhaskar Joshi*, Dhruv Kapur*, Harikumar Kandath ^† March 30, 2023 ======================================================================================================= Cold gas in galaxies provides a crucial test to evaluate the realism of cosmological hydrodynamical simulations. To extract the atomic and molecular hydrogen properties of the simulated galaxy population, postprocessing methods taking the local UV field into account are required. We improve upon previous studies by calculating realistic UV fields with the dust radiative transfer code SKIRT to model the atomic-to-molecular transition in TNG50, the highest-resolution run of the IllustrisTNG suite. Comparing integrated quantities such as the mass function, we study to what detail the UV field needs to be modelled in order to calculate realistic cold gas properties. We then evaluate new, spatially resolved comparisons for cold gas in galaxies by exploring synthetic maps of atomic hydrogen at redshift zero and compare them to 21-cm observations of local galaxies from the WHISP survey. In terms of non-parametric morphologies, we find that TNG50 maps are less concentrated than their WHISP counterparts (median Δ C≈0.3), due in part to central deficits related to the ejective character of supermassive black hole feedback in TNG. In terms of the column density distribution function, we find discrepancies between WHISP and IllustrisTNG that depend on the total abundance in these datasets as well as the postprocessing method. To fully exploit the synergy between cosmological simulations and upcoming deep /H_2 data, we advocate the use of accurate methods to estimate the UV radiation field and to generate mock maps. galaxies: ISM – galaxies: structure – ISM: molecules – radio lines: ISM – methods: numerical § INTRODUCTION The physical processes regulating the interstellar medium (ISM) are integral to understand the evolution of galaxies. The most important process in galaxy evolution, star formation, occurs in the cold and dense molecular ISM. For molecular hydrogen to form, enough atomic gas (potentially supported by the presence of dust) is required to cool and shield the molecular clouds against energetic ultraviolet (UV) radiation. This leads to the empirical correlation between the surface densities of gas and star-formation rate (the Kennicutt-Schmidt law, ; ), which is significantly tighter when using the surface density of molecular gas on sub-kpc scales (e.g. ). On the other hand, the ISM is actively shaped by various processes of galaxy evolution such as momentum and energy injections from supernovae and active galactic nuclei, chemical evolution due to metal return from evolved stars, large-scale galactic inflows from the gas residing in the halo, and interactions with other galaxies like ram-pressure stripping. These processes shape the ISM into a complex, multi-phase environment which is a prime target for observational campaigns targeted at understanding galaxy evolution. As a dominant fraction of the mass of the ISM exists in the form of atomic hydrogen (), observations targeted at the 21-cm line of are useful to study the structure of nearby galaxies. Unlike Ly-α radiation, photons emitted due to the spin-flip transition which gives rise to the 21-cm line penetrate both interstellar dust clouds and Earth's atmosphere. exists in the ISM mostly in two equilibrium states, the warm (T∼5000K) and the cold (T∼100K) neutral media (WNM and CNM, e.g. ). 21-cm emission from the WNM is generally assumed to be optically thin which gives a trivial relation between the observable 21-cm flux density and the physical quantity of interest, the mass surface density. 21-cm observations of the Milky Way indicate that the WNM contains most of the mass (). The CNM is dense and cold enough for self-absorption to become important which complicates the relation between 21-cm flux and mass. For 21-cm emission observations, correction factors to account for the `hidden' mass in the CNM are rather uncertain. While earlier measurements in the LMC, M31 and M33 indicated a correction factor of ≈35 % (), the more recent study by <cit.> finds for the same objects a correction factor of ≈10 %, in line with the ALFALFA survey of local galaxies (). Observational campaigns using 21-cm observations to measure the galactic content of local galaxies include blind H i-selected surveys (HIPASS, ; ALFALFA, ) and surveys targeting samples representative of the galaxy population (GASS, ; xGASS, ). These spatially unresolved surveys made use of the superior sensitivity of the single-dish Parkes and Arecibo telescopes to measure galactic masses. With these data, multiple scaling relations between the atomic gas mass and other galactic properties such as stellar mass and star formation rate have been established (see for a review), with far-reaching implications for galaxy evolution. Interferometric observations obtained spatially resolved 21-cm fluxes to map the structure and kinematics of (e.g. WHISP, ; THINGS, ; BLUEDISK, ). Such resolved data has been used to study radial surface density profiles (), morphologies in the context of mergers (), and the Tully-Fisher relation () constructed from spatially resolved kinematics (). These observational insights into the cold gas content of galaxies provide a vital test for simulations of galaxy evolution. Modelling atomic and molecular hydrogen within galaxies numerically requires resolving the molecular clouds in the CNM to capture the formation of H_2 within a chemical network. For cosmological hydrodynamical simulations that follow the evolution of baryons and dark matter of large volumes (∼ 100^3 cMpc^3), modelling this chemical network on the required spatial scales is not feasible at present. Moreover, this chemical network depends on the local radiation field, which for current large-volume cosmological simulations run to z=0 is not explicitly followed. Hence, as a critical test for cosmological simulations, postprocessing of the simulation output is required to extract the atomic and molecular hydrogen content and compare to existing or future observational data. Such postprocessing studies (EAGLE: ; ; ; ; AURIGA: ; IllustrisTNG: ; ; ; ; ; ; ; ; ; SIMBA: ; FireBOX: ) mostly focus on redshift zero, where observational data is most available. Despite qualitative agreement, inferred discrepancies include the presence of spurious holes in the EAGLE simulation (), an excess of and too large disks in the AURIGA simulation () and an overabundance of at z=0 for TNG50 (). For molecular hydrogen, a robust comparison between simulations and observations is much more challenging (e.g. ; ). In contrast to the direct detectability of H i, molecular hydrogen is very difficult to observe directly due to the lack of a permanent dipole moment and high excitation temperatures. The molecular gas is usually observed through transitions of the second most abundant molecule, CO, introducing a rather uncertain factor to convert the CO line luminosity to the H_2 mass (see for a review). Key to H i/H_2 postprocessing studies is a scheme to partition the neutral hydrogen simulation output into its atomic and molecular phases. Partitioning schemes based on analytical arguments or high-resolution galaxy simulations both consider the formation of H_2 on dust grains and H_2 photodissociation from UV radiation in the Lyman-Werner band (11.2-13.6 eV, 912-1108Å). Hence, applying such hydrogen partitioning models to cosmological simulations requires estimating the local radiation field strength at 1000Å since the partitioning models use this value as a proxy for the radiation field strength in the Lyman-Werner band. Two different approximations to estimate the UV field exist in the literature: <cit.> introduced a scaling with the local star-formation rate of gas cells. Cosmological simulations typically use a Kennicutt-Schmidt type relation above a specific density threshold to model star formation, while lower-density cells are not star-forming (see Section <ref>). This leads to a strong discontinuity in the estimated UV field. To overcome this limitation, <cit.> introduced a method which spreads the UV flux from star-forming gas cells without attenuation (but with an escape fraction of 10%) throughout the galaxy. Since attenuation by dust is significant in the UV these radiation field estimates introduce large modelling uncertainties. In this study we improve upon previous methods by using the Monte Carlo dust radiative transfer code SKIRT (; ; ) to obtain a realistic estimate for the UV field. We describe our simulation methods and the simulation and observational datasets in Section <ref>. Using the highest-resolution installment of the IllustrisTNG simulation suite, TNG50-1, we explore how the different UV field estimates affect the H i/H_2 statistics of the simulated galaxy population in terms of mass functions and average radial profiles in Section <ref>. Exploiting the realistic UV fields and the high resolution of TNG50-1, we evaluate the realism of cold gas in cosmological simulations by generating spatially resolved maps. We compare the simulated maps to 21-cm data from the WHISP (Westerbork survey of Spiral and Irregular Galaxies) survey in terms of their non-parametric morphologies in Section <ref>. As complementary metric to non-parametric morphologies, we consider column density distribution functions (tracing how much exists per column density instead of how the columns are spatially distributed) in Section <ref>. We discuss and contextualize our results in Section <ref>, and conclude in Section <ref>. § METHODS §.§ The IllustrisTNG simulations The IllustrisTNG suite (; ; ; ; ) is a set of cosmological, magnetohydrodynamical simulations run using the moving-mesh code AREPO (). The simulation suite consists of three different volumes with box sizes of approximately 50, 100, and 300 comoving Mpc, each realized with three to four different resolutions. All of these simulations were run with the same physical model, which means that their subgrid parameters were not recalibrated (unlike the EAGLE simulation, ). For the cosmological parameters, the simulations use the 2015 results measured by the Planck satellite (). Since we explore synthetic maps which are sensitive to the spatial resolution we use the simulation with the highest resolution, TNG50-1, hereafter referred to as TNG50 (; ). We also consider the lower-resolution TNG50-2 and the larger-volume TNG100-1 (hereafter referred to as TNG100) runs to test the convergence of some of our results against resolution and box size. We summarize the box sizes and resolutions of the different IllustrisTNG simulations considered in this study in Table <ref>. In the following, we briefly describe the aspects of IllustrisTNG and its galaxy formation model (; ) that are most relevant to this study. TNG50 simulates a cube with box size of 51.7 comoving Mpc from z=127 to z=0. This volume is resolved with 2160^3 baryonic and dark matter particles, corresponding to a mean particle mass of 8.5×10^4 M_⊙ and 4.5×10^5 M_⊙, respectively. This mass resolution enables a spatial resolution of 70-140 pc for the densest star-forming regions of galaxies. Galaxies are identified using the SUBFIND algorithm (). The IllustrisTNG model incorporates gas radiative processes including metal-line cooling, evolution of stellar populations and chemical enrichment, and feedback from supernovae and black holes that drives galactic outflows and winds. Since molecular clouds cannot be resolved in the simulation, star formation is modelled stochastically for gas with n_H>0.106 cm^-3 according to the two-phase model of <cit.>. The ISM above this density threshold is emulated as cold, star-forming clouds with T=1000 K embedded in hot, ionized gas. This model prescribes an effective equation of state by calculating effective pressures and temperatures as averages over the cold clouds and the hot gas. As we are interested in the cold gas properties of local galaxies we select all subhalos in TNG50 at z=0 with a gas mass larger than 10^7 M_⊙, such that the galaxies are resolved by at least ≈100 gas cells. Furthermore, the SKIRT calculation requires star particles to estimate the radiation field. We choose a minimum total stellar mass of 10^7 M_⊙ such that galaxies are resolved by at least ≈100 stellar particles. These criteria lead to a sample of 12'431 galaxies, comprising both centrals and satellites. This provides a broad base sample which contains the vast majority of galaxies detectable in 21-cm observations. To facilitate the comparison between the different simulation runs we apply the same galaxy sample criteria (M_gas>10^7 M_⊙, M_⋆>10^7 M_⊙) to TNG50-2 and TNG100 as well (the sizes of the galaxy samples are given in Table <ref>). For all calculations in this study we consider the subhalos in the simulation individually, meaning that when modelling the H i/H_2 content of a galaxy we only select gas cells and star particles that are bound to this specific galaxy as identified by the SUBFIND algorithm. <cit.> find that for TNG100 at z=0, ≈98 % of the gas mass is bound to subhalos (for H_2 we expect an even higher fraction), hence we do not miss substantial amounts of H i/H_2 gas with our methodology. §.§ UV field estimates Radiation in the Lyman-Werner band (912-1108Å) can photodissociate molecular hydrogen. Consequently, the radiation field strength at 1000 Å (generally used as a proxy for the radiation of the entire Lyman-Werner band) is a key parameter for many partitioning schemes. We use three different methods to calculate the U_MW parameter, which is the UV field at 1000 Å normalized to the average Milky Way value of 3.43×10^-8 photons s^-1cm^-2Hz^-1 (). All methods rely on estimating the flux from stars and/or star-forming regions. For all gas cells we set a floor on U_MW of 0.00137 which corresponds to the homogeneous UV background (UVB) of <cit.> (in the updated 2011 version) at z=0 (the same UVB model is implemented in IllustrisTNG for the calculation of the ionization state of the gas). Dense gas could be self-shielded against the UVB which would lower the floor value of U_MW. However, such gas cells are typically near high star-formation areas so that the local UV radiation surpasses the UVB, rendering the actual floor value irrelevant. We consider three different schemes to estimate the UV field, summarized in Table <ref>. The simplest method to estimate the UV field follows <cit.>. The `Lagos' approximation is based on the insight that the largest fraction of the UV flux typically comes from very young stars. U_MW is calculated by scaling the local star formation rate surface density by the typical Milky Way value: U_MW=SFR·ρ/m·λ_J/ 10^-9 M_⊙ pc^-2, where ρ is the total gas mass density, m the gas cell mass, SFR its star-formation rate, and λ_J the Jeans length which approximates the size of a self-gravitating gas cloud (; ): λ_J=√(γ(γ-1)u/Gρ). In this equation, γ=5/3 is the ratio of specific heat capacities, u the internal energy per unit mass, and G the gravitational constant. Due to temporal and mass resolution limits in the simulation, it is impossible to resolve arbitrarily small star-formation rates for the gas cells. This leads to a minimum value for U_MW in the Lagos approach for the star-forming gas cells on the order of unity, while all gas cells with zero star-formation rate have U_MW=0.00137 (the UV background). Hence, this UV field approximation creates a substantial, unphysical dichotomy in the UV field distribution of the gas cells. <cit.> introduce an improved UV field calculation to overcome this limitation. In the `Diemer' approximation, star-forming gas cells are assigned a UV flux based on a Starburst99 () calculation of a continuously forming population of stars following a Kroupa IMF (). This UV flux is scaled to the star-formation rate of the gas cell, the flux of a 1 M_⊙yr^-1 cell at a distance of 1 kpc is 3.3×10^-6 photons s^-1cm^-2Hz^-1 corresponding to U_MW=96.2. It is assumed that a certain fraction of this radiation is absorbed by dust within the star-forming region and the remaining fraction propagates through a transparent medium. <cit.> calibrated this escape fraction to 10 % based on the SFR-UV relation in the solar neighbourhood. The propagation of the UV flux is calculated via a Fourier transform on a regular cubic grid. We refer the reader to Appendix A of <cit.> for details. While the `Diemer' method models in-situ absorption of UV flux generated from star-forming regions, it does not capture the significant absorption by diffuse dust in the ISM. Furthermore, older stellar populations can also contribute significantly to the galactic UV flux (; ; ). To generate realistic UV fields taking the complex galactic star-dust geometries into account, 3D radiative transfer modelling is required. Here we use the Monte Carlo dust radiative transfer code (; ). We run SKIRT in the oligochromatic (i.e. single wavelength) mode at 1000 Å including dust attenuation. A SKIRT simulation requires to define a box to set the domain for the radiative transfer calculation. We use a cube with side length 2· r_A[The same cube size is used to compute the Diemer UV field.], with the aperture radius r_A chosen such that a sphere with this radius contains at least 99.9 % of the neutral hydrogen gas mass bound to a specific subhalo (see Section <ref> for the calculation of the neutral hydrogen fraction). We choose the different components in a SKIRT simulation following <cit.>; <cit.> as follows: * Evolved stellar populations: We select star particles bound to the subhalo within the spherical aperture to model the UV emission. Star particles with ages above 10 Myr are treated as evolved stellar populations and modelled with a Bruzual-Charlot spectral energy distribution () with Chabrier initial mass function (). We have verified that the choice of an alternative stellar population library (BPASS, ; ) does not affect our results. * Star-forming regions: Bound star particles within the spherical aperture and with ages below 10 Myr are treated as star-forming regions and modelled with Mappings-III templates (). These templates contain the emission of the young stellar population and its subsequent attenuation by the surrounding dusty birth cloud. We refer to <cit.> for the determination of the various parameters required for the Mappings-III templates such as the compactness of the star-forming region or the ISM pressure. * Diffuse dust: To track the radiation field and its attenuation by diffuse dust we select all gas cells bound to the subhalo within a cube of side length 2· r_A. We allocate dust only in relatively dense and cold gas cells that are in line with the criterion from <cit.>. We assume that 40 % of the metallic mass in these gas cells exists as dust ( and references therein), which we model with the THEMIS dust mix (). The spatial discretization within SKIRT uses the imported Voronoi cell centers of the gas cells (). This allows SKIRT to reconstruct the same Voronoi mesh that was used in IllustrisTNG. Tracking the photon packages through a Voronoi mesh is computationally more expensive than octree grids, but the advantage of importing a grid instead of calculating one outweighs the additional cost here (see also ). We find converged results when running SKIRT with 10^5 photon packages, which runs two minutes for a large galaxy with M_⋆≈10^11 M_⊙ and eight seconds for a smaller galaxy with M_⋆≈10^8 M_⊙ on a 16-core machine. With this `SKIRT' method we can calculate realistic 3D UV fields for IllustrisTNG galaxies taking dust attenuation into account. The largest modelling uncertainty in this approach lies in the modelling of the star-forming regions, since they require estimates of multiple parameters that are not directly available from the output of the cosmological simulation (see also <cit.>; ). We also tested an alternative template library for star-forming regions (described in Kapoor et al. in prep.) which requires parameters that are more readily estimated from the TNG50 output. We found that using this alternative template library typically yields integrated masses that agree with the MAPPINGS-III templates within one percent. Hence, we consider the MAPPINGS-III templates to be sufficiently accurate for this work. For the UV modelling of TNG50-2/TNG100 galaxies, we note that these lower-resolution simulations can have very few star particles with our sample selection criteria (down to seven star particles for TNG100). For these barely resolved objects SKIRT cannot reproduce a meaningful UV field. Hence we use the Diemer UV field as replacement for the SKIRT UV field for galaxies with stellar masses below 10^8 M_⊙ for TNG50-2 and TNG100. §.§ Atomic and molecular hydrogen fractions The z=0 snapshot data of IllustrisTNG contain the total and neutral hydrogen fraction in gas cells. Ideally, the neutral hydrogen fraction in the simulation is determined using on-the-fly ionizing radiative transfer calculations. For IllustrisTNG, the hydrogen ionization is approximated using the ionizing background () and gas self-shielding with the fitting formula from <cit.>. Note that the neutral fractions for star-forming gas cells are for the effective gas temperature and must be recomputed. Hence, we use the intrinsic IllustrisTNG value for the neutral hydrogen fraction only for non star-forming cells. For star-forming cells, we follow <cit.>[See Eqns. A1-A6 in <cit.>.] and calculate the mass fraction of cold gas in the two-phase model of <cit.>. This cold gas at a fixed T=1000 K is assumed to be fully neutral while the hot phase is fully ionized. The breakdown of neutral hydrogen gas into its atomic and molecular constituents requires the use of partitioning schemes. A variety of partitioning schemes exist and have been used to postprocess cosmological simulations (see for a detailed comparison between the different partitioning schemes and how that influences the and H_2 properties of IllustrisTNG galaxies). Empirical partitioning schemes based on observations of local galaxies (; ) find that the ratio between molecular and atomic surface densities scales with the hydrostatic midplane pressure as a power law. As shown by <cit.>, estimates of the midplane pressure are only physically reasonable when computed from a projection (in face-on orientation). The hydrogen partitioning then happens on this 2D grid on a pixel-by-pixel basis, hence 3D information is lost and it is only possible to create maps in face-on orientation (it is not possible to rotate the galaxies back into a random orientation). Since we consider maps in random orientation (as the WHISP data to which we compare the simulated maps consist of randomly oriented maps), we omit the empirical partitioning schemes completely from this study. Analytical studies of <cit.>[The partitioning scheme of <cit.> is incorporated in the MUFASA and SIMBA simulations, splitting the hydrogen gas into its atomic and molecular phases at simulation runtime.] (KMT09), <cit.> (K13), and <cit.> (S14) consider idealized gas geometries with H_2 formation-dissociation balance to obtain the atomic and molecular hydrogen fractions. The main input parameters for these partitioning schemes are the surface density of neutral hydrogen, the dust-to-gas ratio, and (only for K13)[For the partitioning schemes of KMT09 and S14 the UV field strength drops out of the equations due to assumptions about the CNM-WNM pressure equilibrium.] the UV field strength at 1000 Å. Lastly, <cit.> (GK11) and <cit.> (GD14) use high-resolution (mass resolution up to 10^3 M_⊙) ISM simulations of isolated galaxies, coupled to a chemical network including H_2 formation on dust grains and photodissociation from Lyman-Werner photons, to partition the hydrogen based on the same three main parameters as the analytical partitioning scheme of K13. More specifically, the GK11 and GD14 approach consists of two steps: The first step consists of a cosmological simulation that is ran until z=4. This simulation extends over five virial radii of a system that would evolve to a typical M≈10^12 M_⊙ system. In a second step, the UV radiation field at 1000 Å and the dust-to-gas ratio are fixed to a grid of different values and the simulation is continued for 600 Myr. This setup allows for fresh gas infall into the galaxy, but cannot model effects from the larger environment such as mergers or filamentary accretion streams. In this study we use the simulation-calibrated GD14 recipe as our default partitioning scheme as it is an update of the GK11 recipe (adding self-shielding of H_2 from line overlap) and has the UV field as an input parameter. Furthermore, it is more directly applicable to cosmological simulations as it is a hydrodynamical simulation that models the various ISM phases and outputs the and H_2 fractions averaged over larger patches of the galaxy, while the analytical partitioning schemes consider simpler geometries such as a spherical cloud of gas. For our results we display the default partitioning scheme and treat the other UV-dependent partitioning schemes (GK11, K13) as uncertainties, shown as shaded areas. For some results we also consider the UV-independent KMT09 and S14 models which are plotted separately. The application of all of these partitioning schemes to IllustrisTNG is described in detail in Appendix C of <cit.>, except for KMT09 which we present in Appendix <ref>. The analytical and simulation-calibrated partitioning schemes used here have up to three main input parameters; the surface density of neutral hydrogen (Σ_H i+H_2), the dust-to-gas ratio relative to the average Milky Way value (D_MW), and the UV field strength at 1000 Å relative to the average Milky Way value (U_MW). Our estimates for U_MW are described in Section <ref>. As in previous studies (e.g. ), we assume that the dust-to-gas ratio scales with metallicity (as found observationally, e.g. ; ; ) such that D_MW=Z/Z_⊙ with Z the metal mass fraction and Z_⊙=0.0127 the solar metallicity used in IllustrisTNG (). To convert the mass density of neutral hydrogen (ρ_HI+H_2, calculated using the mass fraction of neutral hydrogen) into a surface density a certain length scale is needed. Two strategies to estimate surface densities from the simulation output have emerged: Volumetric modelling using the Jeans approximation and projected modelling using line-of-sight integration (). In the volumetric modelling approach, the surface density of neutral gas is obtained using the Jeans length λ_J: Σ_H i+H_2=ρ_H i+H_2·λ_J=ρ_H i+H_2·√(γ(γ-1)u/Gρ). In theory, the value of γ changes depending on the molecular fraction. Since the Jeans approximation by itself is already a crude approach we refrain from iterating the partitioning scheme to reach a self-consistent value of γ as in <cit.>. In the projected modelling approach the surface density is estimated by rotating the galaxies into face-on position and projecting the gas densities onto a 2D grid. The hydrogen partitioning is then performed on this 2D grid, making it impossible to rotate the galaxy back into random orientation as the 3D information is lost. As we consider maps in random orientation, we stick to the volumetric approach in this study and discuss projected modelling in the context of the different UV field estimates in Appendix <ref>. We remark that the Jeans approximation to calculate surface densities breaks down in some regimes. However, the Jeans approximation gives the same H i/H_2 results on average as the projected modelling approach (). The main output of our H i/H_2 modelling consists of U_MW parameters for the SKIRT, Diemer and Lagos method as well as 11 different molecular mass fractions f_mol[We use f_mol to denote the mass ratio of molecular to neutral hydrogen.] (three different UV fields for GK11, GD14 and K13, plus the UV-independent schemes of KMT09 and S14). We store these values for all gas cells within the spherical aperture[Technically, only the SKIRT UV field calculation requires an aperture. To simplify the comparison between the different models we employ the same spherical aperture to store U_MW and f_mol for all our models. This aperture is defined as the radius that contains 99.9 % of the neutral hydrogen gas mass bound to the subhalo.], for all 12'431 galaxies within our TNG50 base sample. We emphasize that our default H i/H_2-model consists of the GD14 partitioning scheme with the SKIRT UV field. §.§ Observational data: The WHISP survey We use observational data from the WHISP survey to compare to simulated maps from IllustrisTNG. The WHISP survey targets the 21-cm line of some 300 local galaxies across the Hubble sequence (), selected by their 21-cm flux density and their stellar disc size. This interferometric survey at the Westerbork Synthesis Radio Telescope (WSRT) provides spatially resolved 21-cm line profiles such that surface density maps and kinematics can be extracted. In this study we only focus on the surface density maps. In the WHISP survey, each object is observed in three different resolutions, with an average FWHM of the beam of 14, 30, and 60 arcsec, respectively. The sensitivity of the observations increased from a typical root-mean square error of 3 mJy/beam to 0.8 mJy/beam when the instrument was upgraded, leading to an average surface density sensitivity of 0.18 M_⊙pc^-2 for the medium-resolution maps (). The observational characteristics that are relevant for this work are summarized in Table <ref>. The field-of-view (FOV) of the WHISP observations is typically 0.25 square degrees, centered on the object of interest. A substantial amount of maps contain other galaxies, galaxy interactions or mergers. While a study of these objects would be interesting on its own (see ; ), we choose to avoid this additional complexity for the present study and analyze galaxies in IllustrisTNG one-by-one (i.e. processing each subhalo individually, see Section <ref>). Hence, we want to exclude observations that feature ongoing interactions or have other galaxies in their FOV. Recently, <cit.> compiled a catalogue of 228 galaxies observed within WHISP, containing only isolated objects (in their FOV, this does not refer to the environment in general) with good data quality. This sample is representative of the full WHISP sample in terms of galactic morphology. <cit.> also supplemented the WHISP data with infrared data from the WISE survey, enabling estimates of the stellar masses and star-formation rates. This is very advantageous for our purpose as we can select WHISP-like galaxies in IllustrisTNG based on their mass, stellar mass, and star-formation rate. Hence, we use the WHISP sample defined in <cit.> for this study. Additionally, we remove two objects (UGC4483 and UGC9128) as their stellar masses are below our IllustrisTNG selection threshold of 10^7 M_⊙, ending up with 226 galaxies in the observational galaxy sample. We retrieve the moment-zero maps for these 226 galaxies in all three available resolutions from the online Westerbork database[<http://wow.astron.nl/>]. The moment-zero maps, ∑_ν S_ν, are stored in mJy/beam (S_ν denotes the flux in a specific channel). We convert them to surface density maps assuming optically thin 21-cm emission (e.g. ): Σ_H i/M_⊙ pc^-2=8.85∑_ν S_ν/mJy/beamΔ v/km s^-1(Δα·Δδ/arcsec^2)^-1, where Δ v is the velocity channel width and Δα (Δδ) denote the FWHM of the major (minor) axes of the beam. We remark that for 47 WHISP galaxies the beam FWHM information was missing, for these galaxies we used the information from the online database for the highest-resolution observations and average values of 30×30 (60×60) arcsec^2 for the medium (low) resolution observations, respectively. To select WHISP-like galaxies from IllustrisTNG, we need some additional parameters to describe the WHISP galaxies like the mass, stellar mass, and star-formation rate. Calculating these quantities requires a distance estimate. While <cit.> added distance estimates from the NASA Extragalactic Database (NED[<https://ned.ipac.caltech.edu/>]) to their galaxy catalogue, a substantial fraction is derived from redshifts, which can be unreliable for the local WHISP galaxies with a median distance of ≈20 Mpc (). Hence, we opt to compile our own distances from NED, using the most recent redshift-independent distance estimate for each galaxy. If such estimates are not available we use the distance estimated from the redshift, corrected for the influence of the Virgo cluster, the Great Attractor, and the Shapley Supercluster. For each WHISP galaxy we compute the mass from the distance estimate and the surface density map (Eqn. <ref>). We use the medium-resolution maps of the WHISP galaxies to compute their masses, we note that the map resolution hardly affects the masses. For the stellar masses and star-formation rates, we exactly follow the steps outlined in section 3 of <cit.> using their compilation of WISE fluxes but our own distance estimate compilation. §.§ Creation of HI maps Here we describe how we generate maps for the simulation and observational galaxy samples. Part of the analysis of this paper compares different simulation postprocessing methods against each other. For these parts, we use a fairly basic algorithm to create `plain' maps for IllustrisTNG galaxies, described in Section <ref>. When we compare IllustrisTNG against observational data, we use a more advanced algorithm to generate `mock' maps (described in Section <ref>) with the aim of emulating the most important steps of the observations. Lastly, we also make some minor adjustments to the observational WHISP maps, described in Section <ref>. The main steps of the algorithm to create WHISP and mock IllustrisTNG maps are visualized in Figure <ref>. §.§.§ IllustrisTNG: Plain maps (i) Orientation: When creating plain maps in random orientation, we leave the simulated galaxy in its intrinsic orientation. For plain maps in face-on orientation, we rotate the galaxies such that its angular momentum vector aligns with the z-axis of the simulation box. Following <cit.> we calculate the galactic angular momentum using all gas cells within the 3D gas half-mass radius. If there are fewer than 50 gas cells within this radius we use all star particles within two stellar half-mass radii instead. (ii) Projection: We always project the gas cells along the z-axis of the simulation box. We employ a consistent projection algorithm for these tasks, specifically we use the method that was introduced by <cit.> which we briefly summarize here. Since the moving-mesh cells of IllustrisTNG have complex shapes we apply an adaptive Gaussian smoothing kernel to the gas cells. For the smoothing kernel, we use a width of σ=0.5(m/ρ)^1/3. For the calculation of the surface density, the algorithm simply sums up the masses of all smoothed gas cells in each pixel and divides by the pixel area. The map size is set to lie within the spherical aperture such that the map side length is √(2)· r_A. We resolve the plain maps by a fixed number of pixels (128×128). In some cases we are also interested in maps of other quantities than surface density. We employ the same projection algorithm in these cases. For quantities that are averaged (instead of summed) within each pixel (e.g. U_MW, D_MW) we weigh each gas cell by its total mass unless otherwise noted. §.§.§ IllustrisTNG: Mock maps A fair comparison between observations and cosmological simulations needs the two galaxy samples to be broadly comparable. Furthermore, it requires some sort of mock-observation routine which mimics the observational procedure. We do not attempt to fully emulate the interferometric observations of WHISP but include the three most important effects: Mimicking the angular pixel sizes of the observations, smoothing with the beam of the instrument, and adding noise. This is a similar strategy as is applied in the MARTINI tool to create mock datacubes from cosmological simulations (), although there is no noise incorporated in MARTINI. (i) Selection: For each WHISP galaxy, we select IllustrisTNG galaxies with similar stellar masses (±0.2 dex) and star-formation rates (±0.3 dex). The tolerances for these quantities correspond to the 2-σ error bars adopted by <cit.>. For ≈20 % of the WHISP galaxies in the sample, no SFR estimate is available because they were not detected in the WISE W3 band. For these galaxies we only used the stellar mass criterion. In a second step, we then select the five IllustrisTNG galaxies that are closest to the WHISP galaxy in terms of log_10M_HI out of the galaxies that are similar in terms of stellar mass and star-formation rate. The factor of five is chosen to increase the statistics of the simulated galaxy sample while keeping the mass distribution similar. As we find at least five similar IllustrisTNG galaxies (in terms of stellar mass and SFR) for each WHISP galaxy (typically dozens of IllustrisTNG galaxies conform to the stellar mass-SFR selection criteria), the `mock' TNG50 sample consists of 5·226=1130 galaxies. We remark that ≈20 % of this sample is just selected by M_⋆ and M_H i due to some WHISP galaxies being undetected in the WISE W3 band. We have verified that these TNG50 galaxies typically have low star-formation rates, as expected from their observational counterparts for which we only have an upper bound on the SFR. (ii) Orientation: We match the inclinations of the IllustrisTNG galaxies to their corresponding WHISP galaxies, by rotating IllustrisTNG galaxies such that the angle between the z-axis of the simulation box and the angular momentum vector matches the WHISP galaxy inclination. We retrieve the inclinations of the WHISP galaxies from the galaxy catalogue of <cit.>. (iii) Projection: We put the IllustrisTNG galaxies at the distances of their corresponding WHISP galaxies. The galaxies are then projected along the z-axis of the simulation onto a 2D grid. Each map is created in three different resolutions as in the observational data, the angular pixel sizes (5, 10, and 20 arcsec, respectively) are chosen to match the WHISP data. We use the same map size as in Section <ref> of √(2)· r_A. (iv) Smoothing: To emulate the effect of the WHISP beam we then convolve these maps with a circular 2D Gaussian kernel, using the average WHISP beam FWHM (14, 30, and 60 arcesonds, respectively) for the three different resolutions. (v) Noise: After beam convolution, we add Gaussian noise to the maps. We use the average 1-σ WHISP noise levels (see Table <ref>), which vary depending on the map resolution. (vi) Segmentation: For the calculation of the morphological parameters and the CDDF, the main object needs to be identified and separated from the background in the map. While we always project only one subhalo per IllustrisTNG map, it is possible that a subhalo consists of a neutral gas disk and additional disconnected neutral gas structures. For our analysis, we only want to consider the main object (in this case the neutral gas disk), and hence need to segment the map. We use the [We use the function with the average 1-σ noise (see Table <ref>) as threshold and five pixels as the minimum number of connected pixels.] package to find the object which belongs to the central pixel. If the central pixel is assigned to the background we instead use the largest object in the map. To make the segmentation map which cuts out the galaxy of interest more regular we smooth it with a uniform boxcar filter measuring five pixels in each dimensions. By visual inspection, we find that this algorithm reliably identifies the galaxy in the map. §.§.§ WHISP: Observational maps The extraction of maps for the WHISP galaxies is described in Section <ref>. We perform two additional steps to generate the final WHISP maps that we analyze in this study. (i) Noise: The WHISP maps are inherently 3-σ-clipped, meaning that they do not contain any background (all background pixels are set to NaN values). This is problematic for the calculation of the morphological parameters, as some of the parameters need a background in the image in order to be accurately computed. Hence, we replace all NaN pixels from the WHISP maps with Gaussian noise, using the average 1-σ WHISP noise levels depending on map resolution. (ii) Segmentation: We create segmentation maps for the WHISP galaxies using exactly the same algorithm as described in Section <ref>. Again, we find by visual inspection that this algorithm reliably identifies the main galaxy in the FOV of the observation. § UV FIELD COMPARISON We start our analysis by assessing the differences in the SKIRT, Diemer, and Lagos UV field estimates, and explore how this affects the and H_2 properties of the simulated galaxy population in TNG50. To qualitatively compare the different UV field estimates we show 2D projections of the U_MW parameter (calculated using the plain map algorithm in random orientation, see Section <ref>) for an example TNG50 galaxy in the top panels of Figure <ref>. This galaxy has a stellar mass of 9.3×10^10 M_⊙, a high gas-to-stellar mass ratio of 2.1, and a star-formation rate of 9.0 M_⊙ yr^-1. The 3D stellar and gas half-mass radii are shown as solid and dashed circles, respectively. The angular momentum vector of the galaxy is almost parallel to the z-axis of the simulation, hence the random projection in Figure <ref> displays an almost face-on orientation of the galaxy. For the SKIRT UV field (top left panel), some clumpy structure in the spiral arms due to the star-forming regions (young star particles with ages less than 10 Myr) that strongly emit UV is visible. In the outskirts (outside the gas half-mass radius), the attenuation and scattering by dust grains imprints some structure on the UV field. We remark that in the SKIRT UV field, radiation from the evolved stellar population is also modelled, unlike in the other UV field models. For the Diemer UV field (top center panel), radiation is emitted from all star-forming gas cells. These cells are more common but less luminous than the star-forming regions in the SKIRT recipe (star particles with ages below 30 Myr), leading to a less clumpy UV field in the spiral arms for the Diemer UV field. Since the radiation is transported without considering absorption/scattering the radiation field is very smooth in the galactic outskirts as well. We remark that there is a minor discontinuity in the Diemer UV field at the edges of a square enclosing the gas half-mass radius. This is due to numerical resolution issues in the calculation of the Diemer UV field as the algorithm splits the galaxy into a high-resolution region within the gas half-mass radius and a low-resolution region outside of it, allowing a significant speedup of the algorithm. We do not expect that this affects any of our results. Lastly, the Lagos UV field (top right panel) scales the UV flux to the star-formation rate of gas cells. Since there is no transport of radiation in this approach there is a strong bimodality between star-forming and quiescent gas cells. The gas cells without star formation only receive the homogeneous background UV field of U_MW=0.00137 (), note that this floor is applied to all three UV field estimates. We also show maps of surface densities and molecular fractions calculated with the three different UV field estimates in Figure <ref> (using our default partitioning scheme of GD14). Despite significant differences in the UV field, the Σ_H i and f_mol maps for this example galaxy appear almost identical for the SKIRT and Diemer UV field estimates. We remark that the neutral gas hole in the center of this galaxy is due to feedback from the supermassive black hole (see Section <ref>). In Section <ref> and Section <ref> we examine whether the different UV field estimates lead to significant statistical discrepancies for ensemble quantities of the entire galaxy population (within our base sample of 12'431 galaxies) in TNG50. §.§ Radial profiles To assess the impact of the UV field on the H i-H_2-transition on the galaxy population in a more statistical fashion, we calculate 1D projected radial profiles of UV fields and molecular gas fractions for the TNG50 base sample in this section. Following <cit.>, we only consider galaxies with M_H_2>10^8 M_⊙ (we use our default model here to calculate the mass for the galaxy selection) in this section as the f_mol profiles of galaxies with less molecular gas are noisy. The exact value of this cutoff does not affect our conclusions. We show median U_MW radial profiles in Figure <ref>. The interquartile ranges are shown as hatched areas. For the calculation of the radial profiles we first rotated each galaxy into a face-on orientation (see step (i) of Section <ref>) and considered 2D radial profiles (i.e. using the projected distance to the galaxy center, not the 3D distance). For each radial bin, we then compute the median of the U_MW parameter of all gas cells within this bin (using the mean instead of the median would significantly bias U_MW for the Lagos UV field, as star-forming gas cells receive UV fluxes that are 4-6 orders of magnitude larger than non star-forming cells). If there are no gas cells for a galaxy within a radial bin U_MW is set to the uniform UV background. To compute a single average profile for the entire galaxy sample we stacked the galaxies by their gas half-mass radii and computed the median profile. At the center the Lagos UV field exceeds the other estimates by more than an order of magnitude due to the large unattenuated UV field in star-forming gas. As the Lagos UV field is scaled to the gas cell star-formation rate and most of the gas cells at larger radii do not form stars, the Lagos UV field quickly decreases to the uniform UV background at U_MW=0.00137. The Diemer UV field has a steeper UV profile than the SKIRT field, which is due to the centrally concentrated star-forming gas cells being the only UV sources in the Diemer UV field. The SKIRT UV field is more extended as we treat all star particles as UV sources (see Table <ref>). We compare the different calculations for the H_2 mass fractions by plotting the median projected radial profiles of f_mol in Figure <ref>. For each galaxy, we calculate its profile by computing the mean molecular fraction in each radial bin, similar to the calculation of the UV field profiles. If there are no gas cells for a galaxy within a radial bin, f_mol is set to zero. We find that if we instead ignore these radial bins for the calculation of the median profiles f_mol increases nonphysically for large radii. As for the UV field profiles, we stack the profiles of all galaxies by their half-mass radius and compute the median f_mol profile of the galaxy population. Applying the mean instead of the median to the stacked profiles (as is done in ) gives higher f_mol at larger radii and reduces the scatter between the different models substantially. As the mean profiles are dominated by few galaxies that have high molecular fractions at larger radii and for consistency with Figure <ref> we opt for median profiles in Figure <ref>. For the three UV models (SKIRT, Diemer, Lagos) we show the default hydrogen partitioning scheme (GD14) as solid lines and the spread between the UV-dependent schemes as shaded areas. We also show the UV-independent partitioning schemes (KMT09 and S14). The differences in the molecular fractions relative to the default model (GD14 with the SKIRT UV field) are shown in the bottom panels. The significantly lower Lagos UV field (except in the center) manifests in higher molecular fractions. For the SKIRT and Diemer f_mol profiles, the spread due to the different partitioning schemes is much larger than the difference due to changing the UV field. The discrepancy in the f_mol profiles comparing the different UV fields reaches its maximum at ≈0.4 R_gas^1/2. This is not particularly surprising because f_mol is 20-40 % at this radius, so this is broadly the region at which the H i-H_2-transition takes place. The H i-H_2-transition occurs typically over a very narrow range in densities, and other parameters like U_MW can slightly shift this density range (see e.g. Figure 1 of ). This means that the gas density in this region is in a range where the partitioning schemes are particularly susceptible to changes in the UV field. Lastly, we remark that the default GD14 partitioning scheme minimizes the differences in the molecular fractions upon variation of the UV field. Hence, we expect more significant discrepancies in the H i/H_2 properties of galaxies between different UV field estimates when using other partitioning schemes (GK11 or K13). §.§ HI and H2 mass functions We consider H i/H_2 mass functions for our base sample in the upper panels of Figure <ref>. As in Figure <ref> we display the UV-dependent partitioning schemes by showing the default partitioning scheme and spread between schemes for each UV field estimate individually (solid lines and shaded areas). Differences in the molecular fraction relative to the default model are shown in the bottom panel. The mass function (HIMF) is very robust against variations of the partitioning schemes and UV field estimates, except for the high-mass end. masses calculated using the Lagos UV field are lower compared to the SKIRT and Diemer UV fields, with discrepancies in the HIMF up to 20 %. For the H_2 mass function (H2MF), discrepancies are similar but begin at much lower masses. The differences between the partitioning schemes are generally larger for H_2 than for . For both mass functions, the default GD14 partitioning scheme is the model with the least tension between the different UV models (as seen in Figure <ref>). Still, the shaded areas between the different UV models do not overlap for high or intermediate H_2 masses, indicating that the discrepancy between different UV models is robust to the choice of the partitioning scheme. For the UV-independent partitioning schemes, we remark that the S14 model is in good agreement while the KMT09 model significantly exceeds other HIMF model predctions at the high-mass end (and vice versa for the H2MF). We note that our mass functions are not directly comparable to the results of <cit.> due to different sample definitions: <cit.> consider all galaxies above a certain gas or stellar mass threshold, while we need a minimal stellar mass for all galaxies for SKIRT to work properly. This means we miss galaxies with low stellar mass (M_⋆<10^7 M_⊙) and high gas fractions when calculating the gas mass functions. Since the H i-to-stellar mass fraction in TNG50 at a stellar mass of ∼10^7 M_⊙ reaches values up to ∼10 (, figures including the TNG50 simulation can be found online[<benediktdiemer.com/data>]), the calculated mass function is complete only above an mass of ∼10^8 M_⊙. Indeed, we find that our HIMF agrees with the one from <cit.> above masses of 2×10^8 M_⊙. For molecular hydrogen the stellar mass cutoff is irrelevant: Because the H_2-to-stellar mass fractions are always smaller than unity, a stellar mass cutoff at M_⋆=10^7 M_⊙ does not exclude any galaxies with M_H_2>10^7 M_⊙. Hence, our H_2 mass function calculated with the Diemer UV field matches the result from <cit.>. § NON-PARAMETRIC MORPHOLOGIES OF HI MAPS To test our model and to evaluate the realism of the spatial distribution of gas within TNG50 galaxies, we compare simulated surface density maps to 21-cm maps from the observational WHISP dataset. We use the full mock map algorithm (Section <ref>) to generate a WHISP-like sample of 1130 TNG50 maps, to be compared to 226 WHISP maps. To quantitatively compare the maps of the two samples we consider non-parametric morphologies (concentration, asymmetry, smoothness, Gini and M_20 statistics). We calculate these statistics for the TNG50 and WHISP maps consistently with the tool (). §.§ Statmorph is a tool for parametric and non-parametric analyses of astronomical images, and has already been used to study mock images in optical bands for IllustrisTNG (; ). Besides the image, the surface density map in our case, requires a segmentation map and an estimate of the noise as input. A segmentation map separates all objects in the image from the background. As we process galaxies individually in IllustrisTNG and use a galaxy sample from WHISP that only contains isolated objects, we have only one object per image to analyze. The segmentation map is calculated as described in Section <ref>. also requires an estimate of the noise on the maps. We use the simpler gain option of which calculates Poisson noise, the non-parametric morphologies are independent of the actual gain value. For ≈5% of the galaxies[This fraction is very similar for the WHISP and the TNG50 galaxy samples individually.], does not finish properly, mostly due to irregular segmentation maps. Such galaxies are omitted from our analysis. We find that the results at medium resolution are most reliable, as they pick up more features than the low-resolution maps and successfully runs more often than for the high resolution. Hence, we consider the medium-resolution maps of 30 arcsec as default in the following analysis. §.§ Morphological results We show histograms of the non-parametric morphologies for the TNG50 and WHISP maps in varying angular resolution in Figure <ref>. To avoid overcrowded figures we only show our default partitioning scheme (GD14 plus SKIRT UV field) in this section, using other partitioning schemes or UV fields yields comparable results. We remark that the number of galaxies plotted in the histograms has a small dependency on the angular resolution, since is not always able to measure the morphological parameters. We have verified that this does not affect the distributions in Figure <ref>, i.e. the galaxies that are lost in the high-resolution TNG50 maps do not differ significantly in terms of the morphological statistics from the overall galaxy sample. Comparing the WHISP and TNG50 medium-resolution (30 arcsec) maps, we find that the asymmetry statistic matches well, the TNG50 smoothness statistic is slightly lower (i.e. the TNG50 maps are slightly smoother than the WHISP maps), and the distribution of the Gini statistic is a bit more skewed in the WHISP sample. For the TNG50 maps, the scaling with angular resolution seems to be weaker for the smoothness statistic and reversed for the asymmetry statistic. As the maps at resolutions other than 30" are less reliable we do not investigate this discrepancy further. More significant differences arise in the concentration and M_20 statistics. Denoting the difference of the TNG50 and WHISP medians of a morphological statistic k as Δk̃, we have more concentrated WHISP maps (ΔC̃≈-0.32) and find ΔM̃_20≈0.30. We further explore the deviations in the concentration and M_20 statistics in Figure <ref> by plotting their correlation for the medium-resolution maps. This is also the strongest cross-correlation between non-parametric morphologies found by <cit.> and can be used to identify interacting galaxies as outliers in the top right corner. Due to the definitions of C (increasing with brighter pixels in the galaxy center) and M_20 (increasing with brighter pixels in the outskirts), these statistics are usually anticorrelated. Such an anticorrelation is found in the WHISP data (see Figure <ref> or ), in synthetic optical images of IllustrisTNG galaxies (; ), and in images across the UV-submm wavelength range for both observed and simulated galaxies (; ; ). We find that the two samples broadly follow the same anticorrelation, with the TNG50 galaxies shifted to the bottom right corner. Visual inspection of the maps with low concentration and high M_20 reveals that those are exclusively face-on galaxies, while the TNG50 outliers at intermediate concentration and high M_20 are mostly edge-on. These angular momentum trends are less pronounced for the WHISP galaxies. We do not find any other obvious characteristic such as mass that distinguishes the TNG50 outliers from the WHISP galaxies. §.§ Central HI holes in TNG50 By visual inspection of the TNG50 and WHISP maps, we find that many TNG50 galaxies feature large central holes, while such holes are much less prevalent in the WHISP data[On the other hand, by visual inspection of the 34 maps from the THINGS survey () it seems that central holes are prevalent in observed galaxies. This difference compared to WHISP could be related to different spatial resolutions (the THINGS beam has a FWHM of 6 "). We point out that if central holes are washed out in WHISP due to the lower resolution, then the same should happen to the TNG50 maps as they are mock-observed at the WHISP resolution.] (see Figure <ref> for a typical example). Central holes could simultaneously explain the offsets in the concentration and M_20 statistics, with face-on galaxies deviating the most from observations as the impact of the central hole is maximized. In the following we discuss some possible sources of the apparently unrealistic prevalence of large central holes in TNG50. An obvious mechanism to create large central holes is to ionize the neutral hydrogen or expel it from the galaxy center, for instance by feedback from active galactic nuclei (AGN). Indeed, <cit.> showed that related to AGN feedback, star-formation rates of TNG50 galaxies at z∼1 are centrally suppressed, in agreement with observational data. In IllustrisTNG, AGN feedback occurs along a thermal and a kinetic mode, depending on the black hole accretion rate. For galaxies with stellar masses roughly below 10^10.5 M_⊙, the AGN feedback is typically in the thermal mode with a continuous energy injection. For more massive galaxies, the kinetic mode kicks in and energy is injected in a pulsed, directed fashion into the surroundings (see for more details). Furthermore, AGN radiation can ionize a substantial fraction of the hydrogen gas (see figure A6 of ). Recently, <cit.> found that the kinetic AGN feedback in IllustrisTNG is very effective at redistributing neutral gas from the central to the outer galaxy regions compared to the SIMBA simulation, hinting that AGN feedback could indeed create central holes. Upon examination of the neutral gas reservoirs of central and satellite galaxies in TNG100, <cit.> and <cit.> also speculate that some tension with observational data could arise due to AGN feedback ejecting/depleting neutral gas from the center. As a first ad-hoc experiment to check if AGN feedback is responsible for the central holes in TNG50, we exclude all galaxies that experienced AGN feedback in the kinetic mode (≈25 % of the TNG50 sample of 1130 galaxies). This slightly reduces the tension in the morphological statistics to ΔC̃≈-0.23 and ΔM̃_20≈0.25. If we instead discard all TNG50 galaxies with an above-average energy injection in the thermal mode, the tension reduces to ΔC̃≈-0.18 and ΔM̃_20≈0.18. While it seems plausible that AGN feedback is at least partially responsible for the tension in the concentration statistic, we caution that the AGN feedback properties of the TNG50 galaxies are probably cross-correlated with other galaxy properties (e.g. stellar mass, gas-to-stellar mass ratio) which could also influence the morphological statistics. To unequivocally pinpoint the impact of AGN feedback, different cosmological simulation runs varying the feedback parameters need to be considered, which is beyond the scope of this paper. The impact of AGN feedback is also somewhat depending on the numerical mass resolution of the cosmological simulation. For the IllustrisTNG suite, the feedback parameters are calibrated for TNG100, and are then kept the same for all IllustrisTNG runs (). Increasing the resolution of the simulation also increases the black hole accretion rate, especially in the thermal mode (see figure B1 of ). We test the impact of this numerical effect by repeating the construction of maps and morphological analysis for TNG100, which has a mass resolution twenty times coarser than TNG50 (see Table <ref>). We find that the tension is reduced to ΔC̃≈-0.16 and ΔM̃_20≈0.22 when using TNG100. Lastly, it is also possible that there is neutral hydrogen in galaxy centers but the gas is predominantly molecular. To assess this, we check if the choice of the model affects the central holes. We find almost no changes in the morphological statistics upon variation of the UV field model or partitioning scheme, except for the UV-independent partitioning scheme S14 (which increases the tension in the C and M_20 statistics). However, when using `local' hydrogen partitioning schemes based on volume instead of surface densities (described in more detail in Section <ref>), we find that the tension reduces to ΔC̃≈-0.19 and ΔM̃_20≈0.20 (using the local GD14 scheme with the SKIRT UV field, this choice hardly affects the result). This could be related to the Jeans approximation breaking down for the dense galactic centers, as discussed in more detail in Section <ref>. We point out that the unusual central holes could also be due to the AGN feedback in IllustrisTNG being too weak. If AGN activity would evacuate the cold gas from the entire galaxy instead of just the central region, the concentration statistics of such galaxies would increase and better match the WHISP data. It is also possible that such a reduction of cold gas in the galaxy leads to the object not being selected in the TNG50 sample, as it would be unobservable with WHISP. Indeed, results from <cit.> indicate that for TNG100, the AGN feedback quenches the star formation but only redistributes the cold gas instead of decreasing the galactic cold gas reservoir as is observed (). On the other hand, we caution that <cit.> did not mimic the observational steps to mock-observe the simulated galaxies. <cit.> used the same observational data and hydrogen postprocessing method as <cit.> but a careful mock-observation routine, finding excellent agreement for the cold gas reservoirs of TNG100 and observations. The prevalence of central holes is reminiscent of the finding of <cit.>, who discover unrealistic holes in the disks of galaxies in the EAGLE simulation. <cit.> attribute this to the implementation of the stellar feedback in the physical model of EAGLE. Along similar lines, we partly attribute the large central holes in TNG50 to the AGN feedback implementation. The effect is stronger in the higher-resolution simulation TNG50 compared to TNG100 due to the IllustrisTNG calibration scheme. <cit.> found results along similar lines as the surface density profiles of IllustrisTNG galaxies exhibit central deficits compared to observational data from the Bluedisk survey (), with larger mismatches for higher resolution simulations. However, we note that the usage of local partitioning schemes reduces the prevalence of central holes. We conclude that these different effects (and potentially other systematics that we have not uncovered here) conspire to lead to the central holes and eventually the mismatch in the C and M_20 statistics. § HI COLUMN DENSITY DISTRIBUTION FUNCTION The column density distribution function is a widely used metric to quantify the distribution of atomic hydrogen in the Universe (e.g. ; ; ; ). In this study, we use the CDDF as the second metric (apart from the morphologies) to compare TNG50 to observational data. The CDDF is insensitive to the spatial distribution of the H i, it only probes how much atomic hydrogen per column density interval exists in the galaxy population. Hence, this CDDF comparison is somewhat complementary to the analysis of morphologies. We describe how we extract the CDDF from TNG50 (Section <ref>) and WHISP (Section <ref>), and compare them in Section <ref>. We only discuss the default model (GD14 + SKIRT UV field) for TNG50 in this section and compare to other partitioning schemes in Section <ref>. §.§ CDDF measurement from TNG50 As we intend to compare the CDDF from TNG50 to observational data, we use the full mock map algorithm (Section <ref>) to generate 1130 synthetic maps (in the three different angular resolutions of the WHISP survey) for TNG50 galaxies. We then follow the observational procedure to compute the CDDF, which consists of summing the area covered by columns within a specific column density interval over all galaxies (e.g. ; ): CDDF_H i=c/H_0∑_jϕ(M_H i^j) w(M_H i^j) A(N_H i)^j/N_H i ln10 Δ(log_10N_H i), where c denotes the speed of light, N_H i is the column density, and Δ(log_10N_H i) is the constant logarithmic column density bin spacing. This equation is applicable when the sample from which the CDDF is derived (in our case the 1130 mock-selected TNG50 galaxies) is drawn from a broader underlying galaxy population (the base TNG50 sample of 12431 galaxies). The CDDF of the Universe can then be computed by scaling the CDDF from the smaller sample using some galaxy property (e.g. stellar mass or mass) in which the underlying broader sample is assumed to be complete[We remark that CDDFs calculated in this fashion are not necessarily representative of the broader galaxy population since only the M_H i distributions are matched. This does not affect our comparison between TNG50 and WHISP, as we follow the same methodology for the simulated and observed CDDF. Hence, the mock TNG50/WHISP samples are equally (un-)representative of their broader galaxy populations.]. For the CDDF, the mass is a natural choice for this galaxy property (also adopted by e.g. ). Hence, the sum in Eqn. <ref> runs over bins of mass. The area function A(N_H i)^j denotes the area covered by columns within the (logarithmic) column density interval summing over all galaxies (in the smaller sample) within the j-th mass bin. The weight factor w(M_HI^j) corresponds to one over the number of galaxies in the smaller sample within the j-th mass bin, thereby normalizing the area function. To scale the CDDF to the entire galaxy population, the normalized area function (w(M_H i)A(N_H i)) measured from the smaller sample is then multiplied with the abundance of objects in the broader sample in the j-th mass bin per volume, ϕ(M_HI^j). Hence, ϕ(M_H i) corresponds to the mass function (derived from the broader sample), but without dividing by the (logarithmic) mass bin width. For our calculation of the CDDF in TNG50, we compute the area function and the masses for the weight factor from the 1130 mock maps[Instead of computing the masses from the mock maps we could also use the `intrinsic' masses of the galaxies, simply summing the masses of the gas cells. However, this introduces inconsistencies with the observational procedure and the normalization of the CDDF.]. ϕ(M_H i) is calculated from the base TNG50 sample of 12431 galaxies. This approach is equivalent to observational determinations of the CDDF, where a smaller galaxy sample is observed with high spatial resolution to derive the area function. The area function is then scaled using the mass function derived from blind galaxy surveys which measure the broader galaxy population. For cosmological simulations, we have the option to generate spatially resolved maps directly for the broad galaxy population, circumventing the scaling of the CDDF from a smaller galaxy sample. As we want to emulate the observational procedure with realistic mock maps (which requires a mock sample selection and putting the galaxies at the expected distances), we find the approach using Eqn. <ref> more appropriate for this study. We test the simpler approach of measuring the CDDF from plain maps of the broader galaxy population in Section <ref>. §.§ CDDF measurement from WHISP The CDDF at redshift zero can be measured with interferometric 21-cm surveys. A precise determination of the CDDF is presented in <cit.> based on WHISP data (see Section <ref> for details about the WHISP survey). Although we could measure the CDDF from the WHISP maps ourselves, we refrain from doing so as the <cit.> measurement constitutes the gold standard for the observational CDDF at redshift zero and their result is used in many studies (e.g. ; ; ; ). The CDDF measurement from <cit.> is based on the full WHISP survey (in its three available resolutions). Since the smaller WHISP sample of <cit.> that we use throughout this study is representative of the full WHISP sample, we do not expect that the usage of different WHISP samples affects our results. <cit.> have scaled their measurement of the WHISP CDDF to the broader galaxy population (see Eqn. <ref>) using the mass function from the blind HIPASS survey (). We add a cautionary note on H_0 correction factors. While the IllustrisTNG simulations were run with h=0.6774, the observational result from <cit.> was obtained with h=0.75. It would be desirable to correct the observational result to the more recent value of H_0. For the CDDF, however, this is an involved calculation due to the scaling factor ϕ (which is the HIMF modulo the mass bin width) in Eqn. <ref>. This HIMF is parametrized as a Schechter function with parameters that themselves depend on the value of H_0. Hence, H_0 propagates in a non-polynomial way into the CDDF result from <cit.>. We have run a test calculation of the CDDF based on the WHISP maps and the HIPASS mass function from <cit.> to emulate the <cit.> calculation (the actual CDDF calculation is more complex as split the WHISP galaxies by morphological type and use type-specific mass functions). We find that updating h from 0.75 to 0.6774 decreases the CDDF by ≈10-20 %, except for the highest columns (N_H i>1.5×10^22 cm^-2) where the CDDF increases by ≈10 %. Since the H_0 correction factors that we find are much smaller than the difference between the original <cit.> result and our test calculation, we keep using the <cit.> result but multiply it with 0.85 to correct it to our value of h=0.6774. Furthermore, the various biases in the HIMF from <cit.> (e.g. selection effects, self-absorption) are not taken into account by <cit.> when calculating the WHISP CDDF. <cit.> find that the cumulative effect of these biases reduces the total abundance in the Universe by 11.6 %. As the CDDF propagates linearly into the abundance (see Eqn. <ref>), we additionally decrease the WHISP CDDFs uniformly by 11.6 %. For studies of the CDDF, which extends over nine orders of magnitude for N_H i∼10^19-10^22 cm^-2, these correction factors are completely negligible. For our study, we also consider the abundance (which effectively corresponds to the normalization of the CDDF), which scales linearly with any correction factors to the CDDF. Hence, the H_0 and HIMF correction factors are not negligible for our purpose. §.§ CDDF results Since the CDDF measures the area covered by columns per column density interval, the total abundance in the Universe can be derived from the CDDF. Denoting the abundance Ω_H i as a fraction of the critical density of the Universe, Ω_H i=ρ_H i/ρ_c with ρ_H i the mass density in the Universe and ρ_c=1.274×10^11 M_⊙Mpc^-3, we can write the fractional contribution to Ω_H i per (logarithmic) column density interval as follows: ΔΩ_H i/Δ(log_10N_H i) =m_H i/ρ_cVN_H i∑_jϕ(M_H i^j) w(M_H i^j) A(N_H i)^j/Δ(log_10N_H i) = m_H iH_0ln(10)/ρ_cc(N_H i)^2·CDDF_H i(N_H i), where m_H i is the mass of a hydrogen atom, V is the volume of the survey or the simulation box, and the sum in the first line of Eqn. <ref> runs over bins of mass. We show our main CDDF result, Figure <ref>, as the conventional CDDF (left panel) and as fractional contribution to Ω_H i (right panel). Observational results from <cit.> (slightly modified as described in Section <ref>) are shown in grey to black markers for the three available WHISP resolutions. The TNG50 CDDFs for the different angular resolutions are displayed as colored lines. For our comparison it is more convenient to analyze ΔΩ_H i/Δ(log_10N_H i) instead of the conventional CDDF for two reasons: ΔΩ_H i/Δ(log_10N_H i) spans much fewer orders of magnitude which enables a more precise visual comparison for the different column density distributions. Furthermore, it also relates directly to the abundance as Ω_H i effectively corresponds to the integral of ΔΩ_H i/Δ(log_10N_H i) over the column density. To visualize this integral we also show ΔΩ_H i/Δ(log_10N_H i) with a linear y-axis in the inset in Figure <ref>, where the areas under the curves directly correspond to the abundance of TNG50 or the measured abundance in the Universe. We caution that the comparison of the column density distributions becomes unreliable below the sensitivity limits, indicated by the arrows at the top of Figure <ref>. The segmentation for our TNG50 maps is more aggressive (the main object cutout is smaller) than the 3-σ-clipped WHISP maps used by <cit.> such that the TNG50 CDDFs decrease quickly for column densities below the sensitivity limits (note that the median sensitivity limits of WHISP and the TNG50 mock maps are equal due to our mock map generation algorithm, see Section <ref>). Intermediate column densities (N_H i≈10^20-10^21 cm^-2) are significantly more abundant in TNG50 than in WHISP. This leads to a substantial mismatch in the abundance comparing TNG50 and observational data. We point out that the abundance (equivalent to the normalization of the CDDF) as visualized in the inset of Figure <ref> is actually completely independent of the WHISP data or the TNG50 mock maps due to the scaling factor ϕ(M_H i) in Eqn. <ref>. Instead, the normalizations of the CDDFs are given by Ω_H i as measured by the blind HIPASS survey (for the WHISP CDDFs) or the total amount of in the TNG50 base galaxy sample of 12431 galaxies (for the TNG50 CDDFs), respectively. Hence, the tension in the CDDF at these intermediate column densities points towards an excess of in TNG50 or an underestimation of Ω_H i in blind surveys, which we discuss in more detail in Section <ref>. At the highest column densities (N_H i>10^21 cm^-2), the effect of beam smearing becomes substantial both for WHISP and TNG50, which manifests in smearing out the high-density peaks. In this column density range, TNG50 significantly underpredicts the CDDF compared to WHISP. We find that there is neutral hydrogen in TNG50 in these column densities (see also ), but the hydrogen partitioning that we perform is very effective at converting into H_2 such that the abundance of high column densities in the simulation is lower than observed. We discuss this tension and a potential resolution in more detail in Section <ref>. Lastly, we point out that our result contrasts with other studies of the CDDF (; ; ) who find a good agreement between WHISP and cosmological simulations, especially at the high-column end. We attribute this primarily to the beam smoothing of the mock maps, which is not performed in these studies (see Section <ref> for more details). § DISCUSSION §.§ Normalization of the CDDF: HI abundance in IllustrisTNG As shown in the inset of Figure <ref>, the CDDF normalization (equivalent to Ω_H i) in TNG50 significantly exceeds the observational measurement. The observational value of Ω_H i at redshift zero is most reliably determined with blind surveys (see figure 14 of for a data compilation). For HIPASS, <cit.> find Ω_H i=(3.9±0.4^stat±0.4^sys)×10^-4, consistent with the result for ALFALFA of Ω_H i=(4.0±0.1^stat±0.6^sys)×10^-4 (). Both of these measurements are corrected for various biases such as selection bias or self-absorption. Furthermore, we corrected these abundances to our adopted value of the Hubble constant, h=0.6774. As the WHISP CDDF is scaled with the HIPASS HIMF, we expect the exactly same value of Ω_H i=3.9×10^-4 when calculating the abundance from the WHISP CDDF. With the raw CDDF from <cit.> we obtain Ω_H i=5.1×10^-4, but with the Hubble and HIMF biases correction factors (see Section <ref>) the abundance is reduced to the expected value of Ω_H i=3.9×10^-4 (with very little impact on the WHISP resolution). Hence, we consider our correction factors somewhat realistic, at least at the level of total abundance. Studies investigating the abundance in IllustrisTNG (; , ) found higher values of Ω_H i compared to blind surveys. The magnitude of this discrepancy depends on the simulation version of the IllustrisTNG suite and the postprocessing routines. Summing all in galaxies in the TNG100 simulation, <cit.> report Ω_H i=(5.4±0.5)×10^-4, in tension with the observational results. For TNG50, <cit.> report Ω_H i=(7.2±1.1)×10^-4, in significant excess of the data. The error bars on these values correspond to the scatter due to the different partitioning schemes considered by <cit.>. Using our default model (GD14 plus SKIRT UV field) for our base sample of 12431 galaxies in TNG50, we obtain Ω_H i= 6.0×10^-4. This is at the lower edge of the value reported by <cit.>, since the GD14 scheme typically predicts the highest molecular fractions (together with the S14 scheme) and because we miss a few gas-rich galaxies with very low stellar mass due to our galaxy sample selection. §.§.§ Observational considerations A natural explanation for this excess of in IllustrisTNG is that Ω_H i is underestimated by the observations. gas in low-column-density gas could be difficult to detect, while very high columns are prone to self-absorption and hence an underestimation of the actual column density. Indeed, <cit.> find based on observations of M31, M33 and the LMC that self-absorption hides a significant fraction of gas. Correcting for this self-absorption, <cit.> claim that Ω_H i=(5.9±0.9)×10^-4. However, we remark that this study is questioned by more recent findings of <cit.> who use a more thorough model to measure the self-absorption correction. In any case, the distribution of in terms of column densities provides a convenient means to explore if we miss gas when measuring Ω_H i from 21-cm emission data. From the linear inset of Figure <ref>, we learn that Ω_H i is dominated by an intermediate column density regime of N_H i≈10^19-4×10^21 cm^-2. From our mock maps it is impossible to assess if column densities below N_H i≈10^19 contribute substantially to Ω_H i as the mock maps are segmented at approximately this column density (for the low-resolution maps). Generating plain maps (without noise and segmentation) for the full TNG50 base sample (see Section <ref> for more details), we find that column densities in the range 10^17-10^19 cm^-2 only contribute 2.2 % to the total abundance. At the other extreme, column densities above ≈4×10^21 cm^-2 hardly contribute to Ω_H i both for TNG50 and the WHISP data, indicating that self-absorption does not hide a significant amount of gas ( find that self-absorption only becomes important above N_H i≈10^22 cm^-2). It is also possible that blind surveys simply miss a significant population of objects. If the additional gas in TNG50 stemmed from galaxies with very small masses, these objects could be missed, while very heavy galaxies could be rare and hence difficult to observe. For TNG50, we find from the HIMF (Figure <ref>) that the dominating objects for Ω_H i have masses within 3×10^8-3×10^10 M_⊙. We do not expect that our base sample selection (imposing M_gas>10^7 M_⊙ and M_⋆>10^7 M_⊙) or the limited simulation volume affects this finding. Galaxies in this mass range are easy to detect with blind surveys, for ALFALFA the sensitivity is good enough such that the HIMF can be precisely determined for masses larger than 10^7 M_⊙ (figure 2 of ). We conclude that the mismatch in Ω_H i (i.e. the CDDF normalization) between TNG50 and WHISP is due to an excess of atomic hydrogen in the simulation and not an underestimation by the observational data. §.§.§ Simulation considerations A possible resolution of this Ω_H i discrepancy consists in mimicking the blind surveys like HIPASS or ALFALFA in IllustrisTNG, taking sensitivity limits and other observational effects into account (which were neglected in previous studies of Ω_H i in IllustrisTNG). While it is beyond the scope of this paper to fully emulate the ALFALFA or HIPASS observational setup, we assess the impact of the observational effect which probably affects the inferred abundance most, which is beam smoothing. An ALFALFA-like beam smoothing was used in <cit.> (based on ) to emulate the mass function measurement in IllustrisTNG. We follow their approach and apply a Gaussian beam with σ=70 kpc, i.e. we multiply the mass of each gas cell by exp[-r^2/(2σ^2)], with r being the 2D distance of the gas cell to the galaxy center in face-on projection (see (i) in <ref> for the algorithm to rotate galaxies into face-on projection). For our default model, we find that applying this beam smoothing effect reduces the abundance in TNG50 by ten percent to Ω_H i=5.4×10^-4. Hence, mimicking blind surveys reduces the tension in abundance and the CDDF normalization, but does not resolve it for TNG50. The IllustrisTNG simulation resolution also affects the abundance. For the base galaxy samples of TNG50-2 (TNG100) and the default model, we find Ω_H i=5.1×10^-4 (Ω_H i=4.8×10^-4). Applying the ALFALFA-like beam smoothing further reduces these values by approximately ten percent. All observational and simulated inferences of Ω_H i are summarized in Table <ref>. The abundances of the lower-resolution simulations TNG50-2 and TNG100, taking the ALFALFA beam into account, are consistent with the observational values. We show the CDDF for various IllustrisTNG simulations (calculated as described in Section <ref>) in Figure <ref>. Note that the scaling factor ϕ(M_H i) was computed as in Section <ref>, i.e. without ALFALFA-like beam smoothing. This means that the normalization of the IllustrisTNG CDDFs corresponds to the larger values in Table <ref> (e.g. 4.8×10^-4 for TNG100). Differences in the CDDF (left panel) for the various IllustrisTNG runs appear marginal, except at the highest column densities which are dominated by very few high-column-density pixels. In the linear inset in the right panel, it becomes clear that TNG100 is closer to the observational data than TNG50, which is due to the lower abundance in TNG100. Furthermore, we note that the effect of varying the simulation volume is negligible compared to varying the simulation resolution, as TNG100 and TNG50-2 have similar CDDFs (the TNG100 volume is roughly eight times larger than the TNG50-2 one, while the TNG100 mass resolution is comparable to TNG50-2, see Table <ref>). The Ω_H i excess in TNG50 is reminiscent of the luminosity functions analyzed in <cit.>, who find that the lower-resolution TNG50-2 provides a substantially better match to observational data than TNG50. This is expected given that the physical model calibration for IllustrisTNG was undertaken at the resolution of roughly TNG100 (TNG50-2 has a comparable resolution). Indeed, <cit.> show that the stellar mass function in IllustrisTNG is converging, but not fully converged, as a function of numerical resolution. For a given dark matter halo, galaxies form slightly more stars at higher resolutions, shifting the stellar mass function. It is not surprising that these higher stellar masses lead to higher masses[This argument assumes that the H i-to-stellar mass ratio is independent of numerical resolution. Indeed, <cit.> find that this is true to first order and the fractions of various IllustrisTNG simulations agree well with an observational compilation from <cit.>.] and ultimately too much atomic hydrogen in the TNG50 galaxy population. §.§ Hydrogen partitioning based on volume densities The simulation-calibrated partitioning schemes of GK11 and GD14 also exist in `local' variants that are based on volume instead of surface densities. At first sight this is a very attractive feature because it completely circumvents the difficulties of assigning surface densities to gas cells in simulations. According to <cit.>, a cosmological simulation using their local partitioning scheme needs to have a spatial resolution of L_cell∼100 pc. Additionally, <cit.> introduce a correction factor for the spatial resolution which is tested up to L_cell∼500 pc, well in excess of the median size of star-forming gas cells in TNG50 of 138 pc (). For the morphologies, we find that the usage of local partitioning schemes provides a slightly better match to the concentration and M_20 indices between TNG50 and WHISP galaxy samples. Here, we test the local partitioning schemes in terms of the CDDF in Figure <ref>. The application of these schemes to IllustrisTNG with the relevant equations is described in Appendix <ref> and <ref>. Since the predictions of the local partitioning models are similar upon variation of the UV field scheme we only consider the SKIRT UV field here. For completeness we also check the CDDF using the two UV-independent partitioning schemes (KMT09 and S14) in Figure <ref>. We find that the KMT09 model (as well as the UV-dependent GK11 and K13 schemes which are not shown here) produces results comparable to the default GD14 partitioning scheme, while S14 underestimates the abundance of high-column gas even more than the other models. The local partitioning schemes provide a perfect match at the high-column end, which is a substantial improvement over all other models. The higher concentration values of the maps using local partitioning schemes are related to this, as the central holes are less prevalent and more is formed in the dense galactic central regions (see Section <ref>). However, the abundance is very high with the local partitioning schemes (visible in the linear inset in the right panel of Figure <ref>), and reaches almost Ω_H i=10×10^-4 for TNG50, 2.5 times more than what is observed. This excess of using local partitioning schemes persists to TNG100, where we find Ω_H i=6.9×10^-4. This provides an a posteriori justification for refraining from using the local partitioning schemes, albeit they are conceptually appealing as they circumvent the need to estimate column densities for the 3D cosmological simulations. A naive explanation for the difference between conventional, column-density based and local partitioning schemes is that the Jeans approximation breaks down at high column densities. For the presumably star-forming gas cells in these high-density regions, the internal energy of the gas cell which is recorded by IllustrisTNG is an average over cold and hot phases in the context of the two-phase model of <cit.>. This means that the internal energy of the cold, neutral gas phase is overestimated, artificially increasing the Jeans length (see Eqn. <ref>) for the neutral star-forming gas. This unrealistically boosts neutral hydrogen column densities estimated from the Jeans approximation and molecular hydrogen formation, thereby explaining the low abundance of in high-density regions. <cit.> tested this issue by computing neutral hydrogen surface densities with the Jeans approximation, both with the internal energy recorded by IllustrisTNG and the internal energy of the cold gas (using T=1000 K for the cold gas). Surprisingly, they find that the Jeans approximation with the internal energy recorded by IllustrisTNG fits the true surface densities (estimated from projecting the gas cells) better. Instead of tweaking the Jeans approximation, a resolution for discrepancies in the morphological statistics and the CDDF is to use local partitioning schemes, under the assumption that there is simply too much neutral hydrogen in IllustrisTNG. For local partitioning schemes applied to TNG100, the abundance is approximately 1.75 times higher than the observational value. Scaling down the CDDF (computed from local partitioning schemes) by this factor yields an excellent agreement to the WHISP data, even at the high-column-density end. If the local partitioning schemes realistically split the gas into atomic and molecular phases, then the only way to reconcile the observational data with IllustrisTNG is to reduce the amount of neutral hydrogen in the simulation. To test if IllustrisTNG galaxies exhibit too much neutral gas requires mimicking both and H_2 abundance measurements using consistent postprocessing methods (i.e. a consistent hydrogen partitioning scheme). While this is beyond the scope of this paper, we note that results from <cit.> for TNG100 tentatively indicate that there is indeed an overabundance of molecular hydrogen (using conventional, column-density based partitioning schemes). It would be interesting to test if this H_2 overabundance persists when using local partitioning schemes. §.§ TNG-intrinsic CDDF and comparison to other studies §.§.§ CDDF from plain maps Our results for the CDDF (Figure <ref>), especially the mismatch between simulations and observations for high column densities, contrasts with similar studies of the CDDF (; ; ). We attribute the discrepancies to our mock map creation (see Section <ref>) which takes observational effects into account, while other studies of the CDDF in cosmological simulations adopted simpler prescriptions to generate synthetic maps. To verify this and to assess the impact of the observational effects modelled in the context of our mock map algorithm, we construct the CDDF for TNG50 using simpler plain maps (see Section <ref>) in this section. We generate plain maps for the TNG50 base sample of 12431 galaxies, and calculate the CDDF directly from this broad sample without the need of a scaling factor as in Eqn. <ref> (`TNG-intrinsic CDDF'): CDDF_H i=c/H_0∑_jA(N_H i)_j/N_H i ln10 Δ(log_10N_H i) V, where the sum runs over all galaxies in the (broad) sample. Typically, the approach to extract the CDDF from cosmological simulations is to project the full simulation box onto a 2D grid and measure the CDDF there. However, since we focus on the high-density end of the column density distribution which traces gas within galaxies, we find Eqn. <ref> more appropriate here. The intrinsic CDDF generated from plain maps is shown in Figure <ref> (dashed red line). The beam smoothing effect on the mock maps is clearly visible in Figure <ref>, as the CDDF from the mock maps (solid red line) has the highest column density peaks smoothed out and moved to intermediate densities, thereby increasing the tension with the WHISP CDDF. In fact, the CDDFs generated from plain maps (especially from the lower-resolution TNG50-2 and TNG100 runs which are not shown in Figure <ref>) show a good agreement with the observational data, as found in previous studies of the CDDF. We have verified that the choice of the map resolution of the plain maps (128×128 by default) does not impact our CDDF result. §.§.§ Other studies of the CDDF in cosmological simulations A few studies which analyzed CDDFs in cosmological simulations exist. Here, we compare our CDDF to their findings to provide some additional context for our results. <cit.> computed the CDDF for a set of cosmological simulations which explicitly track collisional ionization, radiation from hydrogen recombination and the UV background which sets the hydrogen ionization state. The CDDF itself is calculated by projecting the entire simulation box onto one side, contrary to our approach of summing up the column density distributions of individual galaxies. Molecular hydrogen formation is modelled with the empirical partitioning scheme of <cit.>. <cit.> find an excellent agreement at redshift zero between their CDDF (green line in Figure <ref>) and the WHISP measurement by <cit.>, as their CDDF at the low-column end is lower and hence more in line with observational data. This difference probably arises due to the usage of different cosmological simulations (for instance, the reference simulation of has baryonic particle masses two orders of magnitude larger than TNG50). With radiation from local stellar sources added in <cit.>, the CDDF sharply falls by 0.6 dex for N_H i>10^21 cm^-2 due to a larger ionized hydrogen fraction at large column densities (blue line in Figure <ref>). This leads to a significant underestimation of the CDDF at these high columns. H_2 formation is not considered in <cit.>, leading to the upturn of the CDDF at the highest column densities. We remark that the calculation of the neutral hydrogen fraction in IllustrisTNG is based on <cit.>, without the correction of local stellar sources. Since we base the calculation of the neutral hydrogen fraction in star-forming gas cells on the fraction of gas in the cold phase instead of the IllustrisTNG output, it is not directly clear how the inclusion of local ionizing sources would affect the CDDF. More recently, <cit.> postprocessed the TNG100 simulation with the UV-independent partioning scheme of KMT09 and computed the CDDF in a similar fashion as <cit.>, i.e. by projecting the full simulation box onto one side with 20`000×20`000 points. A notable difference to our method is the usage of the cell size instead of the Jeans length for the calculation of surface densities which are used for the KMT09 partitioning scheme. We find that for star-forming gas cells, the cell sizes are smaller than the Jeans lengths by approximately one order of magnitude. This should significantly lower the molecular fractions for the gas cells. Furthermore, their calculation of the CDDF consists of modelling the gas cells as uniform density spheres, and computing the column density from the length of the line-of-sight segment intersecting the spheres. We find that their CDDF (purple line in Figure <ref>) is lower than ours for all columns below ≈5×10^21 cm^-2. The discrepancy rises to almost one order of magnitude at N_H i≈10^21 cm^-2, and this result holds even when we apply the KMT09 partitioning scheme to the TNG100 simulation for our result. We attribute this discrepancy to the usage of uniform density spheres in <cit.> when calculating the CDDF. We find that for N_H i∼10^21 cm^-2, the radius of a gas cell (modelled as a sphere) does not exceed R_sphere=290 pc: R_sphere=(3M_cell/(4πρ_gas))^1/3, where M_cell≈1.4×10^6 M_⊙ is the mass of a TNG100 gas cell and ρ_gas its gas mass density. The column density in <cit.> is calculated according to N_H i=d_segmentρ_H i/m_H i, with d_segment being the segment through the spherically modelled gas cell. Since d_segment<2R_sphere and ρ_HI<ρ_gas, we have R_sphere<√(3M_cell/(2π m_H iN_H i))≈290 pc for the TNG100 gas cell mass and N_H i=10^21 cm^-2. For the 20`000×20`000 grid used by <cit.> to calculate the CDDF, the different line-of-sights are separated by ≈5 kpc, which means it is very unlikely for the high-density gas to be picked up by the line-of-sights. Since we smooth the Voronoi gas cells when projecting them to a 2D grid we can more accurately take the high-density gas into account. We stress that for our approach for the TNG-intrinsic CDDF (using plain maps), the map resolution (with a default 128×128 grid for each galaxy) does not affect the CDDF. The discrepancy in the CDDF between our result and <cit.> diminishes at the highest column densities, which we attribute to an underestimate of molecular hydrogen formation in <cit.>. This effect increases the CDDF at the highest column densities and hence counters the effect of missing high-density gas cells in <cit.>. Lastly, <cit.> examined the H_2 CDDF in TNG50/TNG100 and its dependency on simulation and map resolution. The H_2 fraction is taken from <cit.> with different partitioning schemes, and the CDDF is calculated by summing over the contributions from individual galaxies as in the present study. In line with our results, <cit.> find that the map resolution (changing from 150 pc to 1 kpc maps) hardly affects the CDDF. The simulation resolution (comparing TNG50 and TNG100) affects the H_2 CDDF in the very high column density regime (N_H_2>10^22 cm^-2), with more molecular hydrogen formed in TNG100. This contrasts to our finding in Figure <ref> where there is more in TNG100 for N_H i>5×10^21 cm^-2 compared to TNG50. § CONCLUSIONS We postprocessed the IllustrisTNG simulations at redshift zero, partitioning the neutral hydrogen content into its atomic and molecular fractions. We explored how the UV radiation from different stellar populations and the propagation of the UV field through the dusty interstellar medium affect the molecular fractions with the radiative transfer code SKIRT. We generated WHISP-like mock maps for IllustrisTNG galaxies and compared them to 21-cm data from the WHISP survey. To compare these resolved maps, we used non-parametric morphologies and the column density distribution function. Our main findings are summarized as follows: * Realistic UV fields taking dust attenuation into account can substantially affect the distribution for a subset of individual galaxies, but for statistical averages such as the H i/H_2 mass functions (Figure <ref>) and average radial profiles (Figure <ref>) the dust attenuation effect is negligible compared to the optically thin (Diemer) scheme. * If the UV radiation is not propagated at all (Lagos scheme), significant statistical differences compared to the SKIRT/Diemer schemes arise. For instance, the mass function is underestimated by 25 % at the high-mass end, while the H_2 mass function is overestimated by 25 % for M_H_2>3×10^8 M_⊙ (Figure <ref>). We remark that our default partitioning scheme (GD14) minimizes the differences between the various UV fields. * For the 30"-resolution maps (which are most reliable), the non-parametric morphologies of maps of WHISP observational data and mock maps of TNG50 galaxies are in good agreement for the asymmetry, smoothness and Gini statistics (Figure <ref>). On the other hand, the TNG50 galaxies feature lower C and higher M_20 values (Figure <ref>). Visual inspection of the TNG50 maps reveals that a substantial amount of TNG50 galaxies exhibits large central holes, which are not seen in the WHISP data. The TNG50 galaxies with a very low concentration statistic are exclusively face-on, where the impact of the central hole on C is maximized. * We attribute the prevalence of central holes in TNG50 galaxies mostly to feedback from AGN, which ionizes and/or expels the neutral hydrogen gas from galaxy centers. Excluding galaxies that experienced feedback from the kinetic channel or above-average energy injection from the thermal channel reduces the tension in the C and M_20 statistics. Switching to the lower-resolution TNG100 simulation (the resolution at which the IllustrisTNG physical model is calibrated) also lowers the tension. * The column density distribution function (CDDF) of TNG50 differs from the one based on WHISP data. TNG50 contains more at intermediate column densities (N_H i≈10^20-20^21 cm^-2), but less high-column-density gas compared to WHISP (Figure <ref>). As the bulk of the abundance (Ω_H i) stems from intermediate column densities, TNG50 also has Ω_H i exceeding observational estimates based on blind surveys (e.g. ALFALFA). The lack of high-column in TNG50 (and IllustrisTNG in general) is due to the beam smoothing in the mock map creation, an effect that was neglected in previous studies of the CDDF in cosmological simulations (; ; ). * As the bulk of the gas resides in a sweet spot for observational detection according to TNG50 (both in terms of masses and column densities), we attribute the Ω_H i tension to an overabundance of atomic hydrogen in TNG50 (and not to an underestimation of the observational value). For the lower-resolution TNG100 simulation, the abundance agrees with the observed value, leading to a better match in the CDDF as well. This result is expected as a TNG50 galaxy will generally have higher stellar and gas masses compared to a TNG100 galaxy in a similar dark matter halo, due to the IllustrisTNG model being calibrated at the TNG100 resolution. * The mismatch in the CDDF at the high-column-density end can be remedied by using local partitioning schemes (based on number instead of column densities). Local partitioning schemes also increase the atomic fractions in galaxy centers, which increases the concentration statistic of TNG50 maps and reduces the tension with the WHISP morphologies. Hence, we hypothesize that the conventional column-based partitioning schemes fail at the highest column densities as the Jeans approximation could break down. The drawback of the local partitioning schemes is that they significantly overpredict the overall abundance of atomic hydrogen. Hence, albeit conceptually appealing, we refrain from using local partitioning schemes for the main results of this paper. Based on the various partitioning schemes and UV fields explored in this study, we recommend to use the column-based (not the local) GD14 model for hydrogen partitioning in cosmological simulations. The GD14 model is an update of the GK11 model, contains a factor to take the cell size into account to mitigate resolution dependencies, and nicely couples conceptually to cosmological simulations as it is itself a simulation-calibrated partitioning scheme. Furthermore, it predicts the largest H_2 fractions of all partitioning schemes considered in this study (except for S14, but this partitioning scheme leads to the most significant tensions in the non-parametric morphologies and the CDDF comparing TNG50 and WHISP), such that it provides the best match to the observational value of Ω_H i. For the UV fields, we recommend to always spread the UV flux, i.e. to use the Diemer or SKIRT schemes. The effect of dust attenuation is negligible for statistical properties of the galaxy population like the CDDF or the HIMF. However, the dust attenuation effect can be significant for individual galaxies and becomes more pronounced for molecular hydrogen properties. Studies that consider H_2 in smaller galaxy samples or zoom-simulations could obtain significantly biased results if not correcting for the effect of dust attenuation. § ACKNOWLEDGMENTS We wish to express our gratitude towards Ana Trčka who helped setting up the SKIRT analysis of the present study. We also wish to thank Benne Holwerda and Nick Gnedin for fruitful discussions about non-parametric morphologies and partitioning schemes, and are grateful for feedback from Annalisa Pillepich and Matthew Smith. AG gratefully acknowledges financial support from the Fund for Scientific Research Flanders (FWO-Vlaanderen, project FWO.3F0.2021.0030.01). DN acknowledges funding from the Deutsche Forschungsgemeinschaft (DFG) through an Emmy Noether Research Group (grant number NE 2441/1-1). This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 882793 `MeerGas'). This study made extensive use of the programming language, especially the (), (), and () packages. The WHISP observation were carried out with the Westerbork Synthesis Radio Telescope, which is operated by the Netherlands Foundation for Research in Astronomy (ASTRON) with financial support from the Netherlands Foundation for Scientific Research (NWO). The WHISP project was carried out at the Kapteyn Astronomical Institute by J. Kamphuis, D. Sijbring and Y. Tang under the supervision of T.S. van Albada, J.M. van der Hulst and R. Sancisi. § DATA AND CODE AVAILABILITY The IllustrisTNG data used in this work is publicly available at <https://www.tng-project.org/> as described by <cit.>. The WHISP moment-zero maps in all resolutions are publicly available at <http://wow.astron.nl/>. WISE fluxes for the WHISP galaxies are publicly available at <https://academic.oup.com/mnras/article/502/4/5711/6095720> (). Distances to WHISP galaxies are taken from the publicly available NED database (<https://ned.ipac.caltech.edu/>). The SKIRT code (version 9) is publicly available at <https://skirt.ugent.be/root/_home.html> and described by <cit.>. The code () is publicly available at its github repository (<https://github.com/vrodgom/statmorph>). The algorithm to spread the UV field in an optically thin fashion (Diemer UV field) as well as the projection algorithm is part of the private repository of Benedikt Diemer. We are happy to share all other parts of the code and generated data of this work upon request. mnras § IMPLEMENTATION OF PARTITIONING SCHEMES §.§ Krumholz, McKee & Tumlinson 2009 (KMT09) <cit.> (KMT09) consider an analytical model of molecular hydrogen formation and dissociation akin to S14. The model is equivalent to K13 except for the estimation of χ which quantifies the balance between UV photodissociation and recombination of H_2. While this parameter depends on U_MW in <cit.>, it is estimated from the metal mass fraction in the gas cell, Z (Eqn. 7 in or Eqn.4 in ): χ=3.1(1+3.1(Z/Z_⊙)^0.365/4.1). The molecular hydrogen fraction as well as the s and τ_c parameter are identically calculated as in the K13 model described in appendix C.4 of <cit.> (their Eqns. 18, 19 and 26). We remark that this partitioning scheme is used to partition the hydrogen in the cosmological simulations MUFASA () and Simba (), but with a clumping factor of ≈10 and the Sobolev approximation when calculating τ_c (we use a clumping factor of five and the Jeans approximation following ). §.§ Gnedin & Kravtsov 2011 (local version) This partitioning scheme is the local version (having number instead of column densities as input) of the GK11 scheme. We refer to appendix C2 of <cit.> for the description of the g parameter (their equations 11 and 12), and write only the new equations for this local, UV-dependent partitioning scheme: f_mol=1/1+exp(-4x-3x^3), x=Λ^3/7ln(D_MWn_H/Λ n_∗), where n_H is the total hydrogen number density in protons per cm^3, n_∗=25 cm^-3, and Λ is Λ=ln(1+gD_MW^3/7(U_MW/15)^4/7). Note that f_mol in this formalism technically corresponds to the mass fraction in molecular hydrogen compared to total hydrogen, while we use f_mol for the mass fraction in H_2 compared to neutral hydrogen. We use the two quantities interchangeably as under all conditions of interest, the ionized hydrogen fraction is negligible when molecular hydrogen forms. §.§ Gnedin & Draine 2014 (local version) Equivalently, this is the local version of the GD14 partitioning scheme. As this varies quite substantially from their partitioning scheme based on surface densities presented in appendix C3 of <cit.> we note the full set of equations here: f_mol=1/1+exp(-x(1-0.02x+0.001x^2)), x=wln(n_H/n_1/2), where n_1/2 is the hydrogen number density at which f_mol=0.5. n_1/2=n_∗Λ/g, w=0.8+Λ^1/2/S^1/3, Λ=ln(1+(0.05/g+U_MW)^2/3g^1/3/U_∗), n_∗=14 cm^-3D_∗^1/2/S, U_∗=9D_∗/S, D_∗=0.172+S^5/1+S^5, and S=L_cell/100 pc is a factor to correct for the spatial resolution of the simulation. Following <cit.> we approximate L_cell=(m/ρ)^1/3 for each gas cell. We remark that <cit.> calibrated this resolution correction up to cell sizes of 500 pc, which should be sufficient to reach the TNG50 cell sizes at which molecular hydrogen formation becomes important. § PROJECTED MODELLING A troublesome aspect of hydrogen partitioning in cosmological simulations is the usage of surface (or column) densities in partitioning schemes, which are difficult and somewhat ambiguous to extract from the simulations. The approach adopted here and in most other studies is to calculate surface densities of gas cells with the Jeans approximation (`volumetric modelling', Eqn. <ref>). <cit.> find that the Jeans approximation introduces significant biases in the estimated surface densities. A possible resolution to overcome this issue is the usage of local partitioning schemes as discussed in Section <ref>. Another alternative suggested by <cit.> is to estimate surface densities by rotating galaxies into a face-on position, projecting the gas cells on a 2D grid to compute surface densities, and performing the hydrogen partitioning directly on this 2D grid (`projected modelling'). For the analytical and simulation-calibrated partitioning schemes considered here, <cit.> find that the volumetric and projected approaches lead to similar results on average for the entire galaxy population, but with large scatter for individual galaxies (see their figure 4). It is not obvious which approach should be preferred, as surface densities have different meanings across the various partitioning schemes. For instance, S14 consider an idealized interstellar gas cloud as a 1D slab, while GK11 consider columns calculated from lower-resolution gas cells with a size of ≈500 pc corresponding to the average thickness of galactic disks in their simulations. An advantage of projected modelling is that it is significantly faster than volumetric modelling. For our purposes, however, an important shortcoming of the projected approach is that one can only generate maps in face-on projection. While this is fine to calculate masses and radial profiles, for the mock maps that we compare to the WHISP survey (which contains galaxies in random orientation) we cannot use the projected approach there. We test here if the finding of <cit.> that the volumetric and projected approaches agree on average (for each partitioning scheme) extends to the SKIRT and Lagos UV fields. To this end we calculate and H_2 masses both in the volumetric and projected incarnation with our default partitioning scheme for the base TNG50 sample. For the projected modelling we generate the face-on maps of all quantities required for the hydrogen partitioning using the plain map algorithm (Section <ref>). Figure <ref> shows the deviation in galaxy H i/H_2 mass from the default model (volumetric modelling with the SKIRT UV field) when changing the UV field and/or switching to projected modelling. The lines indicate the running medians, the hatched areas correspond to the interquartile range. Unsurprisingly, the Lagos models deviate most from the default model, furthermore they also feature the largest difference between the volumetric and projected version (solid and dashed orange lines). For intermediate H_2 masses of ∼10^9 M_⊙, projected modelling leads to H_2 masses lower by a factor of two compared to volumetric modelling (with the Lagos UV field). For the SKIRT and Diemer UV fields the discrepancies between the volumetric and projected models are mostly not significant as the interquartile ranges overlap. For our estimate of Ω_H i with the default model (volumetric SKIRT model), switching to the projected SKIRT model would slightly increase the abundance in IllustrisTNG by ≈5 %. Hence, this systematic uncertainty arising from volumetric/projected modelling does not affect the conclusions from Section <ref>. We note that the default model seems to underestimate the largest masses by ∼20 % compared to all other five models. Furthermore, we remark that although the mass functions between the different UV field estimates are mostly consistent (Figure <ref>), individual galactic H i/H_2 masses can differ a lot between different models, especially when comparing the Lagos UV field with the SKIRT/Diemer UV fields.

http://arxiv.org/abs/2303.06711v1

20230312172337

Large-scale homogeneity and isotropy versus fine-scale condensation. A model based on Muckenhoupt type densities

math.AP

]Large-scale homogeneity and isotropy versus fine-scale condensation. A model based on Muckenhoupt type densities ]Hugo Aimar ]Federico Morana In this brief note we aim to provide, through a well known class of singular densities in harmonic analysis, a simple approach to the fact that the homogeneity of the universe on scales of the order of a hundred millions light years is completely compatible with the fine-scale condensation of matter and energy. We give precise and quantitative definitions of homogeneity and isotropy on large scales. Then we show that Muckenhoupt densities have the ingredients required to a model for the large-scale homogeneity and the fine-scale condensation of the universe. In particular, these densities can take locally infinitely large values (black holes) and at once, in the large scales they are independent of the location. We also show some locally singular densities satisfying the large scale isotropy property. [ [ March 30, 2023 ================== § INTRODUCTION The cosmological principle of the universe states that the universe is homogeneous and isotropic at large scales. Large means here lengths larger than the diameter of galaxy clusters (see Chapter 27 of <cit.>). But, of course, no homogeneity could be expected at small scales. Planets, stars, galaxies are real inhomogeneities in which there is concentration of mass and energy. From the point of view of mathematics, homogeneity has also been considered from several points of view. Perhaps the most simple but non-quantitative meaning is that of topological homogeneity. In this approach the expression “the space looks the same no matter where you are” is interpreted from the topological perspective. Hence a topological space X is homogeneous if for every couple of points x and y in X there exists a homeomorphism f:X→ X such that f(x)=y. This means that the universe looks topologically the same about x than about y. Nevertheless this concept does not take into account the local heterogeneities and it is not quantitative. For simplicity, assume that the “substance” of an n-dimensional Euclidean universe is distributed according to a density ρ(x). Here x=(x_1,⋯,x_n) is a point of ^n and ρ(x)≥ 0. Let us proceed to give a plausible definition of homogeneity in the large scales for a general density ρ(x) of our n-dimensional space. Assume that our universe has two astronomers, located at different points x^1 and x^2. Assume that their astronomical skills grow with the same rhythm. In particular they are able to weigh their own neighborhoods. Set B(x^i, R), i=1,2, to denote the euclidean ball centered at x^i with radius R>0. Precisely B(x^i,R)={x∈^n:|x-x^i|<R}. At the beginning their observations will be disjoint, when R increases they will start sharing part of their observable universe. When R is very large, the will share most of their measurements but the two observed regions will never coincide. In our elementary interpretation, the two observers are drawing the increasing functions M_1(R)=∫_B(x^1,R)ρ(x) dx and M_2(R)=∫_B(x^2,R)ρ(x) dx respectively. We say that the universe, modeled by ρ(x), is homogeneous in the large scales if no matter what the locations x^1 and x^2 are, the two observers are weighing the same mass for R large. More precisely if M_1(R)/M_2(R)1 for every choice x^1, x^2 of points in . It is worthy noticing that not every locally integrable density satisfies this property. In fact, as it is easy to prove, exponential growth is not allowed. The question is whether or not there are nontrivial densities satisfying (<ref>). Non triviality should allow very strong local heterogeneities. Not only objects of large mass but also points of infinitely large density in order to include, at least in an heuristic way, an elementary model for black holes. We aim to show that there is a huge class of densities ρ in ^n fulfilling this two apparently contradictory conditions: homogeneity in the large scales, in the sense of (<ref>), and strong local heterogeneity allowing heavy objects and locally unbounded density. The class of densities that we shall consider is the well known, in harmonic analysis, class of Muckenhoupt weights. See <cit.>, <cit.>, <cit.>. Section <ref> is devoted to a brief introduction of this class of densities and to exhibit some nontrivial examples, including densities that are singular on sets of positive dimension. In Section <ref> we prove that the densities defined by Muckenhoupt weights provide the desired homogeneity at large scales in the sense of (<ref>). In Section <ref> we briefly discuss the corresponding isotropy condition with these types of mass distributions. § MUCKENHOUPT DENSITIES Let us start by introducing the Muckenhoupt classes in the Euclidean setting . There are extensive treatments of the subject in harmonic analysis such as <cit.>, <cit.>, <cit.>. These approaches focus usually more on operator theory, than into the geometrical properties associated to these densities. With B we shall denote the Euclidean open balls of . We shall write B(x,r) when the center x∈ and the radius r>0 of B have to be explicit. A positive measurable and locally integrable function ρ defined in satisfies always, by Schwarz inequality, the following upper estimate for the volume |B| of a ball B |B|^2 = (∫_B dx)^2= (∫_Bρ^1/2ρ^-1/2 dx)^2 ≤(∫_Bρ dx)(∫_Bρ^-1 dx). for every ball B. More generally, from Hölder inequality with 1<p<∞ and 1/q+1/p=1, we have also |B|^p = (∫_B dx)^p = (∫_Bρ^1/pρ^-1/p dx)^p ≤{(∫_Bρ dx)^1/p(∫_Bρ^-q/p dx)^1/q}^p = (∫_Bρ dx)(∫_Bρ^-1/p-1 dx)^p-1 again for every ball B and every density ρ. It is easy to provide examples of densities in which the above estimates for the volume of B are far away from being optimal. On the other hand it is clear that the above inequalities become equations for balls in regions where the density ρ is constant. And that |B|^p and (∫_Bρ dx) (∫_Bρ^-1/p-1 dx)^p-1 are of the same order if there exist positive and finite constants 0<c_1≤ρ(x)≤ c_2<∞ for every x∈. In other words, these two magnitudes become equivalent when ρ is almost constant. On the other hand, from the point of view of the model for a density of the universe, the property 0<c_1≤ρ(x)≤ c_2<∞ is unrealistic because it prevents having very localized and massive objects and even for having some points with vanishing density. Precisely, Muckenhoupt densities deserve analysis and particular consideration because the quantities |B|^p and (∫_Bρ dx)(∫_Bρ^-1/p-1 dx)^p-1 are equivalent and ρ can still have singularities, points of infinity density and points of vanishing density. Let ρ be a density in and 1<p<∞. We shall say that ρ is a Muckenhoupt density, of the class A_p, if there exists a constant C such that (∫_Bρ dx)(∫_Bρ^-1/p-1 dx)^p-1≤ C|B|^p for every ball B in . A density ρ is said to belong to A_∞ if ρ∈ A_p for some 1<p<∞. Benjamin Muckenhoupt introduced in <cit.> these densities as the exact classes for the boundedness of Hardy-Littlewood and singular integrals operators in Lebesgue spaces. In this brief introduction we shall only collect the measure theoretical aspects of Muckenhoupt densities allowing us to prove, in the next section, the homogeneity in the large scales of a universe modeled by this type of densities ρ. The next proposition contains the mains properties of Muckenhoupt densities related with the subsequent development of our model. Let ρ be a density in the n-dimensional Euclidean space. Then (a) for 1<p<q and ρ∈ A_p we have that ρ∈ A_q; (b) for p>1 and ρ∈ A_p there exists a constant C such that the inequality |E|/|B|≤ C(ρ(E)/ρ(B))^1/p, ρ(A)=∫_Aρ dx, holds for every ball B and every measurable subset E of B; (c) for p>1, ρ∈ A_p and C as in (b) the mass satisfies the following doubling property ρ(Bx,2R))≤ 2^npC^pρ(B(x,R)) for every x∈ and every R>0; (d) let ρ be any A_∞ density, then there exist two constants C>0 and γ>0 such that the inequality ρ(E)/ρ(B)≤C(|E|/|B|)^γ, holds for every ball B and every measurable subset E of B. The proof of the above proposition can be found in <cit.> or <cit.>. The first three properties are simple consequences of the definition of A_P. Property (d) instead relies on subtle arguments such as the Calderón-Zygmund decomposition technique. On the other hand Property (d) shall be crucial in our proof of the large scale of homogeneity in the next section. Let us give some examples of locally quite heterogeneous Muckenhoupt densities. The first class of non trivial A_∞ densities which are radial with respect to some point x_0∈ are the powers of the distance. In fact ρ(x)=|x-x_0|^β belongs to the Muckenhoupt class A_p() if and only if -n<β<n(p-1). Hence ρ(x)=|x-x_0|^β belongs to A_∞ if and only if β>-n. When β=0 we recover the trivial constant weights but, when -n<β<0 we have a singularity in x_0 and ρ can be arbitrarily large in any neighborhood of x_0. On the other hand, when β>0, we have densities that are unbounded at infinity, but they are still A_∞ densities. An extension of the above is due to Ricci and Stein <cit.> (see also <cit.>). The singularities of the density can actually be concentrated on sets which are quite irregular and large in the sense of Hausdorff dimension. In <cit.>, <cit.> and references therein, the problem of the construction of densities of Muckenhoupt type which are powers (positive or negative) of the distance to a closed set of is considered. The results are of the following type. Let F be a compact set with Hausdorff dimension n>s≥ 0, then ρ(x)=d^β(x,F) belongs to A_p() for -n(n-s)<β<(n-s)(p-1), where d(x,F)=inf{d(x,y):y∈ F} is the distance of x to F. It is important to realize that n-1 is a possible value for s and, then ρ(x)=1/√(d(x,F)) belongs to A_∞ if F is the boundary of some open domain in . Hence we can have densities with infinitely large values on very large sets of Hausdorff dimension less than n, for example on a surface of ^3 . Several important results of harmonic analysis provide tools for building singular weights. In particular the relation of Muckenhoupt classes with the space of functions of bounded mean oscillation. In some sense these results revels that power laws are in some sense the extremals of the A_∞ class of densities. No exponential behavior for a density, as illustrated in Section <ref>, will produce the desired homogeneity in the large scales allowing strong heterogeneities in any bounded region of the space. § THE MAIN RESULT Let us formally state and prove the main result of this note. Let ρ(x) be an A_∞() Muckenhoupt density. Then with the density ρ(x) is homogeneous at the large scales. Let x^1 and x^2 be two points in and let R>0 be given. With the notation introduced in Section <ref> we have to prove that ρ(B(x^1,R))/ρ(B(x^2,R))1 . With the standard notation for intersections and differences of sets we may write ρ(B(x^1,R))/ρ(B(x^2,R)) = ρ(B(x^1,R)∩ B(x^2,R))+ρ(B(x^1,R)∖ B(x^2,R))/ρ(B(x^2,R)∩ B(x^1,R))+ρ(B(x^2,R)∖ B(x^1,R)) = 1+ρ(A_1)ρ(B(x^1,R)∩ B(x^2,R))1+ρ(A_2)ρ(B(x^2,R)∩ B(x^1,R)) with A_1=B(x^1,R)∖ B(x^2,R) and A_2=B(x^2,R)∖ B(x^1,R). Notice that * A_1⊂ E_1= B(x^1,R)∖ B(x^1,R-|x^1-x^2|)⊂ B(x^1,R); * A_2⊂ E_2= B(x^2,R)∖ B(x^2,R-|x^1-x^2|)⊂ B(x^2,R); * B(x^1,R/2)⊂ B(x^1,R)∩ B(x^2,R) for R large enough; * B(x^2,R/2)⊂ B(x^1,R)∩ B(x^2,R) for R large enough. Now use (3), (4) and the doubling property (c) in Proposition <ref> and obtain ρ(A_i)/ρ(B(x^1,R)∩ B(x^2,R))≤ρ(A_i)/ρ(B(x^i,R/2))≤ 2^npC^p ρ(A_i)/ρ(B(x^i,R)) for i=1,2 and p such that ρ∈ A_p(). The constant C is that in (c) of Proposition <ref>. Let us now apply (1) and (2) above and then (d) in Proposition <ref>. Doing so, since E_i⊂ B(x_i,R) (i=1,2) we get ρ(A_i)/ρ(B(x^1,R)∩ B(x^2,R)) ≤ 2^npC^pρ(E_i)/ρ(B(x^i,R)) ≤ 2^npC^pC(|E_i|/|B(x^i,R)|)^γ = 2^npC^pC[R^n-(R-|x^1-x^2|)^n/R^n]^γ = 2^npC^pC[1-(1-|x^1-x^2|/R)^n]^γ for i=1,2. Since the above expression tends to zero for R→∞, we have the desired homogeneity in the large scales, i.e. ρ(B(x^1,R))/ρ(B(x^2,R))1 . Actually the last estimate provides a velocity of convergence in terms of the distance |x^1-x^2| between the two observers. § ISOTROPY The second component of the cosmological principle is the large scale isotropy of the universe. Roughly speaking, isotropy entails that from every point of the space the universe looks, at large scales, the same in any direction. Assume as before that the density of the universe of n dimensions is distributed according to the nonnegative density function ρ(x)=ρ(x_1,…,x_n). A quantitative version of the above qualitative formulation can be given by the fact that lim_R→∞ λ(x_1,v⃗_⃗1⃗,R)/λ(x_2,v⃗_⃗2⃗,R) = 1 for every x_1,x_2 ∈ and v⃗_⃗1⃗,v⃗_⃗2⃗∈ S^n-1, where λ(x,v⃗,R) = ∫_0^Rρ(x+tv⃗) dt for x∈ and v⃗∈ S^n-1 the unit sphere in . Notice that in order to have the function λ well defined for every x∈, every R>0 and every v⃗∈ S^n-1 we need more than the local integrability of ρ. Nevertheless some local singularities of ρ are still possible with power laws weaker than those providing homogeneity. The next result shows a simple example of isotropy and homogeneity with a singularity. Let x_0 be a fixed point in and 0<α<1. Let ρ(x)=x-x_0^-α, then 1/1-α[ (x-x_0+R)^1-α - x-x_0^1-α] ≤λ(x,v⃗,R) ≤x-x_0^1-α/1-α + 1/1-α(R-x-x_0)^1-α for every x∈, v⃗∈ S^n-1 and R>x-x_0. Let us first estimate pointwise from above and from below the restriction of the density ρ to the half line {x+tv⃗: t>0}. Since -x-x_0≤ (x-x_0)·v⃗≤x-x_0, we have, for t>0 that -2tx-x_0≤ 2t (x-x_0)·v⃗≤ 2tx-x_0. Hence (x-x_0) - tx-x_0/x-x_0^2≤(x-x_0)+tv⃗^2≤(x-x_0) + tx-x_0/x-x_0^2. Then for ρ(x+tv⃗)=(x+tv⃗)-x_0^-α, we have ρ(x+tx-x_0/x-x_0) ≤ρ(x+tv⃗)≤ρ(x-tx-x_0/x-x_0). So that, integrating for t∈ [0,R] we get ∫_0^R (x-x_0) + tx-x_0/x-x_0^-α dt ≤λ(x,v⃗,R)≤∫_0^R (x-x_0) - tx-x_0/x-x_0^-α dt. Then ∫_0^R (t+x-x_0)^-α dt ≤λ(x,v⃗,R)≤∫_0^R x-x_0 - t^-α dt, or, for R large 1/1-α [(x-x_0+R)^1-α - x-x_0^1-α] ≤λ(x,v⃗,R) ≤∫_0^x-x_0 (x-x_0-t)^-α dt + ∫_x-x_0^R (t-x-x_0)^-α dt = x-x_0^1-α/1-α + 1/1-α(R-x-x_0)^1-α, as desired. For x_0∈ and 0<α<1, the density ρ(x)=x-x_0^-α is homogeneous and isotropic. Take x_1, x_2∈; v⃗_⃗1⃗, v⃗_⃗2⃗∈ S^n-1 and R large. Applying Lemma <ref> we have 1/1-α[(x_1-x_0+R)^1-α-x_1-x_0^1-α]/x_2-x_0^1-α1-α+ 11-α(R-x_2-x_0)^1-α ≤λ(x_1,v⃗_⃗1⃗,R)λ(x_2,v⃗_⃗2⃗,R) ≤x_1-x_0^1-α1-α+ 11-α(R-x_1-x_0)^1-α/1/1-α[(x_2-x_0+R)^1-α-x_2-x_0^1-α]. The first and the last terms in the above inequalities tend to one for R→∞ and we are done, since ρ is a Muckenhoupt weight. §.§ Funding This work was supported by the Ministerio de Ciencia, Tecnología e Innovación-MINCYT in Argentina: Consejo Nacional de Investigaciones Científicas y Técnicas-CONICET; and Universidad Nacional del Litoral-UNL. 1 AiCaDuTo14 H. Aimar, M. Carena, R. Durán, and M. Toschi, Powers of distances to lower dimensional sets as Muckenhoupt weights, Acta Math. Hungar. 143 (2014), no. 1, 119–137. 3215609 CoifmanFefferman74 R. R. Coifman and C. Fefferman, Weighted norm inequalities for maximal functions and singular integrals, Studia Math. 51 (1974), 241–250. 358205 DuLoG10 Ricardo G. Durán and Fernando López García, Solutions of the divergence and analysis of the Stokes equations in planar Hölder-α domains, Math. Models Methods Appl. Sci. 20 (2010), no. 1, 95–120. 2606245 GarciaCuervaRubioFranciaBook José García-Cuerva and José L. Rubio de Francia, Weighted norm inequalities and related topics, North-Holland Mathematics Studies, vol. 116, North-Holland Publishing Co., Amsterdam, 1985, Notas de Matemática [Mathematical Notes], 104. 807149 MTWgravitation Charles W. Misner, Kip S. Thorne, and John Archibald Wheeler, Gravitation, W. H. Freeman and Co., San Francisco, Calif., 1973. 0418833 Muckenhoupt72 Benjamin Muckenhoupt, Weighted norm inequalities for the Hardy maximal function, Trans. Amer. Math. Soc. 165 (1972), 207–226. 293384 RicciStein87 Fulvio Ricci and E. M. Stein, Harmonic analysis on nilpotent groups and singular integrals. I. Oscillatory integrals, J. Funct. Anal. 73 (1987), no. 1, 179–194. 890662 SteinBook93 Elias M. Stein, Harmonic analysis: real-variable methods, orthogonality, and oscillatory integrals, Princeton Mathematical Series, vol. 43, Princeton University Press, Princeton, NJ, 1993, With the assistance of Timothy S. Murphy, Monographs in Harmonic Analysis, III. 1232192 Affiliation. Instituto de Matemática Aplicada del Litoral, UNL, CONICET. Address: E-mail address. Hugo Aimar (corresponding author), , Federico Morana,

http://arxiv.org/abs/2303.07186v1

20230313152847

Audio-based Roughness Sensing and Tactile Feedback for Haptic Perception in Telepresence

cs.RO

Audio-based Roughness Sensing and Tactile Feedback for Haptic Perception in Telepresence Bastian Pätzold, Andre Rochow, Michael Schreiber, Raphael Memmesheimer, Christian Lenz, Max Schwarz, and Sven Behnke Autonomous Intelligent Systems University of Bonn, Germany March 30, 2023 ========================================================================================================================================================================================= Haptic perception is incredibly important for immersive teleoperation of robots, especially for accomplishing manipulation tasks. We propose a low-cost haptic sensing and rendering system, which is capable of detecting and displaying surface roughness. As the robot fingertip moves across a surface of interest, two microphones capture sound coupled directly through the fingertip and through the air, respectively. A learning-based detector system analyzes the data in real-time and gives roughness estimates with both high temporal resolution and low latency. Finally, an audio-based haptic actuator displays the result to the human operator. We demonstrate the effectiveness of our system through experiments and our winning entry in the ANA Avatar XPRIZE competition finals, where impartial judges solved a roughness-based selection task even without additional vision feedback. We publish our dataset used for training and evaluation together with our trained models to enable reproducibility. Haptics, Audio, Telepresence, Machine Learning § INTRODUCTION Sensing surface properties through haptics is one of the fundamental ways humans perceive their environment. Humans are able to perform a variety of exploratory movements with their hands and fingertips to discern aspects such as roughness, hardness, and shape of objects they manipulate <cit.>. It is widely understood that integrating haptics into VR, AR and teleoperation systems is a key step towards increasing realism and acceptance of such systems <cit.>. Consequently, numerous methods for sensing <cit.> and displaying <cit.> haptic sensations have been developed. These systems are often very complex, though, and require too much space, especially on the side, where, if we aim at using humanoid-like robot hands, the volume of a fingertip has to suffice. In this work, we showcase the haptic system our team NimbRo developed for the ANA Avatar XPRIZE competition[<https://www.xprize.org/prizes/avatar>] <cit.>. The competition focused on immersive telepresence in a mobile robot including social interaction as well as manipulation capabilities. To evaluate the intuitiveness of the developed telepresence systems, judges had to solve a sequence of increasingly difficult tasks after very little training (30min). The most difficult task to be performed in the finals was to discern two stone types based on their surface roughness. In contrast to previous works, our haptic sensing and display system achieves roughness sensing at very low cost by using off-the-shelf audio components. Both sensing and display components are of very small size and are easily integrated into teleoperation systems. The audio signal captured by two microphones is analyzed by a neural network, which was trained on a dataset of exemplary surfaces. The operator is notified about the presence and roughness of the perceived surface through low-latency vibratory feedback. The system was successfully evaluated at the ANA Avatar XPRIZE finals, where three different operator judges solved all tasks in the fastest time, winning our team NimbRo the $5M grand prize. In summary, our contributions include: * Low-cost hardware design of both sensing and display components, * a learning-based method for online and low-latency roughness analysis, and * an evaluation of this system in the competition as well as offline experiments. § RELATED WORK §.§ Tactile Sensing Tactile sensors are based on a wide range of sensing principles, including capacitance, resistance, magnetism, and optics. For example, <cit.> introduced the BioTac tactile sensor, based on an incompressible liquid as an acoustic conductor. In addition to its capability of measuring shear forces, skin stretch and temperature, it detects vibrations with up to 1040 using a pressure sensor. By using only a single sensor per fingertip, it has a low spatial resolution, though. GelSight, proposed by <cit.>, is capable of measuring high-resolution geometry as well as local/shear forces by visually observing the deformation of an elastomer sensor surface. Despite the promising capabilities of such devices, they suffer from two drawbacks: First, their size is too large for deployment in large quantities with high spatial resolution, or integration in existing hardware solutions. Second, their high cost limits their feasibility for many applications. Our work focuses on deploying considerably smaller and lower-cost audio-based hardware. In a similar manner, <cit.> describe the utilization of microphones to classify road surfaces by capturing the tire-pavement interaction noise. They convert the audio signals to time-frequency RGB images and feed them to a CNN. Even though the captured audio signals also depend on other factors than the road surface, such as the car speed, tire type and wheel torque, they demonstrate the effectiveness of their approach for classifying snow and asphalt surfaces. <cit.> use a piezo acoustical sensor to analyze irregularities in materials, such as aluminum or stainless steel, occurring during manufacturing. They capture the friction sounds when moving a stylus with a diamond tip over the surface of the specimen. This controlled environment allows the determination of roughness parameters <cit.> using classical signal processing approaches. Microphones have also been used to identify touch and swiping gestures in smartphone contexts <cit.>. Most similar to our use case, <cit.> help prosthetic users feel textures by stroking a microphone across object surfaces. In contrast to our system, the signal is filtered in a fixed manner, extracting the median frequency, which is then applied to the user using electrostimulation. This limits the scope to regular textures (such as mesh, rubber, etc), where a frequency is easily extracted. In contrast, our method works on irregular surfaces such as natural stones. §.§ Tactile Rendering Tactile actuators are integrated into numerous devices, such as phones or game controllers. Due to the rising interest in telepresence and VR applications, a wide range of haptic displays – typically referred to as Haptic Gloves – emerged in recent years <cit.>, promising to convey realistic kinesthetic and tactile feedback. Typically, tactile feedback is achieved by utilizing eccentric rotating masses, linear resonant actuators or piezoelectric actuators to display vibrations. Their limitations often include supporting only a single resonant frequency, poor intensity resolution, as well as slow response times. Instead, we utilize an acoustic actuator to address these issues while maintaining comparable size and costs. § METHOD We now describe our method in detail. <Ref> gives an overview of the data flow. §.§ Sensing The surface point of the avatar robot to be provided with roughness sensing capabilities (e.g. the tip of an index finger) is equipped with two microphones. A piezo microphone is attached directly to the inside of the chosen surface to measure vibrations within the robot structure, while a MEMS-type microphone is placed in close proximity (∼2) to the outside of the chosen surface measuring vibrations in the air around it. Once the surface makes contact with and slides over the unknown texture of an object, a sound gets induced into both microphones. These audio signals are then leveraged to classify the unknown texture as either smooth or rough. §.§ Classical Detection Initially, we attempted to find a direct mapping between the piezo microphone and our haptic display, utilizing classical DSP approaches like dynamics processors and filters. While it seemed easy for us to classify the considered textures from hearing the piezo microphone signal, we could not find a suitable transformation to accommodate our haptic perception and the properties of the considered display. Our attempts focused on isolating frequency regions that seemed critical for this perception, enhancing their transients and pitch-shifting them into lower registers supported by the actuator. Next, we tried to first classify the piezo microphone signal using similar DSP techniques, to then generate a new audio signal using a sine oscillator that conveys the haptic perception associated with the classification result. In general, rough textures induce louder signals into the piezo microphone than smooth textures. However, the ambiguous amount of pressure applied by the operator masks this effect. Likewise, a hand-held solid natural stone induces a signal that differs strongly in level and frequency spectrum from a hollow artificial stone mounted inside a box, although their textures are very similar. In summary, while we found a functioning configuration for a limited number of objects and scenarios, the classical approach lacked generalization across situations and users. §.§ Learned Detection Instead of hand-designed filters, we opted for a learning-based approach. As the teleoperation task demands low-latency haptic feedback, we update the prediction with every received audio buffer (∼10) by constructing chunks that have access to 256 of the past. After low-pass filtering and reducing the sampling frequency, we calculate the FFT and concatenate the norm of both signals. Experiments showed a sampling frequency of 2 to be sufficient for representing the relevant features for the described task. Classification is then performed by an MLP with 15 hidden layers of which ten layers, with 256 hidden units each, are equipped with residual connections for a better gradient flow. Experiments have shown that the classification accuracy is increased when the unnormalized input of both the piezo and MEMS microphones are used. When sliding over smooth surfaces—in contrast to rough ones—the MEMS microphone's level tends to be significantly quieter than that of the piezo microphone. Such patterns can easily be learned by a neural network and should therefore not be discarded through normalization. In fact, they constitute our motivation for deploying a MEMS microphone in addition to the piezo microphone. During inference, we detect the loudness of the piezo microphone in real-time and compare it against a preset threshold that slightly exceeds its noise floor. This allows distinguishing between contact and no contact situations, which is used to either pass or block the network classification output. §.§ Actuation The classification results are used to update the amplitude and frequency parameters of a simple sine oscillator. In the case of a smooth classification, we set the oscillator to a low amplitude and a high frequency (e.g. 120), while for a rough classification result, we set a higher amplitude and a lower frequency (e.g. 60). For the no contact case, the amplitude is set to zero. The audio signal generated by the oscillator is fed into a compact loudspeaker with a special voice-coil design, capable of reproducing frequencies between 10 and 250 in the form of intense but mostly inaudible vibrations. The speaker is attached to the operator station matching the sensor position on the avatar robot (e.g. a fingertip of a hand exoskeleton). The latency of the end-to-end haptic feedback is defined by the chunk size of the network, the buffer size of the avatar- and operator audio systems, as well as the transmission latency between them. The haptic feedback allows the operator to intuitively distinguish between no contact, contact with a smooth texture and contact with a rough texture situations. The operator's haptic perception of smooth textures can be described as fizzy, while the perception of rough textures might be described as bumpy. § IMPLEMENTATION §.§ Avatar Robot Both microphones are attached to the left index finger of the avatar robot's Schunk SIH hand (<ref>). We replaced the original index finger with two 3D printed components: The distal phalanx and the finger pad (<ref>). Both feature hollow cylinders inside which allow a silicone layer to connect both components. The piezo microphone is glued to the inside surface of the finger pad. The MEMS microphone is placed on the side of the finger, where it is close to the fingertip, but does not interfere during manipulation tasks. In order to avoid any implications caused by the proximity effect on the network's classification performance, we choose this microphone to have an omnidirectional polar pattern. The silicone layer is slightly compliant and decouples the finger pad from the rest of the robot, preventing vibrations to spill over into the piezo microphone. The finger pad shape is designed to allow for sliding over a wide range of textures without getting stuck, while also producing a suitable amount of vibrations in the finger to allow for reliable classification. The microphones are connected to a Focusrite Scarlett 2i2 interface, which is used for pre-amplification and A/D conversion. We set both inputs to high impedance mode, which maximizes the microphones' frequency response. The digital audio signals are forwarded and processed using the JACK Audio Connection Kit which operates on top of the Advanced Linux Sound Architecture. Both signals are transmitted from the avatar robot to the operator station by a low-latency UDP transmitter utilizing the OPUS audio codec <cit.>. §.§ Operator Station The operator station receives the audio signals of both microphones within a similar JACK environment as described for the avatar robot. First, the signal originating from the piezo microphone is routed directly to the headphones of the Valve Index head-mounted display that the operator is wearing. This helps to convey a more detailed and complete haptic perception of the texture in question. Then, both audio signals are fed to the classification network as described in <ref>. The confidence of the rough class is used to modulate the amplitude of a sine oscillator set to a frequency of 60 and a maximum level of 0. We apply a low-pass filter on this modulation to prevent artifacts in the generated waveform arising from immediate jumps in confidence. A second sine oscillator is set to a frequency of 120 and a level of -25. Before mixing the two generated signals, we use a side-chain gate to mute the latter oscillator in no contact situations. This is achieved by feeding the audio signal originating from the piezo microphone to the respective side-chain input. The low noise floor and mechanical decoupling of the piezo microphone allow us to easily find a suitable threshold parameter to be set within the sidechain gate, in order to facilitate a reliable and sensitive way of sensing contact. Again, nonzero attack and release parameters prevent the gate from inducing artifacts arising from fast amplitude modulations. The generated audio signal is forwarded to our haptic display shown in <ref>. Its actuator is extracted from a Lofelt Basslet, which is a wearable consumer device designed to convey the sense of bass while listening to music via headphones. Originally, the device is wrist-worn, houses a rechargeable battery and offers audio connectivity via Bluetooth. Instead, we embed its haptic actuator in a small-footprint 3D printed case and drive it using the mainboard-integrated soundcard for D/A conversion and a Fosi Audio TP-02 subwoofer amplifier. The case is attached to the left index finger of the SenseGlove DK1 hand exoskeleton worn by the operator. As the chunk size processed by the MLP matches the buffer size of 512 samples set on both, the Avatar robot's and the Operator station's audio system running with a sampling frequency of 48, the latency of the entire audio system is 21 (omitting further network transmission delays). §.§ Data Acquisition and Network Training §.§.§ Dataset We recorded a custom dataset using our instrumented robot hand, making contact with and sliding over various textures with multiple patterns and intensities. It consists of a training set with 20 labeled objects, as shown in <ref>, and two test objects <ref>. Both sets feature an equal number of rough and smooth textured objects. As the task description in the Avatar XPRIZE finals was clear in defining the requirement of classifying the texture of stones, we include various natural and artificial stones in the dataset. However, to improve generalization, we also included other common textures like ingrain wallpaper and a wooden table surface. Some objects are measured multiple times under varying acoustical scenarios (handheld, on a table, inside a box, etc.) to improve domain robustness. For each object, we obtain seven recordings with a duration of 30, including light, medium, and strong pressure levels, with long and short strokes, as well as a recording where we apply longer continuous wiggles, to support other interactive sensing approaches w.r.t. the operator. As inference is run on the operator station, we encode all recordings using the OPUS audio codec before training, mimicking transmission effects. §.§.§ Training We split the training data into chunks of 256 and adopt the label of the respective file if the RMS loudness of the chunk exceeds a threshold determining a valid contact, similar to the threshold used to forward or block the oscillator output. As the specific value of this threshold varies between experiments we report it in the evaluation, respectively. This ensures that all labeled chunks correspond to surface contacts, but conversely, not every contact is assigned to a labeled chunk. Without consideration of these unlabeled chunks, the network would show unpredictable behavior at inference time. In the worst case, smooth chunks would be assigned to the rough class, which was particularly undesirable in the context of the competition. Therefore, we introduce a third non-valid class which is comprised of chunks below the set threshold. The network is trained for five epochs with a batch size of 6000 chunks. We use negative log-likelihood loss and the Adam optimizer with a learning rate of 1e-4. On each chunk, we randomly add Gaussian distributed noise to prevent overfitting and enhance generalizability. This is particularly important, as external noises and the capturing circuitry might induce disturbances. Both our dataset and the trained models are made public to allow for reproducibility of results[<https://github.com/AIS-Bonn/Roughness_Sensing>]. § EVALUATION Our system has been developed and evaluated in several steps, focusing on a quantitative analysis of the model training as well as the intuitiveness and immersion in a longer integrated mission. §.§ Quantitative Analysis We compare two model variants that differ in the data used during training. First, we propose a general variant trained using the entire dataset and the threshold for distinguishing contact set to -26, slightly exceeding the noise floor of the piezo microphone. <ref> shows the confusion matrices of the general model variant for both the test set and the competition runs. It is to be pointed out that we explicitly tuned the model to produce low false-positive rates. Due to the vibration motor inside the actuator being slow in its response, w.r.t. prediction rate, and the vibration intensity of rough classification results set relatively high, even misclassifications of single chunks can give the operator the false impression of sensing a rough surface. Accordingly, the correct classification of only several chunks suffices to convey the desired impression when sliding the finger over a rough texture. Second, we show a fine-tuned model variant optimized for participation in the competition, using a reduced set of training objects including samples of the stones encountered during the competition runs (<ref>). Here, we set the threshold for distinguishing contact dynamically to 50 of the RMS loudness of all files with the respective label. This increases the amount of non-valid classification results when applying light pressure onto an object at the benefit of further decreasing the false-positive rate. <ref> compares the classification accuracy of both model variants during the competition runs and for the test set. While the accuracy for classifying smooth objects shows to be very high with both model variants, the fine-tuned variant does substantially exceed the general one but falls slightly shorter w.r.t. rough objects. Furthermore, to justify the usage of the additional MEMS microphone, we show the accuracy of the general model using only the piezo microphone during training and inference. The low accuracy for classifying smooth textures and the associated false-positive rates of 3.3 for the competition runs and 15.5 for the test set are insufficient for our application. §.§ Integrated Mission <ref> shows our avatar robot during the last task in the finals testing event of the ANA Avatar XPRIZE competition. During this task, the operator was required to find and retrieve one of the rough stones, purely based on their haptic perception. In particular, there were five stones lined up on an anti-slip mat in a small box, with an opening that blocked the operator's vision through a curtain. Three of the stones had a smooth texture, while two had a rough texture and were highlighted in pink color, which was only relevant for the audience to distinguish the stones. <ref> shows a close-up of sample textures encountered in the competition. In total, the track was encountered up to three times per team during the event. Each time, a different operator was controlling our avatar robot. The operators were members of the XPRIZE jury and impartial in their judgment of task completion. They were trained for 30min, directly before the run, to familiarize themselves with the system. However, only a fraction of this time was allocated to training for this specific task, as nine previous tasks needed to be completed to advance to the final task. All three encounters were successful. <ref> and <ref> show the measured audio and generated feedback signals of Day 1 & 2, respectively. While the first run on Qualification Day was not public and results are not available, the latter two runs on testing Days 1 & 2 were broadcasted by the organizers, allowing for a comparison with all other teams that completed the task. <ref> shows that we completed the task considerably faster than any other team. However, it is worth mentioning that other factors besides the haptic feedback contributed to these times, such as fitting the hand inside the box and successfully grasping the stone after identification. §.§ Operation without Sight During our qualification run, we also demonstrated operation without direct sight, i.e. the operator was not able to see the scene and the stones through the robot's cameras as usual. In this scenario, it is difficult for the operator to even find the stones in order to touch them. For this reason, we developed a 3D visualization based on geometry captured by a depth camera mounted in the left palm of the robot. It allows the operator to locate the stones, hold them in place using the right hand, and move the instrumented index finger of the left hand over them. We note that the system is designed to not give away the surface roughness: The measured depth data is adaptively downsampled to remove surface details, leaving only the rough geometric shape. Using this depth-based guidance, the operator judge was able to solve the task without live color view, which demonstrates that our haptic feedback alone is sufficient for the classification of the stones. § CONCLUSION We demonstrated a haptic teleperception system consisting of a microphone sensing setup, which is low-cost and very compact, and a simple haptic rendering method on the operator side. The system was proven to be very effective at the ANA Avatar XPRIZE competition finals, winning the first prize. Even though the system was mostly trained and tested on stone surfaces, the method can be adapted easily to other surface kinds by collecting the appropriate training data.

http://arxiv.org/abs/2303.06737v1

20230312194156

Non-Trivial Query Sampling For Efficient Learning To Plan

cs.RO

[ Daodao Yang =============== empty empty In recent years, learning-based approaches have revolutionized motion planning. The data generation process for these methods involves caching a large number of high quality paths for different queries (start, goal pairs) in various environments. Conventionally, a uniform random strategy is used for sampling these queries. However, this leads to inclusion of "trivial paths" in the dataset (e.g.,, straight line paths in case of length-optimal planning), which can be solved efficiently if the planner has access to a steering function. This work proposes a "non-trivial" query sampling procedure to add more complex paths in the dataset. Numerical experiments show that a higher success rate can be attained for neural planners trained on such a non-trivial dataset. § INTRODUCTION Motion planning, a core problem in artificial intelligence and robotics, is one of finding a collision free, low cost path connecting a start and goal state in a search-space. Popular discrete-space planners such as A* <cit.> and LPA* <cit.> conduct a prioritized search using heuristics and guarantee resolution-optimal paths. On the other hand, single-query sampling-based motion planning (SBMP) algorithms, such as RRT <cit.>, solve this problem by constructing a connectivity graph online using a set of probing samples. The RRT algorithm is probabilistically complete, while it's asymptotically optimal variants such as RRT* <cit.>, RRT^# <cit.>, BIT* <cit.>, FMT* <cit.> converge to the optimal solution almost surely, as the number to samples tends to infinity. However, these algorithms may suffer from a slow convergence rate, especially in higher dimensional settings. In order to address this issue, several techniques that leverage heuristics and collision information have been suggested to improve the performance of these planners. These include <cit.>, <cit.>, <cit.>, <cit.>, <cit.> among others. The recently proposed Informed Set <cit.>, <cit.>, <cit.> and Relevant Region <cit.>, <cit.> family of algorithms utilize current solution information and heuristics to focus the search onto a subset of the search-space, while still maintaining the theoretical guarantees of asymptotic optimality. Although the above methods can improve the performance of motion planning algorithms, they require handcrafted heuristics for efficacy. Also, these methods do not leverage prior experience or data gathered from expert demonstrations. Deep learning based approaches for motion planning address these two limitations by generating a dataset of high quality paths in various environments. In the offline phase, this dataset is used to train a deep neural network (DNN) model to predict quantities of interest, such as cost-to-go <cit.>, next point along the optimal path <cit.>, or a sampling distribution <cit.>. The DNN model can then be used online to focus search and dramatically increase the efficiency of planning algorithms. The data generation process for DNN-based motion planning methods involves sampling and solving a set of queries (start, goal pairs) in a given environment. For many applications, the map/environment in which the robot needs to operate may be fixed or given apriori. However, the query sampling distribution can be modified in order to extract more informative paths beneficial to the learning process. Conventionally, uniform random sampling is used to generate the start and goal states. Many queries in this uniformly sampled dataset can be solved by greedily connecting the start and goal state using a steering function (if available). Such a steering function provides the optimal path between any two states after relaxing the collision constraint. This work proposes adding more "non-trivial" queries to the dataset, which cannot be solved by a simple greedy connection. The efficiency of the neural planners can be boosted by training deep models on this dataset comprising of relatively more complex paths. This is demonstrated by creating datasets with different degrees of non-triviality and benchmarking the performance of the neural planner on various robotic planning tasks. § RELATED WORK Exciting progress has been made at the intersection of learning and motion planning in recent years. To improve the performance of discrete space planners, methods that leverage reinforcement learning <cit.> and imitation learning <cit.>, <cit.> have been proposed. Neural A* <cit.> presents a data-driven approach that reformulates conventional A* as a differentiable module. Huh et al <cit.> present a higher-order function network that can predict cost-to-go values to guide the A* search. For sampling-based planning, techniques such as <cit.> and <cit.> learn a local sampling strategy for global planning. Zhang et al <cit.> present a deep learning based rejection sampling scheme for SBMP. Ichter et al <cit.> train a conditional variational autoencoder (CVAE) model to learn sampling distributions and generate samples along the solution paths. NEXT <cit.> learns a local sampling policy and cost-to-go function using "meta self-improving learning". Kuo et al <cit.> input the planner state and local environment information into a deep sequential model to guide search during planning. Approaches such as <cit.>, <cit.> focus on identifying critical or bottleneck states in an environment to generate a sparse graph while planning. While the above techniques may differ in terms of their model architectures or outputs, they do not tackle the problem of improving the data-generation process via query sampling for increasing the efficacy of neural planners. .5em .5em In <cit.>, Huh et al extend their previous work to present a query sampling technique for non-holonomic robots. However, this technique is purely based on the dynamics of the robot, does not utilize obstacle information and is only applicable for car-like robots. The recently proposed OracleNet <cit.> and Motion Planning Networks (MPNet) algorithm <cit.> learn a deep model to recursively predict the next state along the solution path, given a query. The authors in <cit.> use "Active Continual Learning (ACL)" to improve the data-efficiency of the training process. Similar to the DAGGER algorithm <cit.> the ACL process involves training a MPNet deep model on a set of expert demonstrations for a N_c >0 number of initial iterations. The ACL algorithm then finds the cases where the MPNet planner fails and invokes the expert planner only to solve them. The solutions generated by the expert planner are stored in a replay buffer to be used during training. ACL also leverages the "Gradient Episodic Memory (GEM)" technique <cit.> to alleviate the problem of catastrophic forgetting during the learning process. While more data-efficient, the ACL process can be tricky to implement and computationally more expensive, as it involves interleaving the training process with running the neural planner multiple times. The performance of ACL models can also be worse than that of batch-offline models in many cases <cit.>. In contrast, this work proposes a modified query sampling procedure for data-generation and trains all models in a batch-offline manner. § PROBLEM DEFINITION §.§ Path Planning Problem Let 𝒳⊂ℝ^d denote the search-space for the planning problem with dimension d ≥ 2. Let 𝒳_obs⊂𝒳 denote the obstacle space and 𝒳_free= 𝒳∖𝒳_obs denote the free space. Let c_π(x_s,x_g) denote the cost of moving from a point x_s∈𝒳 to x_g∈𝒳 along a path π. This path π can be represented as an ordered list of length L≥2, π={π_1,π_2…π_L}. Let π_end denote the last point on the path π. Then, the optimal path planning problem is one of finding the minimum cost, feasible path π^* connecting a start x_s and goal x_g state. π^*=min_π∈Π c_π(x_s,x_g), subject to: π_1=x_s, π_end=x_g π_i∈𝒳_free,     i = 1,2 … L. Classical planners discretize 𝒳_free or build a connectivity graph and then perform a search over this graph to solve the above planning problem (<ref>). §.§ Supervised Learning for Planning Learning-based methods use the data gathered from successful plans to train models in the offline phase. In the online phase, this learned model can be used to solve (<ref>) or assist the classical planners. The data generation process involves creating an environment and sampling a set of K_train>0 queries or (start, goal) pairs in 𝒳_free. Let 𝒬=𝒳_free×𝒳_free denote the "query-space". A classical planner is then used to solve the path planning problem (<ref>) for each of the K_train queries and obtain a good quality solution. A quantity of interest to be learned as output (such as cost-to-go, next state on the optimal path etc) is extracted from these solution paths. The objective of the training process is to learn a function f_θ (usually a deep neural network), on the dataset 𝒟 by minimizing an empirical loss with respect to weight parameters θ §.§ Neural Planner The MPNet procedure involves learning a planning network 𝖯𝖭𝖾𝗍, that predicts the next state on the optimal path, given a current state, a goal state and environment information as the input. A typical neural planning algorithm, in line with the one described in <cit.>, is illustrated in Algorithm <ref>. Given a query x_s,x_g, the path π to be returned is initialized with the start-state. At each iteration, the neural planner attempts a greedy connection to the goal using the 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 function. For length-optimal or geometric planning case, the 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 connects any two points using a straight line. For car-like robots, this steering function can use the Dubins curves for dynamically feasible connections <cit.>. To probe the feasibility of this greedy connection, the 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 procedure discretizes the path and collision-checks the points on it. If the greedy connection π_end to x_g is valid, the goal-state is appended to the path and the planning loop terminates. Else, the learned planning network 𝖯𝖭𝖾𝗍, is used to predict the next state on the optimal path. If the path π is infeasible, a neural replanning procedure is performed on this coarse path in an attempt to repair it. Please see <cit.> for more details about these procedures. § NON-TRIVIAL QUERY SAMPLING .5em .5em §.§ Non-trivial Queries Conventionally, uniform random sampling is used in the data generation process to obtain a query x_s,x_g∈𝒬. However, this may result in the inclusion of "trivial" queries in the dataset, for which 𝗌𝗍𝖾𝖾𝗋𝖳𝗈(x_s,x_g)=𝖳𝗋𝗎𝖾. In case of such trivial queries, the neural planning Algorithm <ref> terminates in the first iteration after processing lines 4-6. Thus, a key observation is that 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 procedure in Algorithm <ref>, line 4 performs an implicit classification of queries, so that only "non-trivial" queries are passed over to the 𝖯𝖭𝖾𝗍. Concretely, the set of non-trivial queries can be defined as 𝒬_nt={x_s,x_g∈𝒬 | 𝗌𝗍𝖾𝖾𝗋𝖳𝗈(x_s,x_g)=𝖥𝖺𝗅𝗌𝖾}. This motivates the proposed data generation Algorithm <ref>, which aims to increase the number of non-trivial data samples in 𝒟. After an environment 𝒳, 𝒳_obs is created, data is generated by solving a total of K_train queries. With probability p_nt, the proposed 𝗇𝗈𝗇𝖳𝗋𝗂𝗏𝗂𝖺𝗅𝖰𝗎𝖾𝗋𝗒𝖲𝖺𝗆𝗉𝗅𝗂𝗇𝗀 procedure is used to obtain x_s,x_g∈𝒬_nt. Else, conventional 𝗎𝗇𝗂𝖿𝗈𝗋𝗆𝖲𝖺𝗆𝗉𝗅𝗂𝗇𝗀 returns a query in 𝒬. A classical planner such as A* or BIT* then solves this query and outputs a good quality solution path π. Samples from this path π are appended to the dataset 𝒟 with the proposed data inclusion procedure 𝗂𝗇𝖼𝗅𝗎𝖽𝖾𝖣𝖺𝗍𝖺. §.§ Non-trivial Query Sampling A rejection sampling algorithm to generate new queries in 𝒬_nt is given in Algorithm <ref>. For N_nt number of attempts, uniform sampling is used to first generate a valid query x_s,x_g∈𝒬. The 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 module then validates the connection between start and goal state. If found invalid, the corresponding non-trivial query is returned. Thus, this procedure intends to filter out trivial paths while maintaining the exploratory/coverage property of uniform sampling. Please see Fig. <ref> and Fig. <ref> for a visualization of queries generated using the proposed non-trivial sampling procedure. §.§ Data Inclusion The data inclusion Algorithm <ref> iterates over the segments of path π and logs the current state, goal state (π_i,π_end) as the input and the next state (π_i+1) as the output/label. However, if the flag 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖳𝗋𝗎𝖾, the algorithm skips including the data-sample {(π_i,π_end),π_i+1} if π_i,π_end∉𝒬_nt. Thus, 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖳𝗋𝗎𝖾 ensures that only the non-trival segments of π are incorporated in 𝒟. Please see Fig. <ref> for an illustration of this step. Depending on the topology of 𝒳_obs, the neural planner may find it relatively harder to predict feasible paths in certain environments. The notion of non-trivial queries can be used to define a metric that captures this level of difficulty. Consider a "non-triviality ratio", which can be defined as, γ_nt=# Non-trivial queries/# Uniformly sampled queries . Thus, γ_nt is the ratio of number of non-trivial queries found in a (large enough) set of uniformly sampled queries. This ratio will be high for complex, cluttered and narrow-passage type environments and low for relatively simpler, single-obstacle type environments. .5em .5em .5em .5em § NUMERICAL EXPERIMENTS In order to benchmark the proposed data generation algorithm, the following procedure was implemented for each planning environment. First, four datasets with different parameter settings were created. These were as follows: 𝒟_0  ( p_nt=0,  𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖥𝖺𝗅𝗌𝖾), 𝒟_1  ( p_nt=0.5, 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖥𝖺𝗅𝗌𝖾), 𝒟_2  ( p_nt=1.0, 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖥𝖺𝗅𝗌𝖾), 𝒟_3  ( p_nt=1.0, 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺=𝖳𝗋𝗎𝖾). Thus, 𝒟_0 represents the dataset generated using the conventional uniform query sampling, whereas 𝒟_3 is created using the proposed non-trivial query sampling and data pruning procedure. Four deep models, 𝖯𝖭𝖾𝗍_0,𝖯𝖭𝖾𝗍_1,𝖯𝖭𝖾𝗍_2,𝖯𝖭𝖾𝗍_3 were then trained on their respective datasets. The neural network shape, size and the training parameters were held constant while learning all four models. Performance of the neural planner <ref> using these four models was evaluated on 1) K_test number of new uniform queries and 2) K_test number of new non-trivial queries. Two performance metrics, namely, success ratio and cost ratio were considered. Success ratio gives the number of times (out of K_test in total) the neural planner was successful in finding a feasible (collision-free) solution. Cost ratio denotes the ratio of the neural planner's solution cost to that of classical planner, averaged over K_test trials. Model training and evaluation was performed using the Python PyTorch API on a 64 bit, 16 GB RAM laptop with Intel i7 processor and a NVIDIA GeForce RTX 2060 GPU. A description of robotic planning tasks along with a discussion of results is given below. Point Robot: Four different 20 × 20 environments, illustrated in Fig. <ref>, were considered for the case of point robot planning. Four datasets {𝒟_i }^3_i=0, as described above, were generated for each environment. A total of K_train=3000 number of queries were sampled for each dataset. An A* planner, followed by post-processing/smoothening, was used to solve these queries and obtain length-optimal paths. A small padding of 0.8 units around the obstacles was propagated during the data generation step, and was relaxed during the final performance evaluation step. This was found to greatly boost the success ratio of the neural planner, while making slight compromise in the cost ratio metric. The performance metrics were logged by solving K_test=500 number of unseen uniform and non-trivial queries with the four learned 𝖯𝖭𝖾𝗍 models. Rigid Body Planning: Fig. <ref> shows the instance of planning for a rigid robot in four 10 × 10 environments. A total of K_train=5000 queries were considered to create each of the four datasets {𝒟_i }^3_i=0. All the queries were solved in the SE(2) space using OMPL's <cit.> implementation of the BIT* planner. An obstacle-padding of 0.4 units was propagated during the data-generation phase. The learned 𝖯𝖭𝖾𝗍 models predicted a three dimensional [x,y,θ] vector representing the robot's pose. These models were evaluated on K_test=500 unseen uniform and non-trivial queries. The BIT* planner was allowed a run-time of 3 seconds during the data-generation phase and 2 seconds during the evaluation phase. n-link Manipulator Planning: To observe performance of the neural planner in higher dimensions, a planning problem for 2,4 and 6-link manipulator robot was considered. Please see Fig. <ref>. The joint angles were constrained to lie between -π and π. Four datasets {𝒟_i }^3_i=0 were created for each of the 2,4 and 6-link case by considering a total of 3000,4000 and 5000 queries respectively. These queries were solved using OMPL's BIT* planner with an padding of 0.8 units around the workspace obstacles. The final performance evaluation was done by solving K_test=500 new queries with the neural planner. The BIT* planner was run for 2,4,6 seconds during the data-generation stage and 1,2,4 seconds during the evaluation stage for the 2,4 and 6-link planning respectively. General trends seen from the results in Table <ref>, <ref>, <ref> are as follows. In most cases, 𝖯𝖭𝖾𝗍_2 and 𝖯𝖭𝖾𝗍_3 outperform 𝖯𝖭𝖾𝗍_0 in terms of success ratio, where as the performance of 𝖯𝖭𝖾𝗍_1 is more sporadic. The cost ratio is relatively lower for the rigid body and 6-link case compared to others, as the BIT* planner may not find a good quality solution in the given planning time for these challenging cases. The success ratio for all 𝖯𝖭𝖾𝗍 models is naturally higher over uniform test queries rather than non-trivial test queries. The success ratio also generally has an inverse relation with γ_nt, as seen strongly in the case of rigid body and n-link manipulator planning. For the case of point robot planning, all models perform well with a success rate of over 90 % (see Table <ref>). However, slight performance gains due to the proposed method can be seen for Environments 0,2,3. These gains are much more noticeable for the rigid body planning case (see Table <ref>). The 𝖯𝖭𝖾𝗍_3 model has the highest success ratio in all cases except one (Environment 3, Non-trivial Query), where its performance is comparable to 𝖯𝖭𝖾𝗍_1. The gradation in performance due to dimensionality and γ_nt can be seen clearly in the n-link planning case (see Table <ref>). The success ratio over uniform queries is in the range of 0.9,0.8 and 0.7 for the case of 2,4 and 6-link planning case respectively. For the relatively simpler 2-link planning case with γ_nt=0.225, only small gains in the success ratio over non-trivial queries can be seen. However, the improvement in performance is much more evident for the higher dimensional 4 and 6-link cases. The 𝖯𝖭𝖾𝗍_3 model shows about a 25% increase in the success ratio over 𝖯𝖭𝖾𝗍_0 for the 6-link (non-trivial queries) case. The neural planning Algorithm <ref> and the corresponding results discussed above assume the availability of a steering function. While this is readily available for cases such as geometric or non-holonomic (car-like) planning <cit.>, it may not be computationally tractable for others. To analyze the performance of 𝖯𝖭𝖾𝗍 models without the 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 function, simulations were performed by only executing the lines 8 and 9 of the neural planner <ref> for maximum N_plan iterations. Instead of lines 4-6 in Algorithm <ref>, the following termination condition was implemented, π_end - x_g_2 ≤δ, with a small δ>0. As illustrated in Fig. <ref>, the 𝖯𝖭𝖾𝗍_3 model, which has no trivial sample in its training dataset, naturally cannot solve a trivial query. Numerical results for the rigid body planning without the 𝗌𝗍𝖾𝖾𝗋𝖳𝗈 function and δ=1.0 are tabulated in Table <ref>. The success ratio of all 𝖯𝖭𝖾𝗍 models is adversely affected in this case. The 𝖯𝖭𝖾𝗍_0 model performs best on uniform queries in all environments, whereas the performance of 𝖯𝖭𝖾𝗍_3 is the worst. However, 𝖯𝖭𝖾𝗍_1 or 𝖯𝖭𝖾𝗍_2 show better performance over non-trivial queries in some cases. Thus, without a steering function, a uniformly sampled training dataset might be the best choice if the test queries are uniformly distributed too. However, a model trained over a dataset with an appropriate value of p_nt may perform better over non-trivial test queries. This makes a case for an ensemble model. § CONCLUSION Many of the previous techniques in the literature have focused on exploring different deep architectures for planning, while using a uniformly sampled dataset for training. This work, on the other hand, investigates the problem of improving the data-generation process while holding the model architecture and planning algorithm constant. The proposed query sampling and data pruning procedures add more complicated paths in the dataset. Numerical experiments show that the success rate of the neural planner can be boosted using the deep models trained on such non-trivial datasets. This work presents many opportunities for future research. An ensemble model can be constructed by combining predictions from different models trained on datasets with varying degrees of non-triviality. Instead of a Boolean 𝗉𝗋𝗎𝗇𝖾𝖣𝖺𝗍𝖺 flag, calling the pruning procedure with a probability of γ_nt can be explored. This can prevent excessive pruning and result in a drastic reduction in the size of the dataset for relatively less cluttered environments. Acknowledgements: Authors would like to thank Prof. Le Song, Binghong Chen and Ethan Wang for insightful discussions on this topic. IEEEtran

http://arxiv.org/abs/2303.07258v1

20230313163253

Structures of optimal discrete gradient vector fields on surface with one or two critical cells

math.DS

Ab initio electron-lattice downfolding: potential energy landscapes, anharmonicity, and molecular dynamics in charge density wave materials Tim O. Wehling March 30, 2023 =========================================================================================================================================== We describe all possible structures of discrete vector field (discrete Morse functions) with minimal number of critical cells on the regular CW-complex for the 2-disk (1 cell), the 2-sphere (2 cells), the cylinder (2 cells) and Mobius band (2 cells). § INTRODUCTION Topological graph theory is useful for describing the topological structure of functions and flows on surfaces. These structures are given by discrete structures – graphs with additional information. These graphs are often embedded in a surface, the resulting invariants are discrete Morse functions. There is a natural equivalence on the set of such functions, first explored in the papers of Forman <cit.>. At the same time, the equivalence class is given by a discrete gradient vector field. In the most general case, such fields and functions are considered on regular cellular complexes. The most useful discrete functions and vector fields are those that have the smallest number of critical cells in a given space, among all possible functions and vector fields on the regular cellular decomposition with minimal number of cells. Such fields are called optimal. Therefore, it is important to study the structure of such fields. Graphs as topological invariants of functions were used in the papers of Kronrod <cit.> and Reeb <cit.> for oriented maniofolds, in <cit.> for non-orientable two-dimensional manifolds and in <cit.> for manifolds with boundary, in <cit.> for non-compact manifolds. In general, Morse vector fields (Morse-Smale vector fields without closed orbits) are gradient field of Morse functions. If we fix the value of functions in singular points the field determinate the topological structure of the function <cit.>. Therefore, Morse–Smale vector fields classification is closely related to the classification of the functions. Topological classification of smooth function on closed 2-manifolds was also investigated in <cit.>, on 2-manifolds with the boundary in <cit.> and on closed 3-manifolds in <cit.>. In <cit.>, the classifications of flows on closed 2- manifolds and <cit.> on manifolds with the boundary were obtained. Topological properties of Morse-Smale vector fields on 3-manifolds was investigated in <cit.>. The main invariants of graphs and their embeddings in surfaces can be founded in <cit.>. The main purpose of this paper is to study the topological properties of the optimal discrete gradient vector fields on surface with one or two critical cells. These fields have and minimal number of critical cell on the regular cellular complex with minimal number of cells. The first section gives the basic definitions and topological properties of discrete Morse functions and discrete gradient vector fields. In the second section we describe all possible structures of optimal discrete vector flows on a two-dimensional disk, in the third section – on a two-dimensional sphere, in the fourth section – on the cylinder, and in the fifth section – on a Mobius strip. § DISCRETE MORSE FUNCTIONS AND VECTOR FIELD If a cell (or simplicial) complex K is given, then a discrete function is a mapping that matches each simplex with a real number. If the structure of a cellular complex (triangulation) is given on the manifold M, then for the function f: M →ℝ we take the value in the center of each cell (simplex). Thus, we construct a discrete Morse function. On the contrary, if a Morse function is given on a triangulated manifold, then it determines the value of the function at the center of each simplex. Consider the first barycentric sub-division. For it, we have the value of the function at each vertex. By linearity, we extend the function to each simplex, thereby obtaining a piecewise linear function on the manifold. With the help of smoothing, you can get a smooth function from it. Similar constructions can be made for regular cage complexes. Forman approach based on the use of discrete gradient flows was proposed in <cit.>.These designs are easier to build algorithms and implement on computers. On the set of simplexes (cells) of the K complex, we introduce the order relation: τ < σ⇔τ⊂σ, τσ. By |A| we denote the number of elements of the set A. The function f: K →ℝ is called a discrete Morse function if for an arbitrary simplex (cell) σ of dimension k: |τ^k-1: τ^k-1<σ, f(τ^k-1) ≥ f(σ)| + |τ^k+1: τ^k+1>σ, f(τ^k+1) ≤ f(σ)| ≤ 1. |τ^k-1: τ^k-1<σ, f(τ^k-1) ≥ f(σ)| + |τ^k+1: τ^k+1>σ, f(τ^k+1) ≤ f(σ)| = 0. Simplexes that are not critical are called regular. A discrete Morse function is called simple if it takes different values on different critical simplexes (cells). A discrete vector field on the complex K is the set of pairs V={ (σ, τ) } for which σ, τ∈ K, σ <τ, dim σ = τ +1 and each simplex (cell) of the complex can belong to no more than one pair. Each such pair is called a vector and is represented by an arrow directed from the center of the first simplex (cell) to the center of the second. The gradient vector field V_f of the discrete Morse function is called the field V_f={σ^(k), τ^(k+1): σ^(k)<τ^(k+1), f( σ^(k)) ≥ f( τ^(k+1)). In fig. <ref> shows the discrete Morse function on the torus and its gradient field. Critical cells are: 1) vertex 0, 2) edges {1,2} and {3,4}, 3) 2- cell [5,6,8,7]. The index of a critical cell (simplex) is its dimension. The gradient path of a discrete vector field V is called a sequence of cells (simplexes) τ_0, σ_1, τ_1, …, σ_m-1 , τ_m-1,σ_m such that (σ_i, τ_i)∈ V, σ_i< τ_i-1, 1 ≤ i ≤ m, σ_0=σ_m. Examples of gradient paths on <ref> b): 1) two paths from {1,2} to {0} – {1,2},{1} ,{0,1},{0} and {1,2},{2},{0,2},{0}; 2) two paths from [5,6,8,7] to {1,2} – [5,6,8,7],{5,6},[1, 2,6,5],{1,2} and [5,6,8,7], {5,6}, [1,2,6,5],{2,6 }, [2,0,3,6],{3,6},[6,3,4,8],{8,4},[8,4,0,2 ],{2,8},[8,2,1,7],{2,1}. τ < σ⇔τ⊂σ, τσ. Two discrete vector field are called isomorphic if there is an isomorphism of the CW-complexes which preserve the pairs. Our aim is to describe all possible up to isomorphism structures of flows with one or two critical cell. § DISCRETE VECTOR FIELDS ON THE 2-DISK Since a 2-disk is homotopically equivalent to a point, any discrete gradient vector field on it will have a critical 0-cell. The optimal cellular partition of the 2-disk is a triangle - 3 vertices, 3 edges and one face (it cannot be less, since there must be at least three vertices and three edges on the boundary of the surface to fulfill the definition of a regular cellular complex). To specify discrete vector fields on a 2-disk, we use the numbering and notation of cells, as in Fig. <ref>. Since the 2-cell is not critical, it belongs to one pair. Without loss of generality, we can assume that this is a pair of cA. Note that there is an axial symmetry that preserves this pair. Taking it into account, there are two possibilities for choosing a critical 0-cell: 0 or 1. In the first case, we will find such a vector field: V_1={1a, 2b, cA}. In the second case, the vector field is V_2={0a, 2b, cA}. Thus, the following is true. The number of non-isomorphic optimal discrete vector fields on the 2-disk is equal 2. § DISCRETE VECTOR FIELDS ON THE 2-SPHERE According to the discrete Morse theory, an optimal discrete gradient vector field on a 2-sphere has one critical 0-cell and one critical 2-cell. An optimal regular CW-complex of a sphere has four 0-cells, six 1-cells, and four 2-cells. Let us denote the cells of the regular cell partition as shown in Fig. <ref>. Without loss of generality, we assume that the critical 2-cell is cell B. If none of the 1-cells (e, d, f) bordering it are in pairs with 2-cells, then they form a cycle, which is impossible. Thus, three cases are possible: 1) one of these cells is paired with a 2-cell (eA), 2) two pairs (eA and fC), 3) three pairs (eA, fC and dD). In the first case, there are 2 variant: V_1-3={…, eA, aD, cC}, V_4-7={…, eA, aD, bC}. By V_1-3 we denote next three vector fields: In the first variant it is possible to select the critical vertex in 3 ways (due to symmetry): V_1={2b,1d,3f, eA, aD, cC}, V_2={0b,2d,3f, eA, aD, cC}, V_3={0b,1d,3f, eA, aD, cC}. In the second variant we have 4 vector fields: V_4={1d,2f,3c, eA, aD, bC}, V_5={0c,2d,3f, eA, aD, bC}, V_6={0c,1d,3f, eA, aD, bC}, V_7={0c,1d,2f, eA, aD, bC}. In the second case V_8-11={…, eA, fC, aD} there are four ways for critical vertex location. In the third case V_12-13={…, eA, dD, fC} there are two ways for critical vertex location. Thus, the following is true. The number of non-isomorphic optimal discrete vector fields on the 2-sphere is equal 13. § DISCRETE VECTOR FIELDS ON THE CYLINDER The optimal structure of the cell complex on a cylinder consists of three 2-cells, quadrilaterals, glued in a circle. All vertices in it are equal. Let's assign number 0 to the critical vertex. We cut the cylinder along the glue, ending at this vertex. we get a rectangle divided into three squares. We number its cells as in Fig. <ref>. Note that an optimal vector field has one critical vertex and a critical edge. There is one non-trivial isomorphism of the cell complex onto itself, which is fixed at the critical vertex. On fig. 3 it corresponds to axial symmetry about the vertical axis. The vector field is determined by the pairs that contain the 2-cell and the location of the critical 1-cell. The following options are possible: V_1-3={…, aA, bB, iC}, V_4-6={…, aA, bB, fC}, V_7-9={…, gA, bB, aC}, V_10-12={…, gA, bB, iC}, V_13-15={…, gA, bB, cC}, V_16-20={…, gA, bB, fC}, V_21-23={…, dA, bB, aC}, V_24-28={…, dA, bB, iC}, V_29-31={…, dA, bB, cC}, V_32-34={…, dA, bB, fC}, V_35-39={…, gA, eB, aC}, V_40-44={…, gA, eB, iC}, V_44-49={…, gA, eB, cC}, V_50-54={…, gA, eB, fC}, V_55-57={…, aA, eB, cC}, V_58-60={…, aA, eB, fC}, V_61-63={…, bA, eB, cC}, V_64-66={…, bA, eB, fC}, V_67-69={…, dA, eB, fC}, V_70-72={…, gA, hB, aC}, V_73-75={…, gA, hB, iC}, V_76-78={…, gA, hB, cC}, V_79-83={…, gA, hB, fC}, V_84-86={…, aA, hB, cC}, V_87-91={…, aA, hB, fC}, V_92-94={…, bA, hB, cC}, V_95-99={…, bA, hB, fC}, V_100-104={…, dA, hB, fC}. Thus, the following is true. The number of non-isomorphic optimal discrete vector fields on the cylinder S^1× [0,1] is equal 104. § DISCRETE VECTOR FIELDS ON THE MOBIUS BAND The case with the Mobius band is similar to the case with the cylinder. The axial symmetry of the rectangle is replaced by central symmetry. Cell designations are shown in fig. 4. The vector field is determined by the pairs that contain the 2-cell and the location of the critical 1-cell. The following options are possible: V_1-4={…, aA, bB, iC}, V_5-8={…, aA, bB, fC}, V_9-12={…, gA, bB, aC}, V_13-16={…, gA, bB, iC}, V_17-20={…, gA, bB, cC}, V_21-24={…, gA, bB, fC}, V_25-28={…, dA, bB, aC}, V_29-32={…, dA, bB, iC}, V_33-36={…, dA, bB, cC}, V_37-40={…, dA, bB, fC}, V_41-44={…, gA, eB, aC}, V_45-50={…, gA, eB, iC}, V_51-54={…, gA, eB, cC}, V_55-58={…, gA, eB, fC}, V_59-62={…, aA, eB, cC}, V_63-66={…, aA, eB, iC}, V_58-70={…, aA, eB, fC}, V_71-74={…, bA, eB, cC}, V_75-78={…, bA, eB, iC}, V_79-82={…, bA, eB, aC}, V_83-86={…, bA, eB, fC}, V_87-90={…, dA, eB, aC}, V_91-94={…, dA, eB, iC }, V_95-98={…, dA, eB, fC}, V_99-102={…, dA, eB, cC}. Thus, the following is true. The number of non-isomorphic optimal discrete vector fields on the cylinder S^1× [0,1] is equal 102. § CONCLUSION The results of the article prove the effectiveness of the proposed invariants (distinguishing graph and code) for classifying flows on a two-dimensional disk with one singular point. We hope that they can be used to construct similar invariants for flows with a large number of singular points on other surfaces as well. 10 akchurin2022three O. Akchurin, S. Bilun, and A. Prishlyak. Three-color graph as the 1-skeleton of the 2-sphere triangulation. arXiv preprint arXiv:2209.05737, 2022. http://dx.doi.org/10.48550/ARXIV.2209.05737 doi: 10.48550/ARXIV.2209.05737. bilun2002closed S. Bilun and A. Prishlyak. The closed morse 1-forms on closed surfaces. Visn., Mat. Mekh., Kyv. Univ. Im. Tarasa Shevchenka, 2002(8):77–81, 2002. bilun2022visualization S. Bilun and A. Prishlyak. Visualization of morse flow with two saddles on 3-sphere diagrams. arXiv preprint arXiv:2209.12174, 2022. http://dx.doi.org/10.48550/ARXIV.2209.12174doi: 10.48550/ARXIV.2209.12174. bilun2022morse S. Bilun, A. Prishlyak, and A. Prus. Morse flows with fixed points on the boundary of 3-manifold. arXiv preprint arXiv:2209.04019, 2022. http://dx.doi.org/10.48550/arXiv.2209.04019 doi: 10.48550/arXiv.2209.04019. Bolsinov2004 A.V. Bolsinov and A.T. Fomenko. Integrable Hamiltonian systems. Geometry, Topology, Classification. A CRC Press Company, Boca Raton London New York Washington, D.C., 2004. 724 p. for98 R. Forman. Morse theory for cell complexes. Advances in Mathematics. 134 (1): 90–145. 1998. http://dx.doi.org/10.1006/aima.1997.1650doi: 10.1006/aima.1997.1650. GT87 J. Gross and T. Tucker. Topological graph theory. Wiley, 1987. Harary69 F. Harary. Graph Theory. Westview Press, 1969. hatamian2020heegaard Ch. Hatamian and A. Prishlyak. Heegaard diagrams and optimal morse flows on non-orientable 3-manifolds of genus 1 and genus 2. Proceedings of the International Geometry Center, 13(3):33–48, 2020. http://dx.doi.org/10.15673/tmgc.v13i3.1779doi: 10.15673/tmgc.v13i3.1779. HW68 P. Hilton and S. Wylie. Homology theory: An introduction to algebraic topology. Cambridge University Press, 1968. hladysh2016functions B. I. Hladysh and A. O. Pryshlyak. Functions with nondegenerate critical points on the boundary of the surface. Ukrainian Mathematical Journal, 68(1):29–41, 2016. http://dx.doi.org/10.1007/s11253-016-1206-5doi: 10.1007/s11253-016-1206-5. hladysh2017topology B.I. Hladysh and A.O. Prishlyak. Topology of functions with isolated critical points on the boundary of a 2-dimensional manifold. SIGMA. Symmetry, Integrability and Geometry: Methods and Applications, 13:050, 2017. http://dx.doi.org/0.3842/SIGMA.2017.050doi: 0.3842/SIGMA.2017.050. hladysh2019simple B.I. Hladysh and A.O. Prishlyak. Simple morse functions on an oriented surface with boundary. Журнал математической физики, анализа, геометрии, 15(3):354–368, 2019. http://dx.doi.org/10.15407/mag15.03.354doi: 10.15407/mag15.03.354. Kronrod1950 A.S. Kronrod. Functions of two variables. Russian Mathematical Surveys, 5:24–134, 1950. Kybalko2018 Z. Kybalko, A. Prishlyak, and R. Shchurko. Trajectory equivalence of optimal Morse flows on closed surfaces. Proc. Int. Geom. Cent., 11(1):12–26, 2018. http://dx.doi.org/10.15673/tmgc.v11i1.916 doi: 10.15673/tmgc.v11i1.916. loseva2016topology M. Losieva and A. Prishlyak. Topology of morse–smale flows with singularities on the boundary of a two-dimensional disk. Pr. Mizhnar. Heometr. Tsentr, 9(2):32–41, 2016. http://dx.doi.org/10.15673/tmgc.v9i2.279doi: 10.15673/tmgc.v9i2.279. lychak2009morse D.P. Lychak and A.O. Prishlyak. Morse functions and flows on nonorientable surfaces. Methods of Functional Analysis and Topology, 15(03):251–258, 2009. Oshemkov1998 A.A. Oshemkov and V.V. Sharko. Classication of morse-smale flows on two-dimensional manifolds. Matem. Sbornik, 189(8):93–140, 1998. Peixoto1973 M.M. Peixoto. On the classication of flows of 2-manifolds. Dynamical Systems (Proc. Symp. Univ. of Bahia, Salvador, Brasil, 1971), pages 389–419, 1973. prishlyak2017morse A. Prislyak and A. Prus. Morse-smale flows on torus with hole. Proc. Int. Geom. Cent., 10(1):47–58, 2017. http://dx.doi.org/10.15673/tmgc.v1i10.549doi: 10.15673/tmgc.v1i10.549. prishlyak1997graphs A.O. Prishlyak. On graphs embedded in a surface. Russian Mathematical Surveys, 52(4):844, 1997. http://dx.doi.org/10.1070/RM1997v052n04ABEH002074doi: 10.1070/RM1997v052n04ABEH002074. prishlyak2002ms A.O. Prishlyak. Morse–smale vector fields without closed trajectories on-manifolds. Mathematical Notes, 71(1-2):230–235, 2002. http://dx.doi.org/10.1023/A:1013963315626doi: 10.1023/A:1013963315626. prishlyak2003sum A.O. Prishlyak. On sum of indices of flow with isolated fixed points on a stratified sets. Zhurnal Matematicheskoi Fiziki, Analiza, Geometrii [Journal of Mathematical Physics, Analysis, Geometry], 10(1):106–115, 2003. prishlyak2005complete A.O. Prishlyak. Complete topological invariants of morse–smale flows and handle decompositions of 3-manifolds. Fundamentalnaya i Prikladnaya Matematika, 11(4):185–196, 2005. prishlyak2007complete A.O. Prishlyak. Complete topological invariants of morse-smale flows and handle decompositions of 3-manifolds. Journal of Mathematical Sciences, 144:4492–4499, 2007. prishlyak2022topological A. Prishlyak and M. Loseva. Topological structure of optimal flows on the girl's surface. Proceedings of the International Geometry Center, 15(3-4):184–202, 2022. prishlyak2021flows A. Prishlyak, A. Prus, and S. Guraka. Flows with collective dynamics on a sphere. Proc. Int. Geom. Cent, 14(1):61–80, 2021. http://dx.doi.org/10.15673/tmgc.v14i1.1902doi: 10.15673/tmgc.v14i1.1902. prishlyak2020topology A. Prishlyak and M. Loseva. Topology of optimal flows with collective dynamics on closed orientable surfaces. Proceedings of the International Geometry Center, 13(2):50–67, 2020. http://dx.doi.org/10.15673/tmgc.v13i2.1731doi: 10.15673/tmgc.v13i2.1731. prishlyak1999equivalence A.O. Prishlyak. Equivalence of morse function on 3-manifolds. Methods of Func. Ann. and Topology, 5(3):49–53, 1999. prishlyak2000conjugacy A.O. Prishlyak. Conjugacy of morse functions on surfaces with values on a straight line and circle. Ukrainian Mathematical Journal, 52(10):1623–1627, 2000. http://dx.doi.org/10.1023/A:1010461319703doi: 10.1023/A:1010461319703. prishlyak2002morse A.O. Prishlyak. Morse functions with finite number of singularities on a plane. Meth. Funct. Anal. Topol, 8:75–78, 2002. Prishlyak2002beh2 A.O. Prishlyak. Topological equivalence of morse–smale vector fields with beh2 on three-dimensional manifolds. Ukrainian Mathematical Journal, 54(4):603–612, 2002. prishlyak2003topological A.O. Prishlyak. Topological classification of m-fields on two-and three-dimensional manifolds with boundary. Ukrainian Mathematical Journal, 55(6):966–973, 2003. prishlyak2019optimal A.O. Prishlyak and M.V. Loseva. Optimal morse–smale flows with singularities on the boundary of a surface. Journal of Mathematical Sciences, 243:279–286, 2019. prishlyak2007classification A.O. Prishlyak and K.I. Mischenko. Classification of noncompact surfaces with boundary. Methods of Functional Analysis and Topology, 13(01):62–66, 2007. prishlyak2020three A.O. Prishlyak and A.A. Prus. Three-color graph of the morse flow on a compact surface with boundary. Journal of Mathematical Sciences, 249(4):661–672, 2020. http://dx.doi.org/10.1007/s10958-020-04964-1doi: 10.1007/s10958-020-04964-1. Reeb1946 G. Reeb. Sur les points singuliers d’une forme de pfaff complétement intégrable ou d’une fonction numérique. C.R.A.S. Paris, 222:847—849, 1946. Sharko1993 V.V. Sharko. Functions on manifolds. Algebraic and topological aspects., volume 131 of Translations of Mathematical Monographs. American Mathematical Society, Providence, RI, 1993. Smale1961 S. Smale. On gradient dynamical systems. Ann. of Math., 74:199–206, 1961. kkp2013 В.М. Кузаконь, В.Ф. Кириченко, and О.О. Пришляк. Гладкi многовиди. Геометричнi та топологiчнi аспекти. Працi Iнcтитуту математики НАН України.—2013.—97.—500 с, 2013. prish1998vek А.О. Пришляк. Векторные поля Морса–Смейла с конечным числом особых траекторий на трехмерных многообразиях. Доповiдi НАН України, (6):43–47, 1998. prish1998sopr А.О. Пришляк. Сопряженность функций Морса. Некоторые вопросы совр. математики. Институт математики АН Украины, Киев, 1998. prish2001top А.О. Пришляк. Топологическая эквивалентность функций и векторных полей Морса—Смейла на трёхмерных многообразиях. Топология и геометрия. Труды Украинского мат. конгресса, pages 29–38, 2001. prish2002vek А.О. Пришляк. Векторные поля Морса–Смейла без замкнутых траекторий на трехмерных многообразиях. Математические заметки, 71(2):254–260, 2002. prish2015top О.О. Пришляк. Топологія многовидів. Київський університет, 2015. Taras Shevchenko National University of Kyiv Svitlana Bilun Email address: bilun@knu.ua Orcid ID: 0000-0003-2925-5392 Maria Hrechko Email: mariyagr2808@gmail.com Orcid ID: 0009-0005-5347-0342 Olena Myshnova Email: myshnova.olena@gmail.com Orcid ID: 0009-0001-8808-0038 Alexandr Prishlyak Email address: prishlyak@knu.ua Orcid ID: 0000-0002-7164-807X

http://arxiv.org/abs/2303.07382v1

20230313180144

Waveguide QED with Quadratic Light-Matter Interactions

quant-ph

red

http://arxiv.org/abs/2303.07169v2

20230313151230

Dynamic Event-based Optical Identification and Communication

cs.CV

Validation of uncertainty quantification metrics: a primer Pascal PERNOT 0000-0001-8586-6222 March 30, 2023 =========================================================== Optical identification is often done with spatial or temporal visual pattern recognition and localization. Temporal pattern recognition, depending on the technology, involves a trade-off between communication frequency, range and accurate tracking. We propose a solution with light-emitting beacons that improves this trade-off by exploiting fast event-based cameras and, for tracking, sparse neuromorphic optical flow computed with spiking neurons. In an asset monitoring use case, we demonstrate that the system, embedded in a simulated drone, is robust to relative movements and enables simultaneous communication with, and tracking of, multiple moving beacons. Finally, in a hardware lab prototype, we achieve state-of-the-art optical camera communication frequencies in the kHz magnitude. Neuromorphic Computing, Event-Based Sensing, Optical Camera Communication, Optical Flow § INTRODUCTION Identifying and tracking objects in a visual scene has many applications in sports analysis, swarm robotics, urban traffic, smart cities and asset monitoring. Wireless solutions have been widely used for object identification, such as radio frequency identification <cit.> or more recently Ultra Wide Band <cit.>, but these do not provide direct localization and require meshes of anchors and additional processing. One efficient solution is to use a camera to detect specific visual patterns attached to the objects. However, spatial pattern detection has a limited range depending on the camera's spatial resolution. With the rise of fast cameras, temporal light patterns can be used instead, where the detection range depends only on the light intensity of emitting beacons. This technique is known as Optical Camera Communication (OCC) and has been developed primarily for communication between static objects <cit.>. In applications such as asset monitoring on a construction site, it is important to track dynamically moving objects. OCC techniques potentially enable simultaneous communication with, and tracking of, beacons. However, two challenges arise in the presence of relative movements: filtering out the noise and tracking the beacons' positions. Increasing the temporal frequency of the transmitted signal, since noise has lower frequencies than the beacon's signal, addresses this problem. Nevertheless, current industrial cameras do not offer a satisfying spatio-temporal resolution trade-off. Biologically-inspired event cameras, operating with temporally and spatially sparse events, achieve pixel frequencies on the order of 10^4 Hz and can be combined with Spiking Neural Networks (SNNs) to build low-latency neuromorphic solutions. In this paper, we propose to exploit the fine temporal and spatial resolution of event cameras to tackle the challenge of simultaneous OCC and tracking, where the latter is based on the optical flow computed from events by an SNN. We evaluate our approach with a simulated drone monitoring assets on a construction site. We further introduce a hardware prototype comprising a beacon and an event camera, which we use for demonstrating an improvement over state-of-the-art OCC range. § RELATED WORK Optical identification is commonly implemented with frame-based cameras, either by recognizing a spatial pattern in each single image – for instance for license plate recognition <cit.> – or by reading a temporal pattern from an image sequence <cit.>. In the latter, blinking beacons encode a number in binary format, similarly to Morse code, to identify cars or road signs. As explained before, because of the resolution/frame rate trade-off, frame-based cameras impose a hard limit on the beacon's frequency. Event cameras overcome this limitation by capturing individual pixel intensity changes extremely fast rather than full frames <cit.>. Early work combined the fine temporal and spatial resolution of an event camera with blinking LEDs at different frequencies to perform visual odometry <cit.>. Recent work also relies on smart beacons to transmit a message with the UART protocol <cit.>, delivering error-free messages at up to 4 kbps indoors and up to 500 bps at 100 m distance outdoors with brighter beacons. On the tracking front – to track moving beacons in our case – a widely used technique is optical flow <cit.>. Model-free techniques relying on event cameras for object detection have been implemented in <cit.> and <cit.>. To handle the temporal and spatial sparsity of an event camera, a state-of-the-art deep learning frame-based approach <cit.> was adapted to produce dense optical flow estimates from events <cit.>. However, a much simpler and more efficient solution is to compute sparse optical flow with inherently sparse biologically-inspired SNNs <cit.>. § SYSTEM DESCRIPTION The system that we propose is composed of an emitter and a receiver. The former is a beacon emitting a temporal pattern with near infrared light (visible to cameras, but not to humans), attached to the object to be identified and tracked. The receiver component is an event-based camera connected to a computer which, in turn, executes decoding and tracking algorithms. The receiver part comprises algorithmic components for clustering and tracking in which an SNN calculates optical flow. The entire process, from low-level event-processing to high-level sequence-decoding, is schematically depicted in Fig. <ref>. b §.§ Event-Based Communication The emitter is synchronously transmitting, with a blinking pattern, a binary sequence S that consists of a start code S_c, a data payload (identification number) S_p and a parity bit f(S_p), where f returns 1 if S_p has an even number of ones, or 0 otherwise. The start code and the parity bit delimit the sequence and confirm its validity, as illustrated in Fig. <ref>. On the receiver side, the event camera asynchronously generates events upon pixel brightness changes, which can be caused by either a change in the beacon's signal or visual noise in the scene. The current state of the beacon (on or off) cannot be detected by the sensor. Rather, the sensor detects when the beacon transitions between these states. The signal frequency being known, the delay between those transitions gives the number of identical bits emitted. In comparison to a similar architecture with a frame-based camera (200 Hz frame rate) as in <cit.>, our setup relies on an event camera and a beacon blinking in kHz frequency, allowing for a short beacon decoding time, better separation from noise and easier tracking since beacon's motions are relatively slower. §.§ Event-Based Optical Flow b Event-based optical flow is used as a given input for tracking. It is computed from the same camera and events that are used for decoding, and delivers a sparse vector field for visible events with velocity and direction. We implemented an SNN architecture with Spiking Neural Units (SNUs) <cit.> and extended the model with synaptic delays that we call ΔSNU. Its state equations are: s_t = g(W d_Δ(x_t) + l(τ) s_t-1 (1-y_t-1) ) y_t = h(s_t - v_th), where W are the weights, v_th is a threshold, s_t is the state of the neuron and l(τ) its decay rate, y_t is the output, g is the input activation function, h is the output activation function, and d is the synaptic delay function. The delay function d is parameterized with a delay matrix Δ that for each neuron and synapse determines the delay at which spikes from each input x_t will be delivered for the neuronal state calculation. Optical flow is computed by a CNN with 5×5 kernels, illustrated in Fig.<ref>a. Each ΔSNU is attuned to the particular direction and speed of movement through its specific synaptic delays, similarly to <cit.>. When events matching the gradient of synaptic delays are observed, a strong synchronized stimulation of the neuron leads to neuronal firing. This results in sparse detection of optical flow. The synaptic delay kernels are visualized in Fig. <ref>b. We use 8 directions and 4 magnitudes, with the maximum delay period corresponding to 10 executions of the tracking algorithm. Weights are set to one and v_th=5, which yielded the best tracking results. §.§ Object Tracking §.§.§ Clustering Camera events are being accumulated in a time window and clustered with the Density-Based Spatial Clustering of Applications with Noise <cit.>, chosen to get rid of noisy, isolated events and to retrieve meaningful objects from the visual scene. Such clusters are filtered according to: N_e/π× |b-d|^2 > r, N_e ∈{N_min;N_max} where N_e is the number of events in the cluster, |b-d| the Euclidean distance between the cluster's barycenter b and d its most distant event, r a shape ratio, and N_min and N_max the minimal/maximal emitter size in pixel. We reduce the remaining clusters to their barycenter, size and their polarity and call these "targets". The polarity of a target P is given by P=(∑_ip_i)/ N_e where p_i=1 for a positive polarity and -1 for a negative one for each event i. §.§.§ Tracking Targets are kept in memory for tracking over time and are then called tracks. A Kalman filter is attributed to each track and updated for every processed time window, as depicted in Fig. <ref>. Similarly to <cit.>, predicted tracks' states are matched to detected targets to minimize the L1-norm between tracks and targets. Unmatched targets are registered as new tracks. b §.§.§ Identification A matched track's sequence is updated using the target's mean event polarity P. * If P≥0.5 then the beacon is assumed to have undergone an on transition. We add n=(t_c-t_t) / f_beacon zeros to the binary sequence where t_c is the current timestamp, t_t is the stored last transition timestamp and f_beacon is the beacon blinking frequency and set t_t=t_c. * If P≤-0.5 then the beacon is assumed to have undergone a transition to the off state. Likewise, we add n ones to the binary sequence. * Otherwise, the paired beacon most likely has not undergone a transition but just moved. Similarly to <cit.>, a confidence value is computed to classify tracks as new, valid or invalid, as illustrated in Alg.<ref>. Indeed, noise can pass clustering filters but will soon be invalidated as its confidence will never rise. To correct for errors (for instance due to occlusions), the confidence increments are larger than decrements. When a track's sequence is long enough to be decoded, this sequence is valid if it complies with the protocol and maintains the same payload (if this track was previously correctly recognized). § RESULTS §.§ Static Identification Our hardware beacon has four infrared LEDs (850 nm) and an ESP32 micro-controller to set the payload S_p=42 and the blinking frequency. To receive the signal, we used a DVXplorer Mini camera, with a resolution of 640×480 and a 3.6 mm focal length lens. In a static indoor setup, the hardware event camera enables us to achieve high data transmission frequencies, plotted in Fig. <ref>. The metric is the Message Accuracy Rate (MAR): the percentage of correct 11-bit sequences decoded from the beacon's signal during a recording. The MAR stays over 94 % up to 2.5 kHz, then decreases quickly, due to the temporal resolution of the camera. Using a 16 mm focal length lens we could identify the beacon at a distance of 11.5 m indoors, with 87 % MAR and a frequency of 1 kHz and obtained 100 % MAR at 100 Hz at 16 m – see Fig <ref>. t0.42 §.§ Dynamic Identification To evaluate our identification approach in a dynamic setup, where tracking is required, a simulated use case was developed in the Neurorobotics Platform <cit.>. A Hector drone model, with an on-board event camera plugin <cit.>, flies over a construction site with assets (packages and workers) to be identified and tracked. These are equipped with blinking beacons. The drone follows a predefined trajectory and the scene is captured from a bird's eye view – see Fig. <ref>. A frame-based camera with the same view is used for visualization. Noise is simulated with beacons of different sizes blinking randomly. For varying drone trajectories, assets were correctly identified at up to 28 m, with drone speeds up to 10 m/s (linear) and 0.5 radian/s (self rotational). Movements were fast, relative to the limited 50 Hz beacon frequency imposed by the simulator. A higher MAR was obtained with a Kalman filter integrating optical flow (<ref>) than without it – see Tab. <ref>. MAR and Bit Accuracy Rate (BAR) are correlated in simulation because they drop together only upon occlusion. Finally, we conducted hardware experiments where a beacon was moved at 2 m/s reaching a 94 % BAR at 5m and a 87 % BAR at 16m. This shows that our system enables accurate identification and data transmission even with moving beacons, which, to our knowledge, is beyond the state-of-the-art. § CONCLUSION We propose a novel approach for identification that combines the benefits of event-based fast optical communication and signal tracking with spiking optical flow. The approach was validated in a simulation of drone-based asset monitoring on a construction site. A hardware prototype setup reached state-of-the-art optical communication speed and range. We propose the first – to the best of our knowledge – system to identify fast moving, variable beacons with an event camera. § ACKNOWLEDGEMENTS The research was carried out within IBM and fortiss Center for AI (C4AI). We thank its team for discussions and assistance. The research at fortiss was supported by the HBP Neurorobotics Platform funded through the European Union’s Horizon 2020 Framework Program for Research and Innovation under the Specific Grant Agreements No. 945539 (Human Brain Project SGA3). IEEEbib

http://arxiv.org/abs/2303.06829v1

20230313030932

Unbounded Chaotic Weighted Pseudo-shift

math.FA

hypercyclic operators]unbounded chaotic weighted pseudo-shift Pengyu Qin]Pengyu Qin1 1 College of Mathematics and Statistics, Chongqing University, Chongqing, 401331, P. R. China ruxi Liang]ruxi Liang2,3 2 National Astronomical Observatories, Chinese Academy of Sciences, Beijing, 100101, P. R. China 3 University of Chinese Academy of Sciences, 19A Yuquanlu, Beijing 100049, P. R. China yonglu Shu]yonglu Shu* * College of Mathematics and Statistics, Chongqing University, Chongqing, 401331, P. R. China shuyonglu@163.com In this paper, based on the work of Vijay K. Srivastava and Harish Chandra, we give a characterization of the unbounded hypercyclic weighted pseudo-shift operator wC_φ on ℓ^p or c_0. Moreover we use the hypercyclicity criterion of Bès,Chan, and Seubert to give a necessary and sufficient condition in order that wC_φ be a chaotic operator. [ [ March 30, 2023 ================== § INTRODUCTION Let X be be a topological vector space over the field 𝕂=ℝ or ℂ, a continuous linear operator T on space X is called hypercyclic if there exists a vector x∈ X whose orbit orb( x,T)={T^nx | n=0,1,2,...... } is dense in X. Such a verctor x is said to be hypercyclic vector for T, and the set of all hypercyclic vectors for T is denoted by HC(T). Finally, T is said to be chaotic if it is hypercyclic and its set of periodic points is dense in X.We refer to <cit.> for a more detailed study of hypercyclicity and chaoticity. The study of linear dynamics has a long history and it is closely connected to the problem of invariant subspaces in Hilbert spaces. In <cit.>, Birkhoff showed the hypercyclicity of the operator of translation on the space H(ℂ)of entire functions, while MacLane’s result<cit.> was on the hypercyclicity of the differentiation operator on H(ℂ). And later the first example of a hypercyclic operator on a Banach space was given by Rolewicz <cit.> in 1969, he showed that for any λ with |λ|>1, the shift λ B( x_n ) _n=( λ x_n+1) _n are hypercyclic on the sequence space ℓ^2. Motivated by these examples, the hypercyclicity and chaoticity of linear operators, as one of the most-studied properties in linear dynamics, has become an active area of research<cit.>. Among these examples, the most-studied class is certainly that of weighted shifts. Following the paper of Rolewicz<cit.>, In 1995, Salas <cit.> characterized hypercyclic and weakly mixing unilateral and bilateral weighted shifts on ℓ^2 and ℓ^2(ℤ) respectively while he characterized the supercyclic bilateral weighted shifts on Fréchet space in <cit.>. The mixing weighted backward shift on ℓ^2 and ℓ^2(ℤ) were characterized by Costakis and Sambarino<cit.>. Furthermore, Martínez and Peris<cit.> gave new examples of chaotic backward shift operators in the special case of Köthe sequence spaces. In <cit.>, the sufficient and necessary condition for the hypercyclicity of the weighted shift operator have been shown: Let T be a continuous weighted backward shift on ℓ^2(ℕ) defined by T(∑_k∈ℕx_ke_k) =∑_k∈ℕw_kx_k+1e_k. Then T is hypercyclic if and only if there is an increasing sequence ( n_k ) _k∈ℕ of positive integers such that the weight sequence { w_k } _k∈ℕ satisfies ∏_j=1^n_k| w_j |→∞ , as n_k→∞. This result is generalized by K.-G. Grosse-Erdmann to Fréchet space in <cit.> and the characterization of chaotic weighted backward shift is provided in the following terms: Let T be a weighted backward shift on a Fréchet space X in which ( e_n ) _n∈ℕ is an unconditional basis. Then the following assertions are equivalent: (1) T is chaotic. (2) T has a nontrivial periodic point. (3) the series ∑_n=1^∞( ∏_k=1^nw_k) ^-1e_n converges in X. Moreover, K.-G. Grosse-Erdmann proposed the weighted pseudo-shift operator and disscussed its hypercyclity in <cit.>. Let w∈ℓ^∞ and φ be a self-map on the set of natural numbers ℕ={1,2,3...... }, the weighted pseudo-shift on ℓ^p is defined by wC_φ( x_n ) _n∈ℕ=( w_nx_φ( n )) _n∈ℕ It is stated in <cit.> that wC_φ is bounded if and only if sup_k∈ℕ| w_k |^p | φ^-1 ( k ) | < ∞ (the symbol | φ^-1 ( k ) | stands for the cardinality of the set {φ ^-1( n ) } for n≥ 1). In <cit.>, Vijay K. Srivastava and Harish Chandra give a characterization of the hypercyclic weighted pseudo-shift as following: The operator wC_φ : ℓ ^p →ℓ ^p is hypercyclic if and only if the following assertions holds: (1) φ is injective. (2) φ ^n has no fixed point for any n ∈ℕ. (3) For each n ∈, there exists an increasing sequence { m_k }_k ∈ℕ of positive integers such that lim_k→∞| w_ n w_ φ( n )⋯ w_ φ ^m_k-1( n )|=∞ . We need to note that the above studies are all for bounded operators. In fact, the unbounded hypercyclic operators have also attracted the interest of scholars. In <cit.>, Bès, Chan, and Seubert have defined the notion of hypercyclicity and chaoticity for an unbounded operator. In <cit.>, R. deLaubenfels, H. Emamirad, and K.-G. Grosse-Erdmann generalize the notion of hypercyclic and chaotic semigroups to families of unbounded operators. Let T be an unbounded operator on a separable infinite dimensional Banach space X, the densely defined operator T is called hypercyclic if there is a vector f ∈ D( T^∞)=⋂_n≥ 1 D( T^n ) such that its orbit { f,Tf,T^2f,T^3f,...... } is dense in X. Every such vector f is called hypercyclic vector for T. If there exist an integer n∈ℕ and a vector f∈ D( T^n ) such that T^nf=f, such vector is called periodic and the operator A is said to be chaotic if it possesses both hypercyclic vector and dense sets of periodic points. Motivated by <cit.>, Our aim in this paper is to extend the above result to the unbounded weighted pseudo-shift for establishing a necessary and sufficient condition in order that a such operator to be hypercyclic or chaotic on ℓ^p and c_0 respectively. For achieving this end, we have to use a Hypercyclicity Criterion for unbounded operators. Such a criterion has been already provided by Bès, Chan, and Seubert <cit.> in the following terms: Let X be a separable infinite dimensional Banach space and let T be a densely defined linear operator on X for which T^n is a closed operator for all positive integers n. If there exist a dense subset Y of the domain of T and a (possibly nonlinear and discontinuous) mapping S: Y → Y such that (1) TSy = y, for all y ∈ Y, (2) lim_n →∞ S^n y = 0, for all y ∈ Y, (3) lim_n →∞ T^n y = 0, for all y ∈ Y, then T is hypercyclic. Theorem <ref> also holds when the entire sequence of positive integers in hypothesis is replaced by a subsequence of positive integers. To prove hypercyclicty, we also need the following theorem given in <cit.>. Let T be a hypercyclic operator. Then its adjoint T^* has no eigenvalues. Equivalently, every operator T - λ I, λ∈𝕂, has dense range. In <cit.>, H. Emamirad and G.S. Heshmati characterized unbounded chaotic backward shift operators on ℓ^2 in the following terms: Let T be a unbounded backward shift operator on ℓ^2. Then the following assertions are equivalent: (1) T is chaotic. (2) T has a nontrivial periodic point. (3) the positive series ∑_n=1^∞∏_j=0^n-11/| w_j |^2<+∞. Basing on the above theorems, we extend theorem <ref> to unbounded weighted pseudo-shift in section 2. And in section 3, we give a characterization of the chaotic weighted pseudo-shift operator on ℓ^p as following: Let wC_φ be a unbounded weighted pseudo-shift operator on ℓ^p, then wC_φ is chaotic if and only if the following assertions hold: (1) φ is injective. (2) φ ^n has no fixed point for any n ∈ℕ. (3) ∀ k ∈, the following inequalities hold ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n-1( k )|^p<+∞ , and ∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p<+∞. In section 3, we also extend the above result to the space c_0 . § HYPERCYCLICITY In <cit.>, the necessary and sufficient condition for the hypercyclity of bounded weighted pseudo-shift operators on ℓ^p is given and in this section we extend it to the unbounded case. If the operator wC_ is unbounded, then sup_n∈ℕ| w( n ) |=∞ or sup_n∈ℕ card{φ ^-1( n ) } =∞. We will see in below that the hypercyclicity of wC_ implies that φ is injective. So for the unbounded operator wC_, we only concern the case when sup_n∈ℕ| w( n ) |=∞. Denote e_n = ( n - 1 0,⋯ ,0, 1, 0,⋯) , n∈. and c_00={ x=( x_k ) _k∈ℕ∈𝕂^ℕ| ∃ N∈ℕ, s.t. ∀ k≥ N, x_k=0 . } the space c_00 is dense in ℓ^p or c_0, we will always use this notation below. Let { w_k } _k∈ℕ be an arbitrary weight sequence, we define the unbounded weighted pseudo-shift wC_ on ℓ^p by wC_ (∑_k=1^∞x_ke_k) = ∑_k=1^∞w_k x_ (k) e_k where sup_k∈ℕ| w(k) |^p card{φ ^-1( n ) } = +∞, with the domain D(wC_ ) = { x ∈ l^p | ∑_k=1^∞w_k x_(k) e_k ∈ l^p . }. By the definition of wC_, (∑_k=1^∞x_ke_k) = ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) x_^n (k) e_k where sup_k∈ℕ| w(k) |^pcard{φ ^-1( n ) } = +∞, with the domain D(wC_ ^n) = { x ∈ l^p | ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) x_^n (k) e_k ∈ l^p . }. for all n ∈ℕ. Before proving the necessary and sufficient condition for the hypercyclicity of unbounded operator wC_ on ℓ^p , we need to show some lemmas. Suppose map φ : ℕ→ℕ is injective and φ ^n has no fixed point for any n ∈ℕ. Then (1) ∀ n ∈ , ^n (k) ∞ (k ∞). (2) ∀ k ∈ , ^n (k) ∞ (n ∞). (1). Firstly, we show that φ ^n is injective for any n ≥ 1. Assume φ ^k-1 is injective, if there exists i,j ∈ℕ such that ^k (i) = ^k (j) i.e ( ^k-1 (i)) = ( ^k-1 (j)). Since φ is injective, then ^k-1 (i) = ^k-1 (j). So by the assumption, i=j. Thus by induction, it follows that ∀ n ∈, ^n is injective. Conversely, suppose there exists n_0 ∈ℕ such that φ ^n_0( k ) ↛∞ as k→∞, that is there exists a bounded subsequence {^n_0(m_k)}_k ∈ of {^n_0(k) }_k∈ℕ. Suppose there exists M ∈ℕ such that ∀ k ∈ , ^n_0(m_k) ≤ M. Consider the following set {^n_0(m_1), ^n_0(m_2), ⋯, ^n_0(m_M), ^n_0(m_M+1) } Note that each ^n_0(k) is a positive integer, by pigeonhole principle, there must exists positive integer i,j ≤ M+1 ( i j ) such that ^n_0 (m_i) = ^n_0 (m_j), which is a contradiction to the fact that ^n_0 is injective. (2). Conversely, suppose there exists k_0 ∈ℕ such that φ ^n( k_0 ) ↛∞ as n→∞, that is there exists a bounded subsequence {^m_k(k_0)}_k ∈ of {^n(k_0) }_n∈ℕ. Suppose there exists M∈ℕ such that ∀ k ∈ , ^m_k(k_0) ≤ M. Consider the following set {^m_1 (k_0), ^m_2 (k_0), ⋯, ^m_M (k_0), ^m_M+1 (k_0) } By pigeonhole principle, there must exists positive integer i,j ≤ M+1 ( i j ) such that ^m_i (k_0) = ^m_j (k_0). This contradicts the fact that φ ^n has no fixed point for any n ∈ℕ. Suppose map φ : ℕ→ℕ is injective and φ ^n has no fixed point for any n ∈ℕ. Then the subspace D( [ wC_]^∞) = ⋂_n≥ 1 D( [ wC_] ^n ) is dense in ℓ^p. ∀ x=∑_k=1^∞x_k e_k∈ c_00, there exists N_1>0 such that ∀ k>N_1, x_k=0. Since φ is injective, by Lemma <ref> (1), there exists N_2>0 such that ∀ k>N_2, ^n(k)>N_1. Take N=max{ N_1,N_2 }, then ∀ k > N, x_k = x_^n (k) = 0. Hence, ∀ n∈ (x) = ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) x_^n (k) e_k . has only finite term coefficients that are not zero, which implies (x)∈ l^p. So c_00⊂ D( [ w C_] ^n ) ⊂ l^p, and in view of the denseness of c_00 in ℓ^p, it follows that D( [ w C_] ^n ) is dense in ℓ^p. By the arbitary of n, the conclusion is proven. This lemma legitimates the research of hypercyclicity of the unbounded operator wC_ in the framework of Definition <ref>. And in order to use the hypercyclicity criterion <ref>, we have to ensure that for each n∈ℕ, the operator is closed. ∀ n∈ℕ, the operator defined by (<ref>) and (<ref>) is closed. suppose that as j →∞, D( [ wC_] ^n ) x_j = ∑_k=1^∞c_k,j e_k →∑_k=1^∞c_k^* e_k = x^*∈ l^p . (x_j) →∑_k=1^∞d_k^* e_k = f^* ∈ l^p. by ( <ref>), we infer that ∀ k ∈ , c_k,j c_k^* as j →∞. So ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) c_^n (k), j∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) c_^n (k)^* . by ( <ref>), (x_j) = ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) c_^n (k), j e_k f^* = ∑_k=1^∞d_k^* e_k . Hence, w_k w_(k)⋯ w_^n-1 (k) c_^n (k), j d_k^*. It follows that w_k w_(k)⋯ w_^n-1 (k) c_^n (k)^* = d_k^*. Thus, x^* = (∑_k=1^∞c_k^* e_k) = ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) c_^n (k)^* e_k = ∑_k=1^∞d_k^* e_k = f^*. This implies that x^*∈ D( [ wC_] ^n ) and x^* = f^*. So we can conclude that the operator is closed for each n∈ℕ. Suppose ∀ n∈ℕ, w(n) 0. If the unbounded weighted pseudo-shift operator wC_ on ℓ^p is hypercyclic, then φ is injective. Conversely, suppose there exists k_1, k_2∈ℕ( k_1 k_2) such that (k_1) = (k_2) = l. Let the operator is hypercyclic, then the range Im( wC_ ) is dense in ℓ^p. So for e_k_1∈ l^p, ∀ m ∈ℕ , there exists  a vector x_m ∈ D( [ wC_]^∞) such that ‖ (x_m) - e_k_1‖≤1/m , that is ‖ ( w_1x_ (1), ⋯ , w_k_1-1x_ (k_1-1), w_k_1x_ (k_1) - 1, w_k_1+1x_ (k_1+1), ⋯) ‖≤1/m. Hence, |w(k_1)x_ (k_1) - 1| ≤1/m , |w(k_2)x_ (k_2)| ≤1/m . By (<ref>), |x_l|=|x_ (k_1)|=≥1/|w_k_1| (1-1/m) 1/|w_k_1| ( m ∞). By (<ref>), |x_l|=|x_ (k_2)|≤1/m|w_k_2| 0 ( m ∞). This is a Contradiction, which completes the proof. If the unbounded weighted pseudo-shift operator wC_ on ℓ^p is hypercyclic, then φ ^n has no fixed point for each n ∈ℕ. Let wC_ is hypercyclic. Conversely, supposes φ ^n has some fixed points for each n ∈ℕ, then there exists l,n∈ℕ such that ^n (l) = l. Let f_l(x) = x_l, then f_l is a bounded linear functional on ℓ^p. By the definition of the adjoint ( wC_φ) ^*, ( ( ( wC_) ^n ) ^∗( f_l ) ) (x) = ( f_l ( ( wC_) ^n ) (x) ) = f_l ( ( wC_) ^n (x) ) = f_l ( ( w_k w_(k)⋯ w_^n-1 (k) x_^n (k))_k ∈) = w_l w_(l)⋯ w_^n-1 (l) x_^n (l) = w_l w_(l)⋯ w_^n-1 (l) x_l = w_l w_(l)⋯ w_^n-1 (l) f_l(x) . Hence, ( ( ( wC_) ^n ) ^∗( f_l ) ) = w_l w_(l)⋯ w_^n-1 (l) f_l(x) . It follows that f_l is an eigenvector of . Since wC_φ is hypercyclic, then ( wC_φ) ^n is also hypercyclic, which is a contradiction to Theorem <ref>. Before proving our conclusion, we need to extend the definition of . ∀ n∈, we define φ ^-n( k ) = l , ∃ l s.t. φ^n(l) = k, 0 , φ ^-n( k ) =∅. Moreover, we specify e_0=(0,0,⋯,0,⋯) and w_0=0. By the definition of wC_, (e_k) = w_^-n(k) w_^-(n-1)(k)⋯ w_^-1(k) e_^-n(k). for all n∈ℕ. Next, our conclusions are given. The unbounded operator wC_ on ℓ^p is hypercyclic if and only if the following assertions holds: (1) is injective. (2) ^n has no fixed point for any n ∈ℕ. (3) For each n ∈, there exists an increasing sequence { m_k }_k ∈ℕ of positive integers such that lim_k→∞| w( n ) w( ( n ) ) ⋯ w( ^m_k-1( n ) ) |=∞ . Assume w C_ is hypercyclic. By Theorem <ref> and Theorem <ref>, we infer that the condition ( 1 ) and( 2 ) hold. Now we need only to show ( 3 ). Since the set of hypercyclic vectors for w C_ is dense in ℓ^p, then ∀δ _k>0, there exists a vector x∈ HC( w C_) (also belongs to D( [ w C_] ^∞) ) such that ‖ x-e_n ‖ <δ _k. Therefore, ∀ k n, | x_k |<δ _k⇒| x_ ^j( n )|<δ _k,  ∀ j∈ N ( φ ^n has no fixed point ) Since x∈ HC( w C_), then orb( x, w C_) dense⊂l^p. So there exists m>0 such that ‖( w C_) ^mx-e_n ‖ <δ_k , Hence, < | w _n w _( n )⋯ w _ ^m-1( n )x_ ^m( n )-1 | < ( |w _n w _( n )⋯ w _ ^m-1( n )x_ ^m( n )-1| ^p + ∑_k ∈, k n|w _n w _( n )⋯ w _ ^m-1( n )x_ ^m( n )|^p ) ^1/p = ‖( w C_) ^mx-e_n ‖ < δ _k. Therefore, 1-| w _n w _( n )⋯ w _ ^m-1( n )x_ ^m( n )|<δ _k , Thus, | w _n w _( n )⋯ w _ ^m-1( n )|>1-δ _k/| x_ ^m( n )|>1-δ _k/δ _k = 1/δ_k - 1. By the arbitary of δ_k, we can choose a sequence {δ _k } which tends to 0 , so the condition ( 3 ) is necessary if w C_ is hypercylic. For the sufficient part of the proof, we will use the Hypercyclicity Criterion stated in Theorem <ref>. The closedness of is already proved in Lemma <ref>. Assume the condition ( 1 ) ,( 2 ) ,( 3 ) hold, now we define the map S: c_00 → c_00 ( e_n ) →e_( n )/ w _n, As is easily seen, ∀ n∈ N ( wC_) ( S( e_n ) ) = w C_( e_( n )/ w _n) =e_n . ∀ x=∑_k=1^∞x_k e_k∈ c_00, there exists N_1>0 such that ∀ k>N_1, x_k=0. In view of the condition ( 1 ) ,( 2 ) and Theorem <ref> ( 2 ), for each k∈ℕ, there exists N_2>0 such that ∀ n>N_2, ^n(k)>N_1. So take N=max{ N_1,N_2 }, we infer that for any n>N, x_ ^n( k )=0. So for each x∈ c_00, ( ( w C_) S ) ( x ) =∑_k=1^∞x_k [ ( wC_) ( S( e_k ) )] =∑_k=1^∞x_k e_k = x and (x) = ∑_k=1^∞w_k w_(k)⋯ w_^n-1 (k) x_^n (k) e_k → 0 ( n→∞) . Further, by the definition of S, ‖ S^nx ‖ = ‖ S^n( ∑_k=1^Nx_ke_k) ‖ =‖∑_k=1^Nx_k e_ ^n( k )/ w _k w _( k )⋯ w _ ^n-1( k )‖ =( ∑_k=1^N| x_k |^p1/| w_kw_φ( k )...w_φ ^n-1( k )|^p) ^1/p ≤‖ x ‖( ∑_k=1^N1/| w_kw_φ( k )...w_φ ^n_k-1( k )|^p) ^1/p. Now in view of the condition on w that for each n ∈, there exists an increasing sequence { m_k }_k ∈ℕ of positive integers such that lim_k→∞| w( n ) w( ( n ) ) ⋯ w( ^m_k-1( n ) ) |=∞ . for each x∈ c_00, we obtain that S^m_k x → 0 ,( k →∞). Hence with (<ref>),(<ref>),(<ref>), w C_ satisfies the Hypercyclicity Criterion stated in Theorem <ref>, thus w C_ is hypercyclic. As is easily seen that the characterization of the hypercyclic weighted pseudo-shift operator w C_ is the same in both the bounded and unbounded cases. So using the same approach, we can extend the conclusion to weighted pseudo-shift operators (whether bounded or unbounded) on the space c_0. The weighted pseudo-shift operator wC_ on c_0 is hypercyclic if and only if the following assertions holds: (1) is injective. (2) ^n has no fixed point for any n ∈ℕ. (3) For each n ∈, there exists an increasing sequence { m_k }_k ∈ℕ of positive integers such that lim_k→∞| w( n ) w( ( n ) ) ⋯ w( ^m_k-1( n ) ) |=∞ . § CHAOTICITY In this section, we will give a necessary and sufficient condition in order that w C_ be a chaotic operator. Some preparation is needed before proving it. And the discussion of chaoticity requires us to pay more attention to some properties of , so in first section we will prove a few lemmas. §.§ Some Lemmas The following lemma is provided as a supplement to Lemma <ref>. Suppose is injective and ^n has no fixed point for any n ∈ℕ, then for any k∈, if ∀ n ∈, ^-n(k) 0, then ^-n (k) ∞ (n ∞). Conversely, suppose there exists k_0 ∈ℕ such that ^-n (k_0)↛∞ (n ∞), then there exists a bounded subsequence {^-m_k(k_0)}_k ∈ of {φ ^-n( k_0 ) } _n ∈. Suppose there exists M∈ℕ such that ∀ k ∈ , ^-m_k(k_0) ≤ M. Consider the following set {^-m_1 (k_0), ^-m_2 (k_0), ⋯, ^-m_M (k_0), ^-m_M+1 (k_0) } By pigeonhole principle, there must exists positive integer i,j ≤ M+1 ( i j ) such that ^-m_i (k_0) = ^-m_j (k_0). This contradicts the fact that φ ^n has no fixed point for any n ∈ℕ. For each n ∈ and any positive integer k, we denote a “^n bi-orbit" (later referred to as “^n-orbit" ) of k by the set A_n,k = orb(k,^n) ∪ orb(k,ϕ^n)= {^in(k) | i ∈} - { 0 }, Where ϕ=^-1. Moreover, we denote A_k = A_1,k and B_n,k = {^in(k) | i ∈ , i 0 } - { 0 } = A_n,k - { k }. The reason for dropping { 0 } from the definition of A_n,k is to satisfy the previous definition in ( <ref>) which will make it easier for us to select the minimum element A_n,k. By definition, it is easy to know the set A_n,k⊆. When φ is injective, for any fixed k, there are two cases : when k is a fixed point of φ for some n, A_k is a finite set; when k is not fixed point of φ for any n, A_k is an infinite set. It is worth noting that there are also two cases in which A_k is an infinite set: one is when there exists n ∈ such that ^-n(k) = 0 i.e there is a point without an preimage under ^n in the orbit A_k (this point is like the “generator" of the orbit); and the other is when ^-n(k) 0 for any n ∈ , that is, there is no point without an preimage under ^n in the orbit A_k. The above statement also holds for A_n,k. Next, we give some examples of φ for better illustration. Consider the mapping φ( n ) = 2, n=1 3, n=2 1, n=3 4, else. In this case, ∀ k=1,2,3, A_k={1,2,3} is a finite set and k is a fixed point of φ. Consider the mapping φ( n )=n+1, ∀ n ∈ℕ. For k=1, ^-1(1) = 0 i.e 1 has no preimage. In this case, A_1=ℕ is an infinite set and there is only one -orbit A_1, that is, for any k ∈, A_k=A_1. Furthermore, ^n has no fixed point for any n ∈ℕ. Consider the mapping φ( n ) = 5k-4, n=5k+1 5k+2, n=5k       (k≥ 1) n+1, else. For any k,n∈ℕ, ^-n(k) 0, that is, there is no point without an preimage. In this case, there is only one -orbit A_1. Moreover, A_1=ℕ is an infinite set and ^n has no fixed point for any n ∈ℕ. Suppose is injective without any periodic point, then for each k and N, (1) ∀ n ∈ orb(k,^N), let φ^n_1N( k ) =n (n_1 ∈), then lim_n∞n_1=∞; (2) ∀ n ∈ orb(k,ϕ^N), let φ ^-n_2N( k ) =n (n_2 ∈), then lim_n∞n_2=∞. (1). For any positive integer M, let L = {^n_1N|n_1≤ M} and N = sup L, then when n ∈ orb(k,^N) and n>N, we have n_1>M. This is because if n_1<M, then n = ^n_1N∈ L, hence n≤sup L= N, which makes a contradiction. Therefore, we have lim_n∞n_1=∞. (2). For any M ∈, let L = {^-n_2N|n_2≤ M} and N = sup L, then when n ∈ orb(k,ϕ^N) and n>N, we have n_2>M. This is because if n_2<M, then n = ^-n_2N∈ L, hence n≤sup L= N, which makes a contradiction. Therefore, we have lim_n∞n_2=∞. The following lemma shows that the different ^n orbits are disjoint when φ is injective. Suppose is injective, then for all distinct k_1, k_2 ∈ , their ^n-orbit either the same or disjoint. That is, if A_n,k_1 A_n,k_2, then A_n,k_1∩ A_n,k_2 = ∅. For any fixed n, suppose there exists k_1, k_2 ∈ such that A_n,k_1∩ A_n,k_2∅,that is, there exists i,j ∈ such that ^in(k_1) = ^jn(k_2) = l. If i=j, then by φ being a injection we infer that k_1 = k_2 , therefore A_n,k_1 = A_n,k_2. If i j, we might assume that i>j, then k_2 = ^(i-j)n(k_1), k_2 is in the orbit of k_1 ,i.e , A_n,k_1 = A_n,k_2. Define g_n(k)=inf{ A_n,k}, k ∈, this is a mapping from to which maps each k ∈ to the smallest element in A_n,k . We may regard the orbit A_n,k to be generated by g_n(k) under ^n. Moreover, we let g(k)=g_1(k)=inf{ A_k}. Denote G_n = Im (g_n) as the range of g_n, which represents the set of “generator" for each ^n orbit, which may be finite or infinite, but must be countable. Moreover, we let G=G_1=Im(g). The main purpose for this definition is to select the “generator" of each orbit. Again, we give some examples for illustration. Consider the mapping defined in (<ref>) and (<ref>). For these two mappings, A_k=A_1=ℕ, and g(k)=1 ( ∀ k∈ℕ). Thus G={ 1 } with | G |=1. Consider the mapping φ( n ) = 5, n=1 5k+3, n=5k-1 5k+5, n=5k       (k≥ 1) 5k-4, n=5k+1 5k-3, n=5k+2 n+1, else. In this case, ℕ is divided into two -orbit, A_1={ 1,5k,5k+1 }( k≥ 1 ) and A_2=ℕ-A_1. Thus G={ 1,2 } with | G |=2. Let P_i={ p_i^n |  p_i is the i-th prime number, ∀ i,n ∈. } and P_0=ℕ-⋃_i≥ 1P_i. Then we arrange the elements in P_i in ascending order and denote P_i( m ) as the m-th element. It is clear that P_i( m ) = p_i^m (i≥ 1). 1.Consider the mapping φ( P_i( k ) ) =P_i( k+1 ) , ∀ i≥ 0. In this case, Since there are infinitely many prime numbers, then ℕ is divided into a countably infinite number of orbits. And there are points in each orbit with no preimage, actually for any i, ^-1(P_i(1))=0. Furthermore, for any k ∈, we can infer that g(P_i(k))=p_i (i∈) and g(P_0(k))=1. In this case, G is the set of all prime numbers and 1, Thus | G |=+∞. 2. Consider the mapping φ( n ) = P_i( 2 ) , n=P_i( 1 ) P_i( 3k ) , n=P_i( 3k-1 ) P_i( 3k+2 ) , n=P_i( 3k ) P_i( 3k-2 ) , n=P_i( 3k+1 ). In this case, ℕ is divided into a countably infinite number of orbits.And there are points in each orbit with no preimage. Moreover, for any k ∈, we infer that g(P_i(k))=p_i (i∈) and g(P_0(k))=1. Thus, G is the set of all prime numbers and 1 with | G |=+∞. The following lemma shows that we can divide the positive integers into some disjoint ^n orbits. Suppose is injective, then for ∀ n ∈ N,we have = _k∈ G_nA_n,k. By Lemma <ref>, we know that each orbit A_n,k is disjoint. Let S = _k∈ G_nA_n,k, it suffice to show that S ⊆ and ⊆ S. Clearly, S ⊆ since each A_n,k⊆. On the other hand, ∀ k ∈, there exists l ∈ such that g_n(k)=l∈ G_n, then k ∈ A_n,l⊆ S, hence ⊆ S. Suppose is injective, and has no periodic point, then for every k,n ∈, A_k = A_2,k⊔ A_2,(k) = A_3,k⊔ A_3,(k)⊔ A_3,^2(k) = _i=0^n-1A_n,φ ^i( k ). So we have G ⊂ G_2 ⊂ G_3 ⊂⋯. By Theorem <ref>, we know that each orbit A_n,k is disjoint. Let S=_i=0^n-1A_n,φ ^i( k ). For any k,n ∈, ∀ l ∈ A_k, there exists m∈ such that l = ^m(k). Also there exists a∈ and p ∈{0,1,2,⋯,n-1 } such that m = an + p, it follows that l = ^m(k) = ^an + p(k) = ^an(^p(k)) ∈ A_n,^p(k). Thus A_k ⊆ S. On the other hand, for any l ∈ S, there exists a∈, p ∈{0,1,2,⋯,n-1 } such that l = ^an(^p(k)), let m = an + p, then l = ^an + p(k) = ^m(k) ∈ A_k, so S ⊆ A_k which completes the proof. Suppose is injective, and has no periodic point, then for any finite set F ⊆, and every k ∈, there exists a N ∈ such that for any n > N, B_n,k∩ F = ∅. For any finite set F ⊆, there exists M>0 such that max{ F } <M. By Lemma <ref> (2), ∀ k∈, there exists N_1 ∈ such that for any n>N_1, ^n(k)>M. Therefore for any n>N_1 and i ∈, ^in(k)>M which means (orb(k,^n) - { k }) ∩ C = ∅. By Lemma <ref>, there exists N_2 ∈ such that for any n>N_2, ^-n(k)>M, then for any n>N_2 and i ∈, ^-in(k)>M which means (orb(k,^-n) - { 0,k }) ∩ F = ∅. Thus, Let N=max{ N_1, N_2 } B_n,k∩ F = ( (orb(k,^n) - { k }) ∪ (orb(k,^-n) - { 0,k }) ) ∩ F = ( (orb(k,^n) - { k }) ∩ C ) ∪( (orb(k,^-n) - { 0,k }) ∩ F ) = ∅. which completes the proof. The above lemma shows that for any finite subset of positive integers and a fixed k, when n is large enough, the ^n orbit of k can always “escape" from this finite subset. §.§ on lp space In this section, we want to characterise the chaoticity of the unbounded weighted pseudo-shift operator wC_φ on the space ℓ^p. and we adjust the notation for x, let x(k) denote the number at the k-th position of x. Suppose is injective, and has no fixed point for any n ∈ℕ. Then a point x_N is a N-periodic points of wC_ if and only if x_N can be written as x_N = ∑_k∈ G_Na_k x_N,k, where a_k∈𝕂 and x_N,k = e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k). Let x_N =∑_k=1^∞a_k e_k be an N-periodic point of wC_, then (wC_)^N (x_N) = (wC_)^N (∑_k=1^∞a_k e_k) = ∑_k=1^∞w_k w_(k)⋯ w_^N-1(k) a_^N (k) e_k = x_N = ∑_k=1^∞a_k e_k. Therefore, for any k ∈, w_k w_(k)⋯ w_^N-1(k) a_^N (k) = a_k, that is, a_^N(k) = 1/w_k w_(k)⋯ w_^N-1(k) a_k Replacing l = ^N(k) by (<ref>) , we get a_^2N(k) = a_^N(l) = 1/w_l w_(l)⋯ w_^N-1(l) a_l = 1/w_^N(k) w_^N+1(k)⋯ w_^2N-1(k)1/w_k w_(k)⋯ w_^N-1(k) a_k = 1/w_k w_(k)⋯ w_^2N-1(k) a_k. By induction, ∀ n ∈, a_^nN(k) = 1/w_k w_(k)⋯ w_^nN-1(k) a_k. We note that l = ^N(k), then k = ^-N(l), Substitute it into (<ref>), a_^-N(l) = [ w_^-N(l) w_^-N+1(l)⋯ w_^-1(l) ] a_l = [ w_^-1(l) w_^-2(l)⋯ w_^-N(l) ] a_l . By induction, ∀ n ∈ a_^-nN(k) = [ w_^-1(k) w_^-2(k)⋯ w_^-nN(k) ] a_k. Hence for any k,n∈, by (<ref>) , (<ref>) and Lemma <ref>, x_N = ∑_k=1^∞a_k e_k = ∑_k ∈ G_N[ a_k e_k + ∑_n=1^∞a_^nN(k) e_^nN(k) + ∑_n=1^∞a_^-nN(k) e_^-nN(k)] = ∑_k ∈ G_N[ a_k e_k + ∑_n=1^∞1/w_k w_(k)⋯ w_^nN-1(k) a_k e_^nN(k) + ∑_n=1^∞[ w_^-1(k) w_^-2(k)⋯ w_^-nN(k) ] a_k e_^-nN(k)] = ∑_k ∈ G_Na_k [e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)] = ∑_k ∈ G_Na_k x_N,k. As to the sufficient part of the proof, for any k∈ G_N wC_ ^N(x_N,k) = wC_ ^N(e_k) + ∑_n= 1^∞[1/w_k w_ (k)⋯ w_^nN-1(k)] wC_ ^N(e_^nN(k) ) + ∑_n= 1^∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] wC_ ^N(e_^-nN(k)) = w_^-N(k) w_^-(N-1)(k)⋯ w_^-1(k) e_^-N(k) + ∑_n= 1^∞[1/w_k w_ (k)⋯ w_^nN-1(k)] [ w_^nN-N(k) w_^nN-(N-1)(k)⋯ w_^nN-1(k)] e_^nN-N(k) + ∑_n= 1^∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] [ w_^-nN-N(k) w_^-nN-(N-1)(k)⋯ w_^-nN-1(k)] e_^-nN-N(k) = w_^-1(k) w_^-2(k)⋯ w_^-N(k) e_^-N(k) + e_k + ∑_n= 2^∞[1/w_k w_ (k)⋯ w_^(n-1)N-1(k)] e_^(n-1)N(k) + ∑_n= 1^∞[w_^-1(k) w_^-2(k)⋯ w_^-(n+1)N(k)] e_^-(n+1)N(k) = e_k + ∑_n= 1^∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k) = x_N,k. It implies that each x_N,k is an N-periodic point of wC_, thus x_N = ∑_k∈ G_Na_k x_N,k is also an N-periodic point of wC_. It is easy to see that for any k∈ G_N and a fixed N, x_N,k( j ) 0 only when j∈ A_N,k. And Lemma <ref> proves that the different orbits A_N,k of different k are disjoint, so that any two distinct x_N,k have different non-zero locations, that is, for any k_1,k_2,n ∈, x_N,k_1(n) x_N,k_2(n) =0. Thus they are always linearly independent. ^n has no periodic point for any n, so for any n_1, n_2 > 0, ^n_1N(k) ^-n_2N(k). For the n-th position of x_N,k, we have x_N,k( n ) =1/|w_k w_ (k)⋯ w_^n_1N-1(k)| , n ∈ orb(k,^N). | w_^-1(k) w_^-2(k)⋯ w_^-n_2N(k)| , n ∈ orb(k,ϕ^N). 0 , others. Where ϕ=^-1. The fact that ^n has no periodic point ensures that these three cases cannot happen simultaneously. In fact, n ∈ orb(k,^N) is equivalent to the existence of n_1 such that φ ^n_1N( k ) =n, and n ∈ orb(k,ϕ^N) is equivalent to the existence of n_2 such that φ ^-n_2N( k ) =n. Theorem <ref> actually illustrates the structure of the set of periodic points of wC_: every periodic points of period N can be expressed as linear combinations of countable x_N,k(k ∈ G_N). This can be interpreted as the existence of a set of “basis" (i.e. x_N,k | k ∈ G_N) for the set of periodic points with a fixed period N. To verify the denseness of the set of all periodic points, is to compare the union of bases for all sets of periodic points (i.e. ⋃_N ∈{ x_N,k | k ∈ G_N}) and the bases of ℓ^p (e.g. { e_n | n∈}) in terms of how closely they are. Next, we will show the necessary and sufficient condition for the chaotic weighted pseudo-shift operator wC_φ on the space ℓ^p. Let wC_φ be a unbounded weighted pseudo-shift operator on ℓ^p, then wC_φ is chaotic if and only if the following assertions hold: (1) φ is injective. (2) φ ^n has no fixed point for any n ∈ℕ. (3) ∀ k ∈, the following inequalities hold ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n-1( k )|^p<+∞ , and ∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p<+∞. For the necessary part of the proof, assume that wC_ is chaotic, it is naturally also hypercyclic, then the condition (1) and (2) has been proved in the Theorem <ref>, so we need only to show the condition (3). Suppose x_N is a periodic point of wC_ with period N, by Theorem <ref>, x_N = ∑_k∈ G_Na_k x_N,k = ∑_k∈ G_Na_k [e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)]. Then for any i=0,1,2,⋯,N-1, wC_ ^i(x_N) = ∑_k∈ G_Na_k [wC_ ^i(x_N,k)] = ∑_k∈ G_Na_k [∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-(i+1)(k)] e_^nN-i(k) + ∑_n= 0^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-(nN+i)(k)] e_^-(nN+i)(k)]. Since x_N is an N-periodic point of wC_, it follows that wC_ ^i(x_N) ∈ l^p. Therefore ∀ k ∈ G_N, ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^nN-(i+1)( k )|^p<+∞, ∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-(nN+i)( k )|^p<+∞. Since (<ref>) and (<ref>) hold for all i=0,1,2,⋯,N-1, then ∀ k ∈ G_N, ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n-1( k )|^p<+∞ , ∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p<+∞. For all l ∈, suppose k=g(l), there exists m∈ℤ such that l=φ ^m( k ). By Lemma <ref>, k ∈ G ⊆ G_N, then ∑_n=1^+∞1/| w_lw_φ( l )...w_φ ^n-1( l )|^p = | w_kw_φ( k )...w_φ ^m-1( k )| ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n+m-1( k )|^p<+∞ , ∑_n=1^+∞| w_φ ^-1( l )w_φ ^-2( l )...w_φ ^-n( l )|^p = 1/| w_kw_φ( k )...w_φ ^m-1( k )|∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n+m( k )|^p<+∞. For the proof of sufficiency, assume that the condition (1),(2) and (3) hold, then wC_ is hypercyclic by Theorem <ref>. Now it suffices to show that wC_ has a dense set of periodic points. For any vector y ∈ c_00, there exists m ∈ℕ such that y( k )=0 for any k>m, that is, y = (y_1,y_2,⋯,y_m,0,⋯) = ∑_k=1^my_k e_k. By lemma <ref>, for each k=1,2,⋯,m, there exists N_k such that B_n,k∩{1,2,⋯,m } = ∅ for any n > N_k. Take N_a=max_k=1,2,⋯,m{N_k}, then for each k=1,2,⋯,m, B_n,k∩{1,2,⋯,m } = ∅ for any n > N_a. Hence, A_n,k∩{1,2,⋯,m } = k, that is, the intersection of {1,2,⋯,m } with the ^n-orbit A_n,k is only { k }. By the condition (3), ∀ε >0, there exists N_b>0 such that ∑_n>N_a^1/| w_kw_φ( k )...w_φ ^n-1( k )|^p<ε/2m, ∑_n>N_b^| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p<ε/2m. Let N=max{N_a,N_b}, consider x_N,k = e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k), the above discussion ensures that only the k-th term among the first m terms of x_N,k is not 0. First we show that x_N,k is well-defined and is a periodic point. We have proved that ( wC_φ)^N(x_N,k) =x_N,k in Theorem <ref>. By the condition (3), ‖ x_N,k‖ =( 1+∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^nN-1( k )|^p+∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-nN( k )|^p ) ^1/p ≤( 1+∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n-1( k )|^p+∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p ) ^1/p < +∞, hence ( wC_φ) ^N( x_N,k) =x_N,k and x_N,k∈ D( [ wC_]^N). Next, We prove the denseness of the set of periodic points. Let x = ∑_k=1^my_k x_N,k, as is easily seen, x is also the N-periodic point of wC_. Suppose y_0=max_k=1,2,⋯,m|y_k|, by ( <ref>),( <ref>) and the choice of N, ‖ x - y ‖ = ‖∑_k=1^my_k [ ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)) ] ‖ ≤‖∑_k=1^my_k ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k)‖+ ‖∑_k=1^my_k∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)) ‖ ≤ y_0∑_k=1^m‖∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k)‖+ y_0∑_k=1^m‖∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)) ‖ <y_0/2ε +y_0/2ε=y_0 ε. Hence, in view of the denseness of c_00 in ℓ^p, we infer that the set of periodic points of wC_ is dense in ℓ^p as well, i.e, the operator wC_ is chaotic. Let the mapping be defined by φ( n ) = 4k+2, n=4k 4k-3, n=4k+1       (k≥ 1) n+1, else and let w=( w_n ) _n∈ℕ with w_n = w^-n, n=4k+1       (k≥ 1) w^n, else where | w |>1. Then one easily verifies that is injective, φ ^n has no fixed point for any n ∈ℕ and ∑_n=1^+∞1/| w_kw_φ( k )...w_φ ^n-1( k )|^p<+∞ , ∑_n=1^+∞| w_φ ^-1( k )w_φ ^-2( k )...w_φ ^-n( k )|^p<+∞. for each k∈ℕ. Hence, by the Theorem <ref>, the unbounded operator wC_ is chaotic. §.§ on a space In this section , we extend the characterization of the unbounded chaotic weighted pseudo-shift operator to the space c_0. Let wC_φ be a unbounded weighted pseudo-shift operator on c_0, then wC_φ is chaotic if and only if the following assertions hold: (1) is injective. (2) ^n has no fixed point for any n ∈ℕ. (3) For each k ∈, lim_n→∞| w( k ) w( ( k ) ) ⋯ w( ^n-1( k ) ) |=∞, and lim_n→∞| w(^-1(k)) w(^-2(k)) ⋯ w(^-n(k)) |=0. For the necessary part of the proof, assume that wC_ is chaotic, it is naturally also hypercyclic, then the condition (1) and (2) has been proved in the Corollary <ref>, so we need only to show the condition (3). Suppose x_N is a periodic point of wC_ with period N, by Theorem <ref>, x_N = ∑_k∈ G_Na_k x_N,k = ∑_k∈ G_Na_k [e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)]. Then for any i=0,1,2,⋯,N-1 wC_ ^i(x_N) = ∑_k∈ G_Na_k [wC_ ^i(x_N,k)] = ∑_k∈ G_Na_k [∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-(i+1)(k)] e_^nN-i(k) + ∑_n= 0^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-(nN+i)(k)] e_^-(nN+i)(k)]. Since x_N is an N-periodic point of wC_ , it follows that wC_ ^i(x_N) ∈ c_0. And we know that for different wC_ ^i(x_N,k) , the non-zero positions are different. By Lemma <ref>, ^nN-i(k) ∞ (n ∞), then ∀ k ∈ G_N 1/ |w_k w_ (k)⋯ w_^nN-(i+1)(k)| 0, ( n ∞) If ∀ j ∈, ^-j(k) 0, we infer by Lemma <ref> that ^-(nN+i)(k) ∞ (n ∞), and since wC_ ^i(x_N,k) ∈ c_0, then ∀ k ∈ G_N w_^-1(k) w_^-2(k)⋯ w_^-(nN+i)(k) 0, ( n ∞) If ∃ j_1 ∈, ^-j_1(k) = 0 (which implies that j >j_1, ^-j(k) = 0 ), by the nonation in (<ref>), w_^-j(k) = w_0 =0. Thus, when n is large enough , it follows that w_^-1(k) w_^-2(k)⋯ w_^-(nN+i)(k) = 0 so that (<ref>) holds. since (<ref>) and (<ref>) hold for all i=0,1,2,⋯,N-1, then ∀ k ∈ G_N lim_n→∞| w( k ) w( ( k ) ) ⋯ w( ^n-1( k ) ) |=∞, lim_n→∞| w(^-1(k)) w(^-2(k)) ⋯ w(^-n(k)) |=0. For all l ∈, suppose k=g(l), then there exists m∈ℤ such that l=φ ^m( k ). By Lemma <ref>, k ∈ G ⊆ G_N, then lim_n→∞| w( l ) w( ( l ) ) ⋯ w( ^n-1( l ) ) | = 1/| w_kw_φ( k )...w_φ ^m-1( k )|lim_n→∞| w( k ) w( ( k ) ) ⋯ w( ^n+m-1( k ) ) | = ∞, lim_n→∞| w(^-1(l)) w(^-2(l)) ⋯ w(^-n(l)) | =| w_kw_φ( k )...w_φ ^m-1( k )| lim_n→∞| w(^-1(k)) w(^-2(k)) ⋯ w(^-n+m(k)) | =0. For the proof of sufficiency, assume that the condition (1),(2) and (3) hold, then wC_ is hypercyclic by Corollary <ref>. Now it suffices to show that wC_ has a dense set of periodic points. For any vector y ∈ c_00, there exists m ∈ℕ such that y( k )=0 for any k>m, that is, y = (y_1,y_2,⋯,y_m,0,⋯) = ∑_k=1^my_k e_k. By lemma <ref>, for each k=1,2,⋯,m, there exists N_k such that B_n,k∩{1,2,⋯,m } = ∅ for any n > N_k. Take N_0=max_k=1,2,⋯,m{N_k}, then B_n,k∩{1,2,⋯,m } = ∅ for each k=1,2,⋯,m and any n > N_0. Hence, A_n,k∩{1,2,⋯,m } = k, that is, the intersection of {1,2,⋯,m } with the ^n orbit A_n,k is only { k }. For any N>N_0, consider x_N,k = e_k + ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k). It is clear that x_N,k is an N-periodic point of wC_ and the above discussion ensures that only the k-th term among the first m terms of x_N,k is not 0. First we show that x_N,k∈ c_0. According to Lemma <ref> and the condition (3), lim_n∈ orb(k,^N) ∞| 1/w_k w_ (k)⋯ w_^n_1N-1(k)| = lim_n_1∞| 1/w_k w_ (k)⋯ w_^n_1N-1(k)| = 0, lim_n∈ orb(k,ϕ^N) ∞| w_^-1(k) w_^-2(k)⋯ w_^-n_2N(k)| = lim_n_2∞| w_^-1(k) w_^-2(k)⋯ w_^-n_2N(k)| = 0. According to Remark <ref>, one can know that both two non-zero subsequences of x_N,k(n) converge to 0, and all non-zero terms of x_N,k(n) belong to one of the two subsequences. Therefore, we have lim_n→∞x_N,k(n)=0, and x_N,k belongs to c_0. Next, We prove the denseness of the set of periodic points. Let x = ∑_k=1^my_k x_N,k, As is easily seen, x is also the N-periodic point of wC_. Suppose y_0=max_k=1,2,⋯,m|y_k|, then when n 0, by the condition (3) ‖ x - y ‖ = ‖∑_k=1^my_k [ ∑_n= 1^+ ∞[1/w_k w_ (k)⋯ w_^nN-1(k)] e_^nN(k) + ∑_n= 1^+ ∞[w_^-1(k) w_^-2(k)⋯ w_^-nN(k)] e_^-nN(k)) ] ‖ ≤ y_0 max{sup_n∈| 1/w_k w_ (k)⋯ w_^nN-1(k)|, sup_n∈| w_^-1(k) w_^-2(k)⋯ w_^-nN(k)|} 0. Hence, in view of the denseness of c_00 in c_0, we infer that the set of periodic points of wC_ is dense in c_0 as well,i.e, the operator wC_ is chaotic. Let the mapping :ℕℕ as in the example <ref> (2) and let w_n = w^-n, n=P_i( 3k+1 ) w^n, else for all k≥1 and i≥ 0 where | w |>1. Then one easily verifies that is injective, φ ^n has no fixed point for any n ∈ℕ and lim_n→∞| w( k ) w( ( k ) ) ⋯ w( ^n-1( k ) ) |=∞, lim_n→∞| w(^-1(k)) w(^-2(k)) ⋯ w(^-n(k)) |=0. for each k∈ℕ. Hence, by the Theorem <ref>, the unbounded operator wC_ is chaotic. unsrt

http://arxiv.org/abs/2303.06656v1

20230312131153

Re-evaluating Parallel Finger-tip Tactile Sensing for Inferring Object Adjectives: An Empirical Study

cs.RO

Color Mismatches in Stereoscopic Video: Real-World Dataset and Deep Correction Method This study was supported by Russian Science Foundation under grant 22-21-00478, <https://rscf.ru/en/project/22-21-00478/> Egor Chistov,1 Nikita Alutis,1 Maxim Velikanov,1 and Dmitriy Vatolin12 Lomonosov Moscow State University1 MSU Institute for Artificial Intelligence2 =================================================================================================================================================================================================================================================================== empty empty Finger-tip tactile sensors are increasingly used for robotic sensing to establish stable grasps and to infer object properties. Promising performance has been shown in a number of works for inferring adjectives that describe the object, but there remains a question about how each taxel contributes to the performance. This paper explores this question with empirical experiments, leading insights for future finger-tip tactile sensor usage and design. § INTRODUCTION In recent years, with the developments of tactile sensors and machine learning techniques, finger-tip tactile sensors are increasingly used for robotic tasks such as adjective classification <cit.>. Promising performance has been shown for adjective classification <cit.> or learning adjective distributions <cit.>. However, there are few works exploring: how each taxel in a finger-tip tactile sensor contributes to the inferring of object adjectives, and how taxel spatial density and measurement temporal density influence the performance. This paper seeks answers to these questions via empirical experiments. The empirical study is conducted using the PHAC-2 dataset <cit.>. The dataset was collected using a Willow Garage PR2 equipped with a pair of SynTouch BioTac sensors to explore a set of objects using a set of finger-tip motions (see Fig. <ref>). To focus on the benefit of tactile signals, rather than measuring classification performance, we propose to evaluate the information quality of the tactile signals, i.e., measurements indicating how tactile signals could potentially perform in downstream tasks such as the adjective classification task, but avoiding the influence of classifiers which might introduce differences or biases due to method selection or parameter tuning. Detailed method design is introduced in Section <ref>, followed by experiments (Section <ref>) seeking answers to the aforementioned questions. This paper makes three major contributions: * a new method is proposed to re-evaluate finger-tip tactile data for inferring object adjectives from an information quality rather than classification performance perspective, * comprehensive experiments are conducted to study the benefit of finger-tip tactile sensing from the perspectives of taxel distribution, taxel spatial density and measurement temporal density, and * important insights are obtained to guide the usage and design of finger-tip tactile sensors in related tasks. § RELATED WORKS There are a number of works exploring tactile sensing for inferring object adjectives <cit.>, many of which use the PHAC-2 dataset <cit.>. Early solutions use hand-crafted features together with an SVM for adjective classification <cit.>, achieving promising performance. Better performance was then achieved by various learning approaches <cit.>. Approaches were also proposed to fuse vision and tactile for adjective classification, showing that the fusion of the two modalities improves the performance <cit.>. The current State-of-The-Art classification performance was achieved by a method using learned features <cit.>. More recently, a learning method was proposed to predict perceptual distributions of haptic adjectives <cit.>, demonstrating the feasibility of modeling both the intensity and the variation of tactile perception. All these methods enabled better adjective classification or perceptual distribution prediction. Some of them also studied how different types of signals and different exploratory procedures contributed to adjective classification. However, no study has been done at the taxel level. It remains unknown how much each taxel contributes to inferring the object adjectives. § METHODOLOGY The exploration of how the finger-tip tactile information contributes to a robotic task is conducted by ablating the tactile signals and evaluating information quality measures such as distinguishability which is critical for downstream tasks such as the adjective classification <cit.>. Signals with a higher distinguishability have a more distinct distribution in their feature space, and are therefore better for inferring object adjectives (e.g., easier to classify). The tactile signals are ablated from the perspectives of taxel distribution, taxel spatial density, and measurement temporal density. §.§ Dataset The study is conducted using the tactile adjective classification dataset Penn Haptic Adjective Corpus (PHAC-2) <cit.>. The dataset contains 600 trials collected using a Willow Garage PR2 equipped with two SynTouch BioTac tactile finger-tip sensors as shown in Fig. <ref>. The trials were recorded by using the robot to perform four exploratory procedures (EPs): “squeeze”, “hold”, “slow slide”, and “fast slide” on 60 different objects, with 10 trials of each. A tactile signal frame 𝐕 contains signals from both left (𝐕_L) and right (𝐕_R) BioTac sensors, each of which comprises 19 spatially distributed electrodes (taxels) 𝐄∈ℝ^19 and four scalars: vibration (P_AC), pressure (P_DC), heat flow (T_AC) and temperature (T_DC), i.e., 𝐕_L, 𝐕_R∈ℝ^23. The P_AC is measured at 2.2 kHz, while the other signals are measured at 100 Hz. The dataset was designed for adjective classification, but in this work it is used for tactile signal ablation and information quality evaluation in this work. The dataset has 19 binary adjective labels 𝐀∈𝔹^19 (the version <cit.> has 24 adjectives, but we follow the setup in <cit.> in this work and exclude those with fewer than three positively labeled objects). We use the same training and test sets as <cit.> and <cit.>. §.§ Metrics The information quality (distinguishability) of tactile signals are evaluated using the commonly used mean Average Precision (mAP) <cit.> which directly indicates how signals could potentially perform in downstream tasks such as classification, but with no influence from classifiers which might introduce differences or biases due to method selection or parameter tuning. mAP is a good indication of how samples of different classes are distinctly distributed in the feature space. A larger mAP means a more distinct distribution of samples across different classes, and is therefore easier to classify. When comparing to existing classification baselines such as <cit.> and <cit.>, 𝐅_1 score <cit.> is used to measure classification performance in Section <ref>. §.§ Feature generation Different from <cit.> and <cit.> where hand-crafted or learned features were used, dimension-reduced raw tactile signals are directly used as features for information quality evaluation in this work to avoid unnecessary biases. The dimension reduction is performed via Principal Component Analysis (PCA) <cit.>. A sequence of tactile signals from one sensor from one exploratory procedure are first concatenated into one vector 𝐙_k = [𝐕_k,1, 𝐕_k,2, ⋯, 𝐕_k,N] ∈ℝ^23N, k ∈ [L,R], where N is the total number of frames in one sequence. Then 𝐙_L and 𝐙_R from the two fingers are concatenated to 𝐙∈ℝ^46I. PCA is then applied to 𝐙∈ℝ^46I. The top 200 components are used as the features representing each sequence for information quality evaluation. Evaluation experiments in Section <ref> show that the top 200 components could maintain >99% of the input's variance. For the convenience of expression, the dimension-reduced features are denoted as 𝐙^200∈ℝ^200. Table <ref> shows the N values and feature dimensions for the tactile sequences collected from the four exploratory procedures. The N values we used are the maximum length of the sequences for each EP. The sequences with a shorter length are padded to the maximum length by repeating their first and last elements in a symmetric manner (the non-repeated fractions are in the middle of the padded signals). §.§ Classification method Although the main focus of this work is to explore the benefit of each tactile signal by ablating raw signals and observing their information quality changes, classification performance is also evaluated on the dimension-reduced features 𝐙^200 to see their feasibility by comparing to the initial method <cit.> (SVM with hand-crafted features, denoted as HF+SVM) and the current State-of-The-Art solution <cit.> (SVM with learned features, denoted as LF+SVM) for adjective classification with PHAC-2. Similarly to <cit.>, SVM is adopted to predict binary adjective labels. Nineteen SVMs are trained for each EP (one for each adjective). Cross-validation and grid search are used to select the SVM regularization parameter C with values ranging from 10^-4 to 10^4, tolerance tol for the stopping criterion with values ranging from 10^-6 to 10, and the kernel either linear or RBF <cit.>. Similar to the settings in <cit.>, each classifier is trained using the L_2 metric <cit.> with the 𝐅_1 score averaged over 100 cross-validation sets created by randomly selecting 3 objects from each training set. § EXPERIMENTAL RESULTS Experiments are first conducted to see the feasibility of the dimension-reduced features by evaluating their classification performance with an SVM as introduced in Section <ref>. This is followed by ablation studies to show the benefits of PCA. Then experiments are conducted to study how the individual tactile signals contribute to or influence the adjective inferring performance. §.§ Classification performance The classification performance of SVM with the dimension-reduced features (denoted as PCA+SVM) was evaluated by comparing to HF+SVM and LF+SVM using 𝐅_1 scores. mAP was not compared as there were no mAP results presented in <cit.> and <cit.>. The setups for the dataset, PCA and SVM were introduced in Section <ref>. Fig. <ref> shows the boxplots of the 𝐅_1 scores of the three methods. We can see that PCA+SVM performs better than HF+SVM, but worse than LF+SVM across all EPs. Considering that the three works all use SVM-based classifiers with similar settings, this result indicates that the features produced by the proposed PCA-based dimension reduction method works better than the hand-crafted features, but worse than the learned features. Detailed 𝐅_1 scores are shown in Fig. <ref>, from which we can see that PCA+SVM has a performance in between in most cases (42/76), with 54 cases better than HF+SVM and 12 cases being the best. This is also indicated by the average 𝐅_1 scores listed in Table <ref> where PCA+SVM has an in-between average score across all EPs. These results demonstrate the feasibility of the dimension-reduced features generated directly from the raw tactile signals via PCA. Similarly to <cit.> and <cit.>, we can also observe from Fig. <ref> that the tactile signals collected by the four EPs are not informative enough for some texture-related adjectives (whose best 𝐅_1 scores are below 0.75) such as: “smooth”, “textured”, “rough”, “fuzzy”, “scratchy”, and “hairy”. As mentioned in <cit.>, this poor performance might be related to the degradation of the ridges on the BioTac sensor skin during data collection. The best scores of “cool” and “springy” are also below 0.75, which might also be related to similar issues (not much useful information was obtained in data collection). From the EP perspective, “squeeze” is the most informative (𝐅_1 scores above 0.75 in 10 adjectives); then “hold” comes second (performs good in 8 adjectives); the two sliding EPs only achieve acceptable performance in 7 adjectives. §.§ Effect of PCA To study how the proposed PCA-based dimension reduction method influences information quality, comparisons were conducted on features produced with different PCA component settings. All the rest experiments are evaluated from information quality perspective using mAP. All dataset and PCA setups were the same as in Section <ref>, except the number of components for PCA. Fig. <ref> shows the results, from which we can see that signals with more components have slightly better mAP performance up to 200 components. PCA with 200 components will be used in the remainder of the experiments due to its lower computational cost (dimensions reduced by more than 99.7% as shown in Table <ref>) and no obvious performance loss. §.§ Benefit from each tactile sensor As introduced in Section <ref>, the dataset was collected using a pair of BioTac sensors. This section evaluates how each tactile sensor contributes to the performance in inferring object adjectives. All setups were the same as in Section <ref> except the signals to which PCA was applied. Fig. <ref> shows the mAP results of left finger (𝐙^200_L), right finger (𝐙^200_R) and both fingers (𝐙^200). We can see that the performance of “Both Fingers” is a combination of “Finger L” and “Finger R” across all EPs. All three cases have very similar performance although with some differences, which can also be observed from the average mAP scores in Table <ref>. This indicates that one tactile sensor is sufficient for symmetric use cases (symmetric interaction motions on symmetric objects/surfaces). §.§ Benefits from taxels and other signals Similar to <cit.>, comparisons were conducted on different types of signals, but this work focuses more on electrode versus other signals. All setups were the same as in Section <ref> except the signals which the PCA was applied to. Fig. <ref> shows the mAP results of signals with all four scalars and taxels (“All-PCA200”) and pure taxel signals (“Ele-PCA200”). The performance of the raw tactile signals with no PCA dimension reduction was also evaluated as a reference (“Original”). From Fig. <ref> we can observe that the performance of “All-PCA200” is even slightly better than “Original” although with greatly fewer dimensions. This indicates that the PCA dimension reduction improves the feature quality. The overall performance of “Ele-PCA200” is comparable to “All-PCA200” in most EPs except for “slow slide” where it is worse than “All-PCA200”. This can be observed more clearly in Fig. <ref>, from which we can also see that: all texture-related (“smooth”, “textured”, “soft”, “absorbent”, “rough”, `slippery”, “fuzzy”, “porous”, “scratchy”, and “hairy”) or temperature-related (“cool” and “metallic”) adjectives have worse performance across at least 3 EPs; the others have better performance across at least 2 EPs. This is unsurprising as the four scalar signals contain vibration and temperature related information which is particularly helpful for those adjectives, but not that helpful for the rest (sometimes distracting, such as for “solid”, “squishy”, “compressible” and “hard”). §.§ Benefits of different taxels To explore how each taxel contributes to inferring object adjectives by providing information with a higher quality, comparisons were made on different taxels by dividing the electrodes into 5 regions shown in Fig. <ref>a. The original electrodes (“All regions”) were also compared as a reference. All setups were the same as in Section <ref> except the signals to which the PCA was applied. The mAP results are shown in Fig. <ref>, from which we can observe that the original electrodes (“All regions”) only worked the best in four cases (“squishy” with “hold”, “soft” with “squeeze” and “slow slide”, and “rough” with “squeeze”); the best results of the rest were obtained by individual regions. The surprisingly poor performance of “All regions” indicates that not all electrodes contribute to inferring object adjectives. This might be caused by noisy or confusing signals from some regions for some adjectives and EPs. In particular, “Region 1 to 3” cover most of the best results, indicating that these regions contribute the most to inferring adjectives. “Region 1” tends to contribute more with “squeeze”, “Region 3” contributes more with the rest of three EPs, while “Region 2” contributes across all EPs. The contribution from “Region 4” is very limited (the best in only three cases: “bumpy” with “squeeze” and “hold” EPs, and “thick” with “hold”). “Region 5” contributes the most in the cases where contacts on either/both sides of the BioTac sensor are critical for inferring certain adjectives, such as “porous”, “scratchy” and “absorbent”. §.§ Influence of taxel spatial density This section evaluates how the taxel spatial density influences the performance in inferring object adjectives. The 19 electrode measurements from each sensor were first arranged into a 7×3 rectangular array using the same method in <cit.> as shown in Fig. <ref>b. The study was conducted by comparing the performance of signals spatially-downsampled from 7×3 (original electrodes) to 4×2 (by 2, denoted as SDS2) and 2×1 (by 4, denoted as SDS4). All setups were the same as in Section <ref> except the signals to which the PCA was applied. Fig. <ref> and Fig. <ref> show the results. From Fig. <ref> we can see that downsampling generally decreases the performance in most cases across all EPs. From adjective perspective (Fig. <ref>), “solid”, “squishy”, “compressible” and “hard” have obviously better performance with downsampled signals across 3 EPs except “hold”; most of the rest have obvious performance decreases across at least 3 EPs, particularly for those texture-related ones. This indicates that denser taxels are helpful for those texture-related adjectives, but can be distracting to those non-texture-related ones. §.§ Influence of measurement temporal density The study of measurement temporal density was conducted by comparing signals temporally downsampled by factors of 20, 50, 100, 200, and 400 (denoted as TDS20 to TDS400). The performance of the original electrodes was also compared as a reference (denoted as TDS1). All setups were the same as in Section <ref> except the signals to which the PCA was applied. Considering the significantly reduced number of frames in the downsampled sequences, the numbers of PCA components were set to 100 for TDS200 and 50 for TDS400, which maintained >99% of the input's variance. Table <ref> shows the average mAP results, from which we can observe that signals collected by “squeeze” have obvious performance degradation (⩾0.01) from TDS400, “slow slide” from TDS200, and “fast slide” from TDS100; while “hold” has no obvious performance changes. This trend can also be observed in Fig. <ref>. When paying attention to the numbers of frames in the downsampled sequences of each EP in Table <ref>, we can also observe that the performance starts decreasing at around 10 for “squeeze” (12) and “slow slide” (9), and 18 for “fast slide”. There is no obvious performance difference for the cases with smaller downsampling factors. All these observations indicate that much smaller frame rates can satisfy the demand of inferring object adjectives, although with differences for some adjectives and EPs. For those exploratory procedures with the possibility of obtaining more information, more frames are required to keep the information. From Fig. <ref> we can also see that the performance changes of “solid”, “squishy”, “compressible” and “hard” with “squeeze” are much more significant (⩾0.1) than the others. This indicates that the contributions from the temporal information in these signals are critical for inferring stiffness-related adjectives, and the minimum numbers of frames are around 24 to keep the information. § CONCLUSIONS AND DISCUSSIONS This paper has presented a comprehensive empirical study on how parallel finger-tip tactile contributes to inferring object adjectives using the PHAC-2 dataset. The study was mainly conducted to explore the influence of taxel spatial density, measurement temporal density, and taxel distribution. Experimental results showed that one tactile sensor is sufficient for symmetric use cases (symmetric interaction motions on symmetric objects/surfaces) such as most of the data collected in the PHAC-2 dataset. The two tactile sensors on the PR2 robot have similar performance individually or combined. Both spatial and temporal information from taxels contributes to inferring object adjectives. Dense taxels are beneficial for texture-related adjectives, but can be distracting to non-texture-related ones. The contributions from the temporal information are critical for inferring stiffness-related adjectives such as “solid”, “squishy”, “compressible” and “hard”. The frame-rate of the BioTac sensor is more than sufficient for inferring object adjectives. Sensors with much smaller frame rates can satisfy the demand. The vibration and temperature related signals from the BioTac sensor have significant contributions to the inferring of textured-related or temperature-related adjectives, but can also be distracting for some adjectives such as stiffness-related features. In terms of taxel distribution, not all taxels benefit the inferring of object adjectives: the electrodes on the belly of the sensor make the largest contribution to inferring the most object adjectives, while the electrodes on both sides of the sensor are critical for inferring the adjectives such as “porous”, “scratchy” and “absorbent”, where deep contact (sensor mostly wrapped by the contact surface) between the sensor and object is needed. § ACKNOWLEDGMENT The authors acknowledge continued support from the Queensland University of Technology (QUT) through the Centre for Robotics, and thank the support and help from Dr Ben Richardson for the use of the PHAC-2 dataset. Support from the GentleMAN (RCN 299757) is also greatly acknowledged. Computational resources and services used in this work were provided by the eResearch Office, Queensland University of Technology, Brisbane, Australia. IEEEtran

http://arxiv.org/abs/2303.06628v1

20230312102807

Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models

cs.CV

Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models Zangwei Zheng^1 Mingyuan Ma^2 Kai Wang^1 Ziheng Qin^1 Xiangyu Yue^3 Yang You^1 ^1National University of Singapore ^2UC Berkeley ^3The Chinese University of Hong Kong 1{zangwei, kai.wang, zihengq, youy}@comp.nus.edu.sg 2mamingyuan2001@berkeley.edu 3xyyue@ie.cuhk.edu.hk ================================================================================================================================================================================================================================================================================================== empty Continual learning (CL) can help pre-trained vision-language models efficiently adapt to new or under-trained data distributions without re-training. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by 9.7% average score. Our code locates at <https://github.com/Thunderbeee/ZSCL>. § INTRODUCTION Most deep learning models can access all the data during training <cit.>. If we want to expand a model's knowledge, such as learning a newly found animal species <cit.>, we can re-train the model by adding new classes to the training dataset. However, re-training a large model is costly. In contrast, continual learning (CL) <cit.> incrementally learns task one after another. It can reduce this cost by only learning the new data. Nonetheless, a model tends to forget previous information catastrophically when learning new tasks <cit.>. The “catastrophic forgetting” phenomenon is a great challenge for CL. Recently, vision-language models have shown powerful zero-shot transfer ability <cit.>. They can give zero-shot predictions without any training examples of a task. However, the performance on some tasks is poor due to insufficient relevant image-text pairs in the pre-training datasets. For example, it is difficult for CLIP <cit.> to distinguish among digital numbers, with an accuracy on MNIST <cit.> below 60% much lower than a naively trained CNN <cit.>. If we want to widen the knowledge in the vision-language model by re-training, the computational cost is too large (e.g., CLIP is pretrained on 400 million image-text pairs). Fine-tuning downstream tasks achieves high performance, but one model for a task takes much memory, and the model is not reusable. Prompt learning <cit.> keeps the backbone parameters unchanged. However, it is only effective with limited training data due to a limited prompt length <cit.>. In contrast, continual learning makes learning new knowledge a lifelong process for the vision-language model. The continually learned model can handle any image-text input and can be further used for downstream tasks <cit.>. We find that existing CL methods hardly prevent the forgetting phenomenon for zero-shot transfer ability in continual learning of a pre-trained vision-language model. As shown in <ref> (a), the CL with a pre-trained vision-language model differs from the traditional one. Besides forgetting previously learned task knowledge, the CLIP-based CL suffers from forgetting pre-training knowledge, namely a degradation of zero-shot transfer ability. For the replay-based CL methods <cit.>, the dataset during pre-training may be private and inaccessible during fine-tuning. For distillation-based CL methods <cit.>, they do not lay enough emphasis on the pre-trained model. On the one hand, a large model state change hinders tasks thereafter from using high-quality feature representations. On the other hand, it significantly degrades zero-shot performance on unseen datasets. Our method ZSCL protects the Zero-Shot transfer ability during Continual Learning. We view the knowledge stored in the pre-training model from two perspectives: a well-learned feature space and a good value in the parameter space. In feature space, we re-design previous distillation loss <cit.> with different loss styles, teacher models, and data sources. We find the original CLIP model, as opposed to the newly acquired model, is the best option for the teacher model. Instead of using data collected from previous tasks <cit.> or current task <cit.>, we find a reference dataset with diverse semantics (e.g., images sampled from ImageNet) is a good option for distillation loss. The reference images need not be labeled or matched with the text. Preserving the relative similarity between reference images and texts makes the feature space deviate little from the original. In the parameter space, WiSE-FT <cit.> proposes interpolating the initial and fine-tuned model for better performance. Inspired by this, we ensemble the weights throughout continuous training to prevent a significant shift from the initial CLIP, which can be seen as interpolating models of different zero-shot transfer and downstream task performance tradeoffs. The weight ensemble method is more stable and not sensitive to hyper-parameters. To better evaluate our method, we propose a new benchmark Multi-domain Task Incremental Learning (MTIL). Previous CL tasks are crafted by separating classes in one dataset <cit.>, where the images and classes are within a single domain. In contrast, MTIL consists of data from different sources requiring different expert knowledge. It comprises 11 tasks ranging from animal species to aircraft series recognition. As displayed in <ref> (b), when sequentially training CLIP on 11 datasets, the drop in the performance of task i after training task i is the traditional forgetting phenomenon. The degradation in the accuracy compared to the original zero-shot one before training task i represents the forgetting in zero-shot transfer ability. Our method better protects the zero-shot transfer ability and preserves the learned knowledge. We outperform previous methods in both conventional class-incremental learning and MTIL settings. In <ref> (b), To summarize, our contributions are as follows: * We investigate continual training with the vision-language model and demonstrate the importance of preserving zero-shot transfer ability. A more challenging benchmark MTIL is proposed to evaluate CL methods where the tasks come from distinct domains. * We propose a novel method ZSCL to mitigate the catastrophic forgetting problem in continual learning of the vision-language model by distillation in the feature space and weight ensemble in the parameter space. * The proposed ZSCL outperforms all state-of-the-art methods across multiple benchmark datasets. On 10 steps CL of CIFAR100 and TinyImageNet, our method outperforms the best of previous ones by 7.7% and 6.0% for the Last accuracy. On MTCL, ZSCL outperforms others by 10.9% on Transfer and 9.7% on Avg. scores. § RELATED WORK Vision-Language Models. Inspired by the success of language foundation models such as GPT-3 <cit.> and T5 <cit.>, a series of work pre-train vision-language models on large-scale image-text datasets <cit.>. Among them, Contrastive Language-Vision Pre-training <cit.> achieves remarkable performance on various downstream tasks. It concentrates on aligning images and texts to acquire a joint embedding space. The CLIP model contains an image encoder <cit.> and a text encoder <cit.>. During pre-training, contrastive learning is performed in which a paired image-text is a positive pair while image and text from different image-text pairs form a negative pair. For inference, the closest text embedding for the image is chosen as the prediction. Vision-language models can give zero-shot predictions on unseen tasks with a robust zero-shot transfer ability on various downstream tasks. Continual Learning Methods. Most existing continual learning methods can be categorized into four groups: parameter expansion, memory replay, distillation loss, and parameter regularization. Parameter expansion methods such as DyTox <cit.> and DEN <cit.> introduce new parameters for new tasks. As we want to achieve a more powerful CLIP model at the end of CL, we do not change the architecture of the CLIP model. Memory replay methods <cit.> including iCaRL <cit.> and SER <cit.> keep a memory for exemplars from previously learned tasks. However, pre-training datasets are too large for choosing exemplars or may not be available at downstream training, and downstream data are not good exemplars for preserving the pre-training knowledge. Distillation loss such as LwF <cit.>, LwM <cit.>, LwF-VR <cit.>, and PodNet <cit.> aligns current output space with previous ones, whereas distillation based on current tasks are not strong enough to maintain foundational knowledge. For the CLIP model, we find that general images, even if never seen by the model, plus sentences with enough semantics, can be a good “replay” for the distillation loss. The parameter regularization loss restricts the flexibility of model parameters by training loss <cit.> or weight averaging <cit.>. Although this type of strategy performs badly compared to other ways in previous research <cit.>, we found it helpful for CLIP continual learning. The limited parameter space prevents the model from diverging significantly from its original state. Vision-Language Models for Downstream Tasks. Many works propose different training strategies of vision-language models for better performance on downstream tasks, such as CoOp <cit.>, CLIP-Adapter <cit.> and WiSE-FT <cit.>. However, very few attempts at continual learning exist. Recently, Thengane  <cit.> shows CLIP zero-shot prediction achieves state-of-the-art performance in CL settings even without any training. LwF-VR <cit.> is a modified LwF method for the CLIP model where random-generated sentences are used for distillation loss. However, it only considers the feature space, and the distillation with random sentences cannot protect the vision backbone. Differently, we re-examine what should be used for distillation in the feature space and combine the parameter space weight ensemble to provide better performance for the vision-language model continual learning. § APPROACH §.§ Preliminaries Continual Learning. Given n tasks [𝒯^1, 𝒯^2, ⋯, 𝒯^n], continual training is conducted in sequence on each task 𝒯^i=(𝒟^i, C^i), i=1,…,n. Here, 𝒟^i represents the task dataset {(x^i_j, y^i_j}_j=1^N_i, where x^i_j is an image, y^i_j is a one-hot vector indicating the ground truth, and N_i is the number of images in the dataset. Class names C^i={c^i_j}_j=1^m_i maps the label of an image to an object name, where m_i is the number of classes for task 𝒯^i. The objective of continual learning is to achieve good performance on all tasks. Two continual learning settings are focused on in this study <cit.>. In task-incremental learning, at inference, the image x to be predicted is given with its task identity t, so the model only needs to distinguish between different classes in C^t. In class-incremental learning, the task identity t is not given. Thus we need to predict with the combined class set C=⋃_i=1^nC^i. CLIP model. The CLIP model contains an image encoder f_i and a text encoder f_t. The inference process of the CLIP model for image classification is as follows. First, for task 𝒯^i, each class c is transformed into a sentence by a template like “a photo of {c}”. Then f_t encodes the classes into text embeddings {t^i_j}_j=1^m_i. An image encoder encodes input image x_k. The similarity score between the image embedding and text embeddings are calculated as s^i_k,j=f_i(x_k),t^i_j, where ·,· denotes the cosine similarity. The class with the largest similarity score is the prediction for the image. To fine-tune the CLIP model for downstream tasks, cross-entropy loss CE is applied to the similarity score with a temperature scaling: ℒ_CE = 1/N∑_i=1^NCE(τ·s_i, y_i), where τ is a parameter learned during the pre-training. §.§ Distillation in Feature Space Well-learned feature space for aligned images and texts enables vision-language models' strong zero-shot transfer ability. It also facilitates the learning of downstream tasks. Compared with the pre-training dataset, downstream datasets lie in a small scope in the feature space (shown in <ref>(b)). Direct fine-tuning of downstream tasks greatly distorts the feature distribution of out-of-distribution data, which leads to a significant drop in zero-shot prediction performance. While the cross-entropy improves the performance by altering fine-tuned feature subspace, we need a new regularization to preserve the potential out-of-distribution feature space. The relative similarity between one image and different texts is: p=Softmax(s_1, ⋯, s_m). We hope the above similarity distribution is stable during fine-tuning for all potential images and texts. Given a teacher model f, distillation loss can be applied to penalize changes from the original distribution: ℒ_dist_img=CE(p,p)=-∑_j=1^mp_j·logp_j, where p is the distribution given by the teacher model. Although the distillation form has been widely used in previous methods <cit.>, they are applied to continual learning from scratch. We investigate different components of the distillation loss for enhancing pre-trained vision-language models. Three components are discussed in this paper in detail: the data source, the teacher model, and the loss design. <ref> shows the performance for the different choices. First, for the data source, LwF <cit.> uses data from the current task, while iCarl <cit.> carefully selects data from previous tasks. However, data from downstream tasks span a small subspace and are not good enough to preserve the whole feature space. An ideal choice is the pre-training dataset. However, the CLIP pre-training dataset is private, and the size is too large to load. To solve this challenge, we introduce the reference dataset. A reference dataset is a publicly available image dataset with enough semantics. Enough semantics can be seen as random sampling in the whole feature space. The texts can be related class names, unrelated sentences, or even random tokens. This is because text semantics are easier to sample, and we need not ground truth to calculate the distance between one image to sufficient text samples spread among feature space. For the teacher model, <cit.> use the model after learning task i-1 and before learning task i as the teacher model. During the continual training of the vision-language model, the feature space deviates gradually from the initial model. Using fine-tuned models as teacher models enlarges this change. In contrast, we find using the pre-trained model as a teacher model not only preserves the zero-shot transfer ability but also takes advantage of well-learned feature space for better downstream performance. Finally, previous distillation loss is applied on traditional backbones, where a classification head gives the probability for different labels. For the vision-language model, the probability is calculated based on the relative distance between images and texts. Thus, in addition to <ref>, we impose regularization ℒ_dist_txt on the distances from a text to a batch of images. The whole framework is shown in <ref> (a) with the following training loss: ℒ = ℒ_ce+λ· (ℒ_lwf_img + ℒ_lwf_txt). §.§ Weight Ensemble in Parameter Space Machine learning models integrate learned knowledge in their parameters. To mitigate the forgetting problem, a series of works <cit.> impose regularization losses on the changes of parameters. Weight consolidation (WC) <cit.> imposes the following loss: ℒ_WC=∑_i(θ_i - θ_i)^2. where θ is the parameters of the current model, and θ is the reference ones. Although this regularization prevents forgetting, it hinders learning new tasks in a challenging CL setting. Apart from regularization losses, another method in parameter space is to ensemble different model weights. Model soup <cit.> averages weights of multiple fine-tuned models to improve the model's robustness but introduces additional training costs. WiSE-FT <cit.> propose a weighted average between fine-tuned model and the original model to improve the out-of-distribution prediction robustness: f(x; (1-α)·θ_0 + α·θ_1), where θ_0 is the original model and θ_1 is the fine-tuned one. However, this method is hyper-parameter-sensitive where different α gives different tradeoffs between zero-shot transfer ability and downstream task performance (blue line in <ref>. Inspired by this, we extend the weighted average to the CL setting. The motivation for the weighted average is to prevent fine-tuning from losing too much knowledge in the original model. As training goes by (green line in <ref>), the model performs better on new tasks while losing accuracy on previous ones. The models among training compose a sequence of different learning-forgetting tradeoffs. Instead of interpolating only between the initial and the final model, our method weight ensemble (WE) averages the weights in the sequence during the training time: θ̂_t = θ_0 t=0 1/t+1θ_t + t/t+1·θ̂_t-1 every I iterations. where model weight sampling happens every I iteration. The method is related to Stochastic Weight Averaging (SWA) <cit.>, but we do not use a modified learning rate schedule here because instead of getting better generalization ability, WE aim to give an improved learning-forgetting tradeoff. As shown in <ref>, WE achieves better performance on downstream tasks than WiSE-FT. In addition, while WiSE-FT is sensitive to different values of α, our method is much more robust under different hyper-parameter (I) choices. § MULTI-DOMAIN TASK INCREMENTAL LEARNING Conventional Continual Learning Benchmark. A benchmark consisting of several tasks is needed to evaluate different methods for continual learning. Most previous benchmarks are built by separating classes in a single dataset, such as MNIST <cit.>, CIFAR100 <cit.>, TinyImageNet <cit.>, and ImageNet <cit.>. We also evaluate our method with traditional benchmarks. In CIFAR100 <cit.>, classes are separated into groups to build tasks. Suppose the dataset has m classes, a k-step setting means we learn m/k classes in each new task. The CIFAR100 contains 100 classes, and the setting of 10, 20, and 50 steps are visited. For TinyImageNet with m=200, the first step learns 100 classes, and the rest is learned with 5, 10, and 20 steps. As for ImageNet-100, we have two settings: ImageNet-100-B0, which includes the same amount of classes for each step, and ImageNet-100-B50, which has 50 classes for the first step, and the remaining 50 classes are observed progressively over the next 10 stages. For ImageNet <cit.>, we investigate a 10-step setting, which learns 100 new classes per task. MTIL Benchmark. Different classes from one dataset share the common image source and a similar style <cit.>. Thus, we propose Multi-domain Task Incremental Learning (MTIL), a cross-domain version of task incremental learning. Different tasks are collected from different domains, requiring different domain knowledge for humans to achieve high accuracy. Our MTIL benchmark consists of 11 tasks (detailed in supplementary materials), as some of the tasks illustrated in <ref> (a). The MTIL benchmark is very challenging with a total number of 1,201 classes. We propose two training orders detailed in the appendix, and by default, the results are given under training order-I. Evaluation Metrics. The metrics of MTIL are illustrated in <ref>(b), where rows represent training steps and a column shows performance for one dataset. For conventional continual learning, only scores under the diagonal are meaningful, since they cannot give zero-shot predictions on unseen tasks. In contrast, the zeros-hot transfer ability enables a vision-language model to provide predictions for all datasets. The average accuracy on all datasets among all timestamps is the “Avg” metric. The “Last” metric is the average performance of all tasks after CL. The “Transfer” metric is the average task performance in the upper-right triangle of the matrix. Every task's performance is first averaged to equal the weight of each dataset. It measures to what extent the zero-shot transfer ability is preserved. Before learning task i, tasks not earlier than i are not fine-tuned. Thus, their performance is an indicator of zero-shot transfer ability. § EXPERIMENTS Implementation. We use CLIP <cit.> model with image encoder ViT-B/16 <cit.>. We conduct training with AdamW <cit.> optimizer and use a label smoothing <cit.> technique for a better baseline result. For multi-domain task continual learning, we train 1K iterations for each task, while for class-incremental learning, we follow the same evaluation setting in <cit.>. More implementation details can be found in the supplementary material. §.§ Main Properties We ablate our method in feature-space in <ref>, and different choices for parameter-space regularization in <ref>. Several interesting characteristics are noted. Continual learning loss. In <ref>, several types of loss for feature space are tested. The Feature Distance penalizing absolute distances achieves a low accuracy. Distillation loss on probability distribution regularizes relative distance in the feature space. With image-only or text-only distillation, the Transfer, Avg., and Last accuracy all improve. A further boost in performance occurs with both of the distillation losses. Data source for replay. <ref> seek the standard for a good reference dataset. As shown in <ref>, distillation on images of current tasks achieves a good Transfer score. However, it hinders the learning on new tasks and results in a low Avg. and Last score. Images in a specific domain (, Flowers) are also not good choices. General images in ImageNet and Conceptual Caption (CC) datasets are examples of good reference datasets. These images are easily available by a web crawler <cit.>. Text with more semantics can improve performance (shown in <ref>). When using sentences from the Conceptual Caption dataset <cit.>, or even sentences randomly generated from the CLIP vocabulary, there is no ground truth target among the texts for the image from the ImageNet dataset. However, they all achieve an improvement due to more semantics. The reference image dataset does not need to be labeled, matched with the text, or seen by the CLIP model. Enough semantics in the reference image dataset boosts the distillation performance. In <ref>, a smaller number of images and classes all lead to a degradation in the performance. Fewer classes of images in the reference dataset have a worse impact on the performance compared with the image numbers. To keep a reasonable memory buffer, we randomly sample 100k images from ImageNet for MTIL and use texts from CC dataset. For class-incremental learning, conceptual caption dataset's validation set (28k images) is used to avoid information leakage. Teacher model. <ref> shows the performance of distillation loss with different teacher models. Unlike conventional continual learning, the teacher model should not be the one trained on the previous task; otherwise, the deviation from the initial CLIP in the previous task may be amplified. In contrast, with the initial CLIP as the teacher model, not only is the zero-shot performance improved but the Mean and Last scores are also boosted, indicating that preserving high-quality feature space facilitates continual learning. Parameter-space regularization. In <ref>, we experiment with three different parameter-space regularizations. We experiment two variants of WiSE-FT <cit.>: the weighted average between the current model with the initial CLIP or the one at the previous task. The result shows the latter one is a better choice because keeping weight averaging with initial CLIP loses the newly learned knowledge. We experiment with different α choices for WiSE, and the best result is reported. While distillation loss improves the whole performance, the parameter-space regularization further protects the zero-shot transfer ability with a higher Transfer. Among the three parameter-space regularization, only WE achieves a better Last score. WC greatly improves the Transfer scores with a lower Last score. A combination of the weight consolidation loss with weight ensemble achieves a better tradeoff between Transfer and Last value. While ZSCL^*, a variant without the WC loss, obtains the highest Last score, the ZSCL with WC loss outperforms it with 2.8% Avg. and 5.9% Transfer scores. §.§ Multi-domain Task Incremental Learning <ref> displays the performance of different methods on the MTIL benchmark, and <ref> presents the detailed Transfer, Avg, and Last metrics on each dataset. Zero-shot denotes the zero-shot prediction performance of the initial CLIP model, and Fine-tune means the direct fine-tuning accuracy on each dataset, which can be seen as an upper-bound where no forgetting phenomenon happens. Continual learning uses cross-entropy loss to learn each dataset sequentially, where there is a significant forgetting issue on both zero-shot predictions (Transfer drops by 24.8%) and newly learned knowledge (Last drops by 9.4%). Previous methods improve the Last performance slightly and cannot maintain a high zero-shot prediction performance. Without WC, ZSCL^* achieves the best Last scores, outperform previous best one by 4.4%. Our method ZSCL improves 9.1% on Transfer accuracy, with only 1.3% drops compared with the initial CLIP model, and achieves a 10.1% gain in the Avg. accuracy. §.§ Class Incremental Learning We evaluate our methods on conventional class incremental learning, which shows our method's ability in learning new tasks and preventing previously learned knowledge. <ref> and <ref> display the results on CIFAR100 and TinyImageNet, respectively. We re-implement some previous methods with a CLIP backbone (after CLIP in the table), while others using a special network design cannot be easily adapted. Although zero-shot CLIP prediction achieves a good result on these benchmarks, continual learning with direct fine-tuning or LwF <cit.> degrades the performance greatly, especially under the setting of a large step number. This demonstrates a severe catastrophic forgetting phenomenon in fine-tuning the CLIP model. Our method consistently improves the performance of the CLIP model on both Avg. and Last scores with a large gap towards previous ones. § CONCLUSION In this paper, we investigate continual learning with the vision-language model. We propose a better continual learning algorithm to protect the zero-shot transfer ability in the vision-language model learned in the pre-training stage. Our algorithm mitigates the catastrophic forgetting in both feature space and parameter space. In feature space, distilling the initial model on a reference dataset significantly boosts the model's performance. In parameter space, weight ensemble among different training stages alleviates the forgetting issue. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate the continual learning methods better. On both conventional and new benchmarks, our method achieves state-of-the-art performance. ieee_fullname Appendix § ADDITIONAL BENCHMARK DESCRIPTION <ref> displays the detailed information for different datasets in our benchmark. Two orders are used for the evaluation; the first one is alphabet order (Order-I): Aircraft, Caltech101, CIFAR100, DTD, EuroSAT, Flowers, Food, MNIST, OxfordPet, StanfordCars, SUN397. And the second one is a random order (Order-II): StanfordCars, Food, MNIST, OxfordPet, Flowers, SUN397, Aircraft, Caltech101, DTD, EuroSAT, CIFAR100. § ADDITIONAL IMPLEMENTATION DETAILS Our implementation is based on PyTorch <cit.>. We use batch size 64 for the MTIL benchmark and 128 for the class-incremental learning benchmark. The learning rates are searched among {10^-5,10^-6,10^-7}. Label smoothing <cit.> can substitute the regularization of weight decay and achieve better performance. the label smoothing strength is searched among {0.1,0.2,0.3}. In general, for MTIL, CIFAR100, and TinyImageNet, weight decay 0 and label smoothing 0.2 are good choices. For ImageNet, weight decay 0.1 and label smoothing 0 are used. We experiment with I∈{1,10,100} and find slight changes in performance and thus fix I to 100. A large λ hinders learning new knowledge, and λ=1 is a good choice. § ADDITIONAL MTIL RESULTS The complete result of the MTIL benchmark with t datasets is a matrix of t× t. It is difficult to compare different matrices between different methods, so we summarize the performance by three indicators in the main text. Here, we show the complete matrix of ZSCL in <ref> and ZSCL^* in <ref>. The result of the MTIL method in Order-II is presented in <ref>. Our method surpasses previous methods in another order setting of the MTIL benchmark. A similar conclusion holds from the results of MTIL Order-II compared with MTIL Order-I. Our method ZSCL outperforms others by 9.2% on the Avg. score and 18.1% on the Last score with only a 1.2% performance loss on the Transfer score. This shows our approach can work for different orders of the multi-domain task incremental learning. In addition, compared with Order-I, previous methods achieve a much lower Last score (e.g., for Continual-Learning, Order-I has 77.3%, while Order-II has 65.3%). With ZSCL, the Last score is similar (83.6% compared with 83.4%). This shows our method is more robust towards different training orders. § ADDITIONAL CONVENTIONAL CLASS INCREMENTAL LEARNING RESULTS We re-implement previous methods based on the CLIP backbone for continual learning. For LwF-based <cit.> methods, we experiment with two choices for the teacher model, the previous one and the initial one. We find the initial one gives out better performance and report this result. For the WiSE-FT <cit.> method, we take the average of models after learning each task. We experiment with the average between the previous and current one and the initial one and the current one. Better results are reported. The result of class-continual learning on ImageNet benchmark is presented in <ref>. On IN100-B10, our method outperforms others by 1.97% for the Avg. score and 1.30% for the Last score. On IN100-B50, our method outperforms others by 3.54% for the Avg. score and 6.62 for the Last score. § LIMITATION AND FUTURE WORK Our proposed method has a limitation that a reference dataset is needed. A promising direction of the work is to preserve the zero-shot transfer ability without the need for an outside dataset. For example, we may generate a synthetic image dataset as the reference dataset. Methods like <cit.> can synthesize datasets from a network. There is a trend in the deep learning community to build large models with a huge dataset <cit.>, including vision-language models <cit.>. These models serve as foundation models <cit.> for downstream tasks and require millions of dollars for training. As re-training cost upsurges, continual learning is an efficient approach for updating these models with custom usage. In some cases, we want to correct wrong information in the pre-training dataset or update old information with the latest one. How to conduct this task with a reference dataset is left as future work.

http://arxiv.org/abs/2303.07131v1

20230313140137

Evolutionary quantum feature selection

quant-ph

anton.simen@kipu-quantum.com Kipu Quantum, Berlin, Germany. Latin American Quantum Computing Center, SENAI CIMATEC, Salvador, Brazil. otto.pires@fbter.org.br Latin American Quantum Computing Center, SENAI CIMATEC, Salvador, Brazil. mauro.neto@fbter.org.br Latin American Quantum Computing Center, SENAI CIMATEC, Salvador, Brazil. erick.sperandio@surrey.ac.uk Surrey Institute for People-Centred Artificial Intelligence, University of Surrey, Guildford, United Kingdom Effective feature selection is essential for enhancing the performance of artificial intelligence models. It involves identifying feature combinations that optimize a given metric, but this is a challenging task due to the problem's exponential time complexity. In this study, we present an innovative heuristic called Evolutionary Quantum Feature Selection (EQFS) that employs the Quantum Circuit Evolution (QCE) algorithm. Our approach harnesses the unique capabilities of QCE, which utilizes shallow depth circuits to generate sparse probability distributions. Our computational experiments demonstrate that EQFS can identify good feature combinations with quadratic scaling in the number of features. To evaluate EQFS's performance, we counted the number of times a given classical model assesses the cost function for a specific metric, as a function of the number of generations. Evolutionary quantum feature selection Erick Giovani Sperandio Nascimento March 30, 2023 ====================================== § INTRODUCTION Quantum Feature Selection (QFS) is a novel approach to Feature Selection (FS) in Machine Learning (ML) that leverages principles of Quantum Computing (QC) to enhance the efficiency and effectiveness of traditional FS methods. The most informative features are typically selected in traditional FS methods based on their correlation with the target variable or their predictive power. However, these methods can struggle with high-dimensional datasets, a phenomenon known as the curse of dimensionality <cit.>. On the other hand, Evolutionary Algorithms (EAs) are a family of optimization algorithms that are inspired by the process of natural selection and evolution. These algorithms use a population of candidate solutions and iteratively improve them over generations through selection, recombination, and mutation operations. <cit.><cit.>. The rapid increase in the amount of data has made it challenging to keep up with the computational demands of traditional FS methods <cit.>. As a result, researchers have explored alternative perspectives such as quantum computing and evolutionary algorithms. A procedure was described for a novel FS algorithm based on a Quadratic Unconstrained Binary Optimization (QUBO) problem for reducing model complexity in Machine Learning <cit.>. The algorithms selects a specific number of features based on their importance and redundancy, and the direct approach used in the algorithm yields higher quality solutions compared to iterative or greddy methods. The QUBO problems are particularly interesting because they can be solved on Quantum Hardware, which is why the proposed algorithm was evaluated using a classical computer, a quantum gate computer, and a quantum annealer. The proposed FS algorithm based on QUBO is a promising approach to address the challenges posed by the growing amount of data in Machine Learning. Other study was realized by <cit.> that describes a variational quantum algorithm designed to solve unscontrained black box binary optimization problems, where the objective function is given as a black box. Unlike typical algorithms for optimization where a classical objetive function is provided as a Quandratic Uncontrained Binary Optimization problem and mapped toa sum of Pauli operators, this algorithm directly handles the black box objective function. The algorithm´s theorical justification is based on convergence guarantees of quantum imaginary time evolution. The authors demonstrated that the quantum method produced competitive, and in certain aspects, even better perfomance compared to traditional FS techniques used in today´s industry. This suggests that quantum algorithms could potentially offer significant advantages over classical methods in FS and other optimization problems. However, further research is necessary to explore the full capabilities of this approach and it´s potential applications in real-world scenarios. This paper focuses on the challenge of effective FS for artificial intelligence models due to exponential time complexity. To address this challenge, we propose an innovative heuristic called Evolutionary Quantum Feature Selection (EQFS) that uses the Quantum Circuit Evolution (QCE) algorithm. The QCE uses shalllow depth circuits to sparse probability distributions, because of this they can be useful to be applied in Noisy Intermediate-Scale Quantum(NISQ) devices, which EQFS harnesses to identify good feature combinations with quadratic scaling in the number of features. We evaluated EQFS´s perfomance by counting the number of times a given classical model assesses the cost function for a specific metric as a function of the number of generations. This work was organized as follows: In part <ref>, a brief description of the model used was made. In part <ref>, the results were discussed and finally in part <ref> the conclusions of this work. § QUANTUM FEATURE SELECTION The procedure that will be described to perform QFS uses a hybrid approach, where a quantum evolutionary algorithm plays the role of feature combination optimizer and works together with a classical algorithm that evaluates feature combinations in a supervised learning model. Let X be a dataset of dimensionality dim(X) = n. Each sample of X can be represented as an n-dimensional vector, 𝐯 with its associated vector of binary values, 𝐱 = (x_0, x_1, x_2, . .., x_n), which plays the role of indicating whether a variable will feed (x_i = 1) or not (x_i=0) the classical model. Let a metric function f(𝐱) evaluate the model quality given the 𝐱 combination of features. Let an initial quantum state be given by |ψ⟩ = |0⟩ ^⊗ n and the unit transformation, U|ψ⟩ = |ϕ⟩ being U generated by a quantum circuit that can be subjected to mutations over the generations (see Fig. <ref>). The QFS objective function can be writen as F(U) = ∑_𝐱|⟨𝐱|ϕ⟩|^2 f(𝐱). It is important to note that here we are considering that |ϕ⟩ is not a proper quantum state, but rather a vector of quasi-probabilities after a polynomial set of measurements on U|ψ⟩. Since we have 𝒪(2^n) possible solutions to the problem, a number of measures m = 𝒪(poly(n)) ensures that the approximate solution is found with a time complexity that scales polynomially with the number of variables. In this work, the algorithm used to evolve U is Quantum Circuit Evolution, proposed by <cit.>. At each generation, λ copies of U are created and a mutation operation with a respective probability is applied to each generation. Possible mutations are insert a new gate; modify a rotation angle of a single or two-qubit gate on the current circuit; delete one of the current gates and swap a two-qubit gate (flip target and control). Each of these mutations has its respective probability of occurring on each of the λ copies of U and the μ best individuals are carried over to the next generation, characterizing an elitist procedure known in the literature as (μ + λ)EA. In order to estimate the number of times the objective function is evaluated, consider Ω to be the set of probability amplitudes derived from |ϕ⟩. If we take m as the number of measurements performed on the quantum circuit, we have dim Ω≥ m. However, because it is a heuristic whose initial and final generations have sparse states - since the quantum circuit starts with a small depth - we can consider that very possibly dim Ω≫ m. Given a number K of generations and λ copies, we empirically observe that, for a small k, the total number of model evaluations in each generation, ∑_i=0^λdimΩ_i, can be approximated by a linear function with dependence on k and with a fixed constant defined for m as ∑_i=0^λdimΩ_i = m/K k. Therefore, we can approximate the number t of times that the objective function, f(x), is evaluated by calculation the Area Under the Curve (AUC), given by t ≈∫_0^Km/K k dk. Note that to approximate the solution in reasonable time, we can choose m and K appropriately. The experiments showed that for m and K being 𝒪(linear(n)), the heuristic can already find better solutions than for 𝐱 containing all features. § RESULTS AND DISCUSSION The results of the proposed feature selection procedure using quantum computing are presented and analyzed in this section. This method aims to address the challenges faced by classical feature selection algorithms in handling high dimensional datasets. The procedure is designed to improve the accuracy and efficiency of feature selection. Experimental results are presented to demonstrate the effectiveness of the proposed method and its comparison to classical methods. To carry out the experiments, a labeled data set of dimension n=13 (number of qubits) which uses chemical features to determine the origin of wines <cit.> was used. We adopted the elitist scheme (6+1)EA with the following mutation probabilities: 50 % to insert; 30 % to modify; 10 % delete and 10 % swap. The total number of measurements performed on the quantum circuit was m=64, that is, m ≈ 5n. Note that for n=13, dim(|ψ⟩) = 2^n = 8192, so m/2^n ≈ 7.8e-3. The metric used for f(𝐱) is the test accuracy of the Support Vector Classifier (SVC) model with a linear kernel function. The unseen labeled data used for testing is 20 % of the total data. Fig. <ref> shows the behavior of the quality of the solutions over the course of K=12 generations. The same experiment was performed 10 times for statiscal analysis purposes. In view of this, it was found that even for a small number, m, of measures and few generations of evolution of the circuit, the EQFS can find several combinations of features whose metric, f(𝐱), exceeds the case where the data set is used entirely. The best final distribution of the best individual among the 10 experiments can be seen in Fig. <ref>. From this distribution, we obtained several different combinations of features with test accuracy superior to the case where all features are used (88.8%). §.§ Model Evaluations The experimental evaluation of the model's performance was based on the total number of times that f(𝐱) is evaluated, since this is a critical point of the algorithm. In Fig. <ref> it can be seen that for a small number of generations, K, the total number of evaluations - for all λ individuals - from f(𝐱) grows linearly as a function of the number of generations (with m=64). The Area Under Curve (AUC) for the mean value of evaluations was AUC=434. This value shows that for a small constant c n, the number of evaluations needed to find values equal to the quality of the presented experiments is 𝒪(c n^2). §.§ Quantum circuit depth The depth of the quantum circuit for the number of generations K obviously cannot exceed this value. Our experiments showed that the circuit depth - considering the base B = {RX, RY, RZ, RXX, RYY, RZZ} - reached the average value of p=3, over all executions. The first run generated the circuit of Fig. <ref>. The Quantum Circuit Evolution heuristic has shown promise for solving combinatorial optimization problems in quantum computers, mainly due to the production of shallow quantum circuits. Fig. <ref> showed that for the feature selection task this heuristic also produced circuits with low depth and a reduced level of entanglement (only 3 operations that generate entanglement between qubits), which further corroborates the suitability of this algorithm for NISQ computers, since two-qubit quantum gates have a lower fidelity than single-qubit gates. Given the mutation probabilities chosen for the experiment, which are 30%, 10%, 10% and 10% for insert, delete, swap and modify, respectively, we can see that the sum of the success rates of the delete, swap and modify operations - which do not contribute to increasing the loop depth - is greater than that of the insert operation, since p =3 is 1/4 of k=12. The worst case for loop depth would be p=K, for a scenario where insert would have a 100% success rate. § CONCLUSIONS In this study, we introduced a novel approach for feature selection based on Quantum Circuit Evolution (QCE) algorithm. Our results demonstrate that our procedure, Evolutionary Quantum Feature Selection (EQFS), can identify good feature combinations with a quadratic number of model evaluations. Additionally, we observed that the depth of the quantum circuits generated by EQFS was shallow and produced circuits with a small entanglement degree compared to their variational counterparts. The effectiveness of our method highlights its potential to pave the way for the practical application of quantum computers in feature selection. With our findings, we hope to encourage further research in this area, as the potential impact of quantum computing on feature selection and other machine learning tasks could be significant. Our work contributes to the growing body of knowledge on quantum algorithms for machine learning and provides a promising new avenue for future research. § ACKNOWLEDGEMENTS The authors would like to thank Banco Votorantim (BV) for providing resources for the project from which this article is derived, as well as for contributions and discussions throughout the work. Acknowledgements also to the Supercomputing Center for Industrial Innovation (CS2i), the Reference Center for Artificial Intelligence (CRIA), and the Latin American Quantum Computing Center (LAQCC), all from SENAI CIMATEC.

http://arxiv.org/abs/2303.06657v1

20230312131305

Color Mismatches in Stereoscopic Video: Real-World Dataset and Deep Correction Method

cs.CV

Color Mismatches in Stereoscopic Video: Real-World Dataset and Deep Correction Method This study was supported by Russian Science Foundation under grant 22-21-00478, <https://rscf.ru/en/project/22-21-00478/> Egor Chistov,1 Nikita Alutis,1 Maxim Velikanov,1 and Dmitriy Vatolin12 Lomonosov Moscow State University1 MSU Institute for Artificial Intelligence2 =================================================================================================================================================================================================================================================================== We propose a real-world dataset of stereoscopic videos for color-mismatch correction. It includes real-world distortions achieved using a beam splitter. Our dataset is larger than any other for this task. We compared eight color-mismatch-correction methods on artificial and real-world datasets and showed that local methods are best suited to artificial distortions and that global methods are best suited to real-world distortions. Our efforts improved on the latest local neural-network method for color-mismatch correction in stereoscopic images, making it work faster and better on both artificial and real-world distortions. color mismatches, stereoscopic video, beam splitter, real-world video dataset, color transfer § INTRODUCTION The left and right views of a stereoscopic image (stereopair) can have color mismatches for various reasons (e.g., illuminated camera filters, glare, and polarized light); Fig. <ref> shows example. These mismatches can decrease the overall stereoscopic-video quality and may cause viewer discomfort and headaches <cit.>. Color-mismatch correction is the task of transferring color from one view of a stereopair to corresponding areas in another where the colors differ incorrectly. A sufficiently large dataset and sampling of color-mismatch types are crucial to comparing color-mismatch-correction methods. Previous datasets for this task either are too small or lack real-world mismatch examples. Therefore, we prepared a dataset using a beam splitter and three cameras that each captured a different view simultaneously: a distorted left view, a ground-truth left view, and a right view. In addition to containing real-world distortions, our dataset is thus far the largest for this task. Despite the rapid development of deep learning, the latest neural-network method for correcting color mismatches <cit.> is still inferior to conventional alternatives. We improved it by making it run faster and better on both artificial and real-world distortions. In this paper, we propose a real-world dataset of stereoscopic videos for color-mismatch correction as well as a method for color-mismatch correction of stereoscopic images. We also compared eight color-mismatch-correction methods on an artificial dataset and our real-world dataset. The code and datasets are at <https://github.com/egorchistov/color-transfer/>. § RELATED WORK This section provides an overview of stereoscopic datasets and methods for color-mismatch correction. §.§ Color-Mismatch-Correction Datasets Few datasets are available for color-mismatch correction. They are either too small or lack real-world mismatches. Niu et al. <cit.> employed 2D videos with parallel camera motion to extract pseudo-stereopairs without color mismatches; they gathered 18 source stereopairs. Using Photoshop CS6, they applied saturation, brightness, contrast, hue, color-balance, and exposure operators—each with three severity levels—to one view of the stereopair. Lavrushkin et al. <cit.> proposed gathering frames from stereoscopic movies produced only via 3D rendering to obtain undistorted stereopairs. They employed 1,000 FullHD frames. After adding simplex noise to one view of the stereopair, they smoothed the result by applying a domain transform filter. Grogan et al. <cit.> chose image pairs captured by one camera under different illumination conditions, camera settings, and color touch-up styles. They aligned reference images to match the target images. Their dataset contains 15 image pairs with real-world mismatches. The work of Croci et al. <cit.> used 1,035 stereopairs from several datasets: Flickr1024 <cit.>, InStereo2K <cit.>, and the Ivy Lab database of stereoscopic 3D images <cit.>. They applied various color operators—brightness, color balance, contrast, exposure, hue, and saturation, each with six severity levels—to one view of the stereopair using Photoshop 2021. §.§ Color-Mismatch-Correction Methods The color-transfer problem is solvable globally or locally. Global methods [Reinhard2001CGA, xiao2006color, Pitie2007IET] estimate a single color transformation for all pixels, and local methods [pitie2007automated, Lavrushkin20183DTV, Grogan2019CVIU, Croci2021ICIP] do so for each pixel. Global methods do poorly on complex mismatches, and local methods produce inconsistent results in low-texture regions. Reinhard et al. <cit.> proposed approximating an image as a three-dimensional Gaussian distributed signal. For each channel in the CIELAB color space, they used a linear model to transfer color from a reference image to the target image. They chose CIELAB because it is uncorrelated and allowed them to independently manipulate all three color channels. Xiao et al. <cit.> computed a covariance matrix between color channels instead of treating them independently, because that matrix is the extension of standard deviation in correlated space. The authors decomposed the covariance matrix using the SVD algorithm. They described the transformation as a scale, rotation, and shift of pixel clusters. Pitié et al. <cit.> generalized all linear color-transfer approaches to the covariance-matrix fitting. They showed that the matrix decomposition this fitting implies can use the Cholesky decomposition, the square-root decomposition, or the solution to the Monge-Kantorovitch problem. Pitié et al. <cit.> proposed using an iterative-distribution fitting to transfer color. This method iteratively projects target and reference images on random one-dimensional axes and performs a probability-density transfer along those axes. The authors reduced grain-noise artifacts by minimizing a special cost function. Lavrushkin et al. <cit.> performed stereo matching with a modified cost function, making it better able to handle color mismatches between views. They applied a guided filter with a confidence-based variable-length kernel, using the left view as the target image and a warped right view as the guidance image. Grogan et al. <cit.> proposed color transfer through estimation of the warping function that minimizes the divergence between two probability-density functions. Croci et al. <cit.> employed a convolutional neural network for color-mismatch correction. First, the network extracts features from the input stereopair. It then feeds the extracted features into the parallax-attention mechanism <cit.>, which performs stereo matching. Matched features pass through six residual blocks to yield the corrected stereoscopic view. § PROPOSED DATASET We propose a new real-world dataset of stereoscopic videos for evaluating color-mismatch-correction methods. We collected it using a beam splitter and three cameras that simultaneously capture three views of a scene: a distorted left view, the ground-truth left view, and a right view. §.§ Beam Splitter A beam splitter is an optical device that divides light into a transmitted beam and a reflected beam. Because some light is lost owing to absorption by the reflective coating, this device introduces real-world mismatches between stereopair views. Similar mismatches can appear in stereoscopic movies filmed with a beam-splitter rig. Our approach used a beam splitter to set a zero stereobase between the left camera and the left ground-truth camera, allowing us to create a distorted ground-truth data pair. A third camera captured the right view. Fig. <ref> shows our setup. We disabled optical stabilization and manually assigned all available camera settings, such as ISO, shutter speed, and color temperature. The cameras all had identical settings, so we obtained only beam-splitter distortions. §.§ Parallax Minimization Filming the left distorted view and left ground-truth view without parallax is crucial to achieving precise ground-truth data. Parallax is a difference in the apparent position of an object when viewed along two lines of sight. Postprocessing can precisely correct affine mismatches between views—namely scale, rotation angle, and translation—but it cannot correct parallax without using complex optical-flow, occlusion-estimation, and inpainting algorithms, which can reduce ground-truth-data quality. To minimize parallax, we first visually aligned the lenses of the left distorted and left ground-truth cameras to obtain a visual zero stereobase. Then, using a photo from one camera and a video stream from another, we visualized the squared error between the views. This visualization allowed us to maximize the overlapping region by manually moving one camera. We achieved zero parallax through this method, but the manual alignment was time consuming. §.§ Postprocessing Pipeline Our research employed the postprocessing pipeline in Fig. <ref>. Given the three videos from our cameras, we matched the horizontally flipped left distorted view to the left ground-truth view and rectified the stereopair using a homography transformation. To estimate the transformation parameters, we employed MAGSAC++ <cit.>. We selected SIFT <cit.> and the Brute-Force matcher to match the left views, as well as the LoFTR matcher <cit.> for rectification because it can better handle large parallax. The technique of Pitié et al. <cit.> perfectly corrected the small loss in the transmitted beam. Without it, our approach would have been unable to obtain accurate ground-truth data. We performed temporal alignment using audio streams captured by the cameras: superimposing one stream on the other in the video editor allowed us to find the offset in frames. §.§ Scene Selection Our research involved scenes in which an object rests on a table with a white background, 1.5 meters from the camera (Fig. <ref>). We filmed various objects including transparent glass, color patterns, and moving objects, choosing those that produced color mismatches. To ensure diversity, we filmed scenes under different lighting conditions and moving at different speeds. The lighting changed dynamically in one scene, and different scenes were shot with different set lighting. We recorded several scenes in 4K resolution using three GoPro Hero 9 cameras. For the rest we recorded FullHD on two Legria HF G-10 cameras, one for the left ground-truth view and one for the right view, as well as a Panasonic HDC-SD800 for the left distorted view. The result was 24 scenes, each 50 frames long. We then cropped the region of interest to 960x720 pixels. § PROPOSED METHOD Color-mismatch correction involves transferring color from one view of a stereopair to another in areas where colors incorrectly differ. Without limiting the task’s generality, we propose a deep neural network that transfers color from the right view to the left view. The left corrected view should have a structure consistent with that of the left original view and colors consistent with those of the right view. §.§ Overview We based our method on that of Croci et al. <cit.>, borrowing ideas from Wang et al. <cit.>. Our contribution is an effective multiscale network structure that works 2.6 times faster than Croci’s neural-network-based method <cit.> and, for artificial distortions, outperforms it by 3.7 dB on PSNR and, for real-world distortions, do so by 1.3 dB. Our method consists of three main modules: feature extraction, cascaded parallax attention, and transfer. Fig. <ref> shows the overall architecture. §.§ Feature-Extraction Module The left and right views feed into the feature-extraction module to yield their color and structural features at scales from 1 to 1/32. This module consists of six encoder blocks and three decoder blocks. Each is a residual block with batch normalization and weighted skip connection. For downsampling we used strided convolutions, and for upsampling we used strided transposed convolutions. For each scale we kept the channel count unchanged, as Fig. <ref> shows. §.§ Cascaded Parallax-Attention Module Features from scales 1/16, 1/8, and 1/4 feed into the cascaded parallax-attention module (CasPAM) to obtain a multiscale warping-attention map. This module consists of three parallax-attention modules (PAMs) and three interpolation layers. The interpolation layers allow us to obtain a warping-attention map at these scales whenever directly calculating an attention map is memory inefficient. Each PAM consist of four parallax-attention blocks (PABs) <cit.>. Every block updates the input features and converts them to a query and a key. After multiplying the query by the key, we added the result to the matching attention map. Our last step was to apply lower triangular softmax to this map and multiply the result by the right view’s features to obtain warped right view’s features. §.§ Transfer Module The transfer module concatenates the valid mask, the left-view, and warped-right-view features at scales from 1/32 to 1. A valid mask indicates pixels that match in another view. Then, for each scale we added upsampled features from the previous scale to features from the current scale and passed the result through residual block without batch normalization. The residual block’s output for scale 1 is the corrected left view. §.§ Loss Function We trained our network using the L1, L2, and SSIM loss functions applied to the corrected left view and ground-truth left view. We trained our cascaded parallax-attention module without ground-truth attention maps using the photometric, smoothness, and cycle-consistency loss functions, as Wang et al. <cit.> described. § EXPERIMENTS §.§ Artificial Dataset To compare our method with others we chose two datasets. The first dataset contained undistored stereopairs from Flickr1024 <cit.>, InStereo2K <cit.>, and the Ivy Lab database of stereoscopic 3D images <cit.>, following the method described by Croci et al. <cit.>. To exclude stereopairs with color mismatches, we analyzed them automatically using the technique proposed by Grokholsky et al. <cit.>. The final result was 1,035 undistorted stereopairs. Instead of employing Photoshop 2021 as Croci did, we applied augmentations to the left view of the stereopair. More specifically, our approach used the RandomBrightnessContrast, RandomGamma, and HueSaturationValue augmentations. We split the dataset into three parts: training (835 stereopairs), validation (100 stereopairs), and test (100 stereopairs). §.§ Real-World Dataset The second dataset, described in Section <ref>, contains the distortions we encountered while shooting stereoscopic videos using a beam splitter. We again divided the dataset into three parts: training (900 stereopairs), validation (150 stereopairs), and test (150 stereopairs). §.§ Training Procedure We trained our network with the Adam optimizer over 100 epochs on the artificial dataset using random patches of size 512x256. The learning rate was 0.0001 and the batch size was 16. The network converged in 1.5 hours on a PC with an Intel Xeon Silver 4216 processor and Nvidia Titan RTX GPU with 24 GB of memory. Next, we fine-tuned all the neural-network-based methods on the real-world dataset over another 50 epochs at a learning rate of 0.0001. Our study found that in order to achieve higher scores on real-world distortions, one needs to fine-tune using artificial distortions with real-world ground truth stereopairs. When fine-tuning degraded the quality of the model, we used the untuned model. §.§ Quality Metrics Our evaluation used two full-reference quality metrics: PSNR (peak signal to noise ratio) and SSIM (structural simmilarity). PSNR is good for capturing changes in brightness, contrast, hue, and saturation, and SSIM is good for estimating content-dependent distortions. Also, we measured the time each method took to process one 512x512 frame. We repeated these measurements three times and in each case reported the shortest time in milliseconds. §.§ Ablation Study We conducted an ablation study (Table <ref>) on the validation part of the artificial dataset. Our conclusion is that the combination of the L1, L2, SSIM, and PAM loss functions is optimal and that all architectural components contribute to our network quality. Our study found that neural networks generalize much better to the unseen distortion type by using the valid mask. We also noticed that the batch-normalization layers in the transfer module yielded strong visible artifacts, so we disabled them in this module. §.§ Comparison Study We compared the proposed local method (see Table <ref> and Fig. <ref>) with three global methods (Reinhard et al. <cit.>, Xiao et al. <cit.>, Pitié et al. <cit.>), one local method (Pitié et al. <cit.>), two correspondence-based local methods (Lavrushkin et al. <cit.>, Grogan et al. <cit.>), and one neural-network-based local method (Croci et al. <cit.>). As Table <ref> shows, local methods perform better on artificial distortions and global ones perform better on real-world distortions. Local methods may do worse on the real-world dataset because of inconsistent results in low-texture regions; global methods may do worse on the artificial dataset because the artificial distortions are nonlinear. Our neural-network-based method achieves a 3.7 dB–higher PSNR than Croci's neural-network-based method <cit.> on the artificial dataset and a 1.3 dB-higher PSNR on the real-world dataset, along with a similar SSIM score, and is 2.6 times faster. That score is not, however, enough to outperform the conventional methods, which warranting further investigation. § CONCLUSION Few datasets are available for evaluating color-mismatch-correction methods. Most of them are either too small or lack real-world mismatches. We created a new large-scale dataset of real-world stereoscopic videos to perform this task. We also showed that global and local methods produce different results on artificial and real-world datasets, and that neural-network-based methods fall short of conventional methods. We improved the latest neural-network method for correcting color mismatches and hope our contribution will enable development of methods that perform better on both artificial and real-world datasets. IEEEtran

http://arxiv.org/abs/2303.06627v1

20230312095332

Dark Matter Cosmology with Varying Viscosity: a Possible Resolution to the $ S_8 $ Tension

astro-ph.CO

ḍ./fig/ Ashoorioon, Davari] Amjad Ashoorioon and Zahra Davari School of Physics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5531, Tehran, Iran Dark Matter Cosmology with Varying Viscosity: a Possible Resolution to the S_8 Tension [ ======================================================================================= We study varying forms of viscous dark matter and try to address the intriguing tensions of the standard model of cosmology with the recent cosmological data, including the Hubble and S_8 tensions. We note that assuming the dark matter viscosity depends on the Hubble parameter, dark matter density, or both, one can improve the statistics. Although the models tend to aggravate the Hubble tension a bit, they tend to reduce the S_8 tension, even in comparison with the constant viscosity case. Dark Matter, Cosmological Parameters § INTRODUCTION New observations on different scales have led us to a more comprehensive understanding of the universe's evolution based on the standard ΛCDM model. However, due to the increasing accuracy of observational data, new issues have emerged on the large cosmic scale, for example, H_0 (the value of Hubble parameter observed today) and σ_8 (the rms fluctuation of density perturbations at 8 h^-1Mpc scale) tensions referring to disagreement between the predicted values of today’s expansion rate of the universe and matter clustering from CMB data in comparison with their locally determined quantities <cit.>. These issues lead to considering alternative approaches beyond the ΛCDM model such as general relativity modifications or a new description for dark energy (DE) <cit.> or different theories of dark matter (DM) <cit.>. Finding the reason for these apparent discrepancies has become the main driver of cosmological research. One of the main pillars and components in the standard ΛCDM model is dark matter. The existence of dark matter in our universe is undoubtedly confirmed by numerous kinds of astrophysical observations on a range of length scales from galaxy rotation curves and gravitational lensing to large scale structure (LSS) and the cosmic microwave background (CMB). Nevertheless, the physical nature of DM particles is unclear and an enigma after decades of research, largely because cosmic observations are only sensitive to the gravitational effects of DM rather than the properties of its particles. What we already know about DM in the ΛCDM model is that it is responsible for about 85% of the universe's matter content and is a non-luminous, dark component of matter which must be largely non-relativistic or cold, massive and collisionless. we consider it as perfect and ideal fluid <cit.>. There are many alternatives to cold dark matter, such as cannibal Dark Matter <cit.>, decaying dark matter <cit.>, dynamical dark matter, interacting dark matter<cit.> and, the interaction between dark energy and dark matter <cit.>. In these models, attempts have been made to address some of the tensions of the concordance model resorting to these alternatives. Among the various models to investigate dark matter from a thermodynamic point of view, an interesting proposal is "an exotic fluid with bulk viscosity". In <cit.> is predicted a viscosity of the order of 10^-6 H_0 M_P^2 by the effective field theory of dark matter fluid on large scales. They claimed that this magnitude of viscosity could resolve the discordance between late time large-scale structural observations and Planck's CMB data. These viscosities in the cold dark matter can be engendered in two different ways. The first type of viscosity is generated due to the self interaction between the dark matter particles, and they are called "fundamental viscosity". The second type of viscosity is known as "effective viscosity". This kind of viscosity is expected to be generated on large scales as the integrated effect of the back-reaction of small scales non-linearities <cit.>. During the 1940s and 1950s, the first theory to study bulk viscosity was developed, and called Eckart's theory <cit.>. But in the late 1960s and 1970s, it became clear that this type of approach suffered from problems of causality and stability <cit.>. Therefore, by including the second-order deviations from equilibrium, it led to the second-order or Müller-Israel-Stewart theory<cit.>. In addition to introducing a new parameter called relaxation time, the second order theory is a more complicated theory than the first order one. Therefore, the first-order theory has received more attention. We will also develop this work using the Eckart framework. In this paper, we investigate the dynamics of the universe and linear structure formation using Λ viscous dark matter model. Contrary to <cit.>, we assume that the viscosity parameter is not constant and depends on the dark matter energy density or Hubble parameter, or both these quantities. We notice that assuming such dependences, the total χ^2 gets reduced. Also except for when the viscosity is function of both the Hubble parameter and dark matter energy density, we can alleviate the σ_8 tension better. This paper is structured as follows: in section <ref>, first introduce the theory of this model and derive the main equations governing the evolution at background (subsection <ref>) and perturbation (subsection <ref>) levels for different scenarios. Then we review the key features of the background evolution of the universe. We modify the public Boltzmann solver [<https://github.com/lesgourg/class_public>](the Cosmic Linear Anisotropy Solving System) <cit.> for each type of dissipative mechanisms to the dark matter to calculate the cosmological evolution and CMB anisotropies. In section <ref>, we perform the Markov chain Monte Carlo (MCMC) scans by [<https://github.com/baudren/montepython_public>] <cit.> with a Metropolis-Hasting algorithm using the Planck 2018 TTTEEE+lowl+lowP+lensing data sets <cit.> in combination with other probes such as the Baryon acoustic oscillations, BAO (SDSS DR7, BOSS DR12 <cit.>, eBOSS Ly-α combined correlations), the Pantheon SNIa catalog <cit.> and the cosmic chronometers (CC) data <cit.>. We review the basic properties and the cosmological effects of viscous dark matter models for the best value of free parameters in Section <ref>. According to the results obtained in Section <ref>, we will discuss cosmic tensions and finally conclude in Section <ref>. § PHENOMENOLOGY OF VISCOUS DARK MATTER MODEL §.§ Background level In the beginning, we recall the phenomenology of the viscous dark matter models and investigate the effects of viscosity of DM on the solution of Einstein’s equation. In the framework of general relativity, we consider that our flat, homogeneous and isotropic universe including radiation, baryons, the viscous dark matter as non-perfect fluid, and dark energy as cosmological constant (Λ) is described by FLRW metric ds^2=-dt^2+a^2(t)(dr^2+r^2dθ^2+r^2sin^2θ dϕ^2), where a(t) is the scale factor at the cosmic time t. We write the energy momentum tensor for non-ideal CDM fluid in the Landau frame, as T_μν=T̃_μν+Δ T_μν, where T̃_μν is the energy–momentum tensor of a perfect fluid, i.e. T̃^μν=(ρ+p)u^μ u^ν+pg^μν. Here ρ is the energy density and p the pressure in the fluid rest frame and Δ T_μν in the first order gradient expansion is a tiny perturbation that corresponds to all dissipative processes (heat flux, anisotropic-stress, and bulk viscosity) <cit.> as Δ T_μν=Π_B(u^μ u^ν+g^μν)+Π^μν, where Π_B and Π^μν represent bulk stress and shear stress tensor respectively defined as Π_B=-ξ u_;γ^γ, Π^μν=2ησ^μν= -2η[1/2(Δ^μαΔ^νβ+Δ^μβΔ^να)-1/3Δ^μνΔ^αβ]▽_αu_β, where η and ξ represent shear and bulk viscosity. As shear viscosity is a directional process, the cosmological principle precludes its existence in the FLRW metric. Hence, only bulk viscosity is a unique process that is allowed to occur in an expansive, homogeneous, and isotropic background. Hence, let us examine the expressions <ref> in more detail u^γ_;γ=1/√(-g)∂_γ(√(-g)u^γ)=1/√(-g)∂_γ(√(-g))u^γ+∂_γ u^γ, where √(-g)=r^2 sin^2θ a^3(t), and (∂_γ(√(-g))u^γ=(∂_0√(-g))u^0=3r^2 sin^2θ a^2(t)ȧ(t). So that 1/√(-g)∂_0(√(-g))=3ȧ(t)/a(t)=3H, means that bulk viscosity modifies the effective pressure as p_ eff=p-3Hξ which consists of a sum of two terms, equilibrium pressure p and bulk viscosity pressure, -3Hξ. Choosing a reference frame in which the hydrodynamics four-velocity u_μ is unitary, we can write Equation <ref> for imperfect viscous dark matter fluid as T̃^μν=(ρ+p_ eff)u^μ u^ν+p_ effg^μν. From the conservation of energy, ▽_μ T^μ_ν=0, assuming that no interactions take place among the cosmic fluid components the continuity equations take the form, ρ̇_r+4Hρ_r=0 ⟹ρ_r=ρ_r0 a^-4, ρ̇_b+3Hρ_b=0⟹ρ_b=ρ_b0 a^-3, ρ̇_ vdm+3H(ρ_ vdm-3Hξ)=0, ρ̇_Λ=0 ⟹ρ_Λ=ρ_Λ 0. The dynamics of the Universe containing these components is given by the Friedmann equation: H^2=(ȧ/a)^2=8π G/3(ρ_r+ρ_b+ρ_ vdm+ρ_Λ). Using the definition of the dimensionless energy density parameter Ω_i,0=ρ_i,0/ρ_ cr,0 , where ρ_ cr,0=3H_0^2/(8π G) is the current critical energy density at a=1, the Hubble parameter takes the following form: H^2=H_0^2(Ω_ r0a^-4+Ω_ b0a^-3+y(a)+Ω_Λ), where we define y(a)=ρ_ vdm/ρ_ cr,0 and set the present density parameter of radiation (photon+relativistic neutrinos) as Ω_r0=Ω_γ0(1 + 0.2271 N_ eff). Here, we put the effective extra relativistic degrees of freedom 3.046 in agreement with the standard model prediction <cit.>. Applying the Friedman equation, which takes the form of Σ_iΩ_i=1, we can write Ω_Λ=1-Ω_ r0-Ω_ b0+y(1). Equation <ref> can be rewritten in terms of the scale factor (d/dt=aHd/da) and the definition y(a) as follows, ady(a)/da+3y(a)-ξ̃/H_0√(Ω_ r0a^-4+Ω_ b0a^-3+y(a)+Ω_Λ)=0, where we have defined the dimensionless parameter ξ̃=24π Gξ. Equation <ref> can be solved numerically by fixing the value y(a=1)=Ω_ dm,0. However, we implement the continuity equation <ref> in of the public Boltzmann solver and use the shooting method described in <cit.> to compute the present-day dark matter density. In the rest of the study, we consider four different functional forms of the bulk viscosity coefficient, ξ(ρ_ vdm,H,ρ_ vdmH): i- Designated as (M1), we consider ξ as constant. This is the model considered by <cit.>. ii- For the second model (M2), where we assume ξ to be proportional to powers of the energy density of viscous DM, ξ(ρ_ vdm/ρ_ vdm,0)^n and we would obtain the equation <ref> as ady(a)/da+3y(a)-ξ̃/H_0√(Ω_ r0a^-4+Ω_ b0a^-3+y(a)+Ω_Λ) ( y(a)/Ω_vdm,0)^n=0. ady(a)/da+3y(a)-ξ̃/H_0√(Ω_ r0a^-4+Ω_ b0a^-3+y(a)+Ω_Λ) ( y(a)/Ω_vdm,0)^n=0. iii- For the third model (M3), we consider ξ to be proportional to powers of the Hubble parameter, ξ(H/H_0)^m and the equation <ref> is given as ady(a)/da+3y(a)-ξ̃/H_0(H/H_0)^m+1=0. iv- For the fourth model (M4), coefficient is proportional to ξ(ρ_ vdm/ρ_ vdm,0)^n(H/H_0)^m, and the equation <ref> is given as ady(a)/da+3y(a)-ξ̃/H_0(y(a)/Ω_ vdm,0)^n(H/H_0)^m+1=0. The values assumed for free parameter, ξ in all of the mentioned models are positive for thermodynamical reasons. The standard cold dark matter model is recovered in the limit ξ̃→ 0 for all models. §.§ Perturbation Level In this subsection, we briefly survey the main mathematical form of the linear perturbation theory within the framework of viscous matter dark cosmologies. We consider the density and pressure in terms of spatially homogeneous and isotropic background fields with small spatially varying perturbations: ρ(x,t)=ρ̅(t)+δρ(x,t)p(x,t)=p̅(t)+δ p(x,t). We assume perturbations in the variables δρ,δ p≪ρ̅, p̅ up to linear order, and we focus on scalar perturbations only. We use the normalized density contrast δ≡δρ/ρ̅ to expand the fluid dynamic equations in dimensionless measures of these perturbations and the velocity divergence θ≡∇⃗·v⃗. Note pressure and its perturbation are not independent quantity and they are related to density as: w=p/ρ, c^2_s=δ p/δρ, c^2_ad=ṗ/ρ̇, where w, c_s^2 and c_ad^2 are the dynamical equation of state, the speed of sound in the medium and the adiabatic sound speed of the viscous DM respectively. Since cold DM is a pressureless ideal fluid, w = c^2_s = c^2_ ad = 0. The dynamical equations which govern the evolution of cosmological perturbations in Fourier space can be written in Synchronous gauge as δ̇=-3H(c_s^2-w)δ-(1+w)(θ+ḣ/2), θ̇=-H(1-3c_ad^2)θ+c_s^2/1+wk^2δ-k^2σ, and in the Conformal Newtonian gauge as δ̇=-3H(c_s^2-w)δ-(1+w)(θ-3ϕ̇), θ̇=-H(1-3c_ad^2)θ+c_s^2/1+wk^2δ-k^2σ+k^2 ψ. One can then calculate w=-3ξ̃(ρ_ vdm,H,ρ_ vdmH)H/ρ and c^2_s=ξ̃(ρ_ vdm,H,ρ_ vdmH)θ/aρδ <cit.>. In <cit.>, it is assumed that c^2_ad is scale-independent and approximately equal to c^2_s, assumptions that we take to be valid in the current study. For the evolution of shear stress, we have implemented Non-Cold Dark Matter Fluid Approximation (NCDMFA) in as σ̇=-3[1/τ+H(2/3-c^2_ad)]σ+4/3c^2_ vis/1+w(2θ+ḣ). c^2_ vis is a new parameter named viscosity speed and in implemenation of is assumed as c^2_ vis=3wc^2_ad <cit.>. We implement the perturbation equations <ref>- <ref> in module of Boltzmann code § MCMC ANALYSIS In this section, we present constraints on four viscous dark matter models that we discussed in the previous section. For MCMC analysis we use the Metropolis-Hastings algorithm of the cosmological sampling package , connected to an altered version of the Boltzmann Solver . We use the following datasets to perform the statistical inference: * CMB: One of the most important observables in cosmology due to its well-understood linear physics and sensibility to cosmological parameters is the CMB. Here, we use Planck 2018 CMB temperature and polarization auto- and cross-correlation, both high-l and low-l and with the full set of nuisance parameters, as well as Planck 2018 CMB lensing in the multipole range 40 ≤l≤ 400<cit.>. * BAO: Another important data for present and future cosmology is the BAO thanks again to its well-understood linear physics. We use the BAO measurements from the Baryon Oscillation Spectroscopic Survey Data Release 12 (BOSS DR12) <cit.>, eBOSS DR14-Ly-α combined correlations <cit.> and, WiggleZ <cit.>. * SNe Ia: The Pantheon Super Novae (SN) sample consists magnitude measurements for 1048 SNe Ia with redshifts 0.01 < z < 2.3 <cit.>. Some recent studies have shown that H_0 tension depends directly on the SN Ia absolute magnitude, M_B <cit.>. Therefor here, we consider the M_B as a nuisance parameter in our numerical analysis. * Cosmic Chronometers (CC): The CC approach is very powerful in detecting the expansion history of the Universe obtained through the measurements of the Hubble parameter. We use the H(z) data points from the most recent, accurate estimates and model-independent data in the redshift range 0.07≤ z ≤ 1.965 based on the relative age of passively evolving galaxies (the differential-age technique). These data are uncorrelated with the BAO data points <cit.>. We employ the χ^2 statistics to constrain our theoretical model, χ^2=(P_ obs-P_ th)^2/σ_P^2. Above P_ obs, P_ th and σ_P^2 indicate the observed value, the predicted value and the standard deviation, respectively. The total likelihood is given by: χ^2_ min≡-2lnL=χ^2_ CMB+χ^2_ BAO+χ^2_ SNIa+χ^2_ CC. Here, in addition to the six free parameters of the standard minimum model, i.e. (Ω_b,Ω_DM,10θ_ MC,ln 10^10A_s,n_s,τ_ reio), the viscous dark matter models introduced in the previous section include at most (ξ, n, m) in the M4 model. Model M1 has one parameter and models M2 and M3 has two parameters. The convergence of chains for each parameter is measured by the Gelman-Rubin criterion and one can obtain acceptable R-1 values (i.e. below 0.01 for every parameter) with an iterative strategy <cit.>. In Table <ref>, we report the best and the mean values and 68%CL intervals for main parameters including the total matter density parameter (Ω_m=Ω_B+Ω_DM), the present-day expansion rate of the Universe or the Hubble constant, H_0 and S_8=σ_8√(Ω_m/0.3) in different scenarios. We also show posterior distributions (1σ and 2σ intervals) as dark and light shaded contours, respectively in Figure <ref>. Some points in the Figure <ref> need to be stressed. First, we can see that the decrease of H_0 values is associated with the increase of Ω_m values and vice versa in both viscous and cold dark matter scenarios but the behavior of Ω_m-S_8 and H_0-S_8 in two scenarios are different. Second, it seems that assuming viscosity improves the S_8 tension for the viscous dark matter scenario. We will discuss this further in section <ref>. Next, to check whether the fit is good or not, and also to choose the best and most compatible model with the observational data, we use the simplest method that is usually used in cosmology, which is called the least squares method, χ^2_ min. In this case, the model with smaller χ^2_ min is taken to be a better fit to the data <cit.>. Comparing different viscous models, we note that models M2, M3 and M4 do better than CDM and even M1 scenario where the viscosity parameter was assumed to be constant. The difference between χ^2_ min of the M4 model and the CDM model is 4.47. However one should keep in mind that the model with the lowest χ^2_ min is not necessarily the best model because since adding more flexibility with extra parameters will normally lead to a lower χ^2. That means that a poorly parameterized model should be penalized. In this work, M1 model has one parameter, M2 and M3 have two parameters and M4 has three parameters more than CDM scenario. In order to deal with model selection a standard approach is to compute the Akaike Information Criterion (AIC), AIC=χ^2_ min+2M+2M(M+1)/N-M+1. Above M is the number of free parameters in the model and N is the number of data points. Therefore in this criterion, we reduce the practical influence of M in χ^2_ min in favor of a model with a lower free parameter. Since N≥ M then we neglect the third term in Equation <ref>. We report the result of MCMC analysis for the best-fit χ^2 per observational data set and compare the viscous DM models to the reference CDM model in Table <ref>. The results of this analysis can be interpreted with the Jeffreys' scale as follows: there is substantial support for all viscous DM models (|Δ AIC|<2) and an indication of the consistency of the two scenarios. Among the viscous DM models, the M3 scenario outperforms others. § OBSERVTIONAL EFFECTS OF VISCOUS DM MODELS §.§ Background level To gain better insight into different models of viscous DM and compare it with the cold DM model, we will examine the evolution of the main cosmological quantities for the parameter values of Table <ref> below. As we know one of the key cosmological parameters and a pivotal quantity in cosmology is the current value of the Hubble parameter since it is used to construct the most basic time and distance cosmological scales. On other hand, because of the impact of Hubble's expansion the growth of matter perturbations, it is important to understand the behavior of H(z) in various viscous DM cosmologies. Therefore, we plot the evolution of H(z) and H(z)/(1+z) in Figures <ref> and <ref>. We present in the bottom panel of Figure <ref>, the relative difference Δ H(z)= 100×[H_ Mi-H_ CDM/H_ CDM] of the Hubble parameters with respect to the CDM. As we see the quantity Δ H(z) is positive (negative) for large (small) redshifts in all scenarios. But it should be noted that this difference is less for models M3 and M4 and which means that the corresponding cosmic expansion is smaller than that of the concordance CDM. We have plotted the evolution of the fractional energy densities of viscous dark matter in Figure <ref>. As we expect all viscous DM models mimic the CDM scenario and the universe evolves from a radiation-dominated phase to a matter-dominated epoch the density parameter of the dark matter decreases and the density parameter of dark energy has increased in the recent universe. Given to reported values in Table  <ref>, the density ratio Ω_DM/Ω_B for M3 and M4 models are ≃ 6 larger than CDM model (≃ 5). The last quantity that we plotted in this section is the equation of state parameter for viscous DM, w = P/ρ=-3 ξ̃(ρ_ vdm,H,ρ_ vdmH)H/ρ. As we see in Figure <ref>, in the past (when a→ 0), the viscous DM models behave like CDM (w→ 0), and currently, the value for w is negative for all models. The most negative w are respectively for the M4, M3, M1 and M2 models. §.§ Perturbation Level In this subsection, we turn to the perturbation spectra. To draw the plots we have used the best-fit values of the model parameters obtained from all the observational datasets that we have used in this work. First, we display the power spectra of the anisotropies in the CMB temperature, calculated for viscous DM models along with that for CDM in Figure <ref>. As we can see in the lower multipoles (for l<50), there is a relative deviation of Δ D_l^ TT between the viscous DM models and CDM scenario. while for higher multipoles (for l > 50), we do not observe any significant differences in the curves. We can explain by noting that the CMB temperature is determined only by the energy density component of the energy-momentum tensor, T_00, whereas the density perturbations also depend upon the diagonal elements, T_ ii, which include the pressure. Since the viscous DM scenarios own pressure, this causes a difference between the viscous DM models and CDM. Hence, as the evolution of density perturbations unfolds, small-scale structures are clearly suppressed in viscous DM compared to CDM, which can be seen in the Figure <ref>. The amount of suppression is larger at smaller k's. Respectively, models M3, M4, M1 and M2 suppress the nonlinear matter power spectrum. As we can see in Figures <ref> and  <ref>, the position of the acoustic peaks has not shifted therefore the viscous dark matter has not had much effect on the background expansion. We also see a difference between models in the top plane of the Figure <ref> for low l. This is due integrated Sachs-Wolfe (ISW) effect at such scales. Also the small l's of CMB (TT and even better EE) give information on the reionization history. By using the best-fit values of free parameters, shown in Tables <ref>, we obtain the redshift of the reionization 7.69, 6.99, 8.42 and 8.65, respectfully for models M1, M2, M3 and M4. M2 and M4 have the biggest differences with the standard model (z_ reio,CDM=7.78). The present galaxy surveys provide powerful observational data for the combination fσ_8(z), where f is the linear growth rate of the density contrast, which is related to the peculiar velocity in the linear theory, being defined by f(a)=a δ'(a)/δ(a) and σ_8(a)=σ_8(a=1)δ(a)/δ(a=1) is the root-mean-square mass fluctuation in spheres with the radius 8h^-1 Mpc. Here, we calculate the total density contrast of the matter (baryons + dark matter) as δ(a)=δ_b ρ_b+δ_ dmρ_ dm/ρ_b+ρ_ dm. The linear growth rate, f, is an excellent tool for distinguishing between various dark matter theories based on the growth of large-scale structures. In Figure <ref> we show the linear growth factor f(a) for the different viscous DM and standard models studied in this work for the best-fit values for modes k=10^-4Mpc^-1 (the galaxy cluster scale) and k=1Mpc^-1. Since the linear growth rate is independent of the wave number k at the large scale factors (the low redshift regime), we see that for a→ 1, two modes behave similarly. The bottom panel shows that for small-scale factors the growth function goes to unity, which corresponds to the matter-dominated Universe and the amplitude of matter perturbations decreases at low redshifts in all scenarios with the dominance of dark energy. We observe that the suppression of the amplitude of matter fluctuations in M3 and M4 models that are proportional to the Hubble parameter starts sooner compared with other scenarios. The change in σ_8 in the viscous DM scenarios can be seen as the relative decrease in P(k) in the Figure <ref> (although σ_8 is computed from the linear rather than non-linear power spectrum), thus improving the tension between these CMB-derived predictions and the actual LSS data. Looking at Figure <ref>, we notice that considering viscosity has the influence of slowing down the evolution rate of the dark matter perturbations. This means that structures cluster slower in the viscous dark matter models as we predicted from Figure <ref>. In particular scenario M4 and M3 suppress the power spectrum more in comparison with the CDM scenario. § LARGE-SCALE COSMIC TENSIONS In this section, we reported the obtained results for the H_0 and S_8 tensions. §.§ H_0 Tension The Hubble tension is one of the most lingering tensions in the cosmology. This tension emerged with the first release of Planck results and has grown in significance in the past few years. The last value of H_0 tension is obtained from the most recent analysis of Planck 2018 for the early-universe observation H_0 = 67.44±0.58 kms^-1Mpc^-1 for the ΛCDM model <cit.>. While the late-universe observation of Cepheid-calibrated SNe Ia distance ladder by SH0ES team gives the result of H_0 = 74.03±1.42 kms^-1Mpc^-1 <cit.>. The big discrepancy between the two H_0 values reaches ∼4σ <cit.>. One of the initial motivations of this work was to investigate this tension manipulating the properties of one of the main components of the universe, i.e. dark matter. As the results of Table <ref> and Figure <ref> demonstrate none of the models of viscous dark matter diminishes the H_0 tension. On the other hand, H_0 tension is equivalent to the mismatch of the Pantheon SnIa absolute magnitudes, which when calibrated using the CMB sound horizon and propagated via BAO measurements to low z have a value M_B=-19.401 ± 0.027 mag <cit.>, while when calibrated using the Cepheid stars have a value M_B=-19.244 ± 0.037 <cit.>. In this study, for all of the models, absolute magnitudes obtained is ∼ -19.41 consistent with CMB observations. Therefore none of these models really help us in addressing the H_0 tension. §.§ S_8 Tension Cosmic structures were formed based on primordial density fluctuations that emerged from the inflation era due to gravitational instability. By studying the large-scale structure of the Universe and its evolution in cosmic epochs, it is possible to trace the growth of cosmic perturbations. The main difference between the standard model and other dissipative approaches lies in the behavior of the cosmological perturbations. While perturbations in standard cosmology are always adiabatic, dissipative models of the dark sector are intrinsically non-adiabatic. Associated with this observable, one of the current cosmological tensions is the growth tension. It has come about as the result of the discrepancy between the Planck value of the cosmological parameters Ω_m and σ_8 from weak gravitational lensing surveys, cluster counts, and redshift space distortion data. Since these two parameters have degenerate effects in the lensing surveys, usually S_8=σ_8√(Ω_m/0.3) (Weighted amplitude of matter fluctuations) is used as a parameter to compare the consistency with other observations. Measurements of the CMB temperature and polarization anisotropies by Planck yield S_8 = 0.834±0.016. On the other hand, weak gravitational lensing surveys provide constraints via cosmic shear, e.g. S_8 = 0.759^+0.024_ -0.021 from the Kilo-Degree Survey (KiDS-1000) <cit.> and S_8 = 0.780^+0.030_-0.033 from Hyper Suprime-Cam (HSC) <cit.>. The combination of shear and galaxy clustering from three-year data of DES yields S_8 = 0.776^+0.017_-0.017 and a combination of shear, clustering and galaxy abundance from KiDS-1000 S_8 = 0.773^+0.028_ -0.030, while galaxy cluster counts from SPT-SZ report S_8 = 0.766 ±0.025 and eROSITA results favour S_8 = 0.791^+0.028_-0.031. We adopt the following estimator in order to quantify the discordance or tension in current determinations T_S_8=|S_8^i-S_8|/√(σ^2_i+σ^2). Considering the reported value of the Kilo-Degree Survey KiDS-1000 as observational data, as we see in Table <ref>, all models of viscous dark matter are able to decrease this tension. More than any of the viscous models, M2 can reduce the tension between Planck and KiDs-1000 data. § DISCUSSION AND CONCLUSION With the aim of finding a model more compatible with the universe and reducing the current controversial tensions in cosmology, this study questioned one of the properties of dark matter in the Standard Model of cosmology. We include a physical dissipation mechanism in the description of dark matter such as viscosity. In addition to the constant bulk viscosity case which was studied previously by <cit.>, we assessed the cases in which the bulk viscosity has the dependence on ρ_vdm, Hubble parameter, or both. We noticed that with such dependences of viscosity, the χ^2 for the cosmological datasets is reduced by roughly ∼4. Noting that we have added more parameters we computed the AIC criterion for these models and notice that for all of them |Δ AIC|<2. The least value for this parameter is obtained for the M3 model, |Δ AIC|≃0.22, where the bulk viscosity is assumed to be proportional to the powers of the Hubble parameter. Our results in Tables <ref> and Figure <ref> show that the assumption of viscous dark matter can only reduce one of these tensions, namely the S_8. In order to relax the observed tensions H_0 and S_8 simultaneously, perhaps in addition to viscous dark matter, the presence of relativistic components or early dark energy, should also be considered. § ACKNOWLEDGMENTS This work is supported by Iran Science Elites Federation Grant No M401543. mnras

http://arxiv.org/abs/2303.08048v1

20230313140006

Active target TPC for study of photonuclear reactions at astrophysical energies

physics.ins-det

Active target TPC for study of photonuclear reactions at astrophysical energies Presented at Zakopane Conference on Nuclear Physics 2022 M. Kuich^1e-mail: mkuich@fuw.edu.pl, M. Ćwiok^1, W. Dominik^1, A. Fijałkowska^1, M. Fila^1, A. Giska^1, Z. Janas^1, A. Kalinowski^1, K. Kierzkowski^1, C. Mazzocchi^1, W. Okliński^1, M. Zaremba^1, D. Grządziel^2, J. Lekki^2, W. Królas^2, A. Kulińska^2, A. Kurowski^2, W. Janik^2, T. Pieprzyca^2, Z. Szklarz^2, M. Scholz^2, M. Turzański^2, U. Wiącek^2, U. Woźnicka^2, A. Caciolli^3, M. Campostrini^4, V. Rigato^4, M. Gai^5, H. O. U. Fynbo^6 ^1Faculty of Physics, University of Warsaw, Warsaw, Poland ^2Institute of Nuclear Physics Polish Academy of Sciences, Cracow, Poland ^3University of Padova and INFN-PD, Padova, Italy ^4Laboratori Nazionali di Legnaro, Legnaro, Italy ^5University of Connecticut, CT, USA ^6Department of Physics and Astronomy, Aarhus University, Aarhus, Denmark March 30, 2023 ========================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================= A setup designed to study photonuclear reactions at astrophysical energies - an active target Time Projection Chamber was developed and constructed at the Faculty of Physics, University of Warsaw. The device was successfully employed in two experiments at the Institute of Nuclear Physics Polish Academy of Sciences in Cracow, in which γ- and neutron-induced reactions with CO_2 gas target were measured. The reaction products were detected and their momenta reconstructed. Preliminary results are shown. § INTRODUCTION One of the most important open questions in nuclear astrophysics concerns the creation of carbon and oxygen in the stars. Synthesis of carbon and oxygen happens in helium-burning thermonuclear reactions in the star’s core, the triple-alpha reaction and ^12C(α,γ)^16O, respectively. The basic observable to resolve the reaction rates is the cross-section, which has to be determined at the relevant energies (at the Gamow peak) and also is required as input information for star evolution models <cit.>. One way to determine the reaction cross-section is to study its inverse reaction by using an active-target together with photon and neutron beams. In this paper, we present preliminary studies to validate this approach. § EXPERIMENTS Studying the γ- and n-induced reactions on ^12C and ^16O requires a detection system that allows for reconstructing the momenta of all the charged reaction products. An ideal tool is an active-target time-projection chamber (active-target TPC), filled with CO_2 gas. The Warsaw active-target TPC was developed for this purpose. It consists of a 33× 20 × 20 cm^3 active volume, immersed in a vacuum vessel equipped with a control system to maintain a constant gas pressure inside. The active volume is surrounded by field-shaping electrodes and terminated with a cathode plate on one end and a stack of 3 Gas Electron Multiplier foils as amplification section followed by a planar, 3-coordinate (U, V, W), redundant readout plane at the other end. The arrays of U-, V-, and W-strips register the charge deposit in 2-dimensions, while time-distribution of the charge collected at the electrode, combined with the drift velocity of the electrons in the given gas mixture and drift field, allows for determining the third coordinate. The device allows for full and unambiguous kinematic reconstruction of multiple-particle events <cit.>. The first commissioning measurements with the Warsaw active-target TPC were conducted in 2021 at the Institute of Nuclear Physics, Polish Academy of Sciences in Cracow (IFJ). In the first experiment, a 1.03 MeV proton beam from the Van de Graaff accelerator with currents of about 10-20 μA was used to produce 13.1 MeV γ-rays in the ^15N(p,γ)^16O reaction. For this purpose, a ^15NCr target (about 1.3× 10^18 atoms/cm^2), produced in reactive ion sputtering on a Ta backing at the National Laboratories of Legnaro, was used. The γ beam intensity was monitored by a NaI detector positioned at the side of the target. The TPC, filled with CO_2 at the absolute pressure of 250 mbar, was placed right behind the target station and the produced γ-rays interacted with the gas of the TPC, where they induced the photo-disintegration of ^12C and ^16O. Charged reaction products were detected. Examples of ^16O(γ,p)^15N and ^16O(γ,α)^12C event candidates are shown in Fig. <ref>. Raw charge distributions in strip (U-, V-, W-) coordinate and time are presented as 2D plots. A manual reconstruction method was used to determine the emitted particle position in the chamber and measure its range. Identification of the two-particle event as due to ^16O(γ,p)^15N or ^16O(γ,α)^12C as well as the energy reconstruction was done by comparing the track length of emitted particles with SRIM simulations <cit.>. Examples of event reconstruction is presented in the 1D plots in Fig. <ref> as charge distribution along the track(s) with fitted Bragg curve(s). A preliminary distribution of the reconstructed energy of the protons is depicted in Fig.<ref> (partial statistics was analyzed). [13]l0.5 < g r a p h i c s > Reconstructed proton energy spectrum from ^16O(γ,p)^15N. The mean energy of the protons, ⟨ E_⟩ = 890 ± 24 keV, is in 3σ agreement with expected 966 keV. The second experiment took place at the Impulse Neutron Generator (IGN-14), where 14 MeV neutrons were produced in the d(t,n) reaction and interacted with the CO_2 gas, which was kept at 80 mbar, in the TPC positioned just behind the tritium target. The neutron flux was monitored with a ^3He counter and additionally estimated with activation of aluminium targets. The typical neutron flux amounted to about 3 × 10^4 n/s/cm^2. Two example events for ^12C(n,n')^12C^* and ^16O(n,α)^9Be reactions are presented in Fig. <ref>. Partial statistics was manually analyzed and the emitted particles identified. The majority of the two-particle events were classified as ^12C(n,α)^9Be, while three-particle events as carbon dissociation into 3 α particles. Preliminary reconstruction of the particles momenta indicates that the ^12C(n,n')^12C^* event in the Fig. <ref> corresponds to the decay of the Hoyle state at 7654 keV by 3α emission. § SUMMARY An active-target TPC dedicated for studying reactions of astrophysical interest at the relevant energies with γ or neutron beams was developed at the University of Warsaw. The first experiments employing it were conducted in 2021 at IFJ, where ^16O(γ,α)^12C, ^16O(γ,p)^15N, ^12C(γ,3α), ^12C(n,α)^9Be and ^12C(n,n')^12C^* were observed. Preliminary analysis shows that the Warsaw active-target TPC is an adequate tool for measuring such reactions induced by non-charged particles. We would like to thank H. Czyrkowski and R. Dąbrowski for their support in the preparation of the equipment. Scientific work was supported by the National Science Centre, Poland, contract no. 2019/33/B/ST2/02176, by the University of Warsaw, Poland, through the Interdisciplinary Centre for Mathematical and Computational Modelling, comp. alloc. no. G89-1286. unsrt

http://arxiv.org/abs/2303.06830v1

20230313031628

Anti-helical edge magnons in patterned antiferromagnetic thin film

cond-mat.mes-hall

yunmeili@xmu.edu.cn Department of Physics, School of Physical Science and Technology, Xiamen University, Xiamen 361005, China Helical edge states in topological insulators give counterpropagating spin current on the two parallel edges. We here propose anti-helical edge states of magnons in patterned antiferromagnetic thin films, which host copropagating spin current on the two parallel edges, where the two magnon modes with opposite chirality act like the spin. The embedded heavy metal dot array in the thin film induces interfacial Dzyaloshinskii-Moriya interactions (iDMIs), drives the magnon bands into nontrivial topological phases, characterized by spin Chern number. The resulting helical edge modes lead to spin current with the direction dependent on the sign of iDMI parameter. In a strip geometry, we combine two subsystems with two embedded metal dot arrays, which give opposite iDMI parameters. Anti-helical edge states emerge, compensated by the counterpropagating bulk confined states. Helical and anti-helical edge states are verified by the micromagnetic simulations. Our work is quite helpful in the field of magnon spintronics based on antiferromagnets. Anti-helical edge magnons in patterned antiferromagnetic thin film Yun-Mei Li March 30, 2023 ================================================================== § INTRODUCTION Chiral or helical edge states are a hallmark feature of two-dimensional topological insulators <cit.>. Chiral edge states emerge in quantum Hall or Chern insulators with time-reversal symmetry breaking and propagate along the edges clockwise or counterclockwise. As a consequence, a typical strip geometry supports opposite propagating edge states along the two parallel edges. Helical edge states can be treated as two copies of chiral edge states, protected by the time-reversal symmetry, holding great potential applications in spintronics. The edge states for opposite spin hold opposite chirality, leading to a vanished charge current and a net spin current, which flow in opposite direction at the two parallel edges in the strip geometry. Recent theory predicted the so-called antichiral edge states <cit.>, which copropagate in the same direction at opposite edges, compensated by counterpropagating bulk modes, distinct from the chiral edge states. The original work is based on the modified Haldane model in electronic systems <cit.>, quite hard to realize experimentally. Then a series of the theoretical even experimental works addressed the realization in many other systems, including the electron system based on graphene <cit.>, exciton polariton <cit.>, magnon <cit.>, photon <cit.> and classical electric circuit <cit.> systems. The experimental works in photonic <cit.> and electric circuit <cit.> systems observed the copropagating behavior of antichiral edge states based on the modified Haldane model. All these works require the broken time-reversal symmetry. A natural question arises. In the time-reversal symmetric systems, is it possible to realize “anti-helical" edge states, where the spin current on the parallel edges propagates in the same direction? For now, quite few works made the explorations theoretically or experimentally. In recent yeas, antiferromagnets have attracted significant attention in spintronics due to the ultrafast spin dynamics and lack of stray fields <cit.>. Importantly, antiferromagnets support both left-handed and right-handed polarized magnon modes, related by the pseudo-time reversal symmetry <cit.>, analogous to the electron spin. This coexistence of both polarizations give rise to many novel spin-related physical phenomena, such as the magnon spin Nernst effect <cit.>, Stern-Gerlach effect <cit.> and Hanle effect <cit.>. Manipulating the polarization may facilitate the chirality-based computing <cit.> and logic devices <cit.>. Therefore, the exploration on topologically protected helical or even anti-helical edge state could be greatly helpful in the field of magnon spintronics based on antiferromagnets <cit.>. In this paper, we theoretically propose the realization of anti-helical edge states of magnons in patterned antiferromagnetic thin film, in which the spin current at parallel edges flow in the same direction, compensated by the counterpropagating bulk confined spin currents, as illustrated in Fig. <ref> (a). Here the degenerate two magnon modes with opposite chirality is treated as the spin degree of freedom, as widely discussed in previous works above. Heavy metal dot array is embedded into the thin film to fold the free dispersion of magnons into bands. Moreover, the inversion symmetry breaking at the interfaces between metal dots and the thin film will generate interfacial Dzyaloshinskii-Moriya interactions (iDMIs), which modify the exchange boundary condition and induce nontrivial topological phase for the magnon bands, characterized by the spin Chern number. As a result, helical edge magnons arise in a finite width strip geometry. The left-handed and right-handed edge magnons flow towards opposite directions at the same edge, leading to a pure spin current, revealed by the micromagnetic simulations. The direction of the spin current is determined by the sign of iDMI parameter D. In a finite width ribbon geometry, we combine two subsystems embedded with different metal dot arrays, which induce iDMIs with opposite sign of D, as shown in Fig. <ref> (a). Interesting, the spin currents along the two edges flow in the same direction, while the spin currents along the domain wall flow in opposite direction. The emergence of the anti-helical edge states can also be verified by the micromagnetic simulations. This paper is organized as follows. In Sec. II, we derive the eigenvalue problem for the magnons in the patterned antiferromagnetic thin film in the presence of metal dot array and the iDMIs. In Sec. III, we discuss the band topological properties due to the iDMIs, the topological phase transition and the construction of strip geometry for anti-helical edge states. We also performed the micromagnetic simulations for verifications. In Sec. IV, we summarize our results. In Sec. V, we present the methods and the parameters we adopted for the calculations. § THE PATTERNED ANTIFERROMAGNETIC THIN FILM We consider an antiferromagnetic thin film with periodically embedded metal dots. The metal dots are arranged in a triangular lattice configuration. We construct a domain along two sides with different embedded metal dot array to get the anti-helical edge states in a strip geometry, as illustrated in Fig. <ref> (a). An infinite system with only one type embedded metal dot array is a two-dimensional artificial magnonic crystal. The corresponding unit cell is shown in Fig. <ref> (b). The strip geometry with one or two embedded metal dot array is one-dimensional artificial magnonic crystal. The magnetization dynamics in antiferromagnets can be described by two coupled Landau-Lifshitz-Gilbert (LLG) equations for each sublattice, ṁ_i=-γ𝐦_i×𝐡_i+α𝐦_i×ṁ_i, where i=1,2 denote the two sublattices. γ is the gyromagnetic ratio, α is the Gilbert damping constant. Here γ𝐡_i=K_zm_i^z𝐳+A∇^2𝐦_i-J_ex𝐦_i̅ (with 1̅=2 and 2̅=1) is the effective magnetic field acting locally on sublattice 𝐦_i, where K_z is the easy-axis anisotropy along 𝐳 direction A and J_ex characterize the Heisenberg exchange coupling constant of intra-sublattice and inter-sublattice, respectively. The equilibrium magnetization is in collinear order for the two sublattices and parallel to the z-direction. To get the magnon dispersions, we can divide the magnetization and effective magnetic fields on each sublattice into the static and dynamical ones. Let 𝐦_i=𝐦_i^0+δ𝐦_i and 𝐡_i=𝐡_i,0+δ𝐡_i with 𝐦_i^0 the equilibrium magnetization and |δ𝐦_i|≪𝐦_i^0. 𝐡_i,0 is the z-component of the effective magnetic field. By neglecting the damping first, we have ∂δ𝐦_i/∂ t =-γ(𝐦_i^0×δ𝐡_i+δ𝐦_i×𝐡_i,0). To solve the dispersion in the artificial crystals, we employ the ansatz δ𝐦_i=δ𝐦_i^𝐤(𝐫)e^i(𝐤·𝐫-ω t), and after a transformation, we can get the eigenvalue equations, given by ωψ_±,𝐤 =∓σ_zĤ_𝐤ψ_±,𝐤, where ψ_±,𝐤=(δ m_1,±^𝐤,δ m_2,±^𝐤)^T, δ m_i,±^𝐤=δ m_i,x^𝐤± iδ m_i,y^𝐤, Ĥ_𝐤=[K_z+J_ex+(∇+i𝐤)^2]σ_0+J_exσ_x. σ_0 is 2×2 identity matrix, σ_x,y,z are the Pauli matrices. The solution for ψ_+,𝐤 denotes left-handed modes while ψ_-,𝐤 is solution for right-handed modes. As discussed in our recent work <cit.>, the two polarized modes are related by a pseudo-time reversal symmetry. Note that the two polarized modes are totally decoupled. We turn to the role of the embedded heavy metal dots. The first effect is to fold the free dispersion into magnon bands due to the scattering by the periodic interfaces from the metals. More importantly, as discussed in many theoretical <cit.> and verified in experimental works <cit.>, the heavy metals (such as Pt, W, Au, Re, Ir, etc) with strong spin-orbit coupling can generate chiral interfacial Dzyaloshinskii-Moriya interactions (iDMI) due to the inversion symmetry breaking at the magnets/heavy metal interfaces. In our system, the iDMI 𝐡_DM is given by γ𝐡_DM=D𝐳×(𝐞_t·∇)𝐦_i by comparing our bent surface to the previous plane surface <cit.>. 𝐞_t=𝐧×𝐳 is the in-plane tangential vector of the interface and 𝐧 the normal vector of the interface, both depicted in Fig. <ref> (b). D may be either positive or negative, depending on the metal materials <cit.>. Without loss of generality, we assume D is uniform at the bent surfaces. The iDMI will give an interface torque <cit.>, change the exchange boundary condition into the formalism below, 𝐧·∇𝐦_i+Dd/A𝐳×(𝐞_t·∇)𝐦_i=0, with d the thickness of interface atomic layer and D_0=Dd is a constant, only depending the materials. This constant D_0 value accounts for the experimentally observed approximately inverse thickness dependence of spin wave frequency shift in ferromagnets/heavy metal heterostructures <cit.>. By solving the eigenvalue equation in Eq. (<ref>) with the boundary condition in Eq. (<ref>), we can obtain the magnon dispersions in the 𝐤-space. The strip geometry illustrated in Fig. <ref> (a) can also be calculated. The results are shown in the subsequent section. § RESULTS AND DISCUSSIONS §.§ Band topological phase transitions Before presenting the construction of anti-helical edge states, we first discuss the topological properties of magnons in AFM thin film embedded with only one type heavy metal dot array. The corresponding Brillouin zone (BZ) is shown in Fig. <ref> (a). We present the band dispersions at four different D_0 values for right-handed modes in Fig. <ref> (b). The left-handed bands are degenerate to the right-handed ones. When D=0, all the bands are topological trivial. The first (fourth) and second (fifth) bands are degenerate at K and K^' points, which gives multiple Dirac points in the spectrum. The second (fifth) and third (sixth) bands intersect at Γ point. A finite D lifts all the above the band degeneracies and open bandgap at Γ, K and K^' points. As the right-handed and left-handed modes are totally decoupled, we can use the Chern number to characterize the topological properties for the bands of each polarized mode. The berry connection of n-th band is given by 𝐀_n,±^𝐤=∓ i∫_ucψ_±,𝐤^n,†σ_z∇_𝐤ψ_±,𝐤^nd𝐫 and Berry curvature 𝐁_n,±^𝐤=∇_𝐤×𝐀_n,±^𝐤. The Chern number C_n,±=1/2π∫_BZB_n,z,±^𝐤d^2𝐤. The integral for the berry connection is done over the unit cell. The Chern number for the corresponding right-handed magnon bands are shown in Fig. <ref> (b) by the number on the bands. We find the Chern numbers for the left-handed magnon bands are opposite to the right-handed ones for each degenerate band. Therefore, we can use the spin Chern number to characterize the topological properties of the total system, C_n,s=(C_n,--C_n,+)/2. We can also define the n-th gap spin Chern number C̅_n,s associated with the gap above the n-th band as C̅_n,s=∑_n^'=1^nC_n^',s. The number C̅_n,s determined the number of helical edge states in the n-th gap according to the bulk-boundary correspondence. When D reverse sign, the Chern number for each mode will reverse the sign accordingly. The spin Chern number will also reverse sign. Therefore, the spin current due to the helical edge states are expected to change the flow direction. As the sign of D is determined by the heavy metal materials, we can design our system by choosing the heavy metal as desired. Based on this feature, we can construct the anti-helical edge magnons by structural design. For the first gap, a larger D will give a larger bandgap at K and K^' points, as shown in Fig. <ref> (c). At any value of D, the first gap spin Chern number is always -1. Interestingly, increasing D will drive a topological phase transition for the third and fourth bands. We find when D_0=D_0^c≃ 0.11A, the bandgap between the third and fourth bands at Γ point will be closed. A larger D_0 will reopen the gap, as shown in Fig. <ref> (b) and (d). The spin Chern number for the third band C_3,s will change from 1 to 0, while for the fourth band, C_4,s changes from 1 to 2. The total (spin) Chern number of the two bands remain unchanged. But this topological phase transition will change the third gap spin Chern number C̅_3,s from 0 to -1. This indicates that the third gap will experience a topologically trivial to nontrivial phase transition. From Fig. <ref> (b), we can see only the first and third bandgap are full gap. §.§ Helical edge magnons The nonvanishing n-th gap spin Chern number is expected to give rise to helical edge states covering the bandgap. To verify this prediction, we calculate the band dispersion in a strip geometry with a finite width along y direction and infinite along x direction. The strip geometry with only the upper edge termination are shown in Fig., <ref> (a). The ribbon band dispersions for the right-handed modes are shown in Fig. <ref> (b) with D_0=0.05A (left) and D_0=0.2A (right). The bands for left-handed modes are degenerate to the right-handed ones. For D_0=0.05A, We can see the edge states cover the first and second bandgap. The spin Chern number of the third gap is 0, indicating a trivial gap without supporting edge states. Edge states are also expected to exist in higher bands. But due to the absence of full gaps, they are not distinguishable from the bulk states, thus not shown. For D_0=0.2A, the edge states in the first and second gaps preserve. Additional edge states emerge in the third bandgap due to the topological phase transition for D_0>D_0^c compared to D_0<D_0^c. From the zoom in bands for D_0=0.05A near the first bandgap in Fig. <ref> (c), we can see that the edge states with the frequency located in the shaded region show a totally chiral feature. These edge states thus propagate unidirectionally. We can also see this feature for the states in the first and third gap for D_0=0.2A. The excitation of the states with other frequencies in the first gap and the states in the second gap will do not show unidirectional flow. In the subsequent simulations, we mainly focus on the states in the first bandgap. Due to the helical feature of the edge states, pure spin current locate at the edges. To verify the helical feature of topologically protected edge magnons, we perform micromagnetic simulations by selectively exciting spin waves with either one polarization. The method for the micromagnetic simulations is present in Sec V. The results are shown in Fig. <ref> (d) and (e). We apply a right-handed microwave source with the frequency located in the first bandgap. The excited right-handed spin waves flow unidirectionally towards the left [Fig. <ref> (d)], consistent with the in-pane magnetization distributions obtained from band calculation. When we changed the microwave source to the left-handed one, the excited spin waves move towards right [Fig. <ref> (e)]. When the excitation microwave source is linearly polarized, both magnon modes will be excited. A higher frequency located in the third bandgap will also give the same behavior. When the excitation source is placed in the opposite parallel edge, the moving direction of each mode will be opposite. These behaviors of excited spin waves clearly demonstrated the helical feature of the edge states. Same to the electronic systems, it will lead to a pure magnon spin current on each edge, counterpropagating along the two parallel edges. These states are topologically protected and are thus robust against the disorders from the nanofabrication process, same to our previous work <cit.>. As the sign of spin Chern number is locked to the sign of iDMI parameter D, which is dependent on the heavy metal materials. When we embed different heavy metal dot array to induce an opposite D, both of the polarized edge states will reverse the propagation direction, so as to the spin current. The feature in our system is the key point for the construction of anti-helical edge states. §.§ Anti-helical edge magnons Based on the locking behavior between spin Chern number and D, we now construct the anti-helical edge magnons in our system to make the edge spin currents on the two parallel edges flow towards the same direction. As the the sign of D is determined by the heavy metal materials, we design an interface in the middle of the finite width strip, illustrated in Fig. <ref> (a). Two different heavy metal materials are embedded along the two sides of the interface. The two metal materials are assumed to induce iDMI with opposite D, as illustrated in Fig. <ref> (a). For simplicity, we assume the absolute value of D is the same as different values give the same physics. We plotted the bands for both polarized modes in Fig. <ref> (b) at |D_0|=0.2A. The edge states are marked by red lines while the green lines denote the modes localized along the interface between the two domains with different embedded heavy metal dot array. These edge and interfacial states are still topologically protected. The edge modes are doubly degenerate for both spin, localized on the upper and lower edges, respectively. For each polarized edge states covering the first and third gaps, the two degenerate states propagate along the same direction, while the the interfacial states flow towards the opposite direction. The left-handed polarized bands can be mirrored to the right-handed ones respect to k_xa=π (time-reversal invariant point) and vice versa. On each edge and on the interface, the two polarized modes still counterflow to each other. As a result, the spin current on the two edges flow towards the same direction, compensated by the counterpropagating interfacial spin current, manifestation of the anti-helical feature for magnons in this designed structure. The anti-helical feature of the edge magnons can also be revealed in the micromagnetic simulations. The results are shown in Fig. <ref> (a) and (b). The excited left-handed spin waves on the parallel edges flow towards right while the interfacial ones flow to the left. On the contrary, the excited right-handed edge spin waves propagate towards left and the interfacial ones propagate to the right. These behaviors demonstrate the successful construction of anti-helical edge magnons. In a cell with a finite width along the x direction, a linearly polarized microwave source will excite both polarized spin waves, as shown in Fig. <ref> (c). When the spin waves meet the left and right vertical edges, they will be totally transmitted without backscattering due to the topological protection. But they can not penetrate into the vertical edges lower than the interface. This is because of the anti-helical feature, either polarized vertical edge magnons above and below the interface move in the opposite directions, different from the chiral and helical edge states in topological insulators. Instead, they will be transmitted into the interface, as shown in Fig. <ref> (c). However, after a longer time when transport to the opposite vertical edges, they will split into two beams, as illustrated in the inset of Fig. <ref> (c). One of the two will be transmitted to the lower horizontal edge. All the edge magnons on each horizontal edge can be transmitted to the other only through the interface. §.§ Discussions Above all, we have discussed the band topology due to the iDMIs, the helical also the anti-helical edge magnons in the patterned antiferromagnetic thin film. In our calculations, the thickness of the film is 4 nm. The other thickness will also give the same band dispersions when only considering the k_z=0 bulk modes. The modes with higher frequency at nonvanishing k_z for any film thickness are also expected to give the same physics as any k_z mode share the same symmetry in the xy plane. We adopt the triangular configuration for the embedded heavy metal array. The other configurations like the square lattice are also expected to give similar results. We here adopted a medium lattice constant (a=50 nm) for the calculations. A larger lattice constant will give a smaller bandgap, but does not alter the topological protection of the edge and interfacial states. Our model is experimentally feasible to construct. The helical, anti-helical and interfacial states can be probed in the spin pumping process as opposite magnon chirality could generate opposite spin polarization, discussed in recent experimental work <cit.>. They can also be detected by the polarization selective spectroscopy <cit.>. In addition, our configuration in a ferromagnetic thin film will lead to the antichiral edge states as there is only one magnon polarization in ferromagnets. § SUMMARY In this work, we proposed a theoretical scheme to realize the anti-helical edge magnons in antiferromagnetic thin film with embedded heavy metal dot array. The iDMIs induce helical edge states with the direction of spin current dependent on the sign of iDMI parameter. By constructing an interface with embedded different metal dots along the two sides in a finite width ribbon, edge spin current on the parallel two edges propagate in the same direction, compensated by conterpropagating spin current along the interface. Our proposal is experimentally feasible and the anti-helical edge states can be probed by the state-of-the-art techniques. Our work provides a new method controlling the magnon spin in antiferromagnets and should be quite helpful in the field of magnon spintronics based on antiferromagnets. § METHODS The nearest distance between the metal dot center (lattice constant) is a=50 nm. The radius of the metal dot is r=15 nm. The parameters in the coupled LLG equation are as follows: <cit.> K_z=8.55 GHz, A=7.25×10^-6 Hz·m^2, J_ex=1.105×10^11 Hz. The saturation magnetization M_s=1.94×10^5 A/m, the gyromagnetic ratio γ=2.21×10^5 Hz·m/A. The values of D_0 in the unit of A are given in the main text according to our need for the discussions. The micromagnetic simulation is based on the Eq. (<ref>). The damping constant is adopted as α=2×10^-4 to get a long decay length. We did not take the dipolar fields into calculation due to the antiferromagnet environment. We can selectively excite either one or both of the polarized modes simultaneously by setting the microwave fields. The effective field of the radio wave is given by 𝐡_rf(t)=h_0[cos(ω t)𝐞_x±sin(ω t)𝐞_y], where + denotes the right-handed source and - denotes the left-handed source. Only one component above along the x or y direction give a linearly polarized source. In our micromagnetic simulation, h_0=2×10^9 Hz, ω=2π×16.2 GHz. § ACKNOWLEDGEMENT This work is supported by the startup funding from Xiamen University. 99 MZHasan M. Z. Hasan and C. L. Kane, Colloquium: Topological insulators, https://doi.org/10.1103/RevModPhys.82.3045Rev. Mod. Phys. 82, 3045 (2010). XLQiX.-L. Qi and S.-C. Zhang, Topological insulators and superconductors, https://doi.org/10.1103/RevModPhys.83.1057Rev. Mod. Phys. 83, 1057 (2011). EColomes E. Colomés and M. Franz, Antichiral Edge States in a Modified Haldane Nanoribbon, https://doi.org/10.1103/PhysRevLett.120.086603Phys. Rev. Lett. 120, 086603 (2018). MMDenner M. M. Denner, J. L. Lado, and O. Zilberberg, Antichiral states in twisted graphene multilayers, https://doi.org/10.1103/PhysRevResearch.2.043190Phys. Rev. Research 2, 043190 (2020). XCheng X. Cheng, J. Chen, L.Zhang, L. Xiao, and S. Jia, Antichiral edge states and hinge states based on the Haldane model, https://doi.org/10.1103/PhysRevB.104.L081401Phys. Rev. B 104, L081401 (2021). SMandal S. Mandal, R. Ge, and T. C. H. Liew, Antichiral edge states in an exciton polariton strip, https://doi.org/10.1103/PhysRevB.99.115423Phys. Rev. B 99, 115423 (2019). DBhowmick D. Bhowmick and P. Sengupta, Antichiral edge states in Heisenberg ferromagnet on a honeycomb lattice, https://doi.org/10.1103/PhysRevB.101.195133Phys. Rev. B 101, 195133 (2020). JChen J. Chen, W. Liang, and Z.-Y. Li, Antichiral one-way edge states in a gyromagnetic photonic crystal, https://doi.org/10.1103/PhysRevB.101.214102Phys. Rev. B 101, 214102 (2020). PZhou P. Zhou, G.-G. Liu, Y. Yang, Y.-H. Hu, S. Ma, H. Xue, Q. Wang, L. Deng, and B. Zhang, Observation of Photonic Antichiral Edge States, https://doi.org/10.1103/PhysRevLett.125.263603Phys. Rev. Lett. 125, 263603 (2020). YangY Y. Yang, D. Zhu, Z. Hang, and Y. Chong, Observation of antichiral edge states in a circuit lattice, https://doi.org/10.1007/s11433-021-1675-0Sci. China Phys. Mech. Astron. 64, 257011 (2021). TJungwirth T. Jungwirth, X. Marti, P. Wadley, and J. Wunderlich, Antiferromagnetic spintronics, https://doi.org/10.1038/nnano.2016.18Nature Nanotech. 11, 231 (2016). VBaltz V. Baltz, A. Manchon, M. Tsoi, T. Moriyama, T. Ono, and Y. Tserkovnyak, Antiferromagnetic spintronics, https://doi.org/10.1103/RevModPhys.90.015005Rev. Mod. Phys. 90, 015005 (2018). YMLi1 Y.-M. Li, Y.-J. Wu, X.-W. Luo, Y. Huang, and K. Chang, Higher-order topological phases of magnons protected by magnetic crystalline symmetries, https://doi.org/10.1103/PhysRevB.106.054403Phys. Rev. B 106, 054403 (2022). RCheng R. Cheng, S. Okamoto, and D. Xiao, Spin Nernst Effect of Magnons in Collinear Antiferromagnets, https://doi.org/10.1103/PhysRevLett.117.217202Phys. Rev. Lett. 117, 217202 (2016). VAZyuzin V. A. Zyuzin and A. A. Kovalev, Magnon Spin Nernst Effect in Antiferromagnets, https://doi.org/10.1103/PhysRevLett.117.217203Phys. Rev. Lett. 117, 217203 (2016). ZWang Z. Wang, W. Bao, Y. Cao, and P. Yan, All-magnonic Stern-Gerlach effect in antiferromagnets, https://doi.org/10.1063/5.0096968Appl. Phys. Lett. 120, 242403 (2022). TWimmer T. Wimmer, A. Kamra, J. Gückelhorn, M. Opel, S. Geprägs, R. Gross, H. Huebl, and M. Althammer, Observation of Antiferromagnetic Magnon Pseudospin Dynamics and the Hanle Effect, https://doi.org/10.1103/PhysRevLett.125.247204Phys. Rev. Lett. 125, 247204 (2020). CJia C. Jia, M. Chen, A. F. Schäffer, and J. Berakdar, Chiral logic computing with twisted antiferromagnetic magnon modes, https://doi.org/10.1038/s41524-021-00570-0npj Comput. Mater. 7, 101 (2021). MWDaniels M. W. Daniels, R. Cheng, W. Yu, J, Xiao, and D. Xiao, Nonabelian magnonics in antiferromagnets, https://doi.org/10.1103/PhysRevB.98.134450Phys. Rev. B 98, 134450 (2018). WYu1 W. Yu, J. Lan, and J. Xiao, Magnetic Logic Gate Based on Polarized Spin Waves, https://doi.org/10.1103/PhysRevApplied.13.024055Phys. Rev. Applied 13, 024055 (2020). ABarman A. Barman, et al., The 2021 Magnonics Roadmap, https://doi.org/10.1088/1361-648X/abec1aJ. Phys.: Condens. Matter 33, 413001 (2021). LUdvardi L. Udvardi and L. Szunyogh, Chiral Asymmetry of the Spin-Wave Spectra in Ultrathin Magnetic Films, https://doi.org/10.1103/PhysRevLett.102.207204Phys. Rev. Lett. 102, 207204 (2009). JHMoon J.-H. Moon, S.-M. Seo, K.-J. Lee, K.-W. Kim, J. Ryu, H.-W. Lee, R. D. McMichael, and M. D. Stiles, Spin-wave propagation in the presence of interfacial Dzyaloshinskii-Moriya interaction, https://doi.org/10.1103/PhysRevB.88.184404Phys. Rev. B 88, 184404 (2013). KYamamoto K. Yamamoto, A.-M. Pradipto, K. Nawa, T. Akiyama, T. Ito, T. Ono, and K. Nakamura, Interfacial Dzyaloshinskii-Moriya interaction and orbital magnetic moments of metallic multilayer films, https://doi.org/10.1063/1.4973217AIP Advances 7, 056302 (2017). MdRKAkanda Md. R. K. Akanda, I. J. Park, and R. K. Lake, Interfacial Dzyaloshinskii-Moriya interaction of antiferromagnetic materials, https://doi.org/10.1103/PhysRevB.102.224414Phys. Rev. B 102, 224414 (2020). MKostylev M. Kostylev, Interface boundary conditions for dynamic magnetization and spin wave dynamics in a ferromagnetic layer with the interface Dzyaloshinskii-Moriya interaction, http://dx.doi.org/10.1063/1.4883181J. Appl. Phys. 115, 233902 (2014). PFerriani P. Ferriani, K. von Bergmann, E. Y. Vedmedenko, S. Heinze, M. Bode, M. Heide, G. Bihlmayer, S. Blügel, and R. Wiesendanger, Atomic-Scale Spin Spiral with a Unique Rotational Sense: Mn Monolayer on W(001), https://doi.org/10.1103/PhysRevLett.101.027201Phys. Rev. Lett. 101, 027201 (2008). KZakeri Kh. Zakeri, Y. Zhang, J. Prokop, T.-H. Chuang, N. Sakr, W. X. Tang, and J. Kirschner, Asymmetric Spin-Wave Dispersion on Fe(110): Direct Evidence of the Dzyaloshinskii-Moriya Interaction, https://doi.org/10.1103/PhysRevLett.104.137203Phys. Rev. Lett. 104, 137203 (2010). JCho J. Cho, N.-H. Kim, S. Lee, J.-S. Kim, R. Lavrijsen, A. Solignac, Y. Yin, D.-S. Han, N. J. J. van Hoof, H. J. M. Swagten, B. Koopmans, and C.-Y.You, Thickness dependence of the interfacial Dzyaloshinskii-Moriya interaction in inversion symmetry broken systems, https://doi.org/10.1038/ncomms8635Nature Commun. 6, 7635 (2015). NembachH H. T. Nembach, J. M. Shaw, M. Weiler, E. Jué, and T. J. Silva, Linear relation between Heisenberg exchange and interfacial Dzyaloshinskii-Moriya interaction in metal films, https://doi.org/10.1038/nphys3418Nature Phys. 11, 825 (2015). KDi K. Di, V. Li Zhang, H. S. Lim, S. C. Ng, M. H. Kuok, J. Yu, J. Yoon, X. Qiu, and H. Yang, Direct Observation of the Dzyaloshinskii-Moriya Interaction in a Pt/Co/Ni Film, https://doi.org/10.1103/PhysRevLett.114.047201Phys. Rev. Lett. 114, 047201 (2015). YMLi Y.-M. Li, J. Xiao, and K. Chang, Topological Magnon Modes in Patterned Ferrimagnetic Insulator Thin Films, https://doi.org/10.1021/acs.nanolett.8b00492Nano Lett. 18, 3032 (2018). YLiu Y. Liu, Z. Xu, Lin Liu, K. Zhang, Y. Meng, Y. Sun, P. Gao, H.-W. Zhao, Q. Niu, and J. Li, Switching magnon chirality in artificial ferrimagnet, https://doi.org/10.1038/s41467-022-28965-7Nature Commun. 13, 1264 (2022). YShiota Y. Shiota, T. Arakawa, R. Hisatomi, T. Moriyama, and T. Ono, Polarization-Selective Excitation of Antiferromagnetic Resonance in Perpendicularly Magnetized Synthetic Antiferromagnets, https://doi.org/10.1103/PhysRevApplied.18.014032Phys. Rev. Applied 18, 014032 (2022). WYu2 W. Yu, J. Lan, and J. Xiao, Polarization-selective spin wave driven domain-wall motion in antiferromagnets, https://doi.org/10.1103/PhysRevB.98.144422Phys. Rev. B 98, 144422 (2018).

http://arxiv.org/abs/2303.06734v1

20230312192931

Environmental variability and network structure determine the optimal plasticity mechanisms in embodied agents

q-bio.NC

[ Daodao Yang =============== The evolutionary balance between innate and learned behaviors is highly intricate, and different organisms have found different solutions to this problem. We hypothesize that the emergence and exact form of learning behaviors is naturally connected with the statistics of environmental fluctuations and tasks an organism needs to solve. Here, we study how different aspects of simulated environments shape an evolved synaptic plasticity rule in static and moving artificial agents. We demonstrate that environmental fluctuation and uncertainty control the reliance of artificial organisms on plasticity. Interestingly, the form of the emerging plasticity rule is additionally determined by the details of the task the artificial organisms are aiming to solve. Moreover, we show that co-evolution between static connectivity and interacting plasticity mechanisms in distinct sub-networks changes the function and form of the emerging plasticity rules in embodied agents performing a foraging task. § INTRODUCTION One of the defining features of living organisms is their ability to adapt to their environment and incorporate new information to modify their behavior. It is unclear how the ability to learn first evolved <cit.>, but its utility appears evident. Natural environments are too complex for all the necessary information to be hardcoded genetically <cit.> and more importantly, they keep changing during an organism’s lifetime in ways that cannot be anticipated <cit.>. The link between learning and environmental uncertainty and fluctuation has been extensively demonstrated in both natural <cit.>, and artificial environments <cit.>. Nevertheless, the ability to learn does not come without costs. For the capacity to learn to be beneficial in evolutionary terms, a costly nurturing period is often required, a phenomenon observed in both biological <cit.>, and artificial organisms <cit.>. Additionally, it has been shown that in some complex environments, hardcoded behaviors may be superior to learned ones given limits in the agent's lifetime and environmental uncertainty <cit.>. The theoretical investigation of the optimal balance between learned and innate behaviors in natural and artificial systems goes back several decades. However, it has recently found also a wide range of applications in applied AI systems <cit.>. Most AI systems are trained for specific tasks, and have no need for modification after their training has been completed. Still, technological advances and the necessity to solve broad families of tasks make discussions about life-like AI systems relevant to a wide range of potential application areas. Thus the idea of open-ended AI agents <cit.> that can continually interact with and adapt to changing environments has become particularly appealing. Many different approaches for introducing lifelong learning in artificial agents have been proposed. Some of them draw direct inspiration from actual biological systems <cit.>. Among them, the most biologically plausible solution is to equip artificial neural networks with some local neural plasticity <cit.>, similar to the large variety of synaptic plasticity mechanisms <cit.> that performs the bulk of the learning in the brains of living organisms <cit.>. The artificial plasticity mechanisms can be optimized to modify the connectivity of the artificial neural networks toward solving a particular task. The optimization can use a variety of approaches, most commonly evolutionary computation. The idea of meta-learning or optimizing synaptic plasticity rules to perform specific functions has been recently established as an engineering tool that can compete with state-of-the-art machine learning algorithms on various complex tasks <cit.>. Additionally, it can be used to reverse engineer actual plasticity mechanisms found in biological neural networks and uncover their functions <cit.>. Here, we study the effect that different factors (environmental fluctuation and reliability, task complexity) have on the form of evolved functional reward-modulated plasticity rules. We investigate the evolution of plasticity rules in static, single-layer simple networks. Then we increase the complexity by switching to moving agents performing a complex foraging task. In both cases, we study the impact of different environmental parameters on the form of the evolved plasticity mechanisms and the interaction of learned and static network connectivity. Interestingly, we find that different environmental conditions and different combinations of static and plastic connectivity have a very large impact on the resulting plasticity rules. § METHODS §.§ Environment We imagine an agent who must forage to survive in an environment presenting various types of complex food particles. Each food particle is composed of various amounts and combinations of N ingredients that can have positive (food) or negative (poison) values. The value of a food particle is a weighted sum of its ingredients. To predict the reward value of a given resource, the agent must learn the values of these ingredients by interacting with the environment. The priors could be generated by genetic memory, but the exact values are subject to change. To introduce environmental variability, we stochastically change the values of the ingredients. More precisely, we define two ingredient-value distributions E_1 and E_2 <cit.> and switch between them, with probability p_tr for every time step. We control how (dis)similar the environments are by parametrically setting E_2 = (1 - 2d_e)E_1, with d_e ∈ [0,1] serving as a distance proxy for the environments; when d_e = 0, the environment remains unchanged, and when d_e = 1 the value of each ingredient fully reverses when the environmental transition happens. For simplicity, we take values of the ingredients in E_1 equally spaced between -1 and 1 (for the visualization, see <ref>a, b). §.§ Static agent The static agent receives passively presented food as a vector of ingredients and can assess its compound value using the linear summation of its sensors with the (learned or evolved) weights, see <ref>. The network consists of N sensory neurons that are projecting to a single post-synaptic neuron. At each time step, an input X_t = (x_1, …, x_N) is presented, were the value x_i, i ∈{1, …, N} represents the quantity of the ingredient i. We draw x_i independently form a uniform distribution on the [0,1] interval (x_i ∼ U(0, 1)). The value of each ingredient w_i^c is determined by the environment (E_1 or E_2). The postsynaptic neuron outputs a prediction of the food X_t value as y_t = g(W X_t^T). Throughout the paper, g will be either the identity function, in which case the prediction neuron is linear, or a step-function; however, it could be any other nonlinearity, such as a sigmoid or ReLU. After outputting the prediction, the neuron receives feedback in the form of the real value of the input R_t. The real value is computed as R_t = W^c X_t^T + ξ, where W^c = (w_1^c, …, w_N^c) is the actual value of the ingredients, and ξ is a term summarizing the noise of reward and sensing system ξ∼𝒩(0, σ). For the evolutionary adjustment of the agent's parameters, the loss of the static agent is the sum of the mean squared errors (MSE) between its prediction y_t and the reward R_t over the lifetime of the agent. The agent's initial weights are set to the average of the two ingredient value distributions, which is the optimal initial value for the case of symmetric switching of environments that we consider here. §.§ Moving Agent As a next step, we incorporate the sensory network of static agents into embodied agents that can move around in an environment scattered with food. To this end, we merge the static agent's network with a second, non-plastic motor network that is responsible for controlling the motion of the agent in the environment. Specifically, the original plastic network now provides the agent with information about the value of the nearest food. The embodied agent has additional sensors for the distance from the nearest food, the angle between the current velocity and the nearest food direction, its own velocity, and its own energy level (sum of consumed food values). These inputs are processed by two hidden layers (of 30 and 15 neurons) with tanh activation. The network's outputs are angular and linear acceleration, <ref>. The embodied agents spawn in a 2D space with periodic boundary conditions along with a number of food particles that are selected such that the mean of the food value distribution is ∼ 0. An agent can eat food by approaching it sufficiently closely, and each time a food particle is eaten, it is re-spawned with the same value somewhere randomly on the grid (following the setup of <cit.>). After 5000 time steps, the cumulative reward of the agent (the sum of the values of all the food it consumed) is taken as its fitness. During the evolutionary optimization, the parameters for both the motor network (connections) and plastic network (learning rule parameters) are co-evolved, and so agents must simultaneously learn to move and discriminate good/bad food. §.§ Plasticity rule parametrization Reward-modulated plasticity is one of the most promising explanations for biological credit assignment <cit.>. In our network, the plasticity rule that updates the weights of the linear sensor network is a reward-modulated rule which is parameterized as a linear combination of the input, the output, and the reward at each time step: Δ W_t = η_p [ R_t · (θ_1 X_t y_t + θ_2 y_t + θ_3 X_t + θ_4)^Reward Modulated + (θ_5 X_t y_t + θ_6 y_t + θ_7 X_t + θ_8)_Hebbian ]. Additionally, after each plasticity step, the weights are normalized by mean subtraction, an important step for the stabilization of Hebbian-like plasticity rules <cit.>. We use a genetic algorithm to optimize the learning rate η_p and amplitudes of different terms θ = (θ_1, …, θ_8). The successful plasticity rule after many food presentations must converge to a weight vector that predicts the correct food values (or allows the agent to correctly decide whether to eat a food or avoid it). To have comparable results, we divide θ = (θ_1, …, θ_8) by θ_max = max_k|θ_k|. So that θ / θ_max = θ^norm∈ [-1,1]^8. We then multiply the learning rate η_p with θ_max to maintain the rule's evolved form unchanged, η_p^norm = η_p ·θ_max. In the following, we always use normalized η_p and θ, omitting ^norm. §.§ Evolutionary Algorithm To evolve the plasticity rule and the moving agents' motor networks, we use a simple genetic algorithm with elitism <cit.>. The agents' parameters are initialized at random (drawn from a Gaussian distribution), then the sensory network is trained by the plasticity rule and finally, the agents are evaluated. After each generation, the best-performing agents (top 10 % of the population size) are selected and copied into the next generation. The remaining 90 % of the generation is repopulated with mutated copies of the best-performing agents. We mutate agents by adding independent Gaussian noise (σ = 0.1) to its parameters. § RESULTS §.§ Environmental and reward variability control the evolved learning rates of the static agents To start with, we consider a static agent whose goal is to identify the value of presented food correctly. The static reward-prediction network quickly evolves the parameters of the learning rule, successfully solving the prediction task. We first look at the evolved learning rate η_p, which determines how fast (if at all) the network's weight vector is updated during the lifetime of the agents. We identify three factors that control the learning rate parameter the EA converges to: the distance between the environments, the noisiness of the reward, and the rate of environmental transition. The first natural factor is the distance d_e between the two environments, with a larger distance requiring a higher learning rate, <ref>c. This is an expected result since the convergence time to the “correct” weights is highly dependent on the initial conditions. If an agent is born at a point very close to optimality, which naturally happens if the environments are similar, the distance it needs to traverse on the fitness landscape is small. Therefore it can afford to have a small learning rate, which leads to a more stable convergence and is not affected by noise. A second parameter that impacts the learning rate is the variance of the rewards. The reward an agent receives for the plasticity step contains a noise term ξ that is drawn from a zero mean Gaussian distribution with standard deviation σ. This parameter controls the unreliability of the agent's sensory system, i.e., higher σ means that the information the agent gets about the value of the foods it consumes cannot be fully trusted to reflect the actual value of the foods. As σ increases, the learning rate η_p decreases, which means that the more unreliable an environment becomes, the less an agent relies on plasticity to update its weights, <ref>c. Indeed for some combinations of relatively small distance d_e and high reward variance σ, the EA converges to a learning rate of η_p ≈ 0. This means that the agent opts to have no adaptation during its lifetime and remain at the mean of the two environments. It is an optimal solution when the expected loss due to ignoring the environmental transitions is, on average, lower than the loss the plastic network will incur by learning via the (often misleading because of the high σ) environmental cues. A final factor that affects the learning rate the EA will converge to is the frequency of environmental change during an agent's lifetime. Since the environmental change is modeled as a simple, two-state Markov process (<ref>a), the control parameter is the transition probability p_tr. When keeping everything else the same, the learning rate rapidly rises as we increase the transition probability from 0, and after reaching a peak, it begins to decline slowly, eventually reaching zero (<ref>d). This means that when environmental transition is very rare, agents opt for a very low learning rate, allowing a slow and stable convergence to an environment-appropriate weight vector that leads to very low losses while the agent remains in that environment. As the rate of environmental transition increases, faster learning is required to speed up convergence in order to exploit the (comparatively shorter) stays in each environment. Finally, as the environmental transition becomes too fast, the agents opt for slower or even no learning, which keeps them near the middle of the two environments, ensuring that the average loss of the two environments is minimal (<ref>d). §.§ The form of the evolved learning rule depends on the task: Decision vs. Prediction The plasticity parameters θ = (θ_1, …, θ_8) for the reward-prediction task converge on approximately the same point, regardless of the environmental parameters (<ref>a). In particular, θ_3 → 1, θ_5 → -1, θ_i → 0 for all other i, and thus the learning rule converges to: Δ W_t = η_p [θ_3 X_tR_t + θ_5 X_ty_t] ≈η_p X_t (R_t - y_t). Since by definition y_t = g(W_t X_t^T) = W_t X_t^T (g(x) = x in this experiment) and R_t = W^c X_t^T + ξ we get: Δ W_t = η_p X_t(W^c - W_t)X_t^T + η_p ξ X_t^T. Thus the distribution of Δ W_t converges to a distribution with mean 0 and variance depending on η_p and σ and W converges to W^c. So this learning rule will match the agent's weight vector with the vector of ingredient values in the environment. We examine the robustness of the learning rule the EA discovers by considering a slight modification of our task. Instead of predicting the expected food value, the agent now needs to decide whether to eat the presented food or not. This is done by introducing a step-function nonlinearity (g(x) = 1 if x ≥ 1 and 0 otherwise). Then the output y(t) is computed as: y_t = 1, if W_tX_t^T ≥ 0, 0, if W_tX_t^T < 0. Instead of the MSE loss between prediction and actual value, the fitness of the agent is now defined as the sum of the food values it chose to consume (by giving y_t = 1). Besides these two changes, the setup of the experiments remains exactly the same. The qualitative relation between η_p and parameters of environment d_e, σ and p_tr is preserved in the changed experiment. However, the resulting learning rule is significantly different (<ref>). The evolution converges to the following learning rule: Δ W_t = η_p X_t [θ_3 R_t + θ_7], y_t = 0, η_p X_t[(θ_1 + θ_3) R_t + (θ_5 + θ_7)], y_t = 1. In both cases, the rule has the form Δ W_t = η_p X_t [α_y R_t + β_y]. Thus, the Δ W_t is positive or negative depending on whether the reward R_t is above or below a threshold (γ = -β_y / α_y) that depends on the output decision of the network (y_t = 0 or 1). Both learning rules (for the reward-prediction and decision tasks) have a clear Hebbian form (coordination of pre- and post-synaptic activity) and use the incoming reward signal as a threshold. These similarities indicate some common organizing principles of reward-modulated learning rules, but their significant differences highlight the sensitivity of the optimization process to task details. §.§ The learning rate of embodied agents depends on environmental variability We now turn to the moving embodied agents in the 2D environment. To optimize these agents, both the motor network's connections and the sensory network's plasticity parameters evolve simultaneously. Since the motor network is initially random and the agent has to move to find food, the number of interactions an agent experiences in its lifetime can be small, slowing down the learning. However, having the larger motor network also has benefits for evolution because it allows the output of the plastic network to be read out and transformed in different ways, resulting in a broad set of solutions. The agents can solve the task effectively by evolving a functional motor network and a plasticity rule that converges to interpretable weights (<ref>a). After ∼ 100 evolutionary steps (<ref>d), the agents can learn the ingredient value distribution using the plastic network and reliably move towards foods with positive values while avoiding the ones with negative values. We compare the dependence of the moving and the static agents on the parameters of the environment: d_e and the state transition probability p_tr. At first, in order to simplify the experiment, we set the transition probability to 0, but fixed the initial weights to be the average of E_1 and E_2, while the real state is E_2. In this experiment, the distance between states d_e indicates twice the distance between the agent's initial weights and the optimal weights (the environment's ingredient values) since the agent is initialized at the mean of the two environment distributions. Same as for the static agent, the learning rate increases with the distance d_e (<ref>b). Then, we examine the effect of the environmental transition probability p_tr on the evolved learning rate η_p. In order for an agent to get sufficient exposure to each environment, we scale down the probability p_tr from the equivalent experiment for the static agents. We find that as the probability of transition increases, the evolved learning rate η_p decreases (<ref>c). This fits with the larger trend for the static agent, although there is a clear difference when it comes to the increase for very small transition probabilities that were clearly identifiable in the static but not the moving agents. This could be due to much sparser data and possibly the insufficiently long lifetime of the moving agent (the necessity of scaling makes direct comparisons difficult). Nevertheless, overall we see that the associations observed in the static agents between environmental distance d_e and transition probability p_tr and the evolved learning rate η_p are largely maintained in the moving agents. Still, more data would be needed to make any conclusive assertions about the exact effect of these environmental parameters on the emerging plasticity mechanisms. §.§ Rule redundancy in the embodied agents A crucial difference between the static and the moving agents is the function the plasticity has to perform. While in the static agents, the plasticity has to effectively identify the exact value distribution of the environment in order to produce accurate predictions, in the embodied agents, the plasticity has to merely produce a representation of the environment that the motor network can evolve to interpret adequately enough to make decisions about which food to consume. To illustrate the difference, we plot the Pearson correlation coefficient between an agent's weights and the ingredient values of the environment it is moving in (<ref>e). We use the correlation instead of the MSE loss (which we used for the static agents in <ref>e) because the amplitude of the weight vector varies a lot for different agents and meaningful conclusions cannot be drawn from the MSE loss. For many agents, the learned weights are consistently anti-correlated with the actual ingredient values (an example of such an agent is shown in <ref>e). This means that the output of the sensory network will have the opposite sign from the actual food value. While in the static network, this would lead to very bad predictions and high loss, in the foraging task, these agents perform exactly as well as the ones where the weights and ingredients values are positively correlated, since the motor network can simply learn to move towards food for which it gets a negative instead of a positive sensory input. This additional step of the output of the plastic network going through the motor network before producing any behavior has a strong effect on the plasticity rules that the embodied agents evolve. Specifically, if we look at the emerging rules the top performing agents have evolved (<ref>a), it becomes clear that, unlike the very well-structured rules of the static agents (<ref>a), there is now virtually no discernible pattern or structure. The difference becomes even clearer if we look at the learned weights (at the end of a simulation) of the best-performing agents (<ref>c). While there is some correlation with the environment's ingredient value distribution, the variance is very large, and they do not seem to converge on the “correct” values in any way. This is to some extent expected since, unlike the static agents where the network's output has to be exactly correct, driving the evolution of rules that converge to the precise environmental distribution, in the embodied networks, the bulk of the processing is done by the motor network which can evolve to interpret the scalar value of the sensory network's output in a variety of ways. Thus, as long as the sensory network's plasticity rule co-evolves with the motor network, any plasticity rule that learns to produce consistent information about the value of encountered food can potentially be selected. To further test this assumption, we introduce a bottleneck of information propagation between the sensory and motor networks by using a step-function nonlinearity on the output of the sensory network (Eq. <ref>). Similarly to the decision task of the static network, the output of the sensory network now becomes binary. This effectively reduces the flow of information from the sensory to the motor network, forcing the sensory network to consistently decide whether food should be consumed (with the caveat that the motor network can still interpret the binary sign in either of two ways, either consuming food marked with 1 or the ones marked with 0 by the sensory network). The agents perform equally well in this variation of the task as before (<ref>d), but now, the evolved plasticity rules seem to be more structured (<ref>b). Moreover, the variance of the learned weights in the best-performing agents is significantly reduced (<ref>d), which indicates that the bottleneck in the sensory network is increasing selection pressure for rules that learn the environment's food distribution accurately. § DISCUSSION We find that different sources of variability have a strong impact on the extent to which evolving agents will develop neuronal plasticity mechanisms for adapting to their environment. A diverse environment, a reliable sensory system, and a rate of environmental change that is neither too large nor too small are necessary conditions for an agent to be able to effectively adapt via synaptic plasticity. Additionally, we find that minor variations of the task an agent has to solve or the parametrization of the network can give rise to significantly different plasticity rules. Our results partially extend to embodied artificial agents performing a foraging task. We show that environmental variability also pushes the development of plasticity in such agents. Still, in contrast to the static agents, we find that the interaction of a static motor network with a plastic sensory network gives rise to a much greater variety of well-functioning learning rules. We propose a potential cause of this degeneracy; as the relatively complex motor network is allowed to read out and process the outputs from the plastic network, any consistent information coming out of these outputs can be potentially interpreted in a behaviorally useful way. Reducing the information the motor network can extract from the sensory system significantly limits learning rule variability. Our findings on the effect of environmental variability concur with the findings of previous studies <cit.> that have identified the constraints that environmental variability places on the evolutionary viability of learning behaviors. We extend these findings in a mechanistic model which uses a biologically plausible learning mechanism (synaptic plasticity). We show how a simple evolutionary algorithm can optimize the different parameters of a simple reward-modulated plasticity rule for solving simple prediction and decision tasks. Reward-modulated plasticity has been extensively studied as a plausible mechanism for credit assignment in the brain <cit.> and has found several applications in artificial intelligence and robotics tasks <cit.>. Here, we demonstrate how such rules can be very well-tuned to take into account different environmental parameters and produce optimal behavior in simple systems. Additionally, we demonstrate how the co-evolution of plasticity and static functional connectivity in different sub-networks fundamentally changes the evolutionary pressures on the resulting plasticity rules, allowing for greater diversity in the form of the learning rule and the resulting learned connectivity. Several studies have demonstrated how, in biological networks, synaptic plasticity heavily interacts with <cit.> and is driven by network topology <cit.>. Moreover, it has been recently demonstrated that biological plasticity mechanisms are highly redundant in the sense that any observed neural connectivity or recorded activity can be achieved with a variety of distinct, unrelated learning rules <cit.>. This observed redundancy of learning rules in biological settings complements our results and suggests that the function of plasticity rules cannot be studied independently of the connectivity and topology of the networks they are acting on. The optimization of functional plasticity in neural networks is a promising research direction both as a means to understand biological learning processes and as a tool for building more autonomous artificial systems. Our results suggest that reward-modulated plasticity is highly adaptable to different environments and can be incorporated into larger systems that solve complex tasks. § FUTURE WORK This work studies a simplified toy model of neural network learning in stochastic environments. Future work could be built on this basic framework to examine more complex reward distributions and sources of environmental variability. Moreover, a greater degree of biological realism could be added by studying more plausible network architectures (multiple plastic layers, recurrent and feedback connections) and more sophisticated plasticity rule parametrizations. Additionally, our foraging simulations were constrained by limited computational resources and were far from exhaustive. Further experiments can investigate environments with different constraints, food distributions, multiple seasons, more complex motor control systems and interactions of those systems with different sensory networks as well as the inclusion of plasticity on the motor parts of the artificial organisms. § ACKNOWLEDGEMENTS This work was supported by a Sofja Kovalevskaja Award from the Alexander von Humboldt Foundation. EG and SK thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for their support. We acknowledge the support from the BMBF through the Tübingen AI Center (FKZ: 01IS18039A). AL is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 – Project number 39072764. apalike

http://arxiv.org/abs/2303.07380v2

20230313180109

Killing horizons and surface gravities for a well-behaved three-function generalisation of the Kerr spacetime

gr-qc

Plots/

http://arxiv.org/abs/2303.07257v2

20230313162902

The Audio-Visual BatVision Dataset for Research on Sight and Sound

cs.RO

[ \begin@twocolumnfalse This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. \end@twocolumnfalse ] Ab initio electron-lattice downfolding: potential energy landscapes, anharmonicity, and molecular dynamics in charge density wave materials Tim O. Wehling March 30, 2023 =========================================================================================================================================== empty empty Vision research showed remarkable success in understanding our world, propelled by datasets of images and videos. Sensor data from radar, LiDAR and cameras supports research in robotics and autonomous driving for at least a decade. However, while visual sensors may fail in some conditions, sound has recently shown potential to complement sensor data. Simulated room impulse responses (RIR) in 3D apartment-models became a benchmark dataset for the community, fostering a range of audiovisual research. In simulation, depth is predictable from sound, by learning bat-like perception with a neural network. Concurrently, the same was achieved in reality by using RGB-D images and echoes of chirping sounds. Biomimicking bat perception is an exciting new direction but needs dedicated datasets to explore the potential. Therefore, we collected the BatVision dataset to provide large-scale echoes in complex real-world scenes to the community. We equipped a robot with a speaker to emit chirps and a binaural microphone to record their echoes. Synchronized RGB-D images from the same perspective provide visual labels of traversed spaces. We sampled modern US office spaces to historic French university grounds, indoor and outdoor with large architectural variety. This dataset will allow research on robot echolocation, general audio-visual tasks and sound phænomena unavailable in simulated data. We show promising results for audio-only depth prediction and show how state-of-the-art work developed for simulated data can also succeed on our dataset. The data can be downloaded at <https://forms.gle/W6xtshMgoXGZDwsE7> § INTRODUCTION Large-scale datasets propelled research in past decades, providing first static images and later an abundance of videos for tasks from object detection to activity recognition. Sounds, correlated with visual data from their source, provide exploitable information about an action or context, often at marginal computational overhead. A novel research direction aims to listen to the environment for improved task performance. Simulated room impulse response (RIR) datasets allow researchers to investigate the interaction of sounds with known space layouts for e.g. depth prediction, obstacle avoidance or to drive towards an alarm beyond the line of sight. They have been used successfully to predict 3D layouts from simulated chirps, similar to how bats find their prey. Robots mastering this echolocation could create instant maps beyond their immediate surroundings without LiDAR, overcoming their field of view and operate even in darkness and smoke. While some datasets provide recorded RIRs within limitations, there exists no large-scale real dataset to investigate audio-visual 3D scene understanding for robots. This hinders domain adaptation from simulation as well as learning to be robust against or even exploit real world background sounds. Bats are capable of navigating and localizing prey in flight using echolocation. We showed successfully how to adopt this feat for machine listening, using only audible echoes of chirps for depth prediction <cit.>. To understand the extend to which this method can complement failing vision sensors and even extend beyond the field of view, more real data in typical robotic scenarios is needed. We therefore present the BatVision Dataset providing publicly available large-scale real data of complex scenes for research on audio-visual 3D scene understanding and robotic echolocation. Sound-augmented Task Performance. Sound arrives omnidirectional and provides rich information about the space it crosses. Due to the correlation of sound and vision it can provide complementary information when visual sensors fail. Cameras and LiDAR struggle with occlusions, reflections and smoke. Depth-from-stereo algorithms produce artifacts in low-light conditions and when objects lack textures. Combining audio-visual information can improve task performance with one modality providing labels for the other schulz2021hearing,valverde2021there. Biomimicking the echolocation principle can even improve 3D scene understanding using sound christensen2020batvision,gao2020visualechoes,zhang2022stereo. Audio-visual data suitable for these tasks is sparse because real-world recordings require complex setups and long sessions <cit.>. In consequence, researchers often create small datasets tailored to their tasks <cit.>). For research on robotic echolocation, only very few simple datasets exists. The BatVision Dataset. Motivated by the success of simulated datasets and the lack of real data for robotic echolocation and audio-visual scene understanding, we recorded the real-life BatVision dataset. We provide large-scale recordings from traversing real campus spaces with a robot, covering a wide range of materials, room shapes and objects, shown in <Ref>. We provide synchronized RGB-D camera images and recorded echoes of chirping sounds emitted with a forward facing speaker. Similar to RIRs, echoes can be used to infer the geometry of the room to allow the robot bat-like perception even in darkness, smoke and fog. Data was collected at UC Berkeley (52,220 instances) and Mines Paris PSL (3,120 instances). Emitted chirp signals range between 20Hz and 20kHz. RGB-D images were recorded with an Intel Realsense camera mounted on a robot (Mines Paris) and a ZED camera (UC Berkeley). In the following we place the BatVision dataset among similar ones and argue its unique utility. We then detail the collection process, data and showcase performance of depth prediction methods trained on it. UC Berkeley data will be referred to as BV1 and Mines Paris data as BV2. § RELATED WORK Audio-visual data is available in many datasets, collected in reality or simulation, suitable for different tasks. However, to our knowledge, no existing dataset supports audio-visual 3D scene understanding with large-scale real data. Acoustic Room Impulse Response Datasets. RIR datasets are used in room acoustics, speech and audio processing, sound localization and separation and virtual reality. RIRs account for the effect of room acoustics on audio signals and describe sound propagation. By convolving with RIRs, acoustic properties, such as reverb, can be transferred onto arbitrary dry sounds. Several datasets exist with RIRs, sampled in real spaces using a complex measurement routine. The Acoustic Multichannel RIR Dataset <cit.> contains samples from one rectangular room. Panels, moved between recording sessions, emulate environments such as a small office, meeting or lecture room. The dEchorate dataset <cit.> provides 1.8k sampled RIRs with annotated early echo timings and 3D positions of microphones and sound sources. For the MIRaGe dataset <cit.> 371k RIRs are measured in a dense grid of 4104 source positions in different rooms. <cit.> estimates RIRs from spherical camera images. RIRs were sampled as ground truth in living room style environments using a custom array of 48 microphones and a soundfield microphone. Object shapes are simple rectangles, aligned towards the main axis of the room. While these datasets were sampled with high quality equipment and allow room geometry and RIRs estimation their environments are rectangular rooms with cuboid objects and simple materials. In contrast, for BatVision we sampled data in real public spaces, selected for their variety of materials, shape and architectural properties. Audio-Visual Simulation. Datasets containing simulated data are widely used, cost-effective, allow large scale generation, controlled conditions and easy data annotation <cit.>, <cit.>). Simulation of sound propagation has been extensively used in games and AR applications. SoundSpaces 1.0 <cit.> allows simulating sound propagation by providing RIR renderings built with bidirectional path-tracing in 3D-scanned apartment models. They were generated for discreet positions and orientations in a grid in two 3D environments, Matterport3D <cit.>) and Replica<cit.>). SoundSpaces 2.0 <cit.> provides continuous on-the-fly rendering and improved the sound propagation. It has since become a widely used benchmark showing impressive performance on a wide range of tasks <cit.>. While it is a huge contribution to the community, bridging the gap between simulation and the real world stays a significant challenge. Models trained in simulation often overfit to characteristics of the simulator in surprising ways kadian2020sim2real. However, for audio-based 3D scene understanding there is no real-life dataset of a size, comparable to SoundSpaces which motivated our data collection. Sound Event Localization and Detection Datasets. Sound event localization and detection (SELD) aims to infer the azimuth, elevation and distance of sounds relative to an observer, with or without additional classification. Datasets in the domain contain one or several, moving or static, clear or noisy sounds in different environments. The TUT Acoustic Scenes 2016 dataset <cit.> consists of 15 real acoustic scenes, such as City center or Metro station, with annotated sound event classes, onset and offset times, and spatial information. The STARSS22 dataset <cit.> is a collection of 22 real-world acoustic scene recordings with sound event annotations, captured with a high-resolution spherical microphone array. Annotations cover 13 classes and direction of arrivals. For 3D scene understanding and especially reconstruction, SELD datasets often lack sufficient spatial ground truth information of the environments the sounds occur in. For the BatVision dataset forward facing camera images and depth information is provided to allow exploring the correlations in the audio and visual modality. Real-World Audio-Visual Datasets. The huge EGO 4D dataset <cit.> contains videos from the perspective of persons performing a wide range of activities. Parts contain audio and 3D meshes of the environment. It is an impressive effort and contribution for many research areas such as activity recognition. In contrast, for BatVision we emit chirps into recorded scenes to allow inferring RIRs and finally the 3D scene. It is magnitudes smaller but better suited for our tasks. <cit.> proposed “The Greatest Hits” dataset. By filming objects being struck with a drumstick, they aim to study physical interactions with a visual scene and synthesize plausible impact sounds. It includes 46,577 actions of hitting and scratching objects. The research is a step towards understanding the link between material, action and emitted sound. Similarly, with our recorded echoes we collect interactions between chirps and scene materials. The authors in <cit.> predict approaching vehicles at blind intersections from sound before they enter the line-of-sight. They captured crossing vehicles at intersections with a custom microphone array and a front-facing camera mounted on a car. <cit.> recorded the "Omni Auditory Perception Dataset" standing next to streets with eight binaural microphones and a 360° camera. Pseudo ground truth depth is predicted from monocular images. The authors emphasize its mid-size and the great effort required to create it. <cit.> introduced the "Quiet Campus Dataset" of ambient sounds from a variety of quiet indoor scenes. Paired with RGB-D images, they predict distances to walls from the whirling of a fan or noise coming through a window. While this unique idea shows impressive task performance with passive observations, we focus on active sounds which are humanly audible for improved performance. Recording interactions of chirps with spaces has a recent history. The BatVision depth-from-binaural audio idea <cit.> inspired <cit.> to collect similar data with added 360° LiDAR scans for depth and four ear-shaped microphones. With 5,000 samples the dataset falls between BV1 and BV2. <cit.> predict an occupancy map from conversations in spaces. Beyond using SoundSpaces they also captured real data in a mock-up apartment, citing the lack of publicly available real world data. To compute RIRs they capture chirps from a speaker with an Eigenmike. The "Studio Dataset" <cit.> records 1478 samples of chirp echoes with four microphones and RGB-D images with a RealSense camera. Everything is fixed into one metal frame facing in one direction. While also testing in SoundSpaces, their real-world recordings are done in one rectangular room showing simple cuboid objects. The recording of so many datasets, tailored to specific needs show the potential impact that more and larger publicly available datasets can have. Presented tasks from depth prediction to occupancy mapping can be investigated with our BatVision dataset. We will describe in the following how and give details on the extend of our data collection. § THE AUDIO-VISUAL BATVISION DATASET §.§ Dataset Overview We collected the Audio-Visual BatVision Dataset at various locations at the UC Berkeley (ICSI), and Ecole des Mines Paris using small robots, carrying all sensors. The sites provide varied architectural styles, room shapes, materials and therefore acoustic impressions. We crossed lecture halls, corridors, offices and cobblestone paths recording while emitting humanly audible sine swept signals (chirps) with a forward facing speaker (<Ref>). We recorded their echoes, forward facing camera images and provide depth-maps from the same perspective. Chirps ascend from 20Hz to 20kHz in 3 ms. Their use is motivated by their counterparts in nature, helping animals echolocalize, but they are also common in sound engineering to record RIRs farina2000simultaneous. Binaural microphones record audio at 44.1kHz with 24 bits to keep the full frequency range of the chirps. Each instance is one chirp synchronized with one 1280x720 RGB-D image. An overview of this dataset and other real-world dataset is available in <Ref> The distribution of the collected depth can be seen at <ref> and <ref>. By comparing the average depth value per instances it shows BV2 is more long-tailed. BV1 depth values are clipped to the BV2 max depth value to remove outliers of the stereovision algorithm used. Pixel-wise average depth shows a more complex scene distribution in BV2. Data collected at the UC Berkeley (BV1). In Berkeley, 52,220 instances were collected at two floors of an institute containing hallways, open areas, conference rooms and offices. At two distinct areas of one floor, 39,564 and 7,618 instances were collected for training and validation and 5,038 instances on another floor for testing. While similar, the floors' spatial layout, furniture, occupancy, and decorations are different, see <Ref>. For details, please see christensen2020batvision showcasing the initial depth prediction-from-audio idea. A JBL Flip4 Bluetooth speaker emitted chirps while two USB Lavalier MAONO AU-410 microphones, embedded into silicone ears, recorded their echoes. They were mounted 23.5cm apart while the speaker sat between (<Ref>). We excluded motor sounds by pushing the robot on a trolley. We used a ZED stereo camera to record images of the scene ahead and calculated depth maps with the camera API. Depth-from-stereo fails for some pixel which are NAN. We provide depth maps with RGB images from the left camera. Designed for smaller spaces, audio recordings were cut at 72.5ms, including echoes from objects at 12m distance. This trades-off perception at a relevant distance while excluding later noisy reverberations. We also clip depth to 12m during training even though further distances are available. Data collected in Mines Paris (BV2). In Ecole des Mines Paris, we collected 3,120 instances with large visual and acoustic variety (See <Ref>). We split data into sets of 1,911 train, 625 validation and 584 test instances. Given the multi-modal scene distribution we aim to balance task difficulty. We split instances by time of recording when moving through rooms, avoiding loops. That way, our incremental coverage of rooms will lead to poses being sufficiently separated in the sets. Even if we revisit parts of rooms we chose different trajectories thereby avoiding repeated poses. Tasks on this split will be harder than if instances were randomly assigned, which could result in neighbouring poses ending up in training and test. Simple interpolation will not yield best performance. Admittedly our split is easier than separating complete rooms as done for BV1. We observed outdoor reconstruction performance is acceptable but sub-par suggesting domain shift. If very different features, i.e. outdoors, carpeted rooms or stone corridors, are not in the training data further domain adaptation would be necessary which is left for future work. We provide monocular RGB images and depth from an Intel RealSense D455 camera. The chirp-emitting speaker is identical to BV1 but we recorded echoes with a binaural 3Dio Free Space microphone. A robot carried all hardware and an Nvidia Jetson TX-1 to run all recording software (see <Ref>). We excluded motor noise by switching between stopping and driving and filtered out instances while moving. Synchronization was achieved using ROS timestamps and chirps detection with manual checks. Audio data instances were cut to be 0.45s long which includes echoes from objects up to 75m away. Because of larger spaces in BV2, we keep the long tail to sense far away reflection. The maximum depth value of the camera is ≈64m. §.§ Limitations Unlike in simulation, physically recording introduces a number of limitations. Echoes were recorded in real rooms next to noisy streets and with typical sounds of busy academic institutions. Some data may therefore contain audible noise, typical for the recording context. Models trained will need to learn to be robust to these noise profiles. Some approaches predicting RIRs need independently changing emitter and microphone positions to record sounds on a direct path carlo2021dechorate, luo2022learning. In our data collection by driving, the robot carries all the equipment so the emitter is fixed close to the microphone and only the echoes change. After filtering instances in motion few remain from the same pose. We kept these to allow learning a noise-model but they can be filtered by thresholding optical flow. Collection with different hardware for BV1 and 2 leads to different formats. Audio instances collected at UC Berkeley are shorter. Some very bright images from Mines Paris shows slight purple discoloration due to an issues with the Intel RealSense D455 RGB-D camera. Finally, our consumer-grade speaker can not produce the full frequency spectrum. § DEPTH PREDICTION ON BATVISION DATA The data can be used for depth-prediction from audio-visual data with approaches developed for real or simulated data. For illustration we trained a U-Net audio-only depth prediction baseline and compared with one state-of-the-art audio-visual approach developed by its authors in simulation. Beyond Image to Depth. Recent work improves audio-visual depth prediction using a pre-trained material classifier to decide which modality to pay attention to. The authors parida2021beyond extract latent material features from SoundSpaces 1.0 camera images and generate attention maps for the visual and audio-stream in the network. They convolve a dry-chirp sound with SoundSpace RIRs to simulate emitting it into spaces from different poses. Captured echoes and RGB-D images make up their dataset with which they predict depth from RGB images and audio. Training their code on the BatVision dataset shows similar performance as on SoundSpaces (see <Ref>). Apart from changing the spectrogram resolution, no further adjustments were needed when switching to real data. We ablate results in a study using the audio signal only by setting the RGB images to zero. That way we keep the architecture unchanged. Qualitatively, fine structures are harder to predict which may stem from our coarser ground truth depth compared to simulation (See <Ref>). Measuring depth in reality, either from stereo (ZED camera) or active stereo (RealSense camera), limits the accuracy depending on the object distance. Comparable results show that an approach developed for simulated data can be adapted and works with the same hyperparameters on our real dataset. U-Net Baseline. The extend to which audio can complement visual sensor data is not clear. Related work showed repeatedly the quality of the visual signal strongly dominates task performance. The compared work parida2021beyond investigates material-based attention to let the network choose the influence of each modality more explicitly. We investigate the claim that audio can help in bad visual conditions with a baseline, based on the audio signal alone. This serves to investigate the contribution of the audio signal in isolation with clear quality attribution. However, in all practical applications, the visual signal should be used as well. We train a U-Net from audio only, similar to related work <cit.> and isolated on BV2 and BV1. We achieve solid results, correctly predicting free space, obstacles and the general room layout (see <Ref>). Without further tweaks such as GCC-Phat features or using a GAN christensen2020batvision,christensen2020gccbatvision, the performance shows the exploitable quality of the data itself. Implementation Details. The only input are spectrograms, generated from waveforms with 512 frequency bins (nfft), 64 window and 16 hop length, resized to 256x256. For BV2, best performance is obtained with depth clipped to 30m and audio cut accordingly. This can be explained as audio energy decays with distance so far traveling echoes are hard to distinguish from noise. For BV1, depth is clipped to 12m and the audio accordingly. Ground truth depth is always normalized using the chosen max depth value. We use an 8 block U-Net with skip connection. Each encoder block is composed of 2D convolutions, batch normalization and leaky ReLUs. Each decoder block is composed of 2D transposed convolution, batch normalization and ReLUs (see <Ref> ). We found skip-connections improve performance though contrary to a segmentation task their contribution is not yet clearly understood. We train with 256 batch size, 0.002 learning rate for BV2 and 0.001 learning rate for BV1 with the AdamW optimizer and L1 loss. Results on Depth Prediction. With this simple baseline we retrieve the general geometry of the space (see <Ref> ). On BV2, complex obstacles such as chairs are visible. When the visual sensors fails (e.g on glass), audio gives correct information about the depth. The network generalizes well between different acoustic environment. It reproduces corridors robustly built with various material (e.g carpet floor and glass wall, stone and tiled floor, see <Ref>). Outdoors the performance is diminished with the network underestimating depth systematically. This shows some inevitable bias of the data having a majority of indoor data. Trained on BV1 data, corridors and obstacles are well reconstructed even though finer structures are equally lost. § CONCLUSION Recent success in using simulated audio-visual data for scene understanding shows the potential of the audio modality. Depth prediction from sound alone or in addition to vision is possible, allowing to perceive the environment like a bat. The BatVision dataset will support the research community with large-scale real audio-visual data to improve task performance and uncover novel uses. We present an audio-only depth-prediction baseline as starting point and obtain good results when training a state-of-the-art approach on our data. Future datasets could include ultrasound chirps to enable inaudible human-robot collaboration. § ACKNOWLEDGMENT We thank all our collaborators, especially Daniel Lin for data collection, Jesper Haahr Christensen for insights into sonar and collaboration on the original idea, Karl Zipser for conception of the robot for BV1. Finally David Mazouz and Jacky Lech for advice and help building the robot for BV2. The authors acknowledges the support of the French Agence Nationale de la Recherche (ANR), under grant ANR-22-CE94-0003 (projet Omni-BatVision).

http://arxiv.org/abs/2303.07170v1

20230313151303

Validation of uncertainty quantification metrics: a primer

physics.chem-ph

Institut de Chimie Physique, UMR8000 CNRS,  Universit Paris-Saclay, 91405 Orsay, France pascal.pernot@cnrs.fr The practice of uncertainty quantification (UQ) validation, notably in machine learning for the physico-chemical sciences, rests on several graphical methods (scattering plots, calibration curves, reliability diagrams and confidence curves) which explore complementary aspects of calibration, without covering all the desirable ones. For instance, none of these methods deals with the reliability of UQ metrics across the range of input features. Based on three complementary concepts, calibration, consistency and adaptivity, the toolbox of common validation methods for variance- and intervals- based metrics is revisited with the aim to provide a better grasp on their capabilities. This study is conceived as an introduction to UQ validation, and all methods are derived from a few basic rules. The methods are illustrated and tested on synthetic datasets and examples extracted from the recent physico-chemical machine learning UQ literature. Validation of uncertainty quantification metrics: a primer Pascal PERNOT 0000-0001-8586-6222 March 30, 2023 =========================================================== § INTRODUCTION The quest for confidence in the predictions of data-based algorithms<cit.> has led to a profusion of uncertainty quantification (UQ) methods in machine learning (ML).<cit.> Not all of these UQ methods provide uncertainties that can be relied on,<cit.> notably if one expects uncertainty to inform us on a range of plausible values for a predicted property.<cit.> In pre-ML computational chemistry, UQ metrics consisted essentially in standard uncertainty, i.e. the standard deviation of the distribution of plausible values (a variance-based metric), or expanded uncertainty, i.e. the half-range of a prediction interval, typically at the 95 % level (an interval-based metric).<cit.> The advent of ML methods provided UQ metrics beyond this standard setup, for instance distances in feature or latent space<cit.> or the recent Δ-metric,<cit.> which have no direct statistical or probabilistic meaning. These metrics might however be converted to variance- or interval-based metrics by calibration methods, such as conformal inference.<cit.> Still, all UQ metrics need to be validated to ensure that they are adapted to their intended use. The validation of UQ metrics is designed around the concept of calibration. The use of uncertainty as a proxy to identify unreliable predictions, as done in active learning, does not require the same level of calibration as its use for the prediction of properties at the molecule-specific level.<cit.> A handful of validation methods exist that explore more or less complementary aspects of calibration. A trio of methods seems to have recently taken the center-stage: the reliability diagram (or RMSE vs RMV plot), the calibration curve and the confidence curve. They implement three different approaches to calibration which are not necessarily independent, but they do not cover the full spectrum of calibration requirements. In particular, none of these methods addresses the essential reliability of predicted uncertainties with respect to the input features. As will be shown below, a UQ metric validated by this trio of methods might still be unreliable for individual predictions. Moreover, it appears that some methods are not always used in an appropriate context, such as the oracle confidence curve for variance-based UQ metrics.<cit.> The aim of this article is to design a principled validation framework based on complementary calibration concepts and review the relevant methods within this framework. The choice of methods is based on two main criteria: (1) the nature of the UQ metric to be validated (distribution, interval, variance, other...) and (2) the calibration target (average calibration, consistency or adaptivity, to be defined below). §.§ Related studies Most of the validation methods presented in this study derive from the seminal work of Gneiting et al.<cit.> on the calibration of probabilistic forecasters. Probabilistic forecasters are models that provide for each prediction a distribution of plausible values. A major lesson from Gneiting's work is that calibration metrics estimated over a validation dataset (average calibration) do not necessarily lead to useful UQ statements and that additional properties are necessary to design reliable UQ methods. An example is sharpness which quantifies the concentration of the prediction uncertainties: among a set of average-calibrated methods, one should prefer the sharpest one. Pernot<cit.> noted that, in the absence of a target value, sharpness might not be useful for the validation of individual UQ metrics, so that alternative statistics are required for UQ validation. The definition of calibration provided by Gneiting et al. was extended by Kuleshov<cit.>. The formulation of Kuleshov, based on coverage probabilities, led to use calibration curves as a validation tool. But Levi et al.<cit.> demonstrated that calibration curves are not reliable and proposed instead to use individual calibration<cit.>, based on the conditional variance of the errors with respect to the prediction uncertainty. Implementation of this conditional calibration equation led to reliability diagrams or RMSE vs RMV plots, also called calibration diagrams by Laves et al.<cit.>. In practice, individual calibration is generally unreachable, and an alternative is to consider local calibration<cit.>, which is reflected in the implementation of conditional statistics through binning schemes, as for instance in reliability diagrams. Pernot<cit.> proposed the concept of tightness to differentiate local calibration from average calibration, and implemented it in a Local Z-Variance (LZV) analysis in which calibration is estimated in subsets of the validation dataset. These subsets can be designed according to any relevant property (predicted value, uncertainty or input feature). He also established the link between LZV analysis in uncertainty space and reliability diagrams. Both reliability diagrams and LZV analysis derive from the concept of conditional calibration, but, for reliability diagrams, the conditioning variable is prediction uncertainty, while the LZV analysis opens a larger palette of conditioning variables. Conditional coverage with respect to input features was proposed by Vovk<cit.> to asses the adaptivity<cit.> of conformal predictors. The adaptivity concept can therefore be linked to conditional calibration in input feature space, a form of tightness. Confidence curves derive from another validation approach, known as sparsification error curves mainly used in computer vision.<cit.> They were not intended to validate calibration, but to estimate the correlation between absolute errors and UQ metrics.<cit.> Pernot<cit.> showed that the standard reference curve (the so-called oracle) used for the evaluation of confidence curves is irrelevant for variance-based UQ metrics, and he introduced a new probabilistic reference that enables to test conditional calibration in uncertainty space. Tran et al.<cit.> and Scalia et al.<cit.> published motivating overviews of some of the methods presented here (reliability diagrams, calibration and confidence curves) applied to ML-UQ in materials sciences. §.§ Contributions Building on the the works of Levi et al.<cit.>, Pernot<cit.> and Angelopoulos et al.<cit.> about conditional calibration, I propose to distinguish two calibration targets besides average calibration, namely * consistency as the conditional calibration with respect to prediction uncertainty, and * adaptivity as the conditional calibration with respect to input features. Consistency derives from the metrological consistency of measurements,<cit.> while adaptivity is borrowed from the conformal inference literature<cit.>. Tightness, as introduced by Pernot<cit.>, should then be understood as covering both consistency and adaptivity. Unless there is a monotonous transformation between input features and prediction uncertainty, consistency and adaptivity are distinct calibration targets. The distinction between consistency and adaptivity is important to better define the objective(s) of each validation method. In fact, adaptivity has not been considered in recent overviews of ML-UQ validation metrics<cit.>, but was practically implemented – albeit unnamed – through the LZV and Local Coverage Probability (LCP) analyses (see for instance Fig. 4 in Pernot<cit.>). Still, most ML-UQ validation studies focus on calibration and consistency, and I will show that this is not sufficient to ensure the reliability of an UQ method across the input features space. Adaptivity is essential to achieve reliable UQ at the molecule-specific level advocated by Reiher<cit.>. §.§ Structure of the article The theoretical bases of the calibration-consistency-adaptivity framework are presented in the next section (Sect. <ref>) for variance- and interval-based UQ metrics. A comprehensive set of validation methods is reviewed next (Sect. <ref>) according to their application range and calibration target, in order to appreciate their merits and limitations. Examples from the recent materials and chemistry ML-UQ literature are treated as case studies in Sect. <ref>. Recommendations are presented as conclusions (Sect. <ref>). § VALIDATION CONCEPTS AND MODELS A major distinction will be made below between variance-based UQ metrics and interval-based UQ metrics. These occur in ML as statistical summaries of empirical distributions or ensembles, or as parameters of theoretical distributions (typically normal). Other UQ metrics, such as distances in feature space or latent space,<cit.> or the Δ-metric,<cit.> which have not the correct dimension to be comparable to errors, need first to be converted to one of those two metrics to be usable for validation.<cit.> §.§ Validation model for variance-based UQ metrics §.§.§ Validation datasets In order to validate variance-based UQ results, one needs a set of predicted values V={V_i}_i=1^M, the corresponding uncertainties u_V={u_V}_i=1^M, reference data to compare with R={R_i}_i=1^M, and possibly their uncertainties u_R={u_R_i}_i=1^M. Most of the validation methods considered below require to transform these to E=R-V and u_E=(u_R^2+u_V^2)^1/2. The minimal validation dataset, C={ E,u_E}, enables to test calibration and consistency but not adaptivity. It is therefore better to include a relevant input feature X in the dataset C={ E,u_E,X}, extending the validation targets to adaptivity in X space. Alternatively, if validation is done a posteriori and no relevant input feature is available, adaptivity can be tested on C={ E,u_E,V}, with some caveats presented in Sect. <ref>. §.§.§ Generative model For variance-based UQ metrics, validation methods are based on a probabilistic model linking errors to uncertainties E_i∼ D(0,u_E_i) where D(μ,σ) is an unspecified probability density function with mean μ and standard deviation σ. This model states that errors should be unbiased (μ=0) and that uncertainty describes their dispersion as a standard deviation, following the metrological standard.<cit.> It is essential to note that the relation between E and u_E is asymmetric: small uncertainties should be associated with small errors, but small errors might be associated with small or large uncertainties, while large errors should be associated with large uncertainties. This considerably diminishes the interest of validation tests based on the ranking of E vs u_E (e.g. correlation coefficients). §.§.§ Validation model Average calibration. According to the generative model, the validation of u_E should be based on testing that it correctly describes the dispersion of E.<cit.> One can for instance check that Var(E)≃<u_E^2> where the average is taken over the validation dataset, and which is valid only if <E>≃0.<cit.> However, this formula ignores the pairing between errors and uncertainties, and a more stringent test is based on z-scores (Z=E/u_E)<cit.>, i.e. Var(Z)≃1 Note that if the elements of E and u_E are obtained as the means and standard deviations of small ensembles (say less than 30 elements) these formulas have to be transformed in order to account for the uncertainty on the statistics.<cit.> Unfortunately, hypotheses need then to be made on the distribution of these small ensembles. For a normal distribution, one should have Var(Z)≃(n-1)/(n-3), where n is the ensemble size.<cit.> The satisfaction of one or both of these tests (Eqns. <ref>-<ref>) validates average calibration, which is a minimal requirement, but does not guarantee the usefulness of individual uncertainties, as average calibration can be satisfied by a compensation of under- and over-estimated values of u_E. Conditional calibration: consistency and adaptivity. Based on Eq. <ref>, one can evaluate the reliability of individual uncertainties by conditional calibration<cit.>, i.e., Var(Z|A=a)≃1, ∀ a∈𝒜 where A is a variable with values a in 𝒜. As detailed below, A can be any relevant quantity, such as the uncertainty u_E or a feature X. The choice of the conditioning variable depends on the question to be answered. Choosing the uncertainty u_E will assess the calibration across the range of uncertainty values, i.e. the consistency between E and u_E, while choosing X will assess adaptivity. Using this terminology, one can see that the individual calibration proposed by Levi et al.<cit.> Var(E|u_E=σ)≃σ^2, ∀σ>0 deriving from Eq. <ref> validates consistency. Although it is a step forward from average calibration, it does not test the adaptivity of the UQ metric under scrutiny. Nor does the local Z-Variance (LZV) analysis<cit.> in u_E space, which can be derived from Eq. <ref> Var(Z|u_E=σ)≃1, ∀σ>0 Practical implementation of conditional calibration tests to variance- and interval- based UQ metrics requires to split the validation set into subsets, generally based on the binning of the conditioning variable. In these conditions, one is more testing local than individual calibration<cit.>, and Eq. <ref> leads to reliability diagrams<cit.>, and Eq. <ref> leads to the local Z-Variance (LZV) analysis in u_E space. Note that the bin size in these methods should be small enough to get as close as possible to individual calibration, but also large enough to ensure a reasonable power for statistical testing. The LZV analysis is easily applicable to any conditioning variable (see examples in Pernot<cit.>), which is not the case of the reliability diagram which would need to superimpose reliability curves for each subset and be difficult to analyze. An ideal variance-based UQ metric, i.e., one which provides reliable individual uncertainties, should satisfy calibration, consistency and adaptivity. Consistency and adaptivity are therefore two complementary aspects of tightness, a term introduced in a previous study<cit.> to characterize local calibration. Calibration is a necessary condition to reach consistency or adaptivity. In fact, consistency/adaptivity expressed as conditional calibration should imply average calibration, but the splitting of the data into subsets makes that the power of individual consistency/adaptivity tests is smaller than for the full validation set. It is therefore better to test average calibration separately, notably for small validation datasets. Note that for homoscedastic datasets (u_E=c^te) consistency is implied by calibration, but not adaptivity. As already mentioned, as X, V and u_E are not necessarily related by monotonous transformations, one should not expect these complementary conditional calibration tests to provide identical results. §.§ Validation model for interval-based UQ metrics Prediction intervals can be extracted from predictive distributions,<cit.> generated by quantile regression or conformal predictors,<cit.> or estimated from a variance-based UQ metric and an hypothetical generative distribution<cit.>. §.§.§ Validation datasets In order to validate interval-based UQ results, one needs a set of reference data R={R_i}_i=1^M, and possibly their uncertainties u_R={u_R_i}_i=1^M, and a series of prediction intervals with prescribed coverage P, I_V={I_V,P={I_V_i,P}_i=1^M}_P∈𝒫, where I_V_i,P=[I_V_i,P^-,I_V_i,P^+] and 𝒫 is a set of coverage values, typically expressed as percentages in ]0,100[. For convenience and consistency with the variance-based approach, one might transform these data as errors and error intervals C={V,I_E}={V_i,{I_E_i,P}_P∈𝒫}_i=1^M where I_E_i,P=[R_i-I_V_i,P^-, R_i-I_V_i,P^+] is the P % interval for the prediction error at point i.[Note that this definition differs from the one in my previous studies, which were based on the existence of a predicted value and an expanded uncertainty, leading to zero-centered intervals that should contain the error. In the present case the intervals are error-centered, but should contain 0.] As for variance-based UQ metrics, the uncertainty on the reference values, if any, should be accounted for, and the inclusion in the dataset of pertinent input values or features X={X_i}_i=1^M might be of interest to test adaptivity. §.§.§ Validation model Average coverage. Within this setup, a method is considered to be calibrated if prediction intervals have the correct empirical coverage. One defines the prediction interval coverage probability (PICP) as ν_p=ℙ(0∈ I_E,P) where ℙ is the probability function, and I_E,P is a P=100p % prediction error interval. Using PICPs, a method is calibrated in average, or marginally, if ν_p=p, ∀100p∈𝒫 which is equivalent to Kuleshov's definition<cit.> when all probability levels can be estimated (for instance if one has access to the full predictive distribution). For a limited set of probabilities, one gets a milder calibration constraint. Application of Eq. <ref> for a series of percentages provides a calibration curve, where the values of ν_p are plotted against the target probabilities p.<cit.> Conditional coverage. Average calibration based on PICPs can be met by validation sets with unsuitable properties<cit.>, and, as for variance-based UQ metrics, it is possible to build more stringent calibration tests based on conditional coverage,<cit.> i.e., ν_p=ℙ(0∈ I_E,P|A=a),∀ a∈𝒜 where A is a property with values a in 𝒜. To get the analog of consistency, one might consider to use U_P, the half-range of I_E_i,P, as a conditioning variable in order to check that errors are consistent at all uncertainty scales. For instance, Pernot used the expanded uncertainty U_95 to this aim.<cit.> For adaptivity testing, the conditioning variable can be any relevant feature X.<cit.> In practice, the PICPs are estimated as frequencies over the validation set ν_p,M=1/M∑_i=1^M1(0∈ I_E_i,P) where 1(x) is the indicator function for proposition x, taking values 1 when x is true and 0 when x is false. Estimating a PICP amounts to count the number of times prediction error intervals contain zero. Implementation of the conditional coverage tests requites binning according to the conditioning variable, with PICP testing within each bin, leading to the Local Coverage Probability (LCP) analysis.<cit.> Note that, as for reliability diagrams it is also possible to consider conditional calibration curves, with the same problem of readability and interpretability of overlapping curves. § VALIDATION METHODS The principles exposed in the previous section are now developed into practical methods grouped by validation target (average calibration, consistency and adaptivity). All methods are illustrated on synthetic datasets designed to reveal their potential limitations. §.§ Synthetic datasets To illustrate the methods presented in this study, six datasets of size M=5000 were designed to illustrate common validation problems. They are summarized in Table <ref>. Case A features a quadratic model with additive heteroscedastic noise R_i =a+X_i^2 u_E_i =b×|R_i| V_i =R_i+u_E_i× N(0,1) where N(0,1) is the standard normal distribution. The errors are thus in agreement with the generative model (Eq. <ref>), and, by construction, the set C={X,E,u_E} is calibrated, consistent and adaptive. It should pass all the corresponding tests. To generate a calibrated, non-consistent and non-adaptive dataset (Case B), one preserves all the data of Case A, except the uncertainties which are derived as perturbations of the mean uncertainty of Case A u_E_i=√(<u_E>_A)× N(1,0.1) This transformation preserves the calibration, but the errors are now inconsistent with u_E and adaptivity is also lost. Case C is issued from Case A with a random ordering of u_E. This preserves average calibration by Eq. <ref>, but breaks calibration by Eq. <ref>. Consistency and adaptivity are also broken. Case D is derived from Case A by a uniform scaling of u_E by a factor two. This set should fail all the tests for calibration, consistency and adaptivity. Note that there is not much sense to test consistency and adaptivity when average calibration is not satisfied. This will be considered as illustrative of the expected diagnostics. Case E is similar to Case A, but the errors have now a Student's-t_ν=4 distribution with four degrees of freedom V_i=R_i+u_E_i× t_4(0,1) where t_4(0,1) is the t distribution scaled to have a unit variance. This dataset is calibrated, consistent and adaptive. However, tests involving a normality hypothesis for the generative distribution should fail. Finally, Case F is an homoscedastic dataset deriving from Case A, where the uncertainties are replaced by a constant value (the root mean square of Case A uncertainties). It is calibrated, and not adaptive. For a homoscedastic dataset, consistency is identical to calibration. §.§ Testing average calibration The simplest way to assess average calibration for variance-based UQ metrics is to test the equality in Eq. <ref> or in Eq. <ref>. Various statistical procedures have been proposed, and it is best to avoid any assumption of normality for the distribution of errors or z-scores. See Pernot<cit.> for details. Application to the synthetic datasets is presented in Table <ref>. For Cases A, B, E and F, both tests are satisfied. Case D is obviously non calibrated. However, datasets with consistency problems (Case C) might still check one of both metrics. For interval-based UQ metrics, validation is done by testing Eq. <ref> for the probability levels that are available. The statistical procedure comes with the same caveats as for variance-based UQ metrics and has been detailed by Pernot<cit.>. When a large set of probability levels is available, one can build a calibration curve, as shown next. §.§.§ Calibration curve In a calibration curve<cit.>, one estimates the probability of errors E to fall within a series of intervals with coverage probability varying in [0,1], and one plots the resulting probability against the target one (Eq. <ref>). Calibration curves are applicable to homoscedastic datasets. The method is also referred to as confidence- or intervals-based calibration<cit.> and formalized as a variant of Eqns. <ref>-<ref> lim_M→∞η_p,M=p, ∀ p∈[0,1] where η_p,M=1/M∑_i=1^M1(E_i<q_p,i) and q_p,i is the quantile for probability p at point i. The ideal calibration curve is the identity line (Fig. <ref>, Cases A and F). For validation, a 95 % confidence interval is plotted<cit.>, either around the empirical curve or the identity line (the latter is preferred when multiple curves are drawn for the conditional case presented later). As the synthetic datasets provide only a variance-based UQ metric (u_E), the quantiles in Eq. <ref> are calculated by assuming a normal generative distribution, as usually done in the literature. In such cases, deviations from the identity line are not straightforward to interpret, as they might have their origin in the non-consistency of uncertainties (Cases B and C), their non-calibration (Case D) or a bad choice of reference distribution (Case E). It is interesting to contrast these results with the statistics for average calibration obtained above (Table <ref>). Case E has satisfying calibration statistics, but presents a slightly problematic calibration curve. This is due to the inadequate choice of a normal generative distribution for this dataset based on a Student's distribution. Also, Cases B and C present notably deviant calibration curves despite the fact that they have at least one correct calibration metric. The calibration curve obviously contains more information than the average calibration metrics, but it does not provide direct information on consistency or adaptivity. For instance, without additional information, it is difficult to differentiate the diagnostics for Cases B and D. For variance-based metrics, its use requires also an hypothesis on the generative distribution to estimate the theoretical quantiles, which complicates the interpretation of deviant curves. §.§ Testing consistency The validation literature for variance- or interval-based UQ metrics has mostly focused on consistency tests, and several methods are available, which are not fully equivalent. The aim of this section is to highlight the main features of these methods. §.§.§ “Error vs. uncertainty” plots Plotting the errors vs. the uncertainties can be very informative, notably for datasets missing consistency <cit.>. Additional plotting of guiding lines and running quantiles is a welcomed complement to facilitate the diagnostic.<cit.> This is obviously not applicable to homoscedastic datasets. Examples for the synthetic datasets are shown in Fig. <ref>. The expected “correct behavior” is to have the running quantile lines for a 95% confidence interval to lie in the vicinity of and parallel to the E=±2× u_E lines. Although the generative distribution is not necessarily normal, one does not expect very large deviations from k=2, as seen for instance for the Student's distribution of Case E. Moreover, the cloud of points should be symmetrical around E=0 (absence of bias). The plots enable to sort out Cases B, C and D without ambiguity. For Cases B and C, the running quantiles are more or less horizontal, indicating an absence of link between the dispersion of E and u_E. For Case D, the running quantiles follow closely the k=1 lines, pointing to a probable overestimation of uncertainties. For cases where consistency cannot be frankly rejected on the basis of the shape or scale of the data cloud, it is imperative to perform more quantitative tests as presented below. One should not conclude on good consistency simply based on this kind of plot. §.§.§ Conditional calibration curves in uncertainty space It is formally possible to construct conditional calibration curves, but, to my knowledge, they have not been proposed in the ML-UQ literature. Examples are given in Fig. <ref> using u_E as the conditioning variable with 15 equal-counts bins. In the ideal case (Case A), all the conditional curves lie within the confidence interval estimated for a sample size corresponding to a single bin. Case B shows that the anomaly observed on the average calibration curve is identical at all uncertainty levels, pointing to a consistency problem or a wrong choice of generative distribution. Case C is interesting, as it shows that the small oscillations of the average calibration curve around the identity line are attenuated compared to the much larger deviations observed for conditional curves. This points to the fact that a seemingly good calibration curve might hide a non-consistent case. The large dispersion of conditional curves is less likely to be due to a poor choice of generative distribution (unless the generative distribution varies strongly with u_E) and points to a lack of consistency. Cases D and E offer the same diagnostic as for case B, with an ambiguity between non-consistency and generative distribution problem. Using a Student's-t_ν=4 generative distribution for Case E solves the problem. §.§.§ Reliability diagram So-called reliability diagrams<cit.> test consistency through Eq. <ref> over data subsets, where the data are ordered by increasing u_E values and split into bins. This method is therefore not usable for homoscedastic datasets. Reliability diagrams can also be found in the literature as error-based calibration plots<cit.> (not to be confounded with calibration curves, Sect. <ref>), RvE plots<cit.>, or RMSE vs. RMV curves<cit.>. For convenience, the square roots of the terms of Eq. <ref> are used, linking the mean uncertainty in each bin (as the root mean squared (RMS) of u_E, also called root mean variance (RMV)) with the RMSE or RMSD of E (both are equivalent for unbiased errors). For consistent datasets, the reliability curve should lie close to the identity line, up to statistical fluctuations due to finite bin counts. For a conclusive analysis of deviations from the identity line, the amplitude of these finite size effects should be estimated, for instance by bootstrapping.<cit.> The binning strategy is important<cit.>: some authors advocate for bins with identical counts,<cit.> other for bins with identical widths<cit.>. Both choices are defensible according to the distribution of uncertainties, notably the absence or presence of heavy tails. It should also be noted that the insensitivity of Eq. <ref> to the pairing between E and u_E might still be a hindrance for large bins, but its effect should decrease when the bins get smaller. However, small bins are affected by large statistical fluctuations. The binning strategy should thus be designed to offer a good compromise. An adaptive binning scheme mixing both strategies is proposed in Appendix <ref>. For the synthetic datasets, I use a default choice of 15 equal-counts bins, leading to about 333 points per bin (Fig. <ref>). The non-consistent cases (B, C, D) are correctly identified, although for case B the deviation from the identity line is not noticeable except for the two extreme bins. §.§.§ Local Z-Variance analysis in uncertainty space LZV analysis was introduced by Pernot<cit.> as a method to test local calibration based on Eq. <ref>. As for reliability diagrams, it is based on a binning of the data according to increasing uncertainties. For each bin, one estimates Var(Z) and compares it to 1. Here again, error bars on the statistic should be provided to account for finite bin counts.<cit.> Assuming a nearly constant uncertainty value within a bin, values of Var(Z) smaller/larger than 1 indicate over-/under-estimated uncertainties (by a factor √(Var(Z))). This suggests an alternative plot of the inverse of the standard deviation of Z, Var(Z)^-1/2, instead of Var(Z), which provides a more direct reading and quantification of the local uncertainty deviation. This Local Z-scores Inverse Standard Deviation (LZISD) analysis is used throughout this study. Application of the LZISD analysis to the synthetic datasets confirms the diagnostics of non-tightness for cases B, C and D (Fig. <ref>). For Case B, the deviation from the reference is more legible than on the corresponding reliability diagram [Fig. <ref>(b)], and the largest local deviations can be directly quantified to about 30 % in default and 60 % in excess (these extreme values are probably underestimated, as they result from a bin averaging). §.§.§ Local Coverage Probability analysis in uncertainty space Conditional coverage with respect to an uncertainty metric is available through the Local Coverage Probability analysis.<cit.> It can be directly built from prediction intervals if available, or estimated from prediction uncertainties and an hypothetical generative distribution. In both cases, the empirical coverages are drawn for a series of data subsets based on the binning of the conditioning variable and for the available target probabilities. The LCP analysis of consistency is not applicable to homoscedastic datasets. For the synthetic datasets (Fig. <ref>), one makes the hypothesis of a normal generative distribution to build prediction intervals from the uncertainties for 𝒫={25,50,75,95}. Case A presents perfect conditional coverages over the full uncertainty range and probability levels. For Case B, the empirical coverages are too large, except at the P=95 % level. The dependency along u_E is weak, except for the two extreme intervals, as was observed on the LZISD analysis. A contrario, Case C presents a strong coverage variation across u_E, with inadequate mean PICP values. For Case D, the overestimated uncertainties produce intervals with excess coverage, uniformly across the uncertainty range. Finally, Case E suffers again from the misidentification of the generative distribution, which is solved by using the correct distribution to generate the intervals [Fig. <ref>(f)]. When a PICP value reaches 1.0, one gets no information about the mismatch amplitude with the target probability. The Local Ranges Ratio analysis (LRR) has been proposed by Pernot<cit.> to solve this problem. It estimates the ratio between the width of the empirical interval and the width of the theoretical interval. This tool is not exploited in the present study, but its availability should be kept in mind, §.§.§ Confidence curve A confidence curve is established by estimating an error statistic S on subsets of C iteratively pruned from the points with the largest uncertainties.<cit.> Technically, it is a ranking-based method, as (1) it is insensitive to the scale of the uncertainties, and (2) the relative ranking of the errors and uncertainties plays a determinant role. It is not applicable to homoscedastic datasets. If one defines u_k as the largest uncertainty left after removing the k % largest uncertainties from u_E (k∈{0,1,…,99}), a confidence statistic is defined by c_S(k;E,u_E)=S(E | u_E<u_k) where S is an error statistic – typically the Mean Absolute Error (MAE) or Root Mean Squared Error (RMSE) – and S(E | u_E<u_k) denotes that only those errors E_i paired with uncertainties u_E_i smaller than u_k are selected to compute S. A confidence curve is obtained by plotting c_S against k. Both normalized and non-normalized confidence curves are used in the literature, but the RMSE-based non-normalized version should be preferred for the validation of variance-based UQ metrics.<cit.> A monotonously decreasing confidence curve reveals a desirable association between the largest errors and the largest uncertainties, an essential feature for active learning. It might also enable to detect unreliable predictions. But, in order to test consistency, one needs to compare c_S to a reference. *Reference curves. As already mentioned, only the order of u_E values is used to build a confidence curve, and any change of scale of u_E leaves c_S unchanged. Without a proper reference, c_S cannot inform us on calibration or consistency. An oracle curve can be generated by reordering u_E to match the order of absolute errors. This can be expressed as O(k;E)=c_s(k;E,|E|) It is evident from the above equation that the oracle is independent of u_E and therefore useless for calibration testing. Recast in the probabilistic framework introduced above, the oracle would correspond to a very implausible error distribution D, such that E_i=± u_E_i. Although it is offered as the default reference curve in some validation libraries, I strongly recommend against its use for variance- or interval-based UQ metrics. Using Eq. <ref>, a probabilistic reference curve P can be generated by sampling pseudo-errors E_i for each uncertainty u_E_i and calculating a confidence curve for {Ẽ,u_E}, i.e., P(k;u_E)=⟨ c_S(k;Ẽ,u_E)⟩ _Ẽ where a Monte Carlo average is taken over samples of E_i∼ D(0,u_E_i) The sampling is repeated to have converged mean and confidence band at the 95 % level, typically 500 times. In contrast to the oracle, which depends exclusively on the errors, the probabilistic reference depends on u_E and a choice of distribution D, but it does not depend on the actual errors E. Comparison of the data confidence curve to P enables to test if E and u_E are correctly linked by the probabilistic model, Eq. <ref>, i.e. to test consistency. For RMSE-based confidence curves, P does not depend on the choice of generative distribution (only the width of the confidence band does). This is not the case for MAE-based confidence curves, which makes them more difficult to interpret.<cit.> Given the choice of a generative distribution D, interval-based metrics can be transformed to variance-based metrics and used to build a confidence curve and probabilistic reference. One can check in Fig. <ref> that consistency problems are well identified by the comparison of a confidence curve to the associated probabilistic reference (Cases B-D). When small deviations are observed, it is however difficult to discriminate between a genuine consistency problem and a bad choice of generative distribution for P (Case E). The choice of the appropriate distribution for Case E leaves the confidence curve unchanged but widens the confidence band of the probabilistic reference, ensuring its validation. §.§ Testing adaptivity Adaptivity is not commonly tested in the ML-UQ literature, except maybe through a binary approach discerning in-distribution and out-of-distribution predictions.<cit.> §.§.§ Z vs Input feature plot The analog of the “E vs u_E” plot for consistency is to plot the z-scores Z as a function of a relevant input feature X to check adaptivity. The distribution of Z should be homogeneous along X and symmetric around Z=0. Here again, one can use guide lines Z=±2 and compare with running quantiles for a 95% confidence interval of Z values. Conditioning on X is shown in Fig.<ref> for the synthetic datasets. Considering the shape of the data clouds, adaptivity can be clearly rejected in cases B and C, while the dispersion of the z-scores is insufficient in Case D. The other cases require a more quantitative validation. §.§.§ Conditional calibration curves in input feature space The same principle as used for consistency testing (Sect. <ref>) can be applied to adaptivity by grouping the data according to the binning of a relevant input feature X. The method can now be applied to homoscedastic datasets. Application to the synthetic datasets is presented in Fig. <ref>. For Cases A and F, one would conclude to a good adaptivity along X . For the other cases, the deviations from the identity line are similar to those observed when testing consistency, except for Case B, where the conditional calibration curves are much more dispersed. §.§.§ Local Z-Variance analysis in input feature space As the LZV/LZISD analysis in u_E space was used to validate consistency, one can perform a LZV of LZISD analysis in X space to validate adaptivity. Here again, the LZISD analysis offers a more direct quantification of deviations. Both methods can be used for homoscedastic uncertainties. The diagnostics provided by the LZISD analysis on the synthetic datasets are non-ambiguous (Fig. <ref>): Cases B, C and D stand out as having strong deviations from the reference line, while Cases A, E and F present a good adaptivity, as observed on the “Z vs X” plots. §.§.§ Local Coverage Probability analysis in input feature space Using the same binning for X as used in the other methods, the LCP analysis (Fig. <ref>) provides results conform with those of the conditional calibration curves (Fig. <ref>). Here also, the hypothesis of a normal generative distribution used to build the probability intervals penalizes Case E, which can be solved by an appropriate choice of distribution (not shown). §.§.§ Using latent distances or V as conditioning variables Although using one or several input features X as conditioning variable is the most direct way to test for adaptivity, it might not always be practical, for instance when input features are strings, graphs or images. In such cases, one might use instead a latent variable or the predicted property value V. X and V are interchangeable only if the latter is a monotonous transformation of the former, but they might provide different adaptivity diagnostics if this is not the case. Using V answers to the question: Are uncertainties reliable over the full range of predictions ? The adaptivity analysis of the synthetic datasets using V as a conditioning variable is presented in Appendix <ref>. It shows that the non-monotonous shape of the V=f(X) model might induce artifacts in the adaptivity analysis that can be difficult to interpret if the functional form of the model is unknown or complex. §.§ Validation metrics The present study focuses on graphical validation tools, but validation metrics<cit.> are widely used in the ML-UQ literature. For instance, metrics have been designed for calibration curves<cit.>, reliability diagrams<cit.>, and confidence curves<cit.>. These metrics are generally used to rank UQ methods, but they do not provide a validation setup accounting for the statistical fluctuations due to finite-sized datasets or bins. This is essential for small validation datasets and is not without interest for the conditional analysis of large ML datasets. Pernot<cit.> introduced confidence bands for calibration curves that enables to identify significant deviations from the identity line, and also error bars for reliability diagrams, LZV and LCP analyses. This provides a basis to define upper limits for calibration metrics that can be used for validation purpose. In the present state of affairs, calibration metrics such as the expected normalized calibration error (ENCE)<cit.> or the area under the confidence-oracle (AUCO)<cit.> cannot be used to validate consistency nor adaptivity and deserve further consideration beyond the purpose of this article. § APPLICATIONS The validation methods introduced in the previous sections are now applied to “real life” data coming from the materials science and physico-chemical ML-UQ literature. The choice of datasets enables to explore various aspects of a posteriori validation. It is not my intent in these reanalyses to criticize the analyzes nor the UQ methods presented in the original studies, but simply to show how the augmented set of validation methods proposed above might facilitate and complete the calibration diagnostics. §.§ Case PAL2022 The data have been gathered from the supplementary information of a recent article by Palmer et al.<cit.>. I retained 8 sets of errors and uncertainties before and after calibration by a bootstrap method, resulting from the combination of two materials datasets (Diffusion, M=2040 and Perovskites, M=3836) and several ML methods (RF, LR, GPR....). The datasets are tagged by the combination of both elements: for instance, Diffusion_RF is the dataset resulting from the use of the Random Forest method on the Diffusion dataset. The reader is referred to the original article for more details on the methods and datasets. As only the errors E and uncertainties u_E are available, it is not possible to test adaptivity for this dataset. My aim here is mainly to compare the performance of reliability diagrams and LZISD analysis against the binning strategy, and to compare their results to confidence curves with the probabilistic reference. §.§.§ Average calibration As a first step, I estimated for each dataset the mean of z-scores and their variance to appreciate the average impact of the calibration method. Fig. <ref> summarizes the results. It can be seen that calibration leaves the mean z-score unchanged or improves it (the calibrated value is closer to zero). After calibration, the mean z-scores are not always null, but small enough to consider the error sets to be unbiased. The calibration effect is more remarkable for Var(Z): all the 95 % confidence intervals for Var(Z) after calibration overlap the target value (1.0), except for Perovskite_LR with a very small residual gap. §.§.§ Consistency Consistency can be checked by reliability diagrams, LZV analysis and confidence curves with probabilistic reference. In absence of information on the generative distribution(s), it is best to avoid conditional calibration curves in this example. Before analyzing the datasets, it is important to point out the impact of the binning strategy on the conclusions drawn from reliability diagrams and LZV analysis. In their study, Palmer et al. define 15 intervals forming a regular grid in u_E space, with some adaptations described in Appendix <ref>. This is different from the usual strategy consisting in designing bins with identical counts.<cit.> Besides, considering the size of the datasets, 15 bins might not be enough to reveal local consistency anomalies. The impact of the binning strategy is explored in Appendix <ref>, where an adaptive binning strategy is designed to combine both approaches. When using the log-scale for uncertainty binning, this adaptive strategy is efficient to reveal local consistency anomalies that do not always appear in the original reliability diagrams. Calibrated Random Forest. The LZISD analysis and reliability diagrams are shown in Fig. <ref> for both datasets. The effect of calibration is well visible on those graphs when comparing the red dots (uncalibrated) to the blue ones (calibrated). For both datasets, consistency after calibration is rather good, except for the small uncertainty values (below 0.25 eV for Diffusion and 0.15 eV for Perovskite). The legibility is better on the LZISD analysis, which provides directly an overestimation value of 50 % of the uncertainties around 0.2 eV for Diffusion and up to 100 % around 0.08 eV for Perovskite. The confidence curves in Fig. <ref> confirm this diagnostic, as they lie close to the reference for large uncertainties and start deviate from it at smaller uncertainties. The non-linear scale of threshold values u_k at the top of the plot helps to locate the anomalies in uncertainty space. Calibrated Linear Ridge regression. Calibration and consistency of linear ridge regression results are assessed in Fig. <ref>. The effect of calibration is noticeable over the whole uncertainty range, except at large uncertainties for the Diffusion dataset, where the situation is worsened by calibration. In this area, the uncertainties are overestimated by a factor about 2. At smaller uncertainties, the consistency is far from perfect, with areas of underestimated uncertainties around 0.3 eV and 0.5 eV, compensating for the overestimated values at large uncertainty. The confidence curve displays the problem at large uncertainty values, but is less legible at smaller uncertainties. For the Perovskite dataset, consistency is not reached either, with a compensation between overestimated and underestimated uncertainty areas. The confidence curve is notably deviating from the reference curve. Calibrated Gaussian Process Regression. One considers here the calibration results of a Gaussian Process regression with its Bayesian UQ estimate. It is possible to go quickly to the conclusion that consistency is not achieved. For the Diffusion dataset, all three graphs (Fig. <ref>) concur to reject consistency. The confidence curve even suggests that it would not be reasonable to base an active learning strategy on these uncertainties. In the case of Perovskite one observes a nugget of data with tiny errors and largely overestimated uncertainties, that was identified by the adaptive binning strategy (Appendix <ref>). A zoom over the remaining data (not shown) reveals an heterogeneous situations with small areas of underestimation (below 0.2 eV) and overestimation (around 0.3 eV and 0.6 eV). The confidence curve deviates considerably from the reference and falls unexpectedly to zero at small uncertainties, a consequence of the aforementioned data nugget. Conclusion. The reanalysis of these data shows that the reliability diagrams presented in the original article with a small number of bins and no error bars are often overoptimistic. Using an adaptive binning strategy in log-uncertainty, conditional calibration appears heterogeneous for all datasets, meaning that consistency is not achieved despite a rather good average calibration. The uncertainties are at worst within a factor two of their ideal value, a level that has to be contrasted with their intended application. The confidence curves show that not all ML methods provide reliable uncertainties for active learning. §.§ Case BUS2022 The data presented in Busk et al.<cit.> have been kindly collated by Jonas Busk for the present study. In their article, Busk et al. extend a message passing neural network in order to predict properties of molecules and materials with a calibrated probabilistic predictive distribution. An a posteriori isotonic regression on data unseen during the training is used to ensure calibration. For UQ validation, these authors used the more or less standard trio of reliability diagrams, calibration curves (named quantile-calibration plot) and MAE-based confidence curves with the oracle reference. Note that they duly express a reserve about using the oracle: “However, we do not expect a perfect ranking...”<cit.>. Considering the validation results published for the QM9 dataset in Fig. 2 of the original article, a few questions arise: (1) what is causing the imperfect calibration curve ?; (2) if this is a wrong distribution hypothesis, how does it affect the MAE-based confidence curve ?; (3) how is the confidence curve analysis modified if one uses the RMSE statistic and the probabilistic reference ?; and, (4) what is the diagnostic for adaptivity? The present reanalysis aims to answers these questions. §.§.§ Reanalysis of the QM9 dataset The QM9 dataset consists of 13 885 predicted (V,u_V), reference (R) atomization energies, and molecular formulas. These data are transformed to C={ X,E,u_E} according to Sect. <ref>, where X is the molecular mass generated from the formulas. One should thus be able to test calibration, consistency and adaptivity. Average calibration is readily assessed (Var(Z)=0.96(2)). It is always instructive to inspect the raw data through “E vs u_E” and “Z vs X” plots to get a global appreciation of the link between these quantities (Fig. <ref>). One sees in Fig. <ref>(a) that the data points are neatly distributed between the E=±3u_E guiding lines, and the 0.025 and 0.975 running quantile lines seem to follow closely the E=±2u_E lines up to uE≃0.05. Above this value, the uncertainties seem overestimated and the errors are somehow biased toward the positive values. However, this concerns a very small population (about 95 points) and filtering them out does not affect significantly the calibration statistics presented below. The problem might be simply due to the sparsity of the data in this uncertainty range. The “Z vs X” plot in Fig. <ref>(b) using molecular masses enables to check if calibration is homogeneous in mass space. The shape of the running quantile lines indicate that uncertainties are probably overestimated for masses smaller than the main cluster around 125-130 Da, evolving to a slight underestimation above this peak. Depending on its amplitude, this systematic effect might be problematic, and it hints at a lack of adaptivity. A quantitative analysis of this feature is presented below. In a first step, calibration analyses are done to reproduce the Fig. 2 of the original article: reliability diagram with 10 equal-counts bins [Fig. <ref>(a)], calibration curve [Fig. <ref>(b)] and MAE-based confidence curve (replacing the oracle by the probabilistic reference) [Fig. <ref>(c)]. Then, further analyses are performed to complement the information provided by the first set: LZISD analysis in uncertainty and V space [Fig. <ref>(d,e)], and RMSE-based confidence curve [Fig. <ref>(f)]. The reliability diagram [Fig. <ref>(a)] does not enable to see any major problem and would lead us to conclude that the data are consistent. However, the calibration curve [Fig. <ref>(b)] is not perfect. Considering the good consistency provided by the reliability diagram, one should conclude that there is probably a distribution problem, i.e., the generative distribution of errors that is used to build the probabilistic reference curve should not be normal. Moreover, the confidence curve [Fig. <ref>(c)] is not very close to the probabilistic reference, which, assuming a good consistency, points also to a problem of distribution, as the MAE-based probabilistic reference is sensitive to the choice of distribution. This distribution problem is analyzed more specifically below (Sect. <ref>). A LZISD analysis with an adaptive binning strategy is performed to compare with the original reliability diagram [Fig. <ref>(d)]. Indeed, one observes some discrepancies: overestimation and underestimation of u_E go locally up to 15 %. There is notably an overestimation trend around 0.01 eV. This feature is also present in the reliability diagram, but it is more difficult to visualize small deviations on a parity plot. The LZISD analysis against the molecular mass [Fig. <ref>(e)] confirms and quantifies what has been observed in [Fig. <ref>(b)]: an overestimation between 40 and 80 % at masses below 120 Da, and a slight underestimation by about 20 % above 130 Da (although there are too few data in this range to conclude decisively). Uncertainties are therefore mostly reliable for the main mass peak between 120 and 130 Da, but not outside of this range. This heterogeneity of calibration and lack of adaptivity is not observable in the other validation plots. Finally, a RMSE-based confidence curve is reported in Fig. <ref>(f). It has been truncated at 0.015 eV, but drops sharply from 0.031 eV because of the small set of large positive errors corresponding to the largest uncertainties [Fig. <ref>(a)]. This confidence curve shows a much better agreement with the probabilistic reference than the MAE-based one, albeit there is still a mismatch due to the consistency defects observed in the LZISD analysis and/or to the distribution problem mentioned above. §.§.§ Non-normality of the generative probabilistic model The deviations of the calibration curve observed in Fig. <ref>(b) correspond neatly to observations made for one of the synthetic datasets (Case E) when drawing error from a Student's-t distribution with four degrees of freedom and comparing the percentiles with a normal reference [Fig. <ref>(e)]. Indeed, if one substitutes the normal reference in the calibration plot by a t_ν=4 reference, one obtains a perfect calibration curve [Fig. <ref>(a)]. Similarly, assuming a t_ν=4 generative distribution for the probabilistic reference of confidence curves improves considerably the agreement with the data curve [Fig. <ref>(b,c)], both for the MAE- and RMSE-based approaches. Note the widening of the confidence area of the probabilistic reference in both cases. There remains small deviations indicating that consistency is not perfect. §.§.§ Conclusions All diagnostics based on E and u_E conclude to a good calibration and consistency, notably if one accounts for a non-normal generative model (t_ν=4). The main feature revealed by this reanalysis is the lack of adaptivity seen in the LZISD analysis in molecular mass space. It leads to a significant underestimation of the quality of predictions for the smaller molecules in the QM9 dataset. By contrast, the confidence curves indicate that the uncertainties can be used for active learning without a second thought. §.§ Case HU2022 These data were kindly collated and provided by Y. Hu to facilitate this reanalysis of the results presented in Figs. 3e and 3f of the paper by Hu et al.<cit.>. These figures present plots of the errors on interatomic potentials for the QM9 dataset vs two distances: a distance in feature space, X_F and a distance in latent space<cit.>, X_L. Both distances have been calibrated to interval-based UQ metrics: the half-ranges of prediction intervals, U_F,P and U_L,P have been obtained for P∈{68,95}, by multiplication of the corresponding distances by probability-specific scaling factors optimized by conformal calibration to ensure average coverage. The percentage of points within the resulting intervals are reported in the original figures and show a successful average calibration in both cases. However, the distribution of points in these figures hints at different conditional calibration statistics for both metrics, which I propose to quantify in this reanalysis. The reader is referred to the original article for details on the dataset and ML-UQ methods. A plot of the errors vs the expanded uncertainties at the 95 % level [Fig. <ref>(a,d)] confirms the features observed in the original figures: in this plot, the running quantiles should follow the y=± x guide lines, which is not the case for U_F,95, while this seems to be much better for U_L,95, at least below 0.2 eV/sys. To quantify this difference, the data are then analyzed with the LCP analysis in feature distance space [Fig. <ref>(b,e)]. As the data are conditioned over the feature distance, one is testing the adaptivity of the conformal calibration for individual predictions. Note that for U_F,95 the distances and uncertainties are proportional, and the adaptivity diagnostic is also a consistency diagnostic! A satisfying average coverage is reached in all cases, in agreement with the conclusions of the original article. However, the local coverage of U_L,P is more homogeneous than for U_F,P at both probability levels. The latter presents strong under- and over-coverage areas. The adaptivity of U_L,P is not far from perfect at the 95 % level, except for the extreme bins, but is not good at the 68 % level. It is also worthwhile to look at confidence curves[Fig. <ref>(c,f)]. The original article presents calibration curves which are very close to the identity line for both metrics. One can thus confidently transform the expanded uncertainties at the 95 % level to standard uncertainties using a normal distribution hypothesis. The confidence curves display a much better behavior for U_L than for U_F. For active learning, uncertainty built on latent distance looks therefore much more reliable than built on feature distance. However, consistency is not perfect for U_L. A simple scaling of the distance metrics is thus unable to ensure consistency/adaptivity. A better calibration would require to use conditional conformal calibration, which is reputed for its difficulty,<cit.> and, as far as I know, has not yet been implemented in the ML-UQ context considered here. § CONCLUSIONS This article introduces a principled framework for the validation of UQ metrics. The studied examples target mostly machine learning in a materials science and chemistry context, but the applicability is general. The concept of conditional calibration enables to define two aspects of local calibration: consistency, which assesses the reliability of UQ metrics across the range of uncertainty values, and adaptivity, which assesses the reliability of UQ metrics across the range of relevant input features. The more or less standard UQ validation methods (calibration curves, reliability diagrams and confidence curves) were recast in the proposed framework, showing that they are not designed to test adaptivity. In consequence, adaptivity is presently a blind spot in the ML-UQ validation practice. It was shown on a few examples that consistency and adaptivity are scarcely reached by modern ML-UQ methods, and that a positive consistency diagnostic does not augur of a positive adaptivity diagnostic. Both aspects of local calibration should be tested. Let us summarize the main points arising from this study. Consistency testing. Beside the exploratory “E vs u_E” plots, three validation methods have some pertinence to test consistency of variance-based UQ metrics: * Error vs Uncertainty plots offer a very informative preliminary analysis, devoid of any artifacts due to a binning strategy or the choice of a generative distribution. However, they do not enable to conclude positively on consistency, which requires more quantitative diagnostics. * Reliability diagrams and LZV/LZISD analysis are direct implementation of the conditional calibration tests. It was shown that the LZV/LZISD analysis offers a visually more discriminant approach, along with a direct quantification of possible uncertainty misestimations. Both methods are sensitive to the choice of a binning scheme, and an adaptive binning scheme is proposed to compensate for the drawbacks of equal-counts or equal-width binning strategies. * The RMSE-based confidence curve is not designed to check consistency, but, augmented with a probabilistic reference, it offers an interesting approach, combining two diagnostics: the eligibility of the UQ method for a reliable active learning and the consistency of the dataset. Moreover, it is less dependent on a binning strategy than the reliability diagrams or LZV analysis. An inconvenience is the dependence of the statistical validation reference curve (more precisely its confidence band) on the choice of a generative distribution. I have also shown that conditional calibration curves could be considered to assess consistency. As for confidence curves, their disadvantage for variance-based UQ metrics is the difficulty to discern between a lack of consistency or a bad choice of generative distribution hypothesis. I would then recommend the Error vs Uncertainty plot, the LZISD analysis and the RMSE-based confidence curve with its probabilistic reference as complementary diagnostics for the consistency of variance-based UQ metrics. For interval-based UQ metrics the LCP analysis seems to be the best choice, as it implements directly the conditional calibration test. Calibration curves have been shown to have possibly ambiguous diagnostics, which can be improved by using conditional calibration curves. In order to avoid undue hypotheses about error generative distribution, interval-based validation methods should be reserved to interval-based UQ metrics. Adaptivity testing. Adaptation of the standard methods to test conditional calibration over some input feature is not ideal but feasible, and one could envision conditional reliability diagrams, conditional calibration curves, and even conditional confidence curves. This approach requires large validation datasets and leads to complex graphs that are not easily legible. In addition to the Z vs Input feature plots, I would then recommend the LZISD analysis and the LCP analysis as general tools to test adaptivity of variance- and interval-based UQ metrics, respectively. It was also noted that using the predicted values V as a conditioning variable to test adaptivity is not without risks, as it might produce spurious features that complicate the diagnostic. This should probably be generalized to any property that is susceptible to be strongly correlated with the errors. Further studies are required to clear this up. § ACKNOWLEDGMENTS I warmly thank J. Busk for providing me the data for the BUS2022 case, and Y. Hu and A. J. Medford for the data in the HU2022 case. § AUTHOR DECLARATIONS §.§ Conflict of Interest The authors have no conflicts to disclose. § CODE AND DATA AVAILABILITY The code and data to reproduce the results of this article are available at <https://github.com/ppernot/2023_Primer/releases/tag/v1.0> and at Zenodo (<https://doi.org/10.5281/zenodo.6793828>). The ,<cit.> https://github.com/ppernot/ErrViewLibErrViewLib package implements the validation functions used in the present study, under version (<https://github.com/ppernot/ErrViewLib/releases/tag/v1.7.0>), also available at Zenodo (<https://doi.org/10.5281/zenodo.7729100>). The graphical interface to explore the main UQ validation methods provided by is also freely available (<https://github.com/ppernot/UncVal>). unsrturlPP § APPENDICES § USING THE PREDICTED VALUE TO TEST ADAPTIVITY A posteriori validation of adaptivity on datasets that do not contain input features X might still be attempted by using the predicted values V as proxy. Conditional calibration on V is not necessarily identical to conditional calibration on X, but it might still provide useful diagnostics. Application of “Z vs V” plots to the synthetic datasets is presented in Fig.<ref>(a-f). Z-score values in Case A are globally well distributed (horizontal running quantiles around ±2, but present an anomaly at small V values. This can be traced back to the correlation of V with E (E=R-V), most visible when R is nearly constant at the bottom of the parabolic model used to generate the data. This warns us that conditioning on V might produce some aberrant features, which are not diagnostic of poor adaptivity. The same artifact can be observed in Cases D, E and F. For Case E, the excessive dispersion of points, with |Z| values above 5, points to a non-normal distribution, but this should not be a valid reason to reject adaptivity: a more quantitative analysis is mandatory. By contrast, the absence of adaptivity is readily visible for cases B, C, and D : cases B and C present heterogeneous distributions along V, while Case D has Z values mostly contained between -1 and 1, pointing to an homogeneous overestimation of uncertainties. Case F is similar to Case A. The translation of these observations to the LZISD analysis is presented in Fig. <ref>(g-l). The artifact appears as local deviations of the LZISD values, while the intrinsic non-adaptivity of Cases B and C is more global. Similar observations can be made on the conditional calibration curves and LCP analyses (not shown). This shows that using V as a substitute for X to test adaptivity might lead to ambiguous conclusions when local anomalies are observed. However, global anomalies are certainly diagnostic of a lack of adaptivity of the tested UQ metric. § IMPACT OF THE BINNING STRATEGY To implement the LZV analysis, the default binning strategy is based on intervals with (nearly) equal populations.<cit.> This ensures equal testing power in all intervals, but might result in intervals with very different ranges. This is also often the strategy chosen for reliability diagrams.<cit.> However, some authors prefer to use bins with similar widths,<cit.> a strategy which faces two problems: bins with insufficient population to derive reliable confidence intervals, and bins containing a large part of the total sample for uncertainty distributions with large tails. These problems were addressed by Palmer et al.<cit.> by labeling the bins with less than 30 points as unreliable and by altering the width of bins in the peak area to ensure that the lowest 90% of the values were spread across at least five bins (assuming that uncertainty distributions can only have a heavy upper tail). The use of this strategy leads to very regular reliability diagrams with deviant points mostly in the upper range of uncertainty. I present here an alternative binning strategy, considering that for a positive variable such as uncertainty, a log-transform might be an efficient way to reduce the skewness of the distribution causing the accumulation problem in a few bins. Starting from a regular grid in log space (n bins), the problem of small counts is solved by merging adjacent bins with insufficient populations, while the problem of excessively large counts is solved by splitting bins having more than M/n points (M is the dataset size). Both operations are iterated until the bins population conforms with the chosen limits (the lower limit is set to 30). Two examples below, taken from Palmer et al.<cit.>, show that this adaptive strategy is more efficient to reveal tightness problems. The Diffusion/LR and Perovskite/GPR_Bayesian datasets were analyzed by reliability diagrams and LZISD analysis for increasing number of bins (10, 20 and 40) for the equal-counts bins and adaptive strategies. The reliability diagrams have also to be compared with those of the original article, based on 15 equal-width bins. Let us consider first the Diffusion/LR case. For both types of analysis, the counts-based binning [Fig. <ref>(a-f)] needs at least 20 bins to reveal an overestimation problem for the largest uncertainties, with a stronger effect for 40 bins. The averaging effect in large bins (n=10) would lend us to believe that calibration is good in this area, whereas using four times smaller bins reveals uncertainty in excess by a factor two around 2 eV. By using the new adaptive strategy [Fig. <ref>(g-l)], the problem is apparent for all the specified number of bins. In fact, for a starting point of 10 bins, the merge/split strategy converges to 20 bins, for 20 bins, one gets 30 and for 40 bins 47. The first case provides a good binning without the useless details that arise from the higher bin numbers. For this dataset, one clearly has tightness problems with under- and over-estimations for uncertainties larger than 0.5. These defects compensate and are not detected when using large bins. This problem is also apparent in the original article [Figure 2(d)], albeit for a series of bins with populations below the 30 limit. Aggregating these bins enables to conclude that the effect is statistically significant. The Perovskite/GPR_Bayesian case is shown in Fig. <ref>. As in the previous example, the counts-based binning needs more than 20 bins to display a problem at small uncertainties, in both LZV and RD analyses [Fig. <ref>(a-f)]. The LZV analysis indicates a strong overestimation of the uncertainties around 0.01 eV, the reliability diagram showing that the corresponding errors have a standard deviation of about 1E-8, which would certainly deserve a closer inspection. The adaptive analysis needs 20 initial bins (30 after adaptation) to reveal the problem [Fig. <ref>(g-l)]. This “data nugget” is undetected by the linear equal-width binning of the original study (see Figure 38 of the https://static-content.springer.com/esm/art%3A10.1038%2Fs41524-022-00794-8/MediaObjects/41524_2022_794_MOESM1_ESM.pdfSupplementary Information to Palmer et al.<cit.>). The adaptive binning strategy in logarithmic uncertainty seems to be able to detect consistency problems more efficiently than binning in linear uncertainty, either with bins of same counts or same width. Of course, it does not apply to variables with negative or null values, where it can be replaced by an adaptive strategy in linear space. The adaptive binning strategy is used for the LZISD, LCP and reliability diagrams in the case studies of the present article with a starting point of n=20 bins.

http://arxiv.org/abs/2303.07235v1

20230313160741

Distance Evaluation to the Set of Defective Matrices

cs.SC

theoremTheorem corCorollary exampleExample Distance Evaluation to the Set of Defective Matrices Alexei Yu. Uteshev[The corresponding author], Elizaveta A. Kalinina, Marina V. Goncharova 1010 St. Petersburg State University 1010Faculty of Applied Mathematics 1010St. Petersburg, Russia 99^1 {alexeiuteshev,ekalinina,marina.yashina}@gmail.com 101099 March 30, 2023 ============================================================================================================================================================================================================================================================================== empty We treat the problem of the Frobenius distance evaluation from a given matrix A ∈ℝ^n× n with distinct eigenvalues to the manifold of matrices with multiple eigenvalues. On restricting considerations to the rank 1 real perturbation matrices, we prove that the distance in question equals √(z_∗) where z_∗ is a positive (generically, the least positive) zero of the algebraic equation ℱ(z) = 0, ℱ(z):= 𝒟_λ( [ (λ I - A)(λ I - A^⊤)-z I_n ] )/z^n and 𝒟_λ stands for the discriminant of the polynomial treated with respect to λ. In the framework of this approach we also provide the procedure for finding the nearest to A matrix with multiple eigenvalue. Generalization of the problem to the case of complex perturbations is also discussed. Several examples are presented clarifying the computational aspects of the approach. Keywords: Wilkinson's problem, defective matrix, multiple eigenvalues, distance equation MSC 2010: 68W30, 15A18 , 12D10, 58C40 § INTRODUCTION The origin of the problem of finding the distance from a matrix A ∈ℂ^n× n to the set 𝔻 of matrices with multiple eigenvalues can be traced back to Wilkinson <cit.> who posed it in relation to the sensitivity analysis of matrix eigenvalues. The desired distance further will be treated with respect to either the 2-norm or to the Frobenius norm in ℂ^n× n and will be denoted d(A, 𝔻). It is usually referred to as the Wilkinson distance of A <cit.>. Alternatively, d(A, 𝔻) can be defined as the infA-B where B belongs to the subset of defective matrices, i.e. those possessing at least one eigenvalue whose geometric multiplicity is less than its algebraic multiplicity. Starting from Wilkinson's works <cit.>, the problem of evaluation of d(A, 𝔻) has been studied intensively in <cit.>. The most recent result is presented in the work <cit.>. We briefly trace the developed approaches. Most of them are in the framework of singular value analysis of appropriate parameter dependent matrices. The following theorem gives the min-max representation of d(A,𝔻) obtained by Malyshev <cit.>. Let A∈ℂ^n× n. Let the singular values of the matrix [[ A-λ I_n γ I_n; 𝕆_n× n A-λ I_n ]] be ordered like σ_1(λ,γ)≥σ_2(λ,γ)≥…≥σ_2n(λ,γ) ≥ 0. Then the 2-norm distance d(A,𝔻) can be evaluated as d(A,𝔻)=min_λ∈ℂmax_γ≥0σ_2n-1(λ,γ) . The straight computation of this distance is quite difficult, so to find this distance, in many works the notion of pseudospectra <cit.> is used. Definition. For both the 2-norm and the Frobenius norm, the ε-pseudospectra of a matrix A is Λ_ε(A)={σ_min<ε} where ε>0 and σ_min stands for the smallest singular value of the matrix A-zI. Equivalently, Λ_ε(A)={ z∈ℂ | (A+E-zI)=0, E∈ℂ^n× nE <ε} . If Λ_ε has n components, then A+E has n distinct eigenvalues for all perturbations E∈ℂ^n× n and hence A+E is not defective. In subsequent papers, the pseudospectra approach is used to find the distance to the nearest defective matrix. In <cit.>, a geometric solution to the problem of finding d(A,𝔻) in Frobenius norm is given. The nearest defective matrix is related to the critical points of the minimal singular value σ_min(x,y) of the matrix A-(x+𝐢 y)I that could be obtained by examination of pseudospectra of A. For an approximation of a multiple eigenvalue of the nearest defective matrix, the averaging heuristic by Puiseux series is proposed. Also an iterative method for finding this eigenvalue together with the minimal perturbation is presented. In <cit.>, it is proposed to find the smallest perturbation E such that the components of the pseudospectra of A+E coalesce. The problem is reformulated as follows. One needs to find z∈ℂ,ε∈ℝ,ε>0 and U,V∈ℂ^n,such that[Hereinafter ^⊤ stands for the transpose while ^𝖧 stands for the Hermitian transpose.] (A-zI)V-ε U=𝕆_n× 1,ε V-(A-zI)^𝖧 U=𝕆_n× 1, U^𝖧 V=0 . The algorithm to solve the system of equations presented in this work is rather expensive because it requires the repeated calculation of pseudospectra. Also any condition of coalescence of two pseudospectral curves is necessary. In <cit.>, a new computational approach to approximating the nearest defective matrix by a variant of Newton's method is suggested. The Implicit Determinant Method based on standard Newton's method is used to solve the system (<ref>) in <cit.>. There are several works considering generalizations of Wilkinson's problem for the cases of prescribed eigenvalues or their multiplicities  <cit.>, and matrix pencils <cit.>. The approaches developed in the above cited papers could be characterized as related to the Numerical Linear Algebra. The present paper aims at solving the stated problem for the case of Frobenius norm within the framework of symbolic computation approach. Namely, we reduce the problem to that of the univariate polynomial equation solving. As a matter of fact, the manifold 𝔻 of matrices with multiple eigenvalues in the ℝ^n^2 space of their entries is an algebraic one, i.e. it is represented by a multivariate polynomial equation. If we slightly modify Wilkinson's problem to that of finding d^2(A,𝔻), then the constrained optimization problem becomes an algebraic one in the sense that both the objective function and the constraint be polynomials. Application of the Lagrange multipliers method reduces the problem to that of system of algebraic equations solving. The latter can be resolved, at least in theory, via the analytical procedure of elimination of variables consisting in the multivariate resultant computation or the Gröbner basis construction. Application of these procedures to the system of equations of the treated problem, complemented with z-d^2(A,𝔻)=0, results in a univariate equation ℱ(z)=0 whose zero set contains all the critical values of the squared distance function. This equation will be further referred to as the distance equation and its computation is the priority of the present paper. This approach has been developed in <cit.>. Unfortunately, soon after that publication, a significant gap in reasoning was discovered. It was assumed that the value d(A, 𝔻) could be provided by only the rank 1 perturbation matrix E_∗ and that the nearest to A matrix B_∗=A+E_∗ in 𝔻 might possess only a double real eigenvalue. In Section <ref>, an example of the order 4 matrix A is given where the nearest in 𝔻 matrix possesses a pair of double complex-conjugate eigenvalues. As yet we failed to manage this scenario for the general statement of the problem; neither do we able to claim that it is a zero probability event. We confine ourselves here to considering the case where the critical values of d^2(A,𝔻) are provided only by the rank 1 perturbation matrices. For this case, the practical implementations of the elimination of variables procedure mentioned above can be reduced to just only two bivariate equations. One of these equations follows quite naturally from the developed in <cit.> approach. This is Φ(λ,z)=0 Φ(λ,z):= [ (λ I - A)(λ I - A^⊤)-z I_n ] . The more difficulties causes the deduction of the second equation. It happens to be ∂Φ(λ,z)/ ∂λ =0 . To obtain the distance equation, it is then sufficient to eliminate the variable λ from the obtained system. This can be managed with the aid of discriminant computation, i.e. the function of the coefficients of a polynomial responsible for the existence of a multiple zero for this polynomial. We recall some basic features of this function in Section <ref>. In Section <ref>, we prove the main result of the paper, namely that the value d^2(A,𝔻) is in the set of non-negative zeros of the distance equation. If A ∉𝔻 then generically d^2(A,𝔻) equals the least positive zero z_∗ of this equation. We also detail here the structure of the matrix B_∗ nearest to A in 𝔻. It appears that the multiple eigenvalue of B_∗ coincide with the multiple zero of the polynomial Φ(λ,z_∗). In Section <ref>, computational aspects of the proposed approach are discussed via solving the problem for the two families of matrices treated in the literature. In Section <ref>, we address to the generalization of Wilkinson's problem to the case of complex perturbations. Here the results are presented in a very concise manner with the potential intention of returning to them in future articles. Notation is kept to correlate with <cit.>. For a matrix A ∈ℝ^n× n, f_A(λ) denotes its characteristic polynomial, d(A, 𝔻) denotes the distance from A to the set 𝔻 of matrices possessing a multiple eigenvalue. E_∗ and B_∗ = A+ E_∗ stand for, respectively, the (minimal) perturbation matrix and the nearest to A matrix in 𝔻 (i.e. d(A,𝔻)=A- B_∗); we then term by λ_∗ the multiple eigenvalue of B_∗. I (or I_n) denotes the identity matrix (of the corresponding order). 𝒟 (or 𝒟_λ) denotes the discriminant of a polynomial (with subscript indicating the variable). Remark. All the computations were performed in CAS Maple 15.0 with those approximate done within the accuracy 10^-40. In the paper they are presented rounded to 10^-6. § ALGEBRAIC PRELIMINARIES It is well-known that in the (N+1)-dimensional space of the polynomial F(x)=a_0x^N + a_1x^N-1 +…+a_N ∈ℂ[x], a_0 0, N ≥ 2 coefficients, the manifold of polynomials with multiple zeros is given by the equation D(a_0,a_1,…,a_N)=0 D:=𝒟_x(F(x)) denotes the discriminant of the polynomial. Discriminant is formally defined as a symmetric function of the zeros {λ_1,…, λ_N } of the polynomial F(x) D_x(F(x))= a_0^2N-2∏_1≤ j < k ≤ N (λ_k - λ_j)^2 . This representation gives rise to further transformation of the discriminant into the homogeneous polynomial D(a_0,a_1,…,a_N) of the order 2N-2 with respect to the coefficients of F(x). Such a transformation can be implemented through a preliminary representation of discriminant in an appropriate determinantal form. We will follow the approach based on the Hankel matrix formalism <cit.>. For this aim, find first the Newton sums s_0,s_1,…,s_2N-2 of the polynomial F(x) with the aid of recursive formulas s_0=N, s_1=-a_1/a_0, s_k={[ -(a_1s_k-1+a_2s_k-2+…+a_k-1s_1+a_kk)/a_0, k≤ N ,; -(a_1s_k-1+a_2s_k-2+…+a_Ns_k-N)/a_0, k > N , ]. and compose the Hankel matrix S=[s_j+k]_j,k=0^N-1 = [[ s_0 s_1 s_2 … s_N-2 s_N-1; s_1 s_2 s_3 … s_N-1 s_N; s_2 s_3 s_4 … s_N s_N+1; … …; s_N-1 s_N s_N+1 … s_2N-3 s_2N-2 ]]_N× N . Denote by S_1,…, S_N= S its leading principal minors. One has 𝒟(F)=a_0^2N-2 S_N . The condition S_N=0, …, S_N-k+1=0, S_N-k 0 is the necessary and sufficient for the polynomial F(x) to possess k common zeros with F^' (x). In particular, if S_N=0, S_N-1 0, then F(x) possesses a unique multiple zero and the multiplicity of this zero equals 2. This zero can be computed via the formula λ = s_1-1/S_N-1| [ s_0 s_1 … s_N-3 s_N-1; s_1 s_2 … s_N-2 s_N; ⋮ ⋮; s_N-2 s_N-1 … s_2N-1 s_2N-3 ]| . The determinant in the right-hand side is constructed by deleting the last row and the last but one column in S. Consequently, the set 𝔻 of matrices with multiple eigenvalues is given by the equation 𝒟_λ( (λ I-B) ) =0 . For the case of polynomials with real coefficients, the sequence of leading principal minors of the matrix S permits one to establish the exact number of real zeros for F(x) <cit.>. Let S_N=0, …, S_N-k+1=0, S_N-k 0,…,S_1 0 Then the number of distinct pairs of complex-conjugate zeros for F(x) ∈ℝ[x] equals 𝒱(1,S_1,…, S_N-k) where 𝒱 denotes the number of variations of sign in the given sequence. In the space ℝ^N+1 of polynomials (<ref>) with real coefficients, the discriminant manifold (<ref>) separates the domains of vectors providing the coefficients of polynomials with the same number of real zeros. The last comment of the present section relates to application of discriminant to one problem from Elimination Theory. Consider a bivariate polynomial F(x,y)∈ℝ[x,y], F ≥ 2. The discriminant furnishes the tool for eliminating the variable x from the system of equations F(x,y)=0, ∂ F(x,y)/ ∂ x=0 . Namely, if (x_0,y_0) is a solution to the system (<ref>), then y_0 is necessarily a zero of the algebraic univariate equation 𝒴(y)=0 𝒴(y):=𝒟_x(F(x,y)) . The reverse statement is subject to an extra assumption. If y_1∈ℂ is a zero for 𝒴(y), then there exists a multiple zero for the polynomial F(x,y_1). Under the assumption that y_1 is a simple zero for 𝒴(y), x_1 is a unique multiple zero and its multiplicity equals 2. Then it can be expressed as a rational function of y_1 using the result of Theorem <ref>. These considerations are valid for all the solutions of the system (<ref>) provided that 𝒟_y(𝒴(y)) 0. § DISTANCE EQUATION In terms of the discriminant manifold referred to in the previous section, the problem of evaluation of d^2 (A,𝔻) is equivalent to that of constrained optimization minB-A^2 𝒟_λ(f_B(λ))=0, B∈ℝ^n× n . Here the constraint is an algebraic equation with respect to the entries of the matrix B. Traditional application of the Lagrange multipliers method reduces the problem to that of solving a system of n^2+1 nonlinear algebraic equations. Under the additional assumption the matrix B_∗∈ℝ^n× n providing a solution to this system possesses only one multiple eigenvalue and its multiplicity equals 2, it is possible to reduce the number of variables in the constrained optimization approach. The following result is presented in <cit.>: The value d^2(A,𝔻) belongs to the set of critical values of the objective function G(U):=U^⊤A A^⊤ U - ( U^⊤AU )^2 for the constrained optimization problem under constraints U^⊤U=1, U ∈ℝ^n . If U_∗ be the point providing d^2(A,𝔻), then the perturbation can be computed as E_∗=U_∗ U_∗^⊤ (κ I-A) κ:= U_∗^⊤ A U_∗ . The new optimization problem still have significant number of variables. We aim to eliminate all of them but introduce an extra one responsible for the critical values of the objective function. Stationary points of the function (<ref>) under the constraints (<ref>) can be found via Lagrange method applied to the function G(U)- μ (U^⊤U-1). This results into the system AA^⊤U-(U^⊤ A U)(A+A^⊤)U-μ U = 𝕆_n× 1 . Denote λ:=U^⊤ A U . Then the equation (<ref>) has a nontrivial solution with respect to U if and only if (AA^⊤-λ (A+A^⊤)-μ I)=0 . Under this condition, multiplication of (<ref>) by U^⊤ yields U^⊤AA^⊤U=2λ^2+μ . Wherefrom it follows that the critical values of the objective function (<ref>) are given by z=λ^2+μ . Substitution this into (<ref>) results in the equation connecting z and λ: Φ(λ,z)=0 where Φ(λ,z):= [ A A^⊤- λ (A+A^⊤)+(λ^2-z) I ] =[ (λ I - A)(λ I - A)^⊤-z I ] Zeros z_1,…, z_n of the polynomial Φ(λ,z) with respect to the variable z are evidently real since they are the squares of the singular values for the matrix λ I-A. Our further task is to deduce an extra equation connecting λ and z. The value d^2(A,𝔻) belongs to the set of non-negative zeros of the polynomial ℱ(z)≡𝒟_λ(Φ(λ,z))/z^n . Proof. Under the condition (<ref>), there exists a nontrivial solution for (<ref>) with respect to the column U (λ I-A)(λ I-A)^⊤U=z U . This equality means that U is the right singular vector for the matrix λ I - A corresponding to the singular value √(z). The corresponding left singular vector for that matrix can be found from the equality √(z)V:=(λ I - A)^⊤U . Dual relationship is valid for U: √(z)U=(λ I - A)V . From the conditions (<ref>) and (<ref>) it follows that U^⊤(λ I - A)U=0 . Multiply (<ref>) from the left by U^⊤. From (<ref>), it follows that √(z)=U^⊤(λ I - A)V . Multiply (<ref>) from the left by V^⊤ and utilize (<ref>): √(z)V^⊤V=V^⊤(λ I - A)^⊤U=√(z) . Wherefrom the two alternatives follow V^⊤V=1 √(z)=0 . Similarly, multiplication of (<ref>) from the left by U^⊤ and further application of (<ref>) yields √(z)U^⊤V=0 . This also leads to two alternatives: U^⊤V=0 √(z)=0 . Ignore the case √(z)=0. V^⊤V=1, U^⊤V=0 . Consider the equation (<ref>) as a definition of the √(z) as the function of λ. Differentiate this relation with respect to λ: d √(z)/d λ=U^⊤V+d U^⊤/d λ (λ I - A)V+U^⊤ (λ I - A) d V/d λ . With the aid of (<ref>) and (<ref>) transform this into U^⊤V+√(z)[ d U^⊤/d λ U + V^⊤d V/d λ] . Due to (<ref>) and (<ref>), we arrive at d √(z)/d λ = 0 . Equation (<ref>) defines implicit function z(λ). Differentiation of the identity Φ(λ,z(λ))≡ 0 with respect to λ yields the identity Φ^'_λ(λ,z)+Φ^'_z(λ,z) d z/d λ≡ 0 . Under the condition (<ref>), the variables λ and z are linked by an extra relationship Φ^'_λ(λ,z)=0 . Together with (<ref>), the deduced condition composes the system of algebraic equations Φ (λ,z)=0, Φ^'_λ(λ,z)=0 . According with the results of Section <ref>, elimination of λ from this system can be implemented with the aid of the discriminant computation, i.e. the variable z should satisfy the equation 𝒟_λ (Φ (λ,z))=0 . To prove the validity of (<ref>), it is necessary to additionally confirm that the left-hand side of the last equation is divisible by z^n. This is indeed the case, since the polynomial Φ(λ,0) possesses n multiple zeros coinciding with the eigenvalues of the matrix A. With ℱ(z) given by (<ref>), the distance equation ℱ(z) =0 is now well-defined and in Section <ref> we discuss some of related features and computational aspects. To conclude the present section, we have to detail the properties of the λ-component for the solution of the system (<ref>). Let the polynomial ℱ(z) defined by (<ref>) possess a positive real zero z_0 and this zero be simple. Then the polynomial Φ (λ,z_0) has a unique multiple zero and multiplicity of this zero equals 2. We denote by λ_0. It is evidently real and can be expressed as a rational function of z_0 via, for instance, formula (<ref>). The less evident conclusion is as follows: this multiple zero coincides with the multiple eigenvalue of the matrix in 𝔻 providing the critical value z_0 for the function d^2(A,𝔻). For any real solution (λ_0,z_0) of the system (<ref>) where z_0 0, there exists the rank 1 perturbation E_0 such that E_0=√(z_0) and the matrix B_0=A+E_0 possesses the multiple eigenvalue λ_0. Proof. The number √(z_0) is a singular value for the matrix λ_0 I - A. We intend to prove that the matrix from the theorem statement is defined by the formula E_0:=√(z_0)U_0 V_0^⊤ , where U_0 and V_0 are respectively the left and the right singular vectors of the unit norm for the matrix λ_0 I-A corresponding to √(z_0). Indeed, the matrix B_0=A+E_0 has λ_0 as the eigenvalue corresponding to the eigenvector V_0: B_0V_0=(A+E_0)V_0 (<ref>)= AV_0+√(z_0)U_0 (<ref>) =AV_0+(λ_0I-A)V_0=λ_0V_0 . If (B_0-λ_0 I)<n-1 then the theorem is proved. Assume that (B_0-λ_0 I)=n-1. Let us prove the existence of a column W such that (B_0-λ_0 I)W=V_0 . The necessary and sufficient condition for resolving this equation consists in the fulfillment of the equality (B_0-λ_0I)(B-λ_0I)^+V_0=V_0 where ^+ stands for the Moore-Penrose inverse of the matrix. It can be easily verified that (B_0-λ_0I)(B_0-λ_0I)^+=I-U_0U_0^⊤ (by assumption, (B_0-λ_0 I)=n-1), and the condition (<ref>) is fulfilled: (B_0-λ_0I)(B_0-λ_0I)^+V_0= (I-U_0U_0^⊤)V_0 (<ref>)= V_0 . The columns V_0 and W are linearly independent. Indeed, if α V_0+ β W=𝕆_n × 1 {α, β}⊂ℝ then on multiplying this equality from the left by B_0-λ_0 I it follows that β V_0 = 𝕆_n × 1, and thus β=0. But then α=0 since V_0 is a nonzero column. Hence, (B_0-λ_0I)^2V_0= 𝕆, (B_0-λ_0I)^2W= 𝕆 for the linear independent V_0 and W. Consequently, (B_0-λ_0 I)^2≤ n-2 and this gives evidence that λ_0 should be a multiple eigenvalue for B_0. If A ∉𝔻, then d(A,𝔻) = √(z_∗) , where z_∗ is the minimal positive zero of the polynomial (<ref>) provided that this zero is not a multiple one. Minimal perturbation is evaluated by the formula E_∗=U_∗U_∗^⊤ (λ_∗I-A) . Here λ_∗ is the multiple zero for the polynomial Φ(λ,z_∗) and U_∗∈ℝ^n, U_∗=1 is the left singular vector of the matrix λ_∗I-A corresponding to the singular value √(z_∗). The significance of condition for simplicity of the minimal positive zero z_∗ can be explained as follows. Since we are looking for only real perturbations, formula (<ref>) yields such a matrix if λ_∗ is real. For the matrices of the order n≥ 4, it might happen that the system (<ref>) possesses a solution (z_∗, λ_∗) with an imaginary λ_∗ (we give an example of such a matrix in Section <ref>). Then the system necessarily possesses the solution (z_∗,λ_∗). This implies (v. the last comment from Section <ref>) that z_∗ should be a multiple zero for (<ref>). Therefore, the condition for simplicity of z_∗is sufficient to prevent such an occasion. Formal verification of this condition can be replaced by a more general one relating the discriminant of ℱ(z): 𝒟_z(ℱ(z)) 0 . § PROPERTIES OF THE DISTANCE EQUATION The distance equation for the matrix A=[a_jk ]_j,k=1^2 is found in the form ℱ(z):=16[ (a_11-a_22)^2+(a_12+a_21)^2 ]·{[4z- 𝒟 (f_A(λ)) ]^2-16(a_12-a_21)^2z }=0 . Polynomial in braces has only real zeros with respect to z since its discriminant equals 256(a_12-a_21)^2[(a_11-a_22)^2+ (a_12+a_21)^2 ] ≥ 0 . Some terms in the canonical representation of the polynomial (<ref>) can be explicitly expressed via the entries of the matrix A: Φ(λ,z)≡λ^2n- 2 (A) λ^2n-1 +(-nz+ (AA^⊤)+p_2 ) λ^2n-2 +… + (AA^⊤-zI) . Here p_2 is the coefficient of λ^n-2 in the characteristic polynomial f_A + A^⊤(λ):= (λ I - A-A^⊤). It happens that this polynomial is also responsible for the order of the distance equation. One has ℱ(z)≡ 4^n [ 𝒟_λ (f_A+A^⊤(λ)) ]^2 z^n(n-1) + z . Proof. Let {μ_1,…, μ_n } be the spectrum of the matrix A+A^⊤ while P∈ℝ^n× n be an orthogonal matrix reducing it to the diagonal form: P^⊤(A+A^⊤) P= (μ_1,…, μ_n) . Apply the same transformation to the determinant (<ref>): Φ(λ,z) ≡[ P^⊤AA^⊤P + (λ^2-μ_1 λ-z,…, λ^2-μ_n λ-z) . ] The leading term of the polynomial 𝒟_λ (Φ(λ,z)) with respect to z coincide with that of 𝒟_λ(∏_j=1^n (λ^2-μ_j λ-z)) . The set of zeros of the polynomial under the discriminant sign is as follows {1/2(μ_j ±√(μ_j^2+4z)) }_j=1^n . Using the definition (<ref>) of the discriminant, one gets 𝒟_λ(∏_j=1^n (λ^2-μ_j λ-z)) = ∏_j=1^n (4 z + μ_j^2) ∏_1≤ j < k ≤ n[ z^2 (μ_k-μ_j)^4] . Coefficient of the monomial z^n^2 in the right-hand side can be recognized, via (<ref>), as the square of the discriminant of the characteristic polynomial of A+ A^⊤. As for the determining the structure of the free term of ℱ(z), our successes are restricted to the following Hypothesis. If computed symbolically with respect to the entries of A, ℱ(0) has a factor [𝒟_λ (f_A(λ)) ]^2. According to Theorem <ref>, the polynomial ℱ(z) can be constructed in the form of determinant of a suitable Hankel matrix. For this aim, compute first the Newton sums {s_j(z)}_j=0^4n-2 for the polynomial Φ(λ,z) treated with respect to λ. Direct utilization of the formulas (<ref>) requires the canonical representation (<ref>) for the polynomial Φ(λ,z) while initially we have just only its representation in the determinantal form (<ref>). Fortunately, the Newton sums can be computed in an alternative way. Indeed, Φ(λ,z) ≡ (λ I_2n- W) W:=[ [ A^⊤ √(z) I_n; √(z) I_n A ]] and it is known that the Newton sums of the characteristic polynomial of a matrix can be computed as the traces of matrix powers: s_j(z) ≡(W^j) j ∈{0,1,…} Thus, one has s_2(z)=2 ((A^2)+nz), s_3(z)=2((A^3)+3 z (A)), … Compose the Hankel matrix S(z):=[ s_j+k(z) ]_j,k=0^2n-1 and compute the sequence of its leading principal minors S_1(z),…,S_2n(z). Due to (<ref>) and (<ref>), S_2n(z)≡𝒟_λ(Φ(λ,z))≡ℱ(z) z^n . Evidently, the polynomial Φ(λ,0) possesses only n double zeros, and they all are distinct provided that A ∉𝔻. Consequently, due to Theorem <ref>, one has S_n+1(0)=0,…, S_2n(0)=0. Polynomial ℱ(z) does not have negative zeros. The number of its positive zeros lying within the interval [0,z_0], z_0>0 is not less than | 𝒱(1,S_1(z_0),…,S_2n(z_0)) - 𝒱(1,S_1(0),…,S_n(0)) | . Proof. The first claim of the theorem follows from the positive definiteness of the matrix (λ I - A)(λ I - A)^⊤-z I for z< 0. By Theorem <ref>, the number 𝒱(1,S_1(z_0),…,S_2n(z_0)) equals the number of complex-conjugate pairs of zeros for the polynomial Φ(λ,z_0). When the parameter z varies from 0 to z_0, the discriminant 𝒟_λ (Φ(λ,z)) vanishes at any value of z where a pair of real zeros of Φ(λ,z) transforms to a pair complex-conjugate ones or vice versa. The discriminant vanishes at these values. Theorem <ref> claims that the degree of the distance equation generically equals n(n-1). One can immediately watch that for the skew-symmetric matrix A this estimation is not valid. Moreover, for this type of matrices, polynomial ℱ(z) vanishes identically. Some other types of matrices that permit explicit representation for the polynomial Φ(λ,z), and, as a consequence, for the value d(A, 𝔻), in terms of the spectrum of A can be found in <cit.>. We summarize those results in the following Let all the eigenvalues λ_1,…,λ_n of A be distinct. One has: Φ(λ,z)≡∏_j=1^n [(λ-c)^2-(λ_j-c)^2-z] A= + cI_n , where c ∈ℝ is an arbitrary scalar; Φ(λ,z)≡∏_j=1^n (λ^2-z+1-2λ(λ_j)) A ; Φ(λ,z)≡∏_j=1^n [(λ-λ_j)^2-z ] A . For the case (<ref>), Φ(λ,z) has a multiple zero if n≥ 2. For the case (<ref>), Φ(λ,z) has a multiple zero if n≥ 3. For the both cases, the distance d(A, 𝔻) is attained at the continuum of matrices in 𝔻 <cit.>. Find d(A,𝔻) for the skew-symmetric matrix A=[ [ 0 -4 2 -1; 4 0 7 3; -2 -7 0 11; 1 -3 -11 0 ]] . Solution. Here Φ(λ,z)≡( λ^4-2 λ^2z+200 λ^2+z^2-200 z+3249 )^2 , and 𝒟_λ (Φ(λ,z)) ≡ 0. However, if we take 𝒟_λ (√(Φ(λ,z)))= 𝒟_λ(λ^4-2 λ^2z+200 λ^2+z^2-200 z+3249) the result is the true distance equation 11667456256 z^2-2333491251200 z+37907565375744=0 . Its least positive zero equals 100-√(6751) =1/4(√(314)-√(86))^2 where ± 1/2 𝐢 (√(314)-√(86)) are the eigenvalues of A. Remark. Similar trick works also for the case of orthogonal matrices. § EXAMPLES AND COMPUTATIONAL ASPECTS Once the canonical form of the distance equation is computed, Wilkinson's problem is nearly solved. Indeed, for a univariate algebraic equation, the exact number of real zeros, as well as their location, could be trustworthy determined via purely algebraic procedures. Remark. Theorem <ref> claims that generically the degree of the distance equation equals n(n-1). The both examples below fall into this genericity. For instance, one has ℱ(z)=870 for n=30. Find d(F_n,𝔻) for Frank's matrix <cit.> F_n=[ [ n n-1 n-2 … 2 1; n-1 n-1 n-2 … 2 1; 0 n-2 n-2 … 2 1; 0 0 n-3 … 2 1; ⋮ ⋮ ⋱ ⋮ ⋮; 0 0 0 … 1 1 ]] . Solution. For n=3, one has Φ(λ,z)= λ^6-12 λ^5+ ( -3 z+48 ) λ^4 + ( 24 z-74 ) λ^3 + ( 3 z^2-73 z+48 ) λ^2+ ( -12 z^2+70 z-12 ) λ- z^3+25 z^2-33 z+1 and ℱ(z)= 23839360000 z^6-476315200000 z^5+3522206312000 z^4- 11668368222400 z^3 +16297635326400 z^2-6895772352000 z+ 230443315200 . Distance equation has only real zeros, namely z_1 ≈ 0.036482, z_2 ≈ 0.648383, z_3 ≈ 2.316991, z_4 ≈ 4.954165, z_5 ≈ 5.274176, z_6 = 27/4=6.75 . Thus, d(F_3,𝔻)=√(z_1)≈ 0.191004. To find the corresponding perturbation via (<ref>), first evaluate the multiple zero for Φ(λ,z_1) via (<ref>): λ_∗≈ 0.602966 . Then evaluate the unit left singular vector of the matrix λ_∗ I - A corresponding to √(z_1): U_∗≈[0.639244, -0.751157, -0.164708]^⊤ Finally, E_∗≈[[ -0.019161 -0.041159 0.113343; 0.022516 0.048365 -0.133186; 0.004937 0.010605 -0.029204 ]] . The nearest to F_3 matrix in 𝔻 B_∗=F_3+E_∗≈[[ 2.980838 1.958840 1.113343; 2.022516 2.048365 0.866813; 0.004937 1.010605 0.970795 ]] possesses the spectrum {λ_∗, λ_∗, 6-2λ_∗≈ 4.794067 }. For n>3, the set of nonreal zeros for the distance equation becomes nonempty, and its cardinality, relative to that of real, increases fastly with n. n d(F_n,𝔻) ≈ coefficient size number of real zeros timing (s) 5 4.499950 × 10^-3 ∼ 10^50 12 - 10 3.925527 × 10^-8 ∼ 10^300 30 - 12 1.849890 × 10^-10 ∼ 10^480 34 0.13 20 3.757912 × 10^-21 ∼ 10^1690 62 5 30 1.638008 × 10^-36 ∼ 10^4450 102 30 . The results for F_10 and F_12 confirm estimations d_10≈ 3.93·10^-8 and d_12≈ 1.85· 10^-10 given in <cit.>. Find d(K_n,𝔻) for Kahan's matrix <cit.> K_n=[ [ 1 -c -c … -c -c; 0 s -sc … -sc -sc; 0 0 s^2 … -s^2c -s^2c; ⋱ … ; 0 0 0 ⋱ s^n-2 -s^n-2c; 0 0 0 … 0 s^n-1 ]] s^2+c^2=1 . Solution. We present computational results for two specialization of parameter values. The first one is s= 3/5, c=4/5: n d(K_n,𝔻) ≈ coefficient size number of real zeros timing (s) 5 1.370032 × 10^-3 ∼ 10^310 8 - 10 5.470834 × 10^-6 ∼ 10^2970 48 - 15 2.246949 × 10^-8 ∼ 10^10590 138 6.7 20 9.245309 × 10^-11 ∼ 10^25730 288 145.4 25 3.984992× 10^-10 ∼ 10^52910 258 218.23 30 1.240748× 10^-11 ∼ 10^92460 464 937.66 The second test series correspond to a specialization s^n-1=1/10 treated in <cit.>. For this case, an extra difficulty results from approximation of the entries of the matrix K_n as rational numbers. This results in increasing the length of the coefficients of the distance equation. Compared with the previous case, the timing increases drastically, i.e. more than 10^2 times for the same specializations of n. n d(K_n,𝔻) ≈ number of real zeros 6 4.704940 × 10^-4 10 10 1.538157 × 10^-5 18 15 4.484974 × 10^-7 28 20 1.904858 × 10^-8 38 The results for K_6,K_15 and K_20 confirm estimations given in <cit.>. It should be emphasized however that computation of the whole sets of real zeros for the distance equation is redundant for evaluation of d(A,𝔻). We need to find just only the least positive zero of ℱ(z). For this aim, the determinantal representation (<ref>) for this polynomial might be sufficient for the real zero localization. According to Theorem <ref>, the lower estimate for the number of real zeros of ℱ(z) lying within the interval [0,z_0], z_0 >0 is given by the number (<ref>). If this number is not zero then at least one real zero for ℱ(z) lies in [0,z_0], and the next step in its localization might be the treatment of the matrix S(z_0/2). Experiments with the Frank's matrix (<ref>) demonstrate the unambiguity of the zero isolation process. For the matrix F_10, one has 𝒱(1,S_1(0),…,S_10(0))=0, i.e. all the eigenvalues of A are real. Then (<ref>) coincides with 𝒱_z_0:= 𝒱(1,S_1(z_0),…,S_10(z_0),…, S_20(z_0)) . Some specializations for z_0 z_0 10^-3 10^-9 2× 10^-15 10^-15 𝒱_z_0 5 3 1 0 demonstrate that the number of real zeros of ℱ(z) lying in any interval [0,z_0] happens to be equal to 𝒱_z_0. For instance, there are precisely 5 zeros within the interval [0,10^-3], namely 1.540976× 10^-15, 7.739368× 10^-15, 7.463686 × 10^-13, 1.403045 × 10^-9, 1.412301 × 10^-5 . However, for the case of the matrix [ [ 1 1 -2; 2 1 0; -3 1 1 ]] variations 𝒱_0.4=0, 𝒱_0.5=1, 𝒱_2.25=0 permit one to locate single zeros within the intervals [0.4, 0.5] and [0.5,2.25] but are unable to detect this number for [0.4, 2.25]. § COUNTEREXAMPLES We exemplify here two cases (a) The minimal positive zero of the distance equation not always provides the value d^2(A,𝔻) even if we restrict ourselves to the rank 1 perturbation matrices; (b) The distance d(A,𝔻) is not always provided by the rank 1 perturbations. For the matrix A(ϵ)= [ [ 0 1 1 0; -1 0 0 1; ϵ 0 0 1; 0 0 -1 0 ]] , find d(A(ϵ),𝔻) for ϵ >0. Solution. Distance equation is provided by the polynomial ℱ(z)≡ 65536ϵ^8 [(ϵ+2)^4 z^2 -2ϵ(ϵ+8)(ϵ+2)^2 z + ϵ^2(ϵ-8)^2]^2 ·[(ϵ+1)z-3ϵ-1]^4 × (z^2-3 z+1) [z^2-( ϵ^2+3 ) z+(ϵ+1)^2] . Its zeros are z_1=ϵ(√(ϵ)-√(8))^2/(ϵ+2)^2, z_2=ϵ(√(ϵ)+√(8))^2/(ϵ+2)^2, z_3=3ϵ+1/ϵ+1, z_4=3-√(5)/2≈ 0.381966, z_5=3+√(5)/2≈ 2.618033 , z_6=1/2(ϵ^2+3-|ϵ-1| √(ϵ^2+2 ϵ^2+5)), z_7= 1/2(ϵ^2+3+|ϵ-1| √(ϵ^2+2 ϵ^2+5)) are all real. Zero z_4 is simple, it coincides with the square of a singular value of the matrix A, and the polynomial Φ(λ,z_4) has the real double zero λ_4=0. The corresponding value of the distance function from A to 𝔻 does not depend on ϵ, it equals[Amazing coincidence with the reciprocal to the golden ratio!] √(z_4)=√(5)-1/2≈ 0.618033 . The corresponding perturbation and matrix in 𝔻 are as follows: E_4=1/10[[ 0 √(5)-5 (3√(5)-5 0; 0 0 0 0; 0 0 0 0; 0 -2 √(5) 5-√(5) 0 ]], B_4=1/10[[ 0 5+√(5) 5+3√(5) 0; -10 0 0 10; 10ϵ 0 0 10; 0 -2√(5) -5-√(5) 0 ]] . Double eigenvalue of B_4 is just 0. Next, we do not need to treat the zeros z_6, z_7 and z_3, since they are greater than z_4. Also z_2 >z_1, therefore, the two zeros that can compete for the distance value are z_1 and z_4. It can be verified that z_1 ≤ z_4 ϵ≤ϵ_2 ϵ_2 = 2√(2)(√(5)+3)√(√(5)+2)+7√(5)+15≈ 61.133652 . It looks like d(A,𝔻)=√(z_1) for ϵ≤ϵ_2. However, this is not true for some subinterval in [0,ϵ_2]. Indeed, z_1 is a double zero for ℱ(z), and polynomial Φ(λ,z_1) possesses two double zeros: λ_1,2=±√(K(ϵ))/ϵ+2 K(ϵ):=√(2)(ϵ-√(2)√(ϵ)+2)(√(ϵ)+1/√(2))(√(ϵ)+√(5)+1/√(2))(√(ϵ)-√(5)-1/√(2)) . These zeros are real only for ϵ≥ϵ_1 ϵ_1:=3-√(5)≈ 0.763932 . For the values ϵ < ϵ_2, the minimal positive zero of the distance equation is not responsible for the distance from A to 𝔻. It seems that d(A,𝔻)=√(z_4) for ϵ < ϵ_2. However, this statement is also invalid for some subinterval of the parameter values. The matrix E(ϵ):= ϵ (8-ϵ)/(ϵ^2+16)^2[ [ 0 -4 ϵ ϵ^2 0; -4 ϵ 0 0 ϵ^2; -16 0 0 4 ϵ; 0 -16 4 ϵ 0 ]] represents a rank 2 perturbation that provides for the matrix A(ϵ) + E(ϵ) a pair of double eigenvalues λ_1,2 =±1/ϵ^2+16√(( ϵ^2+4 ϵ-16 ) ( 3 ϵ^ 2+4 ϵ+16 )) . These eigenvalues are non-real for ϵ < 2(√(5)-1) ≈ 2.472136. For these parameter values, one has E(ϵ) =√(2)ϵ (8-ϵ)/ϵ^2+16 and this value is lesser than √(z_1) for ϵ < ϵ_c where ϵ_c denotes the least positive zero of the polynomial ϵ^8-80 ϵ^7-368 ϵ^6-1024 ϵ ^5+64 ϵ^4-9216 ϵ^3-16384 ϵ^2- 32768 ϵ+65536 ; i.e. ϵ_c ≈ 1.055249. [t]125mm < g r a p h i c s > Figure 1. Summarizing: d(A(ϵ), 𝔻)= {[ √(2)ϵ (8-ϵ)/(ϵ^2+16) ϵ∈ [0, ϵ_c]; √(ϵ)|√(ϵ)-√(8)|/(ϵ+2) ϵ∈ [ ϵ_c, ϵ_2]; (√(5)-1)/2 ϵ > ϵ_2 ]. The plot is displayed in Fig. 1 (the first formula — red, the second one — blue, the third one — green). Remark. As it is mentioned in Introduction, the case where d(A,𝔻) is achieved at the rank 2 matrix (i.e. the nearest in 𝔻 matrix possesses two double imaginary eigenvalues) is beyond our consideration. We are not able even to conjecture whether this is a zero probability event or not. § COMPLEX PERTURBATIONS The method proposed above can be extended to the case of complex perturbations. For a real matrix A, we are now looking for the distance to the nearest complex matrix B with multiple eigenvalue: d_C(A,𝔻):=minB-A 𝒟_λ(f_B(λ))=0, B∈ℂ^n× n . Warning. The present section should be considered as a draft of a separate publication to be prepared sometime afterwards. We skip here the details of algebraic backgrounds, proofs of theoretical results and do not bother ourselves with mentioning that the validity of some of the declared results is subject to several extra assumptions preventing the appearance of troubles similar to those dealt with in the previous section. Consider the polynomial Θ(a,b,z)=[((a+b 𝐢)I-A)((a-b𝐢)I-A^⊤)-zI] and generate the system of algebraic equations Θ=0, ∂Θ /∂ a =0, ∂Θ/ ∂ b=0 . We are looking for the real solutions to this system. Since Θ(a,0,z) (<ref>)≡Φ(a,z) , this solution set includes that for the system (<ref>). If the system (<ref>) possesses a solution (a_0,b_0,z_0) with b_0 0 then it has the solution (a_0,-b_0,z_0). Proof. Polynomial Θ(a,b,z) is even in b: Θ(a,-b,z)=[ ((a+𝐢 b)I-A^⊤)((a-𝐢 b)I-A) - zI ] =[ {((a+𝐢 b)I-A^⊤)((a-𝐢 b)I-A)}^⊤ - zI ] =[((a-𝐢 b)I-A^⊤) ((a+𝐢 b)I-A) - zI ]=Θ(a,b,z) . Consequently Θ^'_a is even in b while Θ^'_b is odd b. The latter becomes even on dividing by b. Our aim is to eliminate the variables a and b from the system (<ref>), i.e. to find the bivariate discriminant 𝒟_a,b(Θ) for the polynomial Θ(a,b,z) treated with respect to these variables. The discriminant 𝒟_x,y(F) of a polynomial F(x,y,z) ∈ℂ[x,y,z] is formally defined as the result of elimination of variables x and y from the system of equations F=0, ∂ F / ∂ x=0, ∂ F / ∂ y=0 . This is a polynomial in z and its vanishment at z=z_0 ∈ℂ is the necessary and sufficient condition for the existence of solution (x_0,y_0,z_0) ∈ℂ^3 to the system (<ref>), or equivalently, for the existence of the multiple zero (x_0,y_0) for the polynomial F(x,y,z_0). Constructive computation of discriminant can be implemented in several ways, and we will exemplify below the procedure based of the Bézout construction of the resultant <cit.>. The discriminant 𝒟_a,b(Θ(a,b,z)) is factorized as follows: 𝒟_a,b(Θ(a,b,z))≡ z^n(n+1)/2ℱ(z) ℱ(z) . Here ℱ(z) is defined by (<ref>), while ℱ(z) ∈ℝ[z], ℱ(z) =n(n-1)(n-2)/2 , (For n=2 polynomial ℱ(z) is just a constant). According to Section <ref>, the distance equation ℱ(z)=0 is responsible for the rank 1 real perturbation that provides the distance d(A,𝔻). It turns out that the equation ℱ(z) = 0 is responsible for the rank 1 imaginary perturbation. Its real zero z_0 corresponds to a pair of multiple zeros of the polynomial Θ(a,b, z_0), and these zeros are either in the form (a_0, ±β_0) or in the form (a_0, ±𝐢β_0) with real β_0. We are definitely interested only in the real solutions for the system (<ref>). Let the system (<ref>) possess a real solution (a_0,b_0,z_0) with z_0 >0,b_0 0. Denote U_0∈ℂ^n, U_0=1 the left singular vector for the matrix (a_0+𝐢 b_0)I-A corresponding to the singular value √(z_0). Then the rank 1 perturbation E_0=U_0 U_0^𝖧 ((a_0+𝐢 b_0) I-A) is such that E_0= √(z_0) and the matrix B_0=A+E_0 ∈ℂ^n× n possesses the double eigenvalue a_0+𝐢 b_0. Remark. Evidently, the matrix E_0 provides for the matrix B_0=A+E_0 the double eigenvalue a_0-𝐢 b_0. In view of Theorem <ref>, the distance d_C(A,𝔻) results from the competition between the least positive zero of ℱ(z) and that minimal positive zero of ℱ(z) that corresponds to the real solution for the system (<ref>). Computation of the polynomial ℱ(z) can be simplified if we take into account Theorem <ref>. Substitute 𝔟:=b^2 in the polynomials of the system (<ref>) and denote Ξ(a,𝔟,z):=Θ(a,b,z), Ξ_a(a,𝔟,z):=Θ^'_a(a,b,z), Ξ_𝔟(a,𝔟,z):=Θ^'_b(a,b,z)/b . The result of elimination of variables a and 𝔟 from the system Ξ=0, Ξ_a=0, Ξ_𝔟=0 is the equation z^n(n-1)/2ℱ(z)=0 . If z_0 is a positive zero of ℱ(z), the corresponding real solution to the system (<ref>) might have the 𝔟-component either positive or negative. We are interested only in the positive variant. Find d_C(A,𝔻) for A= [ [ 0 1 0; 0 0 1; -91 -55 -13 ]] . Solution. First compute the polynomial ℱ(z) via (<ref>): ℱ(z) := 33076090700402342058246544 z^6-377039198861306289080145178864 z^5 +937864902703881321034450183916 z^4-771868276098720970149792503999 z^3 +211070978787821517684022650624 z^2 -510584100140452518540394496 z+319295875259784560640000 . Its real zeros are as follows z_1≈ 0.739336, 0.765571, 0.980468, 11396.658548 . Next compose the polynomial Ξ(a,𝔟,z): Ξ(a,𝔟,z)=-z^3+(3a^2+3𝔟+26a+11477)z^2 -(3 a^4+6 a^2𝔟+3 𝔟^2+52a^3+52a𝔟+11756a^2+11536𝔟+11466 a+19757)z +( a^2+𝔟+14 a+49 ) ( (a^2+𝔟+6 a+13)^2- 16 𝔟) . Now we trace briefly the procedure of elimination of a and 𝔟 from the system (<ref>). Consider the monomial sequence 𝕄:={𝔪_j(a,𝔟)}={1,a,𝔟, 𝔟^2} . It is possible to reduce the polynomial 𝔪_j Ξ modulo Ξ_a and Ξ_𝔟, i.e. to find the polynomials {β_jk(z) }_j,k=1^4 ⊂ℝ[z] and {p_11(a,𝔟,z), p_j2(a,𝔟,z)}_j=1^4 ⊂ℝ[a,𝔟,z] satisfying the identity m_j Ξ≡β_j1(z)+ β_j2(z) a + β_j3(z) 𝔟 + β_j4(z) 𝔟^2 + p_j1Ξ_a + p_j2Ξ_𝔟 j∈{1,2,3,4} . For instance, β_11(z)= -17718805921 z^2+610367232 z+22937600, β_12(z)= -39353600 z+5324800, β_13(z)=146694400 z-512000, β_14(z)=-307200, …, β_44(z)= - 76550493273549926400 z^3+ 162810741053705011200 z^2- 1867736871075840000 z- 50331648000000 . Compose the Bézout matrix 𝔅(z):= [ β_jk(z) ]_j,k=1^4 . Then 𝔅(z) ≡ z^3ℱ(z) where ℱ(z) = 412324266119803814719539025 z^3+ 33923334498676415590177600 z^2 +691077589890510378371072 z- 899669298077697638400 . For any zero z_0 of this polynomial, the corresponding a and 𝔟 components of the solution to the system (<ref>) can be obtained in the following way. Denote by {𝔅_4j}_J=1^4 the cofactors of 𝔅 corresponding to the entries of the last row of the matrix 𝔅. Then the a-component of solution is connected with the z-component as a=𝔅_42/𝔅_41= 43719663040898080379 z^2+2929017747573439808 z+ 29336262189312000/2(624300876564482975 z^2-226254560538037856 z-3469512291865600) while the 𝔟-component as 𝔟=𝔅_43/𝔅_41= 3083432482762007609519 z^3+ 1101690698089389073600 z^2+ 67186386329988787456 z- 129087561954918400/16(624300876564482975 z^2-226254560538037856 z-3469512291865600) Polynomial ℱ(z) possesses a single real zero, namely[All the decimals in the following approximation are error-free.] z_1 ≈ 0.0012268490707391199222512104943 , and substitution of this value into the last formulas yields a= a_1 ≈ -4.403922040624116177182912013601, 𝔟 = 𝔟_1 ≈ 0.750705046015830894563798035515 . Since 𝔟_1>0, one may claim that d_C(A,𝔻)=√(z_1)≈ 0.035026405335676681771543151648 . The two perturbations in ℂ^3× 3 providing this distance correspond to the solutions (a_1,b_1,z_1) (a_1,-b_1,z_1) b_1=√(𝔟_1)≈ 0.866432366671415902596255690462 . of the system (<ref>). Let us compute via (<ref>) the one corresponding to (a_1,-b_1,z_1). The unit left singular vector of (a_1-𝐢 b_1)I-A corresponding to the singular value √(z_1) is as follows U_1 ≈[ 0.930609, 0.360923+ 0.039918 𝐢, 0.045052+ 0.008866 𝐢]^⊤ and the minimal perturbation E_1≈[[ 0.001289-0.000442 𝐢 -0.007120+0.000832 𝐢 0.031666+0.002551 𝐢; 0.000519-0.000116 𝐢 -0.002797+0.000017 𝐢 0.012172+0.002348 𝐢; 0.000067-0.000009 𝐢 -0.000353-0.000028 𝐢 0.001509+0.000425 𝐢 ]] . The spectrum of the matrix A+E_1 is { a_1-𝐢 b_1, a_1-𝐢 b_1 ,-13-2(a_1-𝐢 b_1) ≈ -4.192156-1.732865 𝐢} . To test the performability of the algorithm sketched in the present section, we chose the next matrix from the Matlab gallery('grcar',6). Find d_C(A,𝔻) for A= [ [ 1 1 1 1 0 0; -1 1 1 1 1 0; 0 -1 1 1 1 1; 0 0 -1 1 1 1; 0 0 0 -1 1 1; 0 0 0 0 -1 1 ]] . Solution. Here the minimal zero of ℱ(z) equals z_1≈ 0.116565 and that of ℱ(z) equals z_1 ≈ 0.04630491415327188209539627157 . The latter corresponds to the real solution for the system (<ref>): (a_1,± b_1, z_1) a_1 ≈ 0.753316, b_1 ≈ -1.591155 . Thus, one obtains d_C(A,𝔻) = √(z_1)≈ 0.2151857666140395125353 . This confirms estimation d_C(A,𝔻) ≈ 0.21519 from <cit.>. For the solution (a_1,b_1, z_1), the spectrum of the nearest to A matrix in 𝔻 is as follows { 0.361392-1.944783 𝐢,1.139422-1.239762 𝐢,1.502453-0.616966 𝐢,1.490100+0.619201 𝐢,a_1+𝐢 b_1, a_1+𝐢 b_1 } . § CONCLUSION We have investigated Wilkinson's problem for the distance evaluation from a given matrix to the set of matrices possessing multiple eigenvalues. The proposed approach consists in the construction of distance equation with the zero set containing the critical values of the squared distance function. This construction is realized in the ideology of symbolic computations, i.e. the procedure consists of a finite number of elementary algebraic operations on the entries of the matrix. The representation of the distance equation with the aid of the discriminant function should not be taken as a complete surprise. Indeed, the Wilkinson's problem is the one of evaluation the distance to the discriminant manifold in the space of matrix entries. Hence, in view of this circumstance, the appearance of the discriminant in a solution to the problem is somehow natural. The more astonishing is the emergence of the discriminant in nearly any problem of distance evaluation from a point to an algebraic manifold in a multidimensional space <cit.>. Direction for further research is clearly related the stuff of Section <ref>, i.e. the problem of existence the rank 2 minimal perturbation providing d(A,𝔻). 99 AhmadAlam Ahmad, Sk.S., Alam, R.: On Wilkinson's problem for matrix pencils. ELA, 30, pp. 632–648 (2015) AkFrSp Akinola, R. O., Freitag, M. A., Spence A.: The calculation of the distance to a nearby defective matrix. Numerical Linear Algebra with Applications. 21:3, pp. 403–414 (2014) AlamBora Alam, R., Bora, S.: On sensitivity of eigenvalues and eigendecompositions of matrices. Linear Algebra Appl. 396, pp. 273–301 (2005) AlamBoraByersOverton Alam, R., Bora, S., Byers, R., Overton, M. L.: Characterization and construction of the nearest defective matrix via coalescence of pseudospectral components. Linear Algebra Appl. 435, pp. 494–513 (2011) ArmGraVel Armentia, G., Gracia, J.-M., Velasco, F.-E.: Nearest matrix with a prescribed eigenvalue of bounded multiplicities. Linear Algebra Appl., 592, 188–209 (2020) BikkerUteshev Bikker P., Uteshev A.Yu.: On the Bézout construction of the resultant. J.Symbolic Comput., 1999, 28, No 1. 45–88 (1999) Demmel Demmel, J.W.: Computing stable eigendecompositions of matrices. Linear Algebra Appl., 79, pp. 163–193 (1986) Demmel1 Demmel, J.W.: On condition numbers and the distance to the nearest ill-posed problem, Numer.Math. 51, pp. 251–289 (1987) Frank Frank, W.L.: Computing eigenvalues of complex matrices by determinant evaluation and by methods of Danilewski and Wielandt. J. Soc. Indust. Appl. Math. 6(4), pp. 378–392 (1958). Gantmacher Gantmacher, F.R.: The Theory of Matrices. Chelsea, New York (1959) Gracia Gracia, J.-M.: Nearest matrix with two prescribed eigenvalues. Linear Algebra Appl. 401, pp. 277–294 (2005) Kahan Kahan, W.: Numerical linear algebra. Canad. Math. Bull. 9, pp. 757–801 (1966) Kalinina_Uteshev_CASC22 Kalinina, E., Uteshev, A.: Distance evaluation to the set of matrices with multiple eigenvalues. LNCS, 13366. Springer, Cham, pp.206–224 (2022) KokLogKar Kokabifar, E., Loghmani, G.B., Karbassi, S.M.: Nearest matrix with prescribed eigenvalues and its applications. J. Comput. Appl. Math. 298, pp. 53–63 (2016) Lippert Lippert, R.A.: Fixing multiple eigenvalues by a minimal perturbation. Linear Algebra Appl. 432, pp. 1785–1817 (2010) LipEdel Lippert, R.A., Edelman, A.: The computation and sensitivity of double eigenvalues, in: Z. Chen, Y. Li, C.A. Micchelli, Y. Xu (Eds.), Advances in Computational Mathematics: Proc. Gaungzhou International Symposium, Dekker, New York, pp. 353–393 (1999) Malyshev Malyshev, A.: A formula for the 2-norm distance from a matrix to the set of matrices with multiple eigenvalues. Numer. Math. 83, pp. 443–454 (1999) Mengi Mengi, E.: Locating a nearest matrix with an eigenvalue of prespecified algebraic multiplicity. Numer. Math. 118, pp. 109–135 (2011) Ruhe Ruhe, A.: Properties of a matrix with a very ill-conditioned eigenproblem. Numer. Math. 15, pp. 57–60 (1970) TrEmb Trefethen, L. N., Embree, M.: Spectra and Pseudospectra, Princeton University Press, Princeton, NJ (2005) UteshevCherkasov Uteshev, A.Yu., Cherkasov, T.M.: The search for the maximum of a polynomial.J. Symbolic Comput. 25 (5). pp. 587–618 (1998) UteshevYashina2015 Uteshev, A.Yu., Yashina M.V.: Metric problems for quadrics in multidimensional space. J.Symbolic Comput., 68, Part I, pp. 287–315 (2015) Wilkinson Wilkinson, J.H.: The Algebraic Eigenvalue Problem, Oxford University Press, New York (1965) Wilkinson2 Wilkinson, J.H.: Note on matrices with a very ill-conditioned eigenproblem. Numer. Math. 19, 176–178 (1972) Wilkinson1 Wilkinson, J.H.: On neighbouring matrices with quadratic elementary divisors, Numer. Math. 44, 1–21 (1984) Wilkinson3 Wilkinson, J.H.: Sensitivity of eigenvalues, Util. Math. 25, 5–76 (1984)

http://arxiv.org/abs/2303.06710v2

20230312172254

Decision Making for Human-in-the-loop Robotic Agents via Uncertainty-Aware Reinforcement Learning

cs.RO

[ [ March 30, 2023 ================== empty empty In a Human-in-the-Loop paradigm, a robotic agent is able to act mostly autonomously in solving a task, but can request help from an external expert when needed. However, knowing when to request such assistance is critical: too few requests can lead to the robot making mistakes, but too many requests can overload the expert. In this paper, we present a Reinforcement Learning based approach to this problem, where a semi-autonomous agent asks for external assistance when it has low confidence in the eventual success of the task. The confidence level is computed by estimating the variance of the return from the current state. We show that this estimate can be iteratively improved during training using a Bellman-like recursion. On discrete navigation problems with both fully- and partially-observable state information, we show that our method makes effective use of a limited budget of expert calls at run-time, despite having no access to the expert at training time. § INTRODUCTION Deep Reinforcement Learning (DRL) has shown great progress in learning decision-making for complex robotic skills <cit.> using experiences collected by a robotic agent exploring an environment and receiving reward signals. Traditional RL agents use a policy learned during training in order to act autonomously at deployment. However, even a well-trained agent can encounter situations when deployed that are hard to make decisions for, for reasons such as partial state observability, uncertain dynamics, changes in state distributions between training and testing, etc. The Human-in-the-Loop (HitL) paradigm has been developed in robotics precisely for situations where an agent can act autonomously most of the time, but would still benefit from receiving assistance from an available (tele-)operator, usually assumed to be a human expert. This paradigm is particularly powerful if the agent itself makes the decision of when to request assistance, thus freeing the operator from having to monitor task progress. However, this approach gives rise to a critical decision-making problem on the agent's part: when to request help? Too few such requests can lead to the robot making mistakes, but too many requests will overload the expert and lose the benefit of semi-autonomous operation. In this paper, we propose a DRL-based method for a HitL agent to make the critical determination of when to request expert assistance. We posit that the best moment to request such assistance is when the agent is highly uncertain in the successful outcome of the task. From an RL perspective, we relate this uncertainty to the variance of the return from the current state, as perceived by the agent. Numerous RL algorithms provide methods for the agent to estimate the expected return from a given state. We show that similar methods can be used to estimate the variance of return as well during the training process. At deployment time, the agent can then request expert assistance when its estimate of the return variance from the current state falls below a given threshold. We dub our method HULA, for Human-in-the-loop Uncertainty-aware Learning Agent, and illustrate its operation in Fig. <ref>. Critically, HULA does not need to make any calls to the expert during training. This stands in contrast to a standard RL approach, where an agent could learn how to use an expert simply by making numerous such calls at train time. Nevertheless, our method is able to make effective use of a limited budget of expert calls in order to improve task performance. We summarize our key contributions as follows: * To the best of our knowledge, we are the first to propose a method for an HitL RL agent to learn how to effectively budget its interactions with a human expert at deployment time, without needing any expert calls during training. * We show that the variance of return, which can be estimated during training using Bellman-like equations, is an effective measure for agent uncertainty and can be used at deployment to determine when to request assistance. * Our experiments on discrete navigation problems with both fully- and partially-observable state information show that our method is as effective or better in managing its budget of expert calls compared to a standard learning approach that also makes thousands of expert calls during training. § RELATED WORK Our work extends standard reinforcement learning algorithms to human-in-the-loop policy that solves robotics tasks with the help of humans. Researchers have investigated how to learn such a policy. For example, Arakawa propose DQN-TAMER <cit.>, which is a method that incorporates a human observer model by real-time human feedback during training. Expected Local Improvement (ELI) <cit.> trains a state selector that suggests states requiring to query new actions from human experts. PAINT <cit.> learns a classifier to identify irreversible states and query the expert in case entering such states cannot be avoided. Hug-DRL <cit.> leverages a control transfer mechanism between humans and automation that corrects the agent’s unreasonable actions during training by human intervention, where it has demonstrated significant potential in autonomous driving applications. While these methods exploit human intervention during training, our work relies on the internal uncertainty of an agent and only uses an expert in test time. explicitly estimates the uncertainty of an RL agent and use it to make decisions about requesting an expert's assistance. Prior work has explored representing the uncertainty of outcomes during RL training <cit.>. For instance, Kahn use the variance of the dynamics model to represent the uncertainty of collision to avoid damage from high-speed collisions <cit.>. Ensemble Quantile Networks (EQN) <cit.> estimates both aleatoric and epistemic uncertainties via the combination of quantile networks(IQN) <cit.> and randomized prior function(RPF) <cit.>. These works use the uncertainty of an agent to perform task-specific decision-making, while our work differs in using it to collaborate with a human. Moreover, our work represents an agent's uncertainty using the variance of returns, which is a direct indicator of task performance. The closest work to ours is RCMP, which learns a HitL policy that queries the expert when the epistemic uncertainty estimated by the variance of multiple value functions is high <cit.>. However, RCMP requires expert queries during training, whereas our work learns an uncertainty model based on the variance of return and does not query an expert in train time. § METHOD Consider a problem for an autonomous agent formulated as a standard Markov decision process (MDP). A general MDP is defined as a tuple of (S, A, r, p), where S represents the state space, A represents the action space, r(s, a) is a reward function that evaluates immediate reward of action a ∈ A in a state s ∈ S, and p represent a transition distribution p(s_t+1| s_t, a_t). The goal of solving an MDP is to learn a policy π(a|s) that maximizes its expected return 𝔼_π[R], where R=∑γ^tr(s_t, a_t) and γ is a discount factor. Going beyond the standard MDP formulation, we now assume the availability of an expert that can give instructions to the robotic agent. When queried from a given state s_t, this expert can directly provide an action a_exp(s_t) for the robot to execute. We assume that the expert always provides high-quality advice, i.e. a policy of always following expert-provided actions from all states will achieve a satisfactorily high return on the MDP problem above. However, such a policy would be impractical: we assume that the expert has limited bandwidth or availability, thus it is desirable for the agent to balance the goal of achieving high returns with the goal of not overloading the expert with requests. Such a scenario is typical in HitL robotics applications. Our method aims to determine when the agent should request expert assistance while making the most of a limited number of such calls available during deployment. Furthermore, we will like to limit (or eliminate) the number of calls made to the expert during its training phase. To achieve this goal, our method leverages the uncertainty of the outcomes of the agent during exploration. Specifically, we use the variance of the return from a given state as a measure of uncertainty. Intuitively, a state with high variance of the return is one from which the agent, acting alone, could achieve a range of outcomes, ranging from very poor to very effective. We posit that these are the best states in which to request assistance. In contrast, states where the variance is low are those in which the agent is certain of the outcome of the task, and there are fewer ways in which external assistance can be of help. Most RL algorithms work by estimating the expected return from a given state; this estimate is continuously refined during training time. We modify this training procedure also to maintain and refine an estimate of _π(R), the variance of the returns under the learned policy, as an indicator of the agent's uncertainty. At test time, our agent can use this estimate of variance to decide when to request help. We detail these processes next. §.§ Estimation for the variance of return The expected return from a given state is encapsulated by the state value functions, which are explicitly relied on by most RL algorithms. However, while most RL algorithms are not concerned with the variance of the return, we would like to compute and maintain similar estimates for it during training. By definition, the variance of the return is defined by: (R) = 𝔼[R^2] - 𝔼[R]^2 The expected return of taking action a_t in state s_t is given by the Q-function under policy π: 𝔼[R] = Q^π(s_t, a_t) Since the transition function is not available, it is usually approximated via Bellman update: Q^π(s_t, a_t) ← (1 - α) Q^π(s_t, a_t) + α(r(s_t, a_t) + γmax_a_t+1∈ AQ(a_t+1, s_t+1)) The second moment of return 𝔼[R^2], which we call M^π(s_t, a_t), can be approximated using a sampling-based method by: M^π (s_t, a_t) = 𝔼[R^2] = 𝔼[[r(s_t, a_t) + γ∑_k=t+1^k=Nr(s_k, a_k)]^2] = 𝔼[r^2(s_t, a_t) + 2γ∑_k=t^k=Nr(s_k, a_k) + (∑_k=t+1^k=Nr(s_k, a_k))^2] = r^2(s_t, a_t) + 2γ∑_s_t+1∈ S p(s_t+1|s_t, a_t) Q^π(s_t+1, a_t+1) + γ^2∑_s_t+1∈ S p(s_t+1|s_t, a_t) M^π(s_t+1, a_t+1) Since the transition function is not available for a model-free agent, we can estimate M^π using a Bellman-like formula with sampled data: M^π(s_t, a_t) ← (1 - α)M^π(s_t, a_t) + α M'(s_t, a_t) where M'(s_t, a_t) = r^2(s_t, a_t) + 2γ Q^π(s_t+1, a_t+1) + γ^2 M(s_t+1, a_t+1) Here, a_t+1 = _a_t+1∈ AQ(s_t+1, a_t+1), since we assume a greedy agent that will take the actions that lead to max expected return during deployment. Finally, combining M and Q, we can approximate the variance of returns in a state by: (R) = M(s_t, a_t) - Q^2(s_t, a_t) Hence, the variance of the return given a state and action pair only depends on the Q function as well as the M (second function), which in turn depends on Q and can be updated via a Bellman-like recursion. §.§ HULA: Complete method We are now ready to integrate the variance estimation into a complete learning algorithm. We use algorithms from the Q-learning family, as these already build an explicit estimate of the Q function at train time, and also use a greedy policy w.r.t. Q at run-time. During training, in addition to the function estimator for the Q function (referred to as Q_θ_1), we build and update an estimator for M (referred to as Q_θ_2). This estimator can be of the same type as used for Q: for tabular Q-learning, we can use a table, while for Deep Q-Networks (DQN) <cit.> we use a deep neural network. At every iteration, we use Eqs. (<ref>-<ref>) to update the estimator for M. In the case of tabular Q-learning, Eq. (<ref>) can be used directly, while in the case of DQN it can be transformed in a loss function analogous to the one used for Q. The complete procedure is shown in Alg. <ref>. Finally, during deployment, the agent can use the trained estimators for the Q and M functions to compute the variance of the return for the current state. If this variance exceeds a threshold, the agent will request assistance and execute the action provided by the expert. Otherwise, the agent will execute the action prescribed by its own policy (in this case, a greedy selection based on the trained Q-function). We note a key feature of our method: it does not require the presence of an expert during train time. The only change at train time compared to traditional, expert-less Q-learning is the addition of the function estimator for M, which will be used at run-time to help provide a variance estimator. In turn, the variance estimator is used to decide when to request assistance from the expert, which is only needed at deployment. § EVALUATION §.§ Environments We evaluate HULA in a discrete navigation scenario, where the agent must reach a goal while avoiding obstacles and/or traps. Furthermore, we test our approach both on problems where the agent has exact knowledge of its current state (which is tractable via tabular Q-learning) and problems where the agent only has access to sensor data of limited range, creating ambiguity (which requires more powerful function estimators, and the use of DQN). §.§.§ Fully-observable MDPs These are discrete navigation environments where the agent is provided its exact location as observation. In addition to the goal cell, the environment also contains obstacles and traps: colliding with a trap would terminate the episode, while colliding with an obstacle results in a failure of action, and the agent would stay in its original cell. The agent also receives a small penalty for each step it takes. Available actions for an agent are moving up, down, left, and right, and the observations are the coordinates (x, y) of the agent's location. We use the maps shown in Fig. <ref>, and train and test individually on each map. For this simple class of problems, an unsophisticated agent can easily learn to act optimally on its own. However, we introduce uncertainty in the form of a stochastic transition function: at every step, after selecting an action, the agent moves in the desired direction with probability ψ, and moves in a random direction with probability 1-ψ (in practice, we use ψ=0.45). We think of this setting as a "slippery world". Thus, even with an optimal policy, the robot may not reach its goal. To help, we introduce an expert: when called, the expert provides the optimal action in the direction of the goal, and the action provided by the expert is not subject to transition function stochasticity. In this setting, the expert can thus be thought of as an “action corrector” with a better understanding of the environment dynamics, and whose actions always produce the expected results. Intuitively, we would expect an agent to call the expert in high-risk situations where it needs to avoid a wrong move, such as very close to traps. In this experiment, we test on two maps: 1. trap world that requires an agent to navigate around traps with an expert; 2. shortcut grid where the agent can leverage the expert to take the shortcut for higher returns. §.§.§ Partially-observable MDP's To test how our algorithm performs in more complex environments, we extend to a case of partially observable MDP's where the agent does not have complete information about the system. Instead of observing state information as the (x, y) position in the grid, the observation now only includes a finite patch of grid cells around the agent's current location (we use 5x5 grids in our implementation). As shown in Figure <ref>, due to partial observability, it is not always possible for the agent to uniquely identify its own location in the map from observations. In such ambiguous regions, the same action taken based on identical observations can lead to utterly different results. However, other parts of the map are uniquely identifiable based on sensor observations, and thus an autonomous agent can always select the optimal action. In this experiment, the expert observes the full state and always provides optimal actions to the agent. Intuitively, we would expect the agent to make use of the expert when traversing ambiguous areas of the map. §.§ Evaluation Approach The main evaluation metric for our approach focuses on its ability to make effective use of the expert: we would like to see how performance (measured as average episodic return) changes as a function of the number of expert calls made during deployment. We can measure this performance by varying the value of the variance threshold ϵ used in Alg. <ref>: for large values of ϵ, the agent will never make use of the expert; conversely, if ϵ is very small, the agent will call the expert in every state. By sweeping the value of ϵ between these extremes, we obtain more or less autonomous agents, and can plot performance as a function of the number of expert calls that result in each case. As a baseline, we use a standard RL approach that simply integrates the expert into the training procedure. We refer to this baseline as the penalty-based agent. Specifically, the penalty-based agent treats calling expert as an extra action a_call that can be called alongside the other actions at both training and deployment time. However, since the expert is optimal, the penalty-based agent will learn to always call for help. To learn a nontrivial policy that does not call the expert excessively, we employ a penalty of calling an expert in the reward function at train time: r'(s, a) = r(s, a) + c, where c < 0 is a penalty assessed only when calling an expert. By training this method with varying values of c, we again obtain agents that are more or less autonomous: the penalty-based agents trained with high c will call the expert less often, while those trained will low c will call more often. Again, we sweep the value of c between these extremes, and plot performance as a function of the number of expert calls that result in each case. However, for a fair comparison against HULA, we do not assess the expert penalty at deployment time. § RESULTS §.§ Uncertainty Estimation The first question we would like to answer is whether successfully captures the uncertainty of the agent and allocate the expert calls efficiently. We compare the learned variances by our method with the ground truth variance of returns by Monte Carlo sampling the environment with the same policy without the assistance of an expert. As shown in Figure <ref>, in the fully-observable experiment, successfully estimates a variance map that is similar to the ground truth. Specifically, it is uncertain about its outcome in states close to traps and confident when it is far from traps. In the partially observable case, identifies ambiguous states when the agent cannot localize itself in the map. In these states, the agent fails to produce an action faithfully in that the same action taken can lead to different results (e.g., hitting the boundary or reaching the goal). Our results align with the ground-truth variance map, which also has high uncertainty in a similar region. To further investigate 's efficiency in expert allocation, we compare the states with the highest N variances with those in the ground truth variance map. These states are usually where the agent calls the expert, especially when the agent has a budget of N expert calls. Our results show that can estimate the uncertainty of an RL agent accurately and allocate its expert call based on the return variance efficiently. As shown in Table <ref>, in trap world, 80% of the top-5 and top-10 variance states estimated by are also the top-5 and 10 accordingly in the ground-truth variance. In the shortcut world, although our method only correctly estimates 1 of 5 highest variance states, its performance is higher and reaches 60% in the top-10 case. This indicates that it allocates expert calls efficiently if given more budget. In the partially-observable experiment, our agent also recovers most of the high variance states in both the top-5 and top-10 evaluation. §.§ Task Performance In all experiments, our results show that agents achieve higher performance when requesting help from experts compared to the ones that does not use an expert. In the trap world, as shown in Fig. <ref>, achieves similar performance as the penalty-based agent while using the same number of expert calls. Interestingly, in the shortcut world, we find that the penalty-based method has a higher average return than . This is because our method does not evaluate the effect of calling an expert during training. As shown in Fig. <ref>, the penalty-based agent learns to move towards states that ask for help from an expert since it knows it will call an expert in a future state, whereas navigates around the uncertain states to avoid them. However, in the highest-variance states (e.g., states between traps), both agents are able to call the expert and complete the task. In this experiment, we find that is robust to different expert call budgets compared to the penalty-based agent. For example, in the shortcut world, the penalty-based agent cannot learn a policy that uses 2 expert calls to complete the task even if we perform hyper-parameter sweeping with the expert penalty c. This can be caused by the high expert penalty that modifies the original reward structure and hinders the learning of the task. In contrast, our method can flexibly incorporate different expert call budgets and maintain reasonable performance. In the partially-observed grid world, the agent outperforms the penalty-based agent when the allowed expert call is limited in the range of (2, 8]. When both agents are given enough expert assistance, they achieve similar performance. This indicates that efficiently utilizes the expert to localize itself and improve the task performance. An important feature of our method is that it does not introduce extra complexity in training an RL agent. Practically, we stop training when the Q function converges. Therefore, the training efficiency is similar to its underlying RL algorithm. In contrast, the penalty-based agent requires access to an expert during training time. Our experiments show that training each penalty-based agent for the fully-observable environment results in about 70000 expert calls. The requirement of an expert during train time limits its application on learning policies that interact with humans. § CONCLUSION We introduced , a method to learn human-in-the-loop policies by estimating an RL agent's uncertainty. We proposed a return variance estimation method that captures an RL agent's uncertainty. Our experimental results demonstrate that can capture an RL agent's uncertainty and use it to request assistance from experts to achieve high task performance. An important feature of our method is that it does not require the presence of an expert during training. We envision that our approach can be applied to more complex problems. For example, we plan to extend to continuous RL algorithms (e.g. DDPG <cit.>) and learns HitL policies to solve continuous control problems. We also would like to explore the direction of learning uncertain agents in other domains, such as language-guided navigation. -12cm

http://arxiv.org/abs/2303.07187v1

20230313152920

SOBO: A Feedback Bot to Nudge Code Quality in Programming Courses

cs.SE

Software Engineering Education KTH Royal Institute of Technology KTH Royal Institute of Technology RelationalAI KTH Royal Institute of Technology Software Engineering EducationSoftware Engineering Education -1 Recent research has shown the great potential of automatic feedback in education. This paper presents SOBO, a bot we designed to automatically provide feedback on code quality to undergraduate students. SOBO has been deployed in a course at the KTH Royal Institute of Technology in Sweden with 130+ students. Overall, SOBO has analyzed 1687 GitHub repositories and produced 8443 tailored code quality feedback messages to students. The quantitative and qualitative results indicate that SOBO effectively nudges students into adopting code quality best practices without interfering with pedagogical objectives or adding a teaching burden. From this experience, we provide guidelines into how to design and deploy teaching bots in programming courses. SOBO: A Feedback Bot to Nudge Code Quality in Programming Courses Martin Monperrus March 30, 2023 ================================================================= § INTRODUCTION It is generally accepted that code quality should be taught early and often throughout a computer science (CS) curriculum. Yet, Jansen et al. <cit.> showed that feedback related to code quality tends to be delivered far too late to have any meaningful impact, typically at the very end of an assignment and beyond any student interest. Furthermore, Östlund et al. <cit.> demonstrated in a longitudinal study that code quality is not seriously considered by students. However, it is unreasonable to expect busy teachers and teaching assistants to devote significant time delivering targeted feedback on code quality to every student when resources are already stretched. Whilst tools exist to discover code quality violations, it is challenging to integrate such professional tools into a student working environment without creating confusion. Our key insight is to encapsulate such a code quality tool into a friendly bot that automates the production of tailored, helpful feedback on code quality. In this paper, we present SOBO, an automatic feedback tool, whose aim is to nudge students towards understanding and applying code quality best practices early in their software engineering journey. § NUDGING AND AUTOMATIC FEEDBACK At the highest level, SOBO brings together the ideas of nudging and automation in education. Nudging is a concept that has emerged from behavioral economics. We adopt the definition by Thaler and Sunstein <cit.> of nudging being “any aspect of the choice architecture that alters people's behavior in a predictable way”. In their review on nudging in education, Dammgaard and Nielcen <cit.> show that nudge techniques can gently guide students, parents, and teachers toward better educational decisions and attainment. In light of this, we want students to be more aware of code quality - by communicating violations, their meaning and solution - and we hope to alter their behavior towards code quality in a positive way. Automation in education has become increasingly important to manage ever growing enrollments. Within computer science, introductory courses such as CS1 (programming) and CS2 (algorithms and datastructures) are very popular and as such must find ways to manage large numbers of students, whilst not compromising on quality. For example, Paiva et al. <cit.> finds that manual assessment is a bottleneck task for instructors. As a potential solution, automatic feedback allows for instant and detailed feedback on student attempts without ever overloading the instructor - a rare win-win situation. In the last decade, several projects related to automatic feedback have been implemented. Studies like Jansen et al. <cit.> show that it is a hard task to extract meaningful statistics from experiments related to the usage and impact of automation bots. Haug, Markus et al. <cit.> address one common challenge among feedback tools which is the fact that the content of the feedback itself can be ignored by the user if the message is not clear enough. To solve the problem, they propose to add code examples of how to fix the error in the feedback. A recent important study <cit.> highlights the importance of clear and concise information in automated feedback. Another common challenge among automatic feedback tools is how not to give the direct answer to the task to the students. Zhang et al. <cit.> and Liu et al. <cit.> have dealt with this problem: they categorize 1) bots that provide feedback directly related to the grades and 2) bots that nudge better techniques but do not affect the student's grades. Finally, research on code quality in CS1 and CS2 courses highlight two major challenges: (1) the amount of topics to cover, and (2) the assessment of such topics <cit.>. As Jansen et al. observes, most of the educational effort is spent on the aspects of learning the fundamentals of programming, new languages, and development environments. All this research points to 1) automating the code quality feedback and 2) nudging students towards addressing it without any explicit grading demand. This is what our new bot, called SOBO, does. § SOBO DESIGN Motivated by the research in nudging and automation, we next present the design for SOBO by describing the teaching context in which it is deployed, its workflow, the choice of code quality rules, the feedback design, and interaction with the students. §.§ Teaching Context We design SOBO for a CS1 and CS2 course at KTH Royal Institute of Technology, Stockholm, Sweden. The course is part of the first-year program of the Computer Science (CS) degree and the topics covered include both an introduction to programming in Java (CS1) , as well as an introduction to algorithms and data structures (CS2). The course's workflow is as follows: * At the beginning of the year, the students receive instructions about the course stack and they immediately learn about Git/GitHub as a version control and collaboration technology. * Every week, the students receive a unique repository per student group, containing all the relevant task information. They are free to commit/push as often as they like prior to the deadline. * At the end of each week, they have a session with a Teaching Assistant (TA) where they present. * After each session, TAs grade the students' submissions and students have the opportunity to fix issues in case they did not fulfill the requirements. TAs use the GitHub issue tracker to communicate with the students. Per task, the grades can be PASS, FAIL, or KOMPLËTERING when the last one means there are a few mistakes to be corrected. The course has some features that are relevant to the purpose of this work: Consistent course information in one stack. The whole course happens on GitHub: students receive repositories for the weekly tasks and commit/push their solutions from the beginning to the end. The GitHub issue tracker is the main channel of communication for the whole course. Should they talk to a bot, the students would naturally do it via an issue tracker. Access to historical data. The course has been delivered using the same workflow and stack since 2015. Therefore, there is access to data of each previous offering (200 students working in 4000 task repositories over seven years). Previous research has analyzed the data between 2018 and 2022 showing that students are not taking enough action in terms of code quality in their programming assignments <cit.>. §.§ Bot Workflow As summarized in Figure <ref> SOBO's automatic feedback workflow operates as follows: * When a new programming assignment is announced, a weekly repository is created for each student group and added to SOBO's monitoring list. * Then, SOBO opens a GitHub issue to push feedback messages to come. The first post is a friendly greeting text[<https://github.com/SOBO-bot/templates/blob/main/files/greeting.md>]. * Every time a student produces commits, SOBO performs the following task: * It uses a state-of-the-art code quality tool called SonarQube to list the violations in the students' code, incl. the commit hash with the file where the violation is found, the rule violated, and the line of code in the respective file. * Then, it filters each row to confirm the violation comes from a student and not from the provided template <cit.>. * It selects the most prevalent violation types from the last commit in order to give feedback on one single code quality aspect (we cannot flood the students, it would be counter-productive). * SOBO collects the template for the selected rule and fills it up with a table indicating the location of the violation(s) and pushes the feedback to the corresponding issue. Figure <ref> gives an example of this feedback message. * If there are no violations at all, SOBO posts an encouraging message indicating the submission had no violations. This design is per Wessel et al. <cit.>, SOBO does not need any configuration from the students. Also, in case a student does not want to receive feedback, we provide a simple opt-out option to stop receiving automated feedback. §.§ Code Quality Feedback §.§.§ Violation Reports. SonarSource for Java <cit.> is used to analyze students' code quality. This tool indicates the presence of code quality problems in code. §.§.§ Curation of Rules. To select the relevant code quality rules for students, we categorized all SonarSource rules by using the following criteria: * Actionable. A clear example of how to fix a problem can be extracted from the rule. * Knowledge gap. The information is understandable by undergraduate students in an introductory programming course. * Actual problem. The prevalence of violations from previous course iterations is used as a relevance indicator (see Östlund, et al. <cit.> for details). The curating process was done in collaboration with TAs the course responsible. All rules currently supported by SOBO are described in Table <ref>. §.§ Feedback Design §.§.§ Relevance of Rules. As a nudge tool, SOBO must filter the information in order to successfully alter the student behavior and it is not perceived as a negative element in the course <cit.>. Following Guideline #2 from Wessel et al. <cit.>, the information needs to be centralized to avoid overwhelming the user. For, SOBO focuses on the most common rule, as the rule that presents the largest amount of violations in a commit, the rule that requires most attention by students. Each automatic feedback provides information related to the most common rule only. §.§.§ Feedback Template. Wessel et al. <cit.> mentions “If [feedback] is not actionable, it is not usable”. We make sure that the feedback is clear and self-contained so that there is no need for extra search needed to understand it. Inspired by Paiva et al. <cit.>, SOBO messages present knowledge following a template with 1) a presentation of the rule 2) the actual problem in the student's code, 3) and a code example that is actionable. An example is shown in Fig <ref> §.§.§ Frequency of Feedback. Students require timely feedback during the process of their assignments <cit.>. Since we want students to have feedback as soon as they push to the repository, SOBO analyzes students' code every 5 seconds. §.§.§ Distribution Channel. SOBO gives feedback by posting on a dedicated GitHub issue, named “SOBO - Commit Analyzer”. This has two advantages: * students always know where to find the analysis of the latest version of the code; * they can activate or deactivate notifications for a particular task by `unsubscribing' to the issue; §.§ Command Language Following recommendation #2 from Wessel et al. <cit.>, we bake in SOBO a direct communication channel with the students. A second GitHub issue is created to interact with the teaching bot. In this issue, the students receive a help message with a list of commands to use and interact with SOBO. To send a command, the students write a comment on this issue. The available commands are: * : to get more information about the available commands. * : to stop all automatic feedback messages on the repository where the command is executed. * : to restart SOBO on the repository where the command is executed. * : to receive information about a specific commit on the repository where the command is executed. * : to get all violations of a specific rule among the ones used in the project on the repository's latest version. * : to return synthetic data simulating the result of a code injection into the SOBO database. §.§ Data Collection Automatic Message. To keep track of the influence of SOBO, each time a commit is analyzed, a tuple is stored for each violation made by the student. To identify violations the key of the tuple is the GitHub user, the assignment number, rule, file, line of code, and the commit hash. Command Language. For the command language, every command sent by the student and resolved by SOBO is stored for further analysis of user interaction. The key pair in this case is the student, the timestamp, and the task. §.§ Implementation Technical implementation. SOBO is implemented in Java and runs as a daemon on a university server. Platform support. SOBO assumes that students' repositories are hosted on GitHub, and it supports both public GitHub and enterprise GitHub (KTH uses the latter). SOBO has its own identity, hence its own KTH GitHub account to interact with the students. Configuration The main configuration of SOBO is a file with a list of student repositories and a file of feedback templates. Per the course workflow, each week, new repositories are added to the bot monitoring list for the new assignments (one assignment = one repository). SOBO is publicly available as open-source and can be used by other universities: <https://github.com/eclipse/repairnator>. § EXPERIMENTS & RESULTS SOBO is an end-to-end project, we have deployed it in class at KTH with 130+ undergraduate students. Below are the results. §.§ Preparation Before deploying SOBO, we waited for the students to have already been taught about fields, classes, and collections. This means the bot was actually deployed in the 6th week of the course. The week the bot started operating, a presentation was made in the classroom and an informative email was sent, to explain the project, how it will benefit and affect them, and how to turn it off in case they do not want to use it. §.§ Measuring Effectiveness We want to study the number of code quality violations in students' code, for each user, rule, and repository. For each repository, we split the number of violations per rule. To track the evolution of each user, we sort the violations by timestamp and compare the amount between two consecutive commits. If the number of violations for the same rule is reduced, then some have been fixed <cit.>. We assume that the violations have been fixed thanks to SOBO's actions. In the opposite scenario, if the number increases for the same rule between two commits, we consider that new violations have been added. §.§ Capturing Students' Opinions Following the guidelines of Jansen et al. <cit.>, we made a survey to study the users' reception of the bot. This survey was made and sent 10 weeks after the SOBO integration on the course. The survey was divided into four parts starting with an evaluation of the automatic feedback message, the effectiveness of the provided information, the use of the command language, and the general perception of teaching and/or automatic bots. The survey questions are: * How clear was the explanation of the rule? * How concise was the explanation of the rule? * How clear was the example of what to do? * How concise was the example of what to do? * The explanation section helps me to understand how to fix the violation. * The example section helps me to understand how to fix the violation. * I would like to have SOBO integrated on my personal projects. * I would like to have teaching bots in my other courses §.§ Results Quantitative. Figure <ref> shows the number of added and removed code quality violations overall student groups for the three most prevalent rules (S109, S1155, S1481). On the negative side, students add more violations than they remove. This means that the nudging is not perfectly effective, which is as expected because it does not take part in the mandatory graded objectives. On the positive side, we do see 513 code quality violations fix by students. We cannot affirm that this is causally related to SOBO's feedback messages but we have good reasons to believe so. First, there are 19 commits messages showing mentioning that the changes were made thanks to SOBO. Second, the qualitative survey provides solid evidence. Qualitative. From a total of 131 students, we got 57 answers, reaching 43% of the users. The results of the survey are presented in Figure <ref>. The student's responses on the survey show that the automatic feedback was clear (Q1, Q3) and contained enough information for them to fix the violation (Q5, Q6). The survey also shows that the template is probably too verbose (Q2, Q4). A future iteration of the course will fix that aspect. In the free comments, students mentioned a lot of messages were related to rule S109 (“Magic Number”). We believe that this is due to the fact that it is the rule SOBO reports the most. For this rule in particular, students found the “Instead of doing this” part of the template to be very explicit, and that the whole example section allows them to: “learn how to make the code better”. §.§ Impact of the Bot on Pedagogical Objectives The bot did not intrude in the main course workflow. Whilst some initial classroom time was used to explain the presence and purpose of SOBO to students, little further effort was required. Neither the course responsible nor teaching assistants reported extra negative burdens related to SOBO and could focus on the normal delivery of the course content. To that extent, the code quality aspect of the course was entirely delegated to the SOBO bot. Consequently, we envision expanding its scope to more aspects of code quality and supporting more pedagogically relevant rules. § GUIDELINES FOR BOTS IN EDUCATION Based on our experience, here are important guidelines for educators who would like to integrate nudging bots into their courses. Focus on the Message: Previous work has mentioned that the key to automatic feedback in education lies in the feedback message that will be delivered <cit.>. It is essential not to overestimate the knowledge of the students. SOBO messages were designed to be self-contained and present all the information needed in order to solve the code quality problem. Iterate over the feedback design with focus groups gathering both professors, teaching assistants, and students. Delivering the Message: When selecting the distribution channel for the feedback, it is essential to be very clear with the students about the location of the feedback: which platform? which link? In our case, we had one single channel: a specific SOBO GitHub issue, one per student repository. Avoid changes of applications and channels to deliver automated feedback. Respect the Context: We suggest deploying a teaching bot with careful consideration of the teaching context. This concerns the workflow of the course(s): when is the bot introduced? for which objectives? This also concerns the platform: it is better to be integrated into the platform already being used than introduce an additional one. Spend dedicated time explaining the bot's intention and scope in the classroom. Lower the Configuration Barrier: Ideally, SOBO can be reused by any other course with only a small configuration at the beginning of the course. Extremely custom tools are hard to deploy in new environments, whereas good configurability improves the scalability of automated feedback tools <cit.>. Adopt software engineering reuse best practices when implementing your own nudging bot. § CONCLUSION & FUTURE WORK It has been shown that there is a lack of feedback related to code quality in CS1 & CS2 courses. Leaving code quality unaddressed simply delays the cost, which manifests itself as technical debt in future projects. In this paper, we have presented SOBO, a nudging bot that helps students to understand and adopt good code quality practices by generating feedback at the right place and time for maximum impact. In future iterations, SOBO will track previous student submissions in order to personalize the provided feedback. By tracking previous attempts, the feedback can mention specific ways to help the problem faced by the student. Also, since not all CS1 & CS2 courses are taught using Java, an interesting community effort is to extend SOBO to other programming languages (eg. Python), increasing the impact and relevance of the whole nudging bot project. References plain < g r a p h i c s > Sofia Bobadilla is a research engineer at KTH Royal Institute of Technology, Stockholm, Sweden. Her research interest includes bots in software development, software reliability, data science, and multimedia information retrieval. Bobadilla obtained her bachelor's degree in Computer Science from the University of Chile, Santiago, Chile. Contact her at sofbob@kth.se. < g r a p h i c s > Richard Glassey is a teacher at KTH Royal Institute of Technology, Stockholm, Sweden. His current research interests include computer science education, learning analytics, and sustainable education. Glassey received his Ph.D. in computer science from the University of Strathclyde, Glasgow, Scotland. He is a member of ACM. Contact him at glassey@kth.se. < g r a p h i c s > Alexandre Bergel is a computer scientist at RelationalAI, Switzerland. His current researches are on designing tools and methodologies to improve the overall performance and internal quality of software systems and databases by employing profiling, visualization, and artificial intelligence techniques. Contact him at alexandre.bergel@me.com. < g r a p h i c s > Martin Monperrus is Professor of Software Technology at KTH Royal Institute of Technology. His research lies in the field of software engineering with a current focus on automatic program repair, AI on code and program hardening. He received a Ph.D. from the University of Rennes, and a Master's degree from Compiègne University of Technology. Contact him at monperrus@kth.se.

http://arxiv.org/abs/2303.06891v1

20230313063910

Optimality of the Decay Estimate of Solutions to the Linearised Curl-Free Compressible Navier-Stokes Equations

math.AP

1.1 textwidth=16cm, textheight=21.3cm, marginratio=1:1,5:7 proof mystyle mystyle plain [1] definition thmTheorem[section] cor[thm]Corollary lem[thm]Lemma rem[thm]Remark *rem*Remark *propnProposition *krKnown Results definitionDefinition *propositionProposition prop[thm]Proposition *objObjective *mainpropMain Proposition part..5em part claimClaim[section] stepStep[section] equationsection thmsection equ equ* part..5empartproofpartPart ||| ‖‖“"()()()()[][][] #1 3pt 3pt =16cm #1 3pt 3pt Optimality of the Decay Estimate of Solutions to the Linearised Curl-Free Compressible Navier-Stokes Equations Tsukasa Iwabuchi* DáithíÓ hAodha** Mathematical Institute, Tohoku University, 980-0845 Email: *t-iwabuchi@tohoku.ac.jp, **david.declan.hughes.p6@dc.tohoku.ac.jp 135mm Abstract. We discuss optimal estimates of solutions to the compressible Navier-Stokes equations in Besov norms. In particular, we consider the estimate of the curl-free part of the solution to the linearised equations, in the homogeneous case. We prove that our estimate is optimal in the L^∞-norm by showing that the norm is bounded from below by the same decay rate. § INTRODUCTION This paper is concerned with the barotropic compressible Navier-Stokes system _t ρ + ÷ (ρ u) = 0, in (0,∞) ×^3, _t (ρ u) + ÷ (ρ u ⊗ u) - ÷(2 μ D(u) + λ÷ (u) Id) + ∇ p = 0, in (0,∞) ×^3, (ρ, u) |_t=0 = (ρ_0, u_0), in ^3, where ρ : [0,∞) ×^3 → [0,∞), and u: [0,∞) ×^3 →^3 are unknown functions, representing the density and velocity of a fluid, respectively. p: [0,∞) ×^3 → is the pressure in the fluid, and the barotropic assumption gives us p p(ρ), for some smooth function p(·). μ,λ are viscosity coefficients, taken such that μ > 0, 2μ + λ > 0. We define the deformation tensor D(u) 1/2( Du + Du^T ). In this paper, we examine the large-time behaviour, particularly time-decay estimates, of solutions to a linearised version of the above problem. Matsumura-Nishida showed in matsumura-nishida1979, matsumura-nishida1980 that (<ref>) has global solutions when equipped with data (ρ_0, u_0) that is a small perturbation in L^1 ∩ H^3 of (ρ̅, 0) for any positive constant ρ̅, and proved the following decay result [ ρ(t) - ρ̅; u(t) ]_2≤ C (1+t)^-3/4. This is the decay rate of the solution to the heat equation with initial data in L^1. Ponce then extended these results to other L^p norms. In particular, for p ∈ [2,∞], k∈{0,1,2}, and dimension d=2,3, ∇^k [ ρ(t) - ρ̅; u(t) ]_p≤ C (1+t)^-d/2(1-1/p) - k/2. Our results contribute to the theory of Compressible Navier-Stokes in Besov spaces. Global existence of strong solutions to (<ref>) for initial data (ρ_0, u_0) in critical Besov spaces Ḃ^d/2_2,1×Ḃ^d/2 - 1_2,1 was first proven by Danchin in <cit.> and large-time estimates in Besov norms for p close to 2 were proven by Danchin-Xu in <cit.>. Our goal in this paper is to obtain an optimal bound of the solution to a linearised version of the system (<ref>) which isolates the curl-free part of the velocity. In particular, we prove a bound from above, in terms of time t, of the norm of solutions over space x. We then prove that said bound is optimal (sharp) via a corresponding bound from below. Our proof is entirely self-contained. The bound from below in our main result, Theorem <ref>, is original, and is the first ever proof that the decay rate in L^∞ is optimal, or sharp. The estimates from above are analogous to those found in hoff-zumbrun, kobayashi-shibata. Indeed, in <cit.>, the following estimate is obtained. (<cit.>) Let mρ u, m_0 ρ_0 u_0. Assume that E [ a_0; m_0 ]_1 + [ a_0; m_0 ]_H^ 1 + l is sufficiently small, where l≥3 is an integer. Then the Navier-Stokes system (<ref>) with initial data ρ_0, u_0 has a global solution satisfying the following decay estimate for all for any multi-index with || ≤ (l-3)/2: D^_x [ a(t); m(t) - e^tΔ m_0 ]_p ≤ C(l) E (1+t)^ - 3/2 (1-1/p) - 1/2 ( 1 - 2/p ) - ||/2. Note that, in the norm in the above inequality, we are removing the divergence-free part of the linear term of m. Thus what remain are the nonlinear term and the curl-free part of the linear term. Kobayashi-Shibata in <cit.> also obtain the above decay rate. In fact, they obtained a refined estimate that requires less regularity on the initial data and separates the solution into high and low frequencies (see Definition <ref> below). The above decay rate found in <cit.> is associated with the low-frequency part of solutions, while the high-frequency part decays exponentially with t. In what follows, we shall assume that the density approaches 1 at infinity; and so we are concerned with strong solutions which are small perturbations from a constant state (ρ,u) ≡ (1,0). We shall also assume that μ, λ are constant, and set a ρ - 1. Our system (<ref>) can thus be rewritten into the following linearised problem: _t a + ÷(u) = f in (0,∞) ×^3, _t u - μΔ u - (λ + μ) ∇÷(u) + P'(1)∇ a = g in (0,∞) ×^3, (a,u)|_t=0 = (a_0,u_0) in ^3, for some functions f,g. We apply the orthogonal projections 𝒫 and 𝒬 onto the divergence and curl-free fields, respectively. Then, setting P'(1) and νλ + 2μ, we get the system _t a + ÷ (𝒬 u) = f in _> 0×^3, _t u - νΔ u + ∇ a = g in _> 0×^3, _t u - μΔ u = g in _> 0×^3. We set v |D|^-1÷ ( u), where |D|^s u ^-1[ |ξ|^s û], s∈. We note that one can obtain v from u by a Fourier multiplier of homogeneous degree zero. Thus, bounding v is equivalent to bounding u (see <cit.>, Lemma 2.2). We note that we can set = ν = 1, without loss of generality, since the following rescaling a(t,x) = ã/ν t, √()/ν x , u(t,x) = √() ũ/ν t , √()/ν x ensures that (ã, ṽ) solves (<ref>) with = ν = 1. Thus we get that (a, v) solves the following system: _t a + |D| v = f in _> 0×^3, _t v - Δ v - |D| a = h |D|^-1÷ (g) in _> 0×^3. In this paper, we shall focus on the homogeneous case, where f = h = 0. Before we give our main result, we introduce the concepts of Besov spaces, and high-frequency and low-frequency norms. We denote by P(^3) the set of all polynomials over ^3. We use the Littlewood-Paley decomposition of unity to define homogeneous Besov spaces. Let {ϕ̂_j }_ j ∈ be a set of non-negative measurable functions such that * ∑_ j ∈ϕ̂_j () = 1, for all ∈^3 \{0}, * ϕ̂_j () = ϕ̂_0(2^-j), * ϕ̂_j () ⊆{∈^3 | 2^j-1≤ || ≤ 2^j+1}. For f ∈𝒮'(^3)/P(^3), we write Δ̇_j f ^-1 [ϕ̂_j f̂]. The Besov norm is then defined as follows: for 1 ≤ p,q ≤∞, and s ∈, we define f_Ḃ^s_p,q∑_j ∈ 2^sqjΔ̇_j f^q_p^1/q. The set Ḃ^s_p,q is defined as the set of functions, f ∈𝒮'(^3)/P(^3), whose Besov norm is finite. Finally, we also regularly use the following notation for so-called high and low-frequency norms: f^h_Ḃ^s_p,q∑_j ≥ 3 2^sqjΔ̇_j f^q_p^1/q, f^l_Ḃ^s_p,q∑_j ≤ 2 2^sqjΔ̇_j f^q_p^1/q. We now give our main result. We obtain the same decay rate as the previous proposition, and also prove that this decay rate is optimal in the L^∞-norm. Let s ∈, p ∈ [2,∞],q ∈ [1,∞], and t > 1. Then ^-1[ (t); (t) ]_Ḃ^s_p,q≤ C t^-3/2 (1-1/p) - 1/2 (1 - 2/p) ^-1[ _0; _0 ]^l_Ḃ^s_1,q + C e^-t^-1[ _0; _0 ]^h_Ḃ^s+3(1-1/p)_1,q. For the high-frequency norm, we also have for all t>0, ^-1[ (t); (t) ]^h_Ḃ^s_p,q≤ C e^-t^-1[ _0; _0 ]^h_Ḃ^s_p,q. Also, there exists initial data such that, for all sufficiently large t, ^-1[ (t); (t) ]_∞≥ C t^-2. The derivative loss in the high-frequency part of the L^1 – L^∞ estimate (<ref>) is in fact caused by only one of the two eigenvalues (derived in Section 3) associated with the fundamental solution of our problem. The other eigenvalue has strong enough smoothing properties that no additional derivatives are required when estimating its associated semigroup (see Proposition <ref>). Note, however, that we can avoid the need for more regular initial data by using (<ref>) instead for the high-frequency part of solutions. We offer some commentary on the above decay rate for the low frequency part. In Hoff-Zumbrun <cit.>, it is demonstrated that the curl-free part of the solution to (<ref>) is approximated by a convolution of the Green's function for a parabolic system (whence we obtain the heat-like decay) and a function which decays like the fundamental solution of a wave equation (whence the additional decay). To show that this decay rate is sharp in L^∞, we find that this wave-like part of the decay is slowest along |x| = t. This is a natural result. Indeed, looking at the homogeneous case f=h=0 for (<ref>), we may write ^2_t a - Δ a - |D|^3 v =0, ^2_t v - Δ v - _t Δ v =0. Note that the wave propagation speed above is 1. This is a consequence of our assumption made on the constant . § PRELIMINARIES (The Fourier Transform) For a function, f, we define the Fourier transform of f as follows: [f]() f̂() 1/(2π)^3/2 e^-i x· f(x) x. The inverse Fourier transform is then defined as ^-1[f̂](x) 1/(2π)^3/2 e^i x ·ξf̂() . For the purpose of calculating inequalities, we will frequently omit the factor of 1/(2π)^3/2. (Orthogonal Projections on the divergence and curl-free fields) The projection mapping is a matrix with each component defined as follows for i,j ∈{1, 2, 3}: ()_i,jδ_i,j + (-Δ)^-1_i _j. We then define 1 -. For f ∈ (Ḃ^s_p,q(^3))^3, with s ∈, and p, q ∈ [1,∞], we may write f (1 + (-Δ)^-1∇÷) f. § PROOF OF MAIN RESULT We begin our analysis of (a,v) which solves the system _t a + |D| v = 0 in (0,∞) ×^3, _t v - Δ v - |D| a = 0 in (0,∞) ×^3. Taking the Fourier transform over space x, we can write the above system as / t[ ; ] = M_||[ ; ] , with M_||[ 0 -||; || -||^2 ]. Then we may write the following formula for the solution to (<ref>): [ (t); (t) ] = e^t M_||[ _0; _0 ] . We obtain the following eigenvalues for M_||, which differ between high and low frequencies: λ_±() -||^2/21 ± i √(4/||^2 - 1 ), for || <2, -||^2/21 ±√( 1 - 4/||^2), for || >2, from which we obtain the following formulas for our solution: (t) = e^tλ_-λ_+ - e^tλ_+ λ_- /λ_+ - λ_- _0 + e^tλ_- - e^tλ_+λ_+ /λ_+ - λ_- _0, (t) = e^tλ_+ - e^tλ_-λ_- /λ_+ - λ_- _0 + e^tλ_+λ_+ - e^tλ_-λ_- /λ_+ - λ_- _0. We note that, for high and low frequencies away from || = 2, the eigenvalues can be bounded above by a mere constant. However, both the numerator and the denominator approach 0 as || → 2. We thus must consider how to deal with the L^2 and L^∞ norms of e^tλ_-λ_+ - e^tλ_+ λ_- /λ_+ - λ_- _j and the other similar expressions in (<ref>) and (<ref>), for j ∈{ 0,1,2 }. For details, see Appendix <ref>. Otherwise, it suffices to prove time-decay estimates for the semigroup e^tλ_± without including the eigenvalues in the L^∞ norm for all j∈, and in the L^2 norm for all j ∉{0,1,2}. We will proceed with an interpolation argument after bounding the L^∞ and L^2 norms explicitly. The L^2 norm estimate is identical to that of the heat semigroup, as no additional decay is obtained from the oscillation. The L^∞ norm yields a faster decay, for which we now give the proof. (L^1 – L^∞ estimates for high and low frequencies) For all t > 1, ^-1[ e^ t λ_+ϕ̂_j ]_∞ ≤ C e^-t, for all j ≥ 3, ^-1[ e^ t λ_-ϕ̂_j ]_∞ ≤ C 2^3j e^-t , for all j ≥ 3, ^-1[ e^ t λ_±ϕ̂_j ]_∞ ≤ C t^-2, for all j≤ 2. As we mentioned in Remark <ref>, it is only the estimate of e^tλ_- in (<ref>) that causes a derivative loss. We start with (<ref>). We have j ≥ 3, and so _j ⊆{∈^3 | || > 2 }, and thus we are dealing with the high-frequency definitions of our eigenvalues λ_±. That is, we have ^-1[ e^ t λ_+_j ]_∞ = ^-1[ e^-t||^2/2( 1 + √( 1 - 4/||^2))ϕ̂_j ]_∞ = sup_x ∈^3| e^ix· e^- t||^2/2√( 1 - 4/||^2) e^-t||^2/2ϕ̂(2^-j) | ≤ C | e^-t||^2/2 (2^-j) | ≤ C e^-t∫_{ | || > 2 } e^-t||^2/4 ≤ C e^-t, for all j ≥ 3, t >1. For (<ref>), the proof is slightly different. Before we continue, note that -t||^2/2 1 - √(1 - 4/||^2) = -2t1 + √(1 - 4/||^2)^-1 = - t - 4t/||^21 +√(1 - 4 / ||^2)^-2. Thus, we may estimate as follows, ^-1[ e^-t||^2/2( 1 - √( 1 - 4/||^2))ϕ̂_j ]_∞ ≤ e^-2t 1 + √(1 - 4/||^2)^-1ϕ̂_j () ≤ C 2^3j e^-t . Now we consider (<ref>). For frequencies near ||=2, i.e. 0 ≤ j ≤ 2, an exponential decay follows similarly to the above. Finally, we are left with the low frequencies j ≤ -1. We start by extracting the heat-like decay by a change of variables. sup_x∈^3| e^ix· e^± it||^2/2√(4/||^2 - 1 ) e^-t||^2/2(2^-j) | = t^-3/2sup_x∈^3| e^i (t^-1/2x)· e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) |. Thus it remains to extract a further decay of t^-1/2 from the L^∞ norm above. We may estimate (<ref>) to obtain the desired decay for all sufficiently large t: sup_x∈^3| e^i (t^-1/2x)· e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) | ≤ C t^ - 1/2. We explicitly write the proof of the above in Appendix <ref>. Let s ∈, p ∈ [2,∞], and q ∈ [1,∞]. Then, for all t>1, ^-1[ e^t λ_+f̂]_Ḃ^s_p,q ≤ C t^-3/2 (1-1/p) - 1/2 (1 - 2/p) f ^l_Ḃ^s_1,q + C e^-t f ^h_Ḃ^s_1,q, ^-1[ e^t λ_-f̂]_Ḃ^s_p,q ≤ C t^-3/2 (1-1/p) - 1/2 (1 - 2/p) f ^l_Ḃ^s_1,q + C e^-t f ^h_Ḃ^s+3(1-1/p)_1,q. We start with the L^∞ norm. Note that, by Proposition <ref>, we have for all j ≥ 3 ^-1[ e^t λ_+_j f̂]_∞ ≤ C e^-tΔ̇_j f _1, ^-1[ e^t λ_-_j f̂]_∞ ≤ C 2^3j e^-tΔ̇_j f _1. Next, for j≤ 2, ^-1[ e^t λ_±_j f̂]_∞ ≤^-1[ e^t λ_±Φ̂_j ]_∞Δ̇_j f_1 ≤ C t^-3/2 - 1/2Δ̇_j f_1, where Φ̂_j _j-1 + _j + _j+1. Next the L^2 norm for j≤2 follows simply from Young's convolution inequality and heat kernel estimates: ^-1[ e^t λ_±_j f̂]_2 ≤^-1[ e^t λ_±Φ̂_j ]_2Δ̇_j f _1 ≤ C t^-3/4Δ̇_j f _1. For j ≥ 2, we get ^-1[ e^t λ_+_j f̂]_2 ≤^-1[ e^t λ_+Φ̂_j ]_2Δ̇_j f _1 ≤ C e^-tΔ̇_j f _1, ^-1[ e^t λ_-_j f̂]_2 ≤^-1[ e^t λ_-Φ̂_j ]_2Δ̇_j f _1 ≤ C 2^3/2 j e^-tΔ̇_j f _1. We use Hölder interpolation for 2<p<∞ to obtain L^p norm estimates. Multiplying both sides by 2^s and taking the ℓ^q norm completes the proof. (L^p – L^p estimates for high frequencies) There exists a constant C such that, for all p ∈ [1,∞] and t>0, ^-1[ e^tλ_±_j f̂]_p ≤ C e^-t f _p, for all j ≥ 3. Proof. First, we make use of the same equalities as in Proposition <ref>. That is, -t||^2/2 1 ±√(1 - 4/||^2) = -2t1 ∓√(1 - 4/||^2)^-1 = - t - 4t/||^21 ∓√(1 - 4 / ||^2)^-2. Thus, we can rewrite ^-1[ e^tλ_±_j f̂]_p = e^-t^-1[ e^ - 4t/||^2 1 ∓√(1 - 4/||^2)^-2_j f̂]_p ≤ e^-t^-1[ e^ - 4t/||^2 1 ∓√(1 - 4/||^2)^-2_j ]_1 f_p. We note that, for all j ≥ 3, ^-1[ e^ - 4t/||^2 1 + √(1 - 4/||^2)^-2_j ]_1 = ^-1[ e^ - 4t/ 2^2j ||^2 1 + √(1 - 4/ 2^2j ||^2)^-2_0 ]_1 ≤ C e^ - 4t/ 2^2j ||^2 1 + √(1 - 4/ 2^2j ||^2)^-2_0 _W^2,2 ≤ C. We now discuss the optimality of the estimates in the previous section. In particular, we will prove that the low-frequency bound obtained is optimal. First, we denote _t (_1, t^+1/4_2, t^+1/4_3), _t^-1 (_1, t^-1/4_2, t^-1/4_3). We also take a nonnegative nonzero function ∈ C^∞_0 such that ⊆{∈^3 | ||∈(1/2 , 1), |_1| ≥ 1/2 }, (-) = (), for all ∈^3. (Optimality of Low-Frequency Linear Estimate) There exists a constant C such that, for all t sufficiently large, ^-1 e^tλ_±(t^1/2_t) _∞≥ Ct^-2. We note that this bound from below on the low-frequency estimate is sufficient to prove that for all t sufficiently large, ^-1∑_j≤ 2_j e^t λ_±_∞≥ Ct^-2. Indeed, for all t≥ 1, we get by a simple application of Young's convolution inequality: ^-1 e^t λ_±(t^1/2_t) _∞ ≤^-1(t^1/2_t) _1 ^-1∑_j≤ 2_j e^t λ_±_∞ = ^-1() _1 ^-1∑_j≤ 2_j e^t λ_±_∞ ≤() _W^2,2^-1∑_j≤ 2_j e^t λ_±_∞ ≤ C ^-1∑_j≤ 2_j e^t λ_±_∞. First, note that sup_x∈^3| e^ix· e^± it ||^2/2√(4/||^2 - 1 ) e^-t||^2/2 ( t^1/2_t ) | ≥ t^-3/2| e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) | ≥ t^-3/2|∫_^3 ∩{_1 < 0 } e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) | - t^-3/2|∫_^3 ∩{_1 > 0 } e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) | , where in the above, we chose the specific value x = (± t,0,0), and have split the integral into two parts: one with _1 < 0, which we shall bound from below; and one with _1 > 0, which will decay at a slightly faster rate. We consider the _1 < 0 part. The key step is rewriting the exponent of our oscillation function. Note that in this case, _1 = -|_1|. i t^1/2_1 + ||√( 1 - ||^2/4t) = i t^1/2_1 + || - ||^3/ 4t + 4t √( 1-||^2/4t) = i t^1/2 -_1^2 + ||^2 /|_1| + || - ||^3/ 4t + 4t √( 1-||^2/4t). Then, by the substitution →_t^-1, we obtain |∫_^3 ∩{_1 < 0 } e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) | = t^-1/2|∫_^3 ∩{_1 < 0 } e^± i t^1/2 -_1^2 + |_t^-1|^2 /|_1| + |_t^-1| - |_t^-1|^3/ 4t + 4t √( 1-|_t^-1|^2/4t) e^-|_t^-1|^2/2 ( ) |, and thus, by the Lebesgue dominated convergence theorem, we have the following convergence as t→∞: t^+1/2|∫_^3 ∩{_1 < 0 } e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) | →|∫_^3 ∩{_1 < 0 } e^± i (_2^2 + _3^2) /2|_1| e^-_1^2/2 ( ) |, as t→∞, where the final integral is a positive constant, by the restriction to the support of . We thus conclude that there exists some constant C such that, for all sufficiently large t, |∫_^3 ∩{_1 < 0 } e^± i t^1/2_1 + ||√( 1 - ||^2/4t) e^-||^2/2 ( _t ) |≥ C t^-1/2. It remains to show that the _1 >0 part decays at a faster rate. The proof is similar to that of Proposition <ref> in the Appendix, and so we skip the details. The key difference is that _1>0 prevents the cancellation of _1 in the exponent which we exploited above for the _1 <0 case. In fact, in the _1>0 case, we may extract an extra decay of t^-1/4 compared to the normal bound from above proven in Proposition <ref>. (See also Remark <ref>). We lastly must find some suitable initial data (u_0, a_0) such that ^-1[ (t); (t) ]_∞≥ C t^-2. Recall that v_0 = |D|^-1÷ ( u_0 ). Let u_0 e^-|x|^2[ 1; 1; 1 ], and let a_0 ∈ L^1 ∩Ḃ^0_∞,1. Then there exist constants C,δ>0 such that, if a_0_L^1 ∩Ḃ^0_∞,1≤δ, then for all sufficiently large t, ^-1[ (t); (t) ]_∞≥ C t^-2. Firstly, we note that the smallness of a_0 is required to distinguish between v_0 and a_0 via the following reverse triangle inequality ^-1[ (t); (t) ]_∞ ≥^-1[(t) ]_∞ = ^-1[ e^tλ_-λ_+ - e^tλ_+λ_- /λ_+ - λ_- _0 + e^tλ_- - e^tλ_+λ_+ /λ_+ - λ_- _0 ]_∞ ≥^-1[ e^tλ_- - e^tλ_+λ_+ /λ_+ - λ_- _0 ]_∞ - C δ t^-2. Next, focusing on the norm with v̂_0, we note that, due to our choice of u_0, _0 = i C_1 ( _1 + _2 + _3 ) e^-C_2 ||^2/||. Now, we inspect the L^∞-norm above. As in the proof of the previous proposition, we insert the function to restrict our function to the low-frequency case. ^-1[ e^tλ_- - e^tλ_+λ_+ /λ_+ - λ_- _0 ]_∞≥ C ^-1[ e^tλ_- - e^tλ_+λ_+ /λ_+ - λ_- _0 (t^1/2_t) ]_∞ = C sup_x∈^3| e^ix·_0() (t^1/2_t) e^-t||^2/21 - i √(4/||^2 - 1 ) - e^-t||^2/21 + i √(4/||^2 - 1 )/2-(||^2/2)+(i||/2)√( 4-||^2 )| ≥ C t^-3/2| e^-||^2/2_0(t^-1/2) (_t) e^-it^1/2_1 - || √( 1 - ||^2/4t) - e^-it^1/2_1 + || √( 1 - ||^2/4t)/2-(||^2/2t)+(i||/2t^1/2)√( 4-t^-1||^2 )|. Where, for the last step, we set x=(-t,0,0) and made the substitution → t^-1/2, as we did in the previous proposition. The differing signs in our two exponential functions must be addressed. We separate the two terms into two integrals, and make the substitution → - for one of them. Then, since our _0 is antisymmetric, we get | e^- ||^2/2_0(t^-1/2) ( _t ) e^-it^1/2_1 - || √( 1 - ||^2/4t) - e^-it^1/2_1 + || √( 1 - ||^2/4t)/2-(||^2/2t)+(i||/2t^1/2)√( 4-t^-1||^2 )| = | e^- ||^2/2_0(t^-1/2) ( _t ) e^+it^1/2_1 + || √( 1 - ||^2/4t) + e^-it^1/2_1 + || √( 1 - ||^2/4t)/2-(||^2/2t)+(i||/2t^1/2)√( 4-t^-1||^2 )| = 2 | e^- ||^2/2_0(t^-1/2) ( _t )cos( t^1/2_1 + || √( 1 - ||^2/4t) ) /2-(||^2/2t)+(i||/2t^1/2)√( 4-t^-1||^2 )|. We may then follow the same steps as the previous proof, as cos( · ) is sufficiently similar to e^i · for our purposes. Proof of Theorem 1.2.(<ref>) is proven by Proposition <ref> and Proposition <ref>. Indeed, we have for all s ∈, p ∈ [2,∞], and q ∈ [1,∞], [ a(t); v(t) ]_Ḃ^s_p,q = ∑_j≤-1 2^sjq[ a(t); v(t) ]_p^q + ∑_j = 0^2 2^sjq[ a(t); v(t) ]_p^q + ∑_j≥3 2^sjq[ a(t); v(t) ]_p^q ^1/q ≤ C t^-3/2 (1-1/p) - 1/2 (1 - 2/p) ^-1[ _0; _0 ]^l_Ḃ^s_1,q + C e^-t^-1[ _0; _0 ]^h_Ḃ^s+3(1-1/p)_1,q. (<ref>) is proven by Proposition <ref>. Indeed, there exists a constant C such that, for all p ∈ [1,∞] and t>0, ^-1[ e^tλ_±_j [ _0; _0 ]]_p ≤ C e^-t^-1[ (_j-1 + _j + _j+1) [ _0; _0 ]]_p for all j ≥ 3. We thus obtain ^-1[ (t); (t) ]^h_Ḃ^s_p,q≤ C e^-t^-1[ _0; _0 ]^h_Ḃ^s_p,q. (<ref>) is proven by Proposition <ref>. § BEHAVIOUR OF SOLUTIONS NEAR || =2 Let p ∈{ 2, ∞}, j ∈{ 0,1,2 }. Then there exist constants C,c>0 such that, for all t>0, [ a(t); v(t) ]_p≤ Ce^-ct. Taking directly the L^2 norm of a, v, for j ∈{ 0, 1, 2 }, we use Plancherel's theorem to take the Fourier transform inside the norm. Looking at e^tλ_-λ_+ - e^tλ_+ λ_- /λ_+ - λ_- and the other similar expressions in (<ref>) and (<ref>), we see that the numerator and denominator both approach 0 as ||→ 2. First, we may rewrite e^tλ_-λ_+ - e^tλ_+ λ_- /λ_+ - λ_- = (e^tλ_- - e^tλ_+ ) λ_+ /λ_+ - λ_- + e^tλ_+ (λ_+ - λ_-) /λ_+ - λ_- . The second term is of course just e^tλ_+. Focusing on the first term, we inspect the L^2 norm around ||=2. (e^tλ_- - e^tλ_+ ) λ_+ /λ_+ - λ_- (_j-1 + _j +_j+1 ) ^2_2 ≤ C ∫_{1/4≤ || ≤ 2 }| e^- t ||^2/2 e^+it||^2/2√(4/||^2 - 1 ) - e^-it||^2/2√(4/||^2 - 1 )/√( 4/||^2 - 1 )|^2 + C ∫_{ 2 ≤ || ≤ 8 }| e^- t ||^2/2 e^+t||^2/2√( 1 - 4/||^2) - e^-t||^2/2√( 1 - 4/||^2)/√( 1 - 4/||^2 )|^2 I_1 + I_2. Looking at I_1 first, for all t>0, we have I_1 ≤ C e^-t/32∫_{1/4≤ || ≤ 2 }| e^+it||^2/2√(4/||^2 - 1 ) - e^-it||^2/2√(4/||^2 - 1 )/√( 4/||^2 - 1 )|^2 ≤ C t^2 e^-t/32∫_{1/4≤ || ≤ 2 }|sin(t||^2/2√(4/||^2 - 1 ) ) / t||^2/2√( 4/||^2 - 1 )|^2 ≤ C e^-ct. For I_2, values of || away from 2 are handled as in Proposition <ref>. It remains only to consider the integral over || ∈ (2,2+), for some small >0. To that end, we first extract our time decay ∫_{ 2 ≤ || ≤ 2+}| e^- t ||^2/2 e^+t||^2/2√( 1 - 4/||^2) - e^-t||^2/2√( 1 - 4/||^2)/√( 1 - 4/||^2 )|^2 ≤ e^ -4t ∫_{ 2 ≤ || ≤ 2+}| e^+t||^2/2√( 1 - 4/||^2) - e^-t||^2/2√( 1 - 4/||^2)/√( 1 - 4/||^2 )|^2 , and consider the Taylor expansion of the remaining integrand: e^+t||^2/2√( 1 - 4/||^2) - e^-t||^2/2√( 1 - 4/||^2)/√( 1 - 4/||^2 ) = 1/√( 1 - 4/||^2 )∑^∞_n=1 (t||^2/2)^n (1 - 4/||^2)^n/2 - (-1)^n (1 - 4/||^2)^n/2/n! = 2 /√( 1 - 4/||^2 )∑_n ≥ 1, odd (t||^2/2)^n (1 - 4/||^2)^n/2/n! ≤ t||^2 ∑_n ≥ 0, even (t||^2/2)^n (1 - 4/||^2)^n/2/n! ≤ t ||^2 e^ +t||^2/2√( 1 - 4/||^2) , for all || > 2, t>0. We thus obtain for all t>0 and any sufficiently small I_2 ≤ C e^-4t∫_{ 2 ≤ || ≤ 2+}| t ||^2 e^ +t||^2/2√( 1 - 4/||^2)|^2 + C e^-ct ≤ C e^-ct, and so we have obtained the result for p=2. Using the inequality ^-1 [f̂] _∞≤f̂_1, the same result is proven similarly for the L^∞ norm. § BOUNDS FROM ABOVE FOR LOW FREQUENCIES Let j≤ -1. Then, for all sufficiently large t, sup_x∈^3| e^i (t^-1/2x)· e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) | ≤ C t^- 1/2. First note that, by radial symmetry, we can assume without loss of generality that x = (x_1,0,0). We then split the frequency space into four subsets: B_1 {∈^3 | |_2| , |_3| ≤ t^-1/4} B_2 {∈^3 | |_2| ≥ t^-1/4 , |_3| ≤ t^-1/4} B_3 {∈^3 | |_2| ≤ t^-1/4, |_3| ≥ t^-1/4} B_4 {∈^3 | |_2| , |_3| ≥ t^-1/4} ^3 = B_1 ∪ B_2 ∪ B_3 ∪ B_4. We shall estimate our integral over each of the four subsets. We start with B_1. sup_x∈^3|∫_∈ B_1 e^i t^-1/2 x_1 _1 e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) | ≤∫_∫_{ |_3| ≤ t^-1/4}∫_{ |_2| ≤ t^-1/4} e^-||^2/2_2 _3 _1 ≤ C t^-1/2. For B_2, we first note that e^ i t^1/2||^2/2√(4/||^2 - t^-1) = _j e^ i t^1/2||^2/2√(4/||^2 - t^-1)||^2/2√(4/||^2 - t^-1)/ 2 _j i t^1/2 (2 - ||^2 t^-1) , and thus, sup_x∈^3|∫_∈ B_2 e^i t^-1/2 x_1 _1 e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) | ≤∫_{ |_1| < t^1/2}∫_{ |_3| ≤ t^-1/4} |∫_{ |_2| ≥ t^-1/4}_2 e^ i t^1/2||^2/2√(4/||^2 - t^-1)||^2/2√(4/||^2 - t^-1)(2^-j t^-1/2 ) e^ - ||^2/2/ 2 _2 i t^1/2 (2 - ||^2 t^-1) _2 |_3 _1 ≤∫_{ |_1| < t^1/2}∫_{ |_3| ≤ t^-1/4} Ct^-1/4 + Ct^-1/2∫^∞_t^-1/4|_2 ||^2/2√(4/||^2 - t^-1)(2^-j t^-1/2 ) e^ - ||^2/2/ 2 _2 i (2 - ||^2 t^-1) |_2 _3 _1. The above is accomplished by a simple integration by parts. Routine checks verify that ∫^∞_t^-1/4|_2 ||^2/2√(4/||^2 - t^-1)/ 2 _2 i (2 - ||^2 t^-1) (2^-j t^-1/2 ) e^ - _2^2/2|_2 ≤ Ct^+1/4. An additional time decay of t^-1/4 is obtained from the integral over _3, and boundedness of the integral over _1 is assured by the presence of the Gauss function e^-_1^2/2. Thus sup_x∈^3|∫_∈ B_2 e^i t^-1/2 x_1 _1 e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) |≤ C t^-1/2. The process for bounding the integral over B_3 is identical to that of B_2. We are thus left with the integral over B_4. We note that e^ i t^1/2||^2/2√(4/||^2 - t^-1) = _3 _2 e^ i t^1/2||^2/2√(4/||^2 - t^-1)/4_2 _3 i t^1/2||^6/2√(4/||^2 - t^-1)^3 / 4 i t^1/2 ||^2 √(4/||^2 - t^-1) - it^-1/2 ||^6 √(4/||^2 - t^-1)^3 - 4 , (A) and thus, by integration by parts, sup_x∈^3|∫_∈ B_4 e^i t^-1/2 x_1 _1 e^± i t^1/2||^2/2√(4/||^2 - t^-1) e^-||^2/2(2^-j t^-1/2 ) | ≤∫^1_-1|∫_{ |_3| ≥ t^-1/4}∫_{ |_2| ≥ t^-1/4} (A) (2^-j t^-1/2 ) e^-||^2/2_2 _3 |_1 ≤ Ct^-1/2, by some longer but still routine calculations similar to the integral over B_2. Therefore, we have obtained for all j ≤ -1, and all t > t_0, sup_x∈^3| e^ix· e^± it||^2/2√(4/||^2 - 1 ) e^-t||^2/2ϕ̂(2^-j) |≤ C t^-2. This completes the proof. When we limit ourselves to the _1 > 0 case, as in the proof of Proposition <ref>, we may obtain an extra decay of t^-1/4 by performing an additional integration by parts over _1. unsrt

http://arxiv.org/abs/2303.06674v1

20230312142824

Universal Instance Perception as Object Discovery and Retrieval

cs.CV

Universal Instance Perception as Object Discovery and Retrieval Bin Yan^1This work was performed while Bin Yan worked as an intern at ByteDance. Email: mailto:yan_bin@mail.dlut.edu.cnyan_bin@mail.dlut.edu.cn. ^† Corresponding authors: mailto:jiangyi.enjoy@bytedance.comjiangyi.enjoy@bytedance.com, mailto:wdice@dlut.edu.cnwdice@dlut.edu.cn., Yi Jiang^2, †, Jiannan Wu^3, Dong Wang^1, †, Ping Luo^3, Zehuan Yuan^2, Huchuan Lu^1,4 ^1 School of Information and Communication Engineering, Dalian University of Technology, China ^2 ByteDance ^3 The University of Hong Kong ^4 Peng Cheng Laboratory March 30, 2023 =========================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================== All instance perception tasks aim at finding certain objects specified by some queries such as category names, language expressions, and target annotations, but this complete field has been split into multiple independent sub-tasks. In this work, we present a universal instance perception model of the next generation, termed UNINEXT. UNINEXT reformulates diverse instance perception tasks into a unified object discovery and retrieval paradigm and can flexibly perceive different types of objects by simply changing the input prompts. This unified formulation brings the following benefits: (1) enormous data from different tasks and label vocabularies can be exploited for jointly training general instance-level representations, which is especially beneficial for tasks lacking in training data. (2) the unified model is parameter-efficient and can save redundant computation when handling multiple tasks simultaneously. UNINEXT shows superior performance on 20 challenging benchmarks from 10 instance-level tasks including classical image-level tasks (object detection and instance segmentation), vision-and-language tasks (referring expression comprehension and segmentation), and six video-level object tracking tasks. Code is available at https://github.com/MasterBin-IIAU/UNINEXThttps://github.com/MasterBin-IIAU/UNINEXT. § INTRODUCTION Object-centric understanding is one of the most essential and challenging problems in computer vision. Over the years, the diversity of this field increases substantially. In this work, we mainly discuss 10 sub-tasks, distributed on the vertices of the cube shown in Figure <ref>. As the most fundamental tasks, object detection <cit.> and instance segmentation <cit.> require finding all objects of specific categories by boxes and masks respectively. Extending inputs from static images to dynamic videos, Multiple Object Tracking (MOT) <cit.>, Multi-Object Tracking and Segmentation (MOTS) <cit.>, and Video Instance Segmentation (VIS) <cit.> require finding all object trajectories of specific categories in videos. Except for category names, some tasks provide other reference information. For example, Referring Expression Comprehension (REC) <cit.>, Referring Expression Segmentation (RES) <cit.>, and Referring Video Object Segmentation (R-VOS) <cit.> aim at finding objects matched with the given language expressions like “The fourth person from the left”. Besides, Single Object Tracking (SOT) <cit.> and Video Object Segmentation (VOS) <cit.> take the target annotations (boxes or masks) given in the first frame as the reference, requiring to predict the trajectories of the tracked objects in the subsequent frames. Since all the above tasks aim to perceive instances of certain properties, we refer to them collectively as instance perception. Although bringing convenience to specific applications, such diverse task definitions split the whole field into fragmented pieces. As the result, most current instance perception methods are developed for only a single or a part of sub-tasks and trained on data from specific domains. Such fragmented design philosophy brings the following drawbacks: (1) Independent designs hinder models from learning and sharing generic knowledge between different tasks and domains, causing redundant parameters. (2) The possibility of mutual collaboration between different tasks is overlooked. For example, object detection data enables models to recognize common objects, which can naturally improve the performance of REC and RES. (3) Restricted by fixed-size classifiers, traditional object detectors are hard to jointly train on multiple datasets with different label vocabularies <cit.> and to dynamically change object categories to detect during inference <cit.>. Since essentially all instance perception tasks aim at finding certain objects according to some queries, it leads to a natural question: could we design a unified model to solve all mainstream instance perception tasks once and for all? To answer this question, we propose UNINEXT, a universal instance perception model of the next generation. We first reorganize 10 instance perception tasks into three types according to the different input prompts: (1) category names as prompts (Object Detection, Instance Segmentation, VIS, MOT, MOTS). (2) language expressions as prompts (REC, RES, R-VOS). (3) reference annotations as prompts (SOT, VOS). Then we propose a unified prompt-guided object discovery and retrieval formulation to solve all the above tasks. Specifically, UNINEXT first discovers N object proposals under the guidance of the prompts, then retrieves the final instances from the proposals according to the instance-prompt matching scores. Based on this new formulation, UNINEXT can flexibly perceive different instances by simply changing the input prompts. To deal with different prompt modalities, we adopt a prompt generation module, which consists of a reference text encoder and a reference visual encoder. Then an early fusion module is used to enhance the raw visual features of the current image and the prompt embeddings. This operation enables deep information exchange and provides highly discriminative representations for the later instance prediction step. Considering the flexible query-to-instance fashion, we choose a Transformer-based object detector <cit.> as the instance decoder. Specifically, the decoder first generates N instance proposals, then the prompt is used to retrieve matched objects from these proposals. This flexible retrieval mechanism overcomes the disadvantages of traditional fixed-size classifiers and enables joint training on data from different tasks and domains. With the unified model architecture, UNINEXT can learn strong generic representations on massive data from various tasks and solve 10 instance-level perception tasks using a single model with the same model parameters. Extensive experiments demonstrate that UNINEXT achieves superior performance on 20 challenging benchmarks. The contributions of our work can be summarized as follows. * We propose a unified prompt-guided formulation for universal instance perception, reuniting previously fragmented instance-level sub-tasks into a whole. * Benefiting from the flexible object discovery and retrieval paradigm, UNINEXT can train on different tasks and domains, in no need of task-specific heads. * UNINEXT achieves superior performance on 20 challenging benchmarks from 10 instance perception tasks using a single model with the same model parameters. § RELATED WORK Instance Perception. The goals and typical methods of 10 instance perception tasks are introduced as follows. Retrieval by Category Names. Object detection and instance segmentation aim at finding all objects of specific classes on the images in the format of boxes or masks. Early object detectors can be mainly divided into two-stage methods <cit.> and one-stage methods <cit.> according to whether to use RoI-level operations <cit.>. Recently, Transformer-based detectors <cit.> have drawn great attention for their conceptually simple and flexible frameworks. Besides, instance segmentation approaches can also be divided into detector-based <cit.> and detector-free <cit.> fashions according to whether box-level detectors are needed. Object detection and instance segmentation play critical roles and are foundations for all other instance perception tasks. For example, MOT, MOTS, and VIS extend image-level detection and segmentation to videos, requiring finding all object trajectories of specific classes in videos. Mainstream algorithms <cit.> of MOT and MOTS follow an online "detection-then-association" paradigm. However, due to the intrinsic difference in benchmarks of MOTS <cit.> (high-resolution long videos) and VIS <cit.> (low-resolution short videos), most recent VIS methods <cit.> adopt an offline fashion. This strategy performs well on relatively simple VIS2019 <cit.>, but the performance drops drastically on challenging OVIS <cit.> benchmark. Recently, IDOL <cit.> bridges the performance gap between online fashion and its offline counterparts by discriminative instance embeddings, showing the potential of the online paradigm in unifying MOT, MOTS, and VIS. Retrieval by Language Expressions. REC, RES, and R-VOS aim at finding one specific target referred by a language expression using boxes or masks on the given images or videos. Similar to object detection, REC methods can be categorized into three paradigms: two-stage <cit.>, one-stage <cit.>, and Transformer-based <cit.> ones. Different from REC, RES approaches <cit.> focus more on designing diverse attention mechanisms to achieve vision-language alignment. Recently, SeqTR <cit.> unifies REC and RES as a point prediction problem and obtains promising results. Finally, R-VOS can be seen as a natural extension of RES from images to videos. Current state-of-the-art methods <cit.> are Transformer-based and process the whole video in an offline fashion. However, the offline paradigm hinders the applications in the real world such as long videos and ongoing videos (e.g. autonomous driving). Retrieval by Reference Annotations. SOT and VOS first specify tracked objects on the first frame of a video using boxes or masks, then require algorithms to predict the trajectories of the tracked objects in boxes or masks respectively. The core problems of these two tasks include (1) How to extract informative target features? (2) How to fuse the target information with representations of the current frame? For the first question, most SOT methods <cit.> encode target information by passing a template to a siamese backbone. While VOS approaches <cit.> usually pass multiple previous frames together with corresponding mask results to a memory encoder for extracting fine-grained target information. For the second question, correlations are widely adopted by early SOT algorithms <cit.>. However, these simple linear operations may cause serious information loss. To alleviate this problem, later works <cit.> resort to Transformer for more discriminative representations. Besides, feature fusion in VOS is almost dominated by space-time memory networks <cit.>. Unified Vision Models. Recently, unified vision models <cit.> have drawn great attention and achieved significant progress due to their strong generalizability and flexibility. Unified vision models attempt to solve multiple vision or multi-modal tasks by a single model. Existing works can be categorized into unified learning paradigms and unified model architectures. Unified Learning Paradigms. These works <cit.> usually present a universal learning paradigm for covering as many tasks and modalities as possible. For example, MuST <cit.> presents a multi-task self-training approach for 6 vision tasks. INTERN <cit.> introduces a continuous learning scheme, showing strong generalization ability on 26 popular benchmarks. Unified-IO <cit.> and OFA <cit.> proposes a unified sequence-to-sequence framework that can handle a variety of vision, language, and multi-modal tasks. Although these works can perform many tasks, the commonality and inner relationship among different tasks are less explored and exploited. Unified Model Architectures. These works <cit.> usually design a unified formulation or model architecture for a group of closely related tasks. For example, Mask R-CNN <cit.> proposes a unified network to perform object detection and instance segmentation simultaneously. Mask2Former <cit.> presents a universal architecture capable of handling panoptic, instance, and semantic segmentation. Pix2SeqV2 <cit.> designs a unified pixel-to-sequence interface for four vision tasks, namely object detection, instance segmentation, keypoint detection, and image captioning. GLIP <cit.> cleverly reformulates object detection as phrase grounding by replacing classical classification with word-region alignment. This new formulation allows joint training on both detection and grounding data, showing strong transferability to various object-level recognition tasks. However, GLIP <cit.> supports neither prompts in other modalities such as images & annotations nor video-level tracking tasks. In terms of object tracking, Unicorn <cit.> proposes a unified solution for SOT, VOS, MOT, and MOTS, achieving superior performance on 8 benchmarks with the same model weights. However, it is still difficult for Unicorn to handle diverse label vocabularies <cit.> during training and inference. In this work, we propose a universal prompt-guided architecture for 10 instance perception tasks, conquering the drawbacks of GLIP <cit.> and Unicorn <cit.> simultaneously. § APPROACH Before introducing detailed methods, we first categorize existing instance perception tasks into three classes. * Object detection, instance segmentation, MOT, MOTS, and VIS take category names as prompts to find all instances of specific classes. * REC, RES, and R-VOS exploit an expression as the prompt to localize a certain target. * SOT and VOS use the annotation given in the first frame as the prompt for predicting the trajectories of the tracked target. Essentially, all the above tasks aim to find objects specified by some prompts. This commonality motivates us to reformulate all instance perception tasks into a prompt-guided object discovery and retrieval problem and solve it by a unified model architecture and learning paradigm. As demonstrated in Figure <ref>, UNINEXT consists of three main components: (1) prompt generation (2) image-prompt feature fusion (3) object discovery and retrieval. §.§ Prompt Generation First, a prompt generation module is adopted to transform the original diverse prompt inputs into a unified form. According to different modalities, we introduce the corresponding strategies in the next two paragraphs respectively. To deal with language-related prompts, a language encoder <cit.> Enc_L is adopted. To be specific, for category-guided tasks, we concatenate class names that appeared in the current dataset <cit.> as the language expression. Take COCO <cit.> as an example, the expression can be written as “person. bicycle. ... . toothbrush". Then for both category-guided and expression-guided tasks, the language expression is passed into Enc_L, getting a prompt embedding F_p∈ℝ^L×d with a sequence length of L. For the annotation-guided tasks, to extract fine-grained visual features and fully exploit the target annotations, an additional reference visual encoder Enc^ref_V is introduced. Specifically, first a template with 2^2 times the target box area is cropped centered on the target location on the reference frame. Then the template is resized to a fixed size of 256×256. To introduce more precise target information, an extra channel named the target prior is concatenated to the template image, forming a 4-channel input. In more detail, the value of the target prior is 1 on the target region otherwise 0. Then the template image together with the target prior is passed to the reference visual encoder Enc^ref_V, obtaining a hierarchical feature pyramid {C_3,C_4,C_5,C_6}. The corresponding spatial sizes are 32×32, 16×16, 8×8, and 4×4. To keep fine target information and get the prompt embedding in the same format as other tasks, a merging module is applied. Namely, all levels of features are first upsampled to 32×32 then added, and flattened as the final prompt embedding F_p∈ℝ^1024×d. The prompt generation process can be formulated as F_p={ Enc^ref_L(expression) expression-guided Enc^ref_L(concat(categories)) category-guided merge(Enc^ref_V([template, prior]) annotation-guided . §.§ Image-Prompt Feature Fusion In parallel with the prompt generation, the whole current image is passed through another visual encoder Enc_V, obtaining hierarchical visual features F_v. To enhance the original prompt embedding by the image contexts and to make the original visual features prompt-aware, an early fusion module is adopted. To be specific, first a bi-directional cross-attention module (Bi-XAtt) is used to retrieve information from different inputs, and then the retrieved representations are added to the original features. This process can be formulated as F_p2v, F_v2p = Bi-XAtt(F_v, F_p) F^'_v = F_v + F_p2v; F^'_p = F_p + F_v2p Different from GLIP <cit.>, which adopts 6 vision-language fusion layers and 6 additional BERT layers for feature enhancement, our early fusion module is much more efficient. §.§ Object Discovery and Retrieval With discriminative visual and prompt representations, the next crucial step is to transform input features into instances for various perception tasks. UNINEXT adopts the encoder-decoder architecture proposed by Deformable DETR <cit.> for its flexible query-to-instance fashion. We introduce the detailed architectures as follows. The Transformer encoder takes hierarchical prompt-aware visual features as the inputs. With the help of efficient Multi-scale Deformable Self-Attention <cit.>, target information from different scales can be fully exchanged, bringing stronger instance features for the subsequent instance decoding. Besides, as performed in two-stage Deformable DETR <cit.>, an auxiliary prediction head is appended at the end of the encoder, generating N initial reference points with the highest scores as the inputs of the decoder. The Transformer decoder takes the enhanced multi-scale features, N reference points from the encoder, as well as N object queries as the inputs. As shown in previous works  <cit.>, object queries play a critical role in instance perception tasks. In this work, we attempt two query generation strategies: (1) static queries which do not change with images or prompts. (2) dynamic queries conditioned on the prompts. The first strategy can be easily implemented with . The second one can be performed by first pooling the enhanced prompt features F^'_v along the sequence dimension, getting a global representation, then repeating it by N times. The above two methods are compared in Sec <ref> and we find that static queries usually perform better than dynamic queries. The potential reason could be that static queries contain richer information and possess better training stability than dynamic queries. With the help of the deformable attention, the object queries can efficiently retrieve prompt-aware visual features and learn strong instance embedding F_ins∈ℝ^N×d. At the end of the decoder, a group of prediction heads is exploited to obtain the final instance predictions. Specifically, an instance head produces both boxes and masks of the targets. Besides, an embedding head <cit.> is introduced for associating the current detected results with previous trajectories in MOT, MOTS, and VIS. Until now, we have mined N potential instance proposals, which are represented with gray masks in Figure <ref>. However, not all proposals are what the prompts really refer to. Therefore, we need to further retrieve truly matched objects from these proposals according to the prompt embeddings as demonstrated in the right half of Figure <ref>. Specifically, given the prompt embeddings F^'_p after early fusion, for category-guided tasks, we take the embedding of each category name as a weight matrix W∈ℝ^1×d. Besides, for expression-guided and annotation-guided tasks, the weight matrix W is obtained by aggregating the prompt embedding F^'_p using global average pooling (GAP) along the sequence dimension. W={ F^'_p[i], i∈{0,1,...,C-1} category 1/L∑_i=0^LF_p^'(i,j) expression/annotation . Finally, the instance-prompt matching scores S can be computed as the matrix multiplication of the target features and the transposed weight matrix. S=F_insW^⊤. Following previous work <cit.>, the matching scores can be supervised by Focal Loss <cit.>. Different from previous fixed-size classifiers <cit.>, the proposed retrieval head selects objects by the prompt-instance matching mechanism. This flexible design enables UNINEXT to jointly train on enormous datasets with diverse label vocabularies from different tasks, learning universal instance representations. §.§ Training and Inference Training. The whole training process consists of three consecutive stages: (1) general perception pretraining (2) image-level joint training (3) video-level joint training. In the first stage, we pretrain UNINEXT on the large-scale object detection dataset Objects365 <cit.> for learning universal knowledge about objects. Since Objects365 does not have mask annotations, we introduce two auxiliary losses proposed by BoxInst <cit.> for training the mask branch. The loss function can be formulated as ℒ_stage1 = ℒ_retrieve + ℒ_box + ℒ^boxinst_mask Then based on the pretrained weights of the first stage, we finetune UNINEXT jointly on image datasets, namely COCO <cit.> and the mixed dataset of RefCOCO <cit.>, RefCOCO+ <cit.>, and RefCOCOg <cit.>. With manually labeled mask annotations, the traditional loss functions like Dice Loss <cit.> and Focal Loss <cit.> can be used for the mask learning. After this step, UNINEXT can achieve superior performance on object detection, instance segmentation, REC, and RES. ℒ_stage2 = ℒ_retrieve + ℒ_box + ℒ_mask Finally, we further finetune UNINEXT on video-level datasets for various downstream object tracking tasks and benchmarks. In this stage, the model is trained on two frames randomly chosen from the original videos. Besides, to avoid the model forgetting previously learned knowledge on image-level tasks, we also transform image-level datasets to pseudo videos for joint training with other video datasets. In summary, the training data in the third stage includes pseudo videos generated from COCO <cit.>, RefCOCO/g/+ <cit.>, SOT&VOS datasets (GOT-10K <cit.>, LaSOT <cit.>, TrackingNet <cit.>, and Youtube-VOS <cit.>), MOT&VIS datasets (BDD100K <cit.>, VIS19 <cit.>, OVIS <cit.>), and R-VOS dataset Ref-Youtube-VOS <cit.>. Meanwhile, a reference visual encoder for SOT&VOS and an extra embedding head for association are introduced and optimized in this period. ℒ_stage3 = ℒ_retrieve + ℒ_box + ℒ_mask + ℒ_embed Inference. For category-guided tasks, UNINEXT predicts instances of different categories and associates them with previous trajectories. The association proceeds in an online fashion and is purely based on the learned instance embedding following <cit.>. For expression-guided and annotation-guided tasks, we directly pick the object with the highest matching score with the given prompt as the final result. Different from previous works <cit.> restricted by the offline fashion or complex post-processing, our method is simple, online, and post-processing free. § EXPERIMENTS §.§ Implementation Details We attempt three different backbones, ResNet-50 <cit.>, ConvNeXt-Large <cit.>, and ViT-Huge <cit.> as the visual encoder. We adopt BERT <cit.> as the text encoder and its parameters are trained in the first and second training stages while being frozen in the last training stage. The Transformer encoder-decoder architecture follows  <cit.> with 6 encoder layers and 6 decoder layers. The number of object queries N is set to 900. The optimizer is AdamW <cit.> with weight decay of 0.05. The model is trained on 32 and 16 A100 GPUs for Objects365 pretraining and other stages respectively. More details can be found in the appendix. §.§ Evaluations on 10 Tasks We compare UNINEXT with task-specific counterparts in 20 datasets. In each benchmark, the best two results are indicated in bold and with underline. UNINEXT in all benchmarks uses the same model parameters. Object Detection and Instance Segmentation. We compare UNINEXT with state-of-the-art object detection and instance segmentation methods on COCO 𝚟𝚊𝚕2017 (5k images) and 𝚝𝚎𝚜𝚝-𝚍𝚎𝚟 split (20k images) respectively. As shown in Table <ref>, UNINEXT surpasses state-of-the-art query-based detector DN-Deformable DETR <cit.> by 2.7 box AP. By replacing ResNet-50 <cit.> with stronger ConvNeXt-Large <cit.> and ViT-Huge <cit.> backbones, UNINEXT achieves a box AP of 58.1 and 60.6, surpassing competitive rivals Cascade Mask-RCNN <cit.> and ViTDet-H <cit.> by 3.3 and 1.9 respectively. Besides, the results of instance segmentation are shown in Table <ref>. With the same ResNet-50 backbone, UNINEXT outperforms state-of-the-art QueryInst by 4.3 AP and 6.2 AP_L. When using ConvNeXt-Large as the backbone, UNINEXT achieves a mask AP of 49.6, surpassing Cascade Mask R-CNN <cit.> by 2.0. With ViT-Huge as the backbone, UNINEXT achieves state-of-the-art mask AP of 51.8. REC and RES. RefCOCO <cit.>, RefCOCO+ <cit.>, and RefCOCOg <cit.> are three representative benchmarks for REC and RES proposed by different institutions. Following previous literature, we adopt Precision@0.5 and overall IoU (oIoU) as the evaluation metrics for REC and RES respectively and results are rounded to two decimal places. As shown in Table <ref> and Table <ref>, our method with ResNet-50 backbone surpasses all previous approaches on all splits. Furthermore, when using ConvNeXt-Large and ViT-Huge backbones, UNINEXT obtains new state-of-the-art results, exceeding the previous best method by a large margin. Especially on RES, UNINEXT-H outperforms LAVT <cit.> by 10.85 on average. SOT. We compare UNINEXT with state-of-the-art SOT methods on four large-scale benchmarks: LaSOT <cit.>, LaSOT-ext <cit.>, TrackingNet <cit.>, and TNL-2K <cit.>. These benchmarks adopt the area under the success curve (AUC), normalized precision (P_Norm), and precision (P) as the evaluation metrics and include 280, 150, 511, and 700 videos in the test set respectively. As shown in Table <ref>, UNINEXT achieves the best results in terms of AUC and P among all trackers with ResNet-50 backbone. Especially on TNL-2K, UNINEXT outperforms the second best method TransT <cit.> by 5.3 AUC and 5.8 P respectively. Besides, UNINEXT with stronger backbones obtains the best AUC on all four benchmarks, exceeding Unicorn <cit.> with the same backbone by 3.9 on LaSOT. VOS. The comparisons between UNINEXT with previous semi-supervised VOS methods are demonstrated in Table <ref>. DAVIS-2017 <cit.> adopts region similarity , contour accuracy , and the averaged score as the metrics. Similarly, Youtube-VOS 2018 <cit.> reports and for both seen and unseen categories, and the averaged overall score . UNINEXT achieves the best results among all non-memory-based methods, largely bridging the performance gap between non-memory-based approaches and memory-based ones. Furthermore, compared with traditional memory-based methods <cit.>, UNINEXT does not rely on the intermediate mask predictions. This leads to constant memory consumption, enabling UNINEXT to handle long sequences of any length. MOT. We compare UNINEXT with state-of-the-art MOT methods on BDD100K <cit.>, which requires tracking 8 classes of instances in the autonomous driving scenario. Except for classical evaluation metrics Multiple-Object Tracking Accuracy (MOTA), Identity F1 Score (IDF1), and Identity Switches (IDS), BDD100K additionally introduces mMOTA, and mIDF1 to evaluate the average performance across 8 classes. As shown in Table <ref>, UNINEXT surpasses Unicorn <cit.> by 3.0 mMOTA and 2.7 mIDF1 respectively. MOTS. Similar to MOT, BDD100K MOTS Challenge <cit.> evaluates the performance on multi-class tracking by mMOTSA, mMOTSP, mIDF1, and ID Sw. This benchmark contains 37 sequences with mask annotations in the validation set. As shown in Table <ref>, UNINEXT achieves state-of-the-art performance, surpassing the previous best method Unicorn <cit.> by 6.1 mMOTSA. VIS. We compare UNINEXT against state-of-the-art VIS methods on Youtube-VIS 2019 <cit.> and OVIS <cit.> validation sets. Specifically, Youtube-VIS 2019 and OVIS have 40 and 25 object categories, containing 302 and 140 videos respectively in the validation set. Both benchmarks take AP as the main metric. As shown in Table <ref>, when using the same ResNet-50 backbone, UNINEXT obtains the best results on both datasets. Especially on more challenging OVIS, UNINEXT exceeds the previous best method IDOL <cit.> by 3.8 AP. When using stronger ViT-Huge backbone, UNINEXT achieves state-of-the-art AP of 66.9 on Youtube-VIS 2019 and 49.0 on OVIS respectively, surpassing previous methods by a large margin. R-VOS. Ref-Youtube-VOS <cit.> and Ref-DAVIS17 <cit.> are two popular R-VOS benchmarks, which are constructed by introducing language expressions for the objects in the original Youtube-VOS <cit.> and DAVIS17 <cit.> datasets. As same as semi-supervised VOS, region similarity , contour accuracy , and the averaged score are adopted as the metrics. As demonstrated in Table <ref>, UNINEXT outperforms all previous R-VOS approaches by a large margin, when using the same ResNet-50 backbone. Especially on Ref-DAVIS17, UNINEXT exceeds previous best ReferFormer <cit.> by 5.4 . Furthermore, when adopting stronger ViT-Huge backbone, UNINEXT achieves new state-of-the-art of 70.1 on Ref-Youtube-VOS and 72.5 on Ref-DAVIS17. Besides, different from offline RefFormer, UNINEXT works in a flexible online fashion, making it applicable to ongoing videos in the real world. §.§ Ablations and Other Analysis In this section, we conduct component-wise analysis for better understanding our method. All models take ResNet-50 as the backbone. The methods are evaluated on five benchmarks (COCO <cit.>, RefCOCO <cit.>, Youtube-VOS <cit.>, Ref-Youtube-VOS <cit.>, and Youtube-VIS 2019 <cit.>) from five tasks (object detection, REC, VOS, R-VOS, and VIS). The results are shown in Table <ref>. Fusion. To study the effect of feature fusion between visual features and prompt embeddings, we implement a variant without any early fusion. In this version, prompt embeddings do not have an influence on proposal generation but are only used in the final object retrieval process. Experiments show that early fusion has the greatest impact on VOS, the performance on VOS drops drastically by 21.4 without feature fusion. This is mainly caused by the following reasons (1) Without the guidance of prompt embeddings, the network can hardly find rare referred targets like trees and sinks. (2) Without early fusion, the network cannot fully exploit fine mask annotations in the first frame, causing degradation of the mask quality. Besides, the removal of feature fusion also causes performance drop of 2.3 P@0.5 and 2.8 on REC and RVOS respectively, showing the importance of early fusion in expression-guided tasks. Finally, feature fusion has minimum influence on object detection and VIS. This can be understood because both two tasks aim to find all objects as completely as possible rather than locating one specific target referred by the prompt. Queries. We compare two different query generation strategies: static queries by and dynamic queries conditioned on the prompt embeddings. Experiments show that dynamic queries perform slightly better than static queries on the first four tasks. However, static queries outperform dynamic ones by 2.8 AP on the VIS task, obtaining higher overall performance. A potential reason is that N different object queries can encode richer inner relationship among different targets than simply copying the pooled prompt by N times as queries. This is especially important for VIS because targets need to be associated according to their affinity in appearance and space. Unification. We also compare two different model design philosophies, one unified model or multiple task-specific models. Except for the unified model, we also retrain five task-specific models only on data from corresponding tasks. Experiments show that the unified model achieves significantly better performance than its task-specific counterparts on five tasks, demonstrating the superiority of the unified formulation and joint training on all instance perception tasks. Finally, the unified model can save tons of parameters, being much more parameter-efficient. § CONCLUSIONS We propose UNINEXT, a universal instance perception model of the next generation. For the first time, UNINEXT unifies 10 instance perception tasks with a prompt-guided object discovery and retrieval paradigm. Extensive experiments demonstrate that UNINEXT achieves superior performance on 20 challenging benchmarks with a single model with the same model parameters. We hope that UNINEXT can serve as a solid baseline for the research of instance perception in the future. Acknowledgement. We would like to thank the reviewers for their insightful comments. The paper is supported in part by the National Key R&D Program of China under Grant No. 2018AAA0102001, 2022ZD0161000 and National Natural Science Foundation of China under grant No. 62293542, U1903215, 62022021 and the Fundamental Research Funds for the Central Universities No.DUT22ZD210. ieee_fullname § APPENDIX In this appendix, we present more details about the training process and loss functions in  <ref> and  <ref>, network architecture in  <ref>, as well as more analysis and visualizations for better understanding in  <ref>. §.§ Training Process The detailed hyperparameters during training are shown in Tab <ref>. The whole training process consists of three stages. In each stage, the learning rate scheduler is adopted. The learning rate drops by a factor of 10 after the given steps. For multi-dataset training, we follow the implementation of Detic <cit.>, which randomly samples data from different tasks and then computes them on different GPUs in one iteration. Besides, the multi-scale training technique is used across all datasets in all stages. Take the pre-training on Objects365 <cit.> as an example, the original images are resized such that the shortest side is at least 480 and at most 800 pixels while the longest side is at most 1333. We use this as the default setting except on Youtube-VOS <cit.>, Youtube-VIS-2019 <cit.>, and Ref-Youtube-VOS <cit.>. A lower resolution with the shortest side ranging from 320 to 640 and the longest side not exceeding 768 is applied to these datasets <cit.>, following previous works <cit.>. Specifically, in the first stage, the model is pretrained on Objects365 <cit.> for about 340K iterations (12 epochs) and the learning rate drops on the 11th epoch. In the second stage, we finetune UNINEXT on COCO <cit.> and RefCOCO/g/+ <cit.> jointly for 12 epochs. In the third stage, UNINEXT is further finetuned for diverse video-level tasks. To guarantee balanced performance on various benchmarks, we set the data sampling ratios as (SOT&VOS):(MOT&MOTS):VIS:R-VOS = 1:1:1:1. For each task, 45K iterations are allocated, thus bringing 180K iterations in total for the third stage. Besides, to avoid forgetting previously learned knowledge on image-level tasks, we also generate pseudo videos from COCO <cit.> and RefCOCO/g/+ <cit.> and mix them with training data of VIS <cit.> and R-VOS <cit.> respectively. §.§ Loss Functions We present detailed loss functions described in Sec. <ref> for better readability. First, ℒ_retrieve and ℒ_box are used across all three stages. Second, to learn mask representations from coarse boxes <cit.> and fine mask annotations <cit.>, UNINEXT uses ℒ^boxinst_mask in the first stage and ℒ_mask in the next two stages respectively. Finally, to associate instances on different frames <cit.>, UNINEXT additionally adopts ℒ_embed in the last stage. ℒ_retrieve. Given the raw instance-prompt matching score s, the normalized matching probability p is computed as p=σ(s), where σ is sigmoid function. Then ℒ_retrieve can be written as the form of Focal loss <cit.>. flalphaℒ_retrieve() = - (1 - )^γlog(). pt= p if matched 1 - p otherwise. γ and α are 2 and 0.25 respectively. ℒ_box. Following DETR-like methods <cit.>, ℒ_box consists of two terms, GIoU Loss <cit.> and ℓ_1 loss: loss_boxℒ_box(b,b̂) = λ_giouℒ_giou(b,b̂)+λ_L_1‖b-b̂ ‖. loss_giouℒ_giou(b,b̂)=1-IoU(b,b̂)+A^c(b,b̂)-U(b,b̂)/A^c(b,b̂), where A^c(b,b̂) is the area of the smallest box containing b and b̂. U(b,b̂) is the area of the union of b and b̂. ℒ_mask. For datasets with mask annotations <cit.>, Focal Loss <cit.> and Dice Loss <cit.> are adopted. loss_maskℒ_mask(m,m̂) = λ_focalℒ_focal(m,m̂)+λ_diceℒ_dice(m,m̂). loss_diceℒ_dice(m,m̂)=1-2mm̂+1/m̂+m+1, where m and m̂ are binary GT masks and predicted masks after sigmoid activation respectively. ℒ^boxinst_mask. For Objects365 <cit.> without mask annotations, UNINEXT uses Projection Loss and Pairwise Affinity Loss like BoxInst <cit.>, which can learn mask prediction only based on box-level annotations. loss_boxinstℒ^boxinst_mask(b,m̂) = ℒ_proj(b,m̂)+ℒ_pairwise(b,m̂). loss_proj ℒ_proj(b,m̂)= ℒ_dice(proj_x(b),proj_x(m̂))+ ℒ_dice(proj_y(b),proj_y(m̂)). loss_pairwiseℒ_pairwise = -1/N∑_e ∈E_in1_{S_e ≥τ}logP(y_e = 1). y_e P(y_e = 1) = m̂_i, j ·m̂_k, l + (1 - m̂_i, j) ·(1 - m̂_k, l). S_eS_e = S(c_i, j, c_l, k) = exp(-||c_i, j - c_l, k||/θ), where y_e=1 means the two pixels have the same ground-truth label. S_e is the color similarity of the edge e. c_i,j and c_l,k are respectively the LAB color vectors of the two pixels (i, j) and (l, k) linked by the edge. θ is 2 in this work. ℒ_embed. UNINEXT uses contrastive loss <cit.> to train discriminative embeddings for associating instances on different frames. loss_embedℒ_embed = log[1+∑_𝐤^+∑_𝐤^-exp(𝐯 ·𝐤^- - 𝐯 ·𝐤^+) ], where k^+ and k^- are positive and negative feature embeddings from the reference frame. For each instance in the key frame, v is the feature embedding with the lowest cost. §.§ Network Architecture To transform the enhanced visual features F^'_v and prompt features F^'_p into the final instance predictions, an encoder-decoder Transformer architecture is adopted. Based on the original architecture in two-stage Deformable DETR <cit.>, UNINEXT makes the following improvements: * Introducing a mask head for segmentation. To predict high-quality masks, UNINEXT introduces a mask head <cit.> based on dynamic convolutions. Specifically, first an MLP is used to transform instance embeddings into a group of parameters ω. Then these parameters are used to perform three-layer 1×1 convolutions with feature maps, obtaining masks of instances. * Replacing one-to-one Hungarian matching with one-to-many SimOTA <cit.>. Traditional Hungarian matching forces one GT to be only assigned to one query, leaving most of the queries negative. UNINEXT uses SimOTA <cit.>, which enables multiple queries to be matched with one GT. This strategy can provide more positive samples and speed up convergence. During inference, UNINEXT uses NMS to remove duplicated predictions. * Adding an IoU branch. UNINEXT adds an IoU branch to reflect the quality of the predicted boxes. During training, IoU does not affect the label assignment. During inference, the final scores are the geometric mean of the instance-prompt matching scores (after sigmoid) and the IoU scores. * Adding some techniques in DINO <cit.>. To further improve the performance, UNINEXT introduces some techniques <cit.>, including contrastive DN, mixed query selection, and look forward twice. §.§ Analysis and Visualizations Analysis. We compare UNINEXT with other competitive counterparts, which can handle multiple instance-level perception tasks. The opponents include Cascade Mask R-CNN <cit.> for object detection and instance segmentation, SeqTR <cit.> for REC and RES, VMT <cit.> for MOTS and VIS, and Unicorn <cit.> for SOT, VOS, MOT, and MOTS. As shown in Figure <ref>, UNINEXT outperforms them and achieve state-of-the-art performance on all 10 tasks. Retrieval by Category Names. As shown in Figure <ref>, UNINEXT can flexibly detect and segment objects of different categories by taking the corresponding category names as the prompts. For example, when taking “dining table. wine glass. cake. knife” as the prompts, UNINEXT would only perceive dining tables, wine glasses, cakes, and knives. Furthermore, benefiting from the flexible retrieval formulation, UNINEXT also has the potential for zero-shot (open-vocabulary) object detection. However, open-vocabulary object detection is beyond the scope of our paper and we leave it for future works. Retrieval by Language Expressions. We provide some visualizations for retrieval by language expressions in Figure <ref>. UNINEXT can accurately locate the target referred by the given language expression when there are many similar distractors. This demonstrates that our method can not only perceive objects but also understand their relationships in positions (left, middle, right, etc) and sizes (taller, etc). Retrieval by Target Annotations. Our method supports annotations in formats of both boxes (SOT) and masks (VOS). Although there is only box-level annotation for SOT, we obtain the target prior by filling the region within the given box with 1 and leaving other regions as 0. As shown in Figure <ref>, UNINEXT can precisely track and segment the targets in complex scenarios, given the annotation in the first frame.

http://arxiv.org/abs/2303.07109v1

20230313134359

Transformer-based World Models Are Happy With 100k Interactions

cs.LG

Upcycling Models under Domain and Category Shift Sanqing Qu^1Equal Contribution , Tianpei Zou^1, Florian Röhrbein^2, Cewu Lu^3, Guang Chen^1Corresponding author: guangchen@tongji.edu.cn , Dacheng Tao^4,5, Changjun Jiang^1 ^1Tongji University, ^2Chemnitz University of Technology, ^3Shanghai Jiao Tong University, ^4JD Explore Academy, ^5The University of Sydney March 30, 2023 ================================================================================================================================================================================================================================================================================================================================= Deep neural networks have been successful in many reinforcement learning settings. However, compared to human learners they are overly data hungry. To build a sample-efficient world model, we apply a transformer to real-world episodes in an autoregressive manner: not only the compact latent states and the taken actions but also the experienced or predicted rewards are fed into the transformer, so that it can attend flexibly to all three modalities at different time steps. The transformer allows our world model to access previous states directly, instead of viewing them through a compressed recurrent state. By utilizing the Transformer-XL architecture, it is able to learn long-term dependencies while staying computationally efficient. Our transformer-based world model (TWM) generates meaningful, new experience, which is used to train a policy that outperforms previous model-free and model-based reinforcement learning algorithms on the Atari 100k benchmark. Our code is available at <https://github.com/jrobine/twm>. § INTRODUCTION Deep reinforcement learning methods have shown great success on many challenging decision making problems. Notable methods include DQN <cit.>, PPO <cit.>, and MuZero <cit.>. However, most algorithms require hundreds of millions of interactions with the environment, whereas humans often can achieve similar results with less than 1% of these interactions, i.e., they are more sample-efficient. The large amount of data that is necessary renders a lot of potential real world applications of reinforcement learning impossible. Recent works have made a lot of progress in advancing the sample efficiency of RL algorithms: model-free methods have been improved with auxiliary objectives <cit.>, data augmentation (, ), or both <cit.>. Model-based methods have been successfully applied to complex image-based environments and have either been used for planning, such as EfficientZero <cit.>, or for learning behaviors in imagination, such as SimPLe <cit.>. A promising model-based concept is learning in imagination (; ; ; ): instead of learning behaviors from the collected experience directly, a generative model of the environment dynamics is learned in a manner. Such a so-called world model can create new trajectories by iteratively predicting the next state and reward. This allows for potentially indefinite training data for the reinforcement learning algorithm without further interaction with the real environment. A world model might be able to generalize to new, unseen situations, because of the nature of deep neural networks, which has the potential to drastically increase the sample efficiency. This can be illustrated by a simple example: in the game of Pong, the paddles and the ball move independently. In the best case, a successfully trained world model would imagine trajectories with paddle and ball configurations that have never been observed before, which enables learning of improved behaviors. In this paper, we propose to model the world with transformers <cit.>, which have significantly advanced the field of natural language processing and have been successfully applied to computer vision tasks <cit.>. A transformer is a sequence model consisting of multiple self-attention layers with residual connections. In each self-attention layer the inputs are mapped to keys, queries, and values. The outputs are computed by weighting the values by the similarity of keys and queries. Combined with causal masking, which prevents the self-attention layers from accessing future time steps in the training sequence, transformers can be used as autoregressive generative models. The Transformer-XL architecture <cit.> is much more computationally efficient than vanilla transformers at inference time and introduces relative positional encodings, which remove the dependence on absolute time steps. Our contributions: The contributions of this work can be summarized as follows: * We present a new autoregressive world model based on the Transformer-XL <cit.> architecture and a model-free agent trained in latent imagination. Running our policy is computationally efficient, as the transformer is not needed at inference time. This is in contrast to related works <cit.> that require the full world model during inference. * Our world model is provided with information on how much reward has already been emitted by feeding back predicted rewards into the world model. As shown in our ablation study, this improves performance. * We rewrite the balanced KL divergence loss of <cit.> to allow us to fine-tune the relative weight of the involved entropy and cross-entropy terms. * We introduce a new thresholded entropy loss that stabilizes the policy's entropy during training and hereby simplifies the selection of hyperparameters that behave well across different games. * We propose a new effective sampling procedure for the growing dataset of experience, which balances the training distribution to shift the focus towards the latest experience. We demonstrate the efficacy of this procedure with an ablation study. * We compare our transformer-based world model (TWM) on the Atari 100k benchmark with recent sample-efficient methods and obtain excellent results. Moreover, we report empirical confidence intervals of the aggregate metrics as suggested by <cit.>. § METHOD We consider a partially observable Markov decision process (POMDP) with discrete time steps , scalar rewards , high-dimensional image observations , and discrete actions , which are generated by some policy , where o_1:t and a_1:t-1 denote the sequences of observations and actions up to time steps t and , respectively. Episode ends are indicated by a boolean variable . Observations, rewards, and episode ends are jointly generated by the unknown environment dynamics . The goal is to find a policy π that maximizes the expected sum of discounted rewards , where is the discount factor. Learning in imagination consists of three steps that are repeated iteratively: learning the dynamics, learning a policy, and interacting in the real environment. In this section, we describe our world model and policy, concluding with the training procedure. §.§ World Model Our world model consists of an observation model and a dynamics model, which do not share parameters. <ref> illustrates our combined world model architecture. Observation Model: The observation model is a variational autoencoder <cit.>, which encodes observations o_t into compact, stochastic latent states z_t and reconstructs the observations with a decoder, which in our case is only required to obtain a learning signal for z_t: 0.1em 3 Observation encoder: z_t ∼(z_t o_t) Observation decoder: ô_t ∼(ô_t z_t). We adopt the neural network architecture of DreamerV2 <cit.> with slight modifications for our observation model. Thus, a latent state z_t is discrete and consists of a vector of 32 categorical variables with 32 categories. The observation decoder reconstructs the observation and predicts the means of independent standard normal distributions for all pixels. The role of the observation model is to capture only non-temporal information about the current time step, which is different from <cit.>. However, we include short-time temporal information, since a single observation o_t consists of four frames (aka frame stacking, see also <ref>). Autoregressive Dynamics Model: The dynamics model predicts the next time step conditioned on the history of its past predictions. The backbone is a deterministic aggregation model which computes a deterministic hidden state h_t based on the history of the previously generated latent states, actions, and rewards. Predictors for the reward, discount, and next latent state are conditioned on the hidden state. The dynamics model consists of these components: 0.1em 3 Aggregation model: h_t = (z_t-:t,a_t-:t,r_t-:t-1) Reward predictor: r̂_t ∼(r̂_t h_t) Discount predictor: γ̂_t ∼(γ̂_t h_t) Latent state predictor: ẑ_t+1 ∼(ẑ_t+1 h_t). The aggregation model is implemented as a causally masked Transformer-XL <cit.>, which enhances vanilla transformers <cit.> with a recurrence mechanism and relative positional encodings. With these encodings, our world model learns the dynamics independent of absolute time steps. Following <cit.>, the latent states, actions, and rewards are sent into modality-specific linear embeddings before being passed to the transformer. The number of input tokens is 3 - 1, because of the three modalities (latent states, actions, rewards) and the last reward not being part of the input. We consider the outputs of the action modality as the hidden states and disregard the outputs of the other two modalities (see <ref>; orange boxes vs. gray boxes). The latent state, reward, and discount predictors are implemented as multilayer perceptrons (MLPs) and compute the parameters of a vector of independent categorical distributions, a normal distribution, and a Bernoulli distribution, respectively, conditioned on the deterministic hidden state. The next state is determined by sampling from (ẑ_t+1 h_t). The reward and discount are determined by the mean of (r̂_t h_t) and (γ̂_t h_t), respectively. As a consequence of these design choices, our world model has the following beneficial properties: * The dynamics model is autoregressive and has direct access to its previous outputs. * Training is efficient since sequences are processed in parallel (compared with RNNs). * Inference is efficient because outputs are cached (compared with vanilla Transformers). * Long-term dependencies can be captured by the recurrence mechanism. We want to provide an intuition on why a fully autoregressive dynamics model is favorable: First, the direct access to previous latent states enables to model more complex dependencies between them, compared with RNNs, which only see them indirectly through a compressed recurrent state. This also has the potential to make inference more robust, since degenerate predictions can be ignored more easily. Second, because the model sees which rewards it has produced previously, it can react to its own predictions. This is even more significant when the rewards are sampled from a probability distribution, since the introduced noise cannot be observed without autoregression. Loss Functions: The observation model can be interpreted as a variational autoencoder with a temporal prior, which is provided by the latent state predictor. The goal is to keep the distributions of the encoder and the latent state predictor close to each other, while slowly adapting to new observations and dynamics. <cit.> apply a balanced KL divergence loss, which lets them control which of the two distributions should be penalized more. To control the influences of its subterms more precisely, we disentangle this loss and obtain a balanced cross-entropy loss that computes the cross-entropy (z_t+1 o_t+1),(ẑ_t+1 h_t) and the entropy (z_t o_t) explicitly. Our derivation can be found in <ref>. We call the cross-entropy term for the observation model the consistency loss, as its purpose is to prevent the encoder from diverging from the dynamics model. The entropy regularizes the latent states and prevents them from collapsing to one-hot distributions. The observation decoder is optimized via negative log-likelihood, which provides a rich learning signal for the latent states. In summary, we optimize a self-supervised loss function for the observation model that is the expected sum over the decoder loss, the entropy regularizer and the consistency loss _^Obs. = [][] _t=1^T -ln(o_t z_t)_decoder - (z_t o_t)_entropy regularizer + (z_t o_t),(ẑ_t h_t-1)_consistency, where the hyperparameters , ≥ 0 control the relative weights of the terms. For the balanced cross-entropy loss, we also minimize the cross-entropy in the loss of the dynamics model, which is how we train the latent state predictor. The reward and discount predictors are optimized via negative log-likelihood. This leads to a self-supervised loss for the dynamics model _^Dyn. = [][] _t=1^T (z_t+1 o_t+1),(ẑ_t+1 h_t)_latent state predictor - ln(r_t h_t)_reward predictor - ln(γ_t h_t)_discount predictor , with coefficients , ≥ 0 and where γ_t = 0 for episode ends (d_t = 1) and γ_t = γ otherwise. §.§ Policy Our policy π_θ(a_t ẑ_t) is trained on imagined trajectories using a mainly standard advantage actor-critic <cit.> approach. We train two separate networks: an actor a_t ∼(a_t ẑ_t) with parameters and a critic (ẑ_t) with parameters . We compute the advantages via Generalized Advantage Estimation <cit.> while using the discount factors predicted by the world model γ̂_t instead of a fixed discount factor for all time steps. As in DreamerV2 <cit.>, we weight the losses of the actor and the critic by the cumulative product of the discount factors, in order to softly account for episode ends. Thresholded Entropy Loss: We penalize the objective of the actor with a slightly modified version of the usual entropy regularization term <cit.>. Our penalty normalizes the entropy and only takes effect when the entropy falls below a certain threshold ℒ^Ent._θ =max(0, - H(π_θ)/ln(m)), where 0≤Γ≤ 1 is the threshold hyperparameter, H(π_θ) is the entropy of the policy, m is the number of discrete actions, and ln(m) is the maximum possible entropy of the categorical action distribution. By doing this, we explicitly control the percentage of entropy that should be preserved across all games independent of the number of actions. This ensures exploration in the real environment and in imagination without the need for ϵ-greedy action selection or changing the temperature of the action distribution. We also use the same stochastic policy when evaluating our agent in the experiments. The idea of applying a hinge loss to the entropy was first introduced by <cit.> in the context of supervised learning. In <ref> we show the effect of this loss. Choice of Policy Input: The policy computes an action distribution π_θ(a_t x_t) given some view x_t of the state. For instance, x_t could be o_t, z_t, or [z_t,h_t] at inference time, i.e., when applied to the real environment, or the respective predictions of the world model ô_t, ẑ_t, or [ẑ_t,h_t] at training time. This view has to be chosen carefully, since it can have a significant impact on the performance of the policy and it affects the design choices for the world model. Using x_t = o_t (or ô_t) is relatively stable even with imperfect reconstructions ô_t, as the underlying distribution of observations p(o) does not change during training. However, it is also less computationally efficient, since it requires reconstructing the observations during imagination and additional convolutional layers for the policy. Using x_t = z_t (or ẑ_t) is slightly less stable, as the policy has to adopt to the changes of the distributions (z_t o_t) and (ẑ_t+1 h_t) during training. Nevertheless, the entropy regularizer and consistency loss in <ref> stabilize these distributions. Using (or [ẑ_t,h_t]) provides the agent with a summary of the history of experience, but it also adds the burden of running the transformer at inference time. Model-free agents already perform well on most Atari games when using a stack of the most recent frames (e.g., ). Therefore, we choose and apply frame stacking at inference time in order to incorporate short-time information directly into the latent states. At training time we use x_t = ẑ_t, i.e., the predicted latent states, meaning no frame stacking is applied. As a consequence, our policy is computationally efficient at training time (no reconstructions during imagination) and at inference time (no transformer when running in the real environment). §.§ Training As is usual for learning with world models, we repeatedly (i) collect experience in the real environment with the current policy, (ii) improve the world model using the past experience, (iii) improve the policy using new experience generated by the world model. During training we build a dataset = [(o_1, a_1, r_1, d_1), …, (o_, a_, r_, d_)] of the collected experience. After collecting new experience with the current policy, we improve the world model by sampling sequences of length from and optimizing the loss functions in <ref> using stochastic gradient descent. After performing a world model update, we select of the × observations and encode them into latent states to serve as initial states for new trajectories. The dynamics model iteratively generates these trajectories of length based on actions provided by the policy. Subsequently, the policy is improved with standard model-free objectives, as described in <ref>. In <ref> we present pseudocode for training the world model and the policy. Balanced Dataset Sampling: Since the dataset grows slowly during training, uniform sampling of trajectories focuses too heavily on early experience, which can lead to overfitting especially in the low data regime. Therefore, we keep visitation counts v_1, …, v_, which are incremented every time an entry is sampled as start of a sequence. These counts are converted to probabilities using the softmax function (p_1, …, p_T) = softmax(-v_1, …, -v_), where >0 is a temperature hyperparameter. With our sampling procedure, new entries in the dataset are oversampled and are selected more often than old ones. Setting = ∞ restores uniform sampling as a special case, whereas reducing increases the amount of oversampling. See <ref> for a comparison. We empirically show the effectiveness in <ref>. § EXPERIMENTS To compare data-efficient reinforcement learning algorithms, <cit.> proposed the Atari 100k benchmark, which uses a subset of 26 Atari games from the Arcade Learning Environment <cit.> and limits the number of interactions per game to 100K. This corresponds to 400K frames (because of frame skipping) or roughly 2 hours of gameplay, which is 500 times less than the usual 200 million frames (e.g., ). We compare our method with five strong competitors on the Atari 100k benchmark: (i) SimPLe <cit.> implements a world model as an action-conditional video prediction model and trains a policy with PPO <cit.>, (ii) DER <cit.> is a variant of Rainbow <cit.> fine-tuned for sample efficiency, (iii) CURL <cit.> improves representations using contrastive learning as an auxiliary task and is combined with DER, (iv) DrQ <cit.> improves DQN by averaging Q-value estimates over multiple data augmentations of observations, and (v) SPR <cit.> forces representations to be consistent across multiple time steps and data augmentations by extending Rainbow with a self-supervised consistency loss. §.§ Results r0.45 < g r a p h i c s > Performance profiles on the Atari 100k benchmark based on score distributions <cit.>. It shows the fraction of runs across all games (y-axis) above a human normalized score (x-axis). Shaded regions show pointwise 95% confidence bands. We follow the advice of <cit.> who found significant discrepancies between reported point estimates of mean (and median) scores and a thorough statistical analysis that includes statistical uncertainty. Thus, we report confidence interval estimates of the aggregate metrics median, interquartile mean (IQM), mean, and optimality gap in <ref> and performance profiles in <ref>, which we created using the toolbox provided by <cit.>. The metrics are computed on human normalized scores, which are calculated as . We report the unnormalized scores per game in <ref>. We compare with new scores for DER, CURL, DrQ, and SPR that were evaluated on 100 runs and provided by <cit.>. They report scores for the improved DrQ(ε), which is DrQ evaluated with standard ε-greedy parameters. We perform 5 runs per game and compute the average score over 100 episodes at the end of training for each run. TWM shows a significant improvement over previous approaches in all four aggregate metrics and brings the optimality gap closer to zero. §.§ Analysis In <ref> we show imagined trajectories of our world model. In <ref> we visualize an attention map of the transformer for an imagined sequence. In this example a lot of weight is put on the current action and the last three states. However, the transformer also attends to states and rewards in the past, with only past actions being mostly ignored. The two high positive rewards also get high attention, which confirms that the rewards in the input sequence are used by the world model. We hypothesize that these rewards correspond to some events that happened in the environment and this information can be useful for prediction. An extended analysis can be found in <ref>, including more imagined trajectories and attention maps (and a description of the generation of the plots), sample efficiency, stochasticity of the world model, long sequence imagination, and frame stacking. §.§ Ablation Studies Uniform Sampling: To show the effectiveness of the sampling procedure described in <ref>, we evaluate three games with uniform dataset sampling, which is equivalent to setting = ∞ in <ref>. In <ref> we show that balanced dataset sampling significantly improves the performance in these games. At the end of training, the dynamics loss from <ref> is lower when applying balanced sampling. One reason might be that the world model overfits on early training data and performs bad in later stages of training. No Rewards: As described in <ref>, the predicted rewards are fed back into the transformer. In <ref> we show on three games that this can significantly increase the performance. In some games the performance is equivalent, probably because the world model can make correct predictions solely based on the latent states and actions. In <ref> we perform additional ablation studies, including the thresholded entropy loss, a shorter history length, conditioning the policy on [z,h], and increasing the sample efficiency. § RELATED WORK The Dyna architecture <cit.> introduced the idea of training a model of the environment and using it to further improve the value function or the policy. <cit.> introduced the notion of a world model, which tries to completely imitate the environment and is used to generate experience to train a model-free agent. They implement a world model as a VAE <cit.> and an RNN and learn a policy in latent space with an evolution strategy. With SimPLe, <cit.> propose an iterative training procedure that alternates between training the world model and the policy. Their policy operates on pixel-level and is trained using PPO <cit.>. <cit.> present Dreamer and implement a world model as a stochastic RNN that splits the latent state in a stochastic part and a deterministic part; this idea was first introduced by . This allows their world model to capture the stochasticity of the environment and simultaneously facilitates remembering information over multiple time steps. <cit.> use a VQ-VAE to construct a world model with drastically lower number of parameters. DreamerV2 <cit.> achieves great performance on the Atari 50M benchmark after making some changes to Dreamer, the most important ones being categorical latent variables and an improved objective. Another direction of model-based reinforcement learning is planning, where the model is used at inference time to improve the action selection by looking ahead several time steps into the future. The most prominent work is MuZero <cit.>, where a learned sequence model of rewards and values is combined with Monte-Carlo Tree Search <cit.> without learning explicit representations of the observations. MuZero achieves impressive performance on the Atari 50M benchmark, but it is also computationally expensive and requires significant engineering effort. EfficientZero <cit.> improves MuZero and achieves great performance on the Atari 100k benchmark. Transformers <cit.> advanced the effectiveness of sequence models in multiple domains, such as natural language processing and computer vision <cit.>. Recently, they have also been applied to reinforcement learning tasks. The Decision Transformer <cit.> and the Trajectory Transformer <cit.> are trained on an offline dataset of trajectories. The Decision Transformer is conditioned on states, actions, and returns, and outputs optimal actions. The Trajectory Transformer trains a sequence model of states, actions, and rewards, and is used for planning. <cit.> replace the RNN of Dreamer with a transformer and outperform Dreamer on Hidden Order Discovery tasks. However, their transformer has no access to previous rewards and they do not evaluate their method on the Atari 100k benchmark. Moreover, their policy depends on the outputs of the transformer, leading to higher computational costs during inference time. Concurrent to and independent from our work, <cit.> apply a transformer to sequences of frame tokens and actions and achieve state-of-the-art results on the Atari 100k benchmark. § CONCLUSION In this work, we discuss a reinforcement learning approach using transformer-based world models. Our method (TWM) outperforms previous model-free and model-based methods in terms of human normalized score on the 26 games of the Atari 100k benchmark. By using the transformer only during training, we were able to keep the computational costs low during inference, i.e., when running the learned policy in a real environment. We show how feeding back the predicted rewards into the transformer is beneficial for learning the world model. Furthermore, we introduce the balanced cross-entropy loss for finer control over the trade-off between the entropy and cross-entropy terms in the loss functions of the world model. A new thresholded entropy loss effectively stabilizes the entropy of the policy. Finally, our novel balanced sampling procedure corrects issues of naive uniform sampling of past experience. iclr2023_conference § APPENDIX §.§ Extended Experiments Additional Analysis: * We provide more example trajectories in <ref>. * We present more attention plots in <ref>. All attention maps are generated using the attention rollout method by <cit.>. Note that we had to modify the method slightly, in order to take the causal masks into account. * Sample Efficiency: We provide the scores of our main experiments after different amounts of interactions with the environment in <ref>. After 50K interactions, our method already has a higher mean normalized score than previous sample-efficient methods. Our mean normalized score is higher than DER, CURL, and SimPLe after 25K interactions. This demonstrates the high sample efficiency of our approach. * Stochasticity: The stochastic prediction of the next state allows the world model to sample a variety of trajectories, even from the same starting state, as can be seen in <ref>. * Long Sequence Imagination: The world model is trained using sequences of length = 16, however, it generalizes well to very long trajectories, as shown in <ref>. * Frame Stacking: In <ref> we visualize the learned stacks of frames. This shows that the world model encodes and predicts the motion of objects. Additional Ablation Studies: * Thresholded Entropy Loss: In <ref> we compare (i) our thresholded entropy loss for the policy (see <ref>) with (ii) the usual entropy penalty. For (i) we use the same hyperparameters as in our main experiments, i.e., = 0.01 and = 0.1. For (ii) we set = 0.001 and = 1.0, which effectively disables the threshold. Without a threshold, the entropy is more likely to either collapse or diverge. When the threshold is used, the score is higher as well, probably because the entropy is in a more sensible range for the exploration-exploitation trade-off. This cannot be solved by adjusting the penalty coefficient alone, since it would increase or decrease the entropy in all games. * History Length: We trained our world model with a shorter history and set = 4 instead of = 16. This has a negative impact on the score, as can be seen in <ref>, demonstrating that more time steps into the past are important. * Choice of Policy Input: In <ref> we explained why the input to the policy is only the latent state, i.e., x = z. In <ref> we show that using x = [z, h] can result in lower final scores. We hypothesize that the policy network has a hard time keeping up with the changes of the space of h during training and cannot ignore this additional information. * Increasing the Sample Efficiency: To find out whether we can further increase the sample efficiency shown in <ref>, we train a random subset of games again on 10K, 25K, and 50K interactions with the full training budget that we used for the 100K interactions. In <ref> we see that this can lead to significant improvements in some cases, which could mean that the policy benefits from more training on imagined trajectories, but can even lead to worse performance in other cases, which could possibly be caused by overfitting of the world model. When the performance stays the same even with longer training, this could mean that better exploration in the real environment is required to get further improvements. Wall-Clock Times: For each run, we give the agent a total training and evaluation budget of roughly 10 hours on a single NVIDIA A100 GPU. The time can vary slightly, since the budget is based on the number of updates. An NVIDIA GeForce RTX 3090 requires 12-13 hours for the same amount of training and evaluation. When using a vanilla transformer, which does not use the memory mechanism of the Transformer-XL architecture <cit.>, the runtime is roughly 15.5 hours on an NVIDIA A100 GPU, i.e., 1.5 times higher. We compare the runtime of our method with previous methods in <ref>. Our method is more than 20 times faster than SimPLe, but slower than model-free methods. However, our method should be as fast as other model-free methods during inference. In <ref> we have shown that our method achieves a higher human normalized score than previous sample-efficient methods after 50K interactions. This suggests that our method could potentially outperform previous methods with shorter training, which would take less than 23.3 hours. To determine how time-consuming the individual parts of our method are, we investigate the throughput of the models, with the batch sizes of our main experiments. The Transformer-XL version is almost twice as fast, which again shows the importance of this design choice. The throughputs were measured on an NVIDIA A100 GPU and are given in (approximate) samples per second: * World model training: 16,800 samples/s * World model imagination (Transformer-XL): 39,000 samples/s * World model imagination (vanilla): 19,900 samples/s * Policy training: 700,000 samples/s We also examine how fast the policy can run in an Atari game. We measured the (approximate) frames per second on a CPU (since the batch size is 1). Conditioning the policy on [z, h] is about 3 times slower than z, since the transformer is required: * Policy conditioned on z: 653 frames/s * Policy conditioned on [z,h]: 213 frames/s §.§ Derivation of Balanced Cross-Entropy Loss <cit.> propose to use a balanced KL divergence loss to jointly optimize the observation encoder q_θ and state predictor p_θ with shared parameters θ, i.e., λ q_θp_θ + (1 - λ) q_θp_θ, where · denotes the stop-gradient operation and λ∈ [0,1] controls how much the state predictor adapts to the observation encoder and vice versa. We use the identity qp = q,p - q, where q,p is the cross-entropy of distribution p relative to distribution q, and show that our loss functions lead to the same gradients as the balanced KL objective, but with finer control over the individual components: ∇_θ[ λ q_θp_θ + (1 - λ) q_θp_θ] = ∇_θ[ λ(q_θ,p_θ - q_θ) + (1 - λ) ( q_θ,p_θ - q_θ) ] = ∇_θ[ λ_1 q_θ,p_θ + λ_2 q_θ,p_θ - λ_3 q_θ], since ∇_θq_θ = 0 and by defining λ_1 = λ and λ_2 = 1 - λ and λ_3 = 1 - λ. In this form, we have control over the cross-entropy of the state predictor relative to the observation encoder and vice versa. Moreover, we explicitly penalize the entropy of the observation encoder, instead of being entangled inside of the KL divergence. As common in the literature, we define the loss function by omitting the gradient in <ref>, so that automatic differentiation computes this gradient. For our world model, we split the objective into two loss functions, as the observation encoder and state predictor have separate parameters, yielding <ref>. §.§ Additional Training Details In <ref> we present pseudocode for training the world model and the actor-critic agent. We use the SiLU activation function <cit.> for all models. In <ref> we summarize all hyperparameters that we used in our experiments. In <ref> we provide the number of parameters of our models. Pretraining for Better Initialization: During training we need to correctly balance the amount of world model training and policy training, since the policy has to keep up with the distributional shift of the latent space. However, we can spend some extra training time on the world model with pre-collected data (included in the 100K interactions) at the beginning of training in order to obtain a reasonable initialization for the latent states.

http://arxiv.org/abs/2303.08011v1

20230313030317

Large statistical learning models effectively forecast diverse chaotic systems

cs.LG

wgilpin@utexas.edu Department of Physics, The University of Texas at Austin, Austin, Texas 78712, USA Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, Texas 78712, USA Chaos and unpredictability are traditionally synonymous, yet recent advances in statistical forecasting suggest that large machine learning models can derive unexpected insight from extended observation of complex systems. Here, we study the forecasting of chaos at scale, by performing a large-scale comparison of 24 representative state-of-the-art multivariate forecasting methods on a crowdsourced database of 135 distinct low-dimensional chaotic systems. We find that large, domain-agnostic time series forecasting methods based on artificial neural networks consistently exhibit strong forecasting performance, in some cases producing accurate predictions lasting for dozens of Lyapunov times. Best-in-class results for forecasting chaos are achieved by recently-introduced hierarchical neural basis function models, though even generic transformers and recurrent neural networks perform strongly. However, physics-inspired hybrid methods like neural ordinary equations and reservoir computers contain inductive biases conferring greater data efficiency and lower training times in data-limited settings. We observe consistent correlation across all methods despite their widely-varying architectures, as well as universal structure in how predictions decay over long time intervals. Our results suggest that a key advantage of modern forecasting methods stems not from their architectural details, but rather from their capacity to learn the large-scale structure of chaotic attractors. Large statistical learning models effectively forecast diverse chaotic systems William Gilpin March 30, 2023 ============================================================================== § INTRODUCTION Chaos traditionally implies the butterfly effect: that a small change to a system's inputs grows exponentially over time, complicating efforts to reliably predict the system's long-term evolution. Yet recent efforts to generate long-term forecasts of real-world complex systems using statistical learning undermine this intuition, providing compelling examples of successful data-driven predictions of diverse systems such as cellular signaling pathways <cit.>, hourly precipitation forecasts <cit.>, and tokamak plasma disruptions <cit.>. These successes complement recent works showing that large, overparameterized statistical learning methods can learn latent representations of complex systems that implicitly linearize dynamics, raising the possibility of forecasting even strongly-chaotic systems using models with sufficient capacity <cit.>. Such results potentially elucidate the emergence of reservoir computers as best-in-class forecasting methods for dynamical systems <cit.>; these models use emergent properties of random networks to lift complex time series into a random feature space, substantially reducing training requirements by reducing learning to linear regression <cit.>. However, other recently-introduced hybrid models successfully forecast chaotic systems by encoding dynamical constraints within model formulation <cit.>; among these are neural ordinary differential equations <cit.>, physics-informed neural networks <cit.>, and recurrent neural networks with domain-specific architectures <cit.>. However, there is little consensus regarding the best choice of forecasting model for a given dynamical system, and whether dynamical constraints should be encoded, or learned from data via large domain-agnostic models—the former provides an inductive bias that economizes training, while the latter allows greater flexibility and reduces model selection bias. Complicating comparisons among methods are the widely-varying contexts in which prior methods have been evaluated, which vary from domain-specific weather or medical time series to individual well-known chaotic systems like the Lorenz attractor <cit.>. A larger set of representative systems, and a controlled comparison among methods, is necessary to disentangle the relationship between chaoticity, predictability, and forecasting model architecture, as well as to understand how the properties of different black-box machine learning methods interact with the systems they predict. Here, we seek to systematically quantify the relationship between chaos and empirical predictability in large-scale, controlled experiments. We recently introduced a benchmark dataset containing 135 low-dimensional differential equations describing known chaotic attractors <cit.>. Initially curated from published works to include well-known systems such as the Lorenz, Rössler, and Chua attractors, the dataset has grown through crowdsourcing to include examples spanning diverse domains such as climatology, neuroscience, and astrophysics. Each dynamical system is aligned with respect to its dominant timescale and integration timestep using surrogate significance testing <cit.>, and is annotated with calculations of its invariant properties such as the Lyapunov exponent spectrum, fractal dimension, and metric entropy. While each system has a different largest Lyapunov exponent (λ_max), and thus putative chaoticity, some systems are more closely related than others. For example, the Sprott attractor subfamily (λ_max∈ [0.01, 1.1]) exhibit similar qualitative structure such as paired lobes, owing to the presence of predominantly quadratic nonlinearities in the governing differential equations <cit.>. To identify such relationships across our dataset, we convert each dynamical system into a high-dimensional vector by first generating a long trajectory, and then computing 747 characteristic mathematical signal properties such as the metric entropy, power spectral coefficients, Hurst exponents, et al. that are invariant to the initial conditions and sampling rate <cit.>. We then use uniform manifold approximation and projection (UMAP) to visualize these high-dimensional vectors within a two-dimensional plane (Fig. <ref>) <cit.>. The resulting space of chaotic systems shows clear structure, with the Sprott and other scroll-like subfamilies clustering together, while other systems separate. These results suggest that our dataset contains high dynamical diversity beyond absolute chaoticity λ_max, which correlates only weakly with the embedding (ρ = 0.15 ± 0.03, bootstrapped Spearman rank-order coefficient). Our dataset allows us systematically compare the forecasting ability of different statistical forecasting methods across diverse dynamical systems. Forecasting dynamical systems from observations is a well-established field <cit.>, and we structure our experiments as a standard long-term autoregressive forecasting task <cit.>. For each dynamical system, we define two time series Y_train, Y_test∈ℝ^T × D corresponding to multivariate trajectories emanating from different initial conditions, and we define a splitting time such that T = T_past + T_fut, Y_train = Y_train^past∪ Y_train^fut, etc. The model parameters are first fit using Y_train^past, and the accuracy of the resulting predictions on Y_train^fut are used for model selection and hyperparameter tuning. After completing model selection, the final model from each model class is then fit on Y_test^past, and the resulting predictions on Y_test^fut produce the error score ϵ_ik(t) representing the performance of the i^th forecasting model on the k^th dynamical system at future time t ∈ [0, T_fut] after the end of training data availability. We compute 17 different accuracy metrics, including root mean-squared error, pointwise correlation, mutual information, and Granger causality. We report our results in the main text in terms of the symmetric mean absolute percent error (sMAPE) ϵ(t) ≡ (2/N) ∑_t'=1^t (| y̌(t') - ŷ̌̂(t')|) / (|y̌(t')| + |ŷ̌̂(t')|) due to its common use, conceptual simplicity, and correlation with other metrics <cit.>. Each forecasting method represents a class of possible models parametrized by choices made regarding architecture (model size, number of layers, activation function) or training (optimization epochs, batch sizes). Such choices can strongly affect the performance of different methods <cit.>, yet different methods do not necessarily have equivalent adjustable hyperparameters. In order to fairly compare different methods, we restrict hyperparameter tuning to the equivalent of the input time window for each model—such as the input size for deep neural networks, the order of autoregressive models, or the number of time lags for state space models. We tune hyperparameters separately on each system and forecasting model pair. Our overall experiment design is characterized by several timescales: the lookback window hyperparameter corresponds to the number of past timepoints seen simultaneously by a model at a given time during training or prediction; the history length represents the total number of timepoints available to learn the model's parameters before prediction; the forecast horizon represents the number of unseen timepoints into the future that are predicted autoregressively; and the Lyapunov time λ_max^-1 is an invariant property of each distinct dynamical system, representing the characteristic timescale over which forecasts are expected to lose accuracy due to the butterfly effect. We evaluate 24 statistical forecasting models across all 135 dynamical systems. We choose forecasting methods representative of the broad diversity of methods available in the recent literature <cit.>. Traditional methods include standard linear regression, autoregressive moving averages (ARIMA), exponential smoothing, Kalman filtering, Fourier mode extrapolation, boosted random forest models <cit.>, and newly-introduced linear models that account for trends and distribution shift <cit.>. Current state-of-the-art models for general time series forecasting are based on deep neural networks: the transformer model <cit.>, long-short-term-memory networks (LSTM), block recurrent neural networks (RNN), temporal convolutional neural networks <cit.>, and neural basis expansion/neural hierarchical interpolation (NBEATS/NHiTS) <cit.>. The latter methods generates forecasts hierarchically by aggregating separate forecasts at distinct timescales (NBEATS), and can explicitly coarse-grain the time series to further reduce computational costs (NHiTS). We also consider hybrid physics-motivated methods such as neural ordinary differential equations <cit.>, which approximate the continuous-time differential equation underlying time series, as well as echo-state networks and their generalization, nonlinear vector autoregressive models (nVAR), which use a trainable linear recurrent structure with fixed "reservoir" of random nonlinearities <cit.>. Our main forecasting results are summarized in Figure <ref>. Our long-term forecasting experiments require ∼10^18 floating point operations for training, model selection and hyperparameter tuning, a figure comparable to the scale of other recent large-scale machine learning benchmarks <cit.>. For brevity, we report here only results for the top 14 forecasting models in terms of the sMAPE error, and defer the full tabular results and accuracy metrics to the supplementary material and our open-source repository. We observe that large, domain-agnostic forecasting methods successfully forecast diverse chaotic systems, with the strongest methods succeeding consistently across diverse systems and forecasting horizons. The strong relative performance of machine learning models is highlighted in Fig <ref>C, where NBEATS successfully forecasts the Mackey-Glass equation for ∼ 22 Lyapunov times without losing track of the global phase. This and other strongly-performing methods, such as the transformer and LSTM, comprise large models originally designed for generic sequential datasets, and which do not make any assumptions regarding whether observed time series data arises from a dynamical system. This observation suggests that more flexible, generic architectures may prove preferable for problems where some physical structure is present (e.g. analytic generating functions in the form of ordinary differential equations) but stronger domain knowledge (e.g. symmetries or symplecity) is unavailable to further constrain learning. The strong performance of NBEATS and NHiTS suggest that these models may have structural features favoring the chaotic systems dataset. Because these methods compute forecasts hierarchically, they can flexibly integrate information across multiple timescales in a manner inaccessible to classical statistical models <cit.>. While chaotic systems exhibit continuous spectra and thus contain information relevant to forecasting at a variety of timescales, many systems exhibit topologically-preferred timescales such as unstable periodic orbits—like the "loops" on either side of the Lorenz attractor—that dominate the system's underlying measure <cit.>, and which therefore may represent higher priority motifs for learning. We thus argue that performant model architectures contain implicit inductive biases that advantage them on chaotic systems relative to other time series. Consistent with this finding, we note that the next strongest-performing model, nVAR, comprises a reservoir computing architecture <cit.>, which prior works have shown to perform particularly well on time series from dynamical systems <cit.>. Interestingly, the neural ordinary differential equation model performs poorly, despite putatively learning the underlying generator of the time series. We find that over long forecast horizons, neural ODE tend to settle into prolonged oscillations, favoring complex non-chaotic limit cycles over chaotic attractors given finite training data. While the utility of deep learning methods for forecasting general time series has been questioned <cit.>, our results agree with recent benchmarks suggesting that large models strongly outperform classical forecasting methods on long-horizon forecasting tasks <cit.>. We find that classical methods like exponential smoothing or ARIMA do not appear among the top 14 models, implying that the size and diversity of our chaotic systems dataset, as well as the long duration of the forecasting task, require larger models with greater intrinsic capacity to represent complex nonlinear systems. Relative performance among models remains stable across two orders of magnitude in Lyapunov time, indicating that strong models better approximate the underlying propagator for the flow even at small forecasting horizons. Given the autoregressive nature of forecasting, an initial accuracy advantage compounds over time due to the exponential sensitivity of chaotic systems to early errors. While the best forecasts are generated by domain-agnostic time series methods, we note that the different forecasting methods have different intrinsic model complexities and thus capacities. Fig. <ref>A shows the forecasting error at λ_max^-1 versus the computational walltime required to train each model on one central processing unit. Training walltime measures model efficiency, and we interpret it as a loose proxy for model complexity because model size and parameter count are not directly quantifiable across highly-distinct and regularized architectures <cit.>. We find that error and training time exhibit negative correlation (ρ = -0.21 ± 0.04, bootstrapped Spearman coefficient), which persists within most method groups. The best-performing machine learning models require considerable training times; in contrast, nVAR exhibits competitive performance with two orders of magnitude less training time due to its linear structure. The strong performance of reservoir computers on our chaotic systems task may indicate an inductive bias for learning complex dynamical systems, due to their disordered reservoir allowing them to more readily represent systems with continuous spectra <cit.>. This general tradeoff between performance and training time motivates us to search for universal similarities across different forecasting methods. In Fig. <ref>B we compute the time-resolved Spearman correlation between each method's forecast error and the average error of that method, ρ̃_i(t) = ρ(ϵ_ik(t), ⟨ϵ_ik⟩_t))_k. We find universal non-monotonic behavior, in which nearly all methods exhibit peak correlation at one Lyapunov time λ_max^-1 (Fig <ref>B). This observation underscores that the largest Lyapunov exponent represents an appropriate timescale for comparing different dynamical systems, and that diverse forecasting models interact with this property in a shared manner. Like other disordered complex systems, the trained models exhibit peak correlation at peak variance <cit.>, coinciding with when average forecast errors rapidly increase in Fig <ref>B. The critical forecast horizon λ_max^-1 is sufficiently long to distinguish different dynamical systems based on their invariant properties, but short enough that forecast methods do not accrue instabilities, large phase offsets, and other artifacts that saturate forecast error and mask intrinsic differences among systems. To further investigate model complexity and performance, we next perform a series of experiments in which we titrate history length, which determines the the total amount of training data available to each method before generating a forecast (Fig. <ref>C). Unlike measuring training time or parameter count, these experiments seek to determine how effectively different models utilize additional observations. As expected, all models asymptotically improve given more training data. However, the three best-performing models improve more quickly and reach lower asymptotes as the available history increases, suggesting that they better leverage additional data than less performant models. These results match the intuition that these models combine high intrinsic capacity with inductive biases for chaotic dynamical systems in order to outperform other methods on our dataset. Our results show that recently-developed large, overparameterized statistical forecasting models efficiently leverage long-term observations of chaotic attractors, producing best-in-class forecasts that can remain accurate for dozens of Lyapunov times. Commonalities in predictions across highly distinct model classes suggest that performance arises primarily from model capacity and generalization ability, rather than specific architectural choices, and that performance at long prediction times is ultimately limited by a model's ability to learn long-term properties of a dynamical system's underlying attractor. Our observations echo recent findings from other domains and represent an intuitive consequence of the "no free lunch" theorem for model selection <cit.>. Nonetheless, our results are practically informative for forecasting real-world time series driven by underlying dynamical systems. In the absence of restrictions on data availability or training resources, large domain-agnostic models are likely to produce high-quality forecasts without the need for system-specific knowledge. However, in restricted settings, models encoding strong domain knowledge, particularly recent reservoir computing architectures <cit.>, exhibit the strongest performance relative to their computational requirements. While NBEATS, NHiTS, and nVAR perform particularly well in our experiments, we refrain from endorsing these specific models to the detriment of other methods. Our results may be specific to our chaotic systems dataset and, more importantly, the recent literature contains a broad variety of new forecasting models, as well as infinite variations of each method due to hyperparameter and architectural choices, which could potentially exhibit strong performance. Rather, we have chosen representative set of forecasting models bridging different foci of the literature <cit.>, and aim to highlight general trends and the emerging strength of new models on the classical problem of forecasting chaos. In future work, we plan to relate our empirical forecasting results to invariant properties of the different dynamical systems in our dataset, such as various measures of fractality and entropy <cit.>. Such characterization could improve the interpretability of machine learning-based forecasting models, which ostensibly provide less insight into a time series's structure than classical methods <cit.>. However, the strong empirical performance of machine learning suggests the potential for these methods to reveal new properties of nonlinear dynamics and, ultimately, new bounds on the intrinsic predictability and thus reducibility of chaotic systems. § ACKNOWLEDGMENTS W. G. was supported by the University of Texas at Austin. Computational resources for this study were provided by the Texas Advanced Computing Center (TACC) at The University of Texas at Austin. § CODE AVAILABILITY All code used in this study is available online at <https://github.com/williamgilpin/dysts> naturemag

http://arxiv.org/abs/2303.07046v1

20230313121316

Deploying Offline Reinforcement Learning with Human Feedback

cs.LG

arrows ℝ ℤ 𝒜 ℬ 𝒟 ℰ ℱ 𝒢 ℋ 𝒮 ℐ ℒ ℳ 𝒩 𝒫 𝒮 𝒯 𝒱 𝒲 𝒵 𝕊 ℕ 𝐏 TV MSE 𝐰 𝐚 𝐙 thmTheorem[section] lemLemma[section] corCorollary[section] propProposition[section] asmpAssumption[section] defnDefinition[section] oracleOracle[section] factFact[section] conjConjecture[section] remRemark[section] conditionCondition[section] exerciseExercise[section] messMessage[section] thmTheoremTheorems lemLemmaLemmas corCorollaryCorollaries propPropositionPropositions asmpAssumptionAssumptions defnDefinitionDefinitions oracleOracleOracles factFactFacts conjConjectureConjectures remRemarkRemarks exampleExampleExamples conditionConditionConditions exerciseExerciseExercises algorithmAlgorithmAlgorithms tableTableTables figureFigureFigures sectionSectionSections subsectionSectionSections appendixAppendixAppendices messageMessageMessages @mathaccentsingle[3] 0"0362#1^H 2"03620pt#1^H 0=2 #3#2 kern[1]#1@kerna bar[2]@single#1@bar@#1#21@bar@#1#22 bar@[3] ##1##2 @mathaccent #32 @nucleus@char @@style@nucleus_ @@style@nucleus_ @@ @-@ @ 3 tempdima@ tempdima-tempdima 10 @-tempdima @>@ @0pt@kern0.6-@ #31 @kern-0.6@@nucleus@kern0.4@ @0.4@kerna @kern#2 @<@ @kern1@kern1 -@@kern-0.6@#1 @depthne @bgroupempty @egroup@set@skewchar @ @everymath@group @set@skewchar@nested@a #31 @nested@a111#1 @till@marker##1 @char@till@marker#1@char A@char @nested@a111@char #1figure <ref> #1Figure <ref> #1#2figures <ref> and <ref> #1#2#3#4figures <ref>, <ref>, <ref> and <ref> #1section <ref> #1Section <ref> #1#2sections <ref> and <ref> #1#2#3sections <ref>, <ref> and <ref> (<ref>)1(<ref>) #1Equation <ref> #1<ref> #1chapter <ref> #1Chapter <ref> #1#2chapters<ref>–<ref> #1algorithm <ref> #1Algorithm <ref> #1#2algorithms <ref> and <ref> #1#2Algorithms <ref> and <ref> #1part <ref> #1Part <ref> #1#2parts <ref> and <ref> #1⌈#1 ⌉ #1⌊#1 ⌋ 1 ϵ ϵ θ 𝐚 𝐛 𝐜 𝐝 𝐞 𝐟 𝐠 𝐡 𝐢 𝐣 𝐤 𝐥 𝐦 𝐧 𝐨 𝐩 𝐪 𝐫 𝐬 𝐭 𝐮 𝐯 𝐰 𝐱 𝐲 𝐳 a b c d e f g h i j k l m n o p q r s t u v w x y z 𝐀 𝐁 𝐂 𝐃 𝐄 𝐅 𝐆 𝐇 𝐈 𝐉 𝐊 𝐋 𝐌 𝐍 𝐎 𝐏 𝐐 𝐑 𝐒 𝐓 𝐔 𝐕 𝐖 𝐗 𝐘 𝐙 A B C D E F G H I J K L M N O P Q R S T U V W X Y Z 0 1 μ θ a b c d e f g h i j k l m n o p q r s t u v w x y z α β ϵ λ ω μ ψ σ θ a b c d e f g h i j k l m n o p q r s t u v w x y z A B C D E F G H I J K L M N O P Q R S T U V W X Y Z β Φ Λ Σ msl boldbxn A B C D E F G H I J K L M N O P Q R S T U V W X Y Z 𝒜 ℬ 𝒞 𝒟 ℰ ℱ 𝒢 ℋ ℐ 𝒥 𝒦 ℒ ℳ 𝒩 𝒪 𝒫 𝒬 ℛ 𝒮 𝒯 𝒰 𝒱 𝒲 𝒳 𝒴 𝒵 𝔸 𝔹 ℂ 𝔻 𝔽 𝔾 ℍ 𝕀 𝕁 𝕂 𝕃 𝕄 ℕ 𝕆 ℙ ℚ ℝ 𝕊 𝕋 𝕌 𝕍 𝕎 𝕏 𝕐 ℤ Λ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z Σ Λ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

http://arxiv.org/abs/2303.07306v1

20230313173012

The category of extensions and idempotent completion

math.CT

Bennett-Tennenhaus]Raphael Bennett-Tennenhaus Department of Mathematics Aarhus University 8000 Aarhus C Denmark raphaelbennetttennenhaus@gmail.com Haugland]Johanne Haugland Department of Mathematical Sciences NTNU NO-7491 Trondheim Norway johanne.haugland@ntnu.no Sandøy]Mads Hustad Sandøy Department of Mathematical Sciences NTNU NO-7491 Trondheim Norway mads.sandoy@ntnu.no Shah]Amit Shah Department of Mathematics Aarhus University 8000 Aarhus C Denmark amit.shah@math.au.dk [2020]18E05, 18E10, 18G80, 18N10, 18G99 1 Building on previous work, we study the splitting of idempotents in the category of extensions 𝔼-Ext(𝒞) associated to an n-exangulated category (𝒞,𝔼,𝔰). In particular, we show that idempotents split in 𝔼-Ext(𝒞) whenever they do so in 𝒞, allowing us to prove that idempotent completions and extension categories are compatible constructions in a -theoretic sense. More precisely, we show that the exact category obtained by first taking the idempotent completion of (𝒞,𝔼,𝔰), in the sense of Klapproth–Msapato–Shah, and then considering its category of extensions is equivalent to the exact category obtained by first passing to the extension category and then taking the idempotent completion. These two different approaches yield a pair of 2-functors each taking small n-exangulated categories to small idempotent complete exact categories. The collection of equivalences that we provide constitutes a 2-natural transformation between these 2-functors. Similar results with no smallness assumptions and regarding weak idempotent completions are also proved. The category of extensions and idempotent completion [ March 30, 2023 ==================================================== 1 § INTRODUCTION An additive category is called idempotent complete given that every idempotent morphism splits (see <ref>), or equivalently if every idempotent admits a kernel (see e.g. <cit.>). The study of idempotent complete categories dates back to work by Karoubi <cit.>, in which it was shown that an additive category can be naturally embedded into an idempotent complete category , often called its Karoubi envelope or its idempotent completion (see <ref> and <ref>). The splitting of idempotents plays an important role in contemporary algebraic geometry, homological algebra, representation theory and category theory. Indeed, it is intimately connected to the Krull–Remak–Schmidt property and Krull–Schmidt categories (see <cit.>, <cit.>). Krull–Schmidt categories constitute a particularly nice class of examples of idempotent complete categories. In such a category, every object decomposes, essentially uniquely, into a finite direct sum of indecomposable objects with local endomorphism rings. Splitting of idempotents is often a crucial standing assumption when approaching representation theory of finite-dimensional algebras from a categorical or geometrical perspective (see e.g. <cit.>). Furthermore, a generalisation of the Krull–Remak–Schmidt property was given by Azumaya <cit.>, which has since been used in topological data analysis in the study of persistence homology (see e.g. <cit.>). Abelian, or more generally exact, and triangulated categories appear in various areas of mathematics, including functional analysis and mathematical physics, and are of fundamental interest in representation theory and related areas (see e.g. <cit.>). Recently, Nakaoka–Palu introduced categories as a simultaneous generalisation of exact and triangulated <cit.>, and showed that extension-closed subcategories of triangulated categories, which may fail to be triangulated subcategories, carry an extriangulated structure. Herschend–Liu–Nakaoka <cit.> then introduced categories as a higher-dimensional analogue of extriangulated categories in the context of higher homological algebra. An n-exangulated category for an integer n ≥ 1 is a triplet (,,) consisting of an additive category , a biadditive functor ×→ (where denotes the category of abelian groups), and a realisation of . Note that a category is 1-exangulated if and only if it is extriangulated <cit.>. Important classes of examples of n-exangulated categories for higher n include n-exact categories <cit.> and (n+2)-angulated categories <cit.>. Suppose that (,,) is an n-exangulated category. For each pair of objects A,C, elements of (C,A) are called -extensions. The category of extensions associated to (,,), denoted by , has all -extensions as its objects, and the morphisms are morphisms of -extensions; see Subsection <ref>. In a previous article, the authors showed that this category can be equipped with a natural exact structure , giving rise to an exact category (,); see <cit.>. Moreover, we demonstrated that encodes important structural information. As an example, this perspective leads to a full characterisation of n-exangulated functors between n-exangulated categories; see <cit.>. In the present paper, we improve the understanding of the relationship between an n-exangulated category and its associated category of extensions by studying the splitting of idempotents. Our first prominent result is <ref> below, which asserts that the splitting of idempotents in is inherited from . This plays an important role in the paper as a key step towards <ref>. [See <ref>] If is idempotent complete, then is also idempotent complete. As a consequence of <ref>, we obtain that given certain assumptions on , the Krull–Remak–Schmidt property for implies the same property for ; see <ref>, cf. <cit.>, <cit.>. It is shown in <cit.> that the idempotent completion of admits an n-exangulated structure, giving rise to an idempotent complete n-exangulated category (,,1mu); see <ref>. Our second main result, given as <ref> below, demonstrates that idempotent completions and extension categories are compatible constructions. More precisely, the category obtained by first taking the idempotent completion of an n-exangulated category and then considering its category of extensions is equivalent to first passing to the extension category and then taking the idempotent completion. [See <ref>] The category is equivalent to the idempotent completion of the category . <ref> and <ref> are both used in order to obtain the 2-category-theoretic result <ref>, which builds a bridge between the 2-categorical framework established in <cit.> and the results on the idempotent completion of an n-exangulated category from <cit.>. To discuss a 2-category of n-exangulated categories, we use notions of morphisms between n-exangulated categories and of morphisms between such morphisms. Structure-preserving functors between n-exangulated categories as introduced in <cit.> are known as n-exangulated functors. We viewed the theory of n-exangulated categories from a 2-categorical perspective in <cit.> by defining n-exangulated natural transformations between n-exangulated functors and establishing the 2-category n of small n-exangulated categories <cit.>. Furthermore, we constructed a 2-functor n→ to the category of small exact categories <cit.>, which sends a 0-cell (,,) in n to the 0-cell (,) in ; see <ref>. A consequence of <ref> is that restricts to a 2-functor n→ from the 2-category of small idempotent complete n-exangulated categories to the 2-category of small idempotent complete exact categories. The last observation needed in order to state <ref> is that taking idempotent completions yields 2-functors → and n→n; see <ref> and <ref>, respectively. [See <ref>] Consider the diagram nrd[swap] d nr of 2-categories and 2-functors. There is a 2-natural transformation ⇒ consisting of exact equivalences. Similar results as above hold also for weak idempotent completions; see <ref>. We remark that even though <ref> is an expected analogue of <ref> in this setup, the method of proof is different and relies on a previous result of the authors from <cit.>. We note that <ref> follows from a more general result, namely <ref>, in which no smallness assumption is required. In this article the term `category' does not require the collections of morphisms to form sets. In other words, the categories we consider need not be locally small. Just as explained in <cit.>, the reason for the restriction to small categories in <ref> (and indeed in this introduction entirely) is to be able to use the terminology of 2-categories and 2-functors in a way that is consistent with the existing literature. §.§ Structure of the paper In <ref> we recall the construction of the idempotent completion of an exact category and use this to establish the 2-functor from <ref>. Analogously, the 2-functor is defined in <ref> using the idempotent completion of an n-exangulated category in the sense of <cit.>. In <ref> we recall how to form the category of extensions associated to an n-exangulated category and prove <ref> and <ref>. In <ref> we present the definition of the 2-functor from <cit.> and show how the main results of the previous sections culminate in <ref>. <ref> concerns the weak idempotent completion. §.§ Conventions and notation Throughout this paper, let n≥ 1 denote a positive integer. Given objects X and Y in a category , we write (X,Y) for the collection of morphisms from X to Y in . Functors are always assumed to be covariant. We let denote the category of abelian groups. § IDEMPOTENT COMPLETION OF EXACT CATEGORIES YIELDS A 2-FUNCTOR The aim for this section is to explicitly relate the construction of the idempotent completion of an exact category to a 2-categorical framework, establishing the 2-functor which is part of <ref> in <ref>. We start by following Bühler <cit.> in recalling the idempotent completion (or Karoubi envelope). We also refer to Borceux <cit.>. Throughout the section, let denote an additive category. An idempotent in is a morphism e X→ X for some object X satisfying e^2 = e. Splitting of idempotents, as defined below, plays a central role in this article. (See <cit.>.) An idempotent e X → X in splits if there exist morphisms r X → Y and s Y → X such that sr=e and rs=𝕀Y. The category is idempotent complete, or has split idempotents, if each idempotent in splits. Even though the additive category need not have split idempotents, it can always be embedded into an idempotent complete category. This is due to Karoubi <cit.>. (See <cit.>.) Define a category as follows. The objects of are pairs (X,e) for each object X and each idempotent e(X). Given objects (X,X) and (Y,Y) in , the collection ((X,X),(Y,Y)) of morphisms from (X,X) to (Y,Y) consists of triplets (Y,f,X) such that f (X,Y) satisfies f X = f = Y f. The composition of (Y,f,X)((X,X), (Y,Y)) and (Z,g,Y)((Y,Y), (Z,Z)) is given by (Z,g,Y)∘(Y,f,X) (Z, gf, X). It is clear that this composition is associative. The identity 𝕀(X,e) of (X,e) is the morphism (e,e,e). The category is called the idempotent completion of . The category is additive with biproduct given by . It is also idempotent complete; see <cit.> for details. There is a canonical additive inclusion functor → defined by setting (X) (X, 𝕀X) for X and for f (X,Y). This functor is 2-universal among additive functors from to idempotent complete categories; see <cit.>. Let →→ denote the category of composable morphisms in , and note that a functor → induces a functor →→→→→. Now suppose (,) is an exact category. In particular, the exact structure is a collection of objects in →→. One can define an exact structure on by declaring an object in →→ to be in if it is a direct summand of an object belonging to the image of under the functor →→→→→ induced by →. (See <cit.>.) The pair (,) forms an exact category, and (,) → (,) is a fully faithful exact functor that reflects exactness. Let → be an additive functor. Following <cit.>, there is an induced additive functor → given by (X,e) ( X, e) and(Y,f,X) (Y, f, X). We refer to as the completion of . If (,) → (,) is an exact functor, then (,) → (,) is also exact; see the proof of <cit.>. In order to view the constructions above in a 2-categorical framework, we recall some terminology. A 2-category is a collection of 0-cells, 1-cells and 2-cells satisfying certain axioms; see e.g. <cit.> or <cit.>. One should think of 0-cells, 1-cells and 2-cells as objects, morphisms between objects and morphisms between morphisms, respectively. A has two notions of composition of 2-cells: vertical and horizontal. Using the setup below, we recall these notions in the case of natural transformations. We use the Hebrew letters (beth) and (daleth) for natural transformations of additive functors. For the rest of this section, we consider additive categories ,,, additive functors ,,→ and ,,→, and natural transformations ⇒, '⇒, ⇒ and '⇒ as indicated in the diagram [Rightarrow, shorten <= 6pt, shorten >= 6pt, yshift = -2pt ]d [xshift=2pt, yshift=-1pt] [Rightarrow, shorten <= 10pt, shorten >= 12pt, xshift=0pt, yshift = -2pt ]dd [xshift=1pt, yshift=-1pt] [bend left=50]rr [description, xshift=0pt, yshift=0pt]rr [description][bend right=50]rr [description, xshift=0pt, yshift=0pt] [Rightarrow, shorten <= 3pt, shorten >= 6pt, yshift=4pt ]d [xshift=2pt, yshift=3pt]' [bend left=50]rr [description, xshift=0pt, yshift=0pt] [bend right=50]rr [description, xshift=0pt, yshift=0pt] . (See <cit.>.) The vertical composition of and ' is the natural transformation '⇒ given by *[](')_X*[]_X^'X for each X. The horizontal composition of and is the natural transformation ⇒ defined by (X X∘(X) for each X. As described in <cit.>, the natural transformation ⇒ induces a natural transformation ⇒ as follows. Given (X,e), there are the morphisms (𝕀X,e,e) (X,e) → (X,𝕀X) and (e,e,𝕀X) (X,𝕀X) → (X,e). Put (X,e)(e,e,𝕀X)∘ (𝕀 X, X, 𝕀 X) ∘(𝕀X,e,e) = ( e,( e)X e, e) as indicated in the diagram [column sep=3cm] (X,e) [dotted]r(X,e)d[swap](𝕀X,e,e) (X,e) ( X,𝕀 X) r[swap](𝕀 X, X, 𝕀 X) ( X,𝕀 X). u[swap](e,e,𝕀X) It is straightforward to check that is natural. We refer to as the completion of . We write for the collection of 0-cells, 1-cells and 2-cells consisting of exact categories, exact functors and natural transformations, respectively. For i{0,1,2}, we denote the collection of i-cells by i. Given 0-cells (,) and (,), there is a category ((,), (,)) with 1-cells of the form (,) → (,) as objects, and where morphisms and composition are given by 2-cells and vertical composition. Hence, there is a determined by the 0-cells in which are small categories. We furthermore write and when restricting to idempotent complete 0-cells in and , respectively, and note that also is a 2-category. A 2-functor between two 2-categories is an assignment of i-cells in the domain category to i-cells in the codomain category for i{1,2,3}, satisfying some compatibility conditions; see e.g. <cit.> or <cit.>. We now begin to construct the 2-functor used in <ref> in <ref>. Let = (0,1,2) → be defined by the assignments ii→i, where: 0 (,) (,), 1 () , 2 ( ) . If one ignores the set-theoretic issue described in <ref>, then the theorem below should be interpreted as showing that → is a 2-functor. The following statements hold for the assignments 0, 1 and 2. * The pair (0, 1) defines a functor →. * The pair (1, 2) defines a functor ((,), (,)) →((,), (,)) whenever (,) and (,) are exact categories. * The assignment 2 preserves horizontal composition. In particular, restricting to small categories yields a 2-functor →. It follows from the discussions above that the assignments are well-defined. Checking functoriality in <ref> and <ref> is straightforward, while <ref> follows from <cit.>. § IDEMPOTENT COMPLETION OF N-EXANGULATED CATEGORIES YIELDS A 2-FUNCTOR In this section we describe how taking the idempotent completion of an n-exangulated category in the sense of <cit.> relates to a 2-categorical framework. This is done by constructing the 2-functor from <ref> in <ref>. We start by giving an overview of relevant notions and constructions. Given an additive category and a biadditive functor ×→, an element α(C,A) is called an -extension. A morphism of -extensions from α(C,A) to is a pair (a,c) of morphisms a A→ B and c C→ D in such that (C,a)(α)=(c,A)(β). Recall from <cit.> that an n-exangulated category (,,) consists of * an additive category , * a biadditive functor ×→, and * an exact realisation of in the sense of <cit.>, such that axioms (EA1), (EA2) and (EA*[]2^) stated in <cit.> are satisfied. The realisation associates to each -extension α(C,A) a certain homotopy class (α) = [] = [0r 1r ⋯r n+1] of an (n+2)-term complex in with 0 = A and n+1 = C. The pair , α is then called a (distinguished) n-exangle. A morphism , α→∙, β of n-exangles is given by a morphism (0,…, n+1)→∙ of complexes such that (0,n+1) α→β is a morphism of -extensions. In this case, the tuple (0,…, n+1) is said to be a lift of (0,n+1). Suppose throughout this section that (,,) and (,,) are n-exangulated categories. An additive functor → induces a functor → between the associated categories of complexes. One can define a new biadditive functor , which we will denote by (-,-). (See <cit.>.) Let → be an additive functor and suppose there is a natural transformation Γ = {(C,A)(C,A)0.7×(-,-) ⟹(-,-). We call the pair (,Γ) (,,) → (,,) an n-exangulated functor if, for all A,C and each α(C,A), we have that (α)=[] implies ((C,A)(α))=[]. It was demonstrated in <cit.> that one can compose n-exangulated functors as follows. Suppose (, Γ) (,,) → (,,) and (, Φ) (,,) → (,,) are functors between n-exangulated categories. The composite is the functor (,Φ) ∘ (,Γ)(, ×Γ), where × is the natural transformation × = {( C, A)(C,A)0.7×(-,-) ⟹(-,-). For Γ as above and for α(C,A), we will usually write Γ(α) instead of (C,A)(α). Furthermore, we use the simplified notation α (resp. α) for the -extension (C,a)(α)(C,B) (resp. (d,A)(α)(D,A)) for morphisms a A→ B and d D→ C in . As proved in <cit.>, the idempotent completion of an n-exangulated category (,,) admits a canonical n-exangulated structure. We use the notation (,,1mu) for the n-exangulated category obtained from this construction, and recall the definition of the biadditive functor ×→ and the realisation of below. In the case n=1, the construction was given by Msapato <cit.>. (See <cit.>.) For objects (A,A),(C,C), we let ((C,C),(A,A)) (A, α, C) | α(C,A) and (Aα = α = (Cα. For morphisms (B, a, A) (A,A)→ (B,B) and (C,d,D) (D,D)→ (C,C) in we put ((C,d,D) , (B, a, A)) ((C,C), (A,A)) ⟶((D, D), (B, B)) (A, α, C) ⟼ (B, (d,a)(α), D). The set ((C,C),(A,A)) has an abelian group structure given by (A, α, C) + (A, α', C) (A, α + α', C), and <ref> indeed gives a biadditive functor ×→; see <cit.>. Given a complex in and an idempotent morphism ∙→ of complexes, we follow <cit.> and the discussion immediately thereafter in using the notation (, ∙) to denote the complex [column sep=1.3cm, scale cd=0.9] (X_0,0) rr(1,10,0) (1,1) rr(2,21,1) ⋯r (n,n) rr(n+1,n+1n,n) (n+1,n+1) in , where the maps d_i X_i→ X_i+1 are the differentials of the complex . The realisation of is then defined as follows. (See <cit.>.) Let (A, α, C) ((C,C),(A,A)) be arbitrary. Since α(C,A), one may choose so that (α) = [] = [[column sep=0.6cm] A r 1r ⋯r nr C ]. One may also lift (A, C) α→α to an idempotent endomorphism ∙ of the n-exangle ⟨, α⟩ by <cit.>. Using this, we define by setting (A, α, C) [(, ∙)]. To see that the assignment from <ref> does not rely on the choices involved, see <cit.>. By <cit.>, the triplet (,,1mu) is an n-exangulated category and the inclusion (, ) (,,)→(,,1mu) is an n-exangulated functor, where the natural transformation (-,-)⇒(-,-) is given by α↦ (𝕀A, α, 𝕀C) for α(C,A). Now suppose (,Γ) (,,)→ (,,) is an n-exangulated functor. Recall that there is an induced additive functor → as defined in (<ref>). Our next aim is to show that one obtains an n-exangulated functor (,Γ) (,,1mu) → (,,1mu) between the idempotent completions. We first need to define a natural transformation Γ(-,-)⇒(-,-). Set Γ((C,C),(A,A))((C,C),(A,A))×, where ((C,C),(A,A))((C,C), (A,A)) ⟶((C, C), (A, A)) (A, α, C) ⟼ (A, Γ(α), C). Note that (A,Γ(α),C) indeed lies in ((C,C),(A,A)), because naturality of Γ yields (AΓ(α) = Γ((Aα) = Γ(α) and (CΓ(α) = Γ((Cα) = Γ (α). Since Γ(-,-)⇒(,-,-) is a natural transformation, one might wonder if the definition of Γ above agrees with the description of the completion of a natural transformation of additive functors from <ref>. However, the biadditive functors and (-,-) are not necessarily additive functors ×→, so we cannot form the completions of them as in <ref>. Thus, when we use notation of the form Γ for a natural transformation of biadditive functors, it always refers to the construction from <ref>. The pair (,Γ) is an n-exangulated functor (,,1mu)→ (,,1mu). Given a pair of objects (A,A),(C,C), the map ((C,C),(A,A)) from <ref> is a homomorphism of abelian groups as (C,A) is one. It follows from the naturality of Γ that Γ is a natural transformation (-,-)⇒(-,-). Consider now an -extension (A,α,C)((C,C),(A,A)). Following the definition of , we have where (α) = [] and the idempotent ∙→ is a lift of (A, C) α→α. Notice that (Γ(α)) = [] as (,Γ) is n-exangulated. Moreover, the morphism ∙→ is an idempotent lifting (A, C)Γ(α)→Γ(α). We hence see that (Γ(A,α,C)) is given by the class [[column sep=2.6cm, ampersand replacement=&] ( A,A) r(1,10,A)& (1,1) r(2,21,1)&⋯r(C,Cn,n)& ( C,C) ], which is [(,∙)]. This finishes the proof. To consider n-exangulated categories as 0-cells in a 2-category, we use the notion of a morphism between n-exangulated functors. This is captured by the following definition. (See <cit.>.) Suppose that (,Γ), (,Λ)(,,)→(,,) are n-exangulated functors. An n-exangulated natural transformation (,Γ)⇒(,Λ) is a natural transformation ⇒ of additive functors such that, for all A,C and each α(C,A), the pair (A,C) satisfies (AΓ(α) = (CΛ(α). Notice that equation (<ref>) means that (A,C) is a morphism Γ(α) →Λ(α) of -extensions. For a natural transformation ⇒ of additive functors ,→, recall that the completion ⇒ is given by (X,e) = ( e,( e)X e, e) for (X,e). The proposition below shows that the completion of an n-exangulated natural transformation is again n-exangulated. Suppose (,Γ) ⇒ (,Λ) is n-exangulated. Then is an n-exangulated natural transformation (,Γ) ⇒ (,Λ). Consider an -extension (A,α,C)((C,C),(A,A)). Using that (Aα=α, we get (AΓ(α)=Γ(α) by the naturality of Γ. Similarly, we obtain Λ(α)=(CΛ(α). Since is n-exangulated, we have (AΓ(α)=(CΛ(α), while naturality of yields . Combining these observations gives (A (A (AΓ(α)= (A (AΓ(α)= (A (CΛ(α)= (C (AΛ(α) = (CΛ(α) = (C (CΛ(α) = (CCΛ(α) = (CC (CΛ(α). Hence, we have that ((A,A)Γ(A,α,C) = (A,(A)AA,A (A,Γ(α),C) (definitions of , Γ) = (A,(A (A (AΓ(α),C) (<ref>) = (A,((C) CCΛ(α),C) (as above) = ((C,C) Λ(A,α,C) (definitions of , Λ), as required. We now introduce n-exangulated analogues of the collections described in Notation <ref>. Following <cit.>, we write n for the collection of 0-cells, 1-cells and 2-cells consisting of n-exangulated categories, n-exangulated functors and n-exangulated natural transformations between these functors, respectively. Restricting 0-cells in n to small n-exangulated categories yields the 2-category n; see <cit.>. We furthermore write n and n when only considering idempotent complete 0-cells in n and n, respectively, and note that also n is a 2-category. As before, we use a subscript i {0,1,2} to denote i-cells in the collections described above. We conclude this section by constructing the 2-functor used in <ref> in <ref>. Let = (0,1,2) n→n be defined by the assignments ini→ni, where: 0 (,,) (,,1mu), 1 (,Γ) (,Γ), 2 ( ) . The result below is an n-exangulated analogue of <ref>. The following statements hold for the assignments 0, 1 and 2. * The pair (0, 1) defines a functor n→n. * The pair (1, 2) defines a functor n((,,), (,,)) →n((,,1mu), (,,1mu)) whenever (,,) and (,,) are n-exangulated categories. * The assignment 2 preserves horizontal composition. In particular, restricting to small categories yields a 2-functor n→n. It follows from the discussion and results above, in particular Lemmas <ref> and <ref>, that the assignments are well-defined. Functoriality in <ref> is straightforward to check. For <ref>, notice first that n((,,), (,,)), and hence also n((,,1mu), (,,1mu)), is indeed a category by <cit.>. Moreover, for <ref>, note that n-exangulated natural transformations are closed under horizontal composition by <cit.>. As 2 is a restriction of 2, parts <ref> and <ref> follow from <ref><ref> and <ref>, respectively. § THE CATEGORY OF EXTENSIONS AND IDEMPOTENT COMPLETION For consistency with the rest of the paper, we assume throughout this section that (,,) is an n-exangulated category. We note, however, that the results of <ref> hold under more general assumptions; see <ref>. In Subsection <ref> we first recall the definition of the category of extensions associated to (,,), before proving <ref> from <ref>. Building on this result, our ultimate goal is to show that the category of extensions of the idempotent completion is equivalent to the idempotent completion of the category of extensions. These two categories are described explicitly in Subsection <ref> and Subsection <ref>, respectively, culminating in a proof of <ref> from <ref>. §.§ The category of extensions The category of extensions associated to (,,) is denoted by . The objects of are -extensions, and the morphisms are morphisms of -extensions. Recall from <ref> that this means that an object is an element α(C,A) for some objects A,C, while a morphism from α(C,A) to β(D,B) is given by a pair (a,c) of morphisms a A→ B and c C→ D in satisfying α = β. As shown in <cit.>, one can define an exact structure on as follows. Let α(C,A), β(D,B) and γ(G,E) be objects in . A sequence αr(a,c) βr(b,d) γ of composable morphisms in lies in the class if and only if the morphisms a and c in are both sections with b= a and d= c. By definition, a category is idempotent complete provided any idempotent endomorphism splits. Note also that a morphism of -extensions (A,C)α→α for α(C,A) is an idempotent in if and only if both A A → A and C C → C are idempotents in . Hence, <ref> follows from <ref>. The authors are grateful to Dixy Msapato for pointing out <cit.>, which motivated the proof of the result below. Let α(C,A) and suppose (A,C)(α) is an idempotent. Then A and C split in if and only if (A,C) splits in . (⇒) Assume that A and C split in . As A splits, there exist morphisms r A → B and s B → A such that sr=A and rs=𝕀B. Similarly, there exist u C → D and v D → C with vu=C and uv=𝕀D, because C splits. Consider the -extension α(D,B). We see that (s,v)α→α is a morphism in , as (α) = (srα = (e_Aα = (e_Cα = (vuα = (vuvα = α. Analogously, one can show that (r,u)α→α is a morphism of -extensions. Notice that (s,v)∘(r,u) = (A,C) and (r,u)∘(s,v) = (𝕀B,𝕀D), which is the identity on α. This gives a splitting of (A,C), as required. (⇐) If (A,C) splits, then there exist β(D,B) and morphisms (r,u)α→β and (s,v)β→α in such that (s,v)∘(r,u) = (A,C) and . These equations yield splittings of A and C in . We finish this subsection by deducing <ref>, showing that given certain assumptions on , the Krull–Remak–Schmidt property for implies the same property for . In order to see this, we first recall some terminology. Let be a commutative ring. The additive category is said to be -linear if (X,Y) is an -module for all X,Y, and we have (λ g)f=λ (gf)=g(λ f) for all λ and all composable morphisms f and g in . When is -linear, the bifunctor is -bilinear provided that each abelian group (C,A) has the structure of an -module, and we have (λ c,a)=λ(c,a)=(c,λ a) for all λ and any morphisms a and c in . Recall that an -linear category is called -finite (over ) if each -module (X,Y) has finite length (see e.g. <cit.>). If is -linear and is -bilinear, then is also -linear. If in addition is -finite, then so is . Fix objects α and β in , say where α(C,A) and β(D,B). Given a morphism (a,c)α→β in , we have (λ aα= (𝕀C,λ a)(α)= λ(𝕀C,a)(α)= λ(c,𝕀A)(β)= (λ c,𝕀A)(β) =(λ cβ as is -bilinear. This means that (λ a,λ c)α→β is a morphism in , and we take this to be the action of λ on (a,c). Using that is -linear, it is straightforward to check that the -sets of are -modules under this multiplication. Consider another morphism (b,d)β→γ in . Since is -linear, we have (b,d)(λ a,λ c)=(b(λ a),d(λ c))=(λ (ba),λ(dc))=((λ b)a,(λ d), c)=(λ b,λ d)(a,c), which is the action of λ on (b,d)∘ (a,c). This proves that is -linear. The arguments above show that the collection of morphisms α→β in defines an -submodule of the direct sum (A,B)⊕(C,D). The length of this submodule is bounded above by the sum of the lengths of (A,B) and (C,D), proving the second assertion. Recall that is said to be a Krull–Schmidt category if every object decomposes into a finite direct sum of objects with local endomorphism rings. Suppose that is -linear, -finite and Krull–Schmidt, and that is -bilinear. Then is also -linear, -finite and Krull–Schmidt. Note that an R-linear -finite category is Krull–Schmidt if and only if it is idempotent complete; see e.g. <cit.> or <cit.>. Using this, the result now follows by combining <ref> and <ref>. §.§ The category of extensions of the idempotent completion As recalled in <ref>, the idempotent completion admits an n-exangulated structure by <cit.>. The corresponding n-exangulated category is denoted by (,,1mu). Therefore, as in Subsection <ref>, we can consider the category of extensions associated to (,,1mu). We give an explicit description of and the exact structure below. Objects: The objects of are of the form (A,α,C)((C,C),(A,A)) for (C,C) and (A,A) in . In particular, the morphisms A A → A and C C → C in are idempotents, and α(C,A) is an -extension satisfying (Aα = α = (Cα. Morphisms: A morphism (A,α, C) → (B,β, D) in is a pair ((B,a,A),(D,c,C)), where (B,a,A) (A,A) → (B,B) and (D,c,C) (C,C) → (D,D) are morphisms in and (B,a,Aα = (D,c,Bβ. In other words, this means that (a,c)α→β is a morphism of -extensions, aA = a = Ba and cC=c=Dc. Composition: Composition in is defined component-wise. Explicitly, the composition of ((B,a,A),(D,c,C)) and ((E,b,B),(G,d,D)) is ((E,ba,A),(G,dc,C)). Identity morphisms: The identity on (A,α,C) in is ((A,A,A),(C,C,C)). Preadditivity: The addition of morphisms is component-wise. Explicitly, the addition of ((B,a,A),(D,c,C)) and ((B,a',A),(D,c',C)) is ((B,a+a',A),(D,c+c',C)). Exact structure: The collection consists of kernel-cokernel pairs [column sep=3.7cm] (A, α, C) r((B,a,A),(D,c,C)) (B, β, D) r((E,b,B),(G,d,D)) (E, γ, G) in such that (B,a,A) and (D,c,C) are sections with (E,b,B) = (B,a,A) and (G,d,D) = (D,c,C). These conditions are equivalent to the sequences [column sep=2cm,ampersand replacement=&] (A,A) r(B,a,A)& (B,B) r(E,b,B)& (E,E), [column sep=2cm,ampersand replacement=&] (C,C) r(D,c,C)& (D,D) r(G,d,D)& (G,G), being split exact in . We have an immediate corollary of <ref>. The exact category (, ) is idempotent complete. §.§ The idempotent completion of the category of extensions In contrast to what is done in Subsection <ref>, we may first consider the category of extensions associated to (,,), which forms part of an exact category (,). Then we may take the idempotent completion, resulting in an idempotent complete exact category that we denote by (,); see <ref>. We proceed with an explicit description of and the exact structure . Objects: The objects of are of the form (α, (A,C)), where α(C,A) and (A,C)α→α is an idempotent morphism of -extensions. This means that e_A A → A and C C → C are idempotents in and that (Aα = (Cα. Morphisms: A morphism (α, (A,C)) → (β, (B,D)) in is given by a triple ((B,D),(a,c),(A,C)), where (a,c) α→β is a morphism of -extensions and we have (a,c)(A,C) = (a,c) = (B,D)(a,c). Composition: The composition of two composable morphisms ((B,D),(a,c),(A,C)) and ((E,G),(b,d),(B,D)) in is ((E,G),(ba,dc),(A,C)). Identity morphisms: The identity on (α, (A,C)) in is ((A,C),(A,C),(A,C)). Preadditivity: Let ((B,D),(a,c),(A,C)) and ((B,D),(a',c'),(A,C)) be morphisms from (α,(e_A,e_C)) to (β,(e_B,e_D)) in . The addition of these two morphisms is given by ((B,D),(a+a',c+c'),(A,C)). Exact structure: The elements in are direct summands of images of elements in under the functor →; see the discussion before <ref>. In other words, they are direct summands of kernel-cokernel pairs in of the form [column sep=2cm] (α, (𝕀A,𝕀C)) rp (β, (𝕀B,𝕀D)) rq (γ, (𝕀E,𝕀G)), where [ p ((𝕀B,𝕀D),(a,c),(𝕀A,𝕀C)), q ((𝕀E,𝕀G),(b,d),(𝕀B,𝕀D)) ] and αr(a,c) βr(b,d) γ is an element of . Even if (α,(A,C)) is an object in , the -extension α(C,A) does not necessarily satisfy (Aα = α or (Cα =α. For example, if α≠ 0, then (α,(0,0)) is an object in , but α = α = 0 ≠α. In particular, this means that the objects of are not canonically in one-to-one correspondence with those of . <ref> tells us that we cannot expect the categories and to be isomorphic in general. Despite this, we prove that they are always equivalent. In the following, we use the Hebrew letters (mem), (shin) and (tsadi). Note that the functor (,,) that we define in the proof of <ref> below will be used to construct a natural transformation in <ref>, which is the reason for our choice of notation. The exact categories (, ) and (, ) are equivalent. We establish an exact functor (,,) (, ) → (, ) and an exact quasi-inverse (,,) (, ) → (, ). Define (,,) by (,,)(A,α,C) (α,(A,C)) on objects and (,,)((B,a,A),(D,c,C)) ((B,D),(a,c),(A,C)) on morphisms. By our explicit description of and in Subsection <ref> and Subsection <ref>, respectively, we see that (,,) is a well-defined additive functor. Define (,,) by (,,) (α,(A,C)) (A,(Aα,C) on objects and (,,) ((B,D),(a,c),(A,C)) ((B,a,A),(D,c,C)) on morphisms. Note that (,,) is well-defined on objects, since (C (Aα = (A (Cα = (A (Aα = (Aα. It is straightforward to check that (,,) is well-defined on morphisms, and that it is an additive functor. The composite (,,)∘(,,) is the identity functor 𝕀 of , as (Aα = α whenever (A,α,C). For each object (α,(A,C)) in , set (α,(A,C)) ((A,C),(A,C),(A,C)) (α,(A,C)) → ((Aα,(A,C)). This is an isomorphism in . Checking that (α,(A,C))(α,(A,C))𝕀⟹(,,)∘(,,) is natural is straightforward. It remains to show that (,,) and (,,) are exact functors. Recall that the direct sum of two objects (X,X) and (Y,Y) in is given by (X,X)⊕ (Y,Y) = (X⊕ Y,X⊕Y) = (X⊕ Y,[ X 0; 0 Y ]). We first check that (,,) is exact. Let [column sep=4cm] (A, α, C) r((B,a,A),(D,c,C)) (B, β, D) r((E,b,B),(G,d,D)) (E, γ, G) be an arbitrary element of . The underlying sequences of (<ref>) are split exact in , so we may without loss of generality assume that B = A⊕ E, B = A⊕E, D = C⊕ G, D = C⊕G, (a,c) = ([ A; 0 ], [ C; 0 ]) and (b,d) = ([ 0E ], [ 0 G ]). Applying (,,) to (<ref>) then yields the sequence [column sep=1.5cm] (α, (A,C)) rr (β, (A⊕E,C⊕G)) rs (γ, (E,G)), where r (([ A 0; 0 E ],[ C 0; 0 G ]),([ A; 0 ],[ C; 0 ]),(A,C)), s ((E,G),([ 0E ], [ 0 G ]),([ A 0; 0 E ],[ C 0; 0 G ])). We claim that (<ref>) is a direct summand of the sequence [column sep=1.5cm] (α, (𝕀A,𝕀C)) rt (β, (𝕀A⊕ E, 𝕀C⊕ G)) ru (γ, (𝕀E,𝕀G)), where t ((𝕀A⊕ E, 𝕀C⊕ G),([ 𝕀A; 0 ], [ 𝕀C; 0 ]),(𝕀A,𝕀C)), u ( (𝕀E,𝕀G),([ 0 𝕀E ],[ 0 𝕀G ]),(𝕀A⊕ E, 𝕀C⊕ G) ) . Notice first that ([ 𝕀A; 0 ], [ 𝕀C; 0 ]) is a morphism α→β of -extensions, since *[][ 𝕀A; 0 ]_α = *[][ 𝕀A; 0 ]_ (e_Aα = *[][ A; 0 ]_α = *[][ C; 0 ]^β = *[][ 𝕀C; 0 ]^*[][ C 0; 0 G ]^β = *[][ 𝕀C; 0 ]_β. Similarly, the pair ([ 0 𝕀E ],[ 0 𝕀G ]) is a morphism β→γ in . In particular, this yields that (<ref>) is indeed a sequence in . To verify that (<ref>) is a direct summand of (<ref>), notice that there is a section induced by the morphisms ((𝕀A,𝕀C),(A,C),(A,C)), and ((𝕀E, 𝕀G),(E,G),(E,G)). Thus, to finish the proof that (,,) is exact, it suffices to show that (<ref>) lies in . For this, it is in turn enough to verify that [column sep=3cm] αr([ 𝕀A; 0 ], [ 𝕀C; 0 ]) βr([ 0 𝕀E ],[ 0 𝕀G ]) γ lies in . By the arguments above, we already know that (<ref>) is a sequence of morphisms in . As its underlying sequences are split exact in , we have that (<ref>) lies in . We now show that (,,) is exact. Let (α, (A,C)) r (β, (B,D)) r (γ, (E,G)) be a conflation in . Consequently, we have that (<ref>) is a direct summand of a sequence (α', (𝕀A',𝕀C')) r (β', (𝕀B',𝕀D')) r (γ', (𝕀E',𝕀G')), which is the image under of a kernel-cokernel pair [column sep=2cm] α' r(a',c') β' r(b',d') γ' in . Apply (,,) to (<ref>) to obtain (𝕀A',α',𝕀C') r (𝕀B',β',𝕀D') r (𝕀E',γ',𝕀G'). We claim that (<ref>) lies in . Since (<ref>) belongs to , its underlying sequences are split exact in . As → is an additive functor, the sequences [column sep=3cm,ampersand replacement=&] (A', 𝕀A') r(𝕀B',a',𝕀A')&(B', 𝕀B') r(𝕀E',b',𝕀B')&(E', 𝕀E'), [column sep=3cm,ampersand replacement=&] (C', 𝕀C') r(𝕀D',c',𝕀C')&(D', 𝕀D') r(𝕀G',d',𝕀D')&(G', 𝕀G'), are thus split exact in , and so (<ref>) lies in . Since (<ref>) is a direct summand of (<ref>), we know that (,,)(<ref>) is a direct summand of (,,)(<ref>)=(<ref>). Thus, by <cit.>, we deduce that (,,)(<ref>) belongs to , and hence (,,) is an exact functor. For consistency with the rest of the paper, in <ref> we have assumed that (,,) is an n-exangulated category. However, the results from this section work in a more general setting. We describe this setting in what follows, and leave it to the interested reader to reformulate the claims and their proofs. Let and be additive categories and consider a biadditive functor ×→. The category , is defined by letting objects be elements α(C,A) with C and A, while a morphism from α(C,A) to β(D,B) is a pair (a,c) with a(A,B), c(C,D) and (C,a)(α)=(c,B)(β). This definition generalises the construction of the category of extensions from Subsection <ref>. We finish this section by demonstrating the use of <ref> in a concrete example. Consider a commutative ring R. For integers m,n≥0, let m,n(R) denote the set of m× n matrices which, when m,n>0, have entries in R. Note that m,0(R) consists of a single empty column vector of length m. Likewise, the set 0,n(R) consists of an empty length n row. There is a unique way of extending matrix multiplication to define a function l,m(R)×m,n(R)→l,n(R) for any integers l,m,n≥ 0. Let be the category of rectangular matrices over R, defined as follows. Objects of are rectangular matrices Xm,n(R) for m,n≥0, and a morphism from Xm,n(R) to Yp,q(R) is defined by a pair of matrices (A,B)q,n(R)×p,m(R) such that BX=YA. Composition is defined by component-wise matrix multiplication. The identity of an object Xm,n(R) is the pair (*[]I_n,*[]I_m) of identity matrices. The category is preadditive, where the addition of morphisms is given by component-wise matrix addition. Furthermore, is in fact additive, where the direct sum X⊕ Y of Xm,n(R) and Yp,q(R) is given by the block-diagonal matrix in m+p,n+q(R) formed by taking X and Y in the diagonal blocks and 0 elsewhere. The zero object of is the unique element of 0,0(R). We now use <ref> to compute the idempotent completion . Consider first the category of finitely generated free R-modules, and let ×→ be the biadditive functor given by the -bifunctor (-,-). Using that any finitely generated free R-module is isomorphic to *[]R^n for some n≥ 0, it is straightforward to check that is equivalent to the category of extensions. This uses the fact that commutative rings have an invariant basis number, meaning that the value of n is unique. Applying <ref> now gives ≃. Observing that = (-,-), it follows from <cit.> that is equivalent to the arrow category ^→ of , i.e. the category whose objects and morphisms are given by morphisms and commutative squares in , respectively. It is well-known that the idempotent completion of is the category of finitely generated projective R-modules; see e.g. <cit.>. We can thus conclude that is equivalent to the arrow category ^→. § 2-CATEGORICAL COMPATIBILITY The aim of this section is to prove <ref> in <ref>, which asserts that the constructions and results we have exhibited so far are compatible in a 2-categorical framework. We start by recalling the definition of the 2-functor from <cit.>. Given an n-exangulated functor (,Γ)(,,)→(,,), it follows from <cit.> that there is a corresponding exact functor (,Γ) (,)→ (,). This functor is defined by (,Γ)(α) = Γ(α) on objects and by (,Γ)(a,c) = ( a, c) on morphisms. In addition, given an n-exangulated natural transformation (,Γ) ⇒ (,Λ) of n-exangulated functors (,Γ), (,Λ) (,,)→(,,), one can define a natural transformation (,Γ)⇒(,Λ) given by *[]_α = (A,C) for α(C,A); see <cit.>. (See <cit.>.) Let = (0,1,2) n→ be defined by the assignments ini→i, where: 0 (,,) (,), 1 (,Γ) (,Γ), 2 ( ) . It was shown in <cit.> that defines a functor n→ that satisfies the properties of a 2-functor, and thus restricts to a genuine 2-functor . <ref> allow us to restrict to idempotent complete categories. By abuse of notation, we write for this restriction, where it should be noted that is not the completion of the functor in the sense of <ref> (see (<ref>)). We have = (0,1,2) n→, where the assignment _i is defined as a restriction of i for i {0,1,2} and satisfies the same properties. Again, we obtain a 2-functor n→ when restricting 0-cells to small categories. Recall that the functors and were defined in Definitions <ref> and <ref>, respectively, while the exact equivalence (,,) (, ) → (, ) for (,,)n0 was defined in the proof of . The collection of exact equivalences (,,) for (,,)n0 defines a natural transformation ⇒ as indicated in the diagram [column sep=1cm] nd[swap]r d nr[swap][Rightarrow, shorten <= 18pt, shorten >= 20pt,yshift=-5pt, xshift=2pt]ur . In order to demonstrate the naturality of , we must show that [column sep=2cm] (,) r(,,)d[swap](,Γ) (,) d(,Γ) (,) r(,,) (,) commutes in for any n-exangulated functor (,Γ) (,,) → (,,). That is, we need to show that (,Γ)(,,) and (,,)(,Γ) are equal as functors →. To this end, let (A,α,C) be an object in . On the one hand, we have that (,Γ)(,,)(A,α,C) = (,Γ) (α,(A,C)) = (Γ(α),(A,C)), while on the other hand (,,)(,Γ) (A,α, C) = (,,) (A,Γ(α), C) = (Γ(α),(A,C)). Hence, the functors (,Γ)(,,) and (,,)(,Γ) agree on objects. Consider next a morphism ((B,a,A),(D,c,C)) (A,α, C) → (B,β, D) in . We have (,Γ)(,,)((B,a,A),(D,c,C)) = (,Γ) ((B,D),(a,c),(A,C)) = ((B,D),( a, c),(A,C)) = (,,) ((B, a,A),(D, c,C)) = (,,)(,Γ) ((B,a,A),(D,c,C)), which completes the proof. The next result says that satisfies the defining property of a 2-natural transformation between 2-functors; see <cit.>. Let (,Γ) ⇒ (,Λ) be an n-exangulated natural transformation between n-exangulated functors (,Γ), (,Λ) (,,)→(,,). Then the square ()(,Γ)∘(,,)[equal]r[Rightarrow]d[swap]()()𝕀(,,) (,,)∘ ()(,Γ) [Rightarrow]d[xshift=2pt, yshift=0pt]𝕀(,,)()() ()(,Λ)∘(,,)[equal]r (,,)∘()(,Λ) commutes in ( , ). Note first that we have the horizontal equalities by <ref>. Consider an arbitrary object (A,α,C)((C,C),(A,A)) in = ()(,,). On the one hand, (()() 𝕀(,,)(A,α,C) = ()((,,)(A,α,C) ∘ ()(,Γ)(𝕀(,,)(A,α,C) = ()((α,(A,C))∘(,Γ)(𝕀(α,(A,C))) = ()((α,(A,C)) = ( ()(α,(A,C)) = *[]_(α,(A,C)) = ((,Λ) (A,C),(,Λ) (A,C)*[]_α(,Γ)(A,C),(,Γ) (A,C)) = ((A,C), ((A)AA, (C)CC), (A,C)). On the other hand, we have (𝕀(,,)()()(A,α,C) = (𝕀(,,)()(,Λ)(A,α,C)∘(,,)(()()(A,α,C) = (,,)*[]_(A,α,C) = (,,) ( (A, (A)AA, A) , (C, (C)CC, C) ) = ((A,C), ((A)AA,(C)CC), (A,C)), which finishes the proof. Restricting to small categories, <ref> and <ref> yield the following corollary, demonstrating that idempotent completions and extension categories are compatible constructions in a 2-category-theoretic sense. There is a 2-natural transformation ⇒ of 2-functors from n to consisting of exact equivalences. § THE WEAK IDEMPOTENT COMPLETION The aim of this section is to relate our main results to the setup of weak idempotent completions. Note that the definitions and results in <ref> rely on concepts and notation which should be recalled from previous sections. Many of the proofs in the weakly idempotent complete case are straightforward modifications of those for the idempotent completion. However, notice that our proof of the key result <ref> differs significantly from the proof of <ref> and relies on a result from <cit.>. Recall that an additive category is said to be weakly idempotent complete if every retraction has a kernel or, equivalently, if every section has a cokernel; see <cit.>. Every idempotent complete category is weakly idempotent complete; see e.g. and <cit.>. For more detail on weak idempotent completions, see e.g. <cit.>. (See <cit.>.) The weak idempotent completion of is the full subcategory of that consists of all objects (X,e) for which 𝕀X-e splits in . Note that is an additive subcategory of and that there is a canonical additive inclusion functor → defined by (X) (X,𝕀X) for X and (f) (𝕀Y,f,𝕀X) for f(X,Y). This functor is 2-universal among additive functors from to weakly idempotent complete categories; see <cit.>. The functor → factors through via the canonical inclusion functor →, which is the identity on objects and morphisms. In other words, there is a commutative diagram rrdr[swap] ur[swap] of additive categories and functors. Suppose that (,) is an exact category. One defines an exact structure on as follows. An object of →→ is in if it is a direct summand of an object in the image of under the functor →→→→→ induced by →. In particular, a kernel-cokernel pair lies in if and only if it is a kernel-cokernel pair in in which all three terms lie in . <ref> below shows that (,) is a fully exact subcategory of (,). This means that is extension-closed in (,) and that coincides with the inherited exact structure; see <cit.>. The pair (,) is a fully exact subcategory of (,). The inclusion (,) → (,) is a fully faithful exact functor that reflects exactness. Suppose that [column sep=1.7cm] (A,A) r(B, a, A) (B,B) r(C, b, B) (C,C) is a conflation in (,) with (A,A),(C,C). By <cit.>, there is an object (D,D) and an isomorphism [column sep=1.7cm] (A,A) r(B, a, A)[equals]d (B,B) r(C, b, B)d(D, r, B) (C,C) [equals]d (A,A) r(D, c, A) (D,D) r(C, d, D) (C,C) in the category *[]𝐊^3_(;(A,A),(C,C)) defined in <cit.>. By <cit.>, the morphism (D, r, B) is an isomorphism in , so is extension-closed. Since a sequence lies in if and only if it lies in and has all terms in , the inherited exact structure (see <cit.>) coincides with . This shows that (,) is a fully exact subcategory of (,). As a consequence, we observe that (,) → (,) from (<ref>) is an exact functor. One sees directly that (,) → (,) is fully faithful and exact. That it reflects exactness follows from (,) → (,) reflecting exactness (see <ref>) and the commutative diagram (<ref>). Suppose that (,) is also an exact category. Let , (,)→ (,) be exact functors and consider a natural transformation ⇒. The completions and restrict to give exact functors , (,)→ (,). Moreover restricts to a natural transformation (X,e)(X,e)⇒. We write and for the restrictions to weakly idempotent complete 0-cells in and , respectively, and note that is a 2-category. Let = (0,1,2) → be defined by the assignments ii→i, where: 0 (,) (,), 1 () , 2 ( ) . These assignments are well-defined by the discussion above. As a consequence of <ref>, we see that satisfies the properties of a 2-functor, because and are just restrictions of and , respectively. Thus, one deduces an analogue of <ref>. Throughout the rest of this section, suppose that (,,) is an n-exangulated category. The next result is an analogue of <ref>, but interestingly the proof is very different. If is weakly idempotent complete, then is also weakly idempotent complete. Let (a,c)α→β be a morphism in for α(C,A) and β(D,B). Suppose that this morphism is a section. Thus, there is a retraction (r,s)β→α in satisfying (ra,sc) = (r,s)∘ (a,c) = 𝕀α = (𝕀A,𝕀C). In particular, this implies that a and c are sections in . Since is weakly idempotent complete, these morphisms each admit a cokernel, which we denote by b = a and d = c. We have that (b,d) is a cokernel of (a,c) in by , so is weakly idempotent complete. It was shown in <cit.> that the weak idempotent completion of admits an structure because it is an extension-closed subcategory of (,,1mu) in the sense of <cit.>. We denote the corresponding n-exangulated category by (,,3mu). We now recall how and are defined. (See <cit.>.) The biadditive functor ×→ is the restriction of to ×. The exact realisation of is defined as follows. Suppose that (A,α,C)((C,C),(A,A)) is an -extension. Then (A,α,C) = [ ∙ ] for some complex ∙ in by <cit.>. Thus, it is declared that (A,α,C) = [ ∙ ]. <ref> has the following immediate corollary. (, ) is a weakly idempotent complete exact category. Recall from the proof of <ref> that we have an exact equivalence (,,) (, ) → (, ) with quasi-inverse (,,) for each n-exangulated category (,,). It follows from <ref> that (,,) restricts to a functor (,,) (, ) → (, ). To see this, let (A,α,C)((C,C),(A,A)) and consider (,,)(A,α,C) = (α,(A,C)). Note that 𝕀α - (A,C)α→α is a morphism of -extensions since (A,C)(α). We must show that 𝕀α - (A,C) = (𝕀A-A,𝕀C-C) splits in . This follows from <ref>, as (A,A),(C,C) means that 𝕀A-A and 𝕀C-C split in . A similar argument as above shows that (,,) restricts to a functor (,,) (, ) → (, ). Note that (,,) and (,,) are mutually quasi-inverse as they are restrictions of (,,) and (,,). Since it is straightforward to check that (,,) and (,,) preserve the exact structures, we have the following. There is an exact equivalence (,,) (, ) → (, ) given by the restriction of (,,). Suppose that (,Γ) (,,)→ (,,) is an n-exangulated functor. One can define a natural transformation Γ(-,-) ⇒(-,-) by setting ((C,C),(A,A))(A, α, C) (A, Γ(α), C). Notice that Γ is just a restriction of Γ. We claim that the pair (,Γ) is an n-exangulated functor (,,3mu)→(,,3mu). To verify this, assume that (A,α,C) = [∙], which implies . This yields , as (,Γ) is n-exangulated by <ref>. Since (∙) is a complex in , we obtain (Γ(A,α,C)) = [(∙)] = [(∙)] as required. Let (,Γ) ⇒ (,Λ) be an n-exangulated natural transformation between n-exangulated functors (,Γ),(,Λ) (,,)→ (,,). Using that the completion (,Γ) ⇒ (,Λ) is an n-exangulated natural transformation by <ref>, the same holds for the restriction (,Γ)⇒ (,Λ). We write n for the collections obtained by only considering weakly idempotent complete 0-cells in n. Based on the discussion above, we may thus define = (0,1,2) n→n using assignments , where: 0 (,,) (,,3mu), 1 (,Γ) (,Γ), 2 ( ) . It is straightforward to check that the analogue of <ref> holds for . As an application of <ref>, we can restrict to weakly idempotent complete categories. This restriction is denoted by = (0,1,2) n→, where i is the restriction of i for i {0,1,2} and satisfies the same properties. The proof of <ref> yields the next theorem. Similarly, analogues of <ref> and <ref> follow. The collection ' of exact equivalences (,,) for (,,)n0 defines a natural transformation ⇒. By <cit.>, there is an n-exangulated functor (,) (,,) → (,,1mu), where (α) = (𝕀A,α,𝕀C) for α(C,A). Similarly, it is shown in <cit.> that (,) (,,) → (,,3mu) is n-exangulated, where (α) = (𝕀A,α,𝕀C). Diagram (<ref>) can be augmented to a commutative diagram (,,) rr(,)dr[swap](,) (,,1mu) (,,3mu) ur[swap](,) in n, where (B,β,D) (B,β,D) for (B,β,D)((D,D),(B,B)). Using the functor n→, the diagram (<ref>) induces the commutative diagram (,) rr(,)dr[swap](,) (,) (,) ur[swap](,) in . Building on the work in <ref>, it is straightforward to check that we obtain <ref> below. We note that there is a similar commutative diagram involving the exact equivalence (,,) and its restriction (,,). The diagram [column sep=2cm] (, ) r(,)[equals]d (, ) r(,)d(,,)[swap]≃ (, ) d(,,)[swap]≃ (, ) r (, ) r (, ) in is commutative. 1 The authors are grateful to Dixy Msapato for directing them to a result in their paper <cit.>, which led to <ref>. Parts of this work were carried out while the second author visited Aarhus University, and while the first and fourth author visited NTNU. The authors thank the project “Pure Mathematics in Norway” funded by the Trond Mohn Foundation for supporting these stays. The first author is grateful to have been supported during part of this work by the Alexander von Humboldt Foundation in the framework of an Alexander von Humboldt Professorship endowed by the German Federal Ministry of Education. The third author is grateful to have been supported by Norwegian Research Council project 301375, “Applications of reduction techniques and computations in representation theory”. The fourth author is grateful to have been supported during part of this work by the Engineering and Physical Sciences Research Council (grant EP/P016014/1), and the London Mathematical Society with support from Heilbronn Institute for Mathematical Research (grant ECF-1920-57). In addition, the first and fourth authors gratefully acknowledge support from: the Danish National Research Foundation (grant DNRF156); the Independent Research Fund Denmark (grant 1026-00050B); and the Aarhus University Research Foundation (grant AUFF-F-2020-7-16). 1mybst

http://arxiv.org/abs/2303.06861v1

20230313050932

On the achievable degree of stability in strictly negative imaginary state feedback control

math.OC

Semantically Secure Private Set Intersection over Outsourced Multi-Owner Secret-Shared Databases Dongfang Zhao Received: date / Accepted: date ================================================================================================ empty empty In this paper we study the problem of determining the largest degree of stability that can be achieved for SISO systems using negative imaginary state feedback control. A state feedback result is given for synthesising a controller for a plant such that a given closed-loop transfer function is strictly negative imaginary with a prescribed degree of stability. By varying a parameter, the placement of the closed-loop system poles can be adjusted to give a prescribed degree of stability. We show the achievable degree of stability is related to the zero locations of the transfer function from the control input to the disturbance output of the nominal plant being controlled. Moreover, we offer results that outline the largest degree of stability that can be achieved for systems with distinct eigenvalues. § INTRODUCTION The theory of negative imaginary (NI) systems is broadly applicable to problems of robust vibration control for flexible structures; e.g., see <cit.>. In these control systems, unmodelled spillover dynamics can degrade control system performance or lead to instability if the controller is not designed to be robust against this type of uncertainty. NI systems theory provides a way analyzing robustness and designing robust controllers for such flexible structures in the case of collocated force actuators and position sensors; e.g., see <cit.>. Motivated by the robust stability properties of NI systems, research has been made into controller synthesis results with the aim of creating a closed-loop system with the NI or SNI property <cit.>. In this paper, we concentrate on the state feedback negative imaginary control problem as illustrated in Figure <ref>. In this control problem, the unmodelled flexible dynamics are represented by the plant uncertainty Δ(s) which are assumed to have the NI or SNI property. This is a natural assumption in the control of flexible structures, since it is known that any flexible structure with collocated force inputs and position outputs will have the NI property; e.g., see <cit.>. The (strictly) NI state feedback control problem is concerned with synthesising a state feedback control law u = Kx, such that the corresponding closed-loop system will be guaranteed to have the (strictly) NI property One method of approaching the SNI state feedback control problem has been to form a controller using the solution to a certain algebraic Riccati equation (ARE) <cit.>. However, it has been shown that the associated ARE used in these results does not have a stabilizing solution <cit.>. To avoid the computation complexity of solving this ARE, <cit.> proposed a controller synthesis approach using the solution to an algebraic Riccati equation (ARE) that could be obtained by solving two Lyapunov equations. The approach was computationally efficient, but yielded a closed-loop pole at the origin, ensuring a marginally stable closed-loop system. This pole at the origin was problematic, as in vibration control problems for flexible structures, pole placement is related to the degree of damping achieved for the nominal resonant modes. The papers <cit.> modified the approach of <cit.> using a perturbation applied to the plant matrix of the nominal system in order to ensure asymptotic stability of the closed-loop system. The perturbation achieved closed-loop asymptotic stability. However, it was not proven that the closed-loop system retained the NI property after such a perturbation. The paper <cit.> then showed that under suitable assumptions, the perturbation approach of <cit.> did in fact lead to an SNI closed-loop system and by varying the perturbation parameter, the closed-loop poles of the system could be shifted further into the complex plane, resulting in a prescribed degree of stability. The Lyapunov equations used in forming the controller in this approach also depend on the perturbation parameter. As a consequence, there are choices of this parameter that result in the equations no longer being solvable. This puts a limit on the degree of stability that can be achieved. In this paper, we focus on SISO systems and build on the results of <cit.> to study the achievable degree of stability of an SNI closed-loop system formed using the perturbation approach of <cit.>. This issue is important since closed-loop degree of stability is a key performance measure in vibration control problems for flexible structures. Our main results show that the achievable degree of stability for a given system is limited by the eigenvalues of a certain matrix that is related to the zeros of the transfer function from the control input to the disturbance output of the nominal system. Under suitable assumptions, we quantify how the perturbation approach of <cit.> effects this matrix, and the maximum degree of stability that can be achieved before we can no longer guarantee the SNI property of the closed-loop system. § PRELIMINARY DEFINITIONS In this section, we briefly present definitions for both negative imaginary and strictly negative imaginary systems. First consider the linear time-invariant (LTI) system, ẋ(t) = Ax(t) + Bu(t), y(t) = Cx(t) + Du(t), where A ∈ℝ^n × n, B ∈ℝ^n × m, C ∈ℝ^m × n and D ∈ℝ^m × m. The following two definitions relate to the NI and SNI properties of the transfer function matrix G(s) = C(sI - A)^-1B+D corresponding to the system (<ref>). We do not consider a direct feed-through term from disturbance to output in this work. A square transfer function matrix G(s) is NI if the following conditions are satisfied <cit.>: * G(s) has no pole in Re[s]>0. * For all ω≥ 0 such that jw is not a pole of G(s), j(G(jω) - G(jω)^*) ≥ 0. * If s=jω_0, ω_0 > 0 is a pole of G(s) then it is a simple pole. Furthermore, if s=jω_0, ω_0 > 0 is a pole of G(s), then the residual matrix K = lim_s → jω_0 (s-jω_0)jG(s) is positive semidefinite Hermitian. * If s=0 is a pole of G(s), then it is either a simple pole or a double pole. If it is a double pole, then, lim_s → 0 s^2G(s) ≥ 0. Also, an LTI system (<ref>) is said to be NI if the corresponding transfer function matrix G(s) = C(sI-A)^-1B + D is NI. A square transfer function matrix G(s) is SNI if the following conditions are satisfied <cit.>: * G(s) has no poles in Re[s] ≥ 0. * For all ω > 0 such that jw is not a pole of G(s), j(G(jω) - G(jω)^*) > 0. Also, an LTI system (<ref>) is said to be SNI if the corresponding transfer function matrix G(s) = C(sI-A)^-1B + D is SNI. (<cit.>) Suppose the system (<ref>) is a given state space realization with R = CB + B^TC^T > 0 and D=D^T. If there exists a real P=P^T≥0 such that PA + A^TP + (CA - B^TP)^TR^-1(CA-B^TP) = 0 with σ(A-BR^-1(CA-B^TP))⊂ℂ_≤ 0. Then, the state space realization and corresponding transfer function is NI. § STRICTLY NEGATIVE IMAGINARY STATE FEEDBACK In this paper are concerned with the use of state feedback control to achieve a closed-loop system which is SNI and has a prescribed degree of stability. Moreover, we are concerned with how stable can we make the eigenvalues of the closed-loop system while maintaining the SNI property. In this section, we introduce the SNI state feedback control problem along with the state feedback results of the papers <cit.>. In addition to this we will introduce supporting lemmas that contextualise the perturbation method used in the state feedback theorem of <cit.>. §.§ Problem Formulation For the SNI state feedback method we will consider in this paper, we require the following controlled linear state space system: ẋ = Ax + B_1w + B_2u, z = C_1x, where A ∈ℝ^n × n, B_1 ∈ℝ^n × m, B_2 ∈ℝ^n × m, and C_1 ∈ℝ^m × n. This system corresponds to the nominal plant transfer function matrix G(s) in Figure <ref>. If we apply a state feedback control law u=Kx to this system, the corresponding closed-loop system has state space representation ẋ = (A + B_2K)x + B_1w, z = C_1x, with corresponding closed-loop transfer function G_cl(s) = C_1(sI - A - B_2K)^-1B_1. We assume that the system (<ref>) satisfies the following assumptions. The matrix C_1B_2 is non-singular. The matrix R = C_1B_1 + B_1^TC_1^T > 0. Assumption <ref> is essential for ensuring the existence of the controller in the state feedback theorem that will follow. Assumption <ref> is necessary to ensure the invertability of a certain matrix in several of the proofs that follow. §.§ Preliminary Results We now introduce three results that are needed to understand the SNI state feedback theorem discussed in this paper. Firs,t we will introduce a lemma that details the behavior of NI systems under an ϵI perturbation applied to the plant matrix. Then, we present a Schur decomposition applied to a matrix that is used in controller synthesis to avoid the potential computational difficulties of solving an ARE that does not have a stabilizing solution. In the third result, we remark that the eigenvalues of the matrix used in the Schur decomposition actually corresponds to the zeros of a certain related transfer function. Our main results will show that these eigenvalues dictate the maximum degree of stability that the closed-loop system can achieve under SNI state feedback. Suppose the system (<ref>) is a given state space realization with m=1 and an NI transfer function G(s). Then the perturbed state space realization [ A-ϵ I B C D ] will be SNI for all ϵ > 0. We now introduce the matrix A_r = Q(A+ϵ I) = A_q + ϵ Q. where Q = I-B_2(C_1B_2)^-1C_1, A_q = QA, and ϵ > 0 is a parameter which will determine the degree of stability of the closed-loop system in the SNI state feedback theorem of the next subsection. The matrix A_r is fundamental to the state feedback theorem that follows and is discussed further in Remark <ref>. After applying a real Schur decomposition to the matrix A_r as in (<ref>), we may take the orthogonal matrix U and construct the following matrices: à = U^TA_rU = [ Ã_11 Ã_12 0 Ã_22 ], B̃ = U^TB_1 = [ B̃_11 B̃_22 ], C̃ = U^T(B_2(C_1B_2)^-1 - B_1R^-1) = [ C̃_11 C̃_22 ], Z̃ = U^TZU = B̃R^-1B̃^T - C̃RC̃^T = [ Z̃_11 Z̃_12 Z̃_21 Z̃_22 ], where Ã_11 has all of its eigenvalues in the closed left half of the complex plane and Ã_22 is an anti-stable matrix; i.e., σ(Ã_22) ⊂{s:Re[s]>0}. Here, U is an orthogonal matrix dependent on ϵ and is obtained through the real Schur transformation; see Section 5.4 of <cit.>. Also, Z = B_1(B_2^TC_1^T)^-1B_2^T + B_2(C_1B_2)^-1B_1^T - B_2(C_1B_2)^-1R(B_2^TC_1^T)^-1B_2^T. Note that the choice of matrix A_r for the Schur decomposition above is not arbitrary. The eigenvalues of A_q correspond to the zeros of the transfer function from u to z in (<ref>), plus a zero at the origin. To see this, set w≡ 0 and note that (<ref>) implies ż = C_1Ax+C_1B_2u. Since (C_1B_2) is invertible by assumption, we can rearrange (<ref>) to write u = (C_1B_2)^-1ż - (C_1B_2)^-1C_1Ax. Now, we can substitute (<ref>) into (<ref>) to give ẋ = Ax + B_2(C_1B_2)^-1ż- B_2(C_1B_2)^-1C_1Ax = (A- B_2(C_1B_2)^-1C_1A)x + B_2(C_1B_2)^-1ż = A_qx + B_2(C_1B_2)^-1ż. These equations define the inverse system which maps from ż to u. Considering the Laplace transform of ż, suppose z(0)=0 and observe that ℒ(ż) = sℒ(z) + z(0). Thus, the eigenvalues of A_q will be the zeros of the transfer function from u to z plus a zero at zero. §.§ Negative Imaginary State Feedback In this section we introduce the SNI state feedback theorem that we will be analysing in our main result to determine the degree of stability that can be achieved. This theorem can be found in <cit.>. Consider the system (<ref>) with m=1 satisfying Assumptions A1-A2. For a given ϵ > 0, there exists a static state feedback matrix K such that the closed-loop system (<ref>) is SNI with degree of stability ϵ if there exist T ≥ 0 and S ≥ 0 such that -Ã_22T - TÃ_22^T + C̃_22RC̃_22^T = 0, -Ã_22S - SÃ_22^T + B̃_22R^-1B̃_22^T = 0, T-S > 0, where Ã_22, B̃_22 and C̃_22 are obtained from the Schur decomposition (<ref>). Moreover, if the conditions (<ref>)-(<ref>) are satisfied, then the required state feedback controller matrix K is given by K = (C_1B_2)^-1(B_1^TP - C_1A - ϵ C_1 - R(B_2^TC_1^T)^-1B_2^TP), where P = UP̃U^T and P̃ = [ 0 0 0 (T-S)^-1 ]≥ 0. Here, U is the orthogonal matrix obtained through the Schur transformation (<ref>). When the A_q matrix from the Schur decomposition (<ref>) has no unstable eigenvalues, the dimension of Ã_22 is equal to zero. Therefore, we cannot form (<ref>)-(<ref>). This situation corresponds to the case in which the transfer function from u to y in (<ref>) is minimum phase. This case is also considered in the paper <cit.> for the special case of B_1= B_2 in (<ref>). The following theorem is a minor result that allows us to use SNI state feedback in the case in which the transfer function from u to y in (<ref>) is minimum phase. Consider the system (<ref>) with m=1 that satisfies Assumptions A1-A2. Suppose the matrix A_q, obtained from the Schur decomposition (<ref>) has no unstable eigenvalues. For a given ϵ > 0 such that the corresponding matrix A_r also has no unstable eigenvalues, there exists a static state feedback matrix K such that the closed-loop system (<ref>) is SNI with degree of stability ϵ. The required state feedback controller matrix K is given by K = -(C_1B_2)^-1(C_1A + ϵ C_1 ). Proof of Theorem <ref>: For design purposes we will begin by replacing the open loop plant matrix A with the perturbed matrix A_ϵ = A + ϵI. Now, using Lemma <ref>, we know that the closed-loop system (<ref>) is NI if there exists a real P=P^T≥0 such that P + Â^TP + PB_1R^-1B_1^TP + Q = 0 where  = A_cl - B_1R^-1C_1A_cl, R = C_1B_1 + B_1^TC_1^T, Q = A_cl^TC_1^TR^-1C_1A_cl, A_cl = A_ϵ + B_2K, with σ(A_cl-B_1R^-1(CA_cl-B^TP)) ⊂ℂ_≤ 0. Let P=0. Then, (<ref>) reduces to the condition A_cl^TC_1^TR^-1C_1A_cl = 0. We will consider the matrix C_1A_cl. For our choice of K, and noting that for a SISO system, (C_1B_2) is a scalar, C_1A_cl = A+ϵ I + B_2K reduces to C_1A_cl = C_1A+ϵ C_1 -C_1B_2(C_1B_2)^-1(C_1A + ϵ C_1) = 0 Therefore, P=0 satisfies (<ref>). We will now show that σ(A_cl-B_1R^-1(CA_cl-B^TP)) ⊂ℂ_≤ 0. We begin by considering A_cl-B_1R^-1(CA_cl-B_1^TP). When P=0, and substituting for K = -(C_1B_2)^-1(C_1A + ϵ C_1 ) in this expression, we see that A_cl-B_1R^-1(CA_cl-B_1^TP) = A_cl-B_1R^-1CA_cl = A_ϵ - B_2(C_1B_2)^-1C_1A_ϵ = A_r, where A_r is defined in (<ref>). Thus, (<ref>) implies σ(A_cl-BR^-1(CA_cl-B^TP)) = σ(A_r). Since σ(A_r)⊂ℂ_≤ 0 follows from our assumption that A_r has no unstable eigenvalues, we have σ(A_cl-BR^-1(CA_cl-B^TP)) ⊂ℂ_≤ 0. Therefore, it follows from Lemma <ref> that the closed-loop system is NI. Further, Lemma <ref> then implies that the actual closed-loop system corresponding to the unperturbed system will have all its poles shifted by an amount ϵ to the left in the complex plane. Moreover, the actual closed-loop system is SNI with degree of stability ϵ. ▪ § THE ACHIEVABLE DEGREE OF STABILITY This section contains the main results of this paper and provides a solution to the maximum degree of stability that can be achieved using Theorem <ref> to design a controller for SNI state feedback. This section is broken into two subsections. The first offers preliminary results needed in the proofs of the main results. In the second subsection, we offer two results that outline the maximum degree of stability that can be achieved using Theorem <ref>. There are two results because depending on the zeros of the transfer function from u to z in (<ref>), the matrix A_r may have no unstable poles. As a consequence, Theorem <ref> cannot be used and Theorem <ref> must be evaluated instead. §.§ Preliminary Results In this section we provide lemmas necessary to the proof of our main results. Consider the system (<ref>) and suppose the dimensions are such that (C_1B_2) ≠ 0 is a scalar. In addition, let w = (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2. The vector C_1 is orthogonal to w. We show our result using a simple algebraic manipulation. Observe that w C_1^T = (C_1B_1B_2^T-C_1B_2B_1^T/(C_1B_2)^2)C_1^T = C_1B_1B_2^TC_1^T-C_1B_2B_1^TC_1^T/(C_1B_2)^2 = 0. The vector C_1 defined in (<ref>) is in the left-nullspace of the matrices Q, A_q and A_r, where Q, A_q and A_r are defined in (<ref>), (<ref>) and (<ref>). We can show that C_1 is in the left-nullspace of Q, A_q and A_r using the following routine algebra: C_1Q = C_1(I-B_2(C_1B_2)^-1C_1) = C_1-C_1B_2(C_1B_2)^-1C_1 = C_1-C_1 = 0, C_1A_q = C_1QA = 0, C_1A_r = C_1(QA + ϵQ) = C_1QA + C_1Qϵ = 0 + 0 = 0. Consider the system (<ref>) with m=1 satisfying Assumptions A1-A2. Let Z be defined as in (<ref>). For some non-zero vector y, y^T Z y=0 if and only if (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2y = 0, or B_2^Ty = 0. Note that Assumption A1 is required to ensure C_1B_2 ≠ 0 in the manipulation to follow. Now, suppose there exists a non-zero vector y which we use to pre and post multiply Z to form y^TZy. We may expand y^TZy=0 as 0 = y^T[B_2(C_1B_2)^-1R(B_2^TC_1^T)^-1B_2^T - B_1(B_2^TC_1^T)^-1B_2^T - B_2(C_1B_2)^-1B_1^T]y. Using that fact that R = C_1B_1 + B_1^TC_1^T = 2C_1B_1 and C_1B_2 are scalars, we may further manipulate (<ref>) as follows 0 = y^TB_2(2C_1B_1)B_2^Ty/(C_1B_2)^2 - y^TB_1B_2^Ty/C_1B_2 - y^TB_2B_1^Ty/C_1B_2 = (2C_1B_1)B_2^TyB_2^Ty/(C_1B_2)^2 - B_1^TyB_2^Ty/C_1B_2 - B_1^TyB_2^Ty/C_1B_2 = ((2C_1B_1B_2^T)y/(C_1B_2)^2 - B_1^Ty/C_1B_2 - B_1^Ty/C_1B_2)B_2^Ty = 2(C_1B_1B_2^T/(C_1B_2)^2 - B_1^T/(C_1B_2))yB_2^Ty = 2(C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2yB_2^Ty. Now, there are only two cases where (<ref>) is equal to zero, either: (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2y = 0, or B_2^Ty = 0, as required. Consider the system (<ref>) with m=1 satisfying Assumptions A1-A2. Let ϵ > 0 be given and A_r defined as in (<ref>) and suppose there exist matrices T, S ≥ 0 satisfying: -Ã_22T - TÃ_22^T + C̃_22RC̃_22^T = 0, -Ã_22S - SÃ_22^T + B̃_22R^-1B̃_22^T = 0, where Ã_22, B̃_22 and C̃_22 are obtained from the Schur decomposition (<ref>). Let Z be defined as in (<ref>) and let X = T-S. Suppose, λ≠ 0 is an eigenvalue of Ã_22 with corresponding eigenvector v_2^T ∈ ker(Ã_22^T - λI) and let y=U[ 0 v_2^T]^T ∈ ker(A_r^T - λI), where U is obtained from the Schur decomposition (<ref>). Then y^TZy < 0 if and only if v_2^T X v_2>0. Suppose all assumptions are satisfied and (<ref>)-(<ref>) hold. Subtracting (<ref>) from (<ref>) gives the Lyapunov equation, 0 = -Ã_22S - SÃ_22^T + B̃_22R^-1B̃_22^T - (-Ã_22T - TÃ_22^T + C̃_22RC̃_22^T ) = Ã_22X + XÃ_22^T + Z̃_22. Note that the matrix Ã_22 is the 22 block of the matrix à obtained from the Schur decomposition (<ref>). By assumption, we know there exists a vector v_2 that satisfies v_2^TÃ_22 = λv_2^T, Ã_22^T v_2 = λ̅v_2. Now, pre-multiply (<ref>) by v_2^T and post-multiply by v_2 to form v_2^T Ã_22X v_2 + v_2^T XÃ_22^T v_2 + v_2^TZ̃_22v_2 = 0. We may proceed with the following manipulation: 0 = v_2^T Ã_22X v_2 + v_2^T XÃ_22^T v_2 + v_2^TZ̃_22v_2 = λv_2^TX v_2 + v_2^T Xλ̅v_2 + v_2^TZ̃_22v_2 = (λ + λ̅)v_2^T X v_2 + v_2^TZ̃_22v_2. We now rearrange (<ref>) to form v_2^T X v_2 = - v_2^TZ̃_22v_2/(λ + λ̅), where, (λ + λ̅) > 0. In order to see the relationship between v_2^T X v_2 and y^TZy, observe that v^TZ̃v =[ 0 v_2^T ][ Z̃_11 Z̃_12; Z̃_21 Z̃_22 ][ 0; v_2 ] = v_2^TZ̃_22v_2. Using (<ref>), we can rewrite (<ref>) as v_2^T X v_2 = - v^TZ̃v/(λ + λ̅). We know that y^T = v^TU^T by assumption. Thus we can substitute back into (<ref>) to see that v_2^T X v_2 = - v^TZ̃v/(λ + λ̅) = - y^TZy/(λ + λ̅). Since (λ + λ̅) > 0 and (<ref>) was arrived at using only algebraic manipulation, we can infer that y^TZy < 0 if and only if v_2^T X v_2>0. Consider a SISO system (<ref>) satisfying Assumptions A1-A2. Let Z,A_r and X=T-S be defined as in (<ref>), (<ref>) and (<ref>). Suppose there exists a vector y such that y=U[ 0 v_2]^T ∈ ker(A_r^T - λI), where λ > 0 and U is obtained from the Schur decomposition (<ref>). Under these assumptions, v_2^TXv_2=0 if and only if either of the following conditions holds (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2y = 0, or B_2^Ty = 0. Suppose either (<ref>) or (<ref>) hold. We know from Lemma <ref> that this immediately implies y^T Z y=0. Also, we know from Equation (<ref>) in Lemma <ref> that for a vector y such that y=U[ 0 v_2]^T ∈ ker(A_r^T - λI), then v_2^T X v_2 = - v^T(Z̃)v/(λ + λ̅) = - y^TZy/(λ + λ̅), where, (λ + λ̅) > 0. Since (λ + λ̅) > 0, it follows from (<ref>) that y^T Z y=0 implies v_2^T X v_2=0. Conversely, suppose v_2^T X v_2=0. We can use the relationship (<ref>) and after substituting in v_2^T X v_2=0 we see that - y^TZy/(λ + λ̅) = 0. Since, (λ + λ̅) > 0 by assumption, this implies y^TZy = 0. Now, Lemma <ref> states that y^T Z y=0 if and only if (<ref>)-(<ref>) hold. Consider the system (<ref>) satisfying Assumptions A1-A2. Let Z,A_r and X=T-S be defined as in (<ref>), (<ref>) and (<ref>). Suppose the Schur decomposition (<ref>) is such that dim(Ã_22)=1, and let y=U[ 0 1]^T ∈ ker(A_r^T - λI), where λ = Ã_22 > 0 and U is obtained from the Schur decomposition (<ref>). Then X=0, if and only if either of the following conditions holds (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2y = 0, or B_2^Ty = 0. This is a special case of Lemma <ref> and the proof is the same noting that v_2 in that proof is equal to 1 in this case. We know from Lemma <ref> that for a vector y such that y=U[ 0 v_2]^T ∈ ker(A_r^T - λI), X is related to Z by v_2^T X v_2 = - y^TZy/(λ + λ̅). When dim(Ã_22)=1, v_2 = 1 and we can substitute y into (<ref>) to obtain X = - [ 0 1 ]^TZ̃[ 0 1 ]/(λ + λ̅) = - y^TZy/(λ + λ̅), where, (λ + λ̅) > 0. We know from Lemma <ref> that y^T Z y=0 if and only if (<ref>)-(<ref>) hold. Since (λ + λ̅) > 0, it follows from (<ref>) that y^T Z y=0 if and only if X=0. <cit.> Let matrix A, scalars λ and μ, and non zero vectors x, y be given. Suppose that Ax=λx and y^TA=μy^T. If λ≠μ, then y^Tx = 0. <cit.> Let matrix A, eigenvalue λ and non zero vectors x and y be given. Suppose that λ is an eigenvalue of A such that Ax=λx and y^TAx=λy^T. If λ has algebraic multiplicity 1, then y^Tx≠ 0. Also, if λ has geometric multiplicity 1, then it has algebraic multiplicity 1 if and only if y^Tx≠ 0. The following lemma is important as it tells us how the eigenvalues of the matrix A_q behave under the perturbation A_r = A_q + ϵQ. Remember the eigenvalues of A_q correspond to the zeros of the transfer function from u to z in (<ref>), plus a zero at the origin. Also, the solvability of the Lyapunov equations in Theorem <ref> is also depend on the eigenvalues of A_r. Consider the uncertain system (<ref>) with m=1 satisfying Assumptions A1-A2. Consider the n × n matrix A_q = QA defined in (<ref>). Suppose A_q is not a defective matrix and has the distinct eigenvalues 0 , λ_1 , λ_2 , ... , λ_k, such that dim ( ker( A_q) ) = j , dim ( ker( A_q - λ_i I ) ) = n_i, i = 1,2,...,k, where 1 ≤ j ≤ n. Let a value of ϵ > 0 be given such that λ_i ≠ - ϵ ∀ i where i = 1,2,...,k and the matrix A_r = QA + ϵQ is not defective. Then A_r = QA+ϵQ has the distinct possible eigenvalues 0 , ϵ , λ_1 + ϵ , ... , λ_k + ϵ, such that dim ( ker( A_r - ϵI) ) = j - 1 dim ( ker( A_r - (λ_i + ϵ)I) ) = n_i, i = 1,2,...,k dim ( ker( A_r ) ) = 1 . The proof will be published elsewhere. §.§ Main results In this section we address the problem of determining the maximum degree of stability that can be achieved using the SNI state feedback control laws presented in Theorem <ref> and Theorem <ref>. These results are given under a number of assumptions on the eigenvalues of the matrix A_q. Future research will be directed towards more generalized results. Our results are offered as two separate cases that consider when the system has different zeros of the transfer function from u to z in (<ref>). For the results that follow, we assume the following assumption is satisfied: A_q has distinct eigenvalues λ_1, ⋯ , λ_n. This assumption is made so that we can leverage Lemma <ref>, which does not use generalized eigenvector. This assumption will be relaxed in future work. Case 1: The following case considers when the matrix A_q has a single unstable eigenvalue, a single eigenvalue at the origin, and n-2 stable eigenvalues. We only consider a single unstable eigenvalue here in order to simply the proof and conserve space. The eigenvalues of A_q are ordered such that Re[λ_1] ≤⋯≤ Re[λ_n-2] ≤ 0 ≤λ_n and we will assume that the following conditions hold: * For all i ∈{1,⋯,n-2}, Re[λ_i] < 0; * For i = n-1, λ_i= 0; * For i = n, λ_i> 0. Let γ be defined as γ = -Re[λ_n-2]. In other words, γ is the real part of the maximum eigenvalue of A_q less than zero. In the following theorem we will show that a system with an A_q matrix as above may be perturbed for all ϵ within the range 0 < ϵ < γ and still satisfy the conditions of Theorem <ref>. Consider the system (<ref>) that satisfies the conditions of Theorem <ref>. Suppose, the pair {A,B_2} is controllable, the matrix A_q, (<ref>), has distinct eigenvalues and the matrix A_q, has only one unstable eigenvalue; The closed-loop system formed using Theorem <ref> will be SNI with degree of stability ϵ, for any ϵ within the range 0 < ϵ < γ, where γ = -Re[λ_n-2]. Case 2: The following case considers when the matrix A_q has no unstable eigenvalues, a single eigenvalue at the origin, and n-1 stable eigenvalues. Now, consider a matrix A_q ∈ℝ^n × n with distinct eigenvalues λ_1, ⋯ , λ_n. Suppose the eigenvalues of A_q are ordered such that Re[λ_1] ≤⋯≤ Re[λ_n-1] < 0. Also, assume that the following conditions hold: * For all i ∈{1,⋯,n-1}, Re[λ_i] < 0; * For i = n, λ_i= 0. Let γ be defined as γ = -Re[λ_n-1]. In other words, γ is the real part of the maximum eigenvalue of A_q less than zero. In the following theorem we show that the system with an A_q matrix as above may be perturbed for all ϵ within the range 0 < ϵ < γ and still satisfy the conditions of Theorem <ref>; that is, the system will be SNI for all ϵ within the range 0 < ϵ < γ. Moreover, for ϵ > γ, the system may still be SNI if it satisfies Theorem <ref>. Consider the system (<ref>) with dimension n, that satisfies the conditions of Theorem <ref> when unperturbed. Suppose, the pair {A,B_2} is controllable; the matrix A_q defined in (<ref>) has distinct eigenvalues, and the matrix A_q, has no unstable eigenvalues. Then, the closed-loop system formed using Theorem <ref> will be SNI with degree of stability ϵ, for any ϵ within the range 0 < ϵ < γ, where γ = -Re[λ_n-1] is defined as in (<ref>). In addition, for ϵ in the range -Re[λ_n-1] < ϵ < -Re[λ_n-2], if the open-loop system satisfies the conditions of Theorem <ref> at ϵ, then the system will remain SNI for all ϵ within the range -Re[λ_n-2] < ϵ < -Re[λ_n-2]. Proof of Theorem <ref>: If a system satisfies the conditions of Theorem <ref>, there exists an X>0 such that Ã_22X + XÃ_22^T + B̃_22R^-1B̃_22^T - C̃_22RC̃_22^T = 0 Ã_22X + XÃ_22^T + Z̃_22 = 0, where Ã_22, B̃_22 and C̃_22 are obtained from the Schur decomposition (<ref>). Suppose the conditions of Theorem <ref> are satisfied when ϵ=0 and X_0>0 is a solution to (<ref>). The dimension of X_0 is related to Ã_22 by dim(X_0) = dim(Ã_22)=1, so in this case X_0 is scalar. Our proof methodology will show that for any ϵ within the range 0 < ϵ < -Re[λ_n-2], if there exists a left eigenvector of A_r such that X_ϵ=0, then X_0=0 must also be true, which contradicts our initial assumption. To that end, let γ = -Re[λ_n-2] and suppose there exists an ϵ within the range 0 < ϵ < γ such that X=0. Also, let y_1 be a left eigenvector of A_r such that y_1^TA_r = (μ + ϵ) y_1^T, where A_r is defined in (<ref>) and μ=λ_n is the anti-stable eigenvalue of A_q. According to Lemma <ref>, if X=0, then either (C_1B_1B_2^T-C_1B_2B_1^T)/(C_1B_2)^2y_1 = 0, or B_2^Ty_1 = 0. We now consider each of these conditions separately: Case 1, (<ref>) is satisfied: In this case, (C_1B_1B_2^T-C_1B_2B_1^T/(C_1B_2)^2)y_1 = w^Ty_1 = 0. Since we can easily check that the vector w is orthogonal to C_1, it follows that C_1^T ∈ ker(A_r^T). Moreover, we have distinct eigenvalues, therefore C_1 is also a basis for ker(A_r^T). Since w is orthogonal to the basis of ker(A_r^T), this must mean w∈ range(A_r). Thus, we can describe w using a linear combination of basis vectors for range(A_r), to give w = β_1 x_1 + β_2 x_2 + ⋯ + β_n-1x_n-1, where β_1,⋯,β_n-1∈ℝ and x_1,⋯,x_n-1 are eigenvectors of A_r corresponding to the n-1 non-zero eigenvalues. Using this form of w, we can restate (<ref>) as w^Ty_1 = 0 y_1^Tw = 0 β_1 y_1^Tx_1 + ⋯ + β_n-1y_1^Tx_n-1 = 0. According to Lemma <ref>, every term after β_1 y_1^Tx_1 will be equal to zero. Therefore (<ref>) is reduced to β_1 y_1^Tx_1 = 0. Equation (<ref>) tells us that Condition (<ref>) holds if either β_1=0 or y_1^Tx_1 = 0. Since we only have simple eigenvalues, it follows from Lemma <ref> that y_1^Tx_1 = 0 cannot occur. Therefore, we now consider β_1=0. If we substitute β_1=0 back into (<ref>), the first term is eliminated and we are left with w = β_2 x_2 + ⋯ + β_n-1x_n-1. We know from Lemma <ref> that A_q and A_r will have the same right-eigenvectors corresponding to a particular eigenvalue. Therefore, each vector x_i is an eigenvector of A_q and the following holds: A_q x_i = λ_i x_i, for i = 1,2,⋯, n-2,n. Also, for the eigenvalue μ=λ_n in (<ref>), there will be a left-eigenvector of A_q, y_0, that satisfies y_0^T A_q = μy_0^T. Now, consider w^Ty_0 = β_2 x_2^Ty_0 + ⋯ + β_n-1x_n-1y_0. We can apply Lemma <ref> to show that each term in this sum is orthogonal and therefore zero. Thus, w^Ty_0 = β_2 x_2^Ty_0 + ⋯ + β_n-1x_n-1y_0 = 0. Equation (<ref>) is just (<ref>). Therefore, we can conclude from Lemma <ref> that (<ref>) implies X_0=0. This contradicts our initial assumption that X_0>0. Case 2, (<ref>) is satisfied: In this case, let y_1 be a left eigenvector of A_r such that y_1^TA_r = (μ + ϵ) y_1^T and B_2^Ty_1 = 0. We know (A,B_2) is controllable by assumption. Since is (A,B_2) controllable, it follows immediately that (A + ϵI , B_2) is also controllable. Using this knowledge, we will make a manipulation on A_r to show that it is also controllable. To that end, remember A_r = QA + ϵQ, Q= I - B_2(C_1B_2)^-1C_2, and C_1B_2 is a scalar. Let k = -C_1(A+ϵI)/C_1B_2 and consider the following manipulation: A_r = Q(A + ϵI) = (A+ϵI) - (B_2C_1/C_1B_2)(A+ϵI) = A + ϵI + B_2K. Thus, we can conclude from (<ref>) that (A_r, B_2) is also controllable. An immediate consequence of (A_r, B_2) being controllable is that there can be no left eigenvector of A_r that is orthogonal to B_2. Therefore, (<ref>) can never be satisfied and we have a contradiction. We have shown that both cases lead to a contradiction. ▪ Proof of Theorem <ref>: For ϵ within the range 0 < ϵ < γ, where γ = -Re[λ_n-1], the closed-loop system is SNI follows directly from Theorem <ref>. For -Re[λ_n-2] < ϵ < -Re[λ_n-2], the matrix A_q will have attained an unstable eigenvalue. Consequently, X has dimension one. Therefore, if the system satisfies the conditions of Theorem <ref> for -Re[λ_n-2] < ϵ < -Re[λ_n-2], then the system will remain SNI for all ϵ within the range 0< ϵ < -Re[λ_n-2] follows directly from Theorem <ref>. ▪ § ILLUSTRATIVE EXAMPLE In this section we provide an illustrative example that illustrates Theorem <ref>. Consider an uncertain system A = [ -1 0 -1; 1 0 -1; -1 2 1 ], B_1 = [ 1; 1; 1 ], B_2 = [ 0; 1; 1 ], C_1 = [ 1 1 0 ], D = 0. that satisfies assumptions A1-A2. The A_q matrix for this system is A_q = [ -1 0 -1; 1 0 1; -1 2 3 ], with σ(A_q) = {-1.6458 , 0 , 3.6458 }. We can readily verify that for any ϵ<1.6458 this system satisfies the conditions of Theorem <ref>, and the closed-loop system is SNI. Now, we choose ϵ=1.6458 and reevaluate. With ϵ=1.6458, the matrix A_r = QA+ϵQ has the form A_r = [ 0.6458 0 -1; -0.6458 0 1; -2.6458 0.3542 4.6458 ], with σ(A_r) = {0 , 5.2946 } and the 0 eigenvalue has multiplicity two. At exactly ϵ=1.6458, the matrix X is undergoing a dimension change and cannot be evaluated. Therefore, we analyse X as we approach ϵ=1.6458 from the positive direction on the real axis; i.e., we consider ϵ^+ : ϵ→ 1.6458^+. We can reform X_ϵ^+ for our perturbed system as X_ϵ^+ = [ -5.5431 0.2815; 0.2815 0.2477 ]. It is easily verified that X_ϵ^+ is dimension 2 and no longer positive definite, with a negative eigenvalue at -5.5568. We no longer satisfy the conditions of Theorem <ref> and cannot guaranteed the SNI property. § CONCLUSIONS In this paper we have studied the maximum prescribed degree of stability that can be achieved by SISO closed-loop systems formed using SNI state feedback. These results show that in the cases considered, the achievable degree of stability for the closed-loop system is related to the zeros of the transfer of the nominal plant from the control input to the disturbance output. These results were given under a number of assumptions and future research will be directed towards relaxing these assumptions and extending the results to MIMO systems. -12cm IEEEtran

http://arxiv.org/abs/2303.06978v1

20230313102550

A Newton-like Method based on Model Reduction Techniques for Implicit Numerical Methods

math.NA

Self-supervised based general laboratory progress pretrained model for cardiovascular event detection Li-Chin Chen^1, Kuo-Hsuan Hung^1, Yi-Ju Tseng^2,  Member, IEEE, Hsin-Yao Wang^3, Tse-Min Lu^4,5,6, Wei-Chieh Huang^7,4,6, Yu Tsao^1,  Senior Member, IEEE March 30, 2023 =================================================================================================================================================================================================================== empty empty In this paper, we present a Newton-like method based on model reduction techniques, which can be used in implicit numerical methods for approximating the solution to ordinary differential equations. In each iteration, the Newton-like method solves a reduced order linear system in order to compute the Newton step. This reduced system is derived using a projection matrix, obtained using proper orthogonal decomposition, which is updated in each time step of the numerical method. We demonstrate that the method can be used together with Euler's implicit method to simulate CO_2 injection into an oil reservoir, and we compare with using Newton's method. The Newton-like method achieves a speedup of between 39% and 84% for systems with between 4,800 and 52,800 state variables. § INTRODUCTION Newton's method iteratively approximates the solution to a set of nonlinear algebraic equations using successive linearization. It is ubiquitous in numerical methods for simulation, state and parameter estimation, and optimal control of ordinary, partial, differential-algebraic, and stochastic systems of differential equations. In particular, implicit numerical methods are suitable for stiff systems, and they require the solution of one or more sets of nonlinear equations in each time step. Each iteration of Newton's method requires the solution of a linear system of equations. For large-scale systems with many state variables, this is computationally expensive. Therefore, many variations of Newton-type methods, where the Newton step is approximated, have been proposed <cit.>. Two widely used variations are 1) the simplified Newton method where the Jacobian matrix is not reevaluated in each iteration (such that its factorization can be reused) and 2) inexact Newton methods <cit.> where the linear system in each Newton iteration is solved approximately using an iterative method <cit.>. Large-scale systems often arise as the result of spatial discretization of partial differential equations (PDEs) <cit.> or as a network of interacting subsystems <cit.>. In the former case, it is common that the resulting set of ordinary differential equations (ODEs) is stiff due to fast local changes (compared to the simulation horizon). The stiffness of network systems depends on the dynamics of the individual subsystems and their interactions. Model order reduction, or simply model reduction, methods <cit.> are relevant to any type of large-scale system. They identify a lower order dynamical system whose state variables can be used to approximate the state of the original dynamical system. For linear systems, there exist several model reduction methods, and the theory is well-developed. However, the reduction of general nonlinear systems is an active field of research and remains challenging. In this work, we propose a Newton-like method where a reduced order linear system is solved in each iteration. We use a projection matrix to compute the system matrices in this reduced system and to compute the approximate Newton step from the solution to the reduced system. We use proper orthogonal decomposition (POD) <cit.> to compute the projection matrix. The method is relevant to implicit numerical methods, and we demonstrate its utility using Euler's implicit method. A similar method was proposed by Nigro et al. <cit.>. However, we compute a projection matrix in each time step whereas they only compute it a few times during the simulation. Instead, they update the projection matrix adaptively during both the time steps and Newton iterations. Finally, if convergence is too slow, we solve the original (full order) linear system for the Newton step. The advantage of the method proposed in this work is its simplicity. We test the Newton-like method on numerical simulation of CO_2 injection into an oil reservoir. This process is modeled using four coupled PDEs which are discretized using a finite volume method. The remainder of this paper is structured as follows. In Section <ref>, we discuss numerical simulation of nonlinear systems using Euler's implicit method and Newton's method, and we discuss model reduction and POD in Section <ref>. In Section <ref>, we present the Newton-like method proposed in this paper, and we discuss details of the implementation in Section <ref>. Finally, we present the numerical examples in Section <ref>, and we present conclusions in Section <ref>. § NUMERICAL SIMULATION We consider nonlinear systems in the form ẋ(t) = f(x(t), u(t), d(t), p), where x are the states, u are manipulated inputs, d are disturbance variables, p are parameters, and f is the right-hand side function. We include the dependency of f on u, d, and p for completeness, and we assume a zero-order hold parametrization of u and d: u(t) = u_k, t ∈ [t_k, t_k+1[, d(t) = d_k, t ∈ [t_k, t_k+1[. §.§ Discretization We discretize (<ref>) using Euler's implicit method: x_k+1 - x_k = f(x_k+1, u_k, d_k, p) Δ t_k. Here, x_k ≈ x(t_k) and Δ t_k is the k'th time step size. As the right-hand side function, f, is nonlinear in the states, (<ref>) is a set of nonlinear algebraic equations which must be solved for x_k+1 given x_k, u_k, d_k, p, and Δ t_k. §.§ Newton's method We write the nonlinear algebraic equations (<ref>) in residual form: R_k(x_k+1) = R_k(x_k+1; x_k, u_k, d_k, p) = x_k+1 - x_k - f(x_k+1, u_k, d_k, p) Δ t_k = 0. In Newton's method, an approximation of the solution to (<ref>) is iteratively improved using the update formula x_k+1^(ℓ+1) = x_k+1^(ℓ) + Δ x_k+1^(ℓ), where x_k+1^(ℓ) is the ℓ'th approximation. If R_k(x_k+1^(ℓ)) < τ for some tolerance, τ, the iterations are terminated. The Newton step Δ x_k+1^(ℓ) is the solution to the linear system of equations obtained by linearizing (<ref>) around x_k+1^(ℓ): R_k(x_k+1^(ℓ)) + R_kx_k+1(x_k+1^(ℓ)) Δ x_k+1^(ℓ) = 0. The Jacobian matrix is R_kx_k+1(x_k+1) = I - fx(x_k+1, u_k, d_k, p) Δ t_k, where I is the identity matrix and fx is the Jacobian of f. We rearrange terms in order to write the linear system in the form A_k^(ℓ)Δ x_k+1^(ℓ) = b_k^(ℓ), where the system matrix and the right-hand side are A_k^(ℓ) = R_kx_k+1(x_k+1^(ℓ)), b_k^(ℓ) = -R_k(x_k+1^(ℓ)). § MODEL ORDER REDUCTION In this section, we discuss the challenges of reducing general nonlinear models, and we present the POD method used in the Newton-like method. The purpose of model reduction is to approximate the states, x, by a smaller number of reduced states, x̂, e.g., using a linear relation: x(t) ≈ V x̂(t). Here, x∈ℝ^n_x, V∈ℝ^n_x × n_r, and x̂∈ℝ^n_r where the number of reduced states, n_r is significantly smaller than the number of states, n_x. Direct substitution of (<ref>) into the original system (<ref>) gives V ẋ̂̇(t) = f(V x̂(t), u(t), d(t), p). Next, the differential equations are multiplied by the transpose of W ∈ℝ^n_x× n_r from the left: W^T V ẋ̂̇(t) = W^T f(V x̂(t), u(t), d(t), p). Assuming that W^T V is invertible, the reduced system is ẋ̂̇(t) = f̂(x̂(t), u(t), d(t), p), where the reduced right-hand side function, f̂, is f̂(x̂(t), u(t), d(t), p) = (W^T V)^-1 W^T f(V x̂(t), u(t), d(t), p). For general nonlinear systems, evaluating the reduced right-hand side function in (<ref>) is more expensive than evaluating the right-hand side function in the original system (for linear systems, the reduced system matrices can be computed prior to simulation or analysis). Consequently, using explicit numerical methods for the reduced system (<ref>) is, in most cases, not faster than for the original system. However, for implicit methods, the most computationally expensive step is to solve the linear system of equations in each iteration of the Newton step. Therefore, it is possible to achieve significant speedup by simulating the reduced system. However, it still remains challenging to accurately approximate general nonlinear systems using a reduced system. The matrix W^T V is often the identity matrix or a diagonal matrix (e.g., for clustering approaches). §.§ Proper orthogonal decomposition In the Newton-like method presented in Section <ref>, we use POD <cit.> to compute the projection matrices V and W based on a matrix of snapshots, X. That is, each column of X contains the state vector x_k = x(t_k) for some value of k. The singular value decomposition of X is X = U Σ R, where U and R are matrices of left and right singular vectors, respectively, and Σ is a diagonal matrix with the singular values on the diagonal in descending order. The projection matrices V = W consist of the columns of U corresponding to singular values larger than ϵΣ_11. Based on numerical experiments, we choose ϵ = 50 ϵ_m where ϵ_m = 2^-52 is the machine precision § NUMERICAL SIMULATION USING THE NEWTON-LIKE METHOD In the Newton-like method, we replace the Newton update (<ref>) by x_k+1^(ℓ+1) = x_k+1^(ℓ) + V_k Δx̂_k+1^(ℓ), where V_k is the projection matrix computed using POD, as described in Section <ref>, and Δx̂_k+1^(ℓ) is the reduced Newton step obtained by solving Â_k^(ℓ)Δx̂_k+1^(ℓ) = b̂_k^(ℓ). We derive the system matrix and the right-hand side following the same steps as in Section <ref>. First, we approximate the Newton step by Δ x_k+1^(ℓ) ≈ V_k Δx̂_k+1^(ℓ). Next, as the linear system would otherwise be overdetermined, we multiply the linear system of equations by W_k^T from the left. The resulting set of linear equations is W_k^T A_k^(ℓ) V_k Δx̂_k+1^(ℓ) = W_k^T b_k^(ℓ). Consequently, the reduced system matrix and right-hand side in (<ref>) are Â_k^(ℓ) = W_k^T A_k^(ℓ) V_k, b̂_k^(ℓ) = W_k^T b_k^(ℓ). The dimension of Â_k^(ℓ) is n_r× n_r which is significantly smaller than the dimension of A_k^(ℓ) which is n_x× n_x. Therefore, the solution of the linear system (<ref>) is much less computationally expensive. §.§ Algorithms We implement Euler's implicit method as shown in Algorithm <ref>. We use Newton's method in the first N_b time steps. For each subsequent time step, we create a snapshot matrix, X_k, consisting of up to N_h of the previous time steps. Based on the snapshot matrix, we compute the projection matrices V_k and W_k as described in Section <ref>. Finally, we use the projection matrices in the Newton-like method to solve the residual equations (<ref>) for the next time step, x_k+1. Based on numerical experiments, we choose N_b = ln(n_x) and N_h = √(n_x). We implement the Newton-like method as shown in Algorithm <ref>. The initial guess is x_k and the result is x_k+1. In each iteration, the reduced system (<ref>) is solved unless the previous reduced Newton step was too small. In line <ref>, we ensure that the full order system is not solved twice in a row. The iterations continue until the norm of the residual equations is below the tolerance τ. For convenience, we also use τ in line <ref> to determine when to solve the full order system. § IMPLEMENTATION DETAILS We implement Algorithm <ref> and <ref> in Matlab <cit.>. The Jacobian matrices of f and R_k are represented as sparse. All other quantities are represented as dense. We use Matlab's to compute the left singular vectors corresponding to the N_h largest singular values, and we use Matlab's (i.e., the backslash operator) to solve the sparse linear systems. Furthermore, we implement the evaluation of the right-hand side function and its Jacobian using the C++ library DUNE <cit.> and the thermodynamic library, ThermoLib <cit.>, and we use a Matlab MEX interface <cit.> to call the implementation from Matlab. The MEX interface allocates memory and evaluates the Jacobian of f as sparse before returning it to Matlab, i.e., we do not use Matlab's to convert it from a dense matrix. Finally, we carry out the computations using Windows Subsystem for Linux 2 (WSL2) on a Windows 10 laptop with 12 MB shared level 3 cache, 1,280 KB level 2 cache for each core, and 48 KB and 32 KB level 1 instruction and data cache for each core, respectively. Furthermore, each core contains two 11'th generation Intel Core i7 3 GHz processors. § NUMERICAL EXAMPLE In this section, we present a numerical example of CO_2 injection into an oil reservoir. We provide a brief description of the model in Appendix <ref>, and we use the porosity and permeability fields of the top layer of model 2 from <cit.>. They are shown in Fig. <ref> and <ref>. Clearly, they are very heterogeneous. Initially, the reservoir contains water and oil consisting of methane, n-decane and CO_2, and the water and oil phases do not mix. Furthermore, we assume that there is no gas phase. We inject liquid CO_2, which also contains 1‰ methane and 1‰ n-decane (mole fractions). The CO_2 is injected through two wells located at opposite corners. Each well injects 0.11 m^3 per day. Fig. <ref> shows an example of a simulation of 1,200 grid cells. The reservoir consists of the left-most 20 columns of cells in the fields shown in Fig. <ref> and <ref>. The top row of figures show the oil saturation (ratio between oil and water volume) over time and the bottom row shows the pressure. As CO_2 is injected, it displaces water. Therefore, the oil saturation decreases in the middle area. In this simulation, the oil and water cannot leave the reservoir. Consequently, the pressure increases significantly over. Furthermore, the pressure is almost completely uniform in this simulation. Fig. <ref> shows the computation time of simulating 8 days of CO_2 injection using Euler's implicit method and 1) Newton's method and 2) the Newton-like method described in Section <ref>. We use 40 time steps of 0.1 day and 20 time steps of 0.2 day, and we repeat the simulation for different numbers of grid cells. The number of grid cells changes the simulation scenario because we include a larger part of the reservoir, i.e., we do not refine the discretization. The computation time increases linearly with the number of grid cells, and the increase is lower for the Newton-type method. Fig. <ref> shows the speedup which is between 39% (for the lowest number of grid cells) and 84%. The mean speedup is 66% and the standard deviation is 12%. § CONCLUSIONS In this paper, we propose a Newton-like method based on model reduction methods (specifically, POD). The method can improve the computational efficiency of implicit numerical methods for approximating the solution to general nonlinear ODEs. Such systems are difficult to reduce directly using existing model reduction methods. The method approximates the linear system solved in each Newton iteration by a lower order system which is significantly less computationally expensive to solve. The system matrix and right-hand side of the reduced linear system are computed based on a projection matrix, which is updated in each time step (except for the first few time steps). We demonstrate the utility of the method for numerical simulation of CO_2 injection into an oil reservoir. We model this process by four coupled PDEs, and we discretize them using a finite volume method. Compared to Newton's method, the Newton-like method achieves a speedup of between 39% and 84% for 1,200 to 13,200 grid cells, which correspond to 4,800 and 52,800 state variables. § COMPOSITIONAL FLOW IN POROUS MEDIA In this appendix, we modify an isothermal and compositional model presented in previous work <cit.>. Specifically, we assume that there is no gas phase and that the rock and the water are incompressible. The purpose of these simplifications is to obtain a model which (after spatial discretization) consists of a set of ODEs. The concentrations of water, C_w, and the k'th component in the oil phase, C_k, are described by the PDEs ∂_t C_w = -∇·𝐍^w, ∂_t C_k = -∇·𝐍_k + Q_k, where 𝐍^w and 𝐍_k are the molar fluxes of the water component and of the k'th component, and Q_k is a source term representing the injection of component k. The molar flux of the k'th component is the product of the mole fraction, x_k, and the molar flux of the entire oil phase: 𝐍_k = x_k 𝐍^o. The molar fluxes of both the oil and the water phase are the products of the molar density, ρ^α, and the volumetric flux, i.e., 𝐍^α = ρ^α𝐮^α, α ∈{w, o}, and the volumetric flux is given by a generalization of Darcy's law to multiphase fluids: 𝐮^α = -k_r^α/μ^α𝐊 (∇ P - ρ̅^α g ∇ z), α ∈{w, o}. We use Corey's model <cit.> of the relative permeabilities, k_r^α, and we use the model by Lohrenz et al. <cit.> to describe the viscosity of the oil phase, μ^o. The viscosity of the water phase is μ^w = 0.3 cP. In this work, we assume the permeability tensor, 𝐊, to be a multiple of the identity matrix, i.e., the permeability is the same in all directions. Furthermore, P is pressure, ρ̅^α is the mass density, g is the gravity acceleration constant, and z is depth. As for the flux of component k, the source term is the product of the mole fraction and the source term for the entire oil phase: Q_k = x_k Q^o. We specify both x_k and Q^o. §.§ Volume balance and cubic equation of state When both oil and gas is present, the phase equilibrium problem for each grid cell is an equality-constrained minimization problem. The objective function is the combined Helmholtz' energy of the rock, water, oil and gas phase and the constraints specify the amounts of moles of each component (in both phases) and the combined volume of all four phases. However, when there is no gas phase, the phase equilibrium problem simplifies to finding a pressure which satisfies the volume balance: V^w + V^o + V^r = V. The volumes of the rock, V^r, the water, V^w, and the entire grid cell, V, are independent of the pressure. Consequently, given the amount of moles in the water phase, n_w, the oil volume, V^o, can be isolated in (<ref>). Given the oil volume V^o = V - V^w - V^r, we can compute the corresponding molar volume v^o = V^o/N^o, where N^o is the total amount of moles in the oil phase, and evaluate the pressure using the cubic equation of state P = RT/v^o - b_m - a_m/(v^o + ϵ b_m)(v^o + σ b_m). Here, R is the gas constant, T is temperature (60^∘C in this work), a_m = a_m(T, n^o) and b_m(n^o) are mixing parameters (n^o is a vector of moles of each component in the oil phase), and ϵ and σ are parameters which depend on the specific cubic equation of state used. We use the Peng-Robinson equation of state. After spatial discretization, the flow models described previously <cit.> consist of differential-algebraic equations (DAEs) in the form ẋ(t) = F(y(t), u(t), d(t), p), 0 = G(x(t), y(t), z(t), p), whereas the model described above is in the form ẋ(t) = F(y(t), u(t), d(t), p), y(t) = G̃(x(t), p). Here, y and z are algebraic variables, F is the right-hand side function, and G and G̃ are algebraic functions. As the algebraic equations (<ref>) are explicit in the algebraic variables, the above model is also in the form (<ref>), i.e., ẋ(t) = f(x(t), u(t), d(t), p) = F(G̃(x(t), p), u(t), d(t), p). §.§ Discretization The PDEs (<ref>) are in the form ∂_t C = -∇·𝐍 + Q, and we discretize them using a finite volume method. First, we integrate over each grid cell, Ω_i: ∂_t ∫_Ω_i C V = -∫_Ω_i∇·𝐍 V + ∫_Ω_i Q V. The left-hand side can be evaluated exactly, i.e., ∫_Ω_i C V = n_i, where n_i is the total amount of moles in the i'th grid cell. Next, we use Gauss' divergence theorem, write the surface integral as the sum of the integrals over each face of the grid cell, γ_ij, and approximate the integrals: ∫_Ω_i∇·𝐍 V = ∫_∂Ω_i𝐍·𝐧 A = ∑_j∈𝒩^(i)∫_γ_ij𝐍·𝐧 A ≈∑_j∈𝒩^(i) (A 𝐍·𝐧)_ij. The set 𝒩^(i) contains the indices of the grid cells that are adjacent to the i'th grid cell, A_ij is the area of face γ_ij (the face shared by Ω_i and Ω_j), and 𝐧 is the outward normal vector on the boundary of the grid cell, ∂Ω_i. Similarly, we approximate the integral of the source term as ∫_Ω_i Q V ≈ (Q V)_i. The resulting approximation in (<ref>) contains the flux evaluated at the center of each face of the grid cell. We use a two-point flux approximation <cit.> to approximate this flux. The result is (A 𝐍^w ·𝐧)_ij ≈ -(ΓĤ^w ΔΦ^w)_ij, (A 𝐍_k ·𝐧)_ij ≈ -(x_k ΓĤ^o ΔΦ^o)_ij, where Γ_ij is the geometric part of the transmissibilities, Ĥ^α is the fluid part of the transmissibilities, and ΔΦ^α is difference in potential. The geometric part of the transmissibilities is given by Γ_ij = A_ij(Γ̂_ij^-1 + Γ̂_ji^-1)^-1, Γ̂_ij = (𝐊_i c_ij - c_i/|c_ij - c_i|^2) ·𝐧, where c_ij is the center of γ_ij and c_i is the center of Ω_i. Furthermore, the difference in potential is ΔΦ_ij^α = (Δ P - ρ̅^α g Δ z)_ij, where we approximate the mass density at the face center using an average, i.e., ρ̅_ij^α = (ρ̅_i^α + ρ̅_j^α)/2, and the differences in pressure and depth are Δ P_ij = P_j - P_i, Δ z_ij = z_j - z_i. Finally, the fluid part of the transmissibilities is upwinded based on the difference in potential: Ĥ_ij^α = (ρ^α k_r^α/μ^α)_i if ΔΦ_ij^α < 0, (ρ^α k_r^α/μ^α)_j if ΔΦ_ij^α≥ 0.

http://arxiv.org/abs/2303.07135v1

20230313140357

A diffuse interface approach for vector-valued PDEs on surfaces

math.NA

label1]Michael Nestler [label1]organization=Institute of Scientific Computing, TU Dresden, city=Dresden, postcode=01062, country=Germany label1,label2,label3]Axel Voigt [label2]organization=Center for Systems Biology Dresden (CSBD), addressline=Pfotenhauerstraße 108, city=Dresden, postcode=01307, country=Germany [label3]organization=Cluster of Excellence Physics of Life (PoL), city=Dresden, postcode=01062, country=Germany Approximating PDEs on surfaces by the diffuse interface approach allows us to use standard numerical tools to solve these problems. This makes it an attractive numerical approach. We extend this approach to vector-valued surface PDEs and explore their convergence properties. In contrast to the well-studied case of scalar-valued surface PDEs, the optimal order of convergence can only be achieved if certain relations between mesh size and interface width are fulfilled. This difference results from the increased coupling between the surface geometry and the PDE for vector-valued quantities defined on it. § INTRODUCTION PDEs on surfaces remain an active field of research in applied mathematics and computational science. Due to their coupling with the geometry of such surfaces, PDEs are intrinsically nonlinear. This leads to new challenges in modeling and numerical analysis. Most of these challenges are addressed for scalar-valued surface PDEs, see <cit.> for a review. In the scalar case the coupling between surface geometry and the PDE is relatively weak, and thus numerical approaches established in flat spaces are applicable after small modifications. For vector-valued surface PDEs these approaches are no longer sufficient. Surface vector-fields often need to meet additional constraints. One example is the tangentiality of these fields. In this case, they need to be considered elements of the tangent bundle of the surface. This will lead to a strong nonlinear coupling between the surface geometry and the PDE. One break through which allows us to deal with these new challenges was the idea to express the solution and the surface differential operators in the global coordinate system of the embedding space and to penalize normal components, independently introduced in <cit.> and generalized in <cit.>. This approach essentially allows us to apply established tools for scalar-valued surface PDEs to each component. Popular approaches are surface finite elements (SFEM) and trace finite elements (TraceFEM), which have been applied to various problems in liquid crystal theory <cit.>, fluid mechanics <cit.> and biological physics <cit.>. Numerical analysis results for these methods also exist, but are restricted to the most simple equations of this type, the surface vector-valued Laplace equation <cit.>, the surface vector-valued Helmholtz equation <cit.> and the surface Stokes equations <cit.>. All these results show the necessity for an appropriate approximation of the geometric quantities that enter these equations. Optimal order of convergence can often only be achieved if this approximation is of a higher order than the solution. These results reflect the increased coupling between the surface geometry and the solution for vector-valued surface PDEs if compared with scalar-valued surface PDEs. This increase in complexity is also shown numerically by comparing scalar-, vector- and tensor-valued surface diffusion equations by various numerical methods <cit.>. Another popular method to solve surface PDEs, the diffuse interface method <cit.> is less explored for vector-valued surface PDEs. Two examples, where solutions are compared with other more established methods are <cit.>. The diffuse interface method approximates the surface PDE by a bulk PDE that can be solved with standard numerical tools. This makes the approach attractive to be used in various application areas for more complex problems, especially those where the surface evolves according to physical laws which depend on the vector-valued surface quantity. Such applications can be found in biology, e.g. in morphogenesis, where the method recently led to spectacular results <cit.>. However, knowing the subtleties associated with the approximation of geometric terms in SFEM and TraceFEM, discussed above, and the additional error emerging from the approximation of the diffuse interface method, such results should be considered with care. We here consider a tangential vector-valued surface Helmholtz equations to explore the convergence properties of the diffuse interface methods. This can be considered as a model problem, and the obtained results can be assumed to be applicable also to other more complex surface PDEs. In Section <ref> we introduce the surface model, review the described extension to the embedding space and required penalization of the normal component, which provides the basis for a SFEM discretization and derive from this the diffuse interface formulation. We formulate these approaches in variational form and provide a finite element (FEM) discretization. The diffuse interface approximation is justified by formal matched asymptotics, which directly follow from the results for scalar-valued surface PDEs <cit.>. In Section <ref> we construct an analytical solution and perform various convergence studies. The results indicate the necessity for accurate approximations of the surface normals. These approximations can be obtained from the phase field variable or the signed distance function of the implicit surface description. Depending on the chosen approach specific minimal requirements regarding the relation of interface width and mesh size at the interface have to be meet for optimal convergence. Ignoring these requirements can result in numerical solution procedures that do not converge! These results significantly differ from the known results for the diffuse interface method for scalar-valued surface PDEs. In Section <ref> we draw conclusions. § MODEL PROBLEM We demonstrate the applicability of the diffuse interface approach by considering the surface vector-valued Helmholtz equation on a smooth, oriented two-dimensional surface ⊂^3 without boundaries. The equation reads - + = , for a tangential vector field defined in the tangent bundle of and a compatible right hand side . The considered differential operators and correspond to the metric-preserving covariant derivative ∇_. The diffuse interface formulation will be based on the SFEM formulation for vector-valued surface PDE's, as detailed in <cit.>. The general idea is to express the solution and the operators in the global coordinate system of the embedding space ^3 and to penalize normal components. We denote the Euclidean coordinates of ^3, by uppercase letters I,J,K, … and consider Einstein's sum convention. We consider the outward oriented normal , the surface identity = 𝕀- and the shape operator = -∇_. For we denote the formal extension to the embedding space by , for which we require [ - (·) ]|_=. Finally we define the full tangential projection of such extended fields along all their components by [.], for a vector-valued field [𝐏]_I = _IJP_J and for a 2-tensor-valued field []_IJ = _IKQ_KL_LJ. Given these notations we can express the surface differential operators by their ^3 counterparts. Acting on the extended field this reads = [∇]|_ + (|_·) []_IJ = _IL∂_K _L _KJ + _IJ_L_L, see <cit.> for details. To obtain the embedded variational formulation we use scalar products of pointwise full contraction of the tangential parts of the extended solution and the test function , e.g. = = ∫_ ([],[]) d = ∫__I _IJ_J d. From these ingredients we set up a component wise solution-test space [H^1()]^3 and write the embedded variational problem [∇] + (·)[∇] + (·) + + ·· = ∀∈ [H^1()]^3 with an additional penalty term with prefactor to approximate · = 0. In this formulation each component can be considered by classical SFEM for scalar-valued fields, see <cit.>. Various numerical studies confirm the applicability of this approach. Numerical analysis provides various estimates of solution error convergence for surface mesh size _→ 0. Here, we consider the surface to be approximated by faceted polyhedra . Such first order geometry approximation allows only linear Lagrangian elements [_]_I ∈() ≈ H^1() for the discretization of the component solution-test space. This approach is referred to as the linear isogeometric ansatz. With the penalty prefactor = (1/_^2) one obtains quadratic convergence of the L^2-error ( -[_];_) = ( -[_] - [_])^1/2. Please note, by approximating by a faceted polyhedra we yield _ = 0 which significantly simplifies the discretized weak formulation of the Laplacian in eq. (<ref>). Various numerical studies, also for more complicated problems <cit.>, confirm quadratic order. At first glance it seems straight forward to apply the tools introduced in <cit.> to each component in eq. (<ref>) to obtain a diffuse interface approximation which turns the problem defined on into a problem defined in ^3. However, eq. (<ref>) contains various geometric terms, which are not considered in <cit.> or any subsequent analysis. The sensitivity of the solution to the approximation of these geometric terms is known from <cit.> and contrary to the scalar case <cit.>, higher order approximations for the surface and the solution only lead to better convergence properties if a higher order approximation of the normals is used in the introduced penalization term. So, the question arises if an additional approximation of the normals, resulting from an implicit description of the surface in the diffuse interface formulation, is sufficient to obtain the same convergence properties as the SFEM approach. We will answer this question in the following. In order to formulate the diffuse interface approximation we extend the quantities , , and to the embedding space. We consider this component wise constant in the normal direction and denote the extended quantities by , , and , respectively. With this we can write _IJ=- _IK∂_K _J, and _IJ|_ = _IJ. We now follow <cit.> and embed the surface in a bulk domain ⊂ and use an implicit description of . For this purpose we use a signed distance function () with |∇ | = 1 and = {∈, ()=0} and a phase field function () = 1/2 (1 -tanh(3()/)) with interface width > 0 and = {∈, ()=1/2}. With these functions we can evaluate extended normals. We obtain = ∇, = -∇ / | ∇ The surface delta function is approximated by ≈() = 36/ (1-)^2 ^2, the classical double-well potential in Ginzburg-Landau energies <cit.>. Within this notation we substantiate the previously used constant normal extension by the closest point projection : →, () = -, () = ∘(), which is well defined for sufficient small such that () < 1, ∀∈. With this notation we express the surface scalar product by a scalar product of the embedding space = ≈∫_() _I _IJ_J d = . We use this embedded scalar product to set up the embedded variational problem with a component wise solution space [H^1()]^3. So the diffuse interface approximation of the component wise embedded variational problem eq. (<ref>) reads [[∇] + (·)][∇] + (·) + + [·][·] = ∀∈ [H^1()]^3. Considering such component wise formulation as a coupled system of scalar fields, the matched asymptotic analysis of <cit.> provides formal convergence of eq. (<ref>) to eq. (<ref>) for → 0. This establishes eq. (<ref>) to be a diffuse interface approximation of eq. (<ref>). Using this procedure any surface PDE can be reformulated into the corresponding diffuse interface approximation. In analogy to the SFEM approach, we use a linear ansatz for the implicit geometric description by _, _∈(_) for a tetrahedral mesh _ of , where denotes the bulk mesh size in the interface ∈ [0.05, 0.95]. Within such ansatz the recovered surface = {_≡ 0} or = {_≡ 1/2} is a faceted polyhedra. The same ansatz is considered for the solution _∈ [(_)]^3. This allows us to consider _ = 0 across the interface and we obtain H(_,) + P(_,) + R(_,) = F() ∀ ∈ [(_h)]^3 with the Helmholtz operator H(_,), the normal penalty term P(_,), a bulk stabilization R(_,) and the right hand side F(), defined by: H(_,) = (_)[_]_IJ ∂_K _,I∂_K _J + (_)[_]_IJ _,I_J P(_,) = (_) [_]_I _,I[_]_J_J R(_,) = ∂_K _,I∂_K _I F() = (_)[_]_IJ _I_J, respectively. Here we have used a component wise constant regularization outside the interface of =10^-6 as discussed for the scalar case in <cit.>. Inspired by the normal penalty condition, specified in <cit.> for the isogeometic case, we consider also within the diffuse interface approximation = 10/h^2. § RESULTS We are interested in the convergence properties of the diffuse interface approach. We have to take into consideration that error measures will depend on and . For further discussion we define a relation (, ) to essentially express the number of mesh points across the interface for a given interface width. For our considerations we follow <cit.> and group those relations as linear, where = () so the number of points remains constant within the interface for → 0, and higher order relations, where ^2 = (^3),(^4),(^5) such that the number of points within the interface increases for → 0, see Figure-<ref>-[B]. Within the same setting, a linear ansatz for the geometric description and a linear ansatz for the solution, in <cit.> a numerical study was performed to estimate convergence rates. For a linear relation (, ) quadratic convergence was obtained for the scalar-valued problem, which can be considered optimal. It thus provides an upper limit to our vector-valued problem. However, it remains open if this limit can be reached in typical applications, where the signed distance function _h or the phase field variable _h need to be constructed from a surface mesh or where the phase field variable _h might be a solution to another PDE. In these cases the computation of the normals _h by eq. (<ref>)[A] or [B] adds an additional source of error. §.§ Benchmark formulation To assess the impact of the major sources of approximations in the diffuse interface approach we perform a series of numerical experiments on a torus with radi R=1 and r=0.5. The torus is embedded in = [-2,2]^3. The analytical signed distance function for this geometry is given in ℝ^3, = [x,y,z] by () = (( √(x^2 + y^2) - R )^2 + z^2 )^1/2 - r The analytic phase field function is defined as before and for the analytic normal to we consider = ∇. On we consider the solenoidal tangential vector field = ×∇_(x^2 y - 5 z^3) and its extension to by = ∘, see Figure-<ref>-[C]. This vector field will be used as an analytical solution for estimating the rate of error convergence. A compatible right hand side in eq. (<ref>) is constructed by F() = () [ ] _IJ[∇]_IJ + ()_I[]_I. To obtain results, which are comparable to the SFEM results in <cit.> we estimate - [_]≤ - _ + _· and approximate surface error measured by the following diffuse interface L^2 error measures. ( - _h; , ) = (() ( - _h)·( - _h))^1/2 (_h ·; , ) = (() [_h·][_h ·])^1/2. For the subsequent numerical experiments we consider a set of mesh sizes ∈ [(1/2)^7, (1/2)^2], corresponding to bisections of . For each and (,) relation we evaluate an associated and construct an adaptive tetrahedral mesh with meshsize in the diffuse interface of width . Off the interface, the mesh is only refined to preserve conformity. See Figure <ref>-[A]. We use the toolbox <cit.> to generate these meshes and evaluate _. uses a triangulated surface as input which we generate via <cit.> with vertices on and meshsize _ = /4. Different methods exist to compute signed distance functions, see e.g. <cit.>. In a second order method is implemented. We use the finite element toolbox <cit.> to solve eq. (<ref>), which uses an unpreconditioned BiCGstabL solver provided by <cit.> for the linear system. §.§ Numeric evaluation of diffuse interface variables and normals First we examine the convergence of errors in the chosen method to compute diffuse interface variables _, _ and _ derived from these fields. We compute _ and _ with for all and (,) relations. We evaluate the diffuse interface L^2-errors the analytical results ( - _; , ), and ( - _; , ) and plot these quantities versus , see Figure-<ref>-[A]. We observe the errors in evaluating _h converge quadratic, → 0 for linear (,) relation. Higher order (,) relations increase the rate of convergence. For _h we observe the rates of convergence to be reduced about one order compared to rates for _h. Recalling that |∇ | = 1 while |∇ | ≈ 1/ near the interface such reduced order of convergence is expected. Using the evaluated diffuse interface fields to determine the normals along the approaches described in eq. (<ref>) we evaluate the diffuse interface L^2-errors, ( - ∇_h; , ) and ( + ∇_h / | ∇_h |; , ), see Figure-<ref>-[B]. Using the linear (,) relation, see Figure-<ref>-[B](blue lines), we yield quadratic rate of convergence in ( -∇_; , ) while (^a + ∇_ / | ∇_ |; , ) does not converge for → 0. Considering higher order (,) relations the convergence rates, , for ( - ∇_; , ) could be improved beyond quadratic rate. In the case of ( +∇_ / | ∇_ |; , ) a linear convergence rate could be achieved for ^2 = (^3) and a quadratic rate is achieved with ^2 = (^5) relation. We conclude that, in order to obtain normals with converging errors for → 0, provides sufficient good approximations for _h to evaluate _h = ∇_h even for a linear (,) relation. Yet considering the approach of computing normals from _h, we have to keep in mind that the rate of convergence, → 0, is effectively reduced by two orders. Therefore, errors in normals only converge with higher order (,) relations. §.§ Error convergence for solution with analytic diffuse interface variables Next we consider the analytical forms for , and and solve eq. (<ref>) for a set of and (,) relations. We evaluate the diffuse interface L^2-error ( - _h; , ), see Figure-<ref>-[C]. We observe that for a linear (,) relation a quadratic rate of error convergence → 0. This confirms the result from <cit.> for scalar-valued PDEs. Again, using higher order (,) relations improves the rate of error convergence → 0. But reviewing the error convergence it becomes apparent that is the crucial factor in convergence of ( - _h; , ) and using higher order (,) relations does not improve the rate of convergence for → 0. So, we conclude that an analytical diffuse interface approximation does not impair the convergence of errors → 0, and we observe a quadratic rate of convergence for linear (,) relation. §.§ Error convergence for diffuse interface approximation We now combine both, consider approximations of the diffuse interface variables _, _ and _ obtained with and solve eq. (<ref>) for a set of and (,) relations. For these numeric solutions _h we evaluate the diffuse interface L^2-error ( - _h; , ) and the normal component (_h ·; , ), see Figure-<ref>. For the normal approximation by _h = ∇_h, see Figure <ref>-[A] and a linear (,) relation we observe quadratic rate of convergence for → 0 in the normal component and for the solution error. Due to the linear nature of the (,) relation this behavior translates into quadratic rate of convergence with absolute values very close to the results with analytic diffuse interface variables and normals in Figure <ref>-[C]. For higher order (,) relations the errors and convergence rates are also very close to the results with analytic diffuse interface variables and normals. Quite contrary, we observe while using the approximation of _h = - ∇_h / | ∇_h | the errors in normal approximation to be dominant. Reviewing Figure <ref>-[B](right) we observe that the reduced normal approximation quality of this approach yields reduced convergence rate for controlling the solutions normal component, see Figure <ref>-[B](left) directly impacting the solution error convergence rate in Figure <ref>-[B](middle). In the case of a linear (,) relation =() we observe no convergence in solution error, at ^2=(^3) we obtain a linear rate of convergence and for ^2=(^4) and ^2=(^5) we yield close to quadratic rates of convergence. Comparing to the results for analytic diffuse interface variables and normals we observe only for ^2=(^5) similar errors and rates of convergence for (_h · ; , ) and ( - _h; , ) → 0 and → 0. Therefore, we conclude that the diffuse interface method, based on component-wise FEM, yields rates of error convergence compatible to the results of SFEM for vector-valued surface PDEs only if a high-quality signed-distance function is used to compute the normals. In this case a linear (,) relation is sufficient. Alternatively if the normals are approximated by _h = - ∇_h / | ∇_h | a (,) relation with ^2=(^5) is required to achieve quadratic rate of convergence. A linear (,) relation does not lead to convergence in this case! This significantly differs from the diffuse interface method for scalar-valued surface PDEs introduced in <cit.>. § DISCUSSION AND CONCLUSION We provided a systematic derivation of a diffuse interface method to solve vector-valued surface PDEs. This approach builds on established SFEM formulations <cit.> and allows us to consider each component as a scalar quantity. The resulting system of equations can be solved by standard FEM. The connection to the surface PDE is established by formal matched asymptotic following <cit.> and is applicable to any vector-valued surface PDE. Along the same lines the approach can also be extended to tensor-valued surface PDEs. Our numerical studies show that the same convergence properties as for SFEM can be achieved. With a linear isogeometric approach, which in our approach corresponds to linear Lagrange elements and a simplicial mesh, a quadratic order of convergence can be obtained for the considered vector-valued surface Helmholtz equation on a torus. However, to achieve this optimal order requires an accurate approximation of the surface normals. If they have to be numerically constructed a specific (,) relation is necessary. If the computation of the normals is based on the signed distance function _h a linear relation is sufficient. However, if the computation of the normals is based on the phase field variable _h a linear relation does not lead to convergence and a higher order relation is required. The numerical results suggest ^2=(^5) to obtain a quadratic rate of convergence. This result has severe consequences for applications. While for stationary surfaces, the computation of a signed distance function is not an issue and in such cases the diffuse interface method with normals computed from the signed distance function provides a valuable tool with optimal order of convergence, for moving surfaces this is no longer an option. The increased computational cost resulting from the computation of a signed distance function in each time step makes the approach inefficient. Assuming that the evolution of the surface is governed by a phase field problem which determines _h in each time step, the normals should be computed using _h. Such approaches are common for scalar-valued surface quantities, such as concentration <cit.> or particle density <cit.>. However, for vector-valued surface quantities using this approach does not lead to convergence if a linear (,) relation is considered. We can only expect convergence for higher order relations and quadratic convergence for ^2=(^5). This needs to be respected in applications. Accounting for this issue increases the computational cost. However, the advantages of using basic numerical tools like linear Lagrangian elements and generic iterative solvers and the possibility to deal with topological changes in the evolution of the surface remain. This together with the possibility to couple the surface PDE with equations in the embedding space <cit.>, e.g. in two-phase flow problems with fluidic surfaces <cit.>, makes this approach appealing for various applications. Acknowledgments: This work was supported by the German Research Foundation (DFG) through grant VO 899/22 within the Research Unit ”Vector- and Tensor-Valued Surface PDEs” (FOR 3013). We further acknowledge computing resources provided by ZIH at TU Dresden and by JSC at FZ Jülich, within projects WIR and PFAMDIS, respectively. elsarticle-num

http://arxiv.org/abs/2303.06941v1

20230313091613

Radio spectral properties at 150-5000MHz of star-forming galaxies in the ELAIS-N1 field

astro-ph.GA

firstpage–lastpage Context-Aware Selective Label Smoothing for Calibrating Sequence Recognition Model Shuangping Huang^1,2, Yu Luo^1, Zhenzhou Zhuang^1, Jin-Gang Yu^1,2*, Mengchao He^3, Yongpan Wang^4 =========================================================================================================== By combining high-sensitivity LOFAR 150 MHz, uGMRT 400 MHz, GMRT 610 MHz, and JVLA 5 GHz data in the ELAIS-N1 field, we study the radio spectral properties of radio-detected star-forming galaxies (SFGs) at observer-frame frequencies of 150–5000 MHz. We select ∼3,500 SFGs that have both LOFAR 150 MHz and GMRT 610 MHz detections by removing AGN from the two radio samples, and obtain a median spectral index of α_150^610=-0.51±0.01 with a scatter of σ=0.2. Due to the relatively lower sensitivity of uGMRT 400 MHz data, we apply a flux cut of S_610>300 μJy and obtain the median spectral indices of α_150^385=-0.42^+0.03_-0.02, α_385^610=-0.44^+0.03_-0.04, and α_150^610=-0.42^+0.02_-0.01 for the sample of 258 SFGs that have detections at these three radio frequencies. The JVLA 5 GHz observations only cover the central 0.1 deg^2, where ∼100 SFGs are selected, for which we obtain median α_610^5000=-1.14^+0.04_-0.05, α_385^5000=-1.08^+0.01_-0.02 and α_150^5000=-0.87±0.01, with some selection bias towards the steeper spectrum because of the high sensitivity of the 5GHz data. Overall, the results show that the radio spectrum is flatter if we include a lower frequency dataset when measuring the radio spectral index at 150–5000 MHz. We study the correlations between radio spectral index and physical properties of radio-selected SFGs and find that, on average, the radio spectrum slightly steepens with increasing stellar mass. However, we only find that the radio spectrum flattens with increasing optical depth at V-band at ν<∼1 GHz. We suggest that spectral ageing due to the energy loss of cosmic ray electrons and thermal free-free absorption could be among the most likely physical mechanisms that drive the two correlations respectively. Both of these could be the physical causes of why the radio spectrum is flatter at low frequency than at high frequency. radio continuum: galaxies – methods: observational – galaxies: formation –galaxy: evolution –galaxies: star formation –galaxies: statistics § INTRODUCTION The radio continuum emission from star-forming galaxies (SFGs) consists of the non-thermal synchrotron radiation and thermal free-free emission. In SFGs, massive stars (M>8) explode to form Type II and Type Ib supernovae and accelerate cosmic ray (CR) electrons, which gyrate within large-scale galactic magnetic fields and produce radio emission via a synchrotron process <cit.>. The underlying physical process of free-free emission is powered by the Coulomb scattering between free ions and electrons in the regions, which are ionised by young, massive stars. Therefore, both of these processes are associated with massive star formation in galaxies. These characteristics, as well as the radio emission's insensitivity to dust attenuation, make the radio continuum a powerful tool to study the astrophysical properties of SFGs <cit.>. Although the free-free emission traces a nearly instantaneous star formation <cit.>, it is very challenging to target it at gigahertz frequencies or below due to its flat spectrum (α∼-0.1, where S_ν∝ν^α). At ν>∼30 GHz, thermal free-free radiation is expected to dominate the radio continuum emission, but the observation condition at such high-frequency range is stringent <cit.>. Therefore, although it is a delayed tracer of star formation (∼100 Myr), the non-thermal synchrotron emission, which has a steep spectrum of α∼-0.8, is widely used to study activities of SFGs in both the local and distance Universe <cit.>. On galaxy-integrated scales, the radio continuum based star formation rate (SFR) is most often calibrated by an empirical correlation between far-infrared (FIR) and radio (typically at ∼1.4 GHz) luminosities (the FIRRC), which appears to hold across several orders of magnitude in spatial scale, luminosity, gas surface density and photon/magnetic field density and with a modest evolution with redshift <cit.>. In addition, the FIRRC is also widely used to classify radio sources as either SFGs or radio-loud active galactic nuclei (AGN) because it is characteristic of radio-loud AGN activities to accelerate the CR electrons and cause an `excess' of radio emission <cit.>. Therefore, studies of SFGs based on radio continuum emission rely on a well-determined radio spectrum, especially at low frequency since the rest-frame ∼1.4 GHz is shifted to ν<1.4 GHz at high redshift. However, due to the lack (until recently) of high-sensitivity and large coverage low-frequency radio surveys and the fact that the radio spectra of SFGs are dominated by synchrotron radiation at ν∼1-10 GHz, most of the studies based on rest-frame radio spectra of SFGs assumed a synchrotron spectrum or a synchrotron spectrum with 10% contribution from the free-free emission (α∼-0.7) when k-correcting the observed flux densities at ∼1.4 GHz <cit.>. These assumptions will overestimate the radio flux density at low frequency because of a flattening of radio spectrum at ν<∼1 GHz as suggested by some recent works <cit.>. At frequency ν<∼1 GHz, besides the two emission processes, the spectra of SFGs are expected to be affected by additional physical mechanisms. Notably, the regions become optically thick at ν<∼1 GHz, where free-free absorption causes greater attenuation of lower-frequency synchrotron radiation. This effect has been observed in the Galactic centre and some nearby galaxies <cit.>, and it has been widely used to explain a flattening of radio spectrum at low frequency for both local and high-redshift galaxies <cit.>. However, some other studies suggest that thermal gas absorption might not be the main cause of a flattening of radio spectrum at low frequency, since, for instance, there is no correlation between the inclination of the galaxies and the flatness of their radio spectrum at low frequency <cit.>. Alternatively, the intrinsic steepening of the synchrotron spectrum, due to synchrotron and inverse Compton losses growing stronger at high CR energy, is used to explain why the high-frequency radio spectrum is steeper than the low-frequency in the literature <cit.>. Besides, the Razin-Tsytorich effect <cit.>, and the effects from relativistic bremsstrahlung and synchrotron self absorption are also expected to be more important at low frequency <cit.>. In recent years, the high-sensitivity and wide field extragalactic radio continuum surveys conducted by the Low Frequency Array <cit.>, Giant Metrewave Radio Telescope <cit.>, and Murchison Widefield Array <cit.> are finally allowing us to study the radio spectral properties at ν_ obs<∼1 GHz with a statistically significant sample of SFGs. In this work, we combine high-sensitivity LOFAR 150 MHz data, GMRT 610 MHz data, 400 MHz data from the upgraded GMRT (uGMRT), and 5 GHz data observed by the Karl G. Jansky Very Large Array (VLA) in the ELAIS-N1 field <cit.> to study radio spectral properties at observer-frame frequencies of 150–5000 MHz. Details of these radio data as well as the ancillary data used in this work are described in Section <ref>. We also summarise the main properties of these radio datasets in Table <ref>. We analyse the data and select SFGs from the radio surveys in Section <ref>. The measured radio spectral indices at different frequency ranges and the correlations between radio spectrum and physical properties of SFGs are presented in Section <ref>. Using these results, we discuss the underlying physical mechanisms that drive the radio spectrum at rest-frame ∼150–10,000 MHz in Section <ref>. Our results are summarised in Section <ref>. Throughout this paper, we adopt the AB magnitude system <cit.> and assume a flat ΛCDM cosmological model with the Hubble constant H_0 = 67.27 km s^-1 Mpc^-1, matter density parameter Ω_ m = 0.32, and cosmological constant Ω_Λ = 0.68 <cit.>. § OBSERVATIONS AND DATA §.§ LOFAR 150 MHz data The LOFAR 150 MHz data of the ELAIS-N1 field used in this work are from the LOFAR Two Metre Sky Survey <cit.> Deep Fields <cit.>. The observations were taken with the High Band Antenna (HBA) array with frequencies between 115–177 MHz in the LOFAR observation cycles 0, 2, and 4. The central frequency of the observation is 146.2 MHz. The final Stokes I image for all survey fields have a resolution of 6. The ELAIS-N1 field is the deepest of the LoTSS Deep Fields to date with an effective observing time of 163.7 hours <cit.>. The median RMS of the central 6.7 deg^2, where the highest-quality multi-wavelength data exists <cit.>, is 22.9 μJy beam^-1. Within this region, <cit.> created a final cross-matched and associated catalogue that contains 31,610 LOFAR 150 MHz radio sources (Figure <ref>). We refer the reader to <cit.> for more details about the observations and data reduction of LOFAR 150 MHz data and to <cit.> for the creation of the final cross-matched and associated radio catalogue. We also present a short summary of the cross-matched results in Section <ref>. §.§ uGMRT 400 MHz data The ELAIS-N1 field was observed at 300–500 MHz with the uGMRT in GTAC cycle 32 <cit.>. The observations cover an area of ∼1.8 deg^2. With a total on-source time of ∼13 hours, the observations reach an RMS noise of ∼15 μJy beam^-1 near the centre of the field. The angular resolution of the image is 46×43. <cit.> detected 2,528 >6 σ radio sources from the primary beam-corrected uGMRT 400 MHz image (Figure <ref>). Details of observation, data reduction, and source detection are presented in <cit.>. When analysing the data, <cit.> used the 400 MHz (the central frequency) as the effective frequency of their observations. However, there are some satellite radio-frequency interference (RFI) with about 20 MHz width near 480 MHz, which means the data is not divided equally below and above 400 MHz. We therefore estimate an effective frequency of 385 MHz by considering this effect. We test the reliability of this estimation by measuring the radio spectral indices at observer-frame frequencies of 150–385 MHz, 385–610 MHz and 150–610 MHz for GMRT 610 MHz-detected sources with S_ 610 MHz>300 μJy. As we present in Section <ref>, using 385 MHz as the effective frequency results in the median spectral indices that are consistent with those in the literature. §.§ GMRT 610 MHz data There are two sets of GMRT 610 MHz data of the ELAIS-N1 field used in this work. One is from the ELAIS-N1 610 MHz Deep Survey <cit.>. The other is from a wide-area GMRT 610 MHz survey of the ELAIS-N1 field <cit.>. The GMRT deep data include a set of pointings from the GMRT wide survey <cit.>. §.§.§ GMRT deep The ELAIS-N1 610 MHz Deep Survey covers an area of 1.86 deg^2 in the centre of the field. The survey was carried out with seven closely spaced GMRT pointings with a on-source integration time of ∼18 hours for each pointing <cit.>. The seven pointings were mosaicked to create the final image. To reduce the noise around the edge, <cit.> also includes a set of pointings with on-source time of 2.5 hours per pointing from GMRT wide survey in the final mosaic image. The median RMS noise of the final mosaic image is 19.5 μJy beam^-1 with a minimum noise of 7.1 μJy beam^-1 in the central region. The angular resolution of the GMRT deep image is 6. <cit.> detected 4,290 >5 σ radio sources from the final mosaic image (Figure <ref>). We refer the reader to <cit.> for more details. §.§.§ GMRT wide The wide-area GMRT 610 MHz survey covers an area of 12.8 deg^2 in the ELAIS-N1 field by using a total of 51 GMRT pointings. The on-source time for each pointing is ∼2.5 hours and for the final mosaic data is 153 hours. The final mosaic image reaches an RMS noise of ∼40 μJy beam^-1 and an angular resolution of 6. A total of 6,400 radio sources were detected from the final mosaic image (Figure <ref>). Details of observations, data reduction, and source extraction are presented in <cit.>. §.§ JVLA 5 GHz data The JVLA 5 GHz observations only cover the central 0.13 deg^2 of the ELAIS-N1 field <cit.>. The bandwidth of the observation is 2 GHz and the central frequency is 5 GHz. With a total of 60 hours observation time, the JVLA 5 GHz data reaches an RMS noise of 1.05 μJy beam^-1. The angular resolution of the image is 25. In total, <cit.> detected 387 radio sources with S_ 5GHz>5 μJy from the image. Although <cit.> only focus on the polarized emission from these sources, we include their continuum emission in this work to measure the radio spectral index between the observer-frame frequencies of 150–5000 MHz. §.§ Additional Multi-wavelength Data/Catalogues The ELAIS-N1 field was originally chosen for deep extra-galactic observations with the Infrared Space Observatory (ISO) due to its low infrared background <cit.>. The deep infrared as well as the ultraviolet (UV) and optical observations make this field one of the best-studied astronomical fields. Since this work is based on the radio-selected SFGs, we do not use these multi-wavelength data directly but utilize the cross-matched results of GMRT 610 MHz radio sources from <cit.> and of LOFAR 150 MHz sources from <cit.>. We therefore give a short summary of the multi-wavelength data used in the cross-matches in their works and refer the reader to <cit.> and <cit.> for more details of these data. <cit.> determined the multi-wavelength counterparts of GMRT 610 MHz-detected radio sources by using the Spitzer Extragalactic Representative Volume Survey (SERVS) DR2 firstly <cit.>. The SERVS Data Fusion project included most of the publicly available far-UV to FIR data and homogenized these data to estimate the photometric and spectroscopic redshifts for IRAC 3.6 and/or 4.5 -selected sources <cit.>. Furthermore, if a SERVS match was not found, <cit.> used the J and K-band data from the UKIRT Infrared Deep Sky Survey (UKIDSS) Deep Extragalactic Survey (DXS) <cit.> to identify counterparts to GMRT 610 MHz radio sources. We will be back to their cross-matched results in Section <ref>. In addition to the deep 3.6 and 4.5 from the SERVS and near-infrared (NIR) J and K-band data from UKIDSS-DIS, the following datasets were also used by <cit.>: * UV data from the Deep Imaging Survey (DIS) taken with the Galaxy Evolution Explorer (GALEX) space telescope <cit.>, * u-band data from the Spitzer Adaptation of the Red-sequence Cluster Survey <cit.>, * optical g, r, i, z, and y-band data from the Medium Deep Survey (MDS), one of the Panoramic Survey Telescope and Rapid Response System (Pan-STARRS-1) surveys <cit.>, * the g, r, i, z, y-band, and the narrowband NB921 data from the Hyper-Suprime-Cam Subaru Strategic Program (HSC-SSP) survey <cit.>, and * the mid-infrared (MIR) 3.6 , 4.5 , 5.8 , and 8.0 data from the Spitzer Wide-area Infra-Red Extragalactic (SWIRE) survey <cit.>. In addition, <cit.> included the far-infrared (FIR) data at 24 from the Multi-band Imaging Photometer for Spitzer <cit.> and at 100 , 160 , 250 , 360 , and 520 from Herschel Multi-tiered Extragalactic survey <cit.> in their multi-wavelength photometric catalogue that were used to cross-match with the LOFAR 150 MHz radio sources. We will introduce their cross-matched results in Section <ref>. § ANALYSIS To investigate the radio spectral properties of radio-detected sources, we first check the consistency of the astrometry of sources detected at different radio frequencies. Then, we introduce the selection of SFGs from GMRT 610 MHz-detected sources in <cit.> and from LOFAR 150 MHz-detected sources in Best et al. submitted. §.§ Astrometry To test the relative astrometry of the radio sources detected at different frequencies, we first cross-match the radio sources detected at LOFAR 150 MHz, GMRT 610 MHz, uGMRT 400 MHz, and JVLA 5 GHz by using a matching radius of 10 . As shown in the top panels of Figure <ref>, the astrometry of GMRT 610 MHz data, both GMRT deep and wide, is well-consistent with that of LOFAR 150 MHz data with the median offsets of Δ R.A. ∼01 and the Δ Decl. ∼004. However, the median astrometric offsets between uGMRT 400 MHz and LOFAR 150 MHz data are ∼05 in both right ascension and declination. Such offsets result in relatively larger cross-matching radii between 400 MHz sources and the sources detected at LOFAR 150 MHz and GMRT 610 MHz as described in Section <ref>. The bottom-right panel of Figure <ref> shows that the astrometry of JVLA 5 GHz data is consistent with that of LOFAR 150 MHz data very well with the median offsets of Δ R.A. =-015 and the Δ Decl. =004. §.§ SFGs selection Except the JVLA 5 GHz data, which have a very limited sky coverage, the LOFAR 150 MHz and GMRT 610 MHz (deep) data are the two deepest datasets that are utilised in studying radio spectral properties of the sources in this work. We therefore directly use the cross-matched results as well as the classification of SFGs and AGN in the two radio catalogues, i.e., the LOFAR 150 MHz catalogue presented in <cit.> and the GMRT 610 MHz catalogue presented in <cit.>. §.§.§ Cross-match and SFGs selection of 150 MHz sources To produce a multi-wavelength photometric catalogue that is used for radio-optical/IR cross-match, <cit.> first created the pixel-matched images by using 20 UV to mid-IR bands data in the ELAIS-N1 field. Sources were detected from a χ^2 sum of optical SpARCS-u, PS1-griz, and NIR UKIDSS-DXS J, K imaging data and a separate χ^2 image that using only the Spitzer-IRAC 3.6 and 4.5 data from both the SWIRE and the SERVS (Section <ref>). This resulted in a multi-wavelength catalogue that contains ∼1.5 million sources in the overlapping region of Pan-STARRs, UKIDSS-DXS, and Spitzer-SWIRE surveys. These sources were used for cross-match with the LOFAR radio catalogue. For the 31,610 LOFAR 150 MHz sources within the overlapping region, 97.6% (30,839/31,610) of them have been identified with an optical, and/or NIR, and/or MIR counterpart in <cit.>. The photometric redshift and optical depth of these 150 MHz sources were estimated from the best-fit spectral energy distributions (SEDs) by using the aperture-matched UV to MIR photometries presented in <cit.>. The photometric redshift was estimated by using a hybrid methodology that combines template fitting and machine learning. Details of the estimation are described in <cit.>. For the sources with z<1.5, <cit.> estimated their physical properties, including the optical depth, by adopting the SED fitting code MAGPHYS <cit.>. The fitting was done by determining the best-fit χ^2 between the photometry and all possible models that satisfy the energy balance criterion (which links the energy absorbed at short wavelengths with that re-radiated by dust in the FIR), assuming a two-component dust model from <cit.> and an initial mass function from <cit.>. The optical depth is the natural logarithm of the difference between the unattenuated and attenuated best-fit model spectrum at V-band. To do the source classification for these 150 MHz sources, Best et al. (submitted) fitted the aperture-matched multi-wavelength photometries using four different SED fitting codes, two of which <cit.> included AGN templates in the fitting, and two of which <cit.> did not. For CIGALE and AGNfitter, the output of the codes is the best estimate of the fraction of the mid-infrared light, which is contributed by the AGN (f_ AGN). Best et al. (submitted) used these two values of f_ AGN, combined with information on whether the CIGALE and AGNfitter codes provided a better overall fit to the data than BAGPIPES and MAGPHYS, to determine whether there was evidence for the presence of a radiative AGN in the SED. A significant contribution to this classification is the behaviour of the SED through the mid-IR region of the spectrum, where previous mid-IR colour-colour selection techniques have been used to identify AGN <cit.>. However, the use of the full SED information provides superior results, and is possible in ELAIS-N1 due to the high-quality multi-wavelength datasets. In total, Best et al. (submitted) classified 3,295 sources as hosting optical/IR AGN activity. For a total of 31,610 LOFAR 150 MHz sources, 1,072 of them are unable to be classified due to an absence of critical data, typically a photometric redshift. In addition, Best et al. (submitted) also investigated the presence of radio AGN activity within the sample. Using the outputs of the four SED fitting codes, a consensus stellar mass and consensus SFR was derived for each object (see Best et al. (submitted) for details). Using an SFR to radio luminosity relation defined from the ridge-line of the distribution <cit.>, Best et al. (submitted) were able to identify those sources which showed more than 0.7 dex excess in radio luminosity relative to the value that would be predicted from their consensus SFR. In total, 4,797 sources were identified in this manner to be radio-excess AGN (of which 510 were also optical/IR selected AGN). And 1,314 sources remained unclassifiable. Combining the two sets of classifications, we obtain a sample of 22,720 SFGs, after removing all objects which were either classified as AGN through optical/IR or radio-excess selection, or were unclassifiable due to a lack of relevant data in one of the two diagnostics. For the radio-detected SFGs, except the photometric redshift and optical depth at V-band from <cit.> and <cit.>, we also use the consensus stellar mass from Best et al. submitted in Section <ref>. §.§.§ Cross-match and SFGs selection of 610 MHz sources For the 4,290 GMRT 610 MHz-detected sources in <cit.>, 3,689 of them have a MIR counterpart from the SERVS and 3,542 have a NIR counterpart from the UKIDSS-DXS. Among them, 3,295 have been detected at both MIR and NIR. The multi-wavelength counterparts of these 610 MHz sources were identified by using their MIR and/or NIR positions. Based on these cross-match results, <cit.> carried out a multi-wavelength study using optical, X-ray, infrared, and radio diagnostics to identify AGN in their sample. We refer readers to <cit.> for more details of the cross-match and SFG/AGN classification. Here we briefly summarize the results of their classification. For the 4,290 GMRT 610 MHz sources in <cit.>, 3,490 of them have at least one measured multi-wavelength SFG/AGN classification diagnostic. Among them, * 209 are classified as MIR AGN by adopting the MIR colour-colour selection criteria in <cit.>, * 106 are optical spectroscopic AGN by using the BOSS CLASS and SUBCLASS parameters <cit.>, * 102 are X-ray AGNs with L_ X>10^42 erg s^-1, and * 340 are classified as radio loud AGN by comparing their MIR (MIPS 24 um) to radio flux ratio <cit.>. We, therefore, remove a total of 648 AGNs and obtain a sample of 2,842 SFGs from the GMRT 610 MHz detected sources. For the GMRT 610 MHz SFGs that have a LOFAR 150 MHz counterpart within 20 (the cross-match radius of the two datasets as we discussed in Section <ref>), 70% of them were also classified as SFGs in Best et al. submitted. However, 25% of them were classified as AGN, specifically, 17% were classified as radio AGN in Best et al. submitted. Therefore, the different selections of radio AGN cause the major difference between the source classifications presented in Best et al. submitted and <cit.>. The remaining 5% were unclassifiable due to a lack of measured diagnostic in Best et al. submitted. In this work, to obtain a clean sample of SFGs, we remove all AGN classified in <cit.> or Best et al. from our sample as presented in Section <ref>. § RESULTS Because of the high sensitivity of LOFAR 150 MHz data, we use 610 MHz or 385 MHz as the selection frequency in studying radio spectral properties at observer-frame frequencies of 150–610 MHz and 150–385 MHz (Section <ref> and <ref>) to obtain the relatively complete samples. §.§ GMRT-DEEP 610 MHz as the selection frequency We first use the GMRT-DEEP 610 MHz as the selection frequency to measure the radio spectral index at observer-frame frequencies of 150–610 MHz. To include all possible 150 MHz counterparts of these 610 MHz-selected sources, we first use a cross-match radius of 10. For the 4,290 GMRT 610 MHz sources in <cit.>, 87% (3,748/4,290) of them have a LOFAR 150 MHz counterpart within 10. This fraction increases to 90% if we limit the 610 MHz sources to those with an associated redshift <cit.>. For the 2,842 GMRT 610 MHz-selected SFGs (Section <ref>), these fractions are also 87% (for all SFGs) and 90% (for SFGs with redshift) respectively. §.§.§ Radio spectral index between the observer-frame frequencies of 150–610 MHz We first measure the radio spectral index at observer-frame frequencies of 150–610 MHz, α_ 150^ 610, for a total of 2,462 GMRT 610 MHz-selected SFGs, which have a LOFAR 150 MHz counterpart within 10. The α_ 150^ 610 is measured by performing a linear fit in logarithmic space, and its uncertainty (σ_150^610) is inherited from the uncertainties of the two flux densities. We show the α_ 150^ 610 as a function of the separation between the GMRT 610 MHz and LOFAR 150 MHz detections in Figure <ref>. The distribution of α_ 150^ 610 becomes asymmetric when the separation is greater than 20 as shown in Figure <ref>. We therefore choose r_ 150-610=20 as the cross-match radius when identifying the LOFAR 150 MHz counterparts for the 610 MHz selected SFGs. To reduce the contamination from AGN, we remove additional 590 GMRT 610 MHz sources that have a LOFAR 150 MHz counterpart within 20 and are classified as AGN in Best et al. submitted. The left panel of Figure <ref> shows the measured α_150^610 for the remaining 1,782 SFGs. There are 16 SFGs with r_ 150-610<20 but have the measured α_150^610-σ_150^610>0.0 (very flat radio spectrum at 150–610 MHz). We show their SNR at 150 MHz as a function of α_150^610 in the top-left inset plot in Figure <ref>. We find that those SFGs with a flatter radio spectrum do have a lower SNR at 150 MHz. We also visually inspect both LOFAR 150 MHz and GMRT 610 MHz images and find that four of them are resolved as two components at LOFAR 150 MHz. We therefore remove them from our sample. There are also two sources with α_150^610+σ_150^610<-1.0 (very steep radio spectrum at 150–610 MHz). Visually inspection shows that they are faint at GMRT 610 MHz. There are 1,590 SFGs with r_ 150-610<20 after we remove the four SFGs that are resolved at 150 MHz. We show the distribution of their α_ 150^ 610 in the right plot of Figure <ref>. The median is α_ 150 MHz^ 610 MHz=-0.51±0.01 with a scatter of σ=0.24 for these SFGs. For the 4,290 GMRT 610 MHz-detected sources, 800 of them lack a measured diagnostic to classify them as SFGs or AGN in <cit.>. Among them, 667 have a LOFAR 150 MHz counterparts within 20 and 596/667 are classified as SFGs in Best et al. submitted. As shown in Figure <ref>, the median α_ 150^ 610 of these 596 SFGs is α_ 150 MHz^ 610 MHz=-0.52±0.01 with a standard deviation of σ=0.26, which is highly consistent with the results shown in Figure <ref>. We therefore measure the radio spectral index at observer-frame frequencies of 150–610 MHz for 2,186 GMRT 610 MHz-selected SFGs in total. For the GMRT 610 MHz-selected sources in <cit.>, although the completeness of the detections at LOFAR 150 MHz is ∼90%, we further check these sources without LOFAR 150 MHz counterparts. There are 327 GMRT 610 MHz-selected sources with estimated photometric redshift in <cit.> but not included in the LOFAR final cross-matched catalogue <cit.>. Our visual inspection shows that most of them are within the poor quality areas of the LOFAR map, such as near a bright source. We directly measure the pixel value at their 610 MHz positions and find 92% of them have the value >100 μJy beam^-1, which would correspond to a > 5σ detection in a cleaner part of the LOFAR image. Therefore, although these sources are not included in the LOFAR catalogue because of the inability to fit with a Gaussian profile, there is no reason to expect that this will bias our results. §.§.§ Radio spectral properties at 150–610 MHz For the GMRT 610 MHz-selected SFGs, we show their radio spectral index between observer-frame frequencies of 150–610 MHz as functions of redshift, optical depth at V-band, and stellar mass in Figure <ref>. The photometric redshifts for the total 2,186 SFGs are from the LOFAR multi-wavelength photometric catalogue <cit.>. To reduce the effect from incompleteness of our radio flux-limited sample, we limit SFGs to z<3 when fitting the correlations between α_150^610 and redshift (rest-frame frequency). We perform a linear fit by minimizing the χ^2 and using the uncertainties of α_150^610 as the inverse weights. As shown in the left panel of Figure <ref>, on average, the radio spectral slope at observer-frame 150–610 MHz slightly flattens at low redshift (low frequency) with a ∼3 σ significance. We further divide our sample into the massive (log (M_*/)>10.8) and less-massive (log (M_*/)≤10.8) subsamples with sample sizes of 1,040 and 1,020 respectively. The correlation between α_150^610 and redshift of the two subsamples are consistent with that of the full sample with a linear fit slope of β=-0.03±0.01, although the less-massive subsample is affected more by the incompleteness of our radio flux-limited sample. In our sample, there are 1,481 SFGs with z<1.5, which have estimated optical depth at V-band, τ_V, from <cit.>. As shown in the middle panel of Figure <ref>, the radio spectrum at observer-frame 150–610 MHz flattens with increasing optical depth. The linear fit slope of this correlation is β=-0.06±0.01. We also divide the sample into massive (log (M_*/)>10.7) and less-massive (log (M_*/)≤10.7) subsets with sample sizes of 718 and 763 respectively. We obtain the linear fit slopes of β=0.09±0.01 and β=0.05±0.01 for the two subsets respectively, which confirms that this trend is independent from the stellar mass of SFGs. To reduce the effect of incompleteness, we limit our sample to log (M_*/)>10 when fitting the correlations between α_150^610 and stellar mass as shown in the right panel of Figure <ref>. On average, the radio spectral slope at observer-frame 150–610 MHz slightly steepens with increasing stellar mass for these radio-selected SFGs with a linear fit slope of β=-0.07±0.01. We further divide our sample into high-redshift (z>0.9) and low-redshift (z≤0.9) subsets and find that this trend is independent from redshift, with the linear fit slopes being β=-0.07±0.01 and β=-0.09±0.01 respectively. These results are consistent with our previous work based on the MeerKAT and VLA data in the COSMOS field <cit.>, although for a relatively higher frequency (observer-frame frequencies of 1.3–3 GHz). §.§ GMRT-WIDE 610MHz as the selection frequency We also use the relatively shallower but wide GMRT 610 MHz data from <cit.> to study radio spectral properties between observer-frame frequency of 150–610 MHz for radio-selected SFGs. For the 6,400 GMRT 610 MHz-detected sources <cit.>, 4,542 are within the region that has both LOFAR and multi-wavelength observations (Figure <ref>). Among them, 4,070 (90%) have a LOFAR 150 MHz counterparts within 10 and 1,857 are classified as SFGs in Best et al. submitted. In addition, 1,258/6,400 overlap with GMRT-deep detected sources and 394 are classified as AGNs in <cit.>. The overlap between the two AGN samples is 342. We therefore select a total of 1,805 SFGs and measure their spectral indices between observer-frame frequency of 150–610 MHz as shown in Figure <ref>. We also adopt r_150-610=20 as the cut when cross-matching the GMRT (wide) 610 MHz and LOFAR 150 MHz data as shown in Figure <ref>. There are 48 sources with r<20 that have α_150^610-σ_150^610>0.0 (very flat radio spectrum between observer-frame frequency of 150-610 MHz), and we show their SNR at 150 MHz as a function of α_150^610 in the top inset plot in Figure <ref>. We visually inspect both LOFAR 150 MHz and GMRT 610 MHz images and find that six of these 48 sources are resolved as two components at LOFAR 150 MHz. We therefore remove these six sources in the following analyses. Besides, as shown in the top inset plot of Figure <ref>, the sources with an extremely flatter radio spectrum at 150–610 MHz have lower SNR at 150 MHz. For the remaining 1,711 SFGs with r_150-610<20, we obtain a median radio spectral index of α_150^610=-0.49±0.01 with a scatter of σ=0.26, which are well-consistent with the results based on the GMRT deep data (Section <ref>). We note that 396 of these SFGs are also included in the GMRT-deep sample shown in Figure <ref>. For the 1,711 GMRT-wide-detected SFGs, 1,296 with z<1.5 have estimated optical depth at V-band, τ_V, from <cit.>. As shown in Figure <ref>, on average, the radio spectrum at observed 150–610 MHz for the GMRT-wide-selected SFGs also slightly flattens with increasing optical depth and steepens with increasing stellar mass. However, we do not find the weak correlation between α_150^610 and redshift (rest-frame frequency). This might be caused by the sample of SFGs being not as clean as the one based on GMRT-deep-detected sources (Section <ref>). Overall, the radio spectral properties of GMRT-wide selected SFGs are consistent with those based on the GMRT-deep selected sample. §.§ uGMRT 400 MHz as the selection frequency We use the uGMRT 400 MHz-detected sources from <cit.> to study radio spectrum between observer-frame frequencies of 150–385 MHz and 385–610 MHz respectively. As we describe in Section <ref>, the effective frequency of uGMRT 400 MHz used in this work is 385 MHz. For the 2,528 uGMRT 400 MHz radio sources from <cit.>, 2,219 (88%) have a LOFAR 150 MHz counterpart within 10. Among them, 1,255 are classified as SFGs in Best et al. submitted. In addition, 2,113 of the 2,528 uGMRT 400 MHz sources have a GMRT 610 MHz counterpart within 10. Among them, 1180 are within 25 and 401/1,180 are classified as AGNs in <cit.>. The overlap between the two AGN samples from Best et al. submitted and <cit.> is 340. Therefore, we remove the additional 61 AGN from our sample and obtain a sample of 1,194 SFGs. §.§.§ Radio spectral indices at observed 150–385 MHz and 385–610 MHz We show the radio spectral index between observer-frame frequencies of 150–385 MHz as a function of the separation between LOFAR 150 MHz and uGMRT 400 MHz detections for these 1,194 SFGs in the left panel of Figure <ref>. According to the distribution of α_150^385, we choose r_ 150-385=25 as the cut when cross-matching the uGMRT 400 MHz and LOFAR 150 MHz sources. There are 20 SFGs with r_ 150-385<25 that have α_150^385-σ_150^385>0.3, which means that they have very flat radio spectra at observed 150–385 MHz. Our visual inspection shows that three of them are resolved as two components at LOFAR 150 MHz. There are also two SFGs with α_150^385+σ_150^385<-0.8, i.e., with very steep radio spectra at observed 150–385 MHz. We find that both of them are resolved at LOFAR 150 MHz. We therefore remove these five sources from our sample and show the distribution of α_150^385 for the remaining 1,158 SFGs with r_ 150-385<25 in Figure <ref>. The median radio spectral index between observer-frame frequencies of 150–385 MHz is α_150^385=-0.29±0.01 with a scatter of σ=0.31. For the 1,194 SFGs that have both uGMRT 400 MHz and LOFAR 150 MHz detections, 1,038 also have a GMRT 610 MHz counterpart within 10. We show their radio spectral indices between observer-frame frequencies of 150–610 MHz as a function of the separation between LOFAR 150 MHz and GMRT 610 MHz detections in the middle panel of Figure <ref>. Among them, 896 have r_ 150-610 MHz<20 after we remove the sources that are resolved at LOFAR 150 MHz. The median α_150^610 for these 896 uGMRT 400 MHz-selected SFGs is α_150^610=-0.52±0.01 with a scatter of σ=0.23, which is consistent with the results based on GMRT 610 MHz-detected SFGs (Section <ref> and <ref>). We also show the radio spectral index between observer-frame frequencies of 385–610 MHz as a function of the separation between uGMRT 400 MHz and GMRT 610 MHz detections for the 1,038 uGMRT 400 MHz-selected SFGs that have a GMRT 610 MHz counterpart within 10 in the right panel of Figure <ref>. According to the distribution of α_385^610, we choose r_ 385-610=25 as the cut when cross-matching uGMRT 400 MHz and GMRT 610 MHz sources. There are 926 uGMRT 400 MHz-selected SFGs that have r_ 385-610<25 after we remove the SFGs that are resolved at LOFAR 150 MHz. The median spectral index is α_385^610=-0.91^+0.02_-0.03 with a scatter of σ=0.66. The differences in spectral slope above and below 385 MHz occur because of the relatively shallower uGMRT 400 MHz data, i.e., the radio spectral indices we measured are flatter over the 150–385 MHz and steeper over the 385–610 MHz (Eddington bias). To get a less biased sample in studying radio spectrum at observer-frame frequencies of 150–385 MHz and 385–610 MHz, we instead select the sample for study by applying a flux density cut at GMRT 610 MHz. The coverages of GMRT deep 610 MHz and uGMRT 400 MHz are slightly offset as shown in Figure <ref>. Within the overlapped region, only 47% (2,024/4,269) of GMRT 610 MHz sources have a 400 MHz counterpart within 10. We therefore limit the sources to S_ 610 MHz>300 μJy to guarantee that >80% of GMRT 610 MHz sources have a 400 MHz counterpart. Within the overlapped region, there are 303 SFGs with S_ 610 MHz>300 μJy. Among them, 258 have a 400 MHz counterpart within 25. All of them also have a LOFAR 150 MHz counterpart within 20 thanks to the high sensitivity of LOFAR data. The median spectral index between observer-frame frequency of 150–385 MHz, 385–610 MHz, and 150–610 MHz are α_150^385=-0.42^+0.03_-0.02, α_385^610=-0.44^+0.03_-0.04, and α_150^610=-0.42^+0.02_-0.01, which are consistent with the results from nearby brighter galaxies <cit.>. §.§.§ Radio spectral properties at 150–385 MHz and 385–610 MHz For SFGs with S_ 610 MHz>300 μJy and with detections at GMRT 610 MHz, uGMRT 400 MHz, and LOFAR 150 MHz, we show the correlations between their radio spectral indices and redshift, optical depth at V-band, and stellar mass in Figure <ref>. We find that, on average, the radio spectra at observer-frame frequency of 150–385 MHz and 385–610 MHz are not significantly correlated with redshift, i.e., rest-frame frequency, but flatten with increasing optical depth and steepen with increasing stellar mass. Consequently, the difference between radio spectral indices at observer-frame frequency of 150–385 MHz and 385–610 MHz, i.e., spectral curvature, is not correlated with these physical properties as shown in the third row of Figure <ref>. §.§ JVLA 5 GHz as the selection frequency Although the JVLA 5 GHz observations only cover the central 0.13 deg^2 of the ELAIS-N1 field, the sensitivity reaches ∼1 μJy beam^-1. We therefore use JVLA 5 GHz as the selection frequency to study the radio spectral properties for radio-selected SFGs at observer-frame frequencies of 150–5000 MHz. §.§.§ Radio spectral indices at observed 150–5000 MHz Because of a relatively higher angular resolution of JVLA 5 GHz data, we use a radius of 30 when cross-matching the JVLA 5 GHz sources with the radio sources detected at 610 MHz, 400 MHz, or 150 MHz. We find that applying a smaller matching radius, i.e., 15, does not change the median radio spectral indices at observer-frame frequencies of 385-5000 MHz and 150-5000 MHz, and only makes the median radio spectrum at 610-5000 MHz slighter flatter. We therefore keep all the 610 MHz, 400 MHz, and 150 MHz sources that are matched with the JVLA 5 GHz within 30 in the following analyses. For the 387 JVLA 5 GHz-detected sources in <cit.>, 210 of them are cross-matched with a GMRT 610 MHz sources within 30. Among them, 137 are classified as SFGs in <cit.>. We also remove 72 JVLA 5 GHz sources that have a LOFAR counterparts and are classified as AGN by Best et al. submitted. The overlap of the two AGN samples is 35. For the remaining 100 SFGs, we show the measured radio spectral index at observer-frame frequency of 610–5000 MHz, α_610^5000, as a function of the separation between JVLA 5 GHz and GMRT 610 MHz detections in Figure <ref>. The median is α_610^5000=-1.14^+0.03_-0.05 with a scatter of σ=0.32. There are 168 of the 387 JVLA 5 GHz sources that have an uGMRT 400 MHz counterpart within 30. By removing the 31 and 65 AGN classified in <cit.> and Best et al. submitted respectively, we obtain a median radio spectral index of α_385^5000=-1.08_-0.02^+0.01 for a sample of 81 SFGs (Figure <ref>). The standard deviation of the measured radio spectral indices for these SFGs is σ=0.21. There are 217 JVLA 5 GHz sources that are matched with a LOFAR 150 MHz source within 30. Among them 133 are classified as SFGs in Best et al. submitted. We also remove additional five AGN that are classified in <cit.>. For the remaining 128 SFGs in our sample, the median radio spectral index at observer-frame frequency of 150–5000 MHz is α_150^5000=-0.87^+0.01_-0.02 with a scatter of σ=0.18. §.§.§ Radio spectral properties at observed 150–5000 MHz We show the measured radio spectral indices at observer-frame frequencies of 610–5000 MHz, 385–5000 MHz and 150–5000 MHz as functions of redshift, optical depth at V-band, and stellar mass in Figure <ref>. We find that, on average, the radio spectrum of SFGs slightly steepens with increasing redshift and stellar mass at 150–5000 MHz. However, there is no significant correlation between radio spectral index and optical depth at V-band, which is different from that at low-frequency (ν<∼1 GHz, Figure <ref> and Figure <ref>). For the 100 SFGs that have both JVLA 5 GHz and GMRT 610 MHz (deep) detections, 82 of them also have a LOFAR 150 MHz counterpart within 20 (the cross matching radius between GMRT 610 MHz and LOFAR 150 MHz sources). Although the sample size is very limited and the JVLA 5 GHz data is deeper than the data at the other two frequencies, we find that the difference between radio spectral indices at observer-frame frequencies of 150–610 MHz and 610–5000 MHz is not correlated with redshift (rest-frame frequency) and stellar mass but positively correlated with optical depth at V-band weakly. This is not surprising given the lack of significant correlation between the radio spectral index and the V-band optical depth at ν>∼1 GHz as shown in Figure <ref>. § DISCUSSION By combining the LOFAR 150 MHz, uGMRT 400 MHz, GMRT 610 MHz, and JVLA 5 GHz data in the ELAIS-N1 field, we study the radio spectral properties at observer-frame frequencies of 150–5000 MHz. Here we discuss the possible physical mechanisms that determine the radio spectrum at this frequency range. §.§ Radio spectral indices at observer-frame 150–5000 MHz By using the two radio datasets with the greatest sensitivity and coverage in this work, we obtain a total of ∼3,500 SFGs that have detections at both GMRT 610 MHz and LOFAR 150 MHz and a median spectral index of α_610^150=-0.51±0.01. Our measurement is consistent with the median spectral indices at similar frequency ranges of nearby galaxies <cit.> and submillimeter-selected galaxies <cit.>. We note that <cit.> and <cit.> also include the ∼150 MHz data in their analyses from the LOFAR Multifrequency Snapshot Sky Survey and the LoTSS respectively. However, compared with the median radio spectral index at high frequency, specifically, at rest-frame ∼1.3–10 GHz in <cit.>, as well as the synchrotron spectrum, the radio spectrum of SFGs is flatter at ν_ rest<∼1 GHz. JVLA 5 GHz data have an RMS noise of 1.05μJy, which corresponds to ∼6μJy, ∼8μJy, and ∼17μJy at 610 MHz, 385 MHz, and 150 MHz respectively if we assume a radio spectral index of α=-0.8 (synchrotron spectrum). Therefore, using 5 GHz as the selection frequency biases the sample to SFGs with a relatively steeper radio spectrum, which causes the median radio spectral slope at observer-frame frequencies of 610–5000 MHz, 385–5000 MHz and 150–5000 MHz to be steeper than that of the synchrotron spectrum as shown in Section <ref>. Although including the JVLA 5 GHz data biases the sample to SFGs with a relatively steeper radio spectrum because of the high-sensitivity of the data, the radio spectrum is flatter if we include a lower frequency dataset when measuring the radio spectral index between observer-frame frequency of 150–5000 MHz as shown in Figure <ref>. However, the high-sensitivity of LOFAR 150 MHz data is also the reason for a flatter radio spectrum at 150–5000 MHz. Based on the correlations between radio spectral indices and physical properties of SFGs we find in this work, we will discuss the possible physical mechanisms that cause the radio spectrum to be flatter at low frequency (ν_ rest<∼1 GHz) than at high frequency in Section <ref>. §.§ Radio spectral properties at observer-frame 150–5000 MHz For radio-selected SFGs, we find that, on average, their radio spectrum slightly steepens with increasing stellar mass at any frequency ranges analysed in this work. However, we only find that at ν_ obs<∼1 GHz, the radio spectrum of SFGs slightly flattens with increasing optical depth. This results a positive correlation between spectral curvature (α_150^610-α_610^5000) and optical depth at V-band. At ν_ obs<∼1 GHz, the difference between radio spectra indices (α_150^385-α_385^610) is not significantly correlated with physical properties of SFGs. This phenomenon has been reported in <cit.>, which found that the spectral curvature at observer-frequency of 150–1400 MHz does not dependent on redshift, SFR, and specific SFR of SFGs. In addition, the median radio spectral indices we measured at 150–5000 MHz indicate that the radio spectrum is flatter at low frequency than at high frequency. Here we discuss the possible explanations of these correlations. §.§.§ Radio spectrum steepens with increasing stellar mass At observer-frame frequency of 150–5000 MHz, we find that the radio spectrum slightly steepens with increasing stellar mass. This correlation has been reported in our previous work that studied radio spectral properties at observer-frame frequency of 1.3–3 GHz by combining the MeerKAT 1.3 GHz and VLA 3 GHz data in the COSMOS field <cit.>. In that work, we noted that this correlation could be explained by age-related synchrotron loss. For the relativistic CR electron, the critical frequency, at which the electron emits most of its energy, is proportional to the square of its energy and the strength of the perpendicular component of the magnetic field, i.e, ν_ c∝ E^2 B_. If we only take synchrotron radiation in account, the lifetime of a CR electron is t_ sy∝ν_ c^-0.5 B_^-1.5. Therefore, high-energy CR electrons lose their energy faster than that of low-energy CR electrons. Similarly, the inverse Compton loss of these CR electrons is also greater at high frequency <cit.>. For SFGs, fresh CR electrons are constantly injected in the star-forming regions, which keeps the radio spectral slope constant. We therefore discussed in <cit.> that the slightly increased ratio of aged and young relativistic CR electrons with increasing stellar mass, could be a physical cause leading to radio spectrum steepening for massive SFGs. §.§.§ Radio spectrum flattens with increasing optical depth at ν<∼1 GHz At ν_ obs<∼1 GHz, the radio spectrum of SFGs slightly flattens with increasing optical depth at V-band. However, this correlation is not found when we include the high-frequency JVLA 5 GHz data, although the sample size of 5 GHz-detected sources is very limited because of the small coverage of the data. The effective optical depth at V-band depends on the relative distribution of the diffuse gas and dust along the line of sight. SFGs with higher optical depth at V-band are expected to have higher column density of free electron (n_ e) along the line of sight. Besides, these electrons might be more associated with colder regions around the star formation sites in SFGs. Therefore they are affected more by free-free absorption because the absorption coefficient has the following form <cit.>: κ∝ n_e^2 T_e^-1.35 ν^-2.1, where T_ e is the electron temperature of the emitting region. This equation and other theoretical models as well as previous observations show that the effect of thermal free-free absorption increases at low frequency <cit.>, which explains why we only find the correlation between optical depth at V-band and radio spectrum at low frequency. §.§.§ The low-frequency radio spectrum is flatter than the high-frequency As discussed in Section <ref>, our measured median radio spectral index at observer-frame frequency of 150–50000 MHz show that, on average, the radio spectrum of SFGs is flatter at low frequency than at high frequency. However, when we investigate the measured radio spectral index as a function of rest-frame frequency (redshift) at either frequency range, i.e., 150-610 MHz (Figure <ref>) or 385–610 MHz (Figure <ref>), we only find a very weak or even negligible correlation because of the large scatter of measured radio spectral indices at any given rest-frame frequencies. The correlations of SFGs' radio spectral slope with the stellar mass and with the optical depth at V-band suggest that both spectral ageing of CR electrons and thermal free-free absorption could be the physical causes of why the radio spectrum is flatter at low frequency than at high frequency. These two physical mechanisms are also used as the main explanations for the different radio spectral slopes at low and high frequency in the literature <cit.>. The other mechanisms, such as synchrotron self-absorption and the Razin-Tsytovich effect, may also flatten the radio spectrum at low-frequency. However, synchrotron self-absorption requires extremely high brightness temperatures in order to be significant, and so its impact in SFGs is expected to be weak <cit.>. The Razin-Tsytovich effect can suppress synchrotron emission below the Razin frequency, ν_ R∝ n_eB^-1. This effect may dominate over internal free-free absorption when the thermal gas is considerably hotter than 10^4 K, which is rarely attained by normal SFGs <cit.>. As shown by Equation <ref> to <ref>, the magnetic field (B), column density of free electrons (n_e), and electron temperature (T_ e) in the region are the key features of the physical mechanisms that are responsible for the different radio spectral slopes at low and high frequency discussed above. Studies of radio spectrum of nearby galaxies and theoretical simulations of CR electron propagation discussed the effect of these individual physical parameters as well as some other physical conditions, such as galactic wind <cit.>. In contrast, one advantage of our statistical studies is the ability to suggest possible causes of why the radio spectrum is flatter at low frequency than at high frequency. In particular, we find that the correlations between the SFGs' integrated radio spectrum and their astrophysical properties could lend support to spectral ageing and thermal free-free absorption as two possible main mechanisms for our samples. § CONCLUSION By combining high-sensitivity LOFAR 150 MHz, uGMRT 400 MHz, GMRT 610 MHz, and JVLA 5 GHz data in the ELAIS-N1 field, we measure the radio spectral indices between observer-frame frequency of 150–5000 MHz and study their correlations with physical properties of SFGs. The main conclusions from our work are as follows. 1) We obtain ∼3,500 SFGs that have detections at both GMRT 610 MHz and LOFAR 150 MHz by removing AGN from the two radio samples. The median radio spectral index for these SFGs is α^610_150=-0.51±0.01 with a scatter of σ=0.2. Therefore, the radio spectrum at 150–610 MHz is flatter than that at high frequency (ν_ rest>∼1 GHz) and the synchrotron spectrum. 2) Because of the relatively low-sensitivity of uGMRT 400 MHz data, we limit our sample to the SFGs with S_ 610 MHz>300 μJy, and obtain a median spectral indices of α^385_150=-0.42^+0.03_-0.02, α^610_385=-0.44^+0.03_-0.04 and α^610_150=-0.42^+0.02_-0.01. 3) Although JVLA 5 GHz data has very limited coverage, and its high-sensitivity biases the sample to SFGs with a relatively steeper radio spectrum, we obtain median α_610^5000=-1.14^+0.04_-0.05, α_385^5000=-1.08^+0.01_-0.02 and α_150^5000=-0.87±0.01 for samples of ∼100 SFGs that have detections at the two corresponding frequencies. Therefore, we find that the radio spectrum is flatter if we include a lower frequency dataset, although the high sensitivity of LOFAR 150 MHz data is also the reason for a flatter radio spectrum at 150–5000 MHz. 4) We correlate the measured radio spectral indices with physical properties of SFGs and find that, on average, the radio spectrum steepens with increasing stellar mass at observer-frame frequency of 150–5000 MHz. We discuss that the spectral ageing due to the energy loss of relativistic CR electrons via synchrotron and inverse Compton radiation could explain this trend. 5) At ν_ obs<∼1 GHz, we find that, on average, the radio spectrum slightly flattens with increasing optical depth at V-band of radio-selected SFGs. However, this correlation is not found when we include the high-frequency JVLA 5 GHz data. We suggest that thermal free-free absorption is a possible underlying physical mechanism of this correlation. 6) Our statistical study of the correlation between integrated radio spectrum and physical properties of SFGs suggests the spectral ageing of CR electrons and thermal free-free absorption as two main physical mechanisms that cause the radio spectrum to be flatter at low frequency than at high frequency. FXA and MV acknowledge financial support from the Inter-University Institute for Data Intensive Astronomy (IDIA), a partnership of the University of Cape Town, the University of Pretoria, the University of the Western Cape, and the South African Radio Astronomy Observatory. FXA is grateful for support from the National Science Foundation of China (12233005, 12073078, and 12173088). MV acknowledges the support from the South African Department of Science and Innovation's National Research Foundation under the ISARP RADIOSKY2020 Joint Research Scheme (DSI-NRF Grant Number 113121) and the CSUR HIPPO Project (DSI-NRF Grant Number 121291). PNB and RK acknowledge support from the UK STFC via grant ST/V000594/1. We acknowledge the use of the ilifu cloud computing facility –www.ilifu.ac.za, a partnership of the University of Cape Town (UCT), the University of the Western Cape, the University of Stellenbosch, Sol Plaatje University, the Cape Peninsula University of Technology, and the South African Radio Astronomy Observatory. The ilifu facility is supported by contributions from the IDIA, the Computational Biology division at UCT, and the Data Intensive Research Initiative of South Africa (DIRISA). § DATA AVAILABILITY The LOFAR 150 MHz data used for the analysis of this paper are publicly available at https://lofar-surveys.org/deepfields.html. The GMRT 610 MHz and JVLA 5 GHz data used in this research will be shared upon reasonable request to the corresponding author. mnras

http://arxiv.org/abs/2303.06887v1

20230313061017

Negative transverse magnetoresistance due to negative off-diagonal mass in linear dispersion materials

cond-mat.mes-hall

]Negative transverse magnetoresistance due to negative off-diagonal mass in linear dispersion materials ^1 Department of Engineering Science, University of Electro-Communications, Chofu, Tokyo 182-8585, Japan ^2 Institute for Advanced Science, University of Electro-Communications, Chofu, Tokyo 182-8585, Japan awashima@kookai.pc.uec.ac.jp This study calculated the magnetoresistance (MR) in the Dirac electron system, Dresselhaus–Kip–Kittel (DKK) model, and nodal-line semimetals based on the semiclassical Boltzmann theory, with particular focus on the detailed energy dispersion structure. The negative off-diagonal effective-mass was found to induce negative transverse MR owing to the energy dispersion effect. The impact of the off-diagonal mass was more prominent in case of a linear energy dispersion. Further, Dirac electron systems could realize negative MR even if the Fermi surface was perfectly spherical. The obtained negative MR in the DKK model may explain the long-standing mystery in p-type Si. § INTRODUCTION Since its discovery by Thomson in 1857 <cit.>, magnetoresistance (MR) has garnered attention from both scientific and practical use perspectives, such as mobility measurements and sensitive magnetic sensors <cit.>. The microscopic origin of orbital MR is the Lorentz force. Therefore, the essential properties of orbital MR must be explainable within the semiclassical theory. However, the simple semiclassical theory provided in textbooks <cit.> cannot comprehensively explain the properties of MR even in elementary systems. For example, the simple semiclassical theory states that transverse MR (TMR) (current and magnetic field are perpendicular) is not observed in single carrier systems with parabolic dispersion. However, experiments have confirmed that many single carrier systems exhibit a clear TMR <cit.>. Further, it states that longitudinal MR (LMR) (current and magnetic field are in parallel) should not be realized in every system owing to the absence of the Lorentz force in such a configuration. However, many systems have been experimentally shown to exhibit a finite LMR, and in certain cases, the LMR has been found to be larger than TMR <cit.>. Effective mass is considered a key aspect for understanding such unexpected properties. It is expressed as α_ij(=m_ij^-1)=ħ^-2∂^2E(k)/∂ k_i∂ k_j. Conventional simple semiclassical theories assume a parabolic energy dispersion, where the effective mass is diagonal and constant (k-independent). However, the MR can exhibit unexpected properties in cases where the k-dependence or the off-diagonal components of the effective mass are involved. For example, even for a single carrier system, the TMR has been observed owing to the k-dependent effective mass <cit.>. In addition, the LMR becomes finite when the effective mass contains off-diagonal elements. Moreover, off-diagonal effective mass driven large LMR has been observed in multi carrier systems <cit.>. Recently, another anomalous MR, referred to as the negative MR has garnered renewed interest, particularly in Weyl semimetals from the perspectives of chiral anomaly and the Berry curvature<cit.>. Following this trend, it was recently proposed that the negative off-diagonal mass can induce negative MR in Weyl semimetals, irrespective of the chiral anomaly or the Berry curvature <cit.>. However, whether this concept of negative off-diagonal mass is common or specific to Weyl semimetals is yet to be understood. Thus, this study showed that the concept of negative MR owing to the negative off-diagonal mass is considerably common, particularly in systems with a linear dispersion. This was realized through the examination of three different systems: the Dirac electrons, Dresselhaus–Kip–Kittel (DKK) model, and nodal-line semimetals. These particular systems were selected because their energy dispersions contain a root sign, which easily yields the negative off-diagonal mass. For example, the energy dispersion of the isotropic Dirac electron system considered in this study is expressed as follows <cit.>: E^Dirac_±=±√(Δ^2+Δħ^2 k^2/m^*), where k^2=k_x^2+k_y^2+k_z^2, 2Δ is a band gap between the conduction and valence bands, m^* is the effective-mass, and ħ is the Dirac constant. The off-diagonal component of the inverse mass α_xy can be calculated as follows: α^Dirac_xy=-Δ^2 ħ^2 k_xk_y/m^*2(Δ^2+ Δħ^2 k^2/m^*)^3/2. As evident, α_xy^ Dirac has a negative sign, which is because of appearance of k-dependence in the root sign. (Specifically, the positive or negative sign of the effective mass must be determined by excluding the factor k_x k_y, that is, by considering α̅_ij=α_ijk_ik_j/|k_i||k_j|.) Similarly, the DKK model and the nodal-line semimetals possess the negative off-diagonal mass; however, their k-dependences are different. This study also discussed the influence of the linearity of the energy dispersion on the off-diagonal elements. § THEORY OF MAGNETORESISTANCE The MR was calculated considering a combination of the semiclassical Boltzmann theory and the well-known Chambers formula<cit.>. Using the Chambers formula, the conductivity σ̂ and resistivity ρ̂ are expressed as σ_ij=e^2⟨ v_i(k)v̅_j(k)⟩_F, ρ̂=σ̂^-1. where e (>0) is the elementary charge and ⟨⋯⟩_F=∫(dk/4π^3)⋯(-∂ f_0/∂ E) represents integration along the Fermi surface at low temperature (f_0 is the Fermi-Dirac distribution function at equilibrium). Further, v(k) is the velocity of the electron obtained from the energy dispersion and is expressed as follows: v(k)=1/ħ∂ E(k)/∂k. Moreover, v̅(k) is the velocity averaged over time for electrons moving in a magnetic field. v̅(k)=∫_-∞^0dtv(k(t))exp(t/τ), dk(t)/dt=-e/ħv(k)×B, where B is the magnetic field and τ is the relaxation time. The Chambers formula does not assume any specific material in the derivation process and is thus applicable to any arbitrary energy dispersion. However, the Chambers formula does not include effects derived from topological quantities such as the Berry curvature or chiral anomaly. The following results were obtained based on the detailed structure of the energy dispersion without any topological effects. § DIRAC ELECTRON SYSTEM First, the single-carrier isotropic Dirac electron system was examined. Its energy dispersion is expressed as Eq. (<ref>). Owing to the Fermi surface of this model being perfectly spherical, no MR is expected upon the application of the parabolic dispersion approximation to this model <cit.>. However, a large negative TMR was observed beyond the parabolic approximation. Figure. <ref> (a) shows the TMR calculated using the Chambers formula for an isotropic Dirac electron system with a magnetic field applied in the z-direction, B=(0,0, B). The calculations yielded a negative TMR, -56% at μ B= 1.0 for E_F=15 meV, where μ is the mobility. (maximum permissible value of the negative MR was -100%). Here, μ B=1.0 corresponded to a relatively weak-field in Dirac electrons systems, where the effective mass was very small with a high mobility <cit.>. For example, μ B=1.0 corresponded to B=0.01 T when μ=100 T^-1. This finite TMR was caused by the k-dependence of the effective mass. Furthermore, the negative TMR originated from the negative off-diagonal mass. The Chambers formula Eq. (<ref>) is not appropriate for determining the physical origin of the field-dependence in the resistivity because the contribution from B is hidden. By contrast, to clarify the physical origin of the field-dependence, the Jones-Zener expansion is extremely useful <cit.>. Thus, the magneto-resistivity ρ_xx was expanded up to the second order with respect to B as follows: ρ^(2)_xx=τ^2[R_xx+R_xy] B^2, R_xx=Λ_x^2⟨ v_x^2(α_xxα_yy-α_xy^2)⟩_F +Λ_x^2⟨ v_x^2(v_xγ_xyy) +v_x^2(v_yγ_xxy+v_yγ_xyx) -v_x^2(v_xγ_xxx)⟩_F. R_xy=-Λ_x^2Λ_y⟨(v_xv_yα_xy-v_y^2α_yy)^2⟩_F where Λ_μ=1/⟨ v_μ⟩_F^2 and γ_ijk=ħ^-3∂^3E(k)/∂ k_i∂ k_j∂ k_k. R_xx originates from σ_xx(B), and R_xy originates from σ_xy^2(B). Consequently, it is evident that the coefficient of B^2 is mostly obtained in terms of the (invers) effective mass α_ij. Therefore, the intuitive understandings for the anomalous MR can be obtained upon an in-depth investigation of the properties of the effective-mass. The key to the negative TMR are R_xy and α_xy, where the former negatively contributes to ρ_xx owing to the negative coefficient in Eq. (<ref>). In the ordinary case of α̅_xy >0, the negative contribution from R_xy is small because the α_yy-term offsets the α_xy-term. However, when α̅_xy<0, both the α_xy and α_yy cooperatively boost the negative contribution of R_xy. Consequently, the negative off-diagonal mass can realize the negative TMR. The Dirac electron system is a typical example for such a negative off-diagonal mass induced negative TMR. Figure. <ref> (c) shows a plot α̅_xy as a function of k_x and k_y. The majority of α̅_xy was negative. Further, α̅_xy was vanishingly small when k_x, y=0, although it was not positive. The widespread negative region in α̅_xy causes the significantly negative TMR in the Dirac electron system. Moreover, the TMR of the Dirac electron system becomes zero in the region Δ≃ E_F, which can be approximated as the parabolic dispersion. The energy dispersion Eq. (<ref>) can be approximated as E^Dirac_+≃Δ+ħ^2/2m(k_x^2+k_y^2+k_z^2), for Δ≃ E_F. The off-diagonal mass becomes zero in case of this parabolic dispersion, and the k-dependence of the diagonal element vanishes. Consequently, the negative TMR should be vanishingly small when E_F approaches Δ. § DKK MODEL The DKK model is an effective model for investigating the valence band around the Γ point of Si and Ge. The energy dispersion is expressed as<cit.> E^DKK_±=Ak^2±√(B^2k^4+C^2(k_x^2k_y^2+k_y^2k_z^2+k_z^2k_x^2)). It is expressed in a form that included the root sign similar to that in case of the Dirac electron system; that is, the negative TMR because of the negative off-diagonal mass is expected. In Si, the parameters were A=-4.29, B=0.68, and C=4.87 in units of ħ^2/2m<cit.>. Further, positive and negative signs corresponded to heavy and light hole, respectively <cit.>. Figure. <ref> (a) shows the energy dispersion of DKK model. The single carrier MR was calculated for each heavy and light hole shown in Fig. <ref> (b). The MR was always positive for heavy hole and saturated at 25%. However, for light hole, the MR increased up to μ B=0.2 and exhibited saturation signs at approximately 5%, after which it began to decrease and became negative at approximately μ B=1.0. In the presence of both positive and negative off-diagonal masses, the behaviour of the MR is governed by the one dominant in the entire Fermi surface. The effective mass α_xy in the DKK model is expressed as α^DKK_xy=A_xA_y/(4ħ^2(D^DKK)^3) ± k_xk_y(8B^2+4C^2)/(2ħ^2 D^DKK), A_x=4B^2k^2k_x+C^2k_x(2k_y^2+2k_z^2), A_y=4B^2k^2k_y+C^2k_y(2k_x^2+2k_z^2), D^DKK=(B^2k^2+C^2(k_x^2k_y^2+k_y^2k_z^2+k_z^2k_x^2))^1/2. Figure. <ref> (c) shows a plot of α̅_xy of light hole against several k_z. Negative off-diagonal element was observed over a wide range; however, positive values appeared for certain values of k_z. The entire Fermi surface, including all k_z, was dominated by negatives; however, positive values weakened the negative off-diagonal element. The opposite was true for heavy hole, which exhibited a negative off-diagonal element in certain parts; however, a positive off-diagonal element was observed in the entire Fermi surface. Thus, a heavy hole with a positive dominant off-diagonal element results in a positive MR. By contrast, a light hole with a dominant negative element yields a negative MR. § NODAL-LINE SEMIMETALS In nodal-line semimetals, the band gap is closed along the nodal-line in k space <cit.>. The following Hamiltonian was proposed for nodal-line semimetals as an effective model obtained using the k·p theory<cit.>. H^nodal={[ħ^2/2m(k_x^2+k_y^2)-u]τ_3+λ k_zτ_1}⊗σ_0. where σ̂ and τ̂ are the Pauli matrices in the spin and orbital spaces, respectively, and m, u, and λ are the model parameters. The eigenvalue is expressed as E^nodal_±=±√((ħ^2/2m(k_x^2+k_y^2)-u)^2+λ^2k_z^2). The energy dispersion is shown in Fig. <ref> (a). Parameter u exhibits an energy dimension and corresponds to the band gap size at the Γ point. For u<E_F, the Fermi surface was a drumhead Fermi surface [Fig. <ref> (c)], and for u>E_F, it was a torus shaped Fermi surface [Fig. <ref> (d)]. As MR results from Lorentz forces, the off-diagonal mass perpendicular to the applied magnetic field is essential. For example, α_xy for B=(0, 0, B) and α_yz for B=(B, 0, 0) are important factors that affect the MR. The model used in this study was anisotropic in the k_z and k_x,k_y directions. Therefore, the effective-masses α^nodal_xy and α^nodal_yz differed significantly from each other. The off-diagonal elements α^nodal_xy and α^nodal_yz of E^nodal_+ are α^nodal_xy=λ^24ħ^2k_xk_yk_z^2/(2m)^2D^nodal, α^nodal_yz=-λ^2k_yk_z[ ħ^2(k_x^2+k_y^2)/2m-u]/mD^nodal, D^nodal={[ħ^2/2m(k_x^2+k_y^2)-u]^2+λ^2 k_z^2}^3/2. In this effective model of the nodal-line semimetals, the behaviour of the TMR differed significantly depending on the direction of the magnetic field. Herein, the TMR characteristics were calculated for two magnetic field directions: B∥ z and B∥ x. The TMR for B∥ z was positive MR for both u=0.0, 0.015 eV, particularly for u=0.015 eV, reaching 230% at μ B=3.0 [Fig. <ref> (a)]. Further, α̅^nodal_xy was always positive for the entire region, as shown in Fig. <ref> (b). By contrast, the TMR for B∥ x decreased after reaching 80% for u=0.015 eV and 10% for u=0.0 eV [Fig. <ref> (c)]. The inset in Fig. <ref> (c) shows the calculation results when λ=1 eV.Å. Compared to the case with λ=0.32 eV.Å, the decreasing trend of MR was stronger and a negative MR was obtained at approximately μ B=1.0. In contrast to the case of B∥ z, α̅^nodal_yz exhibited a negative region, particularly for u=0.0 [Fig. <ref> (d)]. This negative off-diagonal mass was the source of the decreasing TMR for B∥ x. § DISCUSSION §.§ Linearity The impact of the negative off-diagonal mass on the MR is observed to be more significant for systems with linear dispersion. This study demonstrated the reason why the linear dispersion induced negative MR compared with the quadratic dispersion, through simple examples. Suppose that the linear and quadratic energy dispersions E^L and E^Q, respectively, are expressed as E^L=√(Δ^2+Δ k^2), E^Q=√(Δ^2+k^4). (Here, Δħ^2/m=1 was set for simplicity.) Each off-diagonal mass is expressed as follows: α_xy^L=-Δ^2 k_xk_y/(Δ^2+Δ k^2)^3/2 , α_xy^Q=-4k^2k_xk_y/(Δ^2+k^4)^3/2+4k_xk_y/(Δ^2+k^4)^1/2 . With linear dispersion, only the negative term appeared. Whereas, with quadratic dispersion, both negative and positive terms were found to coexist, thus offsetting each other. Therefore, the magnitude of the negative MR for the Dirac electrons (linear) was larger than that for the DKK model (quadratic). Further, nodal line semimetals exhibited linear dispersion for one direction (k_z) and quadratic dispersion for two directions (k_x, k_y). Consequently, the nodal line semimetals exhibited behaviour in-between that of the Dirac electrons and the DKK model. §.§ Comparison with experiments A primary finding of the study is that the Dirac electrons exhibited significant negative TMR. However, typical Dirac electron systems, such as Bi and PbTe, have not been experimentally observed to exhibit negative TMR <cit.>. The source of the discrepancy between theory and experiment is the multi-carrier nature. Bi and PbTe have three and four Dirac valley at the L points in their Brillouin zone, respectively<cit.>. To examine the manner in which the results of this study change in multi-carrier systems, the calculations performed on the Dirac electrons were extended to the four-valley system, which is an effective model for PbTe <cit.>. Four valleys of Dirac electrons were considered, where the long-axis of each Fermi surface was oriented along the four body-diagonal directions of the cubic (cf. the inset of Fig. <ref>). The energy dispersion of the four-Dirac-valley model is expressed as E_i^multi=√(Δ^2+Δħ^2/m^*k·Â_i·k) Â_1= [ 1 -β -β; -β 1 -β; -β -β 1 ] ,Â_2= [ 1 -β β; -β 1 β; β β 1 ] , Â_3= [ 1 β β; β 1 -β; β -β 1 ] ,Â_4= [ 1 β -β; β 1 β; -β β 1 ]. where β in Â_i is an anisotropic parameter and acquires the value of 0≤β≤ 1/2. Further, it is fully isotropic at β=0 and fully anisotropic (i.e., a cylindrical Fermi surface) at β=1/2. The total MR in the four-Dirac-valley model is obtained as ρ̂_multi=(∑_i^4σ̂_i)^-1. Figure. <ref> shows the TMR calculated in the four-valley model. When only one Dirac carrier was considered, its MR exhibited a negative TMR. However, when simultaneously considering four Dirac carriers, the total MR became a positive TMR. This transformation from negative to positive TMR can be understood on the basis of the off-diagonal mass. In the multi-ellipsoidal model, the off-diagonal mass becomes finite and positive when the axis of ellipsoid is tilted from the orientation of the electric current <cit.>. This positive off-diagonal mass owing to the multi-carrier contributes positively to the TMR. Consequently, in the multi-Dirac system, the positive off-diagonal mass owing to the multi-carrier nature suppresses the negative off-diagonal mass because of the linear dispersion, thus resulting in the transformation from negative to positive TMR. Therefore, the negative TMR of Dirac electrons should be observed only in the single-carrier Dirac electron systems, where the Dirac electrons are located at the Γ point. The negative TMR has been already observed in p-type Si 60 years ago <cit.>. However, there exists no reasonable explanation for p-type Si exhibiting the negative TMR. Thus, the mechanism of the negative TMR remains a longstanding mystery. The results for the DKK model may entail a reasonable explanation for this mystery. The negative off-diagonal mass of light hole yields negative TMR, and its negative contribution dominates the total TMR because the mobility of the light hole is much larger than that of the heavy hole. Certain nodal-line semimetals exhibit negative LMR and positive TMR <cit.>. However, to the best of our knowledge, no experimental study exhibiting negative TMR has been reported. A reason of this discrepancy between experiments and our theory can be attributed to the sensitivity to the direction of the magnetic field. As shown in Sec. 5, the properties of the TMR change drastically according to the orientation of the magnetic field. Therefore, the negative TMR may be observed via the tuning of the magnetic field direction. Another possible reason for the discrepancy is the multi-carrier nature of actual nodal-line semimetals. The nodal-line semimetals generally have a very complex band structure, and they frequently contain other carriers in addition to the nodal-line carries<cit.>. This multi-carrier nature induces the positive TMR as we have shown for the multi-Dirac model. To comprehensively explain the MR experiments in the actual nodal-line semimetals, their complex band structure must be considered based on the Chambers formula, which will be a future problem. Moreover, if an ideal nodal-line material is discovered, the positive to negative TMR transformation can be observed by controlling the field orientation. § SUMMARY This study showed that the negative TMR can be realized in Dirac electrons, the DKK model, and the nodal-line semimetals based on the semi-classical Boltzmann theory with the Chambers formula. It was emphasized that the negative TMR was induced by the negative off-diagonal mass, considering the detailed band structure and k-dependence of the effective-mass. The effect of the negative off-diagonal mass was enhanced in the case of linear energy dispersion. Further, the proposed negative TMR mechanism easily explained the negative TMR in p-type Si, which has been a long-standing mystery. In addition, in Si and many other systems, the mysterious negative MR properties were found to be explainable by focusing on the negative off-diagonal mass, which could be easily obtained from the density functional theory. Further research is required to clarify the relationship between effective-mass and MR in many other systems and deepen our understanding of negative MR. Moreover, the proposed concept of negative off-diagonal mass is general in nature, and thus, future studies will focus on providing a new route to the negative MR in addition to the chiral anomaly and the Berry curvature. § ACKNOWLEDGEMENTS We would like to thank M. Tokunaga, A. Yamada, K. Behnia, B. Fauqué, and Z. Zhu for the fruitful discussions and helpful comments. This work was supported by JSPS KAKENHI grants 19H01850 and 22K18318. § REFERENCE iopart-num

http://arxiv.org/abs/2303.06985v1

20230313103548

Fermionic quantum processing with programmable neutral atom arrays

quant-ph

Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Department of Physics, Harvard University, Cambridge, MA 02138, USA Department of Physics, Harvard University, Cambridge, MA 02138, USA Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Department of Physics, Harvard University, Cambridge, MA 02138, USA Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Department of Physics, University of Colorado, Boulder, Colorado 80309, USA JILA, University of Colorado and National Institute of Standards and Technology, Boulder, Colorado 80309, USA Department of Physics, Harvard University, Cambridge, MA 02138, USA Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Université Grenoble Alpes, CNRS, Laboratoire de Physique et Modélisation des Milieux Condensés (LPMMC), Grenoble 38000, France Department of Physics, University of Colorado, Boulder, Colorado 80309, USA JILA, University of Colorado and National Institute of Standards and Technology, Boulder, Colorado 80309, USA Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, 6020 Innsbruck, Austria Simulating the properties of many-body fermionic systems is an outstanding computational challenge relevant to material science, quantum chemistry, and particle physics. Although qubit-based quantum computers can potentially tackle this problem more efficiently than classical devices, encoding non-local fermionic statistics introduces an overhead in the required resources, limiting their applicability on near-term architectures. In this work, we present a fermionic quantum processor, where fermionic models are locally encoded in a fermionic register and simulated in a hardware-efficient manner using fermionic gates. We consider in particular fermionic atoms in programmable tweezer arrays and develop different protocols to implement non-local tunneling gates, guaranteeing Fermi statistics at the hardware level. We use this gate set, together with Rydberg-mediated interaction gates, to find efficient circuit decompositions for digital and variational quantum simulation algorithms, illustrated here for molecular energy estimation. Finally, we consider a combined fermion-qubit architecture, where both the motional and internal degrees of freedom of the atoms are harnessed to efficiently implement quantum phase estimation, as well as to simulate lattice gauge theory dynamics. Fermionic quantum processing with programmable neutral atom arrays P. Zoller March 30, 2023 ================================================================== § INTRODUCTION The study of strongly correlated fermionic systems lies at the core of some of the most interesting problems in modern physics. These include profound questions regarding the inner workings of the universe, such as the physics of quark-gluon plasmas <cit.>, as well as technologically pressing challenges in material science and quantum chemistry, from high-temperature superconductivity <cit.> to nitrogen fixation <cit.>. The defining feature of fermionic many-body systems is the fundamental indistinguishability of their constituents, which dictates the anti-symmetry of the wavefunction and the corresponding quantum statistics. Importantly, this indistinguishabilty gives rise to the so-called sign problem, which severely limits the applicability of many numerical approaches, such as Monte Carlo methods, highlighting the innate difficulty of solving fermionic many-body problems on classical computers <cit.>. One of the most promising alternatives to address these problems is provided by quantum computers <cit.>. Traditionally, quantum computing involves distinguishable spin-1/2 particles, where quantum information is stored in superposition states of qubit registers, and processed by the action of quantum gates. The quantum statistics of fermions needs to be encoded in such qubit-based devices on a software level, which incurs overhead in circuit depths <cit.> or qubit numbers <cit.>. This presents a substantial challenge for current experiments where noise limits gate and readout fidelities. Although this approach has been applied to simple fermionic models from quantum chemistry <cit.>, condensed-matter <cit.> and particle physics <cit.>, using quantum processors based on superconducting circuits or trapped ions, experimental studies have so far been restricted to small system sizes. Neutral atom systems provide a route to bypass this issue and construct quantum devices where fermionic statistics are built-in on a hardware level. The natural indistinguishability of atoms, which come as bosons or fermions, is for instance leveraged in celebrated analog quantum simulations of Hubbard models in optical lattices <cit.>. Recently, optical tweezers have emerged as powerful tools to trap and manipulate neutral atoms with an unprecedented level of programmability and scalability <cit.>. So far, these systems have, however, mainly been used to realize spin models with distinguishable constituents <cit.>, where each atomic position is pinned to a specific tweezer, internal electronic or nuclear spin states are used to represent qubit states, and interactions between these qubits are implemented using highly-excited Rydberg states. In this work, we envision the next-generation of such tweezer setups, where not only the internal but also the external degrees of freedom are coherently controlled, and fully integrated in the quantum processing architecture. This is a crucial prerequisite for capitalizing on the indistinguishability of (fermionic) atoms, which requires the possibility for their center-of-mass wave functions to overlap, e.g., by coherently delocalizing atoms across tweezers. Remarkably, this motional control has already been demonstrated in pioneering proof-of-principle experiments with tweezer pairs and double-well potentials <cit.>. Below, we describe a blueprint of the elements required for such a fermionic quantum processor <cit.>, where both quantum hardware and software are co-designed to efficiently simulate fermionic models. More precisely, we describe protocols for the basic set of fermionic quantum gates, including, apart from Rydberg-mediated interacting gates, digital tunneling gates (or `fermionic beam splitters') implemented through Merge or Shuttle protocols. While the present work focuses on a setup with tweezer arrays, we note alternative setups involving optical lattices, as developed originally in the context of bosonic atoms <cit.>. We exemplify our proposal for concrete atomic systems, and discuss the requirements and experimental challenges for their physical implementation. Furthermore, we provide illustrative examples of application of such a fermionic quantum processor in the context of digital quantum simulation for quantum chemistry and for lattice gauge theories (LGT), where we use this fermionic gate set to find efficient circuit decompositions and demonstrate considerable depth reductions compared to traditional qubit-based approaches. § HARDWIRED FERMI STATISTICS We consider fermionic atoms in an array of microtraps that represent the fermionic quantum register [Fig. <ref>(a)]. We write c^_j,σ (c_j,σ^†) for the annihilation (creation) operator of atoms on lattice site j, which we assume to be prepared in the trap ground state, and σ labeling internal atomic states [qudits <cit.>]. The quantum state of the quantum register comprising N atoms occupying L microtraps will thus be a quantum superposition of all possible configurations. To be concrete, we will illustrate the quantum gate set, including in particular digital tunneling gates, for spinless fermions, i.e., dropping the index σ for the moment. In this case, the state of the quantum register is given by a superposition of Fock states |n_1,…,n_L⟩, where n_j=0,1 is the atomic occupation, and ∑_j=1^Ln_j=N. In the last section, we consider again spinful fermions and work with a combined fermion-qubit register to encode more general models, which are not purely fermionic, such as LGTs. In the context of quantum simulation (QS) of many-body systems <cit.>, we are interested in particle-number conserving unitaries acting on this register <cit.>, which can be constructed using the gate set ℬ𝒦 = {e^iπ/4n_i, e^iπ n_in_j, e^iπ/4(c^†_i c^_j + H.c.)}, as shown by Bravyi and Kitaev <cit.>. As we demonstrate in the next section, the circuit depth required to simulate fermionic models can be considerably shortened by considering instead the more general set , where 𝒰^( int)_i, j(θ) ≡ e^-iθ n_i n_j , 𝒰^ (t)_i,j(θ⃗) ≡ e^-i[θ_1/2(e^-iθ_2c^†_i c^_j + H.c.) + θ_3/2(n_i - n_j)] are generalized interaction (int) and tunneling (t) gates, respectively [Fig. <ref>(a)], and θ⃗ = (θ_1,θ_2,θ_3). In the context of qubit-based quantum computation, where single-qubit rotations together with one entangling interaction gate are sufficient to achieve universality <cit.>, fermionic degrees of freedom first need to be encoded into qubits using, e.g., a Jordan-Wigner (JW) transformation <cit.>. Tunneling gates, required to simulate many-body fermionic systems, can be implemented in the case of a JW encoding using 𝒪(L) entangling gates <cit.>. Fermionic atoms trapped in the motional ground state of optical tweezers offer the unique possibility to avoid this overhead by implementing the gate set 𝒢 directly. Specifically, the tunneling gate, 𝒰^ (t)_i,j(θ⃗), can be realized using different approaches that exploit the capability of dynamically rearranging the tweezer positions, two of which we discuss now in more detail. The Merge approach realizes the tunneling gate by temporarily bringing the two tweezers i and j so close together that atoms can tunnel between the corresponding lowest vibrational states [see Fig. <ref>(a)], as it has been demonstrated in recent experiments <cit.>. More generally, one may fully merge (and subsequently separate) the tweezer pair using custom designed merging and splitting protocols. The gate parameters θ⃗ can in either case be completely controlled by the tweezer parameters and the details of the merging protocol, such as the tweezer depths, the time-dependent distance of the tweezer minima, and the duration of this coupling process. In practice, these can be determined via optimal control techniques for given target tunnel parameters θ⃗, and allow gate execution on timescales set by the inverse trapping frequency of the tweezers. This approach is therefore natural for light atoms, such as Lithium, where relatively small trap depths are sufficient for large trap frequencies. The Shuttle approach offers an alternative way to realize the tunnel gate, which is based on the capability to realize state-dependent optical potentials <cit.>, and is thus naturally suited for alkaline-earth atoms. We therefore illustrate this idea for the specific example of ^87 Sr, a fermionic isotope of strontium with nuclear spin I = 9/2 below and in Fig <ref>. The central idea is to use two sets of tweezers: a set of static storage traps, whose occupations define the fermionic register, and a set of transport traps, which serve as a “shuttle” for atoms. Crucially, the wavelength of the storage and transport tweezers is chosen such that they trap two different internal states of an atom, respectively. For instance, for ^87 Sr, one can trap the state ^1 S_0 <cit.> in the storage tweezers, and independently trap the clock state ^3 P_0 <cit.> in transport tweezers, reminiscent of the collisional entangling quantum gate with spin-dependent lattices for bosonic atoms <cit.>. Here, we extend these ideas and design a fermionic shuttle that implements the tunneling gate 𝒰^ (t)_i,j(θ⃗). Importantly, when a storage and transport tweezer overlap and their potential shapes match, atoms can be coherently transferred or coherently split between the two tweezers simply by using laser pulses that implement internal rotations R_i(θ⃗) = e^-i[θ_1/2( cos θ_2 X+ sin θ_2 Y) + θ_3/2 Z], where X, Y and Z are Pauli matrices acting on the atomic subspace spanned by ^1S_0 and ^3P_0. Using this mechanism, one can construct the full tunneling gate, 𝒰^ (t)_i,j(θ⃗), as follows [see also Fig. <ref>(b)]: (1) we first bring a transport tweezer to a storage site i and perform a π-pulse rotation, i.e. , (2) after which we move the transport tweezer to site j. (3) We then perform a second pulse R_j(θ⃗^*), with θ⃗^* = (θ_1,θ_2 + π/2, θ_3), (4), bring the transport tweezer back to site i, and (5) finally undo the initial π-pulse. We note that both the Shuttle and the Merge approach can be fully parallelized. We comment on potential error sources for both approaches in a separate section below. In these setups, the interaction gate 𝒰^ (int)_i,j(θ) is essentially equivalent to a standard qubit entangling gate that has already been implemented for alkaline <cit.> as well as alkaline-earth atoms <cit.> using the Rydberg blockade mechanism. To realize it, we first rearrange the tweezers hosting the modes i and j and bring them to a distance that lies within the Rydberg blockade radius. We then drive the atoms with a laser that couples the atoms internal state to a Rydberg state. Owing to the Rydberg blockade mechanism, properly chosen laser pulses result in a unitary e^iϕ_01(n_i + n_j) + i ϕ_11n_i n_j <cit.>. The phases ϕ_01/11 depend on the shape of the Rydberg laser pulse <cit.> and choosing ϕ_11 = 2ϕ_01 - θ provides the desired interaction gate up to the single-particle phase shift ϕ_01. We note that resonantly coupling the atom to the untrapped Rydberg state could create excitations in the trap, which should be suppressed to preserve the atom's indistinguishability. We note that these effects can be reduced by trapping the atoms also in the Rydberg state <cit.>. Alternatively, one can work in the dressed-Rydberg regime, where decay from the Rydberg state is suppressed and interactions become independent of distance, thus avoiding the repulsive forces between atoms. Below, we will discuss these together with other experimental challenges, including several strategies to mitigate the dominant error sources. § FERMIONIC QUANTUM CIRCUITS Now we employ the set of fermionic gates 𝒢 to construct the quantum circuits required for QS of fermionic systems. Let us first focus on purely fermionic Hamiltonians, which we generalize below to fermion-boson models relevant to high-energy physics. To be explicit, we consider the particle-number-conserving fermionic Hamiltonian: ℋ = ∑_ij h_ij^(1) c^†_i c^_j + ∑_ijkl h_ijkl^(2) c^†_i c^†_j c^_k c^_l with complex parameters h_ij^(1) and h_ijkl^(2). This Hamiltonian is frequently used both in condensed matter <cit.> and quantum chemistry <cit.>, where the indices can denote either the position of electrons in a solid-state crystal or the orbitals of a molecule, respectively. Many QS algorithms use as subroutines unitary operations obtained from exponentiating each term in the Hamiltonian. Apart from the tunneling and interaction gates introduced above, these include density-dependent tunneling (dt) as well as pair-tunneling (pt) processes, 𝒰^ (dt)_ijk(θ_1, θ_2) ≡ e^-i θ_1 (e^-iθ_2 c_i^†n_jc^_k + H.c.), 𝒰^ (pt)_i,j,k,l(θ_1, θ_2) ≡ e^-i θ_1 (e^-iθ_2c_i^†c_j^†c^_k c^_l + H.c.). For both (<ref>) and (<ref>), we derived a specific decomposition in terms of the gate set 𝒢, which are shown in Figs. <ref>(a) and (b), respectively. Both decompositions have a constant circuit depth of 5 gates and are exact and optimal. This provides a clear advantage in terms of depth with respect to the circuit obtained using e.g. JW transformations <cit.>, as well as with approximate decompositions based on the universal ℬ𝒦 set [<ref>]. Moreover, note that the all-to-all connectivity of atom arrays plays a crucial point in the case of quantum chemistry, where the presence of non-local terms in (<ref>) introduces further overhead for architectures with only nearest-neighbor gates <cit.>. These fermionic quantum circuits can then be used as precompiled subroutines, carrying over the polynomial saving of resources to the full QS algorithm. For instance, the real-time evolution under Hamiltonian (<ref>) can be simulated by a first-order Trotter expansion <cit.>, e^-iℋt≈[∏_i,j𝒰^ (t)_i,j(h^(1)_i,jδ t,0,0)∏_i,j,k,l𝒰^ (pt)_i,j,k,l(h^(2)_i,j,k,lδ t,0)]^t/δ t, where δ t is the Trotter time and the local gates can be applied in parallel across the system. Another application are hybrid quantum-classical algorithms such as the variational quantum eigensolver (VQE) <cit.>, used in particular to find the ground-state energy of molecules using near-term devices, where 𝒰^ (pt)_i,j,k,l(θ, π/2) allows us to construct variational states that are both hardware-efficient <cit.> and so-called “chemically-inspired” <cit.>. This includes the disentangled unitary coupled cluster ansatz <cit.>, |ψ(θ⃗)⟩ = U(θ⃗)|ψ_ HF⟩, with U(θ⃗) = ∏_i,α𝒰^ (t)_i,α(θ^(1)_iα, π/2, 0)∏_i,j,α,β𝒰^ (pt)_i,j,α,β(θ^(2)_ijαβ, π/2), where the products run over occupied (i, j) and virtual (α, β) modes with respect to the initial Hartree-Fock product state |ψ_ HF⟩. The energy functional E(θ⃗) = ⟨ψ(θ⃗)|ℋ|ψ⃗(θ⃗)⟩, which is then classically minimized, can be also efficiently constructed in our fermionic processor, by first applying randomized tunneling gates followed by measurements in the occupation basis <cit.>. We illustrate our approach for the LiH molecule. We calculate the energy difference δ E ≡ E_ exp(θ⃗^*) - E_0, where θ⃗^* are the optimal variational parameters that minimize E(θ⃗), E_ exp is obtained by evaluating the latter in the presence of experimental errors and E_0 is the exact ground-state energy [We consider the LiH molecule at a fixed interatomic distance of 1.45 Å in the Born-Oppenheimer approximation, with two electrons and four active orbitals. The Hamiltonian parameters h^(1)_i,j and h^(2)_i,j,k,l are calculated using OpenFermion <cit.>.]. The variational circuit U(θ⃗) is depicted in Fig. <ref>(d), and Fig. <ref>(e) shows the average δ E for random fluctuations in the trapping frequency and tweezer positions (Methods), providing an estimate on the required precision to reach chemical accuracy. Note that δ E can be reduced by lowering the tweezer depth, as well as by increasing the pulse intensity to reduce the implementation time, and the full protocol including tweezer transport can be further improved using optimal control <cit.>. A detailed study for larger molecules using more advanced algorithms such as ADAPT-VQE <cit.>, where the advantage of fermionic processing compared to qubit-based protocols is expected to grow with the number of orbitals, will be included in a separate work <cit.>. § FERMION-QUBIT ARCHITECTURE We now consider spinful fermionic atoms, combining the fermionic register and fermion gates introduced above with a more standard qubit-based architecture. This allows us to encode both qubit ancillas and fermionic modes locally, leading to efficient implementations of more advance QS algorithms such as quantum phase estimation (QPE)<cit.>, as well as to simulate boson-fermion models such as LGTs in a hardware-efficient manner. To be specific, we illustrate the architecture using again Sr as an example. Qubit ancillas can be readily included by considering two hyperfine levels of ^1S_0 [Fig. <ref>(a)], where one level, denoted |1̃⟩, is decoupled from the ^3P_0 clock manifold by a magnetic field, and therefore behaves similar to |0⟩ for gates in 𝒢. Rotations R̃ between the states |1̃⟩ and |1⟩ can be implemented through microwave frequencies [Fig. <ref>(a)], enlarging the gate set to 𝒢̃, including both 𝒢 and R̃. In the fermion-qubit register, the same type of atoms can either encode a qubit ancilla or a fermionic mode. In the case of QPE, the energy of ℋ can be estimated by applying a Trotter time evolution under the terms in ℋ controlled by an ancilla, with a precision that grows with the number of ancillas <cit.>. As an example, we show in Fig. <ref>(a) a fermion-qubit circuit associated to a controlled density-dependent hopping process, and other controlled operations can be decomposed similarly in terms of our fermionic gate set. We note that this decomposition requires the three-body gate C_i𝒰^ (int)_j,k(π) = |1̃⟩⟨1̃|⊗𝕀 + |1⟩⟨1|⊗𝒰^ (int)_j,k(π), which can be directly implemented using the Rydberg blockade mechanism <cit.>. Finally, the fermion-qubit architecture allows us to go beyond purely fermionic models, and consider for instance LGTs, where fermions are coupled to dynamical (bosonic) gauge fields. Consider for simplicity a ℤ_2 LGT described by the Hamiltonian ℋ = λ_E∑_⟨ x, y ⟩σ^x_⟨ x, y ⟩ + λ_B∑_□σ^z_□ +λ_J∑_⟨ x, y ⟩(c^†_xσ^z_⟨ x, y ⟩c^_y + H.c.) + λ_m ∑_x (-1)^s_xn_x, where fermion and spin operators, representing matter and gauge degrees of freedom, respectively, act on the sites x and links ⟨ x, y ⟩ of a D-dimensional lattice. The first row in (<ref>) contains the pure-gauge dynamics, including four-body plaquette operators σ^z_□ acting with σ^z_⟨ x, y ⟩ on each link around a plaquette □ [Fig. <ref>(b)], while the second one includes gauge-matter interactions. The Hamiltonian is invariant under local ℤ_2 transformation, i.e. ℋ = 𝒱^†_x ℋ𝒱_x ∀ x, with 𝒱_x = (-1)^n_x∏_yσ^x_⟨ x, y ⟩. Apart from serving as simplified models to study fermionic confinement <cit.>, ℤ_2 LGTs emerge in condensed-matter systems <cit.>, displaying strongly-correlated phenomena such as high-Tc superconductivity <cit.>, topological order <cit.> and unconventional dynamics <cit.>. For D>1, the model presents a sign problem away from half-filling and can not be solved efficiently using classical methods. The corresponding real-time dynamics can be efficiently simulated using quantum devices <cit.> through e.g. a first-order Trotter expansion. In our fermion-qubit architecture, the gate set 𝒢̃ allows us to construct Trotter steps [shown in Fig. <ref>(b)] with a constant circuit depth <cit.>, thanks to the local structure of (<ref>), the possibility to parallelize fermionic gates and the lack of JW strings. The above protocol can be generalized to non-abelian gauge fields, required to address the full Standard Model of particle physics, by further extending it to a fermion-qudit architecture <cit.>, which will be presented in detail in a separate work <cit.>. § EXPERIMENTAL CHALLENGES AND CONSIDERATIONS We finish by discussing in more detail some of the experimental challenges that should be overcome to build a fermionic quantum processor, including the main error sources for the gates discussed above, as well as different strategies to minimize them. The experimental setup for our proposed fermionic quantum processor is similar to existing reconfigurable tweezer platforms with high-fidelity Rydberg gates <cit.>; however, the main new challenge will be coherently controlling the motional degrees of freedom. This introduces new pathways for decoherence, primarily coming from leakage out of the fermionic register (i.e., heating to motional excited states) and dephasing from inhomogeneity between different tweezer sites. Leakage, or heating, out of the motional ground state can arise from state preparation errors, scattering by the tweezer, pulsing the traps off for the Rydberg gate, and from moving the atoms; we argue here that all of these effects can be greatly suppressed. First, heating from tweezer scattering is negligible for sufficiently large detunings. By utilizing tight radial and axial confinement, 3D motional ground state preparation has been realized at levels of ≈ 95% <cit.>. However, by “spilling out” all motional excited states, we can convert the motional ground state occupancy to nearly 100% provided we can nondestructively check for atom presence, which can be done directly using an ancilla atom and a high-fidelity Rydberg gate to check for atom presence <cit.>. Consider for example the 50 kHz trap depth for the Lithium tweezers, relevant for the Merge gate, used in <cit.>, which would give 0.03 Hz scattering rate and an even smaller heating rate due to the Lamb-Dicke parameter. Pulsing the trap off during the Rydberg gate could cause heating, but the probability of transitioning to higher motional states will be roughly (ω t)^2 / 4, where ω = 2π× 15 kHz is the trap frequency <cit.> and t ≈ 100 ns is the implementation time required for 99.9% fidelity gates. This probability then evaluates to < 10^-4 per gate, and so does not contribute significantly compared to a 99.9% gate fidelity. Finally, moving the atoms can cause heating and, as a conservative estimate, we utilize the heating rates calculated in <cit.>. If the atoms are placed relatively close together, at a distance of several micrometers, then for the move to be 99.9-99.99% fidelity each move has to be ≈ 500  μs. This can be significantly speed-up using optimal control <cit.>. In general, we expect that trap inhomogeneities will be the dominant source of dephasing for degrees of freedom coherently encoded in motional states. With a Rydberg gate fidelity of ∼ 99.9% we would like to perform ∼ 1000 operations, where in general the tweezer geometry is reconfigured between each round. If it takes ∼ 500 μs to move atoms between gates, that sets the total operation time of ∼ 500 ms. For the example of Lithium-6, a trap depth of 50 kHz and a reasonable standard deviation between tweezers of 0.2% leads to a coherence time of T^*_2 ≈ 2 ms, which is only enough time for ≈ 4 moves. Note that this is for shallow trap depths of Lithium-6, a light atom, and that this effect will be even more exacerbated for heavier atoms like strontium, where larger trap depths (laser intensities) are required for the same trap frequencies. Whenever there is static inhomogeneity in a system, such as positional trap-depth dependence, the natural approach is to perform a motional echo procedure <cit.>. We can permute around the various tweezer positions so that all atoms acquire the same average phase due to the positional trap depth dependence. First, consider two tweezers with unequal trap intensities and thus unequal energies at the bottom of the trap, as depicted in Fig. <ref>(a). Due to the different energies, the atoms at these sites experience dephasing. However, this effect can be canceled out by repeatedly swapping the tweezer positions after each application of the tunneling gate (see Methods for details). In Fig. <ref>(b), we illustrate this idea by simulating the Floquet evolution of a 1D chain with nearest-neighbor hopping (Methods). The results show that performing the echo procedure allows us to extend the useful simulation time by two orders of magnitude. In practice, the inhomogeneities might not be exactly static, in which case the echo needs to be applied at a rate faster than that timescale. We also note that there could potentially be other methods to design robust sequences of tunnel gates. For example, circuits could be precompiled in order to minimize the number of necessary swap operations while taking into account the specific distribution of spatial coherence. In addition, trap inhomogeneities could be further reduced by storing the atoms in state-dependent optical lattices in combination with tweezers, which could be used to move the atoms to the desired position in order to apply the corresponding gates <cit.>. In this architecture, the Merge gate could be implemented using superlattices, by first placing the atoms in double wells <cit.>, while the Shuttle gate could be implemented by using the lattice and tweezers as storage and transport potentials, respectively. Finally, for the Shuttle gate based on alkaline-earth atoms like strontium as discussed above, additional error sources include phase shifts and losses during step (3) of the protocol due to elastic and inelastic collisions, respectively, between atoms in the ^3 P_0 and ^1 S_0 states. The former are of the order of a few kHz for ^87Sr <cit.>, and should be removed by an appropriate calibration or measured and compensated by additional phase shift gates. Inelastic collisions are even smaller <cit.> than typical Rabi frequencies and can therefore be safely neglected. Finally, we note that the optical potentials for the storage and transport tweezers need to match to properly transfer atoms between them, and fluctuations in the tweezer location and laser intensity will lead to imperfect transfer processes (Methods). The effect of such fluctuations is taken into account in the variational preparation of a simple molecule, illustrated in Figure <ref>. § CONCLUSIONS AND OUTLOOK We presented a fermionic quantum processor based on fermionic atoms trapped in tweezer arrays, and showed how to locally encode and quantum simulate fermionic models in a hardware-efficient manner. We illustrated the advantages of our approach with respect to qubit-based devices for VQE and QPE in quantum chemistry, as well as for Trotter time evolution of gauge theories. We note that the proposed hardware can be used to run more advance and potentially new fermionic quantum algorithms, further optimizing the required resources and facilitating the reach of quantum advantage in the near term. The fermionic gate set could also be extended to include particle non-conserving processes, using e.g. atom reservoirs, allowing to implement error correction protocols, which we plan to investigate in the future. § METHODS §.§ Error model for VQE We exemplify the fermionic quantum simulation of the LiH molecule using VQE, where we consider in particular the tunneling gate implemented using the Shuttle protocol in strontium. We consider optical tweezers generated by laser fields with the following intensity profile in the radial r and longitudinal z directions, I(r, z) = I_0(w_0/w_z)e^-2(r / w_z)^2, with w_z = w_0√(1 + (z / z_R)^2), where w_0 and z_R = π w_0^2/λ are the waist and Rayleigh length of the tweezer, respectively, and λ is the laser (trapping) wavelength. For a properly chosen wavelength, the latter gives rise to an AC Stark shift on a neutral atom, leading to an optical potential V(r, z) = - Re(α) / (2ϵ_0 c) I(r, z), where α is the atom polarizability for the corresponding energy level, ϵ_0 is the permittivity of free space, and c is the speed of light. The trapping potential V(r, z) can be approximated around its minimum using a second-order Taylor expansion, leading to the following harmonic potential, V(r, z) = V_0 + 1/2 m ω_r^2 r^2 + 1/2 m ω_z^2 z^2, where V_0 ≡ V(0, 0), m is the mass of the atom and the frequencies are given by ω_r = √(4V_0/(mw_0^2)) and ω_z = √(2V_0/(mz_R^2)). For typical experimental parameters these frequencies are related by ω_r /ω_z ≈ 10. Here we consider random fluctuations both in the frequencies ω_r and ω_z and in the relative position between tweezers, given by δ r and δ z, following all these parameters independent Gaussian distributions. These experimental fluctuations introduce errors in the fermionic quantum gates, since they lead to imperfect rotations between the storage (S) and transport (P) tweezers. In particular, both the Rabi coupling and the detuning of a give pulse are modified due to the imperfect overlap between the wavefunctions, Ω(δ r, δ z) = Ω_0 f(δ r, δ z) and Δ(δ r, δ z) = Δ_0 f(δ r, δ z), with f(δ r, δ z) = ∫ d r d z 2π r ψ^*_S(r, z) ψ^_P(r +δ r, z + δ z), where we consider Gaussian wavefunctions for the ground state of the harmonic potential (<ref>), ψ_S/P(r, z) = e^- (r / 2r_ zp)^2/(2 π r_ zp)^1/4e^- (z / 2z_ zp)^2/(2 π z_ zp)^1/4, with zero-point fluctuations given by r_ zp = 1 / √(2mω_r) and z_ zp = 1 / √(2mω_z). Moreover, fluctuations in the trap frequency lead to a non-zero difference between the trap depths, δ V_0 = V^S_0 - V^P_0, introducing an extra unwanted detuning, Δ̃ = Δ + δ V_0, giving rise to an extra angle δ V_0 τ for Z rotations, where τ is the pulse time. §.§ Motional echo scheme The positional trap-depth dependence leading to a static dephasing can be mitigated by an appropriate echo procedure. Due to the positional nature of this noise, the natural approach is to ensure that each fermionic register spends equal time at each tweezer site and, on average, experiences the same disorder pattern. The key insight is that, once the evolution is digital, the atom i does not need to always reside in the tweezer site i. That is, we can permute the atoms around the various tweezer positions either by moving them around or introducing swap operations, which effectively change which atom resides in which tweezer. As described in the text, the dephasing between the two tweezers coming from different trap depths can be canceled out by repeatedly swapping the tweezer positions after each application of the tunneling gate. In a many-body context, the atoms need to be shuffled in an ergodic fashion such that they can spend equal time at every tweezer site; since this process is deterministic it is easy to keep track of the orbital labels in the classical experimental software. This approach is particularly simple in a 1D chain with periodic boundary conditions [Fig. <ref>(b) inset]. Here, the two sublattices of tweezers, labeled with even and odd numbers, will sequentially translate by one site in space after every tunneling gate. This results in each orbital spending the same amount of time at each position and, moreover, experiencing the same history of trap depths as the neighboring site, up to boundary conditions, which further improves the echo performance. More concretely, during the t-th timestep, the i-th atom resides in the σ_t(i)-th tweezer. Since the inhomogeniety is (to first order) static, but varies from tweezer to tweezer, this can be modelled via a time-dependent disorder Hamiltonian D_t = ∑_x h_σ_t(x)n̂_x. The effective dynamics is captured by alternating disorder evolution and target evolution. Moving into an interaction picture with respect to the disorder, and assuming each moving step takes time τ, the hopping terms c_i^† c_j become dressed as (c_i^† c_j)(t) = e^-i τ∑_t' ≤ t (h_σ_t'(j) - h_σ_t'(i)) c_i^† c_j, where j = i+1, For the cyclic-shift strategy depicted in the inset of Fig. <ref>(b) we have σ_t(x) = (x - t) L, where L is the length of the chain. Here the noise between the two sites becomes time-correlated (between Floquet rounds), σ_t+1(i+1) = σ_t(i) and the relative accumulated disorder is ∑_t' ≤ t (h_σ_t'(i+1) - h_σ_t'(i)) = h_σ_0(i+1) - h_σ_t(i), which implies that the phase noise on the hopping term t_ij remains bounded for all time. Compare this to the situation when no echo is performed where the relative phase grows linearly in time. We note the special nature of this echo procedure, which is intrinsically linked to spatial positions of the traps and cannot be performed by applying global operations, in contrast to the typical echo sequences present in spin systems. A similar effect occurs in the disorder cancellation strategy using swaps, since if two sites x_1 and x_2 interact at time-step t, then σ_t(x_1) = σ_t+1(x_2) and σ_t(x_2) = σ_t+1(x_1). The results shown in Fig. <ref>(b) were obtained by simulating 100 tweezers with 20 atoms in a 1D ring configuration evolving under a nearest-neighbor tunneling Hamiltonian with the hopping rate J=1 and the time step τ=0.13, using free-fermion methods. The Hamiltonian evolution is split into two parallel Floquet rounds and a Gaussian phase disorder with σ_θ = 0.035 is applied between each round. We thank I. Bloch for comments on the manuscript, and discussion of the advantages of hybrid tweezer-optical lattice platforms. Work at Innsbruck was supported by the US Air Force Office of Scientific Research (AFOSR) via IOE Grant No. FA9550-19-1-7044 LASCEM, the European Union’s Horizon 2020 research and innovation program under Grant Agreement No. 101079862 (PASQuanS2), by the Simons Collaboration on Ultra-Quantum Matter, which is a grant from the Simons Foundation (651440, P.Z.), and by ERC Starting grant QARA (101041435, H.P.). Work at JILA is supported by NSF QLCI OMA-2016244 and NIST. B.V. and P.N. acknowledge funding from the Austrian Science Foundation (FWF, P 32597 N). Work at Grenoble is supported by the French National Research Agency via the JCJC project QRand (ANR-20-CE47-0005), and via the research programs EPIQ (ANR-22-PETQ-0007, Plan France 2030), and QUBITAF (ANR-22-PETQ-0004, Plan France 2030). Work at Harvard was supported by the US Department of Energy (DE-SC0021013 and DOE Quantum Systems Accelerator Center, contract number 7568717), the National Science Foundation, and the Center for Ultracold Atoms. D.B. acknowledges support from the NSF Graduate Research Fellowship Program (grant DGE1745303) and The Fannie and John Hertz Foundation. N.M. acknowledges support from the Department of Energy Computational Science Graduate Fellowship under Award Number DE-SC0021110.

http://arxiv.org/abs/2303.07371v1

20230313180004

Probing massless and massive gravitons via entanglement in a warped extra dimension

gr-qc

apsrev4-1

http://arxiv.org/abs/2303.06814v1

20230313021433

Implementing static Dyson-like spheres around spherically symmetric black hole spacetime

gr-qc

^1Lanzhou Center for Theoretical Physics, Key Laboratory of Theoretical Physics of Gansu Province, School of Physical Science and Technology, Lanzhou University, Lanzhou 730000, People's Republic of China, ^2Institute of Theoretical Physics & Research Center of Gravitation, Lanzhou University, Lanzhou 730000, People's Republic of China, ^3 Department of Physics & Astronomy, University of Waterloo, Waterloo, Ont. Canada N2L 3G1 Unique features of particle orbits would produce novel signatures of gravitational observable phenomena, and are quite useful in testing compact astrophysical objects in general relativity or modified theories of gravity. In this work, we observe a representative example that static, spherically symmetric dyonic black hole solution admit static points at finite radial distance. Considering the spherical symmetry, each static point shall produce a static sphere, which can be used to implement the well-known Dyson spheres aiming to collect the infrared radiation energy of the stellar sources. More interestingly, employing with the topological argument, we disclose that, the stable and unstable static spheres always come in pairs in an asymptotically flat spacetime if they exist. While the counterpart naked singularity has one more stable static sphere than the unstable one. Our results have potential application in constructing the Dyson spheres around the stellar sources. 04.20.-q, 04.25.-g, 04.70.Bw Implementing static Dyson-like spheres around spherically symmetric black hole spacetime Shao-Wen Wei^1,2 [weishw@lzu.edu.cn], Yu-Peng Zhang^1,2 [zyp@lzu.edu.cn], Yu-Xiao Liu^1,2 [liuyx@lzu.edu.cn], Robert B. Mann^3 [rbmann@uwaterloo.ca] ======================================================================================================================================================== § INTRODUCTION Exploring the nature of a black hole or spacetime in both weak and strong gravity regimes largely relies on the characteristic geodesic motions of test particles. Some characteristic phenomena, such as the ringdown of black hole binaries <cit.> and shadows <cit.>, can also be understood by the circular orbits of photons or massive particles. Further study also shows that semi orbits, pointy petal orbits, and static light points appear in rotating boson stars, hairy black holes, wormholes <cit.>, and higher-dimensional rotating black holes <cit.>, expecting to have potential signatures on observable effects. Recent developments can also be found in Refs. <cit.>. Another intriguing object that has attracted a lot of attention is the Dyson sphere, proposed by Freeman Dyson <cit.>, which was initially constructed around a star and designed to exploit all the infrared radiation energy of stellar sources. Construction of an actual and rigid Dyson sphere is extremely hard to implement; for example, the required elastic modulus of the material is very high <cit.>, exceeding that of the strongest known material, carbyne, by nine orders of magnitude <cit.>. Noting that the spacetime exterior to a stellar source can be described by a black hole metric, here we aim to investigate Dyson-like spheres around black holes. For a class of rotating hairy black holes, it has been shown <cit.> that there is a static point at which a particle can remain at rest. Such an orbit, however, is limited to the equatorial plane of the rotating black hole. Thus only a thin static belt (see left panel of Fig. <ref>) could be formed around this object, with almost all the (Hawking) radiation energy leaking to infinity without being collected. Since the spin of a stellar source is usually small, we shall proceed to construct the Dyson-like sphere around a spherically symmetric black hole. Due to the spherical symmetry, a matter shell can be easily formed with an equatorial circular orbit described by the red circle shown in the right panel of Fig. <ref>. In order to maintain a rotating Dyson-likesphere, this configuration also has its limitations. For example, when one considers the matter shell rotating with the same angular momentum as the equatorial circular orbit, the matter component undergoes non-geodesic motion at non-equatorial latitudes (the blue curves in Fig. <ref>), and thus experience an inner stress density. If its value is beyond the maximal stress tolerance of the material, the matter shell of the Dyson-like sphere will break up. In order to solve this problem, we need to keep the angular momentum of the equatorial circular orbit sufficiently small. This can be done for a static sphere at radius r_sp with vanishing angular velocity Ω(r_sp)=0 in a nonrotating black hole background. Then each point on this sphere is a particular geodesic, the static point. The internal stress thus disappears and an actual Dyson-like sphere can be firmly formed on this static sphere for any material. In general relativity, minimally coupled to Standard Model matter, such a sphere is absent. We shall show that it indeed exists in nonlinear electrodynamics, and always comes in pairs. § STATIC SPHERES AND STRAIGHT ORBITS Let us consider a static, spherically symmetric black hole with the following line element ds^2=-f(r)dt^2+1/f(r)dr^2+r^2(dθ^2+sin^2θ dϕ^2), where f(r) is the metric function and is only dependent of the radial coordinate r. Without loss of generality, one needs only to focus on the equatorial geodesics with θ=π/2 for the study of geodesic motion of a test particle. With the two Killing vectors for the black hole background (<ref>), ξ^μ=(∂_t)^μ and ψ^μ=(∂_φ)^μ, one can obtain two conserved quantities -E = g_μνu^μξ^ν=g_ttṫ, l = g_μνu^μψ^ν=g_φφφ̇, which are respectively the energy and orbital angular momentum per unit mass of a massive test particle, along each geodesic with tangent vector u^μ; the dots denote the derivative with respect to an affine parameter. For a massive test particle, we have g_μνẋ^μẋ^ν=-1, which leads to -f(r)ṫ^2+1/f(r)ṙ^2+r^2ϕ̇^2=-1. With the above three equations (<ref>)-(<ref>), one easily obtains ṫ=E/f(r), ϕ̇=l/r^2, ṙ^2=-V_eff, where the effective potential is given by V_eff = ( l^2/r^2 + 1 ) f(r) - E^2. For a distant static observer at rest, the angular velocity of a massive test particle circling the black hole is Ω=ϕ̇/ṫ=f(r)l/r^2E. Outside the black hole horizon, f(r)>0. So vanishing Ω requires l=0, as expected. The reduced effective potential becomes 𝒱_eff=f(r)-E^2. The static sphere is also a specific circular sphere, which satisfies 𝒱_eff=∂_r𝒱_eff=0. As a result, the following conditions should be satisfied: E=√(f(r_sp)), f'(r_sp)=0, where the prime denotes the derivative with respect to r, and r_sp is the radius of the static sphere. Note that f”(r_sp)>0 and <0 correspond to stable and unstable static spheres, respectively, and so the former is of particular interest. Moreover, for the well known Schwarzschild and Reissner-Nordström black hole, there is no static sphere outside the black hole horizon. We now investigate whether there are static spheres for other static spherically symmetric black holes. We consider the dyonic black hole <cit.>, which is a solution of the field equations that follow from the action S=1/16π∫√(-g)d^4x(R-α_1F^2-α_2 ((F^2)^2-2F^(4))), where the field strength is F^2=-F^μ_νF_μ^ν and F^(4)=F^μ_νF^ν_ρF^ρ_σF^σ_μ. The coupling parameters α_1 and α_2 are for the standard Maxwell and quasi-topological electromagnetic actions, respectively. Under the ansatz of global polarization, this quasi-topological term has no influence on the Maxwell equation and the energy-momentum tensor. We shall see that particular property yields interesting results. The line element of a static and spherically symmetric black hole is given in the form (<ref>) with <cit.> f(r)=1-2M/r+α_1 p^2/r^2 +q^2/α_1r^2 _2F_1 [1/4,1; 5/4; -4p^2α_2/r^4α_1], where p and q are the electric and magnetic charge parameters. In certain parameter regions, for example, p/M=0.14, q/M=1.02, α_1=1, α_2/M^2=2.76. there are four black hole horizons and three photon spheres, indicating an interesting spacetime structure. In this case, outermost horizon of the black hole is located at r_h/M=0.25. Plotting the reduced effective potential (<ref>) in Fig. <ref> with parameter values given in (<ref>), we observe one stable and one unstable static sphere at r_sp1 and r_sp2 for fixed energy E_sp1 and E_sp2 respectively. For each energy E∈ (E_sp1, E_sp2), there is a bound orbit. A neutral massive test particle will undergo straight back-and-forth motion between two turning points, for example r_1 and r_2 for E=0.25. Note that there is no angular motion due to the vanishing angular velocity. Taking energies to be E=0.2, 0.15, 0.14, and E_sp1=0.1393, we exhibit the radial motion of the particle starting at the small turning points in the r-t plane in Fig. <ref>. Obviously, when the energy decreases and tends to E_so1, the radial region of the motion narrows. In particular, when E=E_sp1 shown in Fig. <ref>, the radial distance remains unchanged with coordinate time, indicating there is a static sphere at r=r_sp1. In summary, we observe two static spheres – one stable and one unstable – for the static, spherically symmetric dyonic black hole. Thus a Dyson-like sphere can be appropriately constructed for this nonrotating black hole. § NUMBER OF STATIC SPHERES We now seek to determine the number of the static spheres in the static, spherically symmetric black hole (<ref>). The approach was mainly developed in Refs. <cit.> for the light ring (see also <cit.>), and was then generalized to timelike circular orbits <cit.>. In order to satisfy the conditions for a static sphere given in (<ref>), we construct the vector ϕ=(ϕ^r, ϕ^θ) with ϕ^r=∂ f(r)/∂ r, ϕ^θ=-cosθ/sin^2θ. Note that ϕ^θ is an auxiliary term and whose explicit form is influenced by the location of the static sphere. Obviously a static sphere is exactly located at a zero point of ϕ. For any given static sphere we can calculate its winding number, where a positive (negative) value signifies a stable (unstable) static sphere. However, here we are concerned with the sum of the winding numbers, namely the total topological number W of the static spheres corresponding to the following topological current <cit.> j^μ=1/2πϵ^μνρϵ_ab∂_νn^a∂_ρn^b, μ,ν,ρ=0,1,2, where ∂_ν=∂/∂ x^ν and x^ν=(τ, r, θ) and the unit vector is defined as n^a=ϕ^a/||ϕ|| (a=1, 2). The parameter τ is a time control parameter. It is not hard to check that this current is conserved: ∂_μj^μ=0. To determine W, we need to examine the behavior of the vector ϕ at the boundary of the r-θ plane. It is easy to find that at θ=0 and π the direction of ϕ is outwards. Near the horizon r=r_h, f(r_h)=0 and f(r>r_h)>0; thus ϕ^r is positive, which indicates the direction of ϕ is rightward in the plane, ignoring the specific values of ϕ^θ. At large r, for an asymptotically flat black hole f(r)∼ 1-2M/r+𝒪(1/r^2) and so f'(r→∞)>0. This suggests that the direction of ϕ is also towards the right. Combining these results, we easily obtain the total topological number W=∑_i w_i=0, where w_i denotes the winding number of the i-th zero point of ϕ. For asymptotically flat, static, spherically symmetric black holes, this result strongly suggests that if static spheres exist, they always come in pairs. We shall see that this result changes for naked singularities. As shown in Ref. <cit.>, the black hole charge can be treated as a time control parameter. In order to study the evolution of the static sphere radius as the time control parameter, we choose the parameter values in (<ref>). For q≤ q_2 and q>q_2, the metric (<ref>) respectively describes a dyonic black hole and naked singularity, with f(r) given in (<ref>). The behavior of the radii of static spheres is shown in Fig. <ref>. From the figure, we find that two static spheres emerge only when q≥ q_1 for the black hole. The one with small radius has a negative winding number and so is unstable, whereas the one with large radius has a positive winding number and is stable. The total topological number vanishes for q<q_1, where static spheres are absent. This property continues for all q_1<q<q_2. At extremality, with q=q_2, a new stable static sphere with smaller radius emerges. For q> q_2 the solution describes a naked singularity, and there are three static spheres; the topological number is W=1-1+1=1, notably different from the situation for a black hole. This situation persists for q_2<q<q_3. For q → q_3, the two smaller spheres merge, leaving only one static sphere of large radius for q > q_3. Nevertheless, the topological number W=1 remains unchanged. We summarize our results in Table <ref>. The topological number W clearly has distinct values for the black hole and naked singularity, indicating they are in different topological classes. § SUMMARY We have investigated the construction of Dyson-like spheres about a static, spherically symmetric dyonic black hole and its naked singularity counterpart. We find that the black hole admits a pair of stable and unstable static spheres, at which a massive particle remains at rest with respect to a static asymptotic static observer. Making use of a topological argument for static spheres in an asymptotically flat spacetime, we showed the black hole and naked singularity have distinct topological numbers (W=0 and 1 respectively) for their static spheres. These results confirm that the stable and unstable static spheres always come in pairs for asymptotically flat black holes. It also indicates that the black hole and naked singularity solutions belong to different topological classes. Moreover, if one the naked singularity solution can be the exterior metric of some stellar structure, there would be a topological phase transition at q=q_2 for the static spheres. Besides realizing the static Dyson-like sphere, the presence of static orbits could result in certain novel gravitational phenomena. For example, there may be one extra static or slow velocity accretion disk, in contrast to those formed at the usual innermost stable circular orbits. Since this orbit, which acts as a light source, is closer to the black hole, it will produce a different shadow pattern. Such double accretion disks will yield interesting observable effects, providing unique tests of general relativity. § ACKNOWLEDGEMENTS This work was supported by the National Natural Science Foundation of China (Grants No. 12075103, No. 12105126, No. 11875151, and No. 12247101), the 111 Project (Grant No. B20063), Lanzhou City's scientific research funding subsidy to Lanzhou University, and the Natural Sciences and Engineering Research Council of Canada. 99 Abbott B. P. Abbott et al. (Virgo, LIGO Scientific), Observation of Gravitational Waves from a Binary Black Hole Merger, Phys. Rev. Lett. 116, 061102 (2016), [arXiv:1602.03837 [gr-qc]]. Akiyama1 K. Akiyama et al. (Event Horizon Telescope), First M87 Event Horizon Telescope Results. I. The Shadow of the Supermassive Black Hole, Astrophys. J. 875, L1 (2019), [arXiv:1906.11238 [astro-ph.GA]]. Akiyama2 K. Akiyama et al. (Event Horizon Telescope), First Sagittarius A* Event Horizon Telescope Results. I. The Shadow of the Supermassive Black Hole in the Center of the Milky Way, Astrophys. J. Lett. 930, L12 (2022). Grandclement P. Grandclement, C. Som, and E. Gourgoulhon, Models of rotating boson stars and geodesics around them: new type of orbits, Phys. Rev. D 90, 024068 (2014), [arXiv:1405.4837 [gr-qc]]. Grould M. Grould, Z. Meliani, F. H. Vincent, P. Grandclement, and E. Gourgoulhon, Comparing timelike geodesics around a Kerr black hole and a boson star, Class. Quant. Grav. 34, 215007 (2017), [arXiv:1709.05938 [astro-ph.HE]]. Collodel L. G. Collodel, B. Kleihaus, and J. Kunz, Static spheres in rotating spacetimes, Phys. Rev. Lett. 120, 201103 (2018), [arXiv:1711.05191 [gr-qc]]. Teodoro M. C. Teodoro, L. G. Collodel, J. Kunz, Retrograde Polish Doughnuts around Boson Stars, JCAP 03, 063 (2021), [arXiv:2011.10288 [gr-qc]]. Yazadjiev M. C. Teodoro, L. G. Collodel, D. Doneva, J. Kunz, P. Nedkova, and S. Yazadjiev, Thick toroidal configurations around scalarized Kerr black holes, Phys. Rev. D 104, 124047 (2021), [arXiv:2108.08640 [gr-qc]]. Gibbons G. W. Gibbons and C. A. R. Herdeiro, Supersymmetric rotating black holes and causality violation, Class. Quant. Grav. 16, 3619 (1999), [arXiv:hep-th/9906098]. Herdeiro C. A. R. Herdeiro, Special properties of five-dimensional BPS rotating black holes, Nucl. Phys. B 582, 363 (2000), [arXiv:hep-th/0003063]. Diemer V. Diemer and J. Kunz, Supersymmetric rotating black hole spacetime tested by geodesics, Phys. Rev. D 89, 084001 (2014), [arXiv:1312.6540 [gr-qc]]. Delgado J. F. M. Delgado, C. A. R. Herdeiro, and E. Radu, Equatorial timelike circular orbits around generic ultracompact objects, Phys. Rev. D 105, 064026 (2022), [arXiv:2107.03404 [gr-qc]]. Lehebel A. Lehebel and V. Cardoso, The fate of observers in circular motion, Phys. Rev. D 105, 064014 (2022), [arXiv:2202.08850 [gr-qc]]. Dyson F. J. Dyson, Search for Artificial Stellar Sources of Infrared Radiation, Science 131, 1667 (1960). Papagiannis M. D. Papagiannis, The Search for Extraterrestrial Life: Recent Developme, IAU Symposium 112, 543 (1985). Wright J. T. Wright, Dyson Spheres, Ser. Astr. J. 200, 1 (2020), [arXiv:2006.16734 [astro-ph.EP]]. Liu2020 H.-S. Liu, Z.-F. Mai, Y.-Z. Li, and H. Lü, Quasi-topological Electromagnetism: Dark Energy, Dyonic Black Holes, Stable Photon Spheres and Hidden Electromagnetic Duality, Sci. China Phys. Mech. Astron. 63, 240411 (2020), [arXiv:1907.10876 [hep-th]]. Cunhaa P. V. P. Cunha, E. Berti, and C. A. R. Herdeiro, Light ring stability in ultra-compact objects, Phys. Rev. Lett. 119, 251102 (2017), [arXiv:1708.04211 [gr-qc]]. Cunhab P. V. P. Cunha and C. A. R. Herdeiro, Stationary Black Holes and Light Rings, Phys. Rev. Lett. 124, 181101 (2020), [arXiv:2003.06445 [gr-qc]]. Weisw S.-W. Wei, Topological Charge and Black Hole Photon Spheres, Phys. Rev. D 102, 064039 (2020), [arXiv:2006.02112 [gr-qc]]. Guo M. Guo and S. Gao, Universal Properties of Light Rings for Stationary Axisymmetric Spacetimes, Phys. Rev. D 103, 104031 (2021), [arXiv:2011.02211 [gr-qc]]. Junior H. C. D. L. Junior, J.-Z. Yang, L. C. B. Crispino, P. V. P. Cunha, and C. A. R. Herdeiro, Einstein-Maxwell-dilaton neutral black holes in strong magnetic fields: topological charge, shadows and lensing, Phys. Rev. D 105, 064070 (2022), [arXiv:2112.10802 [gr-qc]]. Weila S.-W. Wei and Y.-X. Liu, Topology of equatorial timelike circular orbits around stationary black holes, Phys. Rev. D 107, 064006 (2023), [arXiv:2207.08397 [gr-qc]]. Duan Y. S. Duan, The structure of the topological current, Stanford Linear Accelerator Center (SLAC), Report No. SLACPUB- 3301, (1984).

http://arxiv.org/abs/2303.07038v1

20230313115716

Compositeness of $T_{cc}$ and $X(3872)$ with decay and coupled-channel effects

hep-ph

[]kinugawa-tomona@ed.tmu.ac.jp Department of Physics, Tokyo Metropolitan University, Hachioji 192-0397, Japan []hyodo@tmu.ac.jp Department of Physics, Tokyo Metropolitan University, Hachioji 192-0397, Japan The compositeness of weakly bound states is discussed with the effective field theory from the viewpoint of the low-energy universality. We introduce a model with the coupling of the single-channel scattering to the bare state, and study the compositeness of the bound state by varying the bare state energy. In contrast to the naive expectation that the near-threshold states are dominated by the molecular structure, we demonstrate that a non-composite state can always be realized even with the small binding energy. At the same time, however, it is shown that a fine tuning is necessary to obtain the non-composite weakly bound state. In other words, the probability to find a model with the composite dominant state becomes larger with the decrease of the binding energy in accordance with the low-energy universality. For the application to the exotic hadrons, we then discuss the modification of the compositeness by the decay and coupled-channel effects. We quantitatively show that these contributions suppress the compositeness, because of the increase of the fraction of other components. Finally, as the examples of the near-threshold exotic hadrons, the structure of T_cc and X(3872) is studied by evaluating the compositeness. We find the importance of the coupled-channel and decay contributions for the structure of T_cc and X(3872), respectively. Compositeness of T_cc and X(3872) with decay and coupled-channel effects Tetsuo Hyodo March 30, 2023 ======================================================================== § INTRODUCTION To clarify the internal structure of exotic hadrons is one of the central aims of hadron physics. The recent observations of the candidates of the exotic hadrons in the heavy hadron sectors provide opportunities for intensive studies on the structure of hadrons <cit.>. The exotic hadrons are considered to have different internal structures from the ordinary hadrons with qqq or qq̅ as described in the quark models. It is remarkable that many candidates of the exotic hadrons have been discovered near two-hadron thresholds. For example, the tetraquark T_cc has been observed slightly below the threshold of D^0D^*+ in the T_cc→ D^0D^0π^+ decay by the LHCb collaboration in 2021 <cit.>. Its minimum quark content ccu̅d̅ indicates that T_cc is a genuine exotic state with charm C=+2. As a charmonium-like state with C=0, X(3872) has been observed near the D^0D̅^*0 threshold in the B^±→ K^±π^+π^-J/ψ decay in 2003 by the Belle Collaboration <cit.>. X(3872) is considered to be exotic because its mass is not in accordance with the corresponding energy predicted by the quark model <cit.>. As the possible internal structures of the exotic hadrons, for example, the hadronic molecule states and the multi-quark states are considered. The hadronic molecule state is a loosely bound composite system of hadrons formed by the hadronic interactions, such as the deuteron. In contrast, the multiquark state is the compact state of at least four quarks. To reveal the internal structure of the exotic hadrons, many studies are actively performed from both the theoretical and experimental sides. The molecular nature of the bound state can be quantitatively studied by using the compositeness <cit.>. The compositeness is defined as the probability to find the hadronic molecule component in the bound state. Theoretically, the compositeness can be evaluated either from the weak-binding relation <cit.> or from the residue of the pole of the scattering amplitude <cit.>. The internal structure of many hadrons has been studied by the compositeness <cit.>. The notion of the compositeness has also been applied to other systems, such as nuclei and atoms <cit.>. The phenomena associated with the near-threshold states are governed by the low-energy universality <cit.>. From the universality argument, it is expected that the near-threshold states are dominated by the molecular component <cit.>. In fact, Ref. <cit.> shows that the s-wave bound states become completely composite in the weak-binding limit. A similar discussion has been given in the cluster phenomena in nuclear physics, such as the ground state of ^8Be and the ^12C Hoyle state <cit.>. From these discussions of the near-threshold states, one may naively expect that T_cc and X(3872) are the composite dominant states. However, the small binding energy is not the only characteristic feature of T_cc and X(3872). First, both T_cc and X(3872) decay strongly and have a finite decay width. Next, the threshold channel [D^0D^*+ for T_cc and D^0D^*0 for X(3872)] has an isospin partner [D^*0D^+ for T_cc and D^*-D^+ for X(3872)] at a slightly higher energy. These features are illustrated in Fig. <ref>. It is shown that these decay and coupled-channel contributions modify the compositeness of the bound state <cit.>. To understand the nature of T_cc and X(3872), we need to quantitatively evaluate the contributions from the decay and channel coupling to the compositeness. In this work, we first demonstrate how the expectation of the molecular nature of the near-threshold states is realized in an explicit model calculation. We show that the shallow bound state can be elementary dominant only with a fine tuning of the model parameter. In most of the parameter region except for the fine-tuned case, the weakly bound state is composite dominant, as expected from the universality. To consider the realistic exotic hadrons, we then examine the effects of the decay and the coupled channel to the compositeness. We quantitatively evaluate the modification of the expectation from the universality by the decay and the coupled-channel effects. We finally apply the model to calculate the compositeness of T_cc and X(3872) to clarify the important effect for these states. This paper is organized as follows. In Sec. <ref>, we introduce the effective field theory, and numerically calculate the compositeness to discuss the nature of the shallow bound state. We then consider the contributions of the four-point contact interaction, the decay and the channel coupling to the compositeness in Sec. <ref>. In Sec. <ref>, we estimate the compositeness of T_cc and X(3872) by focusing on the importance of the decay and the coupled-channel effects. A summary of this work is given in Sec. V. § WEAKLY BOUND STATES AND LOW-ENERGY UNIVERSALITY In this section, we discuss the composite nature of the weakly bound states in relation to the low-energy universality. In Sec. <ref>, we first construct a simple scattering model with the nonrelativistic effective field theory where a bound state originates from the bare state. In Sec. <ref>, we then numerically compare the compositeness of typical and weakly bound states, and discuss the deviation from the expectation of the low-energy universality with a finite binding energy. We also examine the validity of the weak-binding relation in this model in Sec. <ref>. §.§ Effective field theory Let us introduce a nonrelativistic effective field theory to consider the compositeness of the bound state. We construct a model which describes the single-channel scattering of ψ_1 and ψ_2 coupled to the discrete state ϕ without the direct ψ_1 and ψ_2 interactions. The Hamiltonian is ℋ_ free =1/2m_1∇ψ_1^†·∇ψ_1+1/2m_2∇ψ_2^†·∇ψ_2 +1/2M∇ϕ^†·∇ϕ+ν_0ϕ^†ϕ, ℋ_ int =g_0(ϕ^†ψ_1ψ_2+ψ^†_1ψ^†_2ϕ). Here m_1, m_2 and M are the masses of ψ_1, ψ_2 and the discrete (bare) state ϕ, respectively. ν_0 is the energy of the bare state ϕ measured from the ψ_1ψ_2 threshold, and g_0 is the bare coupling constant of the contact three point interaction. For the Hamiltonian in Eq. (<ref>) to be Hermitian, g_0 must be real. This model can also be regarded as the resonance model without the direct ψ_1ψ_2 interaction in Refs. <cit.>. In this paper, we focus on the two-body scattering of ψ_1 and ψ_2. While we have no direct interactions in this model, the ψ_1ψ_2 scattering occurs through the intermediate ϕ state. Regarding the s-channel exchange of ϕ as the effective interaction V(k), we can derive the on-shell T-matrix T_ on(k) of the ψ_1ψ_2 scattering as a function of the on-shell momentum k from the Lippmann-Schwinger equation: T_ on(k) =V(k)+V(k)G(k)T_ on(k), V(k) =g_0^2/k^2/2μ-ν_0, G(k) =∫d^3q/(2π)^31/k^2/2μ-q^2/2μ+i0^+, with the reduced mass of the ψ_1ψ_2 system μ=(1/m_1+1/m_2)^-1. Because V(k) does not depend on the off-shell momenta, the Lippmann-Schwinger equation reduces to an algebraic equation. At the same time, the absence of the angular dependence of the interaction V(k) leads to the s-wave scattering amplitude. To avoid the divergence of the q integration in the loop function G(k), a cutoff Λ is introduced as the upper applicable boundary of the momentum in the effective field theory <cit.>. In this case, regularized G(k) becomes G(k) =-μ/π^2[Λ+ikarctan(-Λ/ik)]. The scattering observables are expressed by the scattering amplitude f(k), which is related with the on-shell T-matrix T_ on(k) as f(k)=-μ/(2π)T_ on(k). From Eq. (<ref>), we obtain the scattering amplitude f(k) as f(k) =-μ/2π[k^2/2μ-ν_0/g_0^2+μ/π^2[Λ+ikarctan(-Λ/ik)]]^-1. For the low-energy scatterings, the inverse of the scattering amplitude 1/f(k) is expanded in powers of the momentum k (the effective range expansion): 1/f(k) =-1/a_0+r_e/2k^2+𝒪(k^4)-ik. The scattering length a_0 and the effective range r_e are defined from the coefficients of the k^0 and k^2 terms in this expansion. By comparing Eq. (<ref>) with the scattering amplitude in Eq. (<ref>), we obtain the scattering length a_0 and the effective range r_e in this model: a_0 =-[2πν_0/g_0^2μ-2/πΛ]^-1, r_e =-2π/g_0^2μ^2. We note that the effective range in Eq. (<ref>) is always negative because g_0^2≥ 0 for real g_0. This is the feature of the resonance model mentioned in Refs. <cit.>. Suppose that this model generates a bound state with the binding energy B. The bound state is expressed as the pole of the scattering amplitude, and therefore the eigenmomentum is obtained by solving the bound state condition f^-1(k)=0. The bound state pole appears in the complex momentum plane at k=iκ with κ=√(2μ B)>0. The composite nature of the bound state can be characterized by the compositeness X <cit.> defined as the weight of the scattering states in the bound state |Φ⟩: X =∫d^3k/(2π)^3|⟨k|Φ||⟩^2, where |k⟩ is the scattering eigenstate of the free Hamiltonian in Eq. (<ref>) with the momentum k. The compositeness X can be expressed by the effective interaction and the loop function as discussed in Ref. <cit.>: X =G'(-B,Λ)/G'(-B,Λ)-[V^-1(-B)]'. Using this expression, we obtain X in this model from Eqs. (<ref>) and (<ref>): X =[1+π^2κ/g_0^2μ^2(arctan(Λ/κ)-Λ/κ/1+(Λ/κ)^2)^-1]^-1. We define the elementarity Z as the overlap of the bound state |Φ⟩ with the bare state |ϕ⟩ which is the discrete eigenstate of the free Hamiltonian created by the bare ϕ field at rest: Z=|⟨ϕ|Φ||⟩^2=-[V^-1(-B)]'/G'(-B,Λ)-[V^-1(-B)]'. Namely, the elementarity Z represents the fraction of the bare state component in the bound state. From the completeness relation with |ψ⟩ and |k⟩, we obtain Z+X=1, which can also be directly seen from Eqs. (<ref>) and (<ref>). For later convenience, here we introduce the typical energy scale E_ typ associated with the model. Because the cutoff Λ gives the momentum scale, we define E_ typ as E_ typ=Λ^2/2μ. If there is a bound state, the typical binding energy is expected to be B∼ E_ typ. If B≪ E_ typ, the state is regarded as a weakly bound state. To ensure that the bound state is in the applicable region of the model, we impose the condition B ≤ E_ typ. §.§ Numerical calculation In this section, we numerically investigate the compositeness of bound states in the model given in Sec. <ref>. Before the concrete calculations, we summarize the relations among the model parameters. In principle, the model parameters, the bare state energy ν_0, the coupling constant g_0 and the cutoff Λ can be arbitrarily chosen. However, for a given binding energy B, the bound state condition f^-1(iκ)=0 leads to the expression of g_0^2 with other two parameters: g_0^2(B;ν_0,Λ) =π^2/μ(B+ν_0)[Λ-κarctan(Λ/κ)]^-1, with κ=√(2μ B). Therefore, we can reduce one degree of freedom by fixing the binding energy B. In addition, when we work with dimensionless quantities using Λ, the result does not depend on the specific value of Λ. The remaining dimensionless parameter ν_0/E_ typ cannot be determined in the framework of the effective field theory.[ν_0 is the energy of the discrete bare state and corresponds to the quark core state in the application to hadrons. The value of ν_0 may be estimated, for instance, by the constituent quark model. ] In this work, we vary ν_0/E_ typ within the allowed region to investigate the model dependence of the compositeness. The parameter region of ν_0/E_ typ is restricted as follows: (i) As we discussed below Eq. (<ref>), the bound state should satisfy the condition κ≤Λ, which leads to Λ-κarctan(Λ/κ)>0. Therefore, the sign of g_0^2 in Eq. (<ref>) coincides with the sign of B+ν_0. The coupling constant square g_0^2 should be positive in Eq. (<ref>) for the Hermitian Hamiltonian. Hence the lower boundary of ν_0/E_ typ is given by -B/E_ typ≤ν_0/E_ typ.[Strictly speaking, the point at ν_0=-B should be discussed with special care, because the coupling constant vanishes and hence the bound state pole decouples from the scattering amplitude <cit.>. It is shown that the compositeness behaves as X→ 1 (X→ 0) with g_0→ 0 for fixed B=0 (B≠ 0). This behavior is confirmed in the present model as we see below.] (ii) Because the effective field theory is applicable up to the energy scale E_ typ, the upper boundary of ν_0/E_ typ is given by E_ typ/E_ typ=1. In summary, the allowed ν_0/E_ typ region is determined as -B/E_ typ≤ν_0/E_ typ≤ 1. In Fig. <ref>, we plot the compositeness X as a function of normalized bare state energy ν_0/E_ typ.[Note that the compositeness X in Eq. (<ref>) depends implicitly on ν_0 through g_0^2 [see Eq. (<ref>)].] First, we focus on the solid line which represents X of a bound state with the typical binding energy B=E_ typ. For the most of the allowed region -1≤ν_0/E_ typ≤ 1, the compositeness X is smaller than 0.5. In other words, the bound state with B=E_ typ is elementary dominant for most of the ν_0/E_ typ region. Because the ν_0 dependence of the compositeness can be regarded as the model dependence, it is probable to obtain the bound state with X<0.5 in a randomly chosen model. It is consistent with a naive expectation for the model in Sec. <ref> because the origin of the bound state is the bare state ϕ which contributes to the elementarity. We then discuss X of a weakly bound state. The dashed line in Fig. <ref> corresponds to the case with B=0.01E_ typ as a representative value of a small binding energy. In this case, the allowed region of ν_0/E_ typ in Eq. (<ref>) is -0.01≤ν_0/E_ typ≤ 1. In contrast to the typical bound state with B=E_ typ, X is larger than 0.5 for most of the allowed region of ν_0/E_ typ. Therefore, the weak-binding state is mostly composite dominant, even though the bound state originates from the bare state. When we focus on the ν_0∼ -B region, however, the compositeness of the shallow bound state is small. This is because the compositeness is fixed to be zero in the ν_0→ -B limit. This means that we can always generate an elementary dominant state by choosing the bare state energy appropriately. However, we need a fine tuning of ν_0 in the small region around ν_0∼ -B to realize an elementary dominant state. In summary, for shallow bound states, the probability to realize the composite dominant state is much higher than the elementary dominant case, although the latter possibility cannot be completely excluded. At B=0, because the coupling constant in Eq. (<ref>) becomes finite (g_0^2=π^2 ν_0/(μΛ)), from Eq. (<ref>), X=1 holds in the whole region of 0≤ν_0/E_ typ≤ 1. Therefore, the plot of the compositeness in Fig. <ref> becomes the step function. This is understood from the low-energy universality <cit.>. It is known that the compositeness X becomes unity in the weak-binding limit B→ 0 (compositeness theorem) <cit.>. The present model also follows this model independent result. From the expectation of the low-energy universality, the microscopic details such as the value of ν_0 become irrelevant, and the same relation X=1 holds for all models in the B→ 0 limit. In contrast to the finite B≠ 0 case, the elementary dominant state cannot be generated with any ν_0. By gradually decreasing the binding energy from B=0.01E_ typ (dashed line in Fig. <ref>), the region of ν_0 with the elementary dominant state becomes smaller and finally vanishes. We search for the critical binding energy B_ cr at which the fraction of the composite dominant and the elementary dominant region of ν_0/E_ typ is precisely half and half. From the numerical calculation, it turns out that B_ cr= 0.243E_ typ, and we plot the compositeness with B_ cr as the dotted line in Fig. <ref>. Namely, we expect that the composite dominant nature of the bound state becomes prominent for the state with B<B_ cr. We however note that B_ cr=0.243E_ typ is a value specific to the present model. The value depends on the choice of the regularization of the function G and the interaction Lagrangian. As a common feature in all cases shown in Fig. <ref>, X increases with ν_0/E_ typ. We can analytically show this behavior from Eq. (<ref>) because g_0^2 monotonically increases with ν_0. In the ν_0→ -B limit, the compositeness X vanishes.[For ν_0<-B, the compositeness X becomes negative. This is because the norm of the bare state becomes negative for g_0^2<0 and the admixture of the negative norm bare state gives Z<0. However, here we do not consider such cases with non-Hermitian Hamiltonian as discussed above.] When g_0^2 increases, the bound state couples more strongly to the scattering states, and the contribution of the scattering states, and hence the compositeness X, becomes larger. This can also be seen in the ν_0 dependence of X in Eq. (<ref>) mentioned above. In Ref. <cit.>, the authors discuss the nature of the bound state in the weak- and strong-coupling limits, and conclude that the bound state is elementary dominant (composite dominant) for the weak-coupling (the strong-coupling) case. Their result is consistent with our analysis of ν_0∼ -B and ν_0∼ E_ typ in Fig. <ref>. Because ν_0∼ -B (ν_0∼ E_ typ) corresponds to g_0∼ 0 (large g_0) as seen in Eq. (<ref>), states are elementary dominant with X∼ 0 (composite dominant with X∼ 1) in the weak-coupling (strong-coupling) case in our model. To quantitatively discuss the probability to find a model with the composite dominant bound state, we define P_ comp as the fraction of ν_0/E_ typ region with X>0.5: P_ comp =1-ν_c/E_ typ/1+B/E_ typ, where ν_c is the value of ν_0 such that X=0.5[If X<0.5 in the whole ν_0 region, we set ν_c=E_ typ and hence P_ comp=0. In the B→ 0 limit, the compositeness is always unity and hence we define ν_c=-B such that P_ comp=1.]. The definition of ν_c for B=E_ typ is illustrated in Fig. <ref>. Because X is plotted in the region -B/E_ typ≤ν_0/E_ typ≤ 1, the denominator of Eq. (<ref>) corresponds to the length of the horizontal axis in Fig. <ref>. The numerator is expressed by the width of the shaded region. Thus, P_ comp in Eq. (<ref>) is defined as the ratio of the shaded region to the all allowed region of ν_0/E_ typ (the length of the horizontal axis). The explicit values for the cases shown in Fig. <ref> are found to be P_ comp=0.25 at B=E_ typ and P_ comp=0.88 at B=0.01E_ typ. Because B_ cr is defined so that the composite dominant case occupies the half of the whole ν_0/E_ typ region, we obtain P_ comp=0.5 at B=B_ cr. In Fig. <ref>, we plot P_ comp by varying the normalized binding energy B/E_ typ. Small P_ comp at B/E_ typ∼ 1 monotonically increases to unity by decreasing the binding energy B. This shows that the probability to find a model with the composite dominant state is small for the typical bound states but gradually increases along with the reduction of the binding energy. In the small B region, P_ comp rapidly grows toward unity. At B=0, we have P_ comp=1; the bound state becomes completely composite dominant for any models as discussed above. Figure <ref> also shows that even when we slightly go away from the weak-binding limit B=0, it is expected that P_ comp is still close to unity. This suggests that it is probable to find the near-threshold composite dominant states with B≪ E_ typ. In fact, the weak-binding hadrons, nuclei and atoms studied in Ref. <cit.> are all composite dominant states. §.§ Weak-binding relation Finally, we discuss the validity of the weak-binding relation in this model. The weak-binding relation gives the compositeness X_ wb from the scattering length a_0 and the radius of the bound state R=1/κ as follows <cit.> X_ wb =X_ wb^c+𝒪(R_ typ/R), X_ wb^c =a_0/2R-a_0, R_ typ=max{1/Λ, |r_e|,...}, where r_e is the effective range, R_ typ is the typical length scale of the system which is estimated as the maximum length scale expect for a_0, and X_ wb^c is the central value of the compositeness in the weak-binding relation. For a weakly bound state with R≫ R_ typ, we can neglect the correction terms 𝒪(R_ typ/R) in Eq. (<ref>). In this case, X_ wb≈ X_ wb^c is obtained only from the observables a_0 and R. Therefore, the weak-binding relation is a model independent method to estimate the compositeness of the shallow bound state. In Fig. <ref>, we plot the compositeness X in this model [Eq. (<ref>)] and the central value of the compositeness from the weak-binding relation X_ wb^c in Eq. (<ref>). We use a_0 in Eq. (<ref>) and R in this model for the calculation of X_ wb^c. Panel (a) shows X (solid line) and X_ wb^c (dashed line) for the typical binding case B=E_ typ, and panel (b) for the weak-binding case B=0.01E_ typ. We see that the difference between X_ wb^c and X is significant for B=E_ typ in panel (a) while that for B=0.01E_ typ is at most 0.1 in panel (b). Therefore, the weak-binding relation gives a good estimation of the compositeness for the shallow bound state. It is remarkable that the weak-binding relation works to estimate X correctly even in the region ν_0∼ -B where X<0.5 in panel (b). In Ref. <cit.>, the validity of the weak-binding relation is demonstrated for composite dominant (X∼ 1) states with a shallow binding energy. In this work, we find that the weak-binding relation works also for shallow but elementary dominant bound states. It is instructive to analytically show that the exact compositeness X coincides with the weak-binding one in Eq. (<ref>) in the small B limit. For a weakly bound state with B≪ E_ typ (κ≪Λ), the arctangent term in the loop function in Eq. (<ref>) can be approximated as arctan(Λ/κ) =π/2+𝒪(κ/Λ). Under this approximation, the loop function G(iκ) becomes G(iκ)≈-μ/π^2(Λ-π/2κ) (B≪ E_ typ) , and the scattering amplitude is given by f(iκ) ≈[2π/μ(κ^2/2μ+ν_0/g_0^2)-2Λ/π+κ]^-1 (B≪ E_ typ) . Note that the approximation in Eq. (<ref>) is not valid for a large binding energy (κ >2Λ/π) where the square of the coupling constant becomes negative, g^2_0<0. From V and approximated G in Eq. (<ref>), the compositeness X in Eq. (<ref>) is given by X≈[1+2π/Rμ^2g_0^2]^-1 (B≪ E_ typ). By using the scattering length a_0 in Eq. (<ref>), the central value of the compositeness in Eq. (<ref>) is obtained as X_ wb^c =[2R(-2πν_0/g_0^2μ+2/πΛ)-1]^-1 =[2R(π/R^2g_0^2μ^2+1/R)-1]^-1 =[1+2π/Rμ^2g_0^2]^-1. In the second line, we use the bound state condition from the scattering amplitude in Eq. (<ref>) written by R=1/κ: -2π/μ(-κ^2/2μ-ν_0/g_0^2)-2Λ/π+κ =0 ⇔2Λ/π-2πν_0/g_0^2μ=π/R^2g_0^2μ^2+1/R. From Eqs. (<ref>) and (<ref>), we show that the exact compositeness X reduces to the central value estimated by the weak-binding relation X_ wb^c in the small B limit. In Ref. <cit.>, it is shown that there are two origins of the deviation of estimated X_ wb^c from exact X. The first one comes from the higher order terms in the derivative of the loop function, and the second one from those in the effective range expansion of the residue of the bound state pole. The derivative of the approximated loop function in Eq. (<ref>) has only the leading order term, and hence the first deviation does not appear. Because the scattering amplitude in Eq. (<ref>) has no higher order terms of 𝒪(k^4), the second deviation does not arise. In this way, we explicitly show that all the deviations disappear in the B→ 0 limit and the estimation of the compositeness by the weak-binding relation becomes exact. In this context, it is worth noting the deviation of X_ wb^c from exact X in the scattering models discussed in Ref. <cit.>. In the zero-range model with the loop function in Eq. (<ref>), X=X_ wb^c can be shown because the inverse scattering amplitude is given up to 𝒪(k). In contrast, X deviates from X_ wb^c in the resonance model because the four-point contact interaction induces the higher order terms of 𝒪(k^4) in the effective range expansion. § EFFECTS OF FOUR-POINT CONTACT INTERACTION, DECAY AND CHANNEL COUPLING As mentioned in the introduction, the actual exotic hadrons have the finite decay width and the coupling to the additional scattering channel. One can also consider the direct interaction in the threshold channel which is absent in the model in Sec. <ref>. In this section, we consider the four-point contact interaction, decay and coupled-channel effect and show how these contributions modify the results in the previous section. In Sec. <ref>, we introduce the four-point contact interaction in addition to the model in the previous section, and study the contribution of the four-point interaction to the compositeness and low-energy universality. As in the same way, the decay contribution and coupled-channel contribution are discussed in <ref> and <ref>, respectively. §.§ Effect of four-point contact interaction In this section, we investigate the effect of the direct interaction of ψ_1 and ψ_2 in addition to the model in Sec. <ref>. For this purpose, we introduce the four-point contact interaction term with the coupling constant λ_0, and the interaction Hamiltonian in Eq. (<ref>) becomes ℋ_ int =λ_0(ψ_1^†ψ_2^†ψ_1ψ_2)+g_0(ϕ^†ψ_1ψ_2+ψ^†_1ψ^†_2ϕ). Positive λ_0>0 (negative λ_0<0) corresponds to a repulsive (attractive) interaction. Because of the addition of the contact interaction term, the effective interaction V(k) in Eq. (<ref>) as a function of the momentum k changes to V(k) =λ_0+g_0^2/k^2/2μ-ν_0, while the loop function G(k) in Eq. (<ref>) remains unchanged. The scattering amplitude f(k) is obtained as f(k) =-μ/2π[(λ_0+g_0^2/k^2/2μ-ν_0)^-1. .+μ/π^2{Λ+ikarctan(-Λ/ik)}]^-1. As in Sec. <ref>, we consider the bound state with the eigenmomentum k=iκ and the binding energy B=κ^2/(2μ). The compositeness X is calculated from V in Eq. (<ref>) and G in Eq. (<ref>): X =[1+g_0^2π^2κ/μ^2/(B+ν_0)^2(λ_0-g_0^2/B+ν_0)^2(arctan(Λ/κ)-Λ/κ/1+(Λ/κ)^2)]^-1. The model parameters are the bare state energy ν_0, the cutoff Λ, and the coupling constants λ_0 and g_0. As in the model in Sec. <ref>, from the bound state condition f(iκ)^-1=0 with a fixed binding energy, g_0^2 is written in terms of the binding energy B and other morel parameters: g_0^2(B; ν_0,λ_0,Λ) =(B+ν_0)(π^2/μ/Λ-κarctan(Λ/κ)+λ_0). Furthermore, the use of the dimensionless parameters can absorb the Λ dependence. Therefore, the remaining parameters ν_0 and λ_0 are varied in the calculation of the compositeness. As in Sec. <ref>, we vary ν_0/E_ typ in the region -B/E_ typ≤ν_0≤ 1. We now consider the relevant parameter region of λ_0. From Eq. (<ref>), we see that g_0^2 becomes negative for large negative λ_0. To avoid this problem, we introduce the lower boundary of λ_0 as λ_0^ b, which is determined by the condition g_0^2=0 in Eq. (<ref>): λ^ b_0=-π^2/μ[Λ-κarctan(Λ/κ)]^-1. Thus, λ_0 should be chosen in the allowed region -|λ_0^ b| ≤λ_0. Note that λ^ b_0 depends on the binding energy B=κ^2/(2μ). We then determine the region of λ_0 for the numerical calculation. We define λ_0^ cr as the critical value of the attractive coupling constant which supports a bound state at B=0 without the bare state contribution: λ^ cr_0=-π^2/μΛ. In fact, with λ_0=λ^ cr_0 and g_0^2=0, the scattering amplitude in Eq. (<ref>) has a pole at B=κ=0. For a stronger attraction than λ_0^ cr, a bound state is formed only by the direct interaction. Because the formation of a bound state is a non-perturbative phenomenon, |λ_0^ cr| can be regarded as the representative strength of the strong coupling. To examine the weak to strong couplings for both repulsive and attractive interactions, we vary λ_0 in the region -|λ_0^ cr|≤λ_0≤ |λ_0^ cr|. Note that the relation -|λ_0^ b|<-|λ_0^ cr| always holds since κarctan(κ/Λ)>0 for any κ>0. Thus, the condition -|λ_0^b|<λ_0 is guaranteed with Eq. (<ref>) for any κ. To observe the effect of the contact interaction with λ_0, we plot the compositeness X as a function of the normalized bare states energy ν_0/E_ typ for the weak-binding case B=0.01E_ typ in Fig. <ref>. The solid, dashed, and dotted lines express X with λ_0=0, λ_0=-|λ^ cr_0|, and λ_0=|λ^ cr_0|, respectively. As shown in Fig. <ref>, the repulsive interaction |λ^ cr_0|>0 decreases X and the attractive interaction -|λ^ cr_0|<0 increases X for fixed ν_0. To understand this behavior, we consider the interaction mechanisms and their implication to the compositeness of the bound state. In the present model, the bound state is originated not only from the bare state pole but also from the attractive four-point interaction. As discussed in Sec. <ref>, the bare pole term contributes to the elementarity Z. In contrast, the attractive four-point interaction provides the composite bound state, and hence contributes to the compositeness X. With both the interactions, the compositeness of the bound state is determined by the interplay between the bare pole term proportional to g_0^2 and the direct interaction proportional to λ_0. Because the binding energy is chosen to satisfy -B< ν_0, Eq. (<ref>) indicates that g_0^2 increases with λ_0 for fixed B and ν_0. Intuitively, negative λ_0 (attractive four-point interaction) tends to increase the binding energy and hence the coupling to the bare pole term g_0^2 should be reduced to keep the binding energy unchanged. In contrast, g_0^2 increases to compensate the reduction of the binding energy by the repulsive four-point interaction with positive λ_0. This relation between g_0^2 and λ_0, together with the origin of the bound state discussed above, explains the behavior of the compositeness with respect to λ_0; the introduction of the repulsive (attractive) four-point interaction with positive (negative) λ_0 increases (decreases) g_0^2 and therefore the compositeness of the bound state X decreases (increases). The λ_0 dependence of the compositeness can be visualized by plotting X as a function of the normalized coupling constant λ_0/|λ_0^ cr| in Fig. <ref>. In this plot, we fix the bare state energy as ν_0=0.5E_ typ and we have checked that the qualitative result does not change for different values of ν_0. The solid line represents X for B=E_ typ, and the dashed line represents for B=0.01E_ typ. In both cases, X decreases with the increase of λ_0/|λ_0^ cr|, as discussed above. In Fig. <ref>, we see that the compositeness X depends on λ_0 more strongly for B=0.01E_ typ than that for B=E_ typ. This tendency originates in the structure of the bound state at λ_0=0; a stronger coupling g_0 is required to generate the deeper bound state with the same ν_0, as indicated by the smaller compositeness X for the typical bound state with B=E_ typ. The deeper bound state is less affected by the introduced four-point interaction and hence the λ_0 dependence becomes milder. Finally, in Fig. <ref>, we plot the fraction of composite dominant state P_ comp as a function of the normalized binding energy B/E_ typ to discuss the low-energy universality with λ_0 contribution. The solid line represents P_ comp with λ_0=0 (same as Fig. <ref>), the dashed line with λ_0=|λ_0^ cr|, and the dotted line with λ_0=-|λ_0^ cr|. For all the λ_0 cases, P_ comp decreases when the binding energy B increases. Because positive λ_0>0 (repulsive interaction) suppresses the compositeness and ν_c becomes smaller (see Fig. <ref>), P_ comp is also suppressed. In contrast, attractive interaction with negative λ_0<0 enhances P_ comp because it induces the increase of X and the decrease of ν_c in Fig. <ref>. At B=0, we see that P_ comp becomes unity for all λ_0 cases. This result indicates that the bound state becomes completely composite dominant in the B→ 0 limit even with the four-point interaction with any strength. It is consistent with the consequence of the low-energy universality. At the same time, the decrease rate of P_ comp depends on the strength of the four-point interaction λ_0. In other words, λ_0 dependence in Fig. <ref> expresses the model dependence of P_ comp away from the B→ 0 limit. §.§ Effect of decay Because the exotic hadrons generally have a decay width, we consider the decay effect to the compositeness in this section. To concentrate on the decay effect, here we do not include the four-point interaction and set λ_0=0. The decay effect can be formally described in the effective field theory by introducing the decay channel in the energy region lower than the binding energy in addition to the threshold channel <cit.>. In this paper, we effectively introduce the decay effect by letting the coupling constant g_0 be a complex number in the Hamiltonian in Eq. (<ref>). Because the Hamiltonian is non-Hermitian, the the eigenenergy becomes complex as E=-B-iΓ/2, with the decay width Γ. In the presence of the decay width Γ, the square of the coupling constant g_0^2 in Eq. (<ref>) is g_0^2 =π^2/μ(B+iΓ/2+ν_0)[Λ-κarctan(Λ/κ)]^-1, κ = √(2μ (B+iΓ/2)) , which is complex for Γ≠ 0. By definition, the compositeness X and elementarity Z are complex for unstable states <cit.>. In fact, the compositeness X in this model, obtained with Eq. (<ref>), is not a real number with complex g_0^2 and κ for the finite Γ≠0 case. However, we cannot interpret complex X and Z as the probabilities as in the case of the bound state with real X and Z. To discuss the structure of unstable states, we need to introduce other real quantities which can be interpreted as the fraction of the composite (elementary) components instead of complex X (Z). Here we employ the quantities X̃ and Z̃ defined as X̃ =|X|/|X|+|Z|, Z̃ =|Z|/|X|+|Z|, which are proposed in Ref. <cit.>. For stable states without the decay width, X̃ and Z̃ reduce to X and Z because |X|=X, |Z|=Z and X+Z=1. It is clear that X̃ and Z̃ satisfy the sum rule: X̃+Z̃=1. In addition, it follows from the definitions in Eqs. (<ref>) and (<ref>) that the relations 0≤X̃≤ 1 and 0≤Z̃≤ 1 hold. Therefore, we can regard X̃ and Z̃ as the probabilities to find the composite and elementary components in a wavefunction instead of complex X and Z. Hence we call X̃ and Z̃ the compositeness and the elementarity, respectively. To observe the effect of the decay, in Fig. <ref>, we plot the compositeness X̃ by the solid lines as a function of the normalized bare state energy ν_0/E_ typ for various B and Γ. The panels (a) and (b) [(c) and (d)] correspond to the weak-binding (typical binding) case, and the panels (a) and (c) [(b) and (d)] represent the state with a narrow (broad) decay width. The bare state energy ν_0 is varied in the region -B/E_ typ≤ν_0/E_ typ≤ 1. For comparison, the dashed lines represent the compositeness X̃ for the same B but with Γ=0 (same as the solid and dashed lines in Fig. <ref>). By comparing the solid and dashed lines, we see that the effect of the decay width generally suppress the compositeness, while X̃ is enhanced at small ν_0∼ -B. Basically, the compositeness of the threshold channel decreases when the decay width is turned on because the coupling to the decay channel increases. This tendency becomes prominent especially in panel (b). The small ν_0 behavior is however governed by the compositeness at ν_0=-B. From Eqs. (<ref>) and (<ref>), without the decay effects, the compositeness becomes zero in the ν_0→ -B limit because g_0^2→ 0. On the other hand, with a finite width Γ≠ 0, g_0^2 does not vanish at ν_0=-B: g_0^2(-ν_0+iΓ/2;ν_0,Λ) =π^2/μ(-iΓ/2)[Λ-κarctan(Λ/κ)]^-1 ≠ 0. From Eq. (<ref>), the complex compositeness X becomes nonzero, and X̃ in Eq. (<ref>) becomes larger than zero. This explains the enhancement of X̃ at ν_0∼ -B. Furthermore, by comparing panels (a) and (c) with (b) and (d), we see that the ν_0 dependence of X̃ becomes smaller for larger decay width. It follows from Eq. (<ref>) that the ν_0 dependence of g_0^2 is negligible for |B+iΓ/2|≫ν_0. Therefore, X̃ is less dependent on ν_0, and the plot of X̃ becomes flat for larger Γ. For more quantitative discussion, let us analytically evaluate X̃ in the large decay width limit, Γ≫ E_ typ. Because ν_0 is varied in the -B≤ν_0≤ E_ typ region and the binding energy is restricted within B≲ E_ typ, the relations ν_0≪Γ and B≪Γ hold under the large width limit. Furthermore, because κ=√(2μ (B+iΓ/2))∼√(iμΓ) and Γ≫ E_ typ=Λ^2/(2μ), we find |κ|≫Λ in this limit. In this case, the coupling constant g_0^2 in Eq. (<ref>) behaves as g_0^2 =3π^2κ^4/2μ^2Λ^3+…, from the expansion of arctan(Λ/κ) for |κ|≫Λ: arctan(z) =z-z^3/3+𝒪(z^5) (|z|≪ 1). By substituting Eq. (<ref>) into the compositeness in Eq. (<ref>) and expanding the terms in the parenthesis by Λ/κ, we obtain X for the large decay width limit as X =1/2+…. Because Z=1/2+…, X̃ is calculated as X̃ =1/2+…. Therefore, in the large width limit, X̃ approaches 1/2 for any ν_0 as expected from panels (b) and (d) in Fig. <ref>. It is worth noting that, in the large width limit Γ≫ E_ typ, the magnitude of the eigenenergy exceeds the applicable region of the model. Therefore, this formal limit should only be used to understand the behavior of X̃ with increasing Γ. To study the decay effect with respect to the binding energy, we compare panel (a) with (c) where the eigenstates have the same decay width. The decay effect (deviation of X̃ with Γ≠ 0 from that with Γ=0) in panel (a) is sizable, whereas the effect in panel (c) is almost negligible. While the half width Γ/2=0.1E_ typ is larger than the binding energy B=0.01E_ typ in panel (a), the same decay width Γ/2=0.1E_ typ is smaller than B=E_ typ in panel (c). Therefore, we conclude that the deviation of X̃ by the decay effect is determined by the ratio of the binding energy to the decay width. To discuss the low-energy universality, we define P_ comp as in Eq. (<ref>) but with X̃=0.5 as the determination of ν_c. In Fig. <ref>, we plot P_ comp as a function of the normalized binding energy B/E_ typ in the presence of the decay width. The solid line stands for P_ comp with Γ=0, the dashed line with Γ/2=0.1E_ typ, and the dotted line with Γ/2=E_ typ. By comparing the solid line with the dashed and dotted lines, we see that P_ comp decreases when the decay width increases. This reason is understood from the X̃ behavior in Fig. <ref>, where X̃ is suppressed by introducing the decay width and ν_c becomes large accordingly. Therefore, the decay effect makes P_ comp smaller than that for the stable states. From the B dependence in Fig. <ref>, the deviation of the dashed and dotted lines from the solid line becomes smaller in the large B region. Through the comparison of panels (a) and (b) with panels (c) and (d) in Fig. <ref>, the deviation of the compositeness with and without decay width becomes smaller for larger binding energy. In contrast, for B/E_ typ≲ 0.2 region in Fig. <ref>, the deviation of the dashed and dotted lines from the solid line becomes larger. This reason is shown in panels (a) and (c) in Fig. <ref>; for smaller B [panel (a)], there is a sizable change of ν_c while ν_c is almost unchanged for larger B [panel (c)]. At B=0, P_ comp≠ 1 with the finite decay width in contrast to the effect of the direct interaction in Fig. <ref>. With Γ≠ 0, κ and g_0^2 in Eq. (<ref>) are finite at B=0. Therefore, the relation X̃=1 does not hold with finite Γ even in the weak-binding limit B→ 0, and hence P_ comp<1. §.§ Effect of channel coupling In the previous section, we have studied the decay effect which arises from the couplings to the lower energy channel than the threshold one. In this section, we consider the effect of the coupling to the higher energy channel. For this purpose, we introduce the scattering of Ψ_1 and Ψ_2 (channel 2) in addition to the ψ_1ψ_2 scattering (channel 1) in the free Hamiltonian in Eq. (<ref>): ℋ_ free =1/2m_1∇ψ_1^†·∇ψ_1+1/2m_2∇ψ_2^†·∇ψ_2 +1/2M_1∇Ψ_1^†·∇Ψ_1+1/2M_2∇Ψ_2^†·∇Ψ_2 +1/2M∇ϕ^†·∇ϕ+ω_1Ψ_1^†Ψ_1+ω_2Ψ_2^†Ψ_2+ν_0ϕ^†ϕ. where M_1 and M_2 are the masses of Ψ_1 and Ψ_2, ω_1 and ω_2 are the energies of Ψ_1 and Ψ_2 measured from the ψ_1ψ_2 threshold. Δω=ω_1+ω_2>0 is the threshold energy difference between channels 1 and 2. For the transition from channel 1 to channel 2, here we introduce the coupling of channel 2 and the bare state ϕ. We employ the same coupling constant g_0 with that for channel 1 and ϕ. The interaction Hamiltonian leads to ℋ_ int =g_0(ϕ^†ψ_1ψ_2+ψ^†_1ψ^†_2ϕ+ϕ^†Ψ_1Ψ_2+Ψ^†_1Ψ^†_2ϕ). We now consider the on-shell T-matrix T_ on(k) of the coupled-channel scatterings. As in the single-channel case in Sec. <ref>, the scatterings occur through the effective interaction with the bare state ϕ exchange. In the coupled-channel scattering, T_ on(k), the effective interaction V(k) and the loop function G(k) are expressed by the matrices in the channel space. In this model, the on-shell T-matrix is given by T_ on(k_1) =V(k_1)+V(k_1)G(k_1)T_ on(k_1), V(k_1) =[ v(k_1) v(k_1); v(k_1) v(k_1) ], G(k_1) =[ G_1(k_1) 0; 0 G_2(k_2(k_1)) ]. Here each component of V(k) and G(k) is v(k_1) =g_0^2/k_1^2/2μ_1-ν_0, G_i(k_i) =-μ_i/π^2[Λ+ik_iarctan(-Λ/ik_i)], where the momentum of each channel at the energy E is k_1 =√(2μ_1E), k_2(k_1) =√(2μ_2(E-Δω))=√(μ_2/μ_1k_1^2-2μ_2Δω), with μ_1=(1/m_1+1/m_2)^-1 and μ_2=(1/M_1+1/M_2)^-1. As before, we assume that there is a bound state. The bound state condition for the coupled-channel scattering is given by (1-GV)=0. This leads to E-ν_0-g_0^2(G_1(k_1)+G_2(k_2))=0, with E=-B. By solving this condition for g_0^2, we obtain the expression of g_0^2 as g_0^2(B;ν_0,Λ)=-B+ν_0/G_1(iκ_1)+G_2(iκ_2), with κ_1=√(2μ_1B) and κ_2=√(2μ_2(B+Δω)) In the coupled-channel scattering, the compositeness is defined for each channel as X_1 and X_2. X_i is interpreted as the probability to find channel i composite state in the bound state. As in the single-channel case, the compositeness X_1, X_2 and the elementarity Z=1-X_1-X_2 are calculated from the effective interaction in Eq. (<ref>) and the loop functions in Eq. (<ref>). As discussed in Ref. <cit.>, the expression of X_1 is obtained by replacing G→ G_1 and V^-1→ [v_ eff]^-1 in Eq. (<ref>), where v_ eff is the effective interaction in channel 1 obtained by eliminating the bare state and channel 2. In the present model, the effective interaction is <cit.>: [v_ eff]^-1(k_1) =1-G_2(k_2)v(k_1)/[1-G_2(k_2)v(k_1)]v(k_1)+G_2(k_2)v^2(k_1) =v^-1(k_1)-G_2(k_2). Then the compositeness X_1 and X_2 are X_1 =G'_1(iκ_1)/G'_1(iκ_1)+G'_2(iκ_2)-[v^-1]', X_2 =G'_2(iκ_2)/G'_1(iκ_1)+G'_2(iκ_2)-[v^-1]', where κ_1=√(2μ_1B), κ_2=√(2μ_2(B+Δω)), and the derivatives of v^-1 and the loop functions G_1,G_2 are given by [v^-1]' =1/g_0^2, G'_i(iκ_i) =-μ_i^2/π^2κ_i(arctan(Λ/κ_i)-Λ/κ_i/1+(Λ/κ_i)^2), with i=1 and 2. For the numerical calculation, we can choose arbitrarily μ_1,2 and Δω by adjusting m_1,2, M_1,2 and ω_1,2. With the dimensionless parameters, the result only depends on the ratio of μ_1 and μ_2. In this section, to focus on the Δω dependence, we assume μ_1=μ_2. To quantitatively study the contribution of the coupled channel, we plot the compositeness as a function of the normalized bare state energy ν_0/E_ typ in Fig. <ref>. The solid lines stand for X_1+X_2, the dotted lines for X_1. Therefore, the difference between the solid and dotted lines corresponds to X_2. To see the coupled-channel effect to the compositeness, we plot X in Eq. (<ref>) for single-channel case with same B and ν_0 by the dashed lines. Panels (a) and (b) [(c) and (d)] correspond to the weak-binding (typical binding) case, and panels (a) and (c) [(b) and (d)] show the results with small (large) threshold energy difference Δω. By comparing panels (a) with (b) and (c) with (d), we see that X_2 becomes smaller for larger threshold energy difference Δω. This is analytically explained by the behavior of X_2 in Eq. (<ref>) in the large Δω limit. When Δω→∞, κ_2 also goes to infinity. This induces that G'_2→ 0 in Eq. (<ref>) and X_2 becomes zero in Eq. (<ref>). Intuitively, this is because the channel 2 contribution vanishes when the threshold is infinitely far away. In the opposite limit Δω→ 0, we can also analytically show that X_1=X_2, because κ_1=κ_2 and then G_1=G_2 under the assumption of this calculation with μ_1=μ_2, so Eq. (<ref>) becomes identical with Eq. (<ref>). This is because the bound state couples to both the channels with an equal weight. This behavior is reflected in panel (c), where Δω is negligibly smaller than B and the dotted line indicates about a half of the solid line. We note that the ratio X_1/X_2 in the Δω→ 0 limit depends on the coupling strengths in channels 1 and 2. In this work, we obtain X_1/X_2=1 because the common coupling constant g_0 to both the channels is adopted in the interaction Lagrangian in Eq. (<ref>). With different coupling strengths for channels 1 and 2, we obtain the ratio X_1/X_2≠ 1 in the Δω→ 0 limit. It is also observed in all panels in Fig. <ref> that the sum X_1+X_2 (the solid line) is close to X in the single-channel model (the dotted line). In our coupled-channel model, the bound state is formed by the dressing of the bare state through the coupling to the scattering states. The dressing induces the two-body composite component to the eigenstate and increases the compositeness. From a fixed bare state energy ν_0, we need the same amount of the dressing to obtain the bound state at E=-B, irrespective of the number of coupled channels. In the coupled-channel model, channels 1 and 2 work cooperatively to achieve the equivalent dressing with that in the single-channel model. In other words, the compositeness X_1+X_2∼ X represents the total amount needed to dress the bare state to the bound state. For the multi-channel case, the low-energy universality indicates that the bound state is completely dominated by the threshold channel, namely, X_1=1, X_2=0, and Z=0 in the B→ 0 limit. To focus on the dominance of X_1, we define P_ comp in Eq. (<ref>) with ν_c which gives X_1=0.5 as the probability to find a model with the ψ_1ψ_2 composite dominant state. In Fig. <ref>, we plot P_ comp as a function of the normalized binding energy B/E_ typ for Δω=E_ typ (dashed line), Δω=10E_ typ (dotted line), and Δω→∞ (solid line) representing the single-channel case. By comparing three lines, we find that P_ comp is suppressed by a finite threshold energy difference Δω from the single-channel case at the same B, and the suppression becomes larger for smaller Δω. The reason for this is seen as the change of ν_c in panels (a) and (b) in Fig. <ref>; ν_c/E_ typ= 0.15 for Δω=E_ typ [panel (b)] changes to ν_c/E_ typ= 0.71 for Δω=0.01E_ typ [panel (a)] so that the fraction of the composite dominant region decreases. In Fig. <ref>, the dashed line becomes zero in the region B/E_ typ≥ 0.35, where the channel 1 compositeness X_1 is always smaller than 0.5 and there is no X_1 dominant region [see panel (d) in Fig. <ref>]. At B=0 in Fig. <ref>, P_ comp becomes unity even with the coupled-channel effect with finite Δω. For arbitrary Δω, one can always consider the small binding energy B such that B≪Δω. In this case, the bound state decouples from channel 2, and X_2 becomes zero as discussed above in the Δω→∞ limit. At the same time, the bound state is completely dominated by the composite component of the threshold channel, X_1→ 1. This is consistent with the consequence of the low-energy universality. § APPLICATION TO T_CC AND X(3872) Based on the properties of the near-threshold states discussed so far in general cases, we now consider the application to hadron physics. As prominent examples of weakly bound exotic hadrons, we discuss the nature of T_cc and X(3872) by calculating the compositeness with the effective field theory. As mentioned in the introduction (see Fig. <ref>), T_cc is observed slightly below the D^0D^*+ threshold, and the coupled channel of the isospin partner D^*0D^+ exists above the threshold channel. Similarly, X(3872) is the weakly bound state near the D^0D̅^*0 threshold, and couples to the D^+D^*- channel above the threshold. Both the states decay through the strong interaction. Therefore, to analyze the structure of T_cc and X(3872), we introduce both contributions of the decay and the channel coupling discussed in Sec. <ref>. As mentioned in Sec. <ref>, for an unstable state, we need to introduce X̃ in Eq. (<ref>) as the compositeness because the complex compositeness is not interpreted as a probability. For the coupled-channel effect, we define X_1 and X_2 as the compositeness of the threshold and coupled channels, respectively, as in Sec. <ref>. To take into account both the decay and the coupled-channel contributions, we employ X̃_1 and X̃_2 proposed in Ref. <cit.>: X̃_j =|X_j|/∑_j|X_j|+|Z|, (j=1,2). X̃_1 and X̃_2 can be interpreted as the probabilities to find the threshold and the coupled-channel components, respectively. For the numerical calculation, the masses of the D mesons are taken from the Particle Data Group (PDG) <cit.>. We employ the binding energy and the decay width of T_cc from central values of the pole parameters in Ref. <cit.> and those of X(3872) from PDG <cit.>. We use the cutoff Λ=140∼ m_π MeV because π can be exchanged between the D mesons. In this case, the typical binding energy scales are obtained as E_ typ=10.13 MeV and E_ typ=10.14 MeV for the T_cc and X(3872) systems, respectively. The coupled-channel and decay effects are characterized by the threshold energy difference Δω and the decay width Γ. In the T_cc case, the energy difference between the threshold channel and the coupled channel is Δω=1.41 MeV, and the decay width is Γ=0.048 MeV. In the X(3872) case, the energy difference is Δω=8.23 MeV, and the decay width is Γ=1.19 MeV. In this way, we have smaller threshold energy difference Δω and decay width Γ for T_cc, and larger Δω and Γ for X(3872) as shown in Fig. <ref>. In Fig. <ref>, we plot the compositeness of T_cc [panel (a)] and X(3872) [panel (b)] as a function of the bare state energy ν_0. The solid lines represent X̃_1+X̃_2, and the dotted lines X̃_1. For comparison, we show by the dashed lines X̃_1+X̃_2 with artificially setting Γ=0. In panel (a), the solid line almost overlaps with the dashed line and the deviation of X̃_1+X̃_2 by the decay width is too small to observe. This is because X̃_1+X̃_2 of T_cc dose not change when the narrow decay width (0.048 MeV) is turned on. In contrast, in panel (b), we find a sizable deviation by the decay effect. Although the decay width of X(3872) (1.19 MeV) is small in hadron physics, it is nevertheless larger than the binding energy (0.04 MeV) and the magnitude of the decay effect reflects the ratio of the binding energy to the decay width as we discussed in Sec. <ref>. Next, we consider the effect of the channel coupling, indicated by the difference between the solid and dotted lines. In Fig. <ref>, we see that the difference between those lines is larger in panel (a) than in panel (b). In other words, the coupled-channel contribution X̃_2 of X(3872) is smaller than that of T_cc. This is because the channel-coupling effect is suppressed for X(3872) with the larger threshold energy difference in comparison with the T_cc case. The concrete value of ν_0 cannot be determined in the effective field theory, unless other physical quantities (such as the scattering length) are given in addition to the eigenenergy. Alternatively, one can employ a specific model to estimate the value of ν_0. For example, the constituent quark model in Ref. <cit.> gives the bare energy of the four-quark state of T_cc as ν_0=7 MeV. From the estimated value of ν_0, one can read off the structure of T_cc and X(3872) from Fig. <ref>. In summary, the molecular nature of states can be modified by both the decay and the coupled-channel effects even if the pole exists near the threshold, and we need to consider these effects for the quantitative discussion of the compositeness. In particular, the coupled-channel effect should be important for T_cc, and decay width for X(3872) as demonstrated in Fig. <ref>. In fact, the probability to find the model with the composite dominant state P_ comp is obtained as P_ comp(T_cc) =0.55, P_ comp[X(3872)] =0.41. This result shows that the substantial coupled-channel and decay effects can reduce the compositeness of T_cc and X(3872) and the molecular nature is not straightforwardly guaranteed, even though both the states exist within the 1 MeV region from the threshold. § SUMMARY We have discussed the structure of weakly-bound states from the viewpoint of the compositeness. Naively, the states near the two-body s-wave threshold are expected to have a molecule-type composite structure. In this paper, we explicitly demonstrate that it is always possible to construct a weakly-bound non-composite state, but only with a significant fine tuning in the system. In other words, the realization of the non-composite state near the threshold is probabilistically suppressed, in accordance with the low-energy universality. In addition, we quantitatively study how this universal nature of the near-threshold bound states can be modified by the various effects, and examine their implications to the structure of exotic hadrons. We first construct an effective field theory model with a bare state coupled to a single two-body scattering channel, and evaluate the compositeness of the bound state in this model within the allowed parameter region. It is shown that the bound state in this model is usually an elementary-dominant state originated from the bare state when the binding energy B is of the order of the typical energy scale of the system, B∼ E_ typ. However, if the binding energy is small (B≪ E_ typ), then the bound state has a high probability of being the composite-dominant state in the parameter region of the model. We quantitatively show that the probability of generating the composite dominant state gradually increases when B decreases, and finally approaches unity in the B→ 0 limit. While this simple model captures the essential features of the near-threshold bound states, there are various effects present in the application to the exotic hadrons. We thus generalize the above mentioned model by including the four-point contact interaction, the decay effect, and the coupling to the additional scattering channel. It is shown that the attractive (repulsive) four-point contact interaction increases (decreases) the compositeness of the bound state, because it helps to enhance (suppress) the generation of the molecule component. We show that the decay and coupled-channel effects decrease the compositeness, as they induce the contribution from the other components. The importance of the decay (coupled-channel) effect is characterized by the ratio of the decay width (the threshold energy difference) with respect to the real part of the eigenenergy. Finally, we consider the structure of T_cc and X(3872) in this perspective. It is known that T_cc and X(3872) appear close to the D^0D^*+ and D^0D̅^*0 thresholds, respectively, but both states have a finite decay width and a nearby coupled channel [D^*0D^+ for T_cc and D^+D^*- for X(3872)]. As expected from the small threshold energy difference in T_cc and the sizable decay width of X(3872), we show that the channel coupling to the D^*0D^+ (the decay effect) largely influences the compositeness of T_cc [X(3872)]. In other words, it is important to consider the coupling to D^*0D^+ (the decay effect) to quantitatively study the internal structure of T_cc [X(3872)]. In this way, we expect that the result of this work provides quantitative guidance to pin down the important effects for the discussion of the structure of near-threshold states. This work has been supported in part by the Grants-in-Aid for Scientific Research from JSPS (Grants No. JP22K03637 and No. JP18H05402). This work was supported by JST, the establishment of university fellowships towards the creation of science technology innovation, Grant No. JPMJFS2139.

http://arxiv.org/abs/2303.07191v1

20230313153201

Transverse single-spin asymmetry of charged hadrons at forward and backward rapidity in polarized $p$+$p$, $p$+Al, and $p$+Au collisions at $\sqrt{s_{NN}}=200$ GeV}

nucl-ex

[PHENIX Spokesperson: ]akiba@rcf.rhic.bnl.gov Deceased PHENIX Collaboration Reported here are transverse single-spin asymmetries (A_N) in the production of charged hadrons as a function of transverse momentum (p_T) and Feynman-x (x_F) in polarized p^↑+p, p^↑+Al, and p^↑+Au collisions at √(s__NN)=200 GeV. The measurements have been performed at forward and backward rapidity (1.4<|η|<2.4) over the range of 1.5<p_T<7.0  GeV/c and 0.04<|x_F|<0.2. A nonzero asymmetry is observed for positively charged hadrons at forward rapidity (x_F>0) in p^↑+p collisions, whereas the p^↑+Al and p^↑+Au results show smaller asymmetries. This finding provides new opportunities to investigate the origin of transverse single-spin asymmetries and a tool to study nuclear effects in p+A collisions. Transverse single-spin asymmetry of charged hadrons at forward and backward rapidity in polarized p+p, p+Al, and p+Au collisions at √(s__NN)=200 GeV L. Zou March 30, 2023 ====================================================================================================================================================== § INTRODUCTION An understanding of transverse single-spin asymmetries (TSSAs) in transversely polarized proton-proton collisions () is crucial for disentangling the spin structure of the proton and parton dynamics within the proton. The analyzing power A_N is defined as the left-right asymmetry of the produced hadrons with respect to the spin direction of the polarized proton where the polarization direction is perpendicular to the beam direction. Since the 1970s, significant TSSAs in hadron (h) production (p^↑+p→h+X) have been measured at large Feynman-x (x_F = 2 p_L/√(s)) for a wide range of collision energies up to 500 GeV <cit.>. Two approaches have been proposed to account for these large asymmetries. First, the transverse-momentum dependent (TMD) approach is based on TMD parton distribution and fragmentation functions. It requires two scales, a hard-scattering energy scale Q and a soft scale k_T describing the transverse momentum of partons in the proton or of hadrons relative to the parent parton in hadronization process, with k_T ≪ Q. In this framework, the possible origins of the asymmetry are the Sivers <cit.> and Collins <cit.> mechanisms. The Sivers mechanism describes the initial-state correlation between the spin of the transversely polarized proton and the parton transverse momentum, while in the final state the Collins mechanism introduces a correlation between the transverse spin of the fragmenting quark and transverse momentum of the final state hadron. The Collins mechanism convolves with the quark transversity distribution, which describes the quark transverse polarization inside a transversely polarized proton. Secondly, in the other framework, the collinear twist-3 factorization is applicable for observables with one large scale Q, which is usually represented by the particle transverse momentum p_T in reactions such as p^↑+p→h+X. In this framework, the asymmetry arises from a twist-3 multi-parton correlation function and a twist-3 fragmentation function. Twist-3 multi-parton correlation functions represent a spin dependence of the transverse motion of the parton inside a polarized proton <cit.>. Twist-3 fragmentation functions denote parton fragmentation effects during the formation of the final state hadrons. Recent calculations of the twist-3 contribution from parton fragmentation are shown to be important to describe the Relativistic Heavy Ion Collider (RHIC) results <cit.>. Also, a recent phenomenological study using both TMD and collinear twist-3 approaches demonstrates that TSSAs in Semi-Inclusive Deep Inelastic Scattering (SIDIS), Drell-Yan, e^+e^- annihilation, and proton-proton collisions have a common origin <cit.>. Recently, an exploration of the interplay of spin physics and small-x physics through the measurement of TSSAs in transversely polarized proton-nucleus collisions () has attracted attention, where x is the momentum fraction of a proton carried by the parton. In collisions, the properties of small-x gluons inside nuclei can be probed by measuring hadron production in the proton-going direction. The gluon density in the small-x region of the target nuclei is expected to increase significantly and it may be described by the color-glass-condensate (CGC) formalism <cit.>. The framework introduces a characteristic saturation scale Q_s which describes the color-charge density fluctuations, where Q_sA∝ A^1/3 for the target nucleus. In collisions, measuring TSSAs can be used as a probe for the saturation scale in the nucleus <cit.>. In addition, TSSAs in the forward region from collisions help to disentangle differing mechanisms and clarify the origin of the TSSA. Previous calculations of TSSAs incorporating gluon saturation at small-x suggest that TSSAs in may or may not be A-dependent, depending on the mechanisms involved <cit.>. On the experimental side, the A_N of positively charged hadrons at 0.1<x_F<0.2 in PHENIX <cit.> and π^0 results at 0.2<x_F<0.7 in STAR in and collisions have been reported recently <cit.>. It should be noted that the kinematic region of the measurements presented here is outside of the expected CGC range, but considerations on the A-dependence of the various contributions might still be relevant. This paper reports on measurements in 2015 of the transverse single-spin asymmetry for the production of charged hadrons (h^±) over the range of 1.5<p_T<7.0 GeV/c and 0.04<|x_F|<0.2 at forward and backward rapidity (1.4<|η|<2.4) from transversely polarized proton-proton (), proton-Aluminum (), and proton-Gold () collisions at √(s__NN)=200 GeV. Section <ref> describes the experimental setup in PHENIX and the polarized beams at RHIC. The details of the analysis procedure are presented in Sec. <ref>, and the result are shown in Sec. <ref> with discussions in Sec. <ref> and finally, the summary in Sec. <ref>. § EXPERIMENTAL SETUP §.§ The PHENIX experiment The PHENIX detector <cit.> is equipped with central arms at midrapidity, and muon arms at backward and forward rapidity. The muon arms cover the pseudorapidity range 1.2<η<2.4 (north arm) and -2.2<η<-1.2 (south arm) and the full azimuthal angle (Δϕ = 2π) <cit.>. The side view of the PHENIX detector including muon arms during the 2015 run is shown in Fig. <ref>. Muons and hadrons produced from the interaction region pass through a hadron absorber of 7.5 nuclear interaction lengths (λ_I). Surviving muons and hadrons reach a Muon Tracker (MuTr) composed of three stations of cathode-strip readout tracking chambers mounted inside conical-shaped muon magnets. The momentum of a charged particle is measured with the MuTr. The Muon Identifier (MuID) behind the MuTr is composed of five layers (labeled from Gap0 to Gap4) of proportional tube planes combined with an absorber plane (≈1λ_I), respectively. The MuID provides identification of hadrons and muons by measuring penetration depth. Muons of momentum larger than 3 GeV/c penetrate all layers while charged hadrons are mostly stopped at the intermediate layers (Gap2 and Gap3) <cit.>. The Beam-Beam Counters (BBC) are located at z=±144  cm from the interaction point and cover the pseudorapidity range 3.1<|η|<3.9 and full azimuthal angle <cit.>. Each BBC comprises 64 quartz Čerenkov counters. The BBC detects charged particles and provides z-vertex position with a resolution about 2 cm in collisions. The BBC is also used to categorize events with energy deposition in terms of centrality for collisions with ion beam. In the case of and collisions, energy deposition in the A-going direction is used for determining centrality. In addition, the BBC serves as a luminosity monitor. The minimum-bias trigger provided by the BBC requires at least one hit in both directions. The BBC trigger efficiency is 55% in , 72% in , and 84% in collisions. The MuID provides a trigger to select events containing hadron or muon tracks by requiring at least one MuID track reaching Gap2, Gap3 or Gap4. For charged hadrons, events with track candidates having a hit at the Gap2 or Gap3 but no hit at the Gap4 are selected. The new readout electronics for the MuTr allow to trigger events with high momentum (p_T>3  GeV/c) tracks by requiring maximum bending at the middle plane less than 3 MuTr cathode strips <cit.>. §.§ RHIC polarized beams RHIC is the world's first and only polarized proton collider, located at Brookhaven National Laboratory, comprising two counter-circulating storage rings. In each ring, as many as 120 bunches of heavy ions or polarized protons can be accelerated up to 100 GeV for heavy ions and 255 GeV for polarized protons. During the 2015 run period, data from the collisions of vertically polarized protons and ions (, , and ) at √(s__NN)=200 GeV were recorded. The polarization direction of the colliding bunches altered every bunch crossing (106 ns interval) to minimize systematic effects due to time dependence of the detector and accelerator performance. In the run, the counterclockwise direction was selected for ion beams. The average polarization of the proton beam was 58% (clockwise beam) and 57% (counterclockwise beam) for collisions, 58% for collisions, and 61% for collisions with 3% relative uncertainty from the polarization normalization <cit.>. § DATA ANALYSIS §.§ Data Set The integrated luminosity of the data in this analysis is 37 pb^-1 from , 593 nb^-1 from , and 112 nb^-1 from collisions. The data was recorded using the hadron trigger in combination with the BBC trigger. The hadron trigger selected high-momentum tracks stopped at MuID Gap2 or Gap3, with the bending at the middle plane of MuTr being less than 3 cathode strips <cit.>. §.§ Charged Hadron Selection Among the reconstructed tracks stopped at MuID Gap2 or Gap3 are low momentum muons from light hadron decays. From full geant4 simulation studies, such background is significantly suppressed (<5%) after applying a cut p_z>3.5 GeV/c <cit.>. The additional p_T dependent track quality cuts are listed in Table <ref> with numerical values shown for Gap3 tracks at 1.25<p_T<1.5  GeV/c. The distance and angular difference between the extrapolated MuTr track and the MuID track at the first MuID plane's z position are called DG0 and DDG0. The distance between the interaction point and a projected position of a MuID track at z=0 is r_ ref. The number of hits in a MuTr track and χ^2 per degree of freedom (χ^2_ MuTr/ndf) are also included in the track quality cut. The polar scattering angle of a track inside the absorber scaled by the momentum is p·(θ_ MuTr - θ_ vtx) where the angle at the vertex is θ_ vtx and the angle at the MuTr Station 1 is θ_ MuTr. Cuts applied to p·(θ_ MuTr - θ_ vtx) and χ^2 at z_ vtx are effective for rejecting muons decayed inside the absorber. An estimation of the particle composition in the measured charged hadron sample was developed in  <cit.>. Based on pythia <cit.> and hijing <cit.> event generators, charged hadron spectra measured at midrapidity in and at RHIC <cit.> were extrapolated to the PHENIX muon arm rapidity (1.2<η<2.4) for , , and collisions. According to a geant4 <cit.> detector simulation (release 10.0.p02), the initial particle composition is modified due to interaction with the detector material including the front absorber. The reconstructed charged hadrons are mostly K^± and π^± where others (p, p̅) are less than 10%. The estimated K/π ratios at the collision vertex and reconstructed ratios from the simulation are shown in Fig. <ref> for collisions and Fig. <ref> for collisions. K/π ratios in and collisions are consistent with each other in the hijing simulation. Among the geant4 physics lists <cit.>, the QGSP-BERT is used as the default. To simulate hadronic interactions it applies the quark gluon string model for high energy and the Bertini cascade model for low energy hadrons. To estimate possible variations, the FTFP-BERT and QGSP-BIC physics lists are used where the FTF model uses the FRITIOF description of string excitation and fragmentation and BIC uses the binary cascade model. The variations from the different physics lists in the geant4 simulation are shown as the bands on the reconstructed K/π ratios. §.§ Determination of the TSSA The unbinned maximum-likelihood method was used to extract the TSSA (A_N) in this study. It was established in the previous study which used the same detectors <cit.>. The method is robust even for low-statistics data compared to binned approaches. The extended log-likelihood logℒ is defined as logℒ = ∑_i log(1 + P · A_N sin(ϕ_ pol - ϕ_h^i)) + constant, where P is the proton beam polarization and ϕ_h^i is the azimuthal angle of the i-th hadron with respect to the incoming polarized-proton-beam direction. The ϕ_ pol is the azimuthal angle for the beam polarization direction, which takes the values of +π/2/-π/2 for ↑/↓ polarized beam bunches in 2015. The A_N is chosen to be the value which maximizes logℒ. The statistical uncertainty of the log-likelihood estimator is calculated from the second derivative of ℒ with respect to A_N, σ^2(A_N) = (-∂^2 ℒ/∂ A_N^2)^-1. In p+p collisions, both beams were polarized. The A_Ns were measured separately for each beam and turned out to be consistent with each other. They were averaged for the final asymmetry. The A_N is checked with the azimuthal-fitting method used in previous analyses <cit.> based on the polarization formula <cit.>: A_N(ϕ)=σ^↑(ϕ)-σ^↓(ϕ)/σ^↑(ϕ)+σ^↓(ϕ) =1/P·N^↑(ϕ) -R· N^↓(ϕ)/N^↑(ϕ) +R· N^↓(ϕ), where A_N(ϕ) indicates the simple-count-based transverse single-spin asymmetry calculated for each of 16 azimuthal ϕ-bins; σ^↑, σ^↓ represent cross sections; and N^↑, N^↓ are yields for each polarization of spin up (↑) or down (↓). R = L^↑/L^↓ is the ratio of luminosities (relative luminosity) between bunches with spin up and spin down where the luminosity is determined by sampled counts from minimum-bias triggers for the corresponding spin orientation. The A_N in this method is calculated from the fit of A_N(ϕ) distributions using a function A_N·sin(ϕ_ pol↑ - ϕ) where ϕ_ pol ↑=π/2 indicates the azimuthal direction of upward polarized bunches. The difference of the A_N's for this method and the maximum-likelihood method is conservatively included in the systematic uncertainty. The interactions between particles and material prior to being detected by the MuTr, and the finite resolution for momentum and azimuthal angle ϕ, may result in a kinematic smearing effect for the A_N results. The full geant4 simulation is used to correct this effect. The effect from ϕ smearing was found to be negligible. The momentum smearing is shown in Fig. <ref>; for each bin of reconstructed x_F is shown the probability distribution of generated x_F that contribute to each reconstructed bin. The momentum smearing effect is corrected by resolving a set of linear equations which connect A_N for reconstructed x_F (p_T) bins (A_N^ reco) and A_N for true (generated) x_F (p_T) bins (A_N^ true): A_N^ reco, m = ∑_i f^i → m· A_N^ true, i, where A_N^ true, i represents A_N for the i-th true momentum (x_F, p_T) bin and A_N^ reco, m is A_N for the reconstructed momentum in the m-th momentum bin. f^i → m is the fraction of charged hadrons reconstructed in the m-th momentum bin from the i-th true (generated) momentum bin in the simulation. A_N^ reco is measured including an underflow and overflow bin for 0.035<x_F<0.3 and 1.25<p_T<15 GeV/c, and then A_N^ true is calculated for 0.04<x_F<0.2 and 1.5<p_T<7 GeV/c. The difference between A_N^ reco and A_N^ true was found to be small, and was conservatively included in systematic uncertainties. The variation of A_N due to muon contribution is minimal (<0.0005) and not included in the systematic uncertainty. § RESULTS §.§ Results in Proton-Proton Collisions The TSSAs in the production of charged hadrons at in transversely-polarized proton-proton collisions () are shown in Fig. <ref> as a function of p_T and in Fig. <ref> as a function of x_F. The vertical bars represent statistical uncertainties and the boxes are systematic uncertainties. A scale uncertainty from the polarization (3%) is not included. The results are listed in Tables <ref>, <ref>, <ref>, and <ref>. In the tables, δ A_N^ stat is the statistical uncertainty, and δ A_N^ syst is the total systematic uncertainty. The systematic uncertainty is obtained from the quadratic sum of two components (δ A_N^ method and δ A_N^ smear). As explained in the previous section, δ A_N^ method is the difference between the two methods of determining A_N, while δ A_N^ smear is the difference caused by the momentum migration correction (Eq. <ref>). Three different physics lists are tested in the correction and the variation among them is negligible. On the panel (a) of Fig. <ref>, A_N at backward rapidity (x_F<0) for charged hadrons is consistent with zero within uncertainty. On the panel (b), A_N at forward rapidity (x_F>0) is positive for positively-charged hadrons. Positive asymmetry is also shown in the x_F-binned result in Fig. <ref>. A_N for positively-charged hadrons at x_F>0 shows an increasing trend with respect to x_F. The result for negatively-charged hadrons shows some indication of negative A_N around x_F>0.07. The previous results at RHIC energies with larger |x_F| and |η| than this measurement showed positive A_N for π^+, K^± and negative A_N for π^- at x_F>0. The size of A_N increases as a function of x_F at x_F>0. In negative x_F region, A_N is consistent with zero. In this measurement, the result for h^+ is comparable with A_N for π^+, K^+. A_N for h^+ is increasing as a function of x_F at x_F>0 and is consistent with zero at x_F<0. π^- and K^- in the previous measurements showed the opposite sign of A_N at forward rapidity. Therefore one may expect the A_N for h^- has been partially canceled and this can explain the smaller size of A_N for h^- at x_F>0 in Fig. <ref>. This is also confirmed in twist-3 model calculations shown in Fig. <ref>. The dotted bands in Fig. <ref> are obtained from K/π ratios in the simulation combined with theoretical A_N for charged pions and kaons from twist-3 calculation for x_F>0.15 <cit.>, which describes previous RHIC results at forward rapidity <cit.>. In addition, the effect of the K/π production ratio variation is tested for a relative ± 30% variation and the width of the band shows the possible variation of A_N mixture for h^+ and h^-. The variation of the K/π production ratio from different geant4 physics lists shown in Fig. <ref> is much smaller than the relative ± 30% that was considered. §.§ Results in Proton-Nucleus Collisions The A_N results of charged hadrons in proton-nucleus collisions (and ) are shown in Fig. <ref>, <ref>, and <ref> with results and listed in Tables <ref>, <ref>,  <ref>, and <ref> for and <ref>, <ref>, <ref>, and <ref> for . Results at x_F<0 in all collision systems (, ) are consistent with zero within total uncertainty. Results for negatively-charged hadrons from and collisions shown in Fig. <ref> and <ref> are also consistent with zero within total uncertainty, while A_N in collisions shows some indication of negative asymmetry at p_T<3.5 GeV/c on panel (b) (x_F>0) of Fig. <ref> and at 0.07<x_F<0.10 on panel (a) of Fig. <ref>. However, the difference of A_N between and is not significant to state any modification of |A_N| in compared to results. The A_N for positively-charged hadrons at x_F>0 in collisions is consistent with zero while the results for represents a positive asymmetry on panel (b) in Fig. <ref> and on panel (b) in Fig. <ref>. The results for h^+ at x_F>0 show some indication of a positive asymmetry, which is smaller than in collisions. § DISCUSSION The results for the production of positively-charged hadrons (h^+) in collisions agree with a trend with other RHIC data where A_N for π^+ and K^+ is positive and increasing with respect to x_F in the forward region (x_F>0) <cit.>. The results for A_N for negatively-charged hadrons (h^-) range from slightly negative to zero at x_F>0; this may be caused by cancellation between positive asymmetry for K^- and negative asymmetry for π^- shown in RHIC data at larger x_F <cit.>. These agree with the trend of the theoretical calculation at x_F>0.15 <cit.>. In all collision systems, A_N at backward rapidity (x_F<0) is consistent with zero asymmetry within total uncertainty, which also has been shown in previous RHIC measurements in collisions  <cit.>. In the A_N of positively charged hadrons measured from and , the nuclear dependence is evident in the range 0.1<x_F<0.2 where the nonzero asymmetry in contrasts to the approximately zero A_N in collisions; likewise, A_N in is smaller than A_N in . The A-dependence of A_N at 0.1<x_F<0.2 was found to be consistent with A_N∝ A^-α/3 in our previous publication <cit.> where α= 1.10^+0.69_-0.36(stat)^+0.06_-0.05(sys). Recent calculations using collinear factorization on polarized proton (p^↑) and the CGC framework on nuclear target (A) <cit.> predicted that the A_N of inclusive hadrons at forward rapidity in collisions has two contributions where one is A-independent and the other is A^-1/3-dependent <cit.>. However, the A^-1/3-dependent term is dominant for p_T less than Q_s and A-independent term is dominant for higher p_T. In the kinematics of this measurement according to a recent calculation <cit.>, Q_s^ Au≈ 0.9 GeV whereas our measurements correspond to ⟨ p_T⟩≈ 2.9 GeV/c at 0.1<x_F<0.2. An approach using lensing mechanism predicts that A_N decreases as the atomic number of the target (A) increases for k_T below or near Q_s <cit.>. A recent STAR result of π^0 in the more forward region at 2.6<η<4.0, p_T>1.5  GeV/c, and 0.2<x_F<0.7 in and collisions shows a smaller A-dependence than what is observed here <cit.>. Given that our measurement and that from STAR used different kinematic ranges and hadron species, further detailed studies of various observables within a wide kinematic range will be informative on the origin of A_N and its interplay with small-x phenomena. § SUMMARY Reported here are the transverse single-spin asymmetry (A_N) of positively and negatively charged hadrons (h^±) in forward and backward rapidity (1.4<|η|<2.4) over the range of 1.5<p_T<7.0 GeV/c and 0.04<|x_F|<0.2 from transversely polarized proton-proton () and proton-nucleus () collisions. The results at x_F<0 are consistent with zero at all systems. In x_F>0, negative charged hadron results show small to zero A_N in collisions and zero A_N in collisions. A_N for positively charged hadrons increases to positive values as x_F increases in collisions, but results show suppression of A_N at 0.1<x_F<0.2 compared to the result. The results will aid in understanding of the origin of A_N and offer a tool to investigate nuclear effects and phenomena in small-x. We thank the staff of the Collider-Accelerator and Physics Departments at Brookhaven National Laboratory and the staff of the other PHENIX participating institutions for their vital contributions. We also thank D. Pitonyak for the theory calculation. We acknowledge support from the Office of Nuclear Physics in the Office of Science of the Department of Energy, the National Science Foundation, Abilene Christian University Research Council, Research Foundation of SUNY, and Dean of the College of Arts and Sciences, Vanderbilt University (U.S.A), Ministry of Education, Culture, Sports, Science, and Technology and the Japan Society for the Promotion of Science (Japan), Natural Science Foundation of China (People's Republic of China), Croatian Science Foundation and Ministry of Science and Education (Croatia), Ministry of Education, Youth and Sports (Czech Republic), Centre National de la Recherche Scientifique, Commissariat à l'Énergie Atomique, and Institut National de Physique Nucléaire et de Physique des Particules (France), J. Bolyai Research Scholarship, EFOP, the New National Excellence Program (ÚNKP), NKFIH, and OTKA (Hungary), Department of Atomic Energy and Department of Science and Technology (India), Israel Science Foundation (Israel), Basic Science Research and SRC(CENuM) Programs through NRF funded by the Ministry of Education and the Ministry of Science and ICT (Korea). Ministry of Education and Science, Russian Academy of Sciences, Federal Agency of Atomic Energy (Russia), VR and Wallenberg Foundation (Sweden), University of Zambia, the Government of the Republic of Zambia (Zambia), the U.S. Civilian Research and Development Foundation for the Independent States of the Former Soviet Union, the Hungarian American Enterprise Scholarship Fund, the US-Hungarian Fulbright Foundation, and the US-Israel Binational Science Foundation. * § DATA TABLES Tables <ref>, <ref>, <ref>, and <ref> list the TSSAs in the production of charged hadrons at in transversely-polarized proton-proton collisions () that are shown in Fig. <ref> as a function of p_T and in Fig. <ref> as a function of x_F. See further details in the subsection “A. Results in Proton-Nucleus Collisions." Tables <ref>, <ref>,  <ref>, and <ref> for and <ref>, <ref>, <ref>, and <ref> for . list the A_N results of charged hadrons in proton-nucleus collisions (and ) that are shown in Figs. <ref>, <ref>, and <ref> with results. See further details in the subsection “B. Results in Proton-Nucleus Collisions." 59 fxundefined [1] ifx#1 fnum [1] #1firstoftwo secondoftwo fx [1] #1firstoftwo secondoftwo noop [0]secondoftwo ref[1]@startlink#1@href href[1]#1@endlink anitize@url [0]` 12`$12`&12`#12`1̂2`_12`%12 startlink[1] endlink[0] rl [1]href #1 @bib@innerbibempty [Klem et al.(1976)Klem, Bowers, Courant, Kagan, Marshak, Peterson, Ruddick, Dragoset, and Roberts]Klem:1976ui author author R. D. Klem, author J. E. Bowers, author H. W. Courant, author H. Kagan, author M. L. Marshak, author E. A. Peterson, author K. Ruddick, author W. H. Dragoset, and author J. B. Roberts, title title Measurement of Asymmetries of Inclusive Pion Production in Proton Proton Interactions at 6 GeV/c and 11.8 GeV/c, https://doi.org/10.1103/PhysRevLett.36.929 journal journal Phys. Rev. Lett. volume 36, pages 929 (year 1976)NoStop [Antille et al.(1980)Antille, Dick, Madansky, Perret-Gallix, Werlen, Gonidec, Kuroda, and Kyberd]Antille:1980th author author J. Antille, author L. Dick, author L. Madansky, author D. Perret-Gallix, author M. Werlen, author A. Gonidec, author K. Kuroda, and author P. Kyberd, title title Spin Dependence of the Inclusive Reaction p p (Polarized) →π^0 X at 24 GeV/c for High p_T π^0 Produced in the Central Region, https://doi.org/10.1016/0370-2693(80)90933-8 journal journal Phys. Lett. B volume 94, pages 523 (year 1980)NoStop [Adams et al.(1991a)Adams et al.]FNAL-E704:1991ovg author author D. L. Adams et al. (collaboration FNAL-E704 Collaboration), title title Analyzing power in inclusive π^+ and π^- production at high x_F with a 200 GeV polarized proton beam, https://doi.org/10.1016/0370-2693(91)90378-4 journal journal Phys. Lett. B volume 264, pages 462 (year 1991a)NoStop [Adams et al.(1991b)Adams et al.]E581:1991eys author author D. L. Adams et al. (collaboration E581, E704 Collaboration), title title Comparison of spin asymmetries and cross-sections in π^0 production by 200 GeV polarized antiprotons and protons, https://doi.org/10.1016/0370-2693(91)91351-U journal journal Phys. Lett. B volume 261, pages 201 (year 1991b)NoStop [Allgower et al.(2002)Allgower et al.]Allgower:2002qi author author C. E. Allgower et al., title title Measurement of analyzing powers of π^+ and π^- produced on a hydrogen and a carbon target with a 22-GeV/c incident polarized proton beam, https://doi.org/10.1103/PhysRevD.65.092008 journal journal Phys. Rev. D volume 65, pages 092008 (year 2002)NoStop [Adams et al.(2004)Adams et al.]STAR:2003lxu author author J. Adams et al. (collaboration STAR Collaboration), title title Cross-sections and transverse single-spin asymmetries in forward neutral pion production from proton collisions at √(s) = 200 GeV, https://doi.org/10.1103/PhysRevLett.92.171801 journal journal Phys. Rev. Lett. volume 92, pages 171801 (year 2004)NoStop [Lee and Videbaek(2007)]Lee:2007zzh author author J. H. Lee and author F. Videbaek (collaboration BRAHMS Collaboration), title title Single-spin asymmetries of identified hadrons in polarized p+p at √(s) = 62.4 and 200 GeV, https://doi.org/10.1063/1.2750837 journal journal AIP Conf. Proc. volume 915, pages 533 (year 2007)NoStop [Abelev et al.(2008)Abelev et al.]STAR:2008ixi author author B. I. Abelev et al. (collaboration STAR Collaboration), title title Forward Neutral Pion Transverse Single-Spin Asymmetries in p+p Collisions at √(s) = 200 GeV, https://doi.org/10.1103/PhysRevLett.101.222001 journal journal Phys. Rev. Lett. volume 101, pages 222001 (year 2008)NoStop [Arsene et al.(2008)Arsene et al.]BRAHMS:2008doi author author I. Arsene et al. (collaboration BRAHMS Collaboration), title title Single Transverse Spin Asymmetries of Identified Charged Hadrons in Polarized pp Collisions at √(s) = 62.4 GeV, https://doi.org/10.1103/PhysRevLett.101.042001 journal journal Phys. Rev. Lett. volume 101, pages 042001 (year 2008)NoStop [Adamczyk et al.(2012)Adamczyk et al.]STAR:2012ljf author author L. Adamczyk et al. (collaboration STAR Collaboration), title title Transverse Single-Spin Asymmetry and Cross-Section for π^0 and η Mesons at Large Feynman-x in Polarized p+p Collisions at √(s)=200 GeV, https://doi.org/10.1103/PhysRevD.86.051101 journal journal Phys. Rev. D volume 86, pages 051101 (year 2012)NoStop [Adare et al.(2014a)Adare et al.]PHENIX:2013wle author author A. Adare et al. (collaboration PHENIX Collaboration), title title Measurement of transverse-single-spin asymmetries for midrapidity and forward-rapidity production of hadrons in polarized p+p collisions at √(s)=200 and 62.4 GeV, https://doi.org/10.1103/PhysRevD.90.012006 journal journal Phys. Rev. D volume 90, pages 012006 (year 2014a)NoStop [Adare et al.(2014b)Adare et al.]PHENIX:2014qwb author author A. Adare et al. (collaboration PHENIX Collaboration), title title Cross section and transverse single-spin asymmetry of η mesons in p^↑+p collisions at √(s)=200 GeV at forward rapidity, https://doi.org/10.1103/PhysRevD.90.072008 journal journal Phys. Rev. D volume 90, pages 072008 (year 2014b)NoStop [Adam et al.(2021a)Adam et al.]STAR:2020nnl author author J. Adam et al. (collaboration STAR Collaboration), title title Measurement of transverse single-spin asymmetries of π^0 and electromagnetic jets at forward rapidity in 200 and 500 GeV transversely polarized proton-proton collisions, https://doi.org/10.1103/PhysRevD.103.092009 journal journal Phys. Rev. D volume 103, pages 092009 (year 2021a)NoStop [Sivers(1990)]Sivers:1989cc author author D. W. Sivers, title title Single-Spin Production Asymmetries from the Hard Scattering of Point-Like Constituents, https://doi.org/10.1103/PhysRevD.41.83 journal journal Phys. Rev. D volume 41, pages 83 (year 1990)NoStop [Sivers(1991)]Sivers:1990fh author author D. W. Sivers, title title Hard scattering scaling laws for single-spin production asymmetries, https://doi.org/10.1103/PhysRevD.43.261 journal journal Phys. Rev. D volume 43, pages 261 (year 1991)NoStop [Collins(1993)]Collins:1992kk author author J. C. Collins, title title Fragmentation of transversely polarized quarks probed in transverse momentum distributions, https://doi.org/10.1016/0550-3213(93)90262-N journal journal Nucl. Phys. B volume 396, pages 161 (year 1993)NoStop [Qiu and Sterman(1999)]Qiu:1998ia author author J. Qiu and author G. F. Sterman, title title Single transverse spin asymmetries in hadronic pion production, https://doi.org/10.1103/PhysRevD.59.014004 journal journal Phys. Rev. D volume 59, pages 014004 (year 1999)NoStop [Kouvaris et al.(2006)Kouvaris, Qiu, Vogelsang, and Yuan]Kouvaris:2006zy author author C. Kouvaris, author J. Qiu, author W. Vogelsang, and author F. Yuan, title title Single transverse-spin asymmetry in high transverse momentum pion production in pp collisions, https://doi.org/10.1103/PhysRevD.74.114013 journal journal Phys. Rev. D volume 74, pages 114013 (year 2006)NoStop [Koike and Tanaka(2007)]Koike:2007rq author author Y. Koike and author K. Tanaka, title title Universal structure of twist-3 soft-gluon-pole cross-sections for single transverse-spin asymmetry, https://doi.org/10.1103/PhysRevD.76.011502 journal journal Phys. Rev. D volume 76, pages 011502 (year 2007)NoStop [Koike and Tomita(2009)]Koike:2009ge author author Y. Koike and author T. Tomita, title title Soft-fermion-pole contribution to single-spin asymmetry for pion production in pp collisions, https://doi.org/10.1016/j.physletb.2009.04.017 journal journal Phys. Lett. B volume 675, pages 181 (year 2009)NoStop [Kanazawa and Koike(2010)]Kanazawa:2010au author author K. Kanazawa and author Y. Koike, title title New Analysis of the Single Transverse-Spin Asymmetry for Hadron Production at RHIC, https://doi.org/10.1103/PhysRevD.82.034009 journal journal Phys. Rev. D volume 82, pages 034009 (year 2010)NoStop [Kang et al.(2011)Kang, Qiu, Vogelsang, and Yuan]Kang:2011hk author author Z. Kang, author J. Qiu, author W. Vogelsang, and author F. Yuan, title title An Observation Concerning the Process Dependence of the Sivers Functions, https://doi.org/10.1103/PhysRevD.83.094001 journal journal Phys. Rev. D volume 83, pages 094001 (year 2011)NoStop [Kanazawa and Koike(2011)]Kanazawa:2011bg author author K. Kanazawa and author Y. Koike, title title A phenomenological study on single transverse-spin asymmetry for inclusive light-hadron productions at RHIC, https://doi.org/10.1103/PhysRevD.83.114024 journal journal Phys. Rev. D volume 83, pages 114024 (year 2011)NoStop [Kang and Prokudin(2012)]Kang:2012xf author author Z. Kang and author A. Prokudin, title title Global fitting of single-spin asymmetry: an attempt, https://doi.org/10.1103/PhysRevD.85.074008 journal journal Phys. Rev. D volume 85, pages 074008 (year 2012)NoStop [Beppu et al.(2014)Beppu, Kanazawa, Koike, and Yoshida]Beppu:2013uda author author H. Beppu, author K. Kanazawa, author Y. Koike, and author S. Yoshida, title title Three-gluon contribution to the single-spin asymmetry for light hadron production in pp collision, https://doi.org/10.1103/PhysRevD.89.034029 journal journal Phys. Rev. D volume 89, pages 034029 (year 2014)NoStop [Metz and Pitonyak(2013)]Metz:2012ct author author A. Metz and author D. Pitonyak, title title Fragmentation contribution to the transverse single-spin asymmetry in proton-proton collisions, https://doi.org/10.1016/j.physletb.2013.05.043 journal journal Phys. Lett. B volume 723, pages 365 (year 2013), note [Phys. Lett. B 762, 549(E) (2016)]NoStop [Kanazawa et al.(2014)Kanazawa, Koike, Metz, and Pitonyak]Kanazawa:2014dca author author K. Kanazawa, author Y. Koike, author A. Metz, and author D. Pitonyak, title title Towards an explanation of transverse single-spin asymmetries in proton-proton collisions: the role of fragmentation in collinear factorization, https://doi.org/10.1103/PhysRevD.89.111501 journal journal Phys. Rev. D volume 89, pages 111501 (year 2014)NoStop [Gamberg et al.(2017)Gamberg, Kang, Pitonyak, and Prokudin]Gamberg:2017gle author author L. Gamberg, author Z. Kang, author D. Pitonyak, and author A. Prokudin, title title Phenomenological constraints on A_N in p^↑ p→π X from Lorentz invariance relations, https://doi.org/10.1016/j.physletb.2017.04.061 journal journal Phys. Lett. B volume 770, pages 242 (year 2017)NoStop [Cammarota et al.(2020)Cammarota, Gamberg, Kang, Miller, Pitonyak, Prokudin, Rogers, and Sato]Cammarota:2020qcw author author J. Cammarota, author L. Gamberg, author Z. Kang, author J. A. Miller, author D. Pitonyak, author A. Prokudin, author T. C. Rogers, and author N. Sato (collaboration Jefferson Lab Angular Momentum Collaboration), title title Origin of single transverse-spin asymmetries in high-energy collisions, https://doi.org/10.1103/PhysRevD.102.054002 journal journal Phys. Rev. D volume 102, pages 054002 (year 2020)NoStop [Gelis et al.(2010)Gelis, Iancu, Jalilian-Marian, and Venugopalan]Gelis:2010nm author author F. Gelis, author E. Iancu, author J. Jalilian-Marian, and author R. Venugopalan, title title The Color Glass Condensate, https://doi.org/10.1146/annurev.nucl.010909.083629 journal journal Ann. Rev. Nucl. Part. Sci. volume 60, pages 463 (year 2010)NoStop [Kang and Yuan(2011)]Kang:2011ni author author Z. Kang and author F. Yuan, title title Single-Spin Asymmetry Scaling in the Forward Rapidity Region at RHIC, https://doi.org/10.1103/PhysRevD.84.034019 journal journal Phys. Rev. D volume 84, pages 034019 (year 2011)NoStop [Kovchegov and Sievert(2012)]Kovchegov:2012ga author author Y. V. Kovchegov and author M. D. Sievert, title title A New Mechanism for Generating a Single Transverse Spin Asymmetry, https://doi.org/10.1103/PhysRevD.86.034028 journal journal Phys. Rev. D volume 86, pages 034028 (year 2012), note [Phys. Rev. D 86, 079906(E) (2012)]NoStop [Schäfer and Zhou(2014)]Schafer:2014zea author author A. Schäfer and author J. Zhou, title title Transverse single-spin asymmetry in direct photon production in polarized pA collisions, https://doi.org/10.1103/PhysRevD.90.034016 journal journal Phys. Rev. D volume 90, pages 034016 (year 2014)NoStop [Zhou(2015)]Zhou:2015ima author author J. Zhou, title title Transverse single-spin asymmetry in Drell-Yan production in polarized pA collisions, https://doi.org/10.1103/PhysRevD.92.014034 journal journal Phys. Rev. D volume 92, pages 014034 (year 2015)NoStop [Hatta et al.(2016)Hatta, Xiao, Yoshida, and Yuan]Hatta:2016wjz author author Y. Hatta, author B. Xiao, author S. Yoshida, and author F. Yuan, title title Single-Spin Asymmetry in Forward pA Collisions, https://doi.org/10.1103/PhysRevD.94.054013 journal journal Phys. Rev. D volume 94, pages 054013 (year 2016)NoStop [Hatta et al.(2017)Hatta, Xiao, Yoshida, and Yuan]Hatta:2016khv author author Y. Hatta, author B. Xiao, author S. Yoshida, and author F. Yuan, title title Single-spin asymmetry in forward pA collisions II: Fragmentation contribution, https://doi.org/10.1103/PhysRevD.95.014008 journal journal Phys. Rev. D volume 95, pages 014008 (year 2017)NoStop [Aidala et al.(2019)Aidala et al.]PHENIX:2019ouo author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Nuclear Dependence of the Transverse Single-Spin Asymmetry in the Production of Charged Hadrons at Forward Rapidity in Polarized p+p, p+Al, and p+Au Collisions at √(s__NN)=200 GeV, https://doi.org/10.1103/PhysRevLett.123.122001 journal journal Phys. Rev. Lett. volume 123, pages 122001 (year 2019)NoStop [Adam et al.(2021b)Adam et al.]STAR:2020grs author author J. Adam et al. (collaboration STAR Collaboration), title title Comparison of transverse single-spin asymmetries for forward π^0 production in polarized pp, pAl and pAu collisions at nucleon pair c.m. energy √(s_NN)= 200 GeV, https://doi.org/10.1103/PhysRevD.103.072005 journal journal Phys. Rev. D volume 103, pages 072005 (year 2021b)NoStop [Adcox et al.(2003)Adcox et al.]PHENIX:2003nhg author author K. Adcox et al. (collaboration PHENIX Collaboration), title title PHENIX detector overview, https://doi.org/10.1016/S0168-9002(02)01950-2 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 499, pages 469 (year 2003)NoStop [Akikawa et al.(2003)Akikawa et al.]PHENIX:2003yhi author author H. Akikawa et al. (collaboration PHENIX Collaboration), title title PHENIX muon arms, https://doi.org/10.1016/S0168-9002(02)01955-1 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 499, pages 537 (year 2003)NoStop [Adare et al.(2012)Adare et al.]PHENIX:2012itj author author A. Adare et al. (collaboration PHENIX Collaboration), title title Nuclear-Modification Factor for Open-Heavy-Flavor Production at Forward Rapidity in Cu+Cu Collisions at √(s_NN)=200 GeV, https://doi.org/10.1103/PhysRevC.86.024909 journal journal Phys. Rev. C volume 86, pages 024909 (year 2012)NoStop [Allen et al.(2003)Allen et al.]PHENIX:2003tlh author author M. Allen et al. (collaboration PHENIX Collaboration), title title PHENIX inner detectors, https://doi.org/10.1016/S0168-9002(02)01956-3 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 499, pages 549 (year 2003)NoStop [Adachi et al.(2013)Adachi et al.]Adachi:2013qha author author S. Adachi et al., title title Trigger electronics upgrade of PHENIX muon tracker, https://doi.org/10.1016/j.nima.2012.11.088 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 703, pages 114 (year 2013)NoStop [RHIC Polarimetry Group()]polarimetry author author RHIC Polarimetry Group, title title RHIC polarization for Runs 9–12, note RHIC/CAD Accelerator Physics Note 490 (2018)NoStop [Aidala et al.(2020)Aidala et al.]PHENIX:2019gix author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Nuclear-modification factor of charged hadrons at forward and backward rapidity in p+Al and p+Au collisions at √(s__NN)=200 GeV, https://doi.org/10.1103/PhysRevC.101.034910 journal journal Phys. Rev. C volume 101, pages 034910 (year 2020)NoStop [Adare et al.(2014c)Adare et al.]PHENIX:2013txu author author A. Adare et al. (collaboration PHENIX Collaboration), title title Cold-Nuclear-Matter Effects on Heavy-Quark Production at Forward and Backward Rapidity in d+Au Collisions at √(s_NN)=200 GeV, https://doi.org/10.1103/PhysRevLett.112.252301 journal journal Phys. Rev. Lett. volume 112, pages 252301 (year 2014c)NoStop [Sjostrand et al.()Sjostrand, Mrenna, and Skands]Sjostrand:2006za author author T. Sjostrand, author S. Mrenna, and author P. Z. Skands, https://doi.org/10.1088/1126-6708/2006/05/026 title PYTHIA 6.4 Physics and Manual, note J. High Energy Phys. 05 (2006) 026NoStop [Gyulassy and Wang(1994)]Gyulassy:1994ew author author M. Gyulassy and author X. Wang, title title HIJING 1.0: A Monte Carlo program for parton and particle production in high-energy hadronic and nuclear collisions, https://doi.org/10.1016/0010-4655(94)90057-4 journal journal Comput. Phys. Commun. volume 83, pages 307 (year 1994)NoStop [Adare et al.(2011)Adare et al.]PHENIX:2011rvu author author A. Adare et al. (collaboration PHENIX Collaboration), title title Identified charged hadron production in p+p collisions at √(s)=200 and 62.4 GeV, https://doi.org/10.1103/PhysRevC.83.064903 journal journal Phys. Rev. C volume 83, pages 064903 (year 2011)NoStop [Agakishiev et al.(2012)Agakishiev et al.]STAR:2011iap author author G. Agakishiev et al. (collaboration STAR Collaboration), title title Identified hadron compositions in p+p and Au+Au collisions at high transverse momenta at √(s__NN) = 200 GeV, https://doi.org/10.1103/PhysRevLett.108.072302 journal journal Phys. Rev. Lett. volume 108, pages 072302 (year 2012)NoStop [Adare et al.(2013)Adare et al.]PHENIX:2013kod author author A. Adare et al. (collaboration PHENIX Collaboration), title title Spectra and ratios of identified particles in Au+Au and d+Au collisions at √(s_NN)=200 GeV, https://doi.org/10.1103/PhysRevC.88.024906 journal journal Phys. Rev. C volume 88, pages 024906 (year 2013)NoStop [Agostinelli et al.(2003)Agostinelli et al.]GEANT4:2002zbu author author S. Agostinelli et al. (collaboration GEANT4 Collaboration), title title GEANT4–a simulation toolkit, https://doi.org/10.1016/S0168-9002(03)01368-8 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 506, pages 250 (year 2003)NoStop [Allison et al.(2016)Allison et al.]Allison:2016lfl author author J. Allison et al., title title Recent developments in Geant4, https://doi.org/10.1016/j.nima.2016.06.125 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 835, pages 186 (year 2016)NoStop [Aidala et al.(2017)Aidala et al.]PHENIX:2017wbv author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Cross section and transverse single-spin asymmetry of muons from open heavy-flavor decays in polarized p+p collisions at √(s)=200 GeV, https://doi.org/10.1103/PhysRevD.95.112001 journal journal Phys. Rev. D volume 95, pages 112001 (year 2017)NoStop [Aidala et al.(2018)Aidala et al.]PHENIX:2018qvl author author C. Aidala et al. (collaboration PHENIX Collaboration), title title Single-spin asymmetry of J/ψ production in p+p, p+Al, and p+Au collisions with transversely polarized proton beams at √(s__NN)=200 GeV, https://doi.org/10.1103/PhysRevD.98.012006 journal journal Phys. Rev. D volume 98, pages 012006 (year 2018)NoStop [Ohlsen and Keaton(1973)]Ohlsen:1973wf author author G. G. Ohlsen and author P. W. Keaton, title title Techniques for measurement of spin-1/2 and spin-1 polarization analyzing tensors, https://doi.org/10.1016/0029-554X(73)90450-3 journal journal Nucl. Instrum. Methods Phys. Res., Sec. A volume 109, pages 41 (year 1973)NoStop [Zhou(2017)]Zhou:2017sdx author author J. Zhou, title title Single-spin asymmetries in forward p-p/A collisions revisited: the role of color entanglement, https://doi.org/10.1103/PhysRevD.96.034027 journal journal Phys. Rev. D volume 96, pages 034027 (year 2017)NoStop [Benić and Hatta(2019)]Benic:2018amn author author S. Benić and author Y. Hatta, title title Single-spin asymmetry in forward pA collisions: Phenomenology at RHIC, https://doi.org/10.1103/PhysRevD.99.094012 journal journal Phys. Rev. D volume 99, pages 094012 (year 2019)NoStop [Kovchegov and Santiago(2020)]Kovchegov:2020kxg author author Y. V. Kovchegov and author M. G. Santiago, title title Lensing mechanism meets small- x physics: Single transverse spin asymmetry in p^↑+p and p^↑+A collisions, https://doi.org/10.1103/PhysRevD.102.014022 journal journal Phys. Rev. D volume 102, pages 014022 (year 2020)NoStop

http://arxiv.org/abs/2303.07330v1

20230313175142

Optical activity and transport in twisted bilayer graphene: the essence of spatial dispersion effects

cond-mat.mes-hall

Preprint not for distribution CONFIDENTIAL, Version of March 30, 2023 Hanoi University of Civil Engineering (HUCE), 55 Giai Phong road, Hanoi 10000, Vietnam nam.dovan@phenikaa-uni.edu.vn Department of Basic Science, Phenikaa Institute for Advanced Studies (PIAS), 25^th Floor, A9 Building, Phenikaa University, Hanoi 10000, Vietnam This study investigates optical activity and quantum transport in twisted bilayer graphene (TBG) systems, demonstrating that the former results from spatial dispersion effects. The transfer matrix method is used to solve the propagation of electromagnetic waves through two graphene layers that act as the coupling surfaces of a dielectric slab. The resulting optical conductivity tensor is decomposed into a local and a drag part, with the drag transverse conductivity σ_xy^(drag) governing the TBG system's optical activity. An effective continuum model is employed to analyze electron state formation and calculate relevant parts of the optical conductivity tensor. Correlation of electron motions leads to incomplete cancellation and a finite σ_xy^(drag) in the chiral TBG lattice. The study also calculates DC conductivity, showing TBG supports quantum conductivity proportional to e^2/h at the intrinsic Fermi energy. Optical activity and transport in twisted bilayer graphene: the essence of spatial dispersion effects V. Nam Do March 30, 2023 ===================================================================================================== § INTRODUCTION The study of two-dimensional (2D) materials has received significant attention in recent years due to its potential applications in electronics, optoelectronics, spintronics, and valleytronics.<cit.> The use of multiple-layer materials with van der Waals (vdW) interlayer coupling offers various means to tune the electronic structure, such as altering the interlayer distance, changing the lattice alignment through twisting or sliding the layers, and applying external fields. Physically, these material engineering solutions often result in breaking the spatial and temporal symmetries that govern the dynamics of electrons within the material layers. For instance, a magnetic field breaks the time-reversal symmetry, leading to exciting phenomena such as the quantum Hall effect, magneto-optical effects, and the Kerr and Faraday rotations of the polarization plane of linearly polarized light and circular dichroism.<cit.> These intriguing optical phenomena have been utilized in the development of various devices for various applications.<cit.> However, the use of external fields has limitations for small-scale device design. As a result, it is crucial to search for materials with intrinsic properties and solutions that break the spatial symmetries of an electronic system to meet technological demands. Twisted-bilayer graphene (TBG) is a 2D van der Waals (2D vdW) material that has garnered significant attention from the research community due to its interesting electronic properties.<cit.> The TBG system is formed by stacking two graphene layers on top of each other with a rotation, or twist, angle between the two hexagonal lattices. The TBG system has been shown to be optically active, meaning that it has the ability to rotate the polarization plane of linearly polarized light by reflection and transmission and to absorb left- and right-handed light differently.<cit.> This, combined with the ability to mechanically tune the twist angle,<cit.> makes TBG a potential candidate for advanced optical applications. The optical activity of a dielectric medium is typically attributed to the anisotropy of its atomic lattice,<cit.> or to the magneto-electric effect.<cit.> External electric, magnetic, and mechanical forces can also induce anisotropy in a medium.<cit.> It has been shown that strained monolayer graphene is optically active due to the presence of the off-diagonal element σ_xy(ω) of the optical conductivity tensor.<cit.> The AA- and AB-stacked configurations of the TBG system are isotropic and characterized by the point groups D_6h and D_3d, respectively.<cit.> The breaking of the lattice symmetry from the D_6h and D_3d groups to the D_6 and D_3 groups, respectively, has been suggested as the mechanism causing the optical activity in TBG.<cit.> However, even with this symmetry, the TBG system is characterized by a total diagonal optical conductivity tensor, i.e., with zero off-diagonal components σ_xy(ω)=0. So, the mechanism supporting the optical activity of TBG remains to be further explored. In this paper, we analyze the optical activity of twisted bilayer graphene by treating simultaneously two essential aspects of the field-matter interaction problem, namely the propagation of the electromagnetic waves through a material layer and the response of a material layer to the action of an external field. We show that the optical activity of the TBG system is a result of spatial dispersion effects. Our theory is based on the following analysis. Working with 2D van der Waals systems, the physical quantities that describe them, such as the dielectric tensor and the conductivity tensor, are often considered as if they were ideal 2D systems, i.e., with zero thickness, even though the interlayer spacing is taken into account in the description of interlayer coupling.<cit.> However, in reality, even for graphene with only one layer of carbon atoms, an effective thickness can be defined based on the width of the quantum well potential caused by the carbon atoms.<cit.> As a result, 2D vdW systems cannot be considered ideal 2D systems but rather quasi-2D (Q2D) systems. This means that when describing the interaction between an electromagnetic field and a Q2D vdW material system at a macroscopic level, special considerations must be made. Typically, field equations are determined by averaging a set of microscopic field equations over infinitesimal volume elements.<cit.> However, this averaging procedure is only valid when the wavelength of the fields is much smaller than the dimensionalities of the physical system. For Q2D materials, averaging cannot be performed along the Oz direction perpendicular to the Oxy plane of the atomic lattice. As a result, we need to use a hybrid basis of reciprocal-real spaces to represent physical quantities and field equations. For example, the equation J_α(z,𝐪_∥,ω) = ∫ dz^'σ_αβ(z,z^',𝐪_∥,ω)E_β(z^',𝐪_∥,ω) describes the relationship between the component J_α(z,𝐪_∥,ω) of an electrical current density induced in the material by an electric field E_β(z,𝐪_∥,ω), where α, β = x,y are the indices of the components of vector fields, 𝐪_∥ is the wave vector in the material plane, ω is the frequency, and σ_αβ(z,z^',𝐪_∥,ω) is a component of the electrical conductivity tensor of the material layer. In the optical limit, where 𝐪_∥a≪ 1 and a is the lattice constant, we can approximate all quantities at 𝐪_∥=0, i.e., effectively ignoring the spatial effects in the material plane but not along the thickness.<cit.> This description highlights the importance of taking into account the spatial dispersion along the Oz direction, which is typically the transmission direction of electromagnetic fields used to study the material's properties. Theoretical studies of optical activity in TBG have been discussed in several references, including Kim_2016, Morell_2017, Stauber-2018. These studies have revealed the role of the spacing a between two graphene layers and explained optical activity as a magneto-electric effect. In this approach, a magnetization 𝐦_∥ and a polarization 𝐩_∥ emerge in the atomic plane due to the difference in current densities in the two graphene layers. To establish a relationship between these two quantities with electric and magnetic fields, a symmetry analysis and some assumptions were made,<cit.> such as the linear change of the fields in the space between the two graphene layers and the zero-frequency limit,<cit.> or the dephasing of the current operators in the two graphene layers.<cit.> The magnetization and polarization affect the direction and magnitude of the electric and magnetic field vectors within the material layer. However, the problem is that the transverse component σ_xy^(drag) of the conductivity tensor, which plays a role in creating the effect of one layer on the other when an electromagnetic field is transmitted through the system, is treated as the total Hall conductivity of the system in this wave propagation problem. This may be misleading as the system does not support a finite Hall conductivity due to time-reversal symmetry. Therefore, it is essential to consider in detail the electromagnetic wave propagation through the system, where it is understood that the electromagnetic wave must pass through one layer before reaching the second layer, and the coupling between the two layers must play a crucial role. In the next section, we will use Eq. (<ref>) without considering symmetry analysis and the mentioned assumptions to obtain the most important theoretical results in Ref. (Stauber-2018), see Eq. (2) therein, and clarify the role of σ_xy^(drag) in the electromagnetic wave propagation through the TBG system. Our study encompasses both macroscopic and microscopic perspectives. At the macroscopic level, we use the transfer matrix method to analyze the light propagation through a 2D van der Waals system composed of multiple material layers, each treated as a boundary between two vacuum layers of finite width.<cit.> The material system's conductivity tensor is decomposed into local and drag terms, with the local term defining the local relationship between current density and electric field in each layer, and the drag term defining the non-local relationship between current density in one layer and the electric field in other layers. Technically, the finite thickness of the material system makes the assumption of the spatial homogeneity along the thickness invalid. It therefore requires evaluating expressions that relate physical quantities through integrals over z and z'. This results in increased computational complexity and the need for a microscopic description of the material surfaces, such as the potential caused by atoms. To simplify the analysis, we assume that each layer of atoms has zero thickness. It allow us to express relevant quantities as sums of delta functions: J_α(z,ω) = ∑_ℓ J_α^(ℓ)δ(z-z_ℓ), where z_ℓ is the position of the ℓth material layer, and J_α^(ℓ) is the current density on that layer. At the microscopic level, we use the Kubo formula to calculate each term of the conductivity tensor required for the macroscopic level of description. We apply an effective model to determine the low-energy states of electrons in the TBG material lattice, and show that the electron eigen-states can be expressed as a hybridization of single-layer states in the two graphene layers. If the atomic lattice has mirror symmetry (the AA-stacked configuration) or glide symmetry (the AB-stacked configuration) in the lattice plane, the drag transverse conductivity will be vanished due to the cancelling contribution of the electron states. However, generic twisted bilayer graphene lattices have a chiral structure and do not possess these symmetries, leading to a non-zero drag transverse conductivity that determines the optical activity of the system. In addition to determining the components of the optical conductivity tensor, the reflection, transmission, absorption coefficients, and the circular dichroism spectrum, we also calculate the DC conductivity of the TBG system. We find that the hybridized states support a quantum value of 2σ_0^DC, where σ_0^DC = 4e^2/π h, at the intrinsic Fermi energy. This holds true even for the TBG configuration with a magic twist angle, which has a high density of states there due to the presence of a flat band. The structure of this paper is divided into three sections. After the “Introduction" section, the next section (Sec. <ref>) provides a solution to the problem of wave transmission through the twisted bilayer graphene (TBG) system in Sec. <ref>. The solution is based on the transfer matrix method, which illustrates the contribution of the conductivity tensor components to the transfer matrix elements such as the reflection and transmission coefficients and the absorption coefficient. In Sec. <ref>, the local and drag components of the optical conductivity tensor are calculated through the use of the Kubo formula, while the DC conductivity of the TBG system is calculated using the Kubo-Greenwood formula. In Sec. <ref>, an effective model for the electrons in the TBG lattices is presented. In Sec. <ref>, numerical results for the electronic band structure of various TBG configurations, the analysis of the optical conductivity tensor elements, and the spectra of circular dichroism as a function of photon frequency are presented. The conclusion is presented in the final section, Sec. <ref>. § THEORY AND CALCULATION METHODS §.§ Optical chiral responses In this section, we investigate the optical response of TBG systems by formulating a theory for the propagation of electromagnetic waves through a generic TBG sheet. Unlike typical dielectric slabs with two separate surfaces, the electric current densities in the two graphene layers of the TBG sheet are interdependent. The current density in one graphene layer is affected by both the electric field within that layer and the electric field in the other layer. This relationship is described by Eq. (<ref>), which we use to solve the problem at hand. We set up the system, following Ref. Szechenyi_2016, with two graphene layers separating the entire space into three regions, each characterized by parameters (ϵ_i,μ_i), where i=1,2,3 labels the mediums. Our goal is to determine the transfer matrix of the system and to develop expressions for the transmission and reflection matrices of the bilayer graphene system. As the inhomogeneity of the space only occurs in the z direction perpendicular to the TBG surface, the Maxwell equations support a plane wave solution of the form: F_i(x,y,z) = [F_i^+e^ik_izz + F_i^-e^-ik_izz]e^ik_ixxe^ik_iyy, where F represents E (the electric field), D (the electric induction), H (the magnetic field), and B (the magnetic induction). The vector coefficients F_i^± depend on the wave vector k_i in general. The conditions for the continuity of the vector fields at the interface between two mediums are given by: n×[E_2(1)-E_1(1)] = 0, n×[H_2(1)-H_1(1)] = J(1), for the first graphene layer surface, and n×[E_3(2)-E_2(2)] = 0, n×[H_3(2)-H_2(2)] = J(2). for the second graphene layer surface. In Eqs. (<ref>) and (<ref>), the distribution of current densities on each graphene layer is linearly related to the electric field by the conductivity tensor given via Eq. (<ref>). To the completeness of the setup, we rewrite this equation in the explicit form: J(1) = σ^(1)E_1(1)+σ^(drag)E_2(2), J(2) = σ^(drag)†E_1(1)+σ^(2)E_2(2) =σ^(drag)†E_2(1)+σ^(2)E_3(2). The continuity conditions of the fields at interfaces allow us to determine the relationship between the vector fields (F_i^+,F_i^-) in two different media at their interface. In particular, the magnetic field H_i is related to the electric field E_i through the Maxwell equation curl(E) = -∂_tB, where B=μH. Utilizing the plane wave solution (<ref>), we obtain the expression for H_i^±: H_i^± = ±√(ϵ_i/μ_i)k_i×E_i^±/|k_i|, where ϵ_i and μ_i represent the absolute permittivity and absolute permeability of medium i, respectively. This equation makes use of the dispersion relation ω = c|k_i|/n_i, where c is the speed of light in vacuum and n_i is the refractive index of medium i. It is worth noting that √(ϵ_i/μ_i) can be expressed through n_i as √(ϵ_i/μ_i)≈ n_iσ_0/2α, where σ_0=e^2/h is the unit of quantum conductivity and α = e^2/(4πϵ_0ħ c) is the fine-structure constant. In this study, the problem is solved for light transmission along the normal direction to the TBG layer, so k_i = k_in. The transfer matrix 𝐌31 (of size 4× 4) is then derived, relating the components of the electric fields in the first and third mediums as follows: ([ E_3^+(2); E3^-(2) ]) = 𝐌31([ E_1^+(1); E_1^-(1) ]). Here, 𝐌_31=𝐌_32𝐌^f_2𝐌_21, where: 𝐌_32 = ([ 1 1; -√(ϵ_3/μ_3)1-σ^(2) √(ϵ_3/μ_3)1-σ^(2) ])^-1([ 1 1; -√(ϵ_2/μ_2)1+σ^(drag)†e^-ik_2zd √(ϵ_2/μ_2)1+σ^(drag)†e^ik_2zd ]), 𝐌_21 = ([ 1 1; -√(ϵ_2/μ_2)1-σ^(drag)e^ik_2zd √(ϵ_2/μ_2)1-σ^(drag)e^-ik_2zd ])^-1([ 1 1; -√(ϵ_1/μ_1)1+σ^(1) √(ϵ_1/μ_1)1+σ^(1) ]), 𝐌_2^f =([ e^ik_2zd1 0; 0 e^-ik_2zd1 ]), with 1 the 2× 2 identity matrix. Let 𝐫 and 𝐭 be the 2× 2 matrices that relate the components of the electric field vector of the incident light to those of the reflected and transmitted light, i.e., E_1^-(1) = 𝐫·E_1^+ and E_3^+=𝐭·E_1^+. These matrices are referred to as the reflection and transmission matrices, respectively. From Eq. (<ref>), we can deduce the following: 𝐫 = -[𝐌_31^22]^-1𝐌_31^21, 𝐭 = 𝐌_31^11-𝐌_31^12[𝐌_31^22]^-1𝐌_31^21=[𝐌_31^-1]^11. From these matrices, the reflectance (R) and transmittance (T) are determined as the ratio of energy fluxes, specifically: R = |(𝐫·E_1^+)^†·(𝐫·E_1^+)/E_1^+†·E_1^+|, T = |n_3/n_1(𝐭·E_1^+)^†·(𝐭·E_1^+)/E_1^+†·E_1^+|. On the base of the conservation of energy flux, the absorptance is then determined by A = 1-(R+T). The values of R, T, and A are dependent not only on the energy, but also on the polarization states of light. It is important to note that the left- and right-handed polarization states of light are characterized by electric field vectors, E_L,R^+ ∝ (1,± i)^T/√(2) (The supersript T here denotes the transpose operation). To measure the dependence of light absorption on these polarization states, we calculate the spectrum of circular dichroism (CD), defined as: CD = A_L-A_R/A_L+A_R. In the following section, we present the method for calculating the conductivity tensors σ^(ℓ) and σ^(drag), and we will use the results obtained to calculate the transmittance, reflectance, absorbance and circular dichroism via Eqs. (<ref>), (<ref>) and (<ref>). §.§ AC and DC conductivities The optical properties of bilayer graphene systems are notable, as evidenced by previous research.<cit.> When taking into account the layer thickness, the effects of spatial dispersion become evident. To describe these effects, we decompose the current response to an external electric field caused by an incident monochromatic light beam perpendicular to the material plane. To start, we rewrite Equation (<ref>) in the real space representation as follows: J_α(x,z, ω) = ∫ dz^'σ_αβ(x,z;x,z^',ω)E_β(x,z^',ω), where 𝐱 = (x,y) represents the position vector in the material plane with x and y as its coordinates. On the assumption of spatial homogeneity in the material plane, we can write σ_αβ(x,z;x,z^', ω) = σ_αβ(z,z^',ω), and: σ_αβ(z,z^',ω) = ∑_ℓ=1^2σ_αβ(z,z^',ω)δ(z^'-z_ℓ). Eq. (<ref>) therefore becomes: J_α(x,z,ω) = ∑_ℓ=1^2σ_αβ(z,z_ℓ,ω)E_β(x,z_ℓ,ω). More specifically, we resolve the current density induced in each graphene layer as follows: J_α^(ℓ)(x,ω) = σ_αβ^(ℓ)(ω)E_β^(ℓ)(x,ω)+σ_αβ^(drag)(ω)E_β^(ℓ^')(x,ω), where the involved quantities are defined by: J_α^(ℓ)(x,ω) = J_α(x,z_ℓ,ω); E_β^(ℓ)(x,ω) = E_β(x,z_ℓ,ω) and σ_αβ^(ℓ)(ω) = σ_αβ(z_ℓ,z_ℓ,ω), and σ_αβ^(drag)(ω) = σ_αβ(z_ℓ,z_ℓ^',ω) where ℓ≠ℓ^'. Eqs. (<ref>) can be cast in the matrix form: ([ J^(1); J^(2) ])_ω = ([ σ^(1) σ^(drag); σ^(drag)† σ^(2) ])_ω([ E^(1); E^(2) ])_ω, where σ^(ℓ)(ω) = σ^(ℓ)(ω)τ_0 is called the local part, and σ^(drag)(ω) = σ^(drag)(ω)τ_0-iσ_xy^(drag)(ω)τ_y = σ^drag(ω)-σ_xy^dragẑ× is called the drag part. Here τ_0 is the 2× 2 identity matrix and τ_2 is the conventional second Pauli matrix. The compact form (<ref>) is identical to results reported in Refs. Stauber-2018,Ochoa_2020. Importantly, it guarantees the time-reversal symmetry, the rotation symmetry and the layer interchange symmetry. The conductivities σ^(ℓ)(ω), σ^(drag)(ω) and σ^(drag)_xy(ω) can be determined from the Kubo formula. There are a number of versions of the Kubo formula for the electrical conductivity suitable for implementing it in different situations. However, the most important ingredient we must specify is the velocity operators v̂_α. In general, these operators are determined from the position operator x̂_α and the Hamiltonian Ĥ via the Heisenberg equation: v̂_α = 1/iħ[x̂_α,Ĥ] → v_α(k)= 1/ħ∂/∂ k_αH(k). According to the linear response theory, we see that the conductivity has the paramagnetic part only. Using the eigen-vectors of single-particle Hamiltonian (without the presence of external vector potential A(x,t)) as the representation basis, Ĥ|n⟩ = E_n|n⟩, the elements of the optical conductivity tensor are given by this formula: <cit.> σ_αβ^(c)(ω) = i e^2/S1/ω+iη∑_m,n(f_m-f_n)O^(c)mn_αβ/E_m-E_n+ħ(ω+iη), where S is the area of material sample; c = {ℓ,drag}; f_n = f(E-μ,k_BT) is the occupation weight of the energy level E_n determined by the Fermi-Dirac function f with μ,k_BT the chemical potential and thermal energy; η is a positive infinitesimal number, and O_αβ;mn^(c) denotes the product of velocity matrix elements: O_αβ^(ℓ)mn = ⟨ m|v̂_α^(ℓ)|n⟩⟨ n|v̂_β^(ℓ)|m⟩, O_αβ^(drag)mn = ⟨ m|v̂_α^(1)|n⟩⟨ n|v̂_β^(2)|m⟩. To the DC conductivity, it might be optimistic to think that the DC conductivity can be simply obtained from the expression of the optical conductivity in the limit of zero frequency. However, it is not, at least in the numerical calculation. This is because for some finite frequency ω, there is always a finite length scale governing the behavior of electron. This length scale is L_ω = 2π v_F/ω. Meanwhile, there is no such length scale for the DC transport. Furthermore, the static transport has the diffusion nature due to the similarity of the Schrödinger equation and the diffusion equation. The DC conductivity can be obtained from the linear response theory. The vector potential is chosen in the form A(t) = (-Et,0,0), where E is the intensity of a static electric field. The DC conductivity is calculated using the Kubo-Greenwood formula:<cit.> σ_αα(μ,kT) = -∫_-∞^+∞dE ∂ f(E-μ,kT)/∂ Eσ_αα(E), where σ_αα(E) = 2π e^2ħ/S∑_m,n|⟨ m|v̂_α|n⟩|^2 δ(E-E_n)δ(E-E_m). This quantity is seen as the DC conductivity at zero temperature. When performing the above formula, we use the Gaussian function to approximate the δ-Dirac function: δ(E-E_n) ≈1/η√(π)exp(-(E-E_n)^2/η^2), where η > 0 is a tiny number that is appropriately chosen to smear the energy levels for the numerical calculation. §.§ Effective continuum model Effective continuum models for the low-energy states of electrons in TBG systems have been developed since 2007 by Lopes et al. <cit.> However, the model proposed by Bistritzer and MacDonald in 2011 is better known and widely used. In this paper, we present our solution to the Bistritzer-MacDonald model. <cit.> We implement this model for TBG lattices in which the first layer is clockwise rotated by a half-twist angle, θ_1 = -θ/2, and the second layer is counterclockwise rotated by θ_2 = +θ/2. The unrotated layers are defined by unit vectors a_1 = a·e_x and a_2 = acosπ/3·e_x + asinπ/3·e_y, where a is the lattice constant of graphene, and e_x and e_y are the unit vectors defining the x and y directions, respectively. The two associated vectors defining the reciprocal lattice are thus a_1^⋆ = (4π/a)(cosπ/6·e_x+sinπ/6·e_y) and a_2^⋆ = (4π/a)·e_y. Under the twisting, these vectors defining the atomic lattice of the two graphene layers as well as their reciprocal lattice are determined by applying the rotation matrix R_z(θ_ℓ), specifically a_i^(ℓ) = R_z(θ_ℓ)·a_i, and a_i^⋆(ℓ) = R_z(θ_ℓ)·a_i^⋆. The first Brillouin zone BZ^(ℓ) of each graphene layer is defined by a set of six corner points K_i^(ℓ). For instance, the point K_1^(ℓ) is determined by the point K_-^(ℓ) with K_ξ^(ℓ) = -ξ(2a_1^⋆(ℓ)+a_2^⋆(ℓ))/3, where ξ = ± 1. For commensurate twisting, the twist angle θ is determined by two integer numbers m and n via the formula: θ = arctan(|n^2-m^2|sin(π/3)/(n^2+m^2)cos(π/3)+2mn). When the absolute value of the difference between m and n is equal to 1, two unit vectors A_i^⋆ = a_i^⋆(1)-a_i^⋆(2) define the moiré reciprocal lattice. The mini-Brillouin zone comprises six K^M_i points. K_1^M = (-A_1^⋆+A_2^⋆)/3, and K^M_6 = (A_1^⋆+2A_2^⋆)/3 define two of these points. The remaining points can be obtained by shifting these points using the reciprocal unit vectors. We begin by representing the Hamiltonian H for electrons using the layer-resolution vector basis set {|ℓ⟩,|,ℓ = 1,2}. Using the identity 1=∑_ℓ=1^2|ℓ⟩⟨ℓ|, we can express H as follows: H = ∑_ℓ,ℓ^'=1^2|ℓ⟩ H_ℓ,ℓ^'⟨ℓ^'|, where H_ℓ,ℓ^' = ⟨ℓ|H|ℓ^'⟩. Next, we can further specify H_ℓ,ℓ^' by using the lattice-resolution vector basis set |k,m⟩ = |k+G_m⟩ ,|,k∈ BZ, m∈ℤ. We can then use the identity 1=∑k∈ BZ∑_m|k,m⟩⟨k,m| to obtain: H = ∑_k∈ BZ∑_ℓ,m∑_ℓ^',n|ℓ,k,m⟩ H_ℓ,m;ℓ^',n(k)⟨ℓ^',k,n|, where |ℓ,k,m⟩ = |ℓ⟩|k,m⟩ and we use the relation ⟨ℓ,k,m|H|ℓ^',k^',n⟩ = H_ℓ,m;ℓ^',n(k)δ_k,k^'. In the case of twisted bilayer graphene (TBG) configurations with tiny twist angles, the low energy states of electrons are distinguished by a quantum index ξ = ±1, which corresponds to the two nonequivalent Dirac valleys K_ξ^(1,2) of the graphene monolayers. As the twist angle becomes smaller, the positions of the two points K_ξ^(1) and K_ξ^(2) become closer to each other. The Bistritzer-MacDonald model is represented by a Hamiltonian in real-space representation as follows:<cit.> Ĥ^ξ = [ H_1^ξ(p̂) T^ξ(r̂); T^ξ†(r̂) H_2^ξ(p̂) ]. Here, H_1^ξ(p̂) and H_2^ξ(p̂) are the blocks for the Hamiltonians of the two monolayers, and T^ξ(r̂) is the interlayer coupling block, which are determined via the momentum and position operators, p̂ and r̂, respectively. For the range of energy around the intrinsic Fermi level the two Hamiltonian blocks H_1^ξ(p̂) and H_2^ξ(p̂) are well approximated by the 2D Dirac Hamiltonian: H_ℓ^ξ(p̂) = - v_F(ξσ_x,σ_y)·[R^z(-θ_ℓ)·(p̂-ħK_ξ^ℓ)], where σ_x,σ_y are two conventional Pauli matrices, R^z(-θ_ℓ) is a matrix that rotates the relevant vectors around the Oz axis to maintain the canonical form of the Dirac Hamiltonian, K_ξ^ℓ is the corner point of type (valley) ξ in the first Brillouin zone of layer ℓ, and v_F is the Fermi velocity. The minus sign in the equation is due to the negative value of the hopping parameter V_ppπ = -2.7 eV. The block term for interlayer coupling is expressed as follows:<cit.> T^ξ(r̂) = ([ u u^'; u^' u ])+([ u u^'ω^-ξ; u^'ω^ξ u ])e^iξδk_2·r̂ +([ u u^'ω^ξ; u^'ω^-ξ u ])e^iξδk_3·r̂, The values of δk_2 and δk_3 are defined as A_1^⋆ and A_1^⋆ + A_2^⋆, respectively, where A_1^⋆ and A_2^⋆ are reciprocal lattice vectors and ω = exp(i2π/3). To account for the effects of lattice reconstruction, the parameters u and u^', which are set to be 0.0797 eV and 0.0975 eV, respectively, are chosen based on the description in reference Koshino_2018. The velocity operator corresponding to the Hamiltonian in Eq. (<ref>) can be expressed as follows: v̂_α = -v_F(ξσ_x,σ_y)· R^z_α(-θ_ℓ), Here, R^z_α(-θ_ℓ) denotes the first or second column of the rotation matrix R^z(-θ_ℓ) depending on whether α=x or α=y. It is noteworthy that, because of the relativistic nature of H^ξ_ℓ(p̂) in Eq. (<ref>), the current operator ĵ_α only comprises a drift current component, ĵ_α = ev̂_α, with no diffusion or diamagnetic terms. The spectrum of the Hamiltonian Ĥ is obtained by solving the secular equation given by: ([ H_1^ξ(p̂) T^ξ(r̂); T^ξ†(r̂) H_2^ξ(p̂) ])|ψ_ξ,k⟩ = E|ψ_ξ,k⟩, where the electron states in the twisted bilayer graphene (TBG) lattice are modeled as Bloch state vectors |ψ_ξ,k⟩. Because of the approximate periodicity of the moire lattice in TBG the state vectors |ψ_ξ,𝐤⟩ can be expanded in terms of the plane-wave vectors |k+G_m⟩ which are the eigenvectors of the momentum operator p̂, i.e., p̂|k+G_m⟩ = (k+G_m)|k+G_m⟩ and ⟨r|k+G_m⟩ = e^i(k+G_m)·r. Specifically, we have: |ψ_ξ,k⟩ = ∑_m([ C_1,ξ,k(G_m); C_2,ξ,k(G_m) ])|k+G_m⟩, where C_1,ξ,k(G_m) and C_2,ξ,k(G_m) are 2D vectors of combination coefficients that need to be found. {G_n | n = 1,2,3, , N_G} is a set of N_G vectors of the moire reciprocal lattice. Substituting the expression (<ref>) into Eq. (<ref>) then left-multiplying both sides with ⟨k+G_n| we obtain a set of linear equations in the form: ∑_m[([ H_1^ξ(k+G_n) T_1^ξ; T_1^ξ† H_2^ξ(k+G_n) ])δ_G_n,G_m+. +.∑_j=2^3([ 0 T_j^ξ; 0 0 ])δ_G_n,G_m+δk_j+∑_j=2^3([ 0 0; T_j^ξ† 0 ])δ_G_n,G_m-δk_j]([ C_1,ξ,k(G_m); C_2,ξ,k(G_m) ]) = E([ C_1,ξ,k(G_n); C_2,ξ,k(G_n) ]), wherein we denote: ⟨k+G_n|H_ℓ^ξ|k+G_m⟩ = H_ℓ^ξ(k+G_n)δ_G_n,G_m, ⟨k+G_n|T^ξ(r̂)|k+G_m⟩ = ∑_j=1^3T_j^ξδ_G_n,G_m+δk_j, ⟨k+G_n|T^ξ†(r̂)|k+G_m⟩ = ∑_j=1^3T_j^ξ†δ_G_n,G_m-δk_j, To obtain a numerical solution, we define a 4N_G× 4N_G Hermitian matrix H_k^ξ in block form for each value of ξ and k∈MBZ: [H_k^ξ]_n,n = ([ H_1^ξ(k+G_n) T_1^ξ; T_1^ξ† H_2^ξ(k+G_n) ]), [H_k^ξ]_n,m_j = ([ 0 T_j^ξ; 0 0 ]), ifG_m_j = G_n-δk_j, [H_k^ξ]_n,m_j^' = ([ 0 0; T_j^ξ† 0 ]), ifG_m_j^' = G_n+δk_j. The eigenvalues E_n^ξ(k) and corresponding eigenvectors, obtained by diagonalizing the matrix H_k^ξ, approximate the states of low-energy electrons in the TBG lattice. The accuracy of the effective description is numerically dependent on the number of reciprocal lattice vectors {G_n | n = 1,2,3,, N_G} considered in the calculation. To ensure validity in the low-energy range near the intrinsic Fermi level, where the energy surfaces of the monolayer graphene take the form of cones, we control the value of N_G by using a cutoff energy E_c. Specifically, we count only the G_n vectors such that |G_n|≤ E_c/ħ v_F. § RESULTS AND DISCUSSIONS §.§ Electronic band structure of TBGs We used numerical methods to solve Eq. (<ref>). Specifically, for each value of ξ=± 1, we constructed a Hermitian matrix H_k^ξ of size 4N_G× 4N_G and diagonalized it for each value of k in the mini-Brillouin zone of the TBG system. The sets of eigenvalues E_m(k) and eigenvectors {𝐂^m_ξ(k)=[C^m_1,ξ,k(G_1),C^m_2,ξ,k(G_1),,C^m_1,ξ,k(G_N_G),C^m_2,ξ,k(G_N_G)]^T} were calculated. The results for three TBG configurations with twist angles of θ = 1.05^∘, 1.89^∘ and 3.89^∘ are shown in Fig. <ref>. The blue curves represent the results obtained when the electronic interlayer coupling is artificially turned off. Both sets of data are plotted on the same figure for comparison, to highlight the impact of the interlayer coupling. The presence of multiple dispersion curves in the figure results from the folding of the energy band structure of two graphene layers due to the enlargement of the unit cell of the TBG lattice and the corresponding shrinkage of the Brillouin zone into the mini-Brillouin zone. However, the electronic band structure of the TBG system is not simply formed in this manner. At the points where the bands of monolayer graphene cross, the interlayer coupling leads to the hybridization of Bloch states and the lifting of energy degeneracy to form new bands. Our results agree with other available data in the literature <cit.>. However, in this study, we wish to emphasize the formation of new electronic states, which are special to TBG systems and play a crucial role in determining its electronic, optical and transport properties. Fig. <ref> can be considered a visual representation of the evolution of the band structure with respect to the twist angle. We observe that for the configuration with θ = 3.89^∘, the dispersion curves in the low energy range around the intrinsic Fermi energy (E_F = 0) have a similar form to the blue curves, although they are not identical. This implies that the states created in the bilayer system share similarities with those in monolayer graphene. For smaller twist angles, particularly θ = 1.05^∘, a band with a very narrow bandwidth forms around the Fermi energy, separated from the lower and upper bands by a narrow gap. These dispersion curves are completely different from the blue curves, as seen in Fig. <ref>(e). We conclude that the electronic interlayer coupling creates special states in TBG lattices with small twist angles. Although the TBG systems are two-dimensional, it can be challenging to analyze their geometrical features of energy surfaces in wave-vector space. However, the density of states (DOS) generally exhibits key features, such as van Hove singularity behavior, as shown in the right panels of Figs. <ref> (a, b, c). The blue curves in these figures were obtained numerically without considering the effects of electronic interlayer coupling, which is equivalent to the DOS of two independent graphene layers. The red curves show significant peaks due to van Hove singularities, indicating the presence of extremal and saddle points of the energy surfaces. It's worth noting that the energy bands with index ξ =+1 (resulting from the hybridization of graphene Bloch states in the K_1 valley of the first layer and K_2 valley of the second layer) differ from those with index ξ = -1 (resulting from the hybridization of Bloch states in the K_1^' valley and K_2^'). This difference is evident along the ΓM direction as shown in Figs. <ref> (d, e, f). However, it does not appear in the DOS picture. We computed the DOS for both ξ = +1 and ξ = -1 separately and obtained the same results. This implies that the energy surfaces with ξ = +1 are simply rotations of the corresponding ones with ξ = -1 by an angle of π/3. §.§ DC conductivity The calculation of the electronic structures of TBG configurations with twist angles greater than 2 degrees shows the presence of two energy bands near the Fermi energy (E_F = 0), which have dispersion curves similar to those of monolayer graphene. However, the Fermi velocity of the linear dispersion law in TBG (v_F^TBG) is smaller than that of monolayer graphene (v_F^MLG). As the twist angle decreases, these two bands become closer and eventually form an isolated band with a narrow bandwidth. This “flat band" near the Fermi energy is expected to have limited impact on the transport properties of TBG sheets because of its small Fermi velocity. In other words, electrons occupying these bands are expected to be spatially confined in the atomic lattice. Although these bands are not completely dispersionless, the study of their transport properties is still of great interest. To this end, within the framework of the non-interacting electron approximation, the DC conductivity was calculated using the Kubo-Greenwood formula as given by Eqs. (<ref>) and (<ref>). The results of the calculation for the longitudinal components (σ_xx^DC and σ_yy^DC) of three TBG configurations at low temperatures are presented in Fig. <ref>. The data indicate that the electrical conductivity of TBG systems is anisotropic, which is reflected in the two different longitudinal elements of the conductivity tensor, σ_xx^DC (represented by solid curves) and σ_yy^DC (represented by dashed curves). The conductivity should reflect the electronic structure with respect to the Fermi energy, and our calculations for the two cases of ξ = ± 1 yielded the same result. This is similar to the picture of the density of states, but the difference between the two longitudinal conductivities highlights the anisotropic nature of the energy surfaces. For the TBG configuration with a twist angle of θ = 3.890^∘, the electronic structure is similar to that of monolayer graphene (except for the value of the Fermi velocity), and the conductivity curves have a characteristic V-shape (or an M-shape when considered over a wider energy range, as shown by the blue curves). For TBG configurations with smaller twist angles, the electronic structure and thus the conductivity picture becomes much more complex. However, our calculations show that the conductivity at the intrinsic Fermi level E_F = 0 is always finite and non-zero. Interestingly, the calculated DC conductivity for TBG configurations with twist angles that are not magic values is about two times the quantum value of the minimal conductivity of monolayer graphene, which is σ^DC_0 = 4e^2/π h. In the case of the TBG configuration with the magic twist angle θ = 1.050^∘, the conductivity curve shows a small hump at the Fermi energy E_F = 0 with σ_αα^DC(0) ≈σ^DC_0. This quantum value of the conductivity of monolayer graphene has been intensively discussed in the past and is rooted in the fact that the valence and conduction bands touch each other at the K points, resulting in always-available free carriers due to fluctuations in the electrostatic profile around the Fermi energy level.<cit.> In our calculation, we approximated the delta-Dirac functions in the Kubo-Greenwood formula with a Gaussian function with a parameter η that defines the finite width of the peak. The value of η was taken to be η < 5 meV to account for both numerical calculation needs and the broadening of energy levels and finite lifetime of the electric charge-carrying quasi-particles. Although the observed DC conductivity value in TBG systems is supported by the characteristic hybridized states in the TBG lattice, we suppose that it has the same underlying physical explanation as in monolayer graphene. §.§ AC conductivity and optical Hall drag conductivity In Fig. <ref>, we present the results of our calculations for the total optical conductivity of three different twisted bilayer graphene (TBG) configurations. The calculation was performed using the Kubo formula for the conductivity tensor, where the velocity operator v̂_α was determined from the Hamiltonian (<ref>) using Eq. (<ref>). The time reversal symmetry and the Onsager reciprocal relations eliminate the off-diagonal elements σ_xy and σ_yx of the conductivity tensor, which we numerically confirmed behave as noise with an amplitude of 10^-3. We also verified all of the symmetry properties of the conductivity tensor according to Eq. (<ref>), demonstrating the accuracy of our numerical data. Figure <ref> displays the longitudinal conductivity curves σ_xx(ω) with significant peaks, which are attributed to dominant interband transition processes. This correlation is supported by its relationship to the electronic band structures of the TBG configurations, shown in Fig. <ref>. For instance, in Fig. <ref>(b), at an energy of 0.5681 eV, there is an optical absorption peak that can be assigned to a transition from the highest valence band to the lowest conduction band around the M points in the mini-Brillouin zone. As the twist angle decreases, the electronic structure becomes more complex and the optical conductivity exhibits more optical absorption peaks. For TBG configurations with twist angles greater than 2 degrees, there is a distinctive feature in the low-energy range of the conductivity curve: the real part of the conductivity is independent of the photon energy ħω and takes a value of 2σ_0, where σ_0 = π e^2/2h. This behavior is similar to that of monolayer graphene and results from the linear dispersion law of electron states. In the low-energy range, the energy band structure of TBG systems is qualitatively similar to that of monolayer graphene, but with a smaller Fermi velocity, as previously analyzed. The value of σ_xx(ω) is equal to 2σ_0, as found in previous studies.<cit.> For the case of the “magic angle" θ = 1.050^∘, our calculation shows a clear dominant contribution from optical transitions from the quasi-flat bands around the Fermi level to the lowest conduction band around the Γ point, as shown in Fig. <ref>(c) and Figs. <ref>(c) and <ref>(f). Thus far, we would like to emphasize that the use of the total DC conductivity tensor is not relevant when studying the wave transmission problem in TBG. As we have presented in Secs. <ref> and <ref>, the required quantities to be utilized are σ^(ℓ)(ω) and σ^(drag)(ω). In Figs. <ref> and <ref>, we present the calculation results for the elements of the conductivity tensor blocks σ^(ℓ)(ω) and σ^(drag)(ω) as analyzed in Sec. <ref>. Our numerical calculation reveals that σ^(1)(ω) = σ^(2)(ω) = σ_xx^(ℓ)(ω)τ_0, verifying the equivalent role of the two graphene layers in the TBG lattice. Figs. <ref>(a), (b), and (c) display the real (red) and imaginary (blue) parts of σ_xx^(ℓ)(ω) for ℓ=1,2. It is worth noting that decreasing the twist angle increases the magnitude of σ_xx^(ℓ), particularly when the imaginary part becomes dominant and varies in accordance with 1/ħω, as seen in Fig. <ref>(c). Fig. <ref> shows the real and imaginary parts of the off-diagonal element σ^(drag)(ω) of the drag part of the conductivity tensor. For large twist angles in the TBG configuration, there is a notable difference between σ_xx^(ℓ)(ω) and σ_xx^(drag)(ω), see Figs. <ref>(a) and <ref>(a). However, for small twist angles (<2^∘), these curves are quite similar but have opposite signs. As a result, although there is significant variation in the magnitude of σ_xx^(ℓ)(ω) and σ_xx^(drag)(ω), they compensate each other when summed to give the total value σ_xx(ω) = 2[σ_xx^(ℓ)(ω)+σ_xx^(drag)(ω)], as seen in Fig. <ref>. Figures <ref>(b), (d), and (f) present the calculated results for the transverse conductivity component σ_xy^(drag)(ω). The behavior of σ_xy^(drag)(ω) is primarily characterized by close-to-zero values across the energy range, with some sharp peaks and dips appearing at energy positions with high optical absorption. This finite value of σ_xy^(drag)(ω) can be attributed to both the correlation between the current densities in the two graphene layers and the chiral structure of the atomic lattices. The former is considered a necessary condition, representing the result of the coupling between the two graphene layers, which leads to the formation of composite states like the superposition of Bloch states between the two layers. This means that when electrons move in one graphene layer along the Ox direction, they induce a motion in the second layer along the transverse Oy direction due to interlayer coupling. The sufficient condition, however, is the contribution of the electronic states in the system to the resulting conductivity. If the system had mirror symmetry, these correlations would cancel each other out completely. However, the TBG lattices lack mirror symmetry, unlike the hexagonal lattice of graphene, which results in a finite, non-zero conductivity of σ_xy^(drag)(ω) ≠ 0. To emphasize this point, the hexagonal mini-Brillouin zone of the TBG lattice was divided into three equal parts, as depicted in the inset of Fig. <ref>. This division was based on the assumption of mirror symmetry through the plane M_yz, with domain (1) containing the symmetry, and domains (2) and (3) being its mirror images. Using the Kubo formula, the value of σ_xy^(drag)(ω) was calculated by summing over all k points in each of the three domains. The results, shown in Fig. <ref>, indicate that the results in domains (2) and (3) are similar, opposite in sign, and much larger than the result in domain (1). Upon summing these results, a strong but incomplete cancellation takes place, leading to the final value of σ_xy^(drag)(ω) displayed in Figs. <ref>(b), (d), and (f). For the AA- and AB-stacked configurations with mirror symmetries, the cancellation was found to be complete, resulting in σ_xy^(drag)(ω) = 0. The symmetry analyses in Refs. Morell_2010 and Do_2020 support this theory. §.§ Optical characteristics On the basis of the theory established in Sec. <ref> and the data of the conductivity tensor already computed, we performed the calculation for the transfer matrix, then the reflection, transmission and absorption spectra of several TBG configurations. The obtained results for transmittance and CD are presented in Figs. <ref> and <ref>, respectively. Since the reflectance is negligible, it is not shown here. It should be noted that while the optical properties of TBGs have been extensively studied, they are usually analyzed through the behavior of the real and imaginary parts of the diagonal components of the total optical conductivity tensor,<cit.> rather than through the transmittance and reflectance, which are quantities that can be directly measured in experiments. From the transmission spectra as shown, we see that the transmittance of all TBG configurations can get the value of 98% and the absorptance of 2% in average in a large energy range. The spectral curves clearly exhibit the peaks and dips consistent with the picture of the longitudinal conductivities. The transmission and absorption spectra for the left- and right-handed lights are different to be distinguished. However, the CD spectra is clearly manifested. From Fig. <ref>, we see that for the TBG configuration with θ = 3.890^∘, the CD curve exhibits essential features of experimental data reported by Kim et. al.,<cit.> for instance, with the peaks and dips of large width, except several sharp peaks at the energy position of the absorption peaks. In order to verify the role of the drag transverse conductivity σ_xy^(drag), we did not include it into the calculation. As expected, it results in the zero CD for the total energy range, but mostly does not change the transmission spectrum. We therefore conclude that this drag transverse conductivity plays the decisive role in governing the chiral optical behaviors of the TBG systems. Last but not least, it is important to note that the CD spectrum shown in Fig. <ref> is extracted from the theory presented in Sec. <ref>, which may differ from the formula used in Refs. Morell_2017,Stauber_2018, where CD is simply proportional to the real part of σ_xy. § CONCLUSION The field of engineering two-dimensional quantum materials for electronic and optoelectronic applications has been the subject of extensive research. In this work, we demonstrate that the optical activity observed in the TBG system results from the spatial dispersion effects of a typical vdW heterostructure of two graphene layers lacking mirror and glide symmetries. Our analysis addresses two fundamental aspects of the field-matter interaction problem: the propagation of electromagnetic waves through a material layer and the material layer's response to an external field, both treated simultaneously using the macroscopic and microscopic descriptions, respectively. Technically, we utilize an effective continuum model to demonstrate how electron states in each graphene layer can hybridize, forming electron states in the bilayer system that allow for the correlation of transverse motions between the two layers. In terms of electromagnetic wave propagation, we clarify how the coupling between the two layers can transform the magnitude and direction of the fields within the material layer. We provide a detailed solution to the electromagnetic wave propagation problem through a TBG sheet by considering each graphene layer as a conducting interface between two dielectric media. The two interfaces are not independent and conduct mutually influenced current densities. Our solution to the wave propagation problem clarifies how the roles of the conductivity components characterized by local and nonlocal spatial effects manifest themselves in the optical behavior of the material system. On the basis of explicitly considering the spatial dispersion effects, our analysis naturally composes the conductivity tensor into the local part and the drag part without relying on symmetry analysis or assumptions about the phase shift between current operators in the two layers. We calculate all elements of these parts using the Kubo formula, and formulate and calculate the transfer matrix numerically. We demonstrate that the transverse drag transverse conductivity σ_xy^(drag) plays a critical role in defining the chiral optical response of the TBG system. This conductivity arises from the chiral structure of the atomic lattice, which leads to hybridized states that lack mirror symmetry and support the correlation of transverse motions between the two graphene layers. We use the results of our calculations of the components of the optical conductivity tensor to calculate the transmission and circular dichroism spectra. We show that the overall transparency of twisted bilayer graphene (TBG) samples, on average, is generally 98%, but it has the potential to absorb light up to 2%. Significantly, we establish that the drag transverse conductivity is a decisive factor in determining the circular dichroism, thereby affirming that the optical activity of the TBG system is a manifestation of spatial effects. Furthermore, we found that the DC conductivity of the TBG system exhibits a quantum conductivity value of ∝ e^2/h at the intrinsic Fermi energy, which is determined by the hybridized states in the bilayer system. This quantum value is obtained from a single-particle approximation. However, a rigorous discussion of the intrinsic transport properties of TBG configurations with magic twist angles should be based on a strong correlation picture of low-energy excited states as Dirac fermions. In summary, our study emphasizes the importance of considering the finite thickness of twisted bilayer graphene when analyzing its optical response and transport properties, owing to the significance of spatial dispersion effects. This theoretical approach can also be extended to other van der Waals material systems with multiple layers. § ACKNOWLEDGEMENTS One of the authors, S.T.H., acknowledges the financial support of Hanoi University of Civil Engineering (HUCE) for his work under grant number 28-2021/KHXD-TD. apsrev4-1

http://arxiv.org/abs/2303.07102v1

20230313134121

Entanglement dynamics with string measurement operators

cond-mat.stat-mech

Entanglement dynamics with string measurement operators Giulia Piccitto1⋆, Angelo Russomanno2, and Davide Rossini1 1 Dipartimento di Fisica dell’Università di Pisa and INFN, Largo Pontecorvo 3, I-56127 Pisa, Italy 2 Scuola Superiore Meridionale, Università di Napoli Federico II, Largo San Marcellino 10, I-80138 Napoli, Italy ^⋆ giulia.piccitto@df.unipi.it March 30, 2023 § ABSTRACT We explain how to apply a Gaussian-preserving operator to a fermionic Gaussian state. We use this method to study the evolution of the entanglement entropy of an Ising spin chain following a Lindblad dynamics with string measurement operators, focusing on the quantum-jump unraveling of such Lindbladian. We find that the asymptotic entanglement entropy obeys an area law for finite-range string operators and a volume law for ranges of the string which scale with the system size. The same behavior is observed for the measurement-only dynamics, suggesting that measurements can play a leading role in this context. 1pt fancy 1pt § INTRODUCTION Understanding the role of entanglement in quantum many-body systems is a subject under intensive study <cit.>. While the ground-state entanglement properties have been widely investigated and are relatively well understood, there are still several open questions regarding the stabilization of quantum correlations during the non-equilibrium dynamics. By exploiting a quasi-particle description it has been shown that in integrable systems, after a sudden quench, the entanglement entropy linearly increases in time, and eventually settles to a stationary value which linearly scales with the system size (volume-law scaling) <cit.>. A similar volume-law scaling occurs in ergodic short-range systems <cit.>, while in the presence of a strong disorder the situation is more involved, due to the appearance of many-body localization (see Ref. <cit.> for a review). A natural extension of the above scenario is for a quantum system coupled to an external environment, a setup which is crucial to applications in quantum technologies and quantum computing, since any experimental platform is subject to noise-induced decoherence <cit.>. In the hypothesis of weak and Markovian coupling with the bath, a good modeling for this setup is to assume the evolution of the reduced density matrix of the system to be ruled by a master equation in a Lindblad form <cit.>. The system-bath framework can be also interpreted as a process in which the environment behaves as a classical stochastic process performing random measurements on the quantum system <cit.>. The outcome of any realization of this process is a pure-state stochastic quantum trajectory and the density matrix obtained by averaging over such trajectories obeys a Lindblad-type dynamic evolution <cit.>. Here we are interested in the so-called measurement-induced phase transitions (MIPTs), i.e., in the possibility to develop discontinuities as in the asymptotic entanglement properties that emerge from the interplay between the unitary dynamics (typically generating entanglement) and the measurement processes (typically destroying it). This kind of transitions have been first proposed and studied in unitary random circuits undergoing random measurements (see, e.g., Refs.<cit.> and citing references), which may display a transition between a phase in which the entanglement grows with the system size and another phase where it undergoes a saturation to a finite value (area-law scaling). Similarly, MIPTs may also arise in free-fermion random circuits with temporal randomness <cit.>, Majorana random circuits <cit.>, Dirac fermions <cit.>, and other Hamiltonian models undergoing random measurements at discrete times <cit.>. When measuring string operators, MIPTs can occur in the presence of a measurement-only dynamics, without the interplay with the Hamiltonian <cit.>. Recently, an increasing number of works have been focusing on MIPTs in systems described by quadratic Hamiltonians on a lattice <cit.>. One of the reasons of this interest relies in the Gaussian structure of the states of such systems that, being entirely determined by two-point correlation functions, are suitable for a semi-analytical treatment up to large lattice sizes (of the order of hundreds of sites) <cit.>. In particular, it has been found that the asymptotic volume-law scaling of the entanglement following a unitary evolution <cit.> is unstable under local measurements, independently of the measurement strength <cit.>, and a MIPT may emerge. For example, the Ising chain exhibits a size-dependent crossover towards a subextensive or an area-law entanglement regime, depending on the system parameters, as well as on the measurement rate and the type of unraveling <cit.>. In the same context, a couple of recent works pointed out the importance of having long-range interactions to stabilize larger amounts of entanglement, if the system undergoes random onsite measurements <cit.>. Inspired by these latter findings, here we follow a complementary strategy and address the role of non-local string measurement operators in the dynamics of entanglement for a quadratic Hamiltonian system. The purpose of this paper is twofold. On the one hand, we collect and review all the technical details to implement the dynamics of Gaussian states in the presence of Gaussian-preserving measuring operators. On the other hand, we discuss the effect of non-local measurement operators on the dynamics of the entanglement entropy, while keeping the Hamiltonian with short-range interactions. To this aim, we implement a quantum-jump unraveling of the Lindblad dynamics <cit.> for the Ising chain (mapped to the fermionic Kitaev chain) using strings of Pauli matrices as Lindblad operators (mapped to non-local two-point fermionic operators). The main result is that, when the string range r is comparable with the system size L, contrary to what happens in the presence of local jump operators, the measurement process can stabilize an asymptotic volume-law for the entanglement entropy. As a consequence, a MIPT is conjectured as a function of r. Quite interestingly, the same qualitative conclusion applies if we switch off the Hamiltonian and consider a measurement-only Lindblad dynamics induced by string operators, similarly to what has been studied in Ref. <cit.> for measurement-only models with nonlocal measurements of Pauli strings. The paper is organized as follows. In Section <ref> we briefly review the theory of quadratic fermionic Hamiltonians on a lattice. In Section <ref> and Appendix <ref> we present the technical details required to implement the dynamics of a Gaussian state under the effect of Gaussian-preserving exponentials of quadratic operators. We first set the general framework, and then shed light on the dynamical equations and on the normalization of the state. In Section <ref> we apply such a machinery to string operators, a particular class of operators which generalize the one discussed in Ref. <cit.>. Subsequently, in Section <ref> we present and discuss numerical results for the Ising chain with non-local string measurement operators. Finally, our conclusions are drawn in Section <ref>. The other appendices contain some details on the quantum-jump measurement protocol (Appendix <ref>) and an alternative approach to determine the evolved state with the string operator (Appendix <ref>). § FERMIONIC GAUSSIAN STATES We consider a generic free-fermion system on a lattice with L sites, described by the quadratic Hamiltonian Ĥ = ∑_i,j=1^L ( Q_i,j ĉ_i^†ĉ_j + P_i,j ĉ_i^†ĉ_j^† + h.c.), where ĉ^(†)_j denotes the anticommuting annihilation (creation) operator on the j-th site. To ensure the Hermiticity of Ĥ, the coefficient matrices obey Q^† = Q and P^T = - P. In this context, a Gaussian state has the generic form [Hereafter we omit extrema in all the summations running from 1 to L.] |ψ⟩ = 𝒩exp{1/2∑_p,q Z_p,q ĉ_p^† ĉ_q^†}|0⟩_ĉ, with Z = - ( u^†)^-1 v^† , and 𝒩 a normalization constant. Here u and v are L× L matrices that obey the following symplectic unitarity conditions u^† u + v^† v = I, u v^† + v^* u^T = 0 , where I denotes the L × L identity matrix. These conditions are equivalent to say that the matrix = [ u v^*; v u^* ] is unitary and preserves the fermionic anticommutation relations: In fact it defines a new set of fermionic anticommuting operators γ̂_κ = ∑_j( u^*_j,κ ĉ_j + v^*_j,κ ĉ_j^†) , that enjoy the property γ̂_κ|ψ⟩ = 0, ∀ κ∈{1, …, L} , meaning that |ψ⟩ is the vacuum of the γ̂ fermions, |ψ⟩≡|0⟩_γ̂ (the matrix is the so-called Bogoliubov transformation. For more details see, e.g., Refs. <cit.>). We remark that, because of the fermionic nature of the ĉ-operators, the matrix Z in Eq (<ref>) is antisymmetric, as can be proved by exploiting Eqs. (<ref>) <cit.>. In this context, the Nambu spinor notation is useful. Defining Ψ̂≡ (ĉ_1, ⋯ , ĉ_L, ĉ^†_1, ⋯ , ĉ_L^†)^T, and Φ̂≡ (γ̂_1, ⋯ , γ̂_L, γ̂^†_1, ⋯ , γ̂_L^†)^T, one can rewrite Eq. (<ref>) as Φ̂= ^†Ψ̂ . The state in Eq. (<ref>) is Gaussian, enjoys the Wick's theorem, and thus it is completely determined by two-point correlation functions. Using Eq. (<ref>) one can show that these correlation functions have a simple expression in terms of the matrix: G_i,j≡⟨ĉ_i^† ĉ_j|=⟩[ v v^†]_i,j, F_i,j≡⟨ĉ_i ĉ_j|=⟩[ v u^†]_i,j . An important case where the state has the Gaussian form in Eq. (<ref>) is the ground state of the Hamiltonian (<ref>). The Hamiltonian reads Ĥ = 1/2 Ψ̂^† ℍ Ψ̂+ const. , with ℍ = [ Q P; - P^* - Q^* ]. The 2L × 2L Hermitian matrix ℍ is called the Bogoliubov-de Gennes Hamiltonian matrix, which can be diagonalized through the implementation of a Bogoliubov transformation of the form (<ref>), such that ^-1 ℍ = diag [ ω_κ, -ω_κ ]_κ=1, ⋯, L. The above transformation defines a new set of fermionic quasi-particles γ̂_κ [Eq. (<ref>)], according to which the Hamiltonian of Eq. (<ref>) can be written as Ĥ = ∑_κω_κγ̂_κ^†γ̂_κ + const. . Its ground state is a Gaussian state of the form in Eq. (<ref>), that corresponds to the vacuum of the γ̂_κ operators [see Eq. (<ref>)]. We note that the above formalism can be easily generalized to the case of time-varying coefficients Q_i,j(t) and P_i,j(t) in Eq. (<ref>), by admitting a time dependence of the Bogoliubov transformation matrix = (t). Moreover, as we shall see below in Section <ref>, the form of a generic Gaussian state Eq. (<ref>) is preserved by the application of the exponential of any operator quadratic in the ĉ fermions. § EVOLVING GAUSSIAN STATES We now discuss the evolution of a generic Gaussian state |ψ⟩ under the effect of an operator of the form M̂(s) = ^ξ s Â, being s ∈ℝ, ξ = {1, i}, and  = ∑_i,j( D_i,j ĉ^†_i ĉ_j + O_i,j ĉ^†_i ĉ_j^† + h.c.), a generic Hermitian quadratic operator. For simplicity of notations, we assume all coefficients in  to be real, but our analysis can be easily generalized to the complex case. Due to the fermionic anticommutation rules, we have D^T = D, O^T = - O. Notice that, if ξ=i, the operator in Eqs. (<ref>) describes the real-time evolution according to a (pseudo)-Hamiltonian Â. On the other hand, if ξ=1, it describes an imaginary-time evolution (related to the construction of the thermal ensemble at inverse temperature β = 1/s), that can be thought of as the analytic continuation of the purely imaginary evolution. As discussed in Section <ref>, in some particular cases, the action played by such kind of operator on a given state can be intended as part of a measurement process. In what follows we discuss how to evaluate, for any (real) symmetric  operator, the s-evolved state |ψ(s)⟩ = M̂(s) |ψ⟩ , taking as initial state |ψ⟩ a generic Gaussian state of the form (<ref>). In Appendix <ref> we show that, because of the quadratic structure of Â, the state |ψ(s)⟩ keeps the Gaussian form of Eq. (<ref>): |ψ(s)⟩∝exp{1/2∑_p,q Z_p,q(s) ĉ_p^† ĉ_q^†}|0⟩_ĉ, with Z(s) = -[ u^†(s) ]^-1 v^†(s) and u(s), v(s) submatrices of the s-evolved Bogoliubov matrix (s). In the next Subsection we are going to show how to evaluate the matrix (s) (an alternative derivation is provided in Appendix <ref>). §.§ Bogoliubov matrix evolution To characterize the evolved state |ψ(s)⟩, one only needs to know how the Bogoliubov matrix (s) evolves with M̂(s). Exploiting the identity [ M̂(s) ]^-1 = M̂(-s), we can write |ψ(s)⟩= M̂(s) |ψ⟩∝^𝒵̂(s)M̂(s) |0⟩_ĉ, where |ψ⟩ is a generic Gaussian state given by Eq. (<ref>). The operator 𝒵̂(s) = 1/2∑_p,q Z_p,q ĉ_p^†(s) ĉ_q^†(s) is defined through the evolved fermions ĉ_q^(†)(s) = M̂(s) ĉ_q^(†) M̂(-s) , obeying anticommutation relations {ĉ_q(s), ĉ_p^†(s) } = δ_p,q. Note that, for ξ=1, the transformation M̂(s) is not unitary and the conjugation operation does not commute with the evolution one, hence ĉ_q^† (s) [ M̂(s) ĉ_q M̂(-s) ]^†. Since M̂(s) |0⟩_ĉ∝|0⟩_ĉ, to determine the evolved state, one has to calculate 𝒵̂(s). This can be done by evaluating ∂_s ĉ_q^(†)(s) = ξ M̂(s) [ Â, ĉ_q^(†)] M̂(-s). Given the quadratic structure of the operator Â, the above commutators read [ Â, ĉ^†_q ] = -2∑_i (D_i,q ĉ_i^† + O_q,i ĉ_i), [ Â, ĉ_q ] = -2∑_i (D_i,q ĉ_i + O_q,i ĉ_i^†). Therefore ∂_s ĉ_q^†(s) = - 2 ξ ∑_i ( O_q,i ĉ_i(s) + D_i,q ĉ_i^†(s) ), ∂_s ĉ_q(s) = -2 ξ ∑_i ( D_i,q ĉ_i(s) + O_q,i ĉ_i^†(s) ). At this point, we make the Ansatz that the fermionic operators depend on s only through a Bogoliubov-like matrix Ψ̂(s) = (s) Φ̂, with (s) = [ (s) (s); (s) (s) ] . The expression for the submatrices of the (s) is given in Eq. (<ref>). Note that, while for ξ=i one has u̅^ev(s) = [u^ev(s)]^* and v̅^ev(s) = [v^ev(s)]^* (so that (s) is a true Bogoliubov transformation), in general this is not true. The above Ansatz is equivalent to assume that the s-evolved state |ψ(s)⟩ is a Bogoliubov vacuum γ̂_μ(s)|ψ(s)⟩ = 0, ∀μ. By substituting this Ansatz in Eq. (<ref>), we obtain a set of differential equations for the elements of the evolved matrix: ∂_s _q,μ(s) = -2 ξ ∑_i ( D_q,i _i,μ(s) + O_q,i _i,μ(s) ), ∂_s _q,μ(s) = -2ξ ∑_i ( D_q,i _i,μ(s) + O_q,i _i,μ(s) ), ∂_s _q,μ(s) = -2 ξ ∑_i ( D_q,i _i,μ(s) + O_q,i _i,μ(s) ), ∂_s _q,μ(s) = -2 ξ ∑_i ( D_q,i _i,μ(s) + O_q,i _i,μ(s) ), with the initial condition: _q,μ(0) = [_q,μ(0)]^*= u_q,μ, and _q,μ(0) = [_q,μ(0)]^* = v_q,μ. In a compact form, these equations read ∂_s (s) = -2 ξ 𝔸 (s), with 𝔸 = [ - D - O; -O - D ], being D^T = D, O^T = -O. This leads to the formal solution (s) = ^-2ξ𝔸s. [For s = it and  = Ĥ, it returns the usual Hamiltonian time evolution of the system <cit.>.] The matrix (s) is not a priori characterizing a fermionic state in that it should be unitary, to guarantee the normalization of the state, and should obey Eqs. (<ref>), to impose the anticommutation relations for the Ψ̂(s). While the former requirement is automatically fulfilled for ξ = i (because of the unitarity of the evolution operator), in the case ξ = 1 it must be forced by performing a QR-decomposition, as discussed in details in the next subsection. Once the unitarity of the matrix is imposed, Eqs. (<ref>) are satisfied if the Bogoliubov matrix preserves the symplectic structure of Eq. (<ref>), and this happens for  being Hermitian. By substituting the normalized s-evolved Bogoliubov matrix in Z(s) [cf. Eq. (<ref>)], we obtain the expression for |ψ(s)⟩. If the above conditions are satisfied, this is a Gaussian fermionic state, as the one in Eq. (<ref>). A more detailed discussion about this point is provided in Appendix <ref>. §.§ QR decomposition We now discuss how to restore the unitarity of the s-evolved Bogoliubov matrix (s), when considering a norm-non preserving evolution with ξ=1, that is in general obtained by QR-decomposing the s-evolved Bogoliubov matrix (s). This procedure keeps the Z(s) matrix in Eq. (<ref>) – and then the Gaussian state – invariant, as we are going to see below. Nevertheless, it is necessary to apply the known formulae <cit.> for evaluating the entanglement entropy and the local observables. The QR-decomposition is a procedure that allows one to decompose a generic matrix as K=U_QR, with U_Q a unitary matrix and R an upper triangular one. In what follows, we provide an argument to show that, if Â=Â^†, the unitary matrix obtained by QR-decomposing the s-evolved Bogoliubov matrix (s) describes a fermionic state, i.e., it has the symplectic form of Eq. (<ref>). The straightforward way would be to check it by construction, a procedure that requires some involved calculations. Here we choose a different strategy. Since R is positive definite (we have checked it numerically), the QR decomposition must be unique. Therefore, we assume U_Q to have the right structure and then we show that it exists a decomposition that is compatible with this assumption. Let us consider a symmetric operator  defined in Eq. (<ref>) and the associated coefficient matrix 𝔸 defined in Eq. (<ref>). Because of the symmetries of D and O, we have that {D, O} = 0. [This simply follows from the fact that ({D, O})^† = - {D, O}. Being D and O real matrices, it implies that the anticommutator vanishes.] We notice that 𝔸^2n = [ X^2n 0; 0 X^2n ], X^2 = D^2-O^2. By exploiting the block diagonal form of Eq. (<ref>) and the identities ∑_n=0^∞x^2n/(2n)! = cosh(x), ∑_n=0^∞x^2n/(2n+1)! = sinh(x)/x, we can write ^-2ξ𝔸 s = ∑_n=0^∞[(-2 ξ𝔸 s)^2n/(2n)! + -2ξ𝔸 s (-2 ξ𝔸 s)^2n/(2n+1)!] = [ ch^+(s) sh(s); -sh(s) ch^-(s) ], where, for shortness of notation, we have posed ch^±(s ) = I cosh(-2ξ X s) ± D sinh(-2 ξ X s) X^-1, sh(s ) = O sinh(-2 ξ X s ) X^-1. Notice that [ch^±(s )]^† = ch^±(s ), while sh^†(s ) = - sh(s ). Moreover, since {D, O} = 0, we have [D, ch^±(s )] = [O, ch^±(s )] = 0 and [D, sh(s )] = [O, sh(s )] = 0. This means that [ u^ev(s) v̅^ev(s); v^ev(s) u̅^ev(s) ] = [ ch^+(s) u + sh (s) v ch^+(s ) v^* + sh (s) u^*; - sh(s) u + ch^-(s) v -sh(s) v^* + ch^-(s) u^* ]. We can QR-decompose (s) (s) = (s) R(s) , making the Ansatz that the (s) preserves the symplectic form of Eq. (<ref>), so that (<ref>) can be rewritten as [ u^ev(s) v̅^ev(s); v^ev(s) u̅^ev(s) ] = [ u(s) v^*(s); v(s) u^*(s) ][ r_1(s) r_3(s); 0 r_2(s) ]. In what follows, we show that there exists a decomposition compatible with this Ansatz, by imposing Eq. (<ref>) and that, because of the unicity, it must be the only possible. Let us assume to have derived r_1(s) by construction. We can explicitly write (s) = ^ev(s) R(s)^-1, exploiting the fact that the inverse of an upper block triangular matrix remains an upper block triangular matrix: [ u(s) v^*(s); v(s) u^*(s) ] = [ u^ev(s) r_1^-1(s) u^ev(s) r_3^-1(s) + v̅^ev(s) r_2^-1(s); v^ev(s) r_1^-1(s) v^ev(s) r_3^-1(s) + v̅^ev(s) r_2^-1(s) ]. We can substitute these expressions in Eq. (<ref>) to obtain u^† u + v^† v = [ r_2^T(s) ]^-1 ( [u̅^ev]^T(s) u^ev(s) + [v̅^ev]^T(s) v^ev(s) ) [ r_1(s) ]^-1 = I. Since [u̅^ev]^T(s) u^ev(s) + [v̅^ev]^T(s) v^ev(s)= I, the above equation provides a relation between two of the blocks of the matrix R(s): r_2(s) = [ r_1^T(s) ]^-1. By imposing the second constraint in Eq. (<ref>), we find the relation for r_3(s) r_3(s) = -r_1^T ([u^ev]^†(s) v̅^ev(s) + [v^ev]^†(s) u̅^ev(s))^-1( [u^ev]^†(s) u^ev(s) + [v^ev]^†(s) v^ev(s) ). As a last comment we notice that the matrix Z(s) in Eq. (<ref>) – and then the Gaussian state – are left invariant by the application of the QR decomposition: Z(s) = [ u^† (s) ]^-1 v(s)^† = {[ u^ev(s) r_1^-1(s) ]^†}^-1 [ v^ev(s) r_1^-1(s) ]^† = { [ u^ev]^† (s) }^-1 [ v^ev]^†(s) . § NUMBER-PRESERVING OPERATORS We now consider a generic number-preserving operator Â_ℐ, 𝒥 = ∑_i ∈ℐ∑_j ∈𝒥ĉ_i^†ĉ_j, where ℐ, 𝒥 are two sets of indices labeling certain sites on the system, and apply = ^β to a Gaussian state. We focus on this particular example of the theory presented in Section <ref>, since it generalizes the construction discussed in Ref. <cit.>. By exploiting the anticommutation relations, it is possible to show that ^n = α^n-1, with α = |ℐ∩𝒥|. In fact ^2 = ∑_i, l ∈ℐ ∑_j, m ∈𝒥ĉ^†_i ĉ_j ĉ^†_l ĉ_m = ∑_i, l ∈ℐ ∑_j, m ∈𝒥(ĉ^†_i ĉ_j δ_l,m+ ĉ_i^†ĉ_j ĉ_m ĉ^†_l) = ∑_i, l ∈ℐ ∑_j, m ∈𝒥[ ĉ^†_i ĉ_j δ_l,m+ 12ĉ_i^† (ĉ_j ĉ_m +ĉ_m ĉ_j)ĉ^†_l] = |ℐ∩𝒥| ∑_i∈ℐ ∑_j∈𝒥ĉ^†_i ĉ_j, where the last equality holds, since j and m run in the same set. Therefore, if ℐ∩𝒥 = ∅, we have ^2 =0 and = 𝕀 +, where 𝕀 denotes the identity operator. Otherwise = ∑_n=0^∞(β)^n/n! = 𝕀 + (^αβ-1) /α. In what follows, without loss of generality, we set β≡log(α+1)/α, so that = 𝕀 + reduces to that studied in Ref. <cit.> for ℐ = 𝒥 = {i}. As discussed in Appendix <ref>, the application of such a can be thought as the measurement step of a quantum-jump dynamics. We want to evaluate the effect of this operator on a Gaussian state |ψ⟩_≡|ψ⟩ = (𝕀 + )|ψ⟩ . Following the procedure discussed in Section <ref>, we write the evolved state as |ψ⟩_∝exp{1/2∑_p,q Z_p,q k̂^†_p k̂^†_q }|0⟩, where we introduced the transformed fermionic operators k̂_q^(†) = ĉ_q^(†) ^-1. By exploiting the fact that ^-1 = 𝕀 - (α+1)^-1, we derive the exact expression of k̂^(†) k̂_q^† = ĉ^†_q + ∑_j∈𝒥δ_q,j∑_i∈ℐĉ^†_i, k̂_q = ĉ_q - (α+1)^-1∑_j∈𝒥δ_q,j∑_i∈ℐĉ_i. This leads to the following expressions for the evolved Bogoliubov matrix: u_q,μ^ev = u_q,μ - (α+1)^-1∑_j∈𝒥δ_q,j∑_i ∈ℐ u_i, μ, u̅_q,μ^ev = u^*_q,μ + ∑_j∈𝒥δ_q,j∑_i ∈ℐ u^*_i, μ, v̅_q,μ^ev = v^*_q,μ -(α+1)^-1∑_j∈𝒥δ_q,j∑_i ∈ℐ v^*_i, μ, v_q,μ^ev = v_q,μ + ∑_j∈𝒥δ_q,j∑_i ∈ℐ v_i, μ. As discussed in Appendix <ref>, an alternative derivation can be obtained from the evolved antisymmetric matrix Z. §.§ Local operator The above results can be applied straightforwardly to local (l) operators, as for the one used in Ref. <cit.>, ^l≡n̂_i = ĉ_i^†ĉ_i, such that ^l≡^log(2)Â^l = 𝕀 + ĉ_i^†ĉ_i. The Bogoliubov matrix, evolves according to Eqs. (<ref>), which read u_q,μ^ev, l = (1 - δ_q,i/2) u_q,μ, u̅^ev, l_q,μ = (1+δ_q,i) u^*_q,μ, v̅^ev, l_q,μ = (1 - δ_q,i/2) v^*_q,μ, v_q,μ^ev, l = (1+δ_q,i) v_q,μ. This evolution is implemented by applying to the a block-diagonal matrix with entries _i+L, i+L = _i,i^-1 = 2 (see Eq. (1) of Ref. <cit.>). §.§ Two-site operator We now consider a non-local (nl) operator with support on two sites (ℐ=𝒥 = { i, j }): ^nl = ( ĉ_i^† + ĉ_j^†) ( ĉ_i + ĉ_j) , Since α = |ℐ∩𝒥| = 2, we have ^nl≡^log(3)/2Â^nl = 𝕀 + ^nl. Without losing in generality, in the following we will focus on one dimensional systems and set j=i+r (with r>0). As we shall see in the next Section, a measurement operator as the one in Eq. (<ref>) may unveil a richer physics than local ones. In fact, in the spin-1/2 language, it corresponds to a string operator. The evolved Bogoliubov matrix can be derived from Eqs. (<ref>) which, in this case, read u_q,μ^ev, nl = u_q,μ - 13(δ_q,i + δ_q,i+r)(u_i, μ + u_i+r, μ), v̅^ev, nl_q,μ = v^*_q,μ - 13(δ_q,i + δ_q,i+r)(v^*_i, μ + v^*_i+r, μ), u̅^ev, nl_q,μ = u^*_q,μ + (δ_q,i + δ_q,i+r)(u^*_i, μ + u^*_i+r, μ), v^ev, nl_q,μ = v_q,μ + (δ_q,i + δ_q,i+r)(v_i, μ + v_i+r, μ). § QUANTUM JUMPS WITH STRING OPERATORS In this section we use the machinery presented above to describe the evolution of a quantum Ising chain under the effect of a quantum-jump dynamics generated by ^nl [cf. Eq. (<ref>) and below], with emphasis on the evolution of the entanglement entropy. We consider the Hamiltonian Ĥ = -J ∑_j σ̂_j^x σ̂_j+1^x - h∑_j σ̂_j^z, where σ̂_j^α (α = x, y, z) are the usual spin-1/2 Pauli matrices acting on the j-th site, while J>0 denotes the ferromagnetic spin-spin coupling and h the transverse magnetic field. In the thermodynamic limit, the ground state exhibits a zero-temperature continuous transition at |h/J|=1, separating a paramagnetic (symmetric) phase, for |h/J| > 1, from a ferromagnetic (symmetry-broken) phase, for |h/J| < 1 <cit.>. Through a Jordan-Wigner transformation <cit.>, the Ising model (<ref>) can be mapped into the so-called Kitaev chain Ĥ = -J∑_j( ĉ_j^†ĉ_j+1 + ĉ_j^†ĉ_j+1^† + h.c.) - 2 h ∑_j ĉ^†_j ĉ_j . This Hamiltonian is of the form (<ref>), with Q_j,j = h, Q_j, j+1 = Q_j+1, j = -J/2, P_j, j+1 = - P_j+1, j = -J/2 . Hereafter we work in units of ħ = 1, set J = 1 as the energy scale, and assume periodic boundary conditions for fermions. We implement the quantum-jump protocol described in Appendix <ref> on the fermionic model above, by initializing the system in the ground state of the Hamiltonian (<ref>) with a given value of h and then Trotterizing the time evolution in steps δ t ∝ (4Nγ)^-1, to ensure convergence, being γ the strength of the measurements. The emerging dynamics is one of the possible unravelings <cit.> of the Lindblad master equation <cit.> ρ̇(t) = - i [ Ĥ, ρ(t) ] - γ∑_j ( ρ - 12{, ρ}) , being =(ĉ_j^† + ĉ_j+r^†)(ĉ_j + ĉ_j+r) the same operator as in Eq. (<ref>), with i≡ j + r. [Note that, with our convention for the boundary conditions, we have ĉ^(†)_j+r≡ĉ^(†)_j+r-L for any j+r >L.] In the spin language, this operator is a sum of strings of Pauli matrices, of length r: = ∏_k < jσ̂_k^z σ̂_j^- ∏_k' < j+rσ̂_k'^z σ̂_j+r^+ + ∏_k < j+rσ̂_k^z σ̂_j+r^- ∏_k' < jσ̂_k'^z σ̂_j^+ + 12( σ̂_j^z + σ̂_j+r^z -2 ) , with σ̂_j^± = σ̂_j^x ± i σ̂_j^y. Hereafter, we refer to it as the string operator. We point out that our framework is similar to the one presented in Ref. <cit.>, although with some important qualitative differences. The former is described by a random sequence of stabilizer-formalism measurements of Pauli string operators. These operators have a fixed range r and are taken from a random distribution of strings, located around different sites, and made of different sets of onsite Pauli operators. In this work, in contrast, we are following an unraveling of the quantum master equation Eq. (<ref>). As detailed in Appendix <ref>, we apply random quantum-jump measurements of string operators 𝕀+Â_j(r), located at different sites, but with the structure fixed as the one in Eq. (<ref>). This is sufficient to achieve a measurement-induced nontrivial dynamics, with important consequences on the entanglement generation. It is important to emphasize that the operators 𝕀+Â_j(r) have the exponential form given by Eqs. (<ref>) and (<ref>) (see also Sec. <ref>). In this way, the state keeps its Gaussian form under the application of the quantum jumps, and a numerical analysis of quite large system sizes is possible. §.§ Entanglement entropy Let us focus our attention on the entanglement entropy. Given a pure state |ψ⟩, there is an unambiguous way to quantify its quantum correlations. We divide the system in two subchains, one with length ℓ and the other with length L-ℓ. After evaluating the density matrix ρ_ℓ=Tr_L-ℓ[|ψ⟩⟨ψ|] reduced to the subsystem ℓ, the entanglement entropy is quantified by its von Neumann entropy <cit.>: S_ℓ [ρ_ℓ] = - ρ_ℓlogρ_ℓ . For the Ising model we are discussing herewith, it is known that after, a sudden quench in a purely Hamiltonian dynamics, the entanglement entropy approaches a stationary value that increases linearly with ℓ, thus obeying a volume-law behavior <cit.>. As discussed in Refs. <cit.>, this dynamical volume law is generally destroyed by any arbitrary small local weak measurement process. In this scenario, the asymptotic entanglement entropy is conjectured to exhibit either a logarithmic scaling or an area-law behavior, depending on the strength of measurements and of the Hamiltonian parameters. We are going to show that the situation is quite different for the case of the nonlocal operators defined in Eq. (<ref>). In practice, to avoid spurious contributions from classical correlations, we evaluate Eq. (<ref>) on each stochastic quantum trajectory |ψ_t⟩ (defined in Appendix <ref>), and then average over different stochastic realizations. [For details on how to evaluate the von Neumann entropy of the reduced density matrix of Gaussian states see, e.g., Ref. <cit.>.] For simplicity of notation, we denote the ensemble-averaged entanglement entropy as S_ℓ(t) ≡⟨⟨ S_ℓ[ρ_ℓ^stoch(t)] ⟩⟩, ρ_ℓ^stoch(t)=Tr_L-ℓ[|ψ_t⟩⟨ψ_t|], and where the symbol ⟨⟨⋯⟩⟩ denotes the average over different stochastic realizations. We also define the asymptotic entanglement entropy as S̅_ℓ = lim_T→∞∫_t^⋆^T dt' S_ℓ(t'). [ Ensemble-averaged entanglement entropies have then been obtained with N_ avg = 10^2 different quantum trajectories. Subsequently, to perform time averages, we fix a given evolution time T and average in the interval t'∈ [0.7 T, T].] We remark that, in all the simulations shown here, we keep fixed the partition to ℓ = L/4 and study the scaling with L. Below we present two different scenarios: the measurement-only dynamics (Section <ref>) and the full quantum-jump dynamics (Section <ref>). §.§ Measurement-only dynamics We first neglect the Hamiltonian contribution, i.e., we implement the quantum-jump protocol described in Appendix <ref> with Ĥ=0. The Hamiltonian is only used to set the initial state as its ground state with h = 0.5. However, we have carefully checked that this choice does not affect the asymptotic values of S_ℓ(t), but only the transient dynamics. [Note that the perfectly paramagnetic state with spins pointing along the z direction (h=∞) is a dark state of the dissipation, therefore, initializing the system close to such state leads to slower convergence.] The measurement-only dynamics allows to isolate the effects of measuring the string operators, without considering the interplay with the unitary dynamics. The upper panels of Figure <ref> report the ensemble-averaged entanglement entropy S_ℓ(t) versus the rescaled time γ t, for three ranges of the string measurement operator: r = 1 (panel a), r = 4 (panel b), and r = L/2 (panel c), and for different system sizes L. For r=1, the measurement-only dynamics cannot generate entanglement and the system stabilizes on a low-entangled asymptotic state. This is confirmed by Fig. <ref>(d), that shows the behavior of the asymptotic entanglement entropy S_ℓ versus L, which tends to saturate to a constant value, thus obeying an area-law behavior. As expected, this result is analogous to the case where onsite operators are measured <cit.>. When the correlation range of the dissipative string operator is increased, as for r=4, we may identify three regimes: the short-time transient, the metastable plateau, and the asymptotic dynamics. At short times the entanglement entropy [cf. inset of Fig. <ref>(b)] grows to reach a maximum, that saturates to a constant value for large L. This is clarified by the red curve in Fig. <ref>(e), showing the maximum value of the entanglement entropy versus the system size (notice the logarithmic scale of the x-axis). After the initial transient, the entanglement entropy attains a metastable plateau where the entanglement entropy exhibits an area-law behavior [green curve in Fig. <ref>(e)]. Finally, in a time τ∼ L^2, the entanglement entropy approaches a non-vanishing asymptotic value still obeying an area law [blue curve in Fig. <ref>(e)]. This scenario dramatically changes when r is forced to be a thermodynamic fraction of L (e.g., r = L/2): In that case, the measurement-only dynamics leads to a stationary state, whose entanglement obeys a volume-law behavior. This is clarified by Fig. <ref>(f), which shows the rescaled asymptotic entanglement entropy S̅_ℓ/L versus L^-1. In this plot, the saturation of S̅_ℓ/L to a finite value for L^-1→ 0 clearly suggests a linear growth S̅_ℓ∝ L for large system sizes. §.§ Effect of the Hamiltonian We now switch on the Kitaev Hamiltonian Ĥ and follow the full quantum-jump dynamics for h=0.1, starting from the ground state with h_i = 0.5. As for the case with no Hamiltonian, we have checked that the asymptotic values reached by the entanglement entropy are independent of h_i, while they may depend on h (although the qualitative behaviors with L are unaffected by this choice). Figure <ref> displays S_ℓ(t) versus γ t (upper panels) and the asymptotic value S̅_ℓ versus L, for the three ranges of r = 1 (left) r = 4 (middle), and r = L/2 (right). In panel (a) we show the local case r=1. The long-time averaged entanglement entropy attains a stationary value which is almost independent of L, as clearly visible from Fig. <ref>(d), thus reflecting an area-law behavior. Such value is generally larger than the one obtained in the measurement-only dynamics, without the entangling action of Ĥ [compare with Fig. <ref>(a)]. In Fig. <ref>(b) we show the case r=4, which displays a convergence in time faster than that for a measurement-only dynamics [compare with Fig. <ref>(b)]. The asymptotic data in Fig. <ref>(e) (x-axis in logarithmic scale) suggest that S̅_ℓ initially grows sub-extensively with L, to eventually attain an area-law regime. Physically, this can be understood by the fact that for L much larger than r, the dissipation becomes effectively local and the extensive growth is destroyed. In contrast with that, if the range r of the jump operator is comparable with the total size L, the measurements can stabilize a volume-law scaling of the entanglement, as we can see in Fig. <ref>(c), that refers to r=L/2. Here S̅_ℓ increases in time and quickly attains an asymptotic value that follows a volume-law scaling [see the data in Fig. <ref>(f) for S̅_ℓ/L versus L^-1, which display convergence to a finite value, in the L →∞ limit]. § CONCLUSIONS We have unveiled some aspects of the entanglement-entropy dynamics of fermionic many-body Gaussian states, in the presence of a Gaussian-preserving evolution, with a special focus on the possible emergence of measurement-induced phase transitions in such systems. The first part of the paper presents a detailed technical discussion on how to maintain and treat Gaussianity, when exponentials of quadratic fermion operators are applied to a Gaussian state. Our purpose is to provide the reader with a comprehensive guide to implement fermionic Gaussian evolutions and study the dynamics (not necessarily unitary) of such systems. Among all the possible treatable evolutions, we focused on the quantum-jump unraveling of a Lindblad dynamics induced by a particular class of quadratic fermion operators, having a simple definition in the fermionic language, but a non-trivial string structure in terms of the Pauli matrices. We called them string operators. In the second part of the paper, after having derived the equation of motions for a quantum-jump dynamics induced by the string operators, we have solved it for the quantum Ising chain, focusing on the time behavior of the entanglement entropy. Our main result is that the scaling of the entanglement entropy with the system size L strongly depends on the range r of the string operators. For short strings (r finite), we see that the asymptotic entanglement entropy obeys an non-extensive scaling behavior. In this case, the measurement operator is effectively local in space and thus it asymptotically destroys quantum correlations that may be generated by the unitary dynamics. For string ranges comparable with the system size (r ∼ L), the measurement process may become highly entangling and is able to stabilize a volume-law scaling in the asymptotic entanglement entropy. Remarkably, the picture above is valid both for a measurement-only dynamics and for a dynamics where also the effect of a short-range Hamiltonian is considered. So, the measurement dynamics of the string operators is strong enough to induce different entanglement scalings by itself. A natural extension of our work could be to understand how the MIPT scenario would change in the mutual presence of unitary dynamics and long-range measurement operators (e.g., by considering some power-law decay of dissipators with the distance). It would be also interesting to go beyond free-fermion models and address the role of interactions in the entanglement dynamics under variable-range monitoring. § ACKNOWLEDGEMENTS We acknowledge fruitful discussions with V. Alba, M. Collura, R. Fazio, M. Schirò, and X. Turkeshi. Funding information This work has been partly funded by the Italian MIUR through PRIN Project No. 2017E44HRF. § ROTATION TO THE DIAGONAL BASIS Here we discuss a different derivation of the s-evolved state |ψ(s)⟩=M̂(s)|ψ⟩, where |ψ⟩ is a generic Gaussian state [Eq. (<ref>)], and the M̂(s)=^ξ s is defined in Eqs. (<ref>). The operator  [Eq. (<ref>)] is quadratic in the ĉ fermions and can be written compactly as  = Ψ̂^†𝔸Ψ̂, where Ψ̂= (ĉ_1, ⋯ , ĉ_L, ĉ^†_1, ⋯ , ĉ_L^†)^T is the Nambu spinor associated to the ĉ fermions and 𝔸 the Bogoliubov-de Gennes matrix associated to Â, as defined in Eq. (<ref>). This matrix can be diagonalized by a Bogoliubov transformation 𝕎, such that 𝕎^†𝔸𝕎 = diag [ϵ_ν, - ϵ_ν]_ν=1, ⋯, L , and Â=∑_νϵ_ν(η̂_ν^†η̂_ν - η̂_νη̂_ν^†) . The unitary matrix 𝕎 defines a new set of η̂ fermions, whose associated Nambu spinor is Θ̂≡ (η̂_1, ⋯, η̂_L, η̂^†_1, ⋯, η̂_L^†)^T = 𝕎^†Ψ̂. Let us now evaluate |ψ(s)⟩. We remind that the initial state |ψ⟩ is is the vacuum of the γ̂ fermions, defined through the Bogoliubov transformation in Eq. (<ref>): Φ̂= ^†Ψ̂, where Φ̂= (γ̂_1, ⋯ , γ̂_L, γ̂^†_1, ⋯ , γ̂_L^†)^T. These fermions can be transformed into those diagonalizing 𝔸 by means of a rotation Φ̂= ^† 𝕎 Θ̂= ℚ^† Θ̂, being ℚ = [ u_η v_η^*; v_η u_η^* ] . Therefore we have the relation = 𝕎 ℚ , with given in Eq. (<ref>). In the η̂ representation, the state (<ref>) takes the form |ψ⟩∝exp{1/2∑_μ,ν Z_μ,ν^(η) η̂_μ^† η̂_ν^†}|0⟩_η̂, with Z^(η) = -[ u_η^†]^-1 v_η^†. Going on analogously as in Sec. <ref>, we can write the s-evolved state as |ψ'⟩∝exp{1/2∑_μ,ν Z_μ,ν^(η) η̂^†_μ(s) η̂^†_ν(s) }M̂(s)|0⟩_η̂ , with η̂^†_ν(s) ≡M̂(s) η̂^†_ν M̂(-s). The operator M̂(s) is number-preserving, therefore we can exploit the methods of Section <ref>. We first notice that M̂(s) |0⟩_η∝|0⟩_η, in fact ^ξ s∑_νϵ_ν(η̂_ν^†η̂_ν-η̂_νη̂_ν^†)|0⟩_η̂ = ^-ξ s∑_νϵ_ν|0⟩_η̂ , because η̂_ν|0⟩_η̂=0, ∀ ν∈{1, …, L}. Then we notice that, since (η̂_ν^†η̂_ν)^n = η̂_ν^†η̂_ν (cf. Section <ref>), we have M̂(± s) = ^±ξ s  = ^∓ξ s∑_νϵ_ν∏_ν[1+ ∑_n=1^∞(± 2 ξ sϵ_ν)^n/n!η̂_ν^†η̂_ν] = ^∓ξ s∑_νϵ_ν∏_ν[1 + (^± 2 ξ s ϵ_ν -1 ) η̂_ν^†η̂_ν] , where we have applied the diagonal form of Â, in Eq. (<ref>). Therefore, according to the fact that [ η̂_μ^†η̂_μ, η̂_ν^†] = δ_μ, νη̂_μ^†, we can write η̂^†_ν(s)= [ 1 + (^2 ξ s ϵ_ν -1 )η̂_ν^†η̂_ν] η̂_ν^†[ 1 + (^- 2 ξ s ϵ_ν -1 )η̂_ν^†η̂_ν] = ^2 ξ s ϵ_νη̂_ν^†. Substituting into Eq. (<ref>) we get |ψ(s)⟩∝exp{1/2∑_μ,νZ̃_μ,ν^(η) η̂^†_μ η̂^†_ν}|0⟩_η̂ , with Z̃^(η)_μ, ν≡ Z_μ,ν^(η) ^2ξ s(ϵ_μ+ϵ_ν). We can also write Z̃^(η) = -[ u_η^ev^†(s) ]^-1v_η^ev^†(s), with [ u_η^ev(s); v_η^ev(s) ] =[ ^-2ξ s ϵ 0; 0 ^2ξ s ϵ ][ u_η; v_η ], being ϵ = diag(ϵ_ν). At this point, if ξ=1, we apply the QR decomposition to [ u_η^ev(s); v_η^ev(s) ], as explained in Sec. <ref>. In this way, the matrix Z̃^(η) remains unchanged, but can be written as Z̃^(η)=-[ u_η^†(s) ]^-1v_η^†(s), with u_η(s), v_η(s) obeying the symplectic unitarity conditions Eq. (<ref>) and written as [ u_η^ev(s); v_η^ev(s) ]=[ u_η(s); v_η(s) ] r_1 , with r_1 some upper triangular matrix L× L. We have therefore reduced to the standard Gaussian form for the evolved state |ψ(s)⟩ in the η̂ representation, with the Z matrix given by Z̃^(η)=-[ u_η^†(s) ]^-1v_η^†(s), and the symplectic unitary matrix ℚ(s) = [ u_η(s) v_η^*(s); v_η(s) u^*_η(s) ] providing the set of fermionic operators γ̂_ν(s) annihilating |ψ(s)⟩ [we have Φ̂(s) = ℚ^†(s) Θ̂, with Φ̂(s) ≡ (γ̂_1(s), ⋯ , γ̂_L(s), γ̂^†_1(s), ⋯ , γ̂_L^†(s))^T]. At this point it is possible to come back to the ĉ representation. We can use Eq. (<ref>), and obtain the Bogoliubov matrix (s) [the one such that Φ̂(s) = 𝕌^†(s)Ψ̂] as (s) = 𝕎 ℚ(s). Applying 𝕎 on the left of both sides of Eq. (<ref>) and writing (s) = [ u(s) v^*(s); v(s) u^*(s) ] , we obtain [ u(s); v(s) ]r_1 = ^-2ξ s 𝔸[ u; v ] . [To get this result, we used the relation (s) = 𝕎 ℚ(s), as well as Eqs. (<ref>), (<ref>), and (<ref>)]. The Z matrix of the evolved state in this representation is thus Z(s)=-[ u^†(s) ]^-1 v^†(s). Notice that Eq. (<ref>) is the same result of Eq. (<ref>), when the QR decomposition discussed in Sec. <ref> is applied. As a last comment we point out that Eq. (<ref>) refers to the first two blocks of the Bogoliubov matrix only. This is consistent with the fact that, in the Nambu spinor notation, the degrees of freedom are doubled and there is some redundant information that can be neglected. § QUANTUM-JUMP PROTOCOL We start from the following master equation in the Lindblad form: ρ̇(t) = ℒ[ ρ(t) ] ≡ - i [ Ĥ, ρ(t) ] - γ∑_j ( m̂_j ρ m̂_j^† - 12{m̂_j^† m̂_j, ρ}), where ρ(t) is the density matrix describing the state of the system at time t and m̂_j^(†) are the Lindblad jump operators, while ℒ denotes the Lindbladian superoperator acting on ρ. The solution of Eq. (<ref>) is, in general, a mixed state that can be thought of as an average over many pure-state quantum trajectories of a suitable stochastic process. Given a Lindbladian, there are many stochastic processes (or unravelings) that produce the same averaged density matrix <cit.>. Among them, in this paper we focus on the so-called quantum-jump unraveling <cit.>, whose implementation is described below. We define the non-Hermitian Hamiltonian Ĥ_eff = Ĥ - iγ/2∑_j m̂_j^† m̂_j, then we rewrite Eq. (<ref>) ρ̇(t) = - i [ Ĥ_eff, ρ(t) ] - γ∑_j m̂_j ρ m̂_j^† . In this form, the Lindblad dynamics ruled by Eq. (<ref>) can be interpreted as a deterministic non-Hermitian time evolution of a pure quantum state |ψ_t⟩ generated by Ĥ_eff, plus a stochastic part given by the possibility of applying m̂_j to the state |ψ_t⟩ [cf. the second term in the right-hand-side of Eq. (<ref>)]. For m̂_j Hermitian, this dynamics can be thought of as an occasional, yet abrupt, measurement process in which, at any time, the system has a chance to be POVM-measured <cit.>, i.e., to undergo the action of one of the operators m̂_j, taken from a distribution (see more details below). In this framework, Eq. (<ref>) can be solved by implementing the following stochastic process. The time evolution is discretized in time slices of step d t. Then, at each step: * With probability π_j = γ⟨m̂^†_j m̂_j|d⟩t, the jump operator is applied: |ψ_t+dt⟩↦m̂_j |ψ_t⟩/||m̂_j |ψ_t⟩||; * With probability π = 1 -∑_j π_j the non-Hermitian evolution occurs: |ψ_t+dt⟩↦^-i Ĥ_eff dt |ψ_t⟩/||^-iĤ_eff dt|ψ_t⟩||. Note that the Lindblad master equation in Eq. (<ref>) is invariant under a shift of the jump operators m̂_j ↦𝕀+m̂_j. In practice, one can implement the protocol described above with the Lindblad operators 𝕀+m̂_j (as we actually do in Sec. <ref>), instead of m̂_j, and obtain the same averaged density matrix, although the single trajectories |ψ_t⟩ may behave differently. We remark that, since the von Neumann entropy of Eq. (<ref>) quantifies the amount of genuine quantum correlations of a pure state, we first evaluate it over the single pure-state quantum trajectory and then perform the ensemble averaging. It is important to fix the order of the two operations of measuring and ensemble averaging, since they are commuting only for linear functions of the state ρ (as for expectation values of observables), but not for quantities as the one in Eq. (<ref>). For a more detailed discussion see, e.g., Refs. <cit.>. § EVOLUTION OF THE STATE: A DIFFERENT DERIVATION In this appendix we provide a different derivation of the analytic expression for the evolved state |ψ⟩_, by evaluating the operator 𝒵̂(1)= 1/2∑_p,q Z_p,q ĉ_p^†(1) ĉ_q^†(1), in Eq. (<ref>). To this purpose we notice that [ , Ẑ] = 1/2∑_p,q∑_i∈ℐ(Z_p,𝒥δ_q,i + Z_q,𝒥δ_p,i) ĉ^†_p ĉ^†_q, where we have defined 𝒵̂ = 1/2∑_p,q Z_p,q ĉ_p^† ĉ_q^† and Z_p, 𝒥 = ∑_j∈𝒥 Z_p,j. Consequently [Ẑ, [, Ẑ] ] = 0. This allows to write ^Ẑ = ^Ẑ( + [ , Ẑ] ), that, together with |0⟩ = |0⟩, leads to |ψ⟩_ = 𝒩^Ẑ(𝕀 + [ , Ẑ] ) |0⟩. Since [ , Ẑ]^2 = 0, we can write exp{ [ , Ẑ ] } = 𝕀 + [ , Ẑ] and, therefore, |ψ⟩_ = 𝒩^Ẑ^[ , Ẑ]|0⟩≡𝒩^Ẑ'|0⟩, being Ẑ' = 1/2∑_i ∈ℐ∑_p,q[ Z_p,q +(Z_p,𝒥δ_q,i+Z_q,𝒥δ_p,i) ] ĉ^†_p ĉ^†_q. For the local case ^l in Subsec <ref>, the jumped state |ψ⟩_^l is described by the antisymmetric matrix [ Z^l]', with [ Z^l]'_p,q = (1 + δ_q,i + δ_p,i)Z_p,q. On the other hand, for the non-local case ^nl in Subsec. <ref>, the antisymmetric matrix [ Z^nl]' reads [ Z^nl]'_p, q = Z_p, q + (δ_p,i + δ_p, i+r)(Z_i,q+Z_i+r, q).

http://arxiv.org/abs/2303.07333v1

20230313175630

Lifted particles from the fast spinning primary of the Near-Earth Asteroid (65803) Didymos

astro-ph.EP

^1 Instituto de Física Aplicada a las Ciencias y la Tecnología, Universidad de Alicante, Spain. P.O. Box 99, 03080 Alicante, Spain ^2 Observatorio Astronómico, Universidad Nacional de Córdoba, Laprida 854, Córdoba X5000GBR, Argentina ^3 Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), Argentina ^4 Departamento de Física, Ingeniería de Sistemas y Teoría de la Señal, Universidad de Alicante, Spain. P.O. Box 99, 03080 Alicante, Spain ^5 Instituto de Astrofísica de Andalucía - CSIC, Glorieta de la Astronomía s/n, 18008 Granada, Spain Asteroids – Asteroids, dynamics – Asteroids, rotation – Near-Earth objects – Regoliths. An increasing number of Near Earth Asteroids (NEAs) in the range of a few hundred meters to a few kilometres in size have relatively high spin rates, from less than 4 h, down to ∼2.2 h, depending on spectral type. For some of these bodies, local acceleration near the equator may be directed outwards so that lift off of near-equatorial material is possible. In particular, this may be the case for asteroid Didymos, the primary of the (65803) Didymos binary system, which is the target of the DART (NASA) and Hera (ESA) space missions. The study of the dynamics of particles in such an environment has been carried out —in the frame of the Hera mission and the EC-H2020 NEO-MAPP project— according to the available shape model, known physical parameters and orbital information available before the DART impact. The presence of orbiting particles in the system is likely for most of the estimated range of values for mass and volume. The spatial mass density of ejected material is calculated for different particle sizes and at different heliocentric orbit epochs, revealing that large particles dominate the density distribution and that small particle abundance depends on observation epoch. Estimates of take off and landing areas on Didymos are also reported. Available estimates of the system mass and primary extents, after the DART mission, confirm that the main conclusions of this study are valid in the context of current knowledge. Lifted particles from the fast spinning primary of the Near-Earth Asteroid (65803) Didymos Paula G. Benavidez^1,4 March 30, 2023 ========================================================================================== § INTRODUCTION Many Near Earth Asteroids (NEAs) have been discovered in the last two decades, including at sizes smaller than a few km and fast spin rates. <cit.> showed that an excess of slow rotators (spin period T > 24 h), as well as fast rotators (2.2 h <T< 2.8 h), is present in the NEA population at size D > 0.2 km. The former is also found in main belt asteroids, while the excess of fast rotators is not, and it seems to be peculiar to NEAs. It is noteworthy that 2/3 of fast spinning NEAs are binaries <cit.> and they correspond to the concentration of fast spin rate of primaries of NEA binaries in front of the “spin barrier” at ≈ 2.2 h. The spin state with which asteroids enter the NEA region is mostly affected by non-catastrophic collisions while in the asteroid belt <cit.>, and partially by the non-gravitational YORP effect. YORP is also the main driving source for NEAs spin up once in the inner planet region. §.§ Asteroid stability limits and spin–up We recall here the spin limits for mass shedding defining an upper threshold on the rotation rate at which a particle at the equator of a spherical body is at neutral equilibrium between gravity and centrifugal force. For a homogeneous sphere, the relation for the critical spin rate is: ω_cr=√(4 π G ρ/3) where ρ is the bulk density of the object. Spin limit depends on asteroid density: e.g., T = 2.2 h is the limit corresponding to a spherical body of density ρ=2250 kg/m^3, which is typical of S-type gravitational aggregate asteroids; instead, spin limit is T=3.0 h for C-type gravitational aggregate asteroids, with ρ=1200 kg/m^3. Such densities differ from the density of their meteorite analogues due to macro–porous structure of gravitational aggregates. As soon as enough spin data were available <cit.>, such a spin limit became evident when plotting size vs spin period. Very rarely asteroids larger than 200–300 m in diameter (D) have been observed with a rotational period smaller than 2.2 h. This has led to interpretations of the internal structure of such asteroids. Fast rotators are instead common at D<200 m. Focusing the discussion on the size range 300 m — 10 km, the interpretation of the spin limit leads to two slightly different regimes <cit.>. For asteroids larger than 3 km, the spin barrier does not constrain whether these are strength–less objects or just cracked but coherent bodies. In fact, the upper limit on the tensile strength –given by the barrier itself– is higher than a scaled tensile strength of cracked but coherent bodies. Instead, for asteroids smaller than 3 km, the maximum possible tensile strength allowed by the spin barrier is too low for these asteroids to be cracked but coherent bodies, so they should have predominantly cohesion–less structures. Non–coherent asteroids are aggregates that have re-accumulated fragments by self–gravity right after shattering events. On the contrary, internally cracked objects may arise due to shattering at the threshold energy for fragmentation, with little kinetic energy left to reshuffle fragments. Another way of producing cracked structures is by series of sub–catastrophic collisions summing up similar damage in the overall structure as one single barely shattering event <cit.>. In that case, the object may be coherent allowing for some tensile stress. Both coherent bodies and gravitational aggregates (GA) (often called rubble piles) may withstand spin rates higher than the critical ones for fluids found by Chandrasekhar <cit.>, and spin ideally up to the spin barrier around 2.2 h before failing apart. In the case of coherent –monolithic– structure, that is due to internal solid state forces, which do withstand spin rates beyond that limit. This is especially the case of bodies smaller than 300 m. Instead, non-coherent asteroids (as gravitational aggregates are) may be spun up and undergo shape change corresponding to a minimum energy configuration –led by dissipative forces (e.g., internal friction)– compatible with increasing angular momentum. That is achieved by rotation about the maximum angular inertia axis. As a result, some of those bodies might become top–shaped <cit.>. Therefore, shear strength may be present due to friction and interlocking between GA components <cit.>, as a reaction to the shear stress due to centrifugal force, increasing structural yield. The presence of inter–particle cohesion has also been suggested <cit.>, though still a matter of debate. Such shear strength, regardless of its nature, may prevent the whole structure from falling apart when the rotation spin rate exceeds the stability limit for fluid bodies. This mechanism is successful until the spin barrier is reached. At that point, the body is no longer able to adjust the exceeding energy and angular momentum by shape change through energy dissipation by friction. Depending on internal stiffness, fission or mass shedding takes place, eventually leading to asteroid binary or pair/clan formation <cit.>. The mechanisms of formation of NEA binary systems are a matter of debate and are beyond the scope of this study. §.§ NEA binary systems with fast spinning primary Many binary systems in the NEA population share a number of common features: (a) small mass ratio (≈ 0.01) for satellite to primary components; (b) fast spin primary; (c) top-shape primary <cit.>. Of all binary systems with fast rotating primary, we identified a handful near the edge of stability, as reported in Table <ref> together with a number of single bodies. In this work, we focus on the Didymos system because of its interest as the target of both DART (NASA) and Hera (ESA) space missions. However, the 1996 FG3 binary system also was the former goal of the MarcoPolo-R mission and is now the sample return target of the Chinese (CNSA) Zheng He space mission. Beyond binary primaries, it is worth mentioning that some lonely top–shape NEAs also show fast spin rates. This is indeed the case of 2008 EV5, which was the target of the un–selected ESA (2014) MarcoPolo-R mission. §.§.§ The Didymos system The NEA binary (65803) Didymos is the S-type target of the DART (NASA) and Hera (ESA) space missions. Model predictions may be tested by combined data from those two missions, which makes this a particularly interesting system to study. This asteroid is classified as an Apollo NEA with a semimajor axis of 1.6442688843 ± 1.6 × 10^-9 au, and a large eccentricity of 0.383882802 ± 3× 10^-9 (pre-DART impact heliocentric ephemeris solution 181). Its perihelion is therefore well inside the inner asteroid belt, where the asteroid spends 1/3 of its orbital period. The pre-–impact estimations of the system main physical characteristics were the values available when this research started, and numerical simulations were run. The discussion on how post–DART impact estimation may affect results is carried out in Sec. 5. Didymos, the primary of the binary, had estimated principal axes extent sizes of 832 (± 3%)× 837 (± 3%)× 786 (± 5%) km, the last of which is the size along its spin axis. A Didymos shape model has been derived using both radar and optical telescope data, clearly indicating a top–shape <cit.> before the DART and LICIACube missions imaging. The mass of the system was estimated as 5.278 × 10^11± 10% kg <cit.> from the orbital period of the secondary, named Dimorphos, which is known, 11.9216289 ± 2.8 × 10^-6 h (S. Naidu and S. Chesley, personal communication). The spin period of Didymos is 2.2600± 0.0001 h <cit.> and its bulk density was estimated to be 2170 kg/m^3, with a 30% uncertainty. The size of Dimorphos was estimated to be 164± 10 m, compatible with oblate to prolate shape with axes ratio between 0 and 1.3. No information on the spin rate of Dimorphos is available, though it is assumed to be synchronous to its orbital period. Separation between components is 1.19± 0.03 km <cit.>. Available data do not allow predictions about the internal structure of any of the two bodies. Nevertheless, updated estimates of some physical parameters critical to this study available during the publication process of this work <cit.> confirm L and LL ordinary chondrites as the best meteorite analogues for Didymos. Considering that typical grain density of such meteorite complexes is in the 3500-3600 kg/m^3 range, the system bulk density is compatible with at least 30% bulk porosity of its components. Therefore, the primary may have a gravitational aggregate structure with unknown size distribution of components. §.§ Asteroids on the edge of stability For some NEAs, the centrifugal force acting on surface particles and boulders at near–equatorial latitudes may slightly overcome the gravitational pull of the asteroid itself in the spinning, non–inertial reference frame of the rotating asteroid. In that case, the radial component of acceleration for surface particles is directed outwards, allowing them to leave the surface and undergo corresponding dynamical evolution. Leaving the surface does not mean that particles are lost from the asteroid. In fact, they start their motion at zero velocity but non-zero acceleration, and as soon as they lift off they move under the gravitational field of the asteroid, the non-inertial apparent forces due to rotation, the Sun's gravity and its radiation, and –in the case of binary systems– the gravitational pull of the secondary. Other forces may act as well on small particles on the surface, like electrostatic or molecular forces (cohesion), with the likely result of sticking them together and potentially undergoing the same dynamical effect as dusty clumps. Moreover, small particles, below 1 mm in size, may be lost from the system under the influence of solar radiation pressure (SRP), but even mm to cm–size particles can have their orbits affected by SRP over a longer span <cit.>. We may expect, instead, that more massive particles potentially levitate for some time, land on the surface and lift off again, repeating such cycles over and over, or just land at latitudes from which further lift off is not possible. Alternatively, they may enter mostly unstable orbits and even transfer to the secondary. <cit.> studied the effect of particle lofting due to YORP spin up in a binary system, namely 1999 KW4 (Moshup). They found that transferring angular momentum from the primary to the mutual orbit is possible. Regulation of primary spin at the rate for which material lofting takes place may happen so that the orbital angular momentum of the secondary grows steadily. Apparently, lofting occurs in fast transient episodes separated by long periods of slow spin-up. The authors argue that the end state of the system evolution is likely the separation into two asteroids on closely related heliocentric orbits. That may be a potential origin for “asteroid pairs”. <cit.> outlined the possibility of mass lifting as a general mechanism of regolith dispersal in fast spinning Near-Earth asteroids forming “dusty” environment around such bodies. <cit.> investigated mass shedding from the surface of the primary of (65803) Didymos by a semi-analytical approach for shedding conditions. They determined unstable surface areas by combining the analyses and numerical results of SSDEM simulations. The authors find a vast majority of the shedding mass is finally transferred to Dimorphos and leads to a cumulative growth which may cause a spiralling–in orbit of the secondary, an effect going in the opposite direction to <cit.> finding. Further work on failure modes and mass shedding processes was carried out by <cit.>. The main goal of this work is to study the general dynamical features of this mechanism that may be acting on some of the NEA binary systems listed in Table 1. We use asteroid Didymos as a case study, assuming the best available knowledge to the time of developing the model. Obviously, the actual shape, volume and mass of that asteroid –as well as other physical parameters of the system– will be constrained in much greater detail only after both the DART and Hera missions will characterize the binary system. Here we analyze the dynamical evolution of lifted particles as well as their preference for take–off and landing areas on Didymos. We provide mass density distributions of the material that may currently be, or may have been present around it, assuming a given emission rate. The model set up to study particle dynamics in this system and its validation is introduced in Sec. 2; results are presented in Sec. 3 and 4, and conclusions are discussed in Sec. 5. § MODEL DESCRIPTION §.§ Mass loss due to fast spin rate To study under what conditions particles may leave the Didymos surface, we considered the available radar-based Didymos shape model, made of 1000 vertices and 1996 facets <cit.> (Future, updated work on this matter shall include the shape model of Didymos available after the DART mission). Test particles were initially assumed to be at rest at the geometric centre of each triangular facet. Particles size distribution follow a differential power law n(r) dn ∝ r^κ dr, with index κ = -3.5. We sampled the following particle size range values: r_1 = 4.7 μm, r_2=0.1 mm, r_3 = 2.3 mm and r_4=5.3 cm. These are central values of the corresponding logarithmic size bins in which the overall size distribution is divided. In this way, we check particle behaviour from micron–size, very sensitive to solar radiation pressure, to cm–size —which are typical values for dust grains ejected from active asteroids <cit.>— to multi–cm size, affected only by gravity. Particle density is assumed to be 3500 kg/m^3, according to the L and LL meteorite analogue to the Didymos S spectral type <cit.>. In the case local acceleration is directed outwards, a particle will take off from the surface and will evolve under the gravitational field of Didymos, the gravitational perturbations generated by the secondary and the Sun, and the solar radiation pressure (SRP), according to the corresponding equation of motion. Under this scheme, at the end of a given integration time, we calculate the total mass of ejected particles in each end state. Based on the trajectories of particles, we have defined four possible end states (ES): ES1, particles that lift off and land again on Didymos’ surface; ES2, particles that remain in orbit; ES3, particles that are accreted onto the secondary; and ES4, particles that escape from the system (see Section <ref> for details). We also estimate the mass density in the Didymos system environment as a function of colatitude, longitude, and distance from the surface of the primary. §.§ Equation of motion The equation of motion of a particle in a spinning reference system fixed to the primary body can be written as: d^2 𝐫_𝐝/dt^2 = - ∇𝐔_𝐏 + W_2𝐫_𝐝-𝐫_𝐏 ⊙/||𝐫_𝐝-𝐫_𝐏 ⊙||^3 + W_3 [ 𝐫_𝐏 ⊙-𝐫_𝐝/||𝐫_𝐏 ⊙-𝐫_𝐝||^3 - 𝐫_𝐏 ⊙/r_P ⊙ ^3] + +W_4 [ 𝐫_𝐝𝐒/r^3_dS-𝐫_𝐏𝐒/r^3_PS] + ω× (𝐫_𝐝×ω) + 2 𝐯_𝐝×ω In that expression, bold characters are vectors, r_d is the primary to particle position vector, r_P ⊙ is the primary to the Sun position vector, r_dS is the position vector from the particle to the secondary, and r_PS is the primary to the secondary position vector, as shown in Figure <ref>. Here r_P ⊙ = r_d + r_d ⊙, where r_d ⊙ is the vector from the dust grain to the Sun. The first term of Eq. <ref> corresponds to the gravitational field per unit mass of the polyhedral shape of the primary, corresponding to the definition in <cit.>; the second term is the SRP contribution, where W_2 = (Q_pr/c)[E_⊙/(4 π)][π d^2/(4m_d)], the following terms are the solar and the secondary gravitational perturbations: W_3 = GM_⊙ and W_4= GM_S; the last two terms are the centripetal and Coriolis force, respectively. The Euler force can be neglected in this scheme because the spin change is an extremely slow process, lasting orders of magnitude longer than any other change in particle motion. Moreover, G is the gravitational constant, M_P is the primary mass, M_S is the secondary mass, M_⊙ is the mass of the Sun, Q_pr is the efficiency of solar radiation pressure, which is Q_pr∼ 1 for large absorbing grains <cit.>, c is the speed of light, E_⊙ = 3.93 × 10^26 W is the total power radiated by the Sun, d is the particle diameter and m_d its mass: m_d=ρ_d (π/6)d^3 , where ρ_d is particle density. ω is the angular velocity of the primary (ω=2π /T, where T is the rotational period) and v_d is the velocity of the particle. §.§ Description of the algorithm Eq. <ref> is integrated numerically using a fourth-order Runge-Kutta method. The primary asteroid can be represented as a sphere or any polyhedral shape. In the first case, the surface of the primary is divided into a grid with N_θ colatitude (θ) bins and N_ϕ longitude (ϕ) bins; this grid is extended radially outwards into space with N_r radii (r) bins, forming a total of N_θ× N_ϕ× N_r cells. At the beginning, each sample particle is located at the geometrical centre of each surface cell. In the second case, each sample particle is located at the geometrical centre of each triangular facet of the polyhedron, and the N_θ× N_ϕ× N_r space grid is defined outside the body, from the point on the surface farthest from the centre of the body, r_F. The mass and particle density are computed in those cells. To check the validity of the calculation, the primary asteroid was also modelled as a sphere, and the gravitational field of the corresponding homogeneous sphere was analytically calculated. In the general case, the gravitational field corresponding to the polyhedral model is computed following the guidelines given by <cit.> at each point in space. In both cases, a point-mass secondary body is assumed. Dimorphos follows a circular orbit on the equatorial plane of Didymos. Moreover, the position and orientation of the binary in the heliocentric ecliptic system are rigorously computed at each time step from its available ephemeris and system pole orientation. The position of the Sun, as seen from the primary, is also computed, and the shadow of the primary on the ejected particles is checked at each time step to set the SRP on each particle either on or off. The shadow produced by Dimorphos on Didymos is not taken into account, as it is less than 5% of the Didymos surface, smaller than other sources of uncertainty in the model. Particles are initially at rest on the surface. A detachment condition is applied to check whether any given particle has an outwards component of local acceleration, which depends strongly on the apparent centrifugal force in the rotating system of the primary. Detachment occurs when ∑_iF_i·n > 0, where F_i stands for each force corresponding to the accelerations in Eq. <ref>, and n is the unit vector normal to the surface facet and directed outwards. Particles, once ejected from the surface of the primary, move in the gravitational field of the two bodies plus the radiation pressure and gravitational perturbing forces of the Sun. Based on the trajectories of the particles, we define four final states, as follows: * ES1, landing particles: particles with radial distance from the centre is equal to or less than r_F, the particle is checked to be outside or inside the surface defined by the shape model. In the first case, the integration process continues, otherwise, the algorithm finds the intersection between the trajectory of the particle and the facet of the shape model. The collision coordinates are recorded, and the particle is labelled as ES1. * ES2, orbiting particles: particles belong to this group if at the end of the integration time they are still in orbit. * ES3, particles accreted on Dimorphos: during numerical integration, the position of a given particle with respect to Dimorphos is checked at every time step. The gravitational field of Dimorphos is considered as a point–mass source. The case in which the distance between the particle and the position of Dimorphos is less than its equivalent radius, is considered as a collision, and the particle is labelled as ES3. * ES4, escaping particles: particles located beyond 10^4 m from the centre of mass of Didymos belong to this group. Even if the Hill's radius of the system is 75 km, the limit is set at 10 km distance from the centre of the system for practical reasons. This is a distance beyond which we found that only a negligible mass density contribution is missed from rare particles orbiting back from outer distance. In fact, mass density, even at the Dimorphos distance, is extremely small, and the density profile keeps decreasing further away. This has the benefit of hugely reducing both the storage load and the computational cost with no effect on the mass density calculation. Figure <ref> shows the surface gravity map of the Didymos shape model of a non-rotating Didymos, built using a mass of 5.229 × 10^11 kg with homogeneous density. This is worked out considering that the overall size of Dimorphos, D_S≈ 0.2 D_P, so that the relationship between masses is close to M_S/M_P∼ 0.01. Lowest gravity regions are clearly located on the equatorial bulge (blue-green colour in Figure <ref>). We assessed the difference between the gravity field of Didymos, taking into account its shape model, against a sphere of equal mass and density, by calculating the radial gravitational field. Obviously, differences are found close to the surface (until approximately 570 m), but both fields converge at larger distances, as expected. In our model, the difference between the spherical and the polyhedral shape of Didymos is in the definition of spatial cells. Given that the shape model represents an irregular surface, it is not possible to simply define a surface grid and extend it into space in a uniform way. Instead, we set the longitude and latitude grid starting at the vertex of the triangular facet at the maximum distance to the centre (r_F). For the Didymos shape model, this spherical inner grid surface starts at r_F=427 m, which is the smallest distance at which mass density outside the body is calculated. This is the minimum radial distance from which spatial density computation is performed. In other words, this is the spherical surface from which the computation of the mass density is made outside Didymos. The drawback of this approach is that no information is available inside the tiny region in between the shape model surface and the beginning of the 3D grid. As for the modelling of the shadow cone produced by Didymos, given the moderate depart of the asteroid shape from sphericity, for the sake of simplicity a profile corresponding to a sphere of radius r_F is assumed instead of the shape model itself. That approach has the advantage to save computing time, with a negligible impact on the volume of the projected shadow cone. In addition, in order to save CPU time, the gravity field of the primary is approximated by its corresponding spherical field at r > 550 m, given the proximity of spherical and polyhedral solutions at that distance. The relative error is less than 3% at that point, decreasing at higher values of r. The code input data are the orbital and physical parameters of the system extracted from the mentioned Hera Didymos Reference Model (ESA internal document). Further inputs are the perihelion epoch, the dates of start and end of integration, the integration time step used in the Runge-Kutta procedure, the mass loss rate, and particles properties, i.e., particles radii, their density and the exponent of their size frequency distribution (SFD). Other running parameters are also given: the number of particle size bins, which are conveniently spaced logarithmically to be consistent with the power-law distribution, the time bins, the radius limit for mass density computation, and the escaping distance, i.e., the distance at which particles are considered to escape the system and no longer contribute to density in the considered space range. See Sec. <ref> and <ref> for details. §.§.§ Time integration procedure When we look at the image of a comet tail, we need to keep in mind that particles were ejected from the comet surface at any time before the time at which the image was taken, let's call that time t ≤ T_obs. A similar situation can be envisioned in the case of Didymos ejected particles. The following numerical procedure for reproducing such a situation is applied. Let's call T_start the time at which the program checks, for the first time, the particle surface detachment condition at any given surface facet. The final T_obs time is instead the time at which the calculation of the particle mass and number density is made (in the analogy above, this corresponds to the time at which the observation of the system environment is made). The total integration time is T_obs-T_start, which is divided by the number of time intervals n_t, so that Δ t=(T_obs-T_start)/n_t is the length of each time integration period. One sample particle of each size is located on every surface facet. At the beginning (T_start), the detachment condition for each particle size is checked, if this is met, the sample particle starts its motion under the action of the corresponding forces by a 4th–order R-K integrator, with an integration time step H. This is done for all particle sizes. Many different values for the R-K integration time step were tried, the value finally adopted corresponds to the largest one safely allowing for convergence in the integration. Once integration is over, that is, at t=T_obs, the same procedure is repeated at the initial time T_start + Δ t (Δ t>>H), and so on. This is necessary because we need to consider the evolution of particles potentially leaving their facet at all n_t equally spaced times between T_start and T_obs. The same process is carried out for all surface facets. In this way, the emission of sample particles is checked along the whole integration time. At time T_obs, the number of particles in orbit (ES2) is counted at each space cell. Instead, particles that hit the surface of the primary or the secondary (ES1, ES3), or reach the escape distance (ES4) at some time, do not contribute to space density (ES2), but their end state is recorded. At the end of the whole integration time for all surface facets, the data corresponding to the end–state of all sample particles are stored, as well as their position in space and mass density as a function of longitude, colatitude and radial distance from the centre of Didymos. §.§.§ Mass density calculation Once the end state of sample particles is known, it is necessary to calculate the absolute mass in each end state. Therefore, we face the problem of how to translate sample particle statistics into the corresponding actual mass in each end state. First of all, we need to make assumptions on the mass emission rate; in addition, the mass and number of particles have to be scaled with the adopted SFD. Active asteroids show a diversity of mass–loss mechanisms including sublimation, impacts, fast rotation, electrostatics, thermal effects, etc. Estimates of rotation mass shedding in fast spinning rubble-pile asteroids range from 10^-7 kg/s, reported for asteroid (101955) Bennu <cit.>, to 2.3 kg/s <cit.>, and 35 kg/s <cit.> in the case of (6478) Gault. Episodic mass losses ranging from 1 to 10 kg/s have been derived for multi–tailed asteroid P/2013 P5 (PANSTARRS) <cit.>, which can be clearly attributed to mass shedding due to rotational instability <cit.>. To date, no observation of this kind of activity has been reported for Didymos, so this parameter is unknown. For this reason, we considered a constant reference value of mass production rate for Didymos, that was set arbitrarily to dM/dt = 1 kg/s for the whole asteroid. A suitable fraction of such value is used only for cells for which the lift off condition is met, no particle will be ejected for most surface cells far from the equator though. The contribution of ejected particles to the mass and number density in each space cell is computed as follows. The mass ejected per unit surface area is calculated as (dM/dt)/(4π R_p^2), where dM/dt is the assumed mass production rate on the whole surface, R_p is the Didymos radius, and the mass ejected per surface cell in a time interval Δ t^' is: M_cell=1/4π R_p^2dM/dt S_cellΔ t^' where S_cell is the area of the surface cell. The total mass is distributed according to the already mentioned SFD in the radius range [r_min, r_max]. In a given size bin [r_i, r_j] within such range, the number of particles emitted from each cell in a given time interval is: N_bin= M_cell3/41/πρ_particle∫_r_i^r_j r^κ dr/∫_r_min^r_max r^κ+3 dr and the corresponding emitted mass is: M_bin=N_bin4/3πρ_particle∫_r_i^r_j r^κ+3 dr/∫_r_i^r_j r^κ dr The mass and the number of particles are updated accordingly in every space cell where sample particles are located at the end of integration. The mass and number of particles are finally divided by the cell volume to get the corresponding mass and number density. §.§ Model validation In order to check the validity of our model when applied to the polyhedral representation of solid bodies, we compared the results obtained for the dynamical evolution of particles departing from analytical and polyhedral representations of a spherical body made of 5120 facets (see Figure <ref>). In both cases, the same mass and equivalent size of Didymos were used, M_P=5.229×10^11 kg and R=400 m, respectively. The density profiles are in very good agreement even at small distances from the body surface. Simulations were run over 30 days. § SUMMARY OF MODEL SIMULATIONS The current Didymos system physical parameters are known with wide uncertainties (see Sec. 1). For this reason, we first mapped the mass–volume parameter space and related each pair of values with the corresponding total orbiting mass obtained by our model. Then, we focused on the nominal parameters for the Didymos mass and size and we performed a detailed analysis of particle detachment and landing process by means of numerical simulations. The wide heliocentric distance of the Didymos system —due to its high eccentricity— led us to study the particle behaviour around very different locations (perihelion and aphelion) of the system, and under the corresponding initial epoch conditions, in the case of full orbit integrations. The whole set of simulations carried out is outlined next. * Simulation 1: Conditions for mass lift-off. The first goal is to study under what conditions it is possible to find mass around the primary. Thus, we combined 7 values of each mass and volume parameters of Didymos within estimated uncertainty. That resulted in 49 numerical runs, one per each corresponding value of bulk density (see Table <ref>). We let the system evolve during 30 days, near the perihelion epoch, starting on August 28, 2020 and ending on September 26, 2020. Performing simulation runs around perihelion is the worst case for particle survival, as SRP is more efficient in perturbing their motion and taking particles away from the system. * Simulation 2: Evolution in one full orbit (starting at perihelion). Here we used the nominal mass and volume of Didymos in order to set its nominal bulk density. The simulation was run over a full heliocentric orbit of Didymos (770 days), starting and ending around the perihelion epoch (from August 28, 2020 to October 7, 2022). The behaviour of small (r_1, r_2 and r_3) and large (r_4) particles were analyzed separately to study the effect of the SRP. * Simulation 3: Evolution in one full orbit (starting at aphelion). We used the same physical parameters as in Simulation 2, but the simulation was run over a full heliocentric orbit of Didymos (770 days), starting and ending around the aphelion epoch (from September 15, 2021 to October 25, 2023). Again, the analysis of small and large particles was done separately. * Simulation 4: Short term evolution around perihelion. Here we report the results of Simulation 1 again for the nominal mass and volume of Didymos primary. The system evolved over 30 days around the perihelion epoch, starting on August 28, 2020 and ending on September 26, 2020. * Simulation 5: Short term evolution around aphelion. Analogous to Simulation 4, but over 30 days near the aphelion epoch, starting on September 15, 2021 and ending on October 15, 2021. We run short term (30 days) simulations to be able to catch the main features of the orbiting mass for the whole set of 49 runs in Simulation 1 and to underline (Simulations 4 and 5) the dependence of the outcome for small particles on the observation epoch. This choice, in the case of Simulation 1, was motivated, on one hand, by the fact that full orbit simulations are very time consuming. On the other hand, the behaviour of particles —as far as mass distribution is concerned— does not change around a given epoch, beyond some 20 days. Therefore, we considered a safe strategy to perform 30–day runs around the least favorable epoch for particle survival. Orbital parameters used in simulations are shown in Table 2. Table 3 summarizes the volume and mass parameter space ranges, and the system physical parameters are introduced in Table 4. § RESULTS §.§ Conditions for mass lift-off (Simulation 1) In Figure <ref>, each square gives the orbiting mass at the end of each simulation after 30 days around perihelion, normalized to the maximum orbiting mass value. The central cross corresponds to the nominal case, which will be analyzed further in the next sections (Simulations 2 to 5). Particle detachment takes place at the nominal spin rate for most values in the chosen Didymos physical parameter space, except for three parameter combinations (see Table <ref>). §.§ Particle evolution after one full orbit (Simulations 2 and 3) In Simulation 2 and 3, we analyze particle end states after they leave the surface and the corresponding lifetimes (Table <ref>). Depending on size, more than 97 % of particles fall back onto the surface of the primary (ES1). The time required to land back also depends on particle size. In general, for any ES, the larger the particle, the longer its lifetime. Particles of radius r_1 spend on average 0.81 h orbiting the asteroid before landing on its surface, while particles r_4 may orbit more than 4 hours before landing. In addition, as expected, most particles that escape the system (ES4) correspond to size r_1. However, that end state probability is only 2.81%. When such particles pass from the shadow cone generated by the asteroid to the illuminated region, they are quickly removed from the system, typically after 6 hours. On the contrary, a small fraction of r_2, r_3 and r_4 size particles are eliminated, and they are able to survive in the environment of the asteroid between 37 and 600 hours, before they are removed. Instead, particles that are accreted onto the secondary (ES3) are mostly r_4, followed by r_3. Such particles stay in orbit for some 100 and 81 hours, respectively, before reaching Dimorphos. r_1 and r_2 particles have lower probabilities of reaching the moonlet, taking 2.5 to 4 hours, driven by SRP. At the end of the integration time, there are mostly large particles orbiting the system. r_3 size particles are able to stay in orbit about 46 h in Simulation 2 and 74 hours in Simulation 3, and this difference is related to still some weak interaction with SRP at such size range. On the other hand, r_4 particles have a median lifetime in orbit of ∼ 100 hours and no influence of the SRP is detectable any longer. These results can be easily understood by comparing the effect of the gravitational force (F_g) due to Didymos on any given particle, with respect to the corresponding solar radiation force (F_rad), as a function of the distance to the centre of the asteroid (a): F_g/F_rad=4/3G M_P c ρ_d d/2/Q_pr F_⊙1/a^2 F_⊙ = E_⊙ / 4 π a^2 is the solar flux at the heliocentric distance of the asteroid system, and the other parameters are the same as in Section <ref>. The ratio between the two force contributions is plotted in Figure <ref>. A horizontal gray line separates the distance to Didymos into two regimes, one is dominated by gravity and the other one by solar radiation force. The gray vertical line shows the distance to Dimorphos and the coloured curves represent the relationship between the two mentioned forces for different particle sizes. Solid and dashed curves correspond to motion around perihelion and aphelion, respectively. Particles of size r_1 evolve almost completely under the action of SRP, whereas particles of size r_2 are affected by both SRP and F_g. Instead, motion of particles of size r_3 and r_4 are dominated by the force of gravity. That implies that there is no smooth transition between the behaviour of the particles of size r_1 and r_3 and the corresponding mean lifetimes in ES2. Lifetimes are shorter around perihelion (Simulation 2) than around aphelion (Simulation 3) due to different SRP force values. These results are comparable with the work by <cit.>. They also show that, in such low-gravity environment, SRP plays an important role in the dynamics of small dust grains. Instead, particles larger than a few millimetres are mostly affected by the Didymos gravitational force within the orbit of Dimorphos. §.§ Density distribution of orbiting mass §.§.§ Small particles Here, we focus on the analysis of the behaviour of small particles (r_1, r_2 and r_3) that are in ES2 at the end of Simulations 2, 3, 4 and 5. Mass density profiles as a function of radial distance from the asteroid centre are shown in Figure <ref> (top). This is obtained by integrating on each cell in colatitude and longitude at any given radial distance. Plots begin at r_F=427 m, the origin of spatial cells. In density profiles, maxima are found at r∼475 m in the case of Simulations 2 and 4, at r∼481 m in Simulation 3, and at r∼458 m in Simulation 5. Therefore, the largest mass density of orbiting particles is 31 to 54 meters above the mentioned reference distance from the centre r_F. Beyond such distance, mass density steadily decreases, as expected. As a result, orbiting mass density in simulation runs around perihelion (Simulation 4) is smaller than for the corresponding simulations around aphelion (Simulation 5). This is due to the strong influence of SRP on small particles. The outcome of density profiles for 30 days integration time are very similar to those obtained integrating over a full orbit of the system around the Sun. This is the case both for perihelion and aphelion epoch starting time. Therefore, the amount of small particles in the system environment at a given time strongly depends on the epoch at which the mass density observation is made. The colatitude density profile (See Figure <ref>, bottom), together with the radial density profile, show that orbiting particles are mostly located in a thin disk, with maximum density on the equatorial plane. However, the disk is not symmetrical at north and south of the equator due to Didymos topographical shape model inhomogeneities. On the northern hemisphere, landing location is spread over 25 degrees, between 65^∘ and 90^∘ of colatitude, while, on the southern hemisphere, the distribution extends up to 30 degrees, between 90^∘ and 120^∘ colatitude. The density mass difference between observation about perihelion and aphelion is also shown in the colatitude profile (Figure <ref>, bottom). The analysis of particle motion and mass density shows that although most ejected particles return to the surface of the primary with short orbit lifetimes (a few hours), a continuous lift off process supplies them into space. §.§.§ Large particles This section extends the analysis of results including particles of radius r_4, not affected by SRP. Figure <ref> shows the corresponding radial and colatitude mass distributions. This also applies to any object of size larger than r_4, as they are only affected by gravitational and non–inertial force terms. When large particles are considered, there is no substantial difference in orbital mass density at aphelion with respect to perihelion, although the radial density profile at aphelion (Simulation 5) is slightly larger than the corresponding density at perihelion (Simulation 4) over the whole considered domain. Large particles are the greatest contributors to orbiting mass density, which is equal around both epochs. Instead, small particles contribute more at aphelion than perihelion, building up the very small difference between densities observed around the two different epochs (see Figure <ref>). The disk–like distribution of particles in orbit at a given time is shown in Figure <ref>, in the case of Simulation 4. It is interesting to notice that the probability (p_2(r)) for small particles to have an orbital end state (ES2) is smaller as size decreases (see Table 5). Similarly, duration of orbits is shorter for small particles with respect to large ones (δ t_2(r), normalized). Both effects combine in the size frequency distribution of particles found to be in orbit at a given time, that can be described as n_o(r)dn= n (r)dn p_2(r) δ t_2(r) dr∝ r^-3.5 p_2(r) δ t_2(r) dr ∝ r^κ_o dr, with κ_o≈ -3. In summary, the SFD of particles observed in orbit (ES2) looks to be shallower than the original asteroid surface population assumption (κ= -3.5), and is therefore skewed towards large particles. §.§ Distribution of landing particles Figure <ref> shows —for Simulation 5, as an example— a mass density colour map distribution of the initial position of particles that are able to lift off the surface. At the beginning of any run, particles lie at the geometrical centre of the shape model triangular facets; hence a discrete distribution shows up in Figure <ref>. The detachment area surrounds the equatorial plane, but it is not symmetrical with respect to it. The maximum detachment latitudes to the north and to the south of the equator are θ_N ∼ 28 ^∘ and θ_S ∼ 19 ^∘, respectively, in all the cases, in Simulations 2 to 5. After detachment, most lifted particles (more than 97%) land back on the surface of the primary. Figure <ref> (bottom) shows landing locations after 30 days of evolution. During this time, the gravitational field of Didymos, the gravitational perturbation of Sun and Dimorphos, and SRP, modify particle trajectories, increasing their orbital inclination and allowing them to reach mid and high landing latitudes, typically higher in comparison with the detachment positions. The colour map shows that the density distribution peak in colatitude lies near the equatorial plane. It is important to note that particles landing at low latitudes can likely leave the surface again. Indeed, such particles are a reservoir for later material ejection. Instead, particles falling at mid and high latitudes would not be dynamically able to lift off again from their new location. However, mass sliding mechanism towards low latitudes are known to have taken place on asteroids Bennu and Ryugu <cit.>, such effect may drive mass back to the equatorial region, re–fuelling the take off and landing process. Figure <ref> (top) shows non-homogeneous landing distribution in longitude. This behaviour is related to the topography of the asteroid in the equatorial region. Particles flight heights are low, so lifted-particles preferentially land on relatively high terrains around the equator. § DISCUSSION AND CONCLUSIONS Primary fast spin up in NEA binary systems is relatively common. We study the possibility that mass is able to take off the surface of such bodies, taking (65803) Didymos NEA binary system as a study case because it is the DART (NASA) and Hera (ESA) space missions target. The latest available shape model for the asteroid at the time of carrying on this study was used. We set up a model to study the dynamical behaviour of the particles that are able to take off the surface due to the primary fast spin. The gravitational field of the asteroid is taken into account, together with all non–inertial forces in the asteroid rotating frame. Perturbations due to the satellite, Dimorphos, and to the Sun are suitably considered, as well as SRP. We explored the Didymos primary mass and volume parameter space to assess under what asteroid density conditions particle detachment from the surface is possible. We find that most values in that parameter space allow for particle lift off. In the nominal case, that is possible in a region limited to the (+28^∘, -19^∘) latitude interval around the equatorial plane, non–symmetric due to the primary irregular shape. Available range of mass and volume estimates is compatible with the possibility of having a disc of particles of different sizes around Didymos. The actual mass density is hard to predict as the rate of mass emission is unknown. Nevertheless, our model finds the distributions in colatitude and distance from the asteroid according to assumed parameters. Maximum values for the mass density distribution show up at very low latitudes and at some 30 to 50 m away from the average primary equatorial radius, decreasing steadily at larger distance. Most lifted particles (>97%) do land back on the surface, with variable lifetimes depending on size, but typically smaller than 5 h. Small particles have shorter lifetimes than large ones because of SRP. In fact, when small particles take off on the day side of the asteroid, they are quickly pushed back to the surface, and their flight time is shorter the stronger the SRP is. When such particles take off in the night side, they follow ballistic trajectories until they typically land back or, in the case they reach the terminator, they are pushed away from the system by SRP. Particle landing locations are not evenly distributed in longitude, concentrating instead in equatorial morphology highs. This is because most particles only reach low altitudes with respect to their original take off location, and they eventually stumble into such highs on their flight path. It is interesting to notice that the latitude landing range distribution is wider than that of lift off distribution location. As a consequence, some particles may land at latitudes from which they can no longer take off. SRP has an important effect on small particle dynamics, to the point that, for such particles, mass density in space happens to be a function of the orbital phase at which it is observed (T_obs). Didymos has a very eccentric orbit, such that heliocentric distance changes from 1.013 to 2.276 au. This results in very different contributions from the solar tide and SRP at different distance from the Sun. Total mass surrounding Didymos at aphelion is more than double (≈ 9/4) than at perihelion for sub–cm size particles. Instead, looking at the overall mass in orbit, including large particles —not affected by SRP— no significant difference is found, due to the contribution of large particles to total mass. The size distribution of particles in orbit should be shallower (κ≈ -3) than their size distribution on the asteroid (κ=-3.5) due to smaller probability and lifetime in orbit of small particles with respect to large ones. In conclusion, our model shows that the presence of mass in orbit around Didymos, or similarly shaped fast spinning NEAs, is possible. That includes the potential presence of boulders around the system, that may eventually reach the secondary (≈ 2% probability) at this stage. In any case, as outlined in Section <ref>, the existence and observability of mass around the system may depend on which primary rotational phase and which secondary orbital phase Didymos is undergoing at the observation time, according to the mechanism described by <cit.>. Right after the completion of this study, the DART spacecraft successfully impacted Dimorphos, and first results derived from the images of both the DRACO camera on board DART and the LEIA and LUKE cameras on board the CubeSat LICIACube (ASI: Agenzia Spaziale Italiana) have been published <cit.>. <cit.> report a set of freshly estimated values for the critical parameters in this study: the mass of the system, (5.6± 0.5)× 10^11 kg; the equatorial extents, 850± 5 m on average; the equivalent diameter of both Didymos, 761± 26 m, and Dimorphos, 151± 5 m (uncertainties are given at 1-σ level). Therefeore, the mass of Didymos can also be easily estimated, assuming equal density for the two bodies. We are conscious that the Didymos shape model is preliminary, however, such set of parameters tends to confirm the possibility of mass lift off from the equatorial region of Didymos. In fact, even in the worst case compatible with available uncertainty ranges (largest system mass, smallest Dimorphos equivalent diameter, largest Didymos equivalent diameter, and smallest Didymos equatorial extent), the ratio between the centripetal and gravitational acceleration at the equator, a_c/g=ω^2r_eq^3/(GM_P) is larger than 1 (1.07), indicating outwards acceleration. a_c/g is the driving magnitude here (rather than density), where ω is the spin rate and r_eq is the equatorial radius of Didymos. The nominal values from <cit.> give a_c/g=1.235, which is very close to the (M_1, V_4) set, providing more mass in orbit than our nominal case. a_c/g=1 for r_eq^'=396 m, which is the minimum distance from the spin axis at which lift off is possible, some 30 m smaller than the estimated equatorial radius extent. It is interesting to notice that the estimated bulk density is reported as 2400± 300 kg/m^3, a value that would inhibit mass shedding on a sphere spinning at the Didymos spin rate. The bulk density of a body is the critical parameter only in the case in which its shape is fixed. Comparing a sphere with an equal mass oblate spheroid with equatorial axis a and shortest axis c, the corresponding critical spin rate is ω_cr^'=√(4 π G ρ^'/3c/a) At equal critical spin rate, the relationship between the density of the oblate shaped and the spherical body is ρ^'/ρ=a/c. E.g., for an oblate body with a/c=1.37 (the case of Didymos, as its extent along spin axis is estimated to be 620± 15 m by <cit.>), the critical density is ρ^'=2924 kg/m^3, while, for an equal mass spherical Didymos, it would be ρ=2134 kg/m^3. Indications of the presence of dust around Didymos can be envisioned by looking at the LICIACube images. However, we prefer to avoid speculation at this point, and we delay further interpretation to a forthcoming companion paper. Once further estimate improvement of the main physical parameters and an updated Didymos shape model –pre-Hera mission– will be available, a refined model will be carried out on the “real” Didymos system. We also plan to study the interplay between mass ejected from Dimorphos by the DART impact and the stationary mass distribution around the system –if any. In fact, that may potentially affect the actual mass density to be estimated by the Hera rendezvous with the Didymos system in 2027 and the calibration of instrumentation on board. § ACKNOWLEDGEMENTS NT, ACB and PBL acknowledge funding by the NEO-MAPP project through grant agreement 870377, in the frame of the EC H2020-SPACE-2019. NT acknowledges funding by CONICET. ACB and PBL acknowledge funding by the Ministerio de Ciencia Innovación (PGC 2018) RTI2018-099464-B-I00. FM acknowledges financial supports from grants MCIN/AEI/10.13039/ 501100011033 (GRANT PID2021-123370OB-I00), European Union NextGeneratioEU/PTR, P18-RT-1854 from Junta de Andalucia, and CEX2021-001131-S funded by MCIN/AEI/10.13039/501100011033. 99 elsarticle-harv [Burns et al.(1979)]Burns1979 Burns, J. A., Lamy, P. L., Soter, S. 1979. Radiation forces on small particles in the solar system. Icarus 40, 1–48. doi:10.1016/0019-1035(79)90050-2 [Campo Bagatin(2013)]CampoB2013 Campo Bagatin, A. 2013. Small Asteroids with “Dusty” Atmospheres? 44th Annual Lunar and Planetary Science Conference. [Chandrasekhar (1969)]Chandra1969 Chandrasekhar, S. 1979. Ellipsoidal Figures of Equilibrium. New Haven, Yale University Press. [Cheng et al.(2021)]Cheng2021 Cheng, B. and 7 colleagues 2021. Reconstructing the formation history of top-shaped asteroids from the surface boulder distribution. Nature Astronomy 5, 134–138. doi:10.1038/s41550-020-01226-7 [Daly et al.(2023)]Daly2023 Daly, R. T. and colleagues 2023. Successful Kinetic Impact into an Asteroid for Planetary Defense. Nature. https://doi.org/10.1038/s41586-023-05810-5 [Dunn et al.(2013)]Dunn2013 Dunn, T. L., Burbine, T. H., Bottke, W. F., Clark, J. P. 2013. Mineralogies and source regions of near-Earth asteroids. Icarus 222, 273–282. doi:10.1016/j.icarus.2012.11.007 [Fahnestock and Scheeres(2009)]Fahnestock2009 Fahnestock, E. G., Scheeres, D. J. 2009. Binary asteroid orbit expansion due to continued YORP spin-up of the primary and primary surface particle motion. Icarus 201, 135–152. doi:10.1016/j.icarus.2008.12.021 [Ferrari et al.(2020)]Ferrari2020 Ferrari, F., Lavagna, M., Blazquez, E. 2020. A parallel-GPU code for asteroid aggregation problems with angular particles. Monthly Notices of the Royal Astronomical Society 492, 749–761. doi:10.1093/mnras/stz3458 [Ferrari and Tanga(2022)]Ferrari2022 Ferrari, F., Tanga, P. 2022. Interior of top-shaped asteroids with cohesionless surface. Icarus 378. doi:10.1016/j.icarus.2022.114914 [Hergenrother et al.(2019)]Hergenrother2019 Hergenrother, C. W. and 24 colleagues 2019. The operational environment and rotational acceleration of asteroid (101955) Bennu from OSIRIS-REx observations. Nature Communications 10. doi:10.1038/s41467-019-09213-x [Hirabayashi et al.(2022)]Hirabayashi2022 Hirabayashi, M. and 13 colleagues 2022. Double Asteroid Redirection Test (DART): Structural and Dynamic Interactions between Asteroidal Elements of Binary Asteroid (65803) Didymos. The Planetary Science Journal 3. doi:10.3847/PSJ/ac6eff [Holsapple(2007)]Holsapple2007 Holsapple, K. A. 2007. Spin limits of Solar System bodies: From the small fast–rotators to 2003 EL61. Icarus 187, 500–509. doi:10.1016/j.icarus.2006.08.012 [Holsapple(2022)]Holsapple2022 Holsapple, K. A. 2022. Main belt asteroid collision histories: Cratering, ejecta, erosion, catastrophic dispersions, spins, binaries, tops, and wobblers. Planetary and Space Science 219. doi:10.1016/j.pss.2022.105529 [Housen(2009)]Housen2009 Housen, K. 2009. Cumulative damage in strength-dominated collisions of rocky asteroids: Rubble piles and brick piles. Planetary and Space Science 57, 142–153. doi:10.1016/j.pss.2008.07.006 [Hui et al.(2019)]Hui2019 Hui, M.-T., Kim, Y., Gao, X. 2019. New active asteroid (6478) Gault. Monthly Notices of the Royal Astronomical Society 488, L143–L148. doi:10.1093/mnrasl/slz112 [Jewitt et al.(2015)]Jewitt2015b Jewitt, D., Agarwal, J., Weaver, H., Mutchler, M., Larson, S. 2015. Episodic Ejection from Active Asteroid 311P/PANSTARRS. The Astrophysical Journal 798. doi:10.1088/0004-637X/798/2/109 [Jewitt et al.(2019)]Jewitt2019a Jewitt, D. and 6 colleagues 2019. Episodically Active Asteroid 6478 Gault. The Astrophysical Journal 876. doi:10.3847/2041-8213/ab1be8 [Jewitt and Hsieh(2022)]Jewitt2022 Jewitt, D., Hsieh, H. H. 2022. The Asteroid-Comet Continuum. arXiv e-prints. doi:10.48550/arXiv.2203.01397 [Lauretta et al.(2019a)]Lauretta2019a Lauretta, D. S. and 29 colleagues 2019. The unexpected surface of asteroid (101955) Bennu. Nature 568, 55–60. doi:10.1038/s41586-019-1033-6 [Lauretta et al.(2019b)]Lauretta2019b Lauretta, D. S. and 57 colleagues 2019. Episodes of particle ejection from the surface of the active asteroid (101955) Bennu. Science 366, 3544. doi:10.1126/science.aay3544 [Moreno et al.(2014)]Moreno2014 Moreno, F., Licandro, J., Álvarez-Iglesias, C., Cabrera-Lavers, A., Pozuelos, F. 2014. Intermittent Dust Mass Loss from Activated Asteroid P/2013 P5 (PANSTARRS). The Astrophysical Journal 781. doi:10.1088/0004-637X/781/2/118 [Moreno et al.(2019)]Moreno2019 Moreno, F. and 13 colleagues 2019. Dust properties of double-tailed active asteroid (6478) Gault. Astronomy and Astrophysics 624. doi:10.1051/0004-6361/201935526 [Naidu et al.(2020)]Naidu2020 Naidu, S. P. and 12 colleagues 2020. Radar observations and a physical model of binary near-Earth asteroid 65803 Didymos, target of the DART mission. Icarus 348. doi:10.1016/j.icarus.2020.113777 [Pravec and Harris(2000)]Pravec2000 Pravec, P., Harris, A. W. 2000. Fast and Slow Rotation of Asteroids. Icarus 148, 12–20. doi:10.1006/icar.2000.6482 [Pravec et al.(2006)]Pravec2006 Pravec, P. and 56 colleagues 2006. Photometric survey of binary near-Earth asteroids. Icarus 181, 63–93. doi:10.1016/j.icarus.2005.10.014 [Pravec et al.(2008)]Pravec2008 Pravec, P. and 30 colleagues 2008. Spin rate distribution of small asteroids. Icarus 197, 497–504. doi:10.1016/j.icarus.2008.05.012 [Pravec et al.(2019)]Pravec2019 Pravec, P. and 49 colleagues 2019. Asteroid pairs: A complex picture. Icarus 333, 429–463. doi:10.1016/j.icarus.2019.05.014 [Richardson et al.(2002)]Richardson2002 Richardson, D. C., Leinhardt, Z. M., Melosh, H. J., Bottke, W. F., Asphaug, E. 2002. Gravitational Aggregates: Evidence and Evolution. Asteroids III 501–515. [Roberts et al.(2021)]Roberts2021 Roberts, J. H. and 20 colleagues 2021. Rotational states and shapes of Ryugu and Bennu: Implications for interior structure and strength. Planetary and Space Science 204. doi:10.1016/j.pss.2021.105268 [Rossi et al.(2022)]Rossi2022 Rossi, A. and 12 colleagues 2022. Dynamical Evolution of Ejecta from the DART Impact on Dimorphos. The Planetary Science Journal 3. doi:10.3847/PSJ/ac686c [Sabuwala et al.(2021)]Sabuwala2021 Sabuwala, T., Chakraborty, P., Shinbrot, T. Bennu and Ryugu: diamonds in the sky. Granular Matter 23, 81 (2021). https://doi.org/10.1007/s10035-021-01152-z [Sánchez and Scheeres(2012)]Sanchez2012 Sánchez, D. P., Scheeres, D. J. 2012. DEM simulation of rotation-induced reshaping and disruption of rubble-pile asteroids. Icarus 218, 876–894. doi:10.1016/j.icarus.2012.01.014 [Sánchez and Scheeres(2016)]Sanchez2016 Sánchez, P., Scheeres, D. J. 2016. Disruption patterns of rotating self-gravitating aggregates: A survey on angle of friction and tensile strength. Icarus 271, 453–471. doi:10.1016/j.icarus.2016.01.016 [Sanchez et al.(2019)]Sanchez2019 Sanchez, J. A. and 6 colleagues 2019. Physical Characterization of Active Asteroid (6478) Gault. The Astrophysical Journal 881. doi:10.3847/2041-8213/ab31ac [Walsh et al.(2012)]Walsh2012 Walsh, K. J., Delbo', M., Mueller, M., Binzel, R. P., DeMeo, F. E. 2012. Physical Characterization and Origin of Binary Near-Earth Asteroid (175706) 1996 FG_3. The Astrophysical Journal 748. doi:10.1088/0004-637X/748/2/104 [Werner(1994)]Werner1994 Werner, R. A. 1994. The Gravitational Potential of a Homogeneous Polyhedron or Don't Cut Corners. Celestial Mechanics and Dynamical Astronomy 59, 253–278. doi:10.1007/BF00692875 [Yu et al.(2017)]Yu2017 Yu, Y., Richardson, D. C., Michel, P. 2017. Structural analysis of rubble-pile asteroids applied to collisional evolution. Astrodynamics 1, 57–69. doi:10.1007/s42064-017-0005-6 [Yu et al.(2019)]Yu2019 Yu, Y., Michel, P., Hirabayashi, M., Richardson, D. C. 2019. The expansion of debris flow shed from the primary of 65803 Didymos. Monthly Notices of the Royal Astronomical Society 484, 1057–1071. doi:10.1093/mnras/sty3515 [Zhang et al.(2017)]Zhang2017 Zhang, Y. and 9 colleagues 2017. Creep stability of the proposed AIDA mission target 65803 Didymos: I. Discrete cohesionless granular physics model. Icarus 294, 98–123. doi:10.1016/j.icarus.2017.04.027 [Zhang et al.(2021)]Zhang2021 Zhang, Y. and 7 colleagues 2021. Creep stability of the DART/Hera mission target 65803 Didymos: II. The role of cohesion. Icarus 362. doi:10.1016/j.icarus.2021.114433

http://arxiv.org/abs/2303.06632v1

20230312110418

Focus on Change: Mood Prediction by Learning Emotion Changes via Spatio-Temporal Attention

cs.HC

op-tical net-works semi-conduc-tor IEEE-Xplore B-.05emi-.025em b-.08em T-.1667em.7exE-.125emX L>l<C>c<R>r<𝐱Ł Le.g.i.e.E.g.et al.etc Pre-print version Shell et al.: Bare Demo of IEEEtran.cls for Signal Processing Society Society JournalsFocus on Change: Mood Prediction by Learning Emotion Changes via Spatio-Temporal Attention Soujanya Narayana, Student Member, IEEE, Ramanathan Subramanian, Senior Member, IEEE, Ibrahim Radwan, Member, IEEE, and Roland Goecke, Senior Member, IEEESoujanya Narayana, Ramanathan Subramanian, Ibrahim Radwan, and Roland Goecke are with the Human-Centred Technology Research Centre, Faculty of Science and Technology, Uni. Canberra, Bruce, ACT, Australia. ========================================================================================================================================================================================================================================================================================================================================================================== While emotion and mood interchangeably used, they differ in terms of duration, intensity and attributes. Even as multiple psychology studies examine the mood-emotion relationship, mood prediction has barely been studied. Recent machine learning advances such as the attention mechanism to focus on salient parts of the input data, have only been applied to infer emotions rather than mood. We perform mood prediction by incorporating both mood and emotion change information. We additionally explore spatial and temporal attention, and parallel/sequential arrangements of the spatial and temporal attention modules to improve mood prediction performance. To examine generalizability of the proposed method, we evaluate models trained on the AFEW dataset with EMMA. Experiments reveal that (a) emotion change information is inherently beneficial to mood prediction, and (b) prediction performance improves with the integration of sequential and parallel spatial-temporal attention modules. Mood, Emotion, Spatial Attention, Temporal Attention, Unimodal, Multimodal § INTRODUCTION Psychology and cognitive science studies reveal that emotions play an essential role in rational decision making, perception, and other cognitive processes <cit.>. There is also a growing body of evidence that a healthy balance of emotions is integral to human intelligence and problem solving <cit.>. Although emotions have been studied by researchers for long, no single definition has been agreed upon. Emotions are considered as neurophysiological changes caused by external stimuli, and associated with feelings and behavioural changes <cit.>. Mood is another affective state, which is often considered synonymous with emotion. Although there are no absolute boundaries between these two affective states, they differ in terms of duration, intensity and attributes. While emotion is a short-term affective state lasting a few seconds or minutes, mood denotes a long-term affective state lasting for minutes, hours or even days <cit.>. Mood akin to emotion is known to influence thought process, creativity and judgement <cit.>. While much research has been devoted towards emotion inference[We use the term inferring instead of recognising emotions/mood.]<cit.>, very few computational studies attempt mood prediction <cit.>. The psychology literature (a) recognizes mood and emotion as distinct mechanisms that repeatedly trigger the arousal of each other <cit.>, and (b) ascribes mood as a higher-level disposition that activates lower-level states such as emotions and beliefs <cit.>. Nevertheless, joint modelling of mood and emotion for affect prediction has been neglected. From an affective computing perspective, mood inference entails learning from data rich in affective state annotations, and assigning them to mood categories. The limited number of mood prediction studies can be attributed to the few databases with mood annotations. Popular affective databases such as HUMAINE <cit.>, SEMAINE <cit.> and DREAMER <cit.> are rich in emotion annotations, whereas barely any dataset exists with mood annotations– these annotations denote the mood perceived by the human annotator rather than the actual mood expressed by the actor <cit.>. Another reason for the limited availability of mood-labeled data is that affective labeling is onerous. While a large body of work leverages machine learning for emotion inference via visual, acoustic, and textual data <cit.>, very few explore mood inference. Recent computer vision studies are considerably inspired by the human ability to naturally focus on salient regions in complex scenes, known as attention. Whilst developed for machine translation tasks <cit.>, attention is popularly incorporated in vision systems to improve their recognition performance <cit.>. The attention mechanism guides models to learn what information is meaningful and where to look for the same for a certain task. Attention has also been applied for emotion inference, and to guide models to focus on emotionally relevant parts of speech <cit.>, video frames <cit.>, . While attention has been applied for detecting mood disorders <cit.>, it has not been employed for mood prediction. Mood inference is critical for applications such as diagnosis of mood-related disorders, mood regulation, and developing affective interfaces. This study explores the use of emotion information for mood prediction on the AFEW-VA dataset <cit.> (subset of AFEW <cit.>) comprising valence annotations. Valence denotes the extent of pleasure or displeasure elicited by an event, and we derive positive, negative and neutral mood labels from valence annotations. We additionally develop emotion change (Δ) labels, which denote the change in emotion over a specific window size. For mood prediction, we preliminarily investigate (1) a single-branch 3-dimensional Convolution Neural Network (3D CNN) which uses only mood labels, (2) a two-branch (mood-Δ) convolutional neural network with feature fusion followed by a Multi-Layer Perceptron, (3) a two-branch (mood-Δ) convolutional neural network with end-to-end training, and (4) a teacher-student network which uses both mood and Δ labels. Empirical results described in our prior work <cit.> reveal that emotion change information is beneficial, and enhances mood prediction performance. This work additionally explores the role of attention in mood prediction. We integrate spatial attention, temporal attention, and sequential/parallel spatial-temporal (SST/PST) attention modules with the above architectures to perform mood prediction. Furthermore, to examine generalisabilty, our trained models are evaluated on the EMMA dataset <cit.>, which includes mood and emotion annotations. Fig. <ref> describes our framework, including the prediction architectures and attention modules. Our research contributions are summarised below: * Incorporating emotional change information is beneficial and improves mood prediction performance as compared to models trained only with mood labels. * Mood prediction improves when spatial, temporal, sequential and parallel spatial-temporal attention modules are integrated within the inference architectures. * Our mood inference framework is generalizable as high prediction accuracies are achieved on EMMA <cit.> by models trained on AFEW-VA <cit.>. Performance is evaluated at both the video and chunk levels, where a chunk denotes a snippet of five contiguous frames extracted from the original video. The paper is organised as follows. Sec. <ref> examines the literature on mood and emotion. Sec. <ref> details the datasets used and the process of computing the mood and Δ labels. Mood inference models and attention modules employed for mood prediction are discussed in Sec. <ref>. Empirical results are discussed in Sec. <ref>, and Conclusions presented in Sec. <ref>. § RELATED WORK We examine the literature in terms of (a) psychology studies relating mood and emotion, and (b) computational studies examining emotion and mood inference. §.§ Mood, Emotion and their Interplay Psychologists consider mood and emotion as closely associated, but also distinguish them in terms of duration, intensity, stability, and attributes. Mood is described as a prolonged affect, lasting from minutes to hours or even days, with no specific trigger <cit.>. Conversely, emotions are considered to be quite brief, lasting from a few seconds to minutes, and involve the dynamic phases of onset-apex-offset <cit.>. Also, while at least some emotions (known as Ekman's basic or universal emotions) are known to manifest via unique facial expressions <cit.>, moods are not known to have unique manifestations. Cause and duration primarily distinguish the two affective states <cit.>. Despite differences between the two affective states, the psychology literature also agrees on the confluence of the two. Ekman claims that, partially, mood is inferred from emotional signals associated with mood <cit.>. E.g., we may deduce that someone is in a cheerful mood due to their joyful behaviour over a period of time. Mood is also known to be instantiated by emotions with varying intensity <cit.>. Furthermore, mood acts as a bias while emotions are elicited, reinforcing similar emotions, while dampening dissimilar ones <cit.>. This is similar to the mood-congruity effect, which states that positive mood facilitates the inference of mood-congruent (positive) emotions, but hampers the inference of mood-incongruent (negative) emotions, and vice-versa <cit.>. The mood-emotion loop theory proposed in <cit.> states that mood is a higher level variable activating lower level latent states, such as emotions and beliefs. This study also regards mood and emotions as mechanisms, which form a loop and repeatedly trigger the arousal of each other. To summarise, though mood and emotions are regarded as distinct affective states, the psychology literature identifies a relationship between the two <cit.>. §.§ Mood & Emotion Classification Most studies focus on devising a computational framework for inferring emotions, while only a few examine mood from a computational perspective <cit.>. Eye-tracking is utilized to observe that mood influences information processing styles <cit.>. Positive mood results in a global information processing style, while negative mood influences attention to detail <cit.>. Upper body behavior such as eye contact, arm openness, shoulder orientation, and head movement patterns are observed to perform mood inference in <cit.>. Participants maintained a vertical head position for longer in a positive mood, as compared to negative mood. Mood prediction from recognised emotions shows that clustered emotions in the valence-arousal space[Arousal denotes the level of physiological excitation ranging from bored to excited following an event.] predict single mood better than multiple moods in a video <cit.>. Machine learning algorithms are popularly used for affect inference due to their ability to learn high-level representations from input features. A preliminary step in mood inference is to compile a dataset with mood annotations. Most affective databases comprise continuous and dimensional emotion annotations, , HUMAINE <cit.>, SEMAINE <cit.>, AFEW <cit.>, . In these corpora, emotions are expressed via facial expressions, bosy postures, and physiological signals. EMMA <cit.> is an affective database with mood labels annotated via crowdsourcing. However, EMMA is intended for monitoring the affect of care center residents, with a focus on detecting negative mood. Given the complexity of emotions and how they manifest, a single modality becomes inadequate for learning and representing emotions. Hence, fusing multiple modalities for emotional inference has proved to be more effective than employing unimodal methods <cit.>. Early fusion aims to integrate features extracted from multiple modalities <cit.>. Whereas, late fusion allows for adopting a modality-specific model, and fuses individual predictions by weighting or voting <cit.>. Hybrid fusion corresponds to a trade-off between early fusion and late fusion, and aims at exploiting their advantages in a unified framework <cit.>. Knowledge Distillation (KD) is the process of distilling knowledge from an ensemble of models into a single model <cit.>. This technique achieves both model compression and learning with privileged information <cit.>. A large teacher network is trained with privileged information and transfers this knowledge to the smaller student network to achieve better inference efficiency. It is shown that knowledge can be effectively distilled using the soft target distribution produced by the teacher <cit.>. KD is employed for facial expression recognition, with the teacher having access to fully visible faces, and the student accessing occluded faces in <cit.>. For micro-expression recognition, a multi-task, multi-label network is designed with a residual network as the teacher and a two-layer CNN as the student <cit.>. §.§ Attention for affect classification Inspired by the way in which human perception focuses on certain parts of a given scene to better process information, the attention mechanism seeks to learn salient parts in the input data. Attention is a dynamic selection process that is realized by adaptively weighting features based on their relative importance in the input <cit.>. Introduced in neural machine translation <cit.>, attention has been widely implemented in several visual tasks like face recognition <cit.>, action recognition <cit.>, pose estimation <cit.>, emotion inference <cit.>, . Attention methods include spatial attention (where to attend), channel attention (what to attend), temporal attention (when to attend), and hybrid approaches. For image classification and object detection tasks, Convolutional Block Attention Module (CBAM) is a simple-yet-efficient module implemented on feed-forward CNNs, with channel and spatial attention <cit.>. Temporal attention modules are introduced in Long-Short Term Memory (LSTM) networks to emphasize the expressive speech parts in speech emotion inference <cit.>. Temporal attention is used to weigh audio-visual time-windows that span short video clips, as temporal attention is a powerful approach for sequence modelling, and can be used to fuse audio-visual cues over time <cit.>. To extract discriminative features from EEG signals and improve emotion inference, authors in <cit.> employ an attention-based convolutional recurrent neural network (CRNN), which adopts a channel-wise attention mechanism to adaptively assign weights to different channels. While attention has been employed for emotion inference tasks, less has been done towards implementing attention for mood inference. Temporal attention is integrated with a CNN model to highlight speech responses in unipolar and bipolar disorder detection <cit.>. Attention-based models are found to improve mood disorder detection accuracy by roughly 11%. §.§ Literature summary and research gaps A close examination of related work reveals that (a) whilst extensive research has focused on emotion inference, very few studies have examined mood prediction or modeled the mood-emotion interplay in their framework; (b) Labeled datasets for mood prediction are sparse, as most affective databases only include continuous/categorical emotion labels; (c) Attention mechanism has been employed largely for emotion inference, but not for mood prediction. Differently, this work explores (i) the utility of incorporating emotion change information for mood classification into the neutral, positive and negative classes, (ii) integrating various arrangements of spatial and temporal attention modules with mood prediction architectures, (iii) model generalizability by evaluating AFEW-VA <cit.>-trained models on the EMMA <cit.> dataset, and (iv) evaluating mood prediction performance at the chunk and video levels. Our results demonstrate that accounting for the mood-emotion interplay, by training a teacher-student network with both mood and Δ labels, enables better mood inference than mood-based models. Incorporating attention in this framework further improves prediction performance and model generalizability. § MATERIALS This section presents datasets used in our study, followed by the procedure adopted for synthesizing mood labels from valence annotations. §.§ Datasets We employ the AFEW-VA dataset <cit.> for compiling mood labels. We further evaluate the trained models on the EMMA dataset <cit.>. The two datasets are described below: §.§.§ AFEW-VA AFEW <cit.> is a corpus containing facial expression videos. AFEW-VA <cit.> is a subset of AFEW containing time-continuous valence-arousal annotations. AFEW-VA clips are 10–145 frames long, and are captured under challenging indoor and outdoor conditions. Each video frame is annotated for valence and arousal intensities in the [-10,10] range. The annotations are produced by two expert annotators (one male, one female). §.§.§ EMMA EMMA <cit.> is an affective database containing 180 acted non-interactive videos typically depicting daily scenarios, and subtle facial and body expressions. The videos are lab-recorded, and depict 15 (6 male, 9 female) actors. The situational context and scenarios are defined to the actors, followed by mood induction procedures to elicit the target mood. Each video is associated with a categorical mood label acquired via crowdsourcing. §.§ Labels §.§.§ Mood labels For each AFEW-VA video, we assigned a mood label based on the per-frame valence level. We assigned a model label of +1 (positive mood) to frames with valence values in the range (3,10], 0 (neutral mood) to the valence range [-3,3], and -1 (negative mood) to the (-3,-10] range. We then synthesized video clips/input samples from the original video considering overlapping sequences of 5 contiguous frames. Each clip was assigned the mode of the per-frame mood labels in a sequence. Via this scheme, we extracted a total of 27,651 clips from the AFEW-VA dataset. As EMMA already contains mood labels, we again consider overlapping five-frame sequences from each video, resulting in a total of 536,063 clips. The mood label for each EMMA sample is the same as the source video. §.§.§ Emotion change (Δ) labels In addition to mood labels, each input sample is assigned an emotion change or Δ label, which refers to the change in emotional valence over k frames. For a video with n frames and a window size of k, it is computed as the difference in the valence between the t^th frame and (t - k + 1)^th frame, for t = k, k+1, ..., n. For example, considering k = 5, if v_5 = -4 and v_1 = -2, where v_5 and v_1 respectively denote the valence of the fifth and the first frames, Δ = v_5 - v_1 = -2 (See Fig. <ref>). We assign the sample Δ label as (Δ) defined below: Δ = -1 if Δ < 0, 0 if Δ = 0, 1 if Δ > 0. Each AFEW-VA sample has mood and Δ labels ∈{0, -1, +1}, which are employed for the model training. Since EMMA is used only for testing, mood labels alone suffice for the EMMA samples. § MOOD CLASSIFICATION ARCHITECTURES As we seek to explore if the use of emotion information is beneficial for mood classification, we train different deep neural networks with mood and Δ labels, and compare their performance against models trained only with mood labels. The trained model architectures are described below followed by a discussion of the attention modules. §.§ Classification Architectures §.§.§ 1-CNN We employ a single branch three-layered convolutional neural network, or 1-CNN for mood classification. The model has three convolutional layers which comprise of 16, 32, and 32 kernels respectively with size 3 × 3 × 3 and each of them convolve the input sample with a stride of 3. Followed by each of the convolutional layer is an average pooling layer with a stride of 2-pixel regions. The output of the last convolutional layer is flattened, followed by batch normalization. Following the dense layer comprising 512 neurons is the Softmax layer with three neurons corresponding to the mood classes. The input dimensionality for the 1-CNN is 5 × 32 × 32 × 3, with each sample comprising 5 frames of size 32 × 32 and 3 channels. The model is optimized with the categorical cross-entropy loss function, with the Adam optimiser used for stochastic gradient descent. Fine-tuned hyper-parameters include learning rate ∈{10^-3, 10^-5}, batch size ∈{64, 128, 256}, and dropout rate ∈{0.4, 0.5}. §.§.§ 2-CNN + MLP Multimodal systems have gained increasing attention from researchers <cit.> by overcoming limitations of unimodal counterparts <cit.>. Multimodal systems fuse complementary information from different modalities at multiple levels. To examine the influence of emotion change (Δ) on mood prediction, we fuse the two employing 2-CNN + Multi-layer Perceptron (MLP) as in Fig. <ref>. The 2-CNN + MLP network is a two-branch model, with two identical 1-CNNs along each branch; mood labels are fed to one branch, and Δ labels fed to the other. From the penultimate layer in each branch, a 512-dimensional vector is obtained, and the two vectors are concatenated to obtain a 1024-feature vector input to the MLP for mood classification. Δ labels are assumed to be available during the train and test phases in this model. Dimensionality of the input samples and hyper-parameters are identical to 1-CNN. §.§.§ 2-CNN As a substitute to model fusion, we trained an end-to-end model for mood classification <cit.> via a 2-CNN model shown in Fig. <ref>. Similar to the 2-CNN + MLP, the top and bottom branches are trained with mood and Δ labels respectively, but both branches are trained simultaneously but not individually. The categorical cross-entropy losses L_m and L_Δ from the two branches are summed up, and the cumulative loss is minimised. Unlike the 2-CNN + MLP, where Δ labels are used both in train and test phases, Δ labels act as auxiliary information for the 2-CNN model, and are required only during the training phase (mood labels suffice for testing). §.§.§ TS-Net Besides model fusion and end-to-end training, we employ Knowledge Distillation, which is the process of transferring knowledge from a large model to smaller ones <cit.>. We employ a Teacher-Student Network (Fig. <ref>(right)) where the teacher (large model) distills knowledge to the student (smaller model), and thereby the student tries to mimic the teacher. The pre-trained teacher model (2-CNN + MLP) uses both mood labels and Δ labels in the training phase to distill knowledge to the student (1-CNN). The student model, whic only utilizes mood labels, is activated for inference. The Softmax layer of the student model involves the temperature (T) hyper-parameter, which controls the smoothness of the output probabilities. As T grows, the output Softmax probabilities become softer, providing more information regarding the inter-class relations than one-hot labels <cit.>. The distillation loss is computed via the Kullback–Leibler (KL) divergence measure, while the student loss is computed via sparse categorical cross-entropy. The overall loss of the TS-Net, L_TS is computed as the weighted sum of the distillation loss L_dist and the student loss L_stu, and is given by: L_TS = α L_stu + (1 - α) L_dis where α is a training hyper-parameter. The other fine-tuned hyper-parameters include batch size ∈{16, 64, 128}, T∈{3, 5, 7} and α∈{0.05, 0.1, 0.15, 0.2, 0.25, 0.3}. §.§ Attention modules §.§.§ Spatial attention In humans, spatial attention allows to selectively process visual information through prioritization of locations within a field. Spatial attention modules have been widely implemented in tasks like action recognition, affect classification, . to improve model performance <cit.>. Inspired by <cit.>, we generate spatial attention maps to emphasize meaningful features. As shown in Fig. <ref> (left), spatial attention is computed by applying max and average-pooling operations along the channel axis, followed by concatenation to generate a feature descriptor. A 3D convolutional layer is applied on the concatenated feature, and the output is multiplied with the input to generate a spatial attention map, which emphasizes where to look for in the video clip. §.§.§ Temporal attention Temporal attention is a dynamic mechanism which determines when to pay attention, and directs visual focus to specific time instants. A video represents a sequence of visual segments with large content variance and complexity, where the amount of valuable information provided by frame segments is generally unequal <cit.>, and discriminative information is provided by only some key frames. To model temporal dynamics in video, we apply LSTM networks, thereby capturing long-term temporal information. As shown in Fig <ref> (right), we use two LSTM layers with 128 hidden units per layer, followed by a dense layer with a single neuron. The input feature is multiplied with the output of the dense layer, facilitating the enhanced model to distinguish the most informative frame(s) for mood classification. §.§.§ Integrating attention modules Given an input clip, the spatial and temporal attention modules generate complementary attention information by respectively identifying salient regions and key frames. We integrate spatial and temporal attention modules into the classification networks, and also arrang these modules sequentially and parallely to synthesize the sequential spatial-temporal (SST) and parallel spatial-temporal (PST) attention architectures. * SST Attention: In SST Attention (Fig. <ref> (left)), the input sample I is sequentially fed into the spatial followed by the temporal attention modules. I is input to the spatial module first, and the attention map F_s is then convolved with I to produce I_s which is fed to the temporal module. The temporal map F_t is then multiplied with I_s to obtain I_st, which is fed to the classifier networks. * PST Attention: In PST Attention (Fig. <ref> (right), the input sample I is fed parallelly to the spatial and temporal attention modules. The spatial (F_s) and temporal (F_t) attention maps are then convolved with I to respectively obtain I_s and I_t, which are multiplied to obtain I_st input to the classification models. §.§ Integration with classification models We integrate the spatial and temporal attention modules individually and in combination with each of the mood classifer networks, as shown in Fig. <ref>. Corresponding results are presented in Sec <ref>. §.§ Evaluating on EMMA dataset As mentioned in Sec <ref>, EMMA is an affective database comprising acted videos with mood and emotion annotations. There are ≈ 200 frames in each EMMA video, while the maximum video frames in AFEW-VA is 145. However, EMMA is compiled under lab conditions and with actors, unlike AFEW-VA which represents an in-the-wild video dataset. To examine if (a) the mood and Δ labels assigned over smaller (5-frame) AFEW-VA snippets were applicable over larger EMMA clips, and (b) the mood prediction models trained on AFEW-VA were generalizable to EMMA, we evaluated all mood classifiers on EMMA utilizing only the mood labels. On performing a 90-10 stratified and random train-test split of the EMMA dataset, the test set comprised 18 videos, with 11 negative, 3 positive, and 4 neutral mood labels. Identical to AFEW-VA, each EMMA clip is of dimensionality 5 × 32 × 32 × 3, resulting in a total of 62,410 samples. We evaluate all trained models, with and without integrating attention modules, on EMMA. While evaluating models on EMMA, each 5-frame sample or chunk from the original video is tested, and we then evaluate models at both the chunk and video levels. This enables comparison of mood prediction performance over longer (video-level) and shorter (chunk-level) episodes depicting visual behavior. §.§ Performance metric Subject-independent 5-fold cross validation is used to evaluate all models to eliminate leakage effects. Mean accuracy over the five folds is reported as our performance metric. Results evaluating performance on the EMMA dataset also report the average performance over the five training models synthesized over the cross-validation process. § RESULTS AND DISCUSSION §.§ Classification Experiments §.§.§ Mood inference without attention modules Table <ref> presents AFEW-VA mood classification results without and with various attention modules. The first column, which tabulates accuracies excluding attention modules, conveys that the 1-CNN (Mood) trained with mood labels performs worst, revealing that mood prediction is a non-trivial problem. As a pre-cursor to the 2-CNN + MLP and 2-CNN models which include a CNN-branch trained with Δ labels, we also evaluated a 1-CNN (Δ) model trained and tested with Δ labels; this model performs better than 1-CNN (Mood) implying that emotion changes are easier detected than mood. Comparing the performance of the 2-CNN + MLP and 2-CNN networks, both of which are trained with both mood and Δ labels, we note that the 2-CNN + MLP network performs considerably than the 2-CNN whilst both achieve superior mood inference vis-á-vis the 1-CNN (Mood) model. These findings cumulatively reveal that including emotion-change information for mood prediction, either implicitly as auxiliary information (2-CNN) or explicitly as additional features (2-CNN + MLP), benefits mood inference. The highest mood inference accuracy of 0.89 is obtained with the TS-Net, where the 2-CNN + MLP denotes the pre-trained teacher model trained with both mood and auxiliary Δ labels, while 1-CNN denotes the student model, trained only with mood labels. Overall, these results affirm our claim that modeling the mood-emotion interplay and incorporating emotion change information for mood inference enables superior mood prediction as compared to utilizing mood information alone. Δ labels capture emotion changes over a short duration (local phenomenon), in comparison to mood labels which are applicable over a longer duration (global phenomenon). Our results reveal that short-term emotion changes play a crucial role in inferring long-term mood. §.§.§ Mood inference with attention Mood prediction results on integrating attention modules to the classification models are shown in Table <ref> (`With attention' columns). We examine if integrating attention into the classification architecture in addition to learning emotion changes improves mood prediction. A horizontal comparison in Table <ref> conveys that irrespective of the attention module, the 1-CNN performance significantly improves as compared to the baseline 1-CNN trained with mood labels. Conversely, the 2-CNN + MLP and TS-Net (with the 2-CNN + MLP as the Teacher model) do not benefit when attention modules are integrated, and their prediction accuracies somewhat reduce. The 1-CNN (Δ) does not benefit with either spatial or temporal attention, but its performance improves with sequence and parallel configurations of the spatial and temporal modules. With the 2-CNN network, where Δ labels are employed as auxiliary information, a significant performance improvement is noted with each of the attention modules. This observation reveals that the attention mechanism enables the 2-CNN to focus on informative pixels and important temporal segments for mood prediction. Fig. <ref> qualitatively supports this finding, where the input sample with a neutral mood label is incorrectly predicted as positive by the 2-CNN without any attention modules. Integration of the spatial attention module enables the classification model to predominantly focus on facial regions and achieve the correct prediction. Overall, empirical results reveal that classification models inherently learning from emotion change information (1-CNN (Δ), 2-CNN +MLP and TS-Net) do not additionally benefit from attention, while mood-based networks immensely benefit as attention facilitates focus on the facial region. To investigate the influence of emotion change (Δ) information on mood prediction, we compare Table <ref> results vertically. With spatial attention, while the 2-CNN and TS-Net perform better than the 1-CNN, the results are not significantly different. Likewise, integrating temporal attention also results in a higher accuracy with the 2-CNN and TS-Net as compared to 1-CNN (Mood). With respect to serial and parallel combinations of the spatial and temporal modules, with SST attention, the 2-CNN achieves significantly higher performance than 1-CNN (Mood) as per a two-sample t-test (t(8)=-4.7, p < 0.05). A significant difference is also observed between TS-Net and 1-CNN networks (t(8)=-2.6, p < 0.05). PST attention also yields significantly higher results with the 2-CNN (t(8)=-5.2, p<0.05), and the TS-Net (t(8)=-3.4, p<0.05) as compared with the 1-CNN. As mentioned above, the attention modules and learning from emotion change information are complementary means to improve mood recognition performance. A qualitative depiction of how learning information pertaining to emotion change is shown in Fig. <ref>. Attention maps for the input sample shown in the first row obtained with the 1-CNN (Mood) network including SST attention (2nd row) show that the model focuses regions in the vicinity of the face for mood inference. However, adding auxiliary information in the form of emotion change labels (3rd row) strongly localizes focus on the facial region to achieve superior mood inference. Cumulatively, Fig. <ref> and Fig. <ref> reinforce the observation that incorporating attention or emotion-change information for mood inference, are complementary means to the same end. §.§.§ Evaluation on EMMA As unlike the AFEW-VA dataset where mood labels are derived from valence annotations, the EMMA dataset comprises mood annotations. Also, mood labels in EMMA correspond to longer video segments as compared to EMMA. To examine if the models trained on AFEW-VA, and specifically to investigate if incorporating emotion change information can positively impact EMMA mood prediction, we evaluated all trained models on the EMMA dataset and results are presented in Table <ref>. With no attention, the TS-Net performs best with 69% accuracy followed by the 1-CNN, while the 2-CNN performs worst on EMMA. Including attention modules improves 2-CNN mood prediction performance, while not benefitting inference with the 1-CNN or TS-Net. Overall, the TS-Net performs best on EMMA; we did not fine-tune the AFEW-VA models on EMMA as our main objective was to test model generalizability. We also evaluated mood prediction performance at the chunk (5-frame episodes) and video levels on EMMA. Table <ref> presents the corresponding results using TS-Net, which achieves optimal performance on AFEW-VA. While TS-Net without attention yields significantly higher results at the chunk level than at the video level, integrating attention modules results in significantly higher accuracies at the video-level. A highest accuracy of 0.89 is achieved with the models involving temporal attention. These results convey that while the AFEW-VA models generalize better to EMMA at the video-level, affirming the validity of our annotation and model training processes. §.§ Discussion Summary The classification experiments convey that (a) incorporating emotion change information in addition to the mood labels positively impacts mood prediction; (b) integration of various attention modules to the architectures enables models to focus on key areas/input frames and improves mood prediction; (c) Qualitative results furthermore demonstrate that attention and emotion-change information play a complementary role towards enhancing mood inference. Additionally, AFEW-VA trained models generalize better to EMMA at the video-level than at the chunk-level, confirming that our methodology for annotating mood labels is valid, and can generalize to videos of longer duration. § CONCLUSION With majority of the contemporary studies focusing on emotion inference using multiple modalities and attention mechanism, negligible work has been done towards mood inference and examining the interplay between mood and emotions. The applicability of attention methods for examining mood is also neglected. As a step towards fulfilling these disparities, this work proposes using emotion change information along with attention mechanism for performing mood prediction. Using the AFEW-VA dataset, a single-branch CNN, two-branch CNN with MLP, two-branch CNN with end-to-end training and teacher-student network are employed as base models for mood prediction. Emotion change (Δ) labels, referring to the difference in emotion over a fixed window size, is used in addition to the mood labels for examining mood. Further, various attention mechanisms are integrated to the base models to enhance the model's performance. These models trained and tested on AFEW-VA are evaluated on EMMA dataset to explore the generalisability of approach. Overall, we observe that without integrating attention modules in two-branch CNN and teacher-student network, incorporating emotion change information along with the mood labels improves mood prediction performance. The integration of parallel spatial-temporal attention modules in these models also yields similar results, further validating our claim. This approach is also generalizable as identical results with teacher-student network are observed when the model is evaluated on EMMA dataset. Better results in the video-level analysis as compared to chunk-level analysis shows that the model is also robust towards videos of longer duration. Future work will investigate the use of semi-supervised and weakly supervised techniques, and substitute emotion-change labels with difference or class-activation maps. § ACKNOWLEDGEMENTS This research is partially funded by the Australian Government through the Australian Research Council’s Discovery Projects funding scheme (project DP190101294). IEEEtran -2 plus -1fil [ < g r a p h i c s > ]Soujanya Narayana received her Bachelor's degree from Regional Institute of Education, Mysore, India in 2015 and Masters degree from Christ University, Bengaluru, India in 2017. She is currently a PhD student in the Human Centered Technology Research Centre at the University of Canberra, Australia. Her research interests include affective computing in human-computer interaction and computer vision. She is a student member of the IEEE. -3 plus -1fil [ < g r a p h i c s > ]Ramanathan Subramanian received his Ph.D. in Electrical and Computer Engg. from NUS in 2008. He is Associate Professor in the School of IT & Systems, University of Canberra. His past affiliations include IIT Ropar, IHPC (Singapore), U Glasgow (Singapore), IIIT Hyderabad and UIUC-ADSC (Singapore). His research focuses on Human-centered computing, especially on modeling non-verbal behavioral cues for interactive analytics. He is an IEEE Senior Member, and an ACM and AAAC member. -3 plus -1fil [ < g r a p h i c s > ]Ibrahim Radwan received the Ph.D. degree in computer science from the University of Canberra in 2015. From 2014 to 2016, he was a researcher in a leading automotive industry warehouse, Research Fellow with The Australian National University and currently an assistant professor at the University of Canberra. His research includes computer vision, machine learning, robotics, and artificial intelligence. -3 plus -1fil [ < g r a p h i c s > ]Roland Goecke received his Ph.D. degree in computer science from The Australian National University, Canberra, Australia, in 2004. He is Professor of Affective Computing with the University of Canberra, where he is serves as Director of the Human-Centred Technology Research Centre. His research interests include affective computing, pattern recognition, computer vision, human–computer interaction and multimodal signal processing. He is a senior member of the IEEE, and an ACM and AAAC member.

http://arxiv.org/abs/2303.07017v2

20230313112742

O'Grady tenfolds as moduli spaces of sheaves

math.AG

Readout of quantum devices with a sideband microwave interferometer immune to systematic noise A. Fuhrer March 30, 2023 ============================================================================================== We give a lattice-theoretic characterization for a manifold of type to be birational to some moduli space of (twisted) sheaves on a K3 surface. We apply it to the Li–Pertusi–Zhao variety of type associated to any smooth cubic fourfold. Moreover we determine when a birational transformation is induced by an automorphism of the K3 surface and we use this to classify all induced birational symplectic involutions. § INTRODUCTION To know birational models of irreducible holomorphic symplectic (ihs) manifolds is a significant stage towards the full comprehension of their geometry. Throughout the paper we refer to ihs manifolds deformation equivalent to the O'Grady ten dimensional example as ihs manifolds of type. Many examples of ihs manifolds are realized as moduli spaces of sheaves on abelian or K3 surfaces. The O'Grady examples are symplectic resolutions of these moduli spaces and form a codimension 3 locus in the space of deformations of ihs manifolds of type. For this reason, given any such manifold it is then natural to investigate when it is birational to a moduli space of sheaves. The first motivation for this paper was to provide an answer to this question. As it is well known, the geometry of an ihs manifold X is encoded in its second integral cohomology group H^2(X,ℤ). This group is torsion free and carries a symmetric non-degenerate bilinear form, called the Beauville–Bogomolov–Fujiki form, which induces a lattice structure on it. The first main result of the paper provides an answer to this question via a lattice-theoretic characterization. To this end, similarly to the already existing terminology <cit.> we introduce the lattice-theoretic notion of numerical moduli space (see <ref>). [<ref>] Let (X,η) be a marked pair of type. The following conditions are equivalent: * There exists a K3 surface S, a Mukai vector v=2w where w is a primitive Mukai vector of square 2, and a v-generic polarization θ on S such that X is birational to ; * The manifold X is a numerical moduli space. While this extends analogous results for the other known deformation types of ihs manifolds <cit.>, the theory of moduli spaces of sheaves on K3 surfaces developed into considering also their twisted analogues. Originally, the motivation for generalization to twisted K3 surfaces was the existence of coarse moduli spaces. However, it has become more and more clear that allowing twists has quite unexpected applications (see for example <cit.> for further details). In <ref> we then consider the twisted picture and extend the above result to this setting. Later we consider the ihs manifold of type X_Y defined by Li, Pertusi, and Zhao (LPZ) which is associated to a smooth cubic fourfold Y <cit.>. We are able to provide a purely lattice-theoretic characterization for X_Y to be birational to a moduli space of sheaves on a (possibly twisted) K3 surface, providing a new proof of <cit.>. [<ref> and <ref>] Let X_Y be an LPZ-variety associated to a smooth cubic fourfold Y. The following conditions are equivalent. * The cubic fourfold Y lies in the Hassett divisor 𝒞_d with d satisfying (**) (resp. (**')). * The LPZ-variety X_Y is a (resp. twisted) numerical moduli space. The same question for Fano varieties of lines of cubic fourfolds and for the so called Lehn–Lehn–Sorger–van Straten symplectic eightfolds has been previously answered in <cit.> and <cit.>. Once we know under which conditions an ihs manifold of type is birational to a moduli space of sheaves on a K3 surface S, it is then natural to investigate which birational transformations are induced by an automorphism of S (see <ref>). To this end we introduce the notion of numerically induced birational transformations (see <ref>), which is a characterization in terms of their action on the second integral cohomology lattice. [<ref>] Let (X,η) be a smooth marked pair of type and let G⊂Bir(X) be a finite subgroup. If G is a numerically induced group of birational transformations, then there exists a K3 surface S with an injective group homomorphism G↪Aut(S), a G-invariant Mukai vector v and a v-generic polarization θ on S such that X is birational to and G⊂Bir(X) is an induced group of birational transformations. For what concerns symplectic actions on type manifolds we know that the unique symplectic regular automorphism of finite order is the identity <cit.>, and symplectic birational involutions are classified <cit.>. Lattice-theoretic constraints for nonsymplectic automorphisms of type are treated in <cit.>. Nonsymplectic involutions will be treated in <cit.>, and so far nothing is known for higher order actions. Applying our criterion we show that there exists a unique induced birational symplectic involution on an ihs manifold of type. More precisely we give the following characterization. [<ref>] Let (X, η) be a marked pair of type. Assume that X is a numerical moduli space and let φ∈(X) be a birational symplectic involution. Then φ is induced if and only if H^2(X,ℤ)^φ≅^⊕3⊕_8(-2) ⊕_2(-1) and H^2(X,ℤ)_φ≅_8(-2). §.§ Outline of the paper In <ref> we recall some preliminary results about lattice theory and briefly describe the lattice structure of the second cohomology group of an ihs manifold. In <ref> we specialize to manifolds of type: after describing their construction and properties, we state and prove <ref>. In <ref> we provide a generalization of it to the twisted case. In <ref> we apply the birationality criteria of <ref> and <ref> to the LPZ variety. Finally in <ref> we examine birational transformations of type to determine which ones are induced. §.§ Acknowledgements The authors wish to thank Luca Giovenzana and Giovanni Mongardi for useful discussions. C.F. was partially supported by PRIN "Moduli spaces and Birational Classifications". F.G. and A.G. were supported by the DFG through the research grant Le 3093/3-2. All authors are supported by INDAM GNSAGA. § PRELIMINARIES IN LATTICE THEORY In this section we recall some basic facts of lattice theory for irreducible holomorphic symplectic manifolds, focusing on the special case of type. §.§ Abstract lattices A lattice is a free finite rank ℤ-module with a non-degenerate symmetric bilinear form ·× →ℤ (e,f) ↦ e· f. With an abuse of notation we denote e· e simply by e^2. Moreover, we say that the lattice is even if e^2 is even for any element e ∈ and we denote the determinant, the rank and the signature of the lattice by (), () and () respectively. Given a lattice , we define the dual lattice of to be the lattice ^∨_ℤ(,ℤ). The dual lattice can be characterized uniquely as ^∨={ x∈⊗ℚ| x· y∈ℤ for all y∈}. Given e∈, we define the divisibility of e in is (e,)gcd{e· f| f∈}. One can show that the quotient A_^∨/ is a finite group, called discriminant group of , of order |()|. We say that a lattice is unimodular if A_ is trivial. We denote the unique unimodular even indefinite rank 2 lattice by . Moreover the symbols _n,_n,_n denote the positive definite ADE lattices of rank n. The notation [m] for some m∈ℤ indicates a rank 1 lattice generated by a vector of square m. For any lattice we denote by (n) the lattice with the same structure as as a ℤ-module and quadratic form multiplied by n. Whenever we have a lattice , we can consider a subgroup G⊂ O() of isometries on . We denote by ^G the invariant sublattice of with respect to the action of G, and by _G (^G)^⊥ its orthogonal complement in , which is called coinvariant sublattice. §.§.§ Hodge structure on Given a lattice with a weight 2 Hodge structure, i.e. a decomposition _ℂ⊗_ℤℂ=_ℂ^2,0⊕_ℂ^1,1⊕_ℂ^0,2, with _ℂ^2,0≅_ℂ^0,2 and _ℂ^1,1≅_ℂ^1,1, we will denote by ^1,1_ℂ^1,1∩, and refer to it as the (1,1) - part of the lattice . An embedding of lattices 𝐌↪ is called primitive if the group /𝐌 is torsion free. In this setting we denote by 𝐌^⊥ the orthogonal complement of 𝐌 in with respect to the bilinear form on . Let 𝐌 be a lattice endowed with a weight 2 Hodge structure. An embedding of lattices 𝐌↪ is called Hodge embedding if it is primitive and is endowed with a weight 2 Hodge structure defined as follows: _ℂ^2,0=𝐌_ℂ^2,0, _ℂ^1,1=(𝐌^1,1⊕𝐌^⊥)⊗ℂ, _ℂ^0,2=𝐌_ℂ^0,2. Moreover if has already a Hodge structure on it, we say that the Hodge embedding 𝐌↪ is compatible if the Hodge structure induced by 𝐌 is the one of . §.§ Lattice structure of irreducible holomorphic symplectic manidolds Suppose X is an irreducible holomorphic symplectic manifold of complex dimension 2n. The second integral cohomology group has a well defined lattice structure: in fact is a torsion free ℤ-module of finite rank endowed with a symmetric bilinear form (-,-)_X, called the Beauville-Bogomolov-Fujiki form, which satisfies the following equality for any α∈ H^2(X,ℤ): ∫_Xα^2n= c_X(α,α)^n_X, for some unique positive rational constant, called the Fujiki constant (see for example <cit.>, <cit.>). Note that both the Beauville–Bogomolov–Fujiki form and the Fujiki constant are invariant up to deformation. §.§.§ Lattice of O'Grady 10 type manifolds A marked pair of type is pair (X,η) where X is an irreducible holomorphic symplectic manifold type and η H^2(X,ℤ) → is a fixed isometry of lattices. In what follows, we denote by the isometry class of H^2(X,ℤ), which is well defined in view of <ref>. In <cit.> Rapagnetta shows that is isometric to ^⊕ 3⊕_8(-1)^⊕ 2⊕_2(-1), which is an even lattice of rank 24 and signature (3,21). Moreover, being the second cohomology group of a compact Kähler manifold, is endowed with a natural weight 2 Hodge structure, whose Hodge numbers are deformation invariant. In particular, if (X,η) is a marked pair of type, the integral lattice ^1,1 is called the Néron–Severi lattice of X and it is denoted by (X). Its orthogonal complement is called trascendental lattice and it is denoted by (X). § MODULI SPACES OF SHEAVES OF O'GRADY 10 TYPE In his seminal paper <cit.>, O'Grady discovers a new 10-dimensional ihs manifold not deformation equivalent to a Hilbert scheme of K3 or a generelized Kummer variety. The example arises as a symplectic resolution of a moduli space of sheaves on a K3 surface with fixed numerical constraints. In <ref> and in <ref> we recall the original construction and its generalizations, focusing also on the lattice-theoretic point of view. In <ref> we are finally ready to state and prove criteria to determine if a manifold of type is birational to a moduli space of sheaves on a K3 surface. §.§ The Mukai lattice Let S be a K3 surface. We denote by H(S,ℤ) the even integral cohomology of S, that is H(S,ℤ)=H^2*(S,ℤ)=H^0(S,ℤ)⊕ H^2(S,ℤ)⊕ H^4(S,ℤ). We can put a symmetric bilinear pairing on it, called the Mukai pairing, in the following way: H(S,ℤ)×H(S,ℤ) →ℤ (r_1,l_1,s_1),(r_2,l_2,s_2) ↦ -r_1s_2+l_1l_2-r_2s_1 with r_i∈ H^0(S,ℤ), l_i∈ H^2(S,ℤ) and s_i∈ H^4(S,ℤ). The lattice H(S,ℤ) together with the Mukai pairing is referred to as the Mukai lattice and it is isometric to ^⊕ 4⊕_8(-1)^⊕ 4. In what follows, any element v∈H(S,ℤ) will represented by a triple v=(v_0,v_1,v_2) with v_i∈ H^i(S) and will be called Mukai vector. The natural Hodge structure on H^2(S,ℤ) can be extended to a weight 2 Hodge structure on H(S,ℤ) by setting H_ℂ^2,0(S) H^2,0_ℂ(S) (resp. H_ℂ^0,2(S) H_ℂ^0,2(S)) and H^1,1_ℂ(S) H^0(S,ℂ)⊕ H_ℂ^1,1(S)⊕ H^4(S,ℂ). For any Mukai vector v∈H(S,ℤ) the orthogonal sublattice with respect to the Mukai pairing v^⊥={w∈H(S,ℤ)| (v,w)=0}⊆H(S,ℤ) inherits a weight 2 Hodge structure from H(S,ℤ) in an obvious fashion. Moreover, given a coherent sheaf ℱ on S, we can naturally associate a Mukai vector to it by setting v(ℱ)(rank (ℱ), c_1(ℱ),c_1(ℱ)^2/2 -c_2(ℱ) + rank(ℱ))∈H(S,ℤ). By construction, v(ℱ)=(r,l,s) is of (1,1)-type and it satisfies one of the following relations: * r>0; * r=0 and l≠ 0 with l effective; * r=l=0 and s>0. A nonzero vector v∈H(S,ℤ) satisfying v^2≥ 2 and one of the conditions above is called positive Mukai vector. §.§ Moduli spaces of sheaves of type Given a Mukai vector v and a v-generic polarization θ, we can consider the moduli spaces of θ-semistable sheaves on S with Mukai vector v. More generally we consider Mukai vectors v of type v=mw where m∈ℕ and w∈H(S,ℤ) is some primitive Mukai vector. It is well known that that if ≠∅ then is smooth of dimension v^2+2 and carries a symplectic form <cit.>. Indeed Mukai shows <cit.> that when v is a primitive Mukai vector then is an IHS manifold. Let S be a K3 surface, v be a primitive Mukai vector and θ be a v-generic polarization. Then = and if v^2≥ -2, then is a irreducible holomorphic symplectic manifold of dimension v^2+2=2n, which is deformation equivalent to the n-th Hilbert scheme of a K3 surface. Moreover there is a Hodge isometry between v^⊥≅ H^2(,ℤ). When v is not primitive can be singular with a holomorphic symplectic form on its smooth locus. It is then natural to investigate under which conditions admits a symplectic resolution of singularities, namely a resolution carrying a holomorphic symplectic form which extends the holomorphic symplectic form on the smooth locus of . O'Grady starts from this consideration and proves that the moduli space ℳ_10 with v=(2,0,-2), admits a symplectic resolution ℳ_10 which is a irreducible holomorphic symplectic manifold of dimension 10 and second Betti number 24. This latter condition implies in particular that cannot be deformation equivalent to a Hilbert scheme, so we have a new deformation class of ihs manifolds. We say that an ihs manifold X is of type if it is deformation equivalent to the O'Grady 10 dimensional example ℳ_10. The construction has been later generalized by Perego and Rapagnetta <cit.> allowing different types of Mukai vectors and polarizations. More precisely, Perego and Rapagnetta introduce the concept of OLS-triple (S,v,θ) where S is a K3 surface, v is a Mukai vector and θ a v-generic polarization (see <cit.>). They prove that whenever (S,v,θ) is an OLS-triple, then the moduli space admits a symplectic resolution which is of type. Let (S,v,θ) be an OLS-triple. The resolution of the moduli space is an irreducible holomorphic symplectic manifold which is deformation equivalent to ℳ_10. Consider a holomorphic symplectic manifold X of type. The second integral cohomology group of irreducible holomorphic symplectic manifolds is endowed with a lattice structure. Moreover Perego and Rapagnetta describe the weight 2 Hodge structure on H^2(,ℤ). Let (S,v,θ) be an OLS-triple. Then the pullback π^*:H^2(,ℤ)→ H^2(,ℤ) via the resolution map is injective, so that H^2(,ℤ) admits a pure weight 2 Hodge structure and a lattice structure induced by restricting the Beauville-Bogomolov-Fujiki form on H^2(,ℤ). Moreover there is an isometry of Hodge structures H^2(,ℤ)≅ v^⊥⊂ H^2(,ℤ). Finally they show that the class E∈ H^2(,ℤ) of the exceptional divisor has square 6 and divisibility 3. Moreover we have the following theorem. Keep the notation as above. There exists a finite index embedding v^⊥⊕ℤE↪ H^2(,ℤ). §.§ Birationality criteria for OG10 type manifolds We provide an entirely lattice-theoretic criterion to determine when a manifold X of type is birational to a moduli space of sheaves on a - possibly twisted - K3 surface S, where (S,v, θ) is an OLS-triple as described in <ref>. Let (X,η) be a projective marked pair of type, where η H^2(X,ℤ)→ is a fixed marking. We say that X is a numerical moduli space of type if there exists a primitive class σ∈^1,1 such that * σ^2=-6 and (σ,)=3, * the Hodge embedding σ^⊥_↪_24 embeds a copy of in _24^1,1 as a direct summand. We are now in a position to state our birationality criterion. Let (X,η) be a marked pair of type. The following conditions are equivalent: * There exists a K3 surface S, a Mukai vector v=2w where w is a primitive Mukai vector of square 2 and a v-generic polarization θ on S such that X is birational to ; * The manifold X is a numerical moduli space. Suppose Φ:X is a birational morphism. Then the induced isometry Φ^*: H^2(,ℤ)→ H^2(X,ℤ) is an isometry of Hodge structures. Recall that the manifold is obtained as a symplectic resolution of singularities of the moduli space . Note that the class of the exceptional divisor E∈ H^2(,ℤ) is primitive of (1,1) type, of square -6 and divisibility 3. Moreover, by <cit.>, we have that E^⊥≅ H^2(,ℤ) and there exists a Hodge embedding H^2(,ℤ)↪H(S,ℤ)≅_24. Let σ:=Φ^*E. We want to prove that X is a numerical moduli space: since Φ^* is an isometry of Hodge structures, condition (a) in <ref> is satisfied. For condition (b), observe that the composition σ^⊥E^⊥≅ H^2(,ℤ)↪H(S,ℤ)≅_24≅^⊕ 4⊕_8(-1)^⊕ 2 is a compatible Hodge embedding. We conclude by noticing that the vectors (1,0,0) and (0,0,1) in H(S,ℤ) lie in the image of this embedding and generate a copy of inside H(S,ℤ)^1,1≅_24^1,1. Thus X is a numerical moduli space. For the other direction, since X is a numerical moduli space, there exists a class σ∈^1,1 such that σ^2=-6 and (σ,)=3, and a Hodge isometry σ^⊥_↪_24 such that _24^1,1 contains a copy of as a direct summand. Observe that the lattice σ^⊥_ is isometric to the lattice ^⊕ 3⊕_8(-1)^⊕ 2⊕ [-2], so that setting w:=(σ^⊥_)^⊥__24, we have that w^2=2. Note that, by definition, [w] is a sublattice of _24 and it is of (1,1) type. Moreover, by construction, we have w^⊥__24≅^⊕ 3⊕_8(-1)^⊕ 2⊕ [-2]≅σ^⊥_ so sgn(w^⊥__24)=sgn(σ^⊥_)=(3,20). Moreover the projectivity of X implies the positive part of the signature of NS(X)≅^1,1 is equal to 1. Since the class σ has negative square, the positive part of sgn((σ^⊥_)^1,1) is equal to the positive part of the signature of ^1,1, which is 1. Hence, the lattices _24^1,1 and (σ^⊥_)^1,1⊕ [w] have the same signature: in particular the positive part of sgn(_24^1,1) is equal to 2. By hypothesis _24^1,1≅⊕T for some even lattice T of signature (1,-). By Torelli theorem for K3 surfaces, there exists a K3 surface S such that NS(S)≅T. Setting v=2w, we can always choose a positive generator θ which is a v-generic polarization on S. Now, by <cit.> since w^2=2 then either w or -w is a positive Mukai vector, hence we may assume that (S,v,θ) is an OLS-triple. By <cit.> there exists a Hodge isometry H^2(,ℤ)≅ v^⊥⊂_24, where the orthogonal complement of v is computed with respect to the Mukai pairing. The singular moduli space admits a symplectic resolution with exceptional divisor E∈ H^2(,ℤ) such that E has divisibility 3 and E^2=-6. We want to show that X is birational to by showing that there exists a Hodge isometry between H^2(X,ℤ) and H^2(,ℤ) In fact, when such an isometry exists, birationality is implied by the maximality of ^2() for type manifolds, see <cit.> and <cit.>. Observe that we have a Hodge isometry α: H^2(,ℤ)⊕ℤ E≅σ^⊥⊕σ and two finite index embeddings H^2(,ℤ)⊕ℤ E i H^2(,ℤ) σ^⊥⊕σ j H^2(X,ℤ). We want to lift the Hodge isometry α to a Hodge isometry β:H^2(,ℤ)≅ H^2(X,ℤ). To do that, we look closer at the embeddings i and j above. Indeed we have that H^2(,ℤ)⊕ℤ≅^⊕ 3⊕_8(-1)^⊕ 2⊕[-2]⊕ [-6] and it embeds in H^2(,ℤ)≅^⊕ 3⊕_8(-1)^⊕ 2⊕_2(-1) via the identity on the first two factor, which correspond to the unimodular part. To define i we need to give a map between the lattices [-2]⊕[-6] and _2(-1). Let {a,b} be a set of generators of [-2]⊕[-6] such that, in this basis, the lattice pairing has associated gram matrix ([ -2 0; 0 -6 ]). Likewise, let {a,b} be generators of _2(-1) such that the lattice pairing has the form ([ -2 1; 1 -2 ]). Then we set i(a) =c+d, i(b) =c-d. In the same way, we can define the embedding j:σ^⊥⊕σ↪ H^2(X,ℤ), by setting a'=α(a) and b'=α(b) and c' and d' so that the associated matrix for the pairing on _2(-1)⊂ H^2(X,ℤ) is ([ -2 1; 1 -2 ]). We can now define β as follows: on the unimodular part we set β to be the identity, while on _2(-1) we have β(2c)=β(i(a+b))=j(α(a+b))=j(a'+b')=j(a')+j(b')=c'+d'+c-d'=2c, thus we can set β(c):=c' and likewise β(d):=d'. Then β is a Hodge isometry by construction and we conclude the proof. § MODULI SPACES OF TWISTED SHEAVES OF O'GRADY 10 TYPE Here we recall basic facts about twisted K3 surfaces and extend the results of <ref>. For more details on twisted K3 surfaces we refer to <cit.>. A twisted K3 surface is a pair (S, α) where S is a K3 surface and α is a Brauer class α∈(S) ≅ H^2(S,𝒪_S^*). Choosing a lift B ∈ H^2(S, ℚ) of α under the natural morphism H^2(S, ℚ) →(S) induced by the exponential sequence allows one to introduce a natural Hodge structure H(S,α,ℤ) of weight two associated with (S,α). As a lattice, this is just H(S, ℤ) ≅ H^0(S,ℤ)⊕ H^2(S,ℤ) ⊕ H^4(S,ℤ), but the (2, 0) - part is now given by H^2,0 (S,α, ℤ) ℂ(ω + ω∧ B) where 0 ≠ω∈ H^2,0(S). This defines a Hodge structure by setting H^0,2(S,α)H^2,0, H^1,1(S, α) (H^2,0(S, α)⊕H^0,2(S,α))^⊥ where orthogonality is with respect to the Mukai pairing. Although the definition depends on the choice of B, the Hodge structures induced by two different lifts B and B' of the same Brauer class α are Hodge isometric <cit.>. We have a well defined notion of coherent sheaf on a twisted K3 surface (S,α), given by replacing identity by α in the cocycle condition, see <cit.>. The twisted version of the derived global Torelli theorem is the statement that the bounded derived categories of twisted coherent sheaves on (S,α) and (S',α') are equivalent if and only if there exists a Hodge isometry H(S, α, ℤ) ≅H(S', α', ℤ) preserving the natural orientation of the four positive directions, see <cit.> and <cit.>: D^b(S,α) ≅ D^b(S', α') ⇔∃ H(S, α, ℤ) ≅H(S', α', ℤ) oriented Hodge isometry. We now want to generalise our criterion to the case of moduli space of sheaves on a twisted K3 surface (S,α). We start by giving the definition of twisted numerical moduli space. Let (X,η) be a projective marked pair of type, where η H^2(X,ℤ)→ is a fixed marking. We say that X is a twisted numerical moduli space of type if there exists a primitive class σ∈^1,1 such that * σ^2=-6 and (σ,)=3, * the Hodge embedding σ^⊥_↪_24 embeds a copy of (n) in _24^1,1 as a direct summand for some positive n∈ℕ. We can prove an analogue version of <ref> for the twisted case. Let (X,η) be a marked pair of type. The following conditions are equivalent: * There exists a twisted K3 surface (S,α), a non-primitive Mukai vector v=2w and a v-generic polarization θ on S such that X is birational to ; * The manifold X is a twisted numerical moduli space. As before, suppose X is birational to a symplectic resolution of a moduli space for some twisted K3 surface S(α), a non-primitive Mukai vector v=2w and a v-generic polarization θ on S. Then we have a Hodge isometry H^2(X,ℤ)≅ H^2(,ℤ), such that H^2(,ℤ) embeds in H^2(,ℤ) as the orthogonal complement of the exceptional divisor E, which is of divisibility 3 and square -6. Moreover, there is a compatible Hodge embedding H^2(,ℤ)↪H(S(α),ℤ)≅_24, where H(S,α,ℤ)≅_24 is endowed with the Hodge structured described in <ref>. Let σ be the image of E under the isometry (<ref>): then σ is of (1,1) type, has divisibility 3 in and σ^2=-6. We want to prove that the Hodge embedding σ^⊥_↪_24 embeds in _24^1,1 a copy of (n) as a direct summand. To this end, observe that the composition σ^⊥→E^⊥≅ H^2(,ℤ)↪H(S(α),ℤ)≅_24≅^⊕ 4⊕_8(-1)^⊕ 2 defines a compatible Hodge embedding. Let B∈ H^2(S,ℚ) be such that exp(B)=α. Note that the vectors n(1,B,B^2) and (0,0,1) generate a copy of (n) in H((S,α),ℤ)^1,1≅_24^1,1 as (<ref>) is a compatible Hodge embedding. For the other direction we know that there exits σ∈^1,1 of square -6 and divisibility 3 such that σ^⊥_↪_24 embeds a copy of (n) in _24^1,1. Let e_n,f_n be the standard isotropic generators of (n). Choosing n minimal, we can assume that e_n= e is primitive in _24. Hence we can complete e to a copy of =⟨ e, f⟩ such that there is an orthogonal decomposition _24≅⊕𝐑. With respect to such decomposition the second basis vector f_n of (n) can be written as f_n=γ+nf+ke, for some γ∈𝐑 and k∈ℤ. Similarly, a generator of the (2,0)-part of the Hodge structure on _24 induced by σ^⊥ is orthogonal to e and hence is of the form ω+le with ω∈𝐑⊗ℂ and l∈ℤ. However, it is also orthogonal to f_n, so we have that γ·ω+nl=0. Setting B=-γn we have that ω+le=ω+B∧ω, where B∧ω stands for (B·ω)e. By the surjectivity of the period map there exists a twisted K3 surface (S,α) with α=exp(B), such that H^2(S,ℤ)≅𝐑, identifying H^2,0(S)=⟨ω⟩. Observe that the above construction defines an Hodge isometry between H(S,α,ℤ) and _24 with the Hodge structure induced by the Hodge embedding σ^⊥↪_24. We consider w ∈H(S,α,θ) such that under the above isometry [w]≅ (σ^⊥_)^⊥__24. By construction w^2=2 and either w or -w is a positive Mukai vector. If we take v=2w we can embed H^2(,ℤ) ≅ v^⊥_H↪_24. Again, note that admits a symplectic resolution whose exceptional divisor E has square -6 and divisibility 3 and [E]^⊥_H^2(M̃)≅ H^2(,ℤ). To construct the isometry between H^2(X,ℤ) and H^2(,ℤ) we proceed as in <ref>. § BIRATIONALITY CRITERIA FOR THE LI-PERTUSI-ZHAO VARIETY We now give an application of <ref> and <ref> in the case of the type variety defined by Li, Pertusi and Zhao associated to a smooth cubic fourfold <cit.>. We first recall briefly the construction and the main properties of this variety. Given a smooth cubic fourfold Y one defines the Kuznetsov component as _Y := ⟨_Y, _Y(1), _Y(2) ⟩^⊥⊂ D^b(Y) which is a triangulated subcategory with many properties in common with the derived category of a K3 surface. Its topological K-theory (_Y) := ⟨ [_Y],[_Y(1)], [_Y(2)] ⟩^⊥⊂(Y) is a lattice endowed with the restriction of the Euler form on (Y). It has a natural pure Hodge structure of weight 2 obtained by pulling back the one from the cohomology of Y, see <cit.>. For any smooth cubic fourfold the lattice (_Y) contains two distinguished Hodge classes defined as λ_i := [_L(i)] for i=1,2, where L is a line on Y and (Y) → (_Y) is the projection. In particular, by <cit.> we have ⟨λ_1,λ_2⟩≅_2 and the Mukai vector v(_Y) → H^*(Y, ℚ) restricts to an isomorphism ⟨λ_1,λ_2⟩^⊥≅ H^4_ (Y, ℤ). In <cit.> Li, Pertusi and Zhao showed that there exists a Bridgeland stability condition σ for which the moduli space X_Y:=M_σ(2(λ_1+λ_2),_Y) of semistable objects in _Y with fixed numerical invariant 2(λ_1+λ_2)∈(_Y) is a singular projective variety. Its singular locus consists of objects in ^2(M_σ(λ_1+λ_2,_Y)) which lie naturally inside the bigger moduli space. The Li-Pertusi-Zhao (LPZ) variety X̃_Y is constructed as a symplectic resolution of singularities of X_Y, obtained by blowing-up X_Y along its singular locus. In particular, X_Y is a Hyperkähler manifold of OG10 type. Crucial property for us is that <cit.> (_Y)⊇⟨λ_1+ λ_2⟩^⊥≅ H^2 (X_Y, ℤ) Let :=H^2(X_Y,ℤ) and let σ∈^1,1 be the class of the exceptional divisor of the blow-up X_Y→ X_Y. By <cit.>, we know that σ^2=-6 and (σ,L)=3. In the following we denote by 𝒞 the moduli space of smooth cubic fourfolds and by 𝒞_d the irreducible Hassett divisor consisting of special cubic fourfolds of discriminant d. These are cubic fourfolds which have a primitively embedded rank 2 lattice K⊂ H^2,2(Y, ℤ) containing the square of the hyperplane section, such that the Gram matrix of K has discriminant d. Recall that 𝒞_d is non-empty if and only if d>6 and d≡ 0,2 6 (see <cit.>). We consider the following two properties for d: [ (**): d divides 2n^2+2n+2 for some n∈ℤ;; (**'): in the prime factorization of d2, primes p ≡ 2,3 appear with even exponents. ] By <cit.>, condition (**) is equivalent to the existence of an associated K3 surface, in the sense that for if Y∈𝒞_d with d satisfying (**), then there exists a polarised K3 surface S with a Hodge isometry H^2_ (S, ℤ) ≃ K^⊥⊂ H^4(Y,ℤ). Similarly, the condition (**') is equivalent to the existence of an associated twisted K3 surface. Also, both conditions can be rephrased in purely lattice theoretic terms (see, <cit.>) as follows. For any primitive vector v∈⟨λ_1,λ_2⟩^⊥_≅_2^⊥_ we define L_v(_2⊕ℤv)^sat to be the saturation of _2⊕ℤv inside (𝒜_Y). By <cit.>, one has d satisfies (**)⇔ ∃ v_d∈_2^⊥ such that L_d:=L_v_d has discriminant d and ↪ L_d primitively, d satisfies (**')⇔ ∃ v_d∈_2^⊥ such that L_d:=L_v_d has discriminant d and (n) ↪ L_d for some n∈ℕ. Let X_Y be an LPZ-variety associated to a smooth cubic fourfold Y. The following conditions are equivalent. * The cubic fourfold Y lies in the Hassett divisor 𝒞_d with d satisfying (**). * The LPZ-variety X_Y is a numerical moduli space. We start from showing (i)⇒ (ii). Given the resolution π:X_Y→ X_Y we denote by σ the class of the exceptional divisor. The pullback π^* embeds H^2(X_Y,ℤ) in H^2(X_Y,ℤ). In particular π^*H^2(X_Y,ℤ)=σ^⊥. We have a Hodge embedding σ^⊥j(𝒜_Y)≅Λ_24 such that j(σ^⊥)=j(π^*H^2(X_Y,ℤ))=⟨λ_1+λ_2⟩^⊥_. Moreover we have a Hodge embedding j':H^4(Y,ℤ)_prim(-1)→(𝒜_Y), where the Hodge structure of H^4(Y,ℤ)_prim has a Tate twist decreasing weight by (-1,-1) and j'(H^4(Y,ℤ)_prim)=⟨λ_1,λ_2⟩^⊥. By hypothesis and (<ref>), there exists v_d∈ H^4(Y,ℤ)_prim of type (2,2) such that the saturation L_d of ⟨λ_1,λ_2,j'(v_d)⟩⊂(𝒜_Y)≅_24 contains a copy of primitively. Moreover, note that by construction L_d is contained in the (1,1)-part. The claim now follows by observing that, since j and j' are compatible Hodge embeddings, the Hodge structure on _24 induced by σ^⊥j_24 is the same as the one induced by the isomorphism _24≅(𝒜_Y). We now show (ii)⇒ (i). Let X_Y be the LPZ variety associated with the cubic fourfold Y. We have the following diagram: ⟨λ_1,λ_2⟩^⊥_≅H^4(Y,ℤ)_(-1) [d, hook] [r, hook] (𝒜_Y) ⟨λ_1+λ_2⟩^⊥_≅H^2(X_Y,ℤ)≅σ^⊥_H^2(X_Y) [r, hook] H^2(X_Y,ℤ) where the horizontal isomorphisms are given by the Hodge embeddings j and j'. In terms of abstract lattices the above diagram reads as: ⟨λ_1,λ_2⟩^⊥__24≅^⊕2⊕_8(-1)^⊕2 ⊕_2(-1) [d, hook] [r, hook] _24 ⟨λ_1+λ_2⟩^⊥__24≅^⊕3⊕_8(-1)^⊕2 ⊕[-2] [r, hook] ^⊕3 ⊕_8(-1)^⊕2 ⊕_2(-1) . By assumption the lattice _24^1,1 contains a copy of as a direct summand. Since the orthogonal complement of σ^⊥_H^2(X_Y)⊂_24 is generated by a positive class of (1,1) type, we have that ((σ^⊥)^1,1)≥ 1, as σ^⊥ must contain a negative class of inside _24^1,1. Moreover σ^⊥=⟨λ_1+λ_2⟩^⊥ contains the (1,1) class λ_1 - λ_2 which is of positive square, hence ((σ^⊥)^1,1)≥ 2. We deduce that ρ(X_Y) ≥ 2 and by the diagrams above ρ(Y)_:= (H^2,2(Y,ℤ)_) ≥ 1, which means that the cubic fourfold Y belongs to a Hassett divisor 𝒞_d for some d ∈ℕ. We now show that d satisfies the (**) condition, i.e. that there exists a primitive vector v_d∈ H^2,2(Y,ℤ)_ associated to d such that L_d:=⟨λ_1, λ_2, v_d ⟩^sat contains primitively a copy of . Note that since ( ⟨λ_1,λ_2 ⟩^⊥__24 )^1,1 is equal to H^2,2(Y, ℤ)_(-1), which is negative definite, then the lattice ( ⟨λ_1,λ_2 ⟩^⊥__24 )^1,1∩ is generated by a negative class that we call v_d. We now need to show that L_d contains a copy of primitively. To do so, observe that the copy of in _24^1,1 intersects the lattice _2 generated by λ_1 and λ_2. In fact _24^1,1⊇⟨λ_1,λ_2⟩⊕ H^2,2(Y,ℤ)_≅_2⊕ H^2,2(Y,ℤ)_. Since H^2,2(Y,ℤ)_ is negative definite, it follows that ∩_2 has rank 1 and it is generated by a positive element. Since is unimodular, the embedding ↪ L_d is primitive. Now, the result by F.Giovenzana, L.Giovenzana and Onorati <cit.> is a straightforward consequence of <ref> and <ref>. Keep the notation as above. The following conditions are equivalent. * The cubic fourfold Y lies in the Hassett divisor 𝒞_d with d satisfying (**). * There exists a K3 surface S, a non-primitive Mukai vector v=2w and a v-generic polarization θ on S such that X_Y is birational to . Let X_Y be an LPZ-variety associated to a smooth cubic fourfold Y. The following conditions are equivalent. * The cubic fourfold Y lies in the Hassett divisor 𝒞_d with d satisfying (**'). * The LPZ-variety X_Y is a twisted numerical moduli space. The proof is identical to that in <ref> provided that one replaces with (n). We are now able to state a result analogous to <ref> in the twisted case, which was conjectured in <cit.> and proved with the additional assumption that the map between X_Y and preserves the stratification of the singular locus <cit.>. Our criterion allows to circumvent this issue by looking just at the lattice theoretic setting. Keep the notation as above. The following conditions are equivalent. * The cubic fourfold Y lies in the Hassett divisor 𝒞_d with d satisfying (**'). * There exists a twisted K3 surface (S,α), a non-primitive Mukai vector v=2w and a v-generic polarization θ on S such that X_Y is birational to . § NUMERICALLY INDUCED BIRATIONAL TRANSFORMATIONS By <ref>, we know that if X is a numerical moduli space there exists a K3 surface S, together with a Mukai vector v and a polarization θ such that X is birational to . We now want to investigate which birational transformations of X come from automorphisms of the K3 surface S. To do that, let us give two preliminary definitions. Then in <ref> we consider symplectic birational involutions and determine the induced ones. Let (X,η) be a marked pair of type and let be G⊂Bir(X) be a finite subgroup of birational transformations of X. The group G is called an induced group of birational transformations if there exists a K3 surface S with an injective group homomorphism i G↪Aut(S), a G-invariant Mukai vector v∈H(S,ℤ)^G and a v-generic polarization θ on S such that the action induced by G on via i coincides with given action of G on X. Let X be a smooth irreducible holomorphic symplectic manifold of type and let η H^2(X,ℤ)→ be a fixed marking. A finite group G⊂Bir(X) of birational transformations is called a numerically induced group of birational transformations if there exists a class σ∈NS(X) such that σ^2=-6 and (σ,)=3 such that: (i) σ is G-invariant; (ii) Given the Hodge embedding σ^⊥↪_24, the induced action of G on _24 is such that the (1,1) part of the invariant lattice (_24)^G contains as a direct summand. Observe that if X admits a finite subgroup G⊆(X) which is numerically induced, then X is automatically a numerical moduli space. In what follows we give all statements assuming G⊆Bir(X). If the statements hold true for G⊆(X), then G will be called an (numerically) induced group of automorphisms. If φ∈(S) is an automorphism of the K3 surface S, v ∈H(S,ℤ) is a φ-invariant Mukai vector on S, and θ is a φ-invariant polarization on S, then φ induces an automorphism on . Moreover the automorphism is numerically induced. From the definition of stability, if a sheaf ℱ is θ-stable then φ^*ℱ is θ-stable as we an see in <cit.>. For this reason we have a regular automorphism φ̂ of . Then the singular locus of is preserved by φ̂ and so we get an automorphism φ̃ of φ̂ on the blow-up . We call σ the class of the exceptional divisor which is fixed by the induced action. Moreover, due to the primitive embedding σ^⊥ = H^2(, ℤ) ↪H(S,ℤ) ≅_24 the induced action of φ on σ^⊥ induces an action on H(S,ℤ) ≅_24 which is the same action induced by φ on H(S,ℤ). Moreover the sublattice H^0 (S,ℤ) ⊕ H^4 (S,ℤ)≅ is preserved by the action of φ̃ and it is contained in the (1, 1) part of the lattice _24. In the assumption of <ref>, we ask that the polarization θ∈(S) is φ-invariant, so that the automorphism of the K3 surface induces an automorphism of . Note that this condition is never verified for a nontrivial symplectic automorphism of S according to the result of <cit.>. On the other hand the case of induced nonsymplectic involutions will be the subject of an upcoming paper <cit.>. If θ is not φ-invariant we show in <ref> that we get at least a birational self map of . Before giving the next proposition, recall that so far we have considered moduli spaces of sheaves on K3 surfaces, referring to Gieseker moduli spaces. In particular we consider a generalization, namely moduli spaces of Bridgeland semistable objects. In particular we recall some results about the birational geometry of these moduli spaces. Fix S a projective K3 surface, Bridgeland constructed stability conditions on D^b(S) and these form a complex manifold denoted by (S). The latter has a wall and chamber decomposition, for which any moduli space of Bridgeland semistable objects with a fixed Mukai vector v=2v_0 of O'Grady type is a singular projective symplectic variety which admits a symplectic resolution of singularities, which is an irreducible holomorphic symplectic manifold of type. For any such moduli space the natural stratification of the singular locus is given by M_v_0⊂^2 M_v_0⊂ M_v. We recall the following result. In the setting above we have: * For any stability condition τ from a distinct chamber of (S) any τ semistable object is Gieseker semistable. We call this the Gieseker chamber. * Let τ, τ'∈(S) be generic and v∈H(S,ℤ) be a Mukai vector of O'Grady type, then there exists a birational map ψ M_v(S,τ) M_v(S,τ'), which is stratum preserving in the sense that ψ is defined on the generic point of each stratum of the singular locus of M_v(S,τ) and maps it to the generic point of the M_v(S,τ'). If φ∈(S) is an automorphism of the K3 surface S, v ∈H(S,ℤ) is a φ-invariant Mukai vector on S, and θ is a polarization on S, then φ induces a birational transformation on . Moreover the birational transformation is numerically induced. Let τ be a Bridgeland stability condition in the Gieseker chamber. If a sheaf ℱ is θ-stable, then it is τ-stable and φ^*ℱ is φ^*(τ)-stable. For this reason we have an isomorphism φ̂ M_v(S,τ) M_v(S,φ^*(τ)) which maps the singular locus isomorphically to the singular locus. Composing φ̂ with the birational stratum preserving transformation M_v (S,φ^*(τ)) M_v(S,τ) provided by <ref>, we get a birational self-transformation φ of M_v(S,τ). As φ preserves the singular locus, it lifts to a birational self-transformation of the smooth irreducible holomorphic symplectic manifold M_v(S,τ)=, that we call again φ̃, which preserves the exceptional divisor, whose class we call σ. Moreover we have the primitive embedding σ^⊥ = H^2(, ℤ) ↪H(S,ℤ) ≅_24 and the induced action of φ on σ^⊥ induces an action on H(S,ℤ) ≅_24 which is the same action induced by φ on H(S,ℤ). Moreover the sublattice H^0 (S,ℤ) ⊕ H^4 (S,ℤ)≅ is preserved by the action of φ̃ and it is contained in the (1, 1) part of the lattice _24. Let (X,η) be a smooth marked pair of type and let G⊂Bir(X) be a finite subgroup. If G is a numerically induced group of birational transformations, then there exists a K3 surface S with an injective group homomorphism G↪Aut(S), a G-invariant Mukai vector v and a v-generic polarization θ on S such that X is birational to and G⊂Bir(X) is an induced group of birational transformations. We first consider the case when G is symplectic. Then the transcendental lattice 𝐓(X) is contained inside ^G. Since G is numerically induced, there exists a class σ∈NS(X) with σ^2=-6 and (σ,)=3 which is G invariant. We consider the Hodge embedding σ^⊥_↪_24 and we call w the generator of the orthogonal complement of the image of σ^⊥_ inside _24. We can extend the action of G on _24 in a way such that the action on the orthogonal complement of σ^⊥_ in _24 is trivial. Note that, by construction, w is of (1,1) type and it is fixed by the induced action of G on _24. Moreover by assumption we have that (_24)^G=⊕T with contained in the (1,1) part of the Hodge structure of (_24)^G. Since σ is G-invariant then _G=(σ^⊥_)=(_24)_G, then _G↪_24 and _G^⊥__24= _24^G=⊕𝐓. The embedding _G↪_24 respects the action of G, i.e. _G=(_24)_G and (_24)^G=(_G)^⊥__24=⊕𝐓, then we have that the embedding _G ↪_24 can be decomposed in _G φ^⊕3⊕_8(-1)^⊕ 2↪_24, where φ is a compatible Hodge embedding and by construction _G^⊥_^⊕3⊕_8(-1)^⊕2 =𝐓. Since _G is of (1,1)-type then φ(_G) ⊆ (^⊕3⊕_8(-1)^⊕2)^1,1. The lattice ^⊕3⊕_8(-1)^⊕2 has an Hodge structure hence, by Torelli for K3 surfaces, there exists a K3 surface S such that H^2(S,ℤ) is Hodge isometric to ^⊕3⊕_8(-1)^⊕ 2. Moreover the action induced by G on H^2(S,ℤ) is a Hodge isometry since by construction H^2(S,ℤ)_G=_G and H^2(S,ℤ)^G=𝐓. Furthermore the induced action of G on H^2(S,ℤ) is a monodromy operator due to the maximality of the monodromy group of a K3 surface in the group of isometries, and the coinvariant lattice H^2(S, ℤ)_G does not contain wall divisors for the K3 surface, i.e. vectors of square -2. In fact the coinvariant lattice H^2(S,ℤ)_G is isometric to the coinvariant lattice _G with respect to the induced action of G⊂(X) on the second integral cohomology lattice of the manifold of type, hence it does not contain prime exceptional divisors, hence by the characterization of prime exceptional divisors for manifolds of type it contains any vectors of square -2. We apply <cit.> and we conclude that G ⊆(S). Since w was invariant by construction, by <ref> we have that X is birational to where v=2w is a Mukai vector on S and v^2=(2w)^2=8. Moreover the induced action of G ⊆(S) on H^2(, ℤ) coincide by construction with the given action of G on H^2(X,ℤ)= which is the action we started with. We deduce that G⊂(X) is an induced group of birational transformations. Now, since σ and w are G-invariant then the bL _G=(^G)^⊥ is isomorphic and Hodge isometric to _G≅ (σ^⊥_)_G≅ (Λ_24)_G. Now, given Λ_24 with the Hodge structure induced by σ^⊥, by Torelli theorem there exists a K3 surface S such that H̃(S,ℤ) is Hodge isometric to Λ_24. Since the bL _G embeds in Λ_24 by (<ref>), it also embeds in H̃(S,ℤ). As a result we can write H̃(S,ℤ) ≅Λ_24⊃(Λ_24)_G⊕ (Λ_24)^G = _G⊕⊕T H̃(S,ℤ) ≅ H^2(S,ℤ)⊕, with both copies of in the (1,1) part. Comparing the two equations, we can define a Hodge embedding _G↪ H^2(S,ℤ) by sending in the first equation in the copy of in the second one, _G to (Λ_24)_G as in (<ref>) and T in its image under the finite index embedding of (Λ_24)_G⊕ (Λ_24)^G inside Λ_24. Since all the maps involved are Hodge isometries, we have that (_G)^⊥_H^2 is Hodge isometric to T. Consider as before w = (σ^⊥_)^⊥_Λ_24=[2] and take v=2w ∈H̃(S,ℤ)≅Λ_24 as Mukai vector on S. By construction, X is a numerical moduli space and by <ref> it is birational to the moduli space . Moreover, the vector v is fixed by the induced action of G on Λ_24. Now, observe that element of G induces an isometry of Hodge structures on H^2(X,ℤ) and hence on v^⊥≅ U^⊕3⊕ E_8(-1)^⊕2⊕ [-2]⊃ H^2(S,ℤ). We are now left with proving that G is an induced group of birational transformation, i.e. that we have an embedding i:G↪ Aut(S) such that the induced action of G on H^2(S,ℤ) via i coincides with the given action of G on H^2(X,ℤ). Let g∈ G and let f_g be the induced isometry on H^2(S,ℤ). Up to changing signs, we have that f_g∈ O^+(H^2(S,ℤ),q)=Mon^2(S): if f_g preservers a Kähler class then by add reference: there exists a unique element h_g∈ Aut(S) such that h_g^*=f_g.[Maybe it is worth adding the definition of h_g^* in the introduction or in a preliminary section.] Now, if G acts in a symplectic way, we have that f_g preserves a Kähler class if and only if the coinvariant bL H^2(S,ℤ)_f_g does not contain any wall divisors for S (i.e. elements in NS(X) of square -2). By addref, when X is of OG_10 type, we have that the square -2 vectors are contained in the set of prime exceptional divisor. Since G⊂ Bir(X), by add ref, we have that _G⊃ H^2(S,ℤ)_f_g does not contain any prime exceptional divisor. As a result f_g preserves a Kähler class and so we have the desired embedding i:G↪ Aut(S). Now the group G⊂ Aut(S) is a group of automorphisms of S and since by construction the Mukai vector v is preserved, then by (versione analoga della prop 3.5 e della remark 3.6 del mio paper per OG10) we have an induced group of birational transformations on . The induced action on the second integral cohomology of is the action we started with. Now we assume that every non-trivial element of G has a non-symplectic action and without loss of generality we can assume that ^G=(X). As before we have that (_24)^G= ⊕T and we can consider the K3 surface associated to the Hodge structure induced on ^⊥⊂_24. Again by <ref> X is birational to a moduli space of sheaves on S. The group G is a group of Hodge isometries of S preserving T=(S): in particular it preserves all Kähler classes and we can conclude as in the symplectic case. Suppose now that G is any numerically induced group of birational transformations and let G_s be its symplectic part. Considering the action of G_s we obtain a K3 surface S as in the first step with G_s⊂(S). We can extend this action to an action of G by applying the second step to the quotient group Ĝ G/G_s. §.§ Induced involutions In this section we give an application of the criterion about induced automorphisms given in <ref>. The lattice-theoretic criterion that we prove allows to determine if a birational transformation of a manifold of type which is at least birational to , is induced by an automorphism of the K3 surface S. More precisely, by <cit.> we know that there are no regular symplectic involutions on a manifold of type, while by <cit.> we have a lattice-theoretic classification of birational symplectic involutions. We consider the classification of symplectic birational involutions on manifolds of type given in <cit.>, where the authors take a manifold of type, a fixed marking η H^2(X,ℤ) → of X, and classify invariant and coinvariant sublattices, denoted by H^2(X,ℤ)_+ ≅^G and H^2(X,ℤ)_- ≅_G respectively. More precisely, they classify the images of symplectic birational involutions with respect to the representation map η_*(X) → O() f ↦η∘ f^*∘η^-1. We prove that there exists a unique induced birational symplectic involution on a manifold of type and we give the following characterization in terms of its lattice-theoretic behaviour. If the action is induced then the action on the discriminant group is trivial. Indeed it is enough to note that a generator of the discriminant group A_ is the class [σ/3] where σ is the class of the exceptional divisor. Let (X, η) be a marked pair of type. Assume that X is a numerical moduli space and let φ∈(X) be a birational symplectic involution. Then φ is induced if and only if H^2(X,ℤ)^φ≅^⊕3⊕_8(-2) ⊕_2(-1) and H^2(X,ℤ)_φ≅_8(-2). If φ is an induced symplectic birational transformation, then by <ref> we have that φ is numerically induced. In <cit.> we have a classification of symplectic birational involutions on manifolds of type. As the action of an induced automorphism on the discriminant is trivial, we deduce that the H^2(X,ℤ)^φ is either ^⊕3⊕_8(-2) ⊕_2(-1) or _6(-2)⊕(2)^⊕2⊕[2]⊕ [-2]. We can exclude the latter thanks to our criterion, let us explain how. Indeed, if the action is induced, then it is numerically induced, in particular there exists a primitive invariant class σ of square -6 and divisibility 3, but this is a contradiction as any element in _6(-2)⊕(2)^⊕2⊕[2]⊕ [-2] has even divisibility. For the converse, let φ be a birational symplectic involution with H^2(X,ℤ)^φ= ^⊕3⊕_8(-2) ⊕_2(-1) and H^2(X,ℤ)_φ≅_8(-2). We first prove that φ is numerically induced. Since X is a numerical moduli space then there exists a class σ∈(X) with σ^2=-6 and (σ,)=3 such that, given the Hodge embedding σ^⊥↪_24, then _24^1,1 contains as a direct summand. Observe that such copy of must be contained in H^2(X,ℤ)^φ as there is no way to realize as a sublattice of _8(-2). Moreover, since H^2(X,ℤ)_φ≅_8(-2) does not contain any divisibility 3 class, then σ must lie in H^2(X,ℤ)^φ i.e. it is invariant and φ is thus numerically induced. The result now follows from <ref>. The general projective manifold of type with an induced birational symplectic involution has the following integral Hodge structure: H^2(X,ℤ)_φ⊆(X) ≅_8(-2) ⊕ [-6] ⊕ (X) ≅_8(-2) ⊕^⊕2⊕ [-2] ⊆ H^2(X,ℤ)^φ. The birational symplectic involution is induced by a Nikulin involution on a K3 surface S such that (S) ≅_8(-2) ⊕ℤ L where L^2=2. Since the involution is symplectic, the Néron–Severi contains the coinvariant lattice. Moreover, by <ref> the involution is numerically induced, so that the Néron–Severi must contain an invariant class σ of square -6 and divisibility 3. By <ref> the invariant (1,1) part of the induced action of φ on _24, with respect to the Hodge embedding σ^⊥↪_24, contains a copy of as a direct summand. Then this copy of must be contained in σ^⊥ thus we get (X)=_8(-2) ⊕ [-6] ⊕. Referring to the construction of <ref> X is birational to a moduli space on a K3 surface S and we have the following Hodge isometry H(S,ℤ)≅_24. The Néron–Severi of S is isometric to the (1,1) part of the orthogonal complement of H^0(S,ℤ)⊕ H^4(S,ℤ)≅ in _24. The claim follows because we have the finite index embedding σ^⊥⊕ [2] ⊂_24, where 2 is of type (1,1) and (σ^⊥)^1,1≅_8(-2) ⊕. It is known that K3 surfaces with this Néron–Severi lattice and such an action have a Nikulin involution and they are described in <cit.>. amsplain

http://arxiv.org/abs/2303.06916v1

20230313083434

Brownian motion with asymptotically normal reflection in unbounded domains: from transience to stability

math.PR

Brownian motion with asymptotically normal reflection]Brownian motion with asymptotically normal reflection in unbounded domains: from transience to stability Department of Statistics, University of Warwick, UK miha.bresar@warwick.ac.uk Department of Statistics, University of Warwick, The Alan Turing Institute, UK a.mijatovic@warwick.ac.uk Department of Mathematical Sciences, Durham University, UK andrew.wade@durham.ac.uk [ Andrew Wade March 30, 2023 ================== We quantify the asymptotic behaviour of multidimensional drifltess diffusions in domains unbounded in a single direction, with asymptotically normal reflections from the boundary. We identify the critical growth/contraction rates of the domain that separate stability, null recurrence and transience. In the stable case we prove existence and uniqueness of the invariant distribution and establish the polynomial rate of decay of its tail. We also establish matching polynomial upper and lower bounds on the rate of convergence to stationarity in total variation. All exponents are explicit in the model parameters that determine the asymptotics of the growth rate of the domain, the interior covariance, and the reflection vector field. Proofs are probabilistic, and use upper and lower tail bounds for additive functionals up to return times to compact sets, for which we develop novel sub/supermartingale criteria, applicable to general continuous semimartingales. Narrowing domains fall outside of the standard literature, in part because boundary local time can accumulate arbitrarily rapidly. Establishing Feller continuity (essential for characterizing stability) thus requires an extension of the usual approach. Our recurrence/transience classification extends previous work on strictly normal reflections, and expands the range of phenomena observed across all dimensions. For all recurrent cases, we provide quantitative information through upper and lower bounds on tails of return times to compact sets (see <cit.> for a short https://www.youtube.com/watch?v=oDDDWdbPx74YouTube video describing the results). Key words: Reflected diffusion; normal reflection; generalized parabolic domains; recurrence; transience; tails of return-times; invariant distribution; convergence rate in total variation. AMS Subject Classification 2020: 60J60 (Primary); 60J25, 60J65 (Secondary). § INTRODUCTION AND MAIN RESULTS We study the asymptotic behaviour of a multidimensional diffusion in an unbounded, generalized parabolic domain, with asymptotically normal reflection from the boundary. Our model includes Brownian motion with normal reflection. We show that the phase transition between recurrence and transience occurs for asymptotically expanding domains. If the domain narrows asymptotically, we identify the phase transition between null and positive recurrence. In the recurrent case we characterise the asymptotic behaviour of the tails of the return times. Moreover, in the positive-recurrent case we prove the existence of the invariant distribution of the reflected diffusion and establish the polynomial rate of decay of its tail. Finally, we establish the polynomial rate of convergence to stationarity by proving matching upper and lower bounds on the total variation distance between the marginal and the invariant distribution (see a short https://www.youtube.com/watch?v=oDDDWdbPx74YouTube video describing these results <cit.>). All the aforementioned phenomena depend on the asymptotic behaviour of the boundary and are exhibited by a normally reflected Brownian motion. In particular, this implies that a normally reflected Brownian motion in an unbounded domain of any dimension (greater than 1) may converge in total variation to its invariant distribution, which has heavy tails. Before stating our results, we briefly describe our setting. For any d∈, define a closed domain := {(x,y) ∈_+×^d: y_d ≤ b(x)} in ^d+1, where ·_d is the standard Euclidean norm on ^d, :=[0,∞) and b: _+ →_+ is a smooth function with b(0)=0 and b> 0 on (0,∞). Let W = (W_t)_t ∈ be a standard Brownian motion in ^d+1 and Σ a matrix-valued function on , taking values in positive-definite square matrices of dimension (d+1). Denote by Σ^1/2 the symmetric square root of Σ and let ϕ: ∂→^d+1 be a vector field on ∂. Let the processes Z = (Z_t)_t ∈ [0,) and L=(L_t)_t∈[0,) with state spaces and , respectively, satisfy the stochastic differential equation (SDE) Z_t = z + ∫_0^t Σ^1/2(Z_s) W_s + ∫_0^t ϕ(Z_s) L_s & L_t = ∫_0^t 1{Z_s ∈∂} L_s, t ∈ [0,), where ∈ [0,∞] is a possibly finite explosion time and L denotes the local time process of Z at the boundary ∂. For any starting point Z_0=z∈, by <cit.>, SDE (<ref>) has a unique strong solution (with law denoted by _z) under the assumptions in ass:covariance1, ass:domain1 and ass:vector1 stated in Section <ref> below. Informally, ass:covariance1 requires Σ to be bounded, Lipschitz and uniformly elliptic, ass:domain1 requires regularity of b at zero to make the domain smooth, and ass:vector1 stipulates that the smooth bounded vector field ϕ points into the interior of . Unlike in the case of oblique reflection <cit.>, asymptotically normal reflection does not exhibit explosions. Theorem <ref>, stated and proved below, asserts that =∞ _z-a.s. for all starting points z∈, which we assume in the remainder of the introduction. §.§ The main results The asymptotic growth of the domain is described by the parameter β := lim sup_x →∞x b'(x)/b(x), typically equal to the limit lim_x →∞x b'(x)/b(x) (e.g. if b(x) = ax^β for x ≥ x_0 > 0 and some a > 0, or, more generally, if b is regularly varying and b' is eventually monotone <cit.>). For local time in SDE (<ref>) to influence the long-time behaviour of the reflected process Z, we have to assume in ass:domain2 below that β in (<ref>) lies in the interval (-∞,1) (cf. Section <ref> below), making the growth of sublinear (see Remark <ref> below) and, possibly, asymptotically narrowing. Assumption ass:covariance2 below permits Σ to vary smoothly with z =(x,y)∈⊂×^d, but asserts that, as x→∞, the diagonal entry of Σ in the x-direction and the sum of the remaining diagonal entries converge to positive values σ_1^2 and σ_2^2, respectively. Assumption ass:covariance2 ensures that the process does not stop interacting with the boundary far away from the origin. Assumption ass:vector2 specifies linear factors s_0, c_0 ∈ (0,∞), which scale the projections of the vector field ϕ in the x and normal directions, so that, as x→∞, the former projection is asymptotic to s_0 b'(x) while the latter converges to c_0. Since b'(x) → 0 (as β < 1), the assumption ass:vector2 makes the vector field ϕ asymptotically normal. Strictly normal reflection has s_0 = c_0. The reflected process Z is transient (resp. recurrent) if lim_t →∞Z_t_d+1 = ∞ (resp. there exists r_0 ∈ satisfying lim inf_t →∞Z_t_d+1≤ r_0) _z-a.s. The recurrence/transience transition occurs at the critical growth rate of the boundary β_c := c_0σ_1^2/s_0σ_2^2. Suppose that Assumptions ass:domain2, ass:covariance2, ass:vector2 hold and the process Z solves SDE (<ref>). Then the following statements hold for all starting points z∈: (a) if β < β_c, then Z is recurrent; (b) if β > β_c, then Z is transient; (c) if β = β_c and Assumptions ass:vector2plus, ass:domain2plus and ass:covariance2plus are satisfied, then Z is recurrent. Write Z = (X,Y) ∈ in coordinates, so that X_t ∈ for t∈_+. For any r∈, let ς_r := inf{ t ∈: X_t ≤ r} (with convention inf∅ :=∞) be the return time of Z to ∩[0,r]×^d. Define m_c := (1-β/β_c)/2. Note that, if β<β_c, then m_c>0. In this case the following result implies that the return time ς_r is a.s. finite, with m_c being the critical moment exponent. Suppose that Assumptions ass:domain2, ass:covariance2, ass:vector2 hold and the process Z solves SDE (<ref>). Then the following statements hold. (a) If β∈(β_c,1), then for any level r∈(0,∞) and starting point z∈∩(r,∞)×^d, there is positive probability that component X does not reach level r, i.e., _z(ς_r = ∞) > 0. (b) If β∈ (-∞,β_c), then for any >0, there exists x_0∈(0,∞), such that for every r>x_0 and z∈∩(r,∞)×^d, there exist constants C_1,C_2∈(0,∞) for which C_1t^-m_c-≤_z(ς_r≥ t) ≤ C_2t^-m_c+ for all t∈(1,∞). We say that the process Z is positive recurrent if _z[ς_r] < ∞ holds for all z ∈ and sufficiently large r. The process Z is null recurrent if it is recurrent but not positive recurrent. Theorem <ref>(b) implies that Z is positive (resp. null) recurrent if β < -β_c (resp. -β_c<β<β_c). By Theorem <ref>(b), positive recurrence of Z occurs if β < -β_c. In this case, we study properties of the invariant distribution of Z on . Recall that, by <cit.>, the process Z is strong Markov. An invariant distribution π of Z is a probability measure on the Borel σ-algebra ℬ() generated by the open subsets of , satisfying ∫__z [f(Z_t)]π( z)=∫_ f(z)π( z) for all bounded measurable functions f:→ and all t∈. A total variation distance between two probability measures φ_1 and φ_2, defined on ℬ, is given by φ_1-φ_2_TV=sup_B∈ℬ|φ_1(B)-φ_2(B)|. Define M_c := -(1+β/β_c)/2, and note that, in the case β < -β_c, we have M_c>0. Suppose ass:domain2, ass:covariance2, ass:vector2 hold, Z solves SDE (<ref>), and β < -β_c. Then the reflected process Z possesses a unique invariant distribution π. Moreover, for any ϵ>0, the following statements hold: (a) there exist constants c_π, C_π∈(0,∞) such that c_π r^-2M_c-ϵ≤π({z∈:z_d+1≥ r})≤ C_π r^-2M_c+ϵ, for all r∈[1,∞); (b) for any starting point z∈ of Z, there exist constants c_TV,C_TV∈(0,∞) such that c_TV t^-M_c - ϵ≤_z(Z_t∈·)-π_TV≤ C_TVt^-M_c + ϵ, for all t∈[1,∞). The total variation distance between two probability measures P̃ and Q̃ on a σ-algebra of the state space Ω̃ is defined via P̃-Q̃_TV := sup_A ∈|P̃(A)-Q̃(A)|. Theorem <ref> shows that, surprisingly, a reflected Brownian motion Z in ⊂^d+1 with normal reflection (i.e., Z in the class of models with σ_1^2=1, σ_2^2=d, and c_0=s_0, so that β_c=1/d) on an unbounded domain may be polynomially ergodic if the domain narrows sufficiently fast, i.e. β<-1/d. In this case the tail π({z∈:z_d+1>r}) of the invariant distribution π decays asymptotically as r^1+dβ when r→∞. Note that if β<-1/d, the domain has finite volume in ^d+1. However, by modifying either the covariance matrix or the reflection vector field so that β_c<1/d, for any β∈(-1/d,-β_c) we obtain polynomial stability of the reflected Brownian motion in a domain with infinite volume. Theorem <ref>(a) characterises the critical moment of the invariant distribution π of the reflected process Z in a domain with β<-β_c: for any α in [0,2M_c) (resp. (2M_c,∞)), the moment ∫_z_d+1^απ( z) is finite (resp. infinite). Moreover, by Remark <ref> below, for every z∈ we have _z(Z_t∈∂)=0 for Lebesgue almost every t∈. By Theorem <ref>(b), this implies π(∂)=0. The polynomial rate of convergence in total variation of Z_t to stationarity, given in Theorem <ref>(b), is half of the rate of decay of the tail of its stationary measure. Differently put, by Theorem <ref>, we have lim_t →∞log_z(Z_t∈·)-π_TV/log t =1/2lim_r→∞logπ({z∈:z_d+1>r})/log r = -M_c. §.§ Discussion of the main results The fact that ϕ is an asymptotically normal reflection implies that, in the case with β > 0, the process accumulates a positive drift in the horizontal direction when it reaches the boundary. Here, we observe phase transitions between recurrence and transience depending on the model parameters. When β < 0, the process accumulates a negative drift in horizontal direction. In this case, the process is always recurrent, and in some cases the invariant distribution exists. We now comment on the comment on the structure of the proofs and discuss features of the theorems in the previous section. A key step in the proofs of Theorems <ref>, <ref> and <ref> consists of reducing the problem to certain super/submartingale conditions that can be verified. We stress that the processes involved (that turn out to be super/submartingales) in all non-critical cases, covering phenomena from transience to stability, are transformations of the reflected process in SDE (<ref>) via a single parametric family of Lyapunov functions. The class of functions we use are not, and should not be, harmonic because the analysis of the return times and quantitative properties of the invariant distribution and rate of convergence require the presence of a sufficiently strong drift. §.§.§ Positive recurrence Theorem <ref> provides detailed information on the ergodicity of the reflected process Z, with lower bounds matching the upper bounds. To the best of our knowledge, this is the first characterisation of the rate convergence to stationarity in the context of reflecting diffusion, including the ones with drift. Upper bounds abound: for example <cit.> provide upper bounds on the rate of convergence for various reflected diffusions with drift via drift conditions in <cit.> (due to the presence of the drift, the upper bounds in this case are sub-exponential). In contrast, the literature for lower bounds is scarce. Our approach to the lower bounds on the rate of convergence is purely probabilistic. It rests on novel continuous super/submartingale methods, based on Lemma <ref> below, which provide a general setting where the full force of the idea behind <cit.> (see Lemma <ref> in Appendix <ref> below) can be exploited. The approach is robust to the underlying stochastic model and appears to be applicable to a general continuous ergodic Markov processes. The first step in the proof of Theorem <ref> consists of establishing the Feller continuity (Theorem <ref> below) and irreducibility (Proposition <ref> below) for the reflected process Z. The key technical challenge in this step consists of controlling the growth of the local time (cf. Section <ref> below), which requires establishing Feller continuity of the stopped process first (see proof of Theorem <ref> for details). Once Feller continuity and irreducibility of Z have been established, the upper bounds in Theorem <ref> are proved using supermartingale conditions together with the classical convergence results in <cit.>, applicable in the subexponential case. This yields finiteness of moments of the invariant distribution π, which is then translated into upper bounds on the tails in Theorem <ref> via Markov's inequality. The lower bounds in Theorem <ref> require a lower bound on the tail of the invariant distribution π. In contrast to the upper bounds, characterising infinite moments of π alone does not yield a lower bound on the tail of π. In order to circumvent this problem, we give a criterion for ∫_ H π=∞ for any non-decreasing function H (not only polynomial). This criterion is obtained by Lemma <ref>. Once established, the criterion yields lower bounds on the tail of π via elementary methods (see proof of Lemma <ref> below). The lower bound on the rate of convergence in total variation in Theorem <ref> follow from a supermartingale property of a transformed reflected diffusion in (<ref>), the lower bounds on the tail of the stationary distribution π and a general result in <cit.> (see Lemma <ref> in Appendix <ref> below) that converts the tails of the stationary distribution to a lower bound on the convergence rate in total variation. Finally we note that, in Theorem <ref>, the critical case β=-β_c is omitted for brevity. Its analysis would require additional assumptions and a new Lyapunov function, analogous to the ones used in the proof of the critical case of Theorem <ref>(c). We expect that, under appropriate assumptions, such analysis would yield ergodicity of the normally reflected Brownian motion with logarithmic decay in any dimension. §.§.§ Return times If the reflected process is transient, return times to compact sets are not finite almost surely (we will thus discuss Theorem <ref>(a) in Section <ref> below). In the recurrent case, the upper bound on the tail of the return time in Theorem <ref>(b) is established via a supermartingale condition of a transformation of the reflected process, which (via <cit.>) implies the finiteness of the moments of return times. The upper bounds in the theorem then follow by Markov's inequality. The lower bound in Theorem <ref>(b) is established via certain submartingale conditions and lower bounds on the tails of return times to bounded sets in Lemma <ref>. As in the proof of Theorem <ref> discussed above, Lemma <ref> is again critical here (as the infinite moment criterion of <cit.> cannot be applied to obtain the lower bounds on the tail). Note that Lemma <ref> is applicable in the entire recurrent regime. This is key in the proof of Theorem <ref>(b), in contrast to the proof of Theorem <ref>, where Lemma <ref> is applied in the positive-recurrent case only. As our modelling assumptions ass:covariance2, ass:domain2, ass:vector2 are asymptotic (i.e., they only specify the limiting behaviour both of the coefficients of SDE (<ref>), the domain and the reflection vector field as x →∞, see Section <ref> below for details), it is natural for Theorem <ref>(b) to assert the tail behaviour of return times for sufficiently large levels r only. However, with some further work, using an argument based on uniform ellipticity and escape from the boundary, extending the ideas of <cit.> and Section <ref> below from first moment to higher moments estimates, this requirement could be removed. Note also that, by Assumption ass:domain2, the function b is sublinear as x →∞, implying that there exist positive constants c < C, such that c x ≤z_d+1≤ Cx for all z=(x,y)∈∩ (1,∞)×^d. Theorem <ref> may thus be restated for return times of Z=(X,Y), given in terms of Z_d+1, instead of the scalar process X. Recall that for the modulus of the scalar Brownian motion, the critical exponent for return times equals 1/2, with moments of order less (resp. greater) than 1/2 being finite (resp. infinite). By Theorem <ref>(b), for domains with asymptotically increasing (resp. decreasing) boundary function b, i.e. β > 0 (resp. β<0), the critical exponent m_c for the return times of the reflected process satisfies 0<m_c<1/2 (resp. 1/2<m_c). Differently put, the reflected process in an asymptotically expanding (resp. narrowing) domain is, due to the asymptotically positive (resp. negative) projection of the reflection vector field in the x-direction, “less” (resp. “more”) recurrent than the modulus of the scalar Brownian motion. In the case β = 0, the boundary function b may (but need not, see Lemma <ref> below) be asymptotically constant, see the discussion in Remark <ref> below. In this case, the projection of the reflection vector field converges to zero sufficiently fast, so that the critical moment exponent of the return time equals m_c= 1/2, regardless of other model parameters. §.§.§ The recurrence/transience dichotomy Theorem <ref> characterises the recurrence/transience dichotomy for the reflected process defined by the SDE in (<ref>) above. Its proof relies on a generalisation to continuous time, given in Section <ref> below, of the classical Foster–Lyapunov criteria for transience and recurrence. We stress that our approach is purely probabilistic: our criteria are phrased in terms of continuous-time supermartingale conditions. In particular, neither the Markov property nor any explicit knowledge of the infinitesimal characteristics of the process are required (cf. discussion about the approach in <cit.> in Section <ref> below). Both of these features are crucial in the proof of Theorem <ref>. The critical asymptotic growth rate β_c is always positive (see Assumptions ass:covariance2 and ass:vector2). By Theorem <ref>, if β_c ≥ 1, then Z is recurrent for all boundary functions b satisfying ass:domain2. This is for example the case for the two-dimensional Brownian motion (σ_1^2=σ_2^2=1 and d=1) with normal reflection (c_0=s_0) in ⊂^2, as in this case we have β_c = 1. Note that, in any dimension d∈, the critical growth rate β_c may be greater than one, implying recurrence for all boundary functions satisfying ass:domain2. Theorem <ref>(a) strengthens transience of Z, stated in Theorem <ref>(b). Theorem <ref>(a) shows that for any r>0 and z ∈∩(r,∞)×^d, the process Z does not visit the set ∩[0,r]×^d with positive probability (under _z), even though x-r>0 can be arbitrarily small (where z=(x,y)). The proof of Theorem <ref>(b) establishes only that, with positive probability, the process Z does not return to ∩[0,r]×^d after reaching the set ∩(r_1,∞)×^d for a sufficiently large r_1∈(r,∞). As our assumptions are asymptotic, this extension requires proving, using basic analytical techniques, that, with positive probability, the reflected process reaches an arbitrarily high level before visiting a neighbourhood of the origin. §.§ A heuristic To give some intuition behind the main results, we describe a heuristic comparison with reflecting Brownian motion in an interval. We try to estimate how much horizontal drift the process accumulates via reflections by heuristic renewal argument <cit.>. Suppose that the process is at the horizontal position x. Over the short time scale we can approximate the of the vertical coordinate by a reflecting diffusion Y_t/b(x) on the unit sphere 𝕊^d-1, where b(x) = r, for some r∈. This diffusion has zero drift and combined infinitesimal variance of σ_2^2/b(x)^2 along with the vertical reflection of magnitude c_0/b(x). For r∈ define τ_r = inf{t ≥ 0: Y_t_d = r}. Let y_r denote the starting point for the process Y satisfying y_r_d = r. Let _y_r denote the expectation with respect to the process Y started at Y_0 = y_r. Recall that b(x)=r. We apply Itô’s formula for diffusions with reflection to the process r^-2Y_t_d^2. Note that from the infinitesimal variance, it follows that the drift of the process is of magnitude r^-2σ_2^2 and the boundary reflection is at -2r^-1c_0. In the first case we start at y_r/2 and terminate at τ_r, and in the second case we start at y_r and terminate at τ_r/2. We obtain 1 = _y_r/2[r^-2Y_τ_r^2] = 4^-1 + r^-2_y_r/2[∫_0^τ_rσ_2^2 t] = 4^-1 + r^-2σ_2^2_y_r/2[τ_r], and 4^-1 = r^-2_y_r[Y_τ_r/2^2] = 1 + r^-2σ_2^2_y_r[τ_r/2] - 2r^-1 c_0 _y_r[L_τ_r/2]. Clearly L_τ_r = 0 _y_r/2-a.s. From the above equation we obtain _z_r/2[L_τ_r] + _z_r[L_τ_r/2]/_z_r[τ_r/2] + _z_r/2[τ_r] = σ_2^2/2rc_0. Since r = b(x), the expected local time accumulated by Y_t at { y _d= r}, over the complete cycle starting at y_r/2_d = r/2 and ending again at y_r/2_d = r/2, is equal to σ_2^2/2c_0b(x). Recall that X denotes the x-component of the reflected process Z in (<ref>). An informative heuristic argument in <cit.>, based on the renewal theorem, estimates that the average local time accumulates as Suppose the process is at horizontal position x. Over short time-scales, imagine we may approximate the behaviour of the vertical coordinate by a diffusion on the interval [-b(x),b(x)]. This diffusion has zero drift and infinitesimal variance about σ^2. The (vertical) reflection is effectively of magnitude c_0. After a transformation, this is equivalent to unit-magnitude reflection for Brownian motion on [-b(x)/c_0,b(x)/c_0] with variance σ^2/c_0^2. By an heuristic renewal argument similar to <cit.>, this process should accrue boundary local time on average at rate about σ^2/2c_0 b(x). L_t≈σ_2^2/(2c_0b(x)) t, when X_t is at level x. By SDE (<ref>) and the definition of β in (<ref>), the total instantaneous drift of X (when X_t=x) is thus approximately equal to s_0 σ_2^2 b'(x)/(2c_0 b(x)) ≈ s_0 σ_2^2 β/(2c_0 x) for large x∈. Consequently, the large-scale behaviour of the horizontal coordinate X of Z is resembles that of the solution of the SDE X̃_t = s_0 σ_2^2 β/(2c_0 X̃_t) t+σ_1W̃_t for large times t and values X̃_t, where β∈(-∞,1) and W̃ is a one-dimensional Brownian motion. After time-changing the SDE for X̃ by t↦ tσ_1^2, we obtain a Bessel-type process whose drift coefficient is determined by the ratio β/β_c. For β > 0, we have a Bessel process of dimension 2β/β_c and the transition between recurrence and transience in the heuristic matches that of the result in Theorem <ref>. In the case β<0, the standard literature on Langevin diffusions <cit.> implies that the invariant distribution exists if β< -β_c, the tail of invariant distribution decays at the rate x^β/β_c+1 (as x→∞) and the convergence to the invariant distribution is polynomial with the critical exponent (1+β/β_c)/2. Again the heuristic coincides with our results in Theorem <ref>. The definition of the asymptotically normal reflection requires the projection of the vector field ϕ in the x-direction to decay as s_0 b'(x)→0, when x→∞. The heuristic in (<ref>) further motivates this definition: allowing a rate of decay of the projection of ϕ in the x-direction different from that of b'(x) would miss phase transitions and other criticality phenomena. Moreover, heuristic (<ref>) and the definition of β in (<ref>) suggest that, in the super-linear case β>1, the drift of X produced by the reflection at the boundary is asymptotically decays as x^-β for large x ∈. Such a drift is too weak to influence the long-term behaviour of the reflected process, because the process does not interact with the boundary sufficiently often. Finally we note that the heuristic in (<ref>) illustrates the difference between our domain , satisfying ass:domain2, and the smooth domains studied in <cit.>. Recall from <cit.> that the domain with asymptotically narrowing boundary does not satisfy the assumptions of <cit.>. This difference is more than just a technical caveat: one of the crucial features of the domains considered in <cit.> is the availability of bounds, uniform in the starting point, of the increments of expected local time. In contrast, for a narrowing domain (i.e., b(x)→ 0 when x→∞), by (<ref>), an increment of local time over a short time period cannot be bounded uniformly in the starting point because its growth is proportional to 1/b(x). We circumvent this issue via localisation, resulting in more involved proofs of fundamental properties such as the Feller continuity (see Theorem <ref>). Moreover, in the case of asymptotically oblique reflection, a narrowing domain may result in an explosive reflected Brownian motion and explosive local time <cit.>, suggesting further that local time in can grow arbitrarily fast. §.§ Related literature The theory of reflecting diffusions began with <cit.>. A large literature is dedicated to reflecting diffusions in bounded domains (see e.g. <cit.>). In the unbounded case, the classical domains are cones (see e.g. <cit.>). Reflecting diffusions have been extensively studied due to their vast applications, including queueing theory <cit.> and mathematical finance <cit.>. Our domain in the case of (strictly) normal reflection has been studied in <cit.>. Theorem <ref> can be viewed as a generalisation of a result in <cit.>, which considers the case where Σ(z) equals the identity matrix and the reflection vector field is given by the unit normal on the entire boundary ∂. In the context of our model, this setting is within the subclass σ_1^2=1, σ_2^2=d and c_0=s_0=1 (recall that these constants specify only the limiting values of Σ(z) and the reflection field ϕ on ∂ as x→∞). By <cit.>, the (d+1)-dimensional Brownian motion with normal reflection has phase transitions between transience and recurrence at β_c = 1/d, cf. (<ref>). In particular, when dimension d is large, recurrence occurs only when the the boundary function grows at rate slower than 1/d, i.e., very slowly. In addition, Theorem <ref>(a) reveals that a (d+1)-dimensional reflected Brownian motion may be recurrent for a boundary function b with growth close to linear if the projection of the reflection vector field ϕ in the x-direction decays as s_0 b'(x)→0, when x→∞, for a sufficiently small constant s_0>0 (note that the projection in the x-direction of the normal reflection decays precisely at the rate b'(x)→0). Moreover, Theorem <ref>(a) and the formula in (<ref>) imply that Z with normal reflection (i.e. s_0=c_0) may be recurrent for the growth rate β (of the boundary function b) arbitrarily close to one, if the instantaneous variance of Z in the x-direction is greater than the sum of the variances in all other directions. A general result in <cit.>, for domains satisfying existence and uniqueness conditions of <cit.>, states that a (d+1)-dimensional Brownian motion with normal reflection is positive recurrent if and only if has finite volume. As in the case of Theorem <ref> above, Theorem <ref> and formula (<ref>) contrast with the result of <cit.>. It is easy to see that the right choice of parameters results in positive (resp. null) recurrence in the domains with infinite (resp. finite) volume, e.g. β<-1/(s_0d) (resp. -1/(s_0d)<β<1/(s_0d)) in the example in the previous paragraph. As discussed in Section <ref> above, some of the technical difficulties in our paper arise due to the fact that our domain with the function b decreasing to zero (e.g. β < 0) violates the smoothness condition of domains in <cit.>. A general framework, via submartingale problems, for a large family of domains that fail to satisfy the assumptions <cit.> was developed in <cit.>. Similar ideas were employed in <cit.> to characterise stationary distribution for large family of reflecting diffusions with piecewise smooth boundaries. Finally we note that the invariant distribution and the rate of convergence of the reflecting Brownian motion with drift have been studied in <cit.>, motivated by applications in particle systems (see Section <ref> above for a brief discussion of the relation between our result and those in <cit.>). The structure of the paper is given in the table of contents above. § MODELLING ASSUMPTIONS AND PRELIMINARY RESULTS §.§ Modelling assumptions For any d ∈ := {1,2,…}, let ·_d denote the Euclidean norm on ^d. Define the closed domain ⊂×^d, where [0,∞), by := { z =(x,y)∈×^d: y _d ≤ b(x)}, where b: → is strictly positive and differentiable on (0,∞) with b(0)=0. Let ∂={z∈:y_d =b(x)} be the boundary of in ^d+1 and denote the unit sphere in ^d by 𝕊^d-1 := { u ∈^d: u _d = 1}. Let e_x := (1,0) ∈𝕊^d⊂×^d denote the unit vector in the x-direction and, for any u ∈𝕊^d-1, define e_u := (0,u) ∈𝕊^d. We can express z=(x,y)∈ as z=xe_x+y_d e_ŷ, where ŷ := y / y _d∈𝕊^d-1 for y _d > 0 (if y _d =0, we may chose ŷ to be any vector in 𝕊^d-1). Recall that the functions Σ: →_d+1^+ and ϕ:∂→^d+1 are the instantaneous variance and the reflection vector field at the boundary of the domain in SDE (<ref>) above. Throughout we denote by _d+1^+ the group of positive definite square matrices of dimension (d+1). ass:domain1(D1) Let b be a continuous function on , with b(0) = 0 and b(x) > 0 for x > 0. Suppose that b is twice continuously differentiable on (0,∞), such that (i) lim inf_x → 0 (b(x)b'(x)) > 0, and (ii) lim_x → 0(b”(x)/b'(x)^3) exists in (-∞,0]. ass:covariance1(C1) Let Σ: →_d+1^+ be bounded, globally Lipschitz, and uniformly elliptic, i.e., there exists δ_Σ > 0 such that, for every u ∈𝕊^d and all z ∈, we have ⟨Σ(z)u,u⟩≥δ_Σ. ass:vector1(V1) Suppose ϕ: ∂→^d+1 is a C^2-vector field, satisfying sup_z ∈∂ϕ(z)_d+1 < ∞ and inf_x > 0inf_ŷ∈𝕊^d-1⟨ϕ(x,b(x) ŷ),n(x,b(x) ŷ) ⟩ > 0, where n(z)=n(x,b(x) ŷ) is the inwards-pointing unit normal vector at z=(x,y)∈∂. Assumption ass:vector1 requires the vector field ϕ to be smooth, of bounded magnitude and have a uniformly positive component in the normal direction (throughout the paper, ⟨·,·⟩ denotes the standard inner product on ^d+1). Assumption ass:domain1 guarantees that the boundary ∂ is sufficiently regular everywhere, including the origin (see <cit.> for details). Assumption ass:covariance1 ensures that the process Z is not locally constrained in any direction. Throughout we use the matrix norm Σ(z)_op := sup_v∈𝕊^dΣ(z) v_d+1. (In particular, since Σ^1/2(z)_op^2 = sup_v ∈𝕊^d⟨Σ(z)v,v⟩=Σ(z)_op equals the largest eigenvalue of Σ, the boundedness of Σ implies the boundedness of Σ^1/2.) Under assumptions ass:vector1, ass:domain1 and ass:covariance1, the process Z may explode with positive probability. In fact, by <cit.>, we may have (<∞)=1. The following additional assumptions preclude explosion (i.e., as we shall see, imply (=∞)=1) and allow us to characterise transience and recurrence of the process Z; see Theorem <ref> above for a detailed statement. ass:domain2(D2) Suppose that ass:domain1 holds, lim_x →∞ b'(x) = lim_x →∞ b”(x) = 0, the limit in (<ref>) exists and satisfies lim_x→∞ xb'(x)/b(x)=β∈ (-∞,1). ass:covariance2(C2) Suppose that ass:covariance1 holds and that there exist σ_1^2,σ_2^2 ∈ (0, ∞) such that ⟨Σ(z)e_x,e_x⟩ = σ_1^2(1+o_(1)) and Σ(z) -σ_1^2 = σ_2^2(1+o_(1)) as x→∞. ass:vector2(V2) Suppose that ass:vector1 holds and that there exist s_0,c_0 ∈ (0,∞) such that ⟨ϕ(z),e_x ⟩ = s_0b'(x)(1+o_∂(1)) and ⟨ϕ(z),-e_ŷ⟩ =c_0(1+o_∂(1)) as x→∞. Here and throughout, for any g:→(0,∞) and ∈{,∂}, o_(g(x)) as x→∞ denotes a function f:→ satisfying lim_x→∞sup_y:(x,y) ∈|f(x,y)|/g(x) = 0. Since Assumption ass:domain2 requires lim_x→∞ xb'(x)/b(x)=β < 1, for any β' ∈ (β,1) we have b(x) < C_0x^β' for some positive constant C_0 and all sufficiently large x ∈, implying that b has sublinear growth as x →∞. Note however that, as x →∞, Assumption ass:domain2 allows b to have any of the following limits: infinity (requiring β≥ 0), a positive finite limit (requiring β = 0) or a limit equal to 0 (requiring β≤ 0). Interestingly, in the case β=0 the boundary function b may exhibit a variety of different behaviours at infinity. For instance, b may grow to infinity (e.g. b(x) ≈log x), converge to 0 (e.g. b(x) ≈ 1/log x), or be asymptotically constant. Furthermore, it is also possible for the function b to oscillate, i.e.  lim sup_x →∞b(x) = ∞ and lim inf_x →∞ b(x) = 0. For example, any function b satisfying b(x) ≈ (1 + (loglog x)^-2 + sinloglog x)loglog x as x→∞, exhibits such behaviour (see Appendix <ref> for proof that such a function satisfies ass:domain2). Assumption ass:covariance2 ensures that the instantaneous covariance of the process Z stabilises at a positive level in the x-direction as x tends to infinity, without taking up all of the volatility of the process (Σ denotes the sum of the diagonal elements of Σ). Since for the unit normal n(x,y) at (x,y)∈∂, the inner products ⟨ n(x,y),e_x⟩ and ⟨ n(x,y),-e_ŷ⟩ are asymptotically equivalent to b'(x) and 1, respectively, as x→∞, Assumption ass:vector2 requires that the vector field ϕ has the same asymptotic behaviour as the unit normal n(x,y) up to positive constants. As we shall see, it is precisely this property that precludes explosions of the process Z and gives rise to the phenomena studied in this paper. Theorem <ref> above gives rise to the critical exponent β_c, defined in (<ref>) above, at which the global behaviour of Z transitions between recurrence and transience. Quantification of the limits in Assumptions ass:vector2, ass:domain2 and ass:covariance2 are required to understand the behaviour of Z if the boundary of the domain grows at the critical rate β=β_c (see (<ref>) above for the link between the growth rate of the boundary and exponent β). ass:domain2plus(D2+) Assume ass:domain1 and that there exists β∈ (0, 1) such that x b'(x) = β b(x)(1+o(b(x)^2x^-2) ) as x →∞. ass:covariance2plus(C2+) Assume ass:covariance1 and that there exist σ_1^2,σ_2^2 ∈ (0, ∞) and ϵ > 0 such that, as x→∞, ⟨Σ(z)e_x,e_x⟩ = σ_1^2(1+o_(x^-)) and Σ(z) - σ_1^2 = σ_2^2(1+o_(x^-ϵ)). ass:vector2plus(V2+) Assume ass:vector1 and that there exist s_0,c_0 ∈ (0,∞) such that, as x→∞, ⟨ϕ(z),e_x ⟩ = s_0b'(x)(1+o_∂(b(x)^2x^-2)) and ⟨ϕ(x,y),-e_ŷ⟩ =c_0(1+o_∂(b(x)^2x^-2)). §.§ Itô's formula for the reflected process and Lyapunov functions We start by noting that under Assumptions ass:covariance1, ass:domain1 and ass:vector1, by <cit.>, SDE (<ref>) has a unique strong solution (Z,L,) for any starting point in a generalised parabolic domain defined in (<ref>). In Section <ref> we prove Theorem <ref> (which shows that =∞, a.s.) and Theorem <ref>. Sections <ref> and <ref> prove Theorems <ref> and <ref>, respectively. A key step in each of these proofs consists of the application of Itô's formula to an appropriate Lyapunov function. More precisely, let f:→ denote a C^2-function on the open domain ∖∂, such that its gradient ∇ f (i.e., the vector of the partial derivatives of f) has a continuous extension to the closed domain (e.g. if f has a C^2-extension to an open set in ^d+1 containing , which is typically the case in applications below). By Itô’s formula <cit.> we obtain f(Z_t) = f(Z_0) + M_t + 1/2∫_0^t Δ_Σ f(Z_s) s + ∫_0^t ⟨∇ f(Z_s),ϕ(Z_s)⟩ L_s, 0≤ t<, where Δ_Σ f := (Σ^1/2H(f)Σ^1/2)=(Σ H(f)) is the Σ-Laplacian of f (recall that the Hessian matrix H(f):→_d+1^+ consists of the second partial derivatives of f). The (scalar) process M and its quadratic variation [M] on the stochastic interval [0,) are given by M_t := ∫_0^t ⟨∇ f(Z_s),Σ^1/2(Z_s) W_s⟩& [M]_t=∫_0^t Σ^1/2(Z_s)∇ f(Z_s)_d+1^2 s,0≤ t<. The strategy of the proofs that follow consists of applying the continuous semimartingale results of Section <ref> below to the process κ=f(Z) for suitable C^2-Lyapunov functions f. In particular, we will use the representation of the quadratic variation [M] in (<ref>) to conclude that the appropriately stopped process M is a true martingale. Pick w ∈∖{0}. Define k_w:=1+ sup_x ∈(|w(1-w)|^1/2b(x)-x). Note that, under ass:domain2, the function b has sublinear growth at infinity (see Remark <ref> above), implying 1 ≤ k_w < ∞. For any z=(x,y)∈ and parameter γ∈ define f_w,1(z):=(x+k_w)(1 + w(1-w)y _d^2/2(x+k_w)^2)^1/w & f_w,γ(z):=f_w,1(z)^γ. Since, for (x,y) ∈, we have y_d ≤ b(x), the definition of k_w implies 1/2≤ 1 + w(1-w)y _d^2/(2(x+k_w)^2)≤ 3/2 for all w∈∖{0}. Thus (x+k_w)2^-1/| w|≤ f_w,1(x,y)≤ (x+k_w)2^1/| w| for all parameter values w∈∖{0}. Moreover, for any (x,y) ∈ we have (x+k_w)^γ2^-|γ|/| w|≤ f_w,γ(x,y)≤ (x+k_w)^γ2^|γ|/| w| for all γ∈. Note that, for γ>0 (resp. γ <0), the function f_w,γ tends to infinity (resp. zero) as x→∞, making it suitable for the application of the results in Section <ref>. Moreover, it is clear that (for all choices of parameters γ and w) f_w,γ is a C^2-function on the open domain ∖∂ and its gradient ∇ f_w,γ has a continuous extension to the closed domain . The Lyapunov function f_w,γ is inspired by a generalisation of a polynomial approximation of the 2-dimensional harmonic function h_w(z) = r^wcos(wθ) (given in polar coordinates z=(r,θ) of the plane), previously used in the analysis of the reflected processes in wedges <cit.>. The following lemma provides asymptotic properties of the relevant derivatives of f_w,γ. Let assumptions ass:vector2, ass:domain2 and ass:covariance2 hold and fix γ∈ and w∈∖{0}. Then Δ_Σ f_w,γ(z) = γ f_w,1(z)^γ-2(σ_1^2(γ-1)+σ_2^2(1-w) +o_(1)) as x→∞. There exists a constant C>0, such that ∇ f_w,γ(z)_d+1^2 ≤ C (x+k_w)^2(γ-1) for all z=(x,y) ∈. Moreover, if γ (s_0 β/c_0 - 1 + w)<0 (resp. >0), then there exists a positive x_0 such that for all z=(x,y) ∈∂∩ [x_0,∞)×^d we have ⟨∇ f_w,γ(z),ϕ(z)⟩ < 0 (resp. > 0). Note that the constants in o_(1), as well as C and x_0, in Lemma <ref> depend on the values of the parameters γ∈ and w∈∖{0}. Denote v(x,y):=y/(x+k_w)∈^d and define the scalar h(z):=1 + w(1-w) v(z)_d^2/2. For any z=(x,y)∈ we have y_d≤ b(x) and, by ass:domain2 (see also Remark <ref> above), b(x)=o(x^β') as x→∞ for any β'∈(β,1). Hence v(z)_d=o_(1) and, for any r∈, h(z)^r = 1 + o_(1) as x→∞. Since h(z)>0 (by (<ref>)) and f_w,1(z)=(x+k_w)h(z)^1/w for any z=(x,y)∈, we obtain ∇ f_w,1(z) =h(z)^1/w( e_x(1-(1-w)v(z)_d^2/h(z)) + e_ŷ (1-w)v(z)_d/h(z)) = (1+ o_(1))(e_x+e_ŷ(1-w)v(z)_d) as x→∞. (See the first paragraph of Section <ref> for the definition of ŷ, e_x and e_ŷ.) Moreover, since v(z)_d=o_(1) as x→∞ and ∇ f_w,γ(z)=γ f_w,γ-1(z)∇ f_w,1(z), by (<ref>) there exists a positive constant C satisfying ∇ f_w,γ(z)_d+1^2 = γ^2f_w,γ-1(z)^2 (1 + o_(1)) ≤ C(x+k_w)^2(γ-1) for all z=(x,y)∈. Recall that f_w,γ-1(z)>0 by (<ref>) for all z∈. Thus, for z∈∂, the signs of ⟨∇ f_w,γ(z),ϕ(z)⟩ and γ⟨∇ f_w,1(z),ϕ(z)⟩ are equal. Assumption ass:vector2 implies ⟨ e_x,ϕ(z)⟩=s_0b'(x)(1+o_∂(1)) and ⟨ e_ŷ,ϕ(z)⟩=-c_0(1 +o_∂(1)) as x→∞. Note that for z=(x,y)∈∂ we have v(z)_d=b(x)/(x+k_w). By (<ref>) we thus obtain γ⟨∇ f_w,1(z),ϕ(z)⟩ = γ (s_0b'(x) -c_0 (1-w)v(z)_d)(1 + o_∂(1)) =γ b(x) (x+k_w)^-1( s_0 (x+k_w) b'(x)/b(x) - c_0 (1-w)) (1+ o_∂(1)) = γ b(x)(x+k_w)^-1( s_0 β - c_0 (1-w) + o_∂(1)) (1+ o_∂(1)) = c_0 b(x)(x+k_w)^-1[γ(s_0 β/c_0 - 1 + w )+ o_∂(1)], where the third equality in the display follows from the definition of β in ass:domain2. By ass:vector2, the model parameter c_0 is positive. Thus for z=(x,y)∈∂ with x sufficiently large, the sign of γ⟨∇ f_w,1(z),ϕ(z)⟩ equals that of γ(s_0 β/c_0 - 1 + w ) as claimed in the lemma. By definition we have f_w,γ(z) = f_w,1(z)^γ. Hence, for any z ∈, the Hessian takes the form H(f_w,γ)(z)=γ(γ-1)f_w,1(z)^γ-2∇ f_w,1(z)(∇ f_w,1(z))^ +γ f_w,1(z)^γ-1H(f_w,1)(z), where (∇ f_w,1(z))^ denotes the (d+1)-dimensional row vector with coordinates given by the first partial derivatives of f_w,1(z). Since Δ_Σ f_w,γ = (Σ^1/2H(f_w,γ)Σ^1/2)=(Σ H(f_w,γ)), for z = (x,y)∈ we have Δ_Σ f_w,γ(z) =γ(γ-1)f_w,1(z)^γ-2⟨Σ(z)∇ f_w,1(z),∇ f_w,1(z)⟩ +γ f_w,1(z)^γ-1Δ_Σ f_w,1(z). By ass:covariance2 and (<ref>), we have ⟨Σ(z)∇ f_w,1(z),∇ f_w,1(z)⟩ =⟨Σ(z)e_x,e_x⟩ + o_(1) = σ_1^2 (1+ o_(1)). Note that ∂_x h(z) = -w(1-w)y_d^2(x+k_w)^-3 and ∂_y_i h(z) = w(1-w)y_i (x+k_w)^-2 for any z=(x,y)∈ (y_i is the i-th coordinate of y∈^d). An elementary (but tedious) calculation, based on the representation f_w,1(z)=(x+k_w)h(z)^1/w, yields ∂_y_i^2f_w,1(z) =(1-w)(x+k_w)^-1h(z)^1/w-1 +o_((x+k_w)^-1) =(1-w)(x+k_w)^-1(1+o_(1)) as x→∞ for every i∈{1,…,d} (recall that h(z)^1/w-1=1+o_(1)). Moreover, all other elements of the Hessian H(f_w,1)(z) are of order o_((x+k_w)^-1) as x→∞. Thus, by definition Δ_Σ f_w,1(z)=(Σ(z)H(f_w,1)(z)) and the fact that Σ is bounded by ass:covariance1 (contained in ass:covariance2) we get Δ_Σ f_w,1(z)=((Σ(z))-⟨Σ(z)e_x,e_x⟩)(1-w)(x+k_w)^-1(1+o_(1)) as x→∞. By Assumption ass:covariance2, it thus follows that Δ_Σ f_w,1(z)=σ_2^2(1-w)(x+k_w)^-1(1+o_(1)) as x→∞. The expression in (<ref>) is now a direct consequence of (<ref>). ∂_x h(z) = -w(1-w)y_d^2 (x+k_w)^-3 ∂_x f_w,1(z) = h(z)^1/w-(1-w)y_d^2(x+k_w)^-2h(z)^1/w-1 ∂_x^2 f_w,1(z) =(1-w)v(z)_d^2(x+k_w)^-1h(z)^1/w-1[1+(1-w)^2v(z)_d^2h(z)^-1] = (1-w)v(z)_d^2(x+k_w)^-1(1+ o_(1)) as x→∞ ∂_y_i∂_x f_w,1(z) = -(1-w)y_i(x+k_w)^-2h(z)^1/w-1[1+(1-w)v(z)_d^2h(z)^-1] ∂_y_i h(z) = w(1-w)y_i (x+k_w)^-2 ∂_y_i f_w,1(z) =(1-w)(x+k_w)^-1y_ih(z)^1/w-1 ∂_x∂_x f_w,1(z) = -(1-w)v(z)_d^2(x+k_w)^-1h(z)^1/w, ∂_x∂_y_i f_w,1(z) = (1-w)y_i(x+k_w)^-2h(z)^1/w-1, ∂_y_i∂_y_jf_w,1(z) =1{i=j}(1-w)(x+k_w)^-1h(z)^1/w-1 + (1-w)^3 y_iy_j(x+k_w)^-3h(z)^1/w-2, we have Δ_Σf_w,1(z) = ((Σ(z))-⟨Σ(z)e_x,e_x⟩)(1-w)(x+k_w)^-1h(z)^1/w-1 + o_((x+k_w)^-1) =σ_2^2(1-w)(x+k_w)^-1(1+o_(1)), where the second equality follows from the assumption ass:covariance2. By (<ref>) can estimate the Δ_Σ f_w,γ(z) for z = (x,y) ∈, Δ_Σ f_w,γ(z) = γ(γ-1)f_w,1(z)^γ-2(σ_1^2 + o_(1))   + γ f_w,1(z)^γ-1(x+k_w)^-1(σ_2^2(1-w) + o_(1)) = γ f_w,1(z)^γ-2(σ_1^2(γ-1)+σ_2^2(1-w) +o_(1)). By Lemma <ref>, the function f_w,γ controls the sign of the inner product ⟨∇ f_w,γ(z),ϕ(z)⟩ for z=(x,y)∈∂ with sufficiently large x∈. Controlling the sign of ⟨∇ f_w,γ(z),ϕ(z)⟩ for all z=(x,y)∈∂ is crucial for analysing the moments of Z_t at a fixed time t (see the proof of Theorem <ref>) as well as establishing drift conditions in the case of positive recurrence (see Lemma <ref> in Section <ref> below). This requires a slight modification of the function f_w,γ, which we now describe. Fix arbitrary x_0,x_1∈(0,∞), satisfying x_0<x_1, and define the function m:×^d→[0,1] as follows: for z=(x,y)∈×^d let m(z) := exp((x_1-x_0)^-2-((x_1-x_0)^2-(x_1-x)^2)^-1)1{x_0<x<x_1}+1{x_1≤ x}. The function m is smooth, ∂_x m(z)≥ 0 for all z∈×^d, and the following holds: for any z_i=(x_i,y)∈×^d, i∈{0,1}, we have m(z_0) = ∂_x m(z_0) = ∂_x^2 m(z_0) = 0 and m(z_1) = 1, ∂_x m(z_1) = ∂_x^2 m(z_1) = 0. For any constant k∈(0,∞), define F_w,γ(z) := f_w,γ(z)m(z)+k(1-m(z)), z∈. The function F_w,γ is clearly a C^2-function on the open domain ∖∂ and its gradient ∇ F_w,γ has a continuous extension to the closed domain (for all parameters γ∈ and w∈∖{0}). Let Assumptions ass:vector2, ass:domain2 and ass:covariance2 hold and fix γ∈ and w∈∖{0}. Then, if γ(β s_0/c_0 -1 + w) < 0, there exist k∈(0,∞) and 0<x_0<x_1 such that ⟨∇ F_w,γ(z),ϕ(z)⟩≤ 0 for all z∈∂. Note that for any constants 0<k<∞ and 0<x_0<x_1, for all z∈ we have ⟨∇ F_w,γ(z),ϕ(z) ⟩ = (f_w,γ(z)-k)∂_x m(z)⟨ e_x,ϕ(z)⟩ + m(z)⟨∇ f_w,γ(z),ϕ(z)⟩. Since γ(β s_0/c_0 -1 + w) < 0, by Lemma <ref>, there exists x_0'>0 such that ⟨∇ f_w,γ(z),ϕ(z)⟩ < 0 for all z∈∂∩ [x_0',∞)×^d. By Assumption ass:vector2, ⟨ϕ(z), e_x⟩ = s_0 b'(x)(1+f(z)) for all z∈∂, where f:∂→ satisfies sup_y:(x,y)∈∂ |f(x,y)|→0 as x→∞. Pick x_0”∈ such that sup_y:(x,y)∈∂ |f(x,y)|<1/2 for all x∈[x_0”,∞). There are two possibilities. (I) There exists x_0∈[max{x_0',x_0”},∞) such that |b'(x_0)|>0. If b'(x_0)>0 (resp. b'(x_0)<0), by the continuity of b', there exists x_1∈(x_0,∞), such that b'(x)>0 (resp. b'(x)<0) for all x∈[x_0,x_1]. Since s_0>0 by Assumption ass:vector2, for any z=(x,y)∈∂ with x∈[x_0,x_1], we have ⟨ϕ(z), e_x⟩ = s_0 b'(x)(1+f(z))>s_0b'(x)/2>0 (resp. ⟨ϕ(z), e_x⟩ = s_0 b'(x)(1+f(z))<s_0b'(x)/2<0). By (<ref>), f_w,γ is a positive bounded function on the set ∩[x_0,x_1]×^d. Thus we may pick k in the interval (sup_(x,y)∈∂:x_0≤ x≤ x_1 f_w,γ(x,y),∞) (resp. (0,inf_(x,y)∈∂:x_0≤ x≤ x_1 f_w,γ(x,y))). Since the function m, defined above, satisfies min{∂_x m(z),m(z)}≥0 for all z∈, ∂_x m(z)=0 for all z=(x,y) with x∈[x_1,∞) and ∂_x m(z)=m(z)=0 for all z=(x,y) with x∈[0,x_0], by (<ref>) it follows ⟨∇ F_w,γ(z),ϕ(z) ⟩≤ 0 for all z∈∂. (II) |b'(x)|=0 for all x∈[max{x_0',x_0”},∞). Thus, for any x_0, x_1∈[max{x_0',x_0”},∞) with x_0<x_1, we have ∂_x m(z)⟨ e_x,ϕ(z)⟩= ∂_x m(z)s_0 b'(x)(1+f(z))=0 for all z∈. Hence, for any k∈(0,∞), by (<ref>) we have ⟨∇ F_w,γ(z),ϕ(z)⟩ = m(z)⟨∇ f_w,γ(z),ϕ(z)⟩≤ 0 for all z∈∂. The function f_w,γ suffices to establish Theorem <ref>(a)–(b) when the asymptotic exponent β is away from the critical value β_c defined in (<ref>). In the critical case β=β_c, logarithmic (rather than polynomial) growth of the Lyapunov function is required. The function we now define for this purpose is inspired by the analysis of the reflecting Brownian motion and random walk in a wedge in <cit.>. Pick an arbitrary constant δ∈(0,∞) and let g_δ:→ (1,∞) be a continuous function, twice differentiable in the interior of , satisfying g_δ(z) = log (x) -x^-δ + σ_1^2/σ_2^2y_d^2/2x^2 (1+δ x^-δ)+1 for z = (x,y) ∈ with x ∈ (,∞), with σ_1^2 and σ_2^2 given in ass:covariance2plus. Since for any (x,y)∈ we have y_d≤ b(x) and b is sublinear by ass:domain2plus (cf. Remark <ref>), there exists a positive constant C_δ∈ (0,∞) such that log x ≤ g_δ(z) ≤ C_δ + log x, for z=(x,y)∈ with x > . The relevant asymptotic properties of the derivatives of g_δ are in the next lemma. Assume β = β_c, where β (resp. β_c) is defined in (<ref>) (resp. (<ref>)). Let Assumptions ass:vector2plus, ass:domain2plus and ass:covariance2plus hold and choose δ∈(0, min{ϵ, 1 - β}), where ϵ > 0 is the rate of decay in Assumption ass:covariance2plus. Then there exists an x_0 > 0 such that Δ_Σ g_δ < 0 on ∩[x_0,∞)×^d and ⟨∇ g_δ,ϕ⟩ < 0 on ∂∩[x_0,∞)×^d. Recall that β=β_c=c_0σ_1^2/(s _0σ_2^2). By (<ref>), for z = (x,y)∈ with x>, we have ∇ g_δ(z) = e_x(x^-1(1+δ x^-δ) -σ_1^2/σ_2^2y_d^2/x^3(1-δ x^-δ(1+δ/2)) ) + e_ŷ(σ_1^2/σ_2^2y_d/x^2(1+δ x^-δ )); see the first paragraph of Section <ref> for the definition of ŷ, e_x and e_ŷ. By ass:vector2plus we have ⟨ e_x,ϕ(z)⟩=s_0b'(x)(1+o_∂(b(x)^2x^-2)) =s_0β b(x)x^-1(1+o_∂(b(x)^2x^-2)) as x→∞, where the second equality follows from ass:domain2plus, and ⟨ e_ŷ,ϕ(z)⟩=-c_0(1+o_∂(b(x)^2x^-2)) as x→∞. Since z = (x,y)∈∂ satisfies y_d =b(x) as x→∞, by (<ref>) we obtain ⟨∇ g_δ(z),ϕ(z)⟩ =s_0β b(x)x^-2 (1+δ x^-δ)o_∂(b(x)^2x^-2) -s_0^2β^2c_0^-1 b(x)^3x^-4(1+o_∂(b(x)^2x^-2)) = -s_0^2c_0^-1β^2 b(x)^3x^-4+ o_∂(b(x)^3x^-4) as x→∞. Thus there exists x_0>0, such that ⟨∇ g_δ(z),ϕ(z)⟩<0 for all z=(x,y)∈∂∩ [x_0,∞)×^d. By Remark <ref> above, for any β'∈(β,1) it holds b(x)=o(x^β') as x→∞, implying y_d/x=o_(1) as x→∞. By assumption δ∈(0,1-β). Thus b(x)=o(x^β+δ), implying b(x)^2x^-4 = o(x^-2-δ) and b(x)x^-3 = o(x^-2-δ) as x→∞. Recall from (<ref>) that, for any i ∈{1,…,d}, the i-th partial derivative ∂_y_i g_δ along the coordinate y_i of y equals ∂_y_i g_δ(z)=β s_0c_0^-1y_i x^-2(1+δ x^-δ) for x>, implying ∂^2_y_i g_δ(z)=β s_0c_0^-1x^-2(1+δ x^-δ). By the representation of the gradient (<ref>) and the fact y_d≤ b(x) for every z=(x,y)∈, we have ∂_x^2 g_δ(z) = -x^-2(1+δ(1+δ)x^-δ)+o_(x^-2-δ), while all mixed derivatives in the Hessian H(g_δ)(z) are of order b(x)x^-3 = o_(x^-2-δ) as x →∞. Thus, by the definition Δ_Σ g_δ(z)=(Σ(z)H(g_δ)(z)) and the fact that ∂^2_y_i g_δ does not depend on the index i ∈{1,…,d} and is bounded for x>, we get Δ_Σ g_δ(z) = ⟨Σ(z)e_x,e_x⟩∂_x^2 g_δ(z) +((Σ(z))-⟨Σ(z)e_x,e_x⟩) ∂^2_y_1 g_δ(z) +o_(x^-2-δ) = -x^-2 (1+δ(1+δ)x^-δ) (σ_1^2 + o_(x^-)) + (σ_2^2+o_(x^-))β s_0c_0^-1 x^-2(1+δ x^-δ) + o_(x^-2-δ) = -σ_1^2δ^2x^-2-δ +o_(x^-2-δ) as x→∞, where the last equality follows form the identity s_0βσ_2^2/c_0=σ_1^2 and the fact δ<ϵ. Hence, the sign of Δ_Σ g_δ(z) is negative for z∈(x,y)∈ with x sufficiently large as claimed in lemma. § NON-EXPLOSION, RECURRENCE/TRANSIENCE CRITERIA, AND RETURN TIMES OF CONTINUOUS SEMIMARTINGALES This section develops certain semimartingale tools for classifying asymptotic behaviour via Foster–Lyapunov criteria. The general theory developed in this section is expected to have broad applicability. In the present paper, it will be applied to study the reflected process Z, given by SDE (<ref>), via the Lyapunov functions constructed and analysed in Section <ref> above. Fix a probability space (Ω,,) and a filtration (_t)_t ∈ satisfying the usual conditions. Consider an (_t)-adapted continuous process κ = (κ_t)_t ∈, taking values in [0,∞]. Let denote the set of all [0,∞]-valued stopping times with respect to (_t)_t∈. For any ℓ,r ∈ and stopping time T ∈, define the first entry times (after T) by λ_ℓ,T := T + inf{s ∈: T < ∞, κ_T+s≤ℓ}, ρ_r,T := T + inf{s ∈: T < ∞,  κ_T+s≥ r}, where we adopt the convention inf∅ := +∞. If T = 0, we write λ_ℓ := λ_ℓ,0 and ρ_r := ρ_r,0. Almost sure limits ρ_∞ := lim_r →∞ρ_r and ρ_∞,T = lim_r →∞ρ_r,T exist by monotonicity. Explosion of the process κ occurs if the event {ρ_∞ < ∞} has positive probability. Since ρ_r,T = ρ_r on the event {T < ρ_r }, we have ρ_∞ = ρ_∞,T on the event { T < ρ_∞}. For r_0≤ r, we define the first exit time from the interval [r_0,r] after some stopping time T ∈ by S_r,T := λ_r_0,T∧ρ_r,T. Here and throughout we denote x∧ y:=min{x,y} and x∨ y:=max{x,y} for any x,y∈[0,∞]. §.§ Non-explosion We first establish criteria for κ not to explode. The main application of this result in the present paper is to prove that the reflected process with asymptotically normal reflection cannot explode. This should be contrasted with the case of the asymptotically oblique reflection, where explosion may occur, see the characterisation in <cit.>. The non-explosion criteria in <cit.> are more delicate than the ones in Lemma <ref>, but require transience of the underlying semimartingale, making them inapplicable to the entire class of processes considered here. The following result is more robust (i.e. with simpler assumptions), has an elementary proof and covers all the models analysed in this paper. Let κ = (κ_t)_t ∈ be an [0,∞]-valued (_t)-adapted continuous process and V: → (0,∞) a continuous function with lim_x →∞ V(x) = ∞. Suppose there exist r_0,η∈, such that for all r ∈ (r_0,∞) and any T ∈, such that [V(κ_T)1{T<ρ_∞}]<∞, the process ζ^T,r = (ζ^T,r_t)_t ∈, defined by ζ^T,r_t := (V(κ_(t+T)∧ S_r,T)-η(t ∧ (S_r,T-T)))1{T < ρ_∞}, is an (_t+T)-supermartingale, i.e., for 0 ≤ s ≤ t < ∞, [ζ^T,r_t|_s + T] ≤ζ^T,r_s , Then (ρ_∞ = ∞) = 1. Furthermore, if for the (possibly random) starting value κ_0, we assume [V(κ_0)]<∞ then [V(κ_t)] < ∞ for all t∈. Choose r_1∈(r_0,∞) and define recursively the upcrossing and downcrossing times of the process κ over the interval [r_0,r_1] as follows: θ_1 := 0, and if θ_k has been defined for some k∈, then θ_k := λ_r_0,θ_k and θ_k+1 := ρ_r_1,θ_k. Thus we have 0=θ_1≤θ_1≤⋯≤θ_k ≤θ_k ≤θ_k+1≤⋯. Moreover, for any t ∈, we denote D(t,r_0,r_1) :=sup{ k ∈; θ_k ≤ t}, the number of downcrossings of the interval [r_0,r_1] up to time t for the process κ. Since any continuous function on the compact interval [0,t] crosses an interval of positive length at most finitely many times, we have D(t,r_0,r_1) < ∞ a.s. Assume now (ρ_∞ < ∞) > 0. Then there exists t_0 ∈ such that (ρ_∞ < t_0) > 0. We will prove by induction that {θ_k<ρ_∞}∩{ρ_∞ < t_0}={ρ_∞ < t_0} a.s. holds for all k ∈. The induction hypothesis holds for k=1 since θ_1=0 < ρ_∞ a.s. Assume the almost-sure equality of events holds for some k∈. Set T := θ_k and note V(κ_T)1{T<ρ_∞}= V(r_1)1{T<ρ_∞}, since the paths of κ are continuous and, on the event {T<ρ_∞}, we have T<∞. In particular, V(κ_T)1{T<ρ_∞} is bounded and hence integrable. Pick any r ∈ (r_1,∞) and consider the supermartingale ζ^T,r defined in (<ref>). Note that (T∨ t_0)-T is a bounded (_T+t)-stopping time since, for any s∈, we have {T ∨ t_0 -T ≤ s} = {t_0 ∨ T≤ T+s }∈_T+s as both t_0∨ T and s+T are (_t)-stopping times. Applying the optional sampling theorem to ζ^T,r at (T∨ t_0)-T yields: (ρ_r,T<λ_r_0,T∧ (T∨ t_0), T<ρ_∞|_T) V(r)- η t_0 ≤[ζ_(T∨ t_0)-T^T,r|_T] ≤ζ^T,r_0 = V(r_1)1{T<ρ_∞}. Taking expectations on both sides, we obtain the following inequality for every r∈(r_1,∞): (ρ_r,T<λ_r_0,T∧ (T∨ t_0), T<ρ_∞)≤ (V(r_1)+ η t_0)/V(r). Thus, by the monotone convergence theorem, we obtain 0≤(ρ_∞≤λ_r_0,T∧ (T∨ t_0),T<ρ_∞ ) = lim_r→∞(ρ_r,T<λ_r_0,T∧ (T∨ t_0),T<ρ_∞ ) ≤lim sup_r→∞(V(r_1)+t_0η)/V(r)=0, implying {T<ρ_∞}= {ρ_∞>λ_r_0,T∧ (T∨ t_0)}∩{T<ρ_∞} a.s. We hence obtain {ρ_∞ < t_0} ={T<ρ_∞}∩{ρ_∞<t_0} = {ρ_∞>λ_r_0,T∧ (T∨ t_0)}∩{T<ρ_∞}∩{ρ_∞<t_0} = {λ_r_0,T<ρ_∞}∩{ρ_∞<t_0} a.s. The first equality in this display holds by the induction hypothesis, the second holds by (<ref>), and the third equality follows from the fact that T≤λ_r_0,T by definition (<ref>). Since the equality {θ_k = λ_r_0,T < ρ_∞}={ρ_r_1,θ_k =θ_k+1 < ρ_∞} holds almost surely, we proved that {θ_k+1<ρ_∞}∩{ρ_∞ < t_0}={ρ_∞ < t_0} almost surely, thus verifying the induction step. We conclude that {θ_1<…<θ_k<ρ_∞<t_0}= ⋂_i=1^k{θ_i < ρ_∞ < t_0} = {ρ_∞ < t_0} a.s. for all k∈. Since {θ_i<θ_i+1}={θ_i<θ_i<θ_i+1} for every i∈ and D(t_0,r_0,r_1) < ∞ a.s., it follows that (ρ_∞ < t_0) = 0, contradicting our assumption (ρ_∞<∞)>0. For any t∈, by the assumptions in the lemma, for any r ∈ (r_0,∞) and any stopping time T∈, satisfying [V(κ_T)1{T<∞}] < ∞, it holds that [V(κ_(t+T) ∧λ_r_0,T∧ρ_r,T)1{T<∞}|_T] ≤η t + V(κ_T)1{T < ∞} a.s. Moreover, since the process κ does not explode we have lim_r→∞ρ_r,T = ∞ a.s. Thus, by Fatou's lemma for conditional expectations, which applies since the function V is positive, and the inequality in (<ref>), we get [V(κ_(t+T)∧λ_r_0,T)1{T<∞}|_T] = [lim inf_r→∞V(κ_(t+T) ∧λ_r_0,T∧ρ_r,T)1{T<∞}|_T] ≤lim inf_r→∞[V(κ_(t+T) ∧λ_r_0,T∧ρ_r,T)1{T<∞}|_T] ≤η t + V(κ_T)1{T < ∞} a.s. To prove that V(κ) is an integrable process, we will again make use of the crossing times over the interval [r_0,r_1], denoted by θ_k,θ_k, for k-th downcrossing and upcrossing times respectively. We pick t∈ and estimate, [V(κ_t)] = [V(κ_t) ∑_k=1^∞( 1{θ_k≤ t ≤θ_k} + 1{θ_k < t <θ_k+1})] ≤sup{V(r):r∈[0,r_1]} + ∑_k=1^∞[V(κ_t) 1{θ_k ≤ t≤θ_k }]. On the event {θ_k ≤ t≤θ_k }, we have V(κ_t) = V(κ_t ∧λ_r_0,θ_k), since λ_r_0,θ_k = θ_k a.s. Moreover, if k = 1, then V(κ_θ_1) = V(κ_0) and if k≥ 2, then V(κ_θ_k) = V(r_1) holds on the event {θ_k ≤ t≤θ_k }∈_θ_k, implying [V(κ_t)^1/21{θ_k ≤ t≤θ_k }] = [[ V(κ_((t-θ_k)+θ_k) ∧λ_r_0,θ_k)^1/21{θ_k ≤ t≤θ_k }|_θ_k]] ≤[[ V(κ_((t-θ_k)+θ_k) ∧λ_r_0,θ_k)|_θ_k] 1{θ_k ≤ t≤θ_k }|_θ_k] ≤[η t1{θ_k < t < θ_k} + V(κ_θ_k)1{θ_k ≤ t≤θ_k }] ≤η t(θ_k ≤ t≤θ_k ) + (V(r_1) + [V(κ_0)] )(θ_k ≤ t ≤θ_k). This inequality, which holds for every k∈, and (<ref>) implies [V(κ_t)] ≤ 2V(r_1) + [V(κ_0)] + η t < ∞ for every t∈(0,∞). §.§ Transience and recurrence criteria for continuous semimartingales Lemmas <ref> and <ref> of the present subsection provide sufficient conditions for recurrence and transience, respectively, for a continuous semimartingale κ. They are continuous-time analogues to the Foster-Lyapunov criteria for discrete-time processes discussed in e.g. <cit.>. Let V: → (0,∞) be a continuous function with lim_x →∞ V(x) = ∞ and κ = (κ_t)_t ∈ an -valued (_t)-adapted continuous process satisfying lim sup_t →∞κ_t = ∞ a.s. If there exists such r_0 > 0, such that for all t_0 ∈ and r ∈ (r_0,∞), the process (V(κ_(t+t_0)∧ S_r,t_0))_t ∈ is an (_t+t_0)-supermartingale (recall that r_0 features in S_r,t_0=λ_r_0,t_0∧ρ_r,t_0 by definition (<ref>)), i.e., [V(κ_t_0)]<∞ and for 0 ≤ s ≤ t < ∞, [V(κ_(t + t_0) ∧ S_r,t_0)|_s + t_0] ≤ V(κ_(s + t_0) ∧ S_r,t_0), then (lim inf_t →∞κ_t ≤ r_0) = 1. Up to requiring verification over a smaller class of stopping times, the hypotheses of Lemma <ref> essentially imply those of Lemma <ref> with η=0. Thus, a Lyapunov function that implies recurrence will also often yield non-explosion. In the case of a transient process, however, η>0 is typically needed for Lemma <ref> to be applicable. Pick t_0 ∈, r ∈ (r_0,∞) and consider the (_t+t_0)-supermartingale (V(κ_(t+t_0)∧ S_r,t_0))_t∈. Note that the assumption lim sup_t →∞κ_t = ∞ a.s. implies S_r,t_0 < ∞ a.s. and hence it holds that lim_t →∞V(κ_(t+t_0) ∧ S_r,t_0) = V(κ_S_r,t_0) a.s. The supermartingale property, Fatou's lemma, definition (<ref>) and the continuity of κ imply ∞>[V(κ_t_0)] ≥ lim inf_t →∞[V(κ_(t + t_0) ∧ S_r,t_0)] ≥[lim inf_t →∞V(κ_(t + t_0) ∧ S_r,t_0)] = [V(κ_S_r,t_0)] ≥(ρ_r,t_0 < λ_r_0,t_0) V(r). Thus, for all r∈(r_0,∞), we obtain (inf_t ≥ t_0κ_t ≤ r_0) ≥(λ_r_0,t_0 < ∞) ≥(λ_r_0,t_0 <ρ_r,t_0) ≥ 1 - [V(κ_t_0)]/V(r). Since lim_r →∞V(r) =∞ by assumption, it follows that (inf_t ≥ t_0κ_t ≤ r_0) = 1 for any fixed t_0 ∈, implying (lim inf_t →∞κ_t ≤ r_0)=(∩_t_0∈{inf_t ≥ t_0κ_t ≤ r_0})=1. Let V: → (0,∞) be a continuous function with lim_x →∞ V(x) = 0 and κ = (κ_t)_t ∈ an -valued (_t)-adapted continuous process satisfying lim sup_t →∞κ_t = ∞ a.s. If there exists r_0 > 0, such that for all T ∈, satisfying T < ∞ a.s., and r ∈ (r_0,∞), the process (V(κ_(t+T)∧ S_r,T))_t ∈, is an (_t+T)-supermartingale, i.e., for 0 ≤ s ≤ t < ∞, [V(κ_(t + T) ∧ S_r,T)|_s + T] ≤ V(κ_(s + T) ∧ S_r,T), then (lim_t →∞κ_t = ∞) = 1. Note that the function V in Lemma <ref> is assumed to be continuous and have limit zero at infinity, making it bounded. Thus, for any T ∈, satisfying T < ∞ a.s., we have [V(κ_T)]<∞. Once we have such a candidate function V, in order to apply Lemma <ref>, we only need to check the (_t+T)-supermartingale property in (<ref>). Pick an arbitrary ℓ∈(r_0,∞). For any r, r_1∈, satisfying ℓ<r_1<r, define T := ρ_r_1. Since lim sup_t →∞κ_t = ∞ a.s., we have T = ρ_r_1<∞ a.s. and ρ_r,T<∞ a.s., implying further by definition (<ref>) that S_r,T=λ_r_0,T∧ρ_r,T≤ρ_r,T<∞ a.s. Define the process ξ = (ξ_t)_t ∈ by ξ_t:= V(κ_(t + T)∧λ_ℓ,T∧ρ_r,T), t ∈. Since λ_ℓ,T∧ρ_r,T≤ S_r,T a.s., the process ξ equals the (_t+T)-supermartingale (V(κ_(t+T)∧ S_r,T))_t ∈ stopped at the (_t+T)-stopping time λ_ℓ,T∧ρ_r,T-T. Thus, by Remark <ref> and <cit.>, the process ξ is a positive (_t+T)-supermartingale. Hence, for all t ∈, we have V(κ_T) = ξ_0 ≥[ξ_t |_T] ≥[ξ_t 1{λ_ℓ,T≤ρ_r,T}|_T]. Since κ is continuous, it holds lim_t→∞ξ_t 1{λ_ℓ,T≤ρ_r,T} =V(ℓ)1{λ_ℓ,T≤ρ_r,T} and V(κ_T)=V(r_1) a.s. The (conditional) Fatou's lemma yields (λ_ℓ,T≤ρ_r,T|_T) V(ℓ) = [lim inf_t →∞ξ_t1{λ_ℓ,T≤ρ_r,T}|_T] ≤lim inf_t →∞[ξ_t1{λ_ℓ,T≤ρ_r,T}|_T]≤ξ_0 = V(κ_T)= V(r_1). Since, by assumption, κ_t∈ a.s. for all t∈, κ does not explode. Thus lim_r →∞ρ_r,T = ∞ a.s. and, since ρ_r,T<∞ a.s. for all r∈(r_1,∞), we get (recall V(ℓ)>0) (λ_ℓ,T < ∞)=(∪_r∈∩(r_1,∞){λ_ℓ,T≤ρ_r,T}) =lim_r→∞(λ_ℓ,T≤ρ_r,T) ≤ V(r_1)/V(ℓ). Recall T=ρ_r_1<∞ a.s. and note {lim inf_t →∞κ_t≤ℓ}⊂{λ_ℓ,ρ_r_1 < ∞} for all r_1∈(ℓ,∞). The inequality (lim inf_t →∞κ_t≤ℓ) ≤(λ_ℓ,ρ_r_1 < ∞) for all r_1∈(ℓ,∞), the upper bound in (<ref>) and the hypotheses V(r_1)→0 as r_1→∞ imply (lim inf_t →∞κ_t≤ℓ) ≤lim sup_r_1 →∞(λ_ℓ,ρ_r_1 < ∞) = 0. Thus lim inf_t →∞κ_t>ℓ a.s. Since ℓ∈(r_0,∞) was arbitrary, transience follows. §.§ Lower bounds on the tails of return times and associated additive functionals In this subsection we establish lower bounds for the tails of the return times of continuous semimartingales and associated additive functionals. In the Markovian setting, there exists a rich theory providing upper bounds for additive functionals considered here, in the context of establishing related upper bounds on the tails of the invariant distribution (see <cit.> and the references therein). Moreover, in the continuous semimartingale setting, <cit.> establishes upper bounds on the return-time moments. Comparatively, the literature dedicated to lower bounds on the tails of the return times is scarce. Some results in this direction can be found in <cit.>, however the assumptions are too restrictive to be used in our model. Our approach, based on the maximal inequality in Proposition <ref> below, is inspired by the discrete-time results in <cit.>. The link between between additive functionals studied in the present subsection and the invariant distributions in the Markovian setting, described in <cit.>, enables the application of Lemma <ref> below in the proofs of lower bounds on the tails of the invariant distribution and the sub-exponential convergence rate in total variation (stated in Theorem <ref>). Lemma <ref> is also crucial for establishing the lower bounds on the tails of return times in Theorem <ref>. Let ξ = (ξ_t)_t∈ be an -valued (_t)-adapted continuous process and f:^2 → a measurable function. For some r > 0, let τ_r:=inf{t∈:ξ_t≥ r} (with inf∅ =∞) and assume (ξ_t∧τ_r- ∫_0^t∧τ_r f(u,ξ_u) u)_t∈ is an (_t)-supermartingale. Then, for any s∈(0,∞), we have (sup_0≤ t ≤ sξ_t≥ r|_0)≤ r^-1(ξ_0 + [∫_0^s∧τ_r f(u,ξ_u) u|_0]). Pick s∈(0,∞) and consider the stopping time τ_r ∧ s, bounded above by s. By assumption we have [ξ_τ_r ∧ s - ∫_0^τ_r ∧ s f(u,ξ_u) u|_0] ≤ξ_0, which implies that [ξ_τ_r ∧ s|_0] ≤ξ_0 + [∫_0^τ_r ∧ s f(u,ξ_u) u |_0]. Moreover, by the definition of τ_r in the proposition we have {τ_r ≤ s}={sup_u∈[0,s]ξ_u≥ r} a.s. Since the equality ξ_τ_r ∧ s = r holds on this event, by (<ref>) we have (sup_0≤ t ≤ sξ_t ≥ r|_0) = r^-1[ξ_τ_r ∧ s1{τ_r ≤ s}|_0] ≤ r^-1(ξ_0 + [∫_0^τ_r ∧ s f(u,ξ_u) u|_0]), implying the proposition. Note that, in the case [∫_0^s∧τ_r f(u,ξ_u) u]=∞, both the statement and the proof of Proposition <ref> are formally correct, but not informative. In particular, when applying Proposition <ref> we need [∫_0^s∧τ_r f(u,ξ_u) u]<∞, which follows easily if, for example, the function f is continuous (and hence bounded on [0,s]×[0,r]). Let κ = (κ_t)_t∈ be an -valued (_t)-adapted continuous process satisfying lim sup_t→∞κ_t =∞ a.s. Suppose that there exist p∈(0,∞), ℓ∈(0,∞), and C ∈ (0,∞), such that the following hold for all r∈(ℓ,∞): * the process (κ_t∧λ_ℓ∧ρ_r^p)_t∈ is an (_t)-submartingale; * for any q∈(0,1) and r_q := (1-q)^-1r, the process (κ_(ρ_r_q+t)∧λ_r,ρ_r_q^-2 - C∫_ρ_r_q^(ρ_r_q + t) ∧λ_r,ρ_r_qκ_u^-4 u)_t∈, is an (_ρ_r_q+t)-supermartingale. Let h:→ be a non-decreasing measurable function. Then for all r∈(ℓ,∞), q∈(0,1) and ∈(0,C^-1q(1-q)], we have (∫_0^λ_ℓ h(κ_s) s ≥ h(r)r^2|_0)≥ qmin{(κ_0^p-ℓ^p) (1-q)^pr^-p, 1}, on the event {κ_0 > ℓ}. In particular, for all t∈(ℓ,∞) (with h≡1 and =C^-1q(1-q)), we have (λ_ℓ≥ t|_0) ≥ qmin{(κ_0^p-ℓ^p)((1-q)^3q/C)^p/2t^-p/2, 1}, on the event {κ_0 > ℓ}. Pick q∈ (0,1), r ∈(ℓ,∞) and note that it suffices to prove the lemma for =C^-1q(1-q). We start by establishing the following inequality: (λ_r,ρ_r_q > ρ_r_q + r^2|_ρ_r_q) ≥ q a.s. Define (ξ_t)_t∈ by ξ_t := κ_ρ_r_q + t^-2 and note that (<ref>) yields τ_r^-2:=inf{t>0:ξ_t ≥ r^-2}=λ_r,ρ_r_q-ρ_r_q. By Assumption <ref>, the process (ξ_t∧τ_r^-2 - ∫_0^t∧τ_r^-2 Cξ_u^2 u)_t∈ is an (_ρ_r_q+t)-supermartingale. By Proposition <ref> applied to ξ and the stopping time τ_r^-2, we obtain (λ_r,ρ_r_q ≤ρ_r_q +t|_ρ_r_q ) = (τ_r^-2≤ t|_ρ_r_q )=(sup_0≤ u ≤ tξ_u ≥ r^-2|_ρ_r_q) ≤ r^2( ξ_0 + [∫_0^t∧τ_r^-2 Cξ_u^2 u|_ρ_r_q] ) = r^2(κ_ρ_r_q^-2 +[ C∫_ρ_r_q^(ρ_r_q+t)∧λ_r,ρ_r_qκ_u^-4 u|_ρ_r_q ]) ≤ r^2(r_q^-2 + Ctr^-4) ≤ (1-q)^2 + Cr^-2t for any t∈(0,∞). It follows that (λ_r,ρ_r_q > ρ_r_q + t) ≥ 1 - ((1-q)^2 + Cr^-2t). Taking t = r^2, we obtain (<ref>). Note that on the event {λ_r,ρ_r_q > ρ_r_q + r^2}, we have h(κ_ρ_r_q + t)≥ h(r) for all t∈[0, r^2] and measurable non-decreasing functions h:→. Since r>ℓ, the inclusion {∫_0^λ_ℓ h(κ_t) t ≥ h(r)r^2 }⊃{ρ_r_q < λ_ℓ}∩{λ_r,ρ_r_q > ρ_r_q + r^2} holds and, by the inequality in (<ref>), we obtain ( ∫_0^λ_ℓ h(κ_t) t ≥ h(r)r^2|_0) ≥[1{ρ_r_q < λ_ℓ}(λ_r,ρ_r_q > ρ_r_q + r^2|_ρ_r) |_0] ≥ q(ρ_r_q <λ_ℓ|_0). By assumption lim sup_t→∞κ_t =∞ a.s. and thus ρ_r_q ∧λ_ℓ<∞ a.s. Since (κ_t∧ρ_r_q∧λ_ℓ^p)_t∈ is a continuous (_t)-submartingale by Assumption <ref>, dominated convergence implies κ_0^p ≤lim_t→∞[κ_t∧λ_ℓ∧ρ_r_q ^p|_0] = [κ_λ_ℓ∧ρ_r_q ^p|_0] ≤ℓ^p + (ρ_r_q <λ_ℓ|_0)r_q^p. On the event {κ_0 > ℓ} we obtain (ρ_r_q<λ_ℓ|_0) ≥ ((κ_0^p-ℓ^p)r_q^-p)∧ 1. Combining this result with (<ref>) implies the general case of the lemma. The special case follows by setting h≡1. § NON-EXPLOSION AND RECURRENCE/TRANSIENCE DICHOTOMY FOR THE REFLECTED PROCESS §.§ Diffusivity and non-explosion under asymptotically normal reflection The non-explosion of Z is essentially due to Assumption ass:vector2, which stipulates that the horizontal projection of the reflection vanishes sufficiently fast when Z_t is far from the origin. In contrast, in the asymptotically oblique case <cit.>, the horizontal projection of the reflection vector field has a strictly positive limit as Z_t_d+1→∞, which in a domain that narrows sufficiently fast, may lead to explosive behaviour τ_ < ∞ a.s. Suppose that Assumptions ass:domain2, ass:covariance2, ass:vector2 hold and let (Z,L) satisfy SDE (<ref>) in the domain on the stochastic interval [0,). For any starting point z∈, the process Z does not explode, i.e. _z(τ_=∞)=1, and the second moment of Z_t is finite and diffusive, i.e., sup_t∈_z Z_t_d+1^2/(1+t)<∞. The finiteness of the first moment of Z_t_d+1, implied by Theorem <ref>, is crucial in the proof of recurrence in Theorem <ref>(a) and Theorem <ref>(c) (see Section <ref> below for details). A minor modification of the final step in the proof of Theorem <ref> would imply that the p-th moment of Z_t is also diffusive for any p∈(0,∞), i.e., sup_t∈_z Z_t_d+1^p/(1+t)^p/2<∞ (see Remark <ref>, after the proof of Theorem <ref> below). Assume ass:vector2, ass:domain2 and ass:covariance2 are satisfied. Consider the process f_w,γ(Z) for any γ∈(-∞,1] and w∈∖{0}, where the function f_w,γ is given in (<ref>). Itô's formula applied to f_w,γ(Z) on the stochastic interval [0,) is given in (<ref>). The quadratic variation of the local martingale M in (<ref>) grows at most linearly in time. Indeed, the representation in (<ref>) and the bound in (<ref>) of Lemma <ref> (recall that γ≤ 1) imply [M]_t ≤∫_0^t Σ(Z_s)_op∇ f_w,γ(Z_s)_d+1^2 s≤ C_γ t, for 0 ≤ t < τ_, for a constant C_γ∈. The inequality in (<ref>) relies on the norm Σ_op=Σ^1/2_op^2 being bounded via Assumption ass:covariance1. Our first task is to prove _z(=∞)=1 for any z∈. Pick w ∈(-∞, 1-β s_0/c_0)∖{0}, ensuring by Lemma <ref> that ⟨∇ f_w,1(z),ϕ(z)⟩ < 0 for all z=(x,y)∈ with sufficiently large x, and define the process κ := f_w,1(Z). Recall the definition of the stopping times ρ_r (for any r∈_+ and T=0) in (<ref>) and ρ_∞=lim_r→∞ρ_r, both given in Section <ref> for the process κ. By (<ref>), the function f_w,1 has linear growth at infinity, implying the equality of the events {<∞}={ρ_∞<∞}. We will apply Lemma <ref> with the identity function V(r) = r for all r ∈ (2^-1/|w|-1,∞) to conclude (ρ_∞<∞)=0. Since γ=1, by (<ref>) in Lemma <ref>, we have Δ_Σ f_w,1(z) = f_w,1(z)^-1σ_2^2(1-w) + o_(1) as x→∞. Thus, by (<ref>), there exists η>0 satisfying |Δ_Σ f_w,1(z)|<η for all z∈. Moreover, by Lemma <ref>, there exists r_0>0 such that ⟨∇ f_w,1(z),ϕ(z)⟩ < 0 for all z=(x,y)∈ satisfying x≥ r_02^-1/|w|-k_w. Note that the upper bound on f_w,1 in (<ref>) implies that any z∈ with f_w,1(z)≥ r_0 must satisfy x≥ r_02^-1/|w|-k_w (and hence ⟨∇ f_w,1(z),ϕ(z)⟩ < 0). For any stopping time T ∈, recall the definition in (<ref>) of the exit time S_r,T of the process κ from the interval (r_0,r) after time T. In order to apply Lemma <ref>, assume that the stopping time T is such that [κ_T1{T < ρ_∞}] < ∞. Then, by Itô's formula in (<ref>) and the choice of the constants η and r_0, the process ζ^T,r=(ζ_t^T,r)_t∈, defined for any t∈ by ζ_t^T,r := (κ_(t+T) ∧ S_r,T - η ( t ∧ ( S_r,T - T ) ) ) 1{T < ρ_∞}, satisfies ζ_t^T,r-ζ_s^T,r≤(M_(t+T) ∧ S_r,T-M_(s+T) ∧ S_r,T)1{T < ρ_∞} a.s. for any 0≤ s≤ t, where M is the local martingale arising in (<ref>) (for the function f_w,1). Since κ is continuous, on the event {T < ρ_∞} it holds that S_r,T < ρ_∞. Thus, by (<ref>), we get [M]_(t+T) ∧ S_r,T-[M]_T ≤ C_1((t+T) ∧ S_r,T-T)≤ C_1t for all t∈, ensuring that (M_(t+T) ∧ S_r,T)_t∈ is a true (_t+T)-martingale and implying [ζ_t^T,r-ζ_s^T,r|_s+T] ≤ 0 for all 0≤ s≤ t. Since ζ^T,r is an (_t+T)-supermartingale for all r∈(r_0,∞), Lemma <ref> yields (ρ_∞=∞)=1. Thus, _z(τ_ = ∞)=1 for all z∈. In the remainder of the section we assume that Z satisfies SDE (<ref>) on the entire time interval . We now prove there exists a constant C_2>0 such that _zZ_t_d+1^2≤ C_2 (t+z_d+1^2+1) holds for any z∈ and t∈. Pick w∈(-∞,1-β s_0/c_0)∖{0}, note β s_0/c_0-1+w<0 and apply Lemma <ref> to find the constants 0<x_0<x_1 and k∈(0,∞) such that the corresponding function F_w,2, defined in (<ref>), satisfies ⟨∇ F_w,2(z),ϕ(z)⟩≤ 0 for all z∈∂. Moreover, by (<ref>), we have Δ_Σ F_w,2(z)=Δ_Σ f_w,2(z) for all z=(x,y)∈∩ [x_1,∞)×^d. Since F_w,2 is smooth on a neighbourhood of and, by (<ref>) in Lemma <ref>, the function z↦|Δ_Σ f_w,2(z)| is bounded on , there exists a constant C_0'∈(0,∞) satisfying |Δ_Σ F_w,2(z)|<C_0' for all z∈. By Assumption ass:domain2, the boundary function b is sublinear (cf. Remark <ref> above): there exist constants C_1',C_2'∈(0,∞) such that b(x)^2<C_1'x^2+C_2' for all x∈. Thus, by the definition of F_w,2 in (<ref>) and the lower bound on f_w,2 in (<ref>), there exist positive constants C_i'∈(0,∞), i∈{3,4,5,6}, satisfying z_d+1^2≤ x^2+b(x)^2≤ (C_1'+1)x^2+C_2'≤ C_3' F_w,2(z)+C_4'≤ C_5' z_d+1^2 +C_6' for all z∈. Recall that the coordinates of Z=(X,Y), taking values in , satisfy X_t∈ and Y_t∈^d for all t∈. Define the passage time of the level r∈ for the process X by ϱ_r := inf{t∈: X_t≥ r}, (with inf∅ =∞). Fix r and assume that the starting point z∈ of the process Z lies in [0,r)×^d or, equivalently, _z(ϱ_r>0)=1. The definition of F_w,2 in the previous paragraph and Itô's formula in (<ref>) applied to the process (F_w,2(Z_t∧ϱ_r))_t∈ yield the following inequalities for all t∈: F_w,2(Z_t∧ϱ_r)≤ F_w,2(z)+ 1/2∫_0^t∧ϱ_rΔ_Σ F_w,2(Z_s) s + M_ϱ_r∧ t+≤ F_w,2(z)+ tC_0' + M_ϱ_r∧ t. Moreover, the quadratic variation of M_ϱ_r∧· in (<ref>) is almost surely bounded and hence integrable, since the gradient ∇ F_w,2 is bounded on compact sets. Thus _zM_t∧ϱ_r=0 for all t,r∈, implying the inequality _z F_w,2 (Z_t∧ϱ_r) ≤ tC_0' + F_w,2(z) for all t,r∈ and z∈∩ [0,r)×^d. Since Z does not explode in finite time and has continuous paths, we have ϱ_r→∞ a.s. as r→∞ and hence Z_t∧ϱ_r→ Z_t a.s. as r→∞. By Fatou's lemma and the inequalities in (<ref>), for all t∈ and z∈, we obtain _zZ_t_d+1^2= _z lim inf_r→∞Z_t∧ϱ_r_d+1^2≤lim inf_r→∞_zZ_t∧ϱ_r_d+1^2 ≤ tC_0'C_3'+ C_5' z_d+1^2 +C_6', concluding the proof of Theorem <ref>. The proof of Theorem <ref> can be modified to get sup_t∈_z Z_t_d+1^2p/(1+t)^p<∞ for any p∈(0,∞). For p∈, we note that F_w,2p can be constructed by Lemma <ref> as above, so that ⟨∇ F_w,2p(z),ϕ(z)⟩≤ 0 for all z∈∂. By (<ref>) in Lemma <ref> there exists C_0'∈(0,∞) satisfying |Δ_Σ F_w,2p(z)|<C_0'F_w,2(p-1)(z) for all z∈. This inequality, combined with the modification of (<ref>) for z_d+1^2p and F_w,2p(z), and induction on p∈ yields the diffusive property of moments for all even powers. The statement for all positive real powers p can be deduced from the even powers via Lyapunov's inequality. As the generalisation of the p=1 is not essential for the development in the present paper, the details are omitted for brevity. §.§ Proof of recurrence/transience classification By Theorem <ref>, in the remainder of the section we may assume without loss of generality that Z satisfies SDE (<ref>) on the entire time interval . The proof of Theorem <ref> starts with a lemma about non-confinement, essential for the applications of Lemmas <ref> and <ref> in the proof of recurrence/transience dichotomy. Under ass:vector1, ass:domain1 and ass:covariance1, the process Z=(X,Y) defined by SDE (<ref>), started at any z∈, satisfies lim sup_t→∞X_t=lim sup_t→∞f_w,1(Z_t)=lim sup_t→∞F_w,1(Z_t)=lim sup_t→∞g_δ(Z_t)=∞ _z-a.s. for any w∈∖{0} and δ∈(0,∞), where the functions f_w,1, F_w,1 and g_δ are given in (<ref>), (<ref>) and (<ref>), respectively. Under ass:vector1, ass:domain1, ass:covariance1, <cit.> implies that the process Z=(X,Y) cannot be confined to a compact set: lim sup_t →∞ X_t = ∞ _z-a.s. for all z∈. In particular, by the inequality in (<ref>) (resp. (<ref>)), for any parameter value w∈∖{0} (resp. δ∈(0,∞)) the process f_w,1(Z) (resp. g_δ(Z)) is also not confined: lim sup_t →∞f_w,1(Z_t) = ∞ (resp. lim sup_t→∞g_δ(Z_t)=∞) a.s. Since, by (<ref>), the functions f_w,1 and F_w,1 coincide on the complement of a neighbourhood of the origin, the lemma follows. Since Assumptions ass:vector1, ass:domain1, ass:covariance1 hold if ass:vector2, ass:domain2, ass:covariance2 are satisfied, we may apply Lemma <ref> in the proofs of this section. (a) Recurrence for β<β_c. The definition of β_c in (<ref>) and the assumption β<β_c imply that β s_0/c_0<σ_1^2/σ_2^2 (recall the the definition of σ_1^2,σ_2^2 and s_0,c_0 in Assumptions ass:covariance2 and ass:vector2, respectively). Pick w∈(1-σ_1^2/σ_2^2,1-β s_0/c_0)∖{0} and note that 0<1-(1-w)σ_2^2/σ_1^2. Choose γ∈(0,min{1,1-(1-w)σ_2^2/σ_1^2}) and observe the inequalities: σ_1^2(γ-1)+σ_2^2(1-w)<0 and γ(s_0β/c_0-1+w)<0. Lemma <ref> and the inequalities in (<ref>) and (<ref>) imply that there exists r_0∈(0,∞) such that for all z=(x,y)∈ with x≥ 2^-1/|w|r_0-k_w we have Δ_Σ f_w,γ(z)<0 and ⟨∇ f_w,γ(z),ϕ(z)⟩ < 0. Consider the process κ = f_w,1(Z) and a continuous function V:→(0,∞), satisfying V(r) = r^γ for all r∈(2^-1/|w|-1,∞). Since, by (<ref>), f_w,1(z)≥ 2^-1/|w| for all z∈, we have V(κ)=f_w,γ(Z). Moreover, by (<ref>) and the fact that 0<γ<1, there exist constants D_1,D_2∈(0,∞) such that the inequality f_w,γ(z) ≤ D_1 z_d+1+D_2 holds for all z∈. Thus, by Theorem <ref>, we get 0<[V(κ_t)]=_z[f_w,γ(Z_t)] ≤ D_1 _z Z_t_d+1+D_2< ∞ for any t∈ and z∈. For any fixed t_0∈ and any r∈(r_0,∞), recall the definition in (<ref>) of the exit time S_r,t_0 of the process κ from the interval (r_0,r) after time t_0. The choice of r_0, the inequalities in (<ref>), and Itô's formula in (<ref>), applied to the process V(κ_(·+t_0) ∧ S_r,t_0)=f_w,γ(Z_(·+t_0) ∧ S_r,t_0) imply f_w,γ(Z_(t+t_0) ∧ S_r,t_0)- f_w,γ(Z_t_0)-(M_(t+t_0) ∧ S_r,t_0-M_t_0)≤ 0 a.s. The local martingale (M_t)_t∈ has integrable quadratic variation by (<ref>), making it a true martingale and implying [V(κ_(t+t_0)∧ S_r,t_0) -V(κ_(s+t_0)∧ S_r,t_0)|ℱ_s+t_0]≤ 0 for all 0 ≤ s≤ t. Since V(κ_t_0∧ S_r,t_0)=V(κ_t_0) is integrable, the process (V(κ_(t+t_0)∧ S_r,t_0))_t∈ is an (_t+t_0)-supermartingale for all r∈(r_0,∞). By Lemma <ref>, we have lim sup_t →∞κ_t = ∞ a.s. Since lim_r →∞V(r) = ∞, we may apply Lemma <ref> to conclude that κ= f_w,1(Z) is recurrent. By (<ref>) and (<ref>), the recurrence of Z follows. (b) Transience for β>β_c. The definition of β_c in (<ref>) and the assumption β>β_c imply that σ_1^2/σ_2^2<β s_0/c_0. Pick w ∈ (1-β s_0/c_0,1-σ_1^2/σ_2^2)∖{0} and note that 1-σ_2^2/σ_1^2(1-w)<0. Choose γ∈ (1-σ_2^2/σ_1^2(1-w),0) and observe the inequalities: γ(σ_1^2(γ-1)+σ_2^2(1-w))<0 and γ(s_0β/c_0-1+w)<0. Lemma <ref>, along with the inequalities in (<ref>) and (<ref>) imply that there exists r_0∈(0,∞) such that for all z=(x,y)∈ with x≥ 2^-1/|w|r_0-k_w we have Δ_Σ f_w,γ(z)<0 and ⟨∇ f_w,γ(z),ϕ(z)⟩ < 0. Consider the process κ = f_w,1(Z) and a continuous function V:→(0,∞), satisfying V(r) = r^γ for all r∈(2^-1/|w|-1,∞). Since, by (<ref>), f_w,1(z)≥ 2^-1/|w| for all z∈, we have V(κ)=f_w,γ(Z). For any stopping time T ∈, satisfying T < ∞ a.s., we have _zV(κ_T)<∞ for all z∈ since the function V is bounded. Pick r∈(r_0,∞) and recall the definition in (<ref>) of the exit time S_r,T of the process κ from the interval (r_0,r) after time T. The choice of r_0 and Itô's formula in (<ref>), applied to the process V(κ_(·+T ∧ S_r,T)=f_w,γ(Z_(·+T) ∧ S_r,T), imply f_w,γ(Z_(t+T) ∧ S_r,T)- f_w,γ(Z_T)-(M_(t+T) ∧ S_r,T-M_T)≤ 0 a.s. Moreover, local martingale (M_t)_t∈ has integrable quadratic variation by (<ref>), making it a true martingale, and for any t∈, [V(κ_t)] < ∞ by the fact that V is bounded. This implies [V(κ_(t+T)∧ S_r,T)-V(κ_(s+T)∧ S_r,T)|_s+T]≤ 0 for all 0 ≤ s ≤ t. Thus, (V(κ_(t+T)∧ S_r,T)_t∈ is an (_t+T)-supermartingale for all r∈(r_0,∞). Since lim sup_t →∞κ_t = ∞ a.s. and lim_r →∞V(r) = 0, Lemma <ref> yields transience of κ=f_w,1(Z). By (<ref>), the transience of Z follows. (c) The critical case β=β_c. Assume β=β_c and ass:domain2plus, ass:vector2plus, ass:covariance2plus. Consider the process κ := g_δ(Z) with g_δ defined in (<ref>) and the parameter δ chosen to satisfy the assumption in Lemma <ref>. Then, by Lemma <ref>, there exists x_0>0 such that the inequalities in (<ref>) hold. Define r_0:=C_δ + log x_0 and note that, by (<ref>), the inequality r_0≥ g_δ(z) (where z=(x,y)) implies x∈[x_0,∞). Set V(r) = r for all r ∈ (1/2,∞) and note V(κ_t) =κ_t= g_δ(Z_t) (recall g_δ > 1 on ). Pick t_0∈ and r∈(r_0,∞) and recall the definition in (<ref>) of the exit time S_r,t_0 of the process κ from the interval (r_0,r) after time t_0. The choice of r_0, the inequalities in (<ref>) and Itô's formula in (<ref>), applied to the process κ_(·+t_0) ∧ S_r,t_0=g_δ(Z_(·+t_0) ∧ S_r,t_0), imply g_δ(Z_(t+t_0) ∧ S_r,t_0)- g_δ(Z_t_0)-(M_(t+t_0) ∧ S_r,t_0-M_t_0)≤ 0 a.s. By continuity, the gradient ∇ g_δ(z)_d+1^2 is bounded on compact sets and Σ_op is bounded by Assumption ass:covariance1. Thus, by the representation in (<ref>), we can bound the quadratic variation [M]_(t+t_0) ∧ S_r,t_0-[M]_t_0≤∫_t_0^(t+t_0) ∧ S_r,t_0Σ(Z_s)_op∇ g_δ(Z_s)_d+1^2ds ≤C̃_1t a.s., where C̃_1∈(0,∞) is a positive constant. Thus the process (M_(t+t_0) ∧ S_r,t_0-M_t_0)_t∈ is a true martingale. Moreover, since g_δ(z) ≤C̃_2z_d+1 + C̃_3 holds for all z∈ for some positive constants C̃_2,C̃_3, Theorem <ref> implies _z[κ_t]=_z[g_δ(Z_t)]≤C̃_2 _ z Z_t_d+1 +C̃_3<∞ for any t∈ and z∈. Thus, [κ_(t+t_0)∧ S_r,t_0 -κ_(s+t_0)∧ S_r,t_0|ℱ_s+t_0]≤ 0 for all 0 ≤ s≤ t, and hence (κ_(t+t_0)∧ S_r,t_0)_t∈ is an (_t+t_0)-supermartingale for all r∈(r_0,∞) and any t_0∈. Moreover, by Lemma <ref>, we have lim sup_t →∞κ_t=∞ a.s. Since lim_r →∞V(r) = ∞, we may apply Lemma <ref> to conclude that κ= g_δ(Z) is recurrent. By (<ref>) and (<ref>), the recurrence of Z follows. § RETURN TIMES, DRIFT CONDITIONS, AND THE PROOF OF THEOREM <REF> The tails of return times are controlled by Propositions <ref> and <ref>, established in this section. The two propositions are crucial in the proof of Theorem <ref>(b); see Remark <ref> below for more details. Moreover, Propositions <ref> is key in obtaining the lower bounds on the tails of the invariant distribution of Z in the positive-recurrent regime: see the proof of Theorem <ref> in Section <ref> below. The drift conditions in Lemma <ref>, proved in the present section, are used for establishing finite moments (and hence upper bounds on the tails) of the invariant distribution and the rate of convergence to stationarity of Z in the positive-recurrent regime (see the proof of Proposition <ref> in Section <ref> below). The common theme of the proofs of the results in this section is that they are all based on the supermartingale property of certain processes. Recall the definitions of the return time ς_r (for r∈(0,∞)) in (<ref>) and of the critical exponent m_c=(1-β/β_c)/2 in (<ref>). Suppose that ass:domain2, ass:covariance2, ass:vector2 hold and β < β_c. Then, for every p ∈(0,m_c), there exists x_0>0 such that for all x_1 ∈(x_0,∞) and z∈ there exists a constant C ∈ (depending only on z and p) for which _z(ς_x_1 > t) ≤ Ct^-p for all t∈(0,∞). Moreover, equivalently, for every p ∈ (0,m_c) and every z ∈, we have _z[ς_x_1^p] < ∞. Pick γ∈ (0,(1-β/β_c)/2). Then we can choose w ∈ (-∞,1-β s_0/c_0)∖{0}, such that σ_1^2(2γ -1) + σ_2^2(1-w)<0. Thus, for any ∈(0,-(σ_1^2(2γ -1) + σ_2^2(1-w))), Lemma <ref> implies the existence of ℓ_0 ∈(0,∞), such that for all z=(x,y)∈ with x > 2^-1/|w|ℓ_0 -k_w (the constant k_w is given above display (<ref>)), the function f_w,2γ defined in (<ref>) satisfies Δ_Σ f_w,2γ(z) + f_w,2γ-2(z) < 0 and ⟨∇ f_w,2γ(z),ϕ(z)⟩ < 0. Having chosen the parameters γ, w, and , consider the process κ = f_w,1(Z). The key step in the proof of the proposition consists of the application of <cit.> to deduce that, for any z∈ with κ_0=f_w,1(z)> ℓ_0, the return time λ_ℓ_0 of κ below the level ℓ_0, defined in (<ref>), has finite γ-moment, i.e., _z[λ_ℓ_0^γ] < ∞. This will hold by <cit.> if we establish that the process (ξ_t)_t∈, given by ξ_t:=κ_t ∧λ_ℓ_0^2γ + ϵ∫_0^t ∧λ_ℓ_0κ_u^2γ - 2 u, is a supermartingale. With this in mind, take an arbitrary r∈(ℓ_0,∞) and consider the stopped process (ξ_t ∧ρ_r)_t∈, where the stopping time ρ_r, defined in (<ref>), is the first time the process κ reaches level r. Since 0≤ξ_t ∧ρ_r≤max{r^2γ,f_w,2γ(z)}+ t max{r^2γ-2,ℓ_0^2γ-2} for all t∈, we have _z[ξ_t ∧ρ_r]<∞. Moreover, by Itô's formula in (<ref>), the inequalities in (<ref>) and the fact that κ_t≥ℓ_0 for all t∈, for any two times 0≤ s ≤ t < ∞ we have ξ_t ∧ρ_r-ξ_s ∧ρ_r-M_t ∧λ_ℓ_0∧ρ_r+M_s ∧λ_ℓ_0∧ρ_r≤ 0 a.s. Since the local martingale M has, by (<ref>) and ass:covariance2, bounded quadratic variation for each t∈ with probability one, the stopped process (ξ_t ∧ρ_r)_t∈ is a supermartingale for any r∈(ℓ_0,∞). By Theorem <ref> we have lim_r→∞ρ_r=∞. Thus ξ_t=lim inf_r→∞ξ_t∧ρ_r for all t∈. Since the process (ξ_t)_t∈ is non-negative, the conditional Fatou lemma implies that it is a supermartingale. Hence, by <cit.>, we have _z[λ_ℓ_0^γ] < ∞. Recall that f_w,1(Z)=κ. Hence the second inequality in (<ref>) implies that, for x_0:= 2^1/|w|ℓ_0-k_w, we have ς_x_0≤λ_ℓ_0 a.s., and thus _z[ς_x_0^γ]<∞. Moreover, for any x_1∈(x_0,∞), we have _z(ς_x_0≥ς_x_1)=1 and hence _z(ς_x_1 > t)≤_z(ς_x_0>t) for all t∈. Since, for any γ∈(0,m_c), Markov's inequality implies _z(ς_x_0>t)≤_z[ς_x_0^γ]/t^γ<∞ for all t∈(0,∞), we have _z(ς_x_1 > t)≤_z(ς_x_0>t) ≤ Ct^-p for all t∈, p∈(0,m_c) and the constant C:=_z[ς_x_0^p]∈(0,∞), independent of x_1∈(x_0,∞). The next proposition provides lower bounds on the tails of return times and related path functionals in the recurrent case. This result is a key ingredient in the proof of Theorems <ref>(b), as well as in the proof of lower bounds in Theorem <ref>. The proof of Proposition <ref> is based on an application of Lemma <ref> of Section <ref> above. Recall m_c=(1-β/β_c)/2 defined in (<ref>). Suppose that ass:domain2, ass:covariance2, ass:vector2 hold with β < β_c. Then, for every p∈(2m_c,∞), there exist x_0∈(0,∞) and constants c_1,c_2∈(1,∞) such that, for every non-decreasing measurable function h:→, q∈(0,1), x_1∈(x_0,∞) and z=(x,y)∈∩ (c_1x_1+c_2,∞)×^d we have _z(∫_0^ς_x_1 h(c_1(X_s + c_2)) s ≥ r^2h(r)) ≥ qmin{(c_1^-p(x-c_2)^p-x_1^p)(1-q)^pr^-p,1}, for all r∈(c_1x_1+c_2,∞) and all sufficiently small >0. In particular, _z(ς_x_1≥ t) ≥ qmin{(c_1^-p(x-c_2)^p-x_1^p)^p/2(1-q)^p t^-p/2,1}, for every t∈(c_1x_1+c_2,∞) and all sufficiently small >0. Propositions <ref> and <ref> provide crucial estimates in the proof of Theorem <ref>(b) in Section <ref> below. The only assertion of Theorem <ref>(b) not contained in Propositions <ref> and <ref> is that the lower bound in Propositions <ref> actually holds for all z∈∩ (x_0,∞)×^d and not only for the starting points z sufficiently far (in the x-direction) from x_0. This generalisation requires uniform ellipticity and will be established in Section <ref>. Pick p∈ (1-β/β_c,∞) and note p>0. Then there exists w ∈(1-β s_0/c_0,∞)∖{0}, such that p > 1 - σ_2^2/σ_1^2(1-w). Lemma <ref> implies that there exist ℓ_0> 0 and a constant C ∈ such that, for all z=(x,y)∈ with x > 2^-1/|w|ℓ_0 -k_w (the constant k_w is defined above display (<ref>)), we have Δ_Σ f_w,p(z) > 0 and ⟨∇ f_w,p(z),ϕ(z)⟩ > 0, Δ_Σ f_w,-2(z) ≤ Cf_w,-4(z) and ⟨∇ f_w_1,-2(z),ϕ(z)⟩ < 0. Define κ := f_w,1(Z) and recall the return time λ_ℓ_0 of κ below the level ℓ_0, defined in (<ref>). By (<ref>), on the event {κ_t=f(Z_t)≥ℓ_0}, the first coordinate X_t of Z_t satisfies X_t > 2^-1/|w|ℓ_0 -k_w. Itô's formula in (<ref>) applied to κ = f_w,1(Z) and the inequalities in (<ref>) imply that, for all r>ℓ≥ℓ_0 and 0≤ s ≤ t < ∞, we have κ_t∧λ_ℓ∧ρ_r ^p - M_t∧λ_ℓ∧ρ_r≥κ_s ∧λ_ℓ∧ρ_r ^p - M_s ∧λ_ℓ∧ρ_r. The process M is a true martingale, since its quadratic variation is bounded by (<ref>) and ass:covariance2. Thus, since _z[κ_t∧λ_ℓ∧ρ_r^p] ≤max{r^p, f_w,p(z)}, the process (κ_t∧λ_ℓ∧ρ_r^p)_t∈ is an (_t)-submartingale. Pick r∈(ℓ_0,∞), q∈(0,1) and set r_q := r/(1-q). Recall that f_w,1(Z_t) = κ and define the process ξ=(ξ_t)_t∈ by ξ_t := κ_(ρ_r_q + t)∧λ_r,ρ_r_q^-2 - C∫_ρ_r_q^t ∧λ_r,ρ_r_qκ_u^-4 u, where the constant C>0 is as in (<ref>). By Itô's formula in (<ref>) and the inequalities in (<ref>), for every 0≤ s ≤ t < ∞, we have ξ_t - ξ_s - (M_(ρ_r_q + t)∧λ_r,ρ_r_q- M_(ρ_r_q + s)∧λ_r,ρ_r_q) ≤ 0 a.s. Since the process ξ is bounded, an analogous argument to the one in the previous paragraph implies that ξ is a supermartingale. We have now proved that Assumptions (a) and (b) of Lemma <ref> are satisfied. Moreover, by Lemma <ref>, we have lim sup_t→∞κ_t=∞ a.s. Hence, by Lemma <ref>, for any z∈ and ℓ>ℓ_0, satisfying f_w,1(z) > ℓ, any non-decreasing measurable function h:→ and all sufficiently small >0, we obtain _z(∫_0^λ_ℓ h(f_w,1(Z_s)) s≥ r^2h(r)) ≥ qmin{(f_w,1(z)^p-ℓ^p)(1-q)^pr^-p,1}, r∈(ℓ,∞). Define x_0:=2^-1/|w|ℓ_0-k_w, c_1:=2^1/|w| and c_2:=k_w. Then, for any x_1∈(x_0,∞) there exists ℓ>ℓ_0, such that x_1=2^-1/|w|ℓ-k_w. The second inequality in (<ref>) implies ς_x_1≥λ_ℓ and h(f_w,1(Z_s))≤ h(2^1/|w|(X_s+k_w)) for all non-decreasing measurable functions h:→ and s∈[0,λ_ℓ]. Thus the inequality ∫_0^λ_ℓ h(f_w,1(Z_s)) s≤∫_0^ς_x_0 h(2^1/|w|(X_s+k_w)) s holds, implying by (<ref>) the inequality in the proposition for all r∈(2^1/|w|x_0+k_w,∞). The special case follows by choosing h≡1. The next result establishes a drift condition (in the positive-recurrent case), used in the proofs of the upper bounds of Theorem <ref> concerning the finite moments of the invariant distribution π of Z and the total variation distance between _z(Z_t∈·) and π. The proof of Lemma <ref> is analogous to the proof of Proposition <ref>. For any r∈, denote ^(r) := ∩ [0,r]×^d. Suppose that ass:covariance2, ass:vector2 and ass:domain2 hold with β < -β_c. Then, for any γ∈(0,1-β/β_c), there exist parameters w ∈(-∞,1-β s_0/c_0)∖{0} and x_0,x_1,k∈, defining the function F_w,γ in (<ref>), and x_2∈, C_1,C_2∈, such that the process ξ = (ξ_t)_t∈, ξ_t := F_w,γ(Z_t) + C_1∫_0^t F_w,γ-2(Z_u) u - C_2∫_0^t1_^(x_2)(Z_u) u, is an (_t)-supermartingale. Note that the process F_w,γ(Z) in Lemma <ref> gets a non-positive push (by Lemma <ref>) when Z hits the boundary ∂. The constant C_1 (resp. C_2) needs to be sufficiently small (resp. large) for the process ξ to have non-positive drift in the entire interior of . Pick γ∈(0,1-β/β_c) and note that 1+σ_1^2/σ_2^2(γ-1)<1-β s_0/c_0, since β_c=c_0σ_1^2/(s_0σ_2^2) by definition (<ref>). Pick w∈(1+σ_1^2/σ_2^2(γ-1),1-β s_0/c_0)∖{0} and note that γ(β s_0/c_0-1+w)< 0. Lemma <ref> implies that there exist x_0,x_1,k∈(0,∞) such that the function F_w,γ defined in (<ref>) satisfies ⟨∇ F_w,γ(z),ϕ(z)⟩ < 0 for all z∈∂. By (<ref>) we have F_w,γ(z)=f_w,γ(z) on z∈∩(x_1,∞)×^d and (<ref>) yields Δ_Σ F_w,γ(z) = γ F_w,1(z)^γ-2(σ_1^2(γ-1)+σ_2^2(1-w) +o_(1)) as x→∞, where o_(1) is defined after Assumption ass:vector2. Since σ_1^2(γ-1)+σ_2^2(1-w)<0, there exists x_2∈(x_1,∞), such that for C_1:=-(σ_1^2(γ-1)+σ_2^2(1-w))γ/4 we have 1/2Δ_Σ F_w,γ(z) + C_1F_w,γ-2(z) ≤ 0 on z ∈∩(x_2,∞)×^d. Thus, since the functions Δ_Σ F_w,γ and Σ are bounded on the compact set _x_2, there exists C_2∈ such that 1/2Δ_Σ F_w,γ(z) + C_1F_w,γ-2(z) ≤ C_21_^(x_2)(z) for all z∈. Recall the definition of ξ in (<ref>) and set κ:=F_w,1(Z). Note that by definition (<ref>), there exist δ_0>0 such that inf_z∈F_w,1(z)>δ_0. For any r>max{1,δ_0}, the stopped process (ξ_t∧ρ_r)_t∈, where the stopping time ρ_r, defined in (<ref>) as the first time the process κ crosses level r, satisfies -C_2 t ≤_z[ξ_t ∧ρ_r] ≤max{F_w,γ(z),r^γ} +C_1 tmax{ r^γ-2,δ_0^γ-2} for all t∈. Thus _z | ξ_t∧ρ_r | <∞ for all t∈ and z∈. Moreover, the inequality (<ref>) and Itô's formula (<ref>) applied to F_w,γ(Z) imply that, for any 0 ≤ s ≤ t<∞, we have ξ_t∧ρ_r-ξ_s∧ρ_r - (M_t∧ρ_r - M_s∧ρ_r) ≤ 0 a.s. Since, by (<ref>) and ass:covariance2, [M]_t∧ρ_r≤ C_0 t a.s. for all t∈ and some constant C_0>0, (ξ_t∧ρ_r)_t∈ is a supermartingale for any r∈. By Theorem <ref> we have lim_r→∞ρ_r = ∞. Thus, ξ_s = lim inf_r→∞ξ_s∧ρ_r for all s∈. Since ξ_s≥ -C_2t for all s∈[0,t], the conditional Fatou lemma implies that for any 0≤ s ≤ t <∞, we have _z[ξ_t |_s] = _z[lim inf_r→∞ξ_t∧ρ_r|_s] ≤lim inf_r→∞_z[ξ_t∧ρ_r|_s] ≤ξ_s, in addition we deduce the integrability of ξ_t by choosing s=0 and noting that _z[ξ_0]= F_w,γ(z), hence ξ is an (_t)-supermartingale. §.§ Proof of Theorem <ref> Recall the definition of the first crossing time ϱ_r (over r) in (<ref>) and the return time ς_r (below r) in (<ref>). We start with the following claim. Claim. For any z=(x,y)∈ and r_0<x<r_1, we have _z(ϱ_r_1<ς_r_0)>0. Proof of Claim. Consider the process started at z = (x,y) ∈ and satisfying r_0<x<r_1. There are two possibilities. (I) z∈∖∂. Fix 0<h'<h<∞. Let _h := {z”∈: ∃ z'∈∂ such that z”-z'_d+1<h} and _r_0,r_1,h := (∖_h) ∩ [r_0,r_1]×^d, such that z∈_r_0,r_1,h, and _r_0,r_1,h⊂_r_0,r_1,h'⊂_r_0,r_1,h', _r_0,r_1,h”∩{r_1}×^d = _r_0,r_1,h'∩{r_1}×^d for some h”∈(h',h), where _r_0,r_1,h' is a closed domain with C^2 boundary. Note that the closed domain _r_0,r_1,h' can be obtained from _r_0,r_1,h' by smoothing its corners appropriately. A Dirichlet problem on _r_0,r_1,h' with boundary condition f: ∂_r_0,r_1,h'→ is given by 1/2Δ_Σu = 0 on _r_0,r_1,h'∖∂_r_0,r_1,h'; u = f, on ∂_r_0,r_1,h'. Choose a continuous function f: ∂_r_0,r_1,h'→, such that f ≡ 1 on _r_0,r_1,h∩{r_1}×^d = ∂_r_0,r_1,h∩{r_1}×^d and f ≡ 0 on ∂_r_0,r_1,h'∖ (_r_0,r_1,h'∩{r_1}×^d ). Then, by <cit.>, the function u(z) := _z[f(Z_τ)], where τ:= inf{t∈:Z_t∈∂_r_0,r_1,h'}, solves the Dirichlet problem in (<ref>)–(<ref>). Moreover, f(Z_τ) ≤1{ϱ_r_1 < ς_r_0}, _z-a.s., implying u(z) ≤_z(ϱ_r_1 < ς_r_0). Since f is continuous, _r_0,r_1,h' has a C^2 boundary and thus satisfies the inside the sphere property (see <cit.> for definition), and coefficients in (<ref>)–(<ref>) are continuous and uniformly elliptic by assumption ass:covariance2, the maximum principle <cit.>, applied to -u, yields u(z) > 0. (II) z∈∂. Define the stopping time v_h := inf{t>0:Z_t∈∖_h}. For the process started at z ∈∂∩ (r_0,r_1)×^d, the continuity of paths implies _z[ς_r_0]>0. Moreover, Assumptions ass:domain2, ass:vector2 and ass:covariance2 imply the conditions of <cit.>, implying that for some δ_Σ>0 and all sufficiently small h > 0, we have _z[v_h ∧ϱ_r_1] ≤ 2h^2/δ_Σ. Hence, there exists h∈(0, (δ_Σ_z[ς_r_0]/2)^1/2) such that _z[v_h ∧ϱ_r_1] < _z[ς_r_0]. In particular, this implies _z(v_h ∧ϱ_r_1 < ς_r_0) > 0. If _z({v_h ∧ϱ_r_1 < ς_r_0}∩{ϱ_r_1 <v_h} ) > 0 the proof is complete. Otherwise, {v_h ∧ϱ_r_1 < ς_r_0} = {v_h ∧ϱ_r_1 < ς_r_0}∩{ϱ_r_1 >v_h}= {v_h < ς_r_0}, _z-a.s. Thus _z(v_h < ς_r_0)=_z(v_h ∧ϱ_r_1 < ς_r_0) > 0. By the strong Markov property at v_h and step (I) of the proof, the claim follows. Proof of (a): in this case we have β > β_c. By Theorem <ref> the process Z is transient. The Lyapunov function f_w,γ(x,y) (with γ<0) tends to zero by (<ref>) as x→∞. Analogous to the proof of Theorem <ref>(b), an application of Lemma <ref> (and in particular the bound in (<ref>) with κ=f_w,1(Z), V(u)=u^γ), implies that for every x_0∈, there exist c(x_0), such that for all z∈∩ [x_0+c(x_0),∞)×^d, we have _z(ς_x_0 = ∞) > 0. Moreover, the Claim implies that for all z∈∩ (x_0,∞)×^d we have _z(ϱ_x_0+c(x_0)<ς_x_0)>0. The strong Markov property at the stopping time ϱ_x_0+c(x_0) concludes the proof of Theorem <ref>(a). Proof of (b): in this case we have β < β_c. Recall the definition of m_c in (<ref>). The upper bound holds by Proposition <ref>. Moreover, Proposition <ref> implies that, for any p∈(m_c,∞), there exist constants x_0∈(0,∞) and c_1,c_2∈(1,∞), such that for all x_1∈[x_0,∞) and z∈∩ (c_1(x_1+1)+c_2,∞)×^d we have _z(ς_x_1 > t) ≥ Ct^-p, for some constant C and all t ≥ 1, say. By the Claim, for any z∈∩ (x_1,c_1(x_1+1)+c_2)×^d we have _z(ϱ_c_1(x_1+1)+c_2<ς_x_1)>0. Hence the strong Markov property applied at the stopping time ϱ_c_1(x_1+1)+c_2 implies Theorem <ref>(b). § FELLER CONTINUITY AND IRREDUCIBILITY OF THE REFLECTED PROCESS AND APPLICATIONS The existence of the invariant distribution of Z requires positive recurrence (see definition preceding Theorem <ref> above). The study of its moments requires certain technical results established in the present section. Section <ref> is dedicated to the proofs of Feller continuity and irreducibility of the process Z. In Section <ref> we apply these two properties to prove that the reflected process Z is Harris recurrent with an irreducible skeleton chain and that the set ∩ [0,r]×^d is petite for any r ∈(0,∞). Moreover, in Section <ref> we will also show that every petite set for Z is bounded. §.§ Feller continuity and irreducibility of the reflected process In this section we prove that the reflected process Z is Feller-continuous (see Theorem <ref> below) and that the (d+1)-dimensional Lebesgue measure _d+1 on the Borel σ-algebra () on is absolutely continuous with respect to its marginals of the reflected process at positive times (Proposition <ref> below). We start with the latter. Let ass:domain2, ass:covariance2, ass:vector2 hold. Then, for any z∈ and t∈(0,∞) and any A∈(), such that _d+1(A)>0, we have _z(Z_t ∈ A)>0. The proofs of Proposition <ref> and Theorem <ref> below require the following lemma. For small h>0, define a “thin” neighborhood of ∂ in by _h := {z∈: ∃ z'∈∂ such that z-z'_d+1<h}. Let ass:domain1, ass:covariance1, ass:vector1 hold. Then there exists a functions G:→, twice-differentiable on a neighbourhood of ⊂^d+1 and strictly positive on the open set ∖∂. Moreover, the function g(z) := G(z)^2 satisfies ∇ g(z) = 0 for all z∈∂ and for any r>0 there exist h_r>0, δ > 0, such that the following hold (a) Σ^1/2(z)∇ G(z)_d+1^2> δ for all z∈_h_r∩[0,r]×^d; (b) ⟨ϕ(z),∇ G(z)⟩ > δ for all z∈∂∩[0,r]×^d. Recall that Assumptions ass:domain2, ass:covariance2, ass:vector2 imply Assumptions ass:domain1, ass:covariance1, ass:vector1. Extend the function G:→, defined by G(z):=b(x)^2-y_d^2 for any z=(x,y)∈, to a C^2-function on a neighbourhood of in ^d+1. Hence ∇ G(z)_d+1^2=4(b(x)^2b'(x)^2+y_d^2) for all z=(x,y)∈ and ∇ g(z) = 2G(z)∇ G(z) = 0 for all z∈∂, since G(z) = 0 on ∂. Recall that ass:domain1 yields lim inf_x→0b(x)b'(x)>0. Thus, for any r>0, there exists a sufficiently small h_r>0, such that 0<inf_z∈_h_r∩[0,r]×^d∇ G(z)_d+1^2=:δ_h_r (we restrict to x∈[0,r] because of functions b with β≤ 0, e.g. example in Lemma <ref>). Moreover, by ass:covariance1, there exists δ_Σ>0 such that Σ^1/2(z)∇ G(z)_d+1^2=⟨Σ(z)∇ G(z),∇ G(z)⟩≥δ_Σ∇ G(z)_d+1^2≥δ_Σ·δ_h_r>0 for all z∈_h_r∩[0,r]×^d, implying (a). Note that gradient ∇ G(z) for any z∈∂ equals n(z)∇ G(z)_d+1, where n(z) is the inwards-pointing unit normal vector to ∂ at z. Hence, by ass:vector1, there exists a δ_ϕ>0 such that ⟨ϕ(z),∇ G(z)⟩≥δ_h_r^1/2⟨ϕ(z), n(z)⟩ > δ_ϕ·δ_h_r^1/2>0 for all z∈∂∩ [0,r]×^d. We start by proving that Z spends no time at the boundary. Claim 0. The equality ∫_0^∞Z_t∈∂ t=0 holds _z-a.s. for any starting point z∈. Proof of Claim 0. Let G be a function whose existence is guaranteed by Lemma <ref>. Define the non-negative continuous semimartingale ξ=(ξ_t)_t∈, ξ_t:=G(Z_t), denote its local time field by (L_t^u(ξ))_t,u∈ (see <cit.> for definition and properties) and note that ξ_t=0 if and only if Z_t∈∂, since G>0 on ∖∂ by Lemma <ref>. Thus the local-time process L in SDE (<ref>) satisfies L^0(ξ)=L. Itô's formula in (<ref>), applied to G(Z), yields that the quadratic variation [ξ] equals [M] given in (<ref>). Pick any t,r∈(0,∞). The occupation times formula in <cit.> applied to the indicator u↦_{0}(u) of zero, the representation of the quadratic variation [M] in (<ref>), the property in Lemma <ref>(a) and the fact that ξ_s=0 is equivalent to Z_s∈∂ for all s∈ yield 0=∫__{0}(u)L_t∧ϱ_r^u(ξ) u = ∫_0^t∧ϱ_r_{0}(ξ_s) [M]_s≥δ∫_0^t∧ϱ_rZ_s∈∂ s≥0, implying 0=∫_0^t∧ϱ_rZ_s∈∂ s. Since ϱ_r, given in (<ref>), satisfies lim_r→∞ϱ_r=∞ a.s. (by Theorem <ref>) and t>0 is arbitrary, our claim follows. Fubini's theorem and Claim 0 yield ∫_0^∞_z(Z_t∈∂) t=_z∫_0^∞Z_t∈∂ t=0 for any z∈ (since Z is continuous, it is progressively measurable, implying the various integrals are well defined and measurable). In particular, denoting := ∖∂, for any z∈ and t>0, it holds that ∫_0^t _z(Z_s∈) s = t. Claim 1. For every z∈, t>0 and A∈() the following holds: if _d+1(A) > 0 then ∫_0^t _z(Z_s∈ A) s > 0. In order to prove Claim 1, we need Claims 2 and 3 below. For z∈ and h>0, define the open ball in by B(z,h) := { z'∈: z-z'_d+1 < h}. Claim 2. Pick any z∈, s∈ and any ball B(z',h)⊂ and A∈(). The inequalities _z(Z_s∈ B(z',h))>0 and _d+1(B(z',h)∩ A)>0 imply _z(Z_v∈ B(z',h)∩ A) >0 for all v∈(s,s+h^2). Proof of Claim 2. Since B(z',h)⊂, the stopping time τ_∂ B(z',h):=inf{t∈: Z_t∉ B(z',h)} is strictly positive _z”-a.s. for all z”∈ B(z',h). Moreover, the process Z on the stochastic interval [0,τ_∂ B(z',h)), started at any z”∈ B(z',h), coincides with a uniformly elliptic diffusion on ^d+1, stopped upon exiting the ball B(z',h). Thus, <cit.> is applicable and, together with the strong Markov property of Z <cit.>, yields the claim. Claim 3. For any z∈, t>0 and A_0∈(), such that _d+1(A_0)>0, there exist z_0∈, h_0>0 and s∈(0,t) satisfying B(z_0,h_0)⊂, _d+1(A_0 ∩ B(z_0,h_0)) > 0 and _z(Z_s ∈ B(z_0,h_0)) > 0. Proof of Claim 3. Since ∫_0^t _z(Z_s ∈) s = t, there exist s < t, z'∈ and h>0, such that B(z',h)⊂ and _z(Z_s∈ B(z',h)) > 0. Moreover, the assumption _d+1(A_0) > 0 implies that there exists a ball B(z_0,h_0) ⊂ such that _d+1(B(z_0,h_0) ∩ A_0) > 0. It remains to prove that for some s∈(0,t) we have _z(Z_s ∈ B(z_0,h_0)) > 0. If B(z_0,h_0)∩ B(z',h)≠∅, then, since _d+1(B(z_0,h_0) ∩ B(z',h))>0, by Claim 2 applied with A:=B(z_0,h_0) and _z(Z_s∈ B(z',h))>0, there exist v∈(s,t) with _z(Z_v∈ B(z_0,h_0))>0. If B(z_0,h_0)∩ B(z',h)=∅, then there exists a sequence of n∈ balls B(z_i,h_i)⊂, where i∈{1,…,n}, such that z_n=z_0, h_n=h_0 and z_1=z', h_1=h and B(z_i,h_i)∩ B(z_i+1,h_i+1)≠∅ for all i∈{1,…,n-1}. Since _z(Z_s∈ B(z_1,h_1))>0, by Claim 2 (applied with A:=B(z_2,h_2)), there exists time v_1∈(s,t), such that _z(Z_v_1∈ B(z_2,h_2))>0. The Markov property at v_1 and Claim 2 imply the existence of v_2∈(v_1,t) such that _z(Z_v_2∈ B(z_3,h_3))>0. Construct inductively the increasing sequence v_1,v_2,…,v_n-1∈(0,t), set s:=v_n-1 and note _z(Z_v_n-1∈ B(z_n,h_n))>0, implying Claim 3. Proof of Claim 1. Assume that Claim 1 does not hold. More precisely, there exist t>0, z∈ and A_0 ∈(), such that _d+1(A_0)>0 and ∫_0^t _z(Z_v∈ A_0) v = 0. By Claim 3 there exist a ball B(z_0,h_0) in and s∈(0,t) such that _d+1(A_0 ∩ B(z_0,h_0)) > 0 and _z(Z_s ∈ B(z_0,h_0)) > 0. Claim 2 (applied with z':=z_0, h:=h_0 and A:=A_0) yields the contradiction: 0=∫_0^t_z(Z_v∈ A_0) v ≥∫_s^t ∧ (s+h_0^2)_z(Z_v∈ B(z_0,h_0)∩ A_0) v>0. To conclude the proof of the proposition, we strengthen Claim 1. Suppose there exist z∈, t>0 and A∈() with _d+1(A)>0, such that _z(Z_t∈ A)=0. Since there exists a ball B(z',h)∈(), such that _d+1(A ∩ B(z',h)) > 0, by Claim 1 applied to A ∩ B(z',h) we have ∫_0^h^2_z(Z_v ∈ A ∩ B(z',h)) v > 0 for all z∈. We may assume h^2∈(0,t). Since Z is Markov, we have ∫_t-h^2^t _z(Z_s ∈ A ∩ B(z',h)) s > 0 and hence _z(Z_s ∈ A ∩ B(z',h)) > 0 for some s∈(t-h^2,t). Thus t∈(s, s+h^2). By Claim 2 we get _z(Z_t∈ A) ≥_z(Z_t ∈ A ∩ B(z',h))>0, completing the proof of the proposition. By Claim 0 in the proof of Proposition <ref> above, for any z∈, the equality _z(Z_t∈∂)=0 holds for Lebesgue almost every t∈. Note also that the proof of Claim 0 uses only the occupation times formula for continuous semimartingales and basic properties of the solution of SDE (<ref>). Domain , defined in (<ref>), with increasing boundary (e.g. β>0, see (<ref>) for definition) satisfies the conditions of <cit.>, which establishes Feller continuity for reflecting processes Z. However, as explained in <cit.>, the assumptions of <cit.> are not satisfied if the boundary function b decreases to zero (e.g. β<0). In the case β=0, the domain , may but need not, satisfy the assumptions of <cit.>, see example in Lemma <ref> below. Since the case β<0 is when positive recurrence occurs, we develop a new approach to Feller continuity of Z, relying on the localisation of the process. This is more involved than the standard approach in the literature (see e.g. <cit.>) due to the difficulty of obtaining a global bound on the growth of the local time in the case β<0, which requires localisation. Let ass:domain2, ass:covariance2, ass:vector2 hold. For a continuous bounded function f:→, t∈, and a convergent sequence (z_n)_n∈ in with limit lim_n→∞z_n= z_∞∈, we have _z_n[f(Z_t)] →_z_∞[f(Z_t)] as n→∞. The proof of Theorem <ref> requires Lemmas <ref> and <ref>. Lemma <ref> provides growth estimates required for the proof of tightness via the Aldus's criterion <cit.>. Let ass:domain2, ass:covariance2, ass:vector2 hold. Fix any r>0, T>0 and θ∈(0,T]. Then there exist positive constants C_i, for i∈{1,2,3,4}, such that for any z∈ and (_t)-stopping times S_1,S_2 ((_t)_t∈ is the Brownian filtration in (<ref>)), satisfying S_1≤ S_2≤ S_1+θ≤ T, for any >0 the following hold: (a)_z(L_S_2∧ϱ_r-L_S_1∧ϱ_r≥) ≤ (C_1 θ + C_2θ^1/2)/; (b)_z(Z_S_2∧ϱ_r - Z_S_1∧ϱ_r_d+1^2≥) ≤ (C_3θ+ C_4θ^1/2)/. Moreover, _z[L_T ∧ϱ_r]<∞ holds. Let G be the function whose existence is guaranteed by Lemma <ref> and let g(z) := G(z)^2, z∈. Since the first and second derivatives of functions G and g are continuous on , and Σ is bounded by ass:covariance2, it follows that Σ^1/2∇ G_d+1^2, Σ^1/2∇ g_d+1^2, |Δ_Σ G| and |Δ_Σ g| are bounded on ^(r)= ∩ [0,r]×^d for any r>0. Pick r>0, T>0 and θ∈ (0,T]. Recall from Lemma <ref> that ∇ g(z) = 0 and ⟨ϕ(z),∇ G(z)⟩ > δ_r for all z∈∂∩[0,r]×^d and some positive constant δ_r. Let the bounded stopping times S_1,S_2 be as in the statement of the lemma. Itô's formula (<ref>) applied to the processes g(Z) and G(Z) on the stochastic interval [S_1∧ϱ_r,S_2∧ϱ_r] (recall the definition of ϱ_r in (<ref>)) yields g(Z_S_2 ∧ϱ_r) = g(Z_S_1 ∧ϱ_r) + M_S_2∧ϱ_r^g-M_S_1∧ϱ_r^g + 1/2∫_S_1∧ϱ_r^S_2∧ϱ_rΔ_Σ g(Z_u) u, _z-a.s., and δ_r (L_S_2 ∧ϱ_r-L_S_1∧ϱ_r)≤ G(Z_S_2∧ϱ_r) - G(Z_S_1∧ϱ_r) - M_S_2 ∧ϱ_r^G+ M_S_1 ∧ϱ_r^G - 1/2∫_S_1∧ϱ_r^S_2∧ϱ_rΔ_Σ G(Z_u) u, _z-a.s. for any z∈. Note that (M^g_t∧ϱ_r)_t∈ and (M^G_t∧ϱ_r)_t∈ are true martingales by (<ref>). The optional sampling theorem at the bounded stopping time S_2∧ϱ_r yields _z[M_S_2∧ϱ_r^g-M_S_1∧ϱ_r^g|_S_1∧ϱ_r]=0, _z-a.s. and _z[M_S_2∧ϱ_r^G-M_S_1∧ϱ_r^G|_S_1∧ϱ_r]=0, _z-a.s. There exists a constant C_1>0 such that |Δ_Σ g(z)|≤ 2C_1 for all z∈^(r), which implies that _z[g(Z_S_2∧ϱ_r) |_S_1∧ϱ_r] ≤ g(Z_S_1∧ϱ_r) + C_1θ, _z-a.s. holds for any z∈. Moreover, there exists a constant C_2 such that |Δ_Σ G(z)|≤ 2C_2 for z∈^(r). Using the fact that g(z) = G(z)^2, and applying the optional sampling theorem at the bounded stopping time S_2∧ϱ_r implies δ_r _z[L_S_2 ∧ϱ_r - L_S_1 ∧ϱ_r] ≤_z[_z[G(Z_S_2 ∧ϱ_r)|_S_1 ∧ϱ_r] - G(Z_S_1 ∧ϱ_r)+ C_2 (S_2-S_1)] ≤_z[_z[g(Z_S_2 ∧ϱ_r)|_S_1 ∧ϱ_r]^1/2 - G(Z_S_1 ∧ϱ_r)+ C_2 θ] ≤_z[(C_1 θ + g(Z_S_1 ∧ϱ_r))^1/2 - G(Z_S_1 ∧ϱ_r) + C_2 θ] ≤ (C_1 θ)^1/2 + C_2 θ, where we used the Cauchy-Schwarz inequality, the inequality in (<ref>) and the triangle inequality in the second, third and fourth inequalities in the display above, respectively. Application of Markov inequality to (<ref>) implies (a). Moreover, since L_0 = 0, setting S_1 = 0, S_2 = T, θ = T in (<ref>), we get _z[L_T ∧ϱ_r]<∞, as claimed. To prove (b), define the function V_z_0:→ by V_z_0(z) := z-z_0_d+1^2 for a parameter z_0∈. Pick r>0, T>0 and θ∈(0,T]. Since first and second derivatives of V_z_0(z) are continuous on in both variable z and parameter z_0, it follows by ass:vector2 and ass:covariance2 that there exist constants C̃_1,C̃_2,C̃_3 such that sup_z_0∈^(r)⟨∇ V_z_0(z),ϕ(z)⟩≤C̃_1, sup_z_0∈^(r)|Δ_Σ V_z_0(z)|≤ 2C̃_2, and sup_z_0∈^(r)Σ(z)^1/2∇ V_z_0(z)_d+1^2≤C̃_3 hold for every z∈^(r). Itô's formula (<ref>) applied to V_z_0(Z), with Z started at z_0, gives V_z_0(Z_S∧ϱ_r) = M_S∧ϱ_r^V_z_0 + ∫_0^S∧ϱ_r⟨∇ V_z_0(Z_u),ϕ(Z_u)⟩ L_u + 1/2∫_0^S∧ϱ_rΔ_Σ V_z_0(Z_u) u, _z_0-a.s., for any stopping time S and any z_0∈. Moreover, by (<ref>) the process (M^V_z_0_t∧ϱ_r)_t∈ is a true martingale. By (<ref>), for S satisfying S≤θ, we have sup_z_0∈^(r)_z_0[L_S∧ϱ_r]≤(C_1 θ)^1/2 + C_2 θ. Hence, the optional sampling theorem at the bounded stopping time S∧ϱ_r yields _z_0[Z_S∧ϱ_r-Z_0_d+1^2] = _z_0[V_z_0(Z_S∧ϱ_r)] ≤C̃_1 _z_0[L_S∧ϱ_r] + C̃_2 θ≤C̃_1((C_1θ)^1/2+C_2θ)+C̃_2 θ, for any S≤θ and z_0∈^(r), where C_1,C_2 come from part (a) of this lemma. It follows from the strong Markov property that for any stopping times S_1,S_2 satisfying assumptions of the lemma, and any z_0∈ we have _z_0[Z_S_2∧ϱ_r-Z_S_1∧ϱ_r_d+1^2] =_z_0[_z_0[Z_S_2∧ϱ_r-Z_S_1∧ϱ_r_d+1^2|_S_1∧ϱ_r]] ≤C̃_1((C_1θ)^1/2+C_2θ)+C̃_2 θ. Application of Markov inequality gives part (b) of the lemma. Let ass:domain2, ass:covariance2, ass:vector2 hold. Suppose that for any t>0, any continuous bounded function f:→, and a convergent sequence z_n→ z_∞∈ (as n→∞), there exists a sequence (r_k)_k∈ satisfying lim_k→∞r_k = ∞, such that for every k∈ the following holds: _z_n[f(Z_t∧ϱ_r_k)]→_z_∞[f(Z_t∧ϱ_r_k)] as n→∞. Then the process Z is Feller continuous, that is, _z_n[f(Z_t)] →_z_∞[f(Z_t)] as n→∞. Fix t>0. For any r>0, define the events _r := {ϱ_r≤ t} (and their complements _r^c). Claim. For any convergent sequence z_n → z_∞∈ and >0, there exists r_0>0 such that for all r∈[r_0,∞) we have sup_n∈_z_n(_r)≤ and _z_∞(_r)≤. Proof of claim. Pick w∈ (-∞,1-β s_0/c_0). Lemma <ref> guarantees the existence of constants k,x_0,x_1∈ in the definition of the function F_w,1 such that ⟨∇ F_w,1(z),ϕ(z)⟩≤0 for all z∈∂. By (<ref>), we have F_w,1(z)=f_w,1(z) for all z=(x,y)∈∩ [x_1,∞)×^d. Since F_w,1 is smooth on a neighbourhood of , by (<ref>) and (<ref>) in Lemma <ref>, the functions z↦|Δ_Σ F_w,1(z)| and z↦Σ(z)^1/2∇ F_w,1(z)_d+1^2 are bounded on . Hence, there exist constants C̃_1,C̃_2∈(0,∞) satisfying |Δ_Σ F_w,1(z)|≤ 2C̃_1 and Σ(z)^1/2∇ F_w,1(z)_d+1^2≤C̃_2 for all z∈. Define the process κ := 2^1/| w|F_w,1(Z), and recall the definition of ρ_r in (<ref>). Since F_w,1(z) = f_w,1(z) for z∈∩ [x_1,∞)×^d, (<ref>) implies that for r∈[x_1,∞) we have ρ_r≤ϱ_r. Moreover, application of Itô's formula (<ref>) implies that the process κ_t∧ρ_r - C(t∧ρ_r) is a supermartingale for C=2^1/|w|C̃_1, and any r>0, where the local martingale appearing in Itô's formula is a true martingale by (<ref>). Thus, applying Proposition <ref> (with ξ=κ and a constant function f≡ C) we infer that for any r≥ x_1 and any z∈, _z(_r)≤_z(ρ_r ≤ t) ≤ r^-1(2^1/|w|F_w,1(z) + C t). Since (2^1/|w|sup_n∈F_w,1(z_n)+Ct) is finite by the continuity of F_w,1, there exists r_0 as claimed. To prove the lemma, pick an arbitrary continuous bounded function f:→, >0, and a convergent sequence z_n→ z_∞. Let the sequence (r_k)_k∈ tend to infinity and satisfy (<ref>) for every k∈. Since f is bounded, the Claim above implies that there exists k∈ such that sup_z∈| f(z)|(_z_n(_r_k)+_z_∞(_r_k))</3 holds for every n∈. Thus, as f(Z_t)=f(Z_t∧ϱ_r_k)1(_r_k^c)+f(Z_t)1(_r_k), for every n∈ we get |_z_n[f(Z_t)]-_z_∞[f(Z_t)]| ≤|_z_n[f(Z_t∧ϱ_r_k)1(_r_k^c)]-_z_∞[f(Z_t∧ϱ_r_k)1(_r_k^c)]| + /3. Next, by (<ref>), for all large n∈ we have |_z_n[f(Z_t∧ϱ_r_k)] - _z_∞[f(Z_t∧ϱ_r_k)]| </3. Since f(Z_t∧ϱ_r_k)1(_r_k^c)=f(Z_t∧ϱ_r_k)-f(Z_t∧ϱ_r_k)1(_r_k), the triangle inequality yields |_z_n[f(Z_t∧ϱ_r_k)1(_r_k^c) ]-_z_∞[f(Z_t∧ϱ_r_k)1(_r_k^c)]| ≤|_z_n[f(Z_t∧ϱ_r_k)] - _z_∞[f(Z_t∧ϱ_r_k)] | + sup_z∈| f(z)|(_z_n(_r_k)+_z_∞(_r_k)) ≤ 2/3, which, together with (<ref>), implies |_z_n[f(Z_t)]-_z_∞[f(Z_t)]|< for all large n∈. By Lemma <ref>, it suffices to prove that for any f:→, T∈ and a convergent sequence z_n→ z_∞∈ (as n→∞), there exists a sequence (r_k)_k∈ tending to infinity, such that for any k∈ we have _z_n[f(Z_T∧ϱ_r_k)] →_z_∞[f(Z_T∧ϱ_r_k)] as n→∞. The first step is to prove that, for a fixed r>0, the laws of (Z_·∧ϱ_r,L_·∧ϱ_r,W_·) under _z_n are tight. The second step consists of proving that every subsequence converges to the law of (Z_·∧ϱ_r,L_·∧ϱ_r,W_·) under _z_∞. Let E^(1) :=E_1× E_2× E_3, where E_1:=, E_2:= and E_3:=^d+1, and denote by D_E^(1)([0,∞)) the space of càdlàg functions (i.e. right-continuous functions with left limits), mapping the interval [0,∞) into the metric space E^(1). Denote by C_E^(1)([0,∞)) ⊂ D_E^(1)([0,∞)) the subspace of continuous functions. Endow D_E^(1)([0,∞)) with the Skorohod topology and its Borel σ-algebra ^1, see e.g. <cit.> for details. Any function ω∈ D_E^(1)([0,∞)) can be expressed as ω = (a^(1),a^(2),a^(3)), for some “coordinate” càdlàg functions a^(i): [0,∞)→ E_i for i=1,2,3. For any t∈[0,∞), define maps A^(i)_t:D_E^(1)([0,∞))→ E_i by A^(i)_t(ω):=a^(i)(t), i∈{1,2,3}. Define a σ-algebra _t^1 := σ{A^(i)_s:0≤ s≤ t, i∈{1,2,3}}⊂^1 generated by continuous maps A^(i)_s. By <cit.> we have _∞^1=^1. For any z∈, by <cit.>, there exists a filtered probability space (Ω,(_t)_t∈,,_z), supporting the processes (Z,L,W), taking values in E^(1) such that, under _z, the SDE in (<ref>) holds and W is a standard (_t)-Brownian motion. We may assume that the filtration (_t)_t∈ is complete (i.e. _0 contains all _z-null sets of ) and right continuous (i.e. _t = ∩_t<s_s). Pick arbitrary r_0∈ and T>0, and recall the stopping time ϱ_r_0, defined in (<ref>), is the first time the coordinate X of Z=(X,Y) reaches level r_0. The stopped process (Z_t ∧ϱ_r_0∧ T,L_t ∧ϱ_r_0∧ T,W_t∧ T)_t∈ produces a measurable map (Z_·∧ϱ_r_0∧ T,L_·∧ϱ_r_0∧ T,W_·∧ T):(Ω,)→(D_E^(1)([0,∞)),^1) and induces the probability measure Q_z^1(·):=_z((Z,L,W)∈·) on ^1. Moreover, using the notation Z^(z)_t(ω):=A^(1)_t(ω) = (A_t^(1,X)(ω),A_t^(1,Y)(ω))∈×^d, L^(z)_t(ω):=A^(2)_t(ω), W^(z)_t(ω):=A^(3)_t(ω), ϱ^(z)_r := inf{t>0:A_t^(1,X)≥ r} for any r∈(0,∞), ω∈ D_E([0,∞)) and t∈[0,∞), the following holds for any r∈(0,∞) and t∈[0,T]: (a) Z^(z)_t∧ϱ_r^(z) = z + ∫_0^t∧ϱ_r^(z)Σ^1/2(Z^(z)_s) W^(z)_s + ∫_0^t∧ϱ_r^(z)ϕ(Z^(z)_s) L^(z)_s, Q_z-a.s. (b) L^(z)_t∧ϱ_r^(z) = ∫_0^t∧ϱ_r^(z)Z^(z)_s∈∂ L^(z)_s, Q_z-a.s. (c) Under Q_z, the process W^(z) is an (_t)-Brownian motion stopped at time T. Moreover, it is easy to see that the equality in (b) above is equivalent to ∫_0^t∧ϱ_r^(z) (b(X^(z)_s)^2 - Y^(z)_s_d^2) L^(z)_s = 0, Q_z-a.s. Indeed, since b(x)^2-y_d^2=0 for all (x,y)∈∂, (b) implies ∫_0^t∧ϱ_r^(z) (b(X^(z)_s)^2 - Y^(z)_s_d^2) L^(z)_s = ∫_0^t∧ϱ_r^(z) (b(X^(z)_s)^2 - Y^(z)_s_d^2) Z^(z)_s∈∂ L^(z)_s= 0. Conversely, since b(x)^2-y_d^2>0 for all (x,y)∈∖∂, the sets _n:={(x,y)∈:b(x)-y_d∈[1/n+1,1/n)} are pairwise disjoint and satisfy ∪_n∈_n=∖∂ and 0≤∫_0^t∧ϱ_r^(z)Z^(z)_s∈_n L^(z)_s≤ (n+1)∫_0^T (b(X^(z)_s)^2 - Y^(z)_s_d^2) Z^(z)_s∈_n L^(z)_s=0. Note L^(z)_t∧ϱ_r^(z) = ∫_0^t∧ϱ_r^(z)Z^(z)_s∈∂ L^(z)_s +∑_n∈∫_0^t∧ϱ_r^(z)Z^(z)_s∈_n L^(z)_s =∫_0^t∧ϱ_r^(z)Z^(z)_s∈∂ L^(z)_s. We will show that if z_n→ z_∞, for all but countably many r∈(0,r_0), the limiting point of laws of the processes (Z_t∧ϱ_r^(z_n)^(z_n))_n∈ coincides with the law of Z_t∧ϱ_r^(z_∞)^(z_∞). Denote by (D_E^(1)([0,∞)),^1) the space of probability measures on the measurable space (D_E^(1)([0,∞)),^1) and endow it with the topology of weak convergence (see <cit.> for details). We say that the sequence of measures (Q_n)_n∈∈(D_E^(1)[0,∞),^1) is C-tight, if every subsequence of (Q_n)_n∈ has a convergent subsequence and the limiting probability measure Q_* charges only the set C_E^(1)([0,∞)), i.e. Q_*(C_E^(1)[0,∞)) = 1. Recall that the measures Q_z^1, for z∈, are defined by the laws of the stopped processes (Z_·∧ϱ_r∧ T,L_·∧ϱ_r∧ T,W_·∧ T) under the measure _z. Lemma <ref>, and the fact that W is a Brownian motion, imply that for any θ>0 there exist positive constants C_1,C_2 such that for any stopping times S_1,S_2∈_T satisfying S_1≤ S_2≤ S_1+θ≤ T, for any n∈ we have _z_n( (Z_S_2∧ϱ_r_0,L_S_2 ∧ϱ_r_0,W_S_2)-(Z_S_1∧ϱ_r_0,L_S_1 ∧ϱ_r_0,W_S_1)_2d+3^2 ≥) ≤ (C_1 θ + C_2θ^1/2)/. There exists a positive constant C>0, such that sup_t∈_+Z_t∧ϱ_r_0_d+1≤ C _z_n-a.s. for all n∈. Since Lemma <ref> bounds the expected growth of the local time L, for any >0, there exists K>0, such that for all n∈ we have _z_n(sup_0≤ s≤ T(Z_s ∧ϱ_r_0,L_s∧ϱ_r_0,W_s)_2d+3≥ K)≤. Thus we may apply Aldous's tightness criterion <cit.> to deduce that the sequence of measures (Q_z_n^1)_n∈ is tight. Moreover, since Q_z_n^1(C[0,∞)) = 1 for every n∈, <cit.> implies that the sequence (Q_z_n^1)_n∈ is C-tight. It follows that there exists a subsequence (Q_z_n_k^1)_k∈ that converges weakly to a probability measure Q_*^1, satisfying Q_*^1(C[0,∞)) = 1. For notational convenience we assume that the sequence (Q_z_n^1)_n∈ itself converges to Q_*^1. Our aim is to prove that (A^(1),A^(2),A^(3)), under the measure Q_*^1, solves SDE (<ref>). Consider the process :=_·∧ T, where :=(Z_·∧ϱ_r_0,L_·∧ϱ_r_0,W_·,∫_0^·Σ^1/2(Z_u) W_u,∫_0^·∧ϱ_r_0ϕ(Z_u) L_u,∫_0^·∧ϱ_r_0(b(X_u)^2 - Y_u_d^2) L_u). The state space of is E^(2) := E^(1)× E_4× E_5× E_6, where E_4 = ^d+1,E_5=^d+1 and E_6=. Let (D_E^(2)([0,∞)),(_t^2)_t∈[0,∞), ^2) be the filtered measurable space with coordinate projections A^(i):D_E^(2)([0,∞))→ D_E_i([0,∞)) for i∈{1,…,6} (note that A^(i), for i∈{1,2,3}, agrees with the definition in the beginning of the second paragraph of this proof). Moreover, a measurable map :(Ω,)→(D_E^(2)([0,∞)),^2) induces the probability measure Q_z^2(·):=_z(∈·) on ^2. Note that Q_z^2((A^(1),A^(2),A^(3))^-1(·)) = Q_z^1(·) for all z∈. Hence the first three coordinates converge weakly under (Q_z_n^2)_n∈. Denote A^(1) = (A^(1,X),A^(1,Y)), where A^(1,X):D_E^(2)([0,∞)) → D_([0,∞)) and A^(1,Y):D_E^(2)([0,∞)) → D_^d([0,∞)). Note that for each z∈, L and W under the measures _z are adapted to the complete, right-continuous filtration (_t)_t∈ and have continuous sample paths, which are thus in D_E_2([0,∞)) and D_E_3([0,∞)), respectively.[We work with càdlàg paths because we apply <cit.> to conclude the stability of the stochastic integrals.] Moreover, the processes W,L are semimartingales, which satisfy _z_n[L_t∧ϱ_r_0∧ T]<∞ (by Lemma <ref>) and _z_n[[W]_t∧ T] ≤ T for any t∈ and n∈. Thus, Assumption <cit.> is satisfied with the deterministic time τ^α_n := α+1, where α∈(0,∞), n∈ are arbitrary parameters and τ^α_n is a sequence of stopping times in <cit.>. Since the functions z↦Σ(z), z↦ϕ(z), z↦ b(x)^2-y_d+1^2 are continuous for z=(x,y)∈ and _z(∈·) = Q_z^2(·), <cit.> implies that there exists a probability measure Q_*^2 on (D_E^(2)([0,∞)),^2), such that A :=(A^(1)_·,A_·^(2),A_·^(3),∫_0^·Σ^1/2(A^(1)_u) A_u^(3),∫_0^·ϕ(A^(1)_u) A^(2)_u,∫_0^· (b(A_u^(1,X))^2 - A^(1,Y)_u_d^2) A^(2)_u), under the measure Q_z_n^2, converges weakly (as n→∞) to A under Q_*^2. In particular,<cit.> ensures that A^(2),A^(3) are semimartingales under Q_*^2. Note that all limiting processes have continuous paths Q_*^2-a.s., which follows from the fact that A^(1),A^(2) and A^(3) have continuous paths Q_*^2-a.s. We now extend (using <cit.>) the weak convergence to stopping times ϱ_r and the corresponding stopped processes. For a^(1,X)∈ D_([0,∞)) and r>0, denote τ_r(a^(1,X)) := inf{ t>0: a^(1,X)_t ≥ r or a^(1,X)_t-≥ r}, where a^(1,X)_t- denotes the left limit of a^(1,X) at time t. Since, for any z∈, the process X has _z-a.s. continuous paths, we have τ_r(X) = ϱ_r, _z-a.s. We now make the final extension of the state space, which will capture the convergence of the stopping times. Denote E^(3) := E^(2)×, and let (D_E^(3)([0,∞)),(_t^3)_t∈[0,∞), ^3) be the filtered measurable space with coordinate projections A^(i), for i∈{1,…,7}. For any r∈ (0,∞), a measurable map (,ϱ_r):(Ω,)→(D_E^(3)([0,∞)),^3) induces the probability measure Q_z^3,r(·):=_z((,ϱ_r)∈·) on ^3. Furthermore, for any r∈(0,∞), it holds that Q_z^3,r(A^-1(·)) = Q_z^2(·). As before, A under Q^3,r_z_n converge to à under some probability measure Q^3,r_*. Moreover, A^(1,X) has Q_*^3,r-a.s. continuous paths. Thus, <cit.> implies that for all but at most countably many r∈(0,r_0), the map D_([0,∞))→, given by a^(1,X)↦τ_r(a^(1,X)), is continuous at A^(1,X), Q_*^3,r-a.s. (recall that r_0>0 is an arbitrary number, fixed at the beginning of the proof). Pick r_1∈(r_0/2,r_0) such that a^(1,X)↦τ_r_1(a^(1,X)) is continuous at a^(1,X), Q_*^r_1-a.s. The Continuous mapping theorem <cit.> implies that (Ã,τ_r_1(A^(1,X))), under Q_z_n^3,r_1, converges weakly to (Ã,τ_r_1(A^(1,X))) under Q_*^3,r_1. The convergence of the stopped processes follows by applying <cit.> to the stochastic integral ∫_0^t 1{s≤ϱ_r_1}_s (condition <cit.> is satisfied with τ^α_n := α+1 as above). Since _z((,ϱ_r_1)∈·) =Q_z^3,r_1(·), it follows that (Ã_·∧τ_r_1(A^(1,X))), under Q_z_n^3,r_1, converges weakly as n→∞ to (Ã_·∧τ_r_1(A^(1,X))) under the probability measure Q_*^3,r_1. To conclude the proof, we have to show that the process (A^(1),A^(2),A^(3)) under Q_*^3,r_1 solves SDE (<ref>). Denote by 0∈ D_E_1([0,∞)) the function mapping every t∈[0,∞) into the origin of ^d+1. For every n∈, we have A^(1)_·∧τ_r_1(A^(1,X)) - z_n -∫_0^·∧τ_r_1(A^(1,X))ϕ(A^(1)_u) A^(2)_u - ∫_0^·∧τ_r_1(A^(1,X))Σ^1/2(A^(1)_u) A^(3)_u≡0, Q_z_n^3,r_1-a.s. Since the set {0} is closed in D_E_1([0,∞)), <cit.> implies that A_·∧τ_r_1(A^(1,X))^(1) -z_∞ - ∫_0^·∧τ_r_1(A^(1,X))Σ^1/2(A^(1)_u) A_u^(3) - ∫_0^·∧τ_r_1(A^(1,X))ϕ(A^(1)_u) A^(2)_u≡0, Q_*^3,r_1-a.s. Note that since A^(3) is a Brownian motion under Q_z_n^3,r_1 for every n∈, it is also a Brownian motion under the weak limit Q_*^3,r_1. It remains to prove that, under Q^3,r_1_*, A^(2) is a local time of A^(1) at the boundary ∂. For any z∈, the local time has to satisfy L_t∧ϱ_r = ∫_0^t∧ϱ_rZ_s∈∂ L_s, _z-a.s. This requirement is equivalent to ∫_0^t∧ϱ_r (b(X_s)^2 - Y_s_d^2) L_s = 0, _z-a.s. Indeed, since b(x)^2-y_d^2=0 for all (x,y)∈∂, we have ∫_0^t∧ϱ_r (b(X_s)^2 - Y_s_d^2) L_s = ∫_0^t∧ϱ_r (b(X_s)^2 - Y_s_d^2) Z_s∈∂ L_s= 0, _z-a.s., by the definition of local time. Conversely, since b(x)^2-y_d^2>0 for all (x,y)∈∖∂, the sets _k:={(x,y)∈:b(x)^2-y_d^2∈[1/(k+1),1/k)} are pairwise disjoint and satisfy ∪_k∈_k=∖∂. Moreover, we have 0≤∫_0^t∧ϱ_rZ_s∈_k L_s≤ (k+1)∫_0^t∧ϱ_r (b(X_s)^2 - Y_s_d^2) Z_s∈_k L_s=0, implying (<ref>) via L_t∧ϱ_r = ∫_0^t∧ϱ_rZ_s∈∂ L_s +∑_k∈∫_0^t∧ϱ_rZ_s∈_k L_s =∫_0^t∧ϱ_rZ_s∈∂ L_s. Note that ∫_0^·∧τ_r_1(A^(1,X)) (b(A^(1,X))^2 - A^(1,Y)_d^2) A^(2)_u ≡0, Q_z_n^3,r_1-a.s., for every n∈, where 0 now denotes the zero function in D_E_6([0,∞)) (recall that E_6=). Thus, by <cit.>, we get ∫_0^·∧τ_r_1(A^(1,X)) (b(A^(1,X))^2 - A^(1,Y)_d^2) A^(2)_u ≡ 0, Q_*^3,r_1-a.s. Hence, the condition in (<ref>) implies that A^(2) is indeed the local time of A^(1) at the boundary ∂. We have thus proved that (A^(1)_·∧τ_r_1(A^(1,X)),A^(2)_·∧τ_r_1(A^(1,X)),A^(3)_·∧τ_r_1(A^(1,X))), under Q_*^3,r_1 (and hence under Q_*^1), solves SDE (<ref>) on the stochastic interval [0,ϱ_r_1]. The pathwise uniqueness <cit.> of solutions of SDE (<ref>) implies that every sub-sequential limit Q_*^1 of the sequence of (Q_z_n^1)_n∈ equals Q_z_∞^1, implying the Feller continuity for the process stopped at ϱ_r_1, i.e. the limit in (<ref>) holds for r_1. Since r_0 was chosen arbitrarily and r_1∈(r_0/2,r_0), we can inductively construct a sequence (r_k)_k∈ with lim_k→∞r_k = ∞, such that (<ref>) holds for every k∈. An application of Lemma <ref> concludes the proof of Theorem <ref>. §.§ Application to petite sets and Harris recurrence The family of probability measures _z(Z_t ∈·) on (), indexed by t∈ and z∈ constitutes a Markov transition kernel by <cit.> and Theorem <ref>. A non-empty set B ∈() is called petite if there exists a probability measure a on (,()), which does not charge zero (i.e.  a({0})=0), and a non-trivial measure φ_a on (,()) satisfying K_a(z, ·) ≥φ_a(·) for all z ∈ B, where the Markov transition function K_a: ×() → is given by K_a(z,·) := ∫__z(Z_t∈·)a( t). The measurability of z↦ K_a(z,A) for any A∈() follows from <cit.>. Let φ be a σ-finite measure on the Borel σ-algebra (). The process Z is Harris recurrent if φ(A) > 0 implies ∫_0^∞1_A(Z_s) s = ∞, _z-a.s., for all A∈() and z ∈. Suppose that ass:covariance2, ass:vector2 and ass:domain2 hold. Then all compact subsets of are petite. Moreover, if β<β_c, the process Z is Harris recurrent. Fix z_0∈∖∂= and h_0>0, such that B(z_0,2h_0)={z∈:z-z_0_d+1<2h_0}⊂. We now prove that the ball B(z_0,h_0) is petite. Recall that the stopping time τ_∂ B(z_0,2h_0):=inf{t∈: Z_t∉ B(z_0,2h_0)} is strictly positive _z-a.s. for all z∈ B(z_0,h_0). Moreover, the process Z on the stochastic interval [0,τ_∂ B(z_0,2h_0)), started at any z∈ B(z_0,h_0), coincides with a uniformly elliptic diffusion on ^d+1, stopped upon exiting the ball B(z_0,2h_0). Thus, by <cit.>, we have inf_z∈ B(z_0,h_0)_z(Z_h_0^2∈·)≥φ(·), where φ is the Lebesgue measure multiplied by a positive scalar and supported in B(z_0,h_0). Hence, for any z∈ B(z_0,h_0), condition (<ref>) holds with a_1( t) = δ_h_0^2( t), where δ_h_0^2 is the Dirac delta concentrated at h_0^2>0, and taking the non-trivial measure φ_a_1:=φ on (,()). We now prove that an arbitrary compact set D in is also petite. Since B(z_0,h_0) is an open set, Theorem <ref> and <cit.> imply that the function z↦_z(Z_1∈ B(z_0,h_0)) is lower semi-continuous on . Moreover, by Proposition <ref>, we have _z(Z_1∈ B(z_0,h_0))>0 for all z∈ D. Hence, by compactness of D and the lower semi-continuity of the function z↦_z(Z_1∈ B(z_0,h_0)), we get inf_z∈ D_z(Z_1∈ B(z_0,h_0))>0. For any z∈ D, the Markov property of Z implies _z(Z_1+h_0^2∈·)≥_z(Z_1∈ B(z_0,h_0)) inf_z'∈ B(z_0,h_0)_z'(Z_h_0^2∈·)≥inf_z”∈ D_z”(Z_1∈ B(z_0,h_0))φ_a_1(·). Thus the set D satisfies condition (<ref>) with the probability measure a_2( t) = δ_1+h_0^2( t) and non-trivial measure φ_a_2 := inf_z∈ D_z(Z_1∈ B(z_0,h_0))φ_a_1 on (,()), making D petite. In particular, the set ∩ [r_0,∞)×^d is a petite set for every r_0∈(0,∞). Moreover, if β<β_c, Proposition <ref> implies that, for all sufficiently large r_0∈(0,∞), we have _z(ς_r_0 < ∞) = 1 for all z ∈. We conclude that the process Z is Harris recurrent by <cit.>. The following proposition is crucial for establishing the lower bounds on the tail of the invariant distribution of the reflected process Z in Theorem <ref> (the proof of Theorem <ref> requires an estimate of the return times to an arbitrary petite set). Suppose that ass:covariance2, ass:vector2 and ass:domain2 hold with β < β_c. Then every petite set is bounded. Let B be an arbitrary petite set and let the probability measure a on _+ and a non-zero measure φ_a on (,()) be such that (<ref>) holds, with K_a as in (<ref>). Denote ^(r)=∩ [0,r]×^d for any r>0. Since φ_a is a non-trivial measure on (,()), there exists r_0∈(0,∞), such that c:=φ_a(^(r_0))> 0. Proposition <ref> implies that for every q∈(0,1), p∈(1-β/β_c,∞) and z=(x,y)∈, there exists r_1∈(0,∞) such that for every r∈(r_1,∞) there exist constants c_1,c_2∈(1,∞) and ∈(0,∞) satisfying _z(ς_r ≥ t_0) ≥ qmin{(c_1^-p(x-c_2)^p-r^p)(1-q)^p^-1/2 t_0^-p/2,1} for all t_0∈(c_1 r_1+c_2,∞) (recall that the return time ς_r_0 is defined in (<ref>)). Since a is a probability measure, there exists t_0∈(c_1 r_1+c_2,∞) satisfying a([t_0,∞)) < c/2. We now show that there exists x_0∈(0,∞), such that _z(ς_r_0 <t_0) < c/2 for all z∈∖^(x_0). Indeed, fix r ≥max{r_0,r_1} (note ς_r≤ς_r_0), q ∈(1-c/2,1) and p∈(1-β/β_c,∞). Pick x_0>0, such that (c_1^-p(x-c_2)^p-r^p)(1-q)^p^-1/2 t_0^-p/2≥ 1 for all x∈(x_0,∞). Note that this choice of x_0 implies x_0>r and, in particular, x_0∈(r_0,∞). For any z=(x,y)∈∖^(x_0), the inequality in (<ref>) implies _z(ς_r_0 < t_0) ≤_z(ς_r < t_0) < 1-q <c/2. By (<ref>), we have K_a(z,^(r_0))≤∫_0^t_0_z(Z_t∈^(r_0))a( t)+a([t_0,∞)). Since, for all z∈∖^(x_0), we have _z(Z_t∈^(r_0))≤_z(ς_r_0 < t_0)<c/2 for all t∈[0,t_0], the inequality φ_a(^(r_0)) =c> K_a(z,^(r_0)) holds for all z∈∖^(x_0). Since the petite set B satisfies (<ref>), we must have B⊂^(x_0), making B bounded. § STABILITY: THE PROOF OF THEOREM <REF> Existence and uniqueness of the invariant distribution of Z, the upper bounds on the tails of the invariant distribution and upper bounds on the rate of convergence of Z to stationarity will be established using the drift condition (i.e. supermartingale property) given in Lemma <ref> (see Section <ref>) and the fact that every compact set in is petite for the reflected process Z (see Proposition <ref> in Section <ref> above). The lower bounds on the tails of the invariant distribution and the rate of convergence to stationarity will follow from the fact that every petite set of Z is bounded (see Proposition <ref> above) and the control we have established on the return time and length of excursions away from bounded sets (see Proposition <ref>). Theorem <ref> follows easily from Propositions <ref> and <ref> proved in this section. §.§ Existence, uniqueness, and upper bounds The upper bounds on the tails of the invariant distribution are obtained by establishing finiteness of certain moments and applying the Markov inequality. In Section <ref> below we show that these bounds cannot be improved. Suppose that ass:covariance2, ass:vector2 and ass:domain2 hold with β < -β_c and recall M_c = -(1+β/β_c)/2>0. Then there exists the unique invariant distribution π on (,()) for the process Z. Pick ϵ>0. There exists a constant C_π∈(0,∞), such that π({z∈:z_d+1≥ r})≤ C_π r^-2M_c+ϵ for all r∈[1,∞). Furthermore, for every z∈ there exists a constant C_TV∈(0,∞), such that _z(Z_t∈·)-π_TV ≤ C_TVt^-M_c + for all t∈[1,∞). By Proposition <ref>, the process Z admits an irreducible skeleton chain. Moreover, by Proposition <ref> the sets ^(r)=∩[0,r]×^d, defined in (<ref>), are petite for every r>0. Pick arbitrary ∈(0,1-β/β_c-2) and note that γ:=1-β/β_c - >2. By Lemma <ref>, the process ξ, defined in (<ref>), is a supermartingale. Note that by definition of F_w,γ in (<ref>), we have F_w,γ(z)^(γ-2)/γ = F_w,γ-2(z) for z∈∩[x_1,∞)×^d, where x_1 is the constant appearing in the definition of the function m in (<ref>). We may thus apply <cit.> (with V=F_w,γ, ϕ(u) = C_1 u^(γ-2)/γ, b=C_2 and the petite set ^(x_2) from Lemma <ref>), to deduce the existence and uniqueness of the invariant distribution π of Z and ∫_F_w,γ-2(z)π( z)<∞. From the definition of F_w,γ(z) in (<ref>) and the lower bound in (<ref>) it follows that C̃_π:=∫_z=(x,y)∈x^2M_c-π( z)<∞. Moreover, Markov's inequality implies π(∩ [r,∞)×^d)≤∫_ (x/r)^2M_c-ϵπ( z)≤C̃_π r^-2M_c+ϵ for all r∈[1,∞). Recall that for any z = (x,y)∈ we have x ≤z_d+1≤ (x^2+b(x)^2)^1/2 and the boundary function b has sublinear growth (cf. Remark <ref>), implying (x^2+b(x)^2)^1/2/x → 1 as x →∞. Thus, the upper bound π(∩ [r,∞)×^d)≤C̃_π r^-2M_c+ϵ implies the existence of the constant C_π∈(0,∞) such that the bound on the tail π({z ∈:z_d+1≥ r})≤ C_π r^-2M_c+ϵ holds for all r∈[1,∞) as claimed in the proposition. Recall that _d+1 is a Lebesgue measure on ^d+1. The process Z admits an _d+1-irreducible skeleton chain, since for every A∈(), such that _d+1(A) > 0, and z ∈, by Proposition <ref>, we have _z(Z_k∈ A) > 0 for every k ∈. In particular, Assumption (i) in <cit.> is satisfied for Z. For γ = 1-β/β_c-, Lemma <ref> ensures that Assumption (ii) of <cit.> is satisfied. By <cit.>, with the pair of functions Ψ = (Id,1), where Id,1:→ denote the identity and the constant functions, respectively, we obtain r_*(t)_z(Z_t∈·)-π_TV≤ F_w,γ(z) for all t≥ 0, where r_*(t) = φ∘ H_φ^-1(t). Here, the function φ is positive, satisfying φ(u) = C̃_1u^(γ-2)/γ for u≥ 1, and H_φ^-1 is the inverse of the increasing function H_φ, satisfying H_φ(u) = ∫_1^u φ(s)^-1 s for u≥ 1. This implies r_*(t) = C̃_2 (t+1)^γ/2-1 for a positive constant C̃_2∈(0,∞) and t∈(0,∞). §.§ Lower bounds The lower bounds on the tails of the invariant distribution π of the reflected process Z are closely related to the tail behaviour of certain additive functionals of the paths of Z until the return time to a petite set. For a measurable subset D ⊂ and δ > 0, define the return time of the process Z to the set D after the time δ>0 by τ_δ(D) := inf{ t > δ: Z_t ∈ D} (with convention inf∅=∞). Suppose that ass:domain2, ass:covariance2, ass:vector2 hold with β < β_c and pick p ∈(1-β/β_c,∞), a bounded measurable set D⊂ and z∈. Then there exist constants δ∈ (0,∞) and C,r_0,c_1,c_2,ϵ∈(0,∞), such that for every non-decreasing continuous function H:→ with r_H:=inf{r'≥:H(r')>0}<∞, we have _z(∫_0^τ_δ(D) H(c_1(X_s +c_2)) s ≥ r) ≥ C/ G(r/ϵ)^p for all r∈(r_0,∞), where G:→ is the inverse of the strictly increasing function defined on [r_H,∞) by the formula v↦ v^2 H(v). (a) The assumption β < β_c in Proposition <ref> covers both the null-recurrent and positive-recurrent cases. However, the main application of Proposition <ref> in the proof of Lemma <ref> below requires only the positive-recurrent case. Since Lemma <ref> is crucial in the proof of Proposition <ref>, the bound in Proposition <ref> is key for the lower bounds in Theorem <ref>. (b) The statement of Proposition <ref> in fact holds for every δ>0. In the proof of Proposition <ref> below, we apply the non-confinement property of Z, given in Lemma <ref>, to conclude that δ>0 exists. However, since by Proposition <ref> Z is irreducible, the event {X_δ > c_1(d_0+1)+c_2} has positive probability for every δ>0. Since the existence of δ>0 is sufficient for our analysis of the lower bound on the tail of the invariant distribution and non-confinement is weaker (and easier to prove) than irreducibility, we use the formulation of Proposition <ref> above. (c) The assumed continuity of the function H in Proposition <ref> is not necessary: measurability would be sufficient but it would complicate the formulation of the proposition. Pick a non-decreasing continuous function H:→ with r_H=inf{r':H(r')>0}<∞. Proposition <ref> implies that for every p ∈ (1-β/β_c,∞) and q∈(0,1), there exist constants x_0∈(0,∞), c_1,c_2∈(1,∞) and (q)∈(0,∞), such that for every x_1∈(x_0,∞) and z=(x,y)∈∩ (c_1x_1+c_2,∞)×^d and function H̅(r):=H(c_1(r+c_2)), r∈, we have _z(∫_0^ς_x_1H̅(X_s) s ≥ v^2H(v)) ≥ qmin{(c_1^-p(x-c_2)^p-x_1^p)v^-p,1}, for all v∈(c_1x_1+c_2,∞). Recall here that ς_x_1, defined in (<ref>) above, is the return time of the first coordinate X (of Z) below the level x_1 and X_0=x _z-a.s. Note that the following elementary inequality holds since the function is monotonically increasing as p>0: c_1^-p(x-c_2)^p-x_1^p≥ (x_1+1)^p-x_1^p for all x∈(c_1(x_1+1)+c_2,∞). Since v↦ v^2H(v) is strictly increasing on [r_H,∞) with range equal to , for any r∈(c_1x_1+c_2,∞) we can define v:=G(r/ϵ), implying r=ϵ v^2 H(v). Thus, for every z=(x,y)∈∩ (c_1(x_1+1)+c_2,∞)×^d and any C_x_1∈(0,q((x_1+1)^p-x_1^p)/G(c_1x_1+c_2/ϵ)^p) we have _z(∫_0^ς_x_1H̅(X_s) s ≥ r) ≥ qmin{((x_1+1)^p-x_1^p)/G(r/ϵ)^p,1} ≥ C_x_1 /G(r/ϵ)^p for all r∈(c_1x_1+c_2,∞), where the first inequality in (<ref>) follows from (<ref>) above. Fix z∈ and a bounded set D⊂. Let m_D:=sup{x:(x,y)∈ D}∈ satisfy D⊂ [0,m_D]×^d∩ (note that m_D<∞ by assumption on D) and denote d_0 := max{m_D,x_1}. Since, by Lemma <ref>, X is not confined to any compact set, there exists δ > 0, such that _z(X_δ > c_1(d_0+1)+c_2) > 0. Note that, on the event {X_δ > c_1(d_0+1)+c_2}, _z-a.s. we have Z_δ∈∩ (c_1(x_1+1)+c_2,∞)×^d. The Markov property of Z and (<ref>) thus imply that for every p∈(1-β/β_c,∞) and r_0:=c_1 d_0+c_2, there exists a constant C_d_0∈(0,q((d_0+1)^p-d_0^p)/G(c_1d_0+c_2/ϵ)^p), such that _z(∫_0^ς_x_1H̅(X_s) s ≥ r) ≥_z( ∫_0^τ_δ(D)H̅(X_s) s ≥ r, X_δ > r_0+c_1) ≥_z[X_δ > r_0+c_1·_Z_δ( ∫_0^τ_δ(D)H̅(X_s) s ≥ r)] ≥_z[X_δ > r_0+c_1·_Z_δ( ∫_0^ς_d_0H̅(X_s) s ≥ r)] ≥_z(X_δ > r_0+c_1) C/G(r/ϵ)^p for all r∈(r_0,∞). We can now establish the lower bounds on the tail of the the invariant distribution of Z. Suppose that ass:domain2, ass:covariance2, ass:vector2 hold with β < -β_c and recall M_c = -(1+β/β_c)/2. Let π be the invariant distribution of the process Z and pick >0. There exists a constant c_π∈(0,∞) such that c_π r^-2M_c-ϵ≤π({z∈:z_d+1≥ r}) for all r∈[1,∞). Furthermore, for any z∈, there exists a constant c_TV∈(0,∞), such that c_TVt^-M_c-ϵ≤_z(Z_t∈·)-π_TV for all t∈[1,∞). The key step in the proof of Proposition <ref> is the following lemma. Suppose that ass:domain2, ass:covariance2, ass:vector2 hold with β < -β_c. For any ϵ>0 there exists a constant c_π∈(0,∞), such that c_π r^1+β/β_c-≤π(∩ [r,∞)×^d) for all r∈[1,∞). The reflected process Z is positive Harris recurrent, i.e., Z is Harris recurrent (by Proposition <ref>) and admits an invariant distribution (by Proposition <ref>). Thus, by <cit.>, a measurable function f: → [1,∞) satisfies ∫_z ∈f(z)π( z) < ∞ if and only if sup_z ∈ D_z[∫_0^τ_δ(D)f(Z_s) s]< ∞ for some closed petite set D⊂ and all δ > 0, where τ_δ(D)= inf{ t > δ: Z_t ∈ D} is the return time to the set D after time δ (defined before the statement of Proposition <ref> above). Since by Proposition <ref> all petite sets are bounded, Proposition <ref> implies that for any closed petite set D of Z there exists δ>0 such that for any non-decreasing continuous function H̃:→[1,∞) we have _z[∫_0^τ_δ(D)H̃(c_1(X_s+c_2)) s]≥ C∫_r_0^∞ G(r/ϵ)^-p r for all z∈ and p∈(1-β/β_c,∞), where G:→ is the inverse of the function v↦ v^2H̃(v) and the positive constants ϵ, r_0, C, c_1,c_2 are as in Proposition <ref>. By the criterion in <cit.> stated above we thus obtain that ∫_r_0^∞ G(r/ϵ)^-p r = ∞ implies ∫_z=(x,y)∈H̃(c_1(x+c_2))π( z)=∞. The proof of Lemma <ref> proceeds by contradiction. Note that the statement in the lemma is equivalent to the following: for every ϵ>0 there exists r_0∈(0,∞) such that r^1+β/β_c-≤π(∩ [r,∞)×^d) for all r∈[r_0,∞). Assume to the contrary that there exists ϵ>0, such that for every r_0>0 there exists r_1∈[r_0,∞) satisfying r_1^1+β/β_c->π(∩ [r_1,∞)×^d). We may pick r_0>1 and r_1>2r_0. Using a recursive construction, we obtain a sequence (r_n)_n∈, such that r_n+1>2r_n and r_n^1+β/β_c->π(∩ [r_n,∞)×^d) for all n∈. Using this sequence, we now construct a function H satisfying ∫_ H(x)π( z)<∞ but which violates the implication in (<ref>). Set α := -(1+β/β_c)+ and define the function μ:→ by μ(x):=1 for x∈[0,r_1) and μ(x) := r_n^-α for x∈ [r_n,r_n+1), n∈. Since the function x↦π(∩ [x,∞)×^d) is non-increasing, we have π(∩ [x,∞)×^d)≤μ(x) for all x∈. Let H:→[1,∞) be a differentiable function such that H(x) = 1 for x∈[0,r_1) and, for x∈[r_n,r_n+1), we have H'(x) = r_n^α-/2, x∈ [r_n,r_n+1); 1/(r_n+1-r_n), x∈ [r_n+1,r_n+1). Since μ is non-increasing by definition, for x∈[r_n,r_n+1), we have μ(x) ≤ r_n^-α and H'(x)μ(x) = r_n^-/2, x∈[r_n,r_n+1); r_n^-α/(r_n+1-r_n), x∈[r_n+1,r_n+1). The inequality π(∩[x,∞)×^d)≤μ(x) and Tonelli's theorem (with H(0)=1) imply ∫_z=(x,y)∈ H(x)π( z) ≤ 1+∫_r_1^∞ H'(x)π(∩ [x,∞)×^d) x ≤ 1+ ∫_r_1^∞ H'(x)μ(x) x =1+∑_n=1^∞(∫_r_n^1+r_nH'(x)μ(x) x + ∫_1+r_n^r_n+1H'(x)μ(x) x ) ≤1+∑_n=1^∞r_n^-/2 + ∑_n=1^∞ r_n^-α<∞, where the final inequality follows from 2^n-1r_1≤ r_n for every n∈ and α>>0. Let >0 and c_1, c_2∈(1,∞) whose existence is guaranteed by Proposition <ref>. Let H̃:→[1,∞) be a non-decreasing differentiable function satisfying H̃(x)=H(x/c_1- c_2) for all x∈(c_1 c_2,∞). Pick p∈(1-β/β_c,1-β/β_c+/2) and note that p>2 and 2-p+α-/2>0. Let G:→ be the inverse of the function u↦ u^2 H̃(u). Introduce the substitution r=ϵ u^2 H̃(u) into the first integral in (<ref>) to obtain ∫_1^∞G(r/)^-p r = ∫_G(1/ϵ)^∞ u^-p(2uH̃(u) + u^2H̃'(u)) u ≥∑_n=n_0^∞∫_c_1(r_n+c_2)^c_1(2r_n+c_2) u^1-pH̃(u) u= ∑_n=n_0^∞H(r_n)∫_c_1(r_n+c_2)^c_1(2r_n+c_2) u^1-p u =/(p-2)∑_n=n_0^∞ r_n^α-/2 c_1^2-p((r_n-c_2)^2-p-(2r_n-c_2)^2-p) ≥/(p-2)∑_n=n_0^∞ r_n^α-/2(r_n-c_2)^2-p (1-((2r_n-c_2)/(r_n-c_2))^2-p) ≥(1-2^2-p)/(p-2)∑_n=n_0^∞ (r_n-c_2)^2-p+α-/2 =∞, where n_0∈ is sufficiently large so that c_1(r_n+c_2)>G(1/ϵ) and r_n>c_2 hold for all n≥ n_0. The first inequality in the previous display uses the fact that H̃ is non-decreasing and positive, while the second follows from c_1>1 and 2-p<0. Note that H̃(c_1(x+c_2))=H(x) for every x∈. Thus, the implication in (<ref>) and the estimate in (<ref>) yield ∞ = ∫_z=(x,y)∈H̃(c_1(x+c_2))π( z) = ∫_z=(x,y)∈H(x)π( z). This contradicts (<ref>) and concludes the proof of the lemma. Note that π({z∈:z_d+1≥ r}) ≥π(∩ [r,∞)×^d) for every r∈. Thus, the lower bound on the invariant distribution π follows form Lemma <ref>. Pick arbitrary '∈(0,1). Lemma <ref> implies that for and γ := 1-β/β_c - ', there exists w∈(-∞,1-β s_0/c_0), some constant x_0,x_1,d∈, defining the function F_w,γ, and a constant C_3∈(0,∞) such that _z[F_w,γ(Z_t)] ≤ F_w,γ(z)+C_3t for all t∈ and z∈. Since F_w,γ(z) = f_w,γ(z) on z=(x,y)∈∩[x_1,∞)×^d, the upper bound in (<ref>) implies that π({z∈:F_w,γ(z) ≥ r}) ≥π({z=(x,y)∈:x ≥ 2^γ/|w|r^1/γ-k_w}) for all r∈(x_1,∞). This inequality and the lower bound on the invariant distribution in Lemma <ref> imply that there exists a constant C_4∈(0,1) such that π({z∈:F_w,γ(z) ≥ r}) ≥ C_4r^(1 + β/β_c - ϵ')/γ= C_4r^(1 + β/β_c - ϵ')/(1-β/β_c-ϵ') for all r∈(x_1,∞). By further reducing C_4>0 if necessary, we may assume that the inequality in the last display holds for all r∈[1,∞). Define the functions f: [1,∞) → (0,1], f(a) := C_4 a^(1 + β/β_c - ϵ')/(1-β/β_c-ϵ') and F(a):=af(a)= C_4 a^(2 -2ϵ')/(1-β/β_c-ϵ') and note that F is strictly increasing with lim_a →∞ F(a) = ∞. By Lemma <ref>, applied with functions G(z) = F_w,γ(z) (recall from the previous display that π({z∈:G(z)≥ r})≥ f(r) for all r∈[1,∞)) and g(z,t) := F_w,γ(z)+C_3 t≥_z[G(Z_t)], yields constants C_2,C_5∈(0,∞), such that for all t∈[1,∞) we get C_2t^(1+β/β_c-')/(2-2')≤C_5/2(2g(z,t))^(1+β/β_c-ϵ')/(2-2ϵ')≤_z(Z_t_n∈·)-π_TV. Pick arbitrary ∈(0,1) and let ϵ'∈(0,1) be such that 0>(1+β/β_c-')/(2-2')>(1+β/β_c)/2-. Then the bound in the proposition follows from (<ref>). § A LOWER BOUND ON THE CONVERGENCE TO STATIONARITY OF A MARKOV PROCESS Fix m∈ and let κ=(κ_t)_t∈ be a Markov process on an unbounded domain _κ in ^m with an invariant distribution π_κ satisfying π_κ(·) = ∫__κ_u(κ_t∈·) π_κ( u) for every t>0. Via a suitable Lyapunov function, the following lemma converts a lower bound estimate on the tail of invariant distribution π_κ into a lower bound on the rate of convergence in total variation of the law of κ_t to the invariant distribution π_κ. The elementary proof of Lemma <ref> below is adapted from <cit.>. This lemma is key in the proof of the lower bound on the rate of convergence in total variation stated in Theorem <ref>(b). Let _κ be an an unbounded domain in ^m and κ = (κ_t)_t∈ a _κ-valued Markov process with invariant distribution π_κ. Assume the function G:_κ→ [1,∞) satisfies: (a) there exists f:[1,∞)→(0,1], such that the function F:y↦ yf(y) is increasing, lim_y↑∞F(y)=∞ and π_κ({v∈_κ:G(v)≥ y})≥ f(y) for all y∈[1,∞); (b) there exists g:_κ×→[1,∞), such that for every u∈_κ the function t↦ g(u,t) is continuous and increasing to infinity and _u[G(κ_t)] ≤ g(u,t) for all t∈. Then, for any starting point u∈_κ we have π_κ(·)-_u(κ_t∈·)_TV≥1/2f(F^-1(2g(u,t))) for all t∈. A good choice for the function G in Lemma <ref> has the following properties: the expectation _u [G(κ_t)] is bounded as a function of the starting point u and time t and the function y↦π_κ(G^-1([y,∞))) satisfies lim_y→∞π_κ(G^-1([y,∞)))=0 and lim_y→∞yπ_κ(G^-1([y,∞)))=∞. The proof of Lemma <ref> shows that if the assumption in (a) holds for y sufficiently large, then the conclusion of the lemma is valid for all t sufficiently large. The definition of the total variation distance (together with assumption (a)) and the Markov inequality (together with assumption (b)) imply that for every u∈_κ and t∈ the following inequalities hold for all y∈[1,∞): π_κ(·)-_u(κ_t∈·)_TV≥π_κ({v∈^d:G(v)≥ y})-_u(G(κ_t)≥ y)≥ f(y) -g(u,t)/y. Since F(y)→∞ (as y→∞) and t↦ g(u,t) is increasing and continuous, for all y∈[F^-1(2g(u,0)),∞) (where F^-1 is the inverse of the increasing function F defined in (a)), there exists a unique t∈ satisfying F(y)=2g(u,t)∈[1,∞). Differently put, for every t∈, there exists y_t∈[F^-1(1),∞)⊂ [1,∞) satisfying y_t=F^-1(2g(u,t)). Thus, for every t∈, we have f(y_t)-g(u,t)/y_t=f(F^-1(2g(u,t)))/2. § ASYMPTOTICALLY OSCILLATING DOMAIN Let b:→ (0,∞) be a C^2 function with b(0)=0, satisfying b(x) = loglog x(1 + (loglog x)^-2 + sinloglog x) for x > exp(+1). Then lim sup_x→∞b(x) = ∞ and lim inf_x →∞b(x) = 0. Moreover the function b satisfies assumption ass:domain2 with β=0, i.e. lim_x→∞xb'(x)/b(x)=β=0, and lim_x→∞b'(x)=lim_x→∞b”(x)=0 To show lim inf_x→∞b(x) = 0, consider ℓ_k := exp(exp(-π/2 + 2kπ)), for any k∈. It follows that lim_k→∞b(ℓ_k) = lim_k→∞(loglogℓ_k)^-1= lim_k→∞(-π/2 + 2kπ)^-1 = 0. Similarly, to show lim sup_x→∞b(x)=∞, consider ℓ̃_k := exp(exp(2kπ)), for any k∈. We obtain lim_k→∞b(ℓ̃_k) = lim_k→∞(ℓ̃_k + (loglogℓ̃_k)^-1)= lim_k→∞2kπ + (2kπ)^-1 =∞. The first two derivatives of b on x > exp(+1) take the form b'(x) = (xlog x)^-1(1-(loglog x)^-1 + sinloglog x + cosloglog x), b”(x) = (xlog x)^-2((1+log x)(1 -(loglog x)^-2 + sinloglog x+cosloglog x) + 2(loglog x)^-3), implying that lim_x→∞ b'(x) =lim_x→∞ b”(x) = 0. For the result about β it is enough to show lim sup_x→∞ |xb'(x)|/b(x) ≤ 0. We estimate lim sup_x→∞xb'(x)/b(x) = lim sup_x→∞(log x)^-1|1-(loglog x)^-1 + sinloglog x + cosloglog x)/loglog x(1 + (loglog x)^-2 + sinloglog x) ≤lim sup_x→∞loglog x(3 + (loglog x)^-1)/log x = 0, this concludes the proof. § ACKNOWLEDGEMENTS MB is funded by the CDT in Mathematics and Statistics at The University of Warwick. AM is supported by EPSRC under grants EP/W006227/1 & EP/V009478/1 and by The Alan Turing Institute under the EPSRC grant EP/X03870X/1. The work of AW is supported by EPSRC grant EP/W00657X/1. amsplain

http://arxiv.org/abs/2303.07166v1

20230313150952

Improved Tree Search for Automatic Program Synthesis

cs.LG

[ Improved Tree Search for Automatic Program Synthesis Aran Carmontau Lior Wolftau tauThe School of Computer Science , Tel Aviv University Lior Wolfliorwolf@gmail.com Aran Carmonarancarmon@mail.tau.ac.il Automatic Program Synthesis, Exploration, MCTS 0.3in ] In the task of automatic program synthesis, one obtains pairs of matching inputs and outputs and generates a computer program, in a particular domain-specific language (DSL), which given each sample input returns the matching output. A key element is being able to perform an efficient search in the space of valid programs. Here, we suggest a variant of MCTS that leads to state of the art results on two vastly different DSLs. The exploration method we propose includes multiple contributions: a modified visit count, a preprocessing procedure for the training dataset, and encoding the part of the program that was already executed. § INTRODUCTION Search is a key part of many machine learning tasks, in which the output is a sequence. We study the specific task of automatic program synthesis, given a specification in the form of input/output pairs <cit.>. In this sequence generation task, similar to other generation tasks, such as machine translation and image captioning, there are multiple correct answers. Different from most sentence generation tasks in NLP, in this task, one is able to directly evaluate the correctness of the output, by running the generated program. This leads to a well-defined and natural reward, when viewed as a reinforcement learning problem <cit.>: either the generated program produces the specified outputs given the matching inputs or not. Other variants of this reward may consider, for example, the length of the program, encouraging the generated program to be more efficient. However, as reported by previous work <cit.>, employing a reinforcement learning approach, as opposed to training using a maximum likelihood loss to generate the single program that is available as the ground truth, either hurts performance or leads to a small increase in performance. This is despite training the MLE approach in a teacher-forcing way, in which, during training and unlike during test time, the partial programs considered are the prefix of the ground truth programs. In this work we, therefore, focus on the MLE approach and consider different tree search strategies. These methods include the classical beam search, which is often used in the program synthesis domain, the CAB variant of it, which was used successfully in the past <cit.>, and various MCTS approaches that we develop and explore. In addition to studying the specific building blocks of the application of MCTS to this problem, we also suggest two other improvements. The first is a pre-processing step that is applied to the training set, and the second is the addition of an encoder of the current program's history. We demonstrate how these techniques can improve the obtained accuracy, both separately and when combined. § RELATED WORK We experiment with two DSLs. The first is the DeepCoder DSL <cit.>, in which integer registers are being manipulated. Each register may contain either an integer or a list of integers, and the program may apply on the registers, functions such as sort, tail, and multiply. <cit.> have used the predictions of a neural network that was applied to the input/output pairs, in order to augment classical program search techniques, such as SMT-solvers and enumerative search. Specifically, the neural network produced a vector of probabilities for the existence of every command or function in the program. Later on, <cit.> used a neural network to predict the probability of the next statement given the execution state, after applying the partial program as part of a beam search. Specifically, the CAB search strategy <cit.> was used. Since, in this DSL, every statement returns one integer or list into the memory, and since this memory is limited, a learned garbage collection mechanism was used to discard the results of previous statements from the memory. Our work on this DSL extends the method of <cit.>, and also employs this garbage collector. The Karel DSL <cit.> acts on a 2D grid world that contains an actor, various markers, and obstacles. The Karel DSL contains conditions and loops, which are not part of the DeepCoder DSL. However, the latter has some additional high-level functions, such as sort or map. <cit.> have demonstrated how to perform RL for improving the results on this DSL. <cit.> have contributed the encoding of the current execution state (similar to <cit.>), as well as the usage of ensembles. In our work, we do not study the effect of ensembles. Beam search is perhaps the most popular method for exploring search trees. It is similar in nature to a best-first search, but instead of keeping in memory just the most promising candidate, beam search keeps several candidate tracks simultaneously. The number of candidates is the beam width, and the number of the derived expansions extracted from each candidate track (prior to pruning back to the beam width) is the extraction width. Complete Anytime Beam (CAB) search <cit.> is an extension of beam search, which is better-suited for searching under a fixed time budget. In beam search, if the beam width and expansion size are configured too low, the beam search will be too weak to reach its destination. If they are too high, the beam search will reach the execution time limit, before achieving its goal. CAB extends beam search by running it repeatedly, starting with weak and cheap settings and increasing the search parameters after each search failure, as long as it still has time to run. Monte Carlo Tree Search (MCTS) <cit.> is a popular method for prioritizing exploration in game search trees. The search is gradually expanded towards a mix of promising nodes and unexplored nodes. It includes the following steps: (1) Selection: traverse the search tree, until a new leaf node is reached. (2) Expansion: add the leaf node to the tree. (3) Simulation: evaluate the value of the leaf node (4) Backpropogation: update the values of the parent nodes on the track leading to the leaf node, according to its value. In our work, we do not employ a value function and the last two steps are not employed. A major milestone in the field of artificial intelligence was the achievement of super-human performance in Go by using deep learning, reinforcement learning, and MCTS <cit.>. Our method shares with it the specific fraction that is based on the visit count. However, we advocate for performing the count at the environment level, while in the previous Go work the count is based on the location in the tree and not on the board configuration. Our method does not use a learned or a simulation (rollout) based value function, as done by <cit.>. We have tried learning such a value function and observed no improvement in the results. This may stem from the fact that the end goal, i.e., the output of our programs, varies, making such learning too challenging. In addition, as noted for the case of solving Rubik's cube with RL <cit.>, the value function mainly reduces the depth of the MCTS and not its breadth. In our case, since the length of the programs is limited, the depth is less of a concern. § METHOD Our method is based on applying MCTS in lieu of beam search or CAB, which served as the search technique for previous work. We employ the same networks that are used in <cit.>. An encoder E embeds the input states of the execution, as well as the output for each sample and mean pooling is employed in order to obtain a single state vector. Based on this vector, a network P predicts the next statement and a network G predicts the variables to be dropped. P and G share most of their layers. In our work, we propose to add, as an additional input to the network P, the embedding of the partial program that was executed up to the current state, see Sec. <ref>. §.§ Our MCTS variant Given a trained network P that can prioritize the most likely next commands, we can explore the programs' space with a search tree. The nodes of the tree correspond to the execution states, after choosing the statements. The search favors nodes that have been less explored in the past (exploration), as well as nodes that result from commands that are predicted to have higher likelihood (exploitation). The score U(s,a) = P(s,a)/N(s,a)+1 is used, where P(s,a) is the output of the trained network (pseudo-probability of choosing action a, i.e., command, given state s), and N(s,a) is the number of times we took action a, as a response to being in state s. Unlike the literature we are aware of, in our variant of the search we do not backtrack. Instead, we start the search each time from the root node. At first, we pick the most promising command according to the network's output, until we reach a solution, an invalid execution state, or a maximum program length limit. If a solution was not found, we start the search over, this time employing the updated score U, which discounts nodes that have been visited. Shared count While the MCTS methods we are aware of count the number of visits per tree node, we found it useful to count the number of times each state is visited. A state is the memory status obtained after running the partial program (from the root node to the current node) on each of the inputs in the input/output specifications. Two states are identical, if the memory status is identical for all of the inputs in the specification. In the DeepCoder DSL, the memory contains the various registers. We sort the registers such that two states are the same, even if the order of the registers is permuted, as long as the underlying values are the same. In the Karel DSL, the memory contains the status of the grid world. §.§ Pruning Since there are multiple execution pathways that result in the same states, learning with an MLE loss could be suboptimal. Previous work <cit.> has suggested employing RL to overcome this; However, as noted in subsequent work, this does not necessarily lead to significantly improved performance. We suggest, instead, to preprocess the training set such that programs are replaced with shorter programs that are consistent with the input/output specifications. For this task, we employ a network that is pretrained on the original training data. If when applying this network, with our variant of MCTS, to the training dataset, a shorter (specification consistent) program is found, we replace the ground truth program. After this pruning, training is repeated from scratch. Pruning has a drastic effect on the length of programs in the training datasets, as can be seen in Fig. <ref>. For example, for the DeepCoder DSL, it shortens 82% of the length 12 programs, which is the maximal length available in the training dataset. §.§ Encoding of Partial Programs The input in PCCoder <cit.> to the network P, which predicts the next statement, is the embedding of the current registers of the execution environments (one environment per each sample input/output), including the end-goal (output) register of each sample. For Karel, we employ a similar architecture to PCCoder. In this case, P receives an embedding of the input/output grids and an embedding of the last command executed, where the latter part was added following <cit.>. In this work, we propose to add an additional input, which encodes the previous commands executed by the partial program that has already been constructed during the search. For the DeepCoder DSL, the partial program is given as a sequence of 4-tuples containing an operator index, argument indices (one or two, if the operator takes one argument, we duplicate it), and an output index. We apply embedding layers on these four indices. LSTM is then applied to the concatenation of the four embeddings. The number of hidden units is taken to be the sum of the number of options in each look-up-tables we employ for the index embedding (66). The hidden state at the end of the sequence is then concatenated to the activations that arise from the other inputs of P, after the 5 densenet layers of P are applied to these inputs. Two fully-connected layers (of the same size) are then applied, followed by a softmax layer of the same dimensionality, as in the original P. In Karel, the input and output grids are encoded, using the same convolutional networks used by <cit.>. The partial program is encoded by running two LSTMs: the first is run on the embedding of the commands (there are no arguments) and the second on the nesting level (is the command executed within an if condition, a while loop, etc). The LUTs contain 38 command options and six nesting options. The LSTMs have a corresponding number of hidden units. The hidden states of the two LSTMs at the end of the sequence are concatenated to the embedded input/output samples (after the CNN encoding) and passed through three linear layers, with 512 hidden units. This is done separately for each input/output sample, and is aggregated by max pooling, followed by a softmax layer. § EXPERIMENTS We test the various search methods and evaluate the effect on performance of each of our contributions. In all of our experiments, the PCCoder architecture is used. Applying this method for Karel involved the same input/output encoding used in <cit.>. Since, unlike this previous work, we do not use an LSTM to generate the program (PCCoder predicts only the next command), the LSTM is replaced with a fully connected network. A garbage collection network G is not required in this DSL. Note that given enough time, one can find, using a naïve search approach, the programs, up to a certain length, that are consistent with the input/output pairs. The shortest such program can then be selected, in order to improve generalization. Therefore, it is important to enforce a maximal allowed run time. This has obvious drawbacks when comparing across contributions. However, given the same standard hardware, a fixed runtime budget provides a reliable performance metric. In our experiments, we employ a commercial cloud infrastructure to perform all experiments. The DeepCoder DSL experiments were done with eight registers and the benchmark of <cit.> that includes training programs of lengths up to 12, and ground truth test programs of length 14 (all test environments are constructed with programs of length 14, which may or may not have shorter equivalents). The results are given in Tab. <ref>. As can be seen, all of our contributions (pruning of the dataset, encoding the partial program, replacing CAB with our variant of MCTS, and the shared count MCTS) improve performance over the baseline. This is true for the three different values of timeout. Our best result, which includes all contributions, presents an accuracy of 85.2%, in comparison to 50.2% of the previous work, for the same timeout of 2000 seconds. Partial program encoding by itself adds 15% to the CAB search. Pruning has a less dramatic effect: it adds 3% to CAB. In addition, pruning does not add to the shared visit count MCTS method by itself. However, when combined with encoding of the partial history, it adds significantly, e.g., 5.2% for the 500 second timeout. In all four groups obtained by applying or not applying pruning or partial program encoding, the MCTS method with the shared counts outperforms the other search method by a significant margin, and our variant of MCTS without the shared counts outperforms CAB. The Karel benchmark used is taken from <cit.> and is also used by <cit.>. However, on this benchmark, we found it challenging to compare directly with previous work, since the previous work did not employ a timeout. Instead, a fixed number of beams was used, which is not compatible with our MCTS variant going back to the root at each failed search. Performance is measured by exposing five input/output samples to the searcher and measuring if the searcher succeeded in finding a program which is not only consistent with the five input/output examples, but also generalizes to a sixth unseen input/output sample. In Tab. <ref>, we present accuracy for both goals (consistency with the five or the six samples). The results are given for timeouts of 1000 or 2000 seconds. Beam search results are given for the best width and expansion parameters that we have found (128 and 5, respectively). The CAB parameters were also optimized to maximize the test accuracy, and starting with a similar beam search, it doubles the beam width and increases the expansion size by one, at each repeated attempt. For this DSL, the combination of MCTS with past encoding and dataset pruning obtains the best results in both timeouts and for both success criteria. Sharing the visit counts between identical world grids did not improve results and the added bookkeeping increased runtime, leading to slightly lower results. § ACKNOWLEDGEMENT This work was supported by an ICRC grant. icml2019

http://arxiv.org/abs/2303.07299v1

20230313172417

One-form symmetries in $\mathcal{N} = 3$ $S$-folds

hep-th

Improvement of Neutron-HP package: Doppler broadening of the neutron elastic scattering kernel and cross sections [ March 30, 2023 ================================================================================================================= § INTRODUCTION In the recent past, a new paradigm for the notion of symmetry in QFTs became dominant. It is based on the necessity to include higher-form symmetries and the corresponding extended objects in the description of quantum field theories <cit.>. Restricting to four-dimensional QFTs, the simplest way to proceed consists in classifying the one-form symmetries in supersymmetric and conformal theories (SCFTs). A seminal paper that allowed for such a classification has been <cit.> where a general prescription was given in terms of the spectrum of mutually local Wilson and 't Hooft lines <cit.>. Such a prescription was initially based on the existence of a Lagrangian description for the SCFT under investigation. In absence of a Lagrangian description it is nevertheless possible to use other tools, coming from supersymmetry, holography and/or branes. These constructions have allowed to figure out the one-form symmetry structure of many 4d non-Lagrangian SCFTs constructed in various ways <cit.>. A class of theories that has not been deeply investigated so far are SCFTs with 24 supercharges, i.e. 𝒩=3 conformal theories. Such models have been predicted in <cit.>, and then found in <cit.>. Many generalizations have been then studied by using various approaches <cit.>. A key role in the analysis of <cit.> is based on the existence, in the string theory setup, of non-perturbative extended objects that generalizes the notion of orientifolds, the S-folds (see <cit.> for their original definition). From the field theory side, the projection implied by such S-folds on 𝒩=4 SYM has been associated to the combined action of an R-symmetry and an S-duality twist on the model at a fixed value of the holomorphic gauge coupling, where the global symmetry is enhanced by opportune discrete factors. Four possible ℤ_k have been identified, corresponding to k=2, 3, 4 and 6. While the ℤ_2 case corresponds to the original case of the orientifolds <cit.>, where actually the holomorphic gauge coupling does not require to be fixed, the other values of k correspond to new projections that can break supersymmetry down to 𝒩=3. The analysis has been further refined in <cit.>, where the discrete torsion, in analogy with the case of orientifolds, has been added to this description. In this way, it has been possible to achieve a classification of such 𝒩=3 S-folds SCFT in terms of the Shephard–Todd complex reflection groups. The goal of this paper consists in classifying one-form symmetries for such theories, constructing the lattices of lines and identifying which models possess non-invertible symmetries. The main motivation behind this expectation is that for the rank-2 S-folds, in absence of discrete torsion, the SCFTs enhance to 𝒩=4 SYM <cit.> where these properties are present. Our strategy adapts the one presented in <cit.> to S-fold setups. There, the spectrum of lines is built from the knowledge of the electromagnetic charges of massive states in a generic point of the Coulomb branch. These charges are read from the BPS quiver, under the assumption that the BPS spectrum is a good representative of the whole spectrum of electromagnetic charges. In the case of S-folds however such a BPS quiver description has not been worked out and we extract the electromagnetic charges of dynamical particles from the knowledge of the (p,q)-strings configurations in the Type IIB setup <cit.>. The main assumption behind the analysis is that such charges are a good representative of the electromagnetic spectrum. We proceed as follows. First we choose an 𝒩=3 theory constructed via an S-fold projection of Type IIB. This consists in having N D3-branes, together with their images, on the background of an S-fold. At a generic point of the Coulomb branch, the corresponding low energy gauge dynamics corresponds to a U(1)^N gauge theory where each U(1) is associated to a D3. Then we list all (p,q)-strings that can be stretched between D3-branes and their images. They have electric and magnetic charges with respect to U(1)^N. Eventually we run the procedure of <cit.>. This consist in finding all the lines that are genuine, i.e. have integer Dirac pairing with the local particles, modulo screening by the dynamical particles. This gives the lattice of possible charges, then the different global structures correspond to maximal sub–lattices of mutually local lines. Our results are summarized in <Ref>. In the first column, one finds the type of S-fold projection that has been considered. Such projections are identified by the two integers k and ℓ in S_k,ℓ. The integer k corresponds to the ℤ_k projection while the second integer ℓ is associated to the discrete torsion. Then, when considering an S_k,ℓ S-fold on a stack of N D3-branes the complex reflection group associated to such a projection is G(k,k/ℓ,N). In the second column, we provide the one-form symmetry that we found in our analysis, and in the third, the number of inequivalent line lattices that we have obtained. The last column specifies whether there exist cases that admit non-invertible symmetries. Indeed, here we find that in some of the cases there exists a zero-form symmetry mapping some of the different line lattices, that are therefore equivalent. Furthermore in such cases we expect the existence of non-invertible symmetries obtained by combining the zero-form symmetry with a suitable gauging of the one-form symmetry. A remarkable observation strengthening our results regards the fact that our analysis reproduces the limiting G(k,k,2) cases, where supersymmetry enhances to 𝒩=4 with 𝔰𝔲(3), 𝔰𝔬(5) and 𝔤_2 gauge groups for k=3, 4 and 6 respectively. Another check of our result is that it matches with the cases G(3,1,1) and G(3,3,3), where an 𝒩=1 Lagrangian picture has been worked out in <cit.>. Note added: When concluding this paper, the reference <cit.> appeared on arXiv. There, they study the classification of zero, one and two-form symmetries in 𝒩=3 S-fold SCFTs. Their analysis is holographic, along the lines of the construction of <cit.> for 𝒩=4 SYM. We have checked that our results are in agreement with their predictions. § GENERALITIES §.§ Global structures from the IR The strategy adopted here, as already discussed in the introduction, is inspired by the one of <cit.>. The main difference is that instead of using BPS quivers, not yet available for our S-folds, we take advantage of the type IIB geometric setups and probe the charge spectrum with (p,q)-strings – the bound state of p fundamental strings F1 and q Dirichlet strings D1.[In order to provide the IR spectrum of line operators of the SCFTs from this UV perspective, we assume the absence of wall-crossing. While such an assumption is a priori motivated by the high degree of supersymmetry, a posteriori it is justified by the consistency of our results with the literature.] Despite this difference, the rest of the procedure is the one of <cit.> which we now summarize. Denote as γ^i = (e^(i)_1, m^(i)_1; …; e^(i)_r, m^(i)_r) a basis vector of the electromagnetic lattice of dynamical state charges under the U(1)_e^r × U(1)_m^r gauge symmetry on the Coulomb branch. The spectrum of lines can be determined by considering a general line ℒ with charge ℓ=(e^(l)_1, m^(l)_1; …; e^(l)_r, m^(l)_r) . This is a genuine line operator if the Dirac pairings with all dynamical states Ψ are integer: ⟨Ψ, ℒ⟩∈ℤ ∀ Ψ . This can be rephrased as the condition ∑_j=1^r e^(i)_j m^(l)_j - m^(i)_j e^(l)_j ∈ℤ ∀ i . Furthermore, inserting a local operator with charge γ_i on the worldline of a line with charge ℓ shifts its charge by γ_i. Therefore if a line with charge ℓ appears in the spectrum then a line with charges ℓ + ∑ k_i γ_i with k_i ∈ℤ must also appear. When classifying the spectrum of charges of the line operators of a QFT it is then useful to consider the charges ℓ modulo these insertions of local states. This gives rise to equivalence classes of charges with respect to the relation: ℓ∼ℓ + γ_i ∀ i . Borrowing the nomenclature of <cit.>, we will refer to such identification as screening and we will work with each equivalence class by picking one representative. The genuine lines after screening form a lattice. In general two such lines are not mutually local and a choice of global structure corresponds to a choice of a maximal sublattice of mutually local lines. §.§ Charged states in S_k,l-folds We aim to determine the electromagnetic charges of the local states generated by (p,q)-strings stretched between (images of) D3-branes in presence of an S-fold. The S-fold background of Type IIB string theory consist of a spacetime ℝ^4 × (ℝ^6/ℤ_k) where the _k quotient involves an S-duality twist by an element ρ_k ∈ SL(2,) of order k, where k=2,3,4,6. For k>2 the value of the axio-dilaton vev is fixed by the requirement that it must be invariant under the modular transformation associated to ρ_k. The matrices ρ_k and the corresponding values[In our convention, an SL(2,ℤ) transformation of the axio-dilaton τ→ (aτ + b) / (cτ +d) relates to a matrix ρ_k = [ d c; b a ]. We also have S = [ 0 -1; 1 0 ] and T = [ 1 0; 1 1 ].] of τ are given in Table <ref>. A stack of N D3-branes probing the singular point of the S-fold background engineer an 𝒩=3 field theory on the worldvolume of the stack of D3-branes. It is useful to consider the k-fold cover of spacetime, and visualize the N D3-branes together with their (k-1)N images under the S_k-fold projection. We are going to label the m-th image of the i-th D3-brane with the index i_m, where i=1,…,N and m=1,…,k. Under the S-fold projection, the two-form gauge fields of the closed string sector B_2 and C_2 transform in the fundamental representation: ([ B_2; C_2 ]) →ρ_k ([ B_2; C_2 ]) . Consistently, the (p,q) strings charged under these potentials are mapped to (p',q') where: (p' q') = (p q)·ρ_k^-1 . We denote a state associated to a (p,q) connecting the i_m-th D3-brane and the j_n D3-brane as: |p,q⟩ _i_m, j_n= |-p,-q⟩ _j_n , i_m , where we identity states with both opposite charges and orientation. First, strings linking branes in the same copy of ℝ^6/ ℤ_2 transform as follows: |p,q⟩ _i_m, j_m→ζ_k^-1 |p',q'⟩_i_m+1,j_m+1 , where (p',q') are related to (p,q) by (<ref>) and ζ_k is the primitive k-th root of unity. These states always collectively give rise to a single state in the quotient theory, with charges: D3_iD3_j : (0,0 ; … ;p,q^i-th; …; -p,-q^j-th; … ; 0,0) . An important ingredient we need to add to our picture is the discrete torsion for B_2 and C_2 <cit.>. In presence of such a discrete torsion, a string going from the i_m-th brane to the j_m+1-th brane should pick up an extra phase which depends only on its (p,q)-charge and the couple (θ_NS,θ_RR). More precisely, one expects that the S-fold action can be written as follows <cit.>:[We thank Shani Meynet for pointing out <cit.> to us.] |p,q⟩_i_mj_m+1→ζ_k^-1 e^2π i (pθ_NS+qθ_RR)|p',q'⟩_i_m+1j_m+2 , where again (p',q') are related to (p,q) by (<ref>). For i≠ j, this always leads to the following state in the projected theory <cit.>:[The action on (p,q) involves ρ_k^-1, see (<ref>). In writing (<ref>) however, we measure the charge with respect to the brane in the chosen fundamental domain, hence the appearance of ρ_k instead of its inverse.] D3_iD3_j^ρ : (0,0 ; … ;p,q^i-th; …; -(p q)·ρ_k^j-th; … ; 0,0) . Note that this is the only case that might not lead to any state in the quotient theory when i=j, i.e. when a string links a brane and its image. When the quotient state exists, it has charges D3_iD3_i^ρ : (0,0 ; … ;(p q)-(p q)·ρ_k^i-th; … ; 0,0) . Analogously, strings twisting around the S-fold locus n-times pick up n-times the phase in (<ref>). A last remark is that discrete torsion allows some strings to attach to the S-fold if the latter has the appropriate NS and/or RR charge. If this is the case, the state is mapped as in (<ref>): |p,q⟩_S_k i_m→|p^',q^'⟩_S_k i_m+1 , and provides the following charge in the projected theory: S_k D3_i : (0,0 ; … ;p,q^i-th; … ; 0,0) . These rules are illustrated and details on discrete torsion are provided in the remaining of this section for orientifolds and S-folds separately. §.§.§ The case with k=2: orientifolds In this subsection we apply the formalism described above for orientifolds and reproduce the spectrum of strings known in the literature. The matrix ρ_2 is diagonal, therefore the two p and q factors can be considered independently. In this case the field theory obtained after the projection is Lagrangian and can be studied in perturbative string theory with unoriented strings. Discrete torsion takes value in (θ_NS,θ_RR)∈_2 ⊕_2, giving four different choices of O3-planes related by SL(2,ℤ) actions <cit.>, see <Ref>. The orientifold action is then recovered from (<ref>) and (<ref>) with ζ_2 = -1. First, we have |p,q⟩_i_1 j_1→ - |-p,-q⟩_ i_2 j_2 = -|p,q⟩_ j_2 i_2 . For the strings that stretch from one fundamental domain of ℝ^6/ℤ_2 to the next, there are four cases depending on the values of θ_NS and θ_RR: [ O3^- : |p,q⟩_i_1 j_2→ -|p,q⟩_ j_1 i_2 ,; O3^+ : |p,q⟩_i_1 j_2→ -e^pπ i |p,q⟩_j_1 i_2 ,; O3^- : |p,q⟩_i_1 j_2→ -e^qπ i |p,q⟩_j_1 i_2 ,; O3^+ : |p,q⟩_i_1 j_2→ -e^(p+q)π i |p,q⟩_j_1 i_2 . ] It is interesting to consider strings connecting one brane to its image, i=j. In the case of trivial discrete torsion, corresponding to the O3^--plane, all such strings are projected out. On the contrary, in the O3^+ case, an F1-string linking mirror branes survives the projection, while a D1-string similarly positioned is projected out. We also find strings that can attach to the different orientifolds following <cit.>: O3^- : none , O3^+ : |0,1⟩_O3^+ i_m , O3^- : |1,0⟩_O3^- i_m , O3^+ : |1,1⟩_O3^+ i_m , as well as bound states of these. §.§.§ The cases with k>2: S-folds The construction discussed above can be applied to S_k>2 in order to obtain the string states in the quotient theory. For k>2, the discrete torsion groups have been computed in <cit.>, the result being θ_NS=θ_RR∈_3 for the S_3-case and θ_NS=θ_RR∈_2 for the S_4-case. The S_6-fold does not admit non-trivial discrete torsion. It was also pointed out that, for the S_3-case, the choices θ_NS=θ_RR=1/3 and θ_NS=θ_RR=2/3 are related by charge conjugation; therefore everything boils down to whether the discrete torsion is trivial or not. Following the notation of <cit.>, we denote as S_k,1 the S-folds with trivial discrete torsion and as S_k,k the S-folds with non-trivial discrete torsion. As before, the only states that might not lead to any state in the quotient theory are the strings linking different covers of ℝ^6/ℤ_k. These transform as follows: [ S_3,1 : |p,q⟩_i_1 j_m+1→ e^-i2π/3|q-p,-p⟩_i_2 j_m+2 ,; S_3,3 : |p,q⟩_i_1 j_m+1→ e^-i2π/3e^im(p+q)2π/3 |q-p,-p⟩_i_2 j_m+2 ,; S_4,1 : |p,q⟩_i_1 j_m+1→ e^-iπ/2 |-q,p⟩_i_2 j_m+2 ,; S_4,4 : |p,q⟩_i_1 j_m+1→ e^-iπ/2e^im(p+q)π |-q,p⟩_i_2 j_m+2 ,; S_6,1 : |p,q⟩_i_1 j_m+1→ e^-iπ/3 |p-q ,p⟩_i_2 j_m+2 .; ] This shows that no state is projected out for S_3,1 and S_3,3. Analogously to the orientifold cases, we project out some strings linking mirror branes: |p,q⟩_i_n i_n+2 in S_4,1 and S_4,4, and |p,q⟩_i_n i_n+3 in S_6,1 respectively. Finally, we get extra strings linking the S-fold to D-branes for the cases with discrete torsion. Following the discussion in <cit.>, we know that these S-folds admit all kinds of p and q numbers: S_3,3 : |p,q ⟩_S_3,3 i_n , S_4,4 : |p,q⟩_S_4,4 i_n . §.§ Dirac pairing from (p,q)-strings Having determined the states associated to (p,q)-strings that survive the S-fold projection we now analyze the electromagnetic charges of these states. It is useful to consider the system of a stack of D3-branes and an S_k,ℓ-fold on a generic point of the Coulomb branch. This corresponds to moving away the D3-branes from the S-plane. On a generic point of the Coulomb branch, the low energy theory on the D3-branes is a U(1)_i^N gauge symmetry, where each U(1)_i factor is associated to the i-th D3-brane. The theory includes massive charged states generated by the (p,q)-strings studied in the previous section. A (p,q)-string stretched between the i-th and j-th D3-brane has electric charge p and magnetic charge q under U(1)_i as well as electric charge -p and magnetic charge -q under U(1)_j, and is neutral with respect to other branes. We organize the charges under the various U(1)s in a vector: (e_1, m_1; e_2,m_2; …; e_N,m_N) where e_i and m_i are the electric and magnetic charge under U(1)_i, respectively. In this notation the charge of a string stretched between the i-th and j-th D3-brane in the same cover of ℝ^6/ℤ_2 has charge: D3_iD3_j: (0,0;…; p,q^i-th;0,0; …; -p,-q^j-th;… ) , where the dots stand for null entries. We will keep using this notation in the rest of the paper. A (p,q)-string stretched between the i-th D3-brane and the l-th image of the j-th D3-brane imparts electromagnetic charges (p,q) under U(1)_i and charges - (p,q)ρ_k^l under U(1)_j. In formulas: D3_iD3_j^ρ^l: (0,0;…; p,q^i-th;0,0; …; -(p q)·ρ_k^l^j-th;… ) . The last ingredient for our analysis is given by the Dirac pairing between two states. Consider a state Ψ with charges e_i,m_i under U(1)_i and a state Ψ' with charges e'_i,m'_i under U(1)_i. The pairing between F1 and D1-strings in Type IIB dictates that the Dirac pairing between these states is given by: ⟨Ψ, Ψ' ⟩ = ∑_i=1^N (e_i m'_i - m_i e'_i) . By using this construction we can reproduce the usual Dirac pairing of 𝒩=4 SYM with ABCD gauge algebras. As an example we now reproduce the Dirac pairing of D_N, engineered as a stack of N D3-branes probing an O3^--plane. In this case the allowed (p,q)-strings have the following charges: D3_iD3_j: (0,0;…; p,q^i-th;0,0; …; -p,-q^j-th;… ) D3_iD3_j^ρ: (0,0;…; p,q^i-th;0,0; …; p,q^j-th;… ) The states associated to (1,0)-strings correspond to the 𝒲 bosons while the states associated to (0,1)-strings correspond to magnetic monopoles ℳ. For each root 𝒲_i of D_N let ℳ_i be the corresponding coroot. More precisely if 𝒲_i is associated to a (1,0)-string connecting two D3-branes, then the coroot ℳ_i corresponds to the string (0,1) stretched between the same pair of D3-branes. The only non-vanishing Dirac pairing is the one between a 𝒲_i boson and an ℳ_j monopole. This pairing between the simple (co)roots 𝒲_i and ℳ_j is given by the intersection between 𝒲_i and 𝒲_j, explicitly: ⟨𝒲_i,ℳ_j ⟩ = (A_D_N )_i,j , where A_D_N is the Cartan matrix of the D_N algebra, corresponding to an 𝔰𝔬(2N) gauge theory. Indeed the intersection between F1 strings in the background of an O3^- reproduces the intersection of the roots of D_N. The Dirac pairing (<ref>) reproduces the Dirac pairing of 𝔰𝔬(2N) 𝒩=4 SYM. Similar constructions for O3^+, O3^-, and O3^+ lead to the B and C cases (while branes in absence of orientifold would give A). The corresponding gauge algebras are summarized in <Ref>. §.§ Lines in O3-planes Before moving to new results, we illustrate our method with well understood O3-planes. Specifically, we consider placing N=2 D3-branes in the background of an O3^+-plane. In this specific example, the F1-strings corresponding to elementary dynamical states in the quotient theory can be chosen to be |1,0 ⟩_1_2 1_1 and |1,0 ⟩_1_1 2_1. The first links the i=1 brane to its mirror (D3_1^ρ D3_1) and the second links the i=1 to the i=2 brane (D3_1 D3_2). A pictorial representation of this setup is shown in Figure <ref>. In the notation of the previous section, they lead to 𝒲_i-bosons in the gauge theory with the following charge basis: D3_1^ρ D3_1 : w_1 = (2,0;0,0) , D3_1 D3_2 : w_2 = (-1,0;1,0) . These generate the algebra 𝔲𝔰𝔭(4) of electric charges. The elementary magnetic monopoles ℳ_i come from the D1-strings |0,1 ⟩_O3^+ 1_1 and |0,1 ⟩_1_1 2_1, and provide the following charges: O3^+ D3_1 : m_1 = (0,1;0,0) , D3_1D3_2 : m_2 = (0,-1;0,1) . This generates the algebra 𝔰𝔬(5) of magnetic charges. Finally, the elementary (1,1)-strings leading to states in the quotient theory can be chosen to be |1,1 ⟩_1_2 1_1 and |1,1 ⟩_1_1 2_1, i.e. D3_1^ρ D3_1 and D3_1D3_2 respectively. They provide dyons 𝒟_i: D3_1^ρ D3_1 : d_1 = (2,2;0,0) , D3_1D3_2 : d_2 = (-1,-1;1,1) , which reproduces an 𝔲𝔰𝔭(4) algebra. We will limit ourselves to considering the 𝒲-bosons and magnetic monopoles ℳ. Indeed, they generate the full lattice of electromagnetic charges admissible in the orientifold theory. See that d_1 = w_1 + 2 m_1 d_2 = w_2 + m_2 . Clearly, all other allowed (p,q)-charges can be reconstructed in this way. The Dirac pairing between these elementary electromagnetic charges reads ⟨𝒲_1, 𝒲_2 ⟩ = ⟨ℳ_1, ℳ_2 ⟩ = 0 , ⟨ℳ_1, 𝒲_2 ⟩ = 1 , ⟨𝒲_1,ℳ_1 ⟩ =⟨ℳ_2,𝒲_1 ⟩ =⟨𝒲_2,ℳ_2 ⟩ = 2 . Now, introduce a line operator ℒ with charge vector ℓ. It is convenient to express it in the basis of dynamical charges: ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2 , where α_i and β_i to be determined. Screening with respect to 𝒲_1 and 𝒲_2 imposes α_1 ∼α_1 + 1 , α_2 ∼α_2 + 1 , respectively, while screening with respect to ℳ_1 and ℳ_2 imposes β_1 ∼β_1 + 1 , β_2 ∼β_2 + 1 . Mutual locality with respect to the dynamical charges requires the quantities [ ⟨ℒ, 𝒲_1 ⟩ = - 2β_1 +2 β_2 , ⟨ℒ, 𝒲_2 ⟩ = β_1 - 2β_2 ,; ⟨ℒ, ℳ_1 ⟩ = 2α_1 - α_2 , ⟨ℒ, ℳ_2 ⟩ = -2 α_1 + 2α_2 , ] to be integers. All these constraints set α_1 = e/2 α_2 = 0 , β_1 = 0 , β_2 = m/2 1 , with e,m=0,1. Linearity of the Dirac pairing then guarantees mutual locality with respect to the full dynamical spectrum. Thus, the charge of the most general line (modulo screening) must read: ℓ_e,m = 1/2(2e, -m ; 0 , m) . A choice of global structure consists in finding a set of mutually local lines. The mutual locality condition between two lines ℒ and ℒ' with charges ℓ_e,m and ℓ_e',m' is given by: ⟨ℒ , ℒ^'⟩ = 1/2 (-em^' + e^' m) ∈ℤ . Equivalently: em'-me' = 0 2 . We find three such sets, each composed of a single line with non-trivial charge: ℓ_1,0, ℓ_0,1, or ℓ_1,1. In agreement with <cit.>, we find that the line with charge ℓ_1,0 transforms as a vector of 𝔲𝔰𝔭(4) and the theory is USp(4). The line with charge ℓ_0,1 transforms as a spinor of 𝔰𝔬(5) and corresponds to the global structure (USp(4)/ℤ_2)_0. The line with charge ℓ_1,1 transforms both as a vector and a spinor, and the gauge group is (USp(4)/ℤ_2)_1. Motivated by the match between our results (obtained through the procedure described above) and the global structures of Lagrangian theories <cit.>, in the next sections we use our method to analyze the line spectra of S-fold theories. § LINES IN S-FOLDS WITH 𝒩=4 ENHANCEMENT We now derive the spectrum of mutually local lines for the gauge theories obtained with N=2 D3-branes in the background of an S_k,1 plane, in each case k=3, 4 and 6. More precisely, exploiting the strategy spelled out in Section <ref>, we first compute the electromagnetic charge lattice of local states generated by (p,q)-strings. From this we extract the possible spectra of lines and compare them with the ones obtained in an 𝒩=4 Lagrangian formalism <cit.>, since these theories have been claimed to enhance to 𝒩=4 SYM <cit.>. Matching the spectra provides an explicit dictionary between the various lattices and corroborates the validity of our procedure. In section <ref> we will then generalize the analysis to the pure 𝒩=3 S_k,ℓ projections for any rank, thus providing the full classification for the one-form symmetries in all such cases. §.§ Lines in 𝔰𝔲(3) from S_3,1 §.§.§ Dynamical states and their charges Two D3-branes probing the singular point of the S_3,1-fold are claimed to engineer 𝔰𝔲(3) 𝒩=4 SYM. The charges of states generated by (p,q)-strings stretching between D3_1 and D3_2 or its first copy (see Figure <ref>) are D3_1 D3_2 : (p,q;-p,-q) , D3_1 D3_2^ρ : (p,q;q, q-p) , D3_1 D3_2^ρ^2 : (p,q;p-q, p) . One may also consider copies of the strings listed in Equation <ref> such as: D3_1^ρ D3_2^ρ : (-q,p-q;q,q-p) , as well as the strings going from one D3-brane to its own copies, for instance[In the absence of discrete torsion, these states have not been considered previously in the literature <cit.>, and we do here for the sake of consistency with the analysis of section <ref>. Note however that since their charge (which is the only feature that matters in order to derive line spectra) can be expressed as linear combinations of the charges of more conventional states, our results are independent of whether we consider them or not. ] D3_1 D3_1^ρ : (2p-q,p+q;0,0) . The charges of a generic string D3_1 D3_2^ρ^2 in (<ref>) can be expressed in terms of D3_1 D3_2 and D3_1 D3_2^ρ charges: [ D3_1D3_2^ρ^2 : (p,q;p-q,p) = q (1,0;-1,0) + (q-p)(0,1;0,-1); +(p-q) (1,0;0,-1) +p (0,1;1,1) , ] where the first two vectors on the RHS come from D3_1 D3_2 with p=1, q=0 and p=0, q=1 respectively, and the last two come from D3_1 D3_2^ρ with p=1, q=0 and p=0, q=1 respectively. Acting with ρ_3, one can express all D3_1^ρ D3_2^ρ and D3_1^ρ^2 D3_2^ρ^2 charges in terms of D3_1 D3_2 charges. The charges D3_i D3_i^ρ can also be expressed as linear combinations of D3_1 D3_2^ρ and D3_2^ρ D3_1^ρ charges. All in all, we find that the charges of the strings D3_1 D3_2 and D3_1D3_2^ρ form a basis of the lattice of dynamical charges. The states corresponding to the 𝒲-bosons generate the 𝔰𝔲(3) algebra. One can take the strings D3_1D3_2 with p=1 and q=0 and D3_1D3_2^ρ with p=0 and q=1 as representing a choice of positive simple roots. Their electromagnetic charge w reads: w_1 = (1,0;-1,0) , w_2 =(0,1;1,1) . Furthermore, one can choose the strings D3_1D3_2 with p=0 and q=1 and D3_1D3_2^ρ with p=-1 and q=-1 as generating the charge lattice of magnetic monopoles ℳ of 𝒩=4 SYM with gauge algebra 𝔰𝔲(3): m_1 = (0,1;0,-1) , m_2 =(-1,-1;-1,0) . The qualification of electric charges 𝒲 and magnetic monopoles ℳ of the 𝒩=4 theory makes sense since the Dirac pairing reads: [ ⟨𝒲_1, 𝒲_2 ⟩ = ⟨ℳ_1, ℳ_2 ⟩ = 0 ,; ⟨𝒲_1, ℳ_1 ⟩ = ⟨𝒲_2, ℳ_2 ⟩ = 2 ,; ⟨𝒲_1,ℳ_2 ⟩ = ⟨𝒲_2,ℳ_1 ⟩ = -1 . ] In <cit.>, it has been shown that these states correspond indeed to BPS states, and this is a strong check of the claim of the supersymmetry enhancement in this case. §.§.§ Line lattices Having identified the electromagnetic lattice of charges of (p,q)-strings we can now construct the spectrum of line operators and the corresponding one-form symmetries. It is useful to consider the charge ℓ = (e_1, m_1 ; e_2, m_2) of a general line ℒ to be parameterized as follows: [ ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2; = (α_1-β_2, α_2 + β_1-β_2 ; -α_1 +α_2-β_2,α_2 -β_1 ) . ] Screening with respect to w_i and m_i translates as the identifications: α_i ∼α_i+1 , β_i ∼β_i+1 . The Dirac pairing between the generic line ℒ with charge ℓ given in (<ref>) and the states 𝒲 and ℳ must be an integer, i.e.: [ ⟨ℒ, 𝒲_1 ⟩ = 2β_1 - β_2 , ⟨ℒ, 𝒲_2 ⟩ = -β_1 + 2β_2 ,; ⟨ℒ, ℳ_1 ⟩ = -2α_1 + α_2 , ⟨ℒ, ℳ_2 ⟩ = α_1 - 2α_2 ] ∈ . Mutual locality with respect to the other states then follows by linearity as soon as (<ref>) holds. Combining (<ref>) and (<ref>) we have α_1 = - α_2 = e/3 , and β_1 = -β_2 = m/3 , for e,m=0, 1, 2. Then, the charge of the most general line compatible with the spectrum of local operators modulo screening reads ℓ_e,m =1/3 ( 2e-m, e+m; -e-m, e-2m ) . These charges form a finite 3 × 3 square lattice. The Dirac pairing between two lines ℒ and ℒ' with charges ℓ_e,m and ℓ_e',m' is ⟨ℒ , ℒ' ⟩ = 2/3 (e m' - e' m) . Two lines ℒ and ℒ' are mutually local if their Dirac pairing is properly quantized. In our conventions this corresponds to the requirement that ⟨ℒ , ℒ' ⟩ is an integer: e' m - e m' = 0 3 . The lattice of lines together with the mutual locality condition obtained in (<ref>) fully specifies the global structure of the S_3,1 SCFT of rank-2. Our result is equivalent to the one obtained in <cit.> from the Lagrangian description of 𝔰𝔲(3) 𝒩=4 SYM theory. Let us first write the charges in (<ref>) as: ℓ_e,m = e w_1- w_2/3 + m m_1 - m_2/3 . Note that (w_1-w_2)/3 (respectively, (m_1 - m_2)/3) is a weight of the electric (respectively, magnetic) algebra 𝔰𝔲(3) with charge 1 under the center ℤ_3 of the simply-connected group SU(3). Therefore, the line ℓ_e,m corresponds to a Wilson-'t Hooft line of charge (e,m) under ℤ_3×ℤ_3. As shown in <cit.>, there are four possible lattices of mutually local Wilson-'t Hooft lines specified by two integers i=0,1,2 and p=1,3. The corresponding gauge theories are denoted (SU(3)/ℤ_p)_i and relate to the line spectra we have obtained as follows: [ SU(3) ↔ {ℓ_0,0, ℓ_1,0, ℓ_2,0} ,; (SU(3)/ℤ_3)_0 ↔ {ℓ_0,0, ℓ_0,1, ℓ_0,2} ,; (SU(3)/ℤ_3)_1 ↔ {ℓ_0,0, ℓ_1,1, ℓ_2,2} ,; (SU(3)/ℤ_3)_2 ↔ {ℓ_0,0, ℓ_2,1, ℓ_1,2} . ] It follows from linearity and screening that each lattice in the S-fold picture is determined by a single non-trivial representative, that can itself be identified by two integers (e,m). For example, a possible choice is (e,m)= (1,0), (0,1), (1,1), (2,1) . §.§ Lines in 𝔰𝔬(5) from S_4,1 §.§.§ Dynamical states and their charges Two D3-branes probing the singular point of the S_4,1-fold are claimed to engineer 𝔰𝔬(5) 𝒩=4 SYM. Following a reasoning similar to one of the S_3,1-fold case, we can write all string charges as linear combinations of two kinds of strings, say D3_1 D3_2 : (p,q;-p,-q) , D3_1 D3_2^ρ : (p,q;-q,p) . States corresponding to the 𝒲-bosons of 𝒩=4 SYM are generated by D3_1D3_2 with p=1 and q=0, and D3_1D3_2^ρ with p=-1 and q=-1. Their charges are w_1 = (1,0;-1,0) , w_2 = (-1,-1;1,-1) . These states generate the algebra 𝔰𝔬(5) with short and long positive simple roots w_1 and w_2, respectively. A possible choice of states corresponding to elementary magnetic monopoles ℳ is D3_1D3_2 with p=-1 and q=1, and D3_1D3_2^ρ with p=1 and q=0. The charges of these strings are: m_1= (-1,1;1,-1) , m_2 = (1,0;0,1) , with m_1 the long and m_2 the short positive simple roots of the Langland dual algebra 𝔲𝔰𝔭(4). The Dirac pairings between 𝒲 and ℳ are as expected: [ ⟨𝒲_1, 𝒲_2 ⟩ = ⟨ℳ_1, ℳ_2 ⟩ = 0 ,; ⟨𝒲_1, ℳ_1 ⟩ = ⟨𝒲_2, ℳ_2 ⟩ = ⟨ℳ_1, 𝒲_2 ⟩ = 2 ,; ⟨ℳ_2 , 𝒲_1 ⟩ = 1 . ] §.§.§ Line lattices We begin by parametrizing the charge ℓ of a general line ℒ as: ℓ = α_1 w_1 + α_2 w_2+β_1 m_1 + β_2 m_2 = (α_1-α_2-β_1+β_2, β_1-α_2 ; -α_1+α_2+β_1, -α_2-β_1+β_2) . Screening with respect to the local states 𝒲 and ℳ translates as: α_i ∼α_i + 1 , β_i ∼β_i +1 . Mutual locality with respect to the dynamical states generated by (p,q)-strings reads: [ ⟨ℒ,𝒲_1 ⟩ = 2β_1-β_2; ⟨ℒ,𝒲_2 ⟩ = -2 β_1 +2β_2; ⟨ℒ,ℳ_1⟩ = -2α_1 +2α_2; ⟨ℒ,ℳ_2 ⟩ = α_1 - 2α_2; ] ∈ . This imposes α_1 = β_2 = 0 and α_2, β_1 ∈1/2, and therefore the charge of the most general line compatible with the spectrum of local states can be written as: ℓ_e,m = e/2 w_2 + m/2 m_1 = 1/2 (-e-m ,-e+m; e+m, -e-m) . The Dirac pairing between two lines ℒ and ℒ' with charges ℓ_e,m and ℓ_e',m' is: ⟨ℒ, ℒ' ⟩ = 1/2 (e'm - em') . Two such lines are mutually local if their Dirac pairing if ⟨ℒ, ℒ' ⟩ is an integer, i.e.: (e'm - em') = 0 2 . Therefore, the allowed lines form a finite 2× 2 square lattice parametrized by e,m=0,1, where the mutual locality condition is given by (<ref>). This reproduces the expected global structures of 𝒩=4 𝔰𝔬(5) SYM. There are three possible choices of maximal lattices of mutually local lines which correspond to the three possible global structures of 𝔰𝔬(5). The explicit mapping can be obtained by comparing the electromagnetic charges of the lines with the charges of the 𝒲 bosons and monopoles ℳ, along the lines of the analysis of above in the 𝔰𝔲(3) case. We obtain the following global structures: [ Spin(5) ↔ {ℓ_0,0, ℓ_1,0} ,; SO(5)_0 ↔ {ℓ_0,0, ℓ_0,1} ,; SO(5)_1 ↔ {ℓ_0,0, ℓ_1,1} . ] §.§ Trivial line in 𝔤_2 from S_6,1 §.§.§ Dynamical states and their charges Two D3-branes probing the singular point of the S_6,1-fold are claimed to engineer 𝔤_2 𝒩=4 SYM. The charges of states generated by (p,q)-strings are: [ D3_1 D3_2 : (p,q;-p,-q) , D3_1 D3_2^ρ : (p,q;-q,p-q) ,; D3_1 D3_2^ρ^2 : (p,q;p-q,p) , D3_1 D3_2^ρ^3 : (p,q;p,q) ,; D3_1 D3_2^ρ^4 : (p,q;q,-p+q) , D3_1 D3_2^ρ^5 : (p,q;-p+q,-p) ,; etc. ] As shown in <cit.> and as before, one can choose a set of strings representing dynamical particles and generating the algebra 𝔤_2. §.§.§ Line lattice The analysis of the charge spectrum in the case of the S_6,1-fold can be carried out along the lines of the previous sections. One can show that the only line that is mutually local with respect to the local states generated by (p,q)-strings modulo screening is the trivial line with charges ℓ = (0,0;0,0). This is consistent with the enhancement to 𝒩=4 with gauge algebra 𝔤_2 because the center of the simply-connected G_2 is trivial, which implies the absence of non-trivial lines <cit.>. There is only one possible global structure, and the one-form symmetry is trivial. § LINES IN 𝒩=3 S-FOLDS In this section, we generalize the procedure spelled out in the previous sections to S-folds theories of arbitrary rank, and later to the cases with non-trivial discrete torsion for the B_2 and C_2 fields. This allows us to classify the line spectrum for every 𝒩=3 S-fold theory, and identify the one-form symmetry group as well as the allowed global structures for a given theory. The basic ingredients needed in the analysis are the lattice of electromagnetic charges of local states and the Dirac pairing, both of which can be inferred from the Type IIB setup along the lines of the rank-2 cases studied in Section <ref>. As already emphasized, we work under the assumption that the states generated by (p,q)-string form a good set of representatives of the electromagnetic charge lattice of the full spectrum. Note that it does not strictly make sense to talk about (p,q)-strings on the ℝ^4×ℝ^6/ℤ_k S-fold background because the S-fold projection involves an SL(2,ℤ) action which mixes F1 and D1 strings. This is analogous to the fact that in the orientifold cases it only makes sense to consider unoriented strings, since the orientifold action reverses the worldsheet parity (equivalently, it involves the element -𝕀_2 ∈ SL(2,)). Nevertheless it makes sense to consider oriented strings (together with their images) on the double cover of the spacetime; this allows the computation of the electromagnetic charge lattice of local states and the Dirac pairing, as reviewed in Section <ref>. Similarly when dealing with S_k-folds we consider (p,q)-strings on the k-cover of the spacetime, and extract from this the charges of local states and the Dirac pairing. The spectrum of lines can then be obtained using the procedure of <cit.> reviewed in Section <ref>. §.§ Lines in S_3,1-fold Let us first determine the lattice of electromagnetic charges of dynamical states. The charges generated by (p,q)-strings on the background of an S_3,1 fold are given by D3_iD3_j^ρ^l : (0,0; … ; p,q^i-th; …; - (p q)·ρ^l_3^j-th; … ; 0,0) . This expression is obtained from a (p,q)-string stretched between the i-th D3-brane and the l-th image of the j-th D3-brane. Recall that ρ_3 generates a _3 subgroup of SL(2,). A possible basis for the lattice of charges generated by (p,q)-strings is given by: [ w_1 = (1,0;-1,0;…) ,; w_2 = (0,1;1,1;…) ,; m_1 = (0,1;0,-1;…) ,; m_2 = (-1,-1;-1,0;…) ,; P_i = (1,0;0,0; … ;-1,0^i-th;0,0;…) ,; Q_i = (0,1;0,0; … ;0,-1^i-th;0,0;…) , ] where w_i and m_i are the charges of the corresponding states in the rank-2 case, with all other entries set to 0. Let 𝒫_i and 𝒬_i be the states with charges P_i and Q_i respectively, for i=3,…,N. Note that when the rank is N>2, it does not make sense to talk about 𝒲-bosons and magnetic monopoles ℳ since the pure 𝒩=3 theories are inherently strongly coupled and do not admit a Lagrangian description. Nevertheless, we will denote 𝒲_i and ℳ_i the states with charges w_i and m_i respectively, by analogy with the above. The charge ℓ of a general line ℒ can be written as the linear combination: ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2 + ∑_i=3^N ( δ_i P_i + γ_i Q_i) . Besides, screening translates into the identifications: α_i ∼α_i + 1 , β_i ∼β_i +1 , δ_i ∼δ_i+1 , γ_i ∼γ_i+1 . Let us now analyze the constraints imposed on this line given by mutual locality with respect to the dynamical states generated by (p,q)-strings. Our results are summarized in Table <ref>. Consider the mutual locality conditions: ⟨ℒ, 𝒫_i - 𝒫_j ⟩ = δ_i - δ_j ∈ ⇒ δ_i = δ_j = δ i,j=3,…,N , and ⟨ℒ, 𝒬_i - 𝒬_j ⟩ =γ_j - γ_i ∈ ⇒ γ_j = γ_i = γ i,j=3,…,N . Furthermore, there are dynamical states with charges: (0,0; … ;1,-1^i-th; …)= (p,q;…; -p,-q^i-th; …)_|[ p=0; q=1 ] + (p,q; …; p-q,p^i-th;…)_|[ p=0; q=-1 ] , (0,0 ; …;2,1^i-th; …) = (p,q; …; -p,-q^i-th; …)_|[ p=-1; q=0 ] + (p,q; …; p-q,p^i-th;…)_|[ p=1; q=0 ] . Mutual locality with respect to these implies: γ = -δ , δ∈1/3 . Therefore, the charge of a general line can be rewritten as: ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2 + δ (P - Q) , where P = ∑_i=3^N p_i = (N-2, 0; 0,0;-1,0;-1,0;…;-1,0) , Q = ∑_i=3^N q_i = (0,N-2;0,0;0,-1;0,-1;…;0,-1) . In (<ref>), we have modified our notation slightly since the dots … now represent a sequence of pairs (-1,0) and (0,-1) for P and Q respectively. Mutual locality between the line ℒ and the generators of the charge lattice of dynamical states imposes the following constraints: [ ⟨ℒ, 𝒫_i ⟩ = (N-1)δ -α_2 - β_1 + β_2 ,; ⟨ℒ, 𝒬_i ⟩ = (N-1)δ + α_1 - β_2 ,; ⟨ℒ, 𝒲_1 ⟩ = (N-2)δ -2β_1 + β_2 ,; ⟨ℒ, 𝒲_2 ⟩ = (N-2)δ - 2β_2 + β_1 ,; ⟨ℒ, ℳ_1 ⟩ = (N-2)δ +2α_1-α_2 ,; ⟨ℒ, ℳ_2 ⟩ = -2(N-2)δ-α_1 + 2α_2 ] ∈ . One can compute the following: [ ⟨ℒ, 𝒲_1 + 2 𝒲_2 ⟩ = 3(N-2) δ -3β_2 ∈ ⇒ β_2 ∈1/3 ,; ⟨ℒ, ℳ_1 + 2 ℳ_2 ⟩ = -3α_1 ∈ ⇒ α_1 ∈1/3 ,; ⟨ℒ, 𝒲_1 - 𝒲_2 ⟩ = 3(β_2 - β_1) ∈ ⇒ β_1 ∈1/3 ,; ⟨ℒ, ℳ_1 - ℳ_2 ⟩ = 3(N-2) δ + 3(α_1 - α_2) ∈ ⇒ α_2 ∈1/3 . ] In brief, we have found that α_i, β_i, δ∈1/3. It is now useful to treat separately three cases, depending on the value of N mod 3. In all these cases we find that the lines modulo screening can be arranged in a finite 3×3 lattice, the one-form symmetry group is _3 and there are four choices of global structure. §.§.§ Case N = 3n The mutual locality conditions in (<ref>) can be written as: [ ⟨ℒ, 𝒫_i ⟩ = -δ -α_2 - β_1 + β_2 ,; ⟨ℒ, 𝒬_i ⟩ = -δ + α_1 - β_2 ,; ⟨ℒ, 𝒲_1 ⟩ = δ -2β_1 + β_2 ,; ⟨ℒ, 𝒲_2 ⟩ = δ - 2β_2 + β_1 ,; ⟨ℒ, ℳ_1 ⟩ = δ +2α_1-α_2 ,; ⟨ℒ, ℳ_2 ⟩ = δ-α_1 + 2α_2 ] ∈ . One computes that: [ ⟨ℒ, 𝒬_i + 𝒲_1 ⟩ = α_1 + β_1 ⇒ β_1 = - α_1 ,; ⟨ℒ, 𝒫_i + 𝒲_2 ⟩ = -α_2 -β_2 ⇒ β_2 = - α_2 ,; ⟨ℒ, 𝒬_i ⟩ = -δ + α_1+α_2 ⇒ δ = α_1 + α_2 , ] and this implies: α_1 = -β_1 = r/3 , α_2 = -β_2 = s/3 , δ = r+s/3 , r,s = 0,1,2 . Therefore the lines form a finite 3×3 lattice parametrized by r and s. Mutual locality between two general lines ℒ and ℒ' with charges ℓ_r,s and ℓ_r',s' reads: ⟨ℒ, ℒ' ⟩ = 2/3 (s r' - r s') ∈ , or equivalently: s r' - r s' = 0 mod 3 . There are four possible choices of maximal lattices of mutually local lines. As in the rank-2 case discussed in section <ref>, each lattice is uniquely identified by one of its element, or equivalently by the pair (r,s) of one of its non-trivial elements: (r,s) = { (1,0) ↔{ℓ_0,0, ℓ_1,0,ℓ_2,0} (0,1) ↔{ℓ_0,0, ℓ_0,1,ℓ_0,2} (1,1) ↔{ℓ_0,0, ℓ_1,1,ℓ_2,2} (1,2) ↔{ℓ_0,0, ℓ_1,2,ℓ_2,1}. . §.§.§ Case N = 3n+1 In this case the mutual locality constraints (<ref>) are: [ ⟨ℒ, 𝒫_i ⟩ = -α_2 - β_1 + β_2; ⟨ℒ, 𝒬_i ⟩ = α_1 - β_2; ⟨ℒ, 𝒲_1 ⟩ = -δ -2β_1 + β_2; ⟨ℒ, 𝒲_2 ⟩ = -δ - 2β_2 + β_1; ⟨ℒ, ℳ_1 ⟩ = -δ +2α_1-α_2; ⟨ℒ, ℳ_2 ⟩ = 2 δ-α_1 + 2α_2 ] ∈ . One computes that: [ α_2 = α_1 -β_1 ,; δ = α_1 + β_1 ,; α_1 = β_2 . ] Therefore the most general α_i,β_i and δ satisfy: α_1 = β_2 = r/3 , β_1 = s/3 , α_2 = r-s/3 , δ = r+s/3 , r,s=0,1,2 . The lines again form a finite 3×3 lattice parametrized by r and s. Mutual locality between two general lines ℒ and ℒ' with charges ℓ_r,s and ℓ_r',s' reads: ⟨ℒ, ℒ' ⟩ = 1/3 (s r' - r s') ∈ , or equivalently: s r' - r s' = 0 3 . Similarly to the case N=3n there are four possible choices of maximal lattices of mutually local lines that can be indexed by one of their element, or equivalently by (r,s)=(1,0), (0,1), (1,1), (1,2). §.§.§ Case N = 3n+2 In this case, the mutual locality constraints (<ref>) are [ ⟨ℒ, 𝒫_i ⟩ = δ -α_2 - β_1 + β_2; ⟨ℒ, 𝒬_i ⟩ = δ + α_1 - β_2; ⟨ℒ, 𝒲_1 ⟩ = -2β_1 + β_2 = β_1 + β_2; ⟨ℒ, 𝒲_2 ⟩ = - 2β_2 + β_1; ⟨ℒ, ℳ_1 ⟩ = 2α_1-α_2 = -α_1 -α_2; ⟨ℒ, ℳ_2 ⟩ = -α_1 + 2α_2 ] ∈ . One can compute that the solution is given by [ β_2 = -β_1 ,; α_2 = -α_1 ,; δ = -α_1 - β_1 . ] Therefore the most general α_i,β_i and δ satisfy: α_1 = -α_2 = r/3 , β_1 = -β_2 = s/3 , δ = - r+s/3 , r,s=0,1,2 . Dirac pairing between two general lines ℒ and ℒ' with charges ℓ_r,s and ℓ_r',s' reads: ⟨ℒ, ℒ' ⟩ = 2/3 (s r' - r s') ∈ . Two such lines are mutually local if they satisfy the constraint: s r' - r s' = 0 mod 3 . As before, there are four possible choices of maximal lattices of mutually local lines that can be indexed by one of their element, or equivalently by (r,s)=(1,0),  (0,1), (1,1), (1,2) . §.§ Lines in S_4,1-fold We now study the spectrum of lines in theories engineered by a stack of D3-branes probing the S_4,1-fold. The charges of states generated by a (p,q)-string on the background of an S_4,1-fold read D3_iD3_j^ρ^l : (0,0; … ; p,q^i-th; …; -(p q)·ρ_4^l^j-th; …;0,0) for a (p,q)-strings stretched between the i-th D3-brane and the l-th image of the j-th D3-brane. One possible basis for the lattice of charges generated by (p,q)-strings is: [ w_1 = (1,0;-1,0;0,0;…) ,; w_2 = (-1,-1;1,-1;0,0;…) ,; m_1 = (-1,1;1,-1;0,0;…) ,; m_2 = (1,0;0,1;0,0;…) ,; P_i = (1,0;0,0; … ;-1,0^i-th;0,0;…) ,; Q_i = (0,1;0,0; … ;0,-1^i-th;0,0;…) , ] where w_i and m_i are the charges of the corresponding states in the rank-2 case, with all other entries set to 0. We denote 𝒲_i, ℳ_i, 𝒫_i and 𝒬_i the states with charges w_i, m_i, P_i and Q_i, respectively. The charge ℓ of a general line ℒ can be written as the linear combination: ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2 + ∑_i=3^N ( δ_i P_i + γ_i Q_i) . Screening translates into the identifications: α_i ∼α_i + 1, β_i ∼β_i +1, δ_i ∼δ_i+1, γ_i ∼γ_i+1 . In the remainder of this section we compute the constraints imposed by mutual locality between the general line ℒ and dynamical states. Our results are summarized in Table <ref>. Consider first the mutual locality conditions: ⟨ℒ, 𝒫_i - 𝒫_j ⟩ = δ_i - δ_j ∈ ⇒ δ_i = δ_j = δ , ⟨ℒ, 𝒬_i - 𝒬_j ⟩ =γ_j - γ_i ∈ ⇒ γ_j = γ_i = γ . Furthermore, there are dynamical states with charges: [ (0,0; … ;1,-1^i-th; …) = (p,q;…; -p,-q^i-th; …)_|[ p=0; q=1 ] + (p,q; …; -q,p^i-th;…)_|[ p=0; q=-1 ] ,; (0 ,0;…; 1,1^i-th; …) = (p,q; …; -p,-q^i-th; …)_|[ p=-1; q=0 ] + (p,q; …; -q,p^i-th;…)_|[ p=1; q=0 ] . ] and mutual locality with respect to these states implies: γ = -δ, δ∈1/2 . Therefore, the charge of a general line can be rewritten as: ℓ = α_1 w_1 + α_2 w_2 + β_1 m_1 + β_2 m_2 + δ (P - Q) , where P and Q are defined in (<ref>). Mutual locality between the line ℒ and the generators of the charge lattice of dynamical states implies: [ ⟨ℒ, 𝒫_i ⟩ = (N-1)δ +α_2 - β_1 ,; ⟨ℒ, 𝒬_i ⟩ = (N-1)δ + α_1 - α_2-β_1+β_2 ,; ⟨ℒ, 𝒲_1 ⟩ = (N-2)δ -2β_1+β_2 ,; ⟨ℒ, 𝒲_2 ⟩ = 2(N-2)δ - 2β_2 + 2β_1 ,; ⟨ℒ, ℳ_1 ⟩ = 2α_1 - 2α_2; ⟨ℒ, ℳ_2 ⟩ = (N-2)δ-α_1 + 2α_2 ] ∈ . One computes the following: [ ⟨ℒ, 𝒲_1+ 𝒲_2- ℳ_1- ℳ_2 ⟩ = -β_2 - α_1∈ ⇒β_2 = -α_1 ,; ⟨ℒ, 𝒬_i + 𝒫_i ⟩ = -2β_1 ∈ ⇒β_1 ∈1/2 ,; ⟨ℒ, 𝒬_i - 𝒫_i ⟩ = -2α_2 ∈ ⇒α_2 ∈1/2 ,; ⟨ℒ, ℳ_1 ⟩ = 2α_1 ∈ ⇒α_1,β_2 ∈1/2 . ] We have thus shown that α_i, β_i, δ∈1/2 and α_1 = -β_2. It is now useful to treat separately the cases of odd and even N. In both cases we find that the lines form a 2×2 lattice, the one-form symmetry is _2 and there are three choices of global structure. §.§.§ Case N = 2n Mutual locality conditions (<ref>) read: [ ⟨ℒ, 𝒫_i ⟩ = -δ - β_1 +α_2; ⟨ℒ, 𝒬_i ⟩ = -δ - α_2-β_1; ⟨ℒ, 𝒲_1 ⟩ = β_2; ⟨ℒ, 𝒲_2 ⟩ = 0; ⟨ℒ, ℳ_1 ⟩ = 0; ⟨ℒ, ℳ_2 ⟩ = -α_1 ] ∈ , and each solution can be written as: α_2 = r/2 , β_1 = s/2 , α_1 = β_2 = 0 , δ = r+s/2 , r,s = 0,1 . Therefore the lines form a 2×2 lattice parametrized by r,s. Mutual locality between two lines ℒ and ℒ' with charges ℓ_r,s and ℓ_r',s' respectively translates into: ⟨ℒ,ℒ'⟩ = 1/2 (r's - r s') ∈ , or equivalently: r's - r s' = 0 2 . The one-form symmetry group is thus _2 and there are three different choices of maximal lattices of mutually local lines parametrized by (r,s) = (1,0), (0,1), (1,1). §.§.§ Case N = 2n+1 The Dirac pairings (<ref>) read: [ ⟨ℒ, 𝒫_i ⟩ = α_2 - β_1 ,; ⟨ℒ, 𝒬_i ⟩ = - α_2-β_1 ,; ⟨ℒ, 𝒲_1 ⟩ = δ +β_2 ,; ⟨ℒ, 𝒲_2 ⟩ = 0 ,; ⟨ℒ, ℳ_1 ⟩ = 0 ,; ⟨ℒ, ℳ_2 ⟩ = δ-α_1 ] ∈ , and the general solution can be written as: α_1 = β_2 = δ = r/2 , α_2 = β_1 = s/2 , r,s = 0,1 . Mutual locality between two lines ℒ and ℒ' with charges ℓ_r,s and ℓ_r',s' respectively translates into: ⟨ℒ,ℒ'⟩ = 1/2 (r's - r s') ∈ , or equivalently: r's - r s' = 0 2 . As in the previous case, the one-form symmetry group is therefore _2 and there are three different choices of maximal lattices of mutually local lines that can be parametrized by: (r,s) = (1,0), (0,1), (1,1) . §.§ Trivial line in S_6,1-fold The analysis of the spectrum of lines in the case of the S_6,1-fold can be carried out along the lines of the previous subsections. One finds that the integer lattice of charges associated to (p,q)-strings is fully occupied. To see this notice that there are two states with the following charges: [ (1,0;0,0;0,0;…) = (p,q;p-q,p;0,0;…)_|[ p=0; q=-1 ] - (p,q;-q,p,0,0;…)_|[ p=1; q=0 ] ,; (0,1;0,0;0,0;…) = (1,0;0,0;0,0;…) - (p,q;-p-q;0,0;…)_|[ p=0; q=1 ]; - (p,q;-q,p;0,0;…)_|[ p=0; q=-1 ] . ] By combining these states with 𝒫_i and 𝒬_i we can obtain states with electric or magnetic charge 1 with respect to the i-th brane, and all other charges set to zero. Let us now consider a general line ℒ with charge ℓ = (e_1,m_1; e_2,m_2,…). Mutual locality with respect to the local states we have just discussed implies: e_i, m_i ∈ ∀ i , and the insertion of the same local states along the lines translates to the identification: e_i ∼ e_i +1, m_i ∼ m_i+1 . Therefore, the only allowed line modulo screening is the trivial line, with charge ℓ = (0,0;0,0;…). This implies that the one form symmetry group is trivial, and accordingly there is only one possible choice of global form. §.§ Trivial line in the discrete torsion cases We generalize the analysis discussed in the previous sections to the cases with non-trivial discrete torsion in the S_3,3-fold and S_4,4-fold. As we argued in Section <ref> all the strings states that are present when the discrete torsion is trivial are also allowed when the discrete torsion is non-zero. Furthermore, there are strings ending on the S-fold itself, as discussed in Section <ref>. Thus, the lattice of charges of local states in the case of the S_3,3-fold and S_4,4-fold are generated by strings stretched between (images of) D3-branes – as in the cases with trivial discrete torsion – together with those additional strings. One can show that the integer lattice of electromagnetic charges of dynamical states is then fully occupied. Therefore, by a similar argument to the one used in the case of the S_6,1-fold in Section <ref>, the only line that is allowed is the trivial one, and the one-form symmetry group is 1 for the S_3,3-fold and S_4,4-fold with non-zero discrete torsion. § NON-INVERTIBLE SYMMETRIES We now discuss the possible presence of non-invertible symmetries in S-fold theories. In the case of 𝒩=4 theories, the presence of S-duality orbits can imply the existence of non-invertible duality defects which are built by combining the action of some element of SL(2,ℤ) and the gauging of a discrete one-form symmetry <cit.>. Similar structures can be inferred from the S-fold construction. Consider moving one of the D3-brane along the non-contractible one-cycle of S^5/ℤ_k until it reaches its original position. The brane configurations before and after this are identical, and therefore the S-fold theories are invariant under this action. Going around the non-contractible one-cycle of S^5/ℤ_k in the case an S_k,l-fold involves an SL(2,)-transformation on the electric and magnetic charges e_i, m_i associated to the D3-brane that has been moved. Let Σ_k^i denote the process of moving the i-th D3-brane along the non-contractible cycle of an S_k,l-fold. The action of Σ_k^i on the charges is: Σ_k^i: ( [ e_j; m_j ])→{ ρ_k ·( [ e_j; m_j ]) j=i ( [ e_j; m_j ]) j≠ i  . . The charge lattice of dynamical states is invariant under Σ_k^i, while the set of line lattices can be shuffled. Consider for example the S_3,1-case with rank N=2. One can compute explicitly the following orbits: (1,0) [<->]r (0,1) [<->]r (1,1) (1,2) [loop above,looseness=8]  , where the pairs (e,m) parametrize the maximal sub-lattice of mutually local lines as discussed in section (<ref>). Two line lattices connected by an arrow in (<ref>) are mapped to each other under proper combinations of Σ_3^i. This theory enhances to 𝔰𝔲(3) 𝒩=4 SYM. Using the mapping (<ref>) between the line lattices parametrized by (e,m) and the global structures of 𝔰𝔲(3), the formula (<ref>) reproduces the 𝒩=4 orbits under the element ST∈ SL(2,ℤ). As shown in the literature <cit.>, this transformation can be combined with a proper gauging of the one-form symmetry to construct the non-invertible self-duality defects of 𝔰𝔲(3) at τ=e^2π i/3. Therefore in our notation we expect the existence of non-invertible symmetries involving Σ_k^i for the lattices labeled by (e,m)=(1,0),(0,1),(1,1), and none in the (e,m)=(1,2) case. Similarly, one can consider the orbits in the case of S_4,1 with N=2, where the SCFT enhances to 𝔰𝔬(5) 𝒩=4 SYM. By using the transformations Σ_4^i as above we find the following orbits (0,1) ⟷ (1,0) (1,1) [loop above,looseness=8]  , where the pairs (e,m) parametrize the maximal sub-lattices of mutually local lines as discussed in section (<ref>). These reproduce the 𝒩=4 orbits under the element S∈ SL(2,). Again this transformation can be combined with a proper gauging of the one-form symmetry to construct the non-invertible self-duality defects of 𝔰𝔬(5) at τ=i. Motivated by this match, one can expect that in the case of general rank, non-invertible symmetries will be present when multiple choices of maximal sub-lattices of mutually local lines are related by the transformations Σ_k^i, as above. The orbits are: S_3,1: (1,0)⟷(0,1)⟷ (1,1) (1,2) [loop right,looseness=5]  , S_4,1: (0,1) ⟷ (1,0) (1,1) [loop right,looseness=5] N=0 2 (1,0)⟷ (1,1) (0,1) [loop right,looseness=5] N=1 2  , where the pairs (r,s) parametrize the maximal sub-lattices of mutually local lines as in section <ref>. In the S_6,1, S_3,3 and S_4,4-cases, there is only one possible global structure that is mapped to itself by the Σ_k^i transformations. By analogy with the cases where there is 𝒩=4 enhancement, we expect the existence of non-invertible symmetries when the transformations Σ_k^i map different line lattices, built by combining this Σ_k^i-action with a suitable gauging of the one-form symmetry. § CONCLUSIONS In this paper, we have exploited the recipe of <cit.> for arranging the charge lattice of genuine lines modulo screening by dynamical particles. We have adapted such strategy, originally designed for BPS quivers, to the case of (p,q)-strings, in order to access to the electromagnetic charges of non-Lagrangian 𝒩=3 S-fold SCFTs. This procedure has allowed us to provide a full classification of the one-form symmetries of every S-fold SCFT. We singled out two cases with a non-trivial one-form symmetry, corresponding to the ℤ_3 and the ℤ_4 S-folds in absence of discrete torsion, denoted here as S_3,1 and S_4,1 respectively. Our results are consistent with the supersymmetry enhancement that takes place when two D3-branes are considered. Lastly, we discuss the possibility of non-invertible duality defects, by recovering the expected results for the cases with supersymmetry enhancement and proposing a generalization at any rank. We left many open questions that deserve further investigations. It would for example be interesting to study in more details the projection of the states generated by the (p,q)-configurations in an S-fold background. In the present article, the only relevant information was the electromagnetic charges carried by these states, but a deeper analysis of the dynamics of these S-fold theories requires more work. This would in turn improve our understanding of their mass spectrum. For instance, a comparison of the BPS spectrum could be made exploiting the Lagrangian descriptions of <cit.>. This could also help finding the origin of the mapping between the multiple lattices found in the S_3,1 and S_4,1-cases. Further investigations in this direction would deepen our geometric understanding of the non-invertible symmetries expected in this class of theories, along the lines of the brane analysis of <cit.>. It would also be of interest to generalize the analysis to other 𝒩=3 SCFTs that are not constructed from S-fold projections, such as the exceptional 𝒩 = 3 theories <cit.>. These theories can be obtained from M-theory backgrounds and one may study the charge lattice with probe M2-branes. One could therefore apply an analysis similar to the one spelled in <cit.>. Regarding the S-fold constructions, the cases of S-folds with 𝒩=2 supersymmetry <cit.> also deserve further investigations (see <cit.> for similar analysis in class S theories). In the absence of BPS quivers, one needs to adapt the UV analysis of <cit.>. In general, one would like to find a stringy description that avoids wall crossing and allows reading the charge lattices and the one-form symmetries for such theories. § ACKNOWLEDGEMENTS We are grateful to Iñaki García Etxebarria for valuable insights on the manuscript, and to Shani Meynet and Robert Moscrop for useful discussions. The work of A.A., D.M., A.P. and S.R. has been supported in part by the Italian Ministero dell’Istruzione, Università e Ricerca (MIUR), in part by Istituto Nazionale di Fisica Nucleare (INFN) through the “Gauge Theories, Strings, Supergravity” (GSS) research project and in part by MIUR-PRIN contract 2017CC72MK-003. V.T. acknowledges funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster). JHEP

http://arxiv.org/abs/2303.07065v1

20230313123959

MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID

cs.CV

MSINet: Twins Contrastive Search of Multi-Scale Interaction for Object ReID Jianyang Gu^1 Kai Wang^2 Hao Luo^3 Chen Chen^4 Wei Jiang^1* Yuqiang Fang^5 Shanghang Zhang^6 Yang You^2 Jian Zhao^7Co-corresponding authors ^1Zhejiang University ^2National University of Singapore ^3Alibaba Group ^4OPPO Research Institute ^5Space Engineering University ^6Peking University ^7Institute of North Electronic Equipment {gu_jianyang, jiangwei_zju}@zju.edu.cn zhaojian90@u.nus.edu March 30, 2023 ============================================================================================================================================================================================================================================================================================================================================================================================================================= Neural Architecture Search (NAS) has been increasingly appealing to the society of object Re-Identification (ReID), for that task-specific architectures significantly improve the retrieval performance. Previous works explore new optimizing targets and search spaces for NAS ReID, yet they neglect the difference of training schemes between image classification and ReID. In this work, we propose a novel Twins Contrastive Mechanism (TCM) to provide more appropriate supervision for ReID architecture search. TCM reduces the category overlaps between the training and validation data, and assists NAS in simulating real-world ReID training schemes. We then design a Multi-Scale Interaction (MSI) search space to search for rational interaction operations between multi-scale features. In addition, we introduce a Spatial Alignment Module (SAM) to further enhance the attention consistency confronted with images from different sources. Under the proposed NAS scheme, a specific architecture is automatically searched, named as MSINet. Extensive experiments demonstrate that our method surpasses state-of-the-art ReID methods on both in-domain and cross-domain scenarios. Source code available in https://github.com/vimar-gu/MSINethttps://github.com/vimar-gu/MSINet. § INTRODUCTION Object re-identification (Re-ID) aims at retrieving specific object instances across different views <cit.>, which attracts much attention in computer vision community due to its wide-range applications. Previous works have achieved great progresses on both supervised <cit.> and unsupervised ReID tasks <cit.>, most of which adopts backbone models originally designed for general image classification tasks <cit.>. Recent literature <cit.> has shown that applying different architectures on ReID leads to large performance variations. Some works employ Neural Architecture Search (NAS) for ReID <cit.>. The proposed optimizing targets and search spaces stably improve the model performance, yet the main search scheme still follows traditional NAS methods designed for general classification tasks <cit.>. As an open-set task, ReID contains different categories in the training and validation sets <cit.>, while the two sets share exactly the same categories in standard classification tasks <cit.>, which is also followed by traditional NAS methods. The incompatibility between search schemes and real-world training schemes makes the searched architecture sub-optimal for ReID. Moreover, ReID is required to distinguish more subtle distinctions among fine-grained instances compared with image-level classification <cit.>. Some previous works <cit.> have manifested that local perspectives and multi-scale features are discriminative for ReID. However, current utilizations of these features are mostly empirically designed, which can be more flexible according to the characteristics of different network layers. In this work, we propose a novel NAS scheme aiming at addressing the aforementioned challenges. In order to simulate the real-world ReID training schemes, a Twins Contrastive Mechanism (TCM) is proposed to unbind the categories of the training and validation sets. An adjustable overlap ratio of categories builds up the compatibility between NAS and ReID, which provides more appropriate supervision for ReID architecture search. Moreover, to search for more rational utilizations of multi-scale features, we design a Multi-Scale Interaction (MSI) search space. The MSI space focuses on interaction operations between multi-scale features along the shallow and deep layers of the network, which guides the features to promote each other. Additionally, to further improve the generalization capability, we propose a Spatial Alignment Module (SAM) to enhance the attention consistency of the model confronted with images from different sources. With the above NAS scheme, we obtain a light-weight yet effective model architecture, denoted as Multi-Scale Interaction Net (MSINet). We visualize the example activation maps of our proposed MSINet and ResNet50 <cit.> trained on VeRi-776 <cit.> in Fig. <ref>. Compared to ResNet50, MSINet focuses on more unique distinctions with specific semantic information to recognize instances. Besides, MSINet largely increases the distance margin between query image and corresponding negative samples, reflecting extraordinary discriminative capability. Extensive experiments demonstrate that MSINet surpasses state-of-the-art (SOTA) ReID methods on both in-domain and cross-domain scenarios. Our source codes are available in the supplementary material. Our contributions are summarized as follows: * To the best of our knowledge, we are the first to build the NAS search scheme according to the real-world ReID training schemes, which provides more appropriate supervision for the ReID architecture search. * We propose a novel search space based on the Multi-Scale Interaction (MSI) operations and a Spatial Alignment Module (SAM) to improve the model performance on in-domain and cross-domain scenarios. * We construct a light-weight yet effective architecture for ReID tasks, denoted as MSINet. With only 2.3M parameters, MSINet surpasses ResNet50 <cit.> by 9% mAP on MSMT17 <cit.> and 16% mAP on MSMT17→Market-1501 <cit.>. § RELATED WORKS Neural Architecture Search. NAS has been increasingly appealing to the computer vision society, due to its automatic architecture designing characteristics. NAS methods can be roughly separated into four categories: reinforcement learning <cit.>, evolutionary algorithms <cit.>, gradient desent <cit.> and performance prediction <cit.>. Liu et al. establish a differentiable architecture search (DARTS) method <cit.>, which improves the practicability of NAS by a large extent. Some later works further improve the structure through sampling strategy <cit.>, network pruning <cit.>, progressive learning <cit.>, collaborative competition <cit.>, etc. Most of NAS works focus on general image classification tasks, where the training and validation sets share the exact same categories. Following the setting, however, leads to incompatibility with the real-world training schemes of object ReID. In this work, we unbind the category bond between the two sets and propose a novel search scheme suitable for ReID. ReID Network Design. Current ReID works mostly adopt backbones designed for image classification <cit.>. Some works <cit.> design attention modules based on the common backbones to unearth their potential on distinguishing local distinctions. However, these methods usually lead to large calculation consumption. There are also several works focusing on designing ReID-specific architectures. Li et al. present a Filter Pairing Neural Network to dynamically match patches in the feature maps <cit.>. Wang et al. separate and regroup the features of two samples with a WConv layer <cit.>. Guo et al. extract multi-scale features to directly evaluate the similarity between samples <cit.>. However, the siamese structure is inconvenient when conducting retrieval on large galleries. Zhou et al. aggregate multi-scale information to achieve high accuracy with small computing consumption <cit.>. Quan et al. introduce a part-aware module into the DARTS search space <cit.>. Li et al. propose a new search space in regard to receptive field scales <cit.>. These methods have excellent performance on limited parameter scales, but fail to surpass those networks with complex structures. Different from previous works, we design a light-weight searching structure focusing on rational interaction operations between multi-scale features. The searched MSINet surpasses SOTA methods on both in-domain and cross-domain tasks. § METHODS Our goal is to construct an effective NAS scheme to search for a light-weight backbone architecture suitable for ReID tasks. Based on the training schemes of ReID, we propose a novel Twins Contrastive Mechanism to provide more appropriate supervision for the search process. Aiming at rational interaction between multi-scale features, we design a Multi-Scale Interaction search space. We further introduce a Spatial Alignment Module to improve the generalization capability with limited parameter growth. §.§ Twins Contrastive Mechanism NAS aims at automatically searching for the optimal network architecture for certain data. Inspired by <cit.>, a basic differentiable architecture search scheme is established. We define the ordinary model parameters as ω, and architecture parameters as α. For network layer i with a search space of 𝒪, α_i controls the weight of each operation o in the space. The features are parallelly passed through all the operations, and the final output is formulated by the softmax-weighted sum of operation outputs: f(𝐱_i)=∑_o∈𝒪exp{α^o_i}/∑_o^'∈𝒪exp{α^o^'_i}· o(𝐱_i). The search process is conducted in an alternative manner. Training data is utilized to update the model parameters, and validation data is then employed to update the architecture parameters. For most NAS methods designed for image classification tasks, the training and validation data share exactly the same categories and a linear classification layer for loss calculation. Different from standard image classification, as an open-set retrieval task, ReID has different categories in the training and validation sets. The incompatibility between search schemes and real-world training schemes might lead to sub-optimal searching results. Accordingly, we propose a novel Twins Contrastive Mechanism (TCM) for NAS ReID training. Specifically, we employ two independent auxiliary memories 𝒞_tr and 𝒞_val to store the embedded features of the training and validation data, respectively. The memories are initialized with the centroid features, which are calculated by averaging the features of each category. At each iteration, the training loss is first calculated with 𝒞_tr for model parameter updating. Given an embedded feature 𝐟 with category label j, the contrastive classification loss is calculated with: ℒ^cls_tr=-logexp(𝐟·𝐜^j_tr/τ)/∑^N^c_tr_n=0exp(𝐟·𝐜^n_tr/τ), where 𝐜^n_tr represents the memorized feature of category n, N^c_tr stands for the total number of categories in the training set, and τ is the temperature parameter, which is set as 0.05 empirically <cit.>. After updating the model parameters, the embedded feature 𝐟 with category label j is integrated into the corresponding memorized feature 𝐜^j_tr by: 𝐜^j_tr←β𝐜^j_tr + (1-β) 𝐟, where β is set as 0.2 empirically <cit.>. Then the updated model is evaluated on the validation data to generate to validation loss with 𝒞_val replacing 𝒞_tr in Eq. <ref>. The architecture parameter is then updated with the validation loss to finish an iteration. As the loss calculation does not rely on the linear classification layer, the categories of the training and validation sets are unbound. We are able to dynamically adjust the category overlap ratio in these two sets. The advantages of a proper overlap ratio are summarized as two folds. Firstly, TCM better simulates the real-world training of ReID and helps the model focus on truly discriminative distinctions. The differences between the training and validation data improves the generalization capability of the model. Secondly, a relatively small proportion of overlapped categories stabilizes the architecture parameter update through a consistent optimizing target with the model parameter update. §.§ Multi-Scale Interaction Space Although the local perspective and multi-scale features have already been investigated in previous ReID works <cit.>, the utilization of these information is mainly empirically designed aggregation, which is monotonous and restrained. We argue that on the one hand, the rational utilization of multi-scale features should be dynamically adjusted along the shallow and deep layers of the network. On the other hand, introducing interaction other than aggregation creates direct information exchange, and makes fuller use of multi-scale features. Therefore, we propose a novel Multi-Scale Interaction (MSI) search space to establish a light-weight architecture suitable for ReID. As shown in Fig. <ref>, the network is mainly grouped with MSI cells and down-sample blocks, which is generally consistent with OSNet <cit.>. In each cell, the input features are passed through two branches with different receptive field scales. To reduce the calculation burden of the network, for the layers inside each branch, we adopt the stack of 1×1 convolution and multiple depth-wise 3×3 convolution to implement specific scales. A scale ratio ρ of 3:1 is selected for the two branches. These two branches do not share model parameters, except for the Interaction Modules (IM). IM introduces information exchange for the two branches. There are 4 operation options for the IM. With the two-branch input features defined as (𝐱_1, 𝐱_2), the operations can be formulated as: None. None operation involves no parameters, and outputs exactly the input features (𝐱_1, 𝐱_2). Exchange. Exchange acts as the strongest interaction among all options. It directly exchanges the features for the two branches and outputs (𝐱_2, 𝐱_1). Exchange contains no extra parameters, as well. Channel Gate. Channel gate introduces a Multi-Layer Perceptron (MLP) to generate a channel-wise attention gate <cit.> as: G(𝐱)=σ(MLP(𝐱))), and returns (G(𝐱_1)·𝐱_1, G(𝐱_2)·𝐱_2). The MLP is composed of 2 fully connected layers and its parameters are shared for both branches. Thereby it achieves interaction by jointly screening discriminative feature channels. Cross Attention. Traditional channel attention module calculates the channel correlation inside a single feature map<cit.>. The original feature map 𝐱∈ℛ^C× H× W is firstly reshaped into the query feature 𝐱̃∈ℛ^C× N, where N=H× W. Then the correlation activation is calculated by performing a matrix multiplication between the query feature 𝐱̃ and the key feature 𝐱̃^⊤. We propose to exchange the keys of the two branches to explicitly calculate the correlation between each other. The correlation activation is then transformed to a mask, and is added up to the original features with a learnable proportion. After interaction, the multi-scale branches are fused through a sum operation. It is worth noting that the extra parameters brought by multiple interaction options are limited, which enables searching for each cell along the whole network independently. At the beginning of the network, we employ the same stem module as that in OSNet <cit.>, containing a 7× 7 convolutional layer and a 3× 3 max pooling with a stride of 2. After the searching process, the interaction operation o with the largest weight α_i^o at each layer is reserved to form the searched architecture. After searching the architecture, the model is validated on various Re-ID tasks. The training is constrained by the classification id loss and the triplet loss, formulated by: ℒ_id=1/N∑_i=1^N-log(exp𝐖^⊤_i𝐟_i/∑_jexp𝐖^⊤_j𝐟_i), where 𝐟_i is a feature vector, the corresponding classifier weight of which is 𝐖_i, and ℒ_tri=[𝒟(𝐟_a,𝐟_p)-𝒟(𝐟_a,𝐟_n)+m]_+, where 𝐟_a, 𝐟_p, 𝐟_n are the embedded features for the anchor, the hardest positive and negative samples in a mini-batch, 𝒟(·,·) is the Euclidean distance, m is the margin parameter, and [·]_+ is the max(·,0) function. §.§ Spatial Alignment Module The retrieval precision of object ReID tasks are largely affected by the variation of appearances such as poses, illumination and occlusion when the camera conditions change. In order that the model correctly and consistently focuses on the discriminative spatial positions, we design a Spatial Alignment Module (SAM) to explicitly align the spatial attention between images, as shown in Fig. <ref>. Specifically, we first calculate the position-wise correlation activation map 𝐀 between the feature maps in a mini-batch. The activation between sample i and j can be formulated as: 𝐀(i, j)=𝐱̃_j^⊤×𝐱̃_i, where 𝐱̃∈ℛ^C× N is reshaped from the original feature 𝐱∈ℛ^C× H× W. Then we take the maximum activation for each position of sample i as: 𝐚(i,j)=max_dim=1𝐀(i,j). The above process is denoted as “Mutual Conv” in Fig. <ref>. We evaluate the consistency between activation vectors with cosine similarity. For negative samples specifically, there can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aligning all the correlations for sample i, we hope that the network can correct some attention bias and consistently focus on discriminative positions. However, through aligning positive sample pairs, the ID-related features are expected to be emphasized, which cannot be achieved by aligning negative pairs. Therefore, we introduce an extra position activation module (PAM) to generate supervision for the alignment between positive pairs. The spatial alignment loss is formulated as: ℒ_sa(i)=1/N_+∑_p∈ℐ_+(1-S(𝐚̂(i), 𝐚(i,p)))+ 1/N_-∑_n_1,n_2∈ℐ_-(1-S(𝐚(i,n_1), 𝐚(i,n_2))), where ℐ_+ contains positive indices for sample i, the total number of which is N_+, and vice versa. 𝐚̂(i) stands for the generated activation vector for positive sample alignment, and S(·,·) is the cosine similarity. § EXPERIMENTS §.§ Datasets and Evaluation Metrics Our proposed method is evaluated on two person ReID datasets Market-1501 <cit.>, MSMT17 <cit.>, and two vehicle ReID datasets VeRi-776 <cit.> and VehicleID <cit.>. For simplicity, the four datasets are denoted as M, MS, VR and VID in the following sections, respectively. Evaluation metrics include Cumulative Matching Characteristic (CMC) and mean average precision (mAP), which are commonly utilized on ReID tasks. §.§ Architecture Search We conduct the searching process on MSMT17. SGD is adopted for model parameter update with an initial learning rate of 0.025. The model is trained for 350 epochs in total. We adopt a warm-up strategy for the first 10 epochs. Then the learning rate is decayed by 0.1 at 150, 225 and 300 epochs, respectively. Adam <cit.> is adopted for the architecture parameter update with an initial learning rate of 0.002. The learning rate is decayed at the same pace. The images are reshaped to 256×128 for person and 256×256 for vehicles. Data augmentation includes random flip, random crop and random erasing <cit.>. The searched architecture is presented in Tab. <ref>. The “MSINet” in the following experiment sections refers to this architecture. We visualize the feature maps extracted by each MSI cell in Fig. <ref>. At the shallow layers of the network, the kernels mainly focus on overall contour information. Channel gate helps to filter out inferior information, such as the background. As we approach deeper layers, the extracted features each have specific semantic information, where cross attention is more likely to be selected for the interaction. It indicates that cross attention is more rational for exchanging high-level semantic information. §.§ Comparison with Other Backbones We first compare our proposed MSINet with ResNet50 and recent proposed light-weight backbones in both in-domain and cross-domain ReID scenarios. In-Domain Tasks. We adopt a two-group supervised evaluation scheme similar to that in <cit.>: training from scratch and fine-tuning ImageNet<cit.> pre-trained models. The training parameters for both schemes are kept the same as that in architecture search, except for an initial learning rate of 0.065. Triplet loss and cross entropy loss are adopted for the parameter update. The margin m in Eq. <ref> is set as 0.3. <cit.> adopts an FBLNeck. We also employ the same structure. The results are shown in Tab. <ref>. ResNet50 is the most commonly utilized backbone network in ReID tasks, yet holds the worst performance. Moreover, ResNet50 largely depends on ImageNet pre-training, while MSINet without pre-training has already surpassed pre-trained ResNet50 on all metrics. Compared with the other datasets, MS contains more variations on illumination, background and camera pose, and brings a large performance gap between ResNet50 and other methods. It also validates the inadequacy of image classification networks on ReID tasks. OSNet <cit.> and CDNet <cit.> are recently proposed architectures designed specifically for ReID tasks. Both architectures focus on fusing multi-scale features to better suit ReID. CDNet employs a traditional NAS scheme to search for the proper receptive field scales for each cell. MSINet fixes the receptive field scale and instead selects optimal interaction operations inside each cell. With only a bit more parameters, MSINet surpasses all the other backbones by a large margin. Cross-Domain Tasks. Cross-domain experiments verify the generalization capability of the model. Following previous domain generalizable ReID works <cit.>, data augmentation is adjusted to random flip, random crop and color jittering. The model is pre-trained and fine-tuned for 250 epochs to avoid over-fitting. The other settings are kept the same as supervision scenes. With no present pre-trained models for CDNet <cit.>, it is excluded from this section. Tab. <ref> shows that ResNet50 can be easily interfered by different image styles confronted with new image domains. OSNet learns multi-scale features with specific semantic information for ReID, which is domain invariant to some extent. Our proposed search scheme also takes into account the generalization capability of the model. By partly separating the categories for training and validation sets, the searched interaction operations generalize well confronted with new image domains. Except for discrimination, MSINet also surpasses the other backbones on cross-domain tasks by a large margin with faster inference speed. Additionally, we introduce SAM into the model, which aligns the spatial correlations between images. A weighted sum of ReID loss and spatial alignment loss is utilized when training the network with SAM. The weight of spatial alignment loss is set as λ_sa=2.0. Without extra inference consumption or damages on the supervised performance, SAM further boosts the generalization capability of MSINet. §.§ Comparison with State-of-the-art Methods Tab. <ref> further illustrates the supervision performance comparison of our proposed MSINet with the SOTA methods on M and MS datasets. With much less parameters than most of the compared methods, MSINet achieves a retrieval accuracy comparable to that of more complicated ones. Auto-ReID <cit.> first designs a NAS scheme for ReID, yet the DARTS-style architecture contains 13M parameters. RGA-SC <cit.> carefully designs a relation-aware global attention module. MSINet achieves even higher performance with less training consumption, which validates the superiority of selecting rational interaction. We also evaluate the model performance replacing the backbone network from ResNet50 to MSINet for SOTA unsupervised ReID methods in Tab. <ref>. For purely unsupervised learning (USL) method GS <cit.> on M dataset, MSINet performs slightly lower on rank-1, yet has a large superiority on mAP. For HDCRL <cit.>, MSINet shows obvious superiority over ResNet50. For unsupervised domain adaptation (UDA) method IDM <cit.> on M→MS task, MSINet surpasses ResNet50 by a large margin, which further proves that the TCM brings outstanding generalization capability to the searched architecture. §.§ Ablation Studies Effectiveness of Architecture Search. To verify the effectiveness, we conduct supervised training on MS with different search schemes in Fig. <ref> (a). Under the standard classification scheme (“CE Overlap”), the searched model performs poor. Replacing the cross entropy loss to contrastive loss (“TCM Overlap”) only brings a slight improvement. The complete TCM framework unbounds the categories between the training and validation set, and thereby improves the performance by a large margin. In Fig. <ref> (b) we compare the performance of different architectures. Firstly, we validate 4 models each with a unique interaction operation from the 4 options in the search space. None and Exchange, with no trainable parameters, achieve poor performance. Channel gate introduces channel-wise attention, whose model performs the best among 4 options. Cross attention exchanges the key features for the two branches. Over-frequent exchange interferes the ordinary feature extraction and degrades the network performance. Through appropriately arranging interaction operations along the architecture, MSINet surpasses all the above 4 models. Random architecture, on the other hand, shows no rational appliance of interaction operations, which validates that the proposed search scheme helps find suitable architectures for ReID. Effectiveness of Spatial Alignment Module. We validate the effectiveness of each components of SAM on the VR→VID cross-domain experiment in Tab. <ref>. Firstly, we introduce the spatial alignment for positive and negative sample pairs, respectively. Each of them brings certain performance improvements. However, a unified alignment for all sample pairs damages ID-related features and degrades the performance instead. Therefore, we separate the alignment of positive and negative samples, which retains some discriminative features and integrates the effect of both aspects. The extra PAM for positive sample alignment further guarantees the focus on ID-related positions and achieves the best performance. We also conduct in-domain experiment on VR to prove that SAM improves the generalization capability without sacrificing the supervision performance. Adding SAM to OSNet receives similar results, which validates the universality of SAM. Fusing Operation. After interaction, the multi-scale features are fused by sum operation. We investigate several fusing options on MS training from scratch in Fig. <ref> (c). Subtracting (“Minus”) a branch from the other leads to about the same results as “Sum” while multiplication (“Mul”) performs poorly. Comparison with Transformer. Transformer, as a new architecture, has recently been continuously making progresses in many computer vision domains <cit.>, including ReID <cit.>. We compare the model performance with some baseline Transformer models in Tab. <ref>. DeiT-B and ViT-B <cit.> achieves higher performance on MSMT17, with much larger calculation burden compared with our proposed MSINet. On VeRi-776, MSINet surpasses all the baseline Transformer methods. It proves that rational interaction operations between multi-scale features are capable to assist light-weighted pure-CNN models to obtain comparable performance with complex Transformers. Parameter Analysis. Firstly, we study the influence of different receptive field scale ratios ρ inside an MSI cell on MS training from scratch in Fig. <ref> (d). Introducing scale differences between branches improves the model performance significantly, and subsequent increases brings more modest impacts. Considering both parameter scales and model performance, the ratio of 3:1 is selected for MSINet. Secondly, the model performance fluctuation influenced by spatial alignment weight is visualized in Fig. <ref> (e). The experiment is conducted on the VR→VID cross-domain scenario. Employing the alignment generally makes a positive impact on the generalization capability of the model. The optimal loss weight λ_sa locates at 2.0. Visualization Results. We visualize the top-15 retrieved sequences and the corresponding distances from an query image on VR in Fig. <ref>. By comparison, ResNet50 mainly focuses on general appearance features, where the top-rated negative samples share similar car bodies. MSINet, oppositely, concentrates on discriminative distinctions, an empty car hopper in this case, and creates an evident distance gap between positive and negative samples. More details can be seen in the supplementation material. § CONCLUSION In this paper, we design a Twins Contrastive Mechanism for NAS to build the compatibility with ReID. The task-specific search scheme provides the searching process with more appropriate supervision. A Multi-Scale Interaction search space is proposed to establish rational and flexible utilization of multi-scale features. With a Spatial Alignment Module, our proposed MSINet achieves SOTA performance on both supervision and cross-domain scenarios with limited parameter amount. We hope the proposed approach could inspire more works focusing on designing network architectures suitable for ReID tasks. § ACKNOWLEDGEMENT This work is partially supported by National Natural Science Foundation of China (62173302, 62006244), China Scholarship Council (202206320302), Young Elite Scientist Sponsorship Program of China Association for Science and Technology (YESS20200140), and AI Singapore Programme (AISG2-PhD-2021-08-008). ieee_fullname § SEARCH ON VERI-776 We select a training-validation ratio of 60%-80% in the searching process on MSMT17 <cit.>. Without changing any specific configurations, we directly search for the rational interaction operations on VeRi-776 <cit.> dataset. The searched architecture is denoted as MSINet-VR. We compare the structure of MSINet and MSINet-VR in Tab <ref>. Generally, the two searched architecture have common characteristics: Channel Gate is preferred in shallow layers, while Cross Attention is employed for more thorough information interaction in deep layers. Quantitatively, we also conduct relevant supervision and cross-domain experiments with MSINet-VR in Tab. <ref>. All the experiment configurations are kept the same as those of MSINet training. Although there are some fluctuations, generally MSINet-VR has similar performance to MSINet, and the retrieval accuracy still surpasses ResNet50 <cit.> by a large margin. § SEARCH WITH DIFFERENT OVERLAP RATIOS With the identities of training and validation sets unbound, we conduct a series of experiments utilizing different data separation ratios in Tab. <ref> to find the appropriate interaction operations for the network. Firstly, we separate the training and validation sets completely with no identity overlaps. It can be observed that a balanced train-validation ratio generally brings better performance. For the two extremes of data distribution, an over-small validation set makes the architecture optimizer stuck in local minima and achieves poor performance. On the contrary, an over-large validation set brings no severe damage to the architecture search process, despite that the model is still not optimal. It demonstrates that abundant validation data is essential for ReID NAS. Secondly, we randomly select part of identities, and evenly divide their images into the training and validation sets. The experiment results suggest that having a relatively small proportion of overlapped identities, whose images have been partly utilized for model parameter update, stabilizes the searching process and leads to a better architecture. However, when the overlap increases to a certain extent, the resemblance between the training and validation sets will bring negative influence to the ReID architecture search. As a comparison, we conduct the search task with traditional NAS scheme where a linear classification layer and cross entropy loss are employed for the training and validation data, the searched model of which performs worse than our proposed TCM. Combined with above rules and the model performance, we select the architecture searched with the train-validation split of 60%-80% as the proposed MSINet. § SEARCH WITH SOFTMAX LOSS We further compare the detailed interaction operations between MSINet and the architecture searched under traditional NAS scheme, where softmax loss and a unified linear classification layer are utilized for the training and validation sets <cit.> (denoted as MSINet-S) in Tab. <ref>. Compared with MSINet and MSINet-VR, where direct information exchange mainly appears at deep layers, MSINet-S contains a large amount of Exchange and Cross Attention along the whole network. The over-frequent information exchange fails to focus on discriminative features. It also validates the effectiveness of our proposed Twins Contrastive Mechanism on searching for architectures suitable for ReID. § VISUALIZATION RESULTS Some additional visualization results are illustrated to further manifest the effectiveness of our proposed architecture. Firstly, we visualize an example comparison of the top-20 retrieved sequences between ResNet50 and MSINet on MSMT17 in Fig. <ref>. ResNet50 mainly focus on general appearance information, while our proposed MSINet concentrates on discriminative distinctions, the hand bag in this case. Even though positive samples have large appearance differences from the query image, MSINet is still capable to distinguish them. Secondly, example activation maps of ResNet50 and our proposed MSINet on Market-1501 <cit.> are visualized in Fig. <ref>. ResNet50 mainly focuses on the right part of the image, including some background areas. Our proposed MSINet, oppositely, is capable to dynamically focus on discriminative distinctions of each image. Thirdly, to intuitively demonstrate the effectiveness of the Spatial Alignment Module (SAM) on enhancing the attention consistency of the model confronted with images from different sources, we visualize example activation maps on the task of VR→VID. As shown in Fig. <ref>, without alignment, the model can have different activated positions on different images of the same identity, even if they share similar appearances. § COMPARISON AND ADVANTAGES TO OSNET (1) OSNet simply sums up the features of each branch, without detailed exploration on the interaction between branches. In comparison, MSINet practically select rational interaction operations for different network layers. Consequently, MSINet surpasses OSNet not only in supervised, but also in domain generalization performance by a large margin. (2) OSNet contains 4 branches with different receptive field scales, where there exists certain parameter redundancy. We validated in the early exploring that removing the branches with receptive field scales of 3 and 5 has little influence to the model performance. MSINet reduces the number of branches, and increases the scale difference between two branches, which increases the parameter amount by a little bit but significantly reduces the inference time. § DETAILED ANALYSIS ON SAM We compare the proposed SAM module to some previous attention-based methods and analyze it in detail. <cit.> regularizes the attention generated at different network layers for the same image; <cit.> explicitly enforces the longitudinal activation distribution to be the same for two images, which may lead to misalignment if the objects are not properly detected. For negative samples, there can be many different hints for recognition, some of which might be inappropriate, such as the backgrounds. By aggregating the information from different negative samples, the network is driven to only focus on discriminative regions. The motivation of SAM is different from the above two works. For the in-domain setting, the camera condition diffrences are directly addressed by supervised learning. Thus, SAM brings limited improvements, yet doesn't defect the performance, compared to techniques like instance normalization. § MORE ABLATION STUDY Additional cross-domain experiments. We add the M→MS and VID→VR experiment results to Tab. <ref>. MSINet surpasses OSNet on all metrics. Ablation study on softmax operation. SAM aligns the activation values in the feature map, where the discriminative positions are actually matched between different samples. As the “Mutual Conv” operation is non-parametric, it is not proper to apply the softmax-squeezed position attention values for direct alignment, which may result in scale inconsistency. The experiment results in Tab. <ref> also suggest slight influence on this detail. § LIMITATIONS AND FUTURE WORK The designed interaction operations only include forward and exchange in the direct and attention forms, which restricts the size of the search space. In the future works, there are still exploration room for more elaborate and complicated interaction operations and search spaces. There are still exploration room for more complicated search schemes and spaces.

http://arxiv.org/abs/2303.06836v2

20230313034637

Label Information Bottleneck for Label Enhancement

cs.LG

Label Information Bottleneck for Label Enhancement Qinghai Zheng^1,  Jihua Zhu^2[1]Corresponding author, E-mail: zhujh@xjtu.edu.cn,  Haoyu Tang^3 ^1College of Computer and Data Science, Fuzhou University, China ^2School of Software Engineering, Xi'an Jiaotong University, Xi'an, China ^3School of Software, Shandong University, Jinan, China Received: date / Accepted: date ======================================================================================================================================================================================================================================================================================================================== In this work, we focus on the challenging problem of Label Enhancement (LE), which aims to exactly recover label distributions from logical labels, and present a novel Label Information Bottleneck (LIB) method for LE. For the recovery process of label distributions, the label irrelevant information contained in the dataset may lead to unsatisfactory recovery performance. To address this limitation, we make efforts to excavate the essential label relevant information to improve the recovery performance. Our method formulates the LE problem as the following two joint processes: 1) learning the representation with the essential label relevant information, 2) recovering label distributions based on the learned representation. The label relevant information can be excavated based on the “bottleneck” formed by the learned representation. Significantly, both the label relevant information about the label assignments and the label relevant information about the label gaps can be explored in our method. Evaluation experiments conducted on several benchmark label distribution learning datasets verify the effectiveness and competitiveness of LIB. Our source codes are available at <https://github.com/qinghai-zheng/LIBLE> § INTRODUCTION Learning with label ambiguity is important in computer vision and machine learning. Different from the traditional Multi-Label Learning (MLL), which employs multiple logical labels to annotate one instance to address the label ambiguity issue <cit.>, Label Distribution Learning (LDL) considers the relative importance of different labels and draws much attention in recent years <cit.>. By distinguishing the description degrees of all labels, LDL annotates one instance with a label distribution. Therefore, LDL is a more general learning paradigm, MLL can be regarded as a special case of LDL <cit.>. Recently, many LDL methods are proposed and achieve great success in practice <cit.>. Instances with exact label distributions are vital for the training process of LDL methods. Nevertheless, annotating instances with label distributions is time-consuming<cit.>. We take the label distribution annotation process of SJAFFE dataset for example here. SJAFFE dataset is the facial expression dataset, which contains 213 grayscale images collected from 10 Japanese female models, each facial expression image is rated by 60 persons on 6 basic emotions, including happiness, surprise, sadness, fear, anger, and disgust, with a five-level scale from 1 - 5, the higher value indicates the higher emotion intensity. Consequently, the average score of each emotion is served as the emotion label distribution <cit.>. Clearly, the above annotation process is costly and it is unpractical to annotate data with label distributions manually, especially when the number of data is large. Fortunately, most existing datasets in the field of computer vision and machine learning are annotated by single-label or multi-labels <cit.>, therefore, a highly recommended promising solution is Label Enhancement (LE), which attempts to recover the desired label distributions exactly from existing logical labels <cit.>. Driven by the urgent requirement of obtaining label distributions and the convenience of LE, some LE methods are proposed in recent years <cit.>. Given a dataset X = {x_1,x_2, ⋯ ,x_n}∈ℝ^q × n, in which q and n denote the number of dimensions and the number of instances, the potential label set is {y_1,y_2, ⋯ ,y_c}. The available logical labels and the desired distribution labels of X are separately indicated by L = {l_1,l_2, ⋯ ,l_n} and D = {d_1,d_2, ⋯ ,d_n}, where l_i and d_i are: l_i = ( l_i^y_1,l_i^y_2, ⋯ ,l_i^y_c)^T, d_i = ( d_i^y_1,d_i^y_2, ⋯ ,d_i^y_c)^T. To be specific, LE aims to recover D based on the information provided by X and L. For most existing LE methods, their objectives can be concisely summarized as follows: min_θ f_θ(X) - L_F^2 + γ reg(f_θ(X)), in which D = f_θ(X), f_θ(·) indicates the mapping from X to D, reg(·) denotes the regularization function, and γ is the trade-off parameter. Most existing LE methods vary in reg(·). For example, GLLE <cit.> calculates the distance-based similarity matrix of data and employs the smoothness assumption <cit.> to construct reg(·); LESC <cit.> considers the global sample correlations and introduces the low-rank constraint as the regularization; PNLR <cit.> leverages reg(·) to maintain positive and negative label relations during the recovery process. Although a remarkable progress can be made by aforementioned methods, they ignore the label irrelevant information contained in X, which prevents the further improvement of recovery results. For example, in the LE task of recovering facial age label distributions, the label irrelevant information, such as specularities information, cast shadows information, and occlusions information, may result in the incorrect mapping process of f_θ(·) and the unsuitable regularization of reg(·), eventually leads to the unsatisfactory recovery performance. To overcome the aforementioned limitation, we present a Label Information Bottleneck (LIB) method for LE. Concretely, the core idea of LIB is to learn the latent representation H, which preserves the maximum label relevant information, from X, and jointly recovers the label distributions based on the latent representation. For the LE problem, the label relevant information is the information that describes the description degrees of labels. It is tough to explore the label relevant information directly. As shown in Fig. <ref>, we decompose the label relevant information into two components, namely the assignments of labels to the instance and the label gaps between label distributions and logical labels. Inspired by Information Bottleneck (IB) <cit.>, LIB utilizes the existing logical labels to explore the information about the assignments of labels to the instance. Unlike simply employing the original IB on the LE task, our method further considers the information about the label gaps between label distributions and logical labels. It is noteworthy that the above two components of the label relevant information are jointly explored in our method, and that is why we term the proposed method Label Information Bottleneck (LIB). The main contributions can be summarized as follows: ∙We decompose the label relevant information into the information about the assignments of labels to instance and the information about the label gaps between logical labels, both of which can be jointly explored during the learning process of our method. ∙We introduce a novel LE method, termed LIB, which excavates the label relevant information to exactly recover the label distributions. Based on the original IB, which explores the label assignments information for LE, LIB further explores the label gaps information. ∙We verify the effectiveness of LIB by performing extensive experiments on several datasets. Experimental results show that the proposed method can achieve the competitive performance, compared to state-of-the-art LE methods. § RELATED WORK §.§ Label Enhancement To recover the label distributions from the existing logical labels, many efforts are made recently <cit.>. In general, most existing LE methods can be roughly divided into two categories, namely, algorithm adaptation and specialized algorithm <cit.>. Algorithm adaptation extends some existing methods to achieve the goal of LE <cit.>. For example, FCM <cit.> recovers the label distributions by utilizing the fuzzy clustering and fuzzy relabeling. To be specific, FCM utilizes the fuzzy C-means clustering to get different clusters and cluster prototypes, then obtains membership degrees of each instance with respect to different cluster prototypes, finally annotates all instances with label distributions by employing the fuzzy composition and softmax normalization. KM <cit.> leverages the fuzzy SVM to achieve the membership function. During the recovery process, KM separates instances into two clusters and employs the nonlinear function to get the radius and distances between centers and kernelized instances, and then gets the label distributions with the help of the softmax normalization. Specialized algorithm is specially designed to deal with the LE problem. Most existing LE methods belong to the category of specialized algorithm and have the basic objective Eq. (<ref>). By using different constraints, different methods adopt different reg(·) in Eq. (<ref>). For example, based on the assumption that instances closed in the feature space are more likely to share the same label, GLLE <cit.> employs the following local graph information in the feature space to boost the recovery performance: q_i,j={ exp( -x_i-x_j^2/2ε^2),if x_j∈ k( i ), 0, otherwise, . where k( i ) denotes the k-nearest neighbours of x_i. reg(·) in GLLE is constructed as follows: reg(f_θ(X)) = ∑_i,jq_i,jf_θ(x_i) - f_θ(x_j)_2^2. Unlike GLLE, LESC <cit.> considers the global graph information and uses the low-rank representation learning <cit.>: G,Emin  G_*+λ_2E_2,1,  s.t., X=XG+E, where G indicates the low-rank representation of instances in the feature space. The regularization function in LESC is written as follows: reg(f_θ(X)) = f_θ(X) - f_θ(X)G_F^2. For these aforementioned LE methods, they all neglect the label irrelevant information contained in X, the negative effect of which can result in the unsatisfactory recovery results. Taking the recovery of facial emotion label distributions for example, the inaccurate graph information would be obtained in GLLE and LESC with the presence of label irrelevant information, such as the identity information, hindering the further improvement of recovery results. §.§ Information Bottleneck Information bottleneck (IB) <cit.> is an information theoretic principle, which describes the relevant information in data formally. To be concrete, IB has the following objective: min_B -I(B,C),  s.t., I(A,B) ⩽ I_c, where I(·,·) measures the mutual information and I_c is the information constraint. Clearly, IB aims to learn the representation B, which preserves the relevant information about C, from A. Considering the scenario of LE, it is natural to get the following formula: min_H -I(H,L),  s.t., I(X,H) ⩽ I_c. As discussed in Section <ref>, the information merely about the assignments of labels to the instance can be explored based on Eq. (<ref>), which neglects the vital information about the label gaps between logical labels and label distributions. Recently, IB has been successfully utilized in many real-world applications <cit.>. To the best of our knowledge, the method introduced in this paper is the first work that leverages IB to deal with the LE problem. More notably, rather than using IB simply (as shown in Eq. (<ref>)), our method conducts more in-depth exploration to exactly recover label distributions based on IB. § THE PROPOSED METHOD §.§ The Objective Construction Generally, the basic idea can be written as follows: min_H ℒ_as + αℒ_gap, s.t.,I(X,H) ⩽I_c, where ℒ_as excavates the information about the assignments of labels to the instance, ℒ_gap investigates the information about the label gaps between the logical labels and distribution labels, α is the trade-off parameter, the constraint aims to remove the label irrelevant information. The framework of our method is depicted in Fig. <ref>. It's worth noting that employing the original IB for LE merely explores the information about the label assignments. While our LIB makes attempts to capture the information about both the label assignments and the description degrees of labels. §.§.§ Label assignmens information modeling For ℒ_as, inspired by IB, we have the following formula: ℒ_as = -I(H,L). According to the concept of mutual information, ℒ_as can be rewritten out in full as follows: ℒ_as = - ∑_h∑_lp(h,l)logp(l|h)/p(l). For the convenience of optimization, we introduce the variational approximation q(l|h) to p(l|h). Since both the Kullback Leibler divergence and the entropy are positive: KL(p(l|h)||q(l|h)) = ∑_lp(l|h)logp(l|h)/q(l|h)⩾ 0 ⇒∑_lp(l|h)logp(l|h)⩾∑_lp(l|h)logq(l|h), 𝔼_p(l)[-log p(l)] = - ∑_lp(l) log p(l)⩾ 0, based on Markov chain that L←X→H, we can get: L_as⩽ - ∑_x∑_l∑_hp(x,l)p(h|x)log q(l|h). §.§.§ Label gaps information modeling To investigate the label-relevant information about the description degrees of labels, we introduce the label gaps between logical labels and label distributions Δ, and consider the conditional self-information, i.e., I(Δ|H). Therefore, we construct ℒ_gap[It can be also interpreted and derived from the view of the probability distribution: max_Δ log p(H,Δ) ⇒max_Δ log p(Δ |H) + log p(H) ⇒max_Δ log p(Δ |H). We appreciate reviewers for their helpful comments.] as follows: ℒ_gap = I(Δ|H) = -log p(Δ |H)          = - ∑_δ∑_hlog p(δ |h)          = - ∑_l∑_hlog p(l - d̂|h). where δ = l - d̂, d̂ is the label distribution recoveried in our method. §.§.§ Label irrelevant information modeling Regarding the label irrelevant information, LIB employs the constraint in Eq. (<ref>) to discard it during the learning process. I(X,H) can be formulated as follows: I(X,H) = ∑_x∑_hp(x,h)logp(h|x)/p(h). Since it is difficult to calculate p(h) directly, we also introduce the variational approximation q(h) to p(h). Similar to Eq. (<ref>), based on KL(p(h)||q(h)) ⩾ 0, we have: ∑_hp(h)logp(h)⩾∑_hp(h)logq(h). Subsequently, the following formula can be written: I(X,H) ⩽∑_x∑_hp(x,h)logp(h|x)/q(h)                = ∑_x∑_lp(x,l)KL(p(h|x)||q(h)). §.§.§ Objective of LIB By employing the Lagrange multiplier method and combining Eq. (<ref>), (<ref>), (<ref>), and (<ref>), we have: ℒ = ℒ_as + αℒ_gap + β I(X,H)    ⩽ - ∑_x∑_l∑_hp(x,l)p(h|x,l)log q(l|h)        - α∑_l∑_hlog p(l - d̂|h)        + β∑_x∑_lp(x,l)KL(p(h|x)||q(h)). where β is the Lagrange multiplier. Considering the bound of ℒ and using the empirical Monte Carlo approximation of sampling <cit.>, we have the following objective of LIB: .887!ℒ_LIB = 1/n∑_i = 1^n [ - ∑_hp(h|x_i)log q(l_i|h) + βKL(p(h|x_i)||q(h))] - α∑_l∑_hlog p(l - d̂|h). §.§ The Optimization of LIB To minimize the objective of ℒ_LIB, we use the reparameterization trick <cit.>. For p(h|x), we assume that: p(h|x) ∼𝒩(μ _h|x,σ _h|x^2I), where μ _h|x and σ _h|x are obtained by using the encoder network f_θ_en(·), i.e., μ _h|x = f_θ_en^μ(x) and σ _h|x = f_θ_en^σ(x). Subsequently, we have that: h = μ _h|x + σ _h|x⊙ϵ, where ϵ∼𝒩(0,I) and ⊙ is the element-wise product. For q(l|h), we assume: q(l|h) ∼𝒩(μ _l|h,I), where μ _l|h is learned by using the decoder network f_θ_de(·), namely, μ _l|h = f_θ_de(h). For q(h), we assume that: q(h) ∼𝒩(0,I). For p(l - d̂|h), the following assumption is adopted: p(l - d̂|h) ∼𝒩(0,σ _δ|h^2I), where σ _δ|h can be achieved by introducing the gap deviation network f_θ_gd(·), i.e., σ _δ|h = f_θ_gd(h). For the recovered label distribution d̂, we introduce the label distribution network f_θ_ld(·) and has the following formula: d̂ = f_θ_ld(h). Consequently, based on Eq. (<ref>)-(<ref>), we have: .887!1pt  min_θ _en,θ _de,θ _gd,θ _ldℒ_LIB ⇒min_θ _en,θ _de,θ _gd,θ _ld1/n∑_l [1/2μ _l|h - l_2^2 + α (1/2(l - d̂)^T(σ _δ|h^-2I)(l - d̂) + log (σ _δ|h^2I))] + β/2∑_x[μ _h|x^Tμ _h|x + tr(σ _h|x^2I) - log (σ _h|x^2I)]. When the problem of Eq. (<ref>) is optimized, we can effectively recover the desired label distributions. To be specific, given {X,L}, we can obtain H according to Eq. (<ref>) and achieve the recovery results based on Eq. (<ref>), namely, D̂ = f_θ_ld(H). §.§ Comparison with Existing LE Methods The main difference between LIB and existing methods is that our method deals with the problem of LE from the perspective of information bottleneck. Considering the first term in Eq. (<ref>), it aims to minimize d - l_2^2 under the assumption that information in the label distributions is inherited from the initial logical labels <cit.>. For LIB, the more reasonable term: 1/2(l - d̂)^T(σ _δ|h^-2I)(l - d̂) + log (σ _δ|h^2I) which can be deduced by excavating the label relevant information about the label gaps between logical labels and label distributions. Besides, we compare our method with the recently proposed LEVI <cit.> further. Although the objectives of LEVI and LIB are somewhat similar in form, they are essentially different as follows: 1) LIB makes attempts from the perspective of information bottleneck, while LEVI from the view of variational inference; 2) The formulas of LEVI and LIB are just partially similar in form, since the variational inference is employed as the optimization tool in LIB. The details of these two formulas are totally different; 3) LEVI requires an extra regularizer, i.e., d - l_2^2, to constrain the recovery process, while LIB achieves d based on the more reasonable term, i.e., Eq. (<ref>). § EXPERIMENTS To verify the effectiveness and competitiveness of LIB, extensive experiments are conducted in this section. §.§ Experimental Setup As shown in Table <ref>, we use both one toy dataset and 13 real-world datasets for evaluation[http://palm.seu.edu.cn/xgeng/LDL/index.htm]. For the toy dataset, i.e., Artificial dataset, it is utilized to vividly show the recovery performance <cit.>. Movie dataset is collected from movies, SBU-3DFE and SJAFFE datasets are two facial expression datasets. Yeast datasets (alpha to spoem) are collected from 10 biological experiments on the budding yeast genes <cit.>. It is important to note that only the ground-truth label distributions are provided by these datasets. Therefore, we adopt the binarization strategy, which is also used in existing LE works <cit.>, to ensure the consistency of evaluation. We compare our method LIB to 7 LE methods, including FCM <cit.>, KM <cit.>, LP <cit.>, ML <cit.>, GLLE <cit.>, LESC <cit.>, and LEVI <cit.>. The first two methods belong to the algorithm adaption, and the rest methods are specialized algorithms. For the sake of fairness, we utilize the parameter settings recommended in their original works. Specifically, for FCM, we set the parameter β = 2. For KM, we leverage the Gaussian kernel. For LP, we set the parameter α=0.5. For ML, we set the number of neighbors k=c+1. For GLLE, we select λ from {0.01,0.1,...,100} and set the number of neighbors k to c+1. For LESC, λ_1 and λ_2 are selected from {0.0001,0.1,...,10}. For LEVI, MLPs with two hidden layers and softplus activation functions are utilized, and the results are reported after 150 training epochs. For LIB, we select α and β from {0.001,0.01,...,10}, and the fully connected networks with 3 layers and sigmoid activation function are leveraged in the proposed method. To evaluate the recovery performance, we adopt 6 metrics, namely Chebyshev, Canberra, Clark, Kullback-Leibler, Cosine, and Intersection <cit.>. Given the ground-truth label distribution d and the recovered label distribution d̂, the first four metrics and the rest two metrics respectively measure the distance and similarity between d and d̂: D_Chebyshev(d,d̂) =max_i| d^y_i-d̂^y_i|, D_Canberra(d,d̂) = ∑_i=1^c| d^y_i-d̂^y_i|/d^y_i+d̂^y_i, D_Clark(d,d̂) = √(∑_i=1^c( d^y_i-d̂^y_i)^2/( d^y_i+d̂^y_i)^2), D_Kullback-Leibler(d,d̂) = ∑_i = 1^c d^y_ilnd^y_i/d̂^y_i, S_Cosine(d,d̂) = ∑_i=1^cd^y_id̂^y_i/√(∑_i=1^c( d^y_i)^2)√(∑_i=1^c( d̂^y_i)^2), S_Intersection(d,d̂) = ∑_i=1^cmin( d^y_i,d̂^y_i). The smaller values of distance metric and similarity metric indicate the better and the worse results, respectively. §.§ Visualization Results on Toy Dataset The recovery results on the Artificial dataset are vividly presented in Fig. <ref>, which shows the three-dimensional label distributions by the RGB color channels separately. The more similar the color patterns of the recovered results and the ground-truth are, the better the recovery results are. It can be seen that FCM, GLLE, LESC, LEVI, and LIB can obtain the similar color pattern, while KM, LP, and ML are incapable to obtain the promising recovery performance on Artificial dataset. Regarding the visualization results of FCM, GLLE, LESC, LEVI, and LIB, the color pattern that is most close to the ground-truth is achieved by our LIB. §.§ Comparison Results on Real-world Datasets We provide the detailed comparison results on 13 real-world datasets in Table <ref>. Overall, LIB has the competitive recovery performance. We have the following observations: 1) Compared with FCM and KM, which belong to the category of algorithm adaption, remarkable improvements can be achieved by our method; 2) Compared with the methods belonging to the category of specialized algorithm, LIB can also obtain better recovery results in most cases. For example, LIB obtains the best recovery results on Movie datasets in all metrics. Moreover, although LESC can obtain slightly favorable results in some cases, the corresponding recovery results of LIB are also promising and competitive. The underlying reason may be that LESC further considers the sample correlations during the recovery process; 3) The recovery performance of all methods can be roughly ranked as LIB>LESC≈LEVI>GLLE>LP≈FCM>ML>KM. We can conclude that LIB is suitable for the LE problem. The underlying reason for the significant improvement is that the label relevant information, including the information about label assignments and the information about label gaps, can be effectively investigated by LIB. §.§ Analysis and Discussion of LIB We analyze the parameter sensitivity of LIB firstly, and then we conduct the ablation studies as well. §.§.§ Sensitivity of LIB In the proposed method, we choose the values of α and β from {0.001,0.01,...,10}. To show the parameter sensitivity of LIB, we conduct experiments on SBU-3DFE datasets with different values of α and β. Regarding the dimension of the learned latent representation, we set it to 256 for all datasets. The experimental results in metrics of Chebyshev distance, Cosine coefficient, and Intersection similarity are provided in Fig. <ref>. It can be observed that LIB method can get promising recovery results and is robust with respect to different values of α and β in a large range. §.§.§ Ablation studies of LIB The ablation studies are conducted to further verify the effectiveness of introducing the label information bottleneck framework for LE. In the proposed objective Eq. (<ref>), ℒ_as and ℒ_gap explore the label assignments information and label gaps information during the recovery process. As can be observed from Eq. (<ref>) and Eq. (<ref>), only the latent representation H can be learned if we employ ℒ_as merely during the recovery process. Consequently, considering the goal of LE, we compare the proposed LIB with the method termed LIB_gap, which only employs ℒ_gap to achieve the recovery results. In other words, LIB_gap investigates the label gaps information in the case of not considering the label assignments information during the recovery process. To be specific, LIB_gap has with the following objective: min_θ _gd,θ _ld1/2∑_x [(l - d̂)^T(σ _δ|x^-2I)(l - d̂) + log (σ _δ|x^2I)]. Notably, σ _δ|x = f_θ_gd(x) and d̂ = f_θ_ld(x), which are different from the objective Eq. (<ref>) utilized in LIB. Only the partial label relevant information, i.e., the information about the label gaps, is explored in LIB_gap. Table <ref> provides the recovery results of LIB_gap and LIB. It can be observed that LIB outperforms LIB_gap in all cases. Compared with LIB, LIB_gap merely makes the effort to explore the label gap information to boost the recovery performance, while LIB excavates both the information about the label assignments and the information about the label gaps jointly. Therefore, the promising recovery performance can be achieved by LIB. Furthermore, according to the results provided in Table <ref> and <ref>, the recovery results of LIB_gap seem to be competitive, which also indicates that the exploration of information about label gaps is beneficial for LE. § CONCLUSION In this paper, we present a new perspective to deal with the Label Enhancement (LE) problem and introduce the novel Label Information Bottleneck (LIB) method. The label relevant information is decomposed into the information about label assignments and the information about label gaps. Consequently, our method transform the LE problem into simultaneously learning the latent representation and modeling the label gaps. Extensive experiments carried on both the toy dataset and real-world datasets verify the competitiveness of LIB. § ACKNOWLEDGMENTS This work was supported by the National Key R&D Program of China under Grant 2020AAA0109602; the Education and Research Foundation for Middle-aged and Young Teacher of Fujian Province under Grant JAT220005; the National Natural Science Foundation of China (NSFC) under Grant 62206156; Alibaba Group through Alibaba Innovative Research Program, No.21169774. ieeefullname

http://arxiv.org/abs/2303.07228v2

20230313160218

Estimate distillable entanglement and quantum capacity by squeezing useless entanglement

quant-ph

These two authors contributed equally. These two authors contributed equally. wangxin73@baidu.com Institute for Quantum Computing, Baidu Research, Beijing 100193, China Entanglement distillation is crucial in quantum information processing. But it remains challenging to estimate the distillable entanglement and its closely related essential quantity, the quantum capacity of a noisy quantum channel. In this work, we propose methods for evaluating both quantities by squeezing out useless entanglement within a state or a quantum channel, whose contributions are expected to be ignored for the distillable entanglement or the quantum capacity, respectively. We first consider a general resource measure called the reverse divergence of resources to quantify the minimum divergence between a target state and the set of free states. We then introduce the reverse max-relative entropy of entanglement and apply it to establish efficiently computable upper bounds on the distillable entanglement. We also extend the reverse divergence of resources to quantum channels and derive upper bounds on the quantum capacity. We further apply our method to investigate purifying the maximally entangled states under practical noises, such as depolarizing and amplitude damping noises, and notably establish improvements in estimating the one-way distillable entanglement. Our bounds also offer useful benchmarks for evaluating the quantum capacities of qubit quantum channels of interest, including the Pauli channels and the random mixed unitary channels. Estimate distillable entanglement and quantum capacity by squeezing useless entanglement Xin Wang March 30, 2023 ======================================================================================== § INTRODUCTION §.§ Background Quantum entanglement is the most nonclassical manifestation of quantum mechanics and has become a key ingredient in many quantum information processing tasks, e.g., teleportation <cit.>, superdense coding <cit.>, and quantum cryptography <cit.>. More specifically, the entanglement resource is usually assumed to be copies of the maximally entangled states. In a practical scenario, noises inevitably occur in quantum states, resulting in some mixed entangled states, which leads to the need for distillation or purification. A natural question is how to obtain the maximally entangled states from a source of less entangled states using well-motivated operations, known as the entanglement distillation. One fundamental measure for characterizing the entanglement distillation is the one-way distillable entanglement <cit.>, denoted by E_D,→, which is also one of the most important entanglement measures motivated by operational tasks. It captures the highest rate at which one can obtain the maximally entangled states from less entangled states by one-way local operations and classical communication (LOCC): E_D,→(ρ_AB)=sup{r:lim_n →∞ [inf_ΛΛ(ρ_AB^ n)- Φ(2^rn)_1]=0}, where Λ ranges over one-way LOCC operations and Φ(d)=1/d∑_i,j=1^d |ii⟩⟨jj| is the standard d⊗ d maximally entangled state. Likewise, the two-way distillable entanglement E_D,↔(ρ_AB) is defined by the supremum over all achievable rates under two-way LOCC. We have for all bipartite states ρ_AB that E_D,→(ρ_AB) ≤ E_D,↔(ρ_AB). Notably, the distillable entanglement turns out to be closely related to the central idea of quantum capacity in quantum communication tasks, which is central to quantum Shannon theory. Consider modelling the noise in transmitting quantum information from Alice to Bob as a quantum channel _A→ B. The quantum capacity Q(_A→ B) is the maximal achievable rate at which Alice can reliably transmit quantum information to Bob by asymptotically many uses of the channel. By the state–channel duality, if the distillation protocol of the Choi state <cit.> J_AB^ of _A→ B yields the maximally entangled states at a positive rate, then Alice may apply the standard teleportation scheme to send arbitrary quantum states to Bob at the same rate. Thus, one has Q(_A→ B) ≥ E_D,→(J_AB^) since classical forward communication in teleportation does not affect the channel capacity. For the teleportation-simulable channels <cit.>, the equality here holds. Many recent works <cit.> have studied the properties of quantum communication over quantum channels, e.g., non-additivity and superactivation, which determines the complexity of calculating the quantum capacity. Despite many efforts that have been made in the past two decades, computing E_D,→(·) and Q(·) still generally remains a challenging task. Even for the qubit isotropic states and the depolarizing channels, it remains unsolved. Therefore, numerous studies try to estimate them by deriving lower and upper bounds (see, e.g., <cit.> for the distillable entanglement, e.g., <cit.> for the quantum capacity). For the distillable entanglement, a well-known lower bound dubbed Hashing bound is established by Devetak and Winter <cit.>, i.e., E_D,→(ρ_AB) ≥ I_(A⟩ B)_ρ, where I_(A⟩ B)_ρ is the coherent information of the bipartite quantum state ρ_AB. Considering upper bounds, the Rains bound <cit.> is arguably the best-known efficiently computable bound for the two-way distillable entanglement of general states, which can consequently be an upper bound of E_D,→(·). Recent works <cit.> utilize the techniques of finding upper bounds by constructing meaningful extended states. For quantum capacity, many useful upper bounds for general quantum channels are studied for benchmarking arbitrary quantum noise <cit.>. When considering some specific classes of quantum channels, useful upper bounds are also developed to help us better understand quantum communication via these channels <cit.>. In specific, due to the regularization in the characterizations of E_D,→(·) and Q(·), one main strategy to establish efficiently computable upper bounds on them is to develop single-letter formulae. For example, one common approach is to decompose a state (resp. a quantum channel) into degradable parts and anti-degradable parts <cit.>, or use approximate degradability (anti-degradability) <cit.>. Another recent fruitful technique called flag extension optimization <cit.> relies on finding a degradable extension of the state or the quantum channel. However, the performance of these methods is limited by the absence of a good decomposition strategy. It is unknown how to partition a general state or quantum channel to add flags or how to construct a proper and meaningful convex decomposition on them. Thus, the flag extension optimization is only effective for the states and channels with symmetry or known structures. §.§ Main contributions This work considers a family of resource measures called reverse divergence of resources. With a specific construction, we define the reverse max-relative entropy of entanglement for quantum states, which has applications for estimating the distillable entanglement. In the meantime, we introduce reverse max-relative entropy of anti-degradability for quantum channels as a generalization of the concept of that for states, which can be applied to bound the quantum capacity. All these bounds can be efficiently computed via semidefinite programming <cit.>. Furthermore, drawing on the idea of <cit.>, we thoroughly analyze different continuity bounds on the one-way distillable entanglement of a state in terms of its anti-degradability. Finally, we investigate the distillation of the maximally entangled states under practical noises and focus on the quantum capacity of qubit channels. We show that the bound obtained by the reverse max-relative entropy of entanglement outperforms other known bounds in a high-noise region, including the Rains bound and the above continuity bounds. The upper bound offered by the reverse max-relative entropy of anti-degradability also provides an alternative interpretation of the no-cloning bound of the Pauli channel <cit.>, and notably outperforms the continuity bounds on random unital qubit channels. The rest of this paper is structured as follows. We start with some preliminaries used throughout the paper in section <ref>. In section <ref>, we introduce our conceptual contribution, a family of resource measures called the reverse divergence of resources. In section <ref>, we present the application of this concept on the distillable entanglement. Specifically, in subsection <ref>, we show how the reverse max-relative entropy of entanglement can be used to derive a useful upper bound on the one-way distillable entanglement. We derive different continuity bounds in subsection <ref>, with which we compare our bounds and the Rains bound by practical distillable states in subsection <ref>. We also extend our method to the two-way distillable entanglement in subsection <ref>. In section <ref>, we show the application of our method in deriving upper bounds on quantum capacity. We establish some analytical results for Pauli channels and compare our bounds with continuity bounds in subsection <ref> for random mixed unitary channels. In section <ref>, we conclude with a summary and some outlooks for future research. § REVERSE DIVERGENCE OF RESOURCES §.§ Preliminaries Let be a finite-dimensional Hilbert space, and () be the set of linear operators acting on it. We consider two parties Alice and Bob with Hilbert space ℋ_A, ℋ_B, whose dimensions are d_A, d_B, respectively. A linear operator ρ∈() is called a density operator if it is Hermitian and positive semidefinite with trace one. We denote the trace norm of ρ as ρ_1 = √(ρ^†ρ) and let () denote the set of density operators. We call a linear map CPTP if it is both completely positive and trace-preserving. A CPTP map that transforms linear operators in (_A) to linear operators in (_B) is also called a quantum channel, denoted as _A→ B. For a quantum channel _A→ B, its Choi-Jamiołkowski state is given by J_AB^≡∑_i, j=0^d_A-1|i⟩⟨j|𝒩_A → B(|i⟩⟨j|), where {|i⟩}_i=0^d_A-1 is an orthonormal basis in _A. The von Neumann entropy of a state ρ_A is S(A)_ρ := - (ρ_A logρ_A) and the coherent information of a bipartite state ρ_AB is defined by I_c(A⟩ B)_ρ := S(B)_ρ - S(AB)_ρ. The entanglement of formation of a state ρ_AB is given by E_F(ρ_AB) = min_{p_i, |ϕ_i⟩}∑_i p_i S(A)_ϕ_i, where ρ_AB=∑_i p_i |ϕ_i⟩⟨ϕ_i|_AB and the minimization ranges over all pure state decomposition of ρ_AB. We introduce the generalized divergence D(ρ_A σ_A) as a map D: (_A)×(_A) ↦ℝ∪{+∞} that obeys: * Faithfulness: D(ρ_A σ_A)=0 iff ρ_A = σ_A. * Data processing inequality: D(ρ_A σ_A) ≥D[𝒩_A→ A'(ρ_A) 𝒩_A→ A'(σ_A)], where 𝒩_A→ A' is an arbitrary quantum channel. The generalized divergence is intuitively some measure of distinguishability of the two states, e.g., Bures metric, quantum relative entropy. Another example of interest is the sandwiched Rényi relative entropy <cit.> of ρ, σ that is defined by D_α(ρσ):=1/α-1logTr[(σ^1-α/2 αρσ^1-α/2 α)^α], if supp(ρ) ⊂supp(σ) and it is equal to +∞ otherwise, where α∈(0,1) ∪(1, ∞). In the case that α→∞, one can find the max-relative entropy <cit.> of ρ with respect to σ by D_max(ρ || σ) = inf{λ∈ℝ: ρ≤ 2^λσ}. §.§ Reverse divergence of resources In the usual framework of quantum resource theories <cit.>, there are two main notions: i) subset of free states, i.e., the states that do not possess the given resource; ii) subset of free operations, i.e., the quantum channels that are unable to generate the resource. Meanwhile, two axioms for a quantity being a resource measure (·) are essential: 1). Vanishing for free states: ρ∈⇒(ρ) = 0. 2). Monotonicity: ((ρ)) ≤(ρ) for any free operation . (·) is called a resource monotone. Let us define a family of resource measures called reverse divergence of resources: _ℱ(ρ_AB) := min_τ∈ℱD(τ_AB||ρ_AB), where ℱ is some set of free states. By the definition of the reverse divergence of resources in Eq. (<ref>), one can easily check it satisfies condition 1). Whenever the free state set is closed, by the data-processing inequality of D(··), condition 2) will be satisfied. Thus _ℱ(·) is a resource measure. Specifically, in the resource theory of entanglement, some representative free state sets are the separable states (SEP) and the states having a positive partial transpose (PPT). Examples of free sets of operations are LOCC and PPT. We note that the "reverse" here means minimizing the divergence over a free state set in the first coordinate, rather than the second one which has helped one define the relative entropy of entanglement <cit.> and the max-relative entropy of entanglement <cit.>. For some divergence of particular interest, e.g., the quantum relative entropy D(·||·), relevant discussion of the coordinate choices can be traced back to work in <cit.>. In <cit.>, the authors further studied properties of the quantity min_τ∈D(τ||ρ). Here, we try to further investigate meaningful applications of some reverse divergence of resources. In the following, we consider the generalized divergence as the max-relative entropy and study a measure called reverse max-relative entropy of resources, _max,ℱ(ρ_AB) := min_τ∈ℱD_max(τ_AB||ρ_AB), where ℱ is some set of free states. If there is no such state τ_AB∈ that satisfies τ_AB≤ 2^λρ_AB for any λ∈ℝ, _max,ℱ(ρ_AB) is set to be 0. This measure bears many nice properties. First, it can be efficiently computed via semidefinite programming (SDP) in many cases which will be discussed in the latter sections. Second, Eq. (<ref>) gives the closest free state τ_AB∈ to ρ_AB, w.r.t. the max-relative entropy. Third, _max,ℱ(·) is subadditive w.r.t the tensor product of states. In fact, _max,ℱ(ρ_AB) is closely related to the weight of resource W(ρ_AB) <cit.> and the free component Γ(ρ_AB) <cit.>, both of which have fruitful properties and applications <cit.>, as follows 2^-_max,ℱ(ρ_AB) = 1-W(ρ_AB) = Γ(ρ_AB). We note that each part of Eq. (<ref>) quantifies the largest weight where a free state can take in a convex decomposition of ρ_AB. When moving on to operational tasks that the free state can be ignored, what is left in a convex decomposition becomes our main concern. Optimization of the weight in the decomposition can be visualized as squeezing out all free parts of the given state. Thus, we further introduce the -squeezed state of ρ_AB as follows. For a bipartite quantum state ρ_AB and a free state set , If _max,ℱ(ρ_AB) is non-zero, the -squeezed state of ρ_AB is defined by ω_AB = ρ_AB-2^-_max,ℱ(ρ_AB)·τ_AB/1-2^-_max,ℱ(ρ_AB), where τ_AB is the closest free state to ρ_AB in terms of the max-relative entropy, i.e., the optimal solution in Eq. (<ref>). If _max,ℱ(ρ_AB)=0, the -squeezed state of ρ_AB is itself. In the following sections, we will illustrate the applications of the reverse max-relative entropy of resources as well as the squeezing idea on example tasks. One is giving upper bounds on the distillable entanglement of arbitrary quantum states. The other is to derive upper bounds on the quantum capacity of channels. § APPLICATIONS ON DISTILLABLE ENTANGLEMENT In this section, we investigate the information-theoretic application of the reverse max-relative entropy of resources in deriving efficiently computable upper bounds on the distillable entanglement. To showcase the advantage of our bounds, we compare the results with different continuity bounds and the Rains bound on the maximally entangled states with practical noises. §.§ Upper bound on the one-way distillable entanglement Recall that the one-way distillable entanglement has a regularized formula <cit.>: E_D,→(ρ_AB) = lim_n→∞1/nE^(1)_D,→(ρ_AB^⊗ n), where E^(1)_D,→(ρ_AB) := max_T I_(A'⟩ M B)_T(ρ_AB), and the maximization ranges over all quantum instruments T:A→ A'M on Alice’s system. The regularization in Eq. (<ref>) for E_D,→(ρ_AB) is intractable to compute in most cases. However, there are some categories of states whose E_D,→ can be reduced to single-letter formulae. Two important classes are called degradable states and anti-degradable states. Let ρ_AB be a bipartite state with purification |ϕ⟩_ABE. ρ_AB is called degradable if there exists a CPTP map ℳ_B→ E such that ℳ_B→ E(ρ_AB) = _B(ϕ_ABE), and is called anti-degradable if there exists a CPTP map ℳ_E→ B such that ℳ_E→ B(ρ_AE) = _E(ϕ_ABE). Equivalently, a state is anti-degradable if and only if it has a symmetric extension <cit.>, thus is also called a two-extendible state. For the degradable states, it is shown that <cit.> E^(1)_D,→(ρ_AB^⊗ n) = n E^(1)_D,→(ρ_AB) = n I_(A⟩ B)_ρ, ∀ n∈ℕ, resulting in E_D,→(ρ_AB) = I_(A⟩ B)_ρ. For the anti-degradable states, consisting of a compact and convex set, it always holds E^(1)_D,→(ρ_AB) = E_D,→(ρ_AB) = 0. Moreover, E_D,→(·) is convex on decomposing a state into degradable and anti-degradable parts <cit.>. To better exploit this convexity, we introduce the reverse max-relative entropy of unextendible entanglement to help identify the anti-degradable (two-extendible) part of a given bipartite state ρ_AB: _max,(ρ_AB) := min_τ∈D_max(τ_AB||ρ_AB), where is the set of all anti-degradable states. In this resource theory, the resource we consider is the extendible entanglement, and the free states are bipartite states that are possibly shareable between A and a third party E, where E is isomorphic to B. Notably, the extendibility of entanglement is a key property in entanglement theory with many existing applications <cit.>. Here, combined with the idea of entanglement of formation, _max,(ρ_AB) can be applied to derive an upper bound on the one-way distillable entanglement of an arbitrary state ρ_AB as shown in Theorem <ref>. For any bipartite state ρ_AB, it satisfies E_D,→(ρ_AB) ≤ E_ rev^u(ρ_AB):=[1-2^-_max,(ρ_AB)]· E_F(ω_AB), where ω_AB is the -squeezed state of ρ_AB, _max,(·) is the reverse max-relative entropy of unextendible entanglement, and E_F(·) is the entanglement of formation. Suppose the ADG-squeezed state of ρ_AB is ω_AB and the optimal solution in Eq. (<ref>) for ρ_AB is τ_AB. It follows ρ_AB = [1-2^-_max,(ρ_AB)]ω_AB + 2^-_max,(ρ_AB)τ_AB, where τ_AB is anti-degradable. Suppose ω_AB admit a pure state decomposition ω_AB= ∑_ip_i |ω_i⟩⟨ω_i|_AB. Since any pure state is degradable, we now can decompose ρ_AB into degradable parts and anti-degradable parts as ρ_AB = [1-2^-_max,(ρ_AB)]∑_i p_i |ω_i⟩⟨ω_i|_AB + 2^-_max,(ρ_AB)τ_AB, According to the essential convexity of E_D,→(·) on decomposing a state into degradable and anti-degradable parts proved by Leditzky et al. <cit.>, we have E_D,→(ρ_AB) ≤ [1-2^-_max,(ρ_AB)]·∑_i p_i E_D,→(ω_i) + 2^-_max,(ρ_AB) E_D,→(τ_AB) ≤ [1-2^-_max,(ρ_AB)]·∑_i p_i I_c(A⟩ B)_ω_i, where the second inequality is due to the fact in Eq. (<ref>) and Eq. (<ref>) of degradable states and anti-degradable states, respectively. After taking the minimization over all possible decomposition of ω_AB, we arrived at E_D,→(ρ_AB) ≤ [1-2^-_max,(ρ_AB)]·min∑_i p_i I_c(A⟩ B)_ω_i =[1-2^-_max,(ρ_AB)]· E_F(ω_AB). The bound E_ rev^u(ρ_AB) has a cartoon illustration as shown in Fig. <ref>. The main insight of it is to squeeze out as much of the free or useless part, the anti-degradable state here, as possible. We point out that squeezing all useless parts out does not necessarily give the tightest upper bound in terms of the one-way distillable entanglement, e.g., isotropic state <cit.>. Instead of squeezing out all the useless parts, there may be room for exploring more appropriate partitions when we try to decompose a specific quantum state. However, the approach we present in Theorem <ref> is an invaluable method for general states as shown in subsection <ref> and can be seen as a variant of the continuity bound in terms of the anti-degradability of the state. For any bipartite state ρ_AB, it satisfies E_D,→(ρ_AB) ≤E_ rev^u(ρ_AB) := [1-2^-_max,(ρ_AB)]·∑_iλ_i S(B)_ψ_i, where ω_AB=∑_iλ_i |ψ_i⟩⟨ψ_i| is the spectral decomposition of the ADG-squeezed state ω_AB of ρ_AB. Corollary <ref> is followed by the fact that E_F(ω_AB) has a trivial upper bound as E_F(ω_AB) ≤∑_iλ_i S(B)_ψ_i. We note that any other good upper bound on the entanglement of formation can also be applied to Theorem <ref>. In particular, the bound E_ rev^u(ρ_AB) is efficiently computable since _max,(ρ_AB) and 2^-_max,(ρ_AB) can be efficiently computed via an SDP. By Slater’s condition <cit.>, the following two optimization programs satisfy the strong duality, and both evaluate to 1-2^-_max,(ρ_AB). We denote P_BE as a permutation operator on the system B and E and remain the derivation of the dual program in Appendix <ref>. Primal Program min_ω_AB, τ_AB, τ_ABE [ω_AB], s.t. ρ_AB =ω_AB+τ_AB, ω_AB≥ 0, τ_AB≥ 0, τ_ABE≥ 0, _E[τ_ABE] = _B[τ_ABE] = τ_AB, Dual Program max_M_AB,N_AB,K_AE [M_ABρ_AB], s.t. M_AB≤ I_AB, M_AB+N_AB+K_AB≤ 0, N_AB⊗ I_E +P_BE(K_AB⊗ I_E)P_BE^†≥ 0. It is worth mentioning that this new bound is related to the bound E_ DA(·) proposed in <cit.> utilizing the convexity of E_D,→(·) on decomposing a state into degradable and anti-degradable parts. Remarkably, such decomposition is challenging to construct since the degradable state set is not necessarily convex. Thus it is difficult to compute E_ DA(·) in <cit.> in practice due to the hardness of tracking all possible decompositions. Whereas, E_ rev^u(ρ_AB) overcomes this difficulty and is efficiently computable. It outperforms the known upper bounds for many maximally entangled states under practical noises in a high-noise region, as shown in subsection <ref>. Furthermore, our method is flexible in selecting other decomposition strategies, i.e., the object functions in the SDP, other than trivially calculating the ratio at the sub-state decomposition. §.§ Continuity bounds of the one-way distillable entanglement Note that the insight of the bound above is considering the distance of a given state between the anti-degradable states set. With different distance measures, the authors in <cit.> derived continuity bounds on quantum capacity in terms of the (anti)degradability of the channel. Thus for self-contained, we introduce some distance measures between a state and the set and prove the continuity bounds for the state version as a comparison with E_ rev^u(ρ_AB). Let ρ_AB be a bipartite quantum state, the anti-degradable set distance is defined by d_ set(ρ_AB) := min_σ_AB∈ ADG1/2ρ_AB - σ_AB_1, where the minimization ranges over all anti-degradable states on system AB. Analogous to the anti-degradable set distance, one also has the anti-degradable map distance as follows. Let ρ_AB be a bipartite quantum state with purification ϕ_ABE, the anti-degradable map distance is defined by d_ map(ρ_AB):= min_𝒟:E→ B1/2ρ_AB - 𝒟(ρ_AE)_1 , where ρ_AE = _B(ϕ_ABE) and the minimization ranges over all CPTP maps 𝒟. Both parameters can be computed via SDP, ranging from [0,1], and are equal to 0 iff ρ_AB is anti-degradable. Similar to the idea in <cit.> for channels and the proof techniques in <cit.>, we utilize the continuity bound of the conditional entropy in Lemma <ref> proved by Winter <cit.> to derive two continuity upper bounds on the one-way distillable entanglement, concerning the distance measures above. The proofs can be found in Appendix <ref>. We denote h(p)=-plog p - (1-p)log(1-p) as the binary entropy and g(p):= (1+p)h(p/1+p) as the bosonic entropy. For any bipartite state ρ_AB with an anti-degradable set distance ε_ set, it satisfies E_D,→(ρ_AB) ≤ E_ SCB(ρ_AB) := 2ε_ setlog(|A|) + g(ε_ set). For any bipartite state ρ_AB with an anti-degradable map distance ε_ map, it satisfies E_D,→(ρ_AB) ≤ E_ MCB(ρ_AB) := 4ε_ maplog(|B|) + 2g(ε_ map). §.§ Examples of less-entangled states We now compare the performance of different continuity bounds and the Rains bound with E_ rev^u(·) by some concrete examples. Due to noise and other device imperfections, one usually obtains some less entangled states in practice rather than the maximally entangled ones. Such a disturbance can be characterized by some CPTP maps appearing in each party. Thus for the task of the distillation of the maximally entangled states under practical noises, we consider Alice and Bob are sharing pairs of maximally entangled states affected by bi-local noisy channels, i.e., ρ_A'B' = _A→ A'⊗_B→ B'(Φ_AB). Qubit system Suppose Alice's qubit is affected by the qubit amplitude damping channel with Kraus operators K_0 = |0⟩⟨0| + √(1-γ)|1⟩⟨1|, K_1 = √(γ)|0⟩⟨1|, and Bob's qubit is affected by the depolarizing channel (ρ) = (1-p)ρ + p I_2/2. Set γ = 0.1 and the noise parameter of depolarizing noise varies in the range p∈[0.15,0.3]. E_D,→(ρ_A'B') has upper bounds as functions of p shown in Fig. <ref> (a). Qutrit system For the system with a dimension d>2, we consider the multilevel versions of the amplitude damping channel (MAD) <cit.> as a local noise for Alice. The Kraus operators of a MAD channel in a d-dimensional system are defined by K̂_i j ≡√(γ_j i)|i⟩⟨j|, ∀ i, j s.t. 0 ≤ i ≤ d-1, K̂_0 ≡|0⟩⟨0|+ ∑_1 ≤ j ≤ d-1√(1-ξ_j)|j⟩⟨j|, with γ_j i real quantities describing the decay rate from the j-th to the i-th level that fulfill the conditions 0 ≤γ_j i≤ 1, ∀ i, j s.t. 0 ≤ i<j ≤ d-1, ξ_j ≡∑_0 ≤ i<jγ_j i≤ 1, ∀ j=1, ⋯, d-1. Suppose Alice qutrit is affected by a MAD channel with γ_10 = γ_20 = 0.1, γ_21 = 0. Bob's qutrit is affected by a qutrit depolarizing channel with the noise parameter p. Then E_D,→(ρ_A'B') has upper bounds as functions of p shown in Fig. <ref> (b). Qudit system For the qudit system, we consider Alice's qudit is affected by a MAD channel with d=4 defined in Eq. (<ref>) and Eq. (<ref>), where γ_10 = γ_20 = γ_30 = γ_21 = 0.1, γ_31 = γ_32 = 0. Let Bob's qudit be affected by a qudit depolarizing channel with noise parameter p, then E_D,→(ρ_A'B') has upper bounds as functions of p shown in Fig. <ref> (c). Noisy Bell state Specifically for a Bell state destroyed by bi-local depolarizing noises, i.e., ρ_A'B'=_A→ A'⊗_B→ B' (Φ_AB), we give an analytical upper bound on E_D,→(ρ_A'B') as shown in Proposition <ref>. Let Φ_AB be a Bell state and _A→ A', _B→ B' be depolarizing channels on system A and B, respectively. The state ρ_A'B'=_A→ A'⊗_B→ B' (Φ_AB) either is anti-degradable with E_D,→(ρ_A'B')=0 or satisfies 2^-_max,(ρ_A'B') =3(p_A + p_B - p_Ap_B) and E_D,→(ρ_A'B') ≤ 1 - 3(p_A + p_B - p_Ap_B), where p_A, p_B are noise parameters of _A→ A' and _B→ B', respectively. In the beginning, we note that if 1 - 3(p_A + p_B - p_Ap_B)≤ 0, the state ρ_A'B' is anti-degradable according to Lemma <ref>, which gives E_D,→(ρ_A'B')=0. Then if 1 - 3(p_A + p_B - p_Ap_B)>0, we will first prove 2^-_max,(ρ_A'B')≥ 3(p_A + p_B - p_Ap_B) use the primal SDP in (<ref>). We show that {ω̂_A'B',τ̂_A'B'E} is a feasible solution to the primal SDP, where ω̂_A'B' = 1-3(p_A+p_B-p_Ap_B)/2(|00⟩⟨00| + |11⟩⟨11|+ |00⟩⟨11| + |11⟩⟨00|), τ̂_A'B'E = ( [ 4ζ 0 0 0 0 2ζ 2ζ 0; 0 ζ ζ 0 0 0 0 2ζ; 0 ζ ζ 0 0 0 0 2ζ; 0 0 0 0 0 0 0 0; 0 0 0 0 0 0 0 0; 2ζ 0 0 0 0 ζ ζ 0; 2ζ 0 0 0 0 ζ ζ 0; 0 2ζ 2ζ 0 0 0 0 4ζ; ]), with ζ = p_A + p_B - p_A p_B/4. When the Bell state shared by Alice and Bob is affected by bi-local depolarizing channels _A→ A' and _B→ B' with noise parameters p_A and p_B, respectively, the resulting state is ρ_A'B' = _A→ A'⊗_B→ B' (Φ_AB) = 2-(p_A+p_B)+p_Ap_B/4(|00⟩⟨00| + |11⟩⟨11|)+ p_A+p_B-p_Ap_B/4(|01⟩⟨01| + |10⟩⟨10|) + 1-(p_A+p_B)+p_Ap_B/2(|00⟩⟨11|+|11⟩⟨00|). Then the unnormalized state τ̂_A'B' = ρ_A'B' - ω̂_A'B' is τ̂_A'B' = 5(p_A+p_B-p_Ap_B)/4(|00⟩⟨00| + |11⟩⟨11|)+ p_A+p_B-p_Ap_B/4(|01⟩⟨01| + |10⟩⟨10|) + (p_A+p_B-p_Ap_B)(|00⟩⟨11|+|11⟩⟨00|). It is straightforward to check that τ̂_A'B'E≥ 0 and _E[τ̂_A'B'E] = _B[τ̂_A'B'E] = τ̂_A'B'. Thus, we have proven that {ω̂_A'B',τ̂_A'B'E} is a feasible solution to the primal SDP, implying that 1 - 2^-_max,(ρ_A'B')≤[ω̂_A'B'] = 1 - 3(p_A + p_B - p_Ap_B) and 2^-_max,(ρ_A'B')≥ 3(p_A + p_B - p_A p_B). Second, we will use the dual SDP in (<ref>) to prove 2^-_max,(ρ_A'B')≤ 3(p_A + p_B - p_Ap_B). We show that { M_AB, N_AB, K_AB} is a feasible solution to the dual problem, where M_AB = ( [ -1 0 0 2; 0 -3 0 0; 0 0 -3 0; 2 0 0 -1; ]), N_AB = K_AB = -1/2M_AB. To check the constraints, we have M_AB≤ I_AB, M_AB + N_AB + K_AB = 0 and N_AB⊗ I_E + P_BE(K_AB⊗ I_E)P_BE^†≥ 0. By direct calculation, we have [M_ABρ_A'B'] = 1 - 3(p_A + p_B-p_A p_B). Thus, we have proven {M_AB, N_AB, K_AB} is a feasible solution, which implies that 1 - 2^-_max,(ρ_A'B')≥[M_ABρ_A'B'] = 1 - 3(p_A + p_B-p_A p_B), and 2^-_max,(ρ_A'B')≤ 3(p_A + p_B - p_A p_B). Combining the dual part with the primal part, we arrive at 2^-_max,(ρ_A'B') = 3(p_A + p_B - p_Ap_B) Since the normalized state ω_A'B' of ω̂_A'B' is a maximally entangled state with E_D,→(ω_A'B')=1, we have E_D,→(ρ_A'B') ≤ 1 - 3(p_A + p_B-p_A p_B). As shown in Fig <ref>, we plot four different upper bounds and the Hashing bound in different dimensional systems in a high-noise region. We can see that E_ rev^u(ρ_AB) is lower than the other two continuity bounds and the Rains bound in the orange dotted line, acting the tightest among all upper bounds. We demonstrate the advantage of our analytic bound in Proposition <ref> by comparing it with the continuity bounds and the Rains bound in Fig. <ref> (d), where the depolarizing noise parameter p_B is set to 0.1 and the noise parameter p_A is varied. §.§ Extending the method to the two-way distillable entanglement Similar to the reverse max-relative entropy of unextendible entanglement, for a given bipartite state ρ_AB, we introduce the reverse max-relative entropy of NPT entanglement as _max,(ρ_AB) := min_τ∈D_max(τ_AB||ρ_AB), where the minimization ranges over all states. Based on the convexity of E_D,↔(·) on decomposing a state into the maximally correlated (MC) states and the PPT states <cit.> (see Appendix <ref> for more details), the reverse max-relative entropy of NPT entanglement can be applied to derive an upper bound on the two-way distillable entanglement as Theorem <ref>. For any bipartite state ρ_AB, it satisfies E_D,↔(ρ_AB) ≤ E_ rev^npt(ρ_AB):=[1-2^-_max,(ρ_AB)]· E_F(ω_AB), where ω_AB is the PPT-squeezed state of ρ_AB and _max,(·) is the reverse max-relative entropy of NPT entanglement. It also follows an efficiently computable relaxation as E_ rev^npt(ρ_AB)= [1-2^-_max,(ρ_AB)] ∑_iλ_i S(B)_ψ_i, where ω_AB=∑_iλ_i |ψ_i⟩⟨ψ_i| is the spectral decomposition of the PPT-squeezed state ω_AB of ρ_AB. In fact, E_ rev^npt(·) can be interpreted as an easily computable version of the bound E_ MP(·) in <cit.>, utilizing the convexity of E_D,↔(ρ_AB) on the convex decomposition of ρ_AB into MC states and PPT states. Since the set of all MC states is not convex, tracking all possible decompositions to compute E_ MP(·) is generally hard. However, E_ rev^npt(·) is efficiently computable by SDP and tightens the upper bound of the example states illustrated in <cit.>, whose details can be found in Appendix <ref>. We note that R(ρ_AB) ≤ E_ MP(ρ_AB)≤E_ rev^npt(ρ_AB) where R(·) is the Rains bound for the two-way distillable entanglement. Nevertheless, E_ rev^npt(·) connects the reverse max-relative entropy of NPT entanglement with the entanglement of formation, and we believe such connection would shed light on the study of other quantum resource theories as well. § APPLICATIONS ON QUANTUM CHANNEL CAPACITY For a general quantum channel _A→ B, its quantum capacity has a regularized formula proved by Lloyd, Shor, and Devetak <cit.>: Q() = lim_n→∞1/nQ^(1)(^ n), where Q^(1)():= max_|ϕ⟩_A'A I_c(A'⟩ B)_()(ϕ) is the channel coherent information. Similar to the one-way distillable entanglement of a state, the regularization in Eq. (<ref>) makes the quantum capacity of a channel intractable to compute generally. Substantial efforts have been made to establish upper bounds. One routine is inspired by the Rains bound in entanglement theory <cit.>. Tomamichel et al. introduced Rains information <cit.> for a quantum channel as an upper bound on the quantum capacity. Then some efficiently computable relaxations or estimations are given in <cit.>. Another routine is to consider the (anti)degradability of the quantum channels and to construct flag extensions, which gives the currently tightest upper bound for quantum channels with symmetry or known structures <cit.>. A channel _A→ B is called degradable if there exits a CPTP map _B→ E such that ^c = ∘, and is called anti-degradable if there exits a CPTP map _E→ B such that = ∘^c. It is known that is (anti)degradable if and only if its Choi state J_ is (anti)degradable. The quantum capacity of an anti-degradable channel is zero and the coherent information of a degradable channel is additive which leads to Q() = Q^(1)(). Also, the quantum capacity admits convexity on decomposing a channel into a convex combination of degradable channels and anti-degradable channels <cit.> which can give a simple upper bound on Q(). Concerning the (anti)degradability of a channel, the authors in <cit.> called a channel ε-degradable channel if there is a CPTP map _B→ E such that || ^c - ∘||_♢≤ 2ε. A channel is called ε-anti-degradable channel if there is a CPTP map _E→ B such that || - ∘^c||_♢≤ 2ε. Based on these, one has continuity bounds of the quantum capacity as follows. Given a quantum channel _A → B, if it is -degradable, then it satisfies Q() ≤ Q^(1)() + log(d_E - 1) + h() + 2log d_E + g(). If _A→ B is -anti-degradable, it satisfies Q() ≤log(|B|-1) + 2log|B| + h() + g(). With a similar spirit of the reverse max-relative entropy of unextendible entanglement in Eq. (<ref>), for a quantum channel _A→ B, we define the reverse max-relative entropy of anti-degradability of the channel as _max,(_A→ B) := min_'_A→ B∈_D_max('_A→ B||_A→ B), where _ is the set of all anti-degradable channels and the max-relative entropy of '_A→ B with respect to _A→ B is defined by D_max('_A→ B||_A→ B):= inf{λ∈ℝ: J_AB^'≤ 2^λ J_AB^}. If there is no such a channel '_A→ B∈_ that satisfies J_AB^'≤ 2^λ J_AB^, _max,(_A→ B) is set to be 0. Similar to the state case, _max,(_A→ B) has a geometric implication analogous to the distance between _A→ B to the set of all anti-degradable channels. We can introduce the -squeezed channel of _A→ B as follows. For a quantum channel _A→ B and the anti-degradable channel set _, if _max,() is non-zero, the -squeezed channel of _A→ B is defined by _A→ B = _A→ B - 2^-_max,()·^'_A→ B/1-2^-_max,() where ^'_A→ B is the closest anti-degradable channel to _A→ B in terms of the max-relative entropy, i.e., the optimal solution in Eq. (<ref>). If _max,() is zero, the -squeezed channel of _A→ B is itself. Notably, _max,(), or 2^-_max,() equivalently, can be efficiently computed via SDP. The following two optimization programs satisfy the strong duality, and both evaluate to 1-2^-_max,(). We remain the derivation of the dual program in Appendix <ref>. Primal Program min_Γ_AB^, Γ_AB^', γ_ABE [Γ_AB^], s.t. J_AB^ = Γ_AB^ + Γ_AB^', Γ_AB^≥ 0, Γ_AB^'≥ 0, γ_ABE≥ 0, _B[Γ_AB^'] = [Γ_AB^']/d_A · I_A, _E[γ_ABE] = _B[γ_ABE] = Γ_AB^', Dual Program max_M_AB,N_AB,K_AB,R_A [M_AB J_AB^], s.t. M_AB≤ (1- R_A/d_A)I_AB+R_A⊗ I_B, M_AB+N_AB+K_AB≤ 0, N_AB⊗ I_E +P_BE(K_AB⊗ I_E)P_BE^†≥ 0. The conceptual idea we used here is similar to that for the state case in Eq. (<ref>) and Definition <ref>, which is to squeeze or isolate out as much part of anti-degradable channel as possible via a convex decomposition of the original channel. The insight here is that one can ignore the contribution from the anti-degradable part for the quantum capacity, and the quantum capacity admits convexity on the decomposition into degradable and anti-degradable parts. In this way, the following Theorem <ref> gives an upper bound Q_ sqz() for the quantum capacity of . Given a quantum channel _A → B, if it has an ADG-squeezed channel _A→ B, we denote _A→ BB' as an extended channel of _A→ B such that _B'[_A→ BB'(ρ_A)] = _A→ B(ρ_A), ∀ρ_A∈(_A). Then it satisfies Q() ≤ Q_ sqz() := [1-2^-_max,()]·min{ Q^(1)() | _A→ BB'}, where the minimization is over all possible extended channels of S_A→ B. If there is no such a degradable _A→ BB' exists, the value of this bound is set to be infinity. By the definition of the -squeezed channel of _A→ B, we have _A→ B = [1-2^-_max,()] ·_A→ B + 2^-_max,()·'_A→ B where '_A→ B is anti-degradable. We write an extended channel of '_A→ B as '_A→ BB'(ρ_A) = _A→ B(ρ_A) |0⟩⟨0|_B', which is obviously anti-degradable. Then we can construct a quantum channel _A→ BB' as _A→ BB' = [1-2^-_max,()] ·_A→ BB' + 2^-_max,()·'_A→ BB', such that _B'[_A→ BB'(ρ_A)] = _A→ B(ρ_A) for any state ρ_A and _A→ BB' is degradable. This means after discarding the partial environment B', the receiver can obtain the original quantum information sent through _A→ B. In this case, can certainly convey more quantum information than the , i.e., Q()≤ Q(). Note that the quantum capacity admits convexity on the decomposition into degradable parts and anti-degradable parts <cit.>. We conclude that Q()≤ Q() ≤ [1-2^-_max,()] · Q() + 2^-_max,()· Q(') = [1-2^-_max,()] · Q^(1)(), where the equality is followed by the quantum capacity is additive on degradable channels and is zero for anti-degradable channels. Considering the freedom of the choice of _A→ BB', we obtain Q_ sqz() := [1-2^-_max,()] ·min{ Q^(1)() | _A→ BB'} as an upper bound on Q(). Theorem <ref> can be seen as a channel version of Theorem <ref>. However, in order to utilize the convexity of the quantum capacity after the squeezing process, it is challenging to decompose the ADG-squeezed channel into the sum of degradable ones. An alternative approach here is to use the idea of the extension channel. For the qubit channels specifically, this bound is efficiently computable and effective, as shown in subsection <ref>. §.§ Quantum capacity of qubit channels For a quantum channel with dimension two in both input and output systems, we prove that the ADG-squeezed channel is always degradable. Thus, we give an efficiently computable upper bound on the quantum capacity using the idea of the reverse max-relative entropy of anti-degradability. For any qubit channel _A→ B, it is either anti-degradable or satisfies Q() ≤ [1-2^-_max,()] ·max_p ∈[0,1] I_c(p|0⟩⟨0|+(1-p)|1⟩⟨1|, _A→ B), where I_c(ρ, 𝒩) ≡ H(𝒩(ρ))-H(𝒩^c(ρ)) and _A→ B is the ADG-squeezed channel of _A→ B. By the definition of the -squeezed channel of _A→ B, we have _A→ B = [1-2^-_max,()] ·_A→ B + 2^-_max,()·'_A→ B where '_A→ B is anti-degradable and _A→ B is not anti-degradable. If _A→ B is also not degradable, we can further decompose _A→ B into = c _0 +(1-c)S_1 such that _0 is degradable and _1 is anti-degradable since the extreme points of the set of all qubit channels have been shown to be degradable or anti-degradable channels <cit.>. This conflicts with the definition of 2^-_max,(), which implies _A→ B is degradable. Thus, Q() ≤ [1-2^-_max,()] · Q() + 2^-_max,()· Q(') = [1-2^-_max,()] · Q^(1)() =[1-2^-_max,()] ·max_p ∈[0,1] I_c(p|0⟩⟨0|+(1-p)|1⟩⟨1|, _A→ B). Note that the last equality is because _A→ B is degradable, and diagonal input states outperform non-diagonal states during the optimization of the channel coherent information <cit.>. Mixed unitary channels To compare the performance of our method with some best-known computable bounds, e.g., the continuity bound in Theorem <ref> and the bound R̂_α <cit.> generalized from the max-Rain information <cit.>, we consider the mixed unitary channel _A→ B(·) as follows: (ρ) = ∑_i=0^k p_i U_i ρ U_i^†, where ∑_i=0^k p_i= 1 and U_i are unitary operators on a qubit system. In specific, we choose some fixed set of parameters and sample 1000 channels with randomly generated unitaries according to the Haar measure. We compute the distance between Q_ sqz and other bounds, then have statistics on the distribution of these channels according to the distance value. The distribution results are depicted in Fig. <ref> where the purple region corresponds to the cases Q_ sqz is tighter, and the pink region corresponds to the cases Q_ sqz is looser. We can see that in Fig. <ref>(a) and Fig. <ref>(c), Q_ sqz always outperforms the continuity bound of anti-degradability and in Fig. <ref>(b) and Fig. <ref>(d), our bound is tighter than R̂_α(10) for many cases. Pauli channels A representative qubit channel is the Pauli channel describing bit-flip errors and phase-flip errors with certain probabilities in qubits. A qubit Pauli channel Λ(·) is defined as: Λ(ρ) = p_0 ρ + p_1 Xρ X + p_2 Yρ Y +p_3 Zρ Z, where X,Y,Z are the Pauli operators and ∑_i=0^3 p_i = 1 are probability parameters. Note that for the quantum capacity, we only need to consider the cases where p_0 dominates. Since if, for example, bit flip error X happens with probability larger than p_0, one can first apply a X flip, mapping that channel back into the case where p_0>p_1. After utilizing our method on Pauli channels, the ADG-squeezed parameter is characterized in Proposition <ref>, whose proof can be found in Appendix <ref>. Thus, combined with Proposition <ref>, we can recover the no-cloning bound <cit.> on the quantum capacity of qubit Pauli channels. For a qubit Pauli channel Λ(·) with p_0≥ p_i>0 (i=1,2,3), it is either anti-degradable or satisfies 2^-_max,(Λ) = (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2 with an ADG-squeezed channel as the identity channel. For a qubit Pauli channel Λ(·), its quantum capacity is either vanishing or satisfies Q(Λ) ≤ 1-(√(p_1)+√(p_2))^2 -(√(p_2)+√(p_3))^2-(√(p_1)+√(p_3))^2 One recent work <cit.> studies the capacities of a subclass of Pauli channels called the covariant Pauli channel, where the parameters are set p_1 = p_2 with p_0+2p_1+p_3=1, i.e., Λ_ cov(ρ) = p_0 ρ + p_1 (Xρ X + Yρ Y)+ p_3 Zρ Z. Applying Theorem <ref> on the covariant Pauli channels, we can bound their quantum capacity as follows. For a covariant Pauli channel Λ_ cov(·), it is either anti-degradable with a zero quantum capacity or satisfies Q(Λ_ cov)≤ Q_ sqz(Λ_ cov), where Q_ sqz (Λ_ cov) = 3p_0 + p_3 -√(8(p_3-p_0p_3-p_3^2))-2. In Fig. <ref>, we compare our bound with the upper bounds given in <cit.> and the continuity bound of anti-degradability in Theorem <ref>. It can be seen that our bound in the orange line, coinciding with the no-cloning bound, outperforms previous bounds, and thus can better characterize the quantum capacity of Λ_ cov(·) when it is close to being anti-degradable. § CONCLUDING REMARKS We have introduced a resource measure called reverse divergence of resources for quantum states. Specifically, by choosing target free state sets and max-relative entropy, we have employed the reverse max-relative entropy of entanglement which is also related to the "weight of resource" in general resource theory. From a conceptual and technical side, these concepts help us to quantify how much useless entanglement we can squeeze out from a state or a channel, which is meaningful in the distillable entanglement and the quantum capacity, respectively. As these quantities can be efficiently determined by semidefinite programming, they are useful in deriving efficiently computable upper bounds on the distillable entanglement and the quantum capacity. To better explore entanglement distillation, we have derived different continuity bounds on the one-way distillable entanglement in terms of the anti-degradability of the state. In particular, our bound derived from the reverse max-relative entropy of unextendible entanglement outperforms the continuity bounds and the Rains bound in estimating the one-way distillable entanglement for the maximally entangled states under certain noises of interest. We also introduced the reverse max-relative entropy of NPT entanglement and established connections to prior results on the two-way distillable entanglement. Furthermore, we have established an upper bound on the quantum capacity based on the reverse max-relative entropy of anti-degradability of quantum channels. Our method delivers improved results for random mixed unitary qubit channels, making it an improved benchmark for the quantum capacities of these noisy channels. Also, the analytical bound obtained from our method also recovers the no-cloning bound on Pauli channels <cit.>. These results open a novel way to connect valuable quantum resource measures with quantum communication tasks. Except for the existing applications of the reverse max-relative entropy of resources <cit.>, we expect the reverse divergence of resources will find more applications in quantum resource theories in both asymptotic and non-asymptotic regimes <cit.>. One may further consider choosing other divergences, e.g., the sandwiched Rényi relative entropy to solve problems in other tasks possibly. In addition, the reverse divergence of resources may potentially be helpful to the study of other quantum resource theories, including asymmetry <cit.>, thermodynamics <cit.>, coherence <cit.>, nonlocality <cit.>, non-stabilizerness <cit.>, and non-Gaussianity <cit.>. It is also an interesting problem to explore the idea of squeezing useless resources to estimate resource purification rates in quantum resource theories. § ACKNOWLEDGEMENTS. Part of this work was done when C. Z. and C. Z. were research interns at Baidu Research. We would like to thank Bartosz Regula and Ludovico Lami for helpful comments. equationsection § PROPERTIES OF THE REVERSE MAX-RELATIVE ENTROPY OF RESOURCES For bipartite states ρ_0 and ρ_1, it satisfies _max,ℱ(ρ_0) + _max,ℱ(ρ_1) ≥_max,ℱ(ρ_0⊗ρ_1). Since the max-relative entropy is additive, w.r.t, tensor products, we have min_τ' ∈ℱD_max(τ' || ρ_0ρ_1) ≤min_τ_0,τ_1 ∈ℱ D_max(τ_0τ_1 || ρ_0 ρ_1) = min_τ_0 ∈ℱ D_max(τ_0 || ρ_0) + min_τ_1 ∈ℱ D_max(τ_1 || ρ_1), which completes the proof. Let Φ be a one-way LOCC, then for a bipartite state ρ, it satisfies _max,ℱ[Φ(ρ)] ≤_max,ℱ(ρ). This can be directly obtained by the monotonicity of the max-relative entropy with respect to the action of channels. § DUAL SDP FOR THE SQUEEZED PARAMETER OF STATES AND CHANNELS The primal SDP for calculating 2^-_max,(ρ_AB) of the state ρ_AB can be written as: 1-2^-_max,(ρ_AB) = min_ω_AB, τ_AB, τ_ABE [ω_AB], s.t. ρ_AB =ω_AB+τ_AB, ω_AB≥ 0, τ_AB≥ 0, τ_ABE≥ 0, _E[τ_ABE] = _B[τ_ABE] = τ_AB, where Eq. (<ref>) corresponds to the anti-degradable condition of τ_AB. The Lagrange function of the primal problem is L(ω_AB,τ_AB,τ_ABE,M, N, K) = [ω_AB] + ⟨ M, ρ_AB-ω_AB-τ_AB⟩ + ⟨ N, _E[τ_ABE]-τ_AB⟩ + ⟨ K, _B[τ_ABE]-τ_AB⟩ = ⟨ M, ρ_AB⟩ + ⟨ I-M, ω_AB⟩ + ⟨ -M-N-K, τ_AB⟩ + ⟨ N⊗ I_E + P_BE(K_AB⊗ I_E)P_BE^†, τ_ABE⟩, where M_AB, N_AB, K_AB are Lagrange multipliers and P_BE is the permutation operator between B and E. The corresponding Lagrange dual function is g(M,N,K) = inf_ω_AB,τ_AB, τ_ABE≥ 0 L(ω_AB,τ_AB,τ_ABE,M, N, K). Since ω_AB≥ 0, τ_AB≥ 0, τ_ABE≥ 0, it must hold that I-M_AB≥ 0, -M-N-K≥ 0, N⊗ I + P_BE(K ⊗ I)P_BE^†≥ 0. Thus the dual SDP is max_M_AB,N_AB,K_AE [M_ABρ_AB], s.t. M_AB≤ I_AB, M_AB+N_AB+K_AB≤ 0, N_AB⊗ I_E +P_BE(K_AB⊗ I_E)P_BE^†≥ 0. The primal SDP for calculating 2^-_max,() of the channel _A→ B is: 1-2^-_max,() = min_Γ_AB^, Γ_AB^', γ_ABE [Γ_AB^], s.t. J_AB^ = Γ_AB^ + Γ_AB^', Γ_AB^≥ 0, Γ_AB^'≥ 0, γ_ABE≥ 0, _B[Γ_AB^'] = [Γ_AB^']/d_A · I_A, _E[γ_ABE] = _B[γ_ABE] = Γ_AB^', where Eq. (<ref>) corresponds to the anti-degradable condition of the unnormalized Choi state Γ_AB^'. The Lagrange function of the primal problem is L(Γ_AB^, Γ_AB^', γ_ABE,M, N, K, R) = [Γ_AB^] + ⟨ M, J_AB^-Γ_AB^-Γ_AB^'⟩ + ⟨ N, _E[γ_ABE]-Γ_AB^'⟩ + ⟨ K, _B[γ_ABE]-Γ_AB^'⟩ + ⟨ R, _B[Γ_AB^]-[Γ_AB^]/d_A · I_A⟩ = ⟨ M, J_AB^⟩ + ⟨ (1- R/d_A)I-M+R⊗ I_B, Γ_AB^⟩ + ⟨ -M-N-K, Γ_AB^'⟩ + ⟨ N⊗ I_E + P_BE(K_AE⊗ I_B)P_BE^†, γ_ABE⟩, where M_AB, N_AB, K_AB are Lagrange multipliers and P_BE is the swap operator between the system B and E. The corresponding Lagrange dual function is g(M,N,K) = inf_Γ_AB^,Γ_AB^', γ_ABE≥ 0 L(Γ_AB^,Γ_AB^',γ_ABE,M, N, K). Since Γ_AB^≥ 0, Γ_AB^'≥ 0, γ_ABE≥ 0, it must hold that (1- R_A/d_A)I_AB-M_AB+R_A⊗ I_B ≥ 0 -M_AB-N_AB-K_AB ≥ 0 N_AB⊗ I_E + P_BE(K_AB⊗ I_E)P_BE^† ≥ 0. Thus the dual SDP is max_M_AB,N_AB,K_AB,R_A [M_AB J_AB^], s.t. M_AB≤ (1- R_A/d_A)I_AB+R_A⊗ I_B, M_AB+N_AB+K_AB≤ 0, N_AB⊗ I_E +P_BE(K_AB⊗ I_E)P_BE^†≥ 0. § TWO-WAY DISTILLABLE ENTANGLEMENT We first start with the definition of the maximally correlated (MC) state. A bipartite state ρ_A B on ℂ^d ×ℂ^d is said to be maximally correlated (MC), if there exist bases {|i⟩_A}_i=0^d-1 and {|i⟩_B}_i=0^d-1 such that ρ_A B=∑_i, j=0^d-1α_i j|i⟩⟨j|_A ⊗|i⟩⟨j|_B, where (α_i j) is a positive semidefinite matrix with trace 1 . We note that every pure state is an MC state. Then, by the following lemma, one can easily arrive at the upper bound in Theorem <ref>. The two-way distillable entanglement is convex on convex combinations of MC and PPT states. We recall the example state given in <cit.>. Consider a d dimensional Hilbert space, the generalized Pauli operators X and Z are defined by their action on a computational basis {|k⟩}_k=0^d-1 as: X|k⟩:=|k+1( d)⟩ Z|k⟩:=ω^k|k⟩, where ω:=exp (2 π i / d) is a d-th root of unity. The generalized Pauli operators satisfy X Z=ω Z X. Then the generalized Bell basis {|Φ_n, m⟩}_n, m=0, …, d-1 are defined as |Φ_n, m⟩:=(I_d ⊗ X^m Z^n)|Φ⟩, where |Φ⟩:=1/√(d)∑_i=0^d-1|i i⟩. Now set d=3 and denote B={1,6,8}. After numbering |Φ_n,m⟩ = |Φ_3n+m+1⟩, we define the state ω_AB = ∑_i,j=1^dα_i,j|Φ_B[i]⟩⟨Φ_B[j]|, where α = |ψ⟩⟨ψ|, |ψ⟩ = 1/3(|0⟩ + |1⟩ + |2⟩). Then consider the states of the form ρ_AB = (1-p) ω_AB + pτ_AB, where p∈[0,1], the state τ_AB is the following PPT entangled state with a=1/2: τ_A B=1/8 a+1( [ a 0 0 0 a 0 0 0 a; 0 a 0 0 0 0 0 0 0; 0 0 a 0 0 0 0 0 0; 0 0 0 a 0 0 0 0 0; a 0 0 0 a 0 0 0 a; 0 0 0 0 0 a 0 0 0; 0 0 0 0 0 0 1+a/2 0 √(1-a^2)/2; 0 0 0 0 0 0 0 a 0; a 0 0 0 a 0 √(1-a^2)/2 0 1+a/2 ]). We plot the different upper bounds on the two-way distillable entanglement E_D,↔(ρ_AB) in Fig. <ref>. It shows that our bound E_ rev^npt(ρ_AB) is tighter than E_ W and E_ MP as a function of p. § PROOF OF PROPOSITION <REF> AND PROPOSITION <REF> For any states ρ_AB and σ_AB such that 1/2ρ_AB-σ_AB_1 ≤ε≤ 1, it satisfies |S(A|B)_ρ - S(A|B)_σ| ≤ 2εlog(|A|) + (1+ε)h(ε/1+ε). For any bipartite state ρ_AB with an anti-degradable set distance ε_ set, it satisfies E_D,→(ρ_AB) ≤ 2ε_ setlog(|A|) + (1+ε_ set)h(ε_ set/1+ε_ set). Since ρ_AB has a anti-degradable set distance ε_ set, we denote σ_AB the anti-degradable state with 1/2ρ_AB - σ_AB_1 ≤ε_ set. Let T:A^n → A'M be an instrument with isometry U_n:A^n → A'MN and denote Δ = 2ε_ setlog(|A|) + (1+ε_ set)h(ε_ set/1+ε_ set). Then we have E^(1)_D,→(ρ_AB^⊗ n) = max_U_n I_c(A'⟩ B^n M)_U_n ρ^⊗ n U_n^† = max_U_n - S(A' | B^n M)_U_n ρ^⊗ n U_n^† ≤max_U_n - S(A' | B^n M)_U_n (σρ^⊗ n-1) U_n^† + Δ ≤max_U_n - S(A' | B^n M)_U_n σ^⊗ n U_n^† + nΔ = nΔ, where Eq. (<ref>) follows by the fact that ρ_0⊗ρ_1 - σ_0⊗σ_1_1 ≤ρ_0 - σ_0_1 + ρ_1 - σ_1_1, and Lemma <ref>. The inequality in Eq. (<ref>) follows by applying the same argument n times considering σ^ iρ^ n-i for i=1,...,n. After dividing Eq. (<ref>) by n and taking the limit n→∞, we arrive at E_D,→(ρ_AB) = lim_n→∞1/nE^(1)_D,→(ρ_AB^⊗ n) ≤Δ. For any bipartite state ρ_AB with an anti-degradable map distance ε_ map, it satisfies E_D,→(ρ_AB) ≤ 4ε_ maplog(|B|) + 2(1+ε_ map)h(ε_ map/1+ε_ map). Let ϕ_ABE be a purification of ρ_AB, :E→ B be the CPTP map such that 1/2ρ_AB - (ρ_AE)_1 ≤ε_ map with an isometry W:E→ B'G. Let T:A^n → A'M be an instrument with isometry U_n:A^n → A'MN and denote Δ = 2ε_ maplog(|B|) + (1+ε_ map)h(ε_ map/1+ε_ map). For t=1,2,...,n, we can define pure states ψ_A^n B^n B'_1 G_1...B'_t G_t E_t+1...E_n^t = (W_1⊗⋯⊗ W_t) ϕ_ABE^n θ_A' M N B^n B'_1 G_1...B'_t G_t E_t+1...E_n^t = U_n ψ^t, ω_A' M N B^n E^n = U_n ϕ_ABE^n We further define ρ̂_AB' = (ρ_AE) which shares the same purification with ρ_AB listed above, thus an anti-degradable state. Then for t=n we have θ^n = U_n (W_1⊗⋯⊗ W_n) U_n^†ω, it yields I_c(A'⟩ B^n M)_ω = I_c(A'⟩ B^n M)_θ^n_A' M N B^n B'^n G^n = S(B^n M)_θ - S(A'B^n M)_θ = S(A' N B'^n G^n)_θ - S(A'B^n M)_θ = S(A' N B'^n G^n)_θ - S(A'B'^n M)_θ + S(A' B'^n M)_θ - S(A' B^n M)_θ = S(G^n|A'B'^n M)_θ + S(A' B'^n M)_θ - S(A' B^n M)_θ where we abbreviate θ = θ^n_A' M N B^n B'^n G^n. Applying the same technique in the proof of Theorem 2.12 in <cit.>, we can bound S(A' B'^n M)_θ - S(A' B^n M)_θ≤ nΔ. Consequently, it follows that I_c(A'⟩ B^n M)_ω ≤ S(G^n|A'B'^n M)_θ + nΔ ≤ S(G^n|B'^n)_θ + nΔ = S(G^n B'^n)_θ - S(B'^n)_θ + nΔ ≤ S(G^n B'^n)_θ - S(B^n)_θ + 2nΔ ≤ S(G^n B'^n)_θ - S(A'M N G^n B'^n)_θ + 2nΔ ≤ I(A'M N ⟩ G^n B'^n)_θ + 2nΔ = I(A^n ⟩ B'^n)_ρ̂_AB'^ n + 2nΔ = n [I(A ⟩ B)_ρ̂_AB'+ 2Δ] = 2nΔ, where the last equality is due to the anti-degradability of ρ̂_AB'. After dividing Eq. (<ref>) by n and taking the limit n→∞, we arrive at E_D,→(ρ_AB) = lim_n→∞1/nE^(1)_D,→(ρ_AB^⊗ n) ≤ 2Δ. § PROOF OF PROPOSITION <REF> A two qubit state ρ_A B is anti-degradable if and only if, Tr(ρ_B^2) ≥Tr(ρ_A B^2)-4 √(det(ρ_A B)). For a qubit Pauli channel Λ(·) with p_0≥ p_i>0 (i=1,2,3), it is either anti-degradable or satisfies 2^-_max,(Λ) = (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2 with an ADG-squeezed channel as the identity channel. We first will prove 2^-_max,(Λ)≥ (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2, by using the SDP in Eq. (<ref>). We show that Γ_AB^ = α/2( [ 1 0 0 1; 0 0 0 0; 0 0 0 0; 1 0 0 1; ]), is a feasible solution where α = 1-[(√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2]. Note that the Choi state of the Pauli channel is J_AB^Λ = 1/2( [ p_0 + p_3 0 0 p_0-p_3; 0 p_1+p_2 p_1-p_2 0; 0 p_1-p_2 p_1+p_2 0; p_0 - p_3 0 0 p_0+p_3; ]), and the unnormalized state Γ_AB^' = J_AB^Λ - Γ_AB^ is Γ_AB^' = 1/2( [ p_0 + p_3 - α 0 0 p_0-p_3- α; 0 p_1+p_2 p_1-p_2 0; 0 p_1-p_2 p_1+p_2 0; p_0 - p_3 - α 0 0 p_0+p_3- α; ]). Recalling that p_0 + p_1 + p_2 + p_3 = 1, it is then straightforward to check that _B[Γ_AB^'] = [Γ_AB^']/d_A · I_A. The constraint in Eq. (<ref>) corresponds to the anti-degradable condition of Γ_AB^'. By direct calculation, we have [(Γ_B^')^2] = 1/2(1-α)^2, (Γ_AB^') = p_1 p_2 p_3(p_0-α), [(Γ_AB^')^2] = α^2 - 2α p_0 + p_0^2 + p_1^2 + p_2^2 + p_3^2. Then Eq. (<ref>) holds and Γ_AB^' is anti-degradable by Lemme <ref>, which satisfy the constraint in Eq. (<ref>). Thus, we have proven that Γ_AB^ is a feasible solution to the primal SDP, implying that 1 - 2^-_max,(Λ)≤[Γ_AB^'] = 1-[(√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2]. It yields 2^-_max,(Λ)≥ (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2. Second, we will use the dual SDP in Eq. (<ref>) to prove 2^-_max,(Λ)≤ (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2. We show that {M_AB, N_AB, K_AB, R_A} is a feasible solution to the dual problem, where M_AB = ( [ η 0 0 -η + 1; 0 ξ ζ 0; 0 ζ ξ 0; -η + 1 0 0 η; ]), N_AB = K_AB = -1/2M_AB, R_A = 0, and η = -√(p_1) + √(p_2)/2√(p_3), ξ= -√(p_1)+√(p_3)/2√(p_2)-√(p_2)+√(p_3)/2√(p_1)-1, ζ=√(p_1)+√(p_3)/2√(p_2)-√(p_2)+√(p_3)/2√(p_1). It is easy to check that when p_0≥ p_i>0 (i=1,2,3), we have M_AB + N_AB + K_AB= 0, M_AB≤ (1- R_A/d_A)I_AB + R_A ⊗ I_B ⇔ I_AB - M_AB ≥ 0, and N_AB⊗ I_E + P_BE(K_AB⊗ I_E)P_BE^† =1/2M_AB⊗ I_E + P_BE(1/2M_AB⊗ I_E)P_BE^† ≥ 0. It also satisfies [M_ABJ_AB^Λ] = 1-[(√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2]. Thus we have proven that {M_AB,N_AB,K_AB,R_A} is a feasible solution to the dual SDP in Eq. (<ref>), which implies 1 - 2^-_max,(Λ)≥[M_AB J_AB^Λ] = 1-[(√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2]. Then it yields 2^-_max,(Λ)≤ (√(p_1)+√(p_2))^2 + (√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2. Thus, we arrive at 2^-_max,(Λ) = (√(p_1)+√(p_2))^2 +(√(p_2)+√(p_3))^2+(√(p_1)+√(p_3))^2. Since Γ_AB^ is the Bell state after normalization, we know the ADG-squeezed channel is the identity channel.

http://arxiv.org/abs/2303.07267v2

20230313164521

CrysFieldExplorer: a software for rapid optimization of crystal field Hamiltonian

cond-mat.str-el

Neutron Scattering Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA,37831 Neutron Scattering Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA,37831 Department of Physics and Astronomy, Louisiana State University , Baton Rouge, Louisiana, USA,70803 Neutron Scattering Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA,37831 Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA,37831 Neutron Scattering Division, Oak Ridge National Laboratory, Oak Ridge, Tennessee, USA,37831 We present a new lite python-based program, CrysFieldExplorer, for fast optimizing crystal electric field (CEF) parameters to fit experimental data. The main novelty of CrysFieldExplorer is the development of a unique loss function, referred to as the Spectrum-Characteristic Loss (L_Spectrum), which is defined based on the characteristic polynomial of the Hamiltonian matrix. Particle Swarm Optimization and Covariance matrix adaptation evolution strategy are used to find the minimum of the total loss function. We demonstrate that CrysFieldExplorer can performs direct fitting of CEF parameters to any experimental data such as neutron spectrum, susceptibility, magnetizations etc. CrysFieldExplorer can handle a large amount of none-zero CEF parameters and reveal multiple local and global minimum solutions. Detailed crystal field theory, description of the loss function, implementation and limit of the program are discussed within context of two examples. CrysFieldExplorer: a software for rapid optimization of crystal field Hamiltonian Huibo Cao* March 30, 2023 ================================================================================= § INTRODUCTION Single-ion magnetic anisotropy is one of the key elements for exotic quantum states in frustrated systems <cit.>. Crystal electric field (CEF) is responsible for single-ion magnetic anisotropy, and it occurs through Coulomb interaction and Pauli exclusion between the central cation and surrounding anions that splits the energy levels of electrons grouped by different orbitals of the central cation. In the absence of the CEF effect or in a spherically symmetrical field, these energy levels would otherwise be degenerate. The CEF effect is known to cause dramatic magnetic anisotropy which collectively produce exotic ground states in a wide range of quantum materials <cit.> and unconventional high temperature superconductors <cit.>. The calculation of excited energy levels of magnetic ions in a crystalline environment has been well studied <cit.>. Stevens and Hutchings have illustrated systematically the process of determination of the perturbing Hamiltonian from the evaluation of the electrostatic potential experienced by the magnetic ion from the surrounding charges <cit.>. One of the common conventions to express the CEF Hamiltonian is using Stevens Operators. The crux of the CEF study is to solve the single-ion Hamiltonian by matrix diagonalization and retrieve eigenvalues and eigenfunctions from the CEF Hamiltonian to fit the corresponding crystal field excitations. Inelastic neutron scattering is a suitable experimental technique because it directly measures transitions between ground states and excited states of magnetic ions. It has gradually become one of the most popular experimental methods in conducting crystal field analysis. After obtaining neutron scattering and other bulk property data, the next step in the CEF analysis is fitting the CEF Hamiltonian to experimental observables. It has always been a challenging optimization problem to find a solution set that best describes the experimentally measured data. Efforts to tackle this problem include programs such as SPECTRE <cit.>, McPhase <cit.>, Mantid <cit.>, SIMPRE <cit.> CFca <cit.>, FOCUS <cit.> and the latest PyCrystalField <cit.>. These programs use various techniques to calculate the CEF parameters but they all share the same approach that uses χ^2, a type of loss function, to find the global minimum. One of the disadvantages associated with choosing the χ^2 loss function is it relies on a set of starting parameters that needs to be adjacent to the true solution. As we will demonstrate in the next chapter, large energy boundaries in the χ^2 type loss function exist that could trap the optimization process on a local minimum. Typically, the point-charge model and Monte Carlo sampling are used to generate starting parameters. However, the point-charge model is a classical approximation by positioning several point charges around a magnetic ion to represent a distribution of valence electrons. This heavily limits the accuracy of the point-charge model. Additionally, building a point-charge model or Monte Carlo simulation can sometimes be none-trivial for inexperienced researchers. More importantly, when dealing with low local symmetry such as C_i, C_1, the number of none-zero crystal field parameters can be well into high 20s. The complexity and cost of computation of the multi-dimensional loss function increases exponentially and can easily overwhelm currently existing software. To tackle these challenges, we have designed a special loss function, Spectrum - Characteristic Loss (L_Spectrum), based on the theory of polynomial characteristics and developed a lite python program called CrysFieldExplorer. CrysFieldExplorer takes advantage of particle swarm optimization (PSO) <cit.> and covariance matrix adaptation evolution strategy (CMA-ES) <cit.> to minimize a combination of Spectrum - Characteristic Loss and traditional χ^2 losses. It is able to quickly fit neutron spectroscopy data and any other experimental results such as susceptibility, magnetization, specific heat, neutron diffraction, etc with CEF model and yields a series of solutions containing information with local and global minima. It bypasses the step of using a point-charge model to estimate starting parameters. In the following chapters, we will show our methods can not only discover the solutions reported in literature, but also uncover multiple other solutions that could fit all provided experimental data equally well. This suggests that the data used in the traditional optimization process may not have enough resolution to distinguish one from another and require further measurements such as local magnetic susceptibility with polarized neutron diffraction <cit.>. Our findings suggest that the traditional standard of determining one set of best CEF parameters may be flawed. We, therefore, suggest that future CEF analysis should include a list of possible solutions and discuss their relative physical meaning and the reasons for the preferred solution. § THEORY §.§ Crystal Electric Field Theory The CEF potential can be constructed from a point-charge perspective detailed in Hutching's work <cit.>. The electrostatic potential due to the surrounding point-charges can be expressed using tesseral harmonic functions in Cartesian coordinates. Following this convention, the Crystal Field (CF) Hamiltonian can be expressed by Stevens operators as <cit.>: H_CF = ∑_n,m B^m_nO^m_n=∑_n,m[A^m_n θ_n]O^m_n, where B^m_n (|m| ≤ n) is the CF parameters fitted from experimental measurements, O^m_n are the Stevens' Operators. The CF parameters can also be expressed in terms of A^m_n and θ_n, where θ_n represents reduced matrix elements and is also called as the Stevens factor. For rare earth ions, θ_n has been tabulated in Table VI of Ref. <cit.>. §.§ Neutron Scattering Cross-section Inelastic neutron scattering spectroscopy is well-suited in studying the crystal field excitations because neutrons can directly excite electron spins from one level to another and measure the difference between two energy levels. The observed intensity and excited energy levels can be used to fit the crystal field parameter B^m_ns in Eq. <ref>. The excited energy levels correspond to the energy difference between the lowest eigenvalues and the corresponding state of the CEF Hamiltonian. The observed intensity is related to the partial differential magnetic cross-section expressed as: d^2σ/dΩ dE' = Ck_fk_i F(|Q|)S(|Q|,ħω), where k_f and k_i are the momentum of scattered and incident neutrons, C is a constant including the Debye-Waller factor, F(|Q|) is the magnetic form factor of the sample. S(|Q|,ħω) is the scattering function, ħω indicate the energy of neutrons. From S(|Q|,ħω) the relative intensities from different states of the CEF exictations can be calculated. At constant |Q|, the S(|Q|, ħω) is, S(Q, ħω)=∑_i,i'(∑_α|⟨i|J_α|i'⟩|^2)e^-β E_i∑_j e^-β E_j L(Δ E+ħω), i → i' indicates the transition between state i and i' for the magnetic ion. β=-1k_B T, k_B is the Boltzman constant, T indicates temperature. L(Δ E+ħω) is a Lorentzian function that guarantees the energy conservation when neutrons induce a transition from state i to i', which poses a finite energy width or lifetime. The ∑_α sums all three x,y,z component of the J_α matrix. Although many CEF experiments are measured at low temperature, high temperature measurements can be important for confirming or extracting the scattering signal originated from CEF excitations that can be calculated with Eq. <ref>. §.§ Magnetization and Susceptibility Magnetization under an external magnetic field 𝐇 can be readily calculated by combining CEF Hamiltonian and the Zeeman term. The overall Hamiltonian can be expressed as: H=H_CF - μ_B g_J H·J. By diagonalizing Eq.<ref>, we can calculate the eigenstates E_n and eigenfunctions |i⟩ of the CF Hamiltonian in a magnetic field 𝐇. The three components (α = x, y, z) of the magnetization in Cartesian coordinate system are given as, M_α(𝐇,T)=g_j∑_n e^-β E_n⟨n|J|n⟩/Z where Z=∑_n e^-βE_n is the partition function. The powder average of the magnetization can be derived from Eq. <ref> by calculating the averaged magnetization on a unit sphere. The magnetic susceptibility can be calculated by taking derivatives of the magnetization M_α, χ_αβ=∂ M_α∂ H_β. The powder averaged magnetic susceptibility can be calculated in the same fashion by substituting M_α with its powder-average version. Detailed information about the calculation can be found in ref.<cit.>. §.§ Loss Function The core of an optimization problem is constructing a proper loss function that is smooth on the parameter space, sensitive to the change of input parameters and does not produce large energy boundaries around local and global minima. Currently, a few popular softwares such as SPECTRE <cit.>, PyCrystalField <cit.> and Mantid <cit.>, etc are able to fit the CEF Hamiltonian using experimental observables. Different to the existing software packages, here we introduce a newly designed loss function, Spectrum - Characteristic Loss. and demonstrate its advantage when deployed in CEF optimization. The foundation of the Spectrum - Characteristic Loss is from the theory of characteristic polynomial - det(λ I - A) = 0, where λ is the eigenvalue of matrix A, I is an identity matrix. The difference between the lowest and excited eigenvalues are neutron observed energy levels. Using this feature, we construct the loss function for the neutron measured energy levels as: L_E = log_10(∑_i=1det{(E_exp[i]+E_cal[0])I -H }^2det{E_exp[i]I }^2), E_exp indicate the observed excitation levels from neutron scattering, the summation of i starts from 1 which is the first excited energy level observed by neutron experiment. E_cal[0] is the ground state eigenvalue from matrix H. E_cal[0] may not be 0. Subsequently, we construct the mean square root deviation for the intensity and the inverse susceptibility as: L_Intensity = √(∑(I_true[i]-I_calc[i])^2)√(∑(I_true[i])^2), and L_1/χ = √(∑(1/χ_true[i]-1/χ_calc[i])^2)√(∑(1/χ_true[i])^2). To construct the Spectrum - Characteristic loss function, we combine Eq.<ref>, <ref> such that L_Spectrum=log_10(L_E)+L_Intensity. To fit other bulk measured physical properties, CrysFieldExplorer sums up L_Spectrum with other mean square root losses computed from the experimental data and conduct global optimization. Figure <ref> compares the Spectrum - Characteristic Loss with traditional mean square losses of the excited energy levels and relative intensities for an example material . It represents an optimization conducted on a 15-dimensional parameter space. More physical details about the fitting results are discussed in a later chapter. Here we show the computed loss along a straight line passing through the true solution of <cit.>. The line is chosen to point in a random direction within the 15-dimensional parameter space to simulate an optimization procedure (a real optimization may not follow a straight line). The loss is then projected onto the B^0_2 dimension for visualization. The upper panel shows the Spectrum - Characteristic Loss function while the lower panel is the standard χ^2 loss function. To compare the loss functions like-to like we also apply a log10 operation on the energy levels such that χ^2=log_10(χ^2_E)+χ^2_I. Comparing the upper panel to the lower panel, one can tell that the Spectrum - Characteristic Loss has a global structure around the minimum and does not produce sharp energy barrier at B^0_2 ∼ -0.02. Such an energy barrier can trick the optimization algorithm assuming it has already reached the global minimum. This can lead to a wrong solution if the starting parameters fall within the energy well. The Spectrum - Characteristic Loss also appears to be more sensitive to slight perturbations in CEF parameters. In the specific direction that the loss is computed, traditional χ^2 loss function is only able to converge at the true solution which is artificially designed such that the energy levels exactly match between calculated and observed, causing both L_Spectrum and χ^2 to appear sharp at the true solution. On the other hand, the Spectrum - Characteristic Loss function can discover at least three other minima in the parameter space. In reality the real global minimum is unknown for a measured material, so it is important to have a complete list including all potential minima for further examination as different CEF parameters from the list could all reproduce results identical to experimentally measured properties. As the number of CEF parameters increases, it requires more constraints to conclusively solve the CEF Hamiltonian. The multiple solutions CrysFieldExplorer discovered is an indication that the input observables do not impose enough constraints to reach the true solution. In addition, CrysFieldExplorer can add any other experimental data such as specific heat, anisotropy g-tensors, magnetization, etc in the optimization process. The loss function for additional data can be added the same way as mean square root deviation. Users can also adjust weights for different losses to fine tune the optimization process. § EXAMPLES §.§ Benchmark with Rare Earth Pyrochlore: Yb_2Ti_2O_7 Geometrical frustration has been of great interest in condensed matter physics due to the ability to host rich order-disorder states. These exotic states of matter are results of collective behaviors that arise from frustrated interactions within the quantum many-body system. Understanding these phenomena is crucial for developing next generation quantum technologies <cit.>. Rare earth pyrochlore is the archetype of magnetic frustration in three dimensions. This system has been reported to host exotic magnetic states such as quantum spin ice <cit.>, quantum spin liquid <cit.> etc. The pyrochlore structure generally possesses chemical composition of A_2B_2O_7 (B= Sn, Ge, Pt, Zr, Ti etc) with a space group Fd3m and point group D_3d at A site. We selected this system as a high symmetry end to benchmark our optimization model. Its local site possesses a 3-fold rotation axis along the local [111] direction. Due to the symmetry constraints, the number of none zero CEF parameters is 6. In this chapter, we demonstrate the use of CrysFieldExplorer by comparing a list of optimized results with observables in typical measurements such as inelastic neutron scattering, susceptibility, magnetizations and anisotropy g-tensors. One of the representative systems is the Yb_2Ti_2O_7, a promising quantum spin ice candidate with XY-type single-ion magnetic anisotropy <cit.>. The ground states and crystal field effect of Yb^3+ have been well studied by Gaudet et al. in Ref.<cit.>. Yb^3+ has an electronic configuration of 4f^13, total angular momentum 𝐉=7/2. The crystal electric field effect originates from the surrounding O^2- ions of the Yb^3+ site. As a result, 2J+1=8 - fold degeneracy is expected to be lifted by the CEF effect from the ground state, resulting in 7 excited levels. Additionally, Yb^3+ is a Kramers' ion, therefore the 8 - fold degeneracy are all doubly degenerate into 4 well separated doublets. A total of 3 excitations are expected to be observed through inelastic neutron scattering spectroscopy. Oxygen atoms around the local A-site form a distorted cubic structure shown in Figure <ref>. The most convenient placement of the local coordination system placing the local z - axis along the local three-fold <111> rotational axis perpendicular to the oxygen plane. The resulting Hamiltonian takes the form of: H_CEF= B^0_2Ô^0_2 + B^0_4Ô^0_4 + B^3_4Ô^3_4 + B^0_6Ô^0_6 + B^3_6Ô^3_6 + B^6_6Ô^6_6. Eq. <ref> follows the convention of Stevens operators for Ô^m_n, B^m_ns are CF parameters used to describe the Coulomb potential generated by the surrounding oxygen atoms. Inelastic neutron scattering results have been reported in Ref.<cit.>. 1.25 Table <ref> lists previous works on the CEF analysis for . Bertin et al.'s results <cit.> are obtained within the point-charge approximation and taken as starting parameters for the work by Gaudet etl al <cit.>. They deployed a typical mean-square-root deviation minimization algorithm that searched through the six dimensional parameter space in the vicinity of the starting values from Ref. <cit.>. By using a point-charge model to generate starting parameters, one assumes these starting values are close to the global minimum. However, Table <ref> shows that the refined value from point-charge model can be different by an order of magnitude (i.e, B^0_6, B^3_6). Now we will demonstrate the results of CrysFieldExplorer in the system. We adapt the searching algorithm using particle swarm optimization with our customized Spectrum - Characteristic Loss function discussed in previous sections. We construct the PSO model with a particle size N=400 and iteration number of 100 times. The details of the hyper-parameters can be found in Ref. <cit.>. The total loss is constructed as L_total=L_Spectrum+L1/χ, where χ in this case represents the susceptibility data. These are the same experimental data used to obtain results in Ref.<cit.>. We present CrysFieldExplorer's results in Figure <ref> and compare it with those reported in Ref.<cit.>. Figure <ref> lists all 150 results from CrysFieldExplorer on Yb_2Ti_2O_7 in a log scale along the y-axis. Each CEF parameter is plotted against a custom defined goodness of fit χ^2=[Expected Value-Observed Value]^2Expected Value^2. Readers should be aware of the definition of present χ^2 is simply the sum of percentage difference between true solution and calculated solution. It is calculated by comparing the excited energy levels and relative intensities, neutron spectroscopy, susceptibility and magnetization using CEF parameters provided in the Ref. <cit.>. This χ^2 is merely an indication of the agreement goodness between calculated and true observables. We chose χ^2 < 1 as a cut off number in present study as we find results below this number shows good overall agreement between fit and true solution. The solid grey circles represent 150 solutions generated using PSO algorithm. The solid red circles represent 13 solutions with χ^2 < 1. The blue circles mark the parameters reported from Ref. <cit.> and is taken as the "true solution" in this example. The χ^2 for the true solution is given an artifically 0 but shifted by 0.03 so it is visible on the log scale. One of the important observations from Figure <ref> is that there exists three local minima in the parameter space of crystal field. Observing Panel a), b), d), f) from Figure <ref> shows two out of the three local minima are concentrated regions where CrysFieldExplorer can easily converge. There is a third region that contains one acceptable solution with χ^2 <1. The true solution reported from Ref. <cit.> marked by blue circle falls in 1 of the minima. The green and black circles are in two other minima. Interestingly, while no constraints were set on B_4^3 and B_6^3, panels c) and e) show symmetric behavior w.r.t to 0 as expected by the symmetry of local crystalline environment. To analyze the physical meaning of the CEF parameters found by CrysFieldExplorer, the main panels in Fig. <ref> a)-d) show the comparison between calculated and true results in terms of inelastic neutron scattering data, inverse susceptibility, magnetization and the ratio of g_z/g_x(y) for the Yb^3+. Unfortunately the experimentally measured susceptibility and magnetization data was not given in Ref. <cit.>. We assume the calculated susceptibility and magnetization using the CEF parameters provided from Ref. <cit.> as "true solution". Each red lines in a)-c) and the red data points in d) indicate the calculated properties. The blue lines and open blue circles in a) - d) represents the true solutions. The black, green color scheme indicate two other minima results from present study corresponding to the same green and black data points in Figure <ref>. We also list our representative results (green and black data) as CEF parameters in Table <ref>. They are marked by a and b respectively. Clearly, Figure <ref> and <ref> demonstrate our PSO optimizer can successfully converge to the true CEF parameters region from Ref.<cit.>. A universal starting range of [-100,100] meV were set for all six B^m_ns in Eq.<ref>. For general users, we feel this range can be appropriately determined based on the magnitudes of excited energy levels. Additionally, as Spectrum - Characteristic Loss is smooth and can effectively avoid producing steep energy barriers seen in Fig. <ref>, a larger range for the CEF parameters should not be problematic. The average time for the PSO to converge to one solution in the 6-dimensional CEF parameter space was approximately 4.7 minutes on a 12th Gen Intel(R) Core i7-1265U laptop with single core processing. The maximum magnitude of the searching range compared to the order of the minimum CEF parameter B^0_6 is on the order of 10^4. CrysFieldExplorer can locate solutions within this vast range of parameter space within minutes. This is an example to show the powerful optimization capability with CrysFieldExplorer. Observing each panel in Figure <ref> reveals more details of CrysFieldExplorer's results. The main panels of a), b) and c) in Fig. <ref> compare the calculated solutions of χ^2 < 1 with the true solution. These solutions are over-plotted in red and form a region which makes it easier to identify the relative positions of the calculated solutions in from the true solutions. The insets of a), b) c) in Figure <ref> compare the true solution with the solutions a) and b) found and listed in Table <ref>. Solution a) and b) are indicated as green and black open circle in panel d). Readers should pay particular attention to solution a) and b) as they are in two different regions in the CEF parameter space. In panel a) of Fig.<ref>, the calculated spectrum shows good agreement with the true spectrum except at ∼ 76 meV. We estimate approximately ± 3% variation on the peak intensities from our calculated solutions compared to the true solution. The excitations at 81.8 meV and 116.2 meV matches well in both peak width and integrated intensity. In panel b), the inverse of susceptibility shows excellent agreement as well. The CEF fit and true solution are indistinguishable by eye. Panel c) shows the 10K magnetization data. The solid grey lines mark an region where the true solution and calculated solutions with χ^2 < 1 reside in. The inset of panel a), b), c) shows a similar story, no significant difference between true solution and solutions a) and b). Panel d) is the most interesting plot as it shows the ratio of the anisotropy g-tensors g_zg_x(y). Only in this panel we can distinguish solution b) from solution a) and the true solution. Observing panel d), most of the calculated anisotropy g-tensors are in one dominant region where the ratio of g_z/g_x(y) is ∼ 0.5 while g_x is close to 3.8. Our best solution, represented by the green open circle, yields a set of g-tensor g_x(y) = 3.73 g_z = 2.07 that are within the uncertainty given from the published result of 3.69 ± 0.15 and 1.92 ± 0.20 <cit.>.The ratio of g_z/g_x(y) is of particular interest. Previous polarized neutron studies from Cao <cit.> reported g_x(y)=4.1 and g_z=2.25 resulting in a ratio of 1.82. Bertin's <cit.> and Gaudet's<cit.> results are listed in Table <ref> with g_z/g_x(y)= 2.0 and 1.92. From our solution a), the ratio can be calculated to be 1.79 closes to Cao 's result. Now we turn our attention to solution b). Despite having a good agreement with the experimental values in the bulk property measurements, it shows a strong ratio of g_z/g_x(y) ∼ 3.5, with g_z=4.51 significantly larger than solution a) and the true solution. This observation suggests solution b) represents a local minimum that fits well with limited constraints. By analyzing the anisotropy g-tensor, we conclude that solution b) does not agree with the physical model of . Now we turn our focus to solution a) which is in a different region with the true solution, but we cannot distinguish solution a) from the true solution by all the given measurements. Table <ref> compares the ground state wavefunctions between solution a) marked by green circle in Figure <ref> and that reported in Ref. <cit.>. Both solutions also produce similar ground state wavefunctions as reflected by the similarity in the anisotropy g-tensors. The dominant term in the ground state wave function from both solutions are the ±12 terms. Combining these results with previous discussions suggests solution a) is a degenerate solution under the existing constraints imposed by the experimental data reported in the system. Only one set of CEF parameters were reported in each paper from previous studies <cit.>, but our discovery implies that these papers may need more data to conclusively determine the CEF parameters. The ground state of has been extensively studied from various perspectives and our results do not imply whether the previous analyses are incorrect. However, it is an alarming discovery that suggests future CEF analysis should be cautious in confirming or denying their findings. In next chapter, we will show that lack of enough constraints may pose serious uncertainty in determining the true CEF parameters. 2.2 Regardless of which set of CEF parameters is correct, one of CrysFieldExplorer's biggest advantages is the capability to directly commence the fitting procedure starting from Eq. <ref> without any a priori knowledge of the starting parameters. This allows users to skip the step of building a point-charge model or conduct Monte Carlo simulation, either of which can be daunting for inexperienced users. CrysFieldExplorer only requires a wide fitting range to be given. All the parameters of were chosen from a range of [-100, 100] meV which can be estimated from the magnitude of the leading B_2^2 term. This range requirement can be easily estimated by users. Too large or too small beyond this range could result in the excited energy levels to be unreasonable large or small. To conclude, in this section we demonstrate that CrysFieldExplorer can quickly produce a list of CEF solutions for including the ones reported in the literature. Our improved loss function L_Spectrum has a higher sensitivity to the change of CEF parameters and a better global structure that allows us to search through a wide range of parameter space. Identifying which one is the real solution is more complicated. On the one hand, there may not be enough constraints to conclude which solution is the correct one. On the other hand, as most of the experimental data are collected from powder sample, the information about the relationship between local coordinates and crystal structure is intrinsically missing, thus rendering the the model unable to reduce further to one result. §.§ Analysis on the Crystal Field Parameters of the Low Symmetry Tripod Kagome Lattice Er_3Mg_2Sb_3O_14 Now we consider a more challenging case - a tripod Kagome lattice that hosts various exotic magnetic ground states depending on the constituent rare earth element <cit.>. We picked Er_3Mg_2Sb_3O_14 as the test case here. The CEF fitting becomes more challenging for a magnetic ion at low point-symmetry site because the number of none zero crystal field parameters increases. Fig. <ref> shows the local structure of an Er atom in the tripod Kagome lattice Er_3Mg_2Sb_3O_14 with a space group of R3m. The rare earth ion Er^3+ is surrounded by eight oxygen atoms similar to that in pyrochlore system. However, the local site symmetry becomes C_2h in the tripod Kagome lattice, much lower than D_3d in the pyrochlore lattice. The twofold rotational C_2 axis lies in the Kagome plane and is displayed in Fig. <ref> as the green arrow on the center Er atom. As a result, this system requires 15 CEF parameters to fully describe the crystal field Hamiltonian, H_CEF= B^0_2Ô^0_2 + B^1_2Ô^1_2 + B^2_2Ô^2_2 + B^0_4Ô^0_4 + B^1_4Ô^1_4 + B^2_4Ô^2_4 + B^3_4Ô^3_4 + B^4_4Ô^4_4 + B^0_6Ô^0_6 + B^1_6Ô^1_6 + B^2_6Ô^2_6 + B^3_6Ô^3_6 + B^4_6Ô^4_6 + B^5_6Ô^5_6 +B^6_6Ô^6_6. According to Hund's rule, the total, spin and orbit angular momentum for Er^3+ are J=15/2, S=3/2 and L=6 respectively. This results in 2J+1=16 fold degeneracy for the ground state of Er^3+. Since Er^3+ is a Kramer ion, all 16 levels are doublets, 7 excited states are expected to be observed from inelastic neutron scattering measurement. Dun et al. <cit.> observed 6 out of 7 CEF transitions at 6.4(2), 10.5(3), 21.6(4), 50(1), 65(1) and 67.5(9) meV. Due to the large amount of none zero CEF parameters, they were not able to directly fit the CEF transitions with currently existing software packages such as PyCrystalField. A list of CEF parameters was obtained from an effective point-charge model which effectively reduced the number of fitted parameters. Now we demonstrate how CrysFieldExplorer can help provide insight to this problem and discuss why directly fitting the CEF parameters with given experimental data was not feasible. Due to the low site symmetry in Er_3Mg_2Sb_3O_14, the parameter space is magnitudes larger than that in and traditional mean-square-root loss function is less sensitive to the 15 CEF parameters comparing to L_Spectrum as demonstrated previously in Fig. <ref>. To handle this case, we need to improve the optimization from our previous particle swarm optimizer. We continue apply the L_Spectrum and deploy the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) <cit.>, which converges rapidly on complex optimization problems where the global minimum is extremely sharp in the parameter phase space. For a local site with low point group symmetry such as monoclinic in the present case, the choice of principle xyz axis can be somewhat arbitrary. In the case of tripod Kagome lattice, the y-axis is chosen to be parallel to the two-fold rotational axis C_2 such that none zero terms of the CEF parameters are expressed in Eq. <ref>. The choice of x and z axis is arbitrarily defined within a perpendicular plane to the C_2 rotational axis. The x-z plane is indicated as the green plane in Fig. <ref>. The dashed and solid arrows indicate x and z axis are not fixed within the plane perpendicular to y-axis. Rudowicz <cit.> has demonstrated that within the 15 CEF parameters, certain rotations exists along y-axis that make one of the CEF parameters 0 . Here we deploy the transformation of rotating the CEF coordination system w.r.t local y-axis by an angle α such that B_2^0 is 0 for all the solutions. The rotational angle α is defined as, tan(2α)=B_2^13 × B_2^0-B_2^2. The rotational transformations of relevant CEF parameters have been tabulated in Table 2 from Ref.<cit.>. The CEF parameters reported here have all been transformed under this standardization rule with each set of parameters rotated by an angle α. A complete list of CEF parameters and the transformation angle α are plotted in SFig. 1 in the supplemental material. We summarize our results in Fig. <ref>. Different to the case, CrysFieldExplorer can find a large number of solutions across the parameter space with no obvious segregation. We first look at Fig. <ref> c). As has 15 (14 independent) CEF parameters, we choose the B_2^0 as an example and plot the custom defined χ^2 loss on a log scale. The solid grey circles indicate all 150 solutions produced by CrysFieldExplorer in a time span of approximately 12 hours. We chose 1 as a cutoff number for the custom defined χ^2 and denote the best and worst 10 solutions with χ^2 <1 using green and blue open circles. Meanwhile, open blue circles indicate the true solution. The χ^2 for the true solution is artificially defined as 0 but shifted by a small number of 0.03 in order to be shown on the log plot. Figure <ref> a) and b) show the neutron data on the energy transfer range from -1 to 33 meV and 25 to 90 meV respectively. Fig. <ref> d) is the inverse susceptibility data. The solid grey lines in these three plots are an overlay of all the solutions with χ^2 < 1 found by CrysFieldExplorer using CMA-ES algorithm optimizing L_Spectrum. The open blue circle indicates the solution taken from Ref.<cit.> which are treated as true solution. Fig. <ref> a) b) and d) are the three data sets used by CrysFieldExplorer during the optimization process. Figure <ref> e) compares the calculated magnetizations with the true solution. The solid red lines and solid red lines mark the top and last 10 solutions with χ^2 < 1. Following the same color scheme, open blue circles are the true solution. We drew two solid grey lines to indicate a region where all the solutions stay within. The determination of the grey lines are guided by eye. Figure <ref> f) shows the ratio of the anistropy g-tensors. The grey squares indicate the ratio between the largest and smallest while the pink triangle indicate the largest and 2nd smallest anisotropy g-tensors for all solutions with χ^2 < 1. The green and red square represent the top and last 10 CEF fits with χ^2 < 1. Figure <ref> reveals several important features of CrysFieldExplorer when applied to the system. CrysFieldExplorer can perform perfect fits to the neutron data in panels a) and b) to the degree that is impossible for the human eyes to distinguish. We performed fitting on both the discrete CEF energy levels and their relative intensity as well as fitting directly on the neutron spectrum by applying a resolution limited Lorentzian function to compute the entire spectrum. There are 73 out of 150 sets of parameters with χ^2 < 1, all of them produced perfect match to human eyes between the true solution and CEF fit. Additional constraints were imposed when adding the inverse susceptibility data to the fit. However, within the χ^2 < 1 range, CrysFieldExplorer was still able to discover multiple solutions that fit all given experimental observables well. The powder averaged magnetization data was not considered during the fit process but used as a benchmark to check the fitting results. Comparing magnetization data in panel e) of Fig. <ref>, we start noticing differences between each solution set. The top 10 solutions show good match to the true solution at all magnetic fields measured. However, the last 10 CEF fitted solutions with χ^2 < 1 begin to show small deviations at 4 T magnetic field. However, we argue our these 10 solutions still fit all the data well and could still be considered acceptable a few decades ago given the computing power at that time. This shows the efficiency and accuracy CrysFieldExplorer exhibits when being deployed on the latest high-performance computers. Figure <ref> f) is particularly interesting because it indicates all of the solutions produced by CrysFieldExplorer are Ising-like, that is one of the three g_x, g_y or g_z are dominant comparing to the other two. We do not distinguish g_x, g_y or g_z here because there is a certain degree of freedom in choosing the local frame. From present data, it is not possible to obtain the relationship between local frame and global frame. Similar to the case, the powder data measured with the mentioned techniques do not have crystal structure resolution. To conclude, our results show that with increased free CEF parameters, the constraints required to narrow down to a single solution also increases drastically. CrysFieldExplorer was able to produce 73 solutions out of 150 searches that matches neutron data and susceptibility data well. The powder averaged magnetization data were not used in the fitting because it is computationally expensive to calculate the powder average. However, single crystal magnetization data could be added in the fitting without much penalty on the computational cost to impose more constraints. Despite not being able narrow down to one or a few solutions, CrysFieldExplorer is able to produce a list of acceptable solutions. These solutions produce excellent agreement between the observed and measured data such as inelastic neutron scattering spectrum, susceptibility, and magnetization measurements. We demonstrate in the case of , neutron spectroscopy and susceptibility do not impose enough constraints to fully describe the problem, thus making it an under-determined case when only considering limited experimental observables. Therefore, we agree with the statement in the original paper that direct fitting of the CEF parameter is not possible <cit.>. Nonetheless, in the system, by wrapping L_Spectrum into CMA-ES, we show the strong capability and high efficiency CrysFieldExplorer exhibits. The approximated time span to generate 150 sets of solutions is approximately 12 hours. We argue 73 of the 150 solutions produce acceptable results. Future updates can further improve the efficiency by leveraging high performance computing on multi-core processors on user-end computers. CrysFieldExplorer's results also suggest that when fitting a large number of CEF parameters, user should exercise extra caution when choosing the correct solution. In some of the previous research, a solution is believed to be valid when the neutron scattering data, susceptibility data, magnetization measurements, anisotropy g-tensors, etc are all consistent. However, utilizing CrysFieldExplorer's efficient algorithm, we show that these condition can all be met with very good agreement between experimental and fitted results in low site symmetry materials such as with 15 (14 independent) none zero CEF parameters. As the number of CEF parameter increases, more experimental data should be acquired to fit the CEF Hamiltonian with experimental model to obtain conclusive results. These can be, but not limited to, polarized neutron diffraction, specific heat etc. Future researchers should be cautious to draw conclusions on CEF analysis when considering limited experimental data. § IMPLEMENTATION We aim to provide the most straight forward implementations for CrysFieldExplorer to the research community. Users first need to specify the magnetic ions and local site symmetry. This allows CrysFieldExplorer to generate a list of none-zero CEF parameters. Then the total, orbital and spin angular momentum J, L and S can be automatically determined. The program will construct the Hamiltonian up to 6th order using the Stevens operators as well as the matrix representation of J_plus, J_minus and J_z. Then users can import the experimentally measured data into the program for fitting. In the present version of CrysFieldExplorer, it is able to fit neutron scattering data, susceptibility, single crystal magnetization, anisotropy g-tensor data all together with different weights. For neutron scattering data, relative intensities of the observed transitions from the ground state need to be specified by users from a data reduction software. The currently required data format is a two-column text file. The experimental data will be loaded into the optimization program. For high-symmetry cases the program will select particle swarm optimization and for low-symmetry system a more efficient CMA-ES will be chosen instead. The loss function will be predetermined as discussed before. Although CrysFieldExplorer does not require a set of specific starting parameters, a suitable range for all CEF parameters still needs to be given. However, this can be estimated from the magnitudes of the transition energy levels from the inelastic neutron scattering. The optimization strategy will minimize Spectrum - Characteristic Loss first. Once Spectrum - Characteristic Loss reaches -10, it will start optimizing magnetic measurements or any other user defined data. Users are able to prioritize different data by fine tuning their respective weights. By default, CrysFieldExplorer will generate 100 converged CEF parameters in .csv file. These parameters can be checked with either CrysFieldExplorer or a 3rd party software for consistency. The detailed implementation of CrysFieldExplorer can be found at https://github.com/KyleQianliMa/CrisFieldExplorer. § USE AND LIMITATIONS The program CrysFieldExplorer presented here provides a fast converging, gradient free method that scans through large CEF parameter phase space to provide satisfactory results based on user input. It bypasses the need of determining initial parameters from point-charge model which can be difficult for inexperienced users. CrysFieldExplorer contains two optimization methods, PSO and CMA-ES, to minimize a set of specially designed loss functions. The choice of each optimization method is generally chosen such that for CEF parameters <6, PSO is chosen, otherwise CMA-ES is preferred. Although CrysFieldExplorer can successfully produce reliable solutions, it cannot interpret the physical meaning of each solution. Furthermore, if the problem is under-defined, that is experimental data does not pose enough constraints to the CEF Hamiltonian, multiple solutions produced by CrysFieldExplorer can be confusing to determine which solution is the correct solution. Although we provide two optimization methods, there has not been a standard way to choose the best optimization methods for different problems. The weight for different losses can also be tricky to balance. In some test cases it is easier to fit one experiment than another. Finally, the current version of CrysFieldExplorer does not provide a user interface. It requires users with basic knowledge of Python programming language. § CONCLUSION We have developed a lite python-based program CrysFieldExplorer to provide an efficient procedure of fitting the CEF parameters from experimental data. CrysFieldExplorer uses two evolutionary algorithms - particle swarm optimization (PSO) and Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to minimize a newly designed loss function - Spectrum Characteristic loss L_Spectrum. Comparing L_Spectrum function with traditional mean-square-root loss, we showed that L_Spectrum is more sensitive to input parameters, significantly reduces the steepness of energy barriers on the parameter-loss space and has a better global structure around local and global minima. This allows CrysFieldExplorer to start the fitting procedure directly from the CEF Hamiltonian without prior knowledge of starting parameters. Although a fitting range of the CEF parameter is required, the range can be generously given such that it matches the magnitudes of the excited CEF energy levels. To demonstrate the use of CrysFieldExplorer, we have provided two examples in two different systems to show that our algorithm can discover multiple solutions that matches experimental data equally well. Our findings suggest that as the number of to-be-fitted CEF parameters increases, traditional experimental observables such as neutron scattering data, susceptibility and magnetization measurement may not impose enough constraints to fully describe the problem. Furthermore, neither the powder neutron spectroscopy nor the powder susceptibility data have crystal structure resolution to define the connection between local ion and crystal structure frame. This arbitrariness could further produce confusion when determining the correct CEF fit. Therefore, the present standard of reporting CEF parameters may not be conclusive as the CEF parameter phase space is highly degenerate. When conducting CEF analysis with limited experimental data, further measurements such as polarized neutron diffraction can provide additional information to help better determine the correct CEF Hamiltonian. § ACKNOWLEDGMENTS The research at Oak Ridge National Laboratory (ORNL) was supported by the U.S. Department of Energy (DOE), Office of Science, Office of Advanced Scientific Computing Research under the contract ERKJ387, and Office of Basic Energy Sciences, Early Career Research Program Award KC0402020, under Contract DE-AC05-00OR22725. A portion of this research used resources at the High Flux Isotope Reactor and the Spallation Neutron Source, DOE Office of Science User Facilities operated by the Oak Ridge National Laboratory. 51 fxundefined [1] ifx#1 fnum [1] #1firstoftwo secondoftwo fx [1] #1firstoftwo secondoftwo noop [0]secondoftwo ref[1]@startlink#1@href href[1]#1@endlink anitize@url [0]` 12`$12`&12`#12`1̂2`_12`%12 startlink[1] endlink[0] rl [1]href #1 @bib@innerbibempty [Gingras et al.(2000)Gingras, den Hertog, Faucher, Gardner, Dunsiger, Chang, Gaulin, Raju, and Greedan]PhysRevB.62.6496 author author M. J. P. Gingras, author B. C. den Hertog, author M. Faucher, author J. S. Gardner, author S. R. Dunsiger, author L. J. Chang, author B. D. Gaulin, author N. P. Raju, and author J. E. Greedan, title title Thermodynamic and single-ion properties of Tb^3+ within the collective paramagnetic-spin liquid state of the frustrated pyrochlore antiferromagnet Tb_2Ti_2O_7, 10.1103/PhysRevB.62.6496 journal journal Phys. Rev. B volume 62, pages 6496–6511 (year 2000)NoStop [Gardner et al.(2010)Gardner, Gingras, and Greedan]Gardener author author Jason S. Gardner, author Michel J. P. Gingras, and author John E. Greedan, title title Magnetic pyrochlore oxides, 10.1103/RevModPhys.82.53 journal journal Rev. Mod. Phys. volume 82, pages 53–107 (year 2010)NoStop [Cao et al.(2009a)Cao, Gukasov, Mirebeau, Bonville, Decorse, and Dhalenne]huiboPRL author author H. Cao, author A. Gukasov, author I. Mirebeau, author P. Bonville, author C. Decorse, and author G. Dhalenne, title title Ising versus XY Anisotropy in Frustrated R_2Ti_2𝐎_7 Compounds as “Seen” by Polarized Neutrons, 10.1103/PhysRevLett.103.056402 journal journal Phys. Rev. Lett. volume 103, pages 056402 (year 2009a)NoStop [Bertin et al.(2012)Bertin, Chapuis, de Réotier, and Yaouanc]Bertin_2012 author author A Bertin, author Y Chapuis, author P Dalmas de Réotier, and author A Yaouanc, title title Crystal electric field in the R_2Ti_2O_7 pyrochlore compounds, 10.1088/0953-8984/24/25/256003 journal journal Journal of Physics: Condensed Matter volume 24, pages 256003 (year 2012)NoStop [Gaudet et al.(2015)Gaudet, Maharaj, Sala, Kermarrec, Ross, Dabkowska, Kolesnikov, Granroth, and Gaulin]JoYb2015 author author J. Gaudet, author D. D. Maharaj, author G. Sala, author E. Kermarrec, author K. A. Ross, author H. A. Dabkowska, author A. I. Kolesnikov, author G. E. Granroth, and author B. D. Gaulin, title title Neutron spectroscopic study of crystalline electric field excitations in stoichiometric and lightly stuffed Yb_2Ti_2O_7, 10.1103/PhysRevB.92.134420 journal journal Phys. Rev. B volume 92, pages 134420 (year 2015)NoStop [Dun et al.(2021)Dun, Bai, Stone, Zhou, and Mourigal]zhilin author author Zhiling Dun, author Xiaojian Bai, author Matthew B. Stone, author Haidong Zhou, and author Martin Mourigal, title title Effective point-charge analysis of crystal fields: Application to rare-earth pyrochlores and tripod kagome magnets R_3Mg_2Sb_3O_14, 10.1103/PhysRevResearch.3.023012 journal journal Phys. Rev. Research volume 3, pages 023012 (year 2021)NoStop [Paddison et al.(2017)Paddison, Daum, Dun, Ehlers, Liu, Stone, Zhou, and Mourigal]paddison2017continuous author author Joseph AM Paddison, author Marcus Daum, author Zhiling Dun, author Georg Ehlers, author Yaohua Liu, author Matthew B Stone, author Haidong Zhou, and author Martin Mourigal, title title Continuous excitations of the triangular-lattice quantum spin liquid YbMgGaO_4, @noop journal journal Nature Physics volume 13, pages 117–122 (year 2017)NoStop [Gaudet et al.(2019)Gaudet, Smith, Dudemaine, Beare, Buhariwalla, Butch, Stone, Kolesnikov, Xu, Yahne, Ross, Marjerrison, Garrett, Luke, Bianchi, and Gaulin]JoPRL author author J. Gaudet, author E. M. Smith, author J. Dudemaine, author J. Beare, author C. R. C. Buhariwalla, author N. P. Butch, author M. B. Stone, author A. I. Kolesnikov, author Guangyong Xu, author D. R. Yahne, author K. A. Ross, author C. A. Marjerrison, author J. D. Garrett, author G. M. Luke, author A. D. Bianchi, and author B. D. Gaulin, title title Quantum Spin Ice Dynamics in the Dipole-Octupole Pyrochlore Magnet Ce_2Zr_2O_7, 10.1103/PhysRevLett.122.187201 journal journal Phys. Rev. Lett. volume 122, pages 187201 (year 2019)NoStop [Bramwell and Gingras(2001)]Gingras2001 author author Steven T. Bramwell and author Michel J. P. Gingras, title title Spin ice state in frustrated magnetic pyrochlore materials, 10.1126/science.1064761 journal journal Science volume 294, pages 1495–1501 (year 2001), http://arxiv.org/abs/https://www.science.org/doi/pdf/10.1126/science.1064761 https://www.science.org/doi/pdf/10.1126/science.1064761 NoStop [Ross et al.(2011a)Ross, Savary, Gaulin, and Balents]kate2011 author author Kate A. Ross, author Lucile Savary, author Bruce D. Gaulin, and author Leon Balents, title title Quantum excitations in quantum spin ice, 10.1103/PhysRevX.1.021002 journal journal Phys. Rev. X volume 1, pages 021002 (year 2011a)NoStop [Porée et al.(2022)Porée, Lhotel, Petit, Krajewska, Puphal, Clark, Pomjakushin, Walker, Gauthier, Gawryluk, and Sibille]cefcehfo author author Victor Porée, author Elsa Lhotel, author Sylvain Petit, author Aleksandra Krajewska, author Pascal Puphal, author Adam H. Clark, author Vladimir Pomjakushin, author Helen C. Walker, author Nicolas Gauthier, author Dariusz J. Gawryluk, and author Romain Sibille, title title Crystal-field states and defect levels in candidate quantum spin ice Ce_2Hf_2O_7, 10.1103/PhysRevMaterials.6.044406 journal journal Phys. Rev. Mater. volume 6, pages 044406 (year 2022)NoStop [Keimer and Moore(2017)]keimer2017physics author author B Keimer and author JE Moore, title title The physics of quantum materials, @noop journal journal Nature Physics volume 13, pages 1045–1055 (year 2017)NoStop [Smith et al.(2022)Smith, Benton, Yahne, Placke, Schäfer, Gaudet, Dudemaine, Fitterman, Beare, Wildes, Bhattacharya, DeLazzer, Buhariwalla, Butch, Movshovich, Garrett, Marjerrison, Clancy, Kermarrec, Luke, Bianchi, Ross, and Gaulin]evanprx author author E. M. Smith, author O. Benton, author D. R. Yahne, author B. Placke, author R. Schäfer, author J. Gaudet, author J. Dudemaine, author A. Fitterman, author J. Beare, author A. R. Wildes, author S. Bhattacharya, author T. DeLazzer, author C. R. C. Buhariwalla, author N. P. Butch, author R. Movshovich, author J. D. Garrett, author C. A. Marjerrison, author J. P. Clancy, author E. Kermarrec, author G. M. Luke, author A. D. Bianchi, author K. A. Ross, and author B. D. Gaulin, title title Case for a U(1)_ Quantum Spin Liquid Ground State in the Dipole-Octupole Pyrochlore Ce_2Zr_2O_7, 10.1103/PhysRevX.12.021015 journal journal Phys. Rev. X volume 12, pages 021015 (year 2022)NoStop [Fulde et al.(1970)Fulde, Hirst, and Luther]fulde1970superconductors author author Peter Fulde, author LL Hirst, and author Alan Luther, title title Superconductors containing impurities with crystal-field split energy levels, @noop journal journal Zeitschrift für Physik A Hadrons and nuclei volume 230, pages 155–177 (year 1970)NoStop [Mesot and Furrer(1997)]mesot1997crystal author author Joël Mesot and author A Furrer, title title The crystal field in rare earth based high-temperature superconductors, @noop journal journal Journal of superconductivity volume 10, pages 623–643 (year 1997)NoStop [Lynn et al.(1980)Lynn, Moncton, Passell, and Thomlinson]hightccef19809 author author J. W. Lynn, author D. E. Moncton, author L. Passell, and author W. Thomlinson, title title Magnetic correlations and crystal-field levels in the superconductor (Ce_0.73Ho_0.27) Ru_2, 10.1103/PhysRevB.21.70 journal journal Phys. Rev. B volume 21, pages 70–78 (year 1980)NoStop [Metoki et al.(2004)Metoki, Kaneko, Matsuda, Galatanu, Takeuchi, Hashimoto, Ueda, Settai, Ōnuki, and Bernhoeft]metoki2004magnetic author author N Metoki, author K Kaneko, author TD Matsuda, author A Galatanu, author T Takeuchi, author S Hashimoto, author T Ueda, author R Settai, author Y Ōnuki, and author N Bernhoeft, title title Magnetic structure and the crystal field excitation in heavy-fermion antiferromagnetic superconductor cept_3si, @noop journal journal Journal of Physics: Condensed Matter volume 16, pages L207 (year 2004)NoStop [Goremychkin et al.(2004)Goremychkin, Osborn, Bauer, Maple, Frederick, Yuhasz, Woodward, and Lynn]CEF2004hightc author author E. A. Goremychkin, author R. Osborn, author E. D. Bauer, author M. B. Maple, author N. A. Frederick, author W. M. Yuhasz, author F. M. Woodward, and author J. W. Lynn, title title Crystal Field Potential of PrOs_4Sb_12: Consequences for Superconductivity, 10.1103/PhysRevLett.93.157003 journal journal Phys. Rev. Lett. volume 93, pages 157003 (year 2004)NoStop [Rubio Temprano et al.(2000)Rubio Temprano, Mesot, Janssen, Conder, Furrer, Mutka, and Müller]hightccef1 author author D. Rubio Temprano, author J. Mesot, author S. Janssen, author K. Conder, author A. Furrer, author H. Mutka, and author K. A. Müller, title title Large Isotope Effect on the Pseudogap in the High-Temperature Superconductor HoBa_2Cu_4O_8, 10.1103/PhysRevLett.84.1990 journal journal Phys. Rev. Lett. volume 84, pages 1990–1993 (year 2000)NoStop [Boothroyd et al.(1992)Boothroyd, Doyle, Paul, and Osborn]boothroydcef author author A. T. Boothroyd, author S. M. Doyle, author D. MK. Paul, and author R. Osborn, title title Crystal-field excitations in Nd_2CuO_4, Pr_2CuO_4, and related n-type superconductors, 10.1103/PhysRevB.45.10075 journal journal Phys. Rev. B volume 45, pages 10075–10086 (year 1992)NoStop [Pickett(2021)]pickett2021dawn author author Warren E Pickett, title title The dawn of the nickel age of superconductivity, @noop journal journal Nature Reviews Physics volume 3, pages 7–8 (year 2021)NoStop [Stevens(1952)]Stevens_1952 author author K W H Stevens, title title Matrix elements and operator equivalents connected with the magnetic properties of rare earth ions, 10.1088/0370-1298/65/3/308 journal journal Proceedings of the Physical Society. Section A volume 65, pages 209–215 (year 1952)NoStop [Hutchings(1964)]hutchings1964point author author M.T. Hutchings, title title Point-charge calculations of energy levels of magnetic ions in crystalline electric fields, (publisher Academic Press, year 1964) pp. pages 227–273NoStop [Lea et al.(1962)Lea, Leask, and Wolf]lea1962raising author author KR Lea, author MJM Leask, and author WP Wolf, title title The raising of angular momentum degeneracy of f-electron terms by cubic crystal fields, @noop journal journal Journal of Physics and Chemistry of Solids volume 23, pages 1381–1405 (year 1962)NoStop [Walter(1984)]walter1984treating author author U Walter, title title Treating crystal field parameters in lower than cubic symmetries, @noop journal journal journal of physics and chemistry of solids volume 45, pages 401–408 (year 1984)NoStop [Loewenhaupt(1990)]loewenhaupt1990crystal author author M Loewenhaupt, title title Crystal fields in low symmetry systems, @noop journal journal Physica B: Condensed Matter volume 163, pages 479–482 (year 1990)NoStop [Boothroyd(2014)]spectre author author AT Boothroyd, title title Spectre, a program for calculating spectroscopic properties of rare earth ions in crystals, @noop (year 2014)NoStop [Rotter(2013)]McPhase author author M. Rotter, title title Mcphase, http://www.mcphase.de/ (year 2013)NoStop [Arnold et al.(2014)Arnold, Bilheux, Borreguero, Buts, Campbell, Chapon, Doucet, Draper, Ferraz Leal, Gigg, Lynch, Markvardsen, Mikkelson, Mikkelson, Miller, Palmen, Parker, Passos, Perring, Peterson, Ren, Reuter, Savici, Taylor, Taylor, Tolchenov, Zhou, and Zikovsky]Mantid author author O. Arnold, author J.C. Bilheux, author J.M. Borreguero, author A. Buts, author S.I. Campbell, author L. Chapon, author M. Doucet, author N. Draper, author R. Ferraz Leal, author M.A. Gigg, author V.E. Lynch, author A. Markvardsen, author D.J. Mikkelson, author R.L. Mikkelson, author R. Miller, author K. Palmen, author P. Parker, author G. Passos, author T.G. Perring, author P.F. Peterson, author S. Ren, author M.A. Reuter, author A.T. Savici, author J.W. Taylor, author R.J. Taylor, author R. Tolchenov, author W. Zhou, and author J. Zikovsky, title title Mantid—Data analysis and visualization package for neutron scattering and μSR experiments, https://doi.org/10.1016/j.nima.2014.07.029 journal journal Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment volume 764, pages 156–166 (year 2014)NoStop [Baldovi et al.(2013)Baldovi, Cardona-Serra, Clemente-Juan, Coronado, Gaita-Ariño, and Palii]baldovi2013simpre author author Jose J Baldovi, author Salvador Cardona-Serra, author Juan M Clemente-Juan, author Eugenio Coronado, author Alejandro Gaita-Ariño, and author Andrew Palii, @noop title Simpre: A software package to calculate crystal field parameters, energy levels, and magnetic properties on mononuclear lanthanoid complexes based on charge distributions, (year 2013)NoStop [R.Osborn(2017)]CFcal author author R.Osborn, title title Cfcal, https://github.com/rayosborn/cfcal (year 2017)NoStop [Fabi(1995)]fabi1995focus author author Peter Fabi, @noop title FOCUS: an interactive crystal electric field parameter fitting package using neutron scattering data (publisher Council for the Central Laboratory of the Research Councils, year 1995)NoStop [Scheie(2021)]scheie2021pycrystalfield author author Allen Scheie, title title Pycrystalfield: software for calculation, analysis and fitting of crystal electric field hamiltonians, @noop journal journal Journal of Applied Crystallography volume 54, pages 356–362 (year 2021)NoStop [Bonyadi and Michalewicz(2017)]PSO author author Mohammad Reza Bonyadi and author Zbigniew Michalewicz, title title Particle Swarm Optimization for Single Objective Continuous Space Problems: A Review, 10.1162/EVCO_r_00180 journal journal Evolutionary Computation volume 25, pages 1–54 (year 2017)NoStop [Hansen(2006)]hansen2006cma author author Nikolaus Hansen, title title The CMA evolution strategy: a comparing review, @noop journal journal Towards a new evolutionary computation: Advances in the estimation of distribution algorithms , pages 75–102 (year 2006)NoStop [Tokura et al.(2017)Tokura, Kawasaki, and Nagaosa]tokura2017emergent author author Yoshinori Tokura, author Masashi Kawasaki, and author Naoto Nagaosa, title title Emergent functions of quantum materials, @noop journal journal Nature Physics volume 13, pages 1056–1068 (year 2017)NoStop [Ball(2017)]ball2017quantum author author Philip Ball, title title Quantum materials: Where many paths meet, @noop journal journal MRS Bulletin volume 42, pages 698–705 (year 2017)NoStop [Giustino et al.(2021)Giustino, Lee, Trier, Bibes, Winter, Valentí, Son, Taillefer, Heil, Figueroa et al.]giustino20212021 author author Feliciano Giustino, author Jin Hong Lee, author Felix Trier, author Manuel Bibes, author Stephen M Winter, author Roser Valentí, author Young-Woo Son, author Louis Taillefer, author Christoph Heil, author Adriana I Figueroa, et al., title title The 2021 quantum materials roadmap, @noop journal journal Journal of Physics: Materials volume 3, pages 042006 (year 2021)NoStop [Cava et al.(2021)Cava, De Leon, and Xie]cava2021introduction author author Robert Cava, author Nathalie De Leon, and author Weiwei Xie, @noop title Introduction: quantum materials, (year 2021)NoStop [Bassman et al.(2020)Bassman, Gulania, Powers, Li, Linker, Liu, Kumar, Kalia, Nakano, and Vashishta]bassman2020domain author author Lindsay Bassman, author Sahil Gulania, author Connor Powers, author Rongpeng Li, author Thomas Linker, author Kuang Liu, author TK Satish Kumar, author Rajiv K Kalia, author Aiichiro Nakano, and author Priya Vashishta, title title Domain-specific compilers for dynamic simulations of quantum materials on quantum computers, @noop journal journal Quantum Science and Technology volume 6, pages 014007 (year 2020)NoStop [Lau et al.(2020)Lau, Xia, and Cao]lau2020emergent author author Chun Ning Lau, author Fengnian Xia, and author Linyou Cao, title title Emergent quantum materials, @noop journal journal MRS Bulletin volume 45, pages 340–347 (year 2020)NoStop [Stanev et al.(2021)Stanev, Choudhary, Kusne, Paglione, and Takeuchi]stanev2021artificial author author Valentin Stanev, author Kamal Choudhary, author Aaron Gilad Kusne, author Johnpierre Paglione, and author Ichiro Takeuchi, title title Artificial intelligence for search and discovery of quantum materials, @noop journal journal Communications Materials volume 2, pages 105 (year 2021)NoStop [Kimura et al.(2013)Kimura, Nakatsuji, Wen, Broholm, Stone, Nishibori, and Sawa]kimura2013quantum author author Kenta Kimura, author S Nakatsuji, author JJ Wen, author C Broholm, author MB Stone, author E Nishibori, and author H Sawa, title title Quantum fluctuations in spin-ice-like pr2zr2o7, @noop journal journal Nature communications volume 4, pages 1934 (year 2013)NoStop [Martin et al.(2017)Martin, Bonville, Lhotel, Guitteny, Wildes, Decorse, Ciomaga Hatnean, Balakrishnan, Mirebeau, and Petit]petit2017pr227 author author N. Martin, author P. Bonville, author E. Lhotel, author S. Guitteny, author A. Wildes, author C. Decorse, author M. Ciomaga Hatnean, author G. Balakrishnan, author I. Mirebeau, and author S. Petit, title title Disorder and quantum spin ice, 10.1103/PhysRevX.7.041028 journal journal Phys. Rev. X volume 7, pages 041028 (year 2017)NoStop [Gingras and McClarty(2014)]gingras2014quantum author author Michel JP Gingras and author Paul A McClarty, title title Quantum spin ice: a search for gapless quantum spin liquids in pyrochlore magnets, @noop journal journal Reports on Progress in Physics volume 77, pages 056501 (year 2014)NoStop [Ross et al.(2014)Ross, Qiu, Copley, Dabkowska, and Gaulin]orderbydisorder author author K. A. Ross, author Y. Qiu, author J. R. D. Copley, author H. A. Dabkowska, and author B. D. Gaulin, title title Order by Disorder Spin Wave Gap in the XY Pyrochlore Magnet Er_2Ti_2O_7, 10.1103/PhysRevLett.112.057201 journal journal Phys. Rev. Lett. volume 112, pages 057201 (year 2014)NoStop [Bhardwaj et al.(2022)Bhardwaj, Zhang, Yan, Moessner, Nevidomskyy, and Changlani]bhardwaj2022sleuthing author author Anish Bhardwaj, author Shu Zhang, author Han Yan, author Roderich Moessner, author Andriy H Nevidomskyy, and author Hitesh J Changlani, title title Sleuthing out exotic quantum spin liquidity in the pyrochlore magnet Ce_2Zr_2O_7, @noop journal journal npj Quantum Materials volume 7, pages 51 (year 2022)NoStop [Cao et al.(2009b)Cao, Gukasov, Mirebeau, and Bonville]Cao_2009 author author H B Cao, author A Gukasov, author I Mirebeau, and author P Bonville, title title Anisotropic exchange in frustrated pyrochlore Yb_2Ti_2O_7, 10.1088/0953-8984/21/49/492202 journal journal Journal of Physics: Condensed Matter volume 21, pages 492202 (year 2009b)NoStop [Ross et al.(2011b)Ross, Savary, Gaulin, and Balents]kate author author Kate A. Ross, author Lucile Savary, author Bruce D. Gaulin, and author Leon Balents, title title Quantum excitations in quantum spin ice, 10.1103/PhysRevX.1.021002 journal journal Phys. Rev. X volume 1, pages 021002 (year 2011b)NoStop [Rudowicz(1986)]rudowicz1986standardization author author Czesław Rudowicz, title title On standardization and algebraic symmetry of the ligand field hamiltonian for rare earth ions at monoclinic symmetry sites, @noop journal journal The Journal of chemical physics volume 84, pages 5045–5058 (year 1986)NoStop [Rudowicz(1985)]rudowicz1985relations author author Czesław Rudowicz, title title Relations between arbitrary symmetry spin-hamiltonian parameters bkq and bkq in various axis systems, @noop journal journal Journal of Magnetic Resonance (1969) volume 63, pages 95–106 (year 1985)NoStop