Datasets:

kjappelbaum

/

chemnlp-arxiv

like 1

# Introduction {#introduction .unnumbered} Where the availability of online information continues to grow, assessing the accuracy of this information can be a challenge, where support from automated fact-checking systems is becoming increasingly necessary to help getting rid of misinformation. One of the key components of an automated fact-checking pipeline is the check-worthy claim detection system, which given a collection of sentences as input, ranks them based on their need to be fact-checked. The ranked output can be used in different ways, among others by providing fact-checkers with the most important claims to work on, by feeding a claim verification system with the top claims to be checked, or by informing the general public that the top claims may be disputed and need verification. In recent years, there has been a body of research developing models for check-worthy claim detection. Many of these systems have been tested against each other by competing on benchmark datasets released as part of the CheckThat! shared tasks. The recent tendency of these models has been to use Transformer models such as BERT and RoBERTa which have led to improved performance for detecting check-worthy claims in different languages including English and Arabic. Existing claim detection models have been developed for and tested in datasets that provide a mix of topics in both training and test data, in such a way that topics encountered in the test set were already seen during training. While this has provided a valuable setting to compare models against each other, it is also arguably far from a realistic scenario where a claim detection model is trained from data pertaining to certain topics to then detect check-worthy claims in a new, unseen topic; for example, where a model trained from claims associated with politics needs to be used for detecting claims associated with COVID-19 when the new topic emerges. The objective of our work is to assess the extent of the problem of detecting check-worthy claims across topics, i.e. simulating the scenario where new topics emerge and for which labelled data is lacking or limited. To undertake this challenge, we use an Arabic language dataset from the CheckThat! shared tasks which comprises a collection of claims and non-claims distributed across 14 different topics. We first conduct experiments in zero-shot settings using a model called AraCW, which has not seen any data associated with the target topic; the low performance of this model for some of the topics demonstrates the importance of this hitherto unstudied problem in check-worthy claim detection. We further propose AraCWA, an improved model that incorporates two additional components to enable few-shot learning and data augmentation. We test AraCWA with a range of few-shot and data augmentation settings to evaluate its effectiveness, achieving improvements of up to 11% over AraCW. To the best of our knowledge, our work is the first to tackle the check-worthy claim detection task across topics. In doing so, we make the following novel contributions: - We establish a methodology for cross-topic claim detection in both zero-shot and few-shot settings. - We assess the extent of the challenge of detecting claims for new, unseen topics through zero-shot experiments. - We propose an improved approach, AraCWA, which leverages few-shot learning in combination with data augmentation strategies to boost the performance on new topics. - We perform a set of ablative experiments to assess the contribution of the data augmentation technique as well as the few-shot strategy. - We perform an analysis of the characteristics of the topics under study, finding that the semantic similarity between topics can be a reasonable proxy to determine the difficulty level of a topic. In presenting the first study in cross-topic claim detection, our work also posits the need for more research in this challenging yet realistic scenario. # Related Work on Check-Worthy Claim Detection {#related-work-on-check-worthy-claim-detection .unnumbered} While a large portion of the research in automated fact-checking has focused on claim verification, work on check-worthy claim detection has been more limited. However, the claim detection task constitutes an important first step that identifies the claims to be prioritised and to be fed to the subsequent verification component. In what follows, we discuss existing work in check-worthy claim detection with a particular focus on the Arabic language and data augmentation. Existing research in check-worthy claim detection has primarily focused on sentences in the English language. This is the case, for example, of one of the first fact-checking systems developed in the field, ClaimBuster, which studied detection of claims from transcripts of US presidential election debates through what they called "claim spotter". In another study, investigated the use of context and discourse features to improve the claim detection component; while this is a clever approach, it does not generalise to other types of datasets such as tweets in our case, given the limited context available typically in social media. More recently, developed a model, namely CNC, that leverages InferSent embeddings along with part-of-speech tags and named entities to identify claims. In addition to the above studies, there have been numerous competitive efforts to develop claim detection systems thanks to the CheckThat! series of shared tasks. These shared tasks have provided a benchmark for researchers to study the task in a competitive manner, including publicly available datasets in multiple languages. In our study, we use the datasets available from these shared tasks in the Arabic language. #### Check-worthy claim detection in the Arabic language. Check-worthy claim detection for the Arabic language was first tackled by ClaimRank when training the neural network model with originally English sentences from political debates translated into Arabic. In addition, other learning models, e.g. Gradient boosting and k-nearest neighbors, were inspected for the task in Arabic of check-worthiness at CheckThat! 2018. The paradigm shift in the field of NLP occurred with the emergence of Transformers, and Bidirectional Encoder Representations (BERT), which became the state-of-art model for several tasks, including a range of text classification tasks. This is also reflected in the works on check-worthy claim detection as most participants of the recent CheckThat Labs fine-tuned BERT models for Arabic such as mBERT, AraBERT, and BERT-Base-Arabic. However, existing models focus on evaluating their performance on test sets whose topics overlap with those seen in the training data, without focusing on the challenges that different topics may pose to the task. As part of this study, we focus on the novel challenge of investigating how claim detection models perform differently for different topics #### **Data augmentation for claim detection.** Data augmentation techniques enable to leverage the (generally limited) existing data to generate new synthetic data through alterations. By carefully altering the original data samples, one can augment the available training data with the newly generated samples; however, data augmentation techniques may also produce noisy samples which lead to performance deterioration, and therefore creation of useful samples is crucial. Data augmentation strategies have also been successfully used to alleviate different problems, e.g. to settle an imbalanced dataset and to reduce model biases. Different data augmentation methods exist, which categorised into three types: (i) paraphrasing, where the new augmented sentence has similar semantics to the original one, (ii) noising, which adds discrete or continuous noise to the sentence, and (iii) sampling, which is similar to paraphrasing but for more specific tasks which need more information about this to this task like data format. Data augmentation techniques have been barely used in claim detection. For example, used a contextual word embedding augmentation strategy in their participation in the 2021 CheckThat! lab. These data augmentation techniques have however not been studied studied in the context of cross-topic claim detection. In our study, we propose the AraCWA model which provides the flexibility of being used with different data augmentation strategies. More specifically, we test three data augmentation strategies to evaluate their effectiveness in the cross-topic claim detection task, namely back-translation, contextual word embedding augmentation and text-generating augmentation. # Methodology {#sec:Methodology .unnumbered} ## Problem Formulation and Evaluation {#sec:Problem formulation .unnumbered} The check-worthy claim detection task consists in determining which of the sentences, out of a collection, constitute claims that should be fact-checked. Given a collection of sentences \(S = \{s_1, s_2,..., s_n\}\), the model needs to classify each of them into one of \(L = \{CW, NCW\}\), where CW = check-worthy and NCW = not check-worthy. The collection of sentences in \(S\) belongs to a number of topics \(T = \{t_1, t_2,..., t_m\}\), where each sentence \(s_j\) only belongs to one topic \(t_i\). Where previous research has experimented with a mix of sentences pertaining to all topics spread across both training and test data, our objective here to experiment with claim detection across topic is to leave a topic \(t_i\) out for the test, using the sentences from the rest of the topics for training. While the task has sometimes been formulated as a classification problem, here and in line with the CheckThat! shared tasks, we formulate it as a ranking task. As a ranking problem, the model needs to produce an ordered list of sentences in \(S\), where the sentences in the top of the list are predicted as the most likely to be check-worthy claims. The task is in turn evaluated through the Mean Average Precision (MAP) metric, which is defined as follows: \[MAP = \frac{1}{k} \sum_{i=1}^k AP_i\] That is the average of the AP's (average precisions) across classes, where in the case of claim detection \(k = 2\). AP is in turn defined as follows: \[AP_c = \frac{1}{N} \sum{j=1}^N P_i\] where N is the number of instances being considered in the evaluation, whose precision values are averaged. Ultimately, \(AP_c\) measures the average precision on the top \(N\) items for class \(c\). After calculating the AP values for both classes, they are averaged to calculate the final MAP. ## Dataset {#sec:dataset .unnumbered} We make combined use of two claim detection datasets in the Arabic language, CT20-AR and CT21-AR, released as part of two different editions of the CheckThat! shared task. The datasets contain tweets which are labelled as check-worthy or not check-worthy. The tweets are in turn categorised into topics, which allows us to set up the cross-topic experiments. We show the main statistics of these two datasets in Table [1](#tab:AR-datasets){reference-type="ref" reference="tab:AR-datasets"}. #### **The CT20-AR dataset.** This was released as part of the CheckThat! 2020 Lab and it comprises 7.5k tweets distributed across 15 topics, with the tweets evenly distributed across the topics, i.e. 500 tweets per topic. #### **The CT21-AR dataset.** Released as part of the CheckThat! 2021 Lab, part of this dataset overlaps with CT20-AR. After removing the overlapping tweets, this dataset provides two new topics with a total of 600 tweets (347 and 253 tweets for each topic). The combination of both datasets originally led to 17 topics (15 + 2). However, during the process of combining both datasets into one, we found that four of the topics are related to COVID-19, namely those with topic IDs "CT20-AR-03", "CT20-AR-28 w1", "CT20-AR-28 w2" and "CT20-AR-29", Given the significant overlap between them, we combined these four topics into one, hence reducing the number of topics from 17 to 14. The final, combined dataset contains 8,100 tweets distributed across 14 topics. See Appendix A for a list and description of these topics. The CT20-AR and CT21-AR datasets differ slightly in the labels. Where the former only provides binary labels for check-worthy (CW) and not check-worthy (NCW), the latter provides two levels of labels: (i) claim (C) or not claim (NC), and (ii) check-worthy (CW) and not check-worthy (NCW). In both cases, we only rely on the CW / NCW labels for consistency. The distribution of labels in the entire dataset is clearly skewed towards the NCW class accounting for 71.6% of the tweets, whereas the CW class accounts for 28.4% of the tweets (see Figure [\[fig:OverallDatasets\]](#fig:OverallDatasets){reference-type="ref" reference="fig:OverallDatasets"}). Further drilling down into the distribution of CW and NCW across the different topics, Figure [\[fig:ClassesofTopics\]](#fig:ClassesofTopics){reference-type="ref" reference="fig:ClassesofTopics"} shows a high degree of variation. With only one topic where the CW class is larger than the NCW class (i.e. CT20-AR-19, "Turkey's intervention in Syria"), all of the other classes have different degrees of prevalence of the NCW class. Some of the topics have a particularly low number of CW instances, which is the case of "CT20-AR-08\" (i.e. "Feminists\") with only 5% of check-worthy claims, and "CT20-AR-30\" (i.e. "Boycotting countries and spreading rumors against Qatar\") with only 3% of check-worthy claims. ## Experiment settings {#sec:Experiment settings .unnumbered} We perform experiments in two different settings: (i) first, in zero-shot detection settings using AraCW, where the target topic is unseen during training, and (ii) then, in few-shot training settings using AraCWA, where a few samples of the target topic are seen during training. To enable a fair comparison of experiments in both settings, we first hold out the data to be used as test sets. From each of the 14 topics in the dataset, we hold out 200 instances which will be used as part of the training data in the few-shot settings. Holding out these instances for all the experiments enables us to have identical test sets for both settings, where the 200 held-out instances are not used. #### **Zero-shot detection across topics.** In this case, we run a separate experiment for each topic. Where \(t_i\) is the target topic in the test set, we train the model from all topics but \(t_i\). The training data does not see any instances pertaining to topic \(t_i\). #### **Few-shot detection across topics.** The setting in this case is similar to the zero-shot detection, where separate experiments are run for each topic \(t_i\). However, the 200 instances held out above are used as part of the training data. Therefore, in this case, the experiments with topic \(t_i\) in the test set use a model trained for all the other topics as well as 200 instances pertaining to the target topic \(t_i\). ## Claim Detection Models: AraCW and AraCWA {#sec:Approaches .unnumbered} We next describe the two different models we propose for our experiments. We propose AraCW as the base model for the zero-shot experiments, as well as the variant AraCWA which incorporates a data augmentation strategy on top of few-shot learning to boost performance across topics. #### **AraCW.** As one of the top-ranked models in the CheckThat! 2021 Lab, the AraCW model is built on top of a fine-tuned AraBERTv0.2-base transformer model. We perform a set of preprocessing steps which proved effective during initial experiments through our participation in the CheckThat! 2021 Lab: - We substitute all URLs, email addresses, and user mentions with \[url\], \[email\] and \[user\] tokens. - We eliminate line breaks and markup written in HTML, repeated characters, extra spaces, and unwanted characters including emoticons. - We correct white spaces between words and digits (non-Arabic, or English), and/or a combination of both, and before and after two brackets. AraCW processes the tweets through Sequence Classification. Then, the results of the neural network output layer are passed into a softmax function in order to acquire the probability distribution for each predicted output class; we use the value output by the softmax function to rank the sentences by check-worthiness, which produces the final ranked output by ordering them by score. #### **AraCWA.** Our proposed AraCWA model builds on top of AraCW to incorporate two key components: few-shot learning and data augmentation. AraCWA aims to make the most of a small number of instances pertaining to the target topic by using a data augmentation strategy to increase its potential. With AraCWA, we also aim to assess the extent to which a small portion of labelled data can boost performance on new topics, where despite the cost of labelling data, it could be affordable given that it is a small portion. AraCWA's additional components for few-shot learning and data augmentation are as follows: #### *Few-shot learning.* This component leverages the 200 instances pertaining to the target topic, which were originally held out from the test set (see Experiment Settings). This enables to fine-tune the model in a broader training set which includes more relevant samples. The 200 instances incorporated into the training set add to the other 13 topics used for the training (i.e. excluding the 14th topic used as the target). #### *Data augmentation.* Once the 200 instances are incorporated through the few-shot learning component, AraCWA leverages a data augmentation strategy to increase its presence by generating new synthetic samples. The data augmentation component of AraCWA is flexible in that it can make use of different data augmentation algorithms. In our experiments, we test AraCWA with three data augmentation methods, as follows: 1. **Back translation:**\ The objective of a back translation strategy is to translate a text twice, first to another language, and then back to the original language. The new, back-translated sample is likely to be different while similar. To achieve this, we use the state-of-art OPUS-MT models for open translation . We use these models to translate the sample data of a target topic into the English language, then translate them back to the Arabic language. These new, back-translated samples represent new training samples that augment the original training data. 2. **Contextual word level augmentation:**\ The objective of this approach is to generate new samples by altering existing samples using a contextualised language model. In this case, we make use of the nlpaug python package to implement the contextual word embedding technique of augmentation on samples of a target topic. We make use of the arabertv02 transformer model to determine the best appropriate word for augmentation and set the proportion of words to be augmented to 0.3. Additionally, the parameter action is set to "substitute,\" which will replace the word based on the results of the contextual embedding calculation. 3. **Text generation:**\ The objective of this approach is to train a model that can then generate new texts from scratch. To achieve this, we use the aragpt2-medium model to implement AraGPT2, an Arabic version of the GTP2 model. First, the samples from the target topics that we intend to generate more text from are prepared by ArabertPreprocessor for aragpt2-medium. We use the following settings with AraGPT2: - num_beam is set to 5 to execute beam search for cleaner text. - maximum length of the generated text is 200. - top_p is set to 0.75 to determine the cumulative probability of most possible tokens to be selected for sampling. - repetition_penalty is set to 3 to penalize repetition in a text and avoid infinite loops. - token length of no_repeat_ngram_size is set to 3 in order to avoid repeating phrases. # Results {#Results .unnumbered} Next, we present results for our check-worthy claim detection experiments across topics, beginning with zero-shot experiments and followed by few-shot experiments, as well as ablation experiments where the data augmentation is removed from the pipeline. ## AraCW: Zero-Shot detection across topics {#sec:Results_zero-shot .unnumbered} Table [2](#tab:one topic-results){reference-type="ref" reference="tab:one topic-results"} shows results for the zero-shot experiments using the AraCW model for the 14 topics separately. We can observe high variation in terms of MAP scores for the different topics, with values ranging from as low as 0.24 to 0.75. Where the average performance for all topics is over 0.54, we observe that five of the topics perform below this average (underlined and coloured in red in the table). The best score is 0.7538 for the topicID "CT20-AR-19" which is about "Turkey's intervention in Syria" while the worst one is 0.2468 for "CT20-AR-08" which includes tweets about "Feminists". In the first case, we assumed that the best performance is because of the nature of the political war on the topic, and there are some topics of the same nature that were used through training such as "Houthis in Yemen\" and "Events in Libya\". Thus, the model already learned some implications about the topic prior to testing. In the second case, the topic of "Feminists" is totally different and unique with respect to the other topics in the datasets most of which are political. Likewise, the model performed poorly with topicIDs "CT20-AR-10" "CT20-AR-23" "CT20-AR-30" where the topics are about: "Waseem Youssef" which refer to religious views, "The case of the Bidoon in Kuwait" indicates a social issue, and "Boycotting countries and spreading rumours against Qatar" contains rumours. The other results fluctuate between 54% and 68%. This indicates that the AraCW model struggles in the zero-shot settings, particularly when the target topic differs substantially from the ones used for training. This in turn motivated the development of AraCWA leveraging few-shot learning and data augmentation, whose results we discuss in the next section. ## Few-shot detection with data augmentation {#sec:Results_few-shot .unnumbered} Table [3](#tab:results-across-topics){reference-type="ref" reference="tab:results-across-topics"} shows results for the AraCWA model leveraging few-shot learning and data augmentation, namely Back-translation (BT), Contextual word Embedding (CWE), and text generation (TxtGen). We observe substantial improvements of the three AraCWA variants over the AraCW baseline model, with overall improvements ranging from 9% with AraCWA/TxtGen to 11% with AraCWA/CWE. These results demonstrate that AraCWA's enhanced ability to incorporate few shots and augment data leads to substantial improvements. When we look at the improvements across the individual topics (indicated as percentages within brackets reflecting the absolute improvement over AraCW), we observe that most of the differences are positive. Exceptions include the topicID "CT20-AR-05\" where performance drops by 1% and the topicID "CT20-AR-23\" about "Bidoon in Kuwait\", where the performance difference is negligible. Apart from these exceptions, the rest of the topics experience a positive impact with the use of AraCWA. Most remarkably, four of the topics experience an improvement above 20%, which are coloured in blue. Looking back at the topics that proved problematic with the use of the AraCW model (highlighted in red in the AraCW column), we observe that the topicID "CT20-AR-10\" ("Waseem Youssef\") improves by 3% with AraCWA/CWE, the topicID "CT20-AR-08\" ("Feminists\") improves by 24% and the topicID "CT20-AR-30\" ("boycotting countries and spreading rumors against Qatar\") improves by 21%. The remainder topic, "CT20-AR-23\" ("Bidoon in Kuwait\") has no improvement (0%). Other topics with remarkable improvements include "CT21-AR-01\" ("Events in Gulf; Saudi-Qatari reconciliation and Gulf summit\") which improves by 25%, and the topicID "CT20-AR-12\" ("Sudan and normalization\") with han improvement of 28%. All in all, the overall 11% improvement of AraCWA/CWE shows the great potential of AraCWA compared to AraCW, which almost certainly guarantess that performance will not drop, but its improvement margin varies substantially. ## Ablation experiments: Few-shot learning without data augmentation {#ablation-experiments-few-shot-learning-without-data-augmentation .unnumbered} As part of the first ablation experiment to test the different components of AraCWA, we test it with the data augmentation component. Table [4](#tab:ablation){reference-type="ref" reference="tab:ablation"} shows the results comparing the best version of AraCWA using the CWE data augmentation with the ablated AraCWA which does not use any data augmentation strategy. We observe that the use of the data augmentation strategy leads to an overall improvement of 1% over the non-augmented baseline. There are three topics where the data augmentation makes a major impact, with improvements of 11% and 14% for CT20-AR-08 and CT21-AR-01 respectively, and a drop of 11% for CT20-AR-23. The rest of the topics experience a lesser impact, with changes ranging from-4% to +5%. ## Ablation experiments: Using fewer shots {#ablation-experiments-using-fewer-shots .unnumbered} We next test another ablated version of AraCWA, where we reduce the number of shots fed to the few-shot component from the original 200 shots. In this case, we test with 50, 100, 150 and 200 shots. Figure [\[fig:fewer-shots\]](#fig:fewer-shots){reference-type="ref" reference="fig:fewer-shots"} shows the results for AraCWA with different numbers of shots. While there is some inconsistency between the models using 100 and 150 shots, we see that the model with 50 shots performs clearly the worst, whereas the model with 200 shots performs the best. These experiments show the importance of the few-shot learning component of AraCWA, which contributes the most to the improvement of AraCWA over AraCW. # Discussion {#sec:Discussion .unnumbered} Despite the overall improvement of AraCWA, our experiments have shown substantial variability in performance across different topics in the claim detection task, suggesting that indeed some topics are much more challenging than others. Despite the clear improvement of AraCWA over AraCW, the improved model still struggles with some of the topics. One of the factors determining how easy or difficult a new topic will be is its similarity or dissimilarity with respect to the topics seen by the model. Motivated by this, we next look at the semantic similarities between topics and we analyse how they may help us explain the trends in performance we observe. We use the AraBERT transformer model to compute pairwise semantic similarities between the 14 topics in the dataset, which helps us quantify how topics resemble / differ between them. Figure [\[fig:Topics_Similarities\]](#fig:Topics_Similarities){reference-type="ref" reference="fig:Topics_Similarities"} shows the pairwise semantic similarity scores for the 14 topics. While the similarity scores range from low 0.80's to mid 0.90's, we observe that the lower values are clustered around particular topics. This is for example the case for topics like "CT20-AR-08\" ("Feminists"), and "CT20-AR-23\" ("The case of the Bidoon in Kuwait\"), both of which have a large number of low similarity scores and also showed lower performance scores in the zero-shot learning experiments with AraCW as shown in Table [2](#tab:one topic-results){reference-type="ref" reference="tab:one topic-results"}. Even if some of the above similarity scores may explain the performance observed in our experiments, there are some exceptions. For example, topicID "CT20-AR-30\" ("Boycotting countries and spreading rumors against Qatar\") obtains consistently high similarity scores above 86% while it proved to be a challenging topic in the experiments. While the semantic similarity may potentially serve as a proxy to predict the challenging nature of a topic before conducting any annotations for the topic, it should be considered carefully as there can indeed be other factors beyond similarity affecting the performance for a topic. Improving predictability of the difficulty of a topic could be an important avenue for future research. # Conclusions {#sec:Conclusions .unnumbered} Our research on check-worthy claim detection highlights the importance of considering a scenario which has been overlooked so far, i.e. tackling new topics which have not been seen before. Through the use of a competitive model, AraCW, we have shown that indeed tackling new topics is especially challenging. Given that this constitutes a realistic scenario where one has certain datasets labelled to train models and needs to deal with new topics for which to identify claims, our research calls for the need of more research in this scenario. As a first attempt to tackle the cross-topic check-worthy claim detection task, we have proposed a novel model called AraCWA, which incorporates capabilities for few-shot learning and data augmentation into the claim detection. This model has led to substantial improvements which can reach up to 11% overall improvements across multiple topics, positing this as a plausible direction to take for improving the cross-topic detection model. Our research also opens important avenues for future work. Available datasets for claim detection are limited, not least in Arabic, which hinder broader investigation into wider sets of topics; expanding the sets of available topics would be highly valuable to enrich this research. Likewise, this could enable broader and deeper investigation into the inconsistent improvements across different topics, which could lead to further understanding into what makes a topic more or less challenging.

{'timestamp': '2022-12-19T02:13:14', 'yymm': '2212', 'arxiv_id': '2212.08514', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08514'}

# Introduction {#sec:introduction} neural networks (CNNs) are state-of-the-art approach for semantic image segmentation. Their large number of trainable parameters make them capable to fit to different kinds of tasks yielding high performance in terms segmentation accuracy . Yet, in real world applications, CNNs seem sometimes unable to capture and generalize from complex, heterogeneous training sets when the amount of training data is limited . Specifically, in the case of medical image segmentation, samples from the background class, which provide essential contextual information for regions of interest (ROIs), make up the majority of the training set while containing diverse sets of structures with heterogeneous characteristics, making it hard for the segmentation model to learned accurate decision boundaries. When trained with datasets with a highly heterogeneous background class, the segmentation model is prone to underfit these contextual samples and fail to separate the ones which share characteristics similar to the ROI samples. The model then produces false positives (FP), yielding systematic over-segmentation. In this study, we argue that underfitting of the contextual information is a main cause of degraded segmentation performance by affecting precision (calculated as \(\frac{\text{TP}}{\text{TP}+\text{FP}}\)). We observe that better context representation where the background class is decomposed into several subclasses, for example, using additional anatomy labels, significantly improves the ROI segmentation accuracy. We show some examples of human-defined context labels in Fig. [\[fig_mainidea\]](#fig_mainidea){reference-type="ref" reference="fig_mainidea"}(a). In the case of liver tumor segmentation, for example, it is beneficial to also have labels for the liver available in addition to the tumor class. Empirically, we find that when training with human-defined context labels, the segmentation model can yield better performance in terms of Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD), as shown in Fig. [\[fig_mainidea\]](#fig_mainidea){reference-type="ref" reference="fig_mainidea"}(b). However, these human-defined context labels are not always available and difficult and time-consuming to obtain. In many applications, only the ROI labels, e.g., tumor class, are available. Here, we propose context label learning (CoLab), which automatically generates context labels to improve the learning of a context representation yielding better ROI segmentation accuracy. As demonstrated in Fig. [\[fig_mainidea\]](#fig_mainidea){reference-type="ref" reference="fig_mainidea"}(b), CoLab can bring similar improvements when compared with training with human-defined context labels, without the need for expert knowledge. The contribution of this study can be summarized as follows: 1) With the observations of six datasets, we conclude that underfitting of the background class consistently degrades the segmentation performance by decreasing precision. 2) We find that better context representations with a decomposition of the background class can improve segmentation performance. 3) We propose CoLab, a flexible and generic method to automatically generate soft context labels. We validate CoLab with extensive experiments and find consistent improvements where the segmentation accuracy is en par and sometimes better compared to the case where human annotated context labels are available. # Related work {#sec:relatedwork} ## Class imbalance CoLab is related to the class imbalance problem as background commonly constitutes the majority class in image segmentation. However, methods to combat class imbalance mostly focus explicitly on improving the performance of the minority categories . Most approaches ignore the characteristics of majority background class as it is not contributing much to the common evaluation metrics of sensitivity, precision, and DSC of the ROI classes. CoLab focuses specifically on the representation of the background class and is complementary to methods tackling class imbalance such as loss reweighting strategies . Previous studies adopt coarse-to-fine strategies to reduce FP in segmentation tasks with class imbalance . However, ROI samples missed in the coarse stage cannot be recovered with later stages. In contrast, CoLab is trained in an end-to-end manner and can reduce varied kinds of FP. ## Multi-task learning CoLab, which is formulated as multi-label classification, can be seen as a form of multi-task learning (MTL). Current MTL methods train the model with different predefined tasks together with the main task using a shared feature representation . Previous works also attempted to incorporate spatial prior  or task prior  into model training with some predefined auxiliary tasks and optimization functions. In contrast, CoLab reformulates the main task by decomposing the background class with context labels and automatically generate the auxiliary task in a self-supervised manner. We argue that CoLab can have a direct impact on the main task by extending the label space. The main methodology of the CoLab strategy is inspired by some recently proposed methods which aim to generate the weights for pre-defined auxiliary tasks or labels through a similar meta-learning framework . In this study, CoLab is specifically designed for semantic segmentation with heterogeneous background classes which is a common scenario in medical imaging. # Context labels in image segmentation {#sec:analysis} ## Preliminaries We consider CNNs for multi-class segmentation with a total of \(c\) classes. Given a training dataset \(\mathcal{D} = \{ (\boldsymbol{x}_i, \boldsymbol{y}_i) \}_{i = 1}^N\) with \(N\) samples, where \(\boldsymbol{y}_i \in \rm I\!R^c\) is the one-hot encoded label for the central pixel in the image sample \(\boldsymbol{x}_i \in \rm I\!R^d\), such that \(\boldsymbol{1} \cdot \boldsymbol{y}_i = 1\) \(\forall\) \(i\). A segmentation model \(f_\phi(\cdot)\) learns class representations of the input sample \(\boldsymbol{x}_i\), noted as \(\boldsymbol{z}_i = f_\phi(\boldsymbol{x}_i) \in \rm I\!R^c\). We obtain the predicted probability \(\boldsymbol{p}_i\) that the real class of \(\boldsymbol{x}_i\) is \(j\) via a softmax function with \(p_{ij} = \mathrm{e}^{z_{ij}} / \sum_{j=1}^c \mathrm{e}^{z_{ij}}\). Typically the model is optimized by minimizing the empirical risk \(R_{\mathcal{L}_{seg}}(f_\phi)=\frac{1}{N}\sum_{i=1}^{N}\mathcal{L}_{seg}(f_\phi(\boldsymbol{x}_i),\boldsymbol{y}_i)\) computed on the training set. The segmentation loss \(\mathcal{L}_{seg}\) can be defined as the sum of losses \(\mathcal{L}\) over \(c\) classes: \[\begin{split} & \mathcal{L}_{seg}(f_\phi(\boldsymbol{x}_i),\boldsymbol{y}_i) = \sum_{j=1}^{c} \mathcal{L}(p_{ij}, y_{ij}) \\ & = \underbrace{\sum_{j=1}^{c-1}\mathcal{L}(p_{ij}, y_{ij})}_{\text{ROI classes}} + \underbrace{\mathcal{L}(p_{ic}, y_{ic})}_{\text{background class}}, \label{eq:CE} \end{split}\] where \(\mathcal{L}\) is a criterion for a specific class, such as cross entropy (CE) or soft DSC . Here, we further decompose \(\mathcal{L}_{seg}\) into two terms, including an ROI loss (computed on \(c-1\) foreground classes) and a background loss (computed on the background class, only). We aim to improve segmentation performance by augmenting the background class with auxiliary context classes. Specifically, we propose to utilize context labels assigned to different and decomposed background regions. In order to divide the background class into \(t>1\) classes, we create another model \(\tilde{f}_{\theta}(\cdot)\) with \(\tilde{\boldsymbol{z}}_i = \tilde{f}_{\theta}(\boldsymbol{x}_i) \in \rm I\!R^{c+t-1}\) and predicted probability \(\tilde{\boldsymbol{p}}_i\). We also require an additional one-hot label \(\tilde{\boldsymbol{y}}_{i} \in \rm I\!R^{c+t-1}\), where we require \(\tilde{y}_{ij} = y_{ij}\) \(\forall\) \(i\), \(j \in \{1, \cdots, c-1\}\) and \(\sum_{j=c}^{c+t-1}\tilde{y}_{ij} = y_{ic}\) \(\forall\) \(i\). With this notion, the segmentation loss can be written as: \[\begin{split} & \mathcal{L}_{seg}(\tilde{f}_\theta(\boldsymbol{x}_i),\tilde{\boldsymbol{y}}_{i}) = \sum_{j=1}^{c+t-1} \mathcal{L}(\tilde{p}_{ij}, \tilde{y}_{ij}) \\ & = \underbrace{\sum_{j=1}^{c-1}\mathcal{L}(\tilde{p}_{ij}, \tilde{y}_{ij})}_{\text{ROI classes}} + \underbrace{\sum_{j=c}^{c+t-1}\mathcal{L}(\tilde{p}_{ij}, \tilde{y}_{ij})}_{\text{background classes}}, \label{eq:context} \end{split}\] In the following method section, we always consider a simplified but common case where we only have one ROI class (\(c=2\)) for simplicity. In other words, we only consider binary ROI segmentation with \(f_\phi(\boldsymbol{x}_i) \in \rm I\!R^2\), although the method can be naturally extended to multi-class ROI segmentation. With this assumption, Eq. [\[eq:context\]](#eq:context){reference-type="ref" reference="eq:context"} can be simplified as: \[\mathcal{L}_{seg}(\tilde{f}_\theta(\boldsymbol{x}_i),\tilde{\boldsymbol{y}}_{i}) = \underbrace{\mathcal{L}(\tilde{p}_{i1}, \tilde{y}_{i1})}_{\text{ROI class}} + \underbrace{\sum_{j=2}^{t+1}\mathcal{L}(\tilde{p}_{ij}, \tilde{y}_{ij})}_{\text{background classes}} \label{eq:simplifiedcontext}\] ## False positives on background class To better understand the effect of the background class on the model learning, we train multiple CNNs on segmentation datasets which contain heterogeneous background. We conduct experiments on challenging tasks including liver tumor segmentation in computed tomography (CT) images , kidney tumor segmentation in CT images , colon tumor segmentation in CT images , vestibular schwannoma (VS) segmentation in T2-weighted magnetic resonance (MR) images , brain stroke lesion segmentation in T1-weighted MR images  and pancreatic tumor segmentation in CT images . We adopt a well-configured 3D U-Net  as the segmentation model for all the experiments, which has been demonstrated to yield competitive results across different medical image segmentation tasks. The detailed data and network configurations are summarized in Section [5](#sec:experiments){reference-type="ref" reference="sec:experiments"}. Visualizations of the segmentation results are shown in Fig. [\[fig_segexample\]](#fig_segexample){reference-type="ref" reference="fig_segexample"}(a). When trained with binary segmentation tasks and without context labels, the models are prone to over-segment the ROIs with many FP. Specifically, the model trained on the binary liver tumor task predicts other organs outside liver as liver tumor; the model trained on the binary kidney tumor task predicts parts of the healthy kidney regions as kidney tumor; while the model trained on the binary brain tumor task predicts the surrounding brain tissue as brain tumor; the model trained on the binary brain lesion task predicts unrelated healthy brain regions as brain lesion. ## Underfitting of background samples To study the model behavior when trained with heterogeneous background classes, we can monitor the logit distribution of samples from ROI and background classes for the test data. Our observations for liver tumor, kidney tumor, and brain tumor segmentation are summarized in Fig. [\[fig_logitmap\]](#fig_logitmap){reference-type="ref" reference="fig_logitmap"}(a, d, g). We find that the CNN models map the ROI samples to a compact cluster in the logit space while background samples form a more dispersed distribution. This indicates that the model cannot easily map all background samples to a single cluster representing the background class. Although the model seems to separate ROI from background samples in the feature space, it builds complex background representations and unable to capture an accurate decision boundary between ROI and immediate context. A possible reason is that the CNN uses most of its capacity to extracted the common features among background regions with different characteristics. Unfortunately, these shared features are not very discriminate. Specifically, we observe that the logit distribution of background samples overlaps with the learned decision boundary. This is the reason why the model predicts many FP leading to over-segmentation of ROI structures when the background class is heterogeneous. We hypothesise that the width of the distribution serves as an indicator of the heterogeneity of a specific class and is a sign of the difficulty during learning. As a result of sample heterogeneity, a CNN may struggle to reduce intra-class variation of the background class and underfit the background samples, failing to recognize the background samples that share similar characteristics with ROI samples. It should be noted that we do not observe significant difference between the logit distribution of training background samples and test background samples, as also shown in previous studies , indicating that the logit shift of background samples is indeed due to underfitting instead of overfitting. ## The effect of context labels The availability of context labels greatly helps with the ROI segmentation, which provide additional signals to CNN to fit the training data of heterogeneous background samples. We confirm this empirically by including human-defined context labels in the model training. Specifically, we adopt liver masks (\(t=2\)) for liver tumor segmentation, kidney masks (\(t=2\)) for kidney tumor segmentation, brain tissue masks including ventricles, deep grey matter, cortical grey matter, white matter and other tissues (\(t=6\)) for brain tumor segmentation. The kidney and liver masks are manually annotated, while the brain tissue masks are generated automatically using the paired T1-weighted MR images and a multi-atlas label propagation with expectation--maximisation based refinement (MALP-EM) . In order to make a fair comparison, we make sure that all the experiments share the same training schedule except the label space. We sample the training patches only considering ROIs. Specifically, we make sure 50% training patches to contain ROIs and sample the other half of the training patches uniformly. The observations on test and training set are summarized in Fig. [\[fig_segperformance\]](#fig_segperformance){reference-type="ref" reference="fig_segperformance"}. We find that the models yield overall better performance when trained with anatomy masks, as indicated by improved DSC (defined as \(DSC=2\frac{\text{sensitivity} \cdot \text{precision}}{\text{sensitivity} + \text{precision}}\)). Furthermore, we observe that the model trained with context labels yields higher precision while preserving similar sensitivity. The observation is consistent with the visualization results in Fig. [\[fig_segexample\]](#fig_segexample){reference-type="ref" reference="fig_segexample"}(b), where we find the models trained with human-defined context labels reduce FP. Similarly, we visualize the corresponding network behaviour in Fig. [\[fig_logitmap\]](#fig_logitmap){reference-type="ref" reference="fig_logitmap"}(b, e, h). As we only consider the segmentation performance of the ROI (class 1), we visualize the logits in the plane of (\(z_1\), max(\(z_2\),\..., \(z_{t+1}\))). We observe that the model trained with anatomy masks maps the background samples to a narrower distribution and reduces the background logits shift across the decision boundary. This indicates that the models fit the training data better with the help of context labels. Instead of building generic filters for all the background samples, the CNNs can dedicate specific filters to model a more homogeneous subparts of the background samples that share common characteristics. The models are faced with a simplified segmentation task with homogeneous background subclasses yielding better overall performance with the same model capacity. Although anatomy masks are found to be effective context labels, they are not always available in real-world applications. Specifically acquiring manually annotated context labels is time-consuming and would require significant efforts from human experts to generate annotations at large scale. Therefore, we propose CoLab which can automatically discover specific soft context labels using a meta-learning strategy. CoLab benefits the segmentation model training by making it to fit the background samples better, achieving comparable or even better performance when compared with models trained on manually defined anatomy masks. # CoLab {#sec:method} ## Overview {#sec:method-overview} Now we consider CNNs for binary semantic segmentation. Typically, we are given a baseline segmentation model \(f_\phi(\cdot)\) that maps the input image \(\boldsymbol{x}_i\) to the label space \(f_\phi(\boldsymbol{x}_i) \in \rm I\!R^2\). In order to fit the label space with context labels, we first extend the classification layer of \(f_\phi(\cdot)\) with additional \(t-1\) output neurons and obtain \(\tilde{f}_{\theta}(\cdot)\) which map \(\boldsymbol{x}_i\) to \(\tilde{f}_{\theta}(\boldsymbol{x}_i) \in \rm I\!R^{t+1}\). We employ another model \(g_\omega(\cdot)\) as the task generator parameterized by \(\omega\) to produce context labels, with the network output \(\boldsymbol{o}_i\) = \(g_\omega(\boldsymbol{x}_i) \in \rm I\!R^t\). We have no requirements for the backbone of \(g_\omega(\cdot)\) and empirically keep it the same with that of \(f_\phi(\cdot)\). We illustrate the training process of the proposed CoLab in Fig. [\[fig_method\]](#fig_method){reference-type="ref" reference="fig_method"}. At the beginning of each iteration, we obtain the context labels termed as distance constrained label \(\hat{\tilde{\boldsymbol{y}}}_i\) by taking use of \(\boldsymbol{o}_i\) and the ground truth \(\boldsymbol{y}\). This process is marked with 1⃝ and will be illustrated in Section [4.2](#sec:method-labelagg){reference-type="ref" reference="sec:method-labelagg"} and [4.3](#sec:method-contcons){reference-type="ref" reference="sec:method-contcons"}. Then, we calculate a new \(\theta^*\) with one step of gradient descent with 2⃝ to access the impact of \(\hat{\tilde{\boldsymbol{y}}}_i\) on model training of ROI segmentation. Next, we optimize \(\omega\) based on the second-order derivatives through a meta learning scheme with 3⃝, which would be demonstrated in Section [4.4](#sec:meta-learning){reference-type="ref" reference="sec:meta-learning"}. Finally, we optimize the segmentation model \(\tilde{f}_{\theta}(\cdot)\) based on the updated context labels with 4⃝. ## Label aggregation {#sec:method-labelagg} We calculate the context probability \((q_{ij})_{j=1}^{t}\) based on \(\boldsymbol{o}_i\) via the softmax function as: \[q_{ij} = \frac{\mathrm{e}^{o_{ij}}}{\sum_{j=1}^{t}\mathrm{e}^{o_{ij}}}. \label{eq:generatedclass}\] The extended label \(\tilde{\boldsymbol{y}}_i \in \rm I\!R^{t+1}\) is calculated by aggregating the original label \(\boldsymbol{y}_i = (y_{i1}, y_{i2})\) and the context probability \(\boldsymbol{q}_i\) with: \[\begin{aligned} \tilde{y}_{ij}=\left\{ \begin{array}{rcl} y_{i1} & & \text{if} \quad j = 1,\\ q_{ij} & & \text{if} \quad j > 1 \quad \text{and} \quad y_{i1}=0,\\ 0 & & \text{otherwise.} \end{array} \right. \label{eq:aggfunc} \end{aligned}\] The label aggregation process is also illustrated in Fig. [\[fig_labelagg\]](#fig_labelagg){reference-type="ref" reference="fig_labelagg"}. By doing so, we can decompose the background class \(y_{i2}\) into \(t\) subclasses while ensure that the \(\tilde{\boldsymbol{y}}_i\) contains sufficient information about the ROI segmentation. For multi-class segmentation with totally \(c\) classes, \(\tilde{\boldsymbol{y}}_i\) can be calculated as: \[\begin{aligned} \tilde{y}_{ij}=\left\{ \begin{array}{rcl} y_{ij} & & \text{if} \quad j < c,\\ q_{ij} & & \text{if} \quad j \geq c \quad \text{and} \quad y_{ij}=0 \quad \forall j<c, \\ 0 & & \text{otherwise.} \end{array} \right. \label{eq:aggfunc2} \end{aligned}\] ## Context constraints {#sec:method-contcons} Compared with background samples which are further away, the ones closer to the ROIs share similar characteristics with ROI samples and are more likely to be misclassified. In order to make the segmentation model focus more on those hard background samples that are close to ROIs, we make an assumption that all samples which are distant from the ROI need less attention and should be safely assigned the same background label. Specifically, we create a hard label \(\boldsymbol{b}_i \in \rm I\!R^{t+1}\) to represent the the background samples that are far from the ROI: \[b_{ij}=\left\{ \begin{array}{rcl} 1 & & \text{if} \quad j = 2,\\ 0 & & \text{otherwise}. \end{array} \right. \label{eq:bglabel}\] By utilizing the ROI label \(\boldsymbol{y}_i\), we calculate the corresponding distance map \(d_i\) which is the Euclidean distance of a pixel to the closest boundary point of the ROI of any class. We set \(d_i\) to be zero for pixels inside the ROI. We then calculate a soft dilated mask \(\mathcal{M}_i\) based on \(d_i\): \[\mathcal{M}_i =\left\{ \begin{array}{rcl} 1 & & \text{if} \quad d_i < m,\\ \mathrm{e}^{\frac{-d_{i}+m}{\tau}} & & \text{otherwise,} \end{array} \right. \label{eq:lossmask}\] where \(m\) is the margin controlling the model's focus on the pixels which are neighbouring to ROI and \(\tau\) is the temperature to control the probabilities of the dilated regions. Empirically, \(m\) is set as 30 and \(\tau\) is set as 20 for all our experiments. \(\mathcal{M}_i\) would be 1 for pixels around ROIs and decrease close to 0 for pixels far from ROIs. We visualize an example of \(\mathcal{M}_i\) with liver tumor in Fig. [\[fig_lossmask\]](#fig_lossmask){reference-type="ref" reference="fig_lossmask"}. The distance constrained label then is calculated as: \[\hat{\tilde{\boldsymbol{y}}}_i=\mathcal{M}_i\tilde{\boldsymbol{y}}_i+(1-\mathcal{M}_i)\boldsymbol{b}_i. \label{eq:masky}\] In this way, only the regions neighbouring to the tumor are considered to be classified as the contextual background class. Specifically, both \(\tilde{f}_{\theta}(\cdot)\) and \(g_\omega(\cdot)\) would be trained to focus on the regions which are close to ROIs. ## Task generator optimization with meta-gradients {#sec:meta-learning} We formulate the optimization of CoLab as a bi-level problem: \[\begin{aligned} &\min_{\omega}\frac{1}{N}\sum_{i=1}^{N} \mathcal{L}_{ROI}(\tilde{f}_{\theta^*}(\boldsymbol{x}_i), \boldsymbol{y}_i) \label{eqn:objectivefunA} \\ &s.t.\quad \theta^* = \argmin_{\theta}\frac{1}{N} \sum_{i=1}^{N}\mathcal{L}_{seg}(\tilde{f}_{\theta}(\boldsymbol{x}_i), \hat{\tilde{\boldsymbol{y}}}_i)), \label{eq:objectivefunB} \end{aligned}\] where \(\mathcal{L}_{ROI}(\tilde{f}_{\theta}(\boldsymbol{x}_i), \boldsymbol{y}_i) = \mathcal{L}_c(\tilde{p}_{i1}, y_{i1})\) is computed on the ROI class. Different from \(\mathcal{L}\), we choose \(\mathcal{L}_c\) to be a criteria that can be calculated in the one-versus-all manner, such as binary cross entropy (BCE) and soft DSC loss, to represent the binary segmentation performance of the ROI class with more than two output logits. We train the model using a batch of training samples with batch size \(n\). For simplicity, we shorten \(\frac{1}{n}\sum_{i=1}^{n} \mathcal{L}_{ROI}(\tilde{f}_{\theta^*}(\boldsymbol{x}_i), \boldsymbol{y}_i)\) as \(\mathcal{L}_{ROI}(\theta^*)\) and \(\frac{1}{n}\sum_{i=1}^{n}\mathcal{L}_{seg}(\tilde{f}_{\theta}(\boldsymbol{x}_i), \hat{\tilde{\boldsymbol{y}}}_i)\) as \(\mathcal{L}_{seg}(\theta, \omega)\) in the following descriptions. The bi-level optimization problem defined in Eq. [\[eqn:objectivefunA\]](#eqn:objectivefunA){reference-type="ref" reference="eqn:objectivefunA"} and [\[eq:objectivefunB\]](#eq:objectivefunB){reference-type="ref" reference="eq:objectivefunB"} can be solved with gradient descent . Specifically, the derivative of \(\mathcal{L}_{ROI}(\theta^*)\) w.r.t. \(\omega\) can be calculated by applying the chain rule: \[\nabla_{\omega}\mathcal{L}_{ROI}(\theta^*) = (\frac{\partial \theta^*}{\partial \omega})^\intercal \nabla_{\theta}\mathcal{L}_{ROI}(\theta^*), \label{eq:funcchain}\] where \(\nabla_{\omega} = (\frac{\partial}{\partial \omega})^\intercal\) and \(\nabla_{\theta} = (\frac{\partial}{\partial \theta})^\intercal\). One can compute \(\frac{\partial \theta^*}{\partial \omega}\) based on implicit function theorem . However, the derived result would contain a Hessian which is computational expensive and not always possible to access for deep neural networks. Among the many heuristics for the gradient approximation , we follow the solution described in  to approximate \(\theta^*\) by a single optimization step. Specifically, we sample a batch of training data and approximate the optimal inner variable \(\theta^*\) with a step of gradient decent: \[\theta^* \approx \theta-\alpha \nabla_\theta \mathcal{L}_{seg}(\theta, \omega), \label{eq:innerupdate}\] where \(\alpha\) is step size, which is kept the same with the learning rate of \(\theta\). We differentiate this equation w.r.t. \(\omega\) from both sides yielding: \[\frac{\partial \theta^*}{\partial \omega} =-\alpha\nabla^2_{\theta,\omega}\mathcal{L}_{seg}(\theta^*, \omega), \label{eq:omegasimple}\] where \(\nabla^2_{\theta,\omega} = \frac{\partial \nabla_{\theta}}{ \partial \omega}\). By substituting Eq. [\[eq:omegasimple\]](#eq:omegasimple){reference-type="ref" reference="eq:omegasimple"} into Eq. [\[eq:funcchain\]](#eq:funcchain){reference-type="ref" reference="eq:funcchain"}, we can obtain the gradient on \(\omega\) and update it with: \[\begin{aligned} & \omega^{t+1} = \omega^t-\beta \nabla_{\omega^t}\mathcal{L}_{ROI}(\theta^*) \\ & = \omega^t + \alpha\beta\nabla^2_{\omega^t,\theta}\mathcal{L}_{seg}(\theta^*, \omega^t) \nabla_{\theta}\mathcal{L}_{ROI}(\theta^*), \label{eq:func3s} \end{aligned}\] where \(\beta\) is the learning rate to update \(\omega\). In this way, the task generator \(g_{\omega}\) is explicitly trained to produce effective context labels with the second-order gradients. Although Eq. [\[eq:func3s\]](#eq:func3s){reference-type="ref" reference="eq:func3s"} contains an expensive vector-matrix product, it is feasible to calculate with prevailing machine learning frameworks such as PyTorch . We find CoLab can be efficient as it costs as little as 30% additional training time. We summarize the implementation details in supplementary material. The full algorithm is summarized in Algorithm [\[alg:CoLab\]](#alg:CoLab){reference-type="ref" reference="alg:CoLab"}. ::: ## Additional visualization of context labels We provide additional examples of generated tasks in Fig. [\[fig_contexttask2\]](#fig_contexttask2){reference-type="ref" reference="fig_contexttask2"}, which is a supplement to Fig. [\[fig_contexttask\]](#fig_contexttask){reference-type="ref" reference="fig_contexttask"}. ## Intensity histograms of ROIs and context labels We analyze the intensity of context labels of different datasets and summarize the intensity histograms in right part of Fig. [\[fig_hist\]](#fig_hist){reference-type="ref" reference="fig_hist"}. K-means based context labels are generated based on pixel intensity, therefore the histograms of different classes generally distribute in separate regions. Different classes of dilated masks based context labels always share similar intensity distributions, as dilated masks only takes spatial information into account and does not contain much semantic meanings. CoLab learns to generate context labels within the dilated regions based on meta-learning. Similar to anatomy masks, CoLab would highlight related semantic regions which similar to ROIs but differ from other background samples. We should note that here we only analyze the context labels learned in the final models for CoLab. As the task generator would also be optimized during the training process, it is hard to interpret how CoLab help the segmentation model learn better. We also look into the intensity histograms of ROIs in left part of Fig. [\[fig_hist\]](#fig_hist){reference-type="ref" reference="fig_hist"}. We find the intensity histograms of ROIs always largely overlap with that of background samples. The model might require external spatial information to enhance the prior knowledge of ROI location. \(t=2\) could make CoLab highlight the surrounding regions which can help the model reduce false positive predictions on samples with similar intensity but distant from ROIs. In this case, the model cannot benefit from more detailed class representation with \(t>2\) which would further focus on decomposing the samples around ROIs. On the contrary, we notice only the intensity distributions of brain tumor would differentiate from that of background samples. In order to further improve the segmentation accuracy, the model might require more specific subregion information to discriminate between ROIs and the regions with similar intensity. This can be a potential factor due to which only the optimal context class \(t\) of brain tumor segmentation is 4 instead of 2. Practitioners could utilize the intensity histograms of ROI to choose the optimal \(t\). According to these findings, we suggest to choose \(t = 2\) for ROIs having similar intensity with the background samples, while choose \(t > 2\) for ROIs with distinct intensity distributions from the background samples.

{'timestamp': '2022-12-19T02:10:47', 'yymm': '2212', 'arxiv_id': '2212.08423', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08423'}

# Perfect Fluids from a variational principle The usual relativistic description of perfect fluids consists in defining the energy momentum tensor \[T_{\mu\nu}^{\textrm{fluid}}=(\rho+p)U_{\mu}U_{\nu}+p \eta_{\mu\nu}, \label{energy-momentum}\] along with the conservation equation \[T^{\mu\nu}_{\textrm{fluid},\nu}=0. \label{conserved}\] In Eq. ([\[energy-momentum\]](#energy-momentum){reference-type="ref" reference="energy-momentum"}), \(\rho\) is the density of total mass-energy, \(p\) is the fluid pressure, \(\eta_{\mu\nu}\) is the Minkowksi metric and \(U^{\mu}(x)\) is the four-velocity field of the fluid. Eq. ([\[conserved\]](#conserved){reference-type="ref" reference="conserved"}) leads to the dynamical equations of motion of the fluid. In a seminal paper , Schutz found an equivalent set of equations for the fluid via a variational principle through the action \[S_{\textrm{fluid}}=\int{pd^4x}. \label{action}\] In Schutz' formalism, the four-velocity is decomposed in terms of six potentials \[U_{\nu}=\mu^{-1}(\phi_{,\nu}+\alpha\beta_{,\nu}+\theta s_{,\nu}), \label{four-velocity}\] where the field \(\mu\) is the specific enthalpy while \(s\) is the specific entropy. Schutz also assumed an equation of state of the form \(p=p(\mu,s)\) together with the first law of thermodynamics in the form \[dp=\rho_0 d\mu-\rho_0 Tds, \label{first law}\] where \(\rho_0\) is the rest mass density of the fluid and \(T\) is the temperature of the fluid. The enthalpy \(\mu=(\rho+p)/\rho_0\) is not an independent field, since the normalization equation \(U^{\nu}U_{\nu}=-1\) implies \[\mu^2=-\eta^{\nu\lambda}(\phi_{,\nu}+\alpha\beta_{,\nu}+\theta s_{,\nu})(\phi_{,\lambda}+\alpha\beta_{,\lambda}+\theta s_{,\lambda}). \label{normalization}\] The remaining five potentials in Eq. ([\[four-velocity\]](#four-velocity){reference-type="ref" reference="four-velocity"}) have their own evolution equations. The variation of the action given by Eq. ([\[action\]](#action){reference-type="ref" reference="action"}) with respect to \(\phi\), \(\theta\), \(s\), \(\alpha\) and \(\beta\) gives the following evolution equations (Eqs. ([\[first law\]](#first law){reference-type="ref" reference="first law"}) and ([\[normalization\]](#normalization){reference-type="ref" reference="normalization"}) are extensively used in this derivation-see Ref.  for details) \[\left.\begin{array}{ccc} \delta S_{\textrm{fluid}}/\delta \phi = 0 &\Rightarrow & (\rho_0 U^{\nu})_{,\nu}=0, \\ \delta S_{\textrm{fluid}}/\delta s = 0 &\Rightarrow & \ \ \ \ U^{\nu}\theta_{,\nu}=T,\\ \delta S_{\textrm{fluid}}/\delta \beta = 0 &\Rightarrow & U^{\nu}\alpha_{,\nu}=0,\\ \delta S_{\textrm{fluid}}/\delta \theta = 0 &\Rightarrow & \ \ \ \ U^{\nu}s_{,\nu}=0,\\ \delta S_{\textrm{fluid}}/\delta \alpha = 0 &\Rightarrow & U^{\nu}\beta_{,\nu}=0. \label{equations} \end{array}\right.\] Notice that the first and the fourth equations in ([\[equations\]](#equations){reference-type="ref" reference="equations"}) imply that mass is conserved and that the fluid is isentropic, respectively. Finally, the normalization condition and the last two equations in ([\[equations\]](#equations){reference-type="ref" reference="equations"}) imply the evolution equation for the remaining potential \(\phi\), \[U^{\nu}\phi_{,\nu}=-\mu. \label{eq phi}\] Clearly, Eqs. ([\[equations\]](#equations){reference-type="ref" reference="equations"}) and ([\[eq phi\]](#eq phi){reference-type="ref" reference="eq phi"}) only make sense if they correspond to the correct equations of motion for the fluid, i.e., if they are equivalent to Eq. ([\[conserved\]](#conserved){reference-type="ref" reference="conserved"}). In fact, Schutz showed that this is true. In this way, the action ([\[action\]](#action){reference-type="ref" reference="action"}), along with the velocity decomposition given by Eq. ([\[four-velocity\]](#four-velocity){reference-type="ref" reference="four-velocity"}) and the first law of thermodynamics ([\[first law\]](#first law){reference-type="ref" reference="first law"}) give the correct prescription for a perfect fluid. # external force from an external potential? A perfect fluid subjected to an external applied force satisfies the following equation, \[\label{forcex} T^{\mu\nu}_{\phantom{\mu\nu};\nu}=\mathcal{F}^{\mu},\] where \(\mathcal{F}^{\mu}\) is the associated four-force density. Two important questions regarding the source \(\mathcal{F}^{\mu}\) arise, namely, what is the origin of \(\mathcal{F}^{\mu}\) and how do we control this external force? In Ref. , \(\mathcal{F}^{\mu}\) is considered as the four-gradient of an external pressure \(p^{\text{ext}}\), i.e., \[T^{\mu\nu}_{\phantom{\mu\nu};\nu}=p^{\textrm{ext},\mu}. \label{external pressure}\] The function \(p^{\textrm{ext}}\) could, in principle, be arbitrarily given by an external agent. This is motivated by the least action principle for a perfect fluid through the addition of an extra potential \(U(x)\) to the fluid Lagrangian density. This potential is assumed to be independent of any dynamical field. The action in this case becomes \[S_{\textrm{modified}}=\int{\left[p(\phi,\alpha,\beta,\theta,s,g^{\mu\nu})-U(x)\right]\sqrt{-g}d^4x}, \label{modified}\] where we emphasize the \(p\) dependence on the independent dynamical fields. The external potential plays the role of (minus) an external pressure \(p^{\textrm{ext}}=-U\). Then, it is argued that this external pressure plays the role of an external force on the background fluid. In fact, the potential \(U(x)\) modifies the stress-energy tensor. A straightforward calculation (see Ref.  for the details) shows that the variation of the action in Eq. ([\[modified\]](#modified){reference-type="ref" reference="modified"}) gives the modified stress-energy tensor \[T_{\mu\nu}=T_{\mu\nu}^{\textrm{fluid}}+p^{\textrm{ext}}g_{\mu\nu}, \label{new stress}\] where \(T_{\mu\nu}^{\textrm{fluid}}\) is given by Eq. ([\[energy-momentum\]](#energy-momentum){reference-type="ref" reference="energy-momentum"}). However, since \(p^{\textrm{ext}}=-U(x)\) does not depend on the dynamical variables, it does not modify the evolution equations in Eq. ([\[equations\]](#equations){reference-type="ref" reference="equations"}). These equations (as shown in Ref. ) are equivalent to the usual equation of motion \(T^{\nu\lambda}_{\textrm{\tiny{fluid}};\lambda}=0\). Hence, the stress-energy tensor in Eq. ([\[new stress\]](#new stress){reference-type="ref" reference="new stress"}) is not conserved, unless \(p^{\textrm{ext}}\) is constant. This leads to the energy-momentum tensor of two noninteracting fluids, one of them with equation of state \(p=-\rho\), which represents a cosmological constant. In order to have an effective interaction, in Eq. ([\[modified\]](#modified){reference-type="ref" reference="modified"}) we should add a potential of the form \(U=U(\phi,\alpha,\beta,\theta,s,g^{\mu\nu},t)\). But this would dramatically change the Euler-Lagrange equations. Hence, we conclude that Eq. (4) in Ref. , which states that \[T^{\mu\nu}_{\phantom{\mu\nu};\nu}=0\Leftrightarrow T^{\mu\nu}_{\textrm{fluid};\nu}=p^{\textrm{ext},\mu} \label{incorrect}\] cannot be obtained from a variational principle unless \(p^{\textrm{ext},\mu}=0\). Consequently, both questions in the beginning of this section remain unanswered. # External force from Newton's second law But how can we model an external force acting on a perfect fluid in a consistent way? To answer this question, we go back to the well-known subject of classical mechanics of particles and look at the generalization of Newton's second law for a relativistic particle . In inertial coordinates with metric \[ds^2=-dt^2+d\vec{x}^2,\] we have \[F^{\alpha}=\frac{d}{d\tau}(m_0 V^{\alpha}), \label{newton}\] with \(V^{\alpha}=\gamma(1,\vec{v})\) being the four-velocity of the particle, \(\gamma=1/\sqrt{1-\vec{v}^2}\) and \(m_0\) the particle's rest mass. The left hand side of Eq. ([\[newton\]](#newton){reference-type="ref" reference="newton"}) is the external force we actually control, while the right hand side is the particle response to this force. As in, from now on, we assume that there is no back-reaction, i.e. the individual particles can be considered as test particles so that their dynamics do not affect the strong potential responsible for \(F^{\alpha}\). Let us also suppose that we are dealing with an ordinary body force which does not change the particle's internal state (this is equivalent to \(m_0=\textrm{constant}\)). This implies (with the help of the normalization condition \(V^{\alpha}V_{\alpha}=-1\)) \[V^{\alpha}F_{\alpha}=0 \label{good force}\] so that \[F^{\alpha}=\gamma(\vec{F}\cdot \vec{v},\vec{F}), \label{four force}\] with \(\vec{F}=\frac{d\vec{p}}{dt}\) and \(\vec{p}=m_0\gamma \vec{v}\). In this way, the four-force density on a proper volume \(\delta V_0\) of the fluid is given by \[\mathcal{F}^{\alpha}=NF^{\alpha}/\delta V_0,\] where \(N\) is the number of fluid elements in the volume \(\delta V_0\). If we define the three-force density as \[\vec{\mathcal{F}}= N\vec{F}/\delta V, \label{force density}\] where \(\delta V\) is a volume element on an arbitrary frame, we have \[\mathcal{F}^{\alpha}=\gamma(\delta V/\delta V_0)(\vec{\mathcal{F}}\cdot \vec{v},\vec{\mathcal{F}})=(\vec{\mathcal{F}}\cdot \vec{v},\vec{\mathcal{F}}),\] where we used the relativistic relation \(\delta V_0=\gamma \delta V\). Notice that, although we can control the vector

{'timestamp': '2022-12-19T02:10:51', 'yymm': '2212', 'arxiv_id': '2212.08431', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08431'}

# Introduction {#intro} Transformer-based models, the current go-to models in terms of state-of-the-art performance, achieve impressive performances in most natural language processing (NLP) tasks. Those parameter-heavy models are trained on immense corpora and retain most of the training data information. Retaining significant knowledge from training data can lead to performance improvements that rely on unknown and even undesirable similarities between training and test data. For example, gender bias can arise when training on historical data. Using this hidden feature, a model can associate gender and occupation together and use this knowledge for prediction. This problem would go undetected by traditional metrics such as accuracy and recall since the evaluation subset also contains this bias. However, the model will exploit the gender feature to improve its performance instead of learning useful text representations that would help infer someone's occupation. Even though unwanted similarities may improve performance, it creates problems down the value chain if the end-user is unaware of this bias. A lesser-known source of bias is the difference of sequence length between observations of the two classes of a binary classification task. Even though sequence length appears to be a legitimate feature to learn and exploit, it implies that any variation of the length distribution of observations, regardless of their content, may result in classification errors. This bias also impacts the performance of explanation mechanisms, such as LIME, as the main exploited feature is not based on textual content. Another problem with sequence length learning is an artificial overperformance to the presence of this bias in both training and test datasets. This performance distortion causes unfair model comparisons, as some models may converge to a local minimum using sequence length rather than exploiting the model's ability to fully represent textual information. Overall sequence length learning should be avoided because the potential gain in performance comes at the expense of the robustness and explainability of models. Although seldom encountered in public and open datasets, this problem may be more common in private datasets, such as in medicine, where extreme and rare cases are associated with longer text descriptions than milder cases. For instance, treating a rare illness will require more appointments and follow-ups than a common cold. In that sense, models trained to detect rare events will underperform in tasks such as early detection if they use sequence length as a meta-feature. We encounter this problem in our work, as we aim to extract risk factors from catastrophic insurance claims to provide early warnings to analysts. However, basic and catastrophic claim files having different length profiles, our modeling efforts are hindered by the sequence length learning problem. In this paper, we present two contributions associated with the problem mentioned above. The first contribution is an empirical demonstration that transformer-based models are affected by sequence length learning. Our second contribution is the evaluation of techniques using the pretrained transformers' capacities to mitigate this bias. The remainder of the paper is as follows. We present in Section [2](#sec:related_work){reference-type="ref" reference="sec:related_work"} how this problem is addressed in the literature. In Section [3](#sec:evaluation_methodology){reference-type="ref" reference="sec:evaluation_methodology"}, we present the methodology we adopt to inject the sequence length meta-feature in two textual datasets and we illustrate that transformer-based binary classifiers are frail to the problem. Finally, we analyze in Section [4](#sec:simple_technique){reference-type="ref" reference="sec:simple_technique"} data engineering techniques to alleviate the impact of sequence length meta-feature learning. # Related Work {#sec:related_work} The Natural Language Processing research community is interested in fairness regarding many bias sources such as gender or ethnicity (see ). Approaches, such as data augmentation, are presented by and to mitigate (gender) bias. The lesser-known sequence length meta-feature has seldom been explored in the literature. In this paper, we are extending the work of to transformer models, where the sequence length problem in a classification task was detected for Recurrent Neural Networks. This problem was also investigated as a generalization problem over out-of-distribution (OOD) examples in sequence-to-sequence models (see and ). In our work, we show that transformer-based classifiers generalize very well to OOD examples and thus cannot be addressed as such. # Sequence Length Meta-Feature Injection and Impact Evaluation {#sec:evaluation_methodology} This section describes the methodology we use to expose models to the sequence length problem using two sentiment analysis datasets. Our approach uses a truncated training set to artificially inject the sequence length meta-feature in the models and modified test sets to evaluate the behavior of the models for different ranges of sequence lengths. ## Datasets Used In our experiment, we apply our methodology to two text classification datasets. The first dataset, introduced by, is **Amazon-Polarity** (AP). This dataset contains five ratings grouped into negative (label 0) and positive (label 1) classes. The training set consists of 3.6M examples equally divided into two classes. The second dataset is **Yelp-Polarity** (YP), also introduced by. YP is another binary classification dataset containing 500K examples carrying either a negative (0) and positive (1) label for each textual observation. ## Evaluation Steps We begin by creating a sequence length imbalance into the datasets. First we identify the classes containing, on average, the shorter examples and the classes with the longest examples. We then remove from the training set the longest examples that belong to the shortest class and the shortest ones belonging to the longest class. We train models on the altered datasets and evaluate them on 3 test subsets to assess their frailty to the sequence length learning problem. ### Identifying Short and Long Classes The first step is to assign the class labels to the short and long classes. From the sequence length of each observation, we compute the average length of both classes and identify the long class as the class with the highest average (the short class being the class label with the lowest average). As we use a transformer with a specific pretrained tokenizer in our experiment, we estimate the sequence length as the number of tokens after tokenization of the dataset. We also assess the extent of overlap of the class sequence length distributions. As discussed by, overlapping between classes offers a simple and efficient estimator of how present the problem is in a dataset. The more the distributions overlap, the lesser the problem. Finally, we calculate a sequence length from the training distribution that yields the best F1 score for the a classifier using the sequence length of observation as its only feature. This value allows us to separates the test dataset into two partitions. Table [1](#tab:baseline_1){reference-type="ref" reference="tab:baseline_1"} and [2](#tab:baseline_2){reference-type="ref" reference="tab:baseline_2"} contains these values for the two datasets we use in this paper. We find that the negative class examples tend to be longer in both datasets, with a median of 88 tokens for AP and 145 for YP. The positive observations of AP of YP have respectively a median length of 77 and 111 tokens. Both datasets have a healthy overlap percentage of 92.4 % (AP) and 88 % (YP). These overlap values suggest that sequence length learning should not be present based on the red flag provided by. Finally, we get a maximal F1 score of 53.4% if we infer that AP's examples below and above a length of 92 tokens respectively belong to the positive and negative classes. Similarly, we get a maximum F1 score of 55.9 % by inferring that YP's examples above and below 127 tokens belong to the positive and negative classes. Again, as both score F1 scores are very close to 50 %, we hardly suspect the presence of the problem. ### Training Set Alteration to Inject the Sequence Length meta-Feature {#ss:alter_training_Set} To inject the sequence length meta-feature in the datasets, we remove examples above and below a certain sequence length to create a discrepancy between the two classes. In our experiment, we omit examples from the long class below the lower limit and drop examples from the short class above the upper threshold. The selection of the lower and upper thresholds allows us to control the overlap ratio of both categories. Table [3](#tab:lower_upper){reference-type="ref" reference="tab:lower_upper"} presents the upper and lower limits for selected overlap proportions. From the bold values of this table, the reader can understand that, for AP, if we keep examples with fewer than 230 tokens from the short class and examples longer than 90 tokens for the long class, we obtain two distributions where 50% of the examples overlap in sequence length. This dataset alteration is illustrated in Figure [\[fig:yelp_distribution\]](#fig:yelp_distribution){reference-type="ref" reference="fig:yelp_distribution"}((b)). This figure also presents the original distribution (88% overlap). As one can see, we created a length distribution difference that we expect to be captured by the transformer-based model instead of salient textual features. ### Generation of Subsets to Test Predictive Behavior for Different Ranges {#sss:tests_subsets} The last step requires partitioning the test dataset, whose sequence lengths are illustrated in Figure [\[fig:yelp_test_distribution\]](#fig:yelp_test_distribution){reference-type="ref" reference="fig:yelp_test_distribution"}. The first partition, **Gap-test**, contains short class and long class examples having respectively a length shorter or longer than the optimal split point. The examples from this partition contain, to some extent, the sequence length meta-feature associated with their respective category. The second partition, **Reverse-test**, is the complement of the first partition and thus contains long observations of the short class and short examples of the long class. We expect this partition to contain test set examples that do not carry the sequence length meta-features. ### Model training on the Altered Training Set and Evaluation on the Test Subsets The final part consists of obtaining the different accuracies on the 3 test subsets (**Original**, **Gap**, and **Reverse**) using a model trained with the altered dataset. Model hyperparameters and network topology are presented in the appendix. ## Evaluation Metrics {#ss:metrics} Based on accuracy values obtained for the 3 test subsets, we analyze the differences in performance to assess the problem of sequence length learning. ### Analysis of Accuracy Values We first analyse the accuracy from the **Original-test** subset. Even though this metric shows the classifier's general performance regarding the classification task, it is of little help to identify the presence of feature-length learning. The (expected) performance decrease can come from either of those two sources: poor training examples subset (as we exclude some data) or overfitting to the sequence length meta feature. The accuracies from the **Gap-test** and **Reverse-tests** subsets are more helpful. If the performance on **Gap-test** increases and **Reverse-test** decreases as the overlapping proportion decreases, we have a strong hint that the model learned to use the sequence length meta-feature. Finally, given the problem, one can analyze accuracy differences between subsets with and without the unwanted meta-feature to compare the effect of sequence length learning between scenarios. In that sense, **Original-test** and **Gap-test** or **Gap-test** and **Reverse-test** can be used to evaluate the impact of an alleviating technique. As shown in Table [4](#tab:baseline_3){reference-type="ref" reference="tab:baseline_3"}, neither of the two original datasets suffers from the sequence length learning problem. This conclusion is expected and essential, as the performance on these two datasets is often used to benchmark newer algorithms (see ). The metric we use to derive this conclusion is the accuracy difference between **Gap-test** and **Reverse-test**. Respectively, we obtain 95.1 % vs 94.1 % and 98.0 % vs 97.1 % for Amazon-Polarity and Yelp-Polarity. We can easily assess that the accuracy of both models on **Gap-test** is significantly higher than on **Reverse-Test**. Another intuition we can validate in the absence of sequence length learning is the similarity between the two partitions' accuracies (test-original and test-reverse). Finally, our focus in this paper is to show that transformers, given a biased dataset, will suffer from the problem. ## Evaluating the impact of Sequence Length Learning on Datasets {#sec:injection_dataset_evaluation} To evaluate the impact of adding the sequence length meta-feature in our datasets, we finetune the model's weights using the engineered training datasets from section [3.2.2](#ss:alter_training_Set){reference-type="ref" reference="ss:alter_training_Set"} and evaluate the models on the 3 test datasets from section [3.2.3](#sss:tests_subsets){reference-type="ref" reference="sss:tests_subsets"}. The accuracy values obtained for the selected overlap scenarios on the test partitions are presented in Table [5](#tab:baseline_results){reference-type="ref" reference="tab:baseline_results"}. The first two columns indicate which version of training set is used to fit the model. The third column contains the accuracy obtained on the original, unaltered test set associated with the train dataset from the first column. The last two columns are the accuracy values obtained on two generated test partitions. We first notice the degrading performance for both datasets on the **Original-test** partition, from 95.0 % to 53.2 % for AP and 97.6 % to 56.0 % for YP. The accuracy of the model on **Original-test** can be decomposed into two parts, as it is a weighted average between **Gap-test** and **Reverse-test**. As we investigate, we see that the source of the degradation is the models performance on the partition without the sequence length meta-feature (**Reverse-test**). This yield to an important observation: the model's accuracy on the **Reverse-test** subset decreases as the overlapping proportion of both classes reduces, suggesting that the sequence learning effect is more potent as the overlapping percentage decreases. More importantly, we see that transformer-based model failed to capture any important textual features when the overlapping is 0 %. We come to this conclusion since the accuracy on that subset reaches 0% even though these examples can be correctly categorized by a classifier trained on the original training dataset. These results confirm the preliminary work from. Even if the sequence learning does not impact the original datasets, unpublished datasets may have as little as 80% overlapping proportion and be exposed to the sequence length learning impact. Using this analysis, we conclude that the transformer-based models suffer from the sequence length learning problem to a large extent whenever this bias is present in the dataset. # Alleviating the Impact of Sequence Length Learning {#sec:simple_technique} There are a few simple solutions to alleviate the impact of sequence length learning. In this section, we present two solutions to reduce the unwanted impact. First, we exploit the pretrained transformer representations in a few-shot learning paradigm. Secondly, we augment the training dataset by using the transformer's language model pretraining to increase the overlap percentage. ## Exclude Examples for Length with Weak Overlap {#ss:few-shot} The first approach exploits the robustness of transformer representations learned during pretraining. These efficient representations allow us to reduce the number of training examples significantly, we believe that, as, pretrained models such as RoBERTa are strong few-shot learners. Reducing the number of training examples allows us to only pick observations that have similar lengths from both classes. To evaluate this hypothesis, we finetune the RoBERTa official model's weights using the overlapping section of the engineered dataset from Section [3.2.2](#ss:alter_training_Set){reference-type="ref" reference="ss:alter_training_Set"} exclusively. This data selection yields an adjusted overlap percentage of around 100 %, thus eliminating the sequence length meta-feature. For example, for AP, using long and short examples containing more than 70 tokens **a**nd less than 85 tokens provides a training subset reduced to 25% of its original size with an overlapping percentage around 100%. We evaluate the models generated on the three test partitions. For example, in AP 25%, we use 25% of the original training data to fit a model that we evaluate on the complete, original, unaltered test set as well as on the other 2 subsets. The accuracies of those models for **Original-test** and **Gap-test** are presented in table [6](#tab:few_shot){reference-type="ref" reference="tab:few_shot"}. Using those two subsets gives us enough information to view the complete picture. As predicted, the pretrained transformer models perform very well in our few-shot learning paradigm. The model's performance using a small portion of the training data almost reaches the benchmarks obtained by state-of-the-art models. The reader should note that we cannot evaluate the performance of 0 % dataset in our few-shot learning paradigm as there is not observations available to use for training purposes. We estimated that 1000 observations were needed to achieve an acceptable performance in the few shot learning paradigm for both datasets. We believe the nature of the task is important to the strong performance of the models in the few-shot learning paradigm. As the pretraining methodology uses general-purpose corpora, we expect the text documents used in training to contain the full spectrum of both positive and negative sentiment. Hence, per design, the model's pretrained weights can be combined efficiently into a representation that can be used for sentiment classification. These results also show that OOD observations (regarding the sequence length) can be classified with high precision. This contrasts the results obtained by and for a seq2seq task. As discussed, we believe transformers can easily transform text input into a representation carrying sentiment information. Even if transformer hidden weights can be adapted to a fundamental task such as sentiment classification, other binary classification tasks using field-specific vocabulary may not use this technique. The following section presents a data augmentation technique that solely relies on the transformers' language model (LM) weights. ## Augment Training Data using LM The second technique exploits the transformer's pretrained language model to reduce the sequence length bias contained in the training data. We reduce the bias by synthetically increasing the overlapping percentage. To achieve this, we extend the examples from the short class and by shorten the examples from the long class with mask tokens. In our experiment, we used the masked language model (MLM) of RoBERTa to achieve both alterations as presented by. ### Augmenting Training Examples To extend the dataset, we obtain examples from the short class. We add a random number of literal `<mask>` tokens in each instance. In our experiment, the random variable used to count the number of `<mask>` additions for a given document follows a binomial distribution with parameter \(q=0.15\) and \(m\) equal to the number of tokens in the example. The parameter \(q\) is selected to replicate the training parameters of the base RoBERTa model. To reduce the dataset, we obtained examples from the long class. For each example, we selected a random number of tokens to remove. To ensure the sentence remains grammatically correct, we replaced two words with a single `mask` token, effectively reducing the length by one word. Combing examples from these two operations create a dataset containing `<mask>` tokens that need to be filled (inferred) by the transformer's MLM. We present in annex some augmented examples from AP. ### Results and Analysis Results obtained for models trained on the augmented dataset are presented in Table [7](#tab:augmented_results){reference-type="ref" reference="tab:augmented_results"} and [8](#tab:augmented_results_2){reference-type="ref" reference="tab:augmented_results_2"}. These tables contain many indicators that show that our data intervention reduced the sequence length learning impact. For both tables, the first column contains the dataset used, and the second and third column contains the distribution overlap value for the initial (ini.) training set and the augmented (aug.) one. These figures show us that our data augmentation scheme was successful. The next two columns of Table [7](#tab:augmented_results){reference-type="ref" reference="tab:augmented_results"} present the accuracy of the model trained using the initial and augmented data on the unaltered **Original-test** dataset. We consider that the data augmentation reduced the impact of the sequence length learning whenever two criteria are met. The first criterion is the accuracy improvement on the **Original-test** partition. This performance enhancement is a sum of two factors. One factor is the improvement due to the model correctly classifying more test examples from **Reverse-test**. The other factor is the expected decrease on the **Reverse-test**, as our data augmentation techniques lessen the magnitude of the meta-feature potential used by the initial model. The second criterion we consider is the preservation of the model's predictive power on the **Gap-test** subset. This figure should remain as high and as close as possible to the performances of the model trained with the original dataset (from Table [4](#tab:baseline_3){reference-type="ref" reference="tab:baseline_3"}). This criterion is important as we don't want to improve the performance of the **Reverse-test** at the expense of the examples from the **Gap-test** subset. In other words, we want a good model for data from all lengths. As we can see, models trained with an augmented dataset with the lower overlapping (0%, 25%, and 50%) improved their predictive power on the overall test set. We also validated that they retain their predictive power on the **Gap-test** subset. These results support our hypothesis that extending and reducing the example from a biased dataset can alleviate the sequence length learning's impact. Other interesting results are presented in Table [8](#tab:augmented_results_2){reference-type="ref" reference="tab:augmented_results_2"}. This table shows the performance difference of models trained on initial and augmented datasets on the **Gap-test** and **Reverse-test** partition. The bigger the difference, the more important the effect of the sequence length learning is. As an example, one should read the bolded line from the table as follow. The model trained with the YP with 25% overlap dataset had a performance gap of 80.6% between datasets with and without the sequence length features. With the addition of our synthetic examples in the training dataset, the gap was reduced to 40.2%, effectively decreasing the gap by a factor of 50%. As we can see, our data augmentation technique reduced the performance difference for every model trained on each set. # Conclusion & Future Works In this paper, we showed that the transformer-based architecture may suffer from sequence length learning, a problem where a model learns to use the difference between the class observation lengths instead of important textual features. We empirically demonstrated that we can inject the sequence length meta-feature into a dataset and force a transformer-base classifier to use it. Finally, we reduced the impact by eliminating examples where no overlap existed and by augmenting the dataset using the language model capacities of the transformers. In the future, we would like to evaluate hierarchical architectures such as HAN and Recurrence/Transformer over BERT (roBERT by ). We believe these models would also suffer from the same problem and the hierarchical topology of the networks would allow for innovative approaches. Adversarial approaches could also be evaluated, such as training a model to classify while preventing another model to use its representation to predict the sequence length. Finally, we would like to evaluate if other tasks, such as regression, are as impacted as classification.

{'timestamp': '2022-12-19T02:09:56', 'yymm': '2212', 'arxiv_id': '2212.08399', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08399'}

# Introduction In the last decade, the research in quantum physics is experiencing a second quantum revolution . Huge efforts have been and are being made in developing and engineering quantum hardware and control  allowing to perform novel quantum tasks in computation , communication  and sensing . When dealing with real-life quantum hardware, one possibly has to take into account the presence of states which are not populated during the dynamics but still affect it via virtual processes. Their number is exponentially growing with the size of the system and may produce various phenomena from the renormalization of coupling constants to leakage from the relevant \"computational\" Hilbert subspace. These non-relevant sectors can be removed if an effective Hamiltonian \(\hat H_\mathrm{eff}\) is determined which describes the same dynamics of the original \(\hat H\) for relevant cases and it is of course much simpler than tho original one. This is specially important for quantum control where a simpler Hamiltonian  may allow for analytical solutions or for faster convergence of numerical intensive algorithms such those based on optimal control theory  or reinforcement learning . Adiabatic elimination (AE) is perhaps the simplest and still successful method to determine \(\hat H_\mathrm{eff}\) eliminating from the dynamics states that are mostly not populated. However, this approach presents ambiguities and limitations (see § [7.1](#sec:AE-ambiguities){reference-type="ref" reference="sec:AE-ambiguities"}) pointed out for instance in Refs.  where they have been tackled using various techniques such as Green's function formalism  and exploiting Markov approximation in a Lippman-Schwinger approach . A canonical method for computing low-energy effective Hamiltonians relies on the Schrieffer-Wolff transformation technique , *i.e.* searching for a suitable unitary transformation that decouples approximately the relevant and the non-relevant subspaces. Here we propose a derivation of an effective coarse-grained Hamiltonian which is naturally free from the ambiguities and limitations mentioned above. The method is based on the Magnus expansion (ME) which expresses the solution of a differential equation into an exponential form . Applied to the time-evolution operator \(U\) in a suitable coarse-graining time \(\tau\), it yields an approximation whose logarithm gives an effective Hamiltonian. In general, the result depends on \(\tau\) which must be chosen, if possible, in a proper way depending on the values of the system parameters and the significance of the relevant dynamics obtained a posteriori. As a benchmark, we apply the method to a three-level system in Lambda configuration. We discuss the systematic improvement of the approximation and compare it with other strategies. We check the validity of the results by using various figures of merit, the long-time fidelity of quantum evolution being the most informative one. # Methods ## Effective Hamiltonian by the Magnus expansion {#sec:effH-magnus} The Magnus expansion is a mathematical tool that allows one to express the solution of a differential equation in an exponential form. We apply that to the time-evolution operator \(U(t, t_0)\) of a quantum system which solves the Schrödinger equation \(\dot{U}(t,t_0) =-i \, H(t) \, U(t,t_0)\) for initial condition \(U(t_0,t_0) = \mathds{1}\). The Magnus expansion allows to write the logarithm of \(U(t,t_0)\) as a series, or, likewise \[U(t,t_0) = \mathrm{e}^{-i \, \sum_i F_i(t,t_0) }\] The two lowest-order terms of the expansion are given by \[\label{eq:magnus-terms} \begin{aligned} F_1 &= \int_{t0}^t \, ds \, H(s), \\ F_2 &=-\frac{i}{2} \int_{t0}^t \int_{t0}^{s_1} ds_1 \, ds_2 \, \comm{H(s_1)}{H(s_2)} \end{aligned}\] with \(s_1 > s_2\). The term \(F_1\) yields the so-called average Hamiltonian (which is used, for example, in Ref.  to implement a super-adiabatic protocol without the need of a new interaction), while \(F_2\) provides already an excellent approximation in most cases. Higher-order terms involve time-ordered integrals of higher-order nested commutators  and are reported in appendix [\[app:ME_high_order\]](#app:ME_high_order){reference-type="ref" reference="app:ME_high_order"}. Coarse-graining of the dynamics is operated by first splitting the evolution operator into time-slices [^1] \[U(t, t_0) = \prod_j U \Big(t_j \, + \,\frac{\tau}{2}, t_j \,-\frac{\tau}{2} \Big) =: \prod_j U_j.\] Then we consider the ME in the \(j\)-th sub-interval and truncate the series at a given order \(n\) \[{i \over \tau} \,\ln U_j \approx {1 \over \tau} \, \sum_{i=1}^n F_i(t_j|\tau) =: H_\mathrm{eff}(t_j|\tau)\] obtaining an effective Hamiltonian (for a brief discussion about convergence see Appendix [\[app:convergence\]](#app:convergence){reference-type="ref" reference="app:convergence"}). The structure of the ME suggests that accuracy is related to the smallness of the commutator \([H(t),H(t^\prime)]\) at different times, and it may increase by choosing a small enough \(\tau\). If at the same time \(\tau\) can be chosen large enough, the fast dynamics in the integrals defining \(F_i\) is averaged out. When successful, this procedure defines a coarse-grained Hamiltonian \(H_\mathrm{eff}(t_j)\) whose explicit dependence on \(\tau\) is negligible. With the same degree of accuracy, we finally obtain the approximate time-evolution operator as \[U(t,t_0) \approx \prod_j \mathrm{e}^{-i H_\mathrm{eff}(t_j) \tau} \approx \mathrm{exp} \Big\{-i \int_{t_0}^t ds \; H_\mathrm{eff}(s) \Big\} =: U_\mathrm{eff}(t,t_0)\] ## Validation of the effective Hamiltonian {#sec:validation} Our main goal is to find a relatively simple \(H_\mathrm{eff}\) describing accurately the dynamics in a suitable \"relevant\" subspace. The dynamics taking outside this subspace is not important, so \(H_\mathrm{eff}\) needs not to be accurate there. Since we are interested to the dynamics it is natural to compare the exact population histories and coherences for the low-energy dynamics with those obtained with \(H_\mathrm{eff}\). More compact and effective quantifiers can be defined by adapting to our problem standard metrics for operators in the Hilbert space, as the operatorial spectral norm or trace norm . We anticipate that the effective Hamiltonian we are going to derive has a block diagonal structure, \(H_\mathrm{eff}= P_0 \, H_\mathrm{eff} \,P_0 + (\mathds{1}-P_0)\, \, H_\mathrm{eff} \, (\mathds{1}-P_0)\) where the projection operator \(P_0\) defines the relevant subspace. Since \([H_\mathrm{eff},P_0]=0\) both the relevant subspace and its orthogonal complement are invariant under the effective dynamics. Then a suitable quantifier is defined as a fidelity \[\label{eq:relevant-fidelity} F = \min_{ \ket{ \psi_0 } } \left\{ |\bra{ \psi_0 } \mathcal{U}^{\dagger} \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }|^2 \right\}\] where \(\ket{\psi_0} = P_0 \ket{\psi_0}\) is a vector belonging to the relevant subspace. This subspace fidelity can be smaller than one either because \(H_\mathrm{eff}\) is not accurate in describing the dynamics in the relevant subspace or because the exact dynamics determines leakage from the relevant subspace, with probability \(L=\braket{\psi_0}{U^\dagger(t)\, [\mathds{1}-P_0] \, U(t)|\psi_0}\). Therefore we could define another figure of merit characterizing procedures where leakage from the subspace has been excluded by post-selection \[\label{eq:postselection-fidelity} F^\prime = \min_{ \ket{ \psi_0 } } \left\{ \left|{\bra{ \psi_0 } \mathcal{U}^{\dagger} P_0 \over \sqrt{1-L}} \, \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }\right|^2 \right\} =\min_{ \ket{ \psi_0 } } \left\{ {|\bra{ \psi_0 } \, \mathcal{U}^{\dagger} \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }|^2 \over 1-L} \right\}\] which is the subspace fidelity between the effective dynamics and the post-selected vector. The impact of leakage will be quantified by approximating \(F^\prime \approx F_m^\prime := F + L_m\) where \(L_m\) is the probability of leakage evaluated for the initial state which enters the minimization determining \(F\) in Eq.([\[eq:relevant-fidelity\]](#eq:relevant-fidelity){reference-type="ref" reference="eq:relevant-fidelity"}). This approximation is justified if both the infidelity and the leakage are small, \(I := 1-F \ll 1\) and \(L\ll 1\), and arguing that while \(F_m^\prime\) is not a lower bound as \(F^\prime\) in Eq.([\[eq:postselection-fidelity\]](#eq:postselection-fidelity){reference-type="ref" reference="eq:postselection-fidelity"}) the worst-case error may be a significant overestimate for many initial states . # Application to adiabatic elimination {#section:ad_el_pr} We now apply the procedure outlined to a three-level system in Lambda configuration modelled by the Hamiltonian \[\hat H =-\frac{\delta}{2} \ketbra{0}{0} + \frac{\delta}{2} \ketbra{1}{1} + \Delta \ketbra{2}{2} + \frac{1}{2} \sum_{k=0,1} \big[ \Omega_k^* \,\ketbra{k}{2} + \mbox{h.c.} \big]. \label{eq:hamiltonian-lambda}\] which describes a quantum network with on-site energies \((\pm \delta/2, \Delta/2)\) and tunneling amplitudes \(\Omega_k\) (see Fig.[\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}a). The same Hamiltonian provides the standard description in a rotating frame of a three-level atom driven by two near-resonant corotating semiclassical ac electromagnetic fields. In this case, \(\Omega_k\) are the amplitudes of the fields while the single-photon detunings between the atomic level splitting \(E_i\) and the frequencies \(\omega_k\) of the fields, \(\delta_0 := E_2-E_0-\omega_0\) and \(\delta_1 := E_2-E_1-\omega_1\), enter the diagonal elements \(\delta := \delta_2-\delta_1\) and \(\Delta = (\delta_2+\delta_1)/2\). This model describes several three-level coherent phenomena used in quantum protocols from Raman oscillations , stimulated Raman adiabatic passage  and hybrid schemes . The dynamics is governed by the Schrödinger equation \(i \, \dot c_i(t) = \sum_{j=0}^{2} \bra{i} \hat H \ket{j} \, c_j(t)\) for \(i,j = 0, \, 1,\,2\). The system is prepared in the subspace spanned by the two lowest energy states, i.e. \(\ket{\psi_0}= c_0 \ket{0} + c_1 \ket{1}\), which is defined as the \"relevant\" subspace. We want to understand under which conditions the dynamics is confined to the relevant subspace, and to determine a Hamiltonian operator \(\hat H_\mathrm{eff}\) whose projection onto this subspace effectively generates the confined dynamics. ## Adiabatic elimination: ambiguities and limitations {#sec:AE-ambiguities} AE offers a simple and handy solution to this problem . The standard procedure  relies on the observation that if \(\Delta \gg \delta, | \Omega_k |\) for \(k=0,1\), transitions from the lowest energy doublet and the state \(\ket{2}\) are suppressed. Then, assuming then \(\dot c_2(t)\) can be neglected in the Schrödinger equation, we find \(c_2 =-\tfrac{ \Omega_0 }{ 2 \Delta } \, c_0-\tfrac{ \Omega_1 }{ 2 \Delta } \, c_1\). Substituting in the equations for \(\{c_0,c_1\}\) we obtain an effective two-level problem \(i \partial_t \ket{\phi} = \hat H_\mathrm{eff} \ket{\phi}\) where \[\hat{H}_\mathrm{eff} =-\Big( \frac{ \delta }{ 2 }-S_0 \Big) \ketbra{0}{0} + \Big( \frac{ \delta }{ 2 } + S_1 \Big) \ketbra{1}{1} + \Big( \frac{ \tilde{\Omega} }{ 2 } \ketbra{1}{0} + \text{h.c.} \Big)\] where \(S_k =-|\Omega_k|^2/(4 \, \Delta)\), \(k = 0, \, 1\) are energy shifts and \(\tilde{ \Omega } =-\Omega_0 \, \Omega^*_1/(2 \, \Delta )\) is the normalized coupling. Since \(\hat{H}_\mathrm{eff}\) is defined in the relevant subspace the state \(\ket{2}\) is not involved in the problem anymore. This procedure may be generalized to \(d > 3\)-level systems yielding an effective Hamiltonian in a \(n < d\)-dimensional relevant subspace. It has been pointed out in the literature   that standard AE suffers from a number of ambiguities and limitations that here we summarize. 1. If we add to \(H\) a term \(\eta \mathds{1}\) which is an irrelevant uniform shift of all the energy levels, the procedure yields an \(H_\mathrm{eff}\) which depends on \(\eta\) in a non-trivial way. Thus the procedure is affected by a gauge ambiguity. By comparing the exact numerical result with an analytic approximation based on the resolvent method a \"best choice\" \(\eta =0\) has been proposed . 2. AE completely disregards the state \(\ket{2}\). However, although apparently confining the dynamics to the relevant subspace the procedure yields that \(c_2(t) \neq 0\) and depends on time. Thus, on the one hand the approximation misses leakage to \(\ket{2}\); on the other, it does not guarantee that the normalization of states of the relevant subspace is conserved. In Ref.  the problem of normalization is overcome by writing separated differential equations in the relevant and in the non-relevant subspaces. 3. The residual population in \(\ket{2}\) as given by the approximate \(|c_2(t)|^2\) may undergo very fast oscillations with angular frequency \(\sim \Delta\). This is not consistent with the initial assumption that \(\dot{c}_2 \approx 0\). In Ref.  the assumption is supported by arguing that it holds at the coarse-grained level which averages out the dynamics at time-scales \(\sim \Delta^{-1}\) or faster. 4. Standard AE is not a reliable approximation for larger two-photon detunings or larger external pulses and it is not clear how to systematically improve its validity. We will show how the methodology outlined in § [6.1](#sec:effH-magnus){reference-type="ref" reference="sec:effH-magnus"} yields an effective formulation which overcomes the whole criticism above leveraging only on coarse-graining of the dynamics. ## Magnus expansion in the regime of large detunings We now turn to coarse-graining via the ME. Having in mind the regime where \(\Delta \gg |\Omega_k|\) we first transform the Hamiltonian Eq.([\[eq:hamiltonian-lambda\]](#eq:hamiltonian-lambda){reference-type="ref" reference="eq:hamiltonian-lambda"}) to the interaction picture \[\tilde{H}(t) = U_0^\dagger \,(H-H_0)\, U_0= \frac{\Omega_0^*}{2} \, \ketbra{0}{2} \, e^{-i \left( \Delta + \delta/2 \right) t} + \, \frac{\Omega_1^*}{2} \, \ketbra{1}{2} \, e^{-i \left( \Delta-\delta/2 \right) t} + \text{h.c.}\] where \(U_0(t) := \mathrm{e}^{-i H_0 t}\), \(H_0\) being the diagonal part of \(H\). The first two terms of the ME in Eqs.([\[eq:magnus-terms\]](#eq:magnus-terms){reference-type="ref" reference="eq:magnus-terms"}) are evaluated using the integrals reported in the Appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"} \[\begin{aligned} \tilde{H}^{(1)}_\mathrm{eff}(t) &= \frac{ \Omega_0^*}{ 2 } \, e^{-i \left( \Delta + \delta/2 \right) t } \, \text{sinc} \left( \tfrac{ \left( \Delta + \delta/2 \right) \tau }{ 2 } \right) \, \ketbra{0}{2} + \\ &+ \, \frac{ \Omega_1^* }{ 2 } \, e^{-i \left( \Delta-\delta/2 \right) t } \, \text{sinc} \left( \tfrac{ \left( \Delta-\delta/2 \right) \tau }{ 2 } \right) \, \ketbra{1}{2} + \text{h.c.} \\ \tilde{H}^{(2)}_\mathrm{eff}(t) &= {S}_0 \ketbra{0}{0} + {S}_1 \ketbra{1}{1}-(S_0+S_1) \ketbra{2}{2} + \frac{ \tilde{\Omega} }{2} \, \ketbra{1}{0} e^{ i \delta t } + \text{h.c.} \end{aligned}\] where the coefficients are given by \[\begin{aligned} {S}_{k} &=-\frac{ |\Omega_{k}|^2 }{4 \left( \Delta + (-1)^k \delta/2 \right)} \left[ 1-\text{sinc} \left( \tfrac{ \Delta + (-1)^k \delta/2 }{ 2 } \tau \right) \right], \\ \tilde{\Omega} &=-\frac{ \Omega_0 \Omega^*_1 }{ 2 } \, \frac{ \Delta }{ \left( \Delta^2-\delta^4 / 4 \right) } \left[ \text{sinc} \tfrac{ \delta \tau }{ 2 }-\text{sinc} \tfrac{ \Delta \tau }{ 2 } \right], \label{eq:eff-coupling0} \end{aligned}\] for \(k=0,1\). The first-order \(\tilde{H}^{(1)}_\mathrm{eff}\) is an averaged version of \(\tilde H\). The second-order \(\tilde{H}^{(2)}_\mathrm{eff}\) contains shifts of the diagonal entries and an off-diagonal term coupling directly \(\ket{0}\) and \(\ket{1}\). At this order that state \(\ket{2}\) is energy-shifted but not coupled to other states. We now focus on the regime \(\Delta \gg \delta, \, |\Omega_k|\) where we can choose a coarse-graining time such that \(2\pi/\Delta \ll \tau \ll 2\pi/\delta\). Then all the \(\mathrm{sinc}(x)\) functions appearing in the terms of \(H_\mathrm{eff}\) above are nearly vanishing except for \(\mathrm{sinc}(\delta \tau /2) \approx 1\). As a result first-order \(\tilde{H}^{(1)}_\mathrm{eff}(t)\) averages out while the coefficients of \(\tilde{H}^{(2)}_\mathrm{eff}(t)\) become \[\begin{aligned} {S}_{k} =-\frac{ |\Omega_{k}|^2 }{4 \left( \Delta + (-1)^k \delta/2 \right)} \quad; \quad \tilde{\Omega} =-\frac{ \Omega_0 \Omega^*_1 }{ 2 } \, \frac{ \Delta }{ \left( \Delta^2-\delta^4 / 4 \right) } \end{aligned}\] Transforming back to the laboratory frame we finally obtain \[\begin{aligned} {H}_\mathrm{eff} = &-\left( \frac{ \delta }{ 2 }-S_0 \right) \ketbra{0}{0} + \left( \frac{ \delta }{ 2 } + S_1 \right) \ketbra{1}{1} \, + \, \left( \Delta \,-\, S_0 \,-\, S_1 \right) \ketbra{2}{2} \, + \nonumber \\ &+ \, \frac{ \tilde{\Omega} }{ 2 } \, \ketbra{1}{0} \, + \, \text{h.c.} \label{Eq:effective_magnus} \end{aligned}\] Before discussing the results in detail we make some general comments. As anticipated we obtain a a block-diagonal effective Hamiltonian. In the relevant subspace, it is similar to the result of standard AE. Concerning the criticism to the standard AE mentioned in § [7](#section:ad_el_pr){reference-type="ref" reference="section:ad_el_pr"} we first observe that our result is not affected by the gauge ambiguity emerging for \(H \to H + \eta \mathds{1}\), the \"best choice\" rule \(\eta =0\) of Ref.  being set naturally. Indeed the uniform shift only changes trivially \(U_0\) and does not enter \(\tilde H\) where only level splittings appear. Secondly, the state \(\ket{2}\) is not eliminated but the block diagonal structure consistently preserves normalization in each subspace. Thus there is no need to assume that \(\dot c_2=0\). Rather, Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) may describe also a three-level dynamics where all coherences oscillate. Finally, the ME obviously allows for systematic improvement of the result of AE which moreover can be extended significantly as we will argue in the next sections. While for \(\delta = 0\) our \(H_\mathrm{eff}\) is identical to the \"best choice\" result of the standard AE differences emerge for \(\delta \neq 0\). Anticipating the quantitative analysis of § [7.2.1](#sec:delta-neq-zero){reference-type="ref" reference="sec:delta-neq-zero"} here we point out that our \(H_\mathrm{eff}\) reproduces correctly the shifts \(S_k\) as given by second-order perturbation theory[^2] including the correct shift of the \"eliminated\" state \(\ket{2}\). This feature is important for three-level of quantum operations and an example will be discussed later. Coming back to coarse-graining there is still another time scale to take into account. The solution may describe Raman oscillations of the populations in the relevant subspace with a period \(\sim 2 \pi/| \tilde{\Omega} |\). Thus \(\tau\) must be chosen small enough not to average out this dynamics while operating coarse-graining, which implies that \(\Delta \gg \max(\delta,|\tilde{\Omega}|)\). In particular, for small \(\delta\) we need \(\Delta \gg |\tilde{\Omega}|\), implying that \(| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2) \ll 1\). Hence, in order the approximation to work, we don't need both amplitudes \(\Omega_k\) to be small separately, provided their product is small. Notice finally that if \(\delta\) increases, say \(\delta \gg |\tilde{\Omega}|\), but we still may choose the coarse-graining time as \(\tau \gg \max \big(2 \pi/\Delta, 2\pi/\delta\big)\), then Eq.([\[eq:eff-coupling0\]](#eq:eff-coupling0){reference-type="ref" reference="eq:eff-coupling0"}) yields \(\tilde{\Omega} \approx 0\). Differently from standard AE, the three-level \(H_\mathrm{eff}\) resulting by ME correctly reduces in this limit to the diagonal energy-shifted form obtained by perturbation theory. ### Comparison of the results at second-order {#sec:delta-neq-zero} Besides overcoming the criticism raised to the standard AE, the ME approach gives a good approximation already at second order even when \(| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)\) is not very small. We here discuss the case \(\delta \neq 0\) where results of the second-order ME differ from those obtained by standard AE. We first look at the population histories considering the dynamics from an initial state of the relevant subspace \(\ket{\psi_0} = \cos \tfrac{ \theta^\prime }{ 2 } \ket{0} + \sin \tfrac{ \theta^\prime }{ 2 } \ket{1}\) with \(\theta^\prime = \theta + \tfrac{\pi}{2}\). For the curves in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}b the mixing angle \(\theta^\prime\) is chosen such that \(\ket{\psi_0}\) is an eigenstate of an observable \"orthogonal\" in the Bloch space to the effective Hamiltonian obtained by standard AE. In other words, if this latter is represented by a vector forming an angle \(\theta\) with \(\sigma_z\) then we take \(\theta^\prime= \theta + \pi/2\). This choice is expected to maximize the differences between the various cases. Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}b shows that coarse-graining by ME yields population histories that approximate better the exact dynamics in the relevant subspace. Notice that the discrepancy between ME and standard AE is very significant since it indicates a systematic error in the energies which accumulate over time. This clearly emerges from the fidelities \(F\) defined in Eq.([\[eq:relevant-fidelity\]](#eq:relevant-fidelity){reference-type="ref" reference="eq:relevant-fidelity"}) and shown in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c where the minimization has been performed numerically. Indeed, \(F\) become small for the standard AE while the discrepancies between the ME result and the exact dynamics have a much smaller impact than what appears from the population histories. Actually, infidelity for the ME result is almost entirely due to leakage from the relevant subspace. These latter errors seem not to accumulate in time (magenta curve) as confirmed by the corresponding curve in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}d which shows that leakage errors can be corrected by post-selection. The residual error in the phase of the ME curves is remarkably small despite the exceedingly large value of \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)}\approx 0.27\) we used and it is correctable by extending the analysis to fourth order as we will show in the next section. Errors due to leakage in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c almost disappear for values \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)} \ll 0.1\) (black curve in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c), even smaller values being routinely used for control of solid-state artificial atoms. It is anyhow remarkable the accuracy of the ME coarse-grained second-order \(H_\mathrm{eff}\) in describing the protocol supplemented by post-selection for values of \(\delta\) and \(\Omega_k\) much beyond the perturbative regime. ### Higher-order effective Hamiltonian We now exploit the systematic improvement of the approximation. For the sake of simplicity, we will consider \(\delta=0\). In this case, the exact eigenvalues can be calculated analytically, the two splittings \(\epsilon_{ij} := \epsilon_i-\epsilon_j\) being \(\epsilon_{10} = \frac{\Delta}{2} \left( \sqrt{ 1 + 4 \, x }-1 \right)\) and \(\epsilon_{21} = \frac{\Delta}{2} \left( \sqrt{ 1 + 4 \, x } + 1 \right)\) where \(x := \big(|\Omega_0|^2 + |\Omega_1|^2 \big)/( 4 \, \Delta^2)\). The second-order \(H_\mathrm{eff}\) in Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) reproduces the lowest-order expansion, \(\epsilon_{10}/\Delta \approx x\) and \(\epsilon_{21}/\Delta \approx 1+x\) and we now evaluate higher-order terms. Due to the algebra of the operators (see Appendix [\[app:algebra\]](#app:algebra){reference-type="ref" reference="app:algebra"}) the important property holds that terms of odd orders have the same structure as \(\tilde{H}_\mathrm{eff}^{(1)}\) while at even orders they have the structure of \(\tilde{H}_\mathrm{eff}^{(2)}\). In particular the third-order term in the laboratory frame reduces to \[{H}_\mathrm{eff}^{(3)} = \alpha (\tau) \,x\, \left( \ketbra{2}{0} \frac{ \Omega_0 }{ 2 } + \ketbra{2}{1} \frac{\Omega_1 }{ 2 } \right) + \mathrm{ h.c. }\] where[^3] \(\alpha( \tau ) = 1 + { 1 \over 3 }\, \mathrm{sinc} \frac{ \Delta \, \tau }{ 2 } \big[ 1-8 \, \cos (\Delta \, \tau / 2) \big]\) up to an irrelevant factor of modulus one depending on the detailed coarse-graining procedure. After coarse-graining over a time \(\tau \gg 2 \, \pi/\Delta \,\) we are left with \(\alpha (\tau) \approx 1\). The resulting term triggers transitions between the relevant and the not relevant subspaces. While \({H}_\mathrm{eff}^{(3)}\) is \(\sim x^{3/2}\) being off-diagonal it contributes at order \(x^3\) to the correction of the splittings as it can be shown by ordinary non-degenerate perturbation theory. We turn to the fourth order of the ME whose full expression is reported in appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"}. After coarse-graining, the contribution to the effective Hamiltonian in the laboratory frame reduces to \[H^{(4)}_\mathrm{eff} = S^{(4)}_0 \ketbra{ 0 }{ 0 } + S^{(4)}_1 \ketbra{ 1 }{ 1 } + ( \Delta-S^{(4)}_0-S^{(4)}_1 ) \ketbra{ 2 }{ 2 } + \frac{ 1 }{ 2 } \,\big[ \tilde{\Omega}^{(4)} \, \ketbra{ 1 }{ 0 } + \text{h.c.} \big ]\] where \(S^{(4)}_k = x \,| \Omega_k |^2/(4 \Delta)\) and \(\tilde{\Omega}^{(4)} = x\, \Omega_0 \, \Omega_1^*/(2 \Delta)\). This term reproduces the expansion of the exact splitting up to order \(x^2\), providing a more relevant correction to \(H^{(2)}_\mathrm{eff}\) Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) than the third-order \(H^{(3)}_\mathrm{eff}\). Therefore we could neglect this latter term and approximate \(H_\mathrm{eff} \approx H^{(2)}_\mathrm{eff}+ H^{(4)}_\mathrm{eff}\). The above effective Hamiltonian reproduces both the exact energy splittings to order \(x^2\). Being block-diagonal it admits no leakage. We may wonder if including \(H^{(3)}_\mathrm{eff}\) may yield useful information on leakage, but the answer is negative. Indeed we checked that its impact on population histories is small. In particular, the ME yields a coarse-grained version of the population in \(\ket{2}\) which is much smaller than the exact one since this latter oscillates on a time scale \(\sim 2 \pi/\Delta\). ### Validation of the results at fourth-order We now validate our result using the same quantifiers as in § [7.2.1](#sec:delta-neq-zero){reference-type="ref" reference="sec:delta-neq-zero"}. Population histories are shown in Fig. [\[fig:pops\]](#fig:pops){reference-type="ref" reference="fig:pops"} for two different sets of parameters. It is seen that the fourth-order ME (black dashed line) yields a coarse-grained version of the exact result (full gray line) which is accurate in reproducing the Raman oscillations between levels of the relevant subspace. On the contrary the second-order ME (dashed magenta line) clearly shows a discrepancy in the oscillation frequency. This error appears clearly in the long-time fidelity \(F\), shown in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(a-b), and the fidelity of post-selected protocols, shown in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(c-d). In the same figures, we also compare the fourth-order ME with the approximation scheme proposed in Ref.  (green curves) and the two approaches are seen to coincide. Therefore the ME provides a systematic approximation scheme, overcoming the last point of the criticism to standard AE mentioned in § [7.1](#sec:AE-ambiguities){reference-type="ref" reference="sec:AE-ambiguities"}. In particular for the symmetric choice \(|\Omega_0|= |\Omega_1| = 0.3 \, \Delta\) (Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}a, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}c) the fourth order fidelity oscillates between one and \(0.8\) (black dashed curve), whereas the second-order result (magenta dashed curve) decays to lower values because the error in frequencies accumulates in time. The same behaviour is obtained for the asymmetric configuration of the drive amplitudes shown in Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}b, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}d. For these time scales the error in the fourth-order \(F\) in Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(a-b) does not accumulate in time and Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(c-d) show that it is entireka:16-vepsalainen-photonics-squtritly due to leakage since it can be corrected by post-selection. Again we notice that in Figs. [\[fig:pops\]](#fig:pops){reference-type="ref" reference="fig:pops"}, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"} we used parameters with values far beyond the perturbative regime to magnify the errors. Still, errors are not so large and in particular, they are remarkably small for the post-selected dynamics. As in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c, errors due to leakage in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}b (black curve) almost disappear already for values \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)} \ll 0.1\). Finally, we stress the agreement of the ME result with the results of the approach of Ref. . This latter achieves a high-level accuracy using an iterated Lippmann-Schwinger equation in a special gauge defining the interaction picture and supplementing the problem by an ad hoc assumption of Markovianity of the dynamics. Our approach based on the ME shows that the correct result only leverages on coarse-graining. # Discussion and conclusions We introduced a technique based on the Magnus expansion that allows deriving a coarse-grained effective Hamiltonian and yielding a simplified model for the low-energy dynamics of the system. We applied the technique to the problem of the three-level lambda system clarifying ambiguities and inconsistencies of standard AE, which have been raised in the literature . Results from ME accurately interpolate all the limiting cases obtained by standard approximations. For instance, the second-order ME yields a result reducing to the usual AE for \(\delta \ll |\Omega_k|\) but reproducing the perturbative Stark shifts for \(\delta \gg |\Omega_k|\). Moreover, the accuracy of the ME can be systematically increased reaching a full agreement with other accurate approximation schemes such as the one developed in Ref. . Since the ME-based approximation we propose only relies on coarse-graining, this latter is identified as the key ingredient underlying all the approximations. Coarse-grained based approximations yield effective Hamiltonians reproducing accurately the dynamics in the relevant subspace for a wide range of parameters, much beyond the perturbation regime. In particular, it is remarkably large the accuracy achieved for protocols supplemented by post-selection. Notice that the ME approach we described yields a block diagonal effective Hamiltonian which does not cancel the not-relevant subspace but treats it consistently. In particular, for the Lambda system we obtain the correct perturbative energy shift for the level \(\ket{2}\). This result allows non-trivial applications to problems involving multiphoton processes in three-level dynamics  which are relevant for newly developed quantum hardware . Finally, the approach can be extended to the search for effective Hamiltonians in time-dependent problems. The extension is straightforward if the parameters of the Hamiltonian are slowly varying on time scales of the order of the coarse-graining time \(\tau\) as in Ref. . # Introduction In the last decade, the research in quantum physics is experiencing a second quantum revolution . Huge efforts have been and are being made in developing and engineering quantum hardware and control  allowing to perform novel quantum tasks in computation , communication  and sensing . When dealing with real-life quantum hardware, one possibly has to take into account the presence of states which are not populated during the dynamics but still affect it via virtual processes. Their number is exponentially growing with the size of the system and may produce various phenomena from the renormalization of coupling constants to leakage from the relevant \"computational\" Hilbert subspace. These non-relevant sectors can be removed if an effective Hamiltonian \(\hat H_\mathrm{eff}\) is determined which describes the same dynamics of the original \(\hat H\) for relevant cases and it is of course much simpler than tho original one. This is specially important for quantum control where a simpler Hamiltonian  may allow for analytical solutions or for faster convergence of numerical intensive algorithms such those based on optimal control theory  or reinforcement learning . Adiabatic elimination (AE) is perhaps the simplest and still successful method to determine \(\hat H_\mathrm{eff}\) eliminating from the dynamics states that are mostly not populated. However, this approach presents ambiguities and limitations (see § [7.1](#sec:AE-ambiguities){reference-type="ref" reference="sec:AE-ambiguities"}) pointed out for instance in Refs.  where they have been tackled using various techniques such as Green's function formalism  and exploiting Markov approximation in a Lippman-Schwinger approach . A canonical method for computing low-energy effective Hamiltonians relies on the Schrieffer-Wolff transformation technique , *i.e.* searching for a suitable unitary transformation that decouples approximately the relevant and the non-relevant subspaces. Here we propose a derivation of an effective coarse-grained Hamiltonian which is naturally free from the ambiguities and limitations mentioned above. The method is based on the Magnus expansion (ME) which expresses the solution of a differential equation into an exponential form . Applied to the time-evolution operator \(U\) in a suitable coarse-graining time \(\tau\), it yields an approximation whose logarithm gives an effective Hamiltonian. In general, the result depends on \(\tau\) which must be chosen, if possible, in a proper way depending on the values of the system parameters and the significance of the relevant dynamics obtained a posteriori. As a benchmark, we apply the method to a three-level system in Lambda configuration. We discuss the systematic improvement of the approximation and compare it with other strategies. We check the validity of the results by using various figures of merit, the long-time fidelity of quantum evolution being the most informative one. # Methods ## Effective Hamiltonian by the Magnus expansion {#sec:effH-magnus} The Magnus expansion is a mathematical tool that allows one to express the solution of a differential equation in an exponential form. We apply that to the time-evolution operator \(U(t, t_0)\) of a quantum system which solves the Schrödinger equation \(\dot{U}(t,t_0) =-i \, H(t) \, U(t,t_0)\) for initial condition \(U(t_0,t_0) = \mathds{1}\). The Magnus expansion allows to write the logarithm of \(U(t,t_0)\) as a series, or, likewise \[U(t,t_0) = \mathrm{e}^{-i \, \sum_i F_i(t,t_0) }\] The two lowest-order terms of the expansion are given by \[\label{eq:magnus-terms} \begin{aligned} F_1 &= \int_{t0}^t \, ds \, H(s), \\ F_2 &=-\frac{i}{2} \int_{t0}^t \int_{t0}^{s_1} ds_1 \, ds_2 \, \comm{H(s_1)}{H(s_2)} \end{aligned}\] with \(s_1 > s_2\). The term \(F_1\) yields the so-called average Hamiltonian (which is used, for example, in Ref.  to implement a super-adiabatic protocol without the need of a new interaction), while \(F_2\) provides already an excellent approximation in most cases. Higher-order terms involve time-ordered integrals of higher-order nested commutators  and are reported in appendix [\[app:ME_high_order\]](#app:ME_high_order){reference-type="ref" reference="app:ME_high_order"}. Coarse-graining of the dynamics is operated by first splitting the evolution operator into time-slices [^4] \[U(t, t_0) = \prod_j U \Big(t_j \, + \,\frac{\tau}{2}, t_j \,-\frac{\tau}{2} \Big) =: \prod_j U_j.\] Then we consider the ME in the \(j\)-th sub-interval and truncate the series at a given order \(n\) \[{i \over \tau} \,\ln U_j \approx {1 \over \tau} \, \sum_{i=1}^n F_i(t_j|\tau) =: H_\mathrm{eff}(t_j|\tau)\] obtaining an effective Hamiltonian (for a brief discussion about convergence see Appendix [\[app:convergence\]](#app:convergence){reference-type="ref" reference="app:convergence"}). The structure of the ME suggests that accuracy is related to the smallness of the commutator \([H(t),H(t^\prime)]\) at different times, and it may increase by choosing a small enough \(\tau\). If at the same time \(\tau\) can be chosen large enough, the fast dynamics in the integrals defining \(F_i\) is averaged out. When successful, this procedure defines a coarse-grained Hamiltonian \(H_\mathrm{eff}(t_j)\) whose explicit dependence on \(\tau\) is negligible. With the same degree of accuracy, we finally obtain the approximate time-evolution operator as \[U(t,t_0) \approx \prod_j \mathrm{e}^{-i H_\mathrm{eff}(t_j) \tau} \approx \mathrm{exp} \Big\{-i \int_{t_0}^t ds \; H_\mathrm{eff}(s) \Big\} =: U_\mathrm{eff}(t,t_0)\] ## Validation of the effective Hamiltonian {#sec:validation} Our main goal is to find a relatively simple \(H_\mathrm{eff}\) describing accurately the dynamics in a suitable \"relevant\" subspace. The dynamics taking outside this subspace is not important, so \(H_\mathrm{eff}\) needs not to be accurate there. Since we are interested to the dynamics it is natural to compare the exact population histories and coherences for the low-energy dynamics with those obtained with \(H_\mathrm{eff}\). More compact and effective quantifiers can be defined by adapting to our problem standard metrics for operators in the Hilbert space, as the operatorial spectral norm or trace norm . We anticipate that the effective Hamiltonian we are going to derive has a block diagonal structure, \(H_\mathrm{eff}= P_0 \, H_\mathrm{eff} \,P_0 + (\mathds{1}-P_0)\, \, H_\mathrm{eff} \, (\mathds{1}-P_0)\) where the projection operator \(P_0\) defines the relevant subspace. Since \([H_\mathrm{eff},P_0]=0\) both the relevant subspace and its orthogonal complement are invariant under the effective dynamics. Then a suitable quantifier is defined as a fidelity \[\label{eq:relevant-fidelity} F = \min_{ \ket{ \psi_0 } } \left\{ |\bra{ \psi_0 } \mathcal{U}^{\dagger} \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }|^2 \right\}\] where \(\ket{\psi_0} = P_0 \ket{\psi_0}\) is a vector belonging to the relevant subspace. This subspace fidelity can be smaller than one either because \(H_\mathrm{eff}\) is not accurate in describing the dynamics in the relevant subspace or because the exact dynamics determines leakage from the relevant subspace, with probability \(L=\braket{\psi_0}{U^\dagger(t)\, [\mathds{1}-P_0] \, U(t)|\psi_0}\). Therefore we could define another figure of merit characterizing procedures where leakage from the subspace has been excluded by post-selection \[\label{eq:postselection-fidelity} F^\prime = \min_{ \ket{ \psi_0 } } \left\{ \left|{\bra{ \psi_0 } \mathcal{U}^{\dagger} P_0 \over \sqrt{1-L}} \, \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }\right|^2 \right\} =\min_{ \ket{ \psi_0 } } \left\{ {|\bra{ \psi_0 } \, \mathcal{U}^{\dagger} \mathcal{U}_\mathrm{eff} \ket{ \psi_0 }|^2 \over 1-L} \right\}\] which is the subspace fidelity between the effective dynamics and the post-selected vector. The impact of leakage will be quantified by approximating \(F^\prime \approx F_m^\prime := F + L_m\) where \(L_m\) is the probability of leakage evaluated for the initial state which enters the minimization determining \(F\) in Eq.([\[eq:relevant-fidelity\]](#eq:relevant-fidelity){reference-type="ref" reference="eq:relevant-fidelity"}). This approximation is justified if both the infidelity and the leakage are small, \(I := 1-F \ll 1\) and \(L\ll 1\), and arguing that while \(F_m^\prime\) is not a lower bound as \(F^\prime\) in Eq.([\[eq:postselection-fidelity\]](#eq:postselection-fidelity){reference-type="ref" reference="eq:postselection-fidelity"}) the worst-case error may be a significant overestimate for many initial states . # Application to adiabatic elimination {#section:ad_el_pr} We now apply the procedure outlined to a three-level system in Lambda configuration modelled by the Hamiltonian \[\hat H =-\frac{\delta}{2} \ketbra{0}{0} + \frac{\delta}{2} \ketbra{1}{1} + \Delta \ketbra{2}{2} + \frac{1}{2} \sum_{k=0,1} \big[ \Omega_k^* \,\ketbra{k}{2} + \mbox{h.c.} \big]. \label{eq:hamiltonian-lambda}\] which describes a quantum network with on-site energies \((\pm \delta/2, \Delta/2)\) and tunneling amplitudes \(\Omega_k\) (see Fig.[\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}a). The same Hamiltonian provides the standard description in a rotating frame of a three-level atom driven by two near-resonant corotating semiclassical ac electromagnetic fields. In this case, \(\Omega_k\) are the amplitudes of the fields while the single-photon detunings between the atomic level splitting \(E_i\) and the frequencies \(\omega_k\) of the fields, \(\delta_0 := E_2-E_0-\omega_0\) and \(\delta_1 := E_2-E_1-\omega_1\), enter the diagonal elements \(\delta := \delta_2-\delta_1\) and \(\Delta = (\delta_2+\delta_1)/2\). This model describes several three-level coherent phenomena used in quantum protocols from Raman oscillations , stimulated Raman adiabatic passage  and hybrid schemes . The dynamics is governed by the Schrödinger equation \(i \, \dot c_i(t) = \sum_{j=0}^{2} \bra{i} \hat H \ket{j} \, c_j(t)\) for \(i,j = 0, \, 1,\,2\). The system is prepared in the subspace spanned by the two lowest energy states, i.e. \(\ket{\psi_0}= c_0 \ket{0} + c_1 \ket{1}\), which is defined as the \"relevant\" subspace. We want to understand under which conditions the dynamics is confined to the relevant subspace, and to determine a Hamiltonian operator \(\hat H_\mathrm{eff}\) whose projection onto this subspace effectively generates the confined dynamics. ## Adiabatic elimination: ambiguities and limitations {#sec:AE-ambiguities} AE offers a simple and handy solution to this problem . The standard procedure  relies on the observation that if \(\Delta \gg \delta, | \Omega_k |\) for \(k=0,1\), transitions from the lowest energy doublet and the state \(\ket{2}\) are suppressed. Then, assuming then \(\dot c_2(t)\) can be neglected in the Schrödinger equation, we find \(c_2 =-\tfrac{ \Omega_0 }{ 2 \Delta } \, c_0-\tfrac{ \Omega_1 }{ 2 \Delta } \, c_1\). Substituting in the equations for \(\{c_0,c_1\}\) we obtain an effective two-level problem \(i \partial_t \ket{\phi} = \hat H_\mathrm{eff} \ket{\phi}\) where \[\hat{H}_\mathrm{eff} =-\Big( \frac{ \delta }{ 2 }-S_0 \Big) \ketbra{0}{0} + \Big( \frac{ \delta }{ 2 } + S_1 \Big) \ketbra{1}{1} + \Big( \frac{ \tilde{\Omega} }{ 2 } \ketbra{1}{0} + \text{h.c.} \Big)\] where \(S_k =-|\Omega_k|^2/(4 \, \Delta)\), \(k = 0, \, 1\) are energy shifts and \(\tilde{ \Omega } =-\Omega_0 \, \Omega^*_1/(2 \, \Delta )\) is the normalized coupling. Since \(\hat{H}_\mathrm{eff}\) is defined in the relevant subspace the state \(\ket{2}\) is not involved in the problem anymore. This procedure may be generalized to \(d > 3\)-level systems yielding an effective Hamiltonian in a \(n < d\)-dimensional relevant subspace. It has been pointed out in the literature   that standard AE suffers from a number of ambiguities and limitations that here we summarize. 1. If we add to \(H\) a term \(\eta \mathds{1}\) which is an irrelevant uniform shift of all the energy levels, the procedure yields an \(H_\mathrm{eff}\) which depends on \(\eta\) in a non-trivial way. Thus the procedure is affected by a gauge ambiguity. By comparing the exact numerical result with an analytic approximation based on the resolvent method a \"best choice\" \(\eta =0\) has been proposed . 2. AE completely disregards the state \(\ket{2}\). However, although apparently confining the dynamics to the relevant subspace the procedure yields that \(c_2(t) \neq 0\) and depends on time. Thus, on the one hand the approximation misses leakage to \(\ket{2}\); on the other, it does not guarantee that the normalization of states of the relevant subspace is conserved. In Ref.  the problem of normalization is overcome by writing separated differential equations in the relevant and in the non-relevant subspaces. 3. The residual population in \(\ket{2}\) as given by the approximate \(|c_2(t)|^2\) may undergo very fast oscillations with angular frequency \(\sim \Delta\). This is not consistent with the initial assumption that \(\dot{c}_2 \approx 0\). In Ref.  the assumption is supported by arguing that it holds at the coarse-grained level which averages out the dynamics at time-scales \(\sim \Delta^{-1}\) or faster. 4. Standard AE is not a reliable approximation for larger two-photon detunings or larger external pulses and it is not clear how to systematically improve its validity. We will show how the methodology outlined in § [6.1](#sec:effH-magnus){reference-type="ref" reference="sec:effH-magnus"} yields an effective formulation which overcomes the whole criticism above leveraging only on coarse-graining of the dynamics. ## Magnus expansion in the regime of large detunings We now turn to coarse-graining via the ME. Having in mind the regime where \(\Delta \gg |\Omega_k|\) we first transform the Hamiltonian Eq.([\[eq:hamiltonian-lambda\]](#eq:hamiltonian-lambda){reference-type="ref" reference="eq:hamiltonian-lambda"}) to the interaction picture \[\tilde{H}(t) = U_0^\dagger \,(H-H_0)\, U_0= \frac{\Omega_0^*}{2} \, \ketbra{0}{2} \, e^{-i \left( \Delta + \delta/2 \right) t} + \, \frac{\Omega_1^*}{2} \, \ketbra{1}{2} \, e^{-i \left( \Delta-\delta/2 \right) t} + \text{h.c.}\] where \(U_0(t) := \mathrm{e}^{-i H_0 t}\), \(H_0\) being the diagonal part of \(H\). The first two terms of the ME in Eqs.([\[eq:magnus-terms\]](#eq:magnus-terms){reference-type="ref" reference="eq:magnus-terms"}) are evaluated using the integrals reported in the Appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"} \[\begin{aligned} \tilde{H}^{(1)}_\mathrm{eff}(t) &= \frac{ \Omega_0^*}{ 2 } \, e^{-i \left( \Delta + \delta/2 \right) t } \, \text{sinc} \left( \tfrac{ \left( \Delta + \delta/2 \right) \tau }{ 2 } \right) \, \ketbra{0}{2} + \\ &+ \, \frac{ \Omega_1^* }{ 2 } \, e^{-i \left( \Delta-\delta/2 \right) t } \, \text{sinc} \left( \tfrac{ \left( \Delta-\delta/2 \right) \tau }{ 2 } \right) \, \ketbra{1}{2} + \text{h.c.} \\ \tilde{H}^{(2)}_\mathrm{eff}(t) &= {S}_0 \ketbra{0}{0} + {S}_1 \ketbra{1}{1}-(S_0+S_1) \ketbra{2}{2} + \frac{ \tilde{\Omega} }{2} \, \ketbra{1}{0} e^{ i \delta t } + \text{h.c.} \end{aligned}\] where the coefficients are given by \[\begin{aligned} {S}_{k} &=-\frac{ |\Omega_{k}|^2 }{4 \left( \Delta + (-1)^k \delta/2 \right)} \left[ 1-\text{sinc} \left( \tfrac{ \Delta + (-1)^k \delta/2 }{ 2 } \tau \right) \right], \\ \tilde{\Omega} &=-\frac{ \Omega_0 \Omega^*_1 }{ 2 } \, \frac{ \Delta }{ \left( \Delta^2-\delta^4 / 4 \right) } \left[ \text{sinc} \tfrac{ \delta \tau }{ 2 }-\text{sinc} \tfrac{ \Delta \tau }{ 2 } \right], \label{eq:eff-coupling0} \end{aligned}\] for \(k=0,1\). The first-order \(\tilde{H}^{(1)}_\mathrm{eff}\) is an averaged version of \(\tilde H\). The second-order \(\tilde{H}^{(2)}_\mathrm{eff}\) contains shifts of the diagonal entries and an off-diagonal term coupling directly \(\ket{0}\) and \(\ket{1}\). At this order that state \(\ket{2}\) is energy-shifted but not coupled to other states. We now focus on the regime \(\Delta \gg \delta, \, |\Omega_k|\) where we can choose a coarse-graining time such that \(2\pi/\Delta \ll \tau \ll 2\pi/\delta\). Then all the \(\mathrm{sinc}(x)\) functions appearing in the terms of \(H_\mathrm{eff}\) above are nearly vanishing except for \(\mathrm{sinc}(\delta \tau /2) \approx 1\). As a result first-order \(\tilde{H}^{(1)}_\mathrm{eff}(t)\) averages out while the coefficients of \(\tilde{H}^{(2)}_\mathrm{eff}(t)\) become \[\begin{aligned} {S}_{k} =-\frac{ |\Omega_{k}|^2 }{4 \left( \Delta + (-1)^k \delta/2 \right)} \quad; \quad \tilde{\Omega} =-\frac{ \Omega_0 \Omega^*_1 }{ 2 } \, \frac{ \Delta }{ \left( \Delta^2-\delta^4 / 4 \right) } \end{aligned}\] Transforming back to the laboratory frame we finally obtain \[\begin{aligned} {H}_\mathrm{eff} = &-\left( \frac{ \delta }{ 2 }-S_0 \right) \ketbra{0}{0} + \left( \frac{ \delta }{ 2 } + S_1 \right) \ketbra{1}{1} \, + \, \left( \Delta \,-\, S_0 \,-\, S_1 \right) \ketbra{2}{2} \, + \nonumber \\ &+ \, \frac{ \tilde{\Omega} }{ 2 } \, \ketbra{1}{0} \, + \, \text{h.c.} \label{Eq:effective_magnus} \end{aligned}\] Before discussing the results in detail we make some general comments. As anticipated we obtain a a block-diagonal effective Hamiltonian. In the relevant subspace, it is similar to the result of standard AE. Concerning the criticism to the standard AE mentioned in § [7](#section:ad_el_pr){reference-type="ref" reference="section:ad_el_pr"} we first observe that our result is not affected by the gauge ambiguity emerging for \(H \to H + \eta \mathds{1}\), the \"best choice\" rule \(\eta =0\) of Ref.  being set naturally. Indeed the uniform shift only changes trivially \(U_0\) and does not enter \(\tilde H\) where only level splittings appear. Secondly, the state \(\ket{2}\) is not eliminated but the block diagonal structure consistently preserves normalization in each subspace. Thus there is no need to assume that \(\dot c_2=0\). Rather, Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) may describe also a three-level dynamics where all coherences oscillate. Finally, the ME obviously allows for systematic improvement of the result of AE which moreover can be extended significantly as we will argue in the next sections. While for \(\delta = 0\) our \(H_\mathrm{eff}\) is identical to the \"best choice\" result of the standard AE differences emerge for \(\delta \neq 0\). Anticipating the quantitative analysis of § [7.2.1](#sec:delta-neq-zero){reference-type="ref" reference="sec:delta-neq-zero"} here we point out that our \(H_\mathrm{eff}\) reproduces correctly the shifts \(S_k\) as given by second-order perturbation theory[^5] including the correct shift of the \"eliminated\" state \(\ket{2}\). This feature is important for three-level of quantum operations and an example will be discussed later. Coming back to coarse-graining there is still another time scale to take into account. The solution may describe Raman oscillations of the populations in the relevant subspace with a period \(\sim 2 \pi/| \tilde{\Omega} |\). Thus \(\tau\) must be chosen small enough not to average out this dynamics while operating coarse-graining, which implies that \(\Delta \gg \max(\delta,|\tilde{\Omega}|)\). In particular, for small \(\delta\) we need \(\Delta \gg |\tilde{\Omega}|\), implying that \(| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2) \ll 1\). Hence, in order the approximation to work, we don't need both amplitudes \(\Omega_k\) to be small separately, provided their product is small. Notice finally that if \(\delta\) increases, say \(\delta \gg |\tilde{\Omega}|\), but we still may choose the coarse-graining time as \(\tau \gg \max \big(2 \pi/\Delta, 2\pi/\delta\big)\), then Eq.([\[eq:eff-coupling0\]](#eq:eff-coupling0){reference-type="ref" reference="eq:eff-coupling0"}) yields \(\tilde{\Omega} \approx 0\). Differently from standard AE, the three-level \(H_\mathrm{eff}\) resulting by ME correctly reduces in this limit to the diagonal energy-shifted form obtained by perturbation theory. ### Comparison of the results at second-order {#sec:delta-neq-zero} Besides overcoming the criticism raised to the standard AE, the ME approach gives a good approximation already at second order even when \(| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)\) is not very small. We here discuss the case \(\delta \neq 0\) where results of the second-order ME differ from those obtained by standard AE. We first look at the population histories considering the dynamics from an initial state of the relevant subspace \(\ket{\psi_0} = \cos \tfrac{ \theta^\prime }{ 2 } \ket{0} + \sin \tfrac{ \theta^\prime }{ 2 } \ket{1}\) with \(\theta^\prime = \theta + \tfrac{\pi}{2}\). For the curves in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}b the mixing angle \(\theta^\prime\) is chosen such that \(\ket{\psi_0}\) is an eigenstate of an observable \"orthogonal\" in the Bloch space to the effective Hamiltonian obtained by standard AE. In other words, if this latter is represented by a vector forming an angle \(\theta\) with \(\sigma_z\) then we take \(\theta^\prime= \theta + \pi/2\). This choice is expected to maximize the differences between the various cases. Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}b shows that coarse-graining by ME yields population histories that approximate better the exact dynamics in the relevant subspace. Notice that the discrepancy between ME and standard AE is very significant since it indicates a systematic error in the energies which accumulate over time. This clearly emerges from the fidelities \(F\) defined in Eq.([\[eq:relevant-fidelity\]](#eq:relevant-fidelity){reference-type="ref" reference="eq:relevant-fidelity"}) and shown in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c where the minimization has been performed numerically. Indeed, \(F\) become small for the standard AE while the discrepancies between the ME result and the exact dynamics have a much smaller impact than what appears from the population histories. Actually, infidelity for the ME result is almost entirely due to leakage from the relevant subspace. These latter errors seem not to accumulate in time (magenta curve) as confirmed by the corresponding curve in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}d which shows that leakage errors can be corrected by post-selection. The residual error in the phase of the ME curves is remarkably small despite the exceedingly large value of \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)}\approx 0.27\) we used and it is correctable by extending the analysis to fourth order as we will show in the next section. Errors due to leakage in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c almost disappear for values \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)} \ll 0.1\) (black curve in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c), even smaller values being routinely used for control of solid-state artificial atoms. It is anyhow remarkable the accuracy of the ME coarse-grained second-order \(H_\mathrm{eff}\) in describing the protocol supplemented by post-selection for values of \(\delta\) and \(\Omega_k\) much beyond the perturbative regime. ### Higher-order effective Hamiltonian We now exploit the systematic improvement of the approximation. For the sake of simplicity, we will consider \(\delta=0\). In this case, the exact eigenvalues can be calculated analytically, the two splittings \(\epsilon_{ij} := \epsilon_i-\epsilon_j\) being \(\epsilon_{10} = \frac{\Delta}{2} \left( \sqrt{ 1 + 4 \, x }-1 \right)\) and \(\epsilon_{21} = \frac{\Delta}{2} \left( \sqrt{ 1 + 4 \, x } + 1 \right)\) where \(x := \big(|\Omega_0|^2 + |\Omega_1|^2 \big)/( 4 \, \Delta^2)\). The second-order \(H_\mathrm{eff}\) in Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) reproduces the lowest-order expansion, \(\epsilon_{10}/\Delta \approx x\) and \(\epsilon_{21}/\Delta \approx 1+x\) and we now evaluate higher-order terms. Due to the algebra of the operators (see Appendix [\[app:algebra\]](#app:algebra){reference-type="ref" reference="app:algebra"}) the important property holds that terms of odd orders have the same structure as \(\tilde{H}_\mathrm{eff}^{(1)}\) while at even orders they have the structure of \(\tilde{H}_\mathrm{eff}^{(2)}\). In particular the third-order term in the laboratory frame reduces to \[{H}_\mathrm{eff}^{(3)} = \alpha (\tau) \,x\, \left( \ketbra{2}{0} \frac{ \Omega_0 }{ 2 } + \ketbra{2}{1} \frac{\Omega_1 }{ 2 } \right) + \mathrm{ h.c. }\] where[^6] \(\alpha( \tau ) = 1 + { 1 \over 3 }\, \mathrm{sinc} \frac{ \Delta \, \tau }{ 2 } \big[ 1-8 \, \cos (\Delta \, \tau / 2) \big]\) up to an irrelevant factor of modulus one depending on the detailed coarse-graining procedure. After coarse-graining over a time \(\tau \gg 2 \, \pi/\Delta \,\) we are left with \(\alpha (\tau) \approx 1\). The resulting term triggers transitions between the relevant and the not relevant subspaces. While \({H}_\mathrm{eff}^{(3)}\) is \(\sim x^{3/2}\) being off-diagonal it contributes at order \(x^3\) to the correction of the splittings as it can be shown by ordinary non-degenerate perturbation theory. We turn to the fourth order of the ME whose full expression is reported in appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"}. After coarse-graining, the contribution to the effective Hamiltonian in the laboratory frame reduces to \[H^{(4)}_\mathrm{eff} = S^{(4)}_0 \ketbra{ 0 }{ 0 } + S^{(4)}_1 \ketbra{ 1 }{ 1 } + ( \Delta-S^{(4)}_0-S^{(4)}_1 ) \ketbra{ 2 }{ 2 } + \frac{ 1 }{ 2 } \,\big[ \tilde{\Omega}^{(4)} \, \ketbra{ 1 }{ 0 } + \text{h.c.} \big ]\] where \(S^{(4)}_k = x \,| \Omega_k |^2/(4 \Delta)\) and \(\tilde{\Omega}^{(4)} = x\, \Omega_0 \, \Omega_1^*/(2 \Delta)\). This term reproduces the expansion of the exact splitting up to order \(x^2\), providing a more relevant correction to \(H^{(2)}_\mathrm{eff}\) Eq.([\[Eq:effective_magnus\]](#Eq:effective_magnus){reference-type="ref" reference="Eq:effective_magnus"}) than the third-order \(H^{(3)}_\mathrm{eff}\). Therefore we could neglect this latter term and approximate \(H_\mathrm{eff} \approx H^{(2)}_\mathrm{eff}+ H^{(4)}_\mathrm{eff}\). The above effective Hamiltonian reproduces both the exact energy splittings to order \(x^2\). Being block-diagonal it admits no leakage. We may wonder if including \(H^{(3)}_\mathrm{eff}\) may yield useful information on leakage, but the answer is negative. Indeed we checked that its impact on population histories is small. In particular, the ME yields a coarse-grained version of the population in \(\ket{2}\) which is much smaller than the exact one since this latter oscillates on a time scale \(\sim 2 \pi/\Delta\). ### Validation of the results at fourth-order We now validate our result using the same quantifiers as in § [7.2.1](#sec:delta-neq-zero){reference-type="ref" reference="sec:delta-neq-zero"}. Population histories are shown in Fig. [\[fig:pops\]](#fig:pops){reference-type="ref" reference="fig:pops"} for two different sets of parameters. It is seen that the fourth-order ME (black dashed line) yields a coarse-grained version of the exact result (full gray line) which is accurate in reproducing the Raman oscillations between levels of the relevant subspace. On the contrary the second-order ME (dashed magenta line) clearly shows a discrepancy in the oscillation frequency. This error appears clearly in the long-time fidelity \(F\), shown in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(a-b), and the fidelity of post-selected protocols, shown in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(c-d). In the same figures, we also compare the fourth-order ME with the approximation scheme proposed in Ref.  (green curves) and the two approaches are seen to coincide. Therefore the ME provides a systematic approximation scheme, overcoming the last point of the criticism to standard AE mentioned in § [7.1](#sec:AE-ambiguities){reference-type="ref" reference="sec:AE-ambiguities"}. In particular for the symmetric choice \(|\Omega_0|= |\Omega_1| = 0.3 \, \Delta\) (Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}a, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}c) the fourth order fidelity oscillates between one and \(0.8\) (black dashed curve), whereas the second-order result (magenta dashed curve) decays to lower values because the error in frequencies accumulates in time. The same behaviour is obtained for the asymmetric configuration of the drive amplitudes shown in Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}b, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}d. For these time scales the error in the fourth-order \(F\) in Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(a-b) does not accumulate in time and Figs. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}(c-d) show that it is entireka:16-vepsalainen-photonics-squtritly due to leakage since it can be corrected by post-selection. Again we notice that in Figs. [\[fig:pops\]](#fig:pops){reference-type="ref" reference="fig:pops"}, [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"} we used parameters with values far beyond the perturbative regime to magnify the errors. Still, errors are not so large and in particular, they are remarkably small for the post-selected dynamics. As in Fig. [\[fig:pops2\]](#fig:pops2){reference-type="ref" reference="fig:pops2"}c, errors due to leakage in Fig. [\[fig:fids\]](#fig:fids){reference-type="ref" reference="fig:fids"}b (black curve) almost disappear already for values \(\sqrt{| \Omega_0 \, \Omega_1 | /( 2 \, \Delta^2)} \ll 0.1\). Finally, we stress the agreement of the ME result with the results of the approach of Ref. . This latter achieves a high-level accuracy using an iterated Lippmann-Schwinger equation in a special gauge defining the interaction picture and supplementing the problem by an ad hoc assumption of Markovianity of the dynamics. Our approach based on the ME shows that the correct result only leverages on coarse-graining. # Discussion and conclusions We introduced a technique based on the Magnus expansion that allows deriving a coarse-grained effective Hamiltonian and yielding a simplified model for the low-energy dynamics of the system. We applied the technique to the problem of the three-level lambda system clarifying ambiguities and inconsistencies of standard AE, which have been raised in the literature . Results from ME accurately interpolate all the limiting cases obtained by standard approximations. For instance, the second-order ME yields a result reducing to the usual AE for \(\delta \ll |\Omega_k|\) but reproducing the perturbative Stark shifts for \(\delta \gg |\Omega_k|\). Moreover, the accuracy of the ME can be systematically increased reaching a full agreement with other accurate approximation schemes such as the one developed in Ref. . Since the ME-based approximation we propose only relies on coarse-graining, this latter is identified as the key ingredient underlying all the approximations. Coarse-grained based approximations yield effective Hamiltonians reproducing accurately the dynamics in the relevant subspace for a wide range of parameters, much beyond the perturbation regime. In particular, it is remarkably large the accuracy achieved for protocols supplemented by post-selection. Notice that the ME approach we described yields a block diagonal effective Hamiltonian which does not cancel the not-relevant subspace but treats it consistently. In particular, for the Lambda system we obtain the correct perturbative energy shift for the level \(\ket{2}\). This result allows non-trivial applications to problems involving multiphoton processes in three-level dynamics  which are relevant for newly developed quantum hardware . Finally, the approach can be extended to the search for effective Hamiltonians in time-dependent problems. The extension is straightforward if the parameters of the Hamiltonian are slowly varying on time scales of the order of the coarse-graining time \(\tau\) as in Ref. . [^1]: An efficient way to exploit the Magnus expansion is to look at the evolution operator \(U(t,t_0)\) in a time slice \(\tau = (t-t_0)/N\). [^2]: If \(H\) describes a three-level atom under the action of corotating external fields \(S_k\) are the perturbative Stark shifts. [^3]: Up to a phase factor depending on the modality of time slicing (see Appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"}). [^4]: An efficient way to exploit the Magnus expansion is to look at the evolution operator \(U(t,t_0)\) in a time slice \(\tau = (t-t_0)/N\). [^5]: If \(H\) describes a three-level atom under the action of corotating external fields \(S_k\) are the perturbative Stark shifts. [^6]: Up to a phase factor depending on the modality of time slicing (see Appendix [\[app:integrals\]](#app:integrals){reference-type="ref" reference="app:integrals"}).

{'timestamp': '2022-12-19T02:13:02', 'yymm': '2212', 'arxiv_id': '2212.08508', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08508'}

# Introduction {#sec:intro} Fluid systems are necessary to provide drinking water, to cool chemical processes, to dispose waste water or to perform dosing tasks in food production. The manufacturers, planners and operators of such systems are confronted with a range of different challenges. The operation should be uninterrupted and energy-efficient, with minimum investment and personnel costs. The functionality should always be guaranteed and a change in the system or the boundary conditions should not affect it. In summary, this results in the five main challenges for fluid systems: 1. *Functionality*: The functionality of fluid systems and their control systems is to ensure the specified demand, i.e. a certain pressure or volume flow at specific points in the system. For this purpose, the speed or on/off state of the pumps and the valve positions can be adapted. The functional quality of the system is determined by the deviation of the actually provided volume flow from the control objective. 2. *Low effort regarding energy consumption*: Pumps and associated fluid systems are responsible for about \(8\,\%\) of the total electrical energy consumption, in industry for up to \(25\,\%\). For this reason, the energy efficiency of fluid systems is crucial to ensure the economic efficiency of processes and to achieve the sociopolitical goal of reducing \(\mathrm{CO}_2\) emissions. In operation, up to \(20\,\%\) of the energy consumption can still be saved by considering the entire system. Thus, the overall control objective is to fulfill the functionality with as little energy effort as possible. In principle, it is favorable if the valves are opened as far as possible to keep the required hydraulic power low. However, the demand must be met at all valves. In addition, pumps should run at the lowest possible speed and, if possible, the pumps with the highest efficiency in the operating point should be active. 3. *Low effort regarding implementation*: Commissioning requires tuning of the system and design of controllers. If there is a change in the system, e.g. due to wear or changed boundary conditions, a new tuning in the system is necessary to ensure the functionality and energy efficiency. Since fluid systems are usually highly customized, these steps have to be carried out individually, which is time-consuming and costly. Thus, when designing fluid systems and their controllers, modelling effort should be minimized. 4. *High availability*: The above examples show that fluid transport is indispensable for many processes. A malfunction often affects entire production plants or leads to the failure of critical infrastructure, which is why robust systems with high availability are necessary. This is generalized as the ability to fulfil the system functionality even in the case where a disruption occurs. 5. *High acceptability*: Since fluid systems are vital for many social needs and industrial purposes, transparency, traceability and comprehensibility of control tasks are critical for acceptability of the systems. Furthermore, the flexibilisation of production, as it is being advanced in the context of Industry 4.0, demands a high adaptability of technical systems. Fluid systems are not exempt from this. Thus, fluid systems will have to react to different and changing boundary conditions in the future. Fluid system components need to be flexible enough for them to be used in a variety of scenarios. Thus, portable, flexible and scalable fluid systems and controller designs are an important challenge. To meet the five challenges, an appropriate control architecture and operation strategy are crucial. In general, three approaches for the control architecture can be distinguished, as shown in Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"}: (a) a decentral, local control, (b) a central, system-wide control, and (c) a distributed system-wide control. Each of these architectures is presented in the following. In conventional control approaches, the control of the pumps and valves is usually separated, i.e. a local control according to Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"} (a). The underlying concept is that the pumps ensure a certain pressure difference (possibly as a function of the volume flow) and the individual valves at the consumers have a local control for a volume flow, filling level or temperature (e.g. in heat exchangers such as thermostatic valves in a heating system). The pressure applied by the pumps always has to be high enough to enable the valves to operate and to reach the desired setpoints. With local control (a), which is also referred to as *decentralized* control, the system is separated into subsystems, each supplied with a local controller. The subsystems may be constituted only of individual components. The optimal control objectives are pursued on the local level, while coupling between the subsystems is disregarded . On the one hand, this reduces the effort required for tuning and a minimum function is often ensured even in the case of malfunctions. On the other hand, however, the potential for energy savings given by including information on the system as a whole cannot be fulfilled when components are controlled locally. In contrast to this, there are system-wide or *centralized* control approaches, which draw on the control inputs from the entire system and calculate the control output for each component to achieve a global optimum for the common objective . An example is model-predictive control, in which an optimization is carried out based on a substitute model and thus the optimal control variables are estimated. As shown in , this can significantly reduce the energy demand. However, this poses a different challenge--the high communication requirements due to the merging of all measured and controlled variables and the development, implementation and maintenance of the model. This is often not worthwhile for highly individualized systems such as a fluid system, which is why these approaches are not very popular in practice. Ideally, a solution is simple to implement and yet efficient and robust, i.e. combines the advantages of variants (a) and (b). One such solution is *distributed* control, where local controllers coordinate actions to enable the system as a whole to meet the global control objective, as shown in Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"} (c) . These local controllers can be modelled as agents which jointly constitute a multi-agent system. In these systems, each component is assigned a so-called virtual agent, which acts as its control and communication unit. Each agent has domain-specific knowledge, decision rules and goals. Furthermore, agents can read sensors and control actuators at the associated component. The agents can thus perceive and influence the environment, for which they communicate with each other and exchange available knowledge on their respective perceptions and actions. In the process, the agents make independent and autonomous decisions to perform actions. The aim is to achieve system-wide targets as a result of the combined effects of these actions without a central control unit. Conflicting goals are resolved through virtual negotiation between the agents. Consequently, distributed control architectures based on multi-agent systems promise great potential for application to fluid systems. However, it is necessary to ascertain the suitability of distributed agent control for addressing the five key challenges presented above. This entails a systematic study and assessment of the performance of distributed multi-agent systems for control of fluid systems. For this, both the system-wide rules and procedures as well as the behavior of the individual agents need to be defined. Furthermore, an appropriately generic fluid system needs to be selected together with a suitable usage scenario. Finally, benchmarks and assessment criteria need to be defined to gauge to what extent the challenges are addressed by multi-agent systems. These steps were taken within the scope of the presented work and are detailed in the following. First, however, the working hypotheses are presented together with the concrete research questions. ## Research Question {#sec:rq} When considering distributed control of fluid systems based on multi-agent systems there is a plethora of variations of implementation. These variations need to be systematically assessed and sorted into a schema determined by their characteristics regarding a set of requirements. Following from the above, the requirement for distributed multi-agent control of fluid systems is to ensure the functionality. While this holds for all control designs of fluid systems, the expected behavior of multi-agent control with regard to the challenges \(2\) through \(4\) differs from that of conventional control. Three distinct hypotheses on the expected deviation are formulated concerning each of the three challenges. **Hypothesis 1:** Minimum energy consumption is reached when all components of a fluid system are controlled centrally using mathematical global optimization methods. Transitioning from central to distributed agent-based control comes at the cost of increased energy consumption. Nevertheless, agent-based control still uses optimization methods and consumes less energy than conventional local control methods. Central optimal control and local control constitute lower and upper boundaries respectively for energy consumption of fluid systems under distributed multi-agent control. **Hypothesis 2:** Creating substitute models for highly individual and evolving fluid systems involves prohibitive effort. Fluid systems can be controlled adequately using control architectures that dispense with manually created substitute models using distributed agents. This allows for adaptability to changed boundary conditions, as well as transfer of agents between different applications and settings. **Hypothesis 3:** In both central and distributed control, the performance of the system's control depends on the communication infrastructure. Failures in the communication infrastructure adversely affect the performance of central control architectures more severely than distributed control architectures. Thus, distributed control increases the availability of fluid systems. Three design approaches for distributed control of fluid systems are selected and examined thoroughly in the light of the requirements addressed above. These design approaches are (i) distributed model predictive control, (ii) multi-agent deep reinforcement learning, and (iii) market mechanism design. Furthermore, three concrete research questions guide the design of this study to achieve an assessment and comparison of the three design approaches. These questions are: > - *How well do the controllers of each of the design approaches fulfill the functionality and the objective of minimum energy consumption?* > > - *What influence does the available information from substitute model and information exchange between agents have with regard to functionality and effort?* > > - *How well do the distributed approaches cope with a disruption in the communication between agents in comparison to centralized approaches?* Focusing on these research questions with the chosen design approaches, the presented work aims to shed light on the hypotheses outlined above. The approaches were chosen after a thorough examination of the literature. Relevant previous studies are presented in the following section. ## Related Work Multi-agent systems and agent control of technical systems have been widely studied as approaches to distributed control. The following section will first give an overview of multi-agent systems applied to technical systems in general. It will then present various studies investigating multi-agent systems applied to specific technical systems. Furthermore, the section will present work related to the three approaches to distributed control investigated in the scope of the presented study. ### Agent-based Control of Technical Systems Studies presented in the 1980s investigated distributed problem-solving , with communication and interaction between agents playing an important role in early studies of multi-agent systems . This was further investigated by Jennings  and applications of the theory of agent technology in practice discussed by Jennings and Wooldridge  while approaches to implementing agents in software and programming were also presented by Genesereth and Ketchpel  as well as Shoam  in the 1990s. Evans et al. formalized the implementation into a methodology for engineering systems of software agents (cited in ). Since then, extensive text books have been published introducing multi-agent systems, distributed artificial intelligence, and applications, summarizing the state of the art . Furthermore, guides to assessing suitability and deciding on use of multi-agent systems have been published by Bogg et al.  and Beydoun et al. . A more recent comprehensive survey introducing agents and multi-agent systems as well as discussing in detail applications and challenges was presented by Dorri et al. . Among the applications of multi-agent systems mentioned by Dorri et al. are cities and the built environment, i.e., infrastructure systems and buildings . Negenborn studied these applications to infrastructure, both in transportation networks (Negenborn et al. ) and power networks . Faced with decentralization connected with the use of renewable energies, power networks are being transformed to smart grids. How to control these using multi-agent systems was explored by Zimmermann et al. . Algarvio et al.  designed a multi-agent system for hydroelectric power plant control in an energy market with high proportion of renewable energies. Multi-agent system application to control of water distribution networks have also been studied . While not explicitly using multi-agent systems, two studies have investigated distributed control of sewage systems and waste water treatment plants and compared performance of distributed and centralized control systems . Turning from infrastructure systems to buildings, multi-agent systems have been studied extensively in the application to heating, ventilation and air conditioning (HVAC) systems. Simulations and a field test of multi-agent control of a heating and ventilation system for a commercial building were carried out and compared by Constantin et al. , while van Pruissen et al.  compared multi-agent control of a heating system in a commercial building to conventional, centralized heating system control. A dissertation by Huber showed that modularized agents could be combined for automated HVAC systems . Looking beyond functionality, two studies focused on energy cost reduction using multi-agent systems in various configurations for control of HVAC systems . Azuatalam et al. considered whole-building HVAC agent control with a special focus on-demand response , while Nagarathinam et al. investigated scaling of multi-agent systems in building HVAC with multiple units including water-side control of air-handling-units and distributed agents . HVAC systems in buildings demonstrate significant similarities to water distribution systems in buildings. Nevertheless, building automation and multi-agent control for pumping stations and valves have not been studied hitherto. A challenge lies in finding an approach that allows studying the performance of various designs of multi-agent systems for fluid system control in a generalized manner. This allows to transfer promising results to other fluid systems, e.g., industrial cooling circuits, booster stations of high-rise buildings, and urban water distribution systems. While distributed control using multi-agent systems in general has been applied to technical systems in numerous studies, approaches to designing the multi-agent system can vary considerably. The presented work compares three such approaches commonly found in the related literature, distributed model predictive control (DMPC) from the domain of control technology, multi-agent deep reinforcement learning (MADRL) from the domain of machine learning, and market mechanism design from the domain of game theory. Prior works of all three approaches will be presented in the following. ### Distributed Model Predictive Control Model predictive control (MPC) has a long history reaching to the 1970s and 1980s with receding horizon feedback control  and generalized predictive control . A survey by Qin and Badgwell  not only gives an idea of industrial applications of MPC but also an overview of notable works throughout the history of MPC, and the interested reader is referred to that work for further detail. An early survey by García et al. describes MPC as using an explicit model that can be identified separately . One of the major challenges connected with MPC is to attain that system model. Various approaches exist, roughly classifiable as analytical and data-driven. Analytical models require physical modelling of the system and make the inner structure of the system transparent. One approach is to identify the transfer function of the system with, e.g., a step response or impulse response . Other approaches are to derive a non-linear dynamic model of the system and linearize it for use in MPC . Data-driven system models do not focus on the inner structure of the system, but rather derive a model from input-output-values of a black-box . How to derive black-box models from input-output measurements is detailed by Ljung . One such approach are ARX models (AutoRegressive models with eXogenous inputs) as they are structurally similar to state space models and the parameters can be identified through convex optimization . A further black-box approach to deriving a system model is Gaussian Process Regression . Both approaches, physical modelling and models derived from input-output data, have also been combined in the use-case of building modeling . Finally, an effective approach has been to train neural networks based on empirical input-output data and use these as surrogate system models . This has been done for chemical processes  as well as industrial process control  and automotive control . It has been argued that applying centralized MPC to large systems can cause problems due to data transmission requirements, computational requirements and considerations of centralized controllers posing a single point of failure . Alternative approaches are decentralization and distribution, where decentralization means dividing a system into subsystems controlled on a local level without strong considerations of coupling whereas distribution entails local controllers of subsystems cooperating for a common, system-wide objective . Distributed approaches stem from studying parallel and distributed computing . Combining multi-agent systems with distributed model predictive control to solve optimization problems in large systems has been reported . To summarize, MPC is a well-studied control approach for centrally computed optimal control strategies. DMPC allows to ensure cooperation between local control units of subsystems while still using optimization methods for calculating the control strategy. Both approaches require a substitute model of the entire system. The modeling effort may be reduced by applying machine learning methods for deriving this model. ### Multi-Agent Deep Reinforcement Learning A different approach has been to avoid modeling the underlying system. Here, machine learning approaches are implemented directly in the agents controlling the system. These agents learn based on samples generated through trial and error. In this approach, reinforcement learning has been used . This technique is extensively presented in the work of Sutton and Barto . Reinforcement learning is a machine learning approach in which the agent is rewarded for executing desirable actions within its environment . An overview is given by Zhang et al.  of various algorithms combining multi-agent systems with reinforcement learning based either on Markov games or extensive-form games in cooperative, competitive or mixed settings. The perception of various agents interacting with one another in a Markov game goes back to Littman . In a survey, Li presents works showing how reinforcement learning can be extended to deep learning . A further enhancement of general machine learning techniques, deep learning is a technique that requires less human engineering input in designing feature layers. Instead, it uses generalized machine learning techniques for learning these features from sample data . Deep learning is extensively described by Goodfellow et al. in the Deep Learning Book . Foerster et al. combined deep learning and reinforcement learning for machine learning of communication protocols for multi-agent systems . An overview and critical assessment of multi-agent deep reinforcement learning (MADRL) is presented by Hernandez-Leal et al. with regard to current state of the art, recent advances, lessons learned, and challenges for future implementations . To make the learning process of agents more efficient, attention weights can be introduced, as shown by Vaswani et al. . A further approach to enhancing the learning process of agents has been to separate them into actors and critics, where actors learn the decision policies and critics learn the value function . This has been applied to multi-agent systems with reinforcement learning with variations . Here, Iqbal and Sha used a configuration for decentralized actors while critics learn a global value function subject to attention weights, the multi-agent actor-attention-critic algorithm (MAAC) . Further adaptations have been to decentralize the critic  as well as giving critics separate hierarchical attention levels for the agent system and the individual agent . In machine learning, agent systems are modelled as Markov games. MADRL is a promising approach for relying on fully data-driven methods for control of fluid systems. MAAC can be used for an effective training process of agents. ### Market Mechanism Design The third approach to distributed control of fluid systems considered in this study is market mechanism design. Mechanism design is a field of research in the domain of game theory and economics , yet has also been presented alongside other approaches of distributed artificial intelligence . While game theory is commonly associated with agent systems as a modelling approach in social sciences, it bears similarities to control systems with multiple distributed controllers . According to Marden and Shamma, the difference in usage in the two disciplines is that the social sciences use it as a \"descriptive\" method whereas in engineering it is a \"design\" method . They further present some examples where a global objective function serves to ensure a desirable emergent behavior resulting from individual agents' permissible actions in engineering problems, such as synchronization, distributed routing, sensor coverage, wind energy harvesting , vehicle target assignment, content distribution, and ad-hoc networks . Furthermore, for distributed control using game theoretic approaches, the utility functions of individual agents need to be defined in such a way that the agents' actions contribute to the global objective function . Finally, when the desirable outcome is defined according to the global objective function, the interaction protocol between the agents needs to be defined in such a way that the agents learn to converge towards that desirable outcome . The process of designing such a protocol is known as mechanism design . There is a variety of choices for the interaction protocol, which can be roughly categorized as negotiations and auctions . According to Wooldridge, negotiations are more generally applicable than auctions as they allow reaching a common agreement in a wide range of settings, whereas auctions only consider the problem of allocating goods . Various agent languages exist for facilitating the required semantic communication between agents during negotiations, e.g., KIF, KQML, one of them being the protocol standardized by the Foundation of Intelligent Physical Agents (FIPA) . Bellfemmine et al. implemented this protocol in a Java software framework  which was also used in the application of multi-agent systems to HVAC control . Implementing this complex common communication framework requires considerable effort, whereas auctions offer the advantage of being simple to implement as their scope is restricted to allocating goods . Nevertheless, they still offer a wide range of possible configurations for ensuring that the desired outcome is the goal towards which the agents converge, such as English, Dutch or Vickrey auctions and the Vickrey-Clarke-Groves mechanism among others . As the problem of fluid system control fundamentally consists of allocating volumes of fluid to various consumers, the approach of designing a market mechanism is promising due to its simplicity. The approach does not require extensive physical modeling of the fluid system. To summarize the above, agent-based systems are a promising approach to distributed control and have been applied to various technical systems related to fluid systems as well as fluid systems themselves, e.g., water distribution systems. Furthermore, a plethora of work has been carried out investigating various approaches to designing multi-agent systems for distributed control out of which three main strands can be identified: (i) distributed model predictive control, (ii) multi-agent reinforcement learning, and (iii) game theory and mechanism design. Accordingly, the presented work will investigate the application of distributed model predictive control, the multi-agent actor-attention-critic algorithm for multi-agent deep reinforcement learning as well as a market mechanism for distributed fluid system control. Each of these approaches will be further detailed in Section [2](#sec:matmet){reference-type="ref" reference="sec:matmet"}. ## Structure of the Paper Having outlined the motivation, illustrated the research question and presented the related work, the remainder of the paper is structured in the following way. Section [2](#sec:matmet){reference-type="ref" reference="sec:matmet"} will first present the model of the fluid system in general terms, as well as the system topology for the use-case. It will then present how fluid system components are modelled as agents. Finally, it will present in detail the methods and design of the implemented control approaches while also discussing the classification of these approaches with regard to transparency, flexibility, observability and modeling effort as well as illustrating the approach to modelling disruptions in the communication network. Section [3](#sec:res){reference-type="ref" reference="sec:res"} will present the results of the study in the following order. First, addressing research question *(i)*, the overall performance of the three approaches to distributed control is presented. Next, findings concerning the influence of model information and information exchange are shown as required by research question *(ii)*. Finally, research question *(iii)* is addressed with results presented for performance of the two distributed systems under disruption. In section [4](#sec:disc){reference-type="ref" reference="sec:disc"} the results are discussed further regarding both the concrete research questions and the hypotheses while also reflecting on lessons learned and further research directions. The paper is completed by a summary and concluding remarks. # Materials and Methods {#sec:matmet} In the following, the modelling of the fluid system and the investigated application will be presented. Subsequently, it will be explained how the agents for the different components are modelled before introducing the different control approaches of the multi-agent system. ## Fluid System Model {#sec:fsm} A fluid system, e.g. a heating circuit, water supply system or industrial cooling system, consists of the following main elements: (i) the active components pumps and valves and (ii) passive components such as pipes and generic resistances. Pumps add hydraulic power to the system, thereby increasing pressure and delivering a volume flow. Valves reduce the pressure and dissipate hydraulic power, mostly for control purposes. As active components, they are adaptive: Pumps are able to switch on and off or adjust their rotational speed; valves are able to adjust the pressure loss by changing the valve position. Pipes, which are passive elements, connect the components and thus determine the topology, whereby pressure losses occur during the transport of the fluid. Generic resistances, such as heat exchangers, pipe elbows, etc., cause pressure losses depending on the volume flow in the system. The system behavior, i.e. the volume flow and pressure at different points as well as the power consumption, can be described depending on the components, topology and control variables (speeds, valve positions) via a non-linear system model. ### Physical-technical description of the system {#sec:phys_mod} The components are described by their characteristic curves, the machine or system curves. These are shown qualitatively in Figure [\[fig:component_models\]](#fig:component_models){reference-type="ref" reference="fig:component_models"}. A pump increases the pressure from the inlet \(p_\textrm{in}\) to the outlet \(p_\textrm{out}\) by \(\Delta p_\textrm{pump}=p_\textrm{out}-p_\textrm{in}\) depending on the volume flow \(Q\), speed \(n\) and its type (expressed by parameter \(\alpha\)). This adds energy to the system. The pressure increase (=head) of a centrifugal pump can be expressed in general by: \[\Delta p_\textrm{pump}= \alpha_1 Q^2+ \alpha_2 Q n + \alpha_3 n^2 \label{eq:pump_char}\] The volume flow thus depends on the necessary pressure increase, i.e. the resistance of the system, which causes an interdependence of the components. The power consumption \(P_\textrm{pump}\) is a function of the volume flow \(Q\) and the speed \(n\): \[P_\textrm{pump}= \beta_1 Q^3+ \beta_2 Q^2 n+ \beta_3 Q n^2 + \beta_4 n^3 + \beta_5 \label{eq:pump_char_power}\] with the parameter \(\beta\) depending on the pump type. If a frequency converter is installed, the speed can be set between minimum and maximum speed or the pump can be switched off: \(n \in \{ 0\} \cup [n_\textrm{min},n_\textrm{max}]\). A valve serves as a variable resistance in the system. The pressure reduction (=pressure losses) in a resistance depends quadratically on the flow velocity \(w\): \[\Delta p_\textrm{loss}=-\frac{1}{2} \varrho \zeta w^2=-\frac{1}{2} \varrho \zeta \left( \frac{Q}{A'}\right)^2=-l Q^2\] Here, \(\varrho\) is the density of the fluid, which in the scope of this study is assumed to be water. \(\zeta\) represents the pressure loss coefficient, which is constant for a component and \(A'\) is the cross-section. The constant terms can be combined to the parameter \(l\). By changing the valve position \(v\), the pressure loss coefficient or cross-section can be adapted, causing \(l\) and thus the pressure loss to change. For a valve, \(l=l(v) \in [ l_\textrm{min},\infty)\) applies, i.e. there is a certain pressure loss even when the valve is fully open. When it is fully closed, \(l(v)\) becomes arbitrarily large. The passive components in the system also act as a resistance, but they are not adjustable, i.e. \(l_\textrm{passive} = \textrm{const.}\) The system behavior is determined by the components: The valves and passive components cause pressure losses, which are compensated by the pumps. The volume flow rates in the system depend on the cumulative resistances. This creates a complex dependency. It can be described by the conservation of energy and mass. In the scope of this study, a model in Dymola, which is based on Modelica, is used to simulate the system's behavior. The control task is to select the pump speeds and valve positions in such a way that all demands (volume flow rates and pressures at specific points in the system) are satisfied and the energy consumption is as low as possible. Due to the complexity of fluid systems, heuristics are necessary for the solution, as they are considered in the context of this manuscript. ### Use-Case As an application, we consider the water supply of a building, for which a decentralized pressure booster station is used. For the scaling of the use case, an existing test rig for decentralized Booster stations is used as a reference, since it is intended to validate the results experimentally in the future, cf. Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (a). An in-detail description of the test rig can be found in. From this, we derive a system with five valves and two pumps, cf. [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (b). Each valve represents a pressure zone, i.e. several floors in the building. The water demand on these floors is summarized in the demand of the valves. For the evaluation, a 30-minute load profile is used, which simulates the real water demand during 24 hours. For this purpose, the demand distribution for central booster stations from is used and disaggregated into the demand of the different floors. The resulting water demand for the five valves is shown in Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (c), (i). Furthermore, a short load profile of 30 seconds is used to analyze the behavior of the agents, cf. Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (c), (ii). ## Modelling Fluid System Components as Agents The aim of a control system is now to select the control variables in such a way that a common goal is achieved in the best possible way: to fulfil the volume flow requirements while consuming as little energy as possible. In distributed control, this is achieved by designing local component controlers as well as rules for their interaction for system-wide control. This section presents the component level control design, while Section [2.3](#sec:mas){reference-type="ref" reference="sec:mas"} will describe the control methods for system-wide control. The active components of the studied system, pumps and valves are each assigned an agent. All agents have sensors for observing the environment and actuators for influencing the environment. The disparate nature of these components requires defining respective types of agents. The following will present the modeling of each of the agent types used in the presented work. ### Pump Agent The pump agent is responsible for the control of a pump. Depending on the application, the pump can have a controller that sets a target volume flow or be controlled directly by setting a value for relative speed. The pump can either be switched off, which corresponds to a manipulated variable of zero, or operated in a range between minimum speed \(n_{i,\text{min}}\) and maximum speed \(n_{i,\text{max}}\). The manipulated variable range is thus \(n_{i} \in \{0\} \cup [n_{i,\text{min}}, n_{i,\text{max}}] = \mathbb{P}\). In case the pump is controlled directly by the agents, the manipulated variable \(u\) corresponds to the pump speed: \[u_{i,t} = n_{i,t}, \quad \forall t \in \mathcal{T}, i \in \mathcal{P},\] with the set of all time steps \(\mathcal{T}\) and the set of all pump agents \(\mathcal{P}\). In case the pump is controlled by a volume flow controller (e.g. a PI-controller) and the agents set the target volume flow, the manipulated variable is \[u_{i,t} = k_{\mathrm{pump}}(Q_{i,t}), \quad \forall t \in \mathcal{T}, i \in \mathcal{P}.\] Here, \(k_{\mathrm{pump}}: \mathbb{R}_{0,+} \rightarrow \mathbb{P}, Q_{i,t} \mapsto u_{i,t}\) is a controller function which maps the required volume flow to a value for the manipulated variable. Other controllers may be necessary for the motors inside the pump to implement the value of \(u_{i,t}\), but are not considered within the scope of this work. The observations of pump agents can be volume flow rate as well as electrical power. The electrical power depends on the speed and the volume flow and usually follows a non-linear characteristic curve (cf. [2.1.1](#sec:phys_mod){reference-type="ref" reference="sec:phys_mod"}). The deliverable volume flow also depends on the pressure difference. Since the aim of the pump is to deliver the required volume flow at minimum electrical power, only the electrical power is mapped with the cost function. A minimum electrical power \(P_{i,\text{min}}\), even if the rotational frequency is zero, comes from features such as a display or standby functionality. The maximum electrical power \(P_{i,\text{max}}\) provides a limit on the pump's performance. In order for the cost of the agents to be in a similar range, the cost is already related to the maximum power in the cost function. Weighting against the costs of the pumps against each other or against the valve agents can then be done with the dimensionless weighting factor \(\lambda_i\). Thus, the costs are expressed as \[\begin{aligned} c_{\text{pump},i,t} & = \lambda_i \cdot \left(\frac{P_{i,t}-P_{i,\text{min}}}{P_{i,\text{max}}} \right) ^2, \quad \forall t \in \mathcal{T}, i \in \mathcal{P}. \label{eq:Kosten_Pumpe} \end{aligned}\] ### Valve Agent The valve agent is responsible for controlling its valve. Similar to the pump agents, the valves can have a volume flow controller or be controlled directly by the valve agent setting the valve opening \(v\). The range of the valve opening is \(v_{i} \in [v_{i,\text{min}}, v_{i,\text{max}}] = \mathbb{V}\). In case the agent directly manipulates the valve opening, the relation is \[u_{i,t} = v_{i,t}, \quad \forall t \in \mathcal{T}, i \in \mathcal{V},\] with the set of all valve agents \(\mathcal{V}\). Analogous to the pump agents, in case the valve is controlled by a volume flow controller (e.g. a PI-controller) and the agents set the target volume flow, the valve opening is \[u_{i,t} = k_{\mathrm{valve}}(Q_{i,t}), \quad \forall t \in \mathcal{T}, i \in \mathcal{V},\] with \(k_{\mathrm{valve}}: \mathbb{R}_{0,+} \rightarrow \mathbb{V}, Q_{i,t} \mapsto u_{i,t}\), the controller function which maps the required volume flow to a value for the manipulated variable. Relevant variables for the valve agents are the volume flow as well as the differential pressure. A volume flow or pressure demand is registered directly with the valve agents.\ Therefore, the cost function consists of the deviation between the target volume flow and the target pressure. Since only the volume flows are considered in this work, the cost function for the valve agents with the volume flow demand \(Q_{i,t,\text{demand}}\) and the actual volume flow \(Q_{i,t,\text{actual}}\) results in \[\begin{aligned} c_{\text{valve},i,t} & = \lambda_i \cdot (Q_{i,t,\text{demand}}-Q_{i,t,\text{actual}})^2, \quad \forall t \in \mathcal{T}, i \in \mathcal{V}. \label{eq:Kosten_Ventil} \end{aligned}\] The cost factor \(\lambda_i\) is given for the valve agents \(i \in \mathcal{V}\) in units \((\frac{1}{m^3/h})^2\) to make the cost a dimensionless quantity. Squaring the deviation from the target flow rate is also intended to ensure that low total costs correspond to a fair solution for all valves. Agents should not choose to disregard a valve whose target fulfilment has a high cost. ## Control Methods of Multi-Agent System {#sec:mas} Shifting the focus now from the component level to the system level, the following section will present the three separate control methods implemented, which ensure the agents interact to reach the target of the control task. It further includes a description of conventional control methods. Finally, various characteristics will be introduced which will serve to assess the implemented approaches with regard to aims discussed in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}. ### Distributed Model Predictive Control {#sec:dmpc-meth} The method for distributed model predictive control implemented in the presented work follows the work of Pannocchia closely and can be found in detail there . The most fundamental equations and basic algorithms are outlined in the following. Two assumptions are made: *(i)* the system state is equal to the output and *(ii)* the system dynamics are neglected, i.e., the system state does not depend on the preceeding system state. Applying the assumptions to the non-linear fluid system described above yields the non-linear steady state system \[\mathbf{y} = g(\mathbf{u}),\] with the steady state model \(g: \mathbb{V}^{\lvert \mathcal{V} \rvert} \times \mathbb{P}^{\lvert \mathcal{P} \rvert} \rightarrow \mathbb{R}_{0+}^{{\lvert \mathcal{V} \rvert} \times {\lvert \mathcal{P} \rvert}}, \mathbf{u} \mapsto \mathbf{y}.\) Here, \(\lvert \mathcal{V} \rvert, \lvert \mathcal{P} \rvert\) describe the cardinality of the set of valve agents \(\mathcal{V}\) and pump agents \(\mathcal{P}\), respectively. The vector \(\mathbf{u} = [u_1, u_2,..., u_A]^T\), with \(A = \lvert \mathcal{A} \rvert\) the number of agents, contains the manipulated varibles of all agents \(\mathcal{A} = \mathcal{V} \cup \mathcal{P}\). The vector \(\mathbf{y} = [y_1, y_2,..., y_A]^T\) contains the output of all agents, i.e. values of volume flow and power for valve and pump agents respectively. For attaining \(g\), the substitute model of the fluid system, a neural network of the type multi-layer perceptron with a sigmoid function was programmed using the python library *scikit-learn*. Using the Dymola model of the use-case test rig, randomized combinations of input data with \(u_{i,\mathrm{min}} \leq u_i \leq u_{i,\mathrm{max}}\) and the corresponding outputs were generated as a set of training data for the neural network, where 80% served as training data and 20% as validation data. Model predictive control requires consideration of system dynamics which contradicts the second assumption made above. However, by introducing a restriction in the values the manipulated variable of each agent can take in each time step \(u_{i,t}-\delta \leq u_{i,t+1} \leq u_{i,t}+\delta\), the concept of prediction can be introduced into the steady state model. This concept is detailed in the following. The vector of manipulated variables of all agents \(i \in \mathcal{A}\) at time \(t \in \mathcal{T}\), the set of all time steps, can be written as \[\mathbf{u}_t = [\mathbf{u}_{i,t}, \mathbf{u}_{-i,t}]^T,\quad i,-i \in \mathcal{A}, t \in \mathcal{T},\] where \(i\) denotes Agent \(i\) and \(-i\) denotes all agents apart from Agent \(i\). The manipulated variable vector for each agent consists of \(L\) entries with \(L\) as the number of predicted control steps (also known as prediction horizon), such that \[\mathbf{u}_{i,t} = [u_{i,t} (0), u_{i,t} (1),..., u_{i,t} (L-1)]^T,\quad i \in \mathcal{A}, t \in \mathcal{T}.\] Only the first value is implemented, though, before a new control step calculates \(\mathbf{u}_{i,t}\) afresh with a receding horizon. The substitute model can then be written as \[\mathbf{\hat{y}}_{t+1} = g([\mathbf{u}_{i,t}, \mathbf{u}_{-i,t}]),\] with the predicted output of the next time step \(\mathbf{\hat{y}}_{t+1}\). Each agent only considers \(\mathbf{u}_{i,t}\) as a variable, while \(\mathbf{u}_{-i,t}\) are considered constants. However, during calculation of each control step, several iterations can be carried out where the agents exchange information about their proposed value of \(\mathbf{u}_{i,t}\). These communication rounds help to improve the overall solution, yet also increase computational effort. All agents have a common objective, with the overall cost function calculated as \[V(\mathbf{\hat{y}}_{t+1}) = \sum_{i=1}^A\ \rho_i \hat{C}_i (\mathbf{\hat{y}}_{i,t+1}),\] with a weighting factor \(\rho_i\), where the cost function of each agent is calculated for the entire prediction horizon, such that \[\hat{C}_i = \sum_{j=1}^L\ c_i(\mathbf{\hat{y}}_{i,t+1}(j))\] with \(c_i\) being the cost function of the pump and valve agents as defined in Eq. [\[eq:Kosten_Pumpe\]](#eq:Kosten_Pumpe){reference-type="eqref" reference="eq:Kosten_Pumpe"} and Eq. [\[eq:Kosten_Ventil\]](#eq:Kosten_Ventil){reference-type="eqref" reference="eq:Kosten_Ventil"} respectively. The overall optimization problem for each agent can then be written as In the presented work, a central model predictive control is also implemented as a benchmark for the distributed model predictive control. The method follows that outlined above with the difference that there is just one agent optimizing the control variables for all components with equal weights of \(\rho_i = 1\). The optimization problem is thus given as ### Multi-Agent Reinforcement Learning The Reinforcement Learning problem is written as a Markov game, consisting of action, states and rewards. The optimization problem of each agent \(i \in \mathcal{A}\) for each time \(t \in \mathcal{T}\) is to find an action that minimizes the costs *States:* Agent \(i\) chooses its action based on observations \(o_i \in \mathcal{O}\), which represents a subset of all states \(s \in \mathcal{S}\). In addition to the states, the agents also observe the volume flow requirements during training. An agent observes volumeflow requirements of valve agents of the set \(\mathcal{V}\). At time \(t\), all observations are \(o_t = \{o_{1,t}, o_{2,t}...,o_{A,t}\}\). Two versions of MADRL are used in this work. In centralized MADRL, each agent \(i \in \mathcal{A}\) has the complete system knowledge. This means that the observed states contain the output variables of all agents, as well as the volume flow requirements of all valve agents: \[o_{i,t} = \{y_{z,t}|_{\forall z\in\mathcal{A}}, Q_{z,t,\text{user}}|_{\forall z\in \mathcal{V}}\}\;.\] In decentralized MADRL, the agents have only limited system knowledge. The observed states are only their own output variables and the volume flow requirements of all valve agents: \[o_{i,t} = \{y_{i,t}, Q_{z,t,\text{user}}|_{\forall z\in \mathcal{V}}\}\;.\] *Actions:* The pump speeds and valve positions are discretized to create a discrete action space, the manipulated variables of an agent \(i\) thus become \[u_{i,t} \in \{u_i^1, u_i^2,..., u_i^M\}, \quad \forall t \in \mathcal{T},\quad i \in \mathcal{A} \;. \label{eq:actions_discrete}\] The number of discrete valve positions or rotational speed is the same for all agents and is denoted by \(M\). The actions of all agents are the control variables \(a_t = \{u_{1,t}, u_{2,t},..., u_{A,t}\}\). *Rewards:* Since training is centralized in MAAC and the rewards are also only necessary for training, the global reward is used for all agents, i.e. \(r_{i,t} = r_{1,t} = r_{2,t} =... = r_t\). To compare the different control methods, the global rewards are formed from the global costs with \(r_t =-c_{\text{global},t}\). All of this provides the training tuple \(\langle s_{t}, a_t, s_{t+1}, r_{t+1} \rangle\). MAAC is an actor-critic method where both policy and value functions are learned. The policy function calculates the action (Actor) and the Value function evaluates that action (Critic). As training progresses, Actor and Critic both improve. The value function is the action value function \(\chi(s_t,a_t)\), which is approximated by a neural network \(\chi_{\psi}(s_t,a_t)\), where the neural network is parameterized by the weights \(\psi\). The policy function is also approximated by a neural network \(\pi_{\theta}(a|s)\) with weights \(\theta\). To stabilize the training, in addition to the neural network just mentioned, so-called target networks \(\chi_{\overline{\psi}}(s_t,a_t)\) and \(\pi_{\overline{\theta}}(a|s)\) are used, whose weights are updated during the training. \(\mathcal{D}\) denotes the experience replay buffer, where all previous experiences (i.e. tuples of states, actions, new states, and rewards) are stored. To make training more efficient, each agent utilizes an attention mechanism, where attention weights are computed that indicate the contribution of the other agent's observations to the action value function. With the attention mechanism, it is possible for the agent to dynamically focus on other agents during training. During the centralized training phase, the critics neural networks minimize the following loss function: \[\begin{aligned} \begin{split} \mathcal{L}_{\chi}(\psi) &= \sum_{i=1}^{A} \mathbb{E}_{(o, a, o', r) \sim \mathcal{D}}\left[\left(\chi_{i}^{\psi}(s, a)- \end{split} \end{aligned}\] \(\gamma\) is the discount factor and \(\varphi\) is the temperature parameter that can be used to adjust the exploration during training.\ For the policy function, a policy gradient method is used with the gradient \[\begin{aligned} \begin{split} \nabla_{\theta_{i}} J(\theta)&=\mathbb{E}_{o \sim \mathcal{D}, a \sim \pi}\left[\nabla_{\theta_{i}} \log \left(\pi_{\theta_{i}}\left(a_{i} \mid o_{i}\right)\right) \varepsilon_{i}\left(o_{i}, a_{i}\right)\right], \\ \varepsilon_{i}\left(o_{i}, a_{i}\right)&=-\varphi \log \left(\pi_{\theta_{i}}\left(a_{i} \mid o_{i}\right)\right) + A_{i}(o, a) \end{split} \end{aligned}\] where \(A_{i}(o, a)\) is the advantage-function, further explained in. In the execution phase, the actions are calculated with the learned policy function from the observations. For a more detailed description of the implementation, the reader may refer to and. ### Market Mechanism In order to achieve the overall system-wide goal while ensuring autonomous decision-making for the agents, a market mechanism is developed as a third approach. A set of market and agent rules are defined which agents have to follow. Within these rules, a market with supply and demand is established. The agents buy and sell guarantees for a certain volume flow. In turn, they must pay for energy costs or are paid to provide a service. For example, after a sale, pump-agents guarantee that they will pump a certain volume flow. To do this, they must estimate how \"expensive\" the necessary pressure increase will be and, if necessary, also purchase flow guarantees from upstream components. The agents act according to the market rules and aim to maximize their own profit. Thus, energy-efficient behavior is encouraged. This corresponds to the basic idea of a free market, which is able to control the complexity of the system even without a perfect system model and central control. The market rules are: 1. volume flow guarantees are the subject of selling/buying 2. each agent must deliver the sold volume flow at its output 3. each agent can only interact with its neighboring agents 4. each agent has a finite budget available per time step, which consists of the costs incurred by not achieving the goal 5. the order of buying is from down-to upstream, starting with the lowest agent In the method, negotiation rounds are triggered regularly. In these rounds, the agents first calculate their demands and offers and second purchase and sell volume flow guarantees. In between the negotiation rounds, there is a control phase in which the agents control the volume flow to the sum of purchased and sold guarantees. The procedure of a negotiation round is shown in Algorithm [\[algo:negotiation\]](#algo:negotiation){reference-type="ref" reference="algo:negotiation"}. In each round, the individual agents buy a volume flow guarantee, for which they request offers from all upstream neighboring agents and buy the best one. Besides the market rules, the behavior of the agents is crucial. The goal for each agent is to maximize its profit. For this purpose, the best possible offers have to be calculated and favorable offers have to be purchased. While the selection of the best offer is trivial, since simply the cheapest offer that meets the demand is selected, the calculation of the offers is more complex. Since only upstream agents are asked for offers and only pump agents are upstream agents in the use-case model, the method for calculating offers is only implemented for pump agents. When a pump agent is asked for an offer, it calculates offers for different volume flows according to Algorithm [\[algo:offers\]](#algo:offers){reference-type="ref" reference="algo:offers"}. On the one hand, the individual costs \(b_{\textrm{ind}}\) and, on the other hand, the costs that arise because guarantees have to be purchased from upstream components \(b_{\textrm{purchased}}\) are taken into account. The individual costs of the pumps arise from the energy required to increase the pressure. The agent estimates the costs for the total guaranteed volume flow. Since the agent sells parts of the total volume flow several times to one or more agents, the buying agent is only charged for its portion of the total volume flow. This means that for the offers, the previous earnings \(b_{\textrm{earnings}}\) have to be subtracted from the costs of the total volume flow. The estimation of the individual pumping costs is crucial for achieving efficient system operation. In this work, it is realized by rule-based estimations of the presumed power demand for the delivery of a certain volume flow. Therefore, the pressure increase and necessary speed are estimated. For the system resistance, a quadratic relationship is assumed, as the pressure loss depends quadratically on the volume flow. The pump agent \(i \in \mathcal{P}\) estimates the system characteristics for the current operation point based on the measured volume flow \(Q_{i}\) and pressure increase \(\Delta p_{i}\): \[\label{eq:cost_pumps_1} l_{\textrm{est}, i}= \Delta p_{i} / Q_{i}^2;\] From this, the necessary pressure increase \(\Delta p_{\textrm{est},i}\) is calculated for the offered volume flow \(Q_{\textrm{offer},i}\): \[\Delta p_{\textrm{est},i} = l_{\textrm{est}, i} Q_{\textrm{offer},i}^2;\] This is equal to the pressure that the pump has to provide according to the characteristic curve, cf. Eq. [\[eq:pump_char\]](#eq:pump_char){reference-type="eqref" reference="eq:pump_char"}: \[l_{\textrm{est}, i} Q_{\textrm{offer}, i}^2 = \alpha_{\textrm{1}, i} Q_{\textrm{offer}, i}^2+ \alpha_{\textrm{2}, i} Q_{\textrm{offer}, i} n_{\textrm{est}, i} + \alpha_{\textrm{3}, i} n_{\textrm{est}, i}^2 ;\] which can be solved for the estimated speed \(n_{\textrm{est}, i}\). Then, using the estimated speed and the power curve, cf. Eq. [\[eq:pump_char_power\]](#eq:pump_char_power){reference-type="eqref" reference="eq:pump_char_power"}, the required power \(P_{\textrm{est}, i}\) is estimated: \[\label{eq:cost_pumps_last} P_{\textrm{est}, i} = \beta_{\textrm{1}, i} Q_{\textrm{offer}, i}^3+ \beta_{\textrm{2}, i} Q_{\textrm{offer}, i}^2 n_{\textrm{est}, i} + \beta_{\textrm{3}, i} Q_{\textrm{offer}, i} n_{\textrm{est}, i}^2 + \beta_{\textrm{4}, i} n_{\textrm{est}, i}^3 +\beta_{\textrm{5}};\] By multiplying with the constant \(d_\textrm{power}\), this is converted into dimensionless individual costs: \(b_{\textrm{ind}, i} = d_\textrm{power} \cdot P_{\textrm{est}, i}\). In order to guarantee the purchased or sold volume flows after the negotiation, a controller is required. For this purpose, a conventional PI controller is used. For the pumps, a special feature is introduced, which is called artificial scarcity: A slightly lower volume flow than sold is set. This ensures that the downstream components behave \"efficiently\". In the case of the valves, they will not generate any unnecessary pressure loss, since they want the highest possible volume flow. This prevents that e.g. all valves have only a very low opening and the pumps have to work against it. In this approach, communication is only necessary during negotiation rounds and only between direct neighbors. A system-wide model is not required, but an estimation of the system characteristics is. A wrong estimation can lead to a bad calculation of the offers and thus low efficiency, but the functionality is not affected. ### Conventional Methods There is typically no communication when conventional methods are used to control the pumps and valves, i.e. there is local control according to Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"}. In most cases, the valves are responsible for controlling the actual process function, e.g. volume flow or temperature. The pumps must ensure that sufficient pressure is supplied at the valves so that they can operate safely. To do this, the pumps anticipate the operation of the valve by controlling the pressure at the pump outlet as a function of the volume flow. The pressure must be high enough to compensate for all pressure losses in the system all the way to the valves, for which a so-called pressure control curve is used. This always results in overfulfilment, which leads to energetic losses. In the context of this work, the valves locally control a flow rate using a PID controller independently of each other. For pumps, the following control approaches are considered for comparison purposes: - non-controlled: all pumps run continuously at their maximum speed - constant pressure control: the pressure difference of the pumps is constant \(\Delta p = \Delta p^\text{set}_\text{const}\) - Proportional pressure control: the pressure difference of the pump is proportional to the volume flow, in addition to a constant part \(\Delta p = l_\text{prop} Q + \Delta p^\text{set}_\text{const}\) The respective values of the parameters are estimated using a model of the fluid system. For this purpose, a design flow rate \(Q^\text{design}\) must be assumed, which can be covered maximally by the system. If a higher value \(Q^\text{design}\) is assumed than the expected maximum volume flow \(Q^\text{max}\), this corresponds to the integration of a safety factor \(S\): \(Q^\text{design} = S \cdot Q^\text{max}\). ### Assessment Criteria This section will serve the purpose of presenting assessment criteria for the control methods based on multi-agent systems. It will further discuss various characteristics of the approaches, helping to categorize them. The assessment criteria are derived from the challenges presented in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}: - fulfilling the functionality of the system, i.e., the control objective of providing a given volume flow to valves, - effort involved for fulfilling the functionality, i.e., consumption of electrical energy during operation as well as modelling effort of generating a substitute model (DMPC), training of agents (MADRL), or designing a mechanism (market mechanism) and information exchange between agents, - availability, i.e., reliability of fulfilling the functionality in case of disruption, and - acceptability, i.e., transparency of control decisions as well as portability, flexibility, and scalability of the approaches. The cost functions for valves and pumps yield numerical values for assessing fulfillment of functionality and effort with regard to energy consumption (cf. Section [3](#sec:res){reference-type="ref" reference="sec:res"}). Modeling effort, by contrast, is harder to gauge. DMPC requires a substitute model, which in the presented work is acquired using machine learning methods. The implementation of MADRL in this study uses model-free rather than model-based approaches, where the system model is learned directly by the agents. Nevertheless, a new substitute model needs to be learned for every new application or change in the system configuration. Market mechanism is inherently model-free, though the pumps do require a model for calculating the cost of providing different volume flows and the market rules need to be designed and tested prior to implementation. Regarding information exchange, observability is a category with which multi-agent systems can be classified. According to Vlassis, full observability is given when each agent perceives the entirety of states that constitute the state of the agents' environment with each observation . Partial observability is given when the agents only receive partial information on the current state of their environment with each observation . This study considered both fully observable, termed centralized, and partially observable MADRL. As stated above, in DMPC there is information exchange between agents on their proposed values for the manipulated variable, with several iterations being possible. In market mechanism, the information exchange is limited to offers and purchases of volume flow and restricted to neighboring agents. Thus, effort of information exchange strongly depends on the control approach. What effect the effort of information exchange has on the fulfillment of the function is the subject of research question (ii). For assessing the availability, the same cost functions are used as for assessing fulfillment of functionality. The values of these cost functions for a disrupted system are compared to those of an undisrupted system. How the disruption is modelled in this study is presented in the subsequent section. For gauging acceptability, the approaches can be arranged on the scales of black box vs. white box and flexibility. In market mechanism, all control decisions and market rules are transparent and easily understandable, i.e. this represents a white box approach. The machine learning methods of MADRL, by contrast, don't allow for control decisions to be comprehended or traced and neither is the learned substitute model discernible, which makes this a black box approach. DMPC is comprehensible in the regard of the employed optimization methods, though the substitute model created using machine learning is not transparent, meaning that this approach is located between white box and black box. Market mechanism is also the most flexible design approach, as it allows agents to be added or removed without changing the system model as long as the communication between upstream and downstream neighbors is ensured. Both DMPC and MADRL depend on the learned substitute model which requires new training rounds once the system is adapted, rendering the approaches less flexible. To summarize, the requirement of a substitute model plays an important role when assessing both effort and acceptability associated with control approaches. Observability has further implications for effort. Transparency of control decisions and flexibility of control approaches with regard to adapting systems are important criteria for assessing acceptability. Cost functions of agents enable the assessment of functional performance and availability. For the latter, modeling disruptions plays an important role which is presented in the following. ## Modelling Disruptions {#sec:disrmod} A prerequisite for transitioning from local control of fluid system components to central or distributed control architectures is the implementation of information and communication infrastructure. This allows information to be passed from components to a central control unit or between individual component agents in the case of distributed control. It has been remarked that increased dependence on communication infrastructure of fluid systems and water distribution systems in particular introduces new vulnerabilities . In consequence, the approaches to distributed control using DMPC and MADRL are investigated with regard to their performance faced with a disruption of the communication between agents. This allows to address the hypothesis regarding reliability of distributed versus centralized control and contributes to answering research question (iii), cf. [1.1](#sec:rq){reference-type="ref" reference="sec:rq"}. Figure [\[fig:undisrupted\]](#fig:undisrupted){reference-type="ref" reference="fig:undisrupted"} shows the communication infrastructure of both the centralized and the distributed control architectures in the undisrupted case. For centralized control as in Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"} (b), the state variables of the components are communicated to the central control unit which in turn communicates the calculated values for the manipulated variables to the components where they are implemented. For distributed control as in Figure [\[fig:concepts\]](#fig:concepts){reference-type="ref" reference="fig:concepts"} (c), the agents calculate and implement their control variables on a local level. They further communicate their state variables to one another. This requires interconnection of communication between all agents. Figure [\[fig:disrupted\]](#fig:disrupted){reference-type="ref" reference="fig:disrupted"} illustrates the implementation of a disruption in the communication infrastructure for both the centralized and distributed control architectures. The centralized control unit loses its connection with the upper pump and two valves and in consequence ceases to consider them when gathering state variables or calculating and implementing control variables. For the distributed control, on the other hand, the result of the disruption is a separation into two subsystems. The agents of each subsystem are able to exchange information on state and control variables amongst themselves yet not across the system boundary. The lower subsystem consists of one pump and three valves. It is situated upstream from the second subsystem, consisting of one pump and two valves. However, the lower subsystem is situated at a lower geodetic height than the second subsystem and is thus in a structurally advantaged position. The disruption scenario is chosen as a result of domain specific considerations. The valves represent consumers with volume flow demands. To satisfy these, a fluid system requires a pressure source, represented by the pumps. Thus, the link chosen to be disrupted connects two units, each of which is a functional fluid system. Accordingly, the disruption scenario follows the concept of a system consisting of viable units comparable to a honeycomb structure where each cell is a stable structure in itself regardless of the surrounding system as proposed for resilient systems by Heinimann . With regard to the hypothesis and research question under investigation, the chosen disruption scenario is the most compelling for providing insights. ## Implementation All experiments are performed and analyzed using Python. For the MADRL, the freely available code from MAAC was used. To use the hydraulic simulation, the interface OpenAI Gym, which is widely used in the reinforcement learning field, was used. The hydraulic simulation model based on a real-world test rig was implemented in Dymola, a simulation and modeling software based on the Modelica modeling language. The final model was exported as a functional mockup unit (FMU) and used with Python via functional mockup interface (FMI) using the Python library *FMPy*. Both the Dymola model of the test rig used for this study and the FMU file can be found in the data repository <https://tudatalib.ulb.tu-darmstadt.de/bitstream/handle/tudatalib/3625/Fluid_Model.zip>. The differential equations mapped using Dymola are solved using the Dymola solver dassl, a solver with step size control, where the step size was set to 0.025 s. The experiments were run on a computer with an IntelCore™ i7-8700 CPU with 3.2 GHz clock speed and 16 GB RAM. For all simulations, the weighting parameters of the cost of the valves \(i \in \mathcal{V}\) are set to \(\lambda_i = 1\) and those of the pumps \(i \in \mathcal{P}\) are set to \(\lambda_i = 2\). For the model of (D)MPC, a dataset with 150000 data points was sampled from the simulation with randomized inputs. Together with the measured outputs, a deep neural network can be trained to act as the system model during optimization. The R²-value of the system model is 0.9873, meaning that the model is quite close to the Dymola model. For optimization, the library `scipy` with the optimizer *SLSQP* is used. To avoid a solution that is the local and not the global minimum, in each step the optimization is repeated with random starting values. The hyperparameters for the MADRL algorithm were manually tuned after several experiments. As will be mentioned later, systematic optimization of hyperparameters offers the greatest potential for further improvement. As in the original implementation of MAAC, the neural networks for both actor and critic consist of two layers with 256 neurons each. Each training episode has a duration of 4 s and consists of randomized volume flow demand held for 0.5 s. This allows a wide range of combinations to be covered during training. In the training, the simulation frequently crashed due to numerical errors caused by the initially very arbitrary actions of the agents. To prevent this, an additional sparse reward of −25 was added to the global reward for each simulation crash. Training was continued until no improvement of the reward was seen. This was the case after approximately 4,000 episodes, which corresponds to 4.45 h of real-life training. Just like the other methods, the algorithms required for the market mechanism approach are implemented entirely in Python. # Results {#sec:res} In the following, the results of the different approaches applied to the example system, shown in Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"}, are presented. First, a comparison of the three approaches is made with regard to functionality and energy efficiency for the 30 min load profile. In addition, the behavior of agents and transparency of control decisions are analyzed using the 30 s load profile. Next, the influence of the models on which the approaches are based and the amount of information exchanged between the agents is examined. Finally, the behavior in case of disruptions is considered. ## Comparison of Approaches to Distributed Control {#sec:comp} All novel control methods presented were operated with the same 30-minute scenario, cf. Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (c) (i). The total resulting costs--normalized to the minimum costs of all approaches--are shown in Figure [\[fig:total_costs\]](#fig:total_costs){reference-type="ref" reference="fig:total_costs"}. The central MPC has the lowest costs, which are used as a normalization factor. The highest costs of 140 % are observed for the central MADRL-the costs thus vary by up to 40 %. As expected, MPC is better than DMPC, although the difference is small. Partially observable MADRL, on the other hand, is unexpectedly better than central MADRL, even though the latter actually has more information available. An interpretation of this communication and information influence will follow later. In the control objective, both the deviation from the desired volume flow and the energy consumption are taken into account in a joint function. The individual components and the trade-off between them are shown for the different approaches in Figure [\[fig:tradeoff\]](#fig:tradeoff){reference-type="ref" reference="fig:tradeoff"}. There is a clear trade-off between volume flow deviation and energy consumption among the different approaches. Each of the novel approaches, except one, are best with respect to some weighting of the objectives, i.e. the approaches show a Pareto front. Central MADRL has the lowest energy consumption, but the highest volume flow deviation. The market mechanism is exactly the opposite. Partially observable MADRL is dominated, i.e. it is worse than the MPC approaches with respect to both criteria. The conventional strategy shows that the constant pressure control has a higher energy consumption compared to the proportional pressure control, but a lower volume flow deviation. This can also be observed by increasing the safety factor. Without a safety factor, the approaches are dominated by MPC and market mechanism. For higher safety factors, the conventional strategies show the lowest achievable volume flow deviation--but a significantly higher energy consumption. The uncontrolled case represents an upper limit for the lowest possible volume flow deviations, which, however, leads to very high energy consumption. In all approaches, the preference in the trade-off can be controlled. In MPC and MADRL this is done by the factors in the objective function, in the market mechanism by the allocation of costs and budgets. However, despite the same weighting factors in MPC and MADRL, they show a different result, which is also evident in the total costs analyzed earlier. The demand profile shown in Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (c) (ii) is a simplified scenario, i.e. valves 1, 3 and 5 have an initiating demand one after the other. The results shown in Figure [\[fig:30sec\]](#fig:30sec){reference-type="ref" reference="fig:30sec"} allow a closer look at the behavior of the individual agents and to check the comprehensibility. Here, it needs to be stated that while the DMPC approach used the loadprofile presented in Figure [\[fig:use_case\]](#fig:use_case){reference-type="ref" reference="fig:use_case"} (c) (ii), the other two approaches required a longer initial phase without demand. Accordingly, a period of 4 s without demand was added before the actual beginning of the load profile as indicated by negative values on the time axis in Figure [\[fig:30sec\]](#fig:30sec){reference-type="ref" reference="fig:30sec"}. In Figure [\[fig:30sec\]](#fig:30sec){reference-type="ref" reference="fig:30sec"} (a), the upper plot shows that the flow demand of the valves is quickly compensated and thus the cost of the valves decrease. However, the costs of the pumps increase due to the higher energy consumption. In the middle plot, it can be seen that the valves--as expected--open as soon as they have a demand. After valve 3 has a demand (starting from second 11) valve 1 closes to avoid overfilling. At some points, obviously energetically inefficient states can be identified. For example, pump 2 (lower plot) never turns off, although it is not needed until second 21. Furthermore, valve 5 is not fully open in the last third and thus the pump and valves work against each other. A possible explanation for the pump not turning off is the local controller of this pump. At this point there is a discontinuity (switch off from 40 % speed to 0 % speed), which is difficult to be taken into account by the local PID controller. The incomplete valve opening may be due to a poor model or only locally optimal solution. When using the MADRL approach (b), the behavior appears to be similar in general. The demands are quickly satisfied, but non-efficient states occur. This is due to a suboptimal policy of the individual agents. It is noticeable that in the last third, the performance of valve 3 is poor (high costs), which improves the performance of valves 1 and 5. Thus, valve 3 supports the other valves. In the market mechanism (c), the demands are also quickly satisfied. Initially, pump 2 is also unnecessarily switched on, however, due to the negotiations, the actions become more comprehensible. The data of the negotiations show, that pump 1 sells no volume flow to pump 2 at this time. The fact that the pump is still not switched off can therefore only be due to the local controller. ## Influence of Available Information This section addresses research question (ii) concerning the available information from substitute models and communication between agents. First, the effect of modelling effort is considered for the overall costs as well as functional quality and effort regarding energy consumption, cf. Figures [\[fig:total_costs\]](#fig:total_costs){reference-type="ref" reference="fig:total_costs"} and [\[fig:tradeoff\]](#fig:tradeoff){reference-type="ref" reference="fig:tradeoff"}. The model-based approaches MPC and DMPC have significantly lower costs. The approaches that use less model knowledge (MADRL and Market Mechanism) have higher costs. Thus, the tradeoff between reaching the objective and system knowledge--which is associated with high modelling effort--becomes clear. MADRL uses data-driven models as a substitute for analytical model knowledge and achieves better energy efficiency, though at the cost of worse functional quality. Market mechanism, on the other hand, uses no model for the system-wide behavior, yet achieves the best values among the distributed approaches for functional quality albeit with higher energy consumption. As presented in Section [2.3.1](#sec:dmpc-meth){reference-type="ref" reference="sec:dmpc-meth"}, in DMPC the agents may exchange information regarding the respective values of their manipulated variables and recalculate their optimal solution based on the updated information. The influence of the number of such rounds of communication was investigated for DMPC, and the results can be found in Figure [\[fig:com\]](#fig:com){reference-type="ref" reference="fig:com"}. The overall costs for MPC were taken as a benchmark, against which the cost of DMPC with rounds of communication varied between 0 and 9. As expected, the costs for DMPC are consistently higher than for MPC. As the number of communication rounds increases, the costs decrease until a minimum is reached at 4 while a sudden drop in costs occurs between 2 and 3. With the number of communication rounds increased beyond 4 the costs again increase which is contrary to expectations. The costs for MPC should represent an asymptotic limit towards which the costs for DMPC would be expected to converge with increasing number of communication rounds. The presented results suggest that rounds of communication beyond an optimal amount cause confusion rather than improved solutions. As seen in Figure [\[fig:total_costs\]](#fig:total_costs){reference-type="ref" reference="fig:total_costs"}, partially observable MADRL achieves lower total costs than central MADRL, despite being a dominated solution (cf. Figure [\[fig:tradeoff\]](#fig:tradeoff){reference-type="ref" reference="fig:tradeoff"}). Intuitively, central MADRL is expected to perform better, as universal availability of target values and control variables for each agent are closer to a centralized control. A possible interpretation of these results is that the amount of information available to agents in the case of central MADRL causes information overload rather than helping coordination among agents. Partially observable MADRL, on the other hand, achieves a better result with each agent only using its own variable values and all current demands in the system. Thus, this approach seems to represent an adequate compromise, with necessary information shared between agents while avoiding distracting agents with surplus information. Similar to the rounds of communication in the case of DMPC, rounds of bidding were investigated for market mechanism control. This proved to have no influence on the overall costs of market mechanism control for the \(30\ \mathrm{min}\) scenario. To summarize, the presented results suggest that excess amounts of information can deteriorate the results of distributed control approaches. Nevertheless, up to a certain point, increased sharing of information between agents improves the performance of distributed control. ## Performance under Disruption Finally, results for answering research question (iii) are presented in Figure [\[fig:resdisr\]](#fig:resdisr){reference-type="ref" reference="fig:resdisr"}. The disruption as described in Section [2.4](#sec:disrmod){reference-type="ref" reference="sec:disrmod"} is triggered at different times during the 30 min scenario. The results for total costs for MPC, DMPC, central and partially observable MADRL as well as market mechanism in the disrupted case are related to the cost of the undisrupted case with MPC. As expected, the cost in the disrupted case exceeds the cost in the undisrupted case for all control approaches. The total cost for DMPC is lower than for MPC for every time of failure occurrence, i.e. DMPC allows for improved control when faced with the given disruption. With later times of failure occurrence, the costs of both approaches converge at the time of low to no demand in the system. The specific time of failure occurrence affects costs for both MPC and DMPC. The overall trend shows lower increases of cost for later times of failure occurrence for both approaches. This behavior is to be expected due to shorter periods of disruption as well as lower overall demand with increasing time. However, when failure occurs at 3 min and 12.5 min higher costs are recorded for both approaches than at 2 min and 7.5 min respectively. This is due to unfavorable demand distributions at the given times of failure occurrence, which are then propagated for the remainder of the 30 min scenario. As expected from the results shown in Figure [\[fig:total_costs\]](#fig:total_costs){reference-type="ref" reference="fig:total_costs"}, the cost for central MADRL exceeds that of partially observable MADRL. The costs for both approaches are increased compared to the undisrupted case, cf. Figure [\[fig:total_costs\]](#fig:total_costs){reference-type="ref" reference="fig:total_costs"}, though the effect on partially observable MADRL is small. The costs in the disrupted case of partially observable MADRL are of the same magnitude as those of MPC and DMPC under disruption. A trend of lower increase in costs due to disruption with later times of failure occurrence is observable especially for central MADRL while the effect is much less pronounced for partially observable MADRL. The trend is not broken by the effect described for MPC and DMPC. This results in costs for partially observable MADRL being as low as those of DMPC or lower than those of MPC for certain times of failure occurrence. The performance of market mechanism is also severely affected in the case of disruption and the costs exceed those of MPC, DMPC and partially observable MADRL. With later times of failure occurrence, the increase in costs is reduced, similar to central MADRL. To summarize, answering research question (iii), it is concluded that disruptions affect centralized control represented by MPC to a greater extent than distributed control as represented by DMPC and partially observable MADRL while the performance of central MADRL and market mechanism deteriorates significantly beyond the range of the other approaches. Among the distributed approaches, DMPC and partially observable MADRL are best suited to cope with disruptions to the communication infrastructure, while market mechanism appears to be more vulnerable. # Discussion {#sec:disc} Moving beyond central, optimal control of fluid systems, the presented work investigated three approaches to distributed control, DMPC, MADRL, and market mechanism, stemming from the domains of control theory, machine learning, and game theory respectively. These approaches were implemented and tested using a generic fluid system model for water supply and a randomly generated demand pattern based on a typical demand profile for a building. Corresponding to the challenges presented in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}, the performance of the approaches is evaluated regarding functionality, effort in terms of energy consumption and modelling, availability and acceptability. Converting deviation from control objective as well as consumed energy to standardized costs allows assessing functional quality, effort in terms of energy, and availability quantitatively, whereas the criteria modelling effort and acceptability are assessed qualitatively based on the experience of implementing the approaches. The quantitative results have been presented in detail in the preceding section. MPC and DMPC find the best trade-off between control objective and energy consumption, as both values are part of the objective function in the implemented optimization problem. In accordance with the first hypothesis, the distributed approach of market mechanism requires higher amounts of energy whereas central MADRL, by contrast, is more efficient in terms of energy consumption, yet this comes at the cost of significantly worse performance regarding functional quality. Implementation efforts for MADRL and MPC/DMPC are extensive and since the substitute model for MPC/DMPC is generated using an artificial neural network, both approaches are placed towards the black box side of the white box vs. black box spectrum, implying less transparency. Modification of the fluid system entails further effort for both approaches since the substitute model needs to be relearned from scratch, impeding flexibility and scalability of the approaches. By contrast, market mechanism functions based on simple comprehensible rules and algorithms, rendering it a transparent approach. System modifications are easily integrated, as the overall communication structure is maintained and the governing rules apply to all market participants. This shows advantages with regard to flexibility and scalability of the approach. Regarding the hypothesis concerned with availability and reliability in the face of failures in the communication, the results support the conclusion that distributed approaches are more reliable than centralized approaches with the exception of market mechanism. The high information exchange requirements of central MADRL are equally detrimental to performance under disruption as the necessity to gather all information centrally in the case of MPC. However, before arriving at a conclusive statement regarding the hypothesis, methods from information and network theory need to be applied to further investigate the hypothesis. Existing studies have shown the feasibility of agent-based control of transport systems in general and fluid and related systems such as HVAC and water distribution systems in particular. Most studies use rule-based mechanisms for designing agents and interaction, while other studies use approaches from the domain of machine learning. Few compare results to benchmarks set by control using optimization methods or conventional control strategies. The presented study considered distributed control approaches from three different domains, all based on multi-agent systems, and applied them to a generic fluid system. A systematic assessment of the approaches allowed to compare them to one another as well as to benchmarks set by conventional control strategies and centralized control using global optimization methods. Advantages and drawbacks of the three approaches have been identified, yet all three approaches were shown to be feasible in simulation. Transferring approaches from the world of numerics to physical systems poses great challenges with regard to uncertainty, delay tolerance and system inertia. In this study, only the reliability in the face of failures in the communication network was considered. Failures may also occur in the technical system which constitutes the environment of the agent system. The reliability of agent-based control against such disruptions should be investigated in future work. Turning from physical applications to refinement of the methods considered in the presented work, there is great potential for further work. DMPC may be improved by studying various solvers and optimizing parameters such as prediction horizon and control variable boundaries to ensure global optimum operation points at all times. Looking at MADRL, hyperparameter optimization may be a first step towards improving the performance. In order to address the aspect of flexibility and scalability, transfer learning could be considered for training basic agents which are then inserted into existing systems after which the entire system would adapt, learning an updated system model which includes the new components. System safety is a further concern, with methods required to be included in the training process that prohibit unsafe operating conditions. The market mechanism investigated here was devised using heuristic, ad-hoc rules which were only modelled for the components as far as necessitated by the use-case. Aside from generalising the algorithms for further system configurations, methods for optimizing mechanism design could be closely studied to ensure optimal and fair allocation of goods. Furthermore, game theoretic approaches beyond markets may provide further insights for devising appropriate mechanisms. Applications of agent-controlled fluid systems may be in bounded environments, such as factory cooling circuits or domestic heating systems. However, water distribution systems, waste-water disposal systems, gas and district heating are large-scale infrastructures that may also be considered for agent-based control. These infrastructures provide vital services to communities, which raises two further aspects. Firstly, the system boundary of the environment of the agent-system would need to be expanded to include the socio-technical system of consumers using the technical system controlled by agents. This will render demand patterns at valves more complex. Secondly, with the vital importance of these infrastructure systems for communities served by them, questions of resilience need to considered in the design approaches in order to provide minimal functionality and the possibility of recovery when critical adverse events occur. # Conclusions The presented work studied three distributed agent-based control approaches stemming from the domains of game theory, control theory and machine learning respectively in application to a generic fluid system with the medium water. The performance of the approaches was studied in simulation and assessed quantitatively and qualitatively, and compared to benchmarks of conventional control strategies and centrally computed optimal control. Future research directions include application of the distributed agent-based control approaches to a physical fluid system and experimental validation, as well as refinement of the design and methods underlying each of the approaches. # Author Contributions {#author-contributions .unnumbered} Contributions following the CRediT taxonomy: **Kevin T. Logan:** Conceptualization, Methodology, Validation, Data Curation, Writing--original draft preparation, Writing--review and editing, Visualization; **J. Marius Stürmer:** Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing--original draft preparation, Writing--review and editing, Visualization; **Tim M. Müller:** Conceptualization, Methodology, Validation, Writing--original draft preparation, Writing--review and editing, Visualization, Project Administration, Funding acquisition; **Peter F. Pelz:** Resources, Writing--review and editing, Supervision, Project Administration, Funding acquisition.

{'timestamp': '2022-12-19T02:11:16', 'yymm': '2212', 'arxiv_id': '2212.08450', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08450'}

# Introduction {#Sec1} The purpose of the present article is to discuss the ideas which led to the introduction of the proxy-SU(3) symmetry model, its microscopic justification, and its applications to nuclear structure problems developed so far. Starting with a review of nuclear structure models based on the SU(3) symmetry in Section II, we discuss nucleon pairs favoring deformation in Section III and show how these lead to the introduction of the proxy-SU(3) symmetry in Section IV. The approximation leading to the proxy-SU(3) symmetry is tested using Nilsson model calculations in Section V and is "translated" into spherical shell model language in Section VI. After pointing out in Section VII the crucial role played by the highest weight irreducible representations, their consequences on providing parameter-free predictions for the collective variables \(\beta\) and \(\gamma\) of even-even nuclei, on explaining the dominance of prolate over oblate shapes in the ground states of even-even nuclei and on the prediction of a prolate to oblate shape/phase transition are examined in Section VIII, while in Section IX a mechanism predicting on the nuclear chart specific islands on which shape coexistence can appear is discussed. Possible further work on parameter-free predictions of B(E2) transition rates, energy spectra, binding energies and nucleon separation energies is outlined in Section X. Literature has been searched up to October 2022. # SU(3) symmetry in nuclear structure {#Sec2} Symmetries have played a major role in nuclear structure since 1937, when Wigner suggested that to a first approximation nuclear forces should be independent of the orientation of the spins and isospins of the nucleons constituting a nucleus, in the framework of what became known as the SU(4) supermultiplet model. In 1949 a major step forward in deciphering the experimental observations has been made by Mayer and Jensen through the introduction of the shell model, who interpreted the appearance of nuclear magic numbers in terms of a three-dimensional (3D) isotropic harmonic oscillator (HO) potential, to which a crucial spin-orbit term has been added. The shells of the 3D isotropic HO are labeled by the number of excitation quanta \(n\), and are characterized by the unitary symmetries U((n+1)(n+2)/2), having SU(3) subalgebras. Mayer and Jensen shared with Wigner the Nobel Prize in Physics in 1963. The shell model suffices for describing near-spherical nuclei in the vicinity of the magic numbers, away of which, however, large quadrupole moments are observed. In order to explain their appearance, Rainwater in 1950 suggested that deformed shapes are energetically favored away from closed shells. Soon thereafter in 1952 the collective model of Bohr and Mottelson has been introduced, in which the collective variables \(\beta\) and \(\gamma\) describe the departure from the spherical shape and from axial symmetry, respectively. Bohr, Mottelson, and Rainwater shared the Nobel Prize in Physics in 1975. In 1955 Nilsson introduced a modified version of the shell model, using a 3D anisotropic HO with cylindrical symmetry, in which axial deformed nuclei with prolate (rugby ball like) or oblate (pancake like) shapes can be described. In 1958 Elliott proved that within the nuclear \(sd\) shell, possessing the U(6) symmetry, deformation can be described in terms of the SU(3) subalgebra, thus building a bridge between the spherical shell model and the collective model. However, this bridge is valid only in the case of light nuclei, in which the consequences of the spin-orbit interaction on the ordering of the single particle nucleon levels are small. Beyond the \(sd\) nuclear shell the spin-orbit force is known to push within each HO shell the orbitals possessing the highest angular momentum \(j\) to the shell below. The orbitals left back in each shell after this removal, called the normal parity orbitals, along with the orbitals invading from the shell above, having the opposite parity and called the intruder orbitals, form a new shell, in which the SU(3) symmetry of the 3D isotropic HO is broken. Efforts of extending the SU(3) symmetry to heavy nuclei, started in 1972, evolved gradually into the introduction of the Vector Boson Model. At the same it has been realized that the Bohr--Mottelson model has an overall U(5) symmetry possessing an O(5) subalgebra. A major step forward in the extention of the SU(3) symmetry to heavy nuclei has been taken in 1973, with the introduction of the pseudo-SU(3) symmetry. Using a unitary transformation, the incomplete set of normal parity orbitals left in a shell is mapped onto the complete set of orbitals of the shell below. In this way the SU(3) symmetry of the 3D isotropic HO is recovered for the normal parity orbitals. This is achieved by assigning to each orbital a pseudo-orbital angular momentum and a pseudo-spin, while keeping the total angular momentum intact. As a result, each shell is formed now by a normal parity part, which possesses a U(n) symmetry and a SU(3) subalgebra, and by an intruder part which does not possess any SU(3) structure and therefore has to be taken into account separately, using shell model techniques. Later it has been understood that the pseudospin symmetry has relativistic mean field roots. In 1974 it was realized that the nuclear quadrupole degree of freedom can be described in terms of a SU(6) algebra formed by five generalized coordinates and conjugated momenta, and their commutators. Next year Arima and Iachello introduced the Interacting Boson Model (IBM), in which, in addition to the quadrupole degree of freedom, described by \(d\)-bosons of angular momentum two, the monopole degree of freedom, described in terms of \(s\)-bosons of zero angular momentum, is also taken into account. IBM is characterized by an overall U(6) symmetry, possessing three limiting symmetries, U(5) for vibrational nuclei, which corresponds to the Bohr-Mottelson collective model, O(6) for \(\gamma\)-unstable nuclei (which are soft towards triaxial deformation), and SU(3) for deformed nuclei. In 1977 the symplectic model was introduced by Rowe and Rostensteel. The symplectic model, having an overall symmetry called Sp(3,R) by Rowe and Rosensteel but Sp(6,R) by other authors, is a generalization of the fermionic Elliott SU(3) model to include many major oscillator shells in addition to core excitations. Its overall symmetry is characterized by the noncompact algebra Sp(6,R), which does possess a compact SU(3) subalgebra along with other ones. A proton-neutron extension of the model, called the proton-neutron symplectic model (PNSM) is also developed. In 1982 the Interacting two-Vector Boson Model has been introduced, described in terms of two vector bosons of angular momentum one. These form the noncompact algebra Sp(12,R), of which the maximal compact subalgebra is U(6). In 1987 the Fermion Dynamical Symmetry Model has been introduced. Instead of using the usual splitting of the total angular momentum of the nucleons into orbital angular momentum and spin parts, in this model a splitting of the total angular momentum into active and inactive parts is assumed, the k-active part containing a SU(3) subalgebra. In 2000 an *ab initio* approach to a no-core shell model has been introduced for light nuclei. It has been soon thereafter realized that a symplectic symmetry is underlying the *ab initio* no core shell model results, thus paving the way towards the development of the symplectic no-core shell model, which has soon been extended to intermediate-mass nuclei. The realization that nuclei are made of only a few equilibrium shapes led to the introduction of the *ab initio* symmetry-adapted no-core shell model. A SU(3)-adapted basis plays a key role in this approach, taking advantage of the Elliott SU(3) symmetry underpining the Sp(3,R) (alias Sp(6,R) ) symplectic model. The use of the SU(3) symmetry in nuclear structure has been reviewed in 2020 by Kota. In 2017 the proxy-SU(3) symmetry has been suggested, which will be the subject of the present review. However, before describing the proxy-SU(3) symmetry itself, we briefly review the physical motivation behind its introduction. # Nucleon pairs favoring deformation {#Sec3} In 1953 deShalit and Goldhaber in their studies of \(\beta\) transition probabilities noticed that the development of nuclear deformation was favored within the proton--neutron pairs of orbitals (1p3/2, 1d5/2), (1d5/2, 1f7/2), (1f7/2, 1g9/2), (1g9/2, 1h11/2), (1h11/2, 1i13/2), since the nucleons of one kind (protons, for example) appeared to have a stabilizing effect on pairs of nucleons of the other kind (neutrons in the example). In the standard shell model notation \(\vert n l j m_j\rangle\), states are labeled by the number of oscillator quanta \(n\), the orbital angular momentum \(l\), the total angular momentum \(j\), and its \(z\)-projection \(m_j\). In this notation these orbitals form pairs differing by \(\vert \Delta n \Delta l \Delta j \Delta m_j\rangle = \vert 0 1 1 0\rangle\), which we are going to call the spherical shell model \(\vert 0 1 1 0\rangle\) pairs, or simply the \(\vert 0 1 1 0\rangle\) pairs. In 1962 Talmi introduced the concept of seniority , corresponding to the number of nucleon pairs coupled to non-zero angular momentum. Seniority explained the linear dependence of neutron separation energies on the mass number within various series of isotopes, being a major step forward in our understanding of effective interactions and coupling schemes in nuclei. In 1977 Federman and Pittel realized that certain proton--neutron pairs play a major role when adding valence protons and valence neutrons to a nucleus. In particular, the proton--neutron pairs (1d5/2, 1d3/2), (1g9/2, 1g7/2), (1h11/2, 1h9/2), and (1i13/2, 1i11/2) are important for the onset of deformation, while further on deformation is reinforced by the proton--neutron pairs (1d5/2, 1f7/2), (1g9/2, 1h11/2), (1h11/2, 1i13/2), and (1i13/2, 1j15/2). In the shell model notation these sets of pairs, which are shown in Table I, correspond to \(\vert \Delta n \Delta l \Delta j \Delta m_j\rangle = \vert 0 0 1 0\rangle\) and \(\vert 0 1 1 0\rangle\) respectively. The latter set coincides with the de Shalit--Goldhaber pairs mentioned above. In 1985 Casten demonstrated the decisive role played by proton-neutron pairs through the introduction of the \(N_p N_n\) scheme. This scheme showed the systematic dependence of several observables on the quadrupole-quadrupople interaction, which is measured through the product \(N_pN_n\), where \(N_p\) (\(N_n\)) is the number of valence protons (neutrons) counted from the nearest closed shell. In 1987 the systematic dependence of several observables on the \(P\)-factor, defined as \(P= N_p N_n / (N_p+N_n)\), has been demonstrated. The \(P\)-factor expresses the competition between the quadrupole deformation and the pairing interaction, "measured" by \(N_pN_n\) and \(N_p+N_n\), respectively. In 1995 the proton--neutron pairs (1g9/2, 2d5/2), (1h11/2, 2f7/2), (1i13/2, 2g9/2) led to the introduction of the quasi-SU(3) symmetry, which leads to enhanced quadrupole collectivity. In the shell model notation these pairs are expressed as \(\vert \Delta n \Delta l \Delta j \Delta m_j\rangle = \vert 1 2 2 0\rangle\). Detailed studies of double differences of experimental binding energies, led in 2010 to understanding that proton-neutron pairs differing by \(\Delta K[ \Delta N \Delta n_z \Delta \Lambda]=0[110]\) in the Nilsson notation \(K [N n_z \Lambda]\), where \(N\) is the total number of oscillator quanta, \(n_z\) is the number of quanta along the \(z\)-axis, and \(\Lambda\), \(K\) are the projections along the \(z\)-axis of the orbital angular momentum and the total angular momentum respectively, play a major role in the development of nuclear deformation. This effect has been attributed to their large spatial overlaps and has been corroborated by nuclear density functional theory calculations. We are going to call these pairs the Nilsson 0\[110\] pairs, or simply the 0\[110\] pairs. Notice the difference in the notation in comparison to the spherical shell model \(\vert 0 1 1 0\rangle\) pairs. As we shall see below, the Nilsson 0\[110\] pairs play a crucial role in the replacements made within the approximation leading to the proxy-SU(3) symmetry. Furthermore, the relation between the Nilsson 0\[110\] pairs and the spherical shell model \(\vert 0 1 1 0\rangle\) pairs will be clarified through the connection of the proxy-SU(3) symmetry to the spherical shell model. Evidence supporting the formation of 0\[110\] pairs has been found recently within Monte Carlo shell model calculations. # The proxy-SU(3) approximation {#Sec4} The proxy-SU(3) symmetry has been born when the desire to reestablish the SU(3) symmetry of the 3D-HO in medium mass and heavy nuclei, described in section [2](#Sec2){reference-type="ref" reference="Sec2"}, met the experimental hint of the 0\[110\] Nilsson pairs, described in section [3](#Sec3){reference-type="ref" reference="Sec3"}. The 0\[110\] Nilsson pairs have been discovered through the experimental observation that proton-neutron pairs of this type maximize the proton-neutron interaction. This maximization has been attributed to the large spatial overlaps of the two orbitals involved in such pairs. In other words, 0\[110\] Nilsson pairs are very similar, since they possess identical angular momentum and spin properties (identical projections of the orbital angular momentum, the spin, and the total angular momentum) and very similar spatial shapes, since they differ only by one excitation quantum in the \(z\)-axis (and therefore also by one quantum in the total number of quanta). The similarity holds equally well if one considers a 0\[110\] pair of protons, or a 0\[110\] pair of neutrons. Such pairs will still have identical angular momentum and spin properties and very similar spatial shapes. Therefore, in the framework of an approximation needed for some reason, each of the members of a 0\[110\] pair could replace the other, i.e., act as its proxy, with minimal changes inflicted in the physical system under study. Such a situation appears in the shells of the shell model beyond 28 nucleons. The intruder orbitals, which have come down from the shell above, form 0\[110\] pairs with the orbitals which have deserted the shell by fleeing into the shell below. We can therefore think of replacing the intruder orbitals by their 0\[110\] counterparts, which are the deserting orbitals, expecting that the changes caused in the physical system under study would be minimal. In other words, the deserting orbitals can act as proxies of the intruder orbitals. But in this way the shell gets back the SU(3) symmetry of the 3D-HO, which it had lost after the deserting of the orbitals which fled to the shell below. An example can be seen in Fig. 1, in which the \(sdg\) shell is depicted, consisting of the \(3s_{1/2}\), \(2d_{3/2}\), \(2d_{5/2}\), \(1g_{7/2}\), \(1g_{9/2}\) orbitals, having a U(15) overall symmetry, which possesses a SU(3) subalgebra. The spin-orbit interaction pushes the orbital with the highest \(j\), i.e., \(1g_{9/2}\), to the shell below, while it brings down the \(1h_{11/2}\) orbital from the shell above. In this way the 50-82 shell of the shell model is formed, in which the SU(3) symmetry has been lost. Looking however into the details of the orbitals, a sign of hope appears. The intruder \(1h_{11/2}\) orbital consists of the Nilsson orbitals 1/2\[550\], 3/2\[541\], 5/2\[532\], 7/2\[523\], 9/2\[514\], 11/2\[505\], while the deserting \(1g_{9/2}\) orbital consists of the Nilsson orbitals 1/2\[440\], 3/2\[431\], 5/2\[422\], 7/2\[413\], 9/2\[404\]. We observe that the first five Nilsson orbitals of \(1h_{11/2}\) are 0\[110\] partners with the five orbitals making up \(1g_{9/2}\). Therefore it is plausible to replace the 1/2\[550\], 3/2\[541\], 5/2\[532\], 7/2\[523\], 9/2\[514\] orbitals by the 1/2\[440\], 3/2\[431\], 5/2\[422\], 7/2\[413\], 9/2\[404\] orbitals, hoping that the changes inflicted in the physical system would be minimal. In other words, the 1/2\[440\], 3/2\[431\], 5/2\[422\], 7/2\[413\], 9/2\[404\] orbitals will act as proxies of the 1/2\[550\], 3/2\[541\], 5/2\[532\], 7/2\[523\], 9/2\[514\] orbitals. It is now time to consider the gains and losses of this replacement. The gain is that the SU(3) symmetry has been reestablished in the 50-82 shell, since the shell will now contain all the orbitals composing the \(sdg\) shell. However, a couple of losses are lurking. First, the 11/2\[505\] Nilsson orbital had no 0\[110\] partner, thus it is left behind within the 50-82 shell, where it stays outside the SU(3) symmetry and in principle it has to be treated separately by shell model techniques. The good news in this case is that the 11/2\[505\] Nilsson orbital lies at the very top of the 50-82 shell, as one can see in the standard Nilsson diagrams. Thus it should be empty for most nuclei and therefore have little influence on the structure of most of them. The second problem arises from the fact that the intruder orbitals and the deserting orbitals have opposite parities, since they belong to adjacent shells differing by one unit in \(N\). For even-even nuclei this should not cause any trouble, since, for example, a pair of 9/2\[514\] particles will be replaced by a pair of 9/2\[404\] particles, therefore no differences caused by parity should be seen. For odd nuclei one would probably have to use projection techniques, a problem which has not been considered yet. # Corroboration of proxy-SU(3) through Nilsson model calculations {#Sec5} The first test of the accuracy of the proxy-SU(3) approximation described in the previous section has been performed in the framework of the Nilsson model. In each shell of the shell model, two calculations have been performed, one with the real orbitals composing the shell, and another one with the intruder orbitals replaced by their 0\[110\] counterparts. Numerical results for the 82-126 shell are shown in Table II. In the upper part of the table, the matrix elements of the Nilsson Hamiltonian, using the standard set of parameters, are shown, while in the lower part the matrix elements occurring within the \(pfh\) shell resulting after the proxy-SU(3) approximation are reported. We remark that in the last 7 columns of the upper part, corresponding to the \(1i_{13/2}\) orbitals, all non-diagonal matrix elements vanish, due to the fact that they connect orbitals of opposite parity. On the contrary, in the last 6 columns of the lower part of the table, in which the proxies of the \(1i_{13/2}\) orbitals, i.e., the \(1h_{11/2}\) orbitals appear, not all off-diagonal matrix elements vanish, since in this case they connect orbitals of the same parity. However, the number of these off-diagonal matrix elements is small, and in addition their sizes are small in comparison to the diagonal matrix elements, thus they are not expected to affect the single particle energy levels substantially. The diagonal matrix elements in the last six columns of the lower part of the table are also slightly modified in comparison to the corresponding matrix elements in the upper half of the table, but again the changes are small and are not expected to affect the single particle energy levels radically. The evolution of the single particle energy levels with deformation, depicted in Fig. 2, shows that the changes caused in the Nilsson diagrams by the proxy-SU(3) approximation are minimal. Neither the order of the orbitals, nor their dependence on the deformation (upsloping or downsloping) are modified. Similar tables and figures for other shells can be found in the supplemental material for Ref.. It is seen there that the quality of the proxy-SU(3) approximation becomes better in higher shells. # Proxy-SU(3) symmetry in the spherical shell model basis {#Sec6} We have seen that the proxy-SU(3) symmetry is established in the nuclear shells beyond the \(sd\) shell in the framework of the Nilsson model by taking advantage of the 0\[110\] Nilsson pairs. The question arises if such a process, allowing the replacement of certain orbitals by their proxies in order to reestablish the SU(3) symmetry, is possible within the framework of the spherical shell model. In order to examine if such a possibility exists, we start with the Elliott model, in which the cartesian basis of the 3D isotropic HO is used. This basis is labeled as \([n_z n_x n_y m_s]\), in which the number of quanta along the \(z\), \(x\), \(y\) directions and the \(z\)-projection of the spin appear. The first step is to transform this basis into the spherical basis \([n l m_l m_s]\) in \(l\)-\(s\) coupling, labeled by the principal quantum number \(n\), the orbital angular momentum \(l\), its \(z\)-projection (\(m_l\)), and the \(z\)-projection of the spin (\(m_s\)). This can be achieved through a unitary transformation \[= R [n l m_l m_s],\] the details of which can be found in Ref.. Furthermore, the spherical basis can be recoupled from the \(l\)-\(s\) coupling to the \(j\)-\(j\) coupling through the use of Clebsch-Gordan coefficients \[= C [n l j m_j],\] where the total angular momentum \(j\) and its \(z\)-projection appear. Combining these two transformations one obtains the desired connection between the cartesian Elliott basis and the spherical shell model basis in \(j\)-\(j\) coupling \[= R C [n l j m_j].\] In Ref. one can find details of the calculations and transformation tables up to the \(N=3\) shell. The above transformation means that the Nilsson 0\[110\] replacements made within the proxy-SU(3) scheme correspond to \(| 0 1 1 0\rangle\) replacements within the spherical shell model basis. In Table III the correspondence between original shell model orbitals and proxy-SU(3) orbitals is summarized, paving the way for using the proxy-SU(3) symmetry in shell model calculations for heavy nuclei, in a way similar to the recently developed for light nuclei symmetry-adapted no-core shell model approach. By analogy to the unitary transformation occurring in the framework of the pseudo-SU(3) symmetry, a unitary transformation connecting the orbitals being replaced within the proxy-SU(3) scheme has been found within the shell model basis and is depicted in Fig. 3. The above transformation indicates that the 0\[110\] Nilsson pairs identified in Ref. and used within the proxy-SU(3) scheme are identical to the above mentioned de Shalit--Goldhaber pairs and the Federman--Pittel pairs, which are expressed as \(| 0 1 1 0\rangle\) within the spherical shell model basis. Further corroboration of the correspondence between Nilsson pairs and shell model pairs has been provided by calculations within the Nilsson model, in which the first justification of the proxy-SU(3) scheme has been found, as mentioned in Sec. [5](#Sec5){reference-type="ref" reference="Sec5"}. The replacements used in proxy-SU(3) work only for the Nilsson orbitals which possess the highest total angular momentum \(j\) within their shell, as one can see in Tables IV and V. # The dominance of the highest weight irreducible representations of SU(3) {#Sec7} In order to start examining the consequences of the existence of the proxy-SU(3) symmetry in medium mass and heavy nuclei, we should first consider a few properties of the irreducible representations (irreps) of SU(3). In Elliott's notation the irreps of SU(3) are labeled by \((\lambda,\mu)\), where the Elliott quantum numbers \(\lambda\) and \(\mu\) are connected to the number of boxes in the corresponding Young diagram through the relations \[\lambda=f_1-f_2, \qquad \mu=f_2,\] where \(f_1\) (\(f_2\)) is the number of boxes in the first (second) line of the relevant Young diagram. Irreps with \(\lambda > \mu\) are known to correspond to prolate (rugby ball like) shapes, while irreps with \(\lambda < \mu\) are known to describe oblate (pancake like) shapes, with \(\lambda=\mu\) irreps corresponding to maximally triaxial shapes. A quantity characterizing the SU(3) irreps is the eigenvalue of the second order Casimir operator of SU(3), given by \[\label{Casimir} C_2(\lambda,\mu) ={2\over 3} (\lambda^2 + \mu^2 + \lambda \mu + 3\lambda + 3\mu).\] This quantity is known to be connected to the eigenvalues of the quadrupole-quadrupole interaction by \[\label{QQC2} QQ= 4 C_2-3 L(L+1),\] where \(L\) is the eigenvalue of the orbital angular momentum. For the ground state bands of even-even nuclei, in which \(L=0\), it is clear that the eigenvalue of \(C_2\) is proportional to the eigenvalue of the quadrupole-quadrupole interaction, \(QQ = 4 C_2\). Since the quadrupole-quadrupole interaction is known to cause nuclear deformation, it is expected that the irrep with the highest value of \(C_2\) should correspond to the preferred ground state with maximum deformation. There is, however, another factor which should be taken into account. The nucleon-nucleon interaction is known to be attractive and of short range, therefore favoring the maximal spatial overlap among the nucleons, which can be achieved for the most symmetric SU(3) irrep. In the Young diagrams it is known that boxes on the same line correspond to symmetrized particles, while boxes in the same column correspond to antisymmetrized particles. The degree of symmetrization of a given SU(3) irrep can therefore be measured by the ratio of the symmetrized boxes over the total number of boxes, which is \[r= {f_1 \over f_1+f_2}= {\lambda +\mu \over \lambda +2\mu}.\] It has been proved that the irreps with the highest value of \(r\), i.e., with the highest degree of symmetry, correspond to what is called in the mathematical language the highest weight (hw) irreps of SU(3). As such these irreps appear favored by the nucleon-nucleon interaction and, therefore, dominate the related nuclear properties. The irreps possessing the highest eigenvalue of the second order Casimir operator, to be called in what follows the highest \(C_2\) irreps for brevity, and the highest hw irreps of SU(3) in the various nuclear shells are shown in Table VI. It is clear that up to midshell the \(C_2\) and hw irreps are identical, while in the upper half of each shell the \(C_2\) and hw irreps are different, with the exception of the last 5 particle numbers (which correspond to states with up to 4 holes at the top of the shell). Further mathematical details on the dominance of the hw irreps can be found in Refs.. # Physical consequences of the dominance of the highest weight irreps {#Sec8} ## Prolate over oblate dominance One of the long standing puzzles in nuclear structure is the dominance of prolate (rugby ball like) shapes over oblate (pancake like) shapes in the ground state bands of even-even nuclei. In Elliott's notation, prolate (oblate) irreps correspond to \(\lambda>\mu\) (\(\lambda<\mu\)), while irreps with \(\lambda=\mu\) correspond to maximal triaxiality. Despite several attempts to resolve this puzzle, the question is still considered open. The dominance of the hw irreps over the highest \(C_2\) irreps offers a simple, parameter-free justification of the prolate over oblate dominance. As an example, in Table VII, the hw irreps corresponding to the rare earth nuclei with valence protons in the 82-126 shell and valence neutrons in the 126-184 shell are shown. These hw irreps are obtained by adding up the hw irrep \((\lambda_\pi, \mu_\pi)\) corresponding to the valence protons and the hw irrep \((\lambda_\nu, \mu_\nu)\) corresponding to the valence neutrons into the most stretched irrep \((\lambda_\pi+\lambda_\nu, \mu_\pi+\mu_\nu)\). We notice that prolate irreps are obtained over most of the table, with the exception of its lower right corner, near which a few oblate irreps (underlined in Table VII) appear in nuclei lying below the top of both the proton valence shell and the neutron valence shell. The same effect appears in other shells as well, as one can see using the irreps appearing in Table VI. For example, results for rare earths with valence protons in the 50-82 shell and valence neutrons within the 50-82 and the 82-126 shells are given in Ref.. The conclusion of this subsection has been recently corroborated by a heuristic method in Ref.. ## Parameter-free predictions for the collective variables \(\beta\) and \(\gamma\) {#subsec82} Further consequences of the dominance of the hw irreps over the highest \(C_2\) irreps become evident if one considers the connection between the Elliott quantum numbers \(\lambda\), \(\mu\) and the collective variables \(\beta\), \(\gamma\) of the Bohr--Mottelson model. This connection is obtained by employing a linear mapping between the invariant quantities of the two theories, which are the invariants \(\beta^2\) and \(\beta^3 \cos 3 \gamma\) of the collective model and the Casimir operators of second and third order of SU(3). This mapping provides for the angle collective variable \(\gamma\) the expression \[\label{g1} \gamma = \arctan \left( {\sqrt{3} (\mu+1) \over 2\lambda+\mu+3} \right),\] while for the square of the deformation parameter \(\beta\) it gives \[\label{b1} \beta^2= {4\pi \over 5} {1\over (A \bar{r^2})^2} (\lambda^2+\lambda \mu + \mu^2+ 3\lambda +3 \mu +3),\] which indicates that \(\beta^2\) is proportional to the second order Casimir operator of SU(3). In Eq. ([\[b1\]](#b1){reference-type="ref" reference="b1"}) \(A\) is the mass number of the nucleus, while \(\bar{r^2}\) is related to the dimensionless mean square radius, \(\sqrt{\bar{r^2}}= r_0 A^{1/6}\), which is obtained by dividing the mean square radius, being proportional to \(A^{1/3}\), by the oscillator length, which grows as \(A^{1/6}\). For the constant \(r_0\) the value 0.87 is used, determined from a fit over a wide range of nuclei, as done in Ref. . A word of warning is in place here. Since in proxy-SU(3) only the valence nucleons are taken into account, the values of \(\beta\) obtained from Eq. ([\[b1\]](#b1){reference-type="ref" reference="b1"}) should be rescaled by multiplying them by a factor \(A/(S_\pi+S_\nu)\), where \(S_\pi\) (\(S_\nu\)) is the size of the proton (neutron) valence shell. The need for this rescaling has been discussed in detail in Sec. V of Ref.. In practice this rescaling means that Eq. ([\[b1\]](#b1){reference-type="ref" reference="b1"}), when used in the proxy-SU(3) framework, should read \[\label{b2} \beta^2= {4\pi \over 5} {1\over ((S_\pi+S_\nu) \bar{r^2})^2} (\lambda^2+\lambda \mu + \mu^2+ 3\lambda +3 \mu +3).\] Numerical results for the collective variable \(\beta\) (\(\gamma\)) for several rare earth nuclei are shown in Fig. 4 (Fig. 5) and are compared to other theoretical approaches, like the D1S-Gogny interaction and relativistic mean field theory, as well as to experimental values. By "Gogny D1S min." we label the values of the deformation variables \(\beta\) and \(\gamma\) at the HFB minimum energy (entries 4 and 5 in Ref. ), while by "Gogny D1S mean" the mean ground state values (entries 11 and 12 in Ref. ). It is seen that the *parameter-free* proxy-SU(3) predictions are in good agreement with both the theoretical approaches and the empirical values. Additional numerical results for \(\beta\) and \(\gamma\) obtained within the proxy-SU(3) approach can be found in. Additional comparisons of proxy-SU(3) predictions to covariant density functional theory can be found in Refs., while comparisons to recent experimental findings can be found in. In Fig. 4 it is clear that the \(\beta\) curve is not symmetric around midshell, but it appears to exhibit higher values in the first half of the shell. The origin of this discrepancy can be traced in Fig. 6, in which the square root of \(C_2\), which is proportional to \(\beta\), according to Eq. ([\[b2\]](#b2){reference-type="ref" reference="b2"}), is shown. We see that the breaking of the symmetry around midshell is due to the fact that in the upper half of the shell the highest weight irreps enter in the place of the highest \(C_2\) irreps, as indicated by Table 6. ## Prolate to oblate shape/phase transition A second important consequence of the hw irreps dominance is seen in Fig. 7, in which the proxy-SU(3) predictions for the collective variable \(\beta\) for the rare earths with valence protons in the 50-82 shell and valence neutrons in the 82-126 shell are collected. The dip seen at \(N=116\) signifies the occurrence of a shape/phase transition from prolate to oblate shapes, for which extended experimental evidence exists, along with microscopic theoretical considerations, and relevant searches within the interacting boson model (IBM) and the Bohr--Mottelson collective model. In the framework of the Bohr Hamiltonian this shape/phase transition has been referred to as Z(5). *A posteriori* corroboration of the findings of Ref. is found in Refs.. The robustness of this result is emphasized by the fact that it also appears in atomic clusters. The valence electrons in alkali metal clusters, in particular, are supposed to be free, thus forming shells with major magic numbers 2, 8, 20, 40, 58, 92, .... Prolate and oblate shapes in alkali metal clusters have been observed experimentally through optical response measurements, finding oblate shapes below cluster sizes 20 and 40, while prolate shapes have been seen above cluster sizes 8, 20, 40. In other words, prolate (oblate) shapes are seen above (below) the magic numbers, exactly as in atomic nuclei. It should be noticed that the shape/phase transition from prolate to oblate shapes is seen equally clearly in the framework of pseudo-SU(3), if the dominance of the hw irreps over the highest \(C_2\) irreps is taken into account, as discussed in detail in Ref.. Although pseudo-SU(3) and proxy-SU(3) are based on different approximations, involving different unitary transformations, they lead to the same physical conclusion, a fact providing evidence for the compatibility of the two approaches. The compatibility of the proxy-SU(3) and pseudo-SU(3) approaches has also been pointed out recently in the framework of the semimicroscopic algebraic quartet model (SAQM), in which quartets consisting of two protons and two neutrons occupying a single orbital are taken into account, their excitation spectra occurring when a quarter jumps as a whole to the next major shell. Based on the SU(3) symmetry, in its initial form the model had been applicable only to light nuclei, based on the Elliott SU(3) symmetry. However, it has recently been extended to heavy nuclei, based on the proxy-SU(3) symmetry and on the pseudo-SU(3) symmetry, with similar results provided by both approaches. The use of proxy-SU(3) could also be extended in a similar way to the Semimicroscopic Algebraic Cluster Model (SACM), in which the internal structure of the clusters is described in terms of the shell model SU(3) symmetry, while their relative motion is described in terms of the phenomenological algebraic vibron model. # Islands of shape coexistence {#Sec9} ## Harmonic oscillator (HO) and spin-orbit (SO) magic numbers Shape coexistence (SC) is called the appearance in a nucleus of two bands lying close in energy but having radically different structures, for example one of them being spherical and the other deformed, or both of them being spherical, but one of them exhibiting prolate (rugby ball like) deformation and the other one oblate (pancake like) deformation. Shape coexistence has first been suggested in 1956 by Morinaga, in relation to the spectrum of \(^{16}\)O. Since then many experimental examples have been found in both odd and even nuclei, as summarized in the relevant review articles. From the theoretical point of view, SC has been attributed to the existence of particle-hole excitations across shell or subshell closures, and has been believed to be able to appear all over the nuclear chart, although in Fig. 8 of the authoritative review article by Heyde and Wood the regions in which SC has been experimentally observed appear to form certain islands on the nuclear chart. As it has already been mentioned, SC is supposed to be due to particle-hole excitations across shell or subshell closures. However, the magic numbers of the shell model, 2, 8, 20, 28, 50, 82, 126, ..., are known to be valid only at zero deformation. These magic numbers originate from the 3D-HO magic numbers 2, 8, 20, 40, 70, 112, 168, ..., (to be called the HO magic numbers in what follows) because of the action of the spin-orbit interaction. As deformation sets in, the energy gaps separating different shells soon disappear, as one can see in the standard Nilsson diagrams (see also Refs. for the evolution of magic numbers away from stability). Furthermore, as seen in Table III (see also Ref. ), in the framework of the proxy-SU(3) symmetry a set of magic numbers 6, 14, 28, 50, 82, 126, 184, ...(to be called the SO magic numbers in what follows) appears, corresponding to the strong presence of the spin orbit interaction everywhere. We remark that the standard shell model magic numbers follow the HO magic numbers up to 20, while they follow the SO magic numbers beyond this point. ## A dual shell mechanism for shape coexistence It has been suggested that the interplay of HO and SO magic numbers offers a simple justification for the appearance of islands of SC on the nuclear chart. Let us see how this is occurring. Because of the collapse of the shell model quantum numbers as deformation sets in, the protons and/or the neutrons of a nucleus can follow either the HO or the SO magic numbers. The same number of protons or neutrons will then correspond to a different irrep in the HO framework and to another irrep in the SO framework. These irreps can be seen in Table VIII. As an example, let us consider 60 nucleons. In the HO framework, 40 is a magic number, thus there are left 20 nucleons in the 40-70 shell, which has the U(15) symmetry, and therefore the 20 nucleons correspond to the (20,0) hw irrep according to Table VI. In the SO framework, 50 is a magic number, thus there are 10 nucleons left in the 50-82 shell, which in the proxy-SU(3) approximation has the U(15) symmetry, therefore the 10 nucleons correspond to the (20,4) hw irrep of SU(3) according to Table VI. Indeed in Table VIII for 60 nucleons the hw irreps given are (20,4) for the SO case and (20,0) for the HO case. For the \(L=0\) bandheads of two coexisting bands one can use the very simple Hamiltonian \[H = H_0-{\kappa\over 2} QQ,\] where \(H_0\) corresponds to the 3D-HO Hamiltonian and \(QQ\) to the quadrupole-quadrupole interaction. Both bands should belong to the same U(n) algebra within the SO and HO schemes. From Table III we see that this is possible for the nucleon number intervals 6-8, 14-20, 28-40, 50-70, 82-112, 126-168, in which both bands belong to the U(3), U(6), U(10), U(15), U(21), U(28) algebra respectively. We remark that the right borders of these regions are the HO magic numbers. The successful parameter-free predictions of the \(\beta\) and \(\gamma\) collective variables for the ground states of nuclei seen in Sec. [8.2](#subsec82){reference-type="ref" reference="subsec82"} imply that the ground-state band will belong to the SO irrep, thus the bandhead of the coexisting band should lie higher in energy. One can easily see that this requirement leads to the condition \[QQ_{SO} \geq QQ_{HO},\] the full details of the argument given explicitly in Sec. 8 of Ref.. From Eq. ([\[QQC2\]](#QQC2){reference-type="ref" reference="QQC2"}) one then sees that this condition is equivalent to the condition \[\label{ineq} C_2(\lambda_{SO},\mu_{SO}) \geq C_2(\lambda_{HO},\mu_{HO}).\] The eigenvalues of the Casimir operator \(C_2\) in the SO and HO frameworks are shown for the various shells in Fig. 8. We see that the condition of Eq. ([\[ineq\]](#ineq){reference-type="ref" reference="ineq"}) starts being fulfilled at the nucleon numbers 7, 17, 34, 59, 96, 146, which will therefore stand for the left borders of the regions in which SC could be possible, the right borders being given by the HO magic numbers mentioned above. Therefore we conclude that SC can occur only within the nucleon intervals 7-8, 17-20, 34-40, 59-70, 96-112, 146-168, bearing the U(3), U(6), U(10), U(15), U(21), U(28) symmetry respectively. These intervals define horizontal and vertical stripes on the nuclear chart, shown in color in Fig. 9. One can easily see that the islands of SC seen in Fig. 8 of Ref. do lie within the stripes predicted by the dual shell mechanism within the proxy-SU(3) framework just found. ## From stripes to islands of shape coexistence The stripes 7-8, 17-20, 34-40, 59-70, 96-112, 146-168 determined in the previous subsection represent a necessary condition for the appearance of SC, but not a sufficient one. Further work is needed in order to narrow down the stripes into islands. A step in this direction has been taken within covariant density functional theory , using the DDME2 functional and the code of Ref.. A systematic search has been made for particle-hole excitations, which are thought for a long time of being the microscopic mechanism behind SC. Indeed specific islands of SC have been located around the proton shell closures Z=82 and Z=50, in which proton particle-hole excitations are caused by the neutrons, therefore characterizing these cases as ones of neutron-induced shape coexistence. In addition, specific islands of SC have been located around the neutron numbers N=90 and N=60, in which neutron particle-hole excitations are caused by the protons, therefore characterizing these cases as ones of proton-induced shape coexistence. Furthermore, an island of SC has been found around \(Z=N=40\), in which both the proton-induced and neutron-induced mechanisms are present simultaneously. All these islands are consistent with the stripes of the dual shell mechanism within the proxy-SU(3) symmetry, as well as with the empirical islands reviewed in Fig. 8 of Ref.. Further corroboration of the predictions of the dual shell mechanism has been provided by various relativistic microscopic calculations, as well as by calculations using the Bohr Hamiltonian and the IBM. It should be noticed that the above predictions of specific islands of SC are based on the particle-hole excitation mechanism. It could be possible that some other microscopic mechanism creates SC in regions of the nuclear chart outside the islands predicted here. This can be the subject of further investigation. ## Multiple shape coexistence Recent experimental evidence exists that multiple shape coexistence of up to four bands can be seen in certain nuclei, while it is also predicted theoretically in others. Multiple shape coexistence can occur within the dual shell mechanism described above, since the protons can follow either the SO or the HO scheme, and so can independently do the neutrons. As a result, four different irreps, based on the proton-neutron combinations SO-SO, SO-HO, HO-SO, HO-HO can occur in general, giving rise to multiple coexistence of four bands, or three bands in the special case in which equal numbers of valence protons and valence neutrons occupy the same shell. This idea, however, still needs to be tested against experimental evidence. # Conclusion and outlook In this article we have discussed the physical ideas which led to the introduction of the proxy-SU(3) symmetry, the calculations proving its validity and its connection to the shell model framework, as well as its first successful applications in predicting in a parameter-free way the values of the collective variables \(\beta\) and \(\gamma\) for even-even nuclei, as well as the dominance of prolate over oblate shapes in the ground states of even-even nuclei, a prolate to oblate shape/phase transition, and the existence of islands on the nuclear chart on which shape coexistence can appear. Several directions for further investigations are calling attention, as we shall briefly discuss below. The proxy-SU(3) symmetry offers the possibility of making parameter-free predictions for B(E2) transition rates. The value of \(B(E2; 2_1^+\to 0_1^+)\) is known to be connected to the collective variable \(\beta\), thus it can be determined in a parameter-free way from the \(\beta\) values calculated as described in Sec. XIII.B. Since ratios of B(E2)s will only depend on angular momentum coupling coefficients of SO(3) and SU(3), parameter-free predictions for all B(E2)s could be obtained in principle. Some first steps in this direction have been taken in. Based on the experience acquired within the pseudo-SU(3) model, it is expected that the description of nuclear spectra within the proxy-SU(3) symmetry will require the use of third-order and fourth-order operators. In particular, the O(3) symmetry-preserving three-body operator \(\Omega\) and four-body operator \(\Lambda\) (their mathematical names being the \(O_l^0\) and \(Q_l^0\) shift operator respectively) will be needed in order to break the degeneracy between the ground state band (gsb) and the \(\gamma_1\) band, which in the proxy-SU(3) approach lie in general within the same SU(3) irrep. This can be seen, for example, in Table VII, where almost all nuclei are characterized by SU(3) irreps with \(\mu \geq 2\). Since \(K\) takes the values \(K=0,2,\dots,\mu\), the hw irrep will contain both the \(K=0\) (gsb) and \(K=2\) (\(\gamma_1\)) bands. The parameter-free reproduction of the empirical observation that the energy differences between the \(\gamma_1\) band and the ground state band decrease as a function of the angular momentum \(L\) in deformed nuclei, with the opposite trend seen in vibrational and \(\gamma\)-unstable nuclei, should be tested. Some first steps in this direction have been taken in Refs.. It should be noticed that the energy scale can be fixed in an even-even nucleus by determining the energy of the first excited state \(2_1^+\) from the value of the \(B(E2; 2_1^+\to 0_1^+)\), determined in a parameter-free way, as described in section VIII.B. This can be achieved through the microscopically derived Grodzins relation, connecting the energy of the \(2_1^+\) state and \(B(E2; 2_1^+\to 0_1^+)\). Nuclear binding energies and nucleon separation energies are basic nuclear structure quantities, for which extended experimental data and theoretical predictions exist. It is an interesting project to examine the degree at which proxy-SU(3) is able to predict these quantities, preferably in a parameter-free way. Some first steps in this direction have been taken in Refs.. The calculation of two-neutron separation energies is of particular interest, since the recently discovered connection between them and the neutron capture cross sections, which are essential for understanding the astrophysical \(s\) and \(r\) processes.

{'timestamp': '2022-12-19T02:12:31', 'yymm': '2212', 'arxiv_id': '2212.08494', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08494'}

# Introduction An \(m\times n\) partially filled (p.f., for short) array on a set \(\Omega\) is an \(m \times n\) matrix whose elements belong to \(\Omega\) and where some cells can be empty. In 2015, Archdeacon (see [@A]), introduced a class of p.f. arrays that have been extensively studied: the *Heffter arrays*. These arrays have been introduced because of their vast variety of applications and links to other problems and concepts, such as orthogonal cycle decompositions and \(2\)-colorable embeddings (briefly *biembeddings*), see for instance. The existence problem of Heffter arrays has been also deeply investigated: we refer to the survey for the known results in this direction. This paper will focus mainly on the connection between p.f. arrays and embeddings. To explain this link, we first recall some basic definitions, see. The connected components of \(\Sigma \setminus \psi(\Gamma)\) are called \(\psi\)-*faces*. Also, with a slight abuse of notation, we say that a circuit \(F\) of \(\Gamma\) is a face (induced by the embedding \(\psi\)) if \(\psi(F)\) is the boundary of a \(\psi\)-face. Then, if each \(\psi\)-face is homeomorphic to an open disc, the embedding \(\psi\) is said to be *cellular*. In this context, we say that two embeddings \(\psi: \Gamma \rightarrow \Sigma\) and \(\psi': \Gamma' \rightarrow \Sigma'\) are *isomorphic* whenever there exists a graph isomorphism \(\sigma: \Gamma\rightarrow \Gamma'\) such that \(\sigma(F)\) is a \(\psi'\)-face if and only if \(F\) is a \(\psi\)-face. Here we say that \(\sigma\) is an embedding isomorphism or, if \(\psi=\psi'\), an embedding automorphism. Archdeacon, in his seminal paper, showed that, if some additional technical conditions are satisfied, Heffter arrays provide explicit constructions of \(\mathbb{Z}_{v}\)-regular biembeddings of complete graphs \(K_v\) into orientable surfaces. Here *\(\mathbb{Z}_{v}\)-regular* means that \(\mathbb{Z}_v\) acts sharply regularly on the vertex set, hence \(\mathbb{Z}_v\) is contained in the automorphism group of \(\psi\), denoted by \(Aut(\psi)\). Following, the embeddings defined using this construction via p. f. arrays will be denoted as *embeddings of Archdeacon type* or, more simply, *Archdeacon embeddings*. Indeed, this kind of embedding can be considered also for variations of the concept of Heffter arrays, such as the *non-zero sum Heffter arrays* discussed in (see also for other variations and generalizations). In, the author provided a generalization of both the *Heffter* and the *non-zero sum Heffter* arrays and showed that the Archdeacon embedding can be also defined in this more general context. Given \(v=2nk+1\) and a group \(G\) of order \(v\), he defined the so-called *quasi*-*Heffter array \(A\) over \(G\)*[^1] (denoted as \(\Q\H(m,n; h,k)\)) by considering an \(m\times n\) p.f. array with elements in \(G\) such that: - each row contains \(h\) filled cells and each column contains \(k\) filled cells, - the multiset \([\pm x \mid x \in A]\) contains each element of \(G\setminus \{0\}\) exactly once. If every row and every column of a quasi-Heffter sum to zero, consistently with the classical Heffter arrays nomenclature, we denote this array by \(\H(m,n; h,k)\) over \(G\) and, if also \(m=k\) and \(n=h\), by \(\H(m,n)\) over \(G\). Using such arrays, the Archdeacon embedding into orientable surfaces is still well defined, with essentially the same construction of. Then, again in, it was proved that for an embedding constructed from a \(\Q\H(m,n; h,k)\) over \(\Z_v\), its full automorphism group, which acts sharply transitively on the vertex set, is **almost always** exactly \(\mathbb{Z}_{v}\). Note that several examples of \(\Z_v\)-regular embeddings whose full automorphism group is larger than \(\Z_v\) are known (see for instance Example 3.4 of or ), but none of them arises from the Archdeacon construction. For this reason, we find it natural to investigate this existence problem. The main result here presented is indeed the existence of an infinite family of such embeddings whose automorphism groups are strictly larger than \(\Z_v\). In Section \(2\) we will provide a formal definition of the Archdeacon embedding, starting from a Heffter array over a group \(G\). Then, in Section 3 we will analyze one of the interesting classes of Heffter arrays over \(\mathbb{F}_q\) recently introduced by Buratti in. We will show that using these arrays it is possible to obtain infinitely many embeddings of Archdeacon type with a rich automorphism group: more precisely the stabilizer of a given point in the automorphism group has size \((v-1)/2\) and we will show that this is the largest possible one for such embeddings. Moreover, since \(\mathbb{F}_p=\Z_p\) (assuming \(p\) is a simple prime), infinitely many of these embeddings arise from classical Heffter arrays. # The Archdeacon Embedding and its Automorphisms Following, we provide an equivalent, but purely combinatorial, definition of a graph embedding into a surface. Here we denote by \(D(\Gamma)\) the set of all the oriented edges of the graph \(\Gamma\) and, given a vertex \(x\) of \(\Gamma\), \(N(\Gamma,x)\) denotes the set of vertices adjacent to \(x\) in \(\Gamma\). Then, as reported in, a combinatorial embedding \(\Pi=(\Gamma,\rho)\) is equivalent to a cellular embedding \(\psi\) of \(\Gamma\) into an orientable surface \(\Sigma\) (see also, Theorem 3.1). Now we recall the definition of Archdeacon embedding. Here we report it in the case of Heffter arrays over a group \(G\), but we remark that the same embedding is still well-defined also for quasi-Heffter arrays (see, Definition 2.4). We first introduce some notation. The rows and the columns of an \(m\times n\) array \(A\) are denoted by \(R_1,\ldots, R_m\) and by \(C_1,\ldots, C_n\), respectively. Also, by \(\E(A)\), \(\E(R_i)\), \(\E(C_j)\) we mean the list of the elements of the filled cells of \(A\), of the \(i\)-th row and of the \(j\)-th column, respectively. Given an \(m\times n\) p.f. array \(A\), by \(\omega_{R_i}\) and \(\omega_{C_j}\) we denote respectively an ordering of \(\E(R_i)\) and \(\E(C_j)\), and we define by \(\omega_r=\omega_{R_1}\circ \cdots \circ \omega_{R_m}\) the ordering for the rows and by \(\omega_c=\omega_{C_1}\circ \cdots \circ \omega_{C_n}\) the ordering for the columns. Now we are ready to recall the definition of Archdeacon embedding, see. Indeed, Archdeacon proved in that, since the orderings \(\omega_r\) and \(\omega_c\) are compatible, the map \(\rho\) is a rotation of \(K_{v}\) (he considered the case \(G=\Z_v\) but, as noted in, the same proof holds in general), and thus the pair \((K_v,\rho)\) is a combinatorial embedding of \(K_v\). More precisely he showed the following theorem. He also described the faces induced by the Archdeacon embedding under the conditions of Theorem [\[HeffterBiemb\]](#HeffterBiemb){reference-type="ref" reference="HeffterBiemb"}. For this purpose, we take a p.f. array \(A\) that is an \(\H(m,n;h,k)\) admitting two compatible orderings \(\omega_r\) and \(\omega_c\). We consider the oriented edge \((x,x+a)\) with \(a \in \E(A)\). Due to Theorem 1.1 of, the directed edge \((x,x+a)\) belongs to the face \(F_1\) whose boundary is \[\label{F1}\left(x,x+a,x+a+\omega_c(a),\ldots,x+\sum_{i=0}^{k-1} \omega_c^i(a)\right).\] Let us now consider the oriented edge \((x,x+a)\) with \(a\not \in \E(A)\). In this case, \((x,x+a)\) belongs to the face \(F_2\) whose boundary is \[\label{F2}\left(x,x+\sum_{i=1}^{h-1}\omega_{r}^{-i}(-a),x+\sum_{i=1}^{h-2}\omega_{r}^{-i}(-a), \dots,x+\omega_{r}^{-1}(-a)\right).\] A priori these faces are circuits but, under suitable conditions, we can prove that they are simple cycles. To be more precise we need to introduce some further definitions. Given a finite subset \(T\) of an abelian group \(G\) and an ordering \(\omega = (t_1, t_2,\dots, t_k)\) of the elements in \(T\), for any \(i\in [1, k]\) let \(s_i =\sum_{j=1}^i t_j\) be the \(i\)-th partial sum of \(\omega\). The ordering \(\omega\) is said to be *simple* if \(s_b\not=s_c\) for any \(1\leq b,c\leq k\). Given an \(m\times n\) p.f. array \(A\) whose entries belong to a given group \(G\), and given an ordering \(\omega_{C_i}\) for any column \(C_i\) where \(1\leq i\leq n\) and an ordering \(\omega_{R_j}\) for any row \(R_j\) where \(1\leq j\leq m\), we say that \(\omega_r=\omega_{R_1}\circ \cdots\circ\omega_{R_m}\) and \(\omega_c=\omega_{C_1}\circ \cdots\circ\omega_{C_n}\) are *simple* if each \(\omega_{C_i}\) and \(\omega_{R_j}\) is simple. If the natural orderings, from top to bottom for each column and from left to right for each row are simple we say that the array is *globally simple*. We can restate the main result of Archdeacon as follows: Now, in order to investigate the automorphism group of an embedding of Archdeacon type, we revisit the necessary conditions stated in. We first recall that also the notions of embedding isomorphism and automorphism can be defined purely combinatorially as follows (see Korzhik and Voss, page 61). Let \(\Pi\) be an Archdeacon embedding of \(K_{v}\) obtained from a Heffter array \(A\). Then, we denote by \(Aut(\Pi)\) the group of all automorphisms of \(\Pi\), and by \(Aut^+(\Pi)\) we mean its subgroup of orientation-preserving automorphisms. Similarly, by \(Aut_0(\Pi)\) and \(Aut_0^+(\Pi)\) we respectively denote the subgroups of \(Aut(\Pi)\) and of \(Aut^+(\Pi)\) of automorphisms that fix \(0\). Here we have that \(Aut^+(\Pi)\) and \(Aut_0^+(\Pi)\) are, respectively, normal subgroups of \(Aut(\Pi)\) and \(Aut_0(\Pi)\) (their index is either \(1\) or \(2\)). Moreover, if for every \(g \in G\) we denote by \(\tau_g\) the translation action by \(g\), i.e. \(\tau_g(x) = x+g\), it has been proved by Archdeacon (see, Theorem 1.1) that \(\tau_g \in Aut^+(\Pi)\), hence \(G\) is a subgroup of \(Aut^+(\Pi)\). Thus, the automorphism group of an embedding of Archdeacon type is never trivial. On the other hand, in the case \(G=\Z_v\) it has been proven in that, if we start from a quasi-Heffer array, \(Aut(\Pi)\) is almost always exactly \(\Z_v\). Hence we find it natural to investigate whether, given a biembedding of Archdeacon type \(\Pi\), the only automorphism of \(\Pi\) fixing \(0\) must be the identity. Now, we recall some necessary conditions (see ) that have to be satisfied by an automorphism of \(\Pi\) that fixes zero. # A class of highly symmetric embeddings In this section we provide for infinitely many values of \(q\) (that here is a prime power) an embedding of Archdeacon kind, defined over the additive group of \(\mathbb{F}_q\) (the field of order \(q\)), whose full automorphism group is of size \({q\choose 2}\): more precisely this embedding admits \(\mathbb{F}_q\) as a regular automorphism group and the stabilizer of \(0\) is isomorphic to \(\Z_{(q-1)/2}\). We begin by recalling some definitions, firstly presented in. Given a prime power of the form \(q=2mn+1\), an \(H(m,n)\) over \(\mathbb{F}_q\) is a *rank-one* Heffter array if every row is a multiple of some non-zero vector \(X = (x_1, \dotsc, x_n)\) of \(\mathbb{F}_q^n\). Notice that this implies that every column of the array is a multiple of some non-zero vector \(Y = (y_1,\dotsc,y_m)\) of \(\mathbb{F}_q^m\). Again in, Buratti provided a construction of rank one Heffter arrays over \(\mathbb{F}_q\), where \(q=2mn+1\) is a prime power in several cases: we consider here its construction when \(m\) and \(n\) are odd and coprime integers. In this case, given \(\[ \begin{aligned} X &= (1, Y &= (1,\epsilon,\epsilon^2,\epsilon^3, \dotsc, \epsilon^{m-1}).\\ \end{aligned} \] Then, defining the\)mn\(array\)A\_m,n=(a\_i,j)\(whose\)(i,j)\(-cell is given by\)a\_i,j:=\^i-1 Buratti also noticed that these arrays have many symmetries (the so-called automorphisms of the array). Our goal is now to consider a family of embeddings that arise from these arrays and to prove that also these embeddings have a lot of symmetries. In particular, we will see that the full automorphism groups of these embeddings are the largest possible one. First of all, in order to define the Archdeacon embeddings induced by such arrays, we need to recall the following result of (see also ) about compatible orderings of totally filled \(m\times n\) arrays. This means that, if \(m\) and \(n\) are coprime, we can choose \(\ell=0\), obtaining that: Now we want to investigate the automorphism group of such an embedding. We begin by considering the orientation-preserving automorphisms. Now we will show that these embeddings do not admit orientation-reversing automorphisms. We begin by proving a technical lemma: We are then able to derive the following result, that exactly determines the size of the full automorphism group of this class of Archdeacon embedding. We also note that in the paper is presented a construction of a \(H(m,n)\) that admits non-trivial symmetries whenever \(q=2mn+1\) is a prime power and \(mn\) is not a power of two. We remark that also from all these arrays (assuming at least one between \(m\) and \(n\) is odd) we can obtain an Archdeacon embedding with non-trivial automorphisms that fix zero. On the other hand, if \(m\) and \(n\) are not both odd or coprime, we have much fewer symmetries and we are not able to determine exactly the full automorphism group of the embedding. For this reason, we concentrated on the case where \(m\) and \(n\) are both odd and coprime. We conclude the paper by showing an example of an Archdeacon embedding of type \(\Pi_{m,n}\) and by discussing its automorphisms.

{'timestamp': '2022-12-19T02:12:25', 'yymm': '2212', 'arxiv_id': '2212.08491', 'language': 'en', 'url': 'https://arxiv.org/abs/2212.08491'}

# Introduction {#sec:intro} In describing the transition from baryonic or hadronic matter to that of its constituents such of up, down and strange quarks in the interiors of neutron stars (NSs), the most commonly employed methods are either the Maxwell or the Gibbs construction . In the Maxwell construction, charge neutrality is achieved locally, whereas in the Gibbs construction the same is achieved globally. The Maxwell construction is applicable when the interface or surface tension between the two phases is very large (or infinite), whereas the Gibbs construction is valid in the opposite limit of vanishing (or zero) surface tension. While the former method is suitable for transitions from a single component system (say, neutrons only), the latter is well-suited when multiple charges such as neutrons, protons, and electrons are present in the transitioning system, particularly to account for separate baryon number conservation and charge neutrality. For intermediate surface tensions, the shape of the phase boundary could vary with density as in the pasta phase at the crust-core transition treated in the Wigner-Seitz approximation . Nevertheless, the magnitude of the quark-hadron interface tension is highly uncertain, ranging from a few to hundreds of MeV/fm\(^2\); see e.g. discussions in the literature.\ In the pressure \(P\) vs energy density \(\varepsilon\) plane, the Maxwell construction in which the pressure and neutron chemical potential equalities, \(P(H)=P(Q)\) and \(\mu_n(H)=\mu_n(Q)\), are established between the hadronic (\(H\)) and quark (\(Q\)) phases, is characterized by a flat region. The range of densities over which these equalities hold can be determined using the methods described in Refs. . A consequence of this flat region is that the squared equilibrium speed of sound \(c_{\rm eq}^2=dP/d\varepsilon\) becomes zero there. The density region over which the flat region occurs as well as the extent of the jump in the energy density depend on the details of the \(P~ {\rm vs}~\varepsilon\) relationships, or the equation of state (EOS), in each of the two phases.\ The description of the mixed phase in the Gibbs construction is achieved by satisfying the rules \(P(H)=P(Q)\) and \(\mu_n(H)=\mu_u+2\mu_d\), where the chemical potentials \(\mu_u\) and \(\mu_d\) refer to those of the up (\(u\)) and down (\(d\)) quarks, respectively. The conditions of global charge neutrality and baryon number conservation are imposed through the relations \[\begin{aligned} Q &=& fQ(H) + (1-f)Q(Q) =0 \nonumber \\ n_{\rm B} &=& fn_{\rm B}(H) + (1-f)n_{\rm B}(Q) \,, \end{aligned}\] where \(f\) denotes the fractional volume occupied by hadrons and is solved for each baryon density \(n_{\rm B}\). Unlike in the pure phases of the Maxwell construction, \(Q(H)\) and \(Q(Q)\) do not separately vanish in the Gibbs mixed phase. The total energy density is given by \[\begin{aligned} \varepsilon = f\varepsilon(H) + (1-f)\varepsilon(Q) \. \end{aligned}\] Relative to the Maxwell construction, the behavior of the pressure vs baryon density is smooth in the case of Gibbs construction. Discontinuities in its derivatives with respect to baryon density, reflected in \(c_{\rm eq}^2=dP/d\varepsilon\), will however, be present at the densities where the mixed phase begins and ends.\ Situations in which neither the Maxwell nor the Gibbs construction can be applied correspond to cases in which the pressure and chemical potential equalities cannot be met for many hadronic and quark EOSs. In such cases, interpolatory techniques which make the transition a smooth crossover have been used in Refs. . In these approaches, the pressure equality between the two phases characteristic of Maxwell and Gibbs constructions is abandoned, but the pressure vs baryon density in the mixed phase is composed of contributions from hadrons and quarks in an externally prescribed proportion. Outside of the mixed phase, pure hadronic and quark phases exist. The onset and ending densities of the mixed phase are chosen suitably for smooth crossover from one to the other phase.\ The model termed quarkyonic matter departs from first-order phase transitions inasmuch as once quarks appear, both nucleons and quarks coexist until asymptotically large baryon densities when the baryon concentrations vanish . The order of the phase transition depends on the implementation of these models. In Ref. , the transition is second order, but other approaches  have yielded higher-order phase transitions. A characteristic feature of the quarkyonic models is that the \(c_{\rm eq}^2 = dP/d\varepsilon\) exhibits a peak before approaching the value of \(1/3\), an attribute of asymptotically free quarks. Depending on the approach adopted, this value may also be reached from below . A drawback of the quarkyonic model with a second-order transition is that the squared adiabatic speed of sound \(c_{\rm ad}^2 = (\partial P/\partial \varepsilon)_{y_p}\), where \(y_p\) is the proton fraction, becomes infinite at the onset of quarks . This feature prevents the calculation of oscillation modes of NSs, particularly the \(g\)-modes (gravity modes) .\ A crossover model for the transition from hadrons to quarks in NSs to mimic the crossover feature of baryon-free finite temperature studies has also been investigated in Ref. . The key feature of this approach is an analytic mixing or switching function that accounts for the partial pressure of each component as a function of a single thermodynamic variable-the baryon chemical potential. As in the mixed phase hadrons/nucleons and quarks both appear explicitly as separate degrees of freedom in this description, it is straightforward to keep track of their individual contributions to the total pressure. In Ref. , this approach was generalized to beta-equilibrated matter in order to explore non-radial \(g\)-mode oscillations of NSs.\ Our goal here is to devise a framework in which the Maxwell and Gibbs constructions are two extremes of a continuous spectrum of possibilities for first-order phase transitions. We accomplish our goal by postulating three distinct electron clouds (with labels \(eN\), \(eQ\), and \(eG\)) whose members, are either strictly in contact with only nucleons (\(eN\)) or quarks (\(eQ\)), or can be shared between the two phases (\(eG\)). Thus, charge neutrality is fulfilled partially locally and partially globally, the ratio being controlled by a new variable \(g\), which stands for the local-to-total electron ratio. Note that, here, we do not posit distinguishable electrons in the sense of intrinsic quantum numbers; instead, we simply group them in relation to the many-body environment in which they are embedded. This grouping is artificial and, at the end of the calculation, we will be interested only in the total number or fraction of electrons required for the system to be charge-neutral.\ The physical picture is as follows. In the case of a large surface tension between the hadron and the quark phases, the boundary between the two is sharp, and the region each phase occupies is well defined. Correspondingly, the electrons ensuring charge neutrality will be unequivocally associated with (or, at the very least, are far more likely to interact with) one or the other phase by virtue of their spatial position; thus charge neutrality is local. In the opposite limit of very-low/zero surface tension, there is no spatial separation between the two phases and thus charge neutrality is accomplished entirely globally.\ For intermediate surface tension, the boundary between the two phases becomes fuzzy and therefore, in addition to the two unambiguous regions from before, we have a third, grey-zone region where the phase of baryonic matter is unclear. Consequently, some electrons will be explicitly attached to one or the other phase, while the rest interact with both. In this case, charge neutrality is fulfilled partially locally (by some electrons) and partially globally (by the remaining electrons).\ The organization of this paper is as follows. In Sec. II, the formalism to obtain a continuous spectrum of possibilities between the Maxwell and Gibbs constructions is detailed. Here, the relevant equations to describe matter with nucleons, quarks, and electrons as well as those including muons are provided. The equations of state for nucleons, quarks, leptons, and the squared equilibrium and adiabatic sound speeds are given in Sec. III. Non-radial \(g\)-mode oscillations are discussed in Sec. IV. Results of our calculations are presented in Sec. V. A summary and conclusions are contained in Sec. VI.\ # Simulating transitions between Maxwell and Gibbs constructions {#sec:framework} In this section, we present the formalism to obtain a continuous spectrum of possibilities between the Maxwell and Gibbs constructions for first-order phase transitions. We begin with matter containing nucleons (\(N\)), quarks (\(Q\)) and electrons (\(e\)) only. Thereafter, the discussion includes muons (\(\mu\)) as well. Relations corresponding to the conservation laws of baryon number and charge neutrality that connect the various particle fractions \(y_i\), with \(i\) covering \(N=n,p\), \(Q=u,d,s\), and the volume fractions \(f\) and \(g\) are presented first. The working equations result from energy density minimization with respect to the list of variables in \(NQe\mu\) and \(f\). Values of the local-to-total electron ratio \(g\) are chosen parametrically in the range \((0,1)\).\ ## \(NQe\) matter The total energy density of the system is given by the sum of appropriately-weighted contributions from the individual components: \[\begin{aligned} \varepsilon &=& f(\varepsilon_n + \varepsilon_p + g\varepsilon_{eN}) \nonumber \\ &+& (1-f)(\varepsilon_u + \varepsilon_d + \varepsilon_s + g\varepsilon_{eQ}) \nonumber \\ &+&(1-g)\varepsilon_{eG} \,, \label{edenmix} \end{aligned}\] where \(f\) is the hadron-to-baryon fraction and \(g\) the ratio of electrons participating in local charge neutrality to the total number of electrons. Baryon and lepton conservation correspond to the equations \[\begin{aligned} 1 &=& f(y_n + y_p) + (1-f)(y_u + y_d + y_s)/3 \label{baryonnum} \\ 0 &=& y_e-fgy_{eN}-(1-f)gy_{eQ}-(1-g)y_{eG} \label{electronnum} ~, \end{aligned}\] whereas charge neutrality is described by the relations \[\begin{aligned} 0 &=& (y_p-y_{eN})g \label{lcnH} \\ 0 &=& [(2y_u-y_d-y_s)/3-y_{eQ}]g \label{lcnQ} \\ 0 &=& [fy_p + (1-f)(2y_u-y_d-y_s)/3-y_{eG}](1-g) \label{gcn} ~. \end{aligned}\] The overall factors of \(g\) and \((1-g)\) in Eqs. ([\[lcnH\]](#lcnH){reference-type="ref" reference="lcnH"})-([\[lcnQ\]](#lcnQ){reference-type="ref" reference="lcnQ"}), and Eq. ([\[gcn\]](#gcn){reference-type="ref" reference="gcn"}), respectively, are not necessary but they have been kept to emphasize the fact that these equations describe *partial* local charge neutrality (LCN) and global charge neutrality (GCN). Equations ([\[baryonnum\]](#baryonnum){reference-type="ref" reference="baryonnum"})-([\[gcn\]](#gcn){reference-type="ref" reference="gcn"}) are then used to eliminate five of the twelve free variables (\(n_{\rm B}\), \(y_n\), \(y_p\), \(y_u\), \(y_d\), \(y_s\), \(y_{eN}\), \(y_{eQ}\), \(y_{eG}\), \(y_e\), \(f\), \(g\)) in this scheme. The choice is arbitrary but the most convenient set (that is, the set that leads to physically-transparent phase-equilibrium conditions in the fewest number of operations) is the following: \[\begin{aligned} y_u &=& \frac{1 + y_e-fy_n-2fy_p}{1-f} \\ y_d &=& \frac{2-y_e-2fy_n-fy_p-y_s(1-f)}{1-f} \\ y_{eN} &=& y_p \\ y_{eQ} &=& \frac{y_e-fy_p}{1-f} \\ y_{eG} &=& y_e \end{aligned}\] For the subsequent calculation, the nonzero partial derivatives of the above are necessary: \[\begin{aligned} \frac{\partial y_u}{\partial y_n} &=& \frac{-f}{1-f}~,~~~ \frac{\partial y_u}{\partial y_p} = \frac{-2f}{1-f}~,~~~ \frac{\partial y_u}{\partial y_e} = \frac{1}{1-f}~,~~~ \nonumber \\ \frac{\partial y_u}{\partial f} &=& \frac{y_u-y_n-2y_p}{1-f} \\ \frac{\partial y_d}{\partial y_n} &=& \frac{-2f}{1-f}~,~~~ \frac{\partial y_d}{\partial y_p} = \frac{-f}{1-f}~,~~~ \frac{\partial y_d}{\partial y_e} = \frac{-1}{1-f}~,~~~ \nonumber \\ \frac{\partial y_d}{\partial y_s} &=&-1~,~~~ \frac{\partial y_d}{\partial f} = \frac{y_d + y_s-2y_n-y_p}{1-f} \\ \frac{\partial y_{eN}}{\partial y_p} &=& 1 \\ \frac{\partial y_{eQ}}{\partial y_p} &=& \frac{-f}{1-f}~,~~~ \frac{\partial y_{eQ}}{\partial y_e} = \frac{1}{1-f}~,~~~ \nonumber \\ \frac{\partial y_{eQ}}{\partial y_e} &=& \frac{y_{eQ}-y_p}{1-f}~,~~~ \\ \frac{\partial y_{eG}}{\partial y_e} &=& 1 \end{aligned}\] The ground state of matter is obtained by minimizing the energy density \(\varepsilon\) with respect to the remaining free variables \[except the baryon density \(n_{\rm B}\) being that we want to retain it as a free variable for the purposes of studying neutron-star matter (NSM)\]:\ (a) The usual condition for neutron strong equilibrium results from minimization with respect to the neutron fraction, \(y_n\). \[\begin{aligned} \frac{\partial \varepsilon}{\partial y_n} &=& f \frac{\partial \varepsilon_n}{\partial y_n} + (1-f)\left(\frac{\partial y_u}{\partial y_n}\frac{\partial \varepsilon_u}{\partial y_u} + \frac{\partial y_d}{\partial y_n}\frac{\partial \varepsilon_d}{\partial y_d}\right) \nonumber \\ &=& fn\mu_n + (1-f)\left(\frac{-f}{1-f}n\mu_u-\frac{2f}{1-f}n\mu_d\right) \nonumber \\ &=& fn(\mu_n-\mu_u-2\mu_d) = 0 \nonumber \\ \Rightarrow ~ \mu_n &=& \mu_u + 2\mu_d \label{muneq} \end{aligned}\] (b) Minimization with respect to the proton fraction \(y_p\) leads to a condition which combines proton strong-and electron EM equilibrium. These two are no longer independent as a result of our having over-specified the system. \[\begin{aligned} \frac{\partial \varepsilon}{\partial y_p} &=& f \left(\frac{\partial \varepsilon_p}{\partial y_p} + g\frac{\partial y_{eN}}{\partial y_p}\frac{\partial \varepsilon_{eN}}{\partial y_{eN}} \right) \nonumber \\ &+& (1-f)\left(\frac{\partial y_u}{\partial y_p}\frac{\partial \varepsilon_u}{\partial y_u} + \frac{\partial y_d}{\partial y_p}\frac{\partial \varepsilon_d}{\partial y_d} + g\frac{\partial y_{eQ}}{\partial y_p}\frac{\partial \varepsilon_{eQ}}{\partial y_{eQ}}\right) \nonumber \\ &=& f(n\mu_p + gn\mu_{eN}) \nonumber \\ &+& (1-f)\left(\frac{-2f}{1-f}n\mu_u-\frac{f}{1-f}n\mu_d-g\frac{f}{1-f}n\mu_{eQ}\right) \nonumber \\ &=& fn(\mu_p + g\mu_{eN}-2\mu_u-\mu_d-g\mu_{eQ}) = 0 \nonumber \\ \Rightarrow ~ \mu_p &=& 2\mu_u + \mu_d-g(\mu_{eN}-\mu_{eQ}) \label{mupeq} \end{aligned}\] By combining Eqs. ([\[muneq\]](#muneq){reference-type="ref" reference="muneq"}) and ([\[mupeq\]](#mupeq){reference-type="ref" reference="mupeq"}), we find \[\begin{aligned} \mu_u &=& 1/3(2\mu_p-\mu_n + 2\Delta_g) \label{muuDelta}\\ \mu_d &=& 1/3(2\mu_n-\mu_p-\Delta_g) \label{mudDelta}\\ \Delta_g &\equiv& g(\mu_{eN}-\mu_{eQ})~. \label{Deltag} \end{aligned}\] (c) A chemical-potential relation corresponding to quark \(\beta\)-equilibrium is obtained by minimizing with respect to the total electron fraction, \(y_e\): \[\begin{aligned} \frac{\partial \varepsilon}{\partial y_e} &=& (1-f)\left(\frac{\partial y_u}{\partial y_e}\frac{\partial \varepsilon_u}{\partial y_u} + \frac{\partial y_d}{\partial y_e}\frac{\partial \varepsilon_d}{\partial y_d} + g\frac{\partial y_{eQ}}{\partial y_e}\frac{\partial \varepsilon_{eQ}}{\partial y_{eQ}}\right) \nonumber \\ &+& (1-g)\frac{\partial y_{eG}}{\partial y_e}\frac{\partial \varepsilon_{eG}}{\partial y_{eG}} \nonumber \\ &=& (1-f)\left(\frac{1}{1-f}n\mu_u-\frac{1}{1-f}n\mu_d-g\frac{1}{1-f}n\mu_{eQ}\right) \nonumber \\ &+& (1-g)(1)n\mu_{eG} \nonumber \\ &=& n[\mu_u-\mu_d + g\mu_{eQ} + (1-g)\mu_{eG}] = 0 \nonumber \\ \Rightarrow ~ \mu_d &=& \mu_u + g\mu_{eQ} + (1-g)\mu_{eG} \.\label{betaQ} \end{aligned}\] This, together with Eq. ([\[mupeq\]](#mupeq){reference-type="ref" reference="mupeq"}) engenders a relation for nucleon \(\beta\)-equilibrium: \[\begin{aligned} \mu_p &=& 2\mu_u + \mu_d-g\mu_{eN} +[\mu_d-\mu_u-(1-g)\mu_{eG}]\nonumber \\ &=& (\mu_u + 2\mu_d)-g\mu_{eN}-(1-g)\mu_{eG}\nonumber \\ \Rightarrow ~ \mu_p &=& \mu_n-g\mu_{eN}-(1-g)\mu_{eG} \label{betaN} \end{aligned}\] (d) Minimization with respect to the strange-quark fraction \(y_s\) gives a condition for quark weak equilibrium: \[\begin{aligned} \frac{\partial \varepsilon}{\partial y_s} &=& (1-f)\left(\frac{\partial y_d}{\partial y_s}\frac{\partial \varepsilon_d}{\partial y_d} + \frac{\partial \varepsilon_s}{\partial y_s}\right) \nonumber \\ &=& (1-f)(-n\mu_d + n\mu_s) = 0 \nonumber \\ \Rightarrow ~ \mu_d &=& \mu_s \label{weakQ} \end{aligned}\] This condition is necessary not only in neutron-star matter but also for supernovae and NS mergers where the relevant dynamical timescales are longer than those of quark flavor-changing processes.\ (e) We get the condition for mechanical equilibrium by minimizing the energy density with respect to \(f\): \[\begin{aligned} \frac{\partial \varepsilon}{\partial f} &=& (\varepsilon_n + \varepsilon_p + g\varepsilon_{eN})-(\varepsilon_u + \varepsilon_d + \varepsilon_s + g\varepsilon_{eQ}) \nonumber \\ &+& (1-f)\left(\frac{\partial y_u}{\partial f}\frac{\partial \varepsilon_u}{\partial y_u} + \frac{\partial y_d}{\partial f}\frac{\partial \varepsilon_d}{\partial y_d} + g\frac{\partial y_{eQ}}{\partial f}\frac{\partial \varepsilon_{eQ}}{\partial y_{eQ}}\right) \nonumber \\ &=& \varepsilon_N + g\varepsilon_{eN}-\varepsilon_Q-g\varepsilon_{eQ} \nonumber \\ &+& (1-f)\left(\frac{y_u-y_n-2y_p}{1-f}n\mu_u \right. \nonumber \\ &+& \left. \frac{y_d + y_s-2y_n-y_p}{1-f}n\mu_d + g\frac{y_{eQ}-y_p}{1-f}n\mu_{eQ}\right) \end{aligned}\] where, in going from the first to the second equality, use of \(\partial \varepsilon/\partial y_i = n \mu_i\) was made, together with the definitions \(\varepsilon_N \equiv \Sigma_{h=n,p} \,\varepsilon_h\) and \(\varepsilon_Q \equiv \Sigma_{q=u,d,s}\,\varepsilon_q\). In the next step, we group chemical potentials according to whether they are multiplied by nucleon or quark particle fractions: \[\begin{aligned} \frac{\partial \varepsilon}{\partial f} &=& \varepsilon_N + g\varepsilon_{eN} + [(-\varepsilon_Q +ny_u\mu_u + ny_d\mu_d + ny_s\mu_d) \nonumber \\ &+& g(-\varepsilon_{eQ} + ny_{eQ}\mu_{eQ})] \nonumber \\ &-& (y_n + 2y_p)n\mu_u-(2y_n + y_p)n\mu_d-gny_p\mu_{eQ} \end{aligned}\] Then, those \(\mu_u\) and \(\mu_d\) that are proportional to \(y_n\) and \(y_p\) are replaced by Eqs. ([\[muuDelta\]](#muuDelta){reference-type="ref" reference="muuDelta"}-[\[mudDelta\]](#mudDelta){reference-type="ref" reference="mudDelta"}). Moreover, the term \(ny_s\mu_d\) becomes \(ny_s\mu_s\) \[using Eq. ([\[weakQ\]](#weakQ){reference-type="ref" reference="weakQ"})\] with the whole parenthesis in which it belongs written as \(P_Q\) (as per the \(T=0\) thermodynamic identity \(P=n\mu-\varepsilon\)): \[\begin{aligned} \frac{\partial \varepsilon}{\partial f} &=& \varepsilon_N + g\varepsilon_{eN} + P_Q + gP_{Qe} \nonumber \\ &-& (y_n + 2y_p)\frac{n}{3}(2\mu_p-\mu_n + 2\Delta_g) \nonumber \\ &-& (2y_n + y_p)\frac{n}{3}(2\mu_n-\mu_p-\Delta_g)-gny_p\mu_{eQ} \end{aligned}\] Subsequently, we expand the products in the second and third lines above and collect similar terms: \[\begin{aligned} \frac{\partial \varepsilon}{\partial f} &=& \varepsilon_N + g\varepsilon_{eN} + P_Q + gP_{Qe}-gny_p\mu_{eQ} \nonumber \\ &-& \frac{n}{3} (2y_n\mu_p-y_n\mu_n + 2\Delta_g y_n + 4y_p-2y_p\mu_n + 4y_p\Delta_g \nonumber \\ &-&2y_n\mu_p + 4y_n\mu_n-2y_n\Delta_g-y_p\mu_p + 2y_p\mu_n-y_p\Delta_g) \nonumber \\ &=& \varepsilon_N + g\varepsilon_{eN} + P_Q + gP_{Qe}-gny_p\mu_{eQ} \nonumber \\ &-& \frac{n}{3}(3y_n\mu_n + 3y_p\mu_p + 3y_p\Delta_g) \end{aligned}\] Finally, we apply Eq. ([\[Deltag\]](#Deltag){reference-type="ref" reference="Deltag"}) to replace \(\Delta_g\) with the electronic chemical potentials \(\mu_{eN}\) and \(\mu_{eQ}\), which leads to an expression involving only the pressures of the various components (using \(P=n\mu-\varepsilon\) where necessary): \[\begin{aligned} \frac{\partial \varepsilon}{\partial f} &=& (\varepsilon_N-ny_n\mu_n-ny_p\mu_p) + P_Q + gP_{Qe} \nonumber \\ &-& ny_p(\Delta_g + g\mu_{eQ}) \nonumber \\ &=&-P_N + P_Q + gP_{Qe} +\varepsilon_{eN}-gny_{eN}\mu_{eN} \nonumber \\ &=&-P_N-gP_{eN} + P_Q + gP_{eQ} = 0 \nonumber \\ &\Rightarrow& P_N + gP_{eN} = P_Q + gP_{eQ} \label{mecheq} \end{aligned}\] (f) For completeness, we also include the result of the minimization with respect to \(g\). However, we will not be implementing this condition because we want \(g\) to remain a free variable (along with \(n_{\rm B}\)) in order to explore the effects of the changing surface tension. \[\begin{aligned} \frac{\partial \varepsilon}{\partial g} &=& f\varepsilon_{eN} + (1-f)\varepsilon_{eQ}-\varepsilon_{eG} = 0 \nonumber \\ \Rightarrow ~ \varepsilon_{eG} &=& f\varepsilon_{eN} + (1-f)\varepsilon_{eQ} \label{surface} \end{aligned}\] In the present approach, \(g=0\) amounts to a Gibbs construction (GCN) and \(g=1\) to a Maxwell construction (LCN). *It has the added benefit of maintaining control over the various particle fractions in the Maxwell mixed phase which has not been the case in previous literature.* Clearly, first-order transitions of intermediate surface tension will have \(0 < g < 1\). Extension to finite temperature is accomplished by minimizing the free energy density instead of the energy density. The conservation laws remain the same as do the formal expressions describing the phase-equilibrium conditions albeit with the use of the corresponding finite-\(T\) pressures and chemical potentials.Applications to supernovae and neutron star mergers require \((n_{\rm B},y_e,T)\) as independent variables; that is, one must also skip minimization with respect to \(y_e\).\ **Crossovers:** These can also be studied in this context. One sets \(g=0\) [^1] and eliminates the mechanical equilibrium condition \[Eq. ([\[mecheq\]](#mecheq){reference-type="ref" reference="mecheq"})\] in favor of an explicit functional form for \(f\) which approaches asymptotically 0 and 1 at high and low densities, respectively, e.g. \(f=1-\exp[-a~(n/n_{\rm sat})^{-b}]\) where \(a\) and \(b\) are fit parameters and \(n_{\rm sat}\) the saturation density of symmetric nuclear matter. The hadron-to-baryon fraction \(f\) can also depend on composition (prior to equilibration) with the added algebraic burden of terms proportional to \(\partial f/\partial y_i\) in the equilibrium equations. ## \(NQe\mu\) matter The inclusion of muons in the calculation comes at the cost of four additional variables (\(y_{\mu N}\), \(y_{\mu Q}\), \(y_{\mu G}\), \(y_{\mu}\)), a muon-number conservation equation which mimics Eq. ([\[electronnum\]](#electronnum){reference-type="ref" reference="electronnum"}) for electrons, and modifications to the total energy density of the system and the charge neutrality equations (baryon number and electron number equations are unaffected): \[\begin{aligned} \varepsilon &=& f[\varepsilon_n + \varepsilon_p + g(\varepsilon_{eN}+\varepsilon_{\mu N})] \nonumber \\ &+& (1-f)[\varepsilon_u + \varepsilon_d + \varepsilon_s + g(\varepsilon_{eQ}+\varepsilon_{\mu Q})] \nonumber \\ &+&(1-g)(\varepsilon_{eG}+\varepsilon_{\mu G}) \label{edenmixmu}\\ 0 &=& (y_p-y_{eN}-y_{\mu N})g \label{lcnHmu} \\ 0 &=& [(2y_u-y_d-y_s)/3-y_{eQ}-y_{\mu Q}]g \label{lcnQmu} \\ 0 &=& [fy_p + (1-f)(2y_u-y_d-y_s)/3 \nonumber \\ &-& y_{eG}-y_{\mu G}](1-g) \label{gcnmu}\\ 0 &=& y_{\mu}-fgy_{\mu N}-(1-f)gy_{\mu Q}-(1-g)y_{\mu G} \label{muonnum} \end{aligned}\] The minimization procedure yields modifications to the mechanical equilibrium and surface-tension optimization conditions \[Eqs. ([\[mecheq\]](#mecheq){reference-type="ref" reference="mecheq"})-([\[surface\]](#surface){reference-type="ref" reference="surface"})\] such that muonic contributions are accounted, while the chemical potential relations \[Eqs. ([\[muneq\]](#muneq){reference-type="ref" reference="muneq"}), ([\[mupeq\]](#mupeq){reference-type="ref" reference="mupeq"}), ([\[betaQ\]](#betaQ){reference-type="ref" reference="betaQ"}) or ([\[betaN\]](#betaN){reference-type="ref" reference="betaN"}), ([\[weakQ\]](#weakQ){reference-type="ref" reference="weakQ"})\] remain unchanged. Moreover, three new constraints are generated corresponding to lepton weak equilibrium in each of the three regions: \[\begin{aligned} &&P_N + g(P_{eN} + P_{\mu N}) = P_Q + g(P_{eQ} + P_{\mu Q}) \\ &&\varepsilon_{eG} + \varepsilon_{\mu G} = f(\varepsilon_{eN} + \varepsilon_{\mu N}) + (1-f)(\varepsilon_{eQ} + \varepsilon_{\mu Q}) \nonumber \\ \\ &&\mu_{eN} = \mu_{\mu N}~;~~\mu_{eQ} = \mu_{\mu Q}~;~~\mu_{eG} = \mu_{\mu G} \end{aligned}\] # Equation of state To demonstrate the workings of the scheme devised above, we describe the EOSs employed for nucleons, quarks, and leptons below. Selected properties of NSs such as their mass-radius curves, equilibrium and adiabatic squared speeds of sound are calculated results of which are shown and discussed. The outer crust EOS described by a uniform background of relativistic degenerate electrons in an ionic lattice is relatively well-understood. Here we use the SLy4 crust EOS for \(n_{\rm B}<0.05\) fm\(^{-3}\). As our focus is on the core \(g\)-modes, the composition information of the crust is ignored in calculating the equilibrium and adiabatic sound speeds (that is, the two speeds are set equal to each other). ## Nucleons For the description of nucleons, we use the ZLA parametrization  of the Zhao-Lattimer (ZL)  EOS. This is consistent with laboratory data at nuclear saturation density \(n_{\rm sat}\simeq 0.16~{\rm fm}^{-3}\), the chiral effective field theory calculations of Ref. , and constraints obtained by Legred et al. , which combined available observations including the radio pulsar mass measurements of PSR J0348+0432 and J0470+6620 , the mass and tidal deformability measurements of GW170817 and GW190425 , and the x-ray mass and radius constraints from latest *NICER* measurements of J0030+0451 and J0470+6620 . The total energy density of nucleons with a common mass \(m_N=939.5\) MeV is given by the density functional \[\begin{aligned} \varepsilon_N &=& \varepsilon_N(n_{\rm B},y_n,y_p) \nonumber \\ &=& \frac{1}{8\pi^2\hbar^3}\sum_{h=n,p}\left\{k_{Fh}(k_{Fh}^2 + m_N^2)^{1/2}(2k_{Fh}^2 + m_N^2)\right. \nonumber \\ &-& \left. m_N^4\ln\left[\frac{k_{Fh}+(k_{Fh}^2 + m_N^2)^{1/2}}{m_N}\right]\right\} \nonumber \\ &+& 4 n_{\rm B}^2 y_n y_p \left\{\frac{a_0}{n_{\rm sat}} +\frac{b_0}{n_{\rm sat}^{\gamma}} [n_{\rm B}(y_n + y_p)]^{\gamma-1}\right\} \nonumber \\ &+& n_{\rm B}^2 (y_n-y_p)^2\left\{\frac{a_1}{n_{\rm sat}} + \frac{b_1}{n_{\rm sat}^{\gamma_1}}[n_{\rm B}(y_n + y_p)]^{\gamma_1-1}\right\} \nonumber \label{edenZL} \,, \end{aligned}\] where \(k_{Fh} = (3\pi^2\hbar^3n_{\rm B} y_h)^{1/3}\) is the Fermi momentum of nucleon species \(h\). The chemical potentials and the pressure are obtained from Eq. ([\[edenZL\]](#edenZL){reference-type="ref" reference="edenZL"}) according to \[\begin{aligned} \mu_h &=& \frac{\partial (\varepsilon_N/n_{\rm B})}{\partial y_h} ~~;~~ h=n,p\\ P_N &=& n_{\rm B}\sum_{h=n,p}\mu_h y_h-\varepsilon_N ~. \end{aligned}\] ::: [\[tab:Parameters\]]{#tab:Parameters label="tab:Parameters"} ## Quarks For the calculation of the quark EOS, we use the vMIT bag model . The total energy density of quarks in this context is \[\begin{aligned} \varepsilon_Q &=& \varepsilon_Q(n_{\rm B},y_u,y_d,y_s) \nonumber \\ &=& \sum_{q=u,d,s}\varepsilon_q + \frac{1}{2}a~\hbar~[n_{\rm B}(y_u+y_d+y_s)]^2 + \frac{B}{\hbar^3} \nonumber \\ \\ \varepsilon_q &=& \frac{3}{8\pi^2\hbar^3}\left\{k_{Fq} (k_{Fq}^2 + m_q^2)^{1/2}(2 k_{Fq}^2 + m_q^2) \right. \nonumber \\ &-& \left. m_q^4\ln\left[\frac{k_{Fq}+(k_{Fq}^2 + m_q^2)^{1/2}}{m_q}\right]\right\} ~, \end{aligned}\] where \(k_{Fq} = (\pi^2\hbar^3n_{\rm B} y_q)^{1/3}\) is the Fermi momentum of quark species \(q\). Similarly to the nucleonic case, the chemical potentials and pressure can be derived from the thermodynamic identities \[\begin{aligned} \mu_q &=& \frac{\partial (\varepsilon_Q/n_{\rm B})}{\partial y_q} ~~;~~ q=u,d,s\\ P_Q &=& n_{\rm B}\sum_{q=u,d,s}\mu_q y_q-\varepsilon_Q ~. \end{aligned}\] ## Leptons Leptons are treated as non-interacting, relativistic particles for which \[\begin{aligned} \varepsilon_L &=&\frac{1}{8\pi^2\hbar^3}\sum_l\left\{k_{Fl}(k_{Fl}^2+m_l^2)^{1/2}(2k_{Fl}^2 + m_l^2)\right. \nonumber \\ &-& \left. m_l^4\ln\left[\frac{k_{Fl}+(k_{Fl}^2 + m_l^2)^{1/2}}{m_l}\right]\right\} \\ \mu_l &=& (k_{Fl}^2 + m_l^2)^{1/2} \\ P_L &=& n_{\rm B} \sum_l y_l \mu_l-\varepsilon_L \\ k_{Fl} &=& (3\pi^2\hbar^3 n_{\rm B} y_l)^{1/3} ~~;~~ l=e,\mu~. \end{aligned}\] At low baryon densities only electrons are present in the system. The muon onset density is such that \(\mu_e-m_{\mu} = 0\). Depending on the parametrization choice, this condition also gives the density at which muons vanish. ## Sound speeds in the pure and mixed phases We begin with pure-phase thermodynamic quantities written as functions of the total baryon density \(n_{\rm B}\), and the individual particle fractions \(y_n\), \(y_p\), \(y_{eN}\), \(y_u\), \(y_d\), \(y_s\), \(y_{eQ}\), \(y_{eG}\): \[\begin{aligned} \varepsilon_N &=& \varepsilon_N(n_{\rm B},y_n,y_p) ~;~ P_N = P_N(n_{\rm B},y_n,y_p) ~; \nonumber \\ \mu_h &=& \mu_h(n_{\rm B},y_n,y_p) \label{pureN}\\ \nonumber \\ \varepsilon_Q &=& \varepsilon_Q(n_{\rm B},y_u,y_d,y_s) ~;~ P_Q = P_Q(n_{\rm B},y_u,y_d,y_s) ~; \nonumber \\ \mu_q &=& \mu_q(n_{\rm B},y_q) ~;~ q = u, d, s \label{pureQ} \\ \nonumber \\ \varepsilon_{eX} &=& \varepsilon_{eX}(n_{\rm B},y_{eX}) ~;~ P_{eX} = P_{eX}(n_{\rm B},y_{eX}) ~; \nonumber \\ \mu_{eX} &=& \mu_{eX}(n_{\rm B},y_{eX}) ~;~ X = N,Q,G\. \end{aligned}\] In terms of these, we express the thermodynamics of the mixed \((^*)\) phase as \[\begin{aligned} \varepsilon^* &=& f \varepsilon_N + (1-f) \varepsilon_Q \nonumber \\ &+& fg\varepsilon_{eN} + (1-f)g\varepsilon_{eQ} + (1-g)\varepsilon_{eG} \\ P^* &=& f P_N + (1-f) P_Q \nonumber \\ &+& fg P_{eN} + (1-f)g P_{eQ} + (1-g)P_{eG} \\ \mu_h^* &=& \mu_h ~~;~~ \mu_q^* = \mu_q \\ y_h^* &=& f y_h ~~;~~ y_q^* = (1-f) y_q \. \label{qnoG} \end{aligned}\] For NSM (denoted by the subscript \(\beta\)), the various conservation laws \[Eqs. ([\[baryonnum\]](#baryonnum){reference-type="ref" reference="baryonnum"})-([\[electronnum\]](#electronnum){reference-type="ref" reference="electronnum"})\] and conditions for phase equilibrium \[Eqs. ([\[muneq\]](#muneq){reference-type="ref" reference="muneq"}), ([\[mupeq\]](#mupeq){reference-type="ref" reference="mupeq"}), ([\[betaQ\]](#betaQ){reference-type="ref" reference="betaQ"}), ([\[weakQ\]](#weakQ){reference-type="ref" reference="weakQ"}), ([\[mecheq\]](#mecheq){reference-type="ref" reference="mecheq"})\] must be applied. The solution of these equations converts the \(y_i\) and \(f\) from independent variables to functions of \(n_{\rm B}\) and \(g\). Thus, the state variables also become functions of \(n_{\rm B}\) and \(g\) according to the rule \[\begin{aligned} Q(n_{\rm B},y_i,y_j...,g) &\rightarrow& Q_{\beta}[n_{\rm B},y_i(n_{\rm B},g),y_j(n_{\rm B},g)...,g] \nonumber \\ &=& Q_{\beta}(n_{\rm B},g)~. \nonumber \end{aligned}\] Note that the upper-and lower-density boundaries of the mixed phase correspond to \(f_{\beta}(n_{\rm B},g) = 0\) and 1, and depend on \(g\). The adiabatic speed of sound in the mixed phase is obtained by first calculating the expression \[c_{\rm ad}^2(n_{\rm B},y_i,f,g) = \left.\frac{\partial P^*}{\partial n_{\rm B}}\right|_{y_i,f,g} \left(\left.\frac{\partial \varepsilon^*}{\partial n_{\rm B}}\right|_{y_i,f,g}\right)^{-1}\] and then evaluating it for NSM \[c_{\rm{ad},\beta}^2(n_{\rm B},g) = c_{\rm ad}^2[n_{\rm B},y_{i,\beta}(n_{\rm B},g),f_{\beta}(n_{\rm B},g),g]~.\] On the other hand, the equilibrium sound speed is given by the total derivatives of the pressure and the energy density with respect to the baryon density *after* the enforcement of NSM-equilibrium, \[c_{\rm eq}^2 = \frac{dP^*_{\beta}}{dn_{\rm B}}\left(\frac{d\varepsilon^*_{\beta}}{dn_{\rm B}}\right)^{-1}~.\] # Nonradial neutron star oscillations Neutron stars are expected to oscillate in many modes corresponding to different restoring forces. Pressure-supported modes including \(f\)-(fundamental) and \(p\)-(pressure) modes are sensitive to stellar structure. The \(f\)-mode frequency approximately scales with the mean density and is universally correlated with the tidal deformability and the moment of inertia . \(p\)-mode oscillations are more confined towards the surface of the NS and are thus sensitive to the EOS at lower density . Both \(f\)-and \(p\)-modes are sensitive to the bulk pressure and not sensitive to detailed chemical composition. We have verified that the novel construction of first-order phase transitions in this work does not play a significant role due to the universal relation between the oscillation frequencies and other NS observables.\ In this paper, we study the \(g\)-mode, the fluid mode with gravity as restoring force. The \(g\)-mode oscillation acquires non-zero frequency because there is a gradient of chemical composition or a first-order phase transition between the two phases. A universal relation between the chemical \(g\)-mode frequency and lepton fraction was discovered recently  providing key information about the nuclear symmetry energy at high density. A \(g\)-mode due to a density discontinuity from a phase transition can be understood as a special version of a \(g\)-mode due to chemical composition changes, since matter on the low-density side can be treated as having a different composition from that on the high-density side. This situation occurs when matter does not instantaneously change phase upon passing through the phase transition boundary . The discontinuity \(g\)-mode is most sensitive to the local gravity and the density discontinuity at phase transition .\ At high temperature relevant to neutron star mergers, the compositional \(g\)-mode can be suppressed . However, another branch of \(g\)-mode can also have non-zero frequency when adiabatic compression of the NS matter is not in thermal equilibrium with the matter in hydrodynamic equilibrium . These are very low-frequency modes because thermal pressure is negligible in the cores of neutron stars when temperature \(T\lesssim 10^7\) K. For \(T\gtrsim 10^{10}\) K, the thermal \(g\)-mode has comparable frequency to the composition \(g\)-mode. Recent core-collapse supernova simulations suggest that the thermal \(g\)-mode could dominate when there is a large entropy gradient .\ In this work, we consider only the zero-temperature EOS for hybrid NSs with the novel framework of a first-order transition. We focus on the lowest order non-radial \(g\)-mode oscillation (\(\ell=2\)) arising from a gradient in the chemical composition. This oscillation mode couples directly to gravitational waves and has a frequency of a few hundred Hz for NSs which lies in the band of gravitational wave observations . Assuming the chemical composition does not change in a period of oscillation, the local \(g\)-mode frequency \(\nu_g\) is determined by the Brunt-Väisälä frequency, \[\begin{aligned} \nu_g^2&=& g^2\left(\frac{1}{c_{eq}^2}-\frac{1}{c_{ad}^2}\right) e^{\nu-\lambda}\,, \label{eq:BV_frequency} \end{aligned}\] where \(\nu\) and \(\lambda\) are the temporal and radial metric functions. The Brunt-Väisälä frequency depends on density and chemical composition which vary across the NS. Figure [\[fig:delta\]](#fig:delta){reference-type="ref" reference="fig:delta"} shows the difference between the inverse squared sound speeds, and the bracket on the right-hand side of Eq. ([\[eq:BV_frequency\]](#eq:BV_frequency){reference-type="ref" reference="eq:BV_frequency"}). With the correct boundary condition and perturbation fluid equations, one can find global oscillation modes, known as \(g\)-modes driven by local buoyancy oscillations. Detailed methods to calculate the \(g\)-modes with and without the Cowling approximation can be found in our previous work . # Results {#sec:result} In this section we demonstrate the effect of changing the local-to-total lepton ratio on the EOS and its composition, associated structural and tidal properties of NSs, the two sound speeds, and the resulting \(g\)-mode frequencies. We also show plots pertaining to the EOS and particle fractions of a crossover application. All results refer to neutron star (\(\beta\)-equilibrated) matter. ## Equation of State The change in the nucleonic content of the mixed phase is shown in Fig. [\[fig:fbeta\]](#fig:fbeta){reference-type="ref" reference="fig:fbeta"} for five different implementations of charge neutrality. The decrease in \(f_{\beta}\) is steeper as the Maxwell limit (of high surface tension and thus LCN) is approached; that is, the mixed phase becomes narrower in terms of density. This indicates that first-order transitions with sharper phase separation (Maxwell-like, "stiff\") undergo a faster compositional change that can impact the \(g\)-mode frequency more severely than transitions where extensive phase mixing occurs (Gibbs-like, "soft\"). The approximately common intersection point of the various curves occurs at, roughly, \(f_{\beta}=2/3\) near the density \(n_t\) at which the energy densities of the pure phases are equal. We note that, for the models and parametrization used herein, the boundaries of the Gibbs and the Maxwell mixed phases are (0.34, 1.63) fm\(^{-3}\) and (0.75, 0.88) fm\(^{-3}\), respectively. Figure [\[fig:eos\]](#fig:eos){reference-type="ref" reference="fig:eos"} is a representation of the effect of varying \(g\) on the EOS in the pressure vs. energy-density plane for \(NQe\mu\) matter. It follows the trends already seen in Fig. [\[fig:fbeta\]](#fig:fbeta){reference-type="ref" reference="fig:fbeta"} of more Maxwell-like behavior with increasing \(g\) and a correspondingly smaller mixed phase. Our EOS includes a wide variety of possibilities between the Maxwell and Gibbs constructions, some of which are very similar to EOSs of the quark-hadron phase calculated using the Wigner-Seitz approximation (WSA), see e.g. . Thus, we may interpret the present framework as one which recasts the complicated Coulomb and surface problem of WSA into an easier form involving only lepton phase space, with local leptons increasing the energy of the system mimicking the effect of surface energy.\ [^1]: Unlike first-order transitions where two distinct phases are in contact, crossovers involve only a single phase whose ground state properties change drastically as some parameter of the system is changed. Therefore, in the present context, electrons will always encounter a mixture of quarks and hadrons regardless of their configuration-space coordinates, and, correspondingly, charge neutrality is achieved globally, i.e. \(g=0\).

{'timestamp': '2023-02-10T02:00:23', 'yymm': '2302', 'arxiv_id': '2302.04289', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04289'}

# INTRODUCTION Video monitoring provides a non-contact method for monitoring neonates in clinical, home and remote environments. Typically, the detection of the patient's facial area as a region of interest (ROI) with associated facial landmarks is an essential first step. After this step, ROI tracking between video frames and subsequent analysis of the tracked ROI would then be performed. Video monitoring using face ROIs has received increasing interest for infant cardio-respiratory monitoring recently . Cardio-respiratory monitoring is crucial for early detection, prediction of prognosis, and continued monitoring of major complications during the neonatal period . It assists clinicians in providing timely and appropriate care to possibly minimise morbidity and mortality . Similarly, video face monitoring has been investigated for the timely detection of neonates in pain . This monitoring allows for treatment to prevent long-term effects on neurological and behavioural development . Additionally, facial video monitoring for the detection of jaundice  and sleep-wake classification  has been investigated in newborns. Face and facial landmark detection for neonates, especially in clinical environments, has proven difficult . Neonates have less well-defined facial features compared to adults , poses/postures are more varied, and common obstructions, such as bottle feeding and pacifiers are prevalent . Additionally, in the clinical environment, background scenes are complicated, medical devices and interventions may be involved, and varying lighting conditions pose additional challenges . Additionally, many face and facial landmark detectors have been trained on predominantly adult data . Overall, this has led to poor performance in neonatal face ROI detection, which results in manual ROI selection of the neonatal face being common . This paper presents an overview of several existing adult and neonatal face/landmark detectors in Section [2](#sec:existing){reference-type="ref" reference="sec:existing"}. Improved deep learning models using transfer learning, image re-orientation and fusion techniques are proposed in Section [3](#sec:methods){reference-type="ref" reference="sec:methods"}. Existing and proposed methods are then evaluated and discussed in Sections [4](#sec:results){reference-type="ref" reference="sec:results"} and [5](#sec:discussion){reference-type="ref" reference="sec:discussion"}. # EXISTING WORKS {#sec:existing} ## Adult Face and Facial Landmark Detection {#sec:adult} Current state-of-the-art methods for face and facial landmark detection have utilised machine learning, especially deep learning. A traditional method for adult face detection is OpenCV's Haar cascade method . In this method, Haar features are extracted and an AdaBoost algorithm is used to detect the face . For deep learning methods, Dlib face detection uses a histogram of oriented gradients and a convolutional neural network (CNN) based on max-margin for face detection . With multi-task cascaded CNN (MTCNN) , the correlation between the detection and alignment of the face and associated landmarks was utilised to boost performance. MTCNN was broken up into three stages to predict the face and five landmarks (*i.e.*, right eye, left eye, nose, right mouth, and left mouth) on a course to a fine level, and trained on the WIDER FACE dataset . More recently, single-shot multibox detectors (SSD) have been popular for face detection . SSD  utilises a set of default bounding boxes over different aspect ratios and scales per feature map location. During face detection, each default box generates an output of if a face is present, and appropriate adjustments to the box to match the face . RetinaFace  is a novel SSD that unified face bounding box prediction and 2D facial landmark detection. The model was trained on the WIDER FACE dataset , and outputs five facial landmarks (*i.e.* right and left eye, nose, right mouth, and left mouth). BlazeFace  is a lightweight face detector tailored for mobile GPU inference with models for front-facing and rear-facing cameras. The feature extraction network is inspired by MobileNetV1/V2, and the anchor scheme is modified from SSD . MediaPipe provides an updated implementation of the BlazeFace architecture and outputs face detection and 6 facial landmarks (*i.e.* right and left eyes, nose, centre mouth, and left and right ears) . ## Neonatal Face and Facial Landmarks Detection {#sec:neonatal} Similarly to adult-based existing work, the majority of past works in the neonatal space have focused on deep learning-based methods. In particular, the YOLO framework has been especially popular. For face detection, Awais *et al.*  proposed a semi-automated intensity-based method. RGB colour space was converted into the CIELAB colour space. Then, an appropriate threshold is then determined by examining intensity values within each CIELAB channels . Finally, the facial region is determined as the region with the highest number of linked portions . Olmi *et al.*  developed a fully automated face detection method for newborns in the neonatal intensive care unit (NICU) from the aggregate channel feature algorithm. The model was trained on newborn data collected at the Neuro-physiopathology and Neonatology Clinical Units of AOU Careggi, Firenze, Italy . For deep learning-based face detectors, Khanam *et al.*  utilised YOLOv3 model , pre-trained on the MS COCO dataset . Transfer learning was then applied using 473 images of neonates obtained from the internet . Kyrollos *et al.*  constructed a deep learning model with a ResNet backbone and RetinaNet head , pre-trained on the ImageNet dataset . This model was further trained on patients admitted to the NICU of the Children's Hospital of Eastern Ontario . Building on , Dosso *et al.*  developed NICUface. NICUface is designed specifically for neonatal face detection in complex NICU environments. Two base models, YOLOv5Face  and RetinaFace  were explored for NICUface development, as these models performed the best in initial testing. These models were then subsequently trained on publicly available datasets; the newborn baby heart rate estimation database (NBHR)  and iCOPEvid dataset , as well as data previously collected from the Children's Hospital of Eastern Ontario . Overall, NICUface with YOLOv5Face base performed the best compared to existing base models and NICUface with RetinaFace base. NICUface is also able to output 5 facial landmarks (*i.e.* right, and left eyes, nose, right mouth and left mouth); however, the updated model was not re-trained specifically for landmark detection and so only qualitative assessment was performed. Similarly, Hausmann *et al.*  developed a face detector based on YOLOv5 model . The model was trained on images taken from Tampa General Hospital with Male 47.3%: Female 52.7% sex breakdown, and an ethnicity breakdown of 72.7% Caucasian, 20% African-American, and 7.3% Asian. Overall, the model in  did significantly better than YOLO model  baselines. For facial landmarks specifically, Wan *et al.*  developed a model from HRNet's facial landmark detection . The model requires the face detection bounding box as its input, and then outputs 68 facial landmarks relating to eyes, eyebrows, nose, mouth and the outline of the face . This model  was further trained on infant images obtained from the internet. # METHODS {#sec:methods} ## Data Acquisition ### Datasets Three Publicly available datasets , , and are used for this study. NBHR  consists of 1,130 videos of 257 neonates at 0-6 days old, totalling 9.6 hours of recordings with synchronous photoplethysmogram and heart rate physiological signals. The newborn infants were recruited from the Department of Obstetrics and Gynaecology, Xinzhou People's Hospital, China. Biological sex was approximately equal (Male 48.6%: Female 51.4%). Camera angle relative to each baby's location varied, facial occlusion was present in a subset of videos from bottle feeding and sleep position, and a variety of natural hospital room illuminations were obtained . Infant Annotated Faces (InfAnFace) , contains 410 images of infants obtained from Google, YouTube and advertisements. These images were annotated with 68 facial landmarks and pose attributes . iCOPEvid dataset  contains 234 videos (20 seconds/video) of 49 neonates experiencing a set of noxious stimuli, a period of rest and an acute pain stimulus. Neonatal videos were taken in the neonatal nursery at Mercy Hospital in Springfield Missouri, USA, from infants aged between 34 to 70 hours . Biological sex was approximately balanced (Male 53.1%: Female 46.9%), and 16 neonates were swaddled compared to 33 free to move . Ethnicity was distributed as 41 Caucasian, 1 African-American, 1 Korean, 2 Hispanic, and 4 interracial . ### Data extraction For NBHR  and iCOPEvid  datasets, the first frame was extracted from each video to be annotated. Additionally, neonatal images of the same subject that were similar between videos were removed to provide a diverse dataset. From the InfAnFace dataset , images were excluded if not representative of the hospital environment or neonatal age period. Overall, 455 images from 324 neonates were obtained for annotation. These images consisted of 352 images from 257 neonates, 18 images from 18 infants, and 85 images from 49 neonates from the NBHR , InfAnFace , and iCOPEvid  datasets, respectively. ### Annotations We manually annotated the provided dataset. The face bounding box captured the area from the forehead to the chin and between the ears. In cases of partial occlusions, if subsections of the face were still visible/partially visible in the area of occlusion, they were annotated as the face. Otherwise, if a complete region of the face was occluded, it was not included in the bounding box, as the size of the face occluded was difficult to infer. Additionally, 6 facial landmarks were identified, namely; right eye, left eye, nose, right mouth, centre mouth and left mouth. ## Transfer Learning {#sec:transfer} ### Base Models Based on the performance of face detection reported in table [\[tab1:face_detect\]](#tab1:face_detect){reference-type="ref" reference="tab1:face_detect"}, YOLOv5 and YOLOv7Face  were chosen as base models. YOLOv5 starting weights were obtained from . YOLOv7Face starting weights were the 'yolov7' model on GitHub <https://github.com/derronqi/yolov7-face>, which was pre-trained on the WIDER FACE dataset . ### Data Augmentation During each training epoch, random horizontal flip (50% probability), photo-metric colour distortion (*i.e.* \(\pm 1\%\), \(\pm 10\%\), \(\pm 10\%\) HSV Hue, Saturation and Value), translation (*i.e.* \(\pm 10\%\)), and scaling (*i.e.* \(\pm 10\%\)) augmentation were performed. ### Hyperparameters For YOLOv5, hyperparameters were set similarly as in . An initial learning rate of 0.0032, final learning rate of 0.12, warm-up momentum of 0.5, final momentum of 0.843, warm-up period of 2 epochs, intersection over union (IoU) threshold of 20%, and optimised using stochastic gradient descent were used during training. For YOLOv7Face, hyperparameters were set similarly to the sample settings in YOLOv7Face . Learning rate of 0.01, warm-up momentum of 0.8, final momentum of 0.937, warm-up period of 3 epochs, IoU threshold of 20%, and optimised using stochastic gradient descent were used during training. ### Training The data was split into approximately 80% training and 20% test, resulting in 366 images from 258 subjects for training, and 89 images from 66 subjects for test evaluation. It was ensured that all images from the same subject were in only the training or the test sets. Within the training set, data was further split for 5-fold cross-validation, while ensuring all subjects' images were within the same fold. The number of epochs to train the model and the number of layers to freeze within the models were then explored. Based on initial cross-validation results, the final YOLOv5 model was trained over 100 epochs with no layers frozen, and the final YOLOv7Face model was trained over 50 epochs with no layers frozen. Both models were trained with a batch size of 16. ## Orientation {#sec:orientation} Existing detectors have been typically trained on images where the face is upright. Since neonatal poses are variable, rotating the image into four different scenarios could improve the likelihood of a more optimal pose for the model. Hence, utilising the same face and facial landmark detector, four images were inputted into the detector. These four images were the 0, 90, 180, and 270 degrees rotated versions of the original image. The image that outputted the highest confidence detection was then used for face and landmark estimation. ## Fusion {#sec:fusion} Based on the results presented in Tables [\[tab1:face_detect\]](#tab1:face_detect){reference-type="ref" reference="tab1:face_detect"} and  [\[tab1:landmark_detect\]](#tab1:landmark_detect){reference-type="ref" reference="tab1:landmark_detect"}, two fusion methods were considered. Firstly, as RetinaNet with re-orientation and the proposed YOLOv7Face produced the lowest mean normalised error (MNE) for facial landmarks, a fusion model based on majority voting was considered. Whereby, the method that has the highest confidence (*i.e.* max confidence in four RetinaFace outputs compared to the proposed YOLOv7Face), was used for facial landmarks estimation. With this fusion model, the average of execution time is approximately 2.09 s per image. To reduce the computational time, MediaPipe front camera model was used to estimate the optimal orientation of the input image for RetinaFace. In this fusion model, the average execution time is reduced to approximately 1.58 s per image. ## Evaluation For face and facial landmark detection, as there is only one neonate present in each image, the confidence threshold was set to 5% to maximise the number of detections, with the highest confidence detection chosen for evaluation. If no face is detected, this was defined as a false negative. ### Face Detection {#sec:detect} Average precision (AP) is defined by [\[eq:precision\]](#eq:precision){reference-type="eqref" reference="eq:precision"}, with a particular IoU threshold. IoU is defined as the area of overlap divided by the area of union between the reference and estimated bounding box. IoU threshold represents when to classify the face detection as a true positive (TP) or a false positive (FP). Two AP metrics are considered, average precision for IoU threshold of 50% (AP50), and mean average precision for IoU thresholds 50% to 95% in 5% increments (mAP). \[\label{eq:precision} \begin{split} Precision = \frac{TP}{TP+FP} \end{split}\] ### Facial Landmark Detection {#sec:landmark} MNE was used for the assessment of landmark detection between reference (ref) and estimated (est) landmarks for all test samples (N). Where the normalisation distance was the reference face bounding box area (ref_bbox) [\[eq:norm_error\]](#eq:norm_error){reference-type="eqref" reference="eq:norm_error"},[\[eq:mne\]](#eq:mne){reference-type="eqref" reference="eq:mne"}. \[\label{eq:norm_error} \begin{split} norm\_error(i) & = \frac{\sqrt{(ref^{i}_x-est^{i}_x)^2+(ref^{i}_y-est^{i}_y)^2}}{\sqrt{ref\_bbox^{i}_{width} \times ref\_bbox^{i}_{height}}}\\ \end{split}\] \[\label{eq:mne} \begin{split} MNE &= \frac{\sum_{i=1}^{N} norm\_error(i)}{N}\\ \end{split}\] [\[tab1:face_detect\]]{#tab1:face_detect label="tab1:face_detect"} [\[tab1:landmark_detect\]]{#tab1:landmark_detect label="tab1:landmark_detect"} ::: ### Execution Time {#sec:time} We evaluated the average execution time per image during the inference/test phase as another assessment metric. A faster method is more suitable for real-time processing and mobile-phone implementation. The average execution time per image was calculated using Python on a MacBook Pro CPU 2.3 GHz 8-Core Intel i9. ### Significance Testing Paired Wilcoxon signed-rank test was used to determine if IoU and normalised errors were significantly better, with p-value threshold of 0.05 used. # RESULTS {#sec:results} Table [\[tab1:face_detect\]](#tab1:face_detect){reference-type="ref" reference="tab1:face_detect"} presents the face detection results. Within the existing methods, OpenCV, SSD, Dlib and MTCNN in particular struggled to detect the neonate's face with more than three-quarters of faces not detected. Out of the existing methods, YOLOv7Face performed significantly the best with AP50 of 100% and mAP of 64.6%. Both proposed methods showed significant improvements compared to existing methods, with the re-trained YOLOv7Face model performing the best with AP50 of 100% and mAP of 84.7%. Re-orienting the image for the highest confidence showed improved results in all methods with the exception of the proposed YOLOv7Face model. In particular, existing methods RetinaFace, YOLOv5, YOLOv7Face, and NICUface showed significant improvements. Even with re-orientation, the proposed models still performed significantly better, with the exception of the proposed YOLOv5 model and the re-orientated YOLOv5 model compared NICUface, where the p-values were 0.3 and 0.1, respectively. Table [\[tab1:landmark_detect\]](#tab1:landmark_detect){reference-type="ref" reference="tab1:landmark_detect"} presents the facial landmark detection results. Similarly with face detection, re-orientation significantly improved the results for existing methods, whereas the proposed YOLOv7Face model performed worse. Compared with all existing and re-oriented existing methods, the proposed YOLOv7Face model shows significant improvement in overall MNE. The proposed YOLOv7Face model produced MNE values of 0.072, 0.077, 0.062, and 0.072 for all, eyes, nose, and mouth landmarks respectively. Compared to using just the proposed YOLOv7Face model, fusion model 1 (RetinaFace + YOLOv7Face\*) produced insignificant changes. Whilst this fusion model produced on average lower errors, 25 landmark estimates were worse compared to 20 that improved. Whereas fusion model 2 (MediaPipe Front + RetinaFace + YOLOv7Face\*) produced overall significant but minor improvements, with MNE of 0.071, 0.078, 0.062, and 0.069 for all, eyes, nose, and mouth landmarks respectively. # DISCUSSION {#sec:discussion} Overall, the proposed models showed improvement in both face and facial landmark detection. One note of caution regarding the results is that the NICUface model was used unaltered from. However, this model was trained on the same iCOPEvid and NBHR datasets. Hence, the results presented for NICUface may be an overstatement. One limitation of these models is the average execution time of 211 ms and 1,364 ms for YOLOv5 and YOLOv7Face, respectively. With execution times being relatively long compared to other methods, this may hinder real-time processing. However, face detection is typically only required in the first step prior to ROI tracking which is computationally less expensive. Therefore real-time processing may still be possible in the overall implementation of a neonatal video monitoring system with these models. Future work looking into the utilisation of BlazeFace for real-time mobile phone processing may be useful. The re-orientation method proved to be a highly effective process for existing methods to both increase the success rate of detecting the face and facial landmarks, as well as improve their accuracy. This highlights a simple process to address the various poses/postures that the neonate may be in. However, this would require four times the execution time to perform this process. For the proposed models, YOLOv5 produced insignificant improvements and YOLOv7Face had deteriorated results using the re-orientation method. A potential explanation for this is these models were trained on raw input images that encompassed the various neonatal poses and postures without re-orientation. Instead, the re-orientation changed the relationship of the newborn with regard to the clinical environment. # CONCLUSIONS This paper presented two re-trained YOLO-based models that showed significant improvement in face and facial landmark detection. Additionally, the re-orientation of images showed marked improvement in existing methods, whereas the proposed fusion of methods showed minor improvements. Future work into neonatal face segmentation to better localise the face should be investigated.

{'timestamp': '2023-02-10T02:02:16', 'yymm': '2302', 'arxiv_id': '2302.04341', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04341'}

# Introduction {#sec:introduction} Diffusion models have shown great success in generating images with both high diversity and high fidelity. Recent work has demonstrated superior performance than state-of-the-art GAN models, which suffer from unstable training. As a class of flexible generative models, diffusion models demonstrate their power in various applications such as image super-resolution, inpainting, shape generation, graph generation, image-to-image translation, and molecular conformation generation. However, the generation process for diffusion models can be slow due to the need for iterative noise estimation using complex neural networks at each time step. Such a requirement creates obstacles for deployment, particularly on devices with resource constraints. The inference normally requires 50 to 1,000 denoising time steps and has high memory consumption. While previous state-of-the-art approaches (e.g. GANs) are able to generate multiple images in under 1 second, it normally takes several seconds for a diffusion model to sample a single image. Consequently, speeding up the image generation process and reducing the memory requirements are one important step towards broadening the applications of diffusion models. The high latency for generating an image through diffusion comes from two places: On one hand, the denoising process consists of iterative steps to estimate the noise to remove from the model input. On the other hand, the model itself is compute-intensive. While previous work has attempted to accelerate the first factor by finding shorter, more effective sampling trajectories, these approaches do not address the issue of the time-consuming noise estimation process in each iteration of neural network inference. In contrast, our work tackles the second factor, where we focus on compressing and accelerating the noise estimation model, U-Net, used in the diffusion denoising process. Among previously explored compression methods, the extensive shortcut connections of U-Net models lead to complex dimensionality constraints across multiple layers, which requires significant engineering effort to find valid structure via pruning  or neural architecture search . Quantization compresses the model without modifying the model architecture. However, the commonly used quantization-aware training  retrains the model at low precision, which may be unstable given the complicated training scheme of diffusion models, and are constrained by the lack of public training data . Therefore, we find the training-free post-training quantization (PTQ)  the most applicable for diffusion model compression. In this work, we propose a solution to leverage PTQ to compress the cumbersome noise estimation network in diffusion models in a data-free, training-free manner, while maintaining comparable performance to the full precision counterparts. Our work provides the first thorough analysis of the novel challenges of performing PTQ on diffusion model. Specifically, we discover that the output distribution of the noise estimation network at each time step can be largely different, and naively applying previous PTQ approaches with an arbitrary time step leads to poor performance. Furthermore, the distinctive U-Net architecture used by diffusion models requires special design on the quantization scheme and calibration objectives. We further analyze the sensitivity of U-Net architecture to quantization and adjust our quantization scheme accordingly for the best efficiency-accuracy trade-off. The contributions of this paper can be summarized as follows: 1. We perform a thorough analysis of the weight and activation distribution in the granularity of time steps and layers, showcasing the challenges of performing PTQ on diffusion models. 2. We propose a cross-time step calibration data sampling method for diffusion model PTQ to balance the calibration quality and the dataset size. 3. Our proposed method enables W4A8 PTQ for both pixel-space and latent-space diffusion models with a FID increment of only \(0.39 {\text-} 1.88\) and can produce qualitatively comparable samples when being applied on Stable Diffusion for text-guided image synthesis. # Background {#sec:background} #### Image Synthesis Image synthesis aims to generate synthetic images that follow the distribution of real images. Deep generative models have made great progress in image synthesis in recent years. Among them the most popular types of generative models are variational auto-encoder (VAE), flow-based models, generative adversarial networks (GAN), and diffusion-based models. GANs have dominated the field of image synthesis for years before recently diffusion models are shown to perform on par or better than state-of-the-art GANs with more stable training procedures and can be applied to a wide range of modalities. #### Diffusion Models Diffusion models model the image generation process with a series of steps on a Markov chain. During the forward diffusion process, each step adds noise to the data, eventually making the data follow a standard normal distribution. The reverse process removes noise from the data, which is leveraged to gradually generate high-fidelity images from standard Gaussian noise. See for a graphical illustration. We refer readers to for a more detailed introduction. Let \(\vec{x}_0\) be a sample from the data distribution \(\vec{x}_0 \sim q(\vec{x})\). A forward diffusion process adds Gaussian noise to the sample for \(T\) times, resulting in a sequence of noisy samples \(\vec{x}_1,..., \vec{x}_T\): \[\begin{aligned} q(\vec{x}_t|\vec{x}_{t-1}) = \mathcal{N}(\vec{x}_{t}; \sqrt{1-\beta_t}\vec{x}_{t-1}, \beta_t\mathbf{I}) \end{aligned}\] where \(\beta_t \in (0,1)\) is the variance schedule and controls the strength of the Gaussian noise in each step. Since each step only depends on the one previous step, the forward diffusion process follows Markov property. Furthermore, when \(T \rightarrow \infty\), \(\vec{x}_T\) approaches an isotropic Gaussian distribution. Diffusion models generate a sample from a Gaussian noise input \(\vec{x}_T \sim \mathcal{N}(\vec{0}, \mathbf{I})\) by reversing the forward process. However, since the real reverse conditional distribution \(q(\vec{x}_{t-1}|\vec{x}_t)\) is unavailable to us, diffusion models sample from a learned conditional distribution \(p_\theta(\vec{x}_{t-1} | \vec{x}_t)\) which approximates the real reverse conditional distribution with a Gaussian distribution: \[\begin{aligned} \label{eq:diff_rev} p_\theta(\vec{x}_{t-1} | \vec{x}_t) &= \mathcal{N}(\vec{x}_{t-1}; \tilde{\vec{\mu}}_{\theta,t}(\vec{x}_t), \tilde{\beta}_t\mathbf{I}) \end{aligned}\] With the reparameterization trick in, the mean \(\tilde{\vec{\mu}}_{\theta,t}(\vec{x}_t)\) and variance \(\tilde \beta_t\) could be derived: \[\begin{aligned} \tilde{\vec{\mu}}_{\theta,t}(\vec{x}_t) &= \frac{1}{\sqrt{\alpha_t}} (\mathbf{x}_t-\frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_t}} \vec{\epsilon}_{\theta,t}) \\ \tilde \beta_t &= \frac{1-\bar{\alpha}_{t-1}}{1-\bar{\alpha}_t} \cdot \beta_t \end{aligned}\] where \(\alpha_t = 1-\beta_t\), \(\bar \alpha_t = \prod_{i=1}^t \alpha_i\). At training time, the goal of optimization is to minimize the negative log-likelihood \(-\log p_\theta(\mathbf{x}_0)\). With variational inference, a lower bound of it could be found, denoted as \(L_\text{VLB}\): \[\begin{aligned} L_\text{VLB} = \mathbb{E}_{q(\vec{x}_{0:T})} [ \log \frac{q(\vec{x}_{1:T}|\vec{x}_0)}{p_\theta(\vec{x}_{0:T})} ] \geq-\log p_\theta(\vec{x}_0) \end{aligned}\] It is found in that using a simplified loss function to \(L_\text{VLB}\) often obtains better performance: \[\begin{aligned} L_\text{simple} &= \mathbb{E}_{t, \vec{x}_0, \vec{\epsilon}_t} [\|\vec{\epsilon}_t-\vec{\epsilon}_\theta(\sqrt{\bar{\alpha}_t}\vec{x}_0 + \sqrt{1-\bar{\alpha}_t}\vec{\epsilon}_t, t)\|^2 ] \end{aligned}\] At inference time, a Gaussian noise tensor \(\vec{x}_T\) is sampled and is denoised by repeatedly sampling the reverse distribution \(p_\theta(\vec{x}_{t-1}|\vec{x}_t)\). \(\tilde{\vec{\mu}}_{\theta,1}(\vec{x}_1)\) is taken as the final generation result, with no noise added in the final denoising step. #### Post-training Quantization In situations with limited training data or computing resources for the time-consuming quantization-aware fine-tuning process , post-training quantization (PTQ) becomes a preferred method of quantization. Most previous works  use the rounding-to-nearest approach, which quantizes deep neural networks by rounding elements \(w\) to the nearest quantization values. The quantization and dequantization can be formulated as: \[\label{eq:1} \hat{w} = \mathrm{s} \cdot \mathrm{clip}(\lfloor\frac{w}{s}\rceil, c_\text{min}, c_\text{max})\] where \(s\) denotes the quantization scale parameters, \(c_\text{min}\) and \(c_\text{max}\) are the lower and upper bounds for the clipping function \(\mathrm{clip}(\cdot)\). The operator \(\lfloor \cdot \rceil\) represents rounding-to-nearest. Researchers have previously focused on using PTQ with convolutional neural networks (CNNs). For example, EasyQuant presents an efficient method to determine appropriate clipping ranges for quantization, and ZeroQ employs a distillation technique to generate proxy input images for PTQ, which utilizes the inherent statistics of batch normalization layers. SQuant adjusts the model to reduce quantization error based on sensitivity determined through the Hessian spectrum. BRECQ introduces Fisher information into the objective, and optimizes layers within a single residual block jointly using a small subset of calibration data from the training dataset. Some of the PTQ methods still need training datasets to calibrate quantized models but are often unavailable due to privacy and commercial reasons, such as medical and confidential scenarios. Fortunately, in the quantization of diffusion models, the calibration dataset can be constructed by sampling the full precision model, as the input for the denoising process is always gaussian noise (with some user-defined conditions in the conditional generation scenarios). Given the diffusion models are fully trained to convergence, it is reasonable to assume that the sampled data distribution will be a nearly exact match to the training data distribution, enabling the creation of a high-quality calibration dataset in a zero-shot manner. This further shows that post-training quantization is a well-suited method for accelerating diffusion models. # Method {#sec:method} We present our method for diffusion model post-training quantization in this section. Different from conventionally studied deep learning models and tasks (e.g. CNNs, VITs for image classification, or object detection), diffusion models are trained and evaluated in a distinctive multi-step manner. This draws notable challenges to the PTQ process. We analyze the challenge brought by the multi-step inference process in Section [3.1](#ssec:time){reference-type="ref" reference="ssec:time"}, and describe the full PTQ pipeline in Section [3.2](#ssec:PTQ){reference-type="ref" reference="ssec:PTQ"}. ## Challenges under the Multi-step Denoising {#ssec:time} Here we identify two major challenges on PTQ specific to the multi-step inference process of diffusion models. Namely, we investigate on the accumulation of quantization error, and the difficulty of sampling a small calibration dataset to reduce the quantization error. #### Challenge 1: Quantization errors accumulate across time steps. Performing quantization on a neural network model introduces noises on the weight and activation of the well-trained model, leading to quantization errors in each layer's output. Previous research has identified that quantization errors are likely to accumulate across layers, making deeper neural networks harder to quantize. In the case of diffusion models, at any time step \(t\), the input of the denoising model (denote as \(\vec{x}_t\)) is derived by \(\vec{x}_{t+1}\), the output of the model at the previous time step \(t+1\) (as depicted by ). This process effectively multiplies the number of layers involved in the computation by the number of denoising steps for the input \(\vec{x}_{t}\) at time step \(t\), leading to an accumulation of quantization errors towards later steps in the denoising process. We run the denoising process of DDIM on CIFAR-10 with a sampling batch size of 64, and compare the MSE differences between the full precision model and the model quantized to INT8, INT5, and INT4 at each time step. As shown in , there is a dramatic increase in the quantization errors when the model got quantized to 4-bit, and the errors accumulate quickly through iterative denoising. This brings difficulty in preserving the performance after quantizing the model down to low precision, which requires the reduction of quantization errors at all time steps as much as possible. #### Challenge 2: Activation distributions vary across time steps. To reduce the quantization errors at each time step, we follow the common practice in PTQ research to calibrate the clipping range and scaling factors of the quantized model with a small set of calibration data. The calibration data should be sampled to resemble the true input distribution so that the activation distribution of the model can be estimated correctly for proper calibration. Given that the diffusion model takes input from all time steps, determining the data sampling policy across different time steps becomes an outstanding challenge. Here we start by analyzing the output activation distribution of each layer in the UNet model across different time steps. We conduct the same CIFAR-10 experiment using DDIM with 100 denoising steps, and draw the activations ranges of 1000 random samples among all time steps. As shows, the activation distributions gradually change, with neighboring time steps being similar and distant ones being distinctive. The fact that output activations distribution across time steps are not always the same further brings challenges to quantization. Calibrating the UNet model using a single time step will cause overfitting to the activation distribution of this specific time step, while not generalizing to other time steps. For instance, here we try to calibrate the quantized DDIM on the CIFAR-10 dataset with the commonly used BRECQ  method, and compare it with the PTQ model without calibration. As shown in , if we naively take 5120 samples produced by the denoising process from the initial input (\(t=T\)) for calibration, significant performance drops will be induced under 8-bit and 4-bit weights quantization. The calibrated 8-bit model even performed worse than the naive PTQ model without calibration. This indicates that calibrating with a single time step would lead to significant overfitting and be harmful to the overall performance. In order to recover the performance of the quantized diffusion models, we need to select calibration data in a way that takes the different time steps outputs distribution into consideration. ## Post-Training Quantization with Timestep-Aware Calibration {#ssec:PTQ} In diffusion models inference, the denoising process can be interpreted as a procedure that starts from a sample taken from an isotropic Gaussian distribution, and gradually removes the noise at each time step to recover the corresponding \"denoised\" image from the training data distribution in the end. This indicates inputs at nearby consecutive time steps will have relatively similar distributions, while inputs at distant time steps are distributed more diversely between one another, as illustrated by and . With this observation, we randomly sample a few intermediate inputs uniformly in a fixed interval across all time steps to generate a small calibration set, in order to balance the size of the calibration set and its representation ability of the distribution across all time steps. We empirically found that the sampled calibration data is able to recover most of the INT4 quantized models' performance after the calibration, thus adapting this sampling scheme to create the calibration dataset for quantization error correction. Activation functions are kept running at full precision. To calibrate the quantized model, we divide the model into several reconstruction blocks, and iteratively reconstruct outputs and tune the clipping range and scaling factors of weight quantizers in each block with adaptive rounding to minimize the mean squared errors between the quantized and full precision outputs. We define a core component that contains residual connections in the diffusion model UNet as a block, such as a Residual Bottleneck Block or a Transformer Block. Other parts of the model that do not satisfy this condition are calibrated in a per-layer manner. This has been shown to improve the performance to fully layer-by-layer calibration due to better addressing the inter-layer dependencies and generalization. For activation quantization, since activations keep changing over inputs, doing adaptive rounding is infeasible and we only adjust step sizes of quantizers according to . The overall calibration workflow for the quantization is described as follows: [\[tab:bedroom_results\]]{#tab:bedroom_results label="tab:bedroom_results"} [\[tab:church_results\]]{#tab:church_results label="tab:church_results"} ## Text-guided Image Generation In this section, we evaluate Q-Diffusion on Stable Diffusion pretrained on subsets of 512 \(\times\) 512 LAION-5B for text-guided image generation. Inspired by, we sample text prompts from the MS-COCO dataset and include them in the inputs of to generate a calibration dataset with texts condition. In this work, we fix the guidance strength to the default 7.5 in Stable Diffusion as the trade-off between sample quality and diversity. Qualitative results are shown in . Compared to Linear Quantization, our Q-Diffusion provides higher-quality images with more realistic details and better reflections of the semantic information. The output of the W4A8 Q-Diffusion model largely resembles the output of the full precision model. Interestingly, we find some diversity in the lower-level semantics between the Q-Diffusion model and the FP models, like the heading of the horse or the shape of the hat. Understanding how quantization contributes to this diversity shall be explored in the future. # Related Work There are several works on improving the efficiency and effectiveness of the diffusion process. Related methods include simulating diffusion process in fewer steps by generalizing it to a non-Markovian process, adjusting the variance schedule, and leveraging high-order solvers to approximate diffusion generation. However, these works mainly focus on sampling with fewer steps rather than accelerating the model in each step. There is also a work that progressively distills the diffusion model into a student model that requires fewer time steps to provide same quality images, which achieves relatively good results but involves extremely expensive retraining procedures. # Conclusion This work studies the use of quantization to accelerate diffusion models. We propose Q-Diffusion, a novel post-training quantization scheme that conducts calibration with multiple time steps in the denoising process and achieves significant improvements in quantized model performance. Q-Diffusion models under 4-bit quantization achieve comparable results to the full precision models. In the future, we will investigate the optimal time steps importance sampling for the calibration process based on the property of the corresponding time step, utilize mixed-precision to further conduct ultra low-bit quantization, and implement customized integer CUDA kernels to measure real runtime speedup and memory reduction. # Extended Experimental Settings ## Implementation Details We describe the implementation and compute details of the experiments in this section. We adapt the official implementation for DDIM  [^1] and Latent Diffusion  [^2]. For Stable Diffusion, we use the CompVis codebase [^3] and its v1.4 checkpoint. We use the torch-fidelity library [^4] to evaluate FID and IS scores as done in. For quantization experiments, we quantize all weights and activations, but leave activation functions and attention matrix multiplications running with full precision. As demonstrated in , we always disable activation quantization for the first three Conv layers with residual connections and use INT8 precision in DDIM experiments on CIFAR-10 as activations of them present as exceedingly large outliers. No special modifications or mixed precision are done for other experiments. ## Hyperparameters Here we provide the hyperparameters used for our Q-Diffusion calibration in . # Layer-wise Activations Distribution in DDIM and LDM {#ssec:act_layers} We study the distribution of activation values across all time steps in DDIM on CIFAR-10 and LDM on LSUN-Church. shows that the first three Conv layers with residual connections in DDIM have much larger activation ranges that go up to around \(600 \text{-} 1200\), while these ranges for the majority layers are \(<50\). On the other hand, all layers in LDM share relatively uniform activation distributions, having a range \(<15\). # Additional Random Samples In this section, we provide our non-cherry-picked samples from our weight-only quantized and fully quantized diffusion models obtained using Q-Diffusion under 4-bit quantization. Results are shown in the figures below. [^1]: <https://github.com/ermongroup/ddim> [^2]: <https://github.com/CompVis/latent-diffusion> [^3]: <https://github.com/CompVis/stable-diffusion> [^4]: <https://github.com/toshas/torch-fidelity>

{'timestamp': '2023-02-13T02:09:12', 'yymm': '2302', 'arxiv_id': '2302.04304', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04304'}

# Introduction {#sec:intro} One of the next frontiers in cosmological science using the Cosmic Microwave Background (CMB) is the observation of large-scale \(B\)-mode polarization, and the consequential potential detection of primordial gravitational waves. Such a detection would let us glance into the early Universe and its very high energy physics, at scales unattainable by any other experiment. Primordial tensor perturbations, which would constitute a stochastic background of primordial gravitational waves, would source a parity-odd \(B\)-mode component in the polarization of the CMB. The ratio between the amplitudes of the primordial power spectrum of these tensor perturbations and the primordial spectrum of the scalar perturbations is referred to as the tensor-to-scalar ratio \(r\). This ratio covers a broad class of models of the early Universe, allowing us to test and discriminate between models that predict a wide range of values of \(r\). These include vanishingly small values, as resulting from models of quantum gravity, as well as those expected to soon enter the detectable range, predicted by models of inflation. An unequivocal measurement of \(r\), or a stringent upper bound, would thus greatly constrain the landscape of theories on the early Universe. Although there is no evidence of primordial \(B\)-modes yet, current CMB experiments have placed stringent constraints on their amplitude, finding \(r < 0.036\) at 95% confidence when evaluated at a pivot scale of 0.05 Mpc\(^{-1}\). At the same time, these experiments have firmly established that the power spectrum of primordial scalar perturbations is not exactly scale-independent, with the scalar spectral index \(n_s-1\sim0.03\). Given this measurement, several classes of inflationary models predict \(r\) to be in the \(\sim10^{-3}\) range. Even though the only source of primordial large-scale \(B\)-modes at linear order are tensor fluctuations, in practice a measurement is complicated by several factors: first, the gravitational deflection of the background CMB photons by the cosmic large-scale structure creates coherent sub-degree distortions in the CMB, known as CMB lensing. Through this mechanism, the nonlinear scalar perturbations from the late Universe transform a fraction of the parity-even \(E\)-modes into \(B\)-modes at intermediate and small scales. Second, diffuse Galactic foregrounds have significant polarized emission, and in particular foreground components such as synchrotron radiation and thermal emission from dust produce \(B\)-modes with a significant amplitude. Component separation methods, which exploit the different spectral energy distributions (SED) of the CMB and foregrounds to separate the different components, are thus of vital importance. Practical implementations of these methods must also be able to carry out this separation in the presence of instrumental noise and systematic effects. Polarized Galactic foregrounds pose a formidable obstacle when attempting to measure primordial \(B\)-modes at the level of \(r \sim 10^{-3}\). Current measurements of Galactic emission demonstrate that the Galactic \(B\)-mode signal is dominant over the cosmological signal on the relevant scales. At the minimum of polarized Galactic thermal dust and synchrotron, around 80 GHz, their \(B\)-mode signal represents an effective tensor-to-scalar ratio with amplitude larger than the target CMB signal, even in the cleanest regions of the sky. Component separation methods are able to clean most of this, but small residuals left after the cleaning could be comparable to the primordial \(B\)-mode signal we want to measure. In recent years, many works have analyzed this problem and made forecasts of how well we could potentially measure \(r\) with different ground-based and satellite experiments. In general, these works have highlighted how, if left untreated, systematic residuals arising from a simplistic characterization of foregrounds will bias an \(r\sim10^{-3}\) measurement by several \(\sigma\). Thus, it is of vital importance to model the required foreground complexity when cleaning the multi-frequency CMB observations and to keep a tight control of systematics without introduction of bias. Multiple upcoming CMB experiments rank the detection of large-scale primordial \(B\)-modes among their primary science targets. Near-future experiments such as the BICEP Array target a detection at the level of \(r \sim 0.01\), while in the following decade, next-generation projects such as LiteBIRD and CMB-S4 aim at \(r \sim 0.001\). The Simons Observatory (SO), like the BICEP Array, targets the detection of primordial gravitational waves at the level of \(r\sim0.01\), and its performance at realizing this goal is the main focus of this paper. SO is a ground-based experiment, located at the Cerro Toco site in the Chilean Atacama desert, which will observe the microwave sky in six frequency channels, from 27 to 280 GHz, due to start in 2023. SO consists of two main instruments. On the one hand, a Large Aperture Telescope (LAT) with a 6m diameter aperture will target small-scale CMB physics, secondary anisotropies and the CMB lensing signal. Measurements of the latter will serve to subtract lensing-induced \(B\)-modes from the CMB signal to retrieve primordial \(B\)-modes. On the other hand, multiple Small Aperture Telescopes (SATs) with 0.5m diameter apertures will make large-scale, deep observations of \(\sim 10\)% of the sky, with the main aim of constraining the primordial \(B\)-mode signal, peaking on scales \(\ell \sim 80\) (the so-called "recombination bump"). See for an extended discussion on experimental capabilities. In this paper, we aim at validating three independent \(B\)-mode analysis pipelines. We compare their performance regarding a potential \(r\) measurement by the SO SATs, and evaluate the capability of the survey to constrain \(\sigma(r=0) \leq 0.003\) in the presence of foreground contamination and instrumental noise. To that end, we produce sky simulations encompassing different levels of foreground complexity, CMB with different values of \(r\) and different amounts of residual lensing contamination, and various levels of the latest available instrumental noise[^1], calculated from the parametric models presented in. We feed these simulations through the analysis pipelines and test their performance, quantifying the bias and statistical uncertainty on \(r\) as a function of foreground and noise complexity. The three pipelines are described in detail in Section [2](#sec:pipelines){reference-type="ref" reference="sec:pipelines"}. Section [3](#sec:sims){reference-type="ref" reference="sec:sims"} presents the simulations used in the analysis, including the models used to produce CMB and foreground sky maps, as well as instrumental noise. In Section [4](#sec:results){reference-type="ref" reference="sec:results"}, we present our forecasts for \(r\), accompanied by the power spectrum products and a channel weights comparison. Section [5](#sec:d10s5){reference-type="ref" reference="sec:d10s5"} shows preliminary results on a set of new, complex foreground simulations. In Section [6](#sec:conc){reference-type="ref" reference="sec:conc"} we summarize and draw our conclusions. Finally, Appendix [\[app:chi_squared\]](#app:chi_squared){reference-type="ref" reference="app:chi_squared"} summarizes the \(\chi^2\) analysis performed on the Cross-\(C_\ell\) cleaning pipeline, while Appendix [\[sec:nilc_gaussian_bias\]](#sec:nilc_gaussian_bias){reference-type="ref" reference="sec:nilc_gaussian_bias"} discusses biases on Gaussian simulations observed with the NILC cleaning pipeline. # Methods, pipelines {#sec:pipelines} In this section we present our three component separation pipelines, that adopt complementary approaches widely used in the literature: power-spectrum-based parametric cleaning (), Needlet Internal Linear Combination (NILC) blind cleaning, and map-based parametric cleaning. In the following, these are denominated pipelines A, B and C, respectively. The cleaning algorithms operate on different data spaces (harmonic, needlet and pixel space) and vary in their cleaning strategy (parametric, meaning that we assume an explicit model for the frequency spectrum of the foreground components, or blind, meaning that we do not model the foregrounds or make any assumptions on how their frequency spectrum should be). Hence, they do not share the same set of method-induced systematic errors. This will serve as an important argument in favor of claiming robustness of our inference results. Table [\[tab:pipelines_overview\]](#tab:pipelines_overview){reference-type="ref" reference="tab:pipelines_overview"} lists the three pipelines and their main properties. Although there are some similarities between these analysis pipelines and the forecasting frameworks that were exploited in, the tools developed for this paper are novel implementations designed to deal with realistic SO data-like inputs, including complex noise and more exotic foreground simulations compared to what was considered in the previous work. We stress again that no filtering or other systematic effects were included in the noise maps. ## Pipeline A: Cross-\(C_\ell\) cleaning {#ssec:pipelines.Cell} Pipeline A is based on a multi-frequency power-spectrum-based component separation method, similar to that used in the latest analysis carried out by the BICEP/*Keck* collaboration. The data vector is the full set of cross-power spectra between all frequency maps, \(C_\ell^{\nu\nu'}\). The likelihood compares this against a theoretical prediction that propagates the map-level sky and instrument model onto the corresponding power spectra. The full pipeline is publicly available[^2], and a schematic overview is provided in Figure [\[fig:schematic_a\]](#fig:schematic_a){reference-type="ref" reference="fig:schematic_a"}. In step 1, power spectra are measured using a pseudo-\(C_\ell\) approach with \(B\)-mode purification as implemented in `NaMaster`. As described in the presence of a sky mask leads to the presence of ambiguous modes contaminated by full-sky \(E\)-modes. These must be removed at the map level to avoid the contribution to the power spectrum uncertainties from the leaked \(E\)-modes. The mask used for this analysis traces the hits count map released in (see Figure [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}), and its edges are apodized using a C1-type kernel with an apodization scale of 10 degrees, yielding an effective sky coverage of \(f_{\rm sky}\sim10\%\). Each power spectrum is calculated in bandpower windows with constant bin width \(\Delta\ell=10\), of which we only keep the range \(30\leq\ell\leq300\). Our assumption is that on real data, larger scales are contaminated by atmospheric noise and filtering, whereas smaller scales, targeted by the SO-LAT and useful for constraining lensing \(B\)-modes, do not contain any significant primordial \(B\)-mode contribution. To avoid a significant bias in the auto-correlations when removing the impact of instrumental noise, a precise noise model is required that may not be available in practice. We address this issue by using data splits, which in the case of real data may be formed by subdividing data among observation periods, sets of detectors, sky patches or by other means, while in this paper, we resort to simulations. We construct simulated observations for each sky realization comprising \(S=4\) independent splits with the same sky but different noise realizations (each with a commensurately larger noise amplitude). We compute \(BB\) power spectra from pairs of maps, each associated with a given data split and a given frequency channel. For any fixed channel pair combination, we average over the corresponding set of \(S(S-1)/2=6\) power spectra with unequal split pairings. For \(N=6\) SAT frequency channels, this results in a collection of \(N(N+1)/2=21\) noise-debiased multi-frequency power spectra, shown in Figure [\[fig:raw_data_a\]](#fig:raw_data_a){reference-type="ref" reference="fig:raw_data_a"}. Note that we could have modeled and subtracted the noise bias explicitly, since we have full control over the noise properties in our simulations. In realistic settings, however, the accuracy of an assumed noise model may be limited. While inaccurate noise modeling would affect the statistical uncertainty \(\sigma(r)\) through the covariance matrix calculated from simulations, the cross-split approach ensures robustness of the inferred value of \(r\) against noise-induced bias. In step 2, we estimate the bandpower covariance matrix from simulations, assuming no correlations between different multipole windows. Note that, to include foreground signal variance in the budget, Gaussian foreground simulations (see Section [3](#sec:sims){reference-type="ref" reference="sec:sims"}) are considered for the covariance computation step, since our realistic foreground templates cannot be used as statistical samples. As we show in Appendix [\[app:chi_squared\]](#app:chi_squared){reference-type="ref" reference="app:chi_squared"}, this covariance matrix is indeed appropriate, as it leads to the theoretically expected empirical distribution of the \(\hat{\chi}^2_{\rm min}\) statistic not only in the case of Gaussian foregrounds, but also for the non-Gaussian foreground simulations. Inaccurate covariance estimates would make this statistic peak at higher or lower values, which we do not observe. Step \(3\) is the parameter inference stage. We use a Gaussian likelihood when comparing the multi-frequency power spectra with their theoretical prediction. Note that, in general, the power spectrum likelihood is non-Gaussian, and provide an approximate likelihood that is able to account for this non-Gaussianity. We explicitly verified that both likelihoods lead to equivalent parameter constraints, and thus choose the simpler Gaussian option. The validity of the Gaussian approximation is a consequence of the central limit theorem, since each measured bandpower consists of effectively averaging over \(N_{\rm mode}\simeq\Delta\ell\times f_{\rm{sky}}\times(2\ell+1) > 61\) independent squared modes on the scales used here. Note that this assumption is valid thanks to the relatively large SO-SAT sky patch and would not hold any longer for BICEP/*Keck*-like patch sizes. The default sky model is the same as that described in. We model the angular power spectra of dust and synchrotron as power laws of the form \(D_\ell=A_c(\ell/\ell_0)^{\alpha_c}\), with \(\ell_0=80\), and \(c=d\) or \(s\) for dust and synchrotron respectively. The dust SED is modeled as a modified black-body spectrum with spectral index \(\beta_d\) and temperature \(\Theta_d\), which we fix to \(\Theta_d=20\,{\rm K}\). The synchrotron SED is modeled as a power law with spectral index \(\beta_s\). Finally, we consider a dust-synchrotron correlation parameter \(\epsilon_{ds}\). Including the tensor-to-scalar ratio \(r\) and a free lensing \(B\)-mode amplitude \(A_{\rm lens}\), this fiducial model has \(9\) free parameters: \[\{A_{\rm lens},r,A_d,\alpha_d,\beta_d,A_s,\alpha_s,\beta_s,\epsilon_{ds}\}.\] We will refer to results using this model as "\(C_\ell\)-fiducial". Table [\[tab:priors_pipeline_a\]](#tab:priors_pipeline_a){reference-type="ref" reference="tab:priors_pipeline_a"} lists the priors on its \(9\) parameters.\ The main drawback of power-spectrum-based pipelines in their simplest incarnation, is their inability to account for spatial variation in the foreground spectra. If ignored, this spatial variation can lead to biases at the level of \(r\sim O(10^{-3})\), which are significant for the SO target. At the power spectrum level, spatially-varying SEDs give rise to frequency decorrelation, which can be included in the model. Here, we will show results for an extended model that uses the moment-based [^3] parameterization of to describe the spatial variation of \(\beta_d\) and \(\beta_s\). The model introduces 4 new parameters \[\label{eq:moment_params} \{B_s,\gamma_s,B_d,\gamma_d\}\,,\] where \(B_c\) parameterizes the amplitude of the spatial variations in the spectral index of component \(c\), and \(\gamma_c\) is their power spectrum slope (see for further details). We will refer to results using this method as "\(C_\ell\)-moments", or "A + moments". The priors in the shared parameter space are the same as in \(C_\ell\)-fiducial, and Table [\[tab:priors_pipeline_a\]](#tab:priors_pipeline_a){reference-type="ref" reference="tab:priors_pipeline_a"} lists the priors on its additional \(4\) parameters. For both methods, we sample posteriors using the `emcee` code. Note that we assume a top-hat prior on \(r\) in the range \([-0.1,\, 0.1]\) instead of imposing \(r>0\). The reason is that we would like to remain sensitive to potential negative biases on \(r\). While negative \(r\) values do not make sense physically, they may result from e.g. volume effects caused by choosing specific priors on other parameters that we have marginalized over. Opening the prior on \(r\) to negative values allows us to monitor these unwanted effects, offering a simple robustness check. On real data, this will be replaced by a positivity prior \(r>0\), but only after ensuring that our specific prior choices on the other parameters do not bias \(r\), which is the focus of a future work. ## Pipeline B: NILC cleaning {#sec:nilc} Our second pipeline is based on the blind Internal Linear Combination (ILC) method, which assumes no information on foregrounds whatsoever, and instead only assumes that the observed data contains one signal of interest (the CMB) plus noise and contaminants. The method assumes a simple model for the observed multi-frequency maps \(\mathbf{d}_{\nu}\) at \(N_{\nu}\) frequency channels (in either pixel or harmonic space) \[\mathbf{d}_{\nu} = a_{\nu} \mathbf{s} + \mathbf{n}_{\nu} \text{,}\] where \(a_{\nu}\) is the black-body spectrum of the CMB, \(\mathbf{s}\) is the amplitude of the true CMB signal and \(\mathbf{n}_{\nu}\) is the contamination in channel \(\nu\), which includes the foregrounds and instrumental noise. ILC exploits the difference between the black-body spectrum of the CMB and the SED(s) of other components that may be present in the data. The method aims at reconstructing a map of the CMB component \(\tilde{\mathbf{s}}\) as a linear combination of the data with a set of weights \(w_{\nu}\) allowed to vary across the map, \[\label{eq:ilc_reconstruction} \tilde{\mathbf{s}} = \sum_{\nu} w_{\nu} \mathbf{d}_{\nu} = \mathbf{w}^{T} \hat{\sf d}\,,\] where both \(\mathbf{w}\) and \(\hat{\sf d}\) are \(N_\nu\times N_{\rm pix}\) matrices,with \(N_{\rm pix}\) being the number of pixels. Optimal weights are found by minimizing the variance of \(\tilde{\bf s}\). The result is given by: \[\label{eq:nilc_weights} \mathbf{w}^T = \frac{\mathbf{a}^T \hat{\sf C}^{-1} }{\mathbf{a}^T \hat{\sf C}^{-1} \mathbf{a} } \text{,}\] where \(\mathbf{a}\) is the black-body spectrum of the CMB (i.e. a vector filled with ones if maps are in thermodynamic temperature units) and \(\hat{\sf C} = \langle \hat{\sf d}\,\hat{\sf d}^T \rangle\) is the frequency-frequency covariance matrix per pixel of the observed data. Note that we assume no correlation between pixels for the optimal weights. In our particular implementation, we use the Needlet Internal Linear Combination method. NILC uses localization in pixel and harmonic space by finding different weights \({\bf w}\) for a set of harmonic filters, called "needlet windows". These windows are defined in harmonic space \(h_{i}(\ell)\) for \(i=0...,n_{\rm windows}-1\) and must satisfy the constraint \(\sum_{i=0}^{n_{\rm windows}-1} h_{i}(\ell)^2 = 1\) in order to preserve the power of the reconstructed CMB. We use \(n_{\rm windows}=5\) needlet windows shown in Figure [\[fig:nilc_windows\]](#fig:nilc_windows){reference-type="ref" reference="fig:nilc_windows"}, and defined by \[h_{i}(\ell) = \begin{cases} \cos(\frac{\pi}{2}(\ell^{\rm peak}_i-\ell)/(\ell^{\rm peak}_i-\ell^{\rm min}_i)) & \text{if \(\ell^{\rm min}_i \leq \ell < \ell^{\rm peak}_i\) }\\ 1 & \text{if \(\ell = \ell^{\rm peak}_i\)} \\ \cos(\frac{\pi}{2}(\ell-\ell^{\rm peak}_i)/(\ell^{\rm max}_i-\ell^{\rm peak}_i)) & \text{if \(\ell^{\rm peak}_i < \ell \leq \ell^{\rm max}_i\)} \end{cases} \text{,}\] with \(\ell_{\rm min}=\{\)`<!-- -->`{=html}0, 0, 100, 200, 350\(\}\), \(\ell_{\rm max}=\{\)`<!-- -->`{=html}100, 200, 350, 500, 500\(\}\), and \(\ell_{\rm peak}=\{\)`<!-- -->`{=html}0, 100, 200, 350, 500\(\}\) for the corresponding 5 needlet windows. Even though we do not use the full 500-\(\ell\) range where windows are defined for the likelihood sampling, we still perform the component separation on all 5 windows up to multipoles beyond our upper limit of \(\ell=300\), in order to avoid edge effects on the smaller scales. Let us now describe the NILC procedure as illustrated in Figure [\[fig:schematic_b\]](#fig:schematic_b){reference-type="ref" reference="fig:schematic_b"}. In step 1, we perform our CMB reconstruction in the \(E\) and \(B\) field instead of \(Q\) and \(U\). We transform the observed maps to \(a_{\ell m}^X\) with \(X \in E,B\). All frequency channels are then brought to a common beam resolution by rescaling the harmonic coefficients with an appropriate harmonic beam window function. The common beam we adopt is the one from the third frequency channel at 93 GHz, which corresponds to a FWHM of 30 arcmin. For each needlet window index \(i\), we multiply \(a_{\ell m}^X\) with \(h_i(\ell)\) as a harmonic filter. Since different frequency channels have different limiting resolutions, we do not use all channels in every needlet window: the first 2 windows use all 6 frequency channels, the third window does not use the 27 GHz channel and the last 2 needlet windows do not use the 27 and 39 GHz channels. The covariance matrix \(\hat{\sf C}\) has dimensions \(N_\nu\times N_\nu\times N_{\rm pix}\). For each pixel \(p\), its corresponding \(N_\nu\times N_\nu\) elements are computed directly from the data, averaging over the pixels inside a given pixel domain \(\mathcal{D}(p,i)\) around each pixel. In practice, the element \(\nu,\nu'\) of the covariance matrix is calculated by multiplying the two filtered maps at channels \(\nu\) and \(\nu'\), then smoothing that map with a Gaussian kernel with FWHM equal to the size of the pixel domain \(\mathcal{D}(p,i)\). The FWHMs for the pixel domain size are 185, 72, 44, 31, and 39 degrees for each needlet window, respectively. [^4] We then proceed to calculate the weights \(\mathbf{w}^T\) (see Equation [\[eq:nilc_weights\]](#eq:nilc_weights){reference-type="ref" reference="eq:nilc_weights"}) for window \(i\), which is an array with shape (2, \(N_{\nu}\), \(N_{\rm pixels}^{i}\)), with the first dimension corresponding to the \(E\) and \(B\) fields. Note that the number of pixels \(N_{\rm pixels}^i\) is different for each needlet window, since we use different pixel resolutions depending on the smallest scale covered by the window. Finally, we apply Equation [\[eq:ilc_reconstruction\]](#eq:ilc_reconstruction){reference-type="ref" reference="eq:ilc_reconstruction"} to obtain an ILC-reconstructed CMB map for window \(i\). The final step is to filter this map in harmonic space for a second time with the \(h_i(\ell)\) window. The final reconstructed CMB map is the sum of these maps for all five needlet windows. In step 2, the reconstructed CMB maps are compressed into power spectra using `NaMaster`, deconvolving to the common beam resolution. We use \(B\)-mode purification as implemented in the software and the mask shown in Figure [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}. We estimate the noise bias \(N_{\ell}\) in the final map by computing the power spectrum of noise-only simulations processed with the needlet weights and windows obtained from the simulated data as described above. \(N_{\ell}\) is averaged over simulations and subtracted from the \(C_{\ell}\) of the reconstructed maps. Finally, in step 3, we run a Monte Carlo Markov Chain (MCMC) over the reconstructed \(BB\) spectrum (we ignore \(EE\) and \(EB\)) with only two free parameters: the tensor-to-scalar ratio \(r\) and the amplitude of the \(BB\) lensing spectrum \(A_{\rm lens}\), using the Python package `emcee`. Both parameters have a top hat prior (between 0 and 2 for \(A_{\rm lens}\), and between \(-0.013\) and infinity for \(r\)). The covariance matrix is calculated directly over 500 simulations with the same setup but with Gaussian foregrounds. As likelihood, we use the same Gaussian likelihood used in pipeline A and restrict the inference to a multipole range \(30 < \ell \leq 300\). While the NILC implementation described above is blind, it can be extended to a semi-blind approach that introduces a certain level of foreground modeling. For example, constrained ILC explicitly nullifies one or more contaminant components in the observed data (such as thermal dust) by including their modeled SED in the variance minimization problem that calculates the weights in Equation [\[eq:nilc_weights\]](#eq:nilc_weights){reference-type="ref" reference="eq:nilc_weights"}. This foreground modeling can be further extended to include the moment expansion of the SED described in Section [2.1](#ssec:pipelines.Cell){reference-type="ref" reference="ssec:pipelines.Cell"}. This method, known as constrained moment ILC, has been shown to be effective for cleaning the large-scale \(B\)-mode contamination for space experiments such as LiteBIRD. While not used in this work, these extensions and others will be considered in future analyses with more complex foregrounds and systematics. ## Pipeline C: map-based cleaning Our third pipeline is a map-based parametric pipeline based on the `fgbuster` code . This approach is based on the following data model: \[\begin{aligned} \centering {\bf d} = \hat{\sf A}{\bf s} + {\bf n} \end{aligned}\] where \({\bf d}\) is a vector containing the polarized frequency maps, \(\hat{\sf A}=\hat{\sf A}(\boldsymbol{\beta})\) is the so-called mixing matrix assumed to be parameterized by a set of spectral indices \(\boldsymbol{\beta}\), \({\bf s}\) is a vector containing the \(Q\) and \(U\) amplitudes of the sky signals (CMB, foregrounds) and \({\bf n}\) is the noise contained in each frequency map. Starting from the input observed data sets \({\bf d}\), Figure [\[fig:schematic_c\]](#fig:schematic_c){reference-type="ref" reference="fig:schematic_c"} shows a schematic of the pipeline, which contains four steps. Step 0 is the preprocessing of input simulations: for each simulation, we combine the simulated noise maps, the foreground and CMB maps and save them on disk. We create a new set of frequency maps, \(\mathbf{\bar{d}}\), smoothed with a common Gaussian kernel of \(100\arcmin\) FWHM. Step 1 is the actual component separation stage. We optimize the spectral likelihood defined as : \[\label{eq:spectral_likelihood} -2\log\left(\mathcal{L}_{\rm spec}(\boldsymbol{\beta})\right) = \left(\hat{\sf A}^T\hat{\sf N}^{-1}\bar{\bf d}\right)^T\left(\hat{\sf A}^T\hat{\sf N}^{-1}\hat{\sf A}\right)^{-1}\left(\hat{\sf A}^T\hat{\sf N}^{-1}\bar{\bf d}\right)\] which uses the common resolution frequency maps, \(\bar{\bf d}\), built during step 0. The right hand side of Equation [\[eq:spectral_likelihood\]](#eq:spectral_likelihood){reference-type="ref" reference="eq:spectral_likelihood"} contains a sum over the observed sky pixels, assumed to have uncorrelated noise--the diagonal noise covariance matrix \(\hat{\sf N}\) is computed from 500 noise-only simulations. Although, in principle, \(\hat{\sf N}\) can be non-diagonal, we do not observe any significant bias of the spectral likelihood due to this approximation in this study. By minimizing Equation [\[eq:spectral_likelihood\]](#eq:spectral_likelihood){reference-type="ref" reference="eq:spectral_likelihood"} we estimate the best-fit spectral indices \(\tilde{\boldsymbol{\beta}}\) and the corresponding mixing matrix \(\tilde{\sf A} \equiv \hat{\sf A}(\tilde{\boldsymbol{\beta}})\). We also estimate the uncertainties on the recovered spectral indices as provided by the minimizer, a truncated-Newton algorithm  as implemented in `scipy` . Having thus obtained estimators of the foreground SEDs, we can recover the sky component maps with the generalized least-square equation \[\label{eq:weights_c} \tilde{\bf s} = \left(\tilde{\sf A}^T\hat{\sf N}^{-1}\tilde{\sf A}\right)^{-1}\tilde{\sf A}^T\hat{\sf N}^{-1}{\bf d} \equiv \hat{\sf W}{\bf d}\;,\] where \({\bf d}\) is the input raw data, and not the common resolution maps. In steps 1 and 2, we have the possibility to use an inhomogeneous noise covariance matrix i.e. \(\hat{\sf N} = \hat{\sf N}(\hat{\bf n})\) and, although this is not exploited in this work, a spatially varying mixing matrix \(\boldsymbol{\beta} = \boldsymbol{\beta}(\hat{\bf n})\). For the latter, one can use the multi-patch or clustering methods implemented in `fgbuster` . Step 2 comprises the calculation of angular power spectra. The recovered CMB polarization map is transformed to harmonic space using `NaMaster`. We estimate an effective transfer function, \({\bf B}^{\rm eff}_\ell=\hat{\sf W}{\bf B}_\ell\) associated with the reconstructed components \(\tilde{\bf s}\), from the channel-specific beams \({\bf B}_\ell\). Correcting for the impact of this effective beam is vital to obtain an unbiased \(BB\) spectrum of the foreground-cleaned CMB, \(\tilde{C}_\ell^{\rm CMB}\). In the second step, we also estimate the noise bias from noise simulations, i.e. \[\tilde{\sf N}_\ell = \frac{1}{\rm N_{sim}}\sum_{\rm sims}\sum_{m=-\ell}^{\ell}\frac{\tilde{\bf n}_{\ell,m} \tilde{\bf n}_{\ell,m'}^\dagger}{2\ell+1}.\] where \(\tilde{\bf n} = \hat{\sf W}{\bf n}^{\rm sim}\) is the noise in the recovered component-separated sky maps. We consider 500 simulations to estimate the noise bias. Step 3 is the cosmological analysis stage. We model the angular power spectrum of the component-separated CMB map, including the noise contribution, as \[C_\ell^{\rm CMB}(r,A_{\rm lens}) \equiv C_\ell^{\rm prim}(r)+C_\ell^{\rm lens}(A_{\rm lens}) + \tilde{N}^{\rm CMB}_\ell \label{eq:cosmo_model}\] and compare data and model with the cosmological likelihood \[\begin{aligned} &-2\log{\mathcal{L}^{\rm cosmo}} \notag \\ & \qquad =\sum_\ell \left(2\ell + 1 \right)f_{\rm sky}\left( \frac{\tilde{C}_\ell^{\rm CMB}}{C_\ell^{\rm CMB}} + \log(C_\ell^{\rm CMB})\right). \label{eq:cosmo_likelihood} \end{aligned}\] It is worth noting that this is only an approximation to the true map-level Gaussian likelihood which approximates the effective number of modes in each multipole after masking and purification as \(f_{\rm sky}(2\ell+1)\), thus neglecting any mode-coupling effects induced by the survey footprint. We grid the likelihood above along the two dimensions \(r\) and \(A_{\rm lens}\). For each simulation we then estimate the maximum-likelihood values and 68% credible intervals from the marginal distributions of \(r\) and \(A_{\rm lens}\). We verified that the distributions of recovered \(\{r,\,A_{\rm lens}\}\) across simulations are well described by a Gaussian, hence supporting the Gaussian likelihood in Equation [\[eq:cosmo_likelihood\]](#eq:cosmo_likelihood){reference-type="ref" reference="eq:cosmo_likelihood"}.\ Pipeline C also offers the option to marginalize over a dust template. The recovered components in \(\tilde{\bf s}\), Equation [\[eq:weights_c\]](#eq:weights_c){reference-type="ref" reference="eq:weights_c"}, include the dust \(Q\) and \(U\) maps which are typically recovered with high signal-to-noise. In the same way that we compute \(\tilde{C}_\ell^{\rm CMB}\) in step 2, we compute the \(BB\) component of the recovered dust map, \(\tilde{C}_\ell^{\rm dust}\). We then update our cosmological likelihood, Equation [\[eq:cosmo_model\]](#eq:cosmo_model){reference-type="ref" reference="eq:cosmo_model"}, by adding a dust term: \[C_\ell^{\rm CMB} = C_\ell^{\rm CMB}(r,\,A_{\rm lens}) + A_{\rm dust}\tilde{C}_\ell^{\rm dust}.\label{eq:cosmo_model_w_dust}\] This is a similar approach to earlier methods . When choosing this approach, the inference of \(r\) during step 3 therefore involves the marginalization over both parameters \(A_{\rm lens}\) and \(A_{\rm dust}\). In principle one could add synchrotron or other terms in Equation [\[eq:cosmo_model_w\_dust\]](#eq:cosmo_model_w_dust){reference-type="ref" reference="eq:cosmo_model_w_dust"} but we limit ourselves to dust as it turns out to be the largest contamination, and, in practice, marginalizing over it allows us to get unbiased estimates of cosmological parameters. In the remainder of this paper, we will will refer to this method as "C + dust marginalization". # Description of input simulations {#sec:sims} Let us start by examining the final constraints on \(r\) obtained by each pipeline applied to 500 simulations. These results are summarized in Figure  [\[fig:r_sigma_r\_compilation\]](#fig:r_sigma_r_compilation){reference-type="ref" reference="fig:r_sigma_r_compilation"} and Table  [\[tab:r_fiducial_results\]](#tab:r_fiducial_results){reference-type="ref" reference="tab:r_fiducial_results"}. Figure  [\[fig:r_sigma_r\_compilation\]](#fig:r_sigma_r_compilation){reference-type="ref" reference="fig:r_sigma_r_compilation"} shows the mean \(r\) and \((16,\,84)\)% credible intervals found by each pipeline as a function of the input foreground model (labels on the \(x\) axis). Results are shown for 5 pipeline setups: pipeline A using the \(C_\ell\)-fiducial model (red), pipeline A using the \(C_\ell\)-moments model (yellow), pipeline B (blue), pipeline C (green) and pipeline C including the marginalization over the dust amplitude parameter (cyan). For each pipeline, we show two points with error bars. The dot markers and smaller error bars correspond to the results found in the best-case instrument scenario (goal noise level, optimistic \(1/f\) component), while the cross markers and larger error bars correspond to the baseline noise level and pessimistic \(1/f\) component. The quantitative results are reported in Table  [\[tab:r_fiducial_results\]](#tab:r_fiducial_results){reference-type="ref" reference="tab:r_fiducial_results"}.\ We will first discuss the nominal pipelines A, B and C without considering any extensions. We find that for the simpler Gaussian and `d0s0` foregrounds, the nominal pipelines obtain unbiased results, as expected. Pipeline B shows a slight positive bias for Gaussian foregrounds, in combination with inhomogeneous noise only. This bias is absent for homogeneous noise and can be traced back to the pixel covariance matrix used to construct the NILC weights. We will discuss this in more detail in Appendix [\[sec:nilc_gaussian_bias\]](#sec:nilc_gaussian_bias){reference-type="ref" reference="sec:nilc_gaussian_bias"}. For now, we show the results using a more constrained mask that is closer to homogeneity. We stress that these results, marked with a \(^{\dagger}\), are not comparable to the rest in Table [\[tab:r_fiducial_results\]](#tab:r_fiducial_results){reference-type="ref" reference="tab:r_fiducial_results"}, since they are calculated on a different mask. The more complex `d1s1` foregrounds lead to a \(\sim1 \sigma\) bias in the goal and optimistic noise scenario. The `dmsm` foregrounds lead to a noticeable increase of the bias of up to \(\sim2\sigma\), seen with pipeline C in all noise scenarios and with pipeline A in the goal-optimistic case, and slightly less with pipeline B. The modifications introduced in the `dmsm` foreground model include a larger spatial variation in the synchrotron spectral index \(\beta_s\) with respect to `d1s1`, and are a plausible reason for the increased bias on \(r\). Remarkably, we find that, in their simplest incarnation, all pipelines achieve comparable statistical uncertainty on \(r\), ranging from \(\sigma(r)\simeq2.1\times10^{-3}\) to \(\sigma(r)\simeq3.6\times10^{-3}\) (a \(70\%\) increase), depending on the noise model. Changing between the goal and baseline white noise levels results in an increase of \(\sigma(r)\) of \(\sim 20-30\%\). Changing between the optimistic and pessimistic \(1/f\) noise has a similar effect on the results from pipelines A and B, although \(\sigma(r)\) does not increase by more than \(10\%\) when changing to pessimistic \(1/f\) noise for pipeline C. These results are in reasonable agreement with the forecasts presented in.\ Now let us have a look at the pipeline extensions, A + moments and C + dust marginalization. Notably, in all noise and foreground scenarios, the two extensions are able to reduce the bias on \(r\) to below \(1\sigma\). For the Gaussian and `d0s0` foregrounds, we consistently observe a small negative bias (at the \(\sim 0.1\sigma\) level for A + moments and \(< 0.5\sigma\) for C + dust marginalization). This bias may be caused by the introduction of extra parameters that are prior dominated, like the dust template's amplitude in the absence of residual dust contamination, or the moment parameters in the absence of varying spectral indices of foregrounds. If those extra parameters are weakly degenerate with the tensor-to-scalar ratio, the marginal \(r\) posterior will shift according to the choice of the prior on the extra parameters. The observed shifts in the tensor-to-scalar ratio and their possible relation with these volume effects will be investigated in a future work. For the more complex `d1s1` and `dmsm`, both pipeline extensions effectively remove the bias observed in the nominal pipelines, achieving a \(\sim 0.5\sigma\) bias and lower. The statistical uncertainty \(\sigma(r)\) increases for both pipeline extensions, although by largely different factors. While C + dust marginalization yields \(\sigma(r)\) between \(3.0\times10^{-3}\) and \(5.9\times10^{-3}\), the loss in precision for A + moments is significantly smaller, with \(\sigma(r)\) varying between \(2.7\times10^{-3}\) and \(4.0\times10^{-3}\) depending on the noise scenario, an average increase of \(\sim25\%\) compared to pipeline A. In any case, within the assumptions made regarding the SO noise properties, it should be possible to detect a primordial \(B\)-mode signal with \(r=0.01\) at the 2--3\(\sigma\) level with no delensing. The impact of other effects, such as time domain filtering or anisotropic noise may affect these forecasts, and will be studied in more detail in the future.\ This analysis was repeated for input CMB maps generated assuming \(r=0\) or \(0.01\) and \(A_{\rm lens}=0.5\) or \(1\). For simplicity, in these cases we considered only the baseline white noise level with optimistic \(1/f\) noise, and the moderately complex `d1s1` foreground model. The results can be found in Figure [\[fig:r_sigma_r\_extra\]](#fig:r_sigma_r_extra){reference-type="ref" reference="fig:r_sigma_r_extra"} and Table [\[tab:r_extra_results\]](#tab:r_extra_results){reference-type="ref" reference="tab:r_extra_results"}. A 50% delensing efficiency results in a reduction in the final \(\sigma(r)\) by 25-30% for pipelines A and B, \(\sim\)`<!-- -->`{=html}10-20% for A + moments, and 0-33% for C + dust marginalization. The presence of primordial \(B\)-modes with a detectable amplitude increases the contribution from cosmic variance to the error budget, with \(\sigma(r)\) growing by up to \(40\%\) if \(r=0.01\), in agreement with theoretical expectations. Using C + dust marginalization, and considering no delensing, we even find \(\sigma(r)\) decreasing, hinting at the possible breaking of the degeneracy between \(r\) and \(A_{\rm dust}\). We conclude that all pipelines are able to detect the \(r=0.01\) signal at the level of \(\sim3\sigma\). As before, we observe a 0.5-1.2\(\sigma\) bias on the recovered \(r\) that is eliminated by both the moment expansion method and the dust marginalization method.\ Finally, we have explored how cosmological constraints and the pipelines' performances are affected by noise inhomogeneity resulting from weighting the noise pixels according to the SO-SAT hits map. The geographical location of SO, and the size of the SAT field of view, place constraints on the possible scanning strategies. In particular, it forces SO to target a patch with a relatively large sky fraction \(f_{\rm sky}\sim0.15\), and surrounded by a \(\sim10\) degree wide boundary with significantly higher noise (see hits map in Figure [\[fig:mask\]](#fig:mask){reference-type="ref" reference="fig:mask"}). The lower panel of Figure [\[fig:r_hom_vs_inhom\]](#fig:r_hom_vs_inhom){reference-type="ref" reference="fig:r_hom_vs_inhom"} shows the ratio between the values of \(\sigma(r)\) found using inhomogeneous noise realizations and those with homogeneous noise in the baseline-optimistic noise model with `d0s0` foregrounds, averaged over 500 simulations. We see that for all pipeline scenarios, \(\sigma(r)\) increases by \(\sim30\%\) due to the noise inhomogeneity. ## Power spectra {#ssec:results.cls} Having presented the constraints on \(r\) found by each pipeline, let us now examine the CMB power spectrum products. Pipelines B and C produce CMB-only maps and base their inference of \(r\) on the resulting power spectra, whereas pipeline A works directly with the cross-frequency power spectra of the original multi-frequency maps. Nevertheless, CMB power spectra are an important data product that every pipeline should be able to provide. Following the methods presented in and, we use a modified version of pipeline A that retrieves CMB-only bandpowers from multi-frequency power spectra, marginalizing over foregrounds with an MCMC sampler as presented in Section [2.1](#ssec:pipelines.Cell){reference-type="ref" reference="ssec:pipelines.Cell"}. Note that this method, originally developed for high-\(\ell\) CMB science, is applicable since we are in the Gaussian likelihood regime. By re-inserting this cleaned CMB spectrum into a Gaussian likelihood with parameters \((r,\,A_{\rm lens})\), we obtain constraints that are consistent with the results shown in Table [\[tab:r_fiducial_results\]](#tab:r_fiducial_results){reference-type="ref" reference="tab:r_fiducial_results"}. Figure [\[fig:cells-cmb\]](#fig:cells-cmb){reference-type="ref" reference="fig:cells-cmb"} shows the CMB power spectra for the three complex foreground simulations `d0s0`, `d1s1` and `dmsm` (upper, middle and lower panel, respectively) while considering the goal-optimistic noise scenario. The various markers with error bars denote the measured CMB power spectra and their \(1\sigma\) standard deviation across 500 simulations, while the black solid line denotes the input CMB power spectrum. Results are shown in gold triangles, blue circles, turquoise diamonds for pipeline A, B and C respectively. The dotted lines show the best-fit CMB model for the three nominal pipelines (using the same color scheme). Only in the `dmsm` foreground scenario, which is the most complex considered here, we also show the results from pipeline C + dust marginalization (dark red squares with error bars) and the best-fit CMB power spectrum from A + moments (pink dot-dashed line) and C + dust marginalization (dark red dot-dashed line). For the nominal pipelines (A, B and C) without extensions, the measured power spectra display a deviation from the input CMB at low multipoles, increasing with rising foreground complexity. For `dmsm` at multipoles \(\lesssim 50\), this bias amounts to about 1.5\(\sigma\) and it goes down to less than 0.5\(\sigma\) at \(80\lesssim \ell\lesssim 250\). The three pipelines agree reasonably well, while pipeline A appears slightly less biased for the lowest multipoles. Pipelines B and C show an additional mild excess of power in their highest multipole bins, with a \(<0.3\sigma\) increase in pipeline C for \(130 \lesssim \ell\lesssim 170\) and up to 1\(\sigma\) for the highest multipole (\(\ell=297\)) in pipeline B. This might indicate power leakage from the multiple operations on map resolutions implemented in pipelines B and C. In pipeline B, these systematics could come from first deconvolving the multi-frequency maps and then convolving them with a common beam in order to bring them to a common resolution, whereas in pipeline C, the leakage is likely due to the linear combination of the multi-resolution frequency maps following Equation [\[eq:weights_c\]](#eq:weights_c){reference-type="eqref" reference="eq:weights_c"}. Other multipole powers lie within the 1\(\sigma\) standard deviation from simulations for all three pipelines. Both extensions, A + moments and C + dust marginalization, lead to an unbiased CMB power spectrum model, as shown by the pink and dark red dot-dashed lines and the square markers in the lower panel of Figure [\[fig:cells-cmb\]](#fig:cells-cmb){reference-type="ref" reference="fig:cells-cmb"}. In the case of pipelines B and C, comparing the best-fit models obtained from the measured power spectra to the input CMB model, we find sub-sigma bias for all bins with \(\ell>100\). We show, however, that the ability to marginalize over additional foreground residuals (e.g. the dust-template marginalization in pipeline C) is able to reduce this bias on all scales, at the cost of increased uncertainties. Implementing this capability in the blind NILC pipeline B would likely allow it to reduce the bias observed in the figure. The SO-SATs are expected to constrain the amplitude of CMB lensing \(B\)-modes to an unprecedented precision. As can be seen from Figure [\[fig:cells-cmb\]](#fig:cells-cmb){reference-type="ref" reference="fig:cells-cmb"}, individual, cleaned CMB bandpowers without delensing at multipoles \(\ell\gtrsim150\) achieve a signal-to-noise ratio of about 10, accounting for a combined precision on the lensing amplitude of \(\sigma(A_{\rm lens})\lesssim0.03\) when considering multipoles up to \(\ell_{\rm max}=300\). This is consistent with the inference results obtained by pipelines A and B. ## Channel weights {#ssec:channel_weights} Our three baseline pipelines differ fundamentally in how they separate the sky components. One common feature among all pipelines is the use of six frequency channels to distinguish components by means of their different SEDs. In Figure [\[fig:channel_weights\]](#fig:channel_weights){reference-type="ref" reference="fig:channel_weights"} we visualize the channel weights as a function of the band center frequency, showing the pipelines in three vertically stacked panels. In the upper panel, we show the effective weights for the CMB applied to the noise-debiased raw power spectra used by pipeline A, distinguishing between weights for each harmonic bin: \[\mathbf{w_{\ell}}^T = \frac{\mathbf{a}^T \hat{\sf C}_{\ell}^{-1} }{\mathbf{a}^T \hat{\sf C}_{\ell}^{-1} \mathbf{a} }.\] Here, \(\hat{\sf C}_{\ell}\) is the \(6\times 6\) matrix of raw cross-frequency power spectra from noisy sky maps, \(\mathbf{a}\) is a vector of length \(6\) filled with ones. This is equivalent to the weights employed by the SMICA component separation method and by ILC as explained in Section [2.2](#sec:nilc){reference-type="ref" reference="sec:nilc"}. The middle panel shows the pixel-averaged NILC weights for the five needlet windows (Figure [\[fig:nilc_windows\]](#fig:nilc_windows){reference-type="ref" reference="fig:nilc_windows"}) used in pipeline B. In the lower panel we show the CMB weights calculated with the map-based component separation, pipeline C, averaged over the observed pixels to yield an array of \(6\) numbers. All channel weights are averaged over 100 simulations containing CMB (\(r=0\), \(A_{\rm lens}=1\)), `d1s1` foregrounds and two noise models: goal white noise with optimistic \(1/f\) noise is shown as dashed lines, whereas baseline white noise with pessimistic \(1/f\) noise is shown as solid lines. Moreover, the gray shaded areas quantify the 1-\(\sigma\) uncertainty region of these weights in the baseline/pessimistic case estimated from \(100\) simulations. We see from Figure [\[fig:channel_weights\]](#fig:channel_weights){reference-type="ref" reference="fig:channel_weights"} that the average channel weights agree well between pipelines A, B and C. Mid-frequency channels at \(93\) and \(145\) GHz are assigned positive CMB weights throughout all pipelines, while high-and low-frequency channels tend to be suppressed owing to a larger dust and synchrotron contamination, respectively. In more detail, the \(280\) GHz channel is given negative weight in all pipelines, while average weights at \(27\), \(39\) and \(225\) GHz are negative with pipeline C and either positive or negative in pipelines A and B, depending on the angular scale. The CMB channel weight tends to consistently increase for pipelines A and B as a function of multipole, a fact well exemplified by NILC at \(225\) GHz, matching the expectation that the CMB lensing signal becomes more important at high \(\ell\). Overall, Figure [\[fig:channel_weights\]](#fig:channel_weights){reference-type="ref" reference="fig:channel_weights"} illustrates that foregrounds at low and high frequencies are consistently subtracted by the three component separation pipelines, with the expected scale dependence in pipelines A and B. Moreover, at every frequency, the channel weights are non-negligible and of similar size across the pipelines, meaning that all channels give a relevant contribution to component separation for all pipelines. # More complex foregrounds: `d10s5` {#sec:d10s5} During the completion of this paper, the new PySM3 Galactic foreground models were made publicly available. In particular, in these new models, templates for polarized thermal dust and synchrotron radiation have been updated including the following changes: 1. Large-scale thermal dust emission is based on the GNILC maps, which present a lower contamination from CIB emission with respect to the `d1` model, based on `Commander` templates. 2. For both thermal dust and synchrotron radiation small scales structures are added by modifying the logarithm of the polarization fraction tensor[^5]. 3. Thermal dust spectral parameters are based on GNILC products, with larger variation of \(\beta_d\) and \(T_d\) parameter at low resolution compared to the `d1` model. Small-scale structure is also added as Gaussian realizations of power-law power spectra. 4. The new template for \(\beta_s\) includes information from the analysis of S-PASS data, in a similar way as the one of the `sm` model adopted in this work. In addition, small-scale structures are present at sub-degree angular scales. These modifications are encoded in the models called `d10` and `s5` in the updated version of PySM. Although these models are still to be considered preliminary, both in terms of their implementation details in PySM[^6] and in general, being based on datasets that may not fully involve the unknown level of foreground complexity, we decided to dedicate an extra section to their analysis. For computational speed, we ran the five pipeline set-ups on a reduced set of 100 simulations containing the new `d10s5` foregrounds template, CMB with a standard cosmology (\(r=0,\, A_{\rm lens}=1\)) and inhomogeneous noise in the goal-optimistic scenario. The resulting marginalized posterior mean and \((16,\,84)\)% credible intervals on \(r\), averaged over 100 simulations, are: \[\begin{aligned} & r = 0.0194 \pm 0.0021 & (\text{pipeline A})\notag \\ & r = 0.0025 \pm 0.0031 & (\text{A + moments})\notag \\ & r = 0.0144 \pm 0.0023 & (\text{pipeline B})\\ & r = 0.0220 \pm 0.0026 & (\text{pipeline C})\notag \\ & r =-0.0015 \pm 0.0051 & (\text{C + dust marg.})\notag \end{aligned}\] Note that the respective bias obtained with pipelines A, B and C are at 9, 6 and 8\(\sigma\), quadrupling the bias of the `dmsm` foreground model. Crucially, this bias is reduced to \(\sim 0.8\sigma\) with the A + moments pipeline, with a \(40\%\) increase in \(\sigma(r)\) compared to pipeline A, and \(\sim0.3\sigma\) with the C + dust-marginalization pipeline, with a \(95\%\) increase in \(\sigma(r)\) compared to pipeline C. This makes A + moments the unbiased method with the lowest statistical error. The \(C_\ell\)-fiducial model achieves minimum \(\chi^2\) values of \(\approx 600\pm 40\). Although this is an increase of \(\Delta\chi^2\sim30\) with respect to the less complex foreground simulations (see Appendix [\[app:chi_squared\]](#app:chi_squared){reference-type="ref" reference="app:chi_squared"}), the associated probability to exceed is \(p\sim0.16\) (assuming our null distribution is a \(\chi^2\) with \(N_{\rm data}-N_{\rm parameters}=567\) degrees of freedom), and therefore it would not be possible to identify the presence of a foreground bias by virtue of the model providing a bad fit to the data. The \(\chi^2\) value we find also confirms that the covariance matrix calculated from Gaussian simulations is still appropriate for the non-Gaussian `d10s5` template. On the other hand, A + moments achieves best-fit \(\chi^2\) values of \(\approx 531 \pm 33\), with a \(< 5\%\) decrease in \(\Delta\chi^2\) with respect to less complex foreground simulations, indicating an improved fitting accuracy. As shown in Figure [\[fig:model_odds_d10s5\]](#fig:model_odds_d10s5){reference-type="ref" reference="fig:model_odds_d10s5"}, the relative model odds between \(C_\ell\)-fiducial and \(C_\ell\)-moments (see Appendix [\[app:chi_squared\]](#app:chi_squared){reference-type="ref" reference="app:chi_squared"} for more details) vary between \(10^{-26}\) and \(10^{-2}\), clearly favoring \(C_\ell\)-moments. Out of 100 `d10s5` simulations, 99 yield model odds below 1% and 78 below \(10^{-5}\). As opposed to the less complex foreground simulations (`d0s0`, `d1s1` and `dmsm`), `d10s5` gives strong preference to using the moment expansion in the power spectrum model. Note that the AIC-based model odds are computed from the differences of \(\chi^2\) values that stem from the same simulation seed, while the \(\chi^2\) analysis only considers the models individually, which explains why the former makes for a much more powerful model comparison test. These results consider only the most optimistic noise scenario. Other cases would likely lead to larger uncertainty and, as a consequence, lower relative biases. In this regard, it is highly encouraging to see two pipeline extensions being able to robustly separate the cosmological signal from Galactic synchrotron and dust emission with this high-level complexity. This highlights the importance of accounting for and marginalizing over residual foreground contamination due to frequency decorrelation for the level of sensitivity that SO and other next-generation observatories will achieve. The contrast between the results obtained on the `dmsm` and `d10s5` simulations gives us an opportunity to reflect on the strategy one should follow when determining the fiducial component separation method to use in primordial \(B\)-mode searches. Although the `dmsm` model leads to a \(\sim2\sigma\) bias on \(r\) under the simplest component separation algorithms, simple model-selection metrics are not able to provide significant evidence that a more sophisticated modeling of foregrounds is needed. The situation changes with `d10s5`. A conservative approach is therefore to select the level of complexity needed for component separation by ensuring that unbiased constraints are obtained for all existing foreground models consistent with currently available data. The analysis methods passing this test can then form the basis for the fiducial \(B\)-mode constraints. Alternative results can then be obtained with less conservative component separation techniques, but their goodness of fit (or any similar model selection metric) should be compared with that of the fiducial methods. These results should also be accompanied by a comprehensive set of robustness tests able to identify signatures of foreground contamination in the data. This will form the basis of a future work. In a follow-up paper, we will also explore the new set of complex PySM3 foreground templates in more detail. # Conclusions {#sec:conc} In this paper, we have presented three different component separation pipelines designed to place constraints on the amplitude of cosmological \(B\)-modes on polarized maps of the SO Small Aperture Telescopes. The pipelines are based on multi-frequency \(C_\ell\) parametric cleaning (Pipeline A), blind Needlet ILC cleaning (Pipeline B), and map-based parametric cleaning (Pipeline C). We have also studied extensions of pipelines A and C that marginalize over additional residual foreground contamination, using a moment expansion or a dust power spectrum template, respectively. We have tested and compared their performance on a set of simulated maps containing lensing \(B\)-modes with different scenarios of instrumental noise and Galactic foreground complexity. The presence of additional instrumental complexity, such as time-domain filtering, or anisotropic noise, are likely to affect our results. The impact of these effects will be more thoroughly studied in future work. We find the inferred uncertainty on the tensor-to-scalar ratio \(\sigma(r)\) to be compatible between the three pipelines. While the simpler foreground scenarios (Gaussian, `d0s0`) do not bias \(r\), spectral index variations can cause an increased bias of 1--2\(\sigma\) if left untreated, as seen with more complex foreground scenarios (`d1s1`, `dmsm`). Modeling and marginalizing over the spectral residuals is vital to obtain unbiased \(B\)-mode estimates. The extensions to pipelines A and C are thus able to yield unbiased estimates on all foreground scenarios, albeit with a respective increase in \(\sigma(r)\) by \(\sim 20\%\) (A + moments) and \(>30\%\) (C + dust marginalization). These results are in good agreement with the forecasts presented in. After testing on simulations with an \(r=0.01\) cosmology, we conclude that under realistic conditions and if the forecasted map-noise levels and characteristics are achieved, SO should be able to detect a \(r=0.01\) signal at \(\sim2\)--\(3\sigma\) after 5 years of observation. Inhomogeneous noise from the SAT map-making scanning strategy brings about \(30\%\) increase in \(\sigma(r)\) as compared to homogeneous noise. Analyzing the per-channel weights for our pipelines, we find all frequency channels to be relevant for the CMB signal extraction and all pipelines to be in good agreement. These forecasts cover the nominal SO survey, and can be considered pessimistic in the light of prospective additional SATs that will further improve the sensitivity on large angular scales. We have also carried out a preliminary analysis of new, more complex, foreground models recently implemented in PySM3, in particular the `d10s5` foreground template. The much higher level of spatial SED variation allowed by this model leads to a drastic increase in the bias on \(r\) by up to \(\sim9\sigma\), when analyzed with the nominal pipelines A, B and C. Fortunately, this bias can be reduced to below \(1\sigma\) when using A + moments and c + dust marginalization. These extensions lead to a 40% and 95% degradation of the error bars, respectively. Our results highlight the importance of marginalizing over residuals caused by frequency decorrelation for SO-like sensitivities. Although we have not analyzed `d10s5` to the same depth as the other foreground models presented here, it is encouraging to confirm that we have the tools at hand to obtain robust, unbiased constraints on the tensor-to-scalar ratio in the presence of such complex Galactic foregrounds. In preparation for the data collected by SO in the near future, we will continue our investigations into Galactic foreground models with other levels of complexity as the field progresses. Nevertheless, the current work has shown that the analysis pipelines in place for SO are able to obtain robust constraints on the amplitude of primordial \(B\) modes in the presence of Galactic foregrounds covering the full range of complexity envisaged by current, state-of-the-art models. [^1]: We note that the pipelines are still agnostic to some aspects of the instrumental noise such as filtering, which may impact the overall forecasted scientific performance. We anticipate studying these in detail in future work. [^2]: See [github.com/simonsobs/BBPower](https://github.com/simonsobs/BBPower). [^3]: See for more details on the moment-expansion formalism in the context of CMB foregrounds, and as an alternative power-spectrum-based description. [^4]: The domain sizes are estimated directly from the needlet window scale. The ILC bias can be minimized by enlarging the pixel domains to be big enough to include a higher number of modes. We choose the resulting ILC bias to not exceed \(0.2\%\), for which we need pixel domain sizes large enough so that each needlet window contains at least 2500 modes. [^5]: See [pysm3.readthedocs.io/en/latest/preprocess-templates](https://pysm3.readthedocs.io/en/latest/preprocess-templates/small_scale_dust_pysm3.html). [^6]: The PySM library is currently under development, with beta versions including minor modifications of the foreground templates being realized regularly. In this part of our analysis we make use of PySM v3.4.0b3.

{'timestamp': '2023-02-10T02:00:14', 'yymm': '2302', 'arxiv_id': '2302.04276', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04276'}

{'timestamp': '2023-03-02T02:21:07', 'yymm': '2302', 'arxiv_id': '2302.04261', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04261'}

# Model The system we are considering can be conceptually split into three regions: two lateral (\(\mathrm{L}\) and \(\mathrm{R}\)), and a central region (\(\mathrm{C}\)), see Fig. [\[fig:sketch\]](#fig:sketch){reference-type="ref" reference="fig:sketch"}(a). The superconductor primarily induces a superconducting pairing potential \(\Delta(x)\) in the semiconductor through the proximity effect. It also contributes to the electrostatic potential landscape \(V(x) =-\mu(x)\). The ferromagnetic insulator induces an exchange field \(\vb{h}(x)\) in both the superconductor and the semiconductor. Given that both the exchange field and the pairing potential are induced in a semiconductor with controllable charge carrier density, there is not a fixed hierarchy of energy scales, and, in principle, all regimes can be realized in the system. We consider that the exchange field is sufficiently weak in the superconductor such that superconductivity persists. This condition is relaxed in the semiconductor, where the induced exchange field can overcome the induced pairing potential. Therefore, ferromagnetic hybrid junctions allow exploring the parameter space beyond the conventional regime (\(\mu \gg \Delta > |\vb{h}|\)). Note that we do not refer to a particular arrangement of the layers in the lateral region, as different combinations of interfaces, for instance, tripartite arrangement  and tunneling arrangement , allows induction of both superconducting pairing and exchange field in the semiconductor. The system Hamiltonian is \(H=\frac{1}{2}\int \psi^\dag \mathcal{H} \psi\), where the Bogoliubov-de Gennes (BdG) Hamiltonian \(\mathcal{H}\) in the Nambu spinor basis \(\psi^\dag = \begin{pmatrix} \psi_\uparrow^\dag & \psi_\downarrow^\dag &-\psi_\downarrow & \psi_\uparrow \end{pmatrix}\) is \[\begin{split} \mathcal{H} = \qty[\frac{\hbar^2 k_x}{2 m^*}-\mu]\tau_z + \vb{h}\cdot \vb*{\sigma} + \Delta \tau_+ + \Delta^\dag \tau_-+ \mathcal{H}_\mathrm{SOC}\. \end{split}\] Here, \(k_x =-i\partial_x\) is the momentum operator (we consider a single mode in the junction), \(m^*\) is the effective mass, and \(\sigma_j\) and \(\tau_j\) are the Pauli matrices in the spin and particle-hole space, respectively. The spin-orbit coupling Hamiltonian \(\mathcal{H}_\mathrm{SOC}\) is given in Eq. [\[eq:soc_hamiltonian\]](#eq:soc_hamiltonian){reference-type="eqref" reference="eq:soc_hamiltonian"} and discussed in Sec. [2.4](#sec:soc){reference-type="ref" reference="sec:soc"}. The proximity-induced exchange field \(\vb{h}\) is due to the coupling to the ferromagnetic insulator and, in principle, can be spatially inhomogeneous due to the micromagnetic configuration. The magnitude of the \(\vb{h}\) field can also vary due to a nonuniform coupling strength. Moreover, recent theoretical investigations of ferromagnetic InAs-Al-EuS nanowires showed that the electrostatic environment is crucial in modulating the effect of the EuS on the InAs, suggesting that, in principle, it is possible to tune the induced exchange field electrostatically. In this work, we consider that the exchange field \(\vb{h}\) takes a constant value \(\vb{h}_j\) in each of the three regions \(i\in \{\mathrm{L}, \mathrm{C}, \mathrm{R}\}\). We will call collectively \(\mathrm{L}\) and \(\mathrm{R}\) lateral regions, with an exchange field \(\vb{h}_\mathrm{L}=\vb{h}_\mathrm{R}=\vb{h}_\mathrm{lat}\) while we will use the symbol \(\vb{h}_\mathrm{all}\) when considering a homogeneous value for the exchange field. We use the same notation for the chemical potential \(\mu\) of the three regions. In addition, we introduce potential barriers with height \(V_\mathrm{B}\) at the interfaces between the central and the lateral regions, tuning the system from the open (\(V_\mathrm{B} = 0\)) to the quantum dot regimes (\(V_\mathrm{B} \gg-\mu_\mathrm{lat}\)). The induced pairing potential is \(\Delta_j = \Delta_{0, j} e^{i \phi_j}\) with the modulus \(\Delta_{0, j}\) taking finite value only in the \(\mathrm{L}\) and \(\mathrm{R}\) regions while the superconducting phase difference between the two leads is \(\phi = \phi_\mathrm{R}-\phi_\mathrm{L}\). In all the simulations we use realistic parameters for InAs-Al-EuS heterostructures, taking \(\Delta_0 = \SI{0.250}{\milli\eV}\) and effective mass \(m^* = 0.026\,m_e\). We consider a nanowire of total length including both the lateral and central regions of \(\ell_\mathrm{W} = 3~\mu\)m. To obtain numerical results, we discretize the Hamiltonian using a finite-differences scheme with a lattice spacing of \(a = \SI{2}{\nano\meter}\) implemented using the `KWANT` package. The code and the results of the simulations are available at . We focus on the short junction limit, and we consider a central region of length \(\ell_\mathrm{C}=\SI{180}{\nano\meter}\), such that the quantization energy is comparable to the other energy scales. We fix the chemical potential in the lateral regions to be \(\mu_\mathrm{L} = \mu_\mathrm{R} = 16~\Delta = \SI{4}{\milli\eV}\). Directly solving for the quasiparticle spectrum of the continuum model allows treating on equal footing the current carried by Andreev bound states (ABSs) and the quasi-continuum of states above the gap. Usually, the continuum current is subdominant, except for strong exchange fields, where its contribution is comparable to or even larger than the ABS one . The supercurrent in a Josephson junction is an equilibrium phenomenon that can be described by a function \(E_\mathrm{J}(\phi)\), called Josephson potential or phase dispersion relation. This can be evaluated from the quasiparticle spectrum, assuming that the quasiparticles are in thermal equilibrium and calculating the free energy at a fixed phase \[\begin{split} E_\mathrm{J}\qty(\phi) &=-k_\mathrm{B} T_\mathrm{p} \ln \tr e^{-\frac{H}{k_\mathrm{B} T_\mathrm{p}}} \\ &=-k_\mathrm{B} T_\mathrm{p} \sum_n \ln \qty[2 \cosh \qty(\frac{\omega_n\qty(\phi)}{2 k_\mathrm{B} T_\mathrm{p}})]\,, \end{split} \label{eq:free_energy}\] where \(\omega_n\) is the quasiparticle spectrum, \(k_\mathrm{B}\) is the Boltzmann constant and \(T_\mathrm{p}\) is the quasiparticle temperature . From \(E_\mathrm{J}\), the CPR is calculated through the thermodynamic relation \[\langle I \rangle = \frac{2 e}{\hbar} \pdv{E_\mathrm{J}(\phi)}{\phi}\. \label{eq:current}\] The maximum current that can flow in the junction in equilibrium is called critical current, \(I_\mathrm{c}\), while we define the critical phase, \(\phi_\mathrm{c}\), as the phase difference where this is reached \[I_\mathrm{c} = |I (\phi_\mathrm{c})|,\qquad \phi_\mathrm{c} \equiv \argmax_{\phi\in[0, \pi]} \abs{I\qty(\phi)}\,,\] where we restricted the search domain to \([0, \pi]\) in the reciprocal case. We define the quantity \(I_0 \equiv2 e \Delta/\hbar\) as the relevant current scale that has the numerical value \(I_0 = \SI{122}{\nano\ampere}\) for Al. The sign of \(I(\phi_\mathrm{c})\) defines the direction of the supercurrent at the critical phase. It is useful to decompose the phase dispersion in its harmonic components \[E_\mathrm{J}(\phi) = \sum_{k=1}^{\infty} \qty[C_k \cos(k \phi) + S_k \sin(k \phi)]\,, \label{eq:harmonic_decomposition}\] since each harmonic corresponds to the tunneling of multiplets of Cooper pairs between the two superconducting regions. To see this, it is necessary to interpret the phase in Eq. [\[eq:harmonic_decomposition\]](#eq:harmonic_decomposition){reference-type="eqref" reference="eq:harmonic_decomposition"} as an operator and rewrite the expression in the charge basis by interpreting \(\exp(i k \phi)\) as a translation operator. Neglecting the constant term, the result is \[\mathcal{H}_{J} = \sum_n \sum_k \frac{C_k + i S_k}{2} \ket{n+k}\bra{n} + \mathrm{H.c.}\] where \(\ket{n}\) is the state with a difference of \(n\) Cooper pairs in the two leads. In this way, it is easy to see how the terms \(\cos(k \phi)\) mediate the tunneling of \(k\) Cooper pairs. A purely \(\cos(2\phi)\) Josephson junction can be used to create a qubit with nearly-degenerate ground states that are a superposition of only states with the same parity of Cooper pair number. . In this setup, \(\cos(\phi)\) and \(\sin(\phi)\) perturbations can be used to implement rotation in the Bloch sphere . If time-reversal and inversion symmetries are not simultaneously broken, the Josephson junction is reciprocal and \(E_\mathrm{J}(\phi) = E_\mathrm{J}(-\phi)\). It subsequently results in the absence of the \({S_k}\) components, called anomalous. These anomalous components are necessary for \(\phi_0\) junctions and the diode effects. In the reciprocal case, assuming all the components \(\{C_k\}\) with \(k>2\) are negligible, the only minima of the Josephson potential are located at \(\phi = 0\) and \(\pi\) only if \(|C_{1}| \geq 4 C_{2}\). For \(|C_{1}| < 4 C_{2}\), minima can be found at \(\phi = \pm \arctan (C_1/\sqrt{16 C_2^2-C_1^2})\). This case, dubbed \(\pm \phi_0\)-junction, does not require an inversion-symmetry breaking mechanism . ## Single level model {#sec:single_level} Before proceeding to the numerical results, we introduce a single-level model in which a level with energy \(\varepsilon\) couples to two high carrier density superconducting lateral regions, see App. [\[sec:app:gfmodel\]](#sec:app:gfmodel){reference-type="ref" reference="sec:app:gfmodel"} and Ref. . To get an expression for the critical lines separating the different phases, we consider the condition where an ABS crosses the Fermi level, obtained from Eq. [\[eq:app:abs\]](#eq:app:abs){reference-type="ref" reference="eq:app:abs"} for \(\omega=0\). When a spin-split ABS crosses the Fermi level, its occupation changes and provides no contribution to the current. Therefore, the residual current is entirely due to the continuum of states and has a characteristic \(\pi\) contribution . In the case of equal exchange fields in the leads, \(h_\mathrm{L} = h_\mathrm{R} = h_\mathrm{lat}\), the Fermi level crossing condition for ABS with spin \(\sigma=\pm1\) is given by the expression \[\begin{split} \frac{\sigma h_\mathrm{C}}{\gamma} + &\frac{\sigma h_\mathrm{lat}}{\sqrt{\Delta^2-h_\mathrm{lat}^2}} = \\ \pm &\sqrt{\frac{\varepsilon^2}{\gamma^2} + \frac{\Delta^2}{\Delta^2-h_\mathrm{lat}^2}[ 1-T \sin^2(\phi/2)]}\,, \end{split} \label{eq:zero_crossings}\] where \(\gamma=\gamma_\mathrm{L}+\gamma_\mathrm{R}\) is the tunnel rate to the leads and \(T= 4 \gamma_\mathrm{L} \gamma_\mathrm{R}/(\gamma_\mathrm{L}+\gamma_\mathrm{R})^2\) is the transparency of the junction. When a spin-split ABS crosses the Fermi level at \(\phi=\pi\), a metastable \(\pi\) phase appears, marking the transition from \(0\) to \(0'\). When such a crossing happens for \(\phi=0\), the \(0\) phase becomes completely unstable, marking the \(\pi'\) to \(\pi\) transition. Finally, the \(0'\) to \(\pi'\) critical line can be approximated by \(\phi=\pi/2\). The quantum point contact limit, in which the intermediate state in the junction is strongly hybridized with the states in the leads, can be obtained from Eq. [\[eq:app:abs\]](#eq:app:abs){reference-type="ref" reference="eq:app:abs"} for \(\gamma\to\infty\). This results in a generalization of Beenakker's formula  for spin-split leads \[\omega = \pm \Delta \sqrt{1-T \sin^2(\phi/2)}-\sigma h_\mathrm{lat}\.\] # Controllable CPR ## Quantum dot regime {#sec:quantum_dot} The quantum dot regime is reached when large barriers at the edge of the central region are introduced, see Fig. [\[fig:sketch\]](#fig:sketch){reference-type="ref" reference="fig:sketch"}. In this regime, electrons are confined in the central regio The quantum dot regime is optimal for the electrostatic controllability of the 0 -- \(\pi\) transition. When the quantum dot levels align with the chemical potential in the leads, the \(\pi\) phase appears at low exchange fields, and the critical line takes the form \(h_\mathrm{C} \propto \varepsilon\), see Fig. [\[fig:dot-phases\]](#fig:dot-phases){reference-type="ref" reference="fig:dot-phases"}(c). This work does not consider the electrostatic repulsion in the central region (Coulomb blockade). We expect the mean-field picture to be valid for large exchange fields, and the Coulomb repulsion to enhance the exchange field in the central region. In this regime, \(I_\mathrm{c}\) is maximal when the dot levels align with the chemical potential of the leads, as shown in Fig. [\[fig:dot-phases\]](#fig:dot-phases){reference-type="ref" reference="fig:dot-phases"}(b). In the off-resonance condition, the critical lines converge to the spectral gap closing point, \(h=\Delta\). The phase diagram can be understood using the single-level model, which predicts the hyperbolic critical lines \[see the dashed lines in Fig. [\[fig:dot-phases\]](#fig:dot-phases){reference-type="ref" reference="fig:dot-phases"}(a)\]. The Josephson potential and CPRs on resonance are shown in Figs. [\[fig:dot-phases\]](#fig:dot-phases){reference-type="ref" reference="fig:dot-phases"}(c) and (d). The fundamental harmonic dominates the junction properties in both the 0 and the \(\pi\) phases (top and bottom panels). In contrast, high-harmonic contributions become important in the \(0'\) and \(\pi'\) phases because of the double minima Josephson potential. To better understand the harmonic composition of the Josephson energy, we present the lowest harmonic components of the CPR in Fig. [\[fig:dot\]](#fig:dot){reference-type="ref" reference="fig:dot"}. For a small value of the exchange field, in the \(0\) and \(0'\) phases, the harmonic component coefficients show a peak corresponding to an energy level in the quantum dot aligning with the Fermi level of the lateral regions. These peaks have widths that decrease for higher-order components, allowing the relative strength of the first two harmonics, \(\delta C_{21} = |C_2|-|C_1|\), to be tuned by slightly changing the energy of the quantum dot levels electrostatically. In contrast, the sensitivity of \(E_\mathrm{J}\) to the chemical is almost negligible in the \(\pi\) phase and the suppression of \(C_2\) is significant, resulting in a sinusoidal CPR, as shown in Fig. [\[fig:dot-phases\]](#fig:dot-phases){reference-type="ref" reference="fig:dot-phases"}(d). When the system is tuned to the vicinity of the \(0'\)--\(\pi'\) transition, the fundamental harmonic is suppressed, leading to a regime dominated by the second harmonic and a double-well Josephson potential. ## Open regime In contrast to the quantum dot regime, the open regime shows a weak dependence on chemical potential, except close to the edge of the band (\(\mu=0\)), Fig. [\[fig:open\]](#fig:open){reference-type="ref" reference="fig:open"} (a). The open regime shows a \(\pi\) phase for \(h_{\rm all}>\Delta\), as predicted by the analytic expression in Eq. [\[eq:zero_crossings\]](#eq:zero_crossings){reference-type="eqref" reference="eq:zero_crossings"}. The system shows extended metastable \(0'\) and \(\pi'\) regions compared to the quantum dot regime. In the open regime, the transition happens for \(h_\mathrm{all}=\Delta/\sqrt{1-T/2}\). For transparent junctions, this critical line coincides with the zero-temperature paramagnetic limit for superconductors meaning that this regime cannot be achieved in materials with intrinsic superconductivity. Instead, semiconductor-superconductor devices are ideal for reaching the \(\pi'\) and \(\pi\) regimes. At the \(0'\)--\(\pi'\) transition point, the CPR is dominated by the \(\sin(2 \phi)\) term, leading to a Josephson potential with two equivalent minima within the \(\phi\in[0,\pi]\) range. We note that the robustness against local fluctuations in \(\mu_\mathrm{C}\) is a unique feature of the open regime. The other two transitions lines, for \(0\)--\(0'\) and \(\pi\)--\(\pi'\), are also almost independent of the chemical potential once \(\mu_C\gtrsim10\), taking place close to \(h_\mathrm{all}=0\) and \(h_\mathrm{all}=\Delta\sqrt{1-T}\). In both open and dot cases, when \(C_1\) reaches the crossover point and becomes zero, \(C_2\) has a negative sign. This satisfies the condition \(|C_1| > 4C_2\) discussed before, resulting in minima at \(0\) and \(\pi\). For what concerns temperature dependence, since the higher harmonic components are dictated mainly by the lowest ABS, the main effect of increasing temperatures is a reduction of the metastable \(0'\) and \(\pi'\) regions. The open and dot regimes have different advantages and disadvantages for practical applications of ferromagnetic hybrid junctions as a \(\cos(2\phi)\) Josephson element. The open regime is insensitive to noise in the gate voltage but requires a relatively high exchange field \(h_\mathrm{all} \simeq \Delta / \sqrt{2}\) for the second harmonic to dominate. Increasing the barriers, and thus moving toward the dot regime, lowers the required exchange field toward the theoretical limit \(h_\mathrm{all}\simeq0\) at the price of a higher sensitivity to gate noise. It also allows electrostatic control of the harmonic content. Indeed, when the \(C_2\) component approaches its maximum location, \(C_1\) vanishes linearly. This opens up the possibility of introducing a gate-controllable \(\cos(\phi)\) component. Additionally, the ability to change from a dot to an open regime is controlled by the electrostatic environment, allowing to tune the system between the two regimes. ## Inhomogeneous exchange field The exchange field in the lateral and central regions affects CPR differently. To reveal this difference, we now analyze the case where the exchange field in the central (\(h_\mathrm{C}\)) and lateral regions (\(h_\mathrm{R}=h_\mathrm{L}=h_\mathrm{lat}\)) have different values while being still aligned in the same direction, see Fig. [\[fig:inhomogenous_h\_and_soc\]](#fig:inhomogenous_h_and_soc){reference-type="ref" reference="fig:inhomogenous_h_and_soc"}(a) and (b). In the case of small magnetization in the lateral regions (\(h_\mathrm{lat}\sim0\)), we find that a strong polarization in the central one \(h_\mathrm{C}\gg\Delta\) is needed to induce the transition to the \(\pi\) state in the open regime. The exchange field strength necessary to induce a \(0\) -- \(\pi\) transition crucially depends on other parameters of the system. In particular, longer junctions and low density in the central region are associated with transitions at lower fields. The length dependence can be understood using a semiclassical model, where \(h_\mathrm{C}\) adds an extra phase accumulated by quasiparticles in a round-trip between the leads. This phase is proportional to \(h_\mathrm{C} \,\ell_\mathrm{C}\) product, explaining why longer junctions exhibit switches from the 0 to the \(\pi\) phase at lower exchange field values. In addition, it leads to a periodic pattern of \(0\) and \(\pi\) phases along the \(h_\mathrm{C}\) axis. A similar effect can be obtained by reducing \(\mu_\mathrm{C}\), which reduces the Fermi velocity. This is further discussed in App. [\[sec:app:semiclassical\]](#sec:app:semiclassical){reference-type="ref" reference="sec:app:semiclassical"}. A sharp transition from \(0\) to \(\pi\) can also be obtained near gap closing (\(\abs{h_\mathrm{lat}}=\Delta)\). In this case, the transition is associated with a strong reduction in the magnitude of \(I_\mathrm{c}\). This behavior can be understood using the simplified one-level model in Eq. [\[eq:zero_crossings\]](#eq:zero_crossings){reference-type="eqref" reference="eq:zero_crossings"}, which explains the transition to the \(\pi\) phase as the disappearance of the contribution of the lowest ABS, thus reducing the total supercurrent in the junction. The fact that the 0 -- \(\pi\) transition is associated with a decrease of critical current only in the case of gap closing can be potentially used to infer the dominant mechanism in experiments. ## Spin-orbit coupling in the semiconductor {#sec:soc} So far, we have neglected the effect of spin-obit coupling. Here, we consider a simple model for linear spin-orbit coupling that, for a quasi-1D system, takes the form \[\mathcal{H}_\mathrm{SOC} = k_x \qty[\alpha_z \sigma_y + \beta \sigma_x] \tau_z = k_x [\vb*{\kappa} \cdot \sigma] \tau_z \,, \label{eq:soc_hamiltonian}\] where we defined a spin-orbit coupling vector \(\vb*{\kappa}= (\beta_x, \alpha_z, 0)\). In the simplest setup, \(\alpha_z\) arises from the Rashba field and \(\beta\) from the Dresselhaus term. We note that the distinction between the two terms is artificial in a one-dimensional model, as the two terms can be mapped onto each other by a unitary transformation \[U = \exp(-i \theta/2 \sigma_z \tau_0)\,,\] that is a rotation in spin space around the \(z\) axis by an angle \(\theta\), potentially inhomogeneous in space. Using \(\theta=\arctan(\beta/\alpha_z)\), we can always remove the term proportional to \(\sigma_x\) and align the spin-orbit vector in the \(y\) direction. This transformation, however, also rotates the exchange field. This result illustrates the equivalence between inhomogeneous spin-orbit fields and exchange fields. We first focus on the homogeneous spin-orbit situation, displayed in Fig. [\[fig:inhomogenous_h\_and_soc\]](#fig:inhomogenous_h_and_soc){reference-type="ref" reference="fig:inhomogenous_h_and_soc"} (c) and (d). Although spin-orbit coupling splits the Fermi surface, Cooper pairs do not acquire a finite momentum unless time-reversal symmetry is broken. Thus, the oscillation between triplet and singlet components is absent, and it is impossible to obtain a \(\pi\) phase. At a finite magnetic field, the Rashba term couples the two spin-split ABSs reopening the gap, unless the field aligns with the spin-orbit vector \(\vb*{\kappa}\). The effect on the CPR of a transverse Rashba field is a substantial reduction of the \(\pi\) regions and an enlargement of the metastable phases. When the exchange field \(\vb{h}\) is instead aligned with \(\vb*{\kappa}\), the spin-rotation symmetry is unbroken. This allows for \(\pi\) phases, but simultaneously, the system remains gapless for \(h>\Delta\). Anomalous Josephson effects can occur when a spin-orbit coupling vector is aligned with a magnetic field . However, it is not observed in the homogeneous case as the combination of various spin-rotation symmetry-breaking effects is necessary for its manifestation . This can occur, for instance, due to finite spin-splitting with a non-zero component in both the junction direction and the transverse one , or multiple modes that can hybridize . The presence of anomalous currents in similar systems has been considered in Refs. . The spin-orbit field depends on the local electrostatic environment. For this reason, the spin-orbit direction can have different magnitudes and directions in different regions of the ferromagnetic hybrid junction. We consider this situation in Fig. [\[fig:aje\]](#fig:aje){reference-type="ref" reference="fig:aje"}. The spin-orbit direction is misaligned in the central region by an angle \(\theta_\mathrm{C}\) with respect to the lateral ones, while we consider a homogeneous exchange field. Since the Rashba field is proportional to the electric field, this scenario might appear in Josephson junctions due to a varying electrostatic environment. This is equivalent to a homogeneous spin-orbit coupling field and a misaligned exchange field in the three regions. In this case, the CPR shows a non-reciprocal behavior, \(I(\phi)\neq I(-\phi)\), due to the presence of anomalous \(\sin(k \phi)\) terms in the Josephson potential. The non-reciprocal supercurrent has been recently reported in superconductor-semiconductor nanowires . The critical phase in the non-reciprocal case reads as \(\phi_\mathrm{c} \equiv \argmax_{[0, 2\pi)} \abs{I}\). To measure the non-reciprocal behavior, we define the critical current in the two directions, \(I_\mathrm{c}^+\equiv\max I\) and \(I_\mathrm{c}^-\equiv-\min I\), and the corresponding diode efficiency as \(\eta \equiv (I_\mathrm{c}^+-I_\mathrm{c}^-)/(I_\mathrm{c}^+ + I_\mathrm{c}^-)\). For the parameters considered, the efficiency can be as high as \(30\%\) in the region close to the 0 -- \(\pi\) transition. For this point, the CPR shows a characteristic form \(I(\phi) \sim \sin(\phi-\phi_0) + \cos(2\phi)\). The rectification effect exhibits a non-monotonic temperature dependence \(\eta(T_p)\) that shows a maximum at finite temperature. This can be explained by the ABS spectrum, see Fig. [\[fig:aje\]](#fig:aje){reference-type="ref" reference="fig:aje"} (d): the lowest state is dominated by the \(\cos(k\phi)\) contribution, with a weak non-reciprocal behavior. The first excited state has a predominant \(\sin(\phi)\) contribution. Therefore, increasing the temperature increases the non-reciprocal supercurrent contribution through an interference process of the different CPR harmonics . At higher temperatures, more states become populated, washing out the contribution from the harmonics and leading to a sinusoidal CPR, Fig. [\[fig:aje\]](#fig:aje){reference-type="ref" reference="fig:aje"}(e). # Conclusions In this work, we have analyzed the higher harmonics in the current-phase relation appearing in ferromagnetic hybrid junctions. The induced exchange field in the semiconductor can overcome the induced pairing potential without causing a transition to the normal state. This results in spin-polarized Andreev bound states with opposite spin crossing the Fermi level, leading to the 0 -- \(\pi\) phase transition and supercurrent reversal. Higher harmonics in the current-phase relation become dominant close to the 0 -- \(\pi\) transition, where the supercurrent changes sign. In the dot regime (weak coupling to the leads), the \(\pi\) depends on the relative position of the dot levels with respect to the leads chemical potential. In case the junction is tuned into the open regime (large coupling to the leads), the onset of the \(\pi\) phase is less sensitive to changes in chemical potential. We find that the spin-orbit coupling increases the coexisting region between \(0\) and \(\pi\) phases with considerable amplitudes of higher-harmonic components. Finally, we find that non-collinear spin-orbit coupling in the junction, due to a varying electrostatic environment, results in a supercurrent rectification effect whose efficiency peaks around 0 -- \(\pi\) transition. The tunability of the harmonic content of CPRs is relevant for a number of applications, including superconducting diodes , ferromagnetic transmon qubits , and parity-protected qubits . In this context, ferromagnetic junctions in the open regime are promising thanks to their robustness to fluctuations in the charge environment.

{'timestamp': '2023-02-09T02:18:05', 'yymm': '2302', 'arxiv_id': '2302.04267', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04267'}

# Background in Diffusion Models Ho et al. proposed a specific parameterization of the diffusion model to simplify the training process using a score matching-like loss that minimizes the mean-squared error between the true noise and the predicted noise. They also note that the sampling process can be interpreted as equivalent to Langevian dynamics, which allows them to relate the proposed denoising diffusion probabilistic models (DDPM) to score-based methods in. Here, we introduce the forward process, the reverse process, and training of DDPM. ## Forward Process Given a data distribution \(\mathbf{x}_0 \sim q(\mathbf{x}_0)\), DDPM will gradually add noise to the original data distribution until it loses all the original information and becomes an entirely noisy distribution (as shown by the \(x_T\) sample in Figure [\[fig:medDIFF\]](#fig:medDIFF){reference-type="ref" reference="fig:medDIFF"}). To be more specific, we will convolve \(q(\mathbf{x}_0)\) with an isotropic Gaussian noise \(\mathcal{N}(0,\sigma_{i}^{2} \mathbf{I})\) in T steps to produce a noise corrupted sequence \(\bx_1,\bx_2,\dots,\bx_T\). \(\mathbf{x}_T\) will converge to an isotropic Gaussian distribution as \(T\rightarrow \infty\). Under the variance schedule that produces \(\bx_1,\bx_2,\dots,\bx_T\), the distribution is derived as \[\begin{aligned} q(\mathbf{x}_{1:T}|\mathbf{x}_0) &= \prod_{t=1}^{T}q(\mathbf{x}_{t}|\mathbf{x}_{t-1}) ,\\ q(\mathbf{x}_{t}|\mathbf{x}_{t-1}) &= \mathcal{N}(\mathbf{x}_t;\sqrt{1-\beta_t}\mathbf{x}_{t-1},\beta_{t}\mathbf{I}) \end{aligned}\] Using the reparameterization trick introduced in, the following nice property holds according to \[\begin{aligned} q(\mathbf{x}_{t}|\mathbf{x}_{0}) = \mathcal{N}(\mathbf{x}_t;\sqrt{\bar{\alpha}_t}\mathbf{x}_{0},(1-\bar{\alpha}_{t})\mathbf{I}),\\ \quad \bar{\alpha}_t =\prod_{i=1}^{T} \alpha_i ~\text{where}~\alpha_t = 1-\beta_t \end{aligned}\] Here \(\bar{\alpha}\) controls the scale of noise added at each time step. At the beginning, noise should be small so that it is possible for the model to learn well, e.g., \(\beta_1<\beta_2<\dots<\beta_T\) and \(\bar{\alpha}_1>\bar{\alpha}_2\dots>\bar{\alpha}_T\). ## Reverse Process Since the forward process defined above is a Markov chain, the true sample can be recreated from Gaussian noise \(\mathbf{x}_T \sim\mathcal{N}(\mathbf{0},\mathbf{I})\) by reversing the forward process: \[\label{eq:reversed-process} \begin{split} p_{\theta}(\mathbf{x}_{0:T})&:= p(\mathbf{x}_T)\prod_{i=1}^{T} p_{\theta}(\mathbf{x}_{t-1}|\mathbf{x}_{t}),\\ p_{\theta}(\mathbf{x}_{t-1}|\mathbf{x}_{t}) &= \mathcal{N}(\mathbf{x}_{t-1};\mathbf{\mu}_{\theta}(\mathbf{x}_t,t),\mathbf{\Sigma}_{\theta})(\mathbf{x}_t,t)) \end{split}\] Ho et al. noticed that using Bayes's rule and conditioning on \(\mathbf{x}_0\), the reverse process has a closed form expression: \[\label{eq:closed-form} \begin{split} &q(\mathbf{x}_{t-1}|\mathbf{x}_{t},\mathbf{x}_0) = \mathcal{N}(\mathbf{x}_{t-1};\Tilde{\mathbf{\mu}}_{t}(\mathbf{x}_{t},\mathbf{x}_0),\Tilde{\beta}_{t}\mathbf{I}),\\ \text{where}&\quad \Tilde{\mathbf{\mu}}_{t}(\mathbf{x}_{t},\mathbf{x}_0) = \frac{\sqrt{\bar{\alpha}_{t-1}}\beta_{t}}{1-\bar{\alpha}_{t}}\mathbf{x}_0 + \frac{\sqrt{\alpha_{t}}(1-\bar{\alpha}_{t})}{1-\bar{\alpha}_{t}}\mathbf{x}_t\\ \text{and}&\quad \Tilde{\beta}_{t} = \frac{(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_{t}}\beta_t \end{split}\] ## Training Based on the reverse process, the training process is based on optimizing the variational lower bound on the negative log likelihood: \[\begin{split} \mathbb{E}[-\log p_{\theta}(\bx_{0})] &\le \mathbb{E}_{q}\Big[-\log \frac{p_{\theta}(\bx_{0:T})}{q(\bx_{1:T}|\bx_0)}\Big] \\ &= \mathbb{E}_{q}\Big[-\log p(\mathbf{x}_{T})-\sum_{t\geq 1}\log \frac{p_{\theta}(\bx_{t-1}|\bx_{t})}{q(\bx_{t}|\bx_{t-1})}\Big] =: L \end{split}\] For variance reduction purpose, the loss can be rewritten as: \[\label{eq:VLB} \begin{split} \mathbb{E}_{q}\Big[\underbrace{D_{KL}(q({\mathbf{x}_{T}|\mathbf{x}_{0}})||p(\mathbf{x}_{T}))}_{L_{T}} &+ \sum_{t\geq 1} \underbrace{D_{KL}(q(\mathbf{x}_{t-1}|\mathbf{x}_{t},\mathbf{x}_0)||p_{\theta}(\bx_{t-1}|\bx_{t}))}_{L_{t-1}}\\ &-\underbrace{\log p_{\theta}(\mathbf{x}_0|\mathbf{x}_1)}_{L_0} \Big] \end{split}\] Based on [\[eq:reversed-process\]](#eq:reversed-process){reference-type="eqref" reference="eq:reversed-process"}, [\[eq:closed-form\]](#eq:closed-form){reference-type="eqref" reference="eq:closed-form"} and [\[eq:VLB\]](#eq:VLB){reference-type="eqref" reference="eq:VLB"}, \(L_{t-1}\) can be computed as: \[L_{t-1} = \mathbb{E}_{q}\Big[\frac{1}{2\sigma_{t}^2}\|\Tilde{\mathbf{\mu}}_{t}(\mathbf{x}_{t},\mathbf{x}_0)-\mathbf{\mu}_{\theta}(\mathbf{x}_t,t)\|^2 \Big] + C\] where \(C\) is a constant independent of \(\theta\). Notice that \(\mathbf{x}_{t}(\mathbf{x}_0,\mathbf{\epsilon}) = \sqrt{\bar{\alpha}{t}\mathbf{x}_0} + \sqrt{1-\bar{\alpha}_{t}}\mathbf{\epsilon}\) where \(\mathbf{\epsilon} \sim \mathcal{N}(\mathbf{0},\mathbf{I})\). Thus if we apply [\[eq:closed-form\]](#eq:closed-form){reference-type="eqref" reference="eq:closed-form"}, the loss can be reformulated as: \[\resizebox{0.91\hsize}{!}{ \(L_{t-1}-C = \mathbb{E}_{\mathbf{x}_{0},\epsilon}\Big[ \frac{1}{2\sigma_{t}^2}\|\Tilde{\mathbf{\mu}}_{t}\Big(\mathbf{x}_{t}(\mathbf{x}_0,\mathbf{\epsilon}), \frac{\beta_t}{\sqrt{1-\bar{\alpha}}}\mathbf{\epsilon}\Big)-\mathbf{\mu}_{\theta}(\mathbf{x}_t,t)\|^2\Big]\)}\] Since \(\mathbf{x}_t\) is accessible during training, DDPM chooses the following parametrization: \[\mathbf{\mu}_{\theta}(\mathbf{x}_t,t) = \frac{1}{\sqrt{\bar{\alpha}}}\Big(\mathbf{x}_{t}-\frac{\beta_t}{\sqrt{1-\bar{\alpha}}}\mathbf{\epsilon}_{\theta}(\mathbf{x}_t,t)\Big)\] Finally, the training is performed based on the following loss function: \[\mathbb{E}_{\mathbf{x}_{0},\mathbf{\epsilon}}\Big[ \frac{\beta_t^2}{2\sigma_{t}^2\alpha_{t}(1-\bar{\alpha})}\|\mathbf{\epsilon}-\mathbf{\epsilon}_{\theta}(\sqrt{\bar{\alpha}}\mathbf{x}_0 + \sqrt{1-\bar{\alpha}}\mathbf{\epsilon},t)\|\Big]\] ## Algorithm Based on above sections, we can summarize the two key algorithms of DDPM. Algorithm [\[suppalg:training-dff\]](#suppalg:training-dff){reference-type="ref" reference="suppalg:training-dff"} displays the complete training procedure with the simplified objective introduced in. Algorithm [\[suppalg:Sampling-dff\]](#suppalg:Sampling-dff){reference-type="ref" reference="suppalg:Sampling-dff"} illustrates the reverse process sampling procedure, where the goal is to predict the error, \(\epsilon\). # Experiments In this section, we answer following questions: **Q1:** Can `MedDiff` generate high quality synthetic EHRs compared to existing methods? **Q2:** Can `MedDiff` accurately generate conditioned samples that match the distribution of the given label? **Q3:** Can our proposed acceleration technique alleviate the issue of slow generation process of a diffusion model? How is its performance under different settings? We first describe the two health datasets used in our experiments. Next, we give an overview of baseline methods and implementation details. We then evaluate the effectiveness of the proposed acceleration technique with an ablation study. ## Datasets We use the following two publicly available datasets. 1. MIMIC-III: A large database containing de-identified health data associated with approximately sixty thousand admissions of critical care unit patients from the Beth Israel Deaconess Medical Center collected between 2001 and 2012. For each patient, we extract the International Classification of Diseases (ICD-9) diagnosis codes. We represent a patient record as a fixed-size vector with 1071 entries for each patient record. The pre-processed dataset is a 46520 \(\times\) 1071 binary matrix and used to evaluate the binary discrete variable generation and the proposed acceleration technique. 2. Patient Treatment Classification[^1]: A dataset collected from a private hospital in Indonesia. It contains the 8 different laboratory test results of 3309 patient used to determine next patient treatment whether in care or out care. We use this datasets to perform continuous synthetic EHR generation and investigate the effectiveness of conditional generation of `MedDiff`. ## Baselines and Implementation We compare `MedDiff`  with following methods. 1. MedGAN [^2]: A GAN-based model that generates the low-dimensional synthetic records and decodes them with an autoencoder. 2. CorGAN \(^2\): A GAN-based model similar to MedGAN but combines Convolutional Generative Adversarial Networks and Convolutional Autoencoders. 3. Noise Conditional Score Network (NCSN) [^3]: Instead of directly learning the probability of the data log \(p(x)\), this method aims to learn the gradients of log \(p(x)\) with respect to \(x\). This can be understood as learning the direction of highest probability at each point in the input space. After training, the sampling process is achieved by applying Langevin dynamics. 4. DDPM [^4]: A class of latent variable models inspired by considerations from nonequilibrium thermodynamics. We implemented `MedDiff` with Pytorch. For training the models, we used Adam with the learning rate set to \(0.001\), and a mini-batch of \(128\) and \(64\) for MIMIC-III and the patient treatment dataset, respectively on a machine equipped with one Nvidia GeForce RTX 3090 and CUDA 11.2. Hyperparamters of `MedDiff` are selected after grid search. We use a timestep \(T\) of 200, a noise scheduling \(\beta\) from \(1 \times 10^{-4}\) to \(1 \times 10^{-2}\) and a table size \(k=3\). The code will be publicly available upon publication. ## Sample Quality Evaluation We first evaluate the effectiveness of `MedDiff` on MIMIC-III. Following previous works CorGAN and MedGAN, we use the dimension-wise probability as a basic evaluation metric to determine if `MedDiff` can learn the distribution of the real data (for each dimension). This measurement refers to the Bernoulli success probability of each dimension (each dimension is a unique ICD-9 code). We report the dimension-wise probability in Figure [\[fig:correlation\]](#fig:correlation){reference-type="ref" reference="fig:correlation"}. We also use the correlation coefficient \(\rho\) and sum of absolute errors (SAE) as our quantitative metrics. From Figure [\[fig:correlation\]](#fig:correlation){reference-type="ref" reference="fig:correlation"}, we can observe `MedDiff` shows the best performance both in terms of highest correlation (\(\rho=0.98)\) and the lowest SAE (4.16). The results also illustrate that naïve application of DDPM yields substandard results, even worse than MedGAN and CorGAN. ## Conditioned Sample Quality Evaluation Next, we answer the question whether `MedDiff` preserves the conditioned sampling of health records with a given label on the Patient Treatment Classification dataset. We compare the underlying probability density function of the original dataset and generated samples using Gaussian Kernel Density Estimation (KDE). We note that none of the baseline methods offer this capability. The results for the unconditional sampling are shown in Figure [\[fig:kdeall\]](#fig:kdeall){reference-type="ref" reference="fig:kdeall"} while the conditional sampling versions are shown in Figures [\[fig:kdein\]](#fig:kdein){reference-type="ref" reference="fig:kdein"} and [\[fig:kdeout\]](#fig:kdeout){reference-type="ref" reference="fig:kdeout"}. From the plots, the generated synthetic samples mostly follow the overall probability function of the original dataset. Furthermore, we can observe that the distribution of in-care patients and out-care patients are different in terms of shape, local modes, and range. This means the generated samples are potentially more informative and offer more utility when used for data sharing. This experiment also demonstrates the flexibility and advantage of `MedDiff` in terms of ability to deal with class-conditional sampling. ## Effects of Accelerated Sampling Next, we evaluate the usefulness of our proposed accelerated sampling algorithm on MIMIC-III. For \(T=200\), we present the evolution process of generating one sample by use of the regular procedure in Figure [\[fig:T200reg\]](#fig:T200reg){reference-type="ref" reference="fig:T200reg"} and `MedDiff` in Figure [\[fig:T200AA\]](#fig:T200AA){reference-type="ref" reference="fig:T200AA"}. Similarly, Figures [\[fig:T100reg\]](#fig:T100reg){reference-type="ref" reference="fig:T100reg"} and [\[fig:T100AA\]](#fig:T100AA){reference-type="ref" reference="fig:T100AA"} depict the evolution process of generating one sample for \(T=100\). The iteration number is defined as \(T-t\). When \(T=200\), we can observe `MedDiff` converges \(\sim80\) iterations whereas the regular sampling process takes \(\sim160\) iterations, a \(2\times\) speed up. From Figures [\[fig:T100AA\]](#fig:T100AA){reference-type="ref" reference="fig:T100AA"} and [\[fig:T100reg\]](#fig:T100reg){reference-type="ref" reference="fig:T100reg"}, `MedDiff` can generate a high-quality sample when the regular sampling process fails to converge. We perform an additional qualitative ablation study of `MedDiff` to investigate whether our model has actually accelerated the generation process with different hyperparameters. We use the root mean squared error of the Bernoulli success probability of each dimension for the generated samples and original dataset, which is defined as follows, \[RMSE = \sqrt{\frac{\sum_{d=1}^{D}\left(p_{d}-\hat{p}_{d}\right)^{2}}{D}},\] where \(D\) is the feature size. We run `MedDiff` using different \(T\) and AA restart dimension (or table size) \(k\) and plot the results in Figure [\[fig:ablationAA\]](#fig:ablationAA){reference-type="ref" reference="fig:ablationAA"}. Figures [\[fig:t100iter\]](#fig:t100iter){reference-type="ref" reference="fig:t100iter"} and [\[fig:t200iter\]](#fig:t200iter){reference-type="ref" reference="fig:t200iter"} indicate an increasing of table size results in faster convergence in terms of iteration number. Figures [\[fig:t100time\]](#fig:t100time){reference-type="ref" reference="fig:t100time"} and [\[fig:t200time\]](#fig:t200time){reference-type="ref" reference="fig:t200time"} show that although there is additional computation, it does not hinder the benefits of adopting the accelerated algorithm with a small \(k\). Moreover, Figures [\[fig:t100iter\]](#fig:t100iter){reference-type="ref" reference="fig:t100iter"} and [\[fig:t200iter\]](#fig:t200iter){reference-type="ref" reference="fig:t200iter"} verify that adopting the accelerated algorithm can help to generate high-quality samples if the model is trained with fewer timesteps. ## Case Study: Discriminative Analysis Classification models are often developed on EHR data to determine whether the next patient treatment is in-care or out-care. To evaluate the utility of our synthetic EHRs, we evaluate on a prediction task. We use an 80/20 training/test split on the patient treatment dataset. We use the training set (2647 patients) to train and generate class-conditional synthetic data. Then, we evaluate the performance of a logistic regression model trained using the real data or the synthetic data on the real test set (662 patients). Unsurprisingly, there was a drop in AUC between the real data (0.766) and the synthetic data (0.742). Synthetic EHRs can also be beneficial for data augmentation purposes to develop a more robust classifier. We also investigate whether the data generated by `MedDiff` can be used for this purpose by varying the amounts of synthetically generated data used to augment the training set, and evaluate the AUC on the test set. The results from this experiment are shown in Figure [\[fig:dataaug\]](#fig:dataaug){reference-type="ref" reference="fig:dataaug"}. It can be observed that augmenting the training set with more than 2000 synthetic records can yield better performance than just the real data. However, augmenting with small amounts of synthetic data can harm the performance as shown by the performance degradation with less than 2000 synthetic records. We posit that this might be that the synthetic sample distribution may not yet reflect the true distribution with insufficient samples. # Introduction Recent digitisation of health records has enabled the training of deep learning models for precision medicine, personalised prediction of risks and health trajectories. However, there are still patient privacy concerns that need to be accounted for in order to aggregate more data to train more robust models. As such, it is still hard for researchers to obtain access to real electronic health records (EHRs). One approach to mitigate privacy risks is through the practice of de-identification such as perturbation and randomization. However, this approach is vulnerable to re-identification. Another approach is to use synthetic datasets that capture as many of the complexities of the original data set (e.g., distributions, non-linear relationships, and noise). Synthetic EHRs can yield a database that is beyond de-identification and hence are immune to re-identification. Thus, generating realistic, but not real data is a key element to advance machine learning for the healthcare community. There have been several distinguished efforts conducted in a variety of domains about synthetic data (EHR) generation. Unfortunately, existing proposed algorithms predominantly adopt a variant of Generative Adversarial Network (GAN), auto-encoder, or a combination of both. While GANs and autoencoders are natural and widely used candidates for generation, there are several noticeable drawbacks of these models, including mode-collapse for GANs or poor sample diversity and quality for autoencoders. In recent years, diffusion (score) based models have emerged as a family of powerful generative models that can yield state-of-the-art performance across a range of domains, including image and speech synthesis. Diffusion models have been shown to achieve high-quality, diverse samples that are superior to their GAN-based counterparts. Other key advantages of diffusion models include ease of training and tractability, in contrast to GANs, and speed of generation, in contrast to autoregressive models. This leads to a natural question: Is a diffusion model promising for generating synthetic EHRs as well? We answer this in the affirmative, by introducing `MedDiff`, a novel denoising diffusion probabilistic model. `MedDiff`, shown in Figure [\[fig:medDIFF\]](#fig:medDIFF){reference-type="ref" reference="fig:medDIFF"}, generates high quality, robust samples while also being simple enough for practitioners to train. We further accelerate the generation process of `MedDiff` using Anderson acceleration, a numerical method that can improve convergence speed of fixed-point sequences. In summary, our contributions are as follows: - We investigate the effectiveness of diffusion based models on generating discrete EHRs. - We propose to accelerate the generation process, a main drawback of diffusion models. - We introduce a novel conditioned sampling technique to generate discriminative synthetic EHRs. - We show that `MedDiff` can generate realistic synthetic data that mimics the real data and provides similar predictive value. # Problem Specification. In this paper, we consider the solution of the \(N \times N\) linear system \[\label{e1.1} A x = b\] where \(A\) is large, sparse, symmetric, and positive definite. We consider the direct solution of ([\[e1.1\]](#e1.1){reference-type="ref" reference="e1.1"}) by means of general sparse Gaussian elimination. In such a procedure, we find a permutation matrix \(P\), and compute the decomposition \[P A P^{t} = L D L^{t}\] where \(L\) is unit lower triangular and \(D\) is diagonal. # Design Considerations. Several good ordering algorithms (nested dissection and minimum degree) are available for computing \(P\) ,. Since our interest here does not focus directly on the ordering, we assume for convenience that \(P=I\), or that \(A\) has been preordered to reflect an appropriate choice of \(P\). Our purpose here is to examine the nonnumerical complexity of the sparse elimination algorithm given in . As was shown there, a general sparse elimination scheme based on the bordering algorithm requires less storage for pointers and row/column indices than more traditional implementations of general sparse elimination. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. Our purpose here is to examine the nonnumerical complexity of the sparse elimination algorithm given in . As was shown there, a general sparse elimination scheme based on the bordering algorithm requires less storage for pointers and row/column indices than more traditional implementations of general sparse elimination. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. Several good ordering algorithms (nested dissection and minimum degree) are available for computing \(P\) ,. Since our interest here does not focus directly on the ordering, we assume for convenience that \(P=I\), or that \(A\) has been preordered to reflect an appropriate choice of \(P\). Our purpose here is to examine the nonnumerical complexity of the sparse elimination algorithm given in . As was shown there, a general sparse elimination scheme based on the bordering algorithm requires less storage for pointers and row/column indices than more traditional implementations of general sparse elimination. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. Our purpose here is to examine the nonnumerical complexity of the sparse elimination algorithm given in . As was shown there, a general sparse elimination scheme based on the bordering algorithm requires less storage for pointers and row/column indices than more traditional implementations of general sparse elimination. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. Several good ordering algorithms (nested dissection and minimum degree) are available for computing \(P\) ,. Since our interest here does not focus directly on the ordering, we assume for convenience that \(P=I\), or that \(A\) has been preordered to reflect an appropriate choice of \(P\). Our purpose here is to examine the nonnumerical complexity of the sparse elimination algorithm given in . As was shown there, a general sparse elimination scheme based on the bordering algorithm requires less storage for pointers and row/column indices than more traditional implementations of general sparse elimination. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.2, we review the bordering algorithm, and introduce the sorting and intersection problems that arise in the sparse formulation of the algorithm. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. For the old approach, we show that the complexity of the intersection problem is \(O(n^{3})\), the same as the complexity of the numerical computations. For the new approach, the complexity of the second part is reduced to \(O(n^{2} (\log n)^{2})\). To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase -, , . ## Robustness. We do not attempt to present an overview here, but rather attempt to focus on those results that are relevant to our particular algorithm. This section assumes prior knowledge of the role of graph theory in sparse Gaussian elimination; surveys of this role are available in and. More general discussions of elimination trees are given in-,. Thus, at the \(k\)th stage, the bordering algorithm consists of solving the lower triangular system \[\label{1.2} L_{k-1}v = c\] and setting \[\begin{aligned} \ell &=& D^{-1}_{k-1}v, \\ \delta &=& \alpha-\ell^{t} v. \end{aligned}\] # Robustness. We do not attempt to present an overview here, but rather attempt to focus on those results that are relevant to our particular algorithm. ## Versatility. The special structure of this problem allows us to make exact estimates of the complexity. For the old approach, we show that the complexity of the intersection problem is \(O(n^{3})\), the same as the complexity of the numerical computations ,. For the new approach, the complexity of the second part is reduced to \(O(n^{2} (\log n)^{2})\). To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.2, we review the bordering algorithm, and introduce the sorting and intersection problems that arise in the sparse formulation of the algorithm. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. For the old approach, we show that the complexity of the intersection problem is \(O(n^{3})\), the same as the complexity of the numerical computations. For the new approach, the complexity of the second part is reduced to \(O(n^{2} (\log n)^{2})\). To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. In §1.3., we analyze the complexity of the old and new approaches to the intersection problem for the special case of an \(n \times n\) grid ordered by nested dissection. The special structure of this problem allows us to make exact estimates of the complexity. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase \[4\]-\[10\], \[5\], \[6\]. This is accomplished by exploiting the m-tree, a particular spanning tree for the graph of the filled-in matrix. To our knowledge, the m-tree previously has not been applied in this fashion to the numerical factorization, but it has been used, directly or indirectly, in several optimal order algorithms for computing the fill-in during the symbolic factorization phase -, , . # MedDiff In this section, we introduce the three components of `MedDiff`. We first introduce the base architecture. Next, we propose a numerical method to accelerate the generation process. We then equip `MedDiff` with the ability to conduct conditioned sampling. ## Base Architecture In our preliminary experiments, we observed that a simple, fully connected diffusion model is sufficient for low-dimensional data. However, applying DDPM developed for images and audio yields unsatisfying results. The current best architectures for image diffusion models are U-Nets, which are a natural choice to map corrupted data to reverse process parameters. However, they are tailored for 2D signals such as images and video frames. Yet, this may not be a viable option in numerous applications over 1D signals especially when the training data is scarce or application specific. As a result, `MedDiff` uses a modified U-Net architecture including larger model depth/width, positional embeddings, residual blocks for up/downsampling, and residual connection re-scale. Additionally, we use the 1d convolutional form of U-Net to generate a vector for each patient. Under this base architecture, `MedDiff` can better capture the neighboring feature correlations. To better understand the intermediate steps of `MedDiff`, we visualize the resulting images at \(T=0, 10, 50, 100,\) and \(200\) using one sample for illustration. Figure [\[fig:forwardandreverse\]](#fig:forwardandreverse){reference-type="ref" reference="fig:forwardandreverse"} depicts the forward and reverse process of a diffusion model. From Figure [\[fig:forward\]](#fig:forward){reference-type="ref" reference="fig:forward"} we can see that the forward process destroys the input by adding scaled random Gaussian noise step by step. At \(T=200\), the generated sample looks like random noise. If we can reverse the above process and sample from \(q(x_{t-1}|x_t)\), we will be able to recreate the true sample from a Gaussian noise input. Since we cannot easily estimate \(q(x_{t-1}|x_t)\), we learn a model \(p_\theta\) using neural networks to approximate these conditional probabilities in order to run the reverse diffusion process. Figure [\[fig:reverse\]](#fig:reverse){reference-type="ref" reference="fig:reverse"} shows that given a random noise, `MedDiff`  is able to generate a new sample. Furthermore, `MedDiff`  can reconstruct the input if it uses a non-perturbed sampling procedure (set the posterior variance \(\sigma_t\) as \(0\)) and the noise input for the reverse process is exactly the last noisy input from a forward process. This confirms that by setting the posterior variance \(\sigma_t\) as a nonzero number and not sharing the destroyed inputs after the forward process, `MedDiff` can achieve sample diversity without leaking the original inputs. ## Accelerated Generation Process Although the above architecture can provide satisfying results, the speed of generation process can still be a bottleneck. The dilemma is that a small \(T\) usually performs worse than a larger \(T\), but a larger \(T\) requires a longer generation process (and time). To alleviate this issue, we propose an acceleration algorithm from the perspective of iterative methods. We utilize Anderson Acceleration (AA) to run the generation process. AA is a method that accelerates the convergence of fixed-point iterations. The idea is to approximate the final solution using a linear combination of the previous \(k\) iterates. Since solving the proper combination of iterates is a nonlinear procedure, AA is also known as a nonlinear extrapolation method. We start with the generic truncated version of the AA prototype (see Algorithm [\[suppalg:AA\]](#suppalg:AA){reference-type="ref" reference="suppalg:AA"}) and explain its implementation. For each iteration \(t \geq 0\), AA solves a least squares problem with a normalization constraint. The intuition is to minimize the norm of the weighted residuals of the previous \(k\) iterates. The constrained linear least-squares problem in Algorithm [\[suppalg:AA\]](#suppalg:AA){reference-type="ref" reference="suppalg:AA"} can be solved in a number of ways. Our preference is to recast it in an unconstrained form suggested in that is straightforward to solve and convenient for implementing efficient updating of QR. We present the idea of the Quick QR-update Anderson Acceleration implementation as described in in the Supplemental Section [9.1](#supsec:QR){reference-type="ref" reference="supsec:QR"}. Here we provide the theoretical motivation and results. There is a long history of applying AA to Picard's iterative method for solving differential equations. We can regard the updating rule of DDIM as the integrator of [\[eq:integrator-ode\]](#eq:integrator-ode){reference-type="eqref" reference="eq:integrator-ode"} which is approximated by \(\frac{1}{2}[\bar{\alpha}_t \mathbf{\epsilon}_{\theta}(\mathbf{x}_t)-\bar{\alpha}_{t}^2\mathbf{x}_{t}]\). Thus, [\[eqn:update\]](#eqn:update){reference-type="eqref" reference="eqn:update"} can be written as the Picard iteration \[\mathbf{x}_s = \mathbf{x}_T + \int_{T}^{s}\underbrace{\frac{1}{2}[\bar{\alpha}_t \mathbf{\epsilon}_{\theta}(\mathbf{x}_t)-\bar{\alpha}_{t}^2\mathbf{x}_{t}]dt}_{F(\mathbf{x}_t,t)}.\] Under this lens, we can apply AA to this sequence \(\{\mathbf{x}_{t}\}\). Assuming \(F(\mathbf{x}_t,t)\) is uniformly Lipschitz continuous in \(\mathbf{x}\) (a common assumption made in neural ODE literature ) and following the results from Theorem 2.3 in, we can derive the following acceleration guarantees for applying AA to iterates (i.e., intermediate samples) of DDIM. Theorem [\[thm:AA\]](#thm:AA){reference-type="ref" reference="thm:AA"} provides the theoretical justification for leveraging AA to predict the next denoised intermediate sample. Since `MedDiff` requires much fewer iterations and reduced computational costs to generate realistic samples, it can dramatically save inference time. ## Conditioned Sampling Sometimes it is insufficient for the model to produce realistic-looking data, it should also ensure the generated examples preserve utility in a down-stream task. As such, if a particular class label is passed to the synthesizer, it should produce a health record that matches the distribution of that label. This is one of the limitations of MedGAN and CorGAN, in that it may not preserve the class-dependent label information. We equip `MedDiff` with this ability by incorporating the idea of a classifier-guided sampling process. From a conceptual level, the estimated noise \({\epsilon_\theta}(\mathbf{x}_t, t)\) in each step is deducted by \(\sqrt{1-\bar{\alpha}_{t}} \nabla_{\mathbf{x}_{t}} \log f_{\phi}\left(y \mid \mathbf{x}_{t}\right)\), where \(f_{\phi}\left(y \mid \mathbf{x}_{t}\right)\) is a trained classifier on the noisy \(\mathbf{x}_t\). This modified updating rule tends to up-weighting the probability of data where the classifier \(f_{\phi}\left(y \mid \mathbf{x}_{t}\right)\) assigns high likelihood to the correct label. Algorithm [\[suppalg:AASampling-dff\]](#suppalg:AASampling-dff){reference-type="ref" reference="suppalg:AASampling-dff"} summaries the corresponding accelerated and conditioned sampling algorithm. # Related Work & Background To the best of our knowledge, `MedDiff` is the first work to leverage the idea of diffusion based modeling to generate EHRs. Since our work casts insight on the effectiveness of generating EHR via diffusion based models, we leave the generation of more complex EHRs (e.g., multiple data sources, temporal data) for future investigation. Here, we discuss the related works on synthetic EHRs generation and related diffusion based models. ## Synthetic EHR generation Closely related to this work are recent efforts that leverage deep generative models for synthesizing EHRs. MedGAN and CorGAN were introduced to generate patient feature matrices. However, these works rely heavily on the performance of a pre-trained autoencoder model to reduce the dimensionality of the latent variable. Without the pre-trained autoencoder, these GAN-based models can fail to generate high-quality samples, highlighting the difficulty of using these models when generalizing to multiple institutions (e.g., smaller clinics or different patient distributions). `MedDiff` builds on diffusion models and does not require a pre-trained encoder. ## Diffusion Models First proposed in, diffusion models are a family of latent variable generative models characterized by a forward and a reverse Markov process. The forward process gradually adds noise to the original data sample, whereas the reverse process undoes the gradual noising process. In the reverse process, the sampling starts with the \(T\)th noise level, \(x_T\), and each timestep produces less-noisy samples, \(x_{T-1}, x_{T-2},..., x_0\). In essence, the diffusion model learns the "denoised\" version from \(x_{t-1}\) to \(x_t\). Diffusion models have several advantages over existing generative modeling families. They do not rely on adversarial training which can be susceptible to mode collapse and are difficult to train. They also offer better diversity coverage and can accommodate flexible model architectures to learn any arbitrary complex data distributions. With respect to image and speech synthesis, diffusion-based models can achieve high-quality, diverse samples that are superior to their GAN counterparts. Ho et al. proposed a specific parameterization of the diffusion model to simplify the training process using a score matching-like loss that minimizes the mean-squared error between the true noise and the predicted noise. They also note that the sampling process can be interpreted as equivalent to Langevian dynamics, which allows them to relate the proposed denoising diffusion probabilistic models (DDPM) to score-based methods in. Denoising diffusion implicit models (DDIM) was recently proposed to offer a more flexible framework based on DDPM. Here, we introduce the forward process, the reverse process, training, and sampling of DDIM. ## DDIM ### Forward Process Given a data distribution \(\mathbf{x}_0 \sim q(\mathbf{x}_0)\), diffusion models gradually add noise to the original data distribution until it loses all the original information and becomes an entirely noisy distribution (as shown by the \(x_T\) sample in Figure [\[fig:medDIFF\]](#fig:medDIFF){reference-type="ref" reference="fig:medDIFF"}). DDPM convolves \(q(\mathbf{x}_0)\) with an isotropic Gaussian noise \(\mathcal{N}(0,\sigma_{i}^{2} \mathbf{I})\) in T steps to produce a noise corrupted sequence \(\bx_1,\bx_2,\dots,\bx_T\). \(\mathbf{x}_T\) will converge to an isotropic Gaussian distribution as \(T\rightarrow \infty\). However, for DDIM, a different variance schedule is used to produce \(\bx_1,\bx_2,\dots,\bx_T\). The explicit input distribution \(q\) of DDIM is derived as: \[q(\mathbf{x}_{1:T}|\mathbf{x}_0) := q(\mathbf{x}_{T}|\mathbf{x}_0)\prod_{t=2}^{T}q(\mathbf{x}_{t-1}|\mathbf{x}_{t},\mathbf{x}_0)\] where \(q(\mathbf{x}_{T}|\mathbf{x}_0) = \mathcal{N}(\sqrt{\bar{\alpha}}\mathbf{x}_0, (1-\bar{\alpha}_T)\mathbf{I})\) and for all \(t>1\), \[\begin{aligned} q(\mathbf{x}_{t-1}|\mathbf{x}_{t},\mathbf{x}_0) =& \mathcal{N} (\sqrt{\bar{\alpha}_{t-1}}\mathbf{x}_0 + \\ &\sqrt{1-\bar{\alpha}_{t-1}-\sigma^2}\frac{\mathbf{x}_t-\sqrt{\bar{\alpha}_t}\mathbf{x}_0}{\sqrt{1-\bar{\alpha}_{t}}}, \sigma^2\mathbf{I}) \nonumber \end{aligned}\] Here \(\bar{\alpha}\) controls the scale of noise added at each time step. At the beginning, noise should be small so that it is possible for the model to learn well, e.g., \(\bar{\alpha}_1>\bar{\alpha}_2\dots>\bar{\alpha}_T\). The DDIM distribution is parameterized to guarantee the marginal density is equivalent to DDPM. However, the key difference between DDPM and DDIM is that the forward process of DDIM is no longer a Markov process. This allows acceleration of the generative process as multiple steps can be taken. Moreover, different reverse samplers can be utilized by changing the variance of the reverse noise. This means DDIM can be compatible with other samplers. ### Backward process For DDPM, the forward process is defined by a Markov chain, and thus the true sample can be recreated from Gaussian noise \(\mathbf{x}_T \sim\mathcal{N}(\mathbf{0},\mathbf{I})\) by reversing the forward process. However, as noted above, DDIM relies on a family of non-Markovian processes. Denote \(p_{\theta}\) as a parameterized neural network, the reverse process with a prior \(p_{\theta}(\mathbf{x}_{T}) = \mathcal{N}(\mathbf{0},\mathbf{I})\) can be computed as \[\begin{split} p_{\theta}(\mathbf{x}_{0}|\mathbf{x}_1) &= \mathcal{N}\Big(\pred{1},\sigma_{1}^2\mathbf{I}\Big) \\ p_{\theta}(\mathbf{x}_{t-1}|\mathbf{x}_t)& = q(\mathbf{x}_{t-1}|\mathbf{x}_{t},\pred{t}),\quad t>1 \end{split}\] ## Training The training process of DDPM and DDIM is based on optimizing the variational lower bound on the negative log likelihood: \[\begin{aligned} \mathbb{E}[-\log p_{\theta}(\bx_{0})] &\le \mathbb{E}_{q}\Big[-\log \frac{p_{\theta}(\bx_{0:T})}{q(\bx_{1:T}|\bx_0)}\Big] \\ &= \mathbb{E}_{q}\Big[-\log p(\mathbf{x}_{T})-\nonumber \\ & ~~~~~\sum_{t\geq 1}\log \frac{p_{\theta}(\bx_{t-1}|\bx_{t})}{q(\bx_{t}|\bx_{t-1})}\Big] =: L \nonumber \end{aligned}\] As shown in Ho et al., this is equivalent to the following loss function: \[\mathbb{E}_{\mathbf{x}_{0},\mathbf{\epsilon}}\Big[ \frac{\beta_t^2}{2\sigma_{t}^2\alpha_{t}(1-\bar{\alpha})}\|\mathbf{\epsilon}-\mathbf{\epsilon}_{\theta}(\sqrt{\bar{\alpha}}\mathbf{x}_0 + \sqrt{1-\bar{\alpha}}\mathbf{\epsilon},t)\|\Big]\] ## Sampling After training, sampling can be done in DDIM using the following equation: \[\begin{aligned} \mathbf{x}_{t-1} =& \sqrt{\bar{\alpha}_{t-1}}\Big(\pred{t}\Big) +\\ & \sqrt{1-\bar{\alpha}_{t-1}-\sigma_{t}^2}\mathbf{\epsilon}_{\theta}(\mathbf{x}_t, t) + \sigma_t \mathbf{z}_t, \nonumber \end{aligned}\] where \(\mathbf{z}_{t} \sim \mathcal{N}(\mathbf{0},\mathbf{I})\). If we denote \(\frac{\sqrt{1-\bar{\alpha}_s}}{\sqrt{\bar{\alpha}_s}}\) by \(\lambda_{s}\), then the updating rule for continuous time is \[\mathbf{x}_{s} = \frac{\lambda_{s}}{\lambda_{t}}[\mathbf{x}_{t}-\bar{\alpha}_{t}\mathbf{\epsilon}_{\theta}(\mathbf{x}_t)] + \bar{\alpha}_{s}\mathbf{\epsilon}_{\theta}(\mathbf{x}_t) \label{eqn:update}\] To generate high-quality samples, it requires repeating the above updating rule (denoising process) multiple times and thus make the sampling procedure much slower than other generative models that only need one pass. A recent paper showed that DDIM is an integrator of the probability flow ordinary differential equation (ODE) defined in: \[\label{eq:integrator-ode} d\mathbf{x} = [f(\mathbf{x},t)-\frac{1}{2}g^{2}(t)\nabla_{\mathbf{x}}\log p_{t}(\mathbf{x})]dt\] where \(d\mathbf{x} = f(\mathbf{x},t)dt + g(t)dW\) is a stochastic differential equation and \(W\) is Brownian motion. In practice \(\nabla_{\mathbf{x}}\log p_{t}(\mathbf{x}) = \frac{\bar{\alpha}_t \mathbf{\epsilon}_{\theta}(\mathbf{x}_t)-\mathbf{x}_t}{\lambda_{t}^2}\) so \(\mathbf{x}_{s}\) can be regarded as the integrator of \(\frac{1}{2}[\bar{\alpha}_t \mathbf{\epsilon}_{\theta}(\mathbf{x}_t)-\bar{\alpha}_{t}^2\mathbf{x}_{t}]\). # Supplemental Material ## Fast and Efficient Implementation of Anderson Acceleration {#supsec:QR} Define \(f_i=g(w_i)-w_i\), \(\triangle f_i= f_{i+1}-f_i\) for each \(i\) and set \(F_t=[f_{t-p_t}, \dots, f_t\)\], \(\mathcal{F}_t=[\triangle f_{t-p_t}, \dots, \triangle f_t]\). Then solving the least-squares problem (\(\min _{\mathbf{\beta}}\left\|F_{t} \mathbf{\beta}\right\|_{2}, \text { s. t. } \sum_{i=0}^{p_{t}} \beta_{i}=1\)) is equivalent to \[\label{eqn:unAA} \min _{\gamma=\left(\gamma_{0}, \ldots, \gamma_{p_{t-1}}\right)^{T}}\left\|f_{t}-\mathcal{F}_{t} \gamma\right\|_{2}\] where \(\alpha\) and \(\gamma\) are related by \(\alpha_{0}=\gamma_{0}, \alpha_{i}=\gamma_{i}-\gamma_{i-1}\) for \(1 \leq i \leq p_{t}-1\), and \(\alpha_{p_{t}}=1-\gamma_{p_{t}-1}\). Now the inner minimization subproblem can be efficiently solved as an unconstrained least squares problem by a simple variable elimination. This unconstrained least-squares problem leads to a modified form of Anderson acceleration \[\begin{aligned} w_{t+1} &= g\left(w_{t}\right)-\sum_{i=0}^{p_{t}-1} \gamma_{i}^{(t)}\left[g\left(w_{t-p_{t}+i+1}\right)-g\left(w_{t-p_{t}+i}\right)\right] \\ &= g\left(w_{t}\right)-\mathcal{G}_{t} \gamma^{(t)} \end{aligned}\] where \(\mathcal{G}_t=[\triangle g_{t-p_t}, \dots, \triangle g_{t-1}]\) with \(\triangle g_i= g(w_{i+1})-g(w_i)\) for each \(i\). To obtain \(\gamma^{(t)}=\left(\gamma_{0}^{(t)}, \ldots, \gamma_{p_{t}-1}^{(t)}\right)^{T}\) by solving [\[eqn:unAA\]](#eqn:unAA){reference-type="eqref" reference="eqn:unAA"} efficiently, we show how the successive least-squares problems can be solved efficiently by updating the factors in the QR decomposition \(\mathcal{F}_t = Q_tR_t\) as the algorithm proceeds. We assume a thin QR decomposition, for which the solution of the least-squares problem is obtained by solving the \(p_t \times p_t\) linear system \(R \gamma=Q^{\prime}*f_{t}\). Each \(\mathcal{F}_t\) is \(n \times p_t\) and is obtained from \(\mathcal{F}_{t-1}\) by adding a column on the right and, if the resulting number of columns is greater than \(p\), also cleaning up (re-initialize) the table. That is,we never need to delete the left column because cleaning up the table stands for a restarted version of AA. As a result, we only need to handle two cases; 1 the table is empty (cleaned). 2 the table is not full. When the table is empty, we initialize \(\mathcal{F}_1=Q_1R_1\) with \(Q_1=\triangle f_{0} /\left\|\triangle f_{0}\right\|_{2}\) and \(R=\left\|\triangle f_{0}\right\|_{2}\). If the table size is smaller than \(p\), we add a column on the right of \(\mathcal{F}_{t-1}\). Have \(\mathcal{F}_{t-1} =QR\), we update \(Q\) and \(R\) so that \(\mathcal{F}_{t}=\left[\mathcal{F}_{t-1}, \Delta f_{t-1}\right]=Q R\). It is a single modified Gram--Schmidt sweep that is described as follows: Note that we do not explicitly conduct QR decomposition in each iteration, instead we update the factors (\(O(k^2n)\)) and then solve a linear system using back substitution which has a complexity of \(O(k^2)\). Based on this complexity analysis, we can find Anderson acceleration with QR-updating scheme has limited computational overhead. ## Additional results We present the evolution of the reverse (sampling) process in Figure [\[fig:kdeouteve\]](#fig:kdeouteve){reference-type="ref" reference="fig:kdeouteve"}. [^1]: <https://www.kaggle.com/manishkc06/patient-treatment-classification> [^2]: Code available at <https://github.com/astorfi/cor-gan> [^3]: Code adopted from <https://github.com/acids-ircam/diffusion_models>. [^4]: Code adopted from <https://github.com/lucidrains/denoising-diffusion-pytorch>

{'timestamp': '2023-02-10T02:02:44', 'yymm': '2302', 'arxiv_id': '2302.04355', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04355'}

# Hyperparameter Tuning of Beta-VAE Following Higgins et al. , we tune the hyperparameters \(dim(z)\) and \(\beta\) of the Beta-VAE for every dataset. To achieve maximum disentanglement, the number of latent dimensions should match the number of factors of variation in a dataset. This number is usually not known a priori. Choosing a latent space of too high dimensionality leads to a lot of uninformative dimensions with low variance. A latent space of too few latent dimensions, on the other hand, leads to entangled representations of features in the latent space. Therefore, it is necessary to identify the optimal number of latent dimensions to achieve maximally disentangled factors, ideally one in each dimension. To obtain the best combination of \(\beta\) and \(dim(z)\) for a given dataset, we first train a Beta-VAE with \(\beta=1\) and \(dim(z)=32\) on all datasets. We then compare the variance of the distribution in each dimension to that of a Gaussian prior. In the presence of dimensions with relatively low variance, we reduce the number of latent dimensions. Once we find a Beta-VAE with consistently informative dimensions, i.e. with the variance comparable to the prior, we fix \(dim(z)\) and focus on fine-tuning \(\beta\). As outlined by Higgings et al. , for relatively low values of \(\beta\), the VAE learns an entangled latent representation since the capacity in the latent space is too high. On the other hand, for relatively high values of \(\beta\), the capacity in the latent space becomes too low. The VAE performs a low-rank projection of the true data generative factors and again learns an entangled latent representation. This renders some of the latent dimensions uninformative. We find the highest possible \(\beta\) for which all dimensions of the VAE remain informative with a variance close to the prior. In line with the findings of Higgings et al. , \(\beta > 1\) is required for all datasets to achieve good disentanglement (see [1](#tab: hyperparameter tuning){reference-type="ref" reference="tab: hyperparameter tuning"}). # Evaluation We provide illustrations of the latent space traversal for the *Waterbirds* and *Colored MNIST* dataset. # Introduction {#s:intro} Machine learning (ML) addresses a wide range of real-world problems such as quality control, medical diagnosis and facial recognition. However, transferring new ML technology from the lab to the real world is often difficult due to the limited capacity of ML models to generalize. Among the reasons for this limitation are shortcuts: features in data \(X\) that correlate only statistically with the target \(Y\), but are inconsequential for the specific ML task. Geirhos et al. define shortcuts as a certain group of decision rules learned by neural networks. Shortcuts perform well on training data and on independent and identically distributed (i.i.d.) test data but fail on out-of-distribution (o.o.d.) data. The learned solution deviates from the intended solution in the presence of a distribution shift.\ Shortcut learning has been particularly evident in the medical field. A recent MIT technology report reveals that hundreds of tools were developed during the COVID-19 pandemic to diagnose the disease from chest X-rays, but none of them was found reliable enough for clinical use. The predictions of these models were often not based on the appearance of the lungs in the X-ray images. As the datasets were acquired from different sources for positive and negative cases, most of the models ended up learning the systematic differences in the data, e.g. the text fonts on the scans or the pose of the patient being scanned. Since the patients scanned while lying down were more likely to be seriously infected compared to those scanned while standing, the ML models wrongly learned to make a diagnosis based on a person's position. Shortcut learning can also be a cause of ethical concerns towards ML. Due to the existence of spurious correlations in the training data, ML models have been found to reinforce gender stereotypes. Detecting shortcuts is thus a major challenge to ensure the reliability and fairness of artificial intelligence (AI). Only if a model correctly learns the semantically significant correlations during training instead of focusing on the prevalent spurious correlations in the data, can it be deployed for real-world applications. Previous approaches for detecting shortcuts often rely on heatmaps and are thus limited to the identification of spatial shortcuts. We propose a novel method that is capable of identifying a variety of spurious features in images including background color, foreground color, object zoom level, and human facial characteristics. While other approaches applied generative models such as GANs for counterfactual explanations, we are, to the best of our knowledge, the first to leverage variational autoencoders (VAEs) for shortcut detection. We utilize a VAE to identify correlations between features of input \(X\) and target \(Y\) in datasets \(D =(X, Y)\). We provide tools for visualization and statistical analysis on the latent space of the VAE which allow a human judge to reliably detect ML shortcuts as those correlations that are not meaningful to the task at hand.\ In summary, our contributions are: - We introduce a VAE-based method to identify spurious correlations in datasets with minimal human supervision. - We demonstrate the applicability of our approach by finding both artificially injected shortcuts as well as previously unknown, real-world shortcuts in publicly available image and audio datasets. - We construct shortcut adversarial examples based on the spurious correlations identified by our method. # Related Work {#s:related_work} **Machine Learning Shortcuts.** Geirhos et al. locate the origin of shortcuts in the data and the learning process of ML models. The inherent contextual bias in datasets provides opportunities for shortcuts. Natural image datasets contain spurious correlations between the target variable and the background, the object poses or other co-occurring distracting features. In discriminative learning, a model uses a combination of features to make a prediction. Following the "principle of least effort", models tend to rely only on the most obvious features, which often correspond to shortcuts. For instance, convolutional neural networks (CNNs) trained on ImageNet were found to be biased towards the texture of the objects instead of their shapes. In another example, it was discovered that CNNs solely used the location of a single pixel to distinguish between object categories. **Identification of Shortcuts.** The identification of shortcuts in supervised machine learning is still in its infancy. Some of the existing Explainable AI (XAI) approaches have limited utility. Zech et al. use activation heatmaps to reveal the spurious features learned by CNNs trained on X-ray images. As outlined by Viviano et al. , saliency maps can only explain spatial shortcuts, e.g. source tags on images, but fail to identify more complicated ones, e.g. people's gender. Singla and Feizi select the highest activations of neurons in the penultimate layer of a CNN classifier and back-project them onto the input images. The resulting heatmaps highlight the features in the image that maximize neural activations. Under human supervision, the highlighted regions, for a subset of images, are labelled as 'core' (part of the object definition) or 'spurious' (only co-occurring with the object). Using this labelled dataset, they train a classifier to automatically identify the core and spurious visual features for a larger dataset. To diagnose shortcut learning, Geirhos et al. suggest performing o.o.d. generalization tests. Evaluating the model on o.o.d. real-world data in addition to the i.i.d. test set reveals whether a model is actually generalizing on the intended features or simply learning shortcuts from the training data. **Robustness against Shortcuts.** To learn representations robust to spurious correlations, Zhang et al. propose a two-stage contrastive approach. The method first identifies training samples from the same class with different model predictions. Contrastive learning then ensures that the hidden representations for samples of the same class with initially different predictions become close to each other. After training a classifier on shortcut-induced data, Kirichenko et al. retrain its last layer with Empirical Risk Minimization on a small dataset without spurious correlations. While the reweighting of the last layer reduces the model's reliance on background and texture information, the requirement of a shortcut-free subset remains a limitation of this approach. **Proposed Improvements.** Our approach overcomes a number of limitations from previous work. Most of the methods for explicit detection of shortcuts in a dataset are built on heatmaps and therefore limited to only identifying spatial shortcuts. Our method, on the other hand, identifies both spatial and non-spatial shortcuts as demonstrated in . Kirichenko et al. require a shortcut-free dataset for making a model robust against spurious correlations. Our method can serve as a preliminary step for this approach by identifying the shortcuts to be addressed. # Methodology {#s:approach} Let \(D = (X, Y)\) be a dataset, where \(X = \{x^{(i)} | x^{(i)} \in \mathbb{R}^{h \times w \times 3}\}_{i=1}^N\) are the images, \(Y = \{y^{(i)} | y^{(i)} \in \{1,..., C\}\}_{i=1}^N\) are the corresponding targets and \(C\) is the number of classes. We train a VAE on such a dataset and perform statistical analysis on its latent space to identify feature-target correlations in the data. The visualization of traversal in the latent space allows a human judge to examine those correlations for shortcuts. ## Variational Autoencoder {#ss:background_vae} Our approach for shortcut detection is based on variational autoencoders (VAEs), which are probabilistic generative models that learn the underlying data distribution in an unsupervised manner. A VAE attempts to model the marginal likelihood of an observed variable \(x\): \[p_{\theta}(x) = \int p_{\theta}(z) p_{\theta}(x|z) \,dz\] [\[eq:integral\]]{#eq:integral label="eq:integral"}The unobserved variable \(z\) lies in a latent space of dimensionality \(d = \dim(z)\), \(d \ll \dim(x)\). Each instance \(x^{(i)}\) has a corresponding latent representation \(z^{(i)}\). The model assumes the prior over the latent variables to be a multivariate normal distribution \(p_{\theta}(z) = \mathcal{N} (z; 0, I)\) resulting in independent latent factors \(\{z_j\}_{j=1}^d\). The likelihood \(p_{\theta}(x|z)\) is modelled as a multivariate Gaussian whose parameters are conditioned on \(z \sim p_{\theta}(z)\) and computed using the VAE decoder. As outlined by Kingma and Welling , the true posterior \(p_{\theta}(z|x)\) is intractable and hence approximated with a variational distribution \(q_{\phi}(z|x)\). This variational posterior is chosen to be a multivariate Gaussian \[\label{eq:posterior} q_{\phi}(z|x^{(i)}) = \mathcal{N} (z; \mu^{(i)}, (\sigma^{(i)})^2 I)\] whose mean \(\mu^{(i)} \in \mathbb{R}^d\) and standard deviation \(\sigma^{(i)} \in \mathbb{R}^d\) are obtained by forwarding \(x^{(i)}\) through the VAE encoder. The objective function is composed of two terms. A Kullback-Leibler (KL) divergence term ensures that the variational posterior distribution remains close to the assumed prior distribution. \[\begin{aligned} D_{KL}(q_{\phi}(z|x^{(i)})||p_{\theta}(z)) \\ =-\frac{1}{2}\sum_{j=1}^{d}\left( 1 + \text{log}((\sigma_j^{(i)})^2)-(\sigma_j^{(i)})^2-(\mu_j^{(i)})^2 \right) \label{eq:kld} \end{aligned}\] A log-likelihood loss, on the other hand, helps to accurately reconstruct an input image \(x^{(i)}\) from a sampled latent variable \(z^{(i)} = \mu^{(i)} + \sigma^{(i)} \odot \epsilon\) where \(\epsilon \sim \mathcal{N}(0, I)\). Thus the encoder and decoder are trained to maximize the following objective function: \[\label{eq:vae_loss} \begin{aligned} \mathcal{L}(\theta, \phi; x^{(i)}) \\ =-D_{KL}(q_{\phi}(z|x^{(i)}) || p_{\theta}(z)) + log~p_{\phi} (x^{(i)} | z^{(i)}) \end{aligned}\] For modelling images, both the encoder and decoder of a VAE consist of CNNs. Higgins et al. introduce the Beta-VAE to better learn independent latent factors. The authors propose to augment the vanilla VAE loss in by weighing the KL term with a hyperparameter \(\beta\). Choosing \(\beta >\) 1 enables the model to learn a more efficient latent representation of the data with better disentangled dimensions. ## Shortcut Detection with VAE {#ss:application} We train a Beta-VAE on an image dataset with potential shortcuts. The model discovers independent latent factors in a dataset and encodes them in its latent dimensions. Hence, each latent variable \(z_j\) of a trained Beta-VAE is likely to represent a descriptive property of the data, e.g. brightness, color, orientation, or shape. To establish feature-target correlations in the data, we measure how predictive each latent variable \(z_j\) is for each value of the target variable \(y\) (cf. ). We perform a statistical analysis on the latent space of a VAE and assess the utility of its latent dimensions for linear classification. For latent variables \(z_j\) that show a strong correlation with the target variable \(y\), we create visualizations that allow a human judge to easily decide whether the property encoded by \(z_j\) is a valid feature or a shortcut (see ). Our approach requires human inspection only for the candidate properties with the strongest feature-target correlations. In line with previous work, we argue that some form of human inspection will always be necessary to determine whether the model bases its decisions on relevant features or shortcuts. In the following, we explain our method by applying it to an exemplary dataset (the *Lemon Quality* dataset, cf. ). In , we demonstrate that our approach is a reliable and easy way to identify shortcuts in various image and audio datasets. ### Identification of Feature-Target Correlations {#ss: identification of feature-target correlations} Forwarding an image \(x^{(i)}\) through the encoder of a trained Beta-VAE yields the parameters \(\mu^{(i)}\)and \(\sigma^{(i)}\) of the posterior \(q_{\phi}(z|x^{(i)})\). The mean \(\mu^{(i)}\) can be considered as the latent representation for \(x^{(i)}\). We forward the entire dataset through the trained VAE, obtaining \(\mu^{(i)}\) for all \(i \in \{1, \ldots, N\}\). We utilize these representations to determine the feature-target correlations in the data. We propose two different methods for this analysis. **Statistical Analysis on the Latent Space.** For each latent dimension \(j\) and all target classes \(c \in \{1,..., C\}\), we analyze the distributions \(p(z_j|y=c)\), c.f. . We are interested in finding those dimensions \(j\) where two different classes result in two highly disparate distribution estimates. This indicates that feature \(z_j\) is highly correlated with the target classes (a necessary, but not sufficient requirement for a shortcut). The separation between any two distributions is quantified in terms of the Wasserstein distance between them. We hypothesize that the maximum pairwise Wasserstein distance (MPWD) for a particular dimension represents its capability to separate classes. The \(k\) dimensions with the largest MPWD are selected as candidates for shortcut evaluation. Additionally, we visualize the distributions for a particular latent dimension using Kernel Density Estimation (KDE) . For each latent dimension \(j\) and class \(c \in \{1,..., C\}\), we estimate \(p(z_j|y=c)\) over the training data \(\{x^{(i)} | y^{(i)} = c\}_{i=1}^N\): \[\begin{aligned} p(z_j|y=c) = \frac{1}{h K_c} \sum_{n=1}^{|K_c|} \frac{1}{\sqrt{2 \pi}} \text{exp}\left({\frac{\left(z_j-\mu_j^{(n)}\right)^2}{2h^2}} \right) \end{aligned}\] where \(h\) is the bandwidth parameter and \(K_c = \{i | y^{(i)} = c\}\) are the indices of all images belonging to class \(c\). illustrates this approach for the latent dimension \(2\) of a Beta-VAE trained on the *Lemon Quality* dataset. **Classification with VAE Encoder.** Another approach to understanding the correlation between \(z_j\) and \(y\) is to employ a classifier to predict \(y\) given \(z\). We pick the encoder of our trained Beta-VAE and append a fully connected layer. After freezing the encoder backbone, we optimize its dense classification head. The weights in the last layer of this model denote the correlation between the latent dimensions and the target classes. For a dense classification head \(h(z) \to y\), we define the predictiveness of feature \(z_j\) as \[\begin{aligned} \text{pred}(z_j) = \sum_c |\theta_{jc}| \end{aligned}\] where \(\theta_{jc}\) is the weight of the neuron in \(h\) which maps input \(z_j\) to output \(h(z_j)_c\). We select the top \(k\) dimensions with the largest predictiveness as candidates for shortcut evaluation. ### Evaluation of Feature-Target Correlations for Shortcuts {#ss: evaluation of feature-target correlations} After identifying the most predictive features in the latent space, we can now evaluate them for shortcuts. To facilitate the distinction between valid and spurious correlations, we provide a human judge with visualizations that convey the meaning of the features. **Visualizing \(z_j\) via Latent Space Traversal**. The first approach to understanding high-level features encoded in a latent dimension \(z_j\) is to visualize the effect of traversal in that dimension using the trained Beta-VAE. Given an instance \(x^{(i)}\), we obtain \(z^{(i)} = \mu^{(i)}\) from the encoder and then visualize the decoder output \(Dec(z^{(i)}+\lambda)\). The linear interpolation in the latent space is specified by \(\lambda\) where \(\lambda_j \in [z_j^{min}, z_j^{max}]\), \(z_j^{min}=\min(\{z_j^{(i)}\}_{i=1}^N)\), \(z_j^{max}=\max(\{z_j^{(i)}\}_{i=1}^N)\) and \(\lambda_k = 0\) for \(k \neq j\). This helps us to understand whether the latent variable \(z_j\) encodes a useful feature or a shortcut. illustrates the latent traversal for the *Lemon Quality* dataset. **Visualizing Images for Extreme \(z_j\)**. To validate the meaning attributed to \(z_j\), we compute the embeddings \(z_j^{(i)}\) for all the instances \(x^{(i)}\) in the dataset, and identify those instances which minimize or maximize \(z_j\). Particularly, we perform \(\arg\operatorname{sort}_{i} (\{z_j^{(i)}\}_{i=1}^N)\) for every dimension \(j\) in the latent space and display the input images \(x^{(i)}\) corresponding to the first \(l\) and the last \(l\) indices of the sorted values. depicts \(l = 27\) images of the *Lemon Quality* dataset with minimum and maximum values in latent dimension \(2\). ### On the Necessity of Human Judges {#ss:human_intervention_necessity} Our method requires a human judge to evaluate if the information encoded by the latent variable \(z_j\) constitutes a feature or a shortcut. We argue that human interventions are inevitable for ML-based shortcut detection. ML models learn relations between input and output based on statistics alone. Unlike humans, they are hardly equipped with prior knowledge of the real world beyond the given dataset and the specified task. However, this prior knowledge may be necessary to evaluate whether a correlation is valid or spurious. Therefore, to assess the candidate feature-target correlations identified by our model, we need a human judge in the last step of our approach. This human supervision is limited to inferring the meaning of a latent dimension from the visualization of its latent traversal. # Evaluation {#s:eval} We perform shortcut detection with our proposed method on synthetic and real-world datasets. Not only does our method rediscover known shortcuts in these datasets, it also reveals a previously unknown shortcut in a real-world dataset. Additionally, we exploit the knowledge about these spurious correlations to generate shortcut adversarial examples. ## Experimental Setup {#ss:experimental_setup} To obtain a low-dimensional latent representation, we use a Beta-VAE (cf. ) with \(dim(z) \in \{10, 32\}\) depending on the data. For our encoder, we employ a ResNet backbone pretrained on the ImageNet dataset. We append two separate linear layers for predicting the parameters \(\mu\) and \(\sigma\) of the posterior distribution. The decoder of our model consists of \(5\) hidden layers, each applying transposed convolutions with padding \(1\), stride \(2\), and kernel size \(3\). The channel dimensions of the hidden layers in the decoder are chosen as follows: \(512\), \(256\), \(128\), \(64\), and \(32\). We apply ReLU activation and batch normalization in the hidden layers and Sigmoid activation in the output layer. Our model is trained using the Adam optimizer with a learning rate of \(0.001\) to optimize the loss described in . We train the model with early stopping (patience of \(10\) epochs) and a batch size of \(32\). Since \(\beta\) determines the tradeoff between reconstruction and sampling quality, we perform hyperparameter tuning on \(\beta\) for each dataset. Details are provided in the Appendix. The generation of high-quality samples indicates that our model is able to discover meaningful latent factors which is a prerequisite for shortcut detection. While our model can operate on varying input sizes, we use images of size \(128 \times 128\) in our experiments. We do not perform any data augmentation so as to retain the original characteristics of the input data including any shortcut. To estimate the predictiveness of latent features, we append a linear layer to the frozen encoder of a trained Beta-VAE. The resulting classifier with cross-entropy loss is trained with early stopping (patience of 10 epochs), the Adam optimizer and a learning rate of \(0.001\) on batches of size \(32\). We ran all experiments on an Nvidia Titan X 12GB GPU. ## Datasets {#ss:datasets} We evaluate our approach on datasets with artificially introduced shortcuts to demonstrate its ability to identify spurious correlations. Our method can also be applied to real-world datasets to reveal previously unknown shortcuts. We perform a train-val-test split in the ratio 80:10:10 on each dataset unless specified differently. **Waterbirds.** Sagawa et al. extract birds from the Caltech-UCSD Birds-200-2011 dataset and combine them with background images from the Places dataset. As a result, 95% of the water birds appear on a water background while 95% of the land birds appear on a land background. Since this spurious correlation is only introduced in the train set of the *Waterbirds* dataset consisting of 4,795 samples, we use the same to create the train-val-test splits for our experiments. **Colored MNIST.** Arjovsky et al. inject color as a spurious attribute into the MNIST dataset. Following this idea, Zhang et al. create a colored MNIST dataset consisting of five subsets with five associated colors. A fraction \(p_{corr}\) of the training samples are assigned colors based on the group they belong to. The remaining samples are assigned a random color. We follow the color assignment in and choose \(p_{corr} = 0.995\) for coloring the 70,000 MNIST samples. **CelebA.** The CelebFaces Attributes Dataset contains 202,599 images of celebrities, each annotated with 40 facial attributes. Sagawa et al. train a classifier to identify the hair color of the celebrities and discover that the target classes (blond, dark) are spuriously correlated with the gender (male, female). We stick to the setup specified in with the official train-val-test splits of the *CelebA* dataset. **Lemon Quality.** To demonstrate the detection of shortcuts in quality control, we apply our method on the *Lemon Quality* dataset. The dataset consists of 2,533 images labelled with one of three classes, namely 'good quality', 'bad quality', and 'empty background'. **COVID-19.** Following, we obtain a dataset with 112,528 samples for COVID-19 detection by combining COVID-19-positive radiographs from the GitHub-COVID repository and COVID-19-negative radiographs from the ChestX-ray14 repository. **ASVspoof.** The ASVspoof 2019 Challenge Dataset is used to train and benchmark systems for the detection of spoofed audio and audio deepfakes. Müller at al. observe that the length of the silence at the beginning of an audio sample differs significantly between benign and malicious data. Previous deepfake detection models have exploited this shortcut. We chose a subset of the training data (benign audio and attack \(A01\)), which results in a binary classification dataset comprising 6,380 samples. We transform the audio samples to CQT spectrograms , and obtain a frequency-domain representation with \(257\) logarithmically spaced frequency bins, capturing up to \(8\) kHz (Nyquist frequency given the input is \(16\) kHz). ## Results {#ss:results_eval} For every dataset under examination, we present a feature that highly correlates with the targets but is not relevant for the actual task and thus considered a shortcut. While the particular dimension representing such a feature can vary between different training runs, any VAE trained with well-chosen hyperparameters should be able to encode the feature in one of its latent dimensions. summarizes our results. The MPWD and the predictiveness of a spurious dimension are reported relative to the other latent dimensions. Across all experiments, we conclude that a human judge has to review only the top \(k = 3\) latent dimensions with the highest MPWD and predictiveness to identify a shortcut if present in the data. Both scores lead to a similar set of candidate dimensions for shortcut evaluation. Our approach correctly identifies the known shortcut in the *CelebA* dataset. The statistical analysis on the latent space of a trained VAE shows that its latent variable \(z_{26}\) has a high MPWD (among top 3) and a high predictiveness \(pred(z_{26})\) (among top 3). The latent traversal illustrated in reveals that the latent variable \(z_{26}\) encodes the gender of a person. Similarly, we detect the color shortcut in the *Colored MNIST* dataset and the background shortcut in the *Waterbirds* dataset. Illustrations of the respective latent traversals can be found in the Appendix. To the best of our knowledge, we are the first to identify the existence of a shortcut in the *Lemon Quality* dataset. As depicted in , our method reveals the correlation between lemon quality and zoom level. The good-quality lemons are mostly photographed from a distance. The bad-quality lemons, on the other hand, are mostly captured in close-up shots. The identification of the position shortcut in the *COVID-19* dataset is illustrated in and confirms the findings of DeGrave et al. . We make a step towards the generalization of our method to other domains by identifying shortcuts in spectrograms from the *ASVspoof* dataset. exhibits the latent traversal for the spurious silence feature in the *ASVspoof* dataset. A perfect disentanglement in the latent space facilitates the identification of the semantic meaning of a dimension. However, the existence of shortcuts in a dataset makes it difficult for a model to fully separate the factors of variation. Thus, we sometimes observe (see ) that the change of a spurious attribute (e.g. position) coincides with the change of the main distinguishing factor (e.g. lung appearance). ## Shortcut Adversarial Examples {#ss:adv_ex} In addition to identifying the underlying shortcuts in a given dataset, our method enables to create shortcut adversarial examples. Adversarial examples are specifically crafted data points \(\hat{x} = x + \delta\) that are recognized by humans as their true class \(y\), but classified by a model \(f\) as a different class \(t \not =y\). Such examples can for instance be found by solving the following minimization problem \[\argmin_\delta L(f(x + \delta), t) + \lambda \| \delta \|_2^2 \label{eq:adv}\] where \(L\) is the training loss and \(\lambda\) constrains the magnitude of the perturbation \(\delta\). is usually solved using gradient-based optimization, which is computationally expensive and either requires access to the model parameters or necessitates the use of surrogate models, e.g. in transfer-based attacks. The existence of shortcuts paves the way for a new approach of generating adversarial examples. Shortcut adversarial examples can be created by modifying an original image such that it carries a shortcut feature that the model associates with another class. This modification of the image, however, does not change its class as perceived by humans. For example, by cropping the centre of an image showing a good-quality lemon, we mimic a close-up shot of the fruit. Since the model associates close-up shots with bad-quality lemons, it classifies the modified image as 'bad quality'. Similarly, we mimic a distant shot by padding images of bad-quality lemons with empty background pixels. illustrates the creation of shortcut adversarial examples for the *Lemon Quality* dataset. Based on the shortcuts discovered, we only perform a simple image manipulation instead of iteratively computing gradients for the optimization of [\[eq:adv\]](#eq:adv){reference-type="ref" reference="eq:adv"}. Furthermore, our method does not require access to the target model parameters and is computationally less expensive than established proxy attacks. Shortcut adversarial examples are related to backdoor attacks where an adversary injects a trigger into the training dataset to cause an ML model to learn certain adversarial patterns. The model under attack behaves correctly on benign test samples while its predictions turn faulty when the embedded backdoor is triggered. Shortcuts can be seen as backdoors inherent to a given dataset. We do not perform backdoor attacks as we do not inject shortcuts into the training dataset. Instead, we exploit their natural occurrence in the data to generate adversarial examples on which a model is likely to fail at test time. Our shortcut adversarial examples are effective and model agnostic. We employ a CNN classifier comprising 5 convolutional layers with 32, 64, 128, 256, and 512 filters of size \(3 \times 3\). Each of these layers is followed by ReLU activation and max pooling. When trained on the *Lemon Quality* train set, this model reaches an overall accuracy of \(98.4\%\) on the unperturbed *Lemon Quality* test set. Applying the model on shortcut adversarial examples created from this test set decreases the test accuracy for good lemons from \(98.3\%\) to \(65\%\) and for bad lemons from \(97.7\%\) to \(86.6\%\). This is because the classifier learns to rely on the spurious correlation between the zoom level and target class for its prediction. ## Limitations A VAE used for shortcut detection is required to learn meaningful, disentangled factors in its latent space. The Beta-VAE employed in our method can encode a disentangled representation only if the underlying factors of a dataset are independent. Our method relies on this assumption. Through our experiments, we discovered that a Beta-VAE with ResNet encoder lacks the ability to model datasets with a large number of visually diverse classes such as ImageNet. More powerful models like NVAE ) could enhance the modelling and sampling quality on such datasets. However, the size of their latent space and the hierarchical entanglement of features limits their usage in shortcut detection. Finally, we note that VAEs are sensitive to the choice of hyperparameters. The number of latent dimensions should be in accordance with the number of factors of variation in a dataset. The weights assigned to the KL divergence and reconstruction loss are vital for the overall performance of a VAE and thus have to be finetuned for every dataset. # Conclusion {#s:conclusion} In this paper, we introduce a novel approach to detect spurious correlations in machine learning datasets. Our method utilizes a VAE to discover meaningful features in a given dataset and enables a human judge to identify shortcuts effortlessly. We evaluate our approach on image and audio datasets and successfully reveal inherent shortcuts. We further demonstrate how the knowledge of these spurious correlations can be exploited to generate shortcut adversarial examples. We hope that our work inspires researchers to explore the potential of VAEs in the field of shortcut learning. It would be interesting to see if our method can be extended to other domains.

{'timestamp': '2023-02-09T02:17:31', 'yymm': '2302', 'arxiv_id': '2302.04246', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04246'}

# Introduction The miniaturization of optical systems, such as microscopes, telescopes, and cameras, has long been a topic of interest for physicists. By designing smaller optical systems, we can increase their versatility and reduce their cost. Most work has aimed at making the lenses thinner, e.g., metalenses [@Moon2022; @Engelberg2020; @Khorasaninejad2016; @Liang2018; @Chen2017; @Overvig2019; @Yu2014; @Li2022], diffractive lenses [@Sweeney1995; @Faklis1995; @Banerji2019; @Meem2018; @Kim2013; @Overvig2022], and Fresnel lenses [@Joo2022]. Instead, our aim is to miniaturize the space *between* lenses, an often neglected element in optical systems. Since this is usually the largest contributor to the total length of an imaging system, its miniaturization promises the largest impact. Recently, the \"spaceplate\" was introduced as an optical element that compresses the effect of propagation on light into a plate [@Reshef2021]. Such a device, by definition, must replace a slab of free-space with thickness \(d_{\text{eff}}\) while only occupying a physical thickness \(d\) less than \(d_{\text{eff}}\). The compression ratio, \(\mathcal{R}=d_{\text{eff}}/{d}\), is used to characterize spaceplate designs [@Reshef2021]. Like free-space propagation, a spaceplate changes the height of a ray but leaves its angle unchanged. Thus, unlike a lens, which does change ray angle, a spaceplate does not alter the magnification of an imaging system. From the view of physical optics, the spaceplate changes the phase of an incident plane wave without altering its angle. A variety of optical devices have been proposed to compress space. The work of Reshef et al. [@Reshef2021] first introduced and experimentally demonstrated the spaceplate. They experimentally tested two designs, one being a uniaxial crystal and the other a low-index medium. The designs were found to have compression ratios of \(\mathcal{R}=1.12\) and \(1.48\), respectively, which resulted in each spaceplate saving less than \(\SI{4}{mm}\) of space. Remarkably, the uniaxial crystal had a large numerical aperture (\(\text{NA}=\sin\theta\)) and functioned at high incident angles \((\text{NA}=0.85)\) while remaining broadband, i.e., highly transmissive across the visible spectrum. In the same paper, the group also proposed and modeled a multilayer thin-film design. This had a compression ratio of \(\mathcal{R}=5.4\), but was only capable of saving \(\SI{50}{\micro m}\) of space and was narrowband. In another, almost simultaneous work, Guo et al. [@Guo2020] numerically demonstrated a much higher compression ratio of \(\mathcal{R}=144\) using the Fano resonances of a photonic crystal slab at wavelength \(\lambda\). Unlike the uniaxial crystal, however, this design was significantly limited in NA \((\text{NA}=0.01)\) and the actual space saved was limited to approximately \(360\lambda\) [@Guo2020]. Subsequent designs, including those made of homogeneous materials [@Reshef2021], improved multilayer stacks [@Page2022], repeated Fabry-Perot cavities [@Chen2021; @Mrnka2022], and others [@Long2022; @Chen2022; @Ivanov2022], made incremental improvements upon compression ratio and NA. None, however, can be scaled up to save space on the order of meters. Additionally, most of the previously demonstrated designs are limited in either polarization or bandwidth [@Reshef2021; @Guo2020; @Page2022; @Chen2021]. In this paper, we present a broadband, polarization-independent three-lens spaceplate capable of meter-scale compression. The three-lens spaceplate \[Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"})\] utilizes the Fourier plane of a 4-\(f\) lens system to impart the same angle-dependent phase that free-space propagation imparts. Our 4-\(f\) lens system consists of two positive lenses of focal length \(f_{\text{ext}}\), the \"exterior lenses\". The first lens maps the complex amplitude distribution of an incident field across the front focal plane to the back focal plane via a Fourier transform [@Goodman2017]. Thus, there is a one-to-one mapping between the angle of an incoming plane wave and the position in the back focal plane, known as the \"Fourier plane\", which is located a distance \(f_{\text{ext}}\) after the lens. The second lens is a distance \(2f_{\text{ext}}\) after the first lens and performs another Fourier transform, undoing the action of the first lens. With a spatially-varying phase element (i.e., a \"phase mask\") in the Fourier plane, the 4-\(f\) system phase shifts each incoming plane wave but otherwise leaves it unchanged. The angle-dependent phase shift required to emulate free-space propagation happens to correspond to the phase mask created by a thin positive lens. Consequently, such a lens placed in the Fourier plane (i.e., a middle lens with focal length \(f_{\text{mid}}\)) will cause the field transmitted by the 4-\(f\) system to appear as though it had propagated through a slab of free-space, potentially longer than the 4-\(f\) system itself. Note that while this three-lens device is neither monolithic nor plate-like, we still refer to it as a spaceplate due to its similar effect on light. Note as well that the use of a 4-\(f\) system results in inversion of the light about the \(\mathbf{z}\)-axis (\(x=0\)); while this does not affect the compression or absolute magnification of the device, it does differentiate the device from other spaceplates. In the next section, we theoretically demonstrate how this three-lens design can be used to attain free-space compression on a much larger scale than previous spaceplate designs. In Section [9](#sec:exp){reference-type="ref" reference="sec:exp"}, we experimentally demonstrate the three-lens spaceplate and characterize its performance for a number of lens choices. We show how it can be straightforwardly inserted into the design of long imaging systems, even if they are broadband. In Section [10](#sec:fundamentalLimits){reference-type="ref" reference="sec:fundamentalLimits"}, we outline some fundamental theoretical limits on the performance of the three-lens spaceplate. We focus on compression ratio, NA, and bandwidth, which recent work has identified trade-offs between [@Chen2021; @Shastri2022] for other spaceplate designs. In the last two sections, we compare and contrast the three-lens design to related optical systems and discuss prospects for its application. # Fourier optics description of a three-lens spaceplate {#sec:FourierOptics} To motivate the design of the three-lens spaceplate, we follow the Fourier optics description of the spaceplate effect introduced by Reshef et al. [@Reshef2021]. Suppose we have a light field propagating through free-space centered on the \(\mathbf{z}\)-axis. We then consider how each transverse spatial Fourier component of this initial field is transformed by this propagation. Each of these components, or plane waves, is described by its momentum vector, \(\mathbf{k} = \left(k_{x},k_{z} \right) = k(\sin \theta, \cos \theta)\), where \(\theta\) is the angle between the \(\mathbf{z}\)-axis and the direction of the plane wave. When propagated over a distance \(z_0\) in free-space, the amplitude and direction of each plane wave is conserved, but the phase is not. Between two points on the \(\mathbf{z}\)-axis separated by \(z_0\), each plane wave accumulates phase \(\varphi = k_{z} z_0 = |\mathbf{k}|\cos(\theta)z_0\). A spaceplate must produce this same phase response, but in a shorter distance \(d\) than \(z_0\). An ideal spaceplate, if inserted into the light path, will produce the phase response \(\varphi_{\text{SP}} = k_{z}z_0 = z_0(k^2-k_x^2)^{1/2}\). In the small angle approximation, \(\varphi_{\text{SP}} = z_0 k(1-k_x^2/2k^2)\). Then, by neglecting global phase terms, we arrive at our target phase, \(\varphi_{\text{SP}} =-z_0 k_{x}^2/2k\). An approach for considering larger incident angles is given in Appendix [\[sec:appendixLargeAngle\]](#sec:appendixLargeAngle){reference-type="ref" reference="sec:appendixLargeAngle"}. In the Fourier plane of the 4-\(f\) system in Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}a), the field at position \(r\) (i.e., axial distance) is the incoming field component with transverse momentum \(k_{x} = rk/f_{\text{ext}}\). Substitution of this relation into the target phase gives the spatial phase mask required to emulate propagation through space, \[\begin{aligned} \varphi_{\text{SP}} =-\frac{z_0{kr^2}}{2f_{\text{ext}}^2}. \label{eq:spaceplatePhase} \end{aligned}\] The phase added by a spherical, thin lens of focal length \(f_{\text{mid}}\) in the paraxial approximation is given by \(\varphi_{\text{mid}} =-kr^2/2f_{\text{mid}}\), which is of the same form as Eq. ([\[eq:spaceplatePhase\]](#eq:spaceplatePhase){reference-type="ref" reference="eq:spaceplatePhase"}). By setting \(\varphi_{\text{SP}}=\varphi_{\text{mid}}\) we find that the spaceplate effectively propagates light a distance \(z_0 = f_{\text{ext}}^2/f_{\text{mid}}\). In this way, we can compress space by placing a positive lens at the Fourier plane. A slight complication is that, as part of the 4-\(f\) system, there already is free-space propagation of distance \(f_{\text{ext}}\) before the first lens and \(f_{\text{ext}}\) after the last lens. We do not include those regions in the length of space replaced and thus, \(d_{\text{eff}}=z_0-2f_{\text{ext}}\). Similarly, we define the spaceplate thickness to solely be the distance *between* the external lenses, \(d=2f_{\text{ext}}\). Therefore, the compression ratio is \[\begin{aligned} \mathcal{R} =\frac{d_{\text{eff}}}{d}= \frac{f_{\text{ext}}}{2f_{\text{mid}}}-1, \label{eq:compressiveRatio1} \end{aligned}\] which only depends on the focal lengths used. If \(\mathcal{R} > 1\) or, equivalently, if \(f_{\text{mid}}<f_{\text{ext}}/4\), more space is replaced than is occupied by the three-lens spaceplate. Further discussion of system length and compression ratio is given in Appendix [\[sec:appendixCompressionRatio\]](#sec:appendixCompressionRatio){reference-type="ref" reference="sec:appendixCompressionRatio"}. # Experimental demonstration of three-lens space compression {#sec:exp} ## Design of the focus advancement, walkoff, and imaging experiments {#sec:expDesign} The setup used to test the three-lens spaceplate is shown in Fig. ([\[fig:expSetUp\]](#fig:expSetUp){reference-type="ref" reference="fig:expSetUp"}). We measure the advance of the focus of a beam, the walk-off of a beam incident at an angle, as well as imaging properties. These three types of measurements require minor setup variations that will be detailed in the following three subsections. We start by describing elements in the setup common to all three measurements. The external lenses, \(f_{\text{ext}}=\SI{150}{mm}\), are the same throughout all the tests and have \(\SI{50.8}{mm}\) diameters and \(\SI{15.0}{mm}\) thicknesses. We test the effect of the following series of middle lenses: \(f_{\text{mid}}= \{40, 30, 25, 19, 18, 11, 10, 4.51\}\,\SI{}{mm}\) whose respective diameters and thicknesses were \(\{25.4,25.4,12.7,12.7,6.5,8.0, 5.5\}\,\SI{}{mm}\) and \(\{12.5, 14.0, 7.0, 6.0, 2.2,5.0, 6.5, 2.94\}\,\SI{}{mm}\). With the exception of the \(f_{\text{mid}}=\{18,11,4.51\}\SI{}{mm}\) aspheric lenses, all lenses used throughout these tests were achromatic doublets, designed and anti-reflective (AR) coated for the visible range (400-\(\SI{700}{nm}\) light). The \(f_{\text{mid}}= \SI{18}{mm}\) aspheric lens was designed for \(\SI{780}{nm}\) light, and the \(f_{\text{mid}}= \{11,4.51\}\,\SI{}{mm}\) aspheric lenses were designed for \(\SI{633}{nm}\) light. All aspheric lenses were AR-coated for 600-\(\SI{1100}{nm}\). After the spaceplate, an image sensor (\(\SI{5.86}{\micro m} \times \SI{5.86}{\micro m}\) pixel size, \(1936\times1216\) pixels, color) was mounted on a motorized translation stage capable of moving both in the transverse (\(x\)) and longitudinal (\(z\)) directions. With this sensor, we recorded \(x\times y\) spatial intensity distributions at chosen \(z\) positions. The electronic settings of the image sensor were constant throughout the various tests. ## Focus advancement of various three-lens spaceplates {#sec:focusAdvancement} We measure the focus advancement of a \(\SI{633}{nm}\) wavelength Helium-Neon laser that is collimated so that its beam has an intensity full-width-half-maximum (FWHM) of \(\SI{3.5}{mm}\) using three different middle lens focal lengths corresponding to three distinct compression ratios. To create a focusing beam, the laser travels through a \(\SI{50.8}{mm}\) diameter positive lens with a thickness of \(\SI{11.3}{mm}\) and a focal length of \(f_{\text{image}}=\SI{1000}{mm}\) \[see Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}a)\]. The three-lens spaceplate is placed between this positive lens and the nominal focus \(z\)-position, then the image sensor records intensity distributions over a range of \(\SI{500}{mm}\) with a \(\SI{0.25}{mm}\) step size. Each \(x\) pixel row of the recorded distribution is summed to determine the row with the maximal intensity. This row becomes the 1D intensity profile at each \(z\), which is plotted in Fig. ([\[fig:2022914_focusComparison\]](#fig:2022914_focusComparison){reference-type="ref" reference="fig:2022914_focusComparison"}). For each middle lens, the peak intensity along \(z\) is taken as the new focal position. We measure longitudinal focal shifts of \(\Delta z = \SI{+94}{mm}\), \(\SI{-49}{mm}\), and \(\SI{-221}{mm}\) corresponding to observed compression ratios \(\mathcal{R_{\text{obs}}}= 0.69\), \(1.16\), and \(1.74\), respectively. The respective nominal compression ratios are \(\mathcal{R} = 0.9, 1.5,\) and \(2.0\). While these are significantly different from the measured values, they demonstrate that the three-lens spaceplate can successfully replace space. Deviations between \(\mathcal{R}\) and \(\mathcal{R_{\text{obs}}}\) result because the three-lens spaceplate relies on a lens located directly at the Fourier plane of the 4-\(f\) system, and any deviation of the thick middle lens from this position results in a reduction of compression and non-unity magnification. Specifically, we observe that upon transmitting an expanded collimated beam with a \(\text{FWHM}=\SI{3.5}{mm}\) through the three-lens spaceplate, the resultant beam width varies (\(1.7-\SI{4.4}{mm}\)) depending on small adjustments (\(0-\SI{5}{mm}\)) to the longitudinal position of the middle and rear lenses in the spaceplates. The effect of these sensitivities are reduced in subsequent walkoff and imaging experiments. Nonetheless, the observed negative longitudinal shift demonstrates that this design is capable of replacing free-space on a scale of three orders of magnitude greater than previous spaceplate designs. This experiment is a qualitative demonstration of space compression-the following experiments are used to characterize the three-lens spaceplate with greater accuracy. ## Transverse beam walkoff of a three-lens spaceplate We next measure the transverse walk-off \(\Delta w = |x_{\text{SP}}|-|x| = d(\mathcal{R}-1)\sin{\theta}\) \[see Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}b) for a diagram\] of a beam transmitted through the spaceplate, incident at angle \(\theta\) to the \(z\)-axis, rotating about the \(y\)-axis. For each spaceplate, the walk-off as function of angle characterizes the NA and aberrations, and determines the compression ratio with greater accuracy than in the focal advance measurement. Specifically, before conducting walkoff experiments, the longitudinal positions of the middle and rear lenses are incrementally adjusted until a beam (\(\text{FWHM}=\SI{3.5}{mm}\)) transmitted through the spaceplate becomes collimated and has unity magnification. That is, until the exiting beam has diameter of \(\text{FWHM}=(3.5\pm0.1)\,\SI{}{mm}\) both immediately after the rear lens of the spaceplate, and one meter after it. To test the beam walkoff, the imaging lens and beam-expander are absent, leaving the laser beam collimated. The rotation stage in Fig. ([\[fig:expSetUp\]](#fig:expSetUp){reference-type="ref" reference="fig:expSetUp"}) varied \(\theta\) from 0 to 2 degrees in 0.02 degree increments. The stage is located \(f_{\text{ext}}\) before the first lens, and we measure the transverse displacement \(z = 4f_{\text{ext}}=\SI{600}{mm}\) from the rotation axis of the stage. The rotation stage is placed here due to our definition of \(\text{NA}=\sin\theta\), where \(\theta\) is the half-angle of a maximum cone accepted by the system, originating \(f_{\text{ext}}\) before the first lens. By putting the stage at this position, we are able to contextualize the experimental angular limitations of the system with an analysis of its numerical aperture. The position at which we measure the transverse displacement, however, does not matter in general. One will measure the same \(\Delta w\) regardless of the position of the imaging plane. First, we measure the transverse displacement \(x(\theta)\) of a beam from the \(z\)-axis in free space, and then perform analogous measurements \(x_{\text{SP}}(\theta)\) of beams transmitted through each spaceplate. Due to inversion, the beams are displaced in opposite directions, but we can still compute the walkoff as: \(\Delta w = |x_{\text{SP}}|-|x|\). We measure \(x\) and \(x_{\text{SP}}\) by fitting Gaussians to the measured transverse intensity profiles. Fig. ([\[fig:2022914_compressiveFactorAndWalkoff\]](#fig:2022914_compressiveFactorAndWalkoff){reference-type="ref" reference="fig:2022914_compressiveFactorAndWalkoff"}a) presents this series of walkoff measurements for each of the eight middle lenses given in Section [9.1](#sec:expDesign){reference-type="ref" reference="sec:expDesign"}, with corresponding observed compression ratios in Fig. ([\[fig:2022914_compressiveFactorAndWalkoff\]](#fig:2022914_compressiveFactorAndWalkoff){reference-type="ref" reference="fig:2022914_compressiveFactorAndWalkoff"}b). Each of the configurations produced compression ratios within error of their nominal compression ratios over the entire measured range. For the highest experimental compression ratio, \(\mathcal{R}=15.6\), \(f_{\text{mid}}=\SI{4.51}{mm}\), the spaceplate replaced 4.39 meters of free-space, which is more than one-thousand times greater than that demonstrated by other spaceplate designs. ## Broadband imaging with a three-lens spaceplate In our final measurement, we investigate the performance of the three-lens spaceplate for imaging with full-visible-spectrum light. We first illuminate a postage stamp by white-light and use the \(f_{\text{image}}\) lens to form a sharp image on the image sensor. The object and image distances are approximately \(\SI{1.9}{m}\) and \(\SI{2.1}{m}\), respectively, and we measure a magnification of 1.1. We record images with and without a \(f_{\text{mid}}=\SI{10}{mm}\) spaceplate \[Fig. ([\[fig:20220915_fullColorImageComp\]](#fig:20220915_fullColorImageComp){reference-type="ref" reference="fig:20220915_fullColorImageComp"}c), left and right panes, respectively\] in the imaging path. Before imaging, the middle and rear lenses of the spaceplate are calibrated using an expanded collimated beam, as was done in the walkoff experiment. The camera is then placed at the nominal longitudinal position relative to the focus position seen without a spaceplate in the imaging path. Then, the middle and rear lens longitudinal positions are adjusted slightly to achieve the same magnification seen without the spaceplate. To reduce aberration and edge-effects, a large iris was positioned immediately before the third lens in the spaceplate and was adjusted until the highest contrast and resolution were seen. This was at an aperture size of \((12\pm2)\,\SI{}{mm}\). While there still is a reduction in contrast and brightness, which we discuss later, the images show that the three-lens spaceplate can advance the image plane of a full-visible-spectrum imaging system by 1.65 m. Next, we image an NBS 1963a resolution target in order to characterize the reduction in resolution and contrast caused by the addition of four different three-lens spaceplates \[see Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"})\]. The addition of the spaceplates does not change the magnification, as expected. However, Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}b) shows that every spaceplate reduces image contrast and resolution in comparison to the image formed in the absence of a spaceplate. We quantify this reduction with the modulation transfer function (MTF) of the imaging system following the method of Saiga et al. [@Saiga2018] in Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}d). For the case of \(f_{\text{mid}} = \SI{10}{mm}\) (\(\mathcal{R} = 6.5\)) and the stamp imaged without an iris, the spaceplate lowers the spatial frequency at half-max of the MTF by 45% (i.e. \(10.9\,\text{lp/mm}\) is reduced to \(6.0\,\text{lp/mm}\)). Three potential causes for this reduction are aberrations, scattering from lens edges, and diffraction by the lens apertures. The first two would be mitigated by a reduced aperture, whereas the effect of diffraction would be worsened. Adjusting the diameter of an iris \([(12\pm2)\,\SI{}{mm}]\) just before the third lens, we take the images shown in Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}c), which are qualitatively improved in contrast and resolution. This improvement is reflected in the MTF measured with an iris in place, which increases by 38% (i.e. \(6.0\,\text{lp/mm}\) to \(8.3\,\text{lp/mm}\)) suggesting that diffraction is not the main cause of the reduction. Equivalently stated, addition of the spaceplate and iris causes a 24% reduction in the MTF (i.e. \(10.9\,\text{lp/mm}\) to \(8.3\,\text{lp/mm}\)). # Fundamental limits on performance of the three-lens spaceplate {#sec:fundamentalLimits} In this section, we briefly discuss limits on the NA and compression inherent to the three-lens design of the spaceplate. In most physically realizable setups, the NA of the three-lens spaceplate will be limited by the diameter of the third lens, \(D_3\). For this common case, the NA can be expressed in terms of the f-number of the third lens, \(N_3 = f_{\text{ext}}/D_3\), and \(\mathcal{R}\): \[\begin{aligned} \text{NA} = \left[1+4N_3^2\left(1+2\mathcal{R}\right)^{2}\right]^{-\frac{1}{2}}. \label{eq:numericalaperturethirdlens} \end{aligned}\] Eq. ([\[eq:numericalaperturethirdlens\]](#eq:numericalaperturethirdlens){reference-type="ref" reference="eq:numericalaperturethirdlens"}) shows that there is trade-off between compression ratio and NA. The lower the f-number of the third lens, the less strict this trade-off is. The third lens used in our experiment has \(N_3=3\), which limits the NA of the three-lens spaceplate be less than 0.06, worsening as \(\mathcal{R}\) increases. This represents the greatest shortcoming of the three-lens spaceplate design: even when using a relatively large third lens diameter (\(D_3=\SI{50}{mm}\)), the maximum incident angle is three degrees from the optical axis. For information on the less common cases where the first or second lenses limit the NA, see Appendix [\[sec:appendixNA\]](#sec:appendixNA){reference-type="ref" reference="sec:appendixNA"}. We have considered two additional physical mechanisms inherent to the three-lens spaceplate that lead to limitations. However, these turn out to be much less restrictive than the above limitation in practical applications with standard lenses. Firstly, the Abbe Sine Condition limits the NA independent of the lens diameters. Secondly, diffraction from the first lens in the spaceplate limits the achievable compression ratio. These two limitations may be relevant for new types of lenses, such as metalenses, and are therefore explored in depth in Appendices ([\[sec:appendixAbbe\]](#sec:appendixAbbe){reference-type="ref" reference="sec:appendixAbbe"}) and ([\[sec:appendixDiffraction\]](#sec:appendixDiffraction){reference-type="ref" reference="sec:appendixDiffraction"}), respectively. # Comparison to other compressive optics A number of distinctions must be made between the three-lens spaceplate and other methods to conserve space in imaging systems. We start by comparing it to other spaceplate designs. All of these exhibit some degree of transverse translation invariance [@Shastri2022], whereas the three-lens spaceplate does not, since lenses are not translationally invariant. Another difference is that the three-lens spaceplate inverts the transmitted field through the axis of the lenses (i.e., \(x\rightarrow-x\) and \(y\rightarrow-y\)), though this is avoided simply by using a pair of spaceplates. Many conventional imaging systems are already designed to minimize the space they occupy. Examples include the telephoto lens and the Cassegrain reflector whose goal is to create an imaging system shorter than its effective focal length \(f_{\text{eff}}\). Thus the compression of space is intertwined with the main purpose of the imaging system-to provide optical magnification. In practice, telephoto lenses are limited to a length \(> 0.8 f_{\text{eff}}\) [@Kingslake1978; @Tremblay2005; @Tremblay2007]. This is equivalent to a non-telephoto lens followed by a three-lens spaceplate with \(\mathcal{R}\geq 1/0.8=1.25\), much lower than the highest value reported here for three-lens spaceplates \(\mathcal{R}=15.7\), albeit with a higher NA. The nature of the components of these three-lens systems means that they are large. In this way, they differentiate themselves from thin spaceplates. Because they use lenses, implementable three-lens spaceplates have limitations on thickness, aperture, and, consequentially, compression ratio. Using off-the-shelf components and manageable system lengths (\(2f_{\text{ext}}\lesssim\SI{1}{m}\)), compression ratios are limited (\(\mathcal{R}\lesssim 20\)); however, due to the large size of these systems, they can save large absolute quantities of space (on the order of meters). This is in contrast to existing spaceplate designs, which only save space on the order of microns or millimeters [@Reshef2021; @Guo2020; @Page2022; @Chen2021]. One advantage of using standard optical components in three-lens spaceplates is that the resulting spaceplate is broadband, polarization-independent, and highly configurable in compression ratio. The compression ratio of the three-lens spaceplate can easily be changed by replacing the middle lens, whereas the compression ratio of other spaceplate designs is fixed upon manufacturing. Similar systems have been used in microscopy to rapidly tune imaging planes [@Zuo2013; @Chen2018; @Kang2020; @Zuo2015] in transport-of-intensity phase microscopy. However, such implementations do not compress space and are instead intended to artificially increase the depth of field. # Conclusion In this paper, we have successfully demonstrated the compression of free-space using a device that we call a \"three-lens spaceplate\", a positive lens placed at the Fourier plane of a 4-\(f\) optical setup. We have shown that our spaceplate design can be used for meter-scale space compression and the miniaturization of broadband imaging systems. Using the three-lens spaceplate, we have achieved experimental compression ratios up to \(\mathcal{R}_{\text{obs}}=15.7\) that replace 4.4 meters of free-space. Until now, replication of the transfer function of free-space over lengths of this scale had not been demonstrated: previously, space compression of only one or two millimeters had been shown [@Reshef2021]. The three-lens spaceplate has not only shown that this is possible, it has accomplished this using off-the-shelf optical components. Moreover, the system was shown to reduce the length of both monochromatic and full-color imaging systems with a 24% loss of resolution as quantified using its modulation transfer function. Finally, we discussed the trade-off between compression ratio and numerical aperture. Specifically, we found that the numerical aperture is highly limited for the three-lens spaceplate and represents its greatest shortcoming. The simplicity and versatility of this spaceplate design motivate its application and future development, though there are certain limitations which differentiate its applications from other spaceplates. The three-lens spaceplate represents a simple, cost-efficient method for reducing the size of axially-symmetric, long focal length systems such as telescopes. Because these three-lens spaceplates are so large and have such limited numerical apertures, however, their application in smaller devices such as smartphone cameras or virtual reality headsets would be difficult to implement. That said, the inherent use of free-space within the system itself implies the potential for even higher compression ratios through the use of metalenses and other space-saving devices. Metalenses could potentially also reduce the numerical aperture limitations of the three-lens spaceplate by reducing lens thickness, and increasing aperture size. Future work includes using such hybrid designs to reduce overall system length and increase numerical aperture, as this could make the implementation of three-lens spaceplates in small devices more realistic. # Introduction The miniaturization of optical systems, such as microscopes, telescopes, and cameras, has long been a topic of interest for physicists. By designing smaller optical systems, we can increase their versatility and reduce their cost. Most work has aimed at making the lenses thinner, e.g., metalenses [@Moon2022; @Engelberg2020; @Khorasaninejad2016; @Liang2018; @Chen2017; @Overvig2019; @Yu2014; @Li2022], diffractive lenses [@Sweeney1995; @Faklis1995; @Banerji2019; @Meem2018; @Kim2013; @Overvig2022], and Fresnel lenses [@Joo2022]. Instead, our aim is to miniaturize the space *between* lenses, an often neglected element in optical systems. Since this is usually the largest contributor to the total length of an imaging system, its miniaturization promises the largest impact. Recently, the \"spaceplate\" was introduced as an optical element that compresses the effect of propagation on light into a plate [@Reshef2021]. Such a device, by definition, must replace a slab of free-space with thickness \(d_{\text{eff}}\) while only occupying a physical thickness \(d\) less than \(d_{\text{eff}}\). The compression ratio, \(\mathcal{R}=d_{\text{eff}}/{d}\), is used to characterize spaceplate designs [@Reshef2021]. Like free-space propagation, a spaceplate changes the height of a ray but leaves its angle unchanged. Thus, unlike a lens, which does change ray angle, a spaceplate does not alter the magnification of an imaging system. From the view of physical optics, the spaceplate changes the phase of an incident plane wave without altering its angle. A variety of optical devices have been proposed to compress space. The work of Reshef et al. [@Reshef2021] first introduced and experimentally demonstrated the spaceplate. They experimentally tested two designs, one being a uniaxial crystal and the other a low-index medium. The designs were found to have compression ratios of \(\mathcal{R}=1.12\) and \(1.48\), respectively, which resulted in each spaceplate saving less than \(\SI{4}{mm}\) of space. Remarkably, the uniaxial crystal had a large numerical aperture (\(\text{NA}=\sin\theta\)) and functioned at high incident angles \((\text{NA}=0.85)\) while remaining broadband, i.e., highly transmissive across the visible spectrum. In the same paper, the group also proposed and modeled a multilayer thin-film design. This had a compression ratio of \(\mathcal{R}=5.4\), but was only capable of saving \(\SI{50}{\micro m}\) of space and was narrowband. In another, almost simultaneous work, Guo et al. [@Guo2020] numerically demonstrated a much higher compression ratio of \(\mathcal{R}=144\) using the Fano resonances of a photonic crystal slab at wavelength \(\lambda\). Unlike the uniaxial crystal, however, this design was significantly limited in NA \((\text{NA}=0.01)\) and the actual space saved was limited to approximately \(360\lambda\) [@Guo2020]. Subsequent designs, including those made of homogeneous materials [@Reshef2021], improved multilayer stacks [@Page2022], repeated Fabry-Perot cavities [@Chen2021; @Mrnka2022], and others [@Long2022; @Chen2022; @Ivanov2022], made incremental improvements upon compression ratio and NA. None, however, can be scaled up to save space on the order of meters. Additionally, most of the previously demonstrated designs are limited in either polarization or bandwidth [@Reshef2021; @Guo2020; @Page2022; @Chen2021]. In this paper, we present a broadband, polarization-independent three-lens spaceplate capable of meter-scale compression. The three-lens spaceplate \[Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"})\] utilizes the Fourier plane of a 4-\(f\) lens system to impart the same angle-dependent phase that free-space propagation imparts. Our 4-\(f\) lens system consists of two positive lenses of focal length \(f_{\text{ext}}\), the \"exterior lenses\". The first lens maps the complex amplitude distribution of an incident field across the front focal plane to the back focal plane via a Fourier transform [@Goodman2017]. Thus, there is a one-to-one mapping between the angle of an incoming plane wave and the position in the back focal plane, known as the \"Fourier plane\", which is located a distance \(f_{\text{ext}}\) after the lens. The second lens is a distance \(2f_{\text{ext}}\) after the first lens and performs another Fourier transform, undoing the action of the first lens. With a spatially-varying phase element (i.e., a \"phase mask\") in the Fourier plane, the 4-\(f\) system phase shifts each incoming plane wave but otherwise leaves it unchanged. The angle-dependent phase shift required to emulate free-space propagation happens to correspond to the phase mask created by a thin positive lens. Consequently, such a lens placed in the Fourier plane (i.e., a middle lens with focal length \(f_{\text{mid}}\)) will cause the field transmitted by the 4-\(f\) system to appear as though it had propagated through a slab of free-space, potentially longer than the 4-\(f\) system itself. Note that while this three-lens device is neither monolithic nor plate-like, we still refer to it as a spaceplate due to its similar effect on light. Note as well that the use of a 4-\(f\) system results in inversion of the light about the \(\mathbf{z}\)-axis (\(x=0\)); while this does not affect the compression or absolute magnification of the device, it does differentiate the device from other spaceplates. In the next section, we theoretically demonstrate how this three-lens design can be used to attain free-space compression on a much larger scale than previous spaceplate designs. In Section [9](#sec:exp){reference-type="ref" reference="sec:exp"}, we experimentally demonstrate the three-lens spaceplate and characterize its performance for a number of lens choices. We show how it can be straightforwardly inserted into the design of long imaging systems, even if they are broadband. In Section [10](#sec:fundamentalLimits){reference-type="ref" reference="sec:fundamentalLimits"}, we outline some fundamental theoretical limits on the performance of the three-lens spaceplate. We focus on compression ratio, NA, and bandwidth, which recent work has identified trade-offs between [@Chen2021; @Shastri2022] for other spaceplate designs. In the last two sections, we compare and contrast the three-lens design to related optical systems and discuss prospects for its application. # Fourier optics description of a three-lens spaceplate {#sec:FourierOptics} To motivate the design of the three-lens spaceplate, we follow the Fourier optics description of the spaceplate effect introduced by Reshef et al. [@Reshef2021]. Suppose we have a light field propagating through free-space centered on the \(\mathbf{z}\)-axis. We then consider how each transverse spatial Fourier component of this initial field is transformed by this propagation. Each of these components, or plane waves, is described by its momentum vector, \(\mathbf{k} = \left(k_{x},k_{z} \right) = k(\sin \theta, \cos \theta)\), where \(\theta\) is the angle between the \(\mathbf{z}\)-axis and the direction of the plane wave. When propagated over a distance \(z_0\) in free-space, the amplitude and direction of each plane wave is conserved, but the phase is not. Between two points on the \(\mathbf{z}\)-axis separated by \(z_0\), each plane wave accumulates phase \(\varphi = k_{z} z_0 = |\mathbf{k}|\cos(\theta)z_0\). A spaceplate must produce this same phase response, but in a shorter distance \(d\) than \(z_0\). An ideal spaceplate, if inserted into the light path, will produce the phase response \(\varphi_{\text{SP}} = k_{z}z_0 = z_0(k^2-k_x^2)^{1/2}\). In the small angle approximation, \(\varphi_{\text{SP}} = z_0 k(1-k_x^2/2k^2)\). Then, by neglecting global phase terms, we arrive at our target phase, \(\varphi_{\text{SP}} =-z_0 k_{x}^2/2k\). An approach for considering larger incident angles is given in Appendix [\[sec:appendixLargeAngle\]](#sec:appendixLargeAngle){reference-type="ref" reference="sec:appendixLargeAngle"}. In the Fourier plane of the 4-\(f\) system in Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}a), the field at position \(r\) (i.e., axial distance) is the incoming field component with transverse momentum \(k_{x} = rk/f_{\text{ext}}\). Substitution of this relation into the target phase gives the spatial phase mask required to emulate propagation through space, \[\begin{aligned} \varphi_{\text{SP}} =-\frac{z_0{kr^2}}{2f_{\text{ext}}^2}. \label{eq:spaceplatePhase} \end{aligned}\] The phase added by a spherical, thin lens of focal length \(f_{\text{mid}}\) in the paraxial approximation is given by \(\varphi_{\text{mid}} =-kr^2/2f_{\text{mid}}\), which is of the same form as Eq. ([\[eq:spaceplatePhase\]](#eq:spaceplatePhase){reference-type="ref" reference="eq:spaceplatePhase"}). By setting \(\varphi_{\text{SP}}=\varphi_{\text{mid}}\) we find that the spaceplate effectively propagates light a distance \(z_0 = f_{\text{ext}}^2/f_{\text{mid}}\). In this way, we can compress space by placing a positive lens at the Fourier plane. A slight complication is that, as part of the 4-\(f\) system, there already is free-space propagation of distance \(f_{\text{ext}}\) before the first lens and \(f_{\text{ext}}\) after the last lens. We do not include those regions in the length of space replaced and thus, \(d_{\text{eff}}=z_0-2f_{\text{ext}}\). Similarly, we define the spaceplate thickness to solely be the distance *between* the external lenses, \(d=2f_{\text{ext}}\). Therefore, the compression ratio is \[\begin{aligned} \mathcal{R} =\frac{d_{\text{eff}}}{d}= \frac{f_{\text{ext}}}{2f_{\text{mid}}}-1, \label{eq:compressiveRatio1} \end{aligned}\] which only depends on the focal lengths used. If \(\mathcal{R} > 1\) or, equivalently, if \(f_{\text{mid}}<f_{\text{ext}}/4\), more space is replaced than is occupied by the three-lens spaceplate. Further discussion of system length and compression ratio is given in Appendix [\[sec:appendixCompressionRatio\]](#sec:appendixCompressionRatio){reference-type="ref" reference="sec:appendixCompressionRatio"}. # Experimental demonstration of three-lens space compression {#sec:exp} ## Design of the focus advancement, walkoff, and imaging experiments {#sec:expDesign} The setup used to test the three-lens spaceplate is shown in Fig. ([\[fig:expSetUp\]](#fig:expSetUp){reference-type="ref" reference="fig:expSetUp"}). We measure the advance of the focus of a beam, the walk-off of a beam incident at an angle, as well as imaging properties. These three types of measurements require minor setup variations that will be detailed in the following three subsections. We start by describing elements in the setup common to all three measurements. The external lenses, \(f_{\text{ext}}=\SI{150}{mm}\), are the same throughout all the tests and have \(\SI{50.8}{mm}\) diameters and \(\SI{15.0}{mm}\) thicknesses. We test the effect of the following series of middle lenses: \(f_{\text{mid}}= \{40, 30, 25, 19, 18, 11, 10, 4.51\}\,\SI{}{mm}\) whose respective diameters and thicknesses were \(\{25.4,25.4,12.7,12.7,6.5,8.0, 5.5\}\,\SI{}{mm}\) and \(\{12.5, 14.0, 7.0, 6.0, 2.2,5.0, 6.5, 2.94\}\,\SI{}{mm}\). With the exception of the \(f_{\text{mid}}=\{18,11,4.51\}\SI{}{mm}\) aspheric lenses, all lenses used throughout these tests were achromatic doublets, designed and anti-reflective (AR) coated for the visible range (400-\(\SI{700}{nm}\) light). The \(f_{\text{mid}}= \SI{18}{mm}\) aspheric lens was designed for \(\SI{780}{nm}\) light, and the \(f_{\text{mid}}= \{11,4.51\}\,\SI{}{mm}\) aspheric lenses were designed for \(\SI{633}{nm}\) light. All aspheric lenses were AR-coated for 600-\(\SI{1100}{nm}\). After the spaceplate, an image sensor (\(\SI{5.86}{\micro m} \times \SI{5.86}{\micro m}\) pixel size, \(1936\times1216\) pixels, color) was mounted on a motorized translation stage capable of moving both in the transverse (\(x\)) and longitudinal (\(z\)) directions. With this sensor, we recorded \(x\times y\) spatial intensity distributions at chosen \(z\) positions. The electronic settings of the image sensor were constant throughout the various tests. ## Focus advancement of various three-lens spaceplates {#sec:focusAdvancement} We measure the focus advancement of a \(\SI{633}{nm}\) wavelength Helium-Neon laser that is collimated so that its beam has an intensity full-width-half-maximum (FWHM) of \(\SI{3.5}{mm}\) using three different middle lens focal lengths corresponding to three distinct compression ratios. To create a focusing beam, the laser travels through a \(\SI{50.8}{mm}\) diameter positive lens with a thickness of \(\SI{11.3}{mm}\) and a focal length of \(f_{\text{image}}=\SI{1000}{mm}\) \[see Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}a)\]. The three-lens spaceplate is placed between this positive lens and the nominal focus \(z\)-position, then the image sensor records intensity distributions over a range of \(\SI{500}{mm}\) with a \(\SI{0.25}{mm}\) step size. Each \(x\) pixel row of the recorded distribution is summed to determine the row with the maximal intensity. This row becomes the 1D intensity profile at each \(z\), which is plotted in Fig. ([\[fig:2022914_focusComparison\]](#fig:2022914_focusComparison){reference-type="ref" reference="fig:2022914_focusComparison"}). For each middle lens, the peak intensity along \(z\) is taken as the new focal position. We measure longitudinal focal shifts of \(\Delta z = \SI{+94}{mm}\), \(\SI{-49}{mm}\), and \(\SI{-221}{mm}\) corresponding to observed compression ratios \(\mathcal{R_{\text{obs}}}= 0.69\), \(1.16\), and \(1.74\), respectively. The respective nominal compression ratios are \(\mathcal{R} = 0.9, 1.5,\) and \(2.0\). While these are significantly different from the measured values, they demonstrate that the three-lens spaceplate can successfully replace space. Deviations between \(\mathcal{R}\) and \(\mathcal{R_{\text{obs}}}\) result because the three-lens spaceplate relies on a lens located directly at the Fourier plane of the 4-\(f\) system, and any deviation of the thick middle lens from this position results in a reduction of compression and non-unity magnification. Specifically, we observe that upon transmitting an expanded collimated beam with a \(\text{FWHM}=\SI{3.5}{mm}\) through the three-lens spaceplate, the resultant beam width varies (\(1.7-\SI{4.4}{mm}\)) depending on small adjustments (\(0-\SI{5}{mm}\)) to the longitudinal position of the middle and rear lenses in the spaceplates. The effect of these sensitivities are reduced in subsequent walkoff and imaging experiments. Nonetheless, the observed negative longitudinal shift demonstrates that this design is capable of replacing free-space on a scale of three orders of magnitude greater than previous spaceplate designs. This experiment is a qualitative demonstration of space compression-the following experiments are used to characterize the three-lens spaceplate with greater accuracy. ## Transverse beam walkoff of a three-lens spaceplate We next measure the transverse walk-off \(\Delta w = |x_{\text{SP}}|-|x| = d(\mathcal{R}-1)\sin{\theta}\) \[see Fig. ([\[fig:20220922-threeLensDiagram\]](#fig:20220922-threeLensDiagram){reference-type="ref" reference="fig:20220922-threeLensDiagram"}b) for a diagram\] of a beam transmitted through the spaceplate, incident at angle \(\theta\) to the \(z\)-axis, rotating about the \(y\)-axis. For each spaceplate, the walk-off as function of angle characterizes the NA and aberrations, and determines the compression ratio with greater accuracy than in the focal advance measurement. Specifically, before conducting walkoff experiments, the longitudinal positions of the middle and rear lenses are incrementally adjusted until a beam (\(\text{FWHM}=\SI{3.5}{mm}\)) transmitted through the spaceplate becomes collimated and has unity magnification. That is, until the exiting beam has diameter of \(\text{FWHM}=(3.5\pm0.1)\,\SI{}{mm}\) both immediately after the rear lens of the spaceplate, and one meter after it. To test the beam walkoff, the imaging lens and beam-expander are absent, leaving the laser beam collimated. The rotation stage in Fig. ([\[fig:expSetUp\]](#fig:expSetUp){reference-type="ref" reference="fig:expSetUp"}) varied \(\theta\) from 0 to 2 degrees in 0.02 degree increments. The stage is located \(f_{\text{ext}}\) before the first lens, and we measure the transverse displacement \(z = 4f_{\text{ext}}=\SI{600}{mm}\) from the rotation axis of the stage. The rotation stage is placed here due to our definition of \(\text{NA}=\sin\theta\), where \(\theta\) is the half-angle of a maximum cone accepted by the system, originating \(f_{\text{ext}}\) before the first lens. By putting the stage at this position, we are able to contextualize the experimental angular limitations of the system with an analysis of its numerical aperture. The position at which we measure the transverse displacement, however, does not matter in general. One will measure the same \(\Delta w\) regardless of the position of the imaging plane. First, we measure the transverse displacement \(x(\theta)\) of a beam from the \(z\)-axis in free space, and then perform analogous measurements \(x_{\text{SP}}(\theta)\) of beams transmitted through each spaceplate. Due to inversion, the beams are displaced in opposite directions, but we can still compute the walkoff as: \(\Delta w = |x_{\text{SP}}|-|x|\). We measure \(x\) and \(x_{\text{SP}}\) by fitting Gaussians to the measured transverse intensity profiles. Fig. ([\[fig:2022914_compressiveFactorAndWalkoff\]](#fig:2022914_compressiveFactorAndWalkoff){reference-type="ref" reference="fig:2022914_compressiveFactorAndWalkoff"}a) presents this series of walkoff measurements for each of the eight middle lenses given in Section [9.1](#sec:expDesign){reference-type="ref" reference="sec:expDesign"}, with corresponding observed compression ratios in Fig. ([\[fig:2022914_compressiveFactorAndWalkoff\]](#fig:2022914_compressiveFactorAndWalkoff){reference-type="ref" reference="fig:2022914_compressiveFactorAndWalkoff"}b). Each of the configurations produced compression ratios within error of their nominal compression ratios over the entire measured range. For the highest experimental compression ratio, \(\mathcal{R}=15.6\), \(f_{\text{mid}}=\SI{4.51}{mm}\), the spaceplate replaced 4.39 meters of free-space, which is more than one-thousand times greater than that demonstrated by other spaceplate designs. ## Broadband imaging with a three-lens spaceplate In our final measurement, we investigate the performance of the three-lens spaceplate for imaging with full-visible-spectrum light. We first illuminate a postage stamp by white-light and use the \(f_{\text{image}}\) lens to form a sharp image on the image sensor. The object and image distances are approximately \(\SI{1.9}{m}\) and \(\SI{2.1}{m}\), respectively, and we measure a magnification of 1.1. We record images with and without a \(f_{\text{mid}}=\SI{10}{mm}\) spaceplate \[Fig. ([\[fig:20220915_fullColorImageComp\]](#fig:20220915_fullColorImageComp){reference-type="ref" reference="fig:20220915_fullColorImageComp"}c), left and right panes, respectively\] in the imaging path. Before imaging, the middle and rear lenses of the spaceplate are calibrated using an expanded collimated beam, as was done in the walkoff experiment. The camera is then placed at the nominal longitudinal position relative to the focus position seen without a spaceplate in the imaging path. Then, the middle and rear lens longitudinal positions are adjusted slightly to achieve the same magnification seen without the spaceplate. To reduce aberration and edge-effects, a large iris was positioned immediately before the third lens in the spaceplate and was adjusted until the highest contrast and resolution were seen. This was at an aperture size of \((12\pm2)\,\SI{}{mm}\). While there still is a reduction in contrast and brightness, which we discuss later, the images show that the three-lens spaceplate can advance the image plane of a full-visible-spectrum imaging system by 1.65 m. Next, we image an NBS 1963a resolution target in order to characterize the reduction in resolution and contrast caused by the addition of four different three-lens spaceplates \[see Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"})\]. The addition of the spaceplates does not change the magnification, as expected. However, Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}b) shows that every spaceplate reduces image contrast and resolution in comparison to the image formed in the absence of a spaceplate. We quantify this reduction with the modulation transfer function (MTF) of the imaging system following the method of Saiga et al. [@Saiga2018] in Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}d). For the case of \(f_{\text{mid}} = \SI{10}{mm}\) (\(\mathcal{R} = 6.5\)) and the stamp imaged without an iris, the spaceplate lowers the spatial frequency at half-max of the MTF by 45% (i.e. \(10.9\,\text{lp/mm}\) is reduced to \(6.0\,\text{lp/mm}\)). Three potential causes for this reduction are aberrations, scattering from lens edges, and diffraction by the lens apertures. The first two would be mitigated by a reduced aperture, whereas the effect of diffraction would be worsened. Adjusting the diameter of an iris \([(12\pm2)\,\SI{}{mm}]\) just before the third lens, we take the images shown in Fig. ([\[fig:20221121_monochromaticimagingupdated\]](#fig:20221121_monochromaticimagingupdated){reference-type="ref" reference="fig:20221121_monochromaticimagingupdated"}c), which are qualitatively improved in contrast and resolution. This improvement is reflected in the MTF measured with an iris in place, which increases by 38% (i.e. \(6.0\,\text{lp/mm}\) to \(8.3\,\text{lp/mm}\)) suggesting that diffraction is not the main cause of the reduction. Equivalently stated, addition of the spaceplate and iris causes a 24% reduction in the MTF (i.e. \(10.9\,\text{lp/mm}\) to \(8.3\,\text{lp/mm}\)). # Fundamental limits on performance of the three-lens spaceplate {#sec:fundamentalLimits} In this section, we briefly discuss limits on the NA and compression inherent to the three-lens design of the spaceplate. In most physically realizable setups, the NA of the three-lens spaceplate will be limited by the diameter of the third lens, \(D_3\). For this common case, the NA can be expressed in terms of the f-number of the third lens, \(N_3 = f_{\text{ext}}/D_3\), and \(\mathcal{R}\): \[\begin{aligned} \text{NA} = \left[1+4N_3^2\left(1+2\mathcal{R}\right)^{2}\right]^{-\frac{1}{2}}. \label{eq:numericalaperturethirdlens} \end{aligned}\] Eq. ([\[eq:numericalaperturethirdlens\]](#eq:numericalaperturethirdlens){reference-type="ref" reference="eq:numericalaperturethirdlens"}) shows that there is trade-off between compression ratio and NA. The lower the f-number of the third lens, the less strict this trade-off is. The third lens used in our experiment has \(N_3=3\), which limits the NA of the three-lens spaceplate be less than 0.06, worsening as \(\mathcal{R}\) increases. This represents the greatest shortcoming of the three-lens spaceplate design: even when using a relatively large third lens diameter (\(D_3=\SI{50}{mm}\)), the maximum incident angle is three degrees from the optical axis. For information on the less common cases where the first or second lenses limit the NA, see Appendix [\[sec:appendixNA\]](#sec:appendixNA){reference-type="ref" reference="sec:appendixNA"}. We have considered two additional physical mechanisms inherent to the three-lens spaceplate that lead to limitations. However, these turn out to be much less restrictive than the above limitation in practical applications with standard lenses. Firstly, the Abbe Sine Condition limits the NA independent of the lens diameters. Secondly, diffraction from the first lens in the spaceplate limits the achievable compression ratio. These two limitations may be relevant for new types of lenses, such as metalenses, and are therefore explored in depth in Appendices ([\[sec:appendixAbbe\]](#sec:appendixAbbe){reference-type="ref" reference="sec:appendixAbbe"}) and ([\[sec:appendixDiffraction\]](#sec:appendixDiffraction){reference-type="ref" reference="sec:appendixDiffraction"}), respectively. # Comparison to other compressive optics A number of distinctions must be made between the three-lens spaceplate and other methods to conserve space in imaging systems. We start by comparing it to other spaceplate designs. All of these exhibit some degree of transverse translation invariance [@Shastri2022], whereas the three-lens spaceplate does not, since lenses are not translationally invariant. Another difference is that the three-lens spaceplate inverts the transmitted field through the axis of the lenses (i.e., \(x\rightarrow-x\) and \(y\rightarrow-y\)), though this is avoided simply by using a pair of spaceplates. Many conventional imaging systems are already designed to minimize the space they occupy. Examples include the telephoto lens and the Cassegrain reflector whose goal is to create an imaging system shorter than its effective focal length \(f_{\text{eff}}\). Thus the compression of space is intertwined with the main purpose of the imaging system-to provide optical magnification. In practice, telephoto lenses are limited to a length \(> 0.8 f_{\text{eff}}\) [@Kingslake1978; @Tremblay2005; @Tremblay2007]. This is equivalent to a non-telephoto lens followed by a three-lens spaceplate with \(\mathcal{R}\geq 1/0.8=1.25\), much lower than the highest value reported here for three-lens spaceplates \(\mathcal{R}=15.7\), albeit with a higher NA. The nature of the components of these three-lens systems means that they are large. In this way, they differentiate themselves from thin spaceplates. Because they use lenses, implementable three-lens spaceplates have limitations on thickness, aperture, and, consequentially, compression ratio. Using off-the-shelf components and manageable system lengths (\(2f_{\text{ext}}\lesssim\SI{1}{m}\)), compression ratios are limited (\(\mathcal{R}\lesssim 20\)); however, due to the large size of these systems, they can save large absolute quantities of space (on the order of meters). This is in contrast to existing spaceplate designs, which only save space on the order of microns or millimeters [@Reshef2021; @Guo2020; @Page2022; @Chen2021]. One advantage of using standard optical components in three-lens spaceplates is that the resulting spaceplate is broadband, polarization-independent, and highly configurable in compression ratio. The compression ratio of the three-lens spaceplate can easily be changed by replacing the middle lens, whereas the compression ratio of other spaceplate designs is fixed upon manufacturing. Similar systems have been used in microscopy to rapidly tune imaging planes [@Zuo2013; @Chen2018; @Kang2020; @Zuo2015] in transport-of-intensity phase microscopy. However, such implementations do not compress space and are instead intended to artificially increase the depth of field. # Conclusion In this paper, we have successfully demonstrated the compression of free-space using a device that we call a \"three-lens spaceplate\", a positive lens placed at the Fourier plane of a 4-\(f\) optical setup. We have shown that our spaceplate design can be used for meter-scale space compression and the miniaturization of broadband imaging systems. Using the three-lens spaceplate, we have achieved experimental compression ratios up to \(\mathcal{R}_{\text{obs}}=15.7\) that replace 4.4 meters of free-space. Until now, replication of the transfer function of free-space over lengths of this scale had not been demonstrated: previously, space compression of only one or two millimeters had been shown [@Reshef2021]. The three-lens spaceplate has not only shown that this is possible, it has accomplished this using off-the-shelf optical components. Moreover, the system was shown to reduce the length of both monochromatic and full-color imaging systems with a 24% loss of resolution as quantified using its modulation transfer function. Finally, we discussed the trade-off between compression ratio and numerical aperture. Specifically, we found that the numerical aperture is highly limited for the three-lens spaceplate and represents its greatest shortcoming. The simplicity and versatility of this spaceplate design motivate its application and future development, though there are certain limitations which differentiate its applications from other spaceplates. The three-lens spaceplate represents a simple, cost-efficient method for reducing the size of axially-symmetric, long focal length systems such as telescopes. Because these three-lens spaceplates are so large and have such limited numerical apertures, however, their application in smaller devices such as smartphone cameras or virtual reality headsets would be difficult to implement. That said, the inherent use of free-space within the system itself implies the potential for even higher compression ratios through the use of metalenses and other space-saving devices. Metalenses could potentially also reduce the numerical aperture limitations of the three-lens spaceplate by reducing lens thickness, and increasing aperture size. Future work includes using such hybrid designs to reduce overall system length and increase numerical aperture, as this could make the implementation of three-lens spaceplates in small devices more realistic.

{'timestamp': '2023-02-10T02:00:33', 'yymm': '2302', 'arxiv_id': '2302.04295', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04295'}

# INTRODUCTION Investigation of gait patterns is a valuable diagnostic tool that provides clinicians the ability to diagnose a range of impairments affecting a patient's mobility, such as trauma, stroke or neurological conditions. Features extracted from these gait patterns, as key-event times, duration, and joint-angle deviations are all quantifiable measures to allow a clinician to benchmark and make informed decisions over the selection of appropriate therapies. In essence, a gait cycle is comprised of two core phases for each limb: stance and swing. This phase differentiation is associated with foot contact with the floor (stance) and swinging motion of the limb through the air, preceding the following step (swing). Taking the example of a right leg dominant subject, Figure [\[GC_stage\]](#GC_stage){reference-type="ref" reference="GC_stage"} describes the gait cycle for a naturalistic walking pattern. For more detailed explanations of the gait cycle, please refer to. Typical gait analysis relies on a trained clinician's visual and qualitative assessment. To aid in the visual inspection, gait analysis often makes use of high-cost, complex optical recording setups with associated force plates, installed in a clinic or laboratory space. These systems record, through a combination of high-speed cameras, the individual's movement via tracking of a predetermined set of retroreflective markers placed on the individual's body. Although accurate and reliable, optical motion capture systems restrict gait studies to confined locations within the camera setup, and incur in high training requirements for examiners to setup, calibrate, collect and analyse gait data. Due to this lack of versatility, optical systems are unable to investigate a patient's daily life mobility. To tackle this issue, researchers pushed towards the development of wearable technologies, through sensors that could still reliably identify gait events, without the typical constraints of optical systems. To best identify which are the most appropriate sensors for gait event detection on Activities of Daily Living (ADLs), and to pair with EMG analysis, a short review and selection of available sensors technologies was conducted. One of the earliest reported wearable technologies for gait event detection uses foot-switches. These sensors are usually embedded into the shoe or insole, located underneath the heel and hallux of the individual. This allows for the recording of two key gait events: Initial-Contact (IC), in healthy scenarios reported as the heel strike (HS); and End-of-Contact (EC), or Terminal Contact (TC), in healthy scenarios reported as the toe-off (TO). With current iterations, as insole pressure sensors (IPS) and force sensitive resistors (FSR), fitted into footwear (e.g. ), this technology is widely employed today and used as a benchmark for other motion capture systems (see for an up-to-date review). A common issue with these IPS and FSR systems is the constant wear-and-tear of the sensors due to associated shear-forces in the sole, or against the walking surface. This leads to short sensor lifespans and to sensing accuracy loss after sustained periods of usage. Nowadays, the most commonly employed wearable technology for gait event detection is accelerometry (). First described in, this method makes use of acceleration sensors, strategically placed near subjects joints on predetermined sites. The choice of placement will greatly influence the gait parameters extracted and takes into consideration possible noise motion artefacts that contaminate the recordings (e.g. poor sensor adhesion or loose clothing). With advances in wearable technology, typical accelerometer sensors were adapted into IMUs (Inertial Measurement Units), sensors that are able to detect both acceleration and angular velocity (through an integrated gyroscope) in all three dimensions. Depending on the desired gait parameters, the IMU sensors are most often placed at the foot () or shank/calf (), but other studies present interesting alternatives as knee/thigh (, waist () or even head () placements. During the aforementioned gait events (i.e. HS and TO), there are distinguishable peaks in the vertical acceleration profiles that can be used to identify the timing of HS and TO. Upon contacting the surface, the heel will experience a vertical force (ground-reaction force) and its respective upward acceleration, which is identifiable when inspecting the recorded signal. In addition to gait event detection, IMUs also presents benefits in the detection of angular displacements and velocities of different joints () as well as pedestrian tracking solutions (). With state-of-the-art sensor synchronisation, current IMU solutions also provide accurate position estimation for the subjects wearing the sensors, as described in. Due to all these features, comparison studies () often place the IMU as a preferable wearable solution over the previously discussed alternatives (i.e. FSR and Optical systems). Even if the IMU is not as accurate as the FSR or as thorough in limb position detection through optical systems, the IMU sensor durability, robustness, cost and applicability outweigh the benefits of these other technologies in ADL gait detection applications. So far, the wearable technologies presented here rely on external interactions of the individual with the sensors (e.g. pressure applied on the IPS) or measurements of external inferred parameters (e.g. the body part acceleration or angular velocity as measured by the IMU) to identify key gait events. An alternative is to make use of the patient's biological signals, as muscle activity recordings through surface Electromyography (sEMG), to estimate the timing of the events. sEMG provides a reliable non-invasive solution to analyse muscle activity levels during varied motor tasks, and in a non-constrictive manner. sEMG can be used to diagnose motor activity, where the recorded muscle activity is analysed to identify muscle deficiencies, and/or establish individual muscle function levels during pre-defined tasks. This has been specifically used for the analysis of gait, using analysis of the sEMG recorded from hip, knee and ankle flexor and extensor muscles, reported in with positive evidence for differentiating activity in different segments of the gait cycle. However, it being a biological signal, the sEMG can be quite noisy and variable, depending on sensor location and other biological factors. To interpret muscle activity features and classify these into gait events, studies often return to the use of machine learning algorithms (). Unfortunately, as mentioned, there are several factors that may interfere with the sEMG quality and the accuracy of these classification algorithms, especially when recognising how volatile and different the recorded signals can be across different individuals, and even in different muscles of the same individual. All these wearable solutions present promising results for gait event detection, through varied technologies and algorithms. The complexity of such tools decides how viable it is for applications of gait assessment during Activities of Daily Living (ADLs). Unfortunately, these solutions seldom come with a clear methodology of how to apply these tools to detect the most simple of gait events, rarely offering the algorithms to do so. In this method paper, we propose a simplified method of heel strike detection for synchronised sEMG recording segmentation through IMU data paired with sEMG recordings, applied to the diverse and unconstrained dataset of, and further test it with samples of data from, looking at Parkinson's Disease (PD) gait. # METHODOLOGY The following section details the steps, and associated scripts, to isolate sEMG activity patterns in different unconstrained gait modalities, based on the kinematic dataset of. This guide contains step-by-step information about using recordings from an IMU system (), focused on a single sensor on the subject's dominant leg ankle, to identify heel strike timestamps during different gait patterns. For ease of processing, each important step of the sEMG segmentation is tackled and discussed in separate scripts, coded in *MATLAB*(). In summary, the dataset provided in contains unconstrained walking data from 10 healthy participants, with kinematic (through IMU system ), sEMG and EEG (Electroencephalography) recordings. The participants walk through a designed walking course, encompassing a ramp, level ground and staircase sections. In each trial the participant walks up and down the ramp and staircase, as well as level ground walking, providing recordings for 5 different walking modalities. For ease of referencing, please consider the following acronyms: - **RA**: ramp ascent; - **RD**: ramp descent; - **SA**: staircase ascent; - **SD**: staircase descent; - **LGW**: level ground walking; To accompany and summarise the following methodology, Figure [\[SArch\]](#SArch){reference-type="ref" reference="SArch"} details the required steps and processes for the kinematic and sEMG data. ## Interpolation and kinematic activity segmentation Kinematic data (KIN) and sEMG data are recorded at different sampling rates, but synchronised in time. To use kinematics as an indicator for segmentation timestamps in the sEMG counterpart, the kinematic signal requires linear interpolation to match the number of sEMG datapoints. \[\begin{aligned} {\frac {y-y_{0}}{x-x_{0}}}={\frac {y_{1}-y_{0}}{x_{1}-x_{0}}} \\ y=y_{0}+(x-x_{0}){\frac {y_{1}-y_{0}}{x_{1}-x_{0}}} \\ y={\frac {y_{0}(x_{1}-x)+y_{1}(x-x_{0})}{x_{1}-x_{0}}} \label{eq:lin_int1} \end{aligned}\] \[y_s(i) = \frac{1}{2N + 1} (y(i+N)+y(i+N-1)+ \\ \label{eq:lin_int2}\] \[...+y(i-N))\\ \label{eq:lin_int3}\] Equations [\[eq:lin_int1\]](#eq:lin_int1){reference-type="ref" reference="eq:lin_int1"} and [\[eq:lin_int3\]](#eq:lin_int3){reference-type="ref" reference="eq:lin_int3"} present a common linear interpolation. With a similar time duration between recordings but different sampling frequencies, the kinematic signal (y) with less samples is interpolated to match the number of samples of sEMG signal (x), estimating missing kinematic values. This process ensures that both KIN and sEMG recordings are synchronised in time and samples. A moving average smoothing filter is then applied to KIN to remove unwanted oscillations and initial motion noise from the signal. The pre-processed signals now consist of two complete walk trials that includes all five walking modalities previously discussed (RA, RD, SA, SD and LWG) for each subject. A minimal energy threshold and window are created to isolate and timestamp meaningful movement identified using the IMU acceleration. Any signal fragment discovered inside a passing window that is less than a certain energy level is categorised as '0'. Similarly, if significant acceleration is generated beyond the threshold regarded as evidence of the subject walking, it is recorded as 1. As a result, a binary signal is generated identifying datapoints of relevant walking activity. ## Filtering small movement artefacts and separating trials from a single recording The core objectives for this section are: filtering movement and/or transition artefacts, considered as noise; to identify the direction of movement, as the walking modalities presented are dependent on the direction of the walking course. The output of the previous section is a duty-cycle type wave, depicted as the grey square wave in Section A of Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"}), presenting the binary values of 1 when the subject is active, and 0 when the subject is standing still. The start and end of active periods define the timepoints at which walking starts and ends. As the participants complete the trials in a continuous manner without taking unnecessary breaks between walking modalities, any detected activity under a threshold of 6000 datapoints (or 6 seconds) is classified as a motion artefact or noise. In the experiment, each recording is comprised of 2 trials. In a single trial a participant completes the course in a forward direction, rests at a designated location (B), and then returns to the starting point (A), walking in the reverse direction of the course. Each of the walking directions (A-B and B-A) is defined as a \"half-trial\". A single recording has then 4 half-trials, where 2 are the participant walking from the starting point (A) to point (B), and 2 where the participant is walking the reverse direction from point (B) to point (A). Four moments of activity are then extracted from the complete recordings and stored as an independent variable to later use to define the direction in which the participant walked the course. ## Initial sEMG segmentation and direction, using the timepoints defined by the 4 moments of activity per complete trial The extracted timestamps for the start and end of each half-trial are then used to segment the sEMG and KIN signals. The sEMG recordings provided in this dataset are in raw format. The objective of this stage is to remove noise from movement artefacts and possible recording interference, and to isolate relevant frequency content from the data. For this purpose, a 4th order Butterworth bandpass filter is designed, to filter any content outside the 10 to 150 Hz band. Additionally, a notch filter is also used to remove powerline interference at 60 Hz. The following objective, as mentioned, is to cut both sEMG and KIN signals, using the previously defined half-trial timestamps. For this purpose, and highly dependent on the dataset in question, a timestamp with a safety margin of 2000 datapoints (2 seconds) is used to ensure the inclusion of all relevant activity. sEMG and KIN are then segmented into bins of activity based on half-trials, for all trials and all participants. As mentioned, per complete trial, 4 moments of walking activity are detected (example in Section A of Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"}), related to walking the course twice in the forward and twice in the reverse directions, as described previously between (A) to (B). Based on the extracted timestamps and the experiment methodology, the sequence of modalities in the forward direction is separated into RD, LGW and SA, and for the reverse course direction, separated into SD, LGW and RA. ## Walking modality identification within each direction of the course and trials At this stage, sEMG and KIN signals for all trials are segmented and separated according to the direction of the course the participant is walking along. The next step is to identify the timestamps for each walking modality within the forward and reverse courses, and isolate these in both the sEMG and KIN signals. Due to the design of the course, participants take a sharp turn at the transition between each modality (example on Section B of Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"}). This is identified via the estimated position changes as recorded by the IMU. Focusing on the y-axis, any time there is a peak transition, a new modality begins, according to the course guide mentioned earlier. This, paired with safety margins, is then used to further segment both sEMG and KIN into the individual walking modalities. ## Using acceleration data to identify the moment of heel strike Having the data segmented into different walking modalities, the next step is to identify gait cycles within each walking modality. Through acceleration data in the vertical direction (z-axis), the moment of HS can be identified. To do so, new filters are designed to isolate the vertical peaks of acceleration witnessed when the heel strikes the ground, and experiences the ground-reaction force. Two separate 7th order Butterworth filters are created: a high-pass at 9 Hz and a low-pass at 6 Hz. These remove a band where specific motion artefacts lie, and isolate the striking moments. Being post-hoc, these filters are ideal and do not impose any delays on the data. The filtering process is as follows:-the high-pass filter is applied to the raw KIN data, removing frequencies below the 20 Hz band, and potential movement artefacts;-a half-wave rectifier is applied to the resulting filtered data, removing unnecessary negative electrical oscillations imposed by the sensor readings;-a low-pass filter is applied to the rectified data, serving as an anti-aliasing filter at 5 Hz, to remove oscillations from the signal. The HS moment is identified as the highlighted peaks in the Filtering and Peak section of Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"}. Extracting the timestamps of each of these peaks allows then for the isolation of a complete stride, or gait cycle, from one heel strike moment to the immediate next of the same limb. These timestamps are then finally used to segment the sEMG activity into separate gait cycles. # RESULTS Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"} demonstrates the algorithm effectively identifying the moments of HS, exemplified in that segment of level ground walking. For the other walking modalities, there is a noticeable shift in the amplitude of the ground-reaction force acceleration and an adjusted cadence, but each HS is equally as recognisable using our algorithm. With the final synchronised sEMG and KIN signals, individual gait cycles were then extracted and separated. In Figure [\[segEMG\]](#segEMG){reference-type="ref" reference="segEMG"}, an example of sEMG activity is presented for the same level ground walking segment of Figure [\[SDiag\]](#SDiag){reference-type="ref" reference="SDiag"}. Within the isolated gait cycles, we can identify the muscle activity patterns from the segmented sEMG data, as seen in Figure [\[MAPs\]](#MAPs){reference-type="ref" reference="MAPs"}. As a proof of concept, this algorithm was applied to another unconstrained walking dataset, to a PD patient, performing a series of walking tasks in a therapy scenario. In this dataset, patients were using an IMU placed on their dominant foot, akin to the experiment focused on in this report. Unlike the core dataset though, this PD data does not provide sEMG recordings. Nevertheless, with fine tuning of cadence and velocity parameters, through adjustments of the peak detection section, all the moments of HS are detected over the entire trial (over 1 minute). # DISCUSSION AND CONCLUSIONS Several methods have been designed for gait event detection, based on an assortment of different technologies, of different complexities (refer to for an up-to-date review). With advancements in sensor design it is now possible to bring these tools to our everyday lives, and gait assessment to activities of daily living. The importance of tracking patients' performance levels at their own homes and through their daily lives is of massive benefit for the design of smarter continued therapies. With the ambition of continued, supervised at-home therapies and improved rehabilitation, it is essential that researchers are provided with simple guides on using the technology and appropriate methodology to provide them the foundation for exploring therapeutic mobility. Even in the ADL scenario, as demonstrated, kinematics alone may not provide sufficient information to therapists to fully assess patients' mobility. Integration of muscle activity levels, associated with recorded movements and gait events will benefit not only diagnosis but performance assessment. With simplification of the acquisition and processing algorithms, as well as selection of appropriate wearable sensors, a platform that integrates both kinematics and sEMG for home use becomes feasible. With this report, a simplified guide for kinematic and sEMG recording segmentation, and gait event detection was presented and fully documented. Codes and algorithms, that work with the dataset of are made available alongside this method. To adapt to other datasets, users can easily adjust expected cadence or filter level variables, improving detection and allowing the study different walking modalities, providing an initial framework for researchers to explore and adjust to their needs and data. Synchronising both sEMG and kinematic information is the step forward to better characterise movement, and it is crucial that both signals start being analysed in pair. Analysing sufficient sEMG recordings in ADLs is the non-invasive key to defining the patterns of muscle activity that combine to generate movement, and should be centre stage in any mobility study. The same level of \"wearable\" exploration that kinematics and gait cycle has received over the past decades, needs to be pursued for the sEMG case. With adjustable parameters, the algorithm can be adapted to further ADLs scenarios, and potentially applied to patient's smartphones, for a continuous kinematic recording. Due to its simplicity and focus to aid newcomers to the gait studies area, this algorithm can also be combined with machine learning in future iterations, to allow ad-hoc heel strike identification.

{'timestamp': '2023-02-09T02:17:28', 'yymm': '2302', 'arxiv_id': '2302.04243', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04243'}

# Introduction The convex layers of a point set in \(\R^n\) are constructed by selecting the vertices of its convex hull, removing them from the set, and repeating. More formally, suppose that \(X\) is a finite point set in \(\R^n\). Setting \(C_1 = X\), we inductively define \(L_i\) as the vertices of the convex hull of \(C_i\) and set \(C_{i+1} = C_i\setminus L_i\). The greatest integer \(k\) for which \(C_k \neq \emptyset\) is called the *layer number* of \(X\) and is denoted \(L(X)\). This process is sometimes called *peeling*, and the set of layers is whimsically referred to as the *onion*. Convex layers were introduced by Barnett as one of several possible methods of ordering high-dimensional data sets. The peeling process was studied algorithmically by Eddy and Chazelle, and has since found applications in outlier detection, recognition of projectively deformed point sets, and fingerprint recognition. More recently, mathematical attention has turned toward calculating the layer number of various classes of point sets. In 2004, Dalal proved that the expected number of layers in an random \(n\)-point set distributed uniformly in a \(d\)-dimensional ball is \(\Theta_d(n^{2/d+1})\). (We use subscripts in order-of-magnitude notation to denote the variables on which the hidden constants depend.) Following this work, Choi, Joo, and Kim studied the layer number of point sets that avoid clustering, and they prove that \(L(X) = O_{\alpha,d}\big(|X|^{(d+1)/2d}\big)\) for any \(\alpha\)-evenly distributed point set in the unit ball of \(\R^d\). (The parameter \(\alpha\) is a measure of the evenness of spread.) Ambrus, Nielsen, and Wilson improved this result, showing that any such set has \(O_{\alpha,d}\big(|X|^{2/d})\) layers when \(d \geq 3\). Given two integers \(a \leq b\), we write \([a,b] = \{a,a+1,\dots,b-1,b\}\) for the integer range from \(a\) to \(b\), and we set \([n] = [1,n]\). The problem of determining the layer number of the integer grid \([n]^d\) has attracted attention as a notable special case of the layer number problem. In 2013, Har-Peled and Lidický proved that \(L([n]^2) = \Theta(n^{4/3})\) as \(n\to\infty\), solving the two-dimensional problem. Ambrus, Hsu, Peng, and Jan proved that \(L([n]^d) = \Omega_d\big(n^{2d/(d+1)}\big)\) by extending Har-Peled and Lidiký's proof, and they conjecture that this is tight. Since the grid \([n]^d\) is evenly distributed, Ambrus--Nielsen--Wilson's upper bound shows that \(L([n]^d) = O_d(n^2)\), where the hidden constant is exponential in \(d\). In this paper, we present a simple argument that significantly improves the dependence on \(d\): [\[thm:upper\]](#thm:upper){reference-type="ref" reference="thm:upper"} extends to a bound on \(L([n]^d)\) for every \(n\) using the fact that \(L(X)\leq L(Y)\) whenever \(X \subseteq Y\hspace{-0.15em}\). This can be proven by showing that the containment continues to hold at each step of the peeling process (see, for example). We also prove a new lower bound on the layer number of the grid: This lower bound is weaker than current lower bounds in the setting where the dimension is fixed and \(n\to\infty\). However, in the opposing regime where \(n\) is fixed and \(d\to\infty\), [\[thm:upper,thm:lower\]](#thm:upper,thm:lower){reference-type="ref" reference="thm:upper,thm:lower"} completely determine the order of magnitude of the layer number. # Proofs In the following proofs, we consider the translated point set \([-n,n]^d\) in place of \([2n+1]^d\). The advantage of this is that the group of symmetries \(\Gamma\) of \([-n,n]^d\) is generated by permutations and negations of coordinates. Moreover, the layers of \([-n,n]^d\) are invariant under the action of \(\Gamma\): If a point \(y\) is peeled at layer \(i\), so are all of its images under \(\Gamma\). This fact is especially important for the proof of [\[thm:lower\]](#thm:lower){reference-type="ref" reference="thm:lower"}. [\[fig:circles\]](#fig:circles){reference-type="ref" reference="fig:circles"} provides a visual interpretation of the argument in the proof of [\[thm:upper\]](#thm:upper){reference-type="ref" reference="thm:upper"}. \ We thank Shubhangi Saraf and Gergely Ambrus for helpful conversations during the preparation of this paper. We are also very grateful to Callie Garst for originally connecting the two authors.

{'timestamp': '2023-02-16T02:05:24', 'yymm': '2302', 'arxiv_id': '2302.04244', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04244'}

# Introduction Convolutional neural networks (CNNs) arose to one of the dominant types of models in deep learning (DL). The most prominent example is the computer vision, e.g. image classification, object detection, or image segmentation. CNN models proved to be effective in certain signal processing tasks, especially where the long-term context is not required such as speech enhancement, sound source separation, or sound event detection. Some authors showed that convolutional models can achieve similar performance that recurrent neural networks (RNN) at a fraction of model parameters. It is also argued that CNNs are easier to parallelize than recurrent neural networks (RNNs). However, unlike RNNs, which can process incoming data one sample at the time, CNNs require a chunk of data to work correctly. The minimum chunk size is equal to the size of the receptive field which depends on the kernel sizes, strides, dilation, and a number of convolutional layers. Additionally, overlaps may be required to reduce the undesired edge effects of padding. Hence, standard CNN models are characterized by a higher latency than RNNs. Algorithmic latency which is related to model requirements and limitations such as the minimal chunk size (e.g. size of a single FFT frame), a model look-ahead, etc. is inherent to the algorithm. It can be viewed as a delay between output and input signals under assumption that all computations are instantaneous. Computation time is the second component of latency. In case of CNNs it is not linearly dependent on the chunk size, due to the fact that the whole receptive field has to be processed regardless of the desired output size. In the case of audio-visual signals, humans are able to spot the lag between audio and visual stimulus above 10 ms. However, the maximum latency accepted in conversations can be up to 40 ms. For best human-device interactions in many audio applications the buffer size is set to match the maximum acceptable latency. ## Related works Many researchers presented solutions addressing the problem of latency minimization in signal processing models. studied a model consisting of bidirectional LSTM (BLSTM), fully connected (FC) and convolutional layers. Firstly, they proposed to use unidirectional LSTM instead of BLSTM and found that it reduces latency by a factor of 2 while having little effect on the performance. Secondly, they proposed to alter the receptive field of each convolutional layer to be causal rather than centered on the currently processed data point. In addition, it was proposed to shift the input features with respect to the output, effectively providing a certain amount of future context, which the authors referred to as *look-ahead*. The authors showed that predicting future spectrogram masks comes at a significant cost in accuracy, with a reduction of 6.6 dB signal to distortion ratio (SDR) with 100 ms shift between input/output, compared to a zero-look-ahead causal model. It was argued, that this effect occurs because the model is not able to respond immediately to changing noise and speech characteristics. further modified the above-mentioned model by removing LSTM and FC layers and replacing the convolutional layers with depth-wise convolutions followed by point-wise convolutions, among other changes in the model. These changes allowed to achieve a 10-fold reduction in the fused multiply-accumulate operations per second (FMS/s). However, the computational complexity reduction was accompanied by a 10% reduction in signal to noise ratio (SNR) performance. The authors also introduced partial caching of input STFT frames, which they referred to as *incremental inference*. When each new frame arrives, it is padded with the recently cached input frames to match the receptive field of the model. Subsequently, the model processes the composite input and yields a corresponding output frame. introduced a new family of CNN for online classification of videos (MoViNets). The authors developed a more comprehensive approach to data caching, namely layer-wise caching instead of input-only caching. MoViNets process videos in small consecutive subclips, requiring constant memory. It is achieved through so-called *stream buffers*, which cache feature maps at subclip boundaries. Using the stream buffers allows to reduce the peak memory consumption up to an order of magnitude in large 3D CNNs. The less significant impact of 2-fold memory reduction was noted in smaller networks. Since the method is aimed at online inference, stream buffers are best used in conjunction with causal networks. The authors enforced causality by moving right (future) padding to the left side (past). It was reported, that stream buffers lead to approximately 1% reduction in model accuracy and slight increase in computational complexity, which however might be implementation dependent. ## Novelty In this work we propose a novel approach to data caching in convolutional layers called Short-Term Memory Convolution (STMC) which allows processing of the arbitrary chunks without any computational overhead (i.e. after model initialization computational cost of a chunk processing linearly depends on its size), thus reducing computation time. The method is model-and task-agnostic as its only prerequisite is the use of stacked convolutional layers. We also systematically address the problem of algorithmic latency (look-ahead namely) by discussing causality of the transposed convolutional layers and proposing necessary adjustments to guarantee causality of the auto-encoder-like CNN models. The STMC layers are based on the following principles: - Input data is processed in chunks of arbitrary size in an online mode. - Each chunk is propagated through all convolutional layers, and the output of each layer is cached with shift registers (contrary to input-only caching as in ), providing so-called *past context*. - The past context is never recalculated by any convolutional layer. When processing a time series all calculations are performed exactly once, regardless of the processed chunk size. - The conversion of convolutional layers into STMC layers is performed after the model is trained, so the conversion does not affect the training time, nor does it change the model output. STMC can be understood as a generalization of the stream buffering technique, although it was developed independently with a goal to reduce latency and peak memory consumption in CNNs with standard, strided, and transposed convolutions as well as pooling layers (as opposed to standard convolutions only in MoViNets). STMC approaches the caching issue differently than MoViNets. It uses shift registers, which cache feature maps of all layers and shift them in the time axis according to the input size. The shift registers enable the processing of data chunks of arbitrary size, where the chunk size may vary in time. Furthermore, shift registers can buffer multiple states of the network, which allows for handling strides and pooling layers. Finally, we propose a transposed STMC, making it suitable for a broader family of networks, including U-nets. # Methods This section presents the Short-Term Memory Convolutions and their transposed counterpart. ## Short-Term Memory Convolutions[\[section:STMC\]]{#section:STMC label="section:STMC"} The STMC and transposed STMC (tSTMC) layers introduce an additional cache memory to store past outcomes. Fig. [\[fig:STMC_tSTMC_blocks\]](#fig:STMC_tSTMC_blocks){reference-type="ref" reference="fig:STMC_tSTMC_blocks"} shows the conceptual diagram of our method including the interaction between introduced memory and actual convolution. The cache memory provides values calculated in the past to subsequent calls of the STMC layers in order to calculate the final output without any redundant computations. The goal is to process all data only once.

{'timestamp': '2023-02-10T02:02:03', 'yymm': '2302', 'arxiv_id': '2302.04331', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04331'}

# Introduction {#md20} Iterated Function Systems (IFS) on the line consist of finitely many strictly contracting self-mappings of \(\mathbb{R}\). It was proved by Hutchinson that for every IFS \(\mathcal{F}=\left\{f_k\right\}_{k=1}^{m}\) there is a unique non-empty compact set \(\Lambda\) which is called the attractor of the IFS \(\mathcal{F}\) and defined by \[\label{cr65} \Lambda=\bigcup\limits_{k=1}^{m}f_k(\Lambda).\] For every IFS \(\mathcal{F}\) there exists a unique "smallest" non-empty compact interval \(I\) which is sent into itself by all the mappings of \(\mathcal{F}\): \[\label{cr22} I:=\bigcap\left\{ J \big\vert\: J\subset \mathbb{R}\mbox{ compact interval}: f_k(J)\subset J, \forall k\in[m] \right\},\] where \([m]:=\left\{ 1,\dots ,m \right\}\). It is easy to see that \[\label{cr15} \Lambda=\bigcap\limits_{n=1}^{\infty} \bigcup\limits_{(i_1,\dots ,i_n)\in[m]^n} I_{i_1\dots i_n},\] where \(I_{i_1\dots i_n}:= f_{i_1\dots i_n}(I)\) are the **cylinder intervals**, and we use the common shorthand notation \(f_{i_1\dots i_n}:=f_{i_1}\circ\cdots \circ f_{i_n}\) for an \((i_1,\dots ,i_n)\in [m]^n\). The IFSs we consider in this paper are consisting of piecewise linear functions. Thus their derivatives might change at some points, but they are linear over given intervals of \(\mathbb{R}\). We always assume that the functions are continuous, piecewise linear, strongly contracting with non-zero slopes, and that the slopes can only change at finitely many points. Let \(\mathcal{F}=\left\{f_k\right\}_{k=1}^{m}\) be a CPLIFS and \(I \subset \mathbb{R}\) be the compact interval defined in [\[cr22\]](#cr22){reference-type="eqref" reference="cr22"}. For any \(k\in [m]\) let \(l(k)\) be the number of breaking points \(\left\{b_{k,i}\right\}_{i=1}^{l(k)}\) of \(f_k\). They determine the \(l(k)+1\) open intervals of linearity \(\left\{J_{k,i}\right\}_{i=1}^{l(k)+1}\). We write \(S_{k,i}\) for the contracting similarity on \(\mathbb{R}\) that satisfies \(S_{k,i}|_{J_{k,i}}\equiv f_k|_{J_{k,i}}\). We define \(\left\{\rho_{k,i}\right\}_{k\in[m],i\in [l(k)+1]}\) and \(\left\{t_{k,i}\right\}_{k\in[m],i\in [l(k)+1]}\) such that \[\label{cr56} S_{k,i}(x)=\rho_{k,i}x+t_{k,i}.\] We say that \(\mathcal{S}_{\mathcal{F}}:=\left\{S_{k,i}\right\}_{k\in[m],i\in [l(k)+1]}\) is the **self-similar IFS generated by the CPLIFS \(\mathcal{F}\)**. We introduce the **natural pressure function** \[\label{cr64} \Phi(s):=\limsup_{n\rightarrow\infty}\frac{1}{n}\log \sum_{i_1\dots i_n} |I_{i_1\dots i_n}|^s.\] In, Barreira showed that \(\Phi(s): \mathbb{R}_+\to \mathbb{R}\) is a strictly decreasing function with \(\Phi(0)>0\) and \(\lim_{s\to\infty} \Phi(s)=-\infty\). Hence we can define the **natural dimension** of \(\mathcal{F}\) as \[\label{cr61} s_{\mathcal{F}}:=(\Phi)^{-1}(0).\] We note that he called \(\Phi(s)\) the non-additive upper capacity topological pressure. He also proved that the upper box dimension is always bounded from above by the natural dimension. Moreover, in the very special case when there are no breaking points, that is \(\mathcal{F} =\left\{ f_k(x)=\rho _kx+t_k \right\}_{k=1}^{m}\) is self-similar, \(s_{\mathcal{F}}\) is the so-called similarity dimension that is \(\sum_{k=1 }^{m }\rho _{k}^{ s_{\mathcal{F}}}=1\). Here \(\dim_{\rm B}\Lambda\) stands for the box dimension of the attractor. For the definition of fractal dimensions we refer the reader to. The inequality [\[cr13\]](#cr13){reference-type="eqref" reference="cr13"} also follows from, but to the best of our knowledge, in this explicit form it was first stated in. It follows that the Hausdorff dimension of the attractor \(\dim_{\rm H}\Lambda\) is also bounded from above by the natural dimension \(s_{\mathcal{F}}\). We will show that in a sense typically, these dimensions are actually equal. A continuous piecewise linear iterated function system \(\mathcal{F}=\{f_k\}_{k=1}^m\) is uniquely determined by the slopes \(\{\rho_{k,1},\dots,\rho_{k,l(k)+1}\}_{k=1}^m\), the breaking points \(\{b_{k,1},\dots,\) \(b_{k,l(k)}\}_{k=1}^m\) and the vertical translations \(\{f_k(0)\}_{k=1}^m\) of its functions. The latter two are called the translation parameters of \(\mathcal{F}\). To prove our main result we need a separation condition that was introduced by M. Hochman for self-similar iterated function systems. Let \(g_1(x)=\rho_1x+\tau_1\) and \(g_2(x)=\rho_2x+\tau_2\) be two similarities on \(\mathbb{R}\) with \(\rho_1,\rho_2\in\mathbb{R}\setminus \{0\}\) and \(\tau_1,\tau_2\in\mathbb{R}\). We define the distance of these two functions as \[\label{md17} \mathrm{dist}(g_1,g_2):=\begin{cases} \vert \tau_1-\tau_2\vert \mbox{, if } \rho_1=\rho_2; \\ \infty \mbox{, otherwise.} \end{cases}\] Hochman proved that the ESC is a \(\dim_{\rm P}\)-typical property of self-similar IFS. Prokaj and Simon extended this result by showing that it is a \(\dim_{\rm P}\)-typical property of a CPLIFS that the generated self-similar system satisfies the ESC. These two results together with the next theorem yields Theorem [\[md10\]](#md10){reference-type="ref" reference="md10"}. We are going to prove this theorem with the help of Markov diagrams. ## Self-similar IFSs on the line {#x99} The first breakthrough result of this field in the last decade was due to Hochman. He proved that ESC implies that the Hausdorff dimension of a self-similar measure (see ) is equal to the minimum of \(1\) and the ratio obtained by dividing the entropy by the Lyapunov exponent of this measure (for the definitions see ). As an immediate application of this result, we get that ESC implies [\[md45\]](#md45){reference-type="eqref" reference="md45"} for self-similar IFS on the line. Shmerkin proved an analogous result for the \(L^q\) dimension of self-similar measures in 2019. Using similar ideas, Jordan and Rapaport extended Hochman's result mentioned above from self-similar measures to the projections of ergodic measures. Using this result, Prokaj and Simon proved that ESC also implies that formula [\[md45\]](#md45){reference-type="eqref" reference="md45"} holds for graph-directed self-similar attractors on the line. This was an essential tool of the proof of Theorem [\[md14\]](#md14){reference-type="ref" reference="md14"} below. ## Earlier results about CPLIFS {#md23} Let \(\mathcal{F}=\{f_k(x)\}_{k=1}^m\) be a CPLIFS on \(\mathbb{R}\) with attractor \(\Lambda\). If \(\Lambda\) does not contain any breaking point, then we call \(\mathcal{F}\) **regular**. Prokaj and Simon showed in that for any regular CPLIFS there is a self-similar graph-directed IFS with the same attractor. This observation led to the following theorem For a \(k\in[m]\), let \(\rho_k:=\max_{j\in[l(k)+1]}\vert\rho_{k,j}\vert\) be the biggest slope of \(f_k\) in absolute value. We say that \(\mathcal{F}\) is **small** if the following two assertions hold: 1. \(\sum_{k=1}^m \rho_k<1\). 2. The second requirement depends on the injectivity of the functions of the system: 1. If \(f_k\) is injective, then \(\rho_k <\frac{1}{2}\); 2. If \(f_k\) is not injective, then \(\rho_k <\frac{1-\max_j\rho_j}{2}\). The dimension theory of some atypical CPLIFS families is discussed in. By combining Proposition [\[md13\]](#md13){reference-type="ref" reference="md13"}, Theorem [\[md14\]](#md14){reference-type="ref" reference="md14"} and the result of M. Hochman, Prokaj and Simon also showed that the equalities [\[md12\]](#md12){reference-type="eqref" reference="md12"} are \(\dim_{\rm P}\)-typical properties of small CPLIFSs. In this paper, we will extend this result by showing that the same holds without restrictions on the slopes. # Markov Diagrams {#md22} P. Raith and F. Hofbauer proved results on the dimension of expanding piecewise monotonic systems using the notion of Markov diagrams. We will define the Markov diagram in a similar fashion for CPLIFSs, and then use it to prove that \(s_{\mathcal{F}}\) equals the Hausdorff dimension of the attractor for non-regular systems as well, under some weak assumptions. ## Building Markov diagrams {#md21} Let \(\mathcal{F}=\{f_k\}_{k=1}^m\) be a CPLIFS, and let \(I\) be the interval defined by [\[cr22\]](#cr22){reference-type="eqref" reference="cr22"}. Writing \(I_k:=f_k(I)\) and \(\mathcal{I}=\cup_{k=1}^m I_k\), we define the **expanding multi-valued mapping associated to** \(\mathcal{F}\) as \[\label{md59} T:\mathcal{I}\mapsto \mathcal{P}(\mathcal{P}(I)),\quad T(y):= \big\{\{x\in I: f_k(x)=y\}\big\}_{k=1}^m.\] That is the image of any Borel subset \(A\subset\mathcal{I}\) is \[T(A)= \big\{\{x\in I: f_k(x)\in A\}\big\}_{k=1}^m.\] For \(k\in[m], j\in[l(k)+1]\), we define \(f_{k,j}:J_{k,j}\to I_k\) as the unique linear function that satisfies \(\forall x\in J_{k,j}: f_k(x)=f_{k,j}(x)\). We call the expansive linear functions \[\begin{aligned} \label{md58} \forall k\in[m], \forall j\in &[l(k)+1]:\quad f_{k,j}^{-1}: f_k(J_{k,j}) \to J_{k,j}, \\ &\forall x\in J_{k,j}: f_{k,j}^{-1}(f_k(x))=x \nonumber \end{aligned}\] the **branches** of the multi-valued mapping \(T\). As the notation suggests, these are the local inverses of the elements of \(\mathcal{F}\). Similarly, we define \(w(\mathcal{Z}_0)\) as the set of the successors of all elements of \(\mathcal{Z}_0\). That is \[\label{md96} w(\mathcal{Z}_0):=\cup_{Z\in \mathcal{Z}_0} w(Z).\] We often use the notation \[\label{md60} \mathcal{D}_n:=\cup_{i=0}^n w^i(\mathcal{Z}_0) \mbox{, where } w^i(\mathcal{Z}_0)=\underbrace{w\circ\cdots\circ w}_{i \mbox{ times}} (\mathcal{Z}_0).\] Obviously, \[\label{md93} \mathcal{D}=\cup_{i\geq 0} w^i(\mathcal{Z}_0).\] If the union in [\[md93\]](#md93){reference-type="eqref" reference="md93"} is finite, we say that **the Markov diagram is finite**. We define recursively the \(\pmb{n}\)**-th level of the Markov diagram** as \[\label{md53} \mathcal{Z}_n:=\omega(\mathcal{Z}_{n-1})\setminus \cup_{i=0}^{n-1} \mathcal{Z}_{i},\] for \(n\geq 1\). One can imagine the Markov diagram as a (potentially infinitely big) directed graph, with vertex set \(\mathcal{D}\). Between \(C,D\in \mathcal{D}\), we have a directed edge \(C\to D\) if and only if \(D\in w(C)\). We call the Markov diagram **irreducible** if there exists a directed path between any two intervals \(C,D\in \mathcal{D}\). In the next lemma we prove that by choosing an appropriate refinement \(\mathcal{Y}_0\) of \(\mathcal{Z}_0\), the Markov diagram \((\mathcal{D}^{'},\to)\) of \(\mathcal{F}\) with respect to \(\mathcal{Y}_0\) always has an irreducible subdiagram. Further, the elements of \(\mathcal{D}^{\prime}\) cover \(\Lambda\). It implies that \((\mathcal{D}^{'},\to)\) is sufficient to describe the orbits of the points of \(\Lambda\). That is, Lemma [\[md91\]](#md91){reference-type="ref" reference="md91"} enables us to assume that the Markov diagram \((\mathcal{D},\to)\) of \(\mathcal{F}\) is irreducible without loss of generality. ## Connection to the natural pressure {#md19} Similarly to graph-directed iterated function systems, we associate a matrix to Markov diagrams, that will help us determine the Hausdorff dimension of the corresponding CPLIFS (see ). We used in the definition, that for a \(D\in \mathcal{D}\) with \(C\to_{(k,j)} D\) the derivative of \(f_{k,j}\) over \(D\) is a constant number. That is each element of \(\mathbf{F}(s)\) is either zero or a sum of the \(s\)-th power of some contraction ratios. This matrix can be defined for any \(\mathcal{C}\subset \mathcal{D}\) as well, by choosing the indices from \(\mathcal{C}\) only. We write \(\mathbf{F}_{\mathcal{C}}(s)\) for such a matrix. It follows that \(\mathbf{F}_{\mathcal{C}}(s)\) is always a submatrix of \(\mathbf{F}(s)\) for \(\mathcal{C}\subset \mathcal{D}\). We write \(\mathcal{E}_{\mathcal{C}}(n)\) for the set of \(n\)-length directed paths in the graph \((\mathcal{C},\to)\). \[\label{md87} \begin{aligned} \mathcal{E}_{\mathcal{C}}(n):= &\{ ((k_1,j_1),\dots ,(k_n,j_n)) \big\vert \exists C_1,\dots ,C_{n+1}\in \mathcal{C}: \\ &\forall q\in [n] \exists k_q\in[m], j_q\in[l(k)]): C_{q}\to_{(k_q,j_q)} C_{q+1} \}. \end{aligned}\] An \(n\)-length directed path here means \(n\) many consecutive directed edges, and we identify each such path with the labels of the included edges in order. Each path in \((\mathcal{D},\to)\) of infinite length represents a point in \(\Lambda\), and each point is represented by at least one path. Similarly, for \(\mathcal{C}\subset \mathcal{D}\) the points defined by the natural projection of infinite paths in \((\mathcal{C},\to)\) form an invariant set \(\Lambda_{\mathcal{C}}\subset \Lambda\). We define the natural pressure of these sets as \[\label{md88} \Phi_{\mathcal{C}}(s):= \limsup_{n\to \infty}\frac{1}{n}\log \sum_{\mathbf{k}} \vert I_{\mathbf{k}}\vert^s,\] where the sum is taken over all \(\mathbf{k}=(k_1,\dots k_n)\) for which \(\exists j_1,\dots j_n: ((k_1,j_1),\) \(\dots ,(k_n,j_n))\in \mathcal{E}_{\mathcal{C}}(n)\), and \(I\) is the interval defined in [\[cr65\]](#cr65){reference-type="eqref" reference="cr65"}. By the definition of \(\mathcal{D}\) it is easy to see that \(\Phi_{\mathcal{D}}(s)=\Phi (s)\). We will show, that the unique zero of the function \(\Phi_{\mathcal{D}}(s)\) can be approximated by the root of \(\Phi_{\mathcal{C}}(s)\) for some \(\mathcal{C}\subset \mathcal{D}\). To show this, we need to connect the function \(\Phi_{\mathcal{C}}(s)\) to the matrix \(\mathbf{F}_{\mathcal{C}}(s)\). As an operator, \((\mathbf{F}_{\mathcal{D}}(s))^n\) is always bounded in the \(l^{\infty}\)-norm. Thus we can define \[\varrho (\mathbf{F}_{\mathcal{C}}(s)) := \lim_{n\to \infty} \lVert (\mathbf{F}_{\mathcal{C}}(s))^n\rVert_{\infty} ^{1/n}.\] Let \((\mathcal{C}_1,\to),(\mathcal{C}_2,\to),\dots\) be an increasing sequence of irreducible subgraphs of \((\mathcal{D},\to)\). It follows from Seneta's results that the so called \(R\)-values of the matrices \(\mathbf{F}_{\mathcal{C}_n}(s)\) converge to the \(R\)-value of \(\mathbf{F}(s)\). For an irreducible finite matrix \(\mathbf{A}\) we always have \(R(\mathbf{A})=\frac{1}{\varrho(\mathbf{A})}\), then the convergence of the spectral radius \(\varrho(\mathbf{F}_{\mathcal{C}_n}(s))\) to \(\varrho(\mathbf{F}(s))\) follows. Altough \(\mathbf{F}(s)\) may not be finite, the relation \(R(\mathbf{F}(s))=\frac{1}{\varrho(\mathbf{F}(s))}\) can still be guaranteed by some assumptions. That is why the following property has a crucial role in our proofs. In the next section we show how being limit-irreducible implies that the Hausdorff dimension of the attractor is equal to the minimum of the natural dimension and \(1\). Later, in Section [3](#md24){reference-type="ref" reference="md24"}, we investigate what makes a CPLIFS limit-irreducible. ## Proof using the diagrams {#md18} We have already shown a connection between the Markov diagram and the natural pressure of a given CPLIFS. Now using this connection, we show that the natural dimension of a limit-irreducible CPLIFS is always a lower bound for the Hausdorff dimension of its attractor, by approximating the spectral radius of the Markov diagram with its submatrices' spectral radius. As in, the following proposition holds. The proof is essentially the same as the proof of. We obtain the following theorem as the combinations of and. The proof is similar to the proof of Theorem 2 in. # What makes a CPLIFS limit-irreducible? {#md24} It is hard to check whether a CPLIFS \(\mathcal{F}=\{f_k\}_{k=1}^m\) is limit-irreducible, that is if it satisfies definition [\[md82\]](#md82){reference-type="ref" reference="md82"}. In this section, we show by a case analysis that the following proposition holds. Theorem [\[md46\]](#md46){reference-type="ref" reference="md46"} is a straightforward consequence of Proposition [\[md47\]](#md47){reference-type="ref" reference="md47"} and Theorem [\[md72\]](#md72){reference-type="ref" reference="md72"}. According to, if all functions in \(\mathcal{F}\) are injective and the first cylinders are not overlapping, then \(\mathcal{F}\) is limit-irreducible. This observation was utilized by Raith in the proof of. In this section we always assume that \(s\in(0,\dim_{\rm H}\Lambda]\). The overlapping structures may induce multiple edges in the Markov diagram. In the associated matrix \(\mathbf{F}(s)\) each multiple edge is represented as an entry of the form \(\rho_{k_1,j_1}^s+\dots+\rho_{k_n,j_n}^s\) for some \(n>1\). Since these entries can be bigger than \(1\) in absolute value, the assumptions of do not hold. We need to investigate under which conditions can help us. For the convenience of the reader we also present Hofbauer's proof here. Lemma [\[md84\]](#md84){reference-type="ref" reference="md84"} implies that for \(s\in (0,\dim_{\rm H}\Lambda)\) we have \(\varrho (\mathbf{F}(s))>1\), where \(\Lambda\) is the attractor of the CPLIFS \(\mathcal{F}\). Therefore, in order to apply Lemma [\[md65\]](#md65){reference-type="ref" reference="md65"}, it is enough to show that \[\label{md61} \lim_{N\to\infty} \varrho \left( \mathbf{F}_{\mathcal{D}\setminus\mathcal{D}_N}(s) \right) =1.\] If [\[md61\]](#md61){reference-type="eqref" reference="md61"} holds, then \(\mathbf{F}_{\mathcal{D}\setminus\mathcal{D}_N}(s)\) can take the place of the submatrix \(S\) in Theorem [\[md65\]](#md65){reference-type="ref" reference="md65"} for a big enough \(N\). In the special case of expansive piecewise monotonic mappings, [\[md61\]](#md61){reference-type="eqref" reference="md61"} was verified by F. Hofbauer. To extend his results to CPLIFS, we need to show the same for our expansive multi-valued mappings \(T\). The only difference between our and his Markov diagrams is the occurence of multiple edges, caused by the possible overlappings. We note that not all of the overlappings induce multiple edges, as monotonicity intervals of the same level might overlap. Note that the graphs of the branches of \(T\) can only intersect at the endpoint of some base interval \(Z\in\mathcal{Z}_0\) (see Definition [\[md56\]](#md56){reference-type="ref" reference="md56"}). We say that the **order of overlapping** is \(K\) if the maximal number of branches of \(T\) that have intersecting domains is \(K\). ## The case of light overlaps Lemma [\[md63\]](#md63){reference-type="ref" reference="md63"} implies that for a CPLIFS with only light overlaps also holds. For the convenience of the reader, we include here a modified version of the proof of. Lemma [\[md65\]](#md65){reference-type="ref" reference="md65"} and Proposition [\[md42\]](#md42){reference-type="ref" reference="md42"} together gives \[\mathcal{F} \mbox{ has only light overlaps } \implies \mathcal{F} \mbox{ is limit-irreducible.}\] That is for a CPLIFS \(\mathcal{F}\) with only light overlaps and with a generated self-similar system satisfying the ESC, we always have \(\dim_{\rm H}\Lambda=\min\{1,s_{\mathcal{F}}\}\), where \(\Lambda\) is the attractor of \(\mathcal{F}\). ## The case of cross overlaps We call the elements of the set \[\left\{ x\in \mathcal{I} \big\vert \exists k_1,k_2\in [m], \exists j_1\in[l(k_1)], \exists j_2\in[l(k_2)]: f_{k_1, j_1}^{-1}(x)=f_{k_2, j_2}^{-1}(x) \right\}\] **intersecting points**. They form a subset of the critical points \(\mathcal{K}\). Let \(w\in I\) be an intersecting point, then the elements of \(\mathcal{D}\) can only contain \(w\) as their endpoint. If \(D\in \mathcal{D}\) ends in \(w\), then we say that \(D\) **is causing cross overlaps** at \(w\).

{'timestamp': '2023-02-10T02:02:05', 'yymm': '2302', 'arxiv_id': '2302.04333', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04333'}

# Introduction The following problem was posed by W. Hengartner at the second international workshop on planar harmonic mappings at the Technion, Haifa, January 7-13, 2000. The problem has been restated several times (see, , and ). The condition that \(p\) is not a constant multiple of \(q\) ensures that the valence is finite, see. The term *logharmonic* refers to the fact, which is apparent, that the log of such an \(f\) is locally harmonic on the set excluding its zeros. Concerning the case \(m=1\), Bshouty and Hengartner made the following conjecture in. In referring to the maximal valence, this conjecture implicitly contains two assertions: (i) in the case \(m=1\), for each \(n \geq 1\) the valence is at most \(3n-1\) and (ii) for each \(n \geq 1\) there exists a polynomial \(p(z)\) of degree \(n\), a linear \(q(z)\), and a \(w \in \mathbb{C}\) such that there are \(3n-1\) preimages of \(w\) under \(f\). In this paper, as our first main result, we prove the following theorem confirming the first assertion (i), i.e., the upper bound that is entailed in Bshouty and Hengartner's conjecture. Regarding the assertion (ii), i.e., the sharpness part of the conjecture, examples are provided in having valence \(3n-3\), which shows that the above result is asymptotically sharp. There is additionally an example presented in for \(n=2\) with valence \(3n-1=5\) showing sharpness in the case \(n=2\). In Section [3](#sec:numerics){reference-type="ref" reference="sec:numerics"} below, we present additional examples that, according to our numerical simulations, attain the upper bound \(3n-1\) in the case \(n=3,4\). It is an enticing open problem to confirm the second statement (ii) implicit in Bshouty and Hengartner's conjecture by producing examples showing that Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} is sharp for each \(n\). Returning to the general setting of Hengartner's Problem, it was observed in that Bezout's Theorem gives an upper bound \((n+m)^2\) on the valence, and the authors conjectured the following. Since \(n^2+m^2 < (m+n)^2\) when \(m,n \geq 1\), the following theorem, our second main result of this paper, improves Bezout's bound for every \(m,n \geq 1\), hence confirming the above conjecture of Abdulhadi and Hengartner. This result admits further improvement, at least in the case \(m=1\), in light of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"}. The question was raised in whether the maximal valence is actually linear in \(m,n\). We suspect that the maximal valence grows linearly in \(n\) for each fixed \(m\) but grows quadratically when \(n=m\) or when \(n\) is asymptotically proportional to \(m\). Our proof of Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} uses a purely algebraic technique based on polynomial resultants applied to the system of equations obtained by considering \(p(z) \overline{q(z)}-1 =0\) and its complex conjugate \(\overline{p(z)} q(z)-1 = 0\). A technical obstacle in using this method is establishing the following result (stated as Lemma [\[lem:coprime1\]](#lem:coprime1){reference-type="ref" reference="lem:coprime1"} in Section [4](#sec:general){reference-type="ref" reference="sec:general"}) showing that the left hand sides of these equations, viewed as bivariate polynomials in \(z, \overline{z}\), are coprime. As a corollary of Lemma [\[lem:coprimeIntro\]](#lem:coprimeIntro){reference-type="ref" reference="lem:coprimeIntro"} we recover the following aforementioned result that was proved in using several complex analytic tools (including the Riemann mapping theorem, Schwarz reflection, analytic continuation, and properties of Blaschke products), whereas our proof of Lemma [\[lem:coprimeIntro\]](#lem:coprimeIntro){reference-type="ref" reference="lem:coprimeIntro"} is purely algebraic and essentially relies on a single property of resultants (namely, that coprimacy is equivalent to the resultant not vanishing identically). Concerning the *minimal* valence we note, as was observed in, that for \(n>m\) it follows from a generalization of the argument principle that the valence is bounded below by \(n-m\). This lower bound is sharp (and hence gives the minimal valence) as it is easy to check (especially in polar coordinates \(z = r e^{i \theta}\)) that the equation \(z^n \overline{z}^m = 1\) has exactly \(n-m\) solutions (located at roots of unity). ## Sheil-Small's valence problem for harmonic polynomials Hengartner's valence problem for logharmonic polynomials is reminiscent of another valence problem posed by T. Sheil-Small, where instead of a product of \(p(z)\) and \(\overline{q(z)}\) a sum \(p(z) + \overline{q(z)}\) is considered. In his thesis, A. S. Wilmshurst used the maximum principle for harmonic functions together with Bezout's theorem to show that the valence is at most \(n^2\), and he constructed \(n^2\)-valent examples with \(m=n-1\). This confirmed a conjecture of Sheil-Small that \(n^2\) is the maximum valence. Wilmshurst conjectured (cf. ) that for each pair of integers \(n>m\geq1\), the maximum valence is \(3n-2 + m(m-1)\). This conjecture was confirmed for \(n=m-1\) by the above mentioned results of Wilmshurst and for \(m=1\) by a result of the first-named author and G. Swiatek together with L. Geyer's proof of the Crofoot-Sarason conjecture. However, counterexamples to Wilmshurst's conjecture for the case \(m=n-3\) were provided in . Further counterexamples were provided in, , and. In spite of these counterexamples, it is still seems likely that, in the spirit of Wilmshurst's conjecture, the maximal valence increases linearly in \(n\) for each fixed \(m\), ,. The latter result used a nonconstructive probabilistic method, thus connecting the study of the extremal problem with a parallel line of research on the average number of zeros of random harmonic polynomials that had been investigated (for a variety of models) in, , ,. Harmonic polynomials and logharmonic polynomials are just two species of so-called polyanalytic polynomials, that is polynomials \(P(z,\overline{z})\) in \(z\) and \(\overline{z}\) that satisfy \(\left(\frac{\partial}{\partial \overline{z}}\right)^m P(z,\overline{z}) = 0\) with \(\frac{\partial}{\partial \overline{z}}:=\frac{1}{2} \left( \frac{\partial}{\partial x} + i \frac{\partial}{\partial y} \right)\) and hence take the form \[P(z,\overline{z}) = \sum_{k=0}^m p_k(z) \overline{z}^k,\] where \(p_k(z)\) are complex-analytic polynomials. The above Problems [\[prob:Hen\]](#prob:Hen){reference-type="ref" reference="prob:Hen"} and [\[prob:SS\]](#prob:SS){reference-type="ref" reference="prob:SS"} thus fit into a broader setting concerning the valence of polyanalytic polynomials. Polyanalytic polynomials have been studied classically and have also arisen in more recent studies in gravitational lensing, , approximation theory, , numerical linear algebra, random matrix theory, and determinantal point processes. Motivated by the valence problems for harmonic and logharmonic polynomials as well as the prominent role played by zero sets in several of the aforementioned studies on polyanalytic polynomials, it seems natural to pose the following general valence problem for polyanalytic polynomials. The proof of Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} in Section [4](#sec:general){reference-type="ref" reference="sec:general"} leads to the following result (see Lemma [\[lem:PRL\]](#lem:PRL){reference-type="ref" reference="lem:PRL"} in Section [4](#sec:general){reference-type="ref" reference="sec:general"}) estimating the valence of polyanalytic polynomials for which \(\deg p_k \leq n\) for each \(k=0,1...,m\). ## Proof techniques The proof of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} begins with a reformulation of the valence problem in terms of rational harmonic functions and then specializes a technique previously used in that setting by the first named author and G. Neumann in, where the key idea is to use a theorem of Fatou from holomorphic dynamics. This indirect method was introduced by the first named author and G. Swiatek in in their above-mentioned proof of the \(m=1\) case of Wilmshurst's conjecture. Subsequent adaptations of this method have led to solutions to additional problems including sharp estimates for the topology of quadrature domains and classification of the number of critical points of Green's function on a torus. The results of, which are directly relevant to the current paper, confirmed astronomer S. Rhie's conjecture in grativational lensing (motivated by the connection to gravitational lensing the zeros of rational harmonic functions were investigated further in, ,, ,, ). The proof of Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} utilizes another technique that has been applied in studies of gravitational lensing, in this instance, going in a separate, purely algebraic direction utilizing Sylvester resultants. The application of Sylvester resultants to bound the valence of certain polynomials in \(z\) and its complex conjugate \(\overline{z}\) was introduced by A.O. Petters in the study of gravitational lensing in. Our proof of Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} particularly resembles the more recent application of resultants in the study of gravitational lensing by multi-plane point mass ensembles carried out by the third-named author in. ## Outline of the paper We present the proof of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} in Section [2](#sec:BHconjUB){reference-type="ref" reference="sec:BHconjUB"} where we begin by reviewing some preliminary results in Section [2.1](#sec:prelim){reference-type="ref" reference="sec:prelim"} before proceeding to the proof in Section [2.2](#sec:proofBHconjUB){reference-type="ref" reference="sec:proofBHconjUB"}. Our proof of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} adapts the method from with an important modification, see Remark [\[rmk:modification\]](#rmk:modification){reference-type="ref" reference="rmk:modification"}. We discuss Bshouty and Hengartner's \((3n-3)\)-valent examples from the perspective of holomorphic dynamics in Section [3](#sec:numerics){reference-type="ref" reference="sec:numerics"}, where we also present the results of some numerical experiments. We present the proof of Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} in Section [4](#sec:general){reference-type="ref" reference="sec:general"}, where we include review of some essential algebraic preliminaries related to Sylvester resultants. # Anti-holomorphic dynamics and the case \(m=1\) {#sec:BHconjUB} ## Preliminaries {#sec:prelim} In this section, we collect some notation and terminology along with several lemmas that will be used in the proof of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"}. ### Anti-holomorphic dynamics We will need the following result that adapts a classical theorem of Fatou from the setting of holomorphic dynamics to the setting of anti-holomorphic dynamics. A point \(z_0 \in \hat{\CC}\) is a *critical point* of \(r\) if the spherical derivative of \(r\) vanishes at \(z_0\). The complex conjugate of a rational function is referred to as an *anti-rational map*. A fixed point \(z_0\) of the anti-rational map \(z \mapsto \overline{r(z)}\) is referred to as an *attracting* fixed point if \(|r'(z_0)|<1\). A fixed point \(z_0\) is said to attract some point \(w \in \hat{\mathbb{C}}\) if the iterates of \(z \mapsto \overline{r(z)}\) starting at \(w\) converge to \(z_0\). The following result is taken from. ### The argument principle for harmonic maps We will also need the following generalization of the argument principle for harmonic functions. We recall that a zero \(z_0\) of a harmonic function is referred to as a *singular zero* of \(F(z)=h(z) + \overline{g(z)}\) if the Jacobian of \(F\), that is easily computed to be \(|h'(z)|^2-|g'(z)|^2\), vanishes at \(z_0\). Moreover, \(F\) is said to be sense-preserving (respectively sense-reversing) at \(z_0\) if the Jacobian is positive (respectively negative) at \(z_0\). The *order* of a sense-preserving zero is defined to be the smallest positive integer \(k\) such that \(h^{(k)}(z_0) \neq 0\). The order of a sense-reversing zero \(z_0\) is defined to be \(-k\) where \(k\) is the order of \(z_0\) as a sense-preserving zero of \(\overline{F(z)}\). If \(F\) is harmonic in a punctured neighborhood of \(z_0\) and \(F \rightarrow \infty\) as \(z \rightarrow z_0\) then \(z_0\) is referred to as a *pole*. As defined in, the *order of a pole* is given by \(\frac{1}{2\pi} \Delta_{T} \arg F\), where \(T\) is a sufficiently small circle centered at the pole, and \(\Delta_{T} \arg F\) denotes the increment of the argument of \(F(z)\) along \(T\). It follows that the order of a pole \(z_0\) is negative when \(F\) is sense-reversing in a punctured neighborhood of \(z_0\), and elaborating on this point we set aside the following remark that will be used later. The following version of the argument principle is taken from. We note that there are topological versions of the argument principle (having classical roots going back to the Poincaré-Hopf Theorem for vector fields) that hold in more general settings ,. ### Tools for reducing the general case to the generic case The generalized argument principle stated in Lemma [\[lem:argprinc\]](#lem:argprinc){reference-type="ref" reference="lem:argprinc"} above can only be applied in the nondegenerate case where the harmonic mapping is free of singular zeros. The next two lemmas will be used to reduce the problem of establishing the upper bound in the general case (where singular zeros are allowed) to establishing it for the nondegenerate case. Harmonicity comes into play in this reduction, and there is no analogous strategy in general smooth settings. To elaborate, the first of these lemmas establishes a genericity result for nondegenerate examples. It is an important principle of differential topology that results of this type hold more generally for smooth maps. On the other hand, the second lemma establishes a stability result for extremal examples, and as pointed out in this type of result does not hold for arbitrary smooth maps. The following result is taken from. The following result is taken from. ## Proof of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} {#sec:proofBHconjUB} The problem is to bound the number of solutions of \[\label{eq:original} p(z) \overline{q(z)} = w\] when \(q(z)\) is of degree \(m=1\) and \(p\) is a polynomial of degree \(n>1\). By factoring out the leading coefficient of \(q\) and assimilating it into \(p\), we may assume without loss of generality that \(q(z) = z+b\), \(b\in\mathbb{C}\). Since we easily get an upper bound of \(n+1\) in the case \(w=0\) (by setting each factor \(p(z)\) and \(\overline{z+b}\) to zero), we may assume \(w \neq 0\). By multiplying \(p\) by \(1/w\) we may further assume without loss of generality \(w=1\). Applying these reductions and rearranging [\[eq:original\]](#eq:original){reference-type="eqref" reference="eq:original"}, the problem is then to bound the number of solutions of the equation \[\label{eq:reduced} \frac{1}{\overline{p(z)}}-b-z = 0,\] i.e., zeros of the harmonic rational function \(z \mapsto \frac{1}{\overline{p(z)}}-b-z\). We first estimate the number of sense-reversing zeros in the following proposition that is a specialized version of. The proof of the proposition essentially follows with one important modification (see Remark [\[rmk:modification\]](#rmk:modification){reference-type="ref" reference="rmk:modification"} below). We simplify the presentation of the proof (in comparison with the proof of ) by making use of Lemma [\[lem:Fatou\]](#lem:Fatou){reference-type="ref" reference="lem:Fatou"} in place of the classical theorem of Fatou and by settling for a slightly weaker (yet sufficient for our purposes) result that avoids consideration of so-called "neutral" fixed points. From the discussion preceding Proposition [\[prop:dynamics\]](#prop:dynamics){reference-type="ref" reference="prop:dynamics"}, it follows that the problem of establishing Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} reduces to proving the following result. # Numerical results {#sec:numerics} The example from with valence at least \(3n-3\) is the following. Let \[\label{eq:example} f(z) = \left( \frac{z^{n}}{n}-z \right) \overline{z} = |z|^2 \left( \frac{z^{n-1}}{n}-1 \right).\] As pointed out in, for \(w=-\frac{n-1}{n}\) the roots of unity \(z_k = e^{2\pi i k / (n-1)}\), (\(k = 1,2...,n-1\)) are solutions of \(f( z) = w\). Following the reformulation we used above, let us rewrite \(f(z)=w\) as \[\label{eq:example2} \overline{r(z)} = z,\] where \[r(z):= \frac{n w}{z^n-nz} =-\frac{n-1}{z^n-nz}.\] Then the roots of unity \(z_k\) are attracting fixed points of \(\overline{r(z)}\). Indeed, we have \[\begin{aligned} r(z_k) &=-\frac{n-1}{z_k (z_k^{n-1}-n)} \\ &=-\frac{n-1}{z_k (1-n)} = \frac{1}{z_k} = \overline{z_k}, \end{aligned}\] where we used \(|z_k| = 1\) in the last equality. This shows that each root of unity \(z_k\) is a fixed point. To see that each such \(z_k\) is attracting, we compute \[r'(z) = (nz^{n-1}-n)\frac{n-1}{(z^{n}-nz)^2},\] and notice that this vanishes when evaluated at a root of unity, i.e., we have \(r'(z_k)=0\). Thus, they are actually super-attracting fixed points. Attracting fixed points of \(\overline{r(z)}\) are orientation-preserving solutions of \(\overline{r(z)}-z = 0\). By the generalized argument principle, this leads to at least \(3n-3\) total solutions. In fact, we can see from inspecting the dynamics of iteration of the map \(z \mapsto \overline{r(z)}\) that there are exactly \(3n-3\) for this example, since the critical point at infinity is not attracted to a fixed point. On the contrary, it is in a period-two orbit \(\infty \rightarrow 0 \rightarrow \infty \rightarrow 0 \cdots\). This diminishes the possible number of attracting fixed points by one, i.e., inspecting the proof of Proposition [\[prop:dynamics\]](#prop:dynamics){reference-type="ref" reference="prop:dynamics"} the upper bound stated in the proposition becomes \(n-1\). This then diminishes the total count by two. Indeed, following the argument in the proof of Theorem [\[thm:reduced\]](#thm:reduced){reference-type="ref" reference="thm:reduced"} using the generalized argument principle, we have \(n_+ \leq n-1\), \(n_-\leq 2n-2\), and \(N = n_+ + n_-\leq 3n-3\), so that the valence of this example is at most \(3n-3\) (and hence exactly \(3n-3\)). In order to show sharpness of the upper bound in Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"}, we would like to modify this example to have an additional attracting fixed point that attracts the critical point at infinity, leading to a total of \(3n-1\) solutions of \(\overline{r(z)}=z\). According to numerics presented in the subsection below, this strategy succeeds at least for \(n=3,4\), but it remains an open problem to verify this analytically and extend it to each \(n\). ## Numerical examples showing sharpness for \(n=3,4\) Let us consider a modified version of Bshouty and Hengartner's example with the following slightly different choice of \(r\) that includes two real parameters \(A>0, C \geq 0\) (adding \(C\) prevents infinity from going to the origin under iteration, and including \(A\) gives some extra flexibility while trying to tune parameters to produce an additional attracting fixed point) \[r(z):= \frac{A}{z\left[z^{n-1}+n\left(\frac{A}{n-1}\right)^{(n-1)/(n + 1)}\right]}+C.\] When \(C=0\) and \(A=1\), this reduces to (a rotation of) the Bshouty-Hengartner example, with attracting fixed points at roots of \((-1)\) instead of roots of unity. When \(C=0\) and \(A\) varies, there are still \(n-1\) attracting fixed points on a common circle, namely, they are at the \((n-1)\)th roots of \((-1)\) scaled by \(R(A,n):=\left(\frac{A}{n-1}\right)^\frac{1}{n+1}\), i.e., for each \(k=1..,n-1\), the point \(z_k= R(A,n) e^{i\pi\frac{1+2k}{n-1}}\) is a critical point for \(r(z)\) that is also a fixed point of \(\overline{r(z)}\). According to numerical experiments done in Mathematica, when \(n=3\), \(A=100\), \(C=0.68\), this example has three attracting fixed points and hence \(8=3n-1\) solutions. The attracting fixed points are located at \(z \approx 1.257 \pm 2.069 i\) and \(z \approx 2.31\). The critical point at infinity is attracted to the latter fixed point, and the symmetric pair of critical points on the imaginary axis are each attracted to the nearest of the other two fixed points. Similarly, we were able to locate parameter values that, according to numerical simulation, produce an example showing sharpness of Theorem [\[thm:BHconjUB\]](#thm:BHconjUB){reference-type="ref" reference="thm:BHconjUB"} in the case \(n=4\), namely with \(A=250\) and \(C=.86\) we find four attracting fixed points and hence \(11 = 3n-1\) solutions. # An improvement on the Bezout bound {#sec:general} Recall that Hengartner's Valence Problem asks for a bound on the number of solutions to \[\label{eq:logharmonic} p(z) \overline{q(z)} = w,\] where \(w\) is an arbitrary complex number, and \(p\) and \(q\) are (analytic) complex polynomials, of degree \(n\) and \(m\) respectively, such that \(p\) is not a constant multiple of \(q\). The case \(w=0\) admits a bound of \(m+n\) on the valence (by separately setting \(p(z)=0\), \(\overline{q(z)}=0\)), so we may assume \(w \neq 0\). Furthermore, without loss of generality (by mulitplying by \(1/w\) and renaming the coefficients of, say, \(p\)) we assume \(w=1\). Let us further reformulate the problem as one of bounding the number of zeros of the polynomial \(P(z,\overline{z}) = p(z)\overline{q(z)}-1\). Writing \(Q(z,\overline{z}) = \overline{P(z,\overline{z})}\), we note that \(P(z,\overline{z})\) and \(Q(z,\overline{z})\) have the same zeros. Hence, the number of zeros \(z \in \mathbb{C}\) of \(P(z,\overline{z})\) is bounded from above by the number of common zeros \((z_1,z_2) \in \mathbb{C}^2\) of \(P(z_1,z_2)\) and \(Q(z_1,z_2)\). An application of Bezout's theorem (which can be justified when \(p\) is not a constant multiple of \(q\) ) then gives an upper bound of \((n+m)^2\) on the number of zeros, an estimate that we will refer to as the *Bezout bound*. The main goal of this section is to improve the Bezout bound by proving Theorem [\[thm:general\]](#thm:general){reference-type="ref" reference="thm:general"} which in light of the above discussion reduces to establishing the following result. Before presenting the proof of this result, we review some preliminaries concerning Sylvester resultants. While our proof of Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"} uses essentially the same ideas as in, , in the current setting we are able to establish coprimacy of the polynomials \(P,Q\) without imposing an additional nondegeneracy condition. It is an open problem in the theory of gravitational lensing to determine whether the nondegeneracy assumption can be removed from the results of, , bounding the number of apparent images of a single background source lensed by a multiplane gravitational lens comprised of point masses. ## Sylvester Resultants Let \(P\in\mathbb{C}[z_1,z_2]\) be a bivariate polynomial with complex coefficients and variables. Written variously \[\label{eqn:bipolyP} P(z_1, z_2) = \sum_{i=0}^n a_i(z_2)z_1^i = \sum_{j=0}^m b_j(z_1)z_2^j = \sum_{i,j} c_{i,j}z_1^iz_2^j\] we denote by \(\deg_{z_1} P = n\), \(\deg_{z_2} P = m\), and \(\deg P = N = \max \{ i+j \;\:|\;\:c_{i,j} \neq 0 \}\) the *degree in \(z_1\)*, the *degree in \(z_2\)*, and the *total degree* of P respectively. Of course, \(\deg P \leq \deg_{z_1} P + \deg_{z_2}P\). Let \(Q \in \mathbb{C}[z_1,z_2]\) be another such polynomial written \[\label{eqn:bipolyQ} Q(z_1, z_2) = \sum_{i=0}^s d_i(z_2)z_1^i = \sum_{j=0}^t e_j(z_1)z_2^j = \sum_{i,j} f_{i,j}z_1^iz_2^j.\] Denote \(\deg_{z_1} Q = s\), \(\deg_{z_2} Q = t\), and \(\deg Q = M\). Writing \(P\) and \(Q\) as polynomials in \(z_1\), their coefficients are polynomials in \(z_2\), namely \(a_0,..., a_n\) and \(d_0...,d_s\) respectively. The *\(z_1\)-Sylvester Matrix* \(\mathcal{S}_{z_1}(P,Q)\) of \(P\) and \(Q\) is an \((n+s) \times (n+s)\) matrix which is constructed by diagonally stacking \(s\) copies of the polynomials \(a_i\) and \(n\) copies of the \(d_i\) with \(0\) used for the other entries as follows. \[\label{eqn:Sylvester} \mathcal{S}_{z_1}(P,Q)= \left[\arraycolsep=3pt\def1.4{1} \begin{array}{ccccc|cccc} a_n & a_{n-1} & a_{n-2} & \cdots & a_1 & a_0 & 0 & \cdots & 0 \\ 0 & a_{n} & a_{n-1} & a_{n-2} & \cdots & a_1 & a_0 & \cdots & 0 \\ \vdots & & \ddots & & & & & \ddots \\ 0 & \cdots & 0 & a_n & a_{n-1} & a_{n-2} & \cdots & a_1 & a_0 \\ \hline d_s & d_{s-1} & \cdots & d_0 & 0 & 0 & \cdots & & 0 \\ 0 & d_s & d_{s-1} & \cdots & d_0 & 0 & \cdots & &0 \\ 0 & 0 & d_s & d_{s-1} & \cdots & d_0 & \cdots & & 0 \\ \vdots & & & \ddots & & & \ddots & \\ 0 & \cdots & & 0 & d_s & d_{s-1} & \cdots & & d_0 \\ \end{array} \right].\] Formally, for \(1\leq i \leq s\), entry \(\mathcal{S}_{z_1}(P,Q)_{i,j} = a_{n-j+i}\) for \(i\leq j \leq i+n\) and 0 otherwise. Likewise for \(s+1 \leq i \leq s+n\) entry \(\mathcal{S}_{z_1}(P,Q)_{i,j} = d_{i-j}\) for \(i-s\leq j \leq i\) and 0 otherwise. Note that the exact location of columns in the top and bottom halves of [\[eqn:Sylvester\]](#eqn:Sylvester){reference-type="eqref" reference="eqn:Sylvester"} depends on the values of \(n\) and \(s\), but the matrix always has the indicated block structure where the top right \(s \times s\) block is lower triangular and the bottom left \(n \times n\) block is upper triangular. Here, the *\(z_1\)-Sylvester resultant* of \(P\) and \(Q\) is the determinant of the \(z_1\)-Sylvester matrix of \(P\) and \(Q\) and will be denoted \[\mathcal{R}_{z_1}(P,Q) = \det \mathcal{S}_{z_1} (P,Q)\] or as \(\mathcal{R}\) where the context is unambiguous. It is itself a polynomial in \(z_2\) whose degree is bounded by \(nt + sm\).. We say that \(P\) and \(Q\) are *coprime* if their greatest common factor has degree zero, i.e. \(P=CP_1\) and \(Q=CQ_2 \implies \deg C = 0\). The following is shown in Proposition 1 of. The next result follows from and the lemma above. The Resultant Theorem is useful when paired with the additional constraint that the second variable is the conjugate of the first (as will be considered in the next subsection). We note in passing that in the absence of such a constraint, Theorem [\[thm:Res\]](#thm:Res){reference-type="ref" reference="thm:Res"} provides no advantage over the Bezout bound. Indeed, the Bezout bound on the number of common zeros of \(P\) and \(Q\) as defined above is \[\#\{(z_1,z_2) \;\: | \:\; P(z_1,z_2) = 0 = Q(z_1,z_2) \} \leq MN.\] The Resultant Theorem, on the other hand, gives \[\# \{ z_2 \;\:|\;\: P(z_1, z_2)=0=Q(z_1,z_2) \;\text{for some}\; z_1\}\leq nt+ms\] or symmetrically: \[\# \{ z_1 \;\:|\;\: P(z_1, z_2)=0=Q(z_1,z_2) \;\text{for some}\; z_2\}\leq sm+tn.\] But since each \(z_1\) solution could, a priori, have the full set of \(z_2\) solutions as partners, the above two estimates together produce the rather poor estimate \[\# \{ (z_1,z_2) \;\:|\;\: P(z_1, z_2)=0=Q(z_1,z_2)\}\leq (nt+ms)^2,\] which is quite a bit larger than the Bezout bound. ## Polynomials in \(z\) and \(\overline{z}\) {#sec:polyanalytic} As in the statement of Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"}, we now consider the additional restriction that \(z_2 = \overline{z_1}\), so that \(P(z_1,z_2)=P(z,\overline{z})\) is a polyanalytic polynomial, and we also restrict to the case that \(Q(z,\overline{z}) = \overline{P(z,\overline{z})}\). Formally treating \(z\) and \(\overline{z}\) as separate variables we have \(\deg_z P = n = \deg_{\overline{z}} Q\) and \(\deg_{\overline{z}} P = m = \deg_z Q\). Let \(N=\deg(P)=\deg(Q)\). Following, the Bezout bound is then \[\#\{(z,\overline{z}) \;\: | \:\; P(z,\overline{z}) = 0 \} \leq N^2.\] Note that \(N\leq n+m\) and hence \(N^2\leq(n+m)^2\). The corresponding bound provided by Theorem [\[thm:Res\]](#thm:Res){reference-type="ref" reference="thm:Res"} is \[\# \{ z_1 \;\:|\;\: P(z_1, z_2)=0 \;\text{for some}\; z_2 \}\leq m^2+n^2.\] But the restriction \(z_2 = \overline{z_1}\) means in this case each solution matches with only one partner. In other words: \[\# \{ (z,\overline{z}) \;\:|\;\: P(z, \overline{z})=0\}\leq m^2+n^2.\] This proves the following lemma: Let us refer to the bound \(n^2+m^2\) provided by this lemma as the *Resultant Bound*. Trivially, the resutant bound improves the Bezout bound \(N^2\) if and only if \[\label{eq:P<B} n^2+m^2 < N^2.\] This necessary and sufficient condition is depicted in Figure [\[fig:RvsB\]](#fig:RvsB){reference-type="ref" reference="fig:RvsB"}. We note that for the case \(P(z,\overline{z}) = p(z) \overline{q(z)}-w\) related to logharmonic polynomials we have \(N=n+m\), so that the condition [\[eq:P\<B\]](#eq:P<B){reference-type="eqref" reference="eq:P<B"} holds for all \(m,n \geq 1\). On the other hand, the condition [\[eq:P\<B\]](#eq:P<B){reference-type="eqref" reference="eq:P<B"} does not hold for any nontrivial harmonic polynomials, i.e., for polynomials of the form \(f(z) + \overline{g(z)}\) having \(\deg f, \deg g >0\). ## Coprimacy of \(P\) and \(\overline{P}\) and proof of Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"} Let \(p\) and \(q\) be complex univariate polynomials of degree \(n\) and \(m\) respectively. We write these as \[p(z)=\sum_{k=0}^n p_k z^k \;\;\;\text{and} \;\;\; q(z)=\sum_{k=0}^m q_k z^k\] where \(p_k,q_k\in \mathbb{C}\) and \(p_n,q_m \neq 0\). As in the statement of Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"}, let us denote \[\label{eqn:HenSimp} P(z,\overline{z}) = p(z)\overline{q(z)}-1,\] and as in the statement of Lemma [\[lem:PRL\]](#lem:PRL){reference-type="ref" reference="lem:PRL"} we denote \[Q(z,\overline{z}) = \overline{P(z,\overline{z})} = \overline{p(z)}q(z)-1.\] In view of Lemma [\[lem:PRL\]](#lem:PRL){reference-type="ref" reference="lem:PRL"}, all that remains to prove Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"} is to establish the following lemma ensuring that \(P\) and \(Q\) are coprime if \(p\) is not a constant multiple of \(q\). This informal representation is valid if \(n=m\), but is perhaps misleading otherwise. If \(n>m\), say \(n=m+k\). A more precise representation is then given by \[\label{eqn:BigSyl} \left[ \begin{array}{cccc|cccc|cccc} p_n \overline{q} & p_{n-1}\overline{q} & \cdots & p_{k+1}\overline{q} & p_{k}\overline{q} & p_{k-1}\overline{q} & \cdots &p_1\overline{q} & p_0\overline{q}-1 & 0 & \cdots & 0 \\ 0 & p_n\overline{q} & \cdots & p_{k+2}\overline{q} & p_{k+1}\overline{q} & p_{k}\overline{q} & \cdots &p_2\overline{q} & p_1\overline{q} & p_0\overline{q}-1 & \cdots & 0 \\ & & \ddots & & & & \ddots & & & \ddots& \\ 0 & 0 & \cdots & p_{n}\overline{q} & p_{n-1}\overline{q} & p_{n-2}\overline{q} & \cdots & p_m \overline{q} & p_{m-1}\overline{q}& p_{m-2}\overline{q} & \cdots & p_0\overline{q}-1 \\ \hline \\ q_m\overline{p} & q_{m-1}\overline{p} & \cdots &q_1\overline{p} & q_0\overline{p}-1 & 0 & \cdots & 0 & 0 & 0 & \cdots & 0 \\ 0 & q_m\overline{p} & \cdots & q_2\overline{p} & q_1\overline{p} & q_0\overline{p}-1 & \cdots & 0 & 0 & 0 & \cdots & 0 \\ & & \ddots & & & & \ddots & & & \ddots& \\ 0 & 0 & \cdots & 0 & 0 & 0 & \cdots & q_m\overline{p} & q_{m-1}\overline{p} & q_{m-2}\overline{p} & \cdots & q_0\overline{p}-1 \\ \end{array} \right].\] The proof of the lemma will use the characterization of coprimacy in terms of resultants provided by Lemma [\[lem:ZeroRes\]](#lem:ZeroRes){reference-type="ref" reference="lem:ZeroRes"}. Note that \(P\) and \(Q\) may be written \[P(z, \overline{z}) =-1 + \sum_{k=1}^n p_k \overline{q(z)} z^k, \quad \quad Q(z, \overline{z}) = q_0\overline{p(z)}-1 + \sum_{k=1}^m q_k\overline{p(z)} z^k,\] where we have assumed, (by precomposing with a translation), \(p_0 = 0\). Having written \(P,Q\) as polynomials in \(z\) with coefficients depending on \(\overline{z}\), recall that the resultant \(\mathcal{R}_{z}(P,Q)\) of \(P,Q\) is expressed in terms of those coefficients as the determinant of the Sylvester matrix [\[eqn:Sylvester\]](#eqn:Sylvester){reference-type="eqref" reference="eqn:Sylvester"}: \[\label{eqn:BigSyl} \left[\arraycolsep=3pt\def1.4{1.4} \begin{array}{cccc|cccc} p_n\overline{q(z)} & p_{n-1} \overline{q(z)} & \cdots & p_1\overline{q(z)} &-1 & 0 & \cdots & 0 \\ 0 & p_n\overline{q(z)} & p_{n-1} \overline{q(z)} & \cdots & p_1\overline{q(z)} &-1 & \cdots & 0 \\ \vdots & & \ddots & & & & \ddots \\ 0 & \cdots & 0 & p_n\overline{q(z)} & & \cdots & &-1 \\ \hline q_m\overline{p(z)} & \cdots & q_1 \overline{p(z)} & q_0\overline{p(z)}-1 & 0 & 0 &\cdots & 0 \\ 0 & q_m\overline{p(z)} & \cdots & q_1 \overline{p(z)} & q_0\overline{p(z)}-1 & 0 & \cdots & 0 \\ \vdots & & \ddots & & & \\ 0 & \cdots & 0 & q_m\overline{p(z)} & q_{m-1}\overline{p(z)} & & \cdots & q_0\overline{p(z)}-1 \\ \end{array} \right].\] To prove Lemma [\[lem:coprime1\]](#lem:coprime1){reference-type="ref" reference="lem:coprime1"}, we first establish the following intermediate step. We now prove Lemma [\[lem:coprime1\]](#lem:coprime1){reference-type="ref" reference="lem:coprime1"} (and hence conclude Theorem [\[thm:genReduced\]](#thm:genReduced){reference-type="ref" reference="thm:genReduced"} as explained above).

{'timestamp': '2023-02-10T02:02:13', 'yymm': '2302', 'arxiv_id': '2302.04339', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04339'}

# Introduction In this paper, we will achieve some stabilization results for solutions to an initial boundary value problem for the *Fractional Porous Medium Equation* (FPME for short), of the form \[\label{FPMEintro} \left\{\begin{array}{rcll} \partial_{t}u&=&-(-\Delta)^{s} (|u|^{m-1}u), & \mbox{ in } Q:=\Omega\times (0,+\infty),\\ u&=&0, & \mbox{ in } \mathbb R^{N}\setminus\Omega\times(0,+\infty),\\ u(\cdot,0)&=&u_0,& \mbox{ in } \Omega. \end{array} \right.\] Here we consider the porous medium regime, *i.e.* \(m>1\), we assume \(0<s<1\), and \(\Omega\) is a bounded open set of \(\mathbb R^N\). A broad theory has been developed for this problem under several aspects (existence, uniqueness, regularity etc.), see for instance. The main result of this paper concerns solutions emanating from initial data \(u_0\) for which the energy functional \[\frac12\int_{\mathbb R^N}\int_{\mathbb R^N} \inter{\varphi}\,dx\,dy-\frac{m+1}{(m-1)m} \int_\Omega |\varphi|^{\frac{m+1}{m}}\,dx\] does not exceed its first excited level when the choice \(\varphi= |u_0|^{m-1}u_0\) is made. By following the methodology of , in which the local case was considered, we compute the large time asymptotic profile of such solutions in this non-local framework. As in the local case, sign-changing initial data are included in the analysis: irrespective of their sign, if their energy is small enough then they give rise to solutions that in the large time limit are asymptotic to functions with a spatial profile arising in the energy minimization. For a precise statement, we need to introduce, for all \(q\in (1,2)\) and \(\alpha\in(0,+\infty)\), the functional defined on \({\mathcal D}^{s,2}_0(\Omega)\) (see Sect. [2](#sec2){reference-type="ref" reference="sec2"} for the precise definition of this space) by \[\label{functional} \ensuremath{{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\!\left(\varphi \right) = \frac12\int_{\mathbb R^N}\int_{\mathbb R^N} \inter{\varphi}\,dx\,dy-\frac{\alpha}{q} \int_\Omega |\varphi|^{q}\,dx\,,\] whose critical points, by definition, are the weak solutions of the Lane-Emden equation \[\label{ELLintro} (-\Delta)^s\varphi = \alpha\ |\varphi|^{q-2}\varphi\,,\qquad\text{in \(\Omega\),}\] with homogeneous Dirichlet boundary conditions. It is known  that the minimal energy \[\label{minimizer} \Lambda_1=\min\{\lyap\varphi\mathbin{\colon}\varphi\in {\mathcal D}^{s,2}_0(\Omega)\}\,,\] is achieved by a solution with constant sign, that it is unique (up to the sign). Also, we set \[\Phi(u)= |u|^{m-1}u\] and we observe that \(\Phi^{-1}(\varphi) = |\varphi|^{q-2}\varphi\) where \(q=(m+1)/m\). Now, following, we define the *second critical energy level*, or *first excited level*, as it follows \[\Lambda_{2}=\inf\Big\{\Lambda>\Lambda_1 \mathbin{\colon}\text{\(\Lambda\) is a critical value of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\)}\Big\}.\] The the solution to problem [\[FPMEintro\]](#FPMEintro){reference-type="eqref" reference="FPMEintro"} has the following stabilization property. Besides on \(\Omega\), the function \(w_\Omega\) that achieves the minimum in [\[minimizer\]](#minimizer){reference-type="eqref" reference="minimizer"} depends on \(m\) (through \(\alpha\) and \(q\)) and on \(s\); we refer to the material in Sect. [3](#sec:ell){reference-type="ref" reference="sec:ell"} for its existence and uniqueness, that are however well known. Recall that, in the case of nonnegative data \(u_{0}\), it is also well known (see ) that the solution \(u\) stabilizes towards the so called \(\emph{Friendly Giant}\), following the denomination due to Dahlberg and Kenig for the standard porous medium equation (see also ), namely \[S(x,t)=t^{-\alpha}w_{\Omega}(x)^{q-1}.\] That is a separate variable solution taking \(+\infty\) as initial value. In particular, in  various interesting results are shown, related to the finer problem of the sharp convergence rate of the relative error, a question that was also faced in the classical paper. As said, Theorem [\[mainthm1\]](#mainthm1){reference-type="ref" reference="mainthm1"} can be proved via the approach used in  to deal with the local problem, thanks to a Lyapunov-type property of the energy functional [\[functional\]](#functional){reference-type="eqref" reference="functional"}: namely, that \[t\longmapsto \lyap{\Phi(v(\cdot,t))} \text{is non-increasing,}\] whenever \(v\) is an energy solution (see Sect. [4](#sec:4){reference-type="ref" reference="sec:4"} for precise definitions) of the initial boundary value problem for the rescaled equation \[\label{FPMErescintro} \partial_t v + (-\Delta )^s\Phi(v) = \alpha v\.\] This property is inferred, in this paper, from an entropy-entropy dissipation estimate in Sect. [4](#sec:4){reference-type="ref" reference="sec:4"}. In order to prove it, we produce solutions by the classical Euler implicit time discretization scheme. That has the advantage of providing a discrete version of the desired inequality in which we can pass to the limit. We prefer this approach to considering exact solutions to a uniformly parabolic approximation, as done in  for the local problem, mainly because that would require \(C^{1,\alpha}\) estimates for the non-local operators that are obtained by regularizing the *signed* FPME; incidentally, we mention that strong results of this type can be found in  in the case of *non-negative* solutions. We recall here in brief the use of the the Lyapunov property for the proof of Theorem [\[mainthm1\]](#mainthm1){reference-type="ref" reference="mainthm1"}. Given a solution \(u\) of [\[FPMEintro\]](#FPMEintro){reference-type="eqref" reference="FPMEintro"}, the equation [\[FPMErescintro\]](#FPMErescintro){reference-type="eqref" reference="FPMErescintro"} for the function \(v(x,t)=e^{\alpha t} u(x,e^t-1)\) describes a system that evolves, irrespective of the starting conditions, to fixed points, *i.e.*, states of the form \(\Phi^{-1}(v)\) with \(v\) being a critical point of the energy functional. Because of the isolation of the energy minimizing solutions \(\pm w_\Omega\), proved in , this and the Lyapunov property imply that the disconnected set \(\{\pm\Phi^{-1}(w_\Omega)\}\) has a non-empty basin of attraction, including all initial states with energy smaller than the first excited level. (The isolation property implies some restriction on boundary regularity: assumptions weaker than those made in the our main statement are also feasible, but that is not the object of this paper.) Eventually, by the compactness of the relevant Sobolev embedding, by energy coercivity, and by the Lyapunov property, orbits are relatively compact; thus, the \(\omega\)-limit is connected and the only possible cluster point of the orbit emanating from an initial state \(u_0\) below the energy threshold is either \(\Phi^{-1}(w_\Omega)=w_\Omega^{q-1}\) or \(\Phi^{-1}(-w_\Omega)=-w_\Omega^{q-1}\). Yet, meeting the threshold requirement in Theorem [\[mainthm1\]](#mainthm1){reference-type="ref" reference="mainthm1"} implies no restriction on the sign that \(u_0\) should take in \(\Omega\), nonetheless. The relevance in energy of the nodal sets, instead, enters in predicting which one of the two possible limit profiles the orbit will accumulate to. The following Proposition, which is the non-local counterpart of an analogous result of , quantifies this idea. # Notations and Assumptions {#sec2} Throughout this paper, we assume \(\Omega\) to be an open bounded set, we take \(s\in(0,1)\), and we let \(m>1\). Then, we denote by \({\mathcal D}^{s,2}_0(\Omega)\) be the completion of \(C^\infty_0(\Omega)\) with respect to the norm \[\label{seminorm} [u]_{s} = \left\{ \int_{\mathbb R^N}\int_{\mathbb R^N} \inter{u}\,dx\,dy \right\}^\frac12\.\] For every \(u\in C^2(\Omega)\), we take \[\label{slap} (-\Delta)^su(x) = \lim_{\varepsilon\to0^+} \int_{\mathbb R^N\setminus B_\varepsilon(x)} \frac{u(x)-u(y)}{|x-y|^{N+2s}}\,dy\] as the definition of the \(s\)-laplacian of \(u\) at point \(x\). As \(s\) is fixed, we are not interested in multiplying the principal value integral by any renormalization factor. By \(L^0(\Omega)\), we shall denote the space of all (equivalence classes of) real valued measurable functions on \(\Omega\), endowed with the topology of the convergence in measure. We also set \[\label{power} \Phi(v) = |v|^{m-1}v\,,\qquad \text{for all \(v\in\mathbb R\).}\] For all \(1<q<2\) let us define \[\label{lambda1} \lambda_1(\Omega,q,s) = \inf_{u\in{\mathcal D}^{s,2}_0(\Omega)}\left\{ [u]_{s}^2 \mathbin{\colon} \int_\Omega |u|^q\,dx=1\right\}\.\] Given \(1<q<2\) and \(\alpha>0\), we consider the functional \[\label{lyap} \lyap{\varphi} = \frac12\int_{\mathbb R^N}\int_{\mathbb R^N} \inter{\varphi}\,dx\,dy-\frac{\alpha}{q} \int_\Omega |\varphi|^q\,dx\,,\] for all \(\varphi\in{\mathcal D}^{s,2}_0(\Omega)\). # Elliptic Toolkit {#sec:ell} The critical points of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\) are the weak solutions \(u\in {\mathcal D}^{s,2}_0(\Omega)\) of the Lane-Emden type equation \[\label{LEeq} (-\Delta)^su =\alpha |u|^{q-2}u\,,\qquad \text{in \(\Omega\),}\] which means \[\label{LEweak} \int_{\mathbb R^N}\int_{\mathbb R^N} \interr{u}{\psi}\, dx\,dy =\alpha\int_\Omega |u|^{q-2}u\psi\,dx\,,\qquad \text{for all \(\psi \in{\mathcal D}^{s,2}_0(\Omega)\).}\] ## The ground state level We collect some properties of the minimal energy level, defined as \[\label{minL} \Lambda_1 = \inf_{\varphi\in{\mathcal D}^{s,2}_0(\Omega)} \lyap{\varphi}\.\] ## Higher energies We then prove some basic properties of higher energy levels. By Lemma [\[lm:coerc\]](#lm:coerc){reference-type="ref" reference="lm:coerc"}, Eq. [\[PS1\]](#PS1){reference-type="eqref" reference="PS1"} implies that \((u_n)_{n\in\mathbb N}\) is bounded in \({\mathcal D}^{s,2}_0(\Omega)\). Thus, thanks to the compactness of the embedding into \(L^q(\Omega)\), a subsequence of \((u_n)_{n\in\mathbb N}\) (that we do not relabel) converges to some limit \(u\) weakly in \({\mathcal D}^{s,2}_0(\Omega)\) and strongly in \(L^q(\Omega)\). By passing to the limit in [\[PS2\]](#PS2){reference-type="eqref" reference="PS2"} we see[^1] that \(u\) is a weak solution of [\[LEeq\]](#LEeq){reference-type="eqref" reference="LEeq"}. Then, we can choose \(\psi=u\) in [\[LEweak\]](#LEweak){reference-type="eqref" reference="LEweak"}, which gives \[\label{nehari} \lyap{u} = \left(\frac{1}{2}-\frac{1}{q}\right) [u]_{s}^2 = \alpha\left(\frac{1}{2}-\frac{1}{q}\right) \int_\Omega|u|^{q}\,dx\.\] Then, \[[u]_{s}^2 = \alpha\int_\Omega |u|^q\,dx =\lim_{n\to\infty}\alpha \int_\Omega |u_n|^q\,dx\ge \limsup_{n\to\infty}\left([u_n]_{s}^2-\tfrac1n[u_n]_{s}\right) =\limsup_{n\to\infty}\,[u_n]_{s}^2\.\] Here, we used [\[nehari\]](#nehari){reference-type="eqref" reference="nehari"} for the first equality, the convergence in \(L^q(\Omega)\) for the second one, Eq. [\[PS2\]](#PS2){reference-type="eqref" reference="PS2"} for the inequality, and Lemma [\[lm:coerc\]](#lm:coerc){reference-type="ref" reference="lm:coerc"} together with [\[PS1\]](#PS1){reference-type="eqref" reference="PS1"} for the last equality. Hence, by the sequential weak lower semicontinuity of \([\,{\cdot}\,]_s^2\) the convergence of the sequence is also strong in \({\mathcal D}^{s,2}_0(\Omega)\). We have proved statement \((i)\) and we consider now \((ii)\), which means \[\inf\left\{ \lyap{\varphi} \mathbin{\colon} \min\big\{\|\varphi-w\|_{{\mathcal D}^{s,2}_0(\Omega)}\mathbin{,}\|\varphi+w\|_{{\mathcal D}^{s,2}_0(\Omega)}\big\}\ge \|w\|_{{\mathcal D}^{s,2}_0(\Omega)} \right\}>\Lambda_1\,,\] where \(w\) is the positive function with energy \(\Lambda_1\). The contrapositive statement is that we have \(\lyap{\varphi_j}\to \Lambda_1\) along a sequence for which is at a distance both to \(w\) and to \(-w\) larger than the half of that between \(w\) and \(-w\), in contradiction with the strong convergence in \({\mathcal D}^{s,2}_0(\Omega)\) either to \(w\) or to \(-w\), that follows by coercivity (see Lemma [\[lm:coerc\]](#lm:coerc){reference-type="ref" reference="lm:coerc"}). So, \((ii)\) is true and we are left to prove \((iii)\). To do so, we observe that if \(u\in{\mathcal D}^{s,2}_0(\Omega)\) is a critical point of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\) then choosing \(\psi=u\) in [\[LEweak\]](#LEweak){reference-type="eqref" reference="LEweak"} yields [\[nehari\]](#nehari){reference-type="eqref" reference="nehari"}. Thus, all energy levels belong to a bounded set, contained in \(]\Lambda_1,0]\). To prove that that their collection is closed, we write [\[nehari\]](#nehari){reference-type="eqref" reference="nehari"} with \(u\) replaced by \(u_n\), an arbitrary sequence of critical points with energy accumulating to a limit value \(\Lambda\). Then, by the compactness of the embedding of \({\mathcal D}^{s,2}_0(\Omega)\) into \(L^q(\Omega)\), the limit \(u\) of the sequence in \({\mathcal D}^{s,2}_0(\Omega)\) satisfies Eq. [\[LEeq\]](#LEeq){reference-type="eqref" reference="LEeq"}, and so \(\lyap{u}=\Lambda\) by construction. ◻ ## First excited level and spectral gap. We set \[\label{Lambda2} \Lambda_2=\Lambda_2(s) = \inf\left\{\Lambda>\Lambda_1 \mathbin{\colon} \Lambda \text{ is a critical value of } {{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\right\}\] We point out that the set in the right hand side is never empty because it always contains \(0\), which is the critical value associated with the critical point \(u\equiv0\). Also, its infimum is in fact a minimum because the critical levels form a closed set, by Lemma [\[basic-ell-2\]](#basic-ell-2){reference-type="ref" reference="basic-ell-2"}. We first notice that if a gap exists between \(\Lambda_1\) and \(\Lambda_2\) then it gives room to energy levels. Now we recall a general consequence of the uniqueness of (positive) energy minimizing functions: if there is a spectral gap, then the mountain pass does not collapse on the global minimum. We end this section by recalling an important consequence of the stability of energy minimizing solution of the elliptic problem, which in turn implies the fundamental gap, proved in  for the non-local problem following the method used in  for the local one. The following statement differs little from the original one in , which simply states the isolation with respect to the \(L^1(\Omega)\) topology instead of the topology of the convergence in measure. Arguing similarly as in one can observe that if \(\left\{\varphi_{n}\right\}\) is minimizing sequence of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\), then it converges, up to subsequence, either to the positive minimizer \(w\) or to \(-w\). We have then the following direct consequence of Theorem [\[isolated\]](#isolated){reference-type="ref" reference="isolated"}, namely the fundamental gap between the first and the second critical energy level of the functional \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\). The proof is similar to the relevant one in. # The (rescaled) parabolic problem {#sec:4} Given a solution \(u\) of [\[FPMEintro\]](#FPMEintro){reference-type="eqref" reference="FPMEintro"}, the rescaled function \(v(x,t)=e^{\alpha t} u(x,e^t-1)\) solves the following Dirichlet initial boundary value problem \[\label{FPMEs} \left\{\begin{array}{rcll} \partial_{t}v&=&-(-\Delta)^{s} \Phi(v)+\alpha v, & \mbox{ in } Q:=\Omega\times (0,+\infty),\\ v&=&0, & \mbox{ in } \mathbb R^{N}\setminus\Omega\times(0,+\infty),\\ v(\cdot,0)&=&u_0,& \mbox{ in } \Omega. \end{array} \right.\] We recall here the definition of weak solutions of the rescaled fractional porous media equation problem to [\[FPMEintro\]](#FPMEintro){reference-type="eqref" reference="FPMEintro"}. ## Well-posedness and entropy-entropy dissipation inequality The inequality proved in the following theorem is a crucial ingredient for the proof of our main result. By Fréchet-Kolmogorov theorem, [\[preaa2\]](#preaa2){reference-type="eqref" reference="preaa2"} implies that \[\label{preaa3} \text{\( \{ \widehat G_h(\cdot,t)\mathbin{\colon} h>0\}\) is relatively compact in \(L^2(\Omega)\), for all \(0\le t\le T\).}\] Thanks to the vector-valued extension of Ascoli-Arzelà theorem , from [\[preaa1\]](#preaa1){reference-type="eqref" reference="preaa1"} and [\[preaa3\]](#preaa3){reference-type="eqref" reference="preaa3"} we can infer [\[preaa\]](#preaa){reference-type="eqref" reference="preaa"}. By lower semicontinuity, [\[ee2\]](#ee2){reference-type="eqref" reference="ee2"} and [\[ee3\]](#ee3){reference-type="eqref" reference="ee3"} imply [\[EED\]](#EED){reference-type="eqref" reference="EED"}. Since \(\widehat G_{h_j}(0)=g(u_0)\) for all \(j\), by [\[postaa1\]](#postaa1){reference-type="eqref" reference="postaa1"} we also have that \(v(0)=u_0\). Then, in order to conclude we are left to prove that \(v\) is a weak solution of [\[FPMEs\]](#FPMEs){reference-type="eqref" reference="FPMEs"} in \(Q_T\). To see this, we first deduce [\[asspt:enweak\]](#asspt:enweak){reference-type="eqref" reference="asspt:enweak"} from [\[postaa1\]](#postaa1){reference-type="eqref" reference="postaa1"}, thanks to . To prove that [\[FMPEsW\]](#FMPEsW){reference-type="eqref" reference="FMPEsW"} holds too, we use the Euler-Lagrange equation [\[Eulerlagr\]](#Eulerlagr){reference-type="eqref" reference="Eulerlagr"} for \(v_k\) and [\[defubarh\]](#defubarh){reference-type="eqref" reference="defubarh"} to get \[\begin{split} \iint_{Q_T} & \frac{\bar v_{h_j}(x,t)-\bar v_{h_j}(x,t-h_j)}{h_j}\ \eta(x,t)\,dx\,dt \\ & + \int_0^T\int_{\mathbb R^N}\int_{\mathbb R^N} \frac{\Phi(\bar v_{h_j}(x,t))-\Phi(\bar v_{h_j}(y,t))}{|x-y|^{N+2s}}(\eta(x,t)-\eta(y,t))\,dx\,dy\,dt=\alpha\iint_{Q_T} \bar v_{h_j} \eta\,, \end{split}\] for all \(\eta\in C^\infty_c(Q_{T})\), and changing variables yields \[-\iint_{Q_T} \bar v_{h_j}\ \partial_t\widehat\eta^{h_j} +\int_0^T\!\!\iint_{\mathbb R^{2N}}\!\! \frac{\Phi(\bar v_{h_j}(x,t))-\Phi(\bar v_{h_j}(y,t))}{|x-y|^{N+2s}}(\eta(x,t)-\eta(y,t))\,dx\,dy\,dt=\alpha\iint_{Q_T} \bar v_{h_j} \eta\.\] Thanks to [\[ee3\]](#ee3){reference-type="eqref" reference="ee3"}, by passing to the limit in the latter we obtain Eq. [\[FMPEsW\]](#FMPEsW){reference-type="eqref" reference="FMPEsW"} and we conclude. ◻ The structure of \(\omega(u_0)\) is easier to understand under the assumptions \[\label{hp-omega} \partial_tg(v)\in L^2([T_0,+\infty),L^2(\Omega))\,,\qquad \Phi(v)\in L^\infty( [T_0,+\infty),{\mathcal D}^{s,2}_0(\Omega))\] on the weak solution \(v\) of [\[FPMEs\]](#FPMEs){reference-type="eqref" reference="FPMEs"} with initial datum \(u_0\), for an appropriate time \(T_0>0\). These assumptions are the non-local counterpart of those considered in  for the local case. # Paths of controlled energy # Proofs of the main results ## Proof of Theorem [\[mainthm1\]](#mainthm1){reference-type="ref" reference="mainthm1"} Weak solutions can be defined for [\[FPMEintro\]](#FPMEintro){reference-type="eqref" reference="FPMEintro"} similarly as done in Definition [\[energyweak\]](#energyweak){reference-type="ref" reference="energyweak"} for the rescaled problem [\[FPMEs\]](#FPMEs){reference-type="eqref" reference="FPMEs"}. By setting \[v(x,t)=e^{\alpha t}u(x,e^t-1)\] the desired conclusion becomes that the weak solution \(v(\cdot,t)\) of [\[FPMEs\]](#FPMEs){reference-type="eqref" reference="FPMEs"} with \(v(0)=u_0\) converges, as \(t\to+\infty\), either to \(\Phi^{-1}(w)\) or to \(-\Phi^{-1}(w)\) in \(L^{m+1}(\Omega)\), where \(w\) is the positive minimiser of the functional \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\) defined by [\[lyap\]](#lyap){reference-type="eqref" reference="lyap"}, that is unique by Lemma [\[lm:basic-elliptic\]](#lm:basic-elliptic){reference-type="ref" reference="lm:basic-elliptic"}. By Theorem [\[exist-ima\]](#exist-ima){reference-type="ref" reference="exist-ima"}, \(v\) is uniquely determined and the estimate [\[EED\]](#EED){reference-type="eqref" reference="EED"} holds. Therefore, by the compactness of the embedding \({\mathcal D}^{s,2}_0(\Omega)\) into \(L^{q}(\Omega)\), it follows that the orbit \(\{v(\cdot,t)\colon t>0\}\) is precompact in \(L^{m+1}(\Omega)\). Then, the omega-limit \(\omega(u_0)\) is connected, and so in order to get the desired conclusion it suffices to prove that \[\label{fin} \omega(u_0)\subseteq\{\Phi^{-1}(w),-\Phi^{-1}(w)\}\.\] Then, we take \(U\in\omega(u_0)\). From [\[EED\]](#EED){reference-type="eqref" reference="EED"} we can infer [\[hp-omega\]](#hp-omega){reference-type="eqref" reference="hp-omega"} and so, by Theorem [\[omega-lim-char\]](#omega-lim-char){reference-type="ref" reference="omega-lim-char"}, \(\Phi(U)\) is a critical point of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\). Therefore, it is enough to make sure that \[\label{fineq} \lyap{\Phi(U)}<\Lambda_{2}\] because by Corollary [\[cor:FG\]](#cor:FG){reference-type="ref" reference="cor:FG"} that entails that \(\Phi(U)\) is either \(w\) or \(-w\), which in turn gives [\[fin\]](#fin){reference-type="eqref" reference="fin"}. We are left to prove [\[fineq\]](#fineq){reference-type="eqref" reference="fineq"}. To do so, we take a sequence \(t_j\nearrow+\infty\) with \(v(\cdot,t_j)\to U\) in \(L^{m+1}(\Omega)\). We raise the Lipschitz estimate \(|\Phi(b)-\Phi(a)|\le m (|a|\vee|b|)^{m-1} |b-a|\), with \(a=U(x)\) and \(b=v(x,t_j)\), to the power \(q=\frac{m+1}{m}\), we integrate the result over \(\Omega\), and hence we arrive at \[\begin{split} \limsup_{j\to\infty} \|\Phi(v(\cdot,t_j)) &-\Phi(U)\|_{L^q(\Omega)} \le m\Big[\|\Phi(U)\|_{L^q(\Omega)}+\sup_{t>0} \|v(\cdot,t)\|_{L^{m+1}(\Omega)}\Big] \lim_{j\to\infty}\| v(\cdot,t_j)-U\|_{L^{m+1}(\Omega)}\,, \end{split}\] where we also used Hölder inequality with exponents \(m/(m-1)\) and \(m\). Hence, \(\Phi(v(\cdot,t_j))\) converges to \(\Phi(U)\) in measure. Then, by Fatou's Lemma, \[\lyap{\Phi(U)}\le\liminf_{j\to\infty} \lyap{\Phi(v(\cdot,t_j))}\.\] On the other hand, by Theorem [\[exist-ima\]](#exist-ima){reference-type="ref" reference="exist-ima"}, \[\lyap{\Phi(v(\cdot,t))} \le \lyap{\Phi(u_0)} \,,\qquad \text{for all \(t>0\).}\] By assumption, we have \[\lyap{\Phi(u_0)} <\Lambda_{2}\] and [\[fineq\]](#fineq){reference-type="eqref" reference="fineq"} follows by pairing this strict inequality with the previous two ones. 0◻ ## Proof of Proposition [\[prop:select\]](#prop:select){reference-type="ref" reference="prop:select"} Assume that one of the two conditions in [\[assumnonloc\]](#assumnonloc){reference-type="eqref" reference="assumnonloc"} holds and set \(q=(m+1)/m\). Then, by Proposition [\[prop:encontr\]](#prop:encontr){reference-type="ref" reference="prop:encontr"} there exists \(\theta\in C([0,1];{\mathcal D}^{s,2}_0(\Omega))\) such that \(\theta(\cdot, 0)\) equals the positive minimizer \(w\) of \({{\mbox{\script F}\,\,}_{\!\!\!\!q,\alpha}^{\!s}}\), moreover \(\theta(\cdot,1)=\Phi(u_0)\) and \[\label{propcntrl1} \lyap{\Phi(\theta(\cdot,t))}<\Lambda_2 \,, \qquad\text{for all \(t\in[0,1]\).}\] Now, we set \[z(\cdot,t)= \begin{cases} \theta(\cdot,t)\,, & \qquad \text{if \(0\le t\le1\)}\\ \Phi(v(\cdot,t-1))\,,& \qquad \text{if \(t>1\),} \end{cases}\] where \(v\) is the unique solution of [\[FPMEs\]](#FPMEs){reference-type="eqref" reference="FPMEs"} with \(v(0)=u_0\). Then, in view of Theorem [\[exist-ima\]](#exist-ima){reference-type="ref" reference="exist-ima"} and of Proposition [\[prop:encontr\]](#prop:encontr){reference-type="ref" reference="prop:encontr"}, we have \[\label{propctrl2} \lyap{z(\cdot,t)} <\Lambda_{2}\] Hence, by coercivity (see Lemma [\[lm:coerc\]](#lm:coerc){reference-type="ref" reference="lm:coerc"}) we deduce that the trajectory \(z(\cdot,t)\) is contained in a bounded subset of \({\mathcal D}^{s,2}_0(\Omega)\). Since \(t\mapsto z(\cdot,t)\) is continuous from \([0,1]\) to \({\mathcal D}^{s,2}_0(\Omega)\), by the compactess of the embedding into \(L^q(\Omega)\) it is continuous as a function with values in \(L^q(\Omega)\), as well; also, it is easily seen that the continuity of \(t\mapsto v(\cdot,t-1)\) from \([1,+\infty)\) to \(L^{m+1}(\Omega)\) implies that of \(t\mapsto z(\cdot,t)= \Phi(v(\cdot,t-1))\) from \([1,+\infty)\) to \(L^q(\Omega)\). Therefore, \(z\) belongs both to \(L^\infty((0,+\infty);{\mathcal D}^{s,2}_0(\Omega))\) and to \(C([0,+\infty);L^0(\Omega))\). Now, we argue by contradiction, and we assume \(v(\cdot,t)\) to converge in measure to \(-w\). Then, so does \(z(\cdot,t)\) and it follows that \(z\) is then eligible for the mountain pass formula of Proposition [\[prop:mszeta\]](#prop:mszeta){reference-type="ref" reference="prop:mszeta"}. Thus, \[\Lambda^\ast=\inf_{z\in\mathfrak{Z}}\sup_{t\in[0,+\infty)} \lyap{z(\cdot,t)}<\Lambda_2\.\] in contradiction with Proposition [\[prop:MPexcited\]](#prop:MPexcited){reference-type="ref" reference="prop:MPexcited"}.0◻ [^1]: Here, in particular, we are using that \[\lim_{n\to\infty}\int_\Omega |u_n|^{q-2}u_n\psi\,dx=\int_\Omega |u|^{q-2}u\psi\,dx\.\] This can be deduced from the fact that \(\||u_n|^{q-2}u_n-|u|^{q-2}u\|_{L^{q/(q-1)}(\Omega)}\to0\), as \(n\to\infty\), and in turn this latter assertion holds because of the convergence to \(u\) of the sequence \((u_n)_n\) in \(L^q(\Omega)\). To see this, for \(q<2\) one can use the Hölder continuity of the function \(\tau\mapsto|\tau|^{q-2}\tau\).

{'timestamp': '2023-02-09T02:18:05', 'yymm': '2302', 'arxiv_id': '2302.04266', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04266'}

# Introduction One of the major thermodynamic costs of many complex systems arises from communication among their separate subsystems. Examples include cellular sensing systems , the human brain , ecosystems , wireless sensor networks , hardware implementations of machine learning algorithms , and digital computers . Similarly, reducing the thermodynamic costs of communication between processing units and memory units of conventional Von Neumann computational architectures is one of the primary motivations of the field of neuromorphic computing . As emphasized in the latest "Physics of Life" report from the National Academy of Sciences, it is crucial to understand the *common* physical principles underlying communication in biological systems, whether that be a set of interacting bacteria, a colony of insects, a flock of birds, or a human social group . In order to uncover and investigate such principles common to all types of communication systems, not just biological ones, we need a minimal model grounded in the features shared by all of those systems. When constructing such a minimal model, it is important to appreciate that the information transmission in these different communication systems are subject to different physical constraints (e.g., diffusive, electronic, acoustic, etc.--see [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(a-c)). In turn, each of these different constraints impose their own, system-specific thermodynamic costs, since they each limit what theoretical efficiencies the system can exploit . However, one feature common to all of these systems is that they involve an "output" component that is biased to change its state to equal the state of a separate "input" component, which is set exogenously ([\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(d)). A minimal model that can apply to diverse types of communication systems should include this shared aspect of copying between two separate components. Another feature common to many biological and artificial communication systems is that they split a stream of information over multiple channels, i.e., they inverse multiplex. Often this occurs even when a single one of the channels could handle the entire communication load. For example, multiple synapses tend to connect adjacent neurons  and multiple neuronal pathways tend to connect brain regions . In engineering, spatial multiplexing techniques have made multiple-input-multiple-output (MIMO) technology the gold standard for modern wireless communication systems . Therefore, analyzing a minimal model of communication should also result in thermodynamic benefits to inverse multiplexing in many scenarios. Our goal is to construct a minimal model with both of these features. Until now, it has been difficult to construct such a minimal model using a physics-based formalism. The difficulty is that communication systems operate very far from thermodynamic equilibrium, which rules out conventional arguments based on the quasi-static limit  or the linear-response regime . However, the development of stochastic thermodynamics  in the past two decades has supplied a theoretical framework that relates the dynamics of information in a system to the free energy it dissipates in a process arbitrarily far from equilibrium. Stochastic thermodynamics can allow us, therefore, to investigate how this thermodynamic cost varies with communication rate in a two-component copying system. Here we start such an investigation. We begin by providing a background on communication theory and stochastic thermodynamics. We then build on this background to motivate a minimal model of communication. Next, we present a preliminary analysis of the thermodynamics of two types of communication encompassed by our model. These two types correspond roughly to (i) wireless communication or inter-cellular communication, and (ii) wired electronic communication. We prove that in many scenarios of both types of communication, the thermodynamic cost is a convex function of the communication rate. Regardless of the convexity of this function, one might expect that increasing the speed of information transmission through a channel requires a monotonic increase in the thermodynamic costs . However, we find that in many cases this function is *not* monotonic. We then investigate the consequences of this result for when and how one should split a single information stream across multiple physical channels. In particular, we derive a Pareto front representing the trade-off between minimizing total thermodynamic cost and maximizing total information transmission rate shared across a fixed number of communication channels. This analysis may explain why inverse multiplexing arises so often in real-world systems. We end by discussing our model in the broader context of the thermodynamics of computation and by suggesting future work. # Background on Shannon information theory and stochastic thermodynamics {#background-on-shannon-information-theory-and-stochastic-thermodynamics .unnumbered} Claude Shannon's channel coding theorem states that one can pass messages through a noisy communication channel with vanishingly low error . A channel approaches this error-free limit if the messages sent through it are encoded with an appropriate codebook into strings of letters from a finite, discrete alphabet. These encoded strings can be decoded to recover the original message at the output of the channel. Importantly, one can implement error-free communication in noisy channels in this way only up to a finite maximum baud rate, called the channel capacity. Exceeding that rate imposes a non-zero probability of error. A channel's capacity therefore serves as the primary measure of its communication capabilities. This capacity equals the maximal mutual information between input and output attainable by varying the input alphabet distribution . The channel coding theorem further states that one can achieve the channel capacity with negligible error using any one of *many* possible optimal encodings of information at the source[^1]. These encodings produce sequences that look as if they were identically and independently drawn (IID) from the input distribution that maximizes the mutual information between input and output . This theorem led to massive engineering efforts toward developing error-correcting codes that achieve this bound . Such coding strategies are now widespread in digital communication systems . Moreoever, the study of codes that approach the Shannon bound has been one of the major areas of research in modern information theory for several decades. The duality between information and Shannon entropy creates a tight link between information theory and thermodynamics . Stochastic thermodynamics has revealed that the entropy produced over any finite time interval in a single realization of a nonequilibrium process[^2] *can* be negative, in defiance of the Second Law. This entropy production (EP) recovers non-negativity only when averaged over all possible realizations (trajectories) [^3]. A primary focus of most studies in stochastic thermodynamics, the EP is the extra free energy expended compared to an idealized, infinitely slow process. Throughout the paper, we refer to the EP as the thermodynamic cost. This measure is crucial in many powerful results, including bounds on the precision of generalized currents  and the speed of changes in a system's probability distribution over states . The EP has also been applied to study the dynamics of information . Unfortunately, most of the prior work in stochastic thermodynamics involving mutual information and "information processing" fails to relate the EP to the channel capacity. In fact, most of these results do not consider a system dedicated to communication. Some studies analyze how components influence one another in multipartite systems of multiple feedback controllers by quantifying the "information flow" between subsystems . Other studies, e.g., on Bayes' Nets , use the "transfer entropy" rather than the channel capacity to characterize information transferred from inputs to outputs. Studies of specific models for biomolecular copying processes  and cellular sensing  have made analogies with information theory and the channel coding theorem, but are limited in scope to only biological systems. Similarly, the analysis of channel capacity in  applies only to the very restricted domain of electronic circuits comprising a set of transistors running sub-threshold. Until now, the literature has lacked an investigation of the quantitative relationship between the channel capacity and the EP that could apply to all types of communication with a finite set of discrete symbols. We conduct exactly this investigation, justifying the features of our minimal model with the rigor of Shannon information theory. # Stochastic thermodynamics of communication channels {#stochastic-thermodynamics-of-communication-channels .unnumbered} We analyze the thermodynamics of communication channels operating at their channel capacity. In Shannon's model of communication, a process external to the system sets the input state so that its dynamics reflect samplings from the input distribution that achieves the channel capacity. This external system effectively acts as a work reservoir, and could follow arbitrary (potentially non-Markovian) dynamics. The input could be set deterministically or stochastically, periodically according to a clock, according to a continuous-time Markov chain (CTMC), etc. This feature is physically motivated by the fact that in many real communication channels, e.g., radio transmission, the input is set by an exogenous process. Therefore, we ignore the thermodynamics of the external system that sets the input. Finally, the coupling between input and output is non-reciprocal, in that the output dynamics depends on the state of the input but not vice-versa. This common assumption is sometimes called "no back-action" . We consider a communication channel in which the input \(A\) sends letters \(x_A\) from the alphabet \(X_A = \{1, 2, \ldots, L \}\) in any way that reflects the distribution \(\pi_{X_A}\). This distribution over input states maximizes the mutual information between input and output, i.e., achieves the channel capacity. The communication channel is defined by how the output \(B\) changes its state, \(x_B \in X_B = \{1, 2, \ldots, L \}\), in response to the input state. At any given time, the rates of state transitions in the output depend on the current state of the input. The energies of different output states may depend on the state of the input according to a Hamiltonian function \(H(X_B = x_B| X_A = x_A)\). To capture the stochasticity resulting from the noisiness of the channel, the dynamics can be modeled as a CTMC ([\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(e,f)). The joint distribution between the input and the output, \(\bm{p}_{X_A,X_B}\), is a vector where each element \(p_{x_A,x_B} = \pi_{x_A}p_{x_B|x_A}\). The joint distribution evolves via a master equation with an \(L^2 \times L^2\) rate matrix \({\overline{K}}^{X'_A,X'_B}_{X_A,X_B}\): where each element \(K^{x'_A,x'_B}_{x_A,x_B}\) of the rate matrix indicates the probability of observing a state transition \((x'_A,x'_B) \to (x_A,x_B)\) at any given time. The input and output together evolve as a bipartite process[^4], so only one subsystem changes state at any given time. Additionally, the output changes its state due to interactions with a set of \(N\) equilibrium reservoirs \({\mathcal{V}} := \{ v_1, v_2, \ldots, v_N \}\), at temperatures \(\{ T_1, T_2, \ldots, T_N \}\). We denote the temperature of reservoir \(v\) as \(T_v\). This means the overall system's rate matrix can be written as a sum over the rate matrices representing each reservoir's effect on the output's state transitions, plus the rate matrix for the input. So, each element of the rate matrix obeys which assures that only one of the subsystems changes state at any given moment in time (\(\delta^{x'}_{x}\) is the Kronecker delta function that equals \(1\) when \(x' = x\), and equals \(0\) otherwise). Each rate matrix specific to reservoir \(v\)'s effect on the output transition rates given the state of the input follows local-detailed balance: That allows each such rate matrix to decompose as where \(\textbf{R}^{X'_B,X_A}_{X_B,X_A}(v)\) is a symmetric matrix with entries \(R^{x'_B,x_A}_{x_B,x_A}(v) = R^{x_B,x_A}_{x'_B,x_A}(v)\). These matrices represent the couplings of the output to the set of reservoirs \({\mathcal{V}}\). \(\bm{\Pi}^{X'_B,X_A}_{X_B,X_A}(v)\) is a diagonal matrix with values along the diagonal equal to \(\pi^*_{x_B, x_A}(v)\), which is the steady-state probability that the output is in state \(x_B\) and the input is in state \(x_A\) *in the counterfactual case* that the output was coupled only to reservoir \(v\). This probability distribution equals the Boltzmann distribution, so \(\pi^*_{x_B, x_A}(v) = \pi_{x_A} e^{-H(x_B | x_A)/T_v}/Z\), where \(Z = \sum_{x_B} e^{-H(x_B | x_A)/T_v}\). Subtracting the column sum (\(\text{colsum}\)) of the product \(\textbf{R}\bm{\Pi}\) ensures that the resulting matrix is normalized. Note that although every reservoir-specific rate matrix is detail-balanced, the overall rate matrix is not detail-balanced in general unless there is only a single reservoir. Within this minimal model, we analyze two general methods that can modulate the output dynamics in real communication systems. In the "energy switching" case, different values of the input modulate the energies of different output values. Intercellular chemical communication in biological organisms can be modeled in this way. For example in cellular sensing, the concentration of a ligand (input) modulates the free energy of receptor binding (output) [^5]. Additionally, we can model wireless communication in terms of energy switching, where electromagnetic waves sent by the transmitter modify the potential energy function of electrons at the receiver. In the "reservoir switching" case, different values of the input modulate the coupling of the output to different reservoirs. We can model wired electronic communication in circuits in this way, where the input voltage modulates the output wire's couplings with different chemical reservoirs of electrons . ## Energy switching {#energy-switching .unnumbered} In the energy switching method of communication, the energy levels of different output states depend on the input state. We analyze the class of channels for which the output achieves its lowest energy state when it matches the current state of the input, and all other "mismatched states" have the same higher value of energy. Such a Hamiltonian function can be written Since the Hamiltonian is symmetric with respect to every letter in the input alphabet, the channel capacity-achieving input distribution \(\pi_{X_A}\) is the uniform distribution (\(\pi_{x_A} = 1/L\)). Additionally, we set all non-zero entries of every \(\textbf{R}(v) = r_v\), so that the couplings of the output to its different reservoirs cannot be modulated by the state of the input. Therefore, every off-diagonal rate matrix element equals In our simulations for the energy switching case, the input dynamics follow a telegraph process , which is consistent with a CTMC. A telegraph process serves as an accurate model for any real communication channel (e.g., synaptic release ) for which the waiting times for the input in any given state follows an exponential distribution. The input switches its state with rate \(f_s\), which is a positive real number. This means that each state transition in the input is equally likely, so that \(K^{x'_A,x_B}_{x_A,x_B} (A) = K^{x'_A}_{x_A} = f_s\). Therefore, the off-diagonal elements of the overall channel's rate matrix are We first analyze the situation where the channel is in a nonequilibrium steady-state (NESS), which means where \(\bm{0}_{L^2}\) is a vector with \(L^2\) entries, each of which equal zero. Solving for the joint steady-state distribution \(\bm{\pi}_{X_A,X_B}\), we obtain This means that the conditional steady-state distribution of the output state given the input state reads and the marginal steady-state distribution over just the output states is uniform. We find that the EP rate in the NESS is where the EP rate due only to the output (input) transitions, \(\lr{\dot{\sigma}_B}\) (\(\lr{\dot{\sigma}_A}\)), is equal to the top (bottom) line in each of [\[epr-gen\]](#epr-gen){reference-type="ref" reference="epr-gen"} and [\[epr-es\]](#epr-es){reference-type="ref" reference="epr-es"}. The bipartite nature of the overall system's evolution gives rise to this decomposition . As an example, if the output fluctuates due to coupling with a single reservoir, and if those fluctuations obey detailed balance, meaning that the output is in an equilibrium steady state, then \(\lr{\dot{\sigma}_B} = 0\). On the other hand, if the input is set at time \(t=0\) and never changes state thereafter, then \(\lr{\dot{\sigma}_A} = 0\). Whereas \(\lr{\dot{\sigma}_A}\) represents the cost of switching the input, \(\lr{\dot{\sigma}_B}\) represents the thermodynamic cost of the copying process. Since the input distribution is chosen so that the system runs at its channel capacity, the value of the channel capacity in the NESS is simply given by the mutual information between the input and the output: where the entropy of the output distribution is \(S^{X_B} = \sum_{x_B}-\pi_{x_B} \ln \pi_{x_B}\) and the entropy of the conditional distribution of the output given the input is \(S^{X_B|X_A} = \sum_{x_B, x_A}-\pi_{x_B|x_A}\pi_{x_A} \ln \pi_{x_B|x_A}\). [\[fig:results\]](#fig:results){reference-type="ref" reference="fig:results"} shows results of analyzing how the thermodynamic cost of copying co-varies with the channel capacity when we adjust the amount of noise in the channel. We restrict attention to the case when the system is out of equilibrium. There are many different processes that could keep a system away from equilibrium. Here, in order to minimize the number of modeling assumptions, we consider the case where the stationary state is out of equilibrium due to the presence of two thermal reservoirs at different temperatures (\(N = 2\)). To adjust the noise level in the channel, we sweep the temperature of one reservoir \(T_1\) while holding fixed the temperature of the other at \(T_2\). We find that \(\lr{\dot{\sigma}_B}\) as a function of \(C\) is convex with a single global minimum that occurs at a positive value of the channel capacity. In particular, this means that the thermodynamic cost is not a monotonic function of the channel capacity. We prove this convex, non-monotonic relationship in the limit of low signaling rate (Materials and Methods). Note that for energy switching, communication is possible even if the system is in contact with a single thermal reservoir. For a single reservoir at temperature \(T\), the system operates out of equilibrium only during the transient relaxation dynamics that follows a change in the input. We show that when the waiting times between successive inputs are long compared to the relaxation timescale of the system, three results hold for arbitrary \(L\) (Materials and Methods): i) the EP rate does increase monotonically with the channel capacity, ii) the EP rate is a convex function of the channel capacity for capacities larger than a threshold value which for large \(L\) is \(\log(L)/2\), i.e., half of the maximum capacity, and iii) the EP rate diverges as the channel capacity approaches its maximum value \(\ln L\), with \(d \dot\sigma/dC \propto 1/(\ln L-C)\) and \(d^2 \dot\sigma/dC^2 \propto 1/(\ln L-C)^2\). ## Reservoir switching {#reservoir-switching .unnumbered} The channel can alternatively modulate the coupling of the output system with its different reservoirs. For this reservoir switching case, the energies \(H(x_B | x_A) = H(x_B)\) of the output states do not depend on the input. So, the matrices \(\bm{\Pi}^{X'_B,X_A}_{X_B,X_A}(v)\) do not depend on the input state, but the components of the matrices \(\textbf{R}^{X'_B,X_A}_{X_B,X_A}(v) = r_v(x_A)\) in [\[R-param\]](#R-param){reference-type="ref" reference="R-param"} do. Importantly, this kind of communication protocol is used to transfer information between different components of modern electronic circuits. For example, consider a CMOS inverter ([\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(b)), which is the electronic implementation of a NOT gate. The output voltage \(v\) of the inverter connects to two chemical reservoirs of electrons at fixed voltages \(V_1\) and \(V_2\) (with \(V_1>V_2\)), through two complementary MOS transistors (a nMOS and a pMOS). Increasing \(v_{\text{in}}\) increases the conductivity of the nMOS transistor and decreases the conductivity of the pMOS transistor. So for high input voltage, the output's interaction with the reservoir at voltage \(V_2\) dominates, making \(v\simeq V_2\) at steady state. The situation reverses for low input voltages, resulting in \(v \simeq V_1\). The energy associated to an output voltage \(v\) stays fixed at \(E(v) = C_o v^2/2\) according to the output capacitance \(C_o\), which is independent of the input. To make analytical progress, we assume that the interactions with different reservoirs can be turned on and off perfectly. For the CMOS inverter, this equates the MOS transistors to ideal switches. Additionally we assume the output couples to only one reservoir for any given input. In the limit of no signaling rate, the output approaches equilibrium with the corresponding reservoir. In this case, the steady-state EP vanishes and conditional distributions \(\pi_{X_B|x_A}\) are Boltzmann distributions. We assume that the signaling rate is low enough that the waiting times between successive inputs is large compared to the relaxation time of the output. Then, the entropy produced after a change \(x'_{A} \to x_A\) in the input equates to where \(D\KLDx{p}{q}\) is the Kullback-Leibler divergence between distributions \(p\) and \(q\). Define \(N_{x'_{A} \to x_{A}}(t)\) as the number of state transitions \(x'_{A} \to x_{A}\) that occur during a time period of length \(t\). Set \(\lambda_{j \to k} = \lim_{t \to \infty} N_{j \to k}/t\). The average EP rate is Define \(\pi_{x_B} = \sum_{x_A} \pi_{x_B|x_A}\) as the marginal probability that the output state equals \(x_B\) in the steady-state. Recall that, given the conditional distributions \(\pi_{X_B|x_A}\) that characterize the channel, the optimal input distribution \(\pi_{X_A}\) forces the relative entropy \(D\KLDx{\pi_{X_B|x_A}}{\pi_{X_B}}\) to be independent of \(x_A\). So the channel capacity is simply \(C = D\KLDx{\pi_{X_B|x_A}}{\pi_{X_B}}\). If the output distribution is uniform[^6], then We find that in the weak noise regime, the EP is a convex function of the channel capacity, and that it diverges as the capacity approaches its maximum achievable value. In this regime, the conditional probabilities equal \[\pi(x_B|x_A) = \begin{cases} 1-\sum_{x_B \neq x_A } \alpha_{x_B,x_A} & \text{if } x_B = x_A \\ \alpha_{x_B, x_A} & \text{if } x_B \neq x_A, \end{cases}\] where \(0 < \alpha_{x_B, x_A} \ll 1/L\). That is, \(\alpha_{x_B, x_A}\) is the probability to have an incorrect output \(x_B\) given the input \(x_A\). This probability is small, but doesn't equal zero. If we scale the constants \(\alpha_{x_B, x_A}\) by a factor \(0< \eta <1\), we find that the channel capacity approaches its maximum value as \(\eta \to 0\) according to \(|\ln L-C| \propto \eta \ln(1/\eta)\), and the EP rate diverges as \(\lr{\dot \sigma} \propto \ln (1/\eta)\). # Thermodynamics informs when and how to inverse multiplex {#thermodynamics-informs-when-and-how-to-inverse-multiplex .unnumbered} We have found that in both types of communication analyzed above, the EP is a convex function of the channel capacity. So by the theory of convex optimization, a thermodynamic benefit to inverse multiplexing must arise. In what follows, we analyze when one can reduce thermodynamic cost by splitting information streams across multiple channels. We additionally analyze how to distribute information transmission rates across multiple channels in order to reach the minimum achievable thermodynamic cost. In doing so, we derive the Pareto front that represents the set of "optimal" tuples \((C, \lr{\dot{\sigma}})\) one can achieve for any given fixed number of channels. ## \(M\)-channel capacity {#m-channel-capacity .unnumbered} Consider concurrently operating \(M\) channels with capacities \(C_1, C_2, \ldots, C_M\); inputs \(A_1, \ldots, A_M\); and outputs \(B_1, \ldots, B_M\). We set the inputs simultaneously at the beginning of each channel use. We know that the mutual information between the inputs and outputs of \(M\) channels is upper bounded by a value that can only be achieved if the inputs are independent  (Materials and Methods). So the capacity of this multi-channel setup equals the sum of the capacities of each channel. The only joint input distribution that achieves this maximum channel capacity is \(\pi_{X_{A_1},\ldots,X_{A_M}} = \prod_{i=1}^M \pi_{X_{A_i}}\). So by the channel-coding theorem, optimal codebooks make it appear as if the inputs were generated in each channel from its capacity-maximizing input distribution, independently from one another. Additionally, since the inputs to all of the channels are statistically independent, there is no unavoidable mismatch cost, as arises for example in the parallel bit erasure of statistically coupled bits . ## \(M\)-channel EP rate and thermodynamic benefits of inverse multiplexing {#m-channel-ep-rate-and-thermodynamic-benefits-of-inverse-multiplexing .unnumbered} Since we model the multi-channel setup as if it is running at its capacity, each of the constituent channels is an independent CTMC, governed by its own master equation. Therefore, the total EP rate of the combined system is simply the sum of the EP rates of each channel : where \(g_i(C_i)\) expresses channel \(i\)'s EP rate \(\lr{\dot{\sigma}_i}\) as a convex function \(g_i\) of its channel capacity, \(C_i\). We can use the water-filling algorithm to minimize the total EP rate of a set of \(M\) channels subject to a desired total channel capacity \(C_d = \sum_{i=1}^M C_i\), via the Lagrangian:[^7] Differentiating with respect to \(C_j\), which means where \(\lambda\) is chosen to satisfy In particular, this suggests that if \(g_i = g \, \forall \, i\), then it is optimal to use \(M\) channels with identical capacity \(\frac{C_d}{M}\). This finding leads to important insights regarding how to choose between different inverse multiplexing setups. For example, a setup of two channels with capacities \(C_A\) and \(C_B\) is identical in an information-theoretic sense to a setup with two identical channels each with capacity \(\frac{1}{2}(C_A + C_B)\). However from a thermodynamic viewpoint, it would be better to use the two identical channels because that would minimize EP rate due to the convexity of the EP rate with respect to the channel capacity. If we know the functional form of \(g(C)\), we can also use the method outlined in  to calculate the energetically-optimal number of independent, identical channels to use in order to achieve a desired total information rate. More generally, for any set of functions \(\{g_i\}\), we can find the \(\lambda\) that satisfies [\[lagrangian_solve\]](#lagrangian_solve){reference-type="ref" reference="lagrangian_solve"} (in addition to the positivity of the channel capacity), and plug it into the Lagrangian. In this way we can obtain the optimal distribution \(\{C_i\}\) of information transmission rates across \(M\) channels that are not necessarily identical. ## Pareto-optimal fronts for inverse multiplexing {#pareto-optimal-fronts-for-inverse-multiplexing .unnumbered} These results suggest that for any number \(M\) of channels, there exists a Pareto front of points \((C, \lr{\dot{\sigma}})\) that each minimizes \(\lr{\dot{\sigma}} = \sum_{i=1}^M \lr{\dot{\sigma}_i}\) and simultaneously maximizes \(C = \sum_{i=1}^M C_i\). We plot examples of these Pareto fronts in [\[fig:pareto\]](#fig:pareto){reference-type="ref" reference="fig:pareto"}. The plots reveal that higher channel capacities benefit from splitting information rates across more channels. More specifically, there exist thresholds, \(C^{(m)}\), above which it reduces the total EP to split information transmission across \(m+1\) channels instead of \(m\) channels. # Discussion {#discussion .unnumbered} A convex relationship between thermodynamic cost and channel capacity, as well as the energetic benefits of inverse multiplexing that follow, has previously been demonstrated in communication channel models with Gaussian noise . However, for those models, the "power" does not correspond to actual dissipated energy per unit time. This relationship has also been justified semi-formally for biological systems . Due to their assumptions, both of these kinds of studies derive that thermodynamic cost increases monotonically with channel capacity. To the best of our knowledge, ours is the first study to use physics to demonstrate that, in many cases, the thermodynamic cost is actually *not* necessarily a monotonic function of the channel capacity. We find that this relationship is in most cases convex even if non-monotonic, so the thermodynamic benefits of splitting an information stream among multiple channels still arises. This paper also presents the first analysis of inverse multiplexing treated with the rigor provided by stochastic thermodynamics. Our results help illuminate observations in biological communication, and may provide heuristics for an engineer designing communication systems. There are several stochastic thermodynamics formulations one can use to continue to investigate the thermodynamic costs of communication. In the main text we use the one in which the dynamics of the system of interest (SOI) is Markovian, due to its coupling with infinite reservoirs that are at thermal equilibrium. Another formulation, called the "inclusive Hamiltonian\", or "strong coupling" model, treats the case where the reservoirs are finite. There, the joint system comprising the SOI and the reservoirs evolves with unitary Hamiltonian dynamics. So the dynamics of the SOI is non-Markovian in general. There has been a very preliminary analysis of thermodynamically-motivated rate distortion functions for communication channels using this formulation . In the main text we have also assumed that the output has "no back-action" onto the input, so they each obey LDB separately. This non-reciprocity approximates most real-world observable dynamics well . However, strictly speaking, many systems violate this approximation when modeled at finer scales due to micro-reversibility and the consequent property that the *entire system* obeys LDB . While they are less efficient than channels with no back-action, systems that obey LDB explicitly and exactly can implement communication in principle. The mathematics of the simplest such systems is analytically tractable. We present one such example in the Supplementary Information. For the purposes of these investigations, we have not considered the thermodynamic cost of the input encoding function and the output decoding function. However since different coding strategies ("codebooks") would likely incur different thermodynamic costs, we suggest that future work should investigate their stochastic thermodynamics. Such analyses would naturally extend to investigations of the thermodynamic costs of different error-correcting codes. **Acknowledgements** This work was supported by the MIT Media Lab Consortium, Santa Fe Institute, US NSF Grant CHE-1648973, and FQXi Grant FQXi-RFP-IPW-1912. [^1]: Although there are countably infinitely many such encodings, it is very difficult for a human to find one . [^2]: In particular, stochastic thermodynamics describes any nonequilibrium system whose state transitions are mediated by energy exchanges with infinite reservoirs. [^3]: Moreoever, as the number of degrees of freedom in the model of the system increases, the probability that the entropy production is negative decreases . This explains the observation that the inviolability of the Second Law at the macroscopic scale is matched by a complete absence of the Second Law at the microscopic scale . [^4]: A bipartite process is a special case of a multipartite process (MPP)  involving only two subsystems. [^5]: The fraction of bound receptors on the cellular membrane reflects the external ligand concentration. [^6]: Note that consistency with the previous assumptions demands the conditional entropy \(S^{X_B|x_A}\) to hold the same value for all choices of \(x_A\). [^7]: This is a convex optimization problem that loosely speaking is the inverse of the one discussed in Section 9.4 of , which maximizes total channel capacity subject to a "power constraint".

{'timestamp': '2023-02-27T02:05:14', 'yymm': '2302', 'arxiv_id': '2302.04320', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04320'}

# Introduction This paper presents a robust MPC methodology for the trajectory optimisation of VTOL aircraft. Although we consider here the problem of tilt-wing aircraft transition, the method described is equally applicable to tilt-rotors and other forms of VTOL aircraft. One of the main challenges associated with VTOL aircraft is stability and control during transition between powered lift and wing-borne flight. This can be problematic as the aircraft experiences large changes in the effective angle of attack during such manoeuvres. Achieving successful transitions requires robust flight control laws along feasible trajectories. The computation of the flight transition trajectory is a difficult NonLinear Program (NLP) as it involves nonlinear flight dynamics. Several attempts were proposed to solve this problem For example, in, the trajectory optimisation for take-off is formulated as a constrained optimisation problem solved using NASA's OpenMDAO framework and the SNOPT gradient-based optimiser. The problem of determining minimum energy speed profiles for the forward transition manoeuvre of the Airbus A\(^3\) Vahana was addressed in, considering various phases of flight (cruise, transition, descent). Forward and backward optimal transition manoeuvres at constant altitude are computed in for a tiltwing aircraft, considering leading-edge fluid injection active flow control and the use of a high-drag device. The main drawback of these approaches is the computational burden associated with solving a NLP, which makes them unsuitable for real-time implementation. Another strategy to compute the transition relies on linearisation and convex optimisation resulting in approximate but computationally tractable algorithms. In, the transition for a tiltwing VTOL aircraft was computed using convex optimisation by introducing a small angle approximation. This provides a computationally efficient optimisation that could potentially be leveraged online, e.g. for collision avoidance or MPC. The obvious limitation of the approach is the assumption of small angles of attack, which restricts considerably the type of achievable manoeuvres. While the method of introduces a linearisation of the dynamics, there is no consideration for the effect of linearisation error on the dynamics. In this work, we propose a solution to this problem based on a DC decomposition of the nonlinear dynamics. This allows us to obtain tight bounds on the linearisation error and treat this error as a disturbance in a robust optimisation framework, exploiting an idea from tube MPC. The main idea is to successively linearise the dynamics around predicted trajectories and treat the linearisation error as a bounded disturbance. Due to the DC form of the dynamics, the linearised functions are convex, and so are their linearisation errors. These errors can thus be bounded tightly since they take their maximum at the boundary of the domain, and the trajectories of model states can be bounded by a set of convex inequalities (or tubes ). These inequalities form the basis of a computationally-tractable convex tube-based optimisation for the trajectory generation of VTOL aircraft. The contribution of this research is twofold: i) we solve an open problem in trajectory optimisation of VTOL aircraft by allowing aggressive transitions at high angle of attack while guaranteeing safety and computational tractability of the scheme; ii) we make a connection between DC decomposition and robust tube based optimisation and demonstrate the applicability and generalisability of the procedure in. This paper is organised as follows. We start by developing a mathematical model of a tiltwing VTOL aircraft in Section [2](#sec:modeling){reference-type="ref" reference="sec:modeling"}. In Section [3](#sub:convex_formulation){reference-type="ref" reference="sub:convex_formulation"}, we formulate the trajectory optimisation problem and discuss a series of simplifications to obtain a convex program, leveraging ideas from DC decomposition and robust tube MPC. Section [4](#sec:results){reference-type="ref" reference="sec:results"} discusses simulation results obtained for a case study based on the Airbus A\(^3\) Vahana. # Modeling {#sec:modeling} Consider a longitudinal point-mass model of a tiltwing VTOL aircraft equipped with propellers as shown in Figure [\[fig:diagram\]](#fig:diagram){reference-type="ref" reference="fig:diagram"} and subject to a wind gust disturbance. The Equations Of Motion (EOM) are given in polar form by \[\begin{aligned} {2} m \dot{V} &= T \cos{\alpha}-D-mg \sin{\gamma}, &\quad V(t_0) &= V_{0}, \label{eq:eom1} \\ m V \dot{\gamma} &= T \sin{\alpha} + L-mg \cos{\gamma}, &\quad \gamma(t_0) &= \gamma_{0}, \label{eq:eom2} \end{aligned}\] \[J_w \ddot{i}_w = M, \quad i_w(t_0) = i_{0}, \quad \dot{i}_w(t_0) = \Omega_{0}, \label{eq:tiltwing_dyna}\] \[\dot{x} = V \cos{\gamma}, \qquad \dot{z} =-V \sin{\gamma},\] where the control inputs are the thrust magnitude \(T\) and the total torque \(M\) delivered by the tilting actuators, and the model states are the aircraft velocity magnitude \(V\), the flight path angle \(\gamma\) (defined as the angle of the velocity vector from horizontal), the tiltwing angle \(i_w\) and its derivative \(\dot{i}_w\), and the position \((x, z)\) with respect to inertial frame \(O_{XZ}\). Additional variables are the lift force \(L\), drag force \(D\) and the angle of attack \(\alpha\). All model parameters are defined in Table 1. The following input and state constraints apply \[\begin{gathered} \label{eq:cstr1} i_w + \theta = \alpha + \gamma, \, \underline{M} \leq M \leq \overline{M}, \\ \label{eq:cstr1bis} 0 \leq T \leq \overline{T}, \quad 0 \leq V \leq \overline{V}, \\ \label{eq:cstr2} V(t_0) = V_0 \text{ and } V(t_f) = V_f, \\ \label{eq:cstr3} \underline{a} \leq \dot{V} \leq \overline{a}. \end{gathered}\] \(\theta\) is the pitch angle, defined as the angle of the fuselage axis from horizontal. For passenger comfort, \(\theta\) is regulated via the elevator to track a constant reference \(\theta^* = 0\). In order to account for the effect of the propeller wake on the wing, the flow velocity downstream is augmented by the induced velocity of the propeller. This allows us to define the effective velocity \(V_e\) and effective angle of attack \(\alpha_e\) seen by the wing as \[\alpha_e = \arcsin{\left( \frac{V}{V_e} \sin{\alpha}\right)},\] \[V_e = \sqrt{V^2 + \frac{2T}{\rho A n}}.\] Assuming that the wing is fully immersed in the wake, and that \(\alpha_e\ll 1\) to avoid operating the wing in dangerous near-stall regimes[^1], the lift and drag are modeled as follows  \[D = \tfrac{1}{2} \rho S(a_2 \alpha_e^2 + a_1 \alpha_e + a_0) V_e^2 \approx \tfrac{1}{2} \rho S(a_1 \alpha_e + a_0) V_e^2,\] \[L = \tfrac{1}{2} \rho S(b_1 \alpha_e + b_0) V_e^2, \label{eq:L}\] where \(S\) is the wing area, \(\rho\) is the air density, \(a_0, a_1, a_2\) and \(b_0, b_1\) are constant parameters. # Convex optimisation {#sub:convex_formulation} This paper considers how to robustly generate minimum power trajectories for the transition between powered lift and cruise flight modes, suggesting the following objective function \[\label{eq:obj} J = \int_{t_0}^{t_f} P/\overline{P}\ \mathrm{d}t,\] where \(P=TV \cos \alpha\) is the drive power and \(\overline{P} = \overline{T} \overline{V}\). The optimisation problem consists of minimising [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"} while satisfying dynamical constraints, input and state constraints [\[eq:eom1\]](#eq:eom1){reference-type="eqref" reference="eq:eom1"}-[\[eq:L\]](#eq:L){reference-type="eqref" reference="eq:L"}. As such, this problem is a NLP and we thus consider below how to reformulate the problem as a sequence of convex programs. We introduce 4 key manipulations to do so: i) assuming that a path is known *a priori*, we introduce a change of differential operator to integrate the EOM over space, thus simplifying the structure of the problem; ii) to reduce the couplings between the optimisation variables, we combine both EOM to separate the optimisation of the velocity and torque from the other variables, allowing us to solve 2 smaller optimisation problems sequentially and accelerate computation; iii) we discretise the problem; iv) we approximate the nonlinear dynamics by a difference of convex functions and exploit the fact that convex functions can be bounded tightly by a combination of convex and linear bounds. ## Change of differential operator Assuming that a path \((x(s), z(s))\) parameterised by the curvilinear abscissa \(s\) is known *a priori* (which is usually the case in a UAM context where flight corridors are prescribed) and applying the change of differential operator \(\frac{\mathrm{d}}{\mathrm{d}t} = V \frac{\mathrm{d}}{\mathrm{d}s}, \forall V \neq 0\), the dynamics in [\[eq:eom1\]](#eq:eom1){reference-type="eqref" reference="eq:eom1"}-[\[eq:tiltwing_dyna\]](#eq:tiltwing_dyna){reference-type="eqref" reference="eq:tiltwing_dyna"} can be reformulated as \[\begin{aligned} \label{eq:eom7} \frac{1}{2}m E' = T \cos{\alpha}-\frac{1}{2}\rho S\left(a_1 \alpha_e + a_0\right)\left(E + \frac{2T}{\rho A n}\right)-mg \sin{\gamma^*}, \\ \label{eq:eom8} m E \gamma^{*\prime} = T \sin{\alpha} + \frac{1}{2}\rho S\left(b_1 \alpha_e + b_0\right)\left(E + \frac{2T}{\rho A n}\right)-mg \cos{\gamma^*}, \\ \label{eq:d2i_w2} J_w ( \frac{1}{2}E' i_w' + E i_w'') = M, \, i_w(s_0) = i_{0}, \, i_w'(s_0) \sqrt{E(s_0)} = \Omega_{0}, \end{aligned}\] where \(\tfrac{\mathrm{d}\,\cdot}{\mathrm{d}s}=\cdot'\) and \(E=V^2\). The flight path angle \(\gamma^* = \arctan{(-\mathrm{d}z/\mathrm{d}x)}\) is known *a priori* from the path. ## Problem separation We next reduce the couplings between the states and inputs in the EOM [\[eq:eom7\]](#eq:eom7){reference-type="eqref" reference="eq:eom7"}-[\[eq:d2i_w2\]](#eq:d2i_w2){reference-type="eqref" reference="eq:d2i_w2"} by eliminating the angle of attack from the formulation and separating the optimisation into two subproblems as follows. \(\lambda = a_1/b_1\), the combination [\[eq:eom7\]](#eq:eom7){reference-type="eqref" reference="eq:eom7"} \(+\) \(\lambda\)[\[eq:eom8\]](#eq:eom8){reference-type="eqref" reference="eq:eom8"} yields \[\begin{gathered} \!\!\!\!\! \frac{1}{2}m E' + \underbrace{(\lambda m \gamma^{*\prime} + \frac{1}{2}\rho S (a_0-\lambda b_0))}_{c(\gamma^{*\prime})}E + \underbrace{mg( \sin{\gamma^*} + \lambda \cos{\gamma^*})}_{d(\gamma^*)} \\ = \underbrace{T \cos{\alpha} + \lambda T \sin{\alpha}-S^\star (a_0-\lambda b_0 ) T}_{\tau}, \label{eq:nocvx} \end{gathered}\] where \(S^\star=\frac{S}{An}\) \[\tau = T \cos{\alpha} + \lambda T \sin{\alpha}-S^\star (a_0-\lambda b_0 ) T. \label{eq:tau}\] The state and input constraints in [\[eq:cstr1bis\]](#eq:cstr1bis){reference-type="eqref" reference="eq:cstr1bis"}-[\[eq:cstr3\]](#eq:cstr3){reference-type="eqref" reference="eq:cstr3"} can be rewritten as \[\begin{gathered} \label{eq:cstr_cvx1} 0 \leq E \leq \overline{V}^2, \quad \underline{a} \leq E'/2 \leq \overline{a}, \quad 0\leq \tau \leq \overline{T}, \\ E(s_0) = V_{0}^2 \text{ and } E(s_f) = V_{f}^2. \label{eq:cstr_cvx2} \end{gathered}\] In the thrust constraint in [\[eq:cstr_cvx1\]](#eq:cstr_cvx1){reference-type="eqref" reference="eq:cstr_cvx1"}, \(\tau\) was chosen as a proxy for \(T\) since . Likewise, the minimum power criterion in [\[eq:obj\]](#eq:obj){reference-type="eqref" reference="eq:obj"} can be approximated by a convex objective function under these conditions. By the change of differential operator we obtain Since \(\gamma\) and \(\gamma'\) are prescribed the path, [\[eq:nocvx\]](#eq:nocvx){reference-type="eqref" reference="eq:nocvx"} is a linear equality constraint and the following convex optimisation problem can be constructed to minimise [\[eq:obj_cvx\]](#eq:obj_cvx){reference-type="eqref" reference="eq:obj_cvx"} subject to [\[eq:nocvx\]](#eq:nocvx){reference-type="eqref" reference="eq:nocvx"}, [\[eq:cstr_cvx1\]](#eq:cstr_cvx1){reference-type="eqref" reference="eq:cstr_cvx1"} and [\[eq:cstr_cvx2\]](#eq:cstr_cvx2){reference-type="eqref" reference="eq:cstr_cvx2"} as follows \[\begin{aligned} & \mathcal{P}_1: \min_{\substack{\tau,\,E,\,a}} & & \int_{s_0}^{s_f} \tau/\overline{P} \, \mathrm{d}s, \\ & \text{ s.t.} & & \frac{1}{2}m E' + c(\gamma^{*\prime}) E + d(\gamma^*) = \tau, \\ & & & 0 \leq \tau \leq \overline{T},\ \underline{a} \leq \frac{1}{2} E' \leq \overline{a}, \\ & & & 0 \leq E \leq \overline{V}^2,\ E(s_0) = V_{0}^2,\ E(s_f) = V_{f}^2. \end{aligned}\] Solving \(\mathcal{P}_1\) yields the optimal velocity profile along the path and provides a proxy for the optimal thrust. However, a tiltwing angle profile that meets the dynamical constraints and follows the desired path with \(\gamma \approx \gamma^*\) . we use the solution of \(\mathcal{P}_1\) to define a new optimisation problem with variables \(\gamma\), \(\alpha\), \(i_w\), and \(M\) satisfying the constraints [\[eq:cstr1\]](#eq:cstr1){reference-type="eqref" reference="eq:cstr1"}, [\[eq:d2i_w2\]](#eq:d2i_w2){reference-type="eqref" reference="eq:d2i_w2"} and, using [\[eq:tau\]](#eq:tau){reference-type="eqref" reference="eq:tau"} to eliminate the thrust in [\[eq:eom8\]](#eq:eom8){reference-type="eqref" reference="eq:eom8"}, \[\begin{aligned} m E \gamma' &= \tau \sin{\alpha}-mg \cos{\gamma} \nonumber \\ &\quad +\! \frac{1}{2}\rho S \biggl[ b_1 \arcsin \biggl( \frac{\sqrt{E} \sin{\alpha}}{\sqrt{E + \smash{\frac{2\tau}{\rho A n}}\rule{0pt}{8.5pt}}}\biggr) \!+\! b_0\biggr] \Bigl(E \!+\! \frac{2\tau}{\rho A n}\Bigr), \nonumber \\ & = f(\alpha, E, \tau)-mg \cos{\gamma}, \label{eq:eom_new} \end{aligned}\] \[\label{eq:obj2} J_{\gamma} = \int_{s_0}^{s_f} \frac{(\gamma-\gamma^*)^2}{\sqrt{E}} \ \mathrm{d}s.\] Note that only the two EOM [\[eq:eom8\]](#eq:eom8){reference-type="eqref" reference="eq:eom8"} and [\[eq:d2i_w2\]](#eq:d2i_w2){reference-type="eqref" reference="eq:d2i_w2"} are needed to construct this new problem since the linear combination [\[eq:eom7\]](#eq:eom7){reference-type="eqref" reference="eq:eom7"} + \(\lambda\) [\[eq:eom8\]](#eq:eom8){reference-type="eqref" reference="eq:eom8"} is enforced with \(\tau\) and \(E\) prescribed from problem \(\mathcal{P}_1\). We thus state the following optimisation problem \[\begin{aligned} & \mathcal{P}_2: \min_{\substack{\alpha,\,\gamma,\,i_w,\, M}} & & \int_{s_0}^{s_f} \frac{(\gamma-\gamma^*)^2}{\sqrt{E}} \ \mathrm{d}s \nonumber\\ & \text{ s.t.} & & m E \gamma' =f(\alpha, E, \tau)-mg \cos{\gamma}, \nonumber\\ & & & J_w (\tfrac{1}{2} E' i_w' + i_w'' E) = M,\ i_w(s_0) = i_{0}, \nonumber \\ & & & i_w'(s_0) \sqrt{E(s_0)} = \Omega_{0}, \nonumber\\ & & & i_w = \alpha + \gamma, \nonumber\\ & & & \underline{M} \leq M \leq \overline{M},\ \underline{\alpha} \leq \alpha \leq \overline{\alpha} \nonumber\\ & & & \underline{\gamma} \leq \gamma \leq \overline{\gamma},\ \underline{i_w} \leq i_w \leq \overline{i_w}, \nonumber\\ \end{aligned}\] and reconstruct the input \(T\) and state \(V\) *a posteriori* using [\[eq:tau\]](#eq:tau){reference-type="eqref" reference="eq:tau"} and \(V = \sqrt{E}\). Given the solution of both problems as functions of the independent variable \(s\), the step is to map the solution to time domain by reversing the change of differential operator and integratin \[t(\] We have now achieved the separation into two subproblems \(\mathcal{P}_1\) and \(\mathcal{P}_2\), as described in. ## Discretisation {#sec:discrete} The decision variables in \(\mathcal{P}_1\) and \(\mathcal{P}_2\) are functions defined on the interval \([s_0,s_f]\). To obtain computationally tractable problems, we consider \(N+1\) discretisation points \(\{s_0, s_1, \ldots, s_{N}\}\) of the path, with spacing \(\delta_k = s_{k+1}-s_{k}\), \(k = 0, \ldots, N-1\) (\(N\) steps). The notation \(\{u_0, \ldots, u_{N}\}\) is used for the sequence of the discrete values of a continuous variable \(u\) evaluated at the discretisation points of the mesh, where \(u_k = u(s_{k})\), \(\forall k \in \{0, \ldots, N\}\). Assuming a path \(s_k \rightarrow (x_k, z_k)\), the prescribed flight path angle and rate are discretised as follows \[\begin{aligned} \gamma_k^* &= \arctan \Bigl(-\frac{z_{k+1}-z_{k}}{x_{k+1}-x_{k}}\Bigr), \quad k \in \{0, \ldots, N-1\}, \label{eq:gamma} \\ \gamma_k^{*\prime} &= \begin{cases} (\gamma_{k+1}^*-\gamma_{k}^*)/{\delta_k}, & k \in \{0, \ldots, N-2\}, \\ \gamma_{N-2}^{*\prime}, & k = N-1. \end{cases} \label{eq:dgamma} \end{aligned}\] The resulting discretised versions of \(\mathcal{P}_1\) and \(\mathcal{P}_2\) are \[\begin{aligned} & \mathcal{P}_1^\dagger: \min_{\substack{\tau,\,E,\,a}} & & \sum_{k=0}^{N-1} \tau_k/\overline{P} \, \delta_k, \nonumber\\ & \text{ s.t.} & & E_{k+1} = E_{k} + \frac{2 \delta_{k}}{m} (\tau_k-c(\gamma_k^{*\prime}) E_k-d(\gamma_k^*)), \nonumber\\ & & & 0 \leq \tau_k \leq \overline{T}, \quad \underline{a} \leq \frac{E_{k+1}-E_{k}}{2 \delta_{k}} \leq \overline{a}, \nonumber\\ & & & 0 \leq E_k \leq \overline{V}^2, \quad E_0 = V_{0}^2, \quad E_{N} = V_{f}^2, \nonumber\\ \end{aligned}\] \[\begin{aligned} & \mathcal{P}_2^\dagger: \min_{\substack{\alpha,\,\gamma,\\i_w,\, \zeta,\, M}} & & \sum_{k=0}^{N-1} \frac{(\gamma_k-\gamma_k^*)^2}{\sqrt{E_k}}\delta_k \nonumber\\ & \text{s.t.} & & \gamma_{k+1} = \gamma_{k} + \frac{\delta_k}{m E} (f_k(\alpha_k, E_k, \tau_k)-f_{\gamma_k}(\gamma_k)), \nonumber\\ & & & i_{w,k} = \alpha_k + \gamma_k, \nonumber\\ & & & i_{w,k+1} = i_{w,k} + \zeta_k \delta_k, \quad i_{w,0} = i_{0}, \nonumber\\ & & & \zeta_{k+1} = \zeta_{k}\Bigl(1-\frac{E_{k+1}-E_k}{2E_k}\Bigr) + \frac{M_k \delta_k}{J_w {E_k}}, \nonumber \\ & & & \zeta_0 \sqrt{E}_0 = \Omega_{0}, \nonumber \\ & & & \underline{M} \leq M_k \leq \overline{M}, \quad \underline{\alpha} \leq \alpha_k \leq \overline{\alpha}, \nonumber\\ & & & \underline{\gamma} \leq \gamma_k \leq \overline{\gamma}, \quad \underline{i_w} \leq i_{w,k} \leq \overline{i_w}. \nonumber\\ \end{aligned}\] where \(f_{\gamma_k} =-mg \cos{\gamma_k}\). The input and state are reconstructed using \[\label{eq:reconstruct1} T_k = \frac{\tau_k}{ \cos{\alpha_k} + \lambda \sin{\alpha_k}-S^\star (a_0-\lambda b_0 )}, \quad V_k = \sqrt{E_k},\] the time \(t_k\) associated with each discretisation point as time series \[\label{eq:back2time} t_k = \sum_{j=0}^{k-1} \frac{\delta_j}{V_j}.\] We now have a pair of finite dimensional problems \(\mathcal{P}_1^\dagger\) and \(\mathcal{P}_2^\dagger\), but the latter is still nonconvex due to the nonlinear functions \(f_k\) and \(f_{\gamma_k}\) in the dynamics. On a restricted domain \(\gamma_k \in [-\pi/2; \pi/2]\), \(f_{\gamma_k}\) is a convex function of \(\gamma_k\), making it possible to derive tight convex bounds on \(f_{\gamma_k}\) (as discussed in Section [3.5](#sec:cvx_relax){reference-type="ref" reference="sec:cvx_relax"}). However, this is not the case with \(f_k\) and we introduce a method to alleviate this limitation in what follows. ## DC decomposition Motivated by the fact that convex functions can be bounded tightly by convex and linear inequalities (as in ), we seek a decomposition of \(f\) as a Difference of Convex (DC) functions: \(f_k = g_k-h_k\), where \(g_k, h_k\) are convex. A DC decomposition always exists if \(f_k \in \mathcal{C}^2\). Note that since \((E_k, \tau_k)\) are obtained from problem \(\mathcal{P}_1^\dagger\), the function \(f_k(\alpha_k, E_k, \tau_k)\) is single-valued (in \(\alpha_k\)) which considerably simplifies the task of finding a DC decomposition, and motivates the above approach of separating the initial problem in two subproblems with fewer couplings between the variables. However, \(f_k\) is also time varying through its dependence on parameters \((E_k, \tau_k)\) generated online. This requires us to find a DC split for every instance of \((E_k, \tau_k), \, \forall k \in [0, 1,..., N]\) which can be intractable if the horizon is large or the sampling interval is small. Instead, we adopt the more pragmatic approach of i) precomputing offline the DC decompositions on a downsampled grid of values \((E_i, \tau_j), \, \forall (i, j) \in [0, 1,..., N_s] \times [0, 1,..., M_s]\) where \(N_s, M_s \ll N+1\) and ii) interpolating the obtained decompositions online using a lookup table. ### Precomputation of the DC decomposition Inspired by, we develop a computationally tractable method for the DC decomposition of a function \(f_k(\alpha)\) based on an approximation[^2] of the function by a polynomial of degree \(2n\): \[\label{eq:pol_approx} f_k (\alpha) \approx p_{k, 2n} \alpha^{2n} +... + p_{k, 1} \alpha + p_{k, 0} = y^\top P_k y,\] where \(y = [1 \quad \alpha \quad... \quad \alpha^n]^\top \in \mathbb{R}^{n+1}\) is a vector of monomials of increasing order and \(P_k = P_k^\top \in R^{n+1 \times n+1}\) is the Gram matrix of the polynomial defined by \(\{P_k\}_{ij} = p_{k, i+j +1}/\lambda(i, j), \, \forall i,j \in [0, 1,..., n]\) where \(\lambda(i, j) = i + j + 1\) if \(i+j \leq n\) and \(\lambda(i, j) = 2n + 1-(i + j)\). Given \(N_s\) samples \(F_{k, s} \, \forall s \in [1,..., N_s]\) of the function \(f_k\), the polynomial approximation can be obtained by solving a least square problem to find the coefficients that best fit the samples. We now seek the symmetric matrices \(Q_k\), \(R_k\) such that \[y^\top P_k y = y^\top Q_k y-y^\top R_k y,\] where \(g_k \approx y^\top Q_k y\) and \(h_k \approx y^\top R_k y\) are convex polynomials in \(\alpha\). Such conditions can be satisfied if the Hessians \(d^2g_k/d\alpha^2 = y^\top H_{g_k} y\) and \(d^2h_k/d\alpha^2 = y^\top H_{h_k} y\) are Positive Semi-Definite (PSD), i.e. if the following Linear Matrix Inequalities (LMI) hold \[H_{g_k} \equiv {D^\top}^2 Q_k + Q_k D^2 + 2 D^\top Q_k D \succeq 0,\] \[H_{h_k} \equiv {D^\top}^2 R_k + R_k D^2 + 2 D^\top R_k D \succeq 0,\] where \(D\) is a matrix of coefficients such that \(dy/d\alpha = D y\). Finding the DC decomposition thus reduces to solving the following Semi Definite Program (SDP) \[\begin{aligned} & \mathcal{SDP}: \min_{\substack{H_{g_k}}} & & \mathrm{tr} \, H_{g_k} \nonumber\\ & \text{ s.t.} & & H_{g_k} \succeq 0, \nonumber\\ & & & H_{g_k}-({D^\top}^2 P_k + P_k D^2 + 2D^\top P_k D) \succeq 0, \nonumber \end{aligned}\] and computing \(H_{h_k} = H_{g_k}-{D^\top}^2 P_k + P_k D^2 + D^\top P_k D\), followed by the double integration \(d^2g_k/d\alpha^2 = y^\top H_{g_k} y\) and \(d^2h_k/d\alpha^2 = y^\top H_{h_k} y\) to recover \(g_k\) and \(h_k\). This operation is repeated at each point \((E_i, \tau_j)\) of the grid to assemble a look-up table of polynomial coefficients. Note that the objective was chosen so as to regularise the solutions for \(g_k\), \(h_k\) by minimising a proxy for their average curvature, in order to minimise linearisation errors later on. In Figure [\[fig:split\]](#fig:split){reference-type="ref" reference="fig:split"}, we illustrate a typical DC decomposition of the nonlinear dynamics for a given \((E_i, \tau_j)\). ### Coefficient interpolation A bilinear interpolation of the coefficients is performed online to obtain the DC decomposition for each \((E_k, \tau_k), \, \forall k \in [0,..., N]\). This operation preserves convexity since the interpolated polynomial coefficients are a weighted sum of the coefficients in the lookup table. ## Convex relaxation {#sec:cvx_relax} Consider again the nonlinear dynamics in problem \(\mathcal{P}_2^\dagger\), using the DC decomposition of \(f_k\) computed in the previous section and eliminating the angle of attack via \(\alpha_k = i_{w, k}-\gamma_k\) to reduce the number of states, we obtain \[\label{eq:dyna_eq_cvx} \begin{split} \gamma_{k+1} = \gamma_{k} + \frac{\delta_k}{m E} (g_k(i_{w, k}-\gamma_k, E_k, \tau_k) \\-h_k(i_{w, k}-\gamma_k, E_k, \tau_k)-mg \cos \gamma_k). \end{split}\] All nonlinearities in equation [\[eq:dyna_eq_cvx\]](#eq:dyna_eq_cvx){reference-type="eqref" reference="eq:dyna_eq_cvx"} above involve convex and concave functions of the states \(i_{w, k}\) and \(\gamma_k\) whose dynamics are given by \[\begin{gathered} \label{eq:lqr1} i_{w,k+1} = i_{w,k} + \zeta_k \delta_k, \\ \label{eq:lqr2} \zeta_{k+1} = \zeta_{k}\Bigl(1-\frac{E_{k+1}-E_k}{2E_k}\Bigr) + \frac{M_k \delta_k}{J_w {E_k}}. \end{gathered}\] In what follows we will exploit the convexity properties of the functions \(g_k, h_k, f_{\gamma_k} =-mg \cos \gamma_k\) in [\[eq:dyna_eq_cvx\]](#eq:dyna_eq_cvx){reference-type="eqref" reference="eq:dyna_eq_cvx"} to approximate the dynamics by a set of convex inequalities with tight bounds on the state trajectories. To do so, we linearise the dynamics successively around feasible guessed trajectories and treat the linearisation error as a bounded disturbance . We use the fact that the linearisation error of a convex (resp. concave) function is also convex (resp. concave) and can thus be bounded tightly since its maximum (resp. minimum) occurs at the boundary of the set on which the function is constrained. This allows us to construct a robust optimisation using the tube-based MPC framework, and to obtain solutions that are robust to the model error introduced by the linearisation. We start by assuming the existence of a set of feasible trajectories \(i_{w, k}\) and \(\gamma_k\) for [\[eq:dyna_eq_cvx\]](#eq:dyna_eq_cvx){reference-type="eqref" reference="eq:dyna_eq_cvx"}-[\[eq:lqr2\]](#eq:lqr2){reference-type="eqref" reference="eq:lqr2"} and consider the perturbed dynamics \[\label{eq:dyna_eq_cvx_perturbed} \begin{split} \gamma_{k+1} = \gamma_{k} + \frac{\delta_k}{m E} (g_k^\circ + \nabla g_k^\circ (i_{w, k}-\gamma_k-(i_{w, k}^\circ-\gamma_k^\circ)) \\ + w_1-h_k^\circ-\nabla h_k^\circ (i_{w, k}-\gamma_k-(i_{w, k}^\circ-\gamma_k^\circ))-w_2 \\-mg \cos \gamma_k^\circ + mg \sin \gamma_k^\circ (\gamma_k-\gamma_k^\circ) + w_3). \end{split}\] where \(g_{k}^{\circ} = g_k(i_{w, k}^\circ-\gamma_k^\circ)\), \(h_k^\circ = h_k(i_{w, k}^\circ-\gamma_k^\circ)\) are the functions \(g_k, h_k\) evaluated along the guessed trajectory, \(\nabla g_k^\circ = dg_k/d\alpha_k (i_{w, k}^\circ-\gamma_k^\circ)\), \(\nabla h_k^\circ = dh_k/d\alpha_k (i_{w, k}^\circ-\gamma_k^\circ)\) are the first order derivatives of \(g_k, h_k\) evaluated along the guessed trajectory, and \(w_1(i_{w}-\gamma, i_{w, k}^\circ-\gamma_k^\circ)\), \(w_2(i_{w}-\gamma, i_{w, k}^\circ-\gamma_k^\circ)\), \(w_3(\gamma, \gamma_k^\circ)\) are the convex linearisation errors of \(g_k, h_k, f_{\gamma_k} =-mg \cos \gamma_k\) respectively. Since these linearisation errors are convex, they take their maximum on the boundary of the set over which the functions are constrained. Moreover, by definition, their minimum on this set is zero (Jacobian linearisation). We thus infer the following relationships \(\forall i = \{1, 2\}\) and noting \(f_1 \equiv g, f_2 \equiv h\) \[\label{eq:wi_min} \min_{\substack{\gamma \in [\underline{\gamma}_k, \overline{\gamma}_k]\\ i_w \in [\underline{i}_{w, k}, \overline{i}_{w, k}]}} w_i(i_w-\gamma, i_{w}^\circ-\gamma^\circ) = 0,\] \[\label{eq:wi_max} \begin{split} \max_{\substack{\gamma \in [\underline{\gamma}_k, \overline{\gamma}_k]\\ i_w \in [\underline{i}_{w, k}, \overline{i}_{w, k}]}} w_i(i_w-\gamma, i_{w}^\circ-\gamma^\circ) = \qquad \qquad \qquad \qquad \qquad\\ \max \{f_{i, k}-f_{i, k}^\circ-\nabla f_{i, k}^\circ (\overline{i}_{w, k}-\underline{\gamma}_k-(i_{w, k}^\circ-\gamma_k^\circ)) ;\\ f_{i, k}-f_{i, k}^\circ-\nabla f_{i, k}^\circ (\underline{i}_{w, k}-\overline{\gamma}_k-(i_{w, k}^\circ-\gamma_k^\circ))\}, \end{split}\] \[\label{eq:w3_max} \begin{split} \min_{\substack{\gamma \in [\underline{\gamma}_k, \overline{\gamma}_k]}} w_3(\gamma, \gamma^\circ) = 0 \quad \text{and} \quad \max_{\substack{\gamma \in [\underline{\gamma}_k, \overline{\gamma}_k]}} w_3(\gamma, \gamma^\circ) =\\ \max \{-mg \cos \underline{\gamma}_k + mg \cos \gamma_k^\circ-mg \sin \gamma_k^\circ (\underline{\gamma}_k-\gamma_k^\circ); \\-mg \cos \overline{\gamma}_k + mg \cos \gamma_k^\circ-mg \sin \gamma_k^\circ (\overline{\gamma}_k-\gamma_k^\circ) \} \end{split}\] where we assumed that the state trajectories \(\gamma_k\) and \(i_{w, k}\) lie within \"tubes\" whose cross-sections are parameterised by means of elementwise bounds \(\gamma_{k} \in [\underline{\gamma}_{k}, \overline{\gamma}_{k}]\) and \(i_{w, k} \in [\underline{i}_{w, k}, \overline{i}_{w, k}] ,\ \forall k\), which are considered to be optimisation variables. Given these bounds on the states at a given time instant and by virtue of equations [\[eq:wi_min\]](#eq:wi_min){reference-type="eqref" reference="eq:wi_min"}-[\[eq:w3_max\]](#eq:w3_max){reference-type="eqref" reference="eq:w3_max"}, the bounds on the states at the next time instant satisfy the following convex inequalities \[\label{eq:dyna_ineq5} \begin{split} \overline{\gamma}_{k+1} \geq \max_{\substack{\gamma \in \{\underline{\gamma}_k; \overline{\gamma}_k\}\\ i_w \in \{\underline{i}_{w, k}; \overline{i}_{w, k}\}}} \Big\{ \gamma + \frac{\delta_k}{m E} (g_k(i_{w}-\gamma)-h_k(i_{w}^\circ-\gamma^\circ) \\-\nabla h_k^\circ (i_{w}-\gamma-(i_{w, k}^\circ-\gamma_k^\circ))-mg \cos \gamma) \Big\}, \end{split}\] \[\label{eq:dyna_ineq6} \begin{split} \underline{\gamma}_{k+1} \leq \min_{\substack{\gamma \in \{\underline{\gamma}_k, \overline{\gamma}_k\}\\ i_w \in \{\underline{i}_{w, k}, \overline{i}_{w, k}\}}} \Big\{ \gamma + \frac{\delta_k}{m E} (g_k(i_{w}^\circ-\gamma^\circ) \\+\nabla g_k^\circ (i_{w}-\gamma-(i_{w, k}^\circ-\gamma_k^\circ))-h_k(i_{w}-\gamma) \\- mg \cos \gamma^\circ + mg \sin \gamma^\circ (i_{w}-\gamma) ) \Big\}, \end{split}\] \[\label{eq:iw} \overline{i}_{w,k+1} \geq \overline{i}_{w,k} + \zeta_k \delta_k, \quad \underline{i}_{w,k+1} \leq \underline{i}_{w,k} + \zeta_k \delta_k.\] These conditions involve only minimisations of linear functions and maximisations of convex functions. Note that the functions to optimise no longer need to be evaluated on continuous intervals but at their boundaries \(\{\underline{\gamma}_k, \overline{\gamma}_k\}\) and \(\{\underline{i}_{w, k}, \overline{i}_{w, k}\}\) which implies that each maximisation and minimisation above reduces to \(2^2 = 4\) convex inequalities. Moreover, this number can be reduced to avoid the curse of dimensionality since the coefficients of the linear functions appearing in each maximisation and minimisation are known. Finally, the computational burden can be further reduced by introducing a low order approximation of the polynomials in [\[eq:dyna_ineq5\]](#eq:dyna_ineq5){reference-type="eqref" reference="eq:dyna_ineq5"}-[\[eq:iw\]](#eq:iw){reference-type="eqref" reference="eq:iw"}. This was obtained by computing, before including the constraints in the optimisation, a series of quadratic polynomials to each \(g_k\), \(h_k\), \(\forall k\) that are a best fit around \(i_{w, k}^\circ-\gamma_k^\circ\). The tube defined by inequalities [\[eq:dyna_ineq5\]](#eq:dyna_ineq5){reference-type="eqref" reference="eq:dyna_ineq5"}-[\[eq:iw\]](#eq:iw){reference-type="eqref" reference="eq:iw"} can be used to replace \(\mathcal{P}_2^\dagger\) by a sequence of convex programs. Given the solution of \(\mathcal{P}_1^\dagger\) and given a set of feasible (suboptimal) trajectories \(i_{w, k}^\circ\), \(\gamma_k^\circ\) satisfying [\[eq:dyna_eq_cvx\]](#eq:dyna_eq_cvx){reference-type="eqref" reference="eq:dyna_eq_cvx"}-[\[eq:lqr2\]](#eq:lqr2){reference-type="eqref" reference="eq:lqr2"}, the following convex problem is solved sequentially \[\begin{aligned} & \mathcal{P}_2^\ddagger: \min_{\substack{\overline{\gamma},\,\underline{\gamma},\, \overline{i}_w,\\ \underline{i}_w,\, \zeta,\, M, \theta}} & & \sum_{k=0}^{N-1} \frac{\theta_k^2}{\sqrt{E_k}}\delta_k \nonumber\\ & \text{s.t.} & & \theta_k \geq |\overline{\gamma}_k-\gamma_k^*|, \quad \theta_k \geq |\underline{\gamma}_k-\gamma_k^*|, \nonumber\\ & & & \overline{\gamma}_{k+1} \geq \max_{\substack{\gamma \in \{\underline{\gamma}_k; \overline{\gamma}_k\}\\ i_w \in \{\underline{i}_{w, k}; \overline{i}_{w, k}\}}} \Big\{ \gamma + \frac{\delta_k}{m E} (g_k(i_{w}-\gamma) \nonumber \\ & & &-h_k(i_{w}^\circ-\gamma^\circ)-\nabla h_k^\circ (i_{w}-\gamma-(i_{w, k}^\circ-\gamma_k^\circ)) \nonumber\\ & & &-mg \cos \gamma) \Big\}, \nonumber\\ & & & \underline{\gamma}_{k+1} \leq \min_{\substack{\gamma \in \{\underline{\gamma}_k, \overline{\gamma}_k\}\\ i_w \in \{\underline{i}_{w, k}, \overline{i}_{w, k}\}}} \Big\{ \gamma + \frac{\delta_k}{m E} (g_k(i_{w}^\circ-\gamma^\circ) \nonumber \\ & & & +\nabla g_k^\circ (i_{w}-\gamma-(i_{w, k}^\circ-\gamma_k^\circ))-h_k(i_{w}-\gamma) \nonumber\\ & & &-mg \cos \gamma^\circ + mg \sin \gamma^\circ (i_{w}-\gamma) ) \Big\}, \nonumber\\ & & & \overline{i}_{w,k+1} \geq \overline{i}_{w,k} + \zeta_k \delta_k, \quad \overline{i}_{w,0} = i_{0}, \nonumber\\ & & & \underline{i}_{w,k+1} \leq \underline{i}_{w,k} + \zeta_k \delta_k, \quad \underline{i}_{w,0} = i_{0}, \nonumber\\ & & & \zeta_{k+1} = \zeta_{k}\Bigl(1-\frac{E_{k+1}-E_k}{2E_k}\Bigr) + \frac{M_k \delta_k}{J_w {E_k}}, \nonumber \\ & & & \zeta_0 \sqrt{E}_0 = \Omega_{0}, \nonumber \\ & & & \underline{M} \leq M_k \leq \overline{M}, \quad \underline{i_w} \leq \underline{i}_{w,k}, \quad \overline{i}_{w,k} \leq \overline{i_w}, \nonumber\\ & & & \underline{\alpha} \leq \underline{i}_{w, k}-\overline{\gamma}_k, \quad \overline{i}_{w, k}-\underline{\gamma}_k \leq \overline{\alpha}, \nonumber\\ & & & \underline{\gamma} \leq \underline{\gamma}_k, \quad \overline{\gamma}_k \leq \overline{\gamma}, \quad \overline{\gamma}_{0} = \gamma_{0}, \quad \underline{\gamma}_{0} = \gamma_{0}. \nonumber\\ \end{aligned}\] . The process is repeated until \(|\overline{\gamma}_k-\underline{\gamma}_k|\) and \(|\overline{i}_{w, k}-\underline{i}_{w, k}|\) have converged. Once \(\mathcal{P}_1^\dagger\) and \(\mathcal{P}_2^\ddagger\) have been solved, we check whether \(|\gamma_k^*-\gamma_k| \leq \epsilon\) \(\forall k \in \{0, \ldots, N\}\), where \(\epsilon\) is a specified tolerance. If this condition is not met (\(\mathcal{P}_2^\ddagger\) may admit solutions that allow \(\gamma_k\) to differ from the assumed flight path angle \(\gamma^*_k\)), the problem is reinitialized with the updated flight path angle and rate \(\gamma^*_k \gets \gamma_k\), \(\gamma^{*\prime}_k \gets \gamma_k'\) and . When (or maximum number of iterations ) the problem is considered solved and the input and state are reconstructed using the equations in [\[eq:reconstruct1\]](#eq:reconstruct1){reference-type="eqref" reference="eq:reconstruct1"} and the time \(t_k\) associated with each discretisation point as time series. The procedure is summarised in Algorithm 1. Remarks on \(\mathcal{P}_2^\ddagger\): i) the angle of attack has been eliminated from the formulation; ii) the slack variable \(\theta_k\) was introduced to enforce the objective \(\sum_k \max_{\gamma_k \in \{\underline{\gamma}_k;\overline{\gamma}_k\}} (\gamma_k-\gamma^*)^2/\sqrt{E_k}\); iii) to ensure convexity, it is important that \(\overline{\gamma} \leq \pi/2\) and \(\underline{\gamma} \geq-\pi/2\); iv) to improve numerical stability, \(M_k\) can be replaced by \(M_k + K_{i, k}(i_{w, k}-i_{w, k}^\circ) + K_{\zeta, k} (\zeta_k-\zeta_{k}^\circ)\) with \(i_{w, k} \in \{\underline{i}_{w, k}; \overline{i}_{w, k} \}\) where \(K_{i, k}\) and \(K_{\zeta, k}\) are gains obtained, e.g. by solving a LQR problem for the time varying linear system in equations [\[eq:lqr1\]](#eq:lqr1){reference-type="eqref" reference="eq:lqr1"} and [\[eq:lqr2\]](#eq:lqr2){reference-type="eqref" reference="eq:lqr2"}; v) order reduction was performed on polynomials \(g_k\), \(h_k\), i.e. quadratic polynomials were fitted to \(g_k\), \(h_k\) around \(i_{w, k}^\circ-\gamma_k^\circ\) for all \(k\) by solving a least squares problem before running the optimisation. # Results {#sec:results} We consider a case study the Airbus A\(^3\) Vahana. The parameters are reported in Table [\[tab:param\]](#tab:param){reference-type="ref" reference="tab:param"}. We run Algorithm 1 the convex programming software package CVX with Mosek to compute the optimal trajectory for 2 different transition manoeuvres, with boundary conditions given in Table [1](#tab:BC){reference-type="ref" reference="tab:BC"}. For the sake of simplicity, and unless otherwise stated, we limit the number of iterations of problem \(\mathcal{P}_1^\dagger\) to 1 and of \(\mathcal{P}_2^\ddagger\) to 3. The average computation time per iteration of \(\mathcal{P}_2^\ddagger\) was 7.3s. The first scenario is a (near) constant altitude forward transition. This manoeuvre is abrupt and requires a zero flight path angle throughout as illustrated in Figure [\[fig:traj1\]](#fig:traj1){reference-type="ref" reference="fig:traj1"}. As the aircraft transitions from powered lift to cruise, the velocity magnitude increases (a) and the thrust decreases (b), illustrating the change in lift generation from propellers to wing. The tiltwing angle drops quickly at the beginning (c), resulting in an increase in the angle of attack (d). The slight discrepancy in the flight path angle curves in (c) illustrates that problem \(\mathcal{P}_2^\ddagger\) needs not necessarily generate a flight path angle profile corresponding to the exact desired path if the latter is not feasible. Note from graph (d) that the effective angle of attack stays within reasonable bounds, indicating that the wing is not stalled. By contrast to the solution presented in, the angle of attack is not constrained to small values and we can thus achieve a more aggressive transition at an almost constant altitude, with a maximum altitude drop of about 4 m, see Figure [\[fig:zoom\]](#fig:zoom){reference-type="ref" reference="fig:zoom"}. Convergence of problem \(\mathcal{P}_2^\ddagger\) after 3 iterations is shown in Figure [\[fig:obj\]](#fig:obj){reference-type="ref" reference="fig:obj"}. The tube bounds and objective converge quickly toward infinitesimal values after just a few iterations. After that, no more progress can be achieved. For completeness, we consider a second consisting of a backward transition with an increase in altitude (Figure [\[fig:traj2\]](#fig:traj2){reference-type="ref" reference="fig:traj2"}). is characterised by decrease in velocity magnitude and increase in thrust. An increase in altitude of about 200 m is needed for this manoeuvre due to strict bounds on the effective angle of attack. A backward transition at constant altitude would require the wing, which is prohibited in the present formulation, illustrating a limitation of our approach. To achieve the backward transition, a high-drag device or flaps . This was modelled by adding a constant term \(d\) to \(c(\gamma^{*\prime})\) in problem \(\mathcal{P}_1^\dagger\) for the backward transition. # Conclusions {#sec:conclusion} This paper addresses the trajectory optimisation problem for the transition of a tiltwing VTOL aircraft, leveraging DC decomposition of the dynamics and robust tube programming. The approach is based on successive linearisation of the dynamics around feasible trajectories and treating the linearisation error as a bounded disturbance. The DC form of the dynamics allows to enforce tight bounds on the disturbance via a set of convex inequalities that form the basis of a computationally tractable robust optimisation. The algorithm can compute safe trajectories that are robust to model uncertainty for abrupt transitions at near constant altitude, extending the results in. Another contribution of the present work is the extension of the robust tube optimisation paradigm presented in to dynamic systems that are not convex, by means of a DC decomposition of the nonlinear dynamics. Limitations of the present approach are: i) to obtain a computationally tractable formulation, quadratic approximations of the DC polynomials are required; ii) the computation time, although relatively low compared to solving a NLP, is still too high to leverage the optimisation in a MPC setting. Future work will alleviate these problems by i) considering other types of basis functions for the nonlinear dynamics approximation, e.g. radial basis functions that have better scalability than a monomial basis; ii) the use of first order solvers such as ADMM to accelerate computations. We will then investigate robust MPC for the transition of tiltwing VTOL aircraft. [^1]: This will be imposed through a constraint in the optimisation and will be verified *a posteriori* from simulation results. [^2]: Note that any continuous function can be approximated arbitrarily closely by a polynomial.

{'timestamp': '2023-02-10T02:03:04', 'yymm': '2302', 'arxiv_id': '2302.04361', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04361'}

# Introduction The Circular Unitary Ensemble (\(CUE\)) is the unitary group \(U\) equipped with the normalized Haar measure. The Circular Orthogonal Ensemble (\(COE\)) contains matrices of the form \(SS^T\), with \(S\) in the CUE, while the Circular Symplectic Ensemble (\(CSE\)) contains matrices of the form \(SS^D\), with \(S\) in the CUE. Here \(S^T\) and \(S^D\) are the transpose and the quaternion dual of \(S\). As a result, matrices in the COE are symmetric and matrices in the CSE are self-dual . Physically, circular ensembles are important as models of random unitary propagators in complex quantum systems, with the particular ensembles corresponding to different symmetry classes (presence or absence of time-reversal and spin rotation invariances). Mathematically, circular ensembles can be seen as examples of symmetric spaces, related to the unitary group \(U(N)\) itself, in the case of \(CUE(N)\), and two of its quotients, namely \(U(N)/O(N)\) in the case of \(COE(N)\) and \(U(2N)/Sp(N)\) in the case of \(CSE(N)\). Statistically, eigenvalues of matrices from circular ensembles are among the simplest models of correlated random variables, because they have a constant density and their joint probability distribution consists only of a Vandermonde term, \(|\Delta(S)|^{\beta}\), where the Dyson index is \(\beta=1\), \(2\), \(4\) for \(COE\), \(CUE\), \(CSE\), respectively. These properties of Dyson's circular ensembles can be contrasted, for example, with Wigner's Gaussian ensembles, introduced in order to model quantum Hamiltonians, for which the spectral density is not constant and the joint probability distribution of eigenvalues contains extra terms besides the Vandermonde. In this work, we are not interested in spectral statistics but in the complementary problem of the joint distribution of matrix elements. Specifically, the problem of computing moments, i.e. the average value of a product of matrix elements. This can be reduced to the calculation of Weingarten functions, all of which are known for classical compact Lie groups and associated symmetric spaces. Our contribution is to uncover a hidden Gaussian ensemble inside each of the three circular ensembles. This allows the introduction of Gaussian diagrammatics, i.e. topological expansions in terms of ribbon graphs, for the calculation of moments of \(C\beta E(N)\) in the form of a series in inverse power of the quantity \(\frac{\beta}{2}(N-1)+1\), i.e. in inverse powers of \(N+1\), \(N\), \(2N-1\) for \(COE\), \(CUE\), \(CSE\), respectively. The peculiarity is that in these Gaussian ensembles \(N\) appears as a parameter, and they do not correspond to matrices of positive integer dimension. Instead, the dimension should be taken as \(1-2/\beta\). Our results provide new expansions for the Weingarten functions of the three circular ensembles and, indirectly, for the classical compact Lie groups. Other versions of these expansions already have been studied, always relating the coefficients with the solution of some combinatorial problem such as counting maps or factorizations of permutations. A direct proof of the equivalence between the present results and previous ones is an open problem. This paper is organized as follows. In Section 2 we present a short review of the combinatorics of Gaussian matrix models, emphasizing the case involving complex symmetric matrices. In Section 3 we uncover the Gaussian model inside circular ensembles and show thy can be used to obtain moments. In Section 4 we compare our diagrammatics with another diagrammatics for circular ensembles, developed in the context of quantum chaos. In Section 5 we compute, as an application of our approach, statistics of traces of submatrices of dimension \(M\) within \(COE(N)\). We conclude in Section 6. # Review of Gaussian matrix combinatorics Let \(0_d\) and \(1_d\) denote the \(d\)-dimensional null matrix and identity matrix, respectively. Let \(A^T\) and \(\overline{A}\) be the transpose and the complex conjugate of \(A\), and \(A^\dagger =\overline{A^T}\). With \(J=\begin{pmatrix}0_d&1_d\\-1_d&0_d\end{pmatrix}\), define the dual \(A^D=JA^TJ^T\). Let \(Z\) denote complex matrices of dimension \(d\), with no symmetry when \(\beta=2\), symmetric (\(Z^T=Z\)) when \(\beta=1\) and self-dual (\(Z^D=Z\)) when \(\beta=4\). Let \[P_\beta(Z)=\frac{1}{\mathcal{G}(d,\beta)}e^{-\Omega_\beta{\rm Tr}(ZZ^\dagger)}\] be the probability distribution of \(Z\), with \[\mathcal{G}(d,\beta)=\int dZ e^{-\Omega_\beta{\rm Tr}(ZZ^\dagger)}.\] According to Wick's theorem, the integral of a product of \(2n\) matrix elements, \(n\) from \(Z\) and \(n\) from \(\overline{Z}\), will be given as a sum over all possible pairings between \(Z\) and \(\overline{Z}\) elements. For \(\beta=2\), this means \[\int dZ P_2(Z)\prod_{k=1}^n z_{i_{k}j_{k}}\overline{z}_{a_{k}b_{k}}= \sum_{\sigma\in S_n}\prod_{k=1}^n \int dZ P_2(Z) z_{i_{k}j_{k}}\overline{z}_{a_{\pi(k)}b_{\pi(k)}},\] where \(S_n\) is the permutation group. For \(\beta=1\) matrices are symmetric, so the pairing is allowed to reverse indices. This leads to \[\int dZ P_1(Z)\prod_{k=1}^{n} z_{i_{2k-1}i_{2k}}\overline{z}_{j_{2k-1}j_{2k}}= \sum_{\sigma\in H_n}\prod_{k=1}^n \int dZ P_1(Z) z_{i_{2k-1}i_{2k}}\overline{z}_{j_{\sigma(2k-1)}j_{\sigma(2k)}},\] where \(H_n\) is the hyperoctahedral group, the wreath product \(S_2\wr S_n\). The situation for \(\beta=4\) is more convoluted, but it has long been known that models with \(\beta=4\) are dual to models with \(\beta=1\). Therefore, in what follows we avoid \(\beta=4\) for simplicity of exposition. For \(\beta=1, 2\) the matrix elements on the diagonal or above it are independent, and the basic covariances are given by \[\int dZ P_2(Z) z_{ij}\bar{z}_{km}=\frac{1}{\Omega_2}\delta_{ik}\delta_{jm},\] and \[\int dZ P_1(Z) z_{ij}\bar{z}_{km}=\frac{1}{2\Omega_1}(\delta_{ik}\delta_{jm}+\delta_{im}\delta_{jk}).\] Wick's rule leads to an elegant diagrammatical formulation of integrals. Matrix elements coming from traces like \({\rm Tr}(ZZ^\dagger)^{k}\) are arranged around vertices, and calculation of basic covariances are represented by edges. The sum over pairings becomes a sum over diagrams or maps. Because every edge must involve one \(z\) and one \(\overline{z}\), we say they are directed, and by convention they go "from" \(z\) "to" \(\overline{z}\). When \(Z\) is symmetric, edges may be twisted, resulting in a map that cannot be embeded in a orientable space. This kind of combinatorics has been extensively investigated. Complex hermitian matrices (Gaussian Unitary Ensemble) give rise to orientable maps, but with undirected edges; generic complex matrices (Complex Ginibre Ensemble) involve orientable and directed maps, which moreover are face-bicolored. For real matrices the maps are neither directed nor orientable, while face-bicoloring holds if the matrices are generic (Real Ginibre Ensemble) and does not hold if the matrices are symmetric (Gaussian Orthogonal Ensemble). The model we presently consider for \(\beta=1\), with complex symmetric matrices, does not seem to have attracted any attention in this context. It involves directed, non-orientable maps without face-bicoloring. Let \[Z_{\vec{i}\vec{j}}=\prod_k z_{i_kj_k}, \quad Z_{\vec{i}}=\prod_k z_{i_{2k-1}i_{2k}}.\] The integrals we are interested in are \[\label{I2} I_2(\vec{i},\vec{j},\vec{a},\vec{b})=\int dZ P_2(Z)e^{-\Omega_2\sum_{q=2}^\infty \frac{1}{q}{\rm Tr}(ZZ^\dagger)^q}Z_{\vec{i}\vec{j}}\overline{Z}_{\vec{a}\vec{b}},\] and \[\label{I1} I_1(\vec{i},\vec{j})=\int dZ P_1(Z)e^{-\Omega_1\sum_{q=2}^\infty \frac{1}{q}{\rm Tr}(ZZ^\dagger)^q}Z_{\vec{i}}\overline{Z}_{\vec{j}}.\] Diagrammatics arises when the exponential is expanded in a Taylor series, \[e^{-\Omega_\beta\sum_{q=2}^\infty \frac{1}{q}{\rm Tr}(ZZ^\dagger)^q}=\sum_{\lambda}\frac{(-\Omega_\beta)^{\ell(\lambda)}}{z_\lambda}\prod_{i=1}^{\ell(\lambda)}{\rm Tr}(ZZ^\dagger)^{\lambda_i},\] where the sum is over integer partitions \(\lambda=(\lambda_1,\lambda_2,\ldots)\) with no part equal to \(1\). The quantity \(z_\lambda\) equals \(\prod_i \lambda_i \widehat{\lambda_i}!\), where \(\widehat{\lambda_i}\) is the multiplicity of \(i\) in \(\lambda\). The quantity \[p_\lambda(ZZ^\dagger)=\prod_{i=1}^{\ell(\lambda)}{\rm Tr}(ZZ^\dagger)^{\lambda_i}\] is a power sum symmetric polynomial in the eigenvalues of \(ZZ^\dagger\). With this notation, we have \[I_\beta=\sum_{\lambda}\frac{(-\Omega_\beta)^{\ell(\lambda)}}{z_\lambda}J_{\beta,\lambda},\] with \[\label{J2} J_{2,\lambda}(\vec{i},\vec{j},\vec{a},\vec{b})=\int dZ P_2(Z)p_\lambda(ZZ^\dagger)\prod_{k=1}^n z_{i_kj_k}\overline{z}_{a_kb_k},\] and \[\label{J1} J_{1,\lambda}(\vec{i},\vec{j})=\int dZ P_1(Z)p_\lambda(ZZ^\dagger)\prod_{k=1}^n z_{i_{2k-1},i_{2k}}\overline{z}_{j_{2k-1},j_{2k}}.\] Using Wick's rule, this leads to a diagrammatic formulation for the integrals, with diagrammatic rules that are a multiplicative combination of weights as follows: The matrix elements in the integrand are vertices of valence 1; A trace \({\rm Tr}(ZZ^\dagger)^q\) is a vertex of valence \(2q\); To every vertex of valence \(2q\) we associate a factor \((-\Omega_\beta)\); Every Wick contraction of a \(z\) with a \(\overline{z}\) is a edge; To every edge we associate a factor \((2\Omega_\beta/\beta)^{-1}\); Edges may be twisted if \(\beta=1\); To every closed cycle we associate a factor \(d\). Notice that, due to the last rule, the contribution of a diagram is proportional to \(d^c\), with \(c\) the number of closed cycles it contains. Diagrams need not be connected. A diagram coming from a certain \(\lambda\) will contribute to order \(\Omega_\beta^{-(n+|\lambda|-\ell(\lambda))}\). The quantity \(|\lambda|-\ell(\lambda)=r(\lambda)\) is called the rank of the partition \(\lambda\). ## Examples Let us consider as an example \(I_1([i_1,i_2],[j_1,j_2])\). The simplest diagrams have no vertices of even valence (\(\lambda\) being the empty partition), and consist of a single edge connecting the two vertices of valence one, as in Figure 1. There are two of these, one taking \(i_1\) to \(j_1\) and the other being twisted in order to take \(i_1\) to \(j_2\). Together they contribute \[J_{1,\emptyset}([i_1,i_2],[j_1,j_2])=\frac{1}{2\Omega_1}(\delta_{i_1j_1}\delta_{i_2j_2}+\delta_{i_1j_2}\delta_{i_2j_1}).\] Since \(\lambda\) cannot have parts equal to \(1\), the next simple case is \(\lambda=(2)\). Use of a computer algebra system shows that \[J_{1,(2)}([i_1,i_2],[j_1,j_2])=\frac{1}{(2\Omega_1)^3}(2d^3+4d^2+10d+8)(\delta_{i_1j_1}\delta_{i_2j_2}+\delta_{i_1j_2}\delta_{i_2j_1}),\] meaning that, containing one vertex of valence \(4\), there are: \(2\) diagrams with three closed cyles, \(4\) diagrams with two closed cycles, 10 diagrams with one closed cycle and 8 diagrams with no closed cycles. We show four of these diagrams in Figure 2. Dashed lines represent Wick connections. Diagram a) has no closed cycles, diagram b) has one closed cycle, diagram c) has two closed cycles, diagram d) has three closed cycles. Clearly, \[J_{1,\lambda}([i_1,i_2],[j_1,j_2])=\frac{1}{(2\Omega_1)^{|\lambda|+1}}j_{1,\lambda}(d)(\delta_{i_1j_1}\delta_{i_2j_2}+\delta_{i_1j_2}\delta_{i_2j_1}),\] where \(j_{1,\lambda}(d)\) is a polynomial in the dimension \(d\). The first such polynomials are, for partitions of rank 2: \[j_{1,(3)}=5d^4+ 16d^3+49d^2 +74d +48,\] \[j_{1,(2,2)}=4d^6+ 16d^5+92d^4 +224d^3 +412d^2+ 688d+ 384;\] for partitions of rank 3: \[j_{1,(4)}=14d^5+ 64d^4+242d^3 +528d^2+ 688d+ 384,\] \[j_{1,(3,2)}=10d^7+ 52d^6+356d^5 +1180d^4 +3410d^3+ 6568d^2+ 7624d+3840.\] \[j_{1,(2,2,2)}=8d^9+48d^8+432d^7+1744d^6+7704d^5+21568d^4+52912d^3+92992d^2+99072d+46080.\] We notice that \(j_{1\lambda}(d)\) is of order \(d^{|\lambda|+\ell(\lambda)}\), and the coefficient of the largest power is a product of Catalan numbers, \[j_{1\lambda}=\prod_{i=1}^{\ell(\lambda)}\frac{1}{\lambda_i+1}{2\lambda_i \choose \lambda_i}.\] The proof of this fact is analogous to the complex hermitian case. In Figure 3 we show a diagram that contributes to \(J_{1,(4,3)}\) with \(n=3\). It has one vertex of valence 6 and one vertex of valence 8. It has two closed cycles. Two of its edges are twisted. Its contribution is proportional to \(\delta_{i_1j_1}\delta_{i_2j_4}\delta_{i_3j_2}\delta_{i_4j_6}\delta_{i_5j_5}\delta_{i_6j_3}\). # Moments of Circular Ensembles We now show how the Gaussian matrix models on the previous Section are related to the Circular Ensembles. Let \(S\) be in \(C\beta E(N)\) and let \(Z\) be its \(d\times d\) top left block. The distribution of \(Z\) is given by \[\frac{1}{\mathcal{V}(d,\beta)}\det(1-ZZ^\dagger)^{\frac{\beta}{2}(N-2d+1)-1},\] with \(\mathcal{V}(d,\beta)=\int dZ \det(1-ZZ^\dagger)^{\frac{\beta}{2}(N-2d+1)-1}\). This normalization can be computed by changing variables to the real positive eigenvalues of \(X=ZZ^\dagger\). The jacobian of this transformation is \(|\Delta(X)|^\beta\), which gives \[\mathcal{V}(d,\beta)=A_d\int dX \det(1-X)^{\frac{\beta}{2}(N-2d+1)-1}|\Delta(X)|^\beta,\] where \(A_d\) comes from integration over the eigenvectors. This is an integral of Selberg type, whose solution is \[\mathcal{V}(d,\beta)=\prod_{j=1}^d\frac{\Gamma(1+\beta(j-1)/2)\Gamma(\beta(N-d-j+1)/2)\Gamma(1+\beta j/2)}{\Gamma(1+\beta(N-j)/2)\Gamma(1+\beta/2)}.\] By definition, as long as all the indices in the matrix elements are smaller than \(d\), we can consider them either as elements of \(Z\) or of \(S\). This means that \[\label{theu} \frac{1}{\mathcal{V}(d,2)}\int dZ \det(1-ZZ^\dagger)^{(N-2d)}Z_{\vec{i}}\overline{Z}_{\vec{j}}=\langle S_{\vec{i}\vec{j}}\overline{S}_{\vec{a}\vec{b}}\rangle_{CUE(N)},\] and \[\label{theo} \frac{1}{\mathcal{V}(d,1)}\int dZ \det(1-ZZ^\dagger)^{\frac{1}{2}(N-2d+1)-1}Z_{\vec{i}}\overline{Z}_{\vec{j}}=\langle S_{\vec{i}}\overline{S}_{\vec{j}}\rangle_{COE(N)},\] and these quantities, after the integrals have been calculated, are actually independent of \(d\). Using \(\det=e^{{\rm Tr}\log}\) we can write, as long as \(f\) only involves elements of \(S\) inside \(Z\), \[\langle f(S,\overline{S})\rangle_{C\beta E(N)}=\frac{1}{\mathcal{V}(d,\beta)}\int dZ e^{-\omega_\beta{\rm Tr}(ZZ^\dagger)}e^{-\omega_\beta\sum_{q=2}^\infty \frac{1}{q}{\rm Tr}(ZZ^\dagger)^q}f(Z,\overline{Z}),\] with \(\omega_\beta(d)=\frac{\beta}{2}(N-2d+1)-1\). This looks very similar to the Gaussian models, Eqs. ([\[I2\]](#I2){reference-type="ref" reference="I2"}) and ([\[I1\]](#I1){reference-type="ref" reference="I1"}), but not quite identical, in particular because the normalizations are different, \(\mathcal{V}(d,\beta)\neq\mathcal{G}(d,\beta)\). The Gaussian normalization is \[\mathcal{G}(d,\beta)=A_d\int dX e^{-\Omega_\beta{\rm Tr}X}|\Delta(X)|^\beta,\] which equals \[\mathcal{G}(d,\beta)=A_d\Omega_\beta^{-\beta d(d-1)/2-d}\prod_{j=1}^d\frac{\Gamma(1+\beta(j-1)/2)\Gamma(1+\beta j/2)}{\Gamma(1+\beta/2)}.\] So \[\frac{\mathcal{V}(d,\beta)}{ \mathcal{G}(d,\beta)}=\Omega_\beta^{\beta d(d-1)/2+d}\prod_{j=1}^d\frac{\Gamma(\beta(N-d-j+1)/2)}{\Gamma(1+\beta(N-j)/2)}.\] Now, the crucial observation is this: it turns out that if we let \[\label{d} d\to 1-2/\beta,\] we get \[\frac{\mathcal{V}(d,\beta)}{ \mathcal{G}(d,\beta)}\to \Omega_\beta^0\prod_{j=1}^d1=1,\] and then we do arrive at the Gaussian model, with \[\Omega_\beta=\omega_\beta(1-2/\beta)=\frac{\beta}{2}(N-1)+1.\] Therefore, we conclude that moments of circular ensembles can be computed using the diagrammatic rules associated with Gaussian models, given in Eqs. ([\[J2\]](#J2){reference-type="ref" reference="J2"}) and ([\[J1\]](#J1){reference-type="ref" reference="J1"}), provided we use \(\Omega_\beta\) as above and set \(d=1-2/\beta\) at the end of the calculation. The map ([\[d\]](#d){reference-type="ref" reference="d"}) gives \(d=-1\), \(0\), \(1/2\) for \(\beta=1\), \(2\), \(4\), respectively. These are not positive integers as one would expect of a dimension. But these values must be used only after the result has been computed for formal \(d\) and found as a polynomial in \(d\). That this approach works for the \(CUE\) has already been discussed in the physics context. In that case, taking \(d\to 0\) ruled out the presence of closed cycles in the diagrams. That does not happen for the \(COE\), in which case the theory prescribes that the contribution of a diagram with vertex structure \(\lambda\) and \(c\) closed cycles is proportional to \[\frac{1}{z_\lambda}\left(\frac{-1}{2}\right)^{\ell(\lambda)}\frac{(-1)^c}{(N+1)^{|\lambda|-\ell(\lambda)+1}}.\] ## Example Let us take again the simplest example, \[\langle S_{i_1i_2}\overline{S}_{j_1j_2}\rangle_{COE(N)}=\frac{\delta_{i_1j_1}\delta_{i_2j_2}+\delta_{i_1j_2}\delta_{i_2j_1}}{N+1}.\] The trivial diagram with empty \(\lambda\) in the Gaussian model, Figure 1, already agrees with the exact result. Therefore, all other diagrams must cancel out and give a vanishing overall contribution, something which is not at all trivial. But indeed, we do have \(j_{1,(2)}(d=-1)=0,\) so the first correction vanishes. We have \(j_{1,(3)}(-1)=12\), and \(j_{1,(2,2)}(-1)=64\). Taking into account the factor \(\frac{1}{z_\lambda}\left(\frac{-1}{2}\right)^{\ell(\lambda)}\) we get \[-\frac{12}{6}+\frac{64}{12}=0,\] so the second correction vanishes. And we have \(j_{1,(4)}(-1)=32,\) \(j_{1,(3,2)}(-1)=240\) and \(j_{1,(2,2,2)}(-1)=2304,\) which leads to a vanishing third correction: \[-\frac{32}{8}+\frac{240}{24}-\frac{2304}{384}=0.\] # The semiclassical diagrammatics Curiously, another diagrammatical formulation of moments in \(C\beta E(N)\) already exists. It was developed by physicists working with semiclassical path integral approximations in quantum chaos. This was indeed the original motivation for the present work. The semiclassical diagrammatics is as follows: any given moment is written as a sum over diagrams, with weights of \(-N\) for each vertex of valence larger than one, \(1/N\) for each edge, with orientability being required when \(\beta=2\). Closed cycles are not allowed. These rules are exactly the ones we obtained, when \(\beta=2\). So our Gaussian model coincides with the semiclassical model of quantum chaos. However, this is not so for \(\beta=1\). There are two differences in the rules themselves, because in our model the weights are \(-(N+1)\) for each vertex of valence larger than one and \(1/(N+1)\) for each edge. There is also another difference, which is that in our model the edges are directed, while this does not hold in the semiclassical model. For comparison, let us sketch the calculation of \(\langle S_{i_1i_2}\overline{S}_{j_1j_2}\rangle_{COE(N)}\) according to the semiclassical model. The leading order is given by the diagrams in Figure 1, except now they contribute \(1/N\). Then there are diagrams like the ones in Figure 4. Their contributions are: a) \(-1/N^2\) (one vertex of valence 4); b) \(-1/N^3\) (one vertex of valence 6); c) \(1/N^3\) (two vertices of valence 4). We have just found the first three terms in the \(1/N\) expansion of \(1/(N+1)\). When all possible diagrams are summed, the exact result is recovered in the form of the geometric series. Notice how in the diagrammatics obtained in the present work, the same calculation requires a single diagram, with the contribution from all others being zero due to nontrivial cancellations, just like happens for \(CUE(N)\). In contrast, in the semiclassical diagrammatics it is the opposite: infinitely many diagrams are required. Actually, the semiclassical diagrammatics for \(\beta=1\) can be related to matrix integrals involving generic real matrices. This is related to the fact that the Weingarten function of \(COE(N)\) is actually equal to the Weingarten function of the real orthogonal group in dimension one higher, \(O(N+1)\). This equality was established by Matsumoto, who pointed out that it must be grounded in the fact that \(COE(N)\sim U(N)/O(N)\), but still remarked that it was "quite mysterious". A Gaussian model for the combinatorics of orthogonal group moments in terms of generic real matrices was indeed derived in. # Application: traces of submatrices In, Jiang and Matsumoto studied the traces of matrices from the \(COE(N)\). In particular, they showed that \[\lim_{N\to \infty}\langle |p_\lambda(Z)|^2\rangle_{COE(N)}=2^{\ell(\lambda)}z_\lambda.\] They actually derive an exact formula for the above quantity, but it is not quite explicit because it depends on Jack characters, the coefficients in the expansion of power sums into Jack polynomials. They also show that \(\langle p_\lambda(Z)p_\mu(\overline{Z})\rangle_{COE(N)}\) decays like \(1/N\) if \(|\mu|=|\lambda|\) but \(\mu\neq \lambda\). As an application of the results we have obtained, we compute averages of traces of a submatrix \(Z\) of dimension \(M\) inside a \(COE(N)\) matrix. Results from correspond to the particular case \(M=N\). Let \(B(M)\) denote the \(M\times M\) upper left block of \(COE(N)\) matrices. When we wish to compute something like \(\langle |p_\lambda(Z)|^2\rangle_{B(M)}\) using our Gaussian model, we write \[\langle p_\lambda(Z)\rangle_{B(M)}=\sum_{i_1...,i_n=1}^M\left\langle\prod_{k=1}^n Z_{i_k,i_{\pi(k)}}\right\rangle_{COE(N)},\] where \(\pi\) is any permutation with cycle type \(\lambda\), then we use the diagrammatic rules of Section 3 to compute the average, and finally we sum over \(\vec{i}\). In this way, closed cycles associated with \(\vec{i}\) indices have a weight of \(M\), while closed cycles within the average have weight \((-1)\). This calculation is carried out in a computer algebra system, so only the simplest partitions can be addressed. Averages of traces of submatrices were studied for the unitary group in, and averages of their Schur polynomials appear in. These approaches rely on the usual characters of the permutation group, which are readly available in computer algebra systems, so we do not address the \(\beta=2\) case. The simplest average, \[\langle |p_{(1)}(Z)|^2\rangle=\frac{2M}{N+1},\] is actually exact and simple to obtain. For \(n=2\) we obtain two corrections to the leading order results, \[\langle |p_{(2)}(Z)|^2\rangle=\frac{4M(M+1)}{(N+1)^2}-\frac{4M(M+3)}{(N+1)^3}+\frac{2M(M+15)}{(N+1)^4}+\cdots\] and \[\langle |p_{(1,1)}(Z)|^2\rangle=\frac{8M^2}{(N+1)^2}-\frac{16M}{(N+1)^3}-\frac{8M(3M-7)}{(N+1)^4}+\cdots.\] Notice that the large-\(N\) asymptotics may be rather rich. \(\langle |p_{(2)}(Z)|^2\rangle\) equals, to leading order, \(8M(M+1)/N^2+O(N^{-3})\) if \(M\) is held fixed, \(4+O(N^{-2})\) if \(M=N\) and \(4 In constrast with the\)CUE\(, quantities like\)p\_(Z)p\_()\_COE(N)\(do not necessarily vanish when\[. For example, \begin{equation} \langle p_{(2)}(Z)p_{(1,1)}(\overline{Z})\rangle=\frac{8M}{(N+1)^2}-\frac{8M(M+1)}{(N+1)^3}+\frac{4M(7M+1)}{(N+1)^4}+\cdots.\end{equation} Again, the asymptotical behavior depends on how\)M\(related to\)N\(. The above quantity becomes\)`<!-- -->`{=html}8M/N\^2+O(N\^-3)\(if\)M\(is held fixed,\)`<!-- -->`{=html}20/N\^2+O(N\^-3)\(if\)M=N\(and\)`<!-- -->`{=html}8 When \(n=3\) we obtain one correction to the leading order results, resulting in \]\langle |p_{(3)}(Z)|^2\rangle=\frac{6M(M^2+3M+4)}{(N+1)^3}-\frac{18M(M^2+7M+8)}{(N+1)^4}+\cdots,\[ \]\langle |p_{(2,1)}(Z)|^2\rangle=\frac{8M(M^2+M+4)}{(N+1)^3}-\frac{8M(M^2+19M+16)}{(N+1)^4}+\cdots,\[ and \]\langle |p_{(1,1,1)}(Z)|^2\rangle=\frac{48M^3}{(N+1)^3}-\frac{288M^2}{(N+1)^4}+\cdots.\[ The off-diagonal covariances in this case are \]\langle p_{(3)}(Z)p_{(2,1)}(\overline{Z})\rangle=\frac{24M(M+1)}{(N+1)^3}-\frac{24M(M^2+4M+7)}{(N+1)^4}+\cdots,\[ \]\langle p_{(3)}(Z)p_{(1,1,1)}(\overline{Z})\rangle=\frac{48M}{(N+1)^3}-\frac{144M(M+1)}{(N+1)^4}+\cdots,\[ and \]\langle p_{(2,1)}(Z)p_{(1,1,1)}(\overline{Z})\rangle=\frac{48M^2}{(N+1)^3}-\frac{48M(M^2+M+4)}{(N+1)^4}+\cdots.\[ # Conclusion We have found Gaussian matrix models that provide novel diagrammatic rules for the calculation of moments in circular ensembles, \(C\beta E(N)\). The matrices involved are generic complex for \(\beta=2\), complex symmetric for \(\beta=1\) and complex self-dual for \(\beta=4\), and their dimension must be set to \(1-2/\beta\). The result is an expansion in inverse powers of \(N\) for \(\beta=2\), of \(N+1\) for \(\beta=1\) and of \(2N-1\) for \(\beta=4\). In view of the fact that \]\label{iden} {\rm Wg}^{COE(N)}={\rm Wg}^{O(N+1)},\quad {\rm Wg}^{CSE(N)}={\rm Wg}^{Sp(N-1/2)},\(\) our results lead to expansions in inverse powers of \(N\) for the moments of \(U(N)\), \(O(N)\), and \(Sp(N)\). The complex symmetric matrix model we developed here for \(COE(N)\) must therefore give the same results as the generic real matrix model developed in for \(O(N+1)\) and the corresponding semiclassical model. However, this equality between very different matrix models is not obvious a priori and a direct proof would be quite interesting. It might shed some light on the identities ([\[iden\]](#iden){reference-type="ref" reference="iden"}), which at the moment are still little more than coincidences.

{'timestamp': '2023-02-10T02:01:19', 'yymm': '2302', 'arxiv_id': '2302.04311', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04311'}

# Introduction Earthquakes pose a serious threat to human life and property. There are countless examples of the devastating power of major earthquakes, and thus, earthquake mitigation efforts have been intensive. The use of building standards and social strategies has been adopted to minimize earthquake damage, especially in tectonically active areas. A reliable method of predicting earthquakes, using an observable signal, has not yet been developed . The current earthquake forecasting is based on known empirical seismic laws. An active direction for earthquake forecasting is using statistical models, such as the Epidemic-Type Aftershocks Sequence (ETAS) model , which aims to determine the aftershock rate after a mainshock. Such models are based on empirical earthquake laws, including the Gutenberg-Richter law and the Omori-Utsu law. These models, however, are not capable of forecasting mainshocks, and their aftershock prediction success is limited. It is clear that even a small improvement in the forecasting skill of major earthquakes and aftershocks is critical. The Omori-Utsu law is an empirical law that determines the rate of events (aftershocks) after a mainshock. According to this law, the temporal decay of the aftershock rate is expressed as \[r(t) = \frac{k}{(c+t)^p}\] where \(t\) is the time after a mainshock, \(c\) is a case-dependent time scale, \(k\) is a productivity parameter that depends on the magnitude of the mainshock, and the power \(p\) is a parameter that quantifies the decay rate where typically it is assumed that \(p\approx1\). According to the Gutenberg-Richter law, the distribution of earthquake magnitudes decays exponentially. A generalized version of the Omori-Utsu law that combines the known statistical laws for the earthquakes of Båth, Gutenberg-Richter, and the Omori-Utsu has been proposed and tested by on four strong mainshocks in Southern California. Here, we will focus on the standard version of this law. A less familiar version of the Omori law is the inverse Omori law which states that the rate of earthquakes increases prior to mainshocks. Those events are determined a posteriori (as foreshocks), and are also referred to as an anomalous foreshock sequence. In principle, such behavior could be used for forecasting strong earthquakes. There is still a controversy regarding the reason for higher event rates before mainshocks. Two debated conceptual hypotheses about earthquake nucleation, the "pre-slip model" and the "cascade model," offer insight into the role played by foreshocks in earthquake predictability. In the former model, foreshocks are triggered by an aseismic slip anticipating the mainshock, while in the latter model, foreshocks are like any other earthquake that triggers one another to ultimately become larger (the mainshock). However, despite remarkable achievements with regard to foreshock rate, an increased earthquake rate prior to large earthquakes has not always been found, limiting its use for forecasting. For example, a recent study by showed that high heat flow regions in Southern California are associated with elevated foreshock rates. However, since it is typically correlated with small mainshock magnitudes, it is less useful for forecasting major events. Here, we aim to revisit this observation based on a high-resolution earthquake catalog. Such a detailed catalog has recently been constructed based on machine learning algorithms. Such earthquake catalogs may improve earthquake forecasting; we aim to test the validity of the inverse Omori law and the conventional Omori-Utsu law based on a detailed earthquake catalog and to determine whether such catalogs can improve the forecasting of impending earthquakes. There are two main goals in our study: a) to study the rate of earthquakes (foreshocks) before mainshocks, and b) to study the rate of earthquakes (aftershocks) after a mainshock to identify the range of validity of the Omori-Utsu law. For this purpose, we use and compare both the detailed and less detailed catalogs of the Southern California region. We first develop a new technique to identify mainshocks and foreshocks. Then, we determine whether the number of events in a fixed time window, \(D_{\mathrm{fs}}\), is significantly higher than the corresponding number in surrogate earthquake catalogs. We also test whether the higher rate found in the data is similar to or higher than that of the ETAS model. In addition, we test the validity of the Omori-Utsu law for aftershocks, both in real catalogs and ETAS model catalogs. Our results suggest that the Omori-Utsu law may be valid statistically, but does not hold for most mainshocks. # Methodology and Data ## Determining a mainshock and defining the aftershock period We first develop a method to identify mainshocks based on their magnitude and the earthquake rate following large events. The specific steps are described below. We consider earthquake events with magnitudes larger than or equal to a minimum magnitude threshold, \(m_c\); the event dates are marked with \((d_1,d_2,..., d_k)\), with their corresponding event magnitudes \((m_1,m_2,..., m_k)\), where \(k\) is the number of events above the magnitude threshold. Next, we calculate the average daily event rate, \(\langle r_d \rangle = \frac{k}{T}\) where \(T = d_k-d_1\) is the time span of the catalog. Then, we set another threshold, \(M_{th}\), typically \(M_{th}=4\) or \(M_{th}=4.5\), and sort all the events with magnitudes greater than or equal to \(M_{th}\), to obtain a sorted list of events \((m_{p_1}\geq m_{p_2}\geq..., \geq m_{p_r})\). These events are now referred to as potential mainshocks. We consider the events starting with the highest magnitude and ending with the lowest. We then calculate the rate of post-mainshock events (now referred to as possible aftershocks) and determine when the daily rate has returned to its baseline value (usually chosen to be the mean earthquake rate) over a 10-day moving window. We refer to the events during this time period as aftershocks and remove them. Sometimes mainshocks fall within the aftershock period and, thus, are deleted from the list of potential mainshocks. In some cases, a mainshock \(M_j\) falls in the aftershock window where its magnitude is larger than the magnitude of the mainshock under consideration \(M_i\); i.e., \(M_j>M_i\). In this case, there are two options: 1) The mainshock with the larger magnitude, \(M_j\), is deleted as it can be considered as an aftershock; this case is referred to as a foreshock flag = 1. 2) Earthquake \(i\) is considered a foreshock, and its corresponding aftershocks remain unaffected; this case is referred to as a foreshock flag = 0. There were only a few such cases, and the foreshock flag did not strongly affect the results described below. After we repeat this process for all the possible mainshocks, we are left with a catalog like the one shown in Fig. [\[fig:Fig1\]](#fig:Fig1){reference-type="ref" reference="fig:Fig1"}(d), i.e., a set of identified mainshocks. We refer to the aftershock period of mainshock \(i\) as \(D_{\mathrm{as}_i}\). We perform the analysis of the rate of earthquakes before mainshocks on the filtered catalog. ## Finding a significant number of events prior to mainshocks {#finding-a-significant-number-of-events-prior-to-mainshocks .unnumbered} After identifying the mainshocks, we count the number of events prior to the mainshock during a varying foreshock duration \(D_{\mathrm{fs}}\). Our null hypothesis is that the earthquake rate prior to a mainshock is not significantly different than the overall rate of the "cleaned" catalog---the null hypothesis is rejected when the actual rate is significantly different than the mean rate of the cleaned catalog. To estimate whether and by how much the rate prior to the mainshocks is larger than the mean rate, we calculate the rate prior to randomly selected times (during the time range spanned by the catalog). We then construct the probability density function (pdf) and the cumulative density function (cdf) of these rates. Then, we calculate how many real events fall above a certain percentile (for example, 95%) of the surrogate data. If the ratio of real events above the specific percentile is more than the expected one (based on the surrogate data), we can conclude that the null hypothesis is rejected and that the rate of events preceding the mainshock is indeed larger than expected; this can have predictive power. We repeated this procedure using different percentiles, such as 50% (median), 80%, 95%, and 99%. See Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"} for an example of the results for a time window of \(D_{\mathrm{fs}}=3\) days. ## Earthquake catalogs In this study, we analyze two catalogs, the and the catalogs. The Hauksson et al. catalog can be obtained from the Southern California Earthquake Data Center (see the Availability of data and materials section). It spans 40 years, 1981-2019, and is considered to be complete for magnitudes \(m\geq2.5\). It contains \(N = 45843\) events with magnitudes larger than or equal to the magnitude threshold of 2.5. The catalog covers the area of Southern California, between 113.77\(^\circ\)W-122.16W and 30.16\(^\circ\)N--37.51\(^\circ\)N. The Ross et al. catalog spans a 10-year period from 01-Jan-2008 to 31-Dec-2017 and is nearly complete for magnitudes above 0.3. To ensure completeness, we chose the magnitude threshold to be \(m_c=1\) and excluded all events below this magnitude, ending up with \(N = 171,296\) events. The catalog covers the area of Southern California, between 31.9\(^\circ\)N--37.03\(^\circ\)N and 114.81\(^\circ\)W--121.71\(^\circ\)W. ## The ETAS Model The ETAS model, first proposed by, describes earthquake aftershocks based on the analogy of an epidemic. The "infected" individual (or the immigrant) enters the observation region and spreads the. Immigrants are analogous to background events, and their offspring to aftershocks. In this sense, the model can also be thought of as an ancestor-offspring system. Likewise, these offspring can reproduce (i.e., spread infection). After most of the population has been infected or died, there are insufficient receptive individuals for the epidemic to continue. According to the seismological analogy, this would be equivalent to an aftershock sequence (such as an immigrant family) initiated after a background event (such as an ancestor) had died out. As of now, the ETAS model is one of the most common operational earthquake models. The ETAS model assumes that seismic events involve a stochastic point process in space-time. Each event above a magnitude \(M_0\) is selected independently from the Gutenberg--Richter distribution (where b = 1). The conditional rate, \(\lambda\), at location \((x, y)\) at time \(t\) is given by \[\lambda(x,y,t | H_t) = \mu(x,y) + \sum_{t<t_i} k(M_i)\exp(\alpha(M_i-M_0))/(t-t_i + c)^p\] where \(c\) and \(p\) are parameters in the Omori-Utsu law, \(k\) and \(\alpha\) control the aftershock productivity by a mainshock and its magnitude sensitivity, respectively. These five parameters, \(\mu, c, p, k\) and \(\alpha\), can be determined using the Maximum Likelihood Estimation (MLE) with an earthquake catalog consisting of hypo-central times \(t_i\) and magnitudes \(M_i\) which is denoted by \(H_t\). The estimated parameters are given in Table [\[Tab:ETAS_params\]](#Tab:ETAS_params){reference-type="ref" reference="Tab:ETAS_params"}. # Results **Number of events prior to a mainshock compared with surrogate catalogs.** We systematically study the number of recorded events \(n\) (foreshocks) prior to identified mainshocks using different time windows of \(D_{\mathrm{fs}} = 1,2,3,5,10,20,30,40\) days. For each mainshock \(i\), let \(n_i\) be the number of foreshocks that occurred during the time duration \(D_{\mathrm{fs}_i}\) prior to mainshock \(i\). Using the surrogate test described above, we compare the distributions of the real and the surrogate counts and test whether the counts before the mainshocks are larger than the corresponding counts of the control surrogate data. In addition to the surrogate data test, we use the Poisson distribution to estimate the significance of the foreshock counts in the cleaned catalog; a Poisson process is often used to model the earthquake background rate. A significant number of events above the level of the expected percentile indicates that the null hypothesis that the earthquake rate prior to the mainshock is similar to the mean catalog rate is rejected, suggesting an increased seismic activity prior to strong events, in comparison to an uncorrelated Poisson process. The Poisson distribution is given by \(P(X=k) = \frac{\lambda^k e^{-\lambda}}{k!}\) where \(\lambda\) is the mean of the entire (cleaned) catalog. In Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"}(a), we show that for fixed time window \(D_{\mathrm{fs}}=3\) days, most of the mainshocks are preceded by more than 100 events in 3 days, with an average of \(166\) events. This number is significantly higher than the number we obtained for the surrogate events (mean value of 102 events); i.e., \(\approx\)`<!-- -->`{=html}60% increase. The results for \(D_{\mathrm{fs}}=1,2,5,10\) days yielded qualitatively similar results. To evaluate the significance of the results, we calculated the \(50\%, 80\%,95\%\), and \(99\%\) quantiles of the surrogate results distribution. In Fig. [\[fig:Fig2\]](#fig:Fig2){reference-type="ref" reference="fig:Fig2"}, we present the summary of results for the Ross et al. and Hauksson et al. catalogs, respectively, where the dashed vertical lines represent the different quantiles, and the colored \"x\" represents the number of events prior to the different mainshocks. The pdf of the surrogate data is marked by the gray bars, and the gray circles mark the cdf of the surrogate data. After calculating the quantiles, we examined how many events fall above each quantile. For example, when considering mainshocks with a magnitude larger than or equal to 4, using the Ross et al. catalog, 89 events out of 92 potential mainshocks were identified as mainshocks (while the other three are now considered as aftershocks). For the \(99\%\) quantile of the surrogate data, we counted three events with more foreshocks than the \(99\%\) of the random events; i.e., 3/89=3.4% real mainshocks, over three times the expected 1% based on the surrogate data. For the 95% quantile, 10 out of 89 detected mainshocks fell above or equal to the quantile, i.e., 10/89 = 11%, which is more than twice as many events as expected \(5\%\). We summarize the results of the Ross et al. catalog in Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"}(a). It is clear that for a time window of fewer than five days, there are 2--3 times more events prior to the mainshocks in comparison to the surrogate data results, for the higher quantiles. This indicates that the null hypothesis that the earthquake rate prior to mainshocks is similar to the mean rate is rejected and that earthquake activity increases prior to mainshocks. We also analyzed the Hauksson et al. catalog whose magnitude threshold is 2.5. Here (Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"}b) the percentages of events above or equal to the \(95\%\) and \(80\%\) quantiles are 0.06 and 0.25, respectively, much lower than the fraction obtained for the Ross et al. catalog (Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"}a). We attribute the more significant results of the Ross et al. catalog to the lower magnitude threshold that provides many more events, thus enabling easy detection of the increased earthquake rate prior to mainshocks. **Comparison against the ETAS model.** We generated 10 synthetic EATS model catalogs, using the parameters fitted for the Hauksson et al. and Ross et al. catalogs, and using the MLE; see Table [\[Tab:ETAS_params\]](#Tab:ETAS_params){reference-type="ref" reference="Tab:ETAS_params"} and <https://www.kaggle.com/code/tokutani/etas-model-analysis-for-japan-region>. Next, we tested our method on these synthetic catalogs, and the results are summarized in Fig. [\[fig:Fig3\]](#fig:Fig3){reference-type="ref" reference="fig:Fig3"}c,d. In general, the probabilities in the ETAS model are smaller than those of the real data, indicating that some earthquake processes are missing in the ETAS model and that the current ETAS model formulation cannot fully reproduce the increased activity prior to mainshocks. We note that the results of the 99% quantile of the real data may be more uncertain due to the few events that fell above this quantile. We performed the same analysis when foreshock flag=1 and obtained similar results; see Fig. [\[fig:SI_Fig1_Omori\]](#fig:SI_Fig1_Omori){reference-type="ref" reference="fig:SI_Fig1_Omori"}. This indicates that the strict definition of foreshocks within the set of potential mainshocks is not important. # Discussion In our study, we focus on the inverse Omori law and show that indeed, statistically, the earthquake rate is significantly larger than usual prior to mainshocks. Furthermore, we analyze the conventional Omori-Utsu law and summarize the results in the Supplementary Material. We show that the Omori-Utsu law does not hold for the vast majority of mainshocks. However, an analysis of ETAS model synthetic catalogs yields results similar to those for the real catalogs. Since the ETAS model is based on the Omori-Utsu law and since this law is valid only for a relatively small portion of the events, we conclude that the Omori-Utsu law is valid in a statistical sense and probably underlies the rate of aftershocks after mainshocks. In summary, we mainly examine the inverse and, to a lesser degree, the conventional versions of the Omori-Utsu law, using a highly detailed catalog. We also develop and implement an algorithm for mainshock detection. The algorithm is composed of sorting the earthquakes according to their magnitude and deleting the aftershocks appearing between the mainshock and the time point at which the earthquake rate returns to the mean catalog rate. This algorithm also allows us to define the foreshocks. We then propose a surrogate test to investigate whether mainshocks are often preceded by more events than occur during random time windows. This is done by comparing the earthquake rate within a certain time window prior to a mainshock to the earthquake rate of randomly selected time windows over the entire "cleaned" catalog. By analyzing two catalogs that differ in their magnitude completeness, we show that the more detailed catalog resulted in a greater probability for higher earthquake rates prior to mainshocks, sometime three times higher than expected. \[Note that it is possible, in principle, to analyze the Ross et al. catalog, instead of the Hauksson et al. catalog, when setting the magnitude threshold to 2.5 instead of 1. However, this will result in a catalog with a small number of events as this catalog spans 10 years, while the catalog of Hauksson et al. spans 40 years of events, thus allowing a much better statistical analysis. This is why we use this catalog when dealing with a higher magnitude threshold.\] In principle, the increased earthquake rate prior to mainshocks can be used to provide some forecasting of a mainshock. For instance, if the probability to be above the 99% percentile of the surrogate data is 3% (instead of 1%), then an increased rate within the upper 1% indicates, with high probability (three times more than expected), a future major earthquake, within the time window of the analysis. An increased foreshock rate before a large seismic event may be associated with the accumulation of damage (stress) before failure (a quake). This is based on both theoretical and lab experiments. Foreshock activity may vary greatly from fault to fault depending on several factors, such as fault maturity and fault heat flux, reflecting the damage accumulation in each fault system prior to the mainshock. This is supported by theoretical studies that predict that there should be some degree of damage before the fault is activated. In addition, we tested and compared synthetic catalogs generated using the ETAS model. We show that this model does not fully reproduce the results obtained for real catalogs, suggesting that some key processes regarding the development of an earthquake are missing. Finally, we tested the conventional version of the Omori-Utsu law for aftershocks, which describes the rate of aftershocks after mainshocks. Our results suggest that this law appears to hold only for a few mainshocks, suggesting that models should not heavily rely on this law. ## Authors' contributions {#authors-contributions .unnumbered} Y.A. conceptualized, developed, and supervised the project. S.H supervised the project. E.E.A. implemented the methodology, performed the analysis, and prepared the figures. E.E.A. and Y.A. wrote the manuscript.

{'timestamp': '2023-02-10T02:01:57', 'yymm': '2302', 'arxiv_id': '2302.04326', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04326'}

# Introduction Active galactic nuclei (AGN) are considered crucial ingredients in baryonic structure formation in the Universe (e.g. @2003ApJ...599...38B [@2006MNRAS.365...11C; @2012ARA&A..50..455F]). Huge efforts have been invested in compiling large and complete AGN samples, including both high-luminosity quasars and low-luminosity Seyfert galaxies, to understand their physical properties and establish their role in galaxy evolution. Based on the distinctive characteristics of AGN (e.g. accretion rate and obscuration state), various techniques have been developed to search for AGN in different spectral bands, from hard X-rays (e.g. @2012ApJ...749...21A) to mid-IR (e.g. @2005ApJ...631..163S [@2012ApJ...753...30S]) and radio (e.g. @2004A&A...416...35W). The easiest systems to identify are the unobscured (type 1) AGN (e.g. @2010AJ....139.2360S), while the selection of obscured (type 2) AGN is more challenging because the bright nuclear emission is not seen. In the optical, narrow emission-line flux ratios and diagnostic diagrams represent a widely used method to identify luminous type 2 AGN in optical spectra. Among these, a commonly used diagnostic is the BPT diagram, based on the combined \[\]/H\(\beta\) and \[\]/H\(\alpha\) line ratios (with \[\]/H\(\alpha\) or \[\]/H\(\alpha\) as alternatives). Recently, an alternative diagnostics was presented by, the so-called diagram, where the \[\]\(\lambda\)`<!-- -->`{=html}5007Å  emission line flux is replaced by \(\lambda\)`<!-- -->`{=html}4686Å  (hereafter \[\] and , respectively) in the ratio with the H\(\beta\) line flux. Given the higher ionisation potential of (\(E_{\text{ion}}(He^{\text{+}})=54.4\) eV) compared to \[\] (\(E_{\text{ion}}(O^{\text{++}})=35.2\) eV), the diagnostics is more sensitive to AGN activity and better differentiates between ionisation due to AGN and star formation. In particular, it allows to identify low-luminosity AGN in SF galaxies, where the AGN emission may be overwhelmed by intense star formation and, therefore, missed by the standard BPT-based selection, as pointed out by. indeed used the diagram to search for nuclear activity in local galaxies in the Sloan Digital Sky Survey Data Release 7 (SDSS DR7; @2000AJ....120.1579Y [@2006AJ....131.2332G; @2009ApJS..182..543A]). Although the line emission is an optimal tracer of AGN activity, there are several limitations associated with its use. First of all, it is a faint emission line, which is the reason why the published studies on the topic are few and all limited to the Local Universe. Moreover, it can also be of stellar origin. The line is indeed frequently observed in HII regions, associated with young stellar populations and, in particular, with evolved and massive Wolf-Rayet (WR) stars. However, in presence of WR stars, the line appears blended with several metal lines, forming the so-called 'blue-bump' around \(\lambda\)`<!-- -->`{=html}4650Å . Indeed, several studies have also found evidence for substantial emission in low-metallicity, star forming dwarf galaxies. In the case of dwarf galaxies some authors have also proposed other ionising sources, such as X-ray binaries and fast radiative shocks. However, recent observations have revealed the presence of AGN in several dwarf galaxies, therefore indicating that AGN ionising photons may be responsible for the emission even in these systems. Most searches for AGN using optical spectroscopy have so far relied on single-fibre observations (such as the SDSS). As pointed out in, the small size of the optical fibres (\(\sim3^{\prime\prime}\) diameter) can cover only the central region of the galaxy, where the AGN might be hidden due to obscuring material or dominated by other ionisation processes (like SF), and may miss extended visible Narrow Line Regions (NLR). In consequence, the central AGN may not be detected in the central integrated spectrum of its host, leading to an incorrect classification of the galaxy as non-active. Thanks to integral field unit (IFU) observations, it is now possible to search for weak (less diluted) and off-centre AGN signatures, such as cases of AGN offset from the centre due to a recent merger with a non-active galaxy, or of a second off-centre AGN in dual AGN systems. Another possibility might also be a recently turned off AGN, whose relic ionisation remains visible at large distance from the centre. In this work, we combine the power of integral field spectroscopy in the MaNGA survey with the use the diagnostics to identify AGN host candidates. By comparing the AGN samples selected by the emission and the standard BPT diagrams, we find that the former is crucial to detect elusive AGN residing in highly SF galaxies, which are completely missed by the standard BPT classification, therefore increasing the total number of AGN in our parent sample. This paper is organised as follows. In Sect. [2](#sec:selection){reference-type="ref" reference="sec:selection"}, we briefly introduce the MaNGA survey and describe how we identify AGN host candidates starting from the two different spatially-resolved diagnostics, that are the and BPT diagrams. In Sect. [3](#sec:host_properties){reference-type="ref" reference="sec:host_properties"}, we compare the main physical properties of the AGN host galaxies selected by the two diagnostics, while in Sect. [4](#sec:discussion){reference-type="ref" reference="sec:discussion"} we discuss and quantify the impact of the detection limit in MaNGA on our overall AGN census. The conclusions from our study are finally drawn in Sect. [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}. Throughout this work we adopt a flat \(\Lambda\)CDM cosmology with \(\Omega_{\text{m,0}} = 0.3\), \(\Omega_{\Lambda,\text{0}} = 0.7\) and \(H_0 = 70\) km s\(^{-1}\) Mpc\(^{-1}\). # Sample selection and classification {#sec:selection} ## MaNGA data {#subsec:data} Our sample is based on the fifteenth data release (DR15) of the MaNGA survey (Mapping Nearby Galaxies at APO; @2015ApJ...798....7B [@2015AJ....149...77D; @2015AJ....150...19L; @2016AJ....152..197Y; @2017AJ....154...86W]). MaNGA is an optical fibre-bundle IFU spectroscopic survey, part of the fourth phase of the SDSS (SDSS-IV; @2017AJ....154...28B). The MaNGA DR15 catalogue consists of spatially resolved data of about 4600 local (\(z\sim0.03\)) galaxies with a spectral resolution \(R\sim3000\) over the wavelength range of \(\sim3600-10300\) Å , using multiple (from 19 to 127) fibre bundles. The median effective spatial resolution of MaNGA data is of \(2.54^{\prime\prime}\) full width at half maximum, corresponding to \(\sim2\) kpc at \(z\sim0.05\), and the pixel scale is of 0.5\(^{\prime\prime}\). In our analysis, we use spatially resolved data and measurements of galaxies (e.g. data and model cubes, emission-line flux maps) provided by the MaNGA data-analysis pipeline (DAP; @2019AJ....158..160B [@2019AJ....158..231W]), and take global properties (e.g. star formation rate, stellar mass) integrated within the field of view, from the MaNGA Pipe3D value added catalog. ## Spatially resolved emission line diagrams {#subsec:eldiag} With the aim of comparing the efficacy of BPT and diagrams in selecting AGN galaxies, we start dealing with spatially-resolved measurements of emission-line flux of all MaNGA spaxels. From the MaNGA DAP we select spaxels with a signal-to-noise ratio (S/N) higher than 3 in \[\], \[\], H\(\beta\) and H\(\alpha\) for the BPT diagram (BPT spaxels), whereas we require S/N \(>5\) in and S/N \(>3\) in \[\], H\(\beta\) and H\(\alpha\) for the diagram ( spaxels). We apply a higher S/N cut for the , compared to that for the other lines, in order to conservatively select only spaxels with a secure detection and to avoid cases of putative line emission that is actually consequence of a strong residual of the stellar continuum or to contamination from a foreground galaxy (these cases are discussed later in Sect. [2.3.1](#subsec:fake){reference-type="ref" reference="subsec:fake"}.) These S/N cuts lead to \(\sim\)`<!-- -->`{=html}2.1 million and \(\sim\)`<!-- -->`{=html}15200 spaxels for the BPT and diagrams, respectively (out of a total of more than 4.3 million MaNGA spaxels). The large discrepancy (i.e. \(\sim\)`<!-- -->`{=html}2 orders of magnitude) in the number of high-S/N spaxels resulting from the two separate S/N cuts (i.e. the BPT-selected versus --selected spaxels) reveals how faint and rare the detection of the emission line is, even in the spectra of local galaxies. To compute line ratios, we use dust-corrected emission line fluxes resulting from the Gaussian line modelling performed within the MaNGA DAP, except for a few galaxies which instead need a complete or partial spectral re-fitting. In the case of these objects, we take single-spaxel emission line fluxes obtained from our re-modelling (see Sect. [2.3.2](#subsec:refitting){reference-type="ref" reference="subsec:refitting"}). ### BPT and diagnostic diagrams {#subsec:bpt_and_heii} The top panels of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} show the \[\]/H\(\beta\) versus \[\]/H\(\alpha\) version of the BPT (left panel) and the (right) diagnostic diagrams of MaNGA spaxels. Regions in the two diagrams corresponding to different ionisation mechanisms are delimited by demarcation lines. In the BPT diagram, the Ke01 and Ka03 lines represent the theoretical extreme starburst line and the empirical SF line, respectively; the S07 line instead separates LINERs from AGN ionising sources according to. Similarly, we plot separating curves in the diagram as determined in: the solid and dashed lines are defined as the limiting curves above which more than 50% and 10% of the flux, respectively, is expected to come from an AGN. We classify spaxels as AGN-like only when they lie in the pure AGN regions of their respective diagrams: in the BPT, these are the spaxels falling above both the Ke01 and S07 lines, referred to as BPT AGN spaxels (empty/filled blue circles); while in the diagram, the AGN spaxels are those above the 50% AGN line (empty/filled red squares). Filled symbols represent BPT and AGN spaxels meeting the additional requirement of having S/N() \(>5\) and S/N(\[\]) \(>3\), respectively. Whereas there is a substantial difference in number between filled (\(\sim\)`<!-- -->`{=html}3440 spaxels) and empty (\(\sim\)`<!-- -->`{=html}37700 spaxels) blue circles, almost all spaxels fulfilling the S/N threshold in (\(\sim\)`<!-- -->`{=html}12000 spaxels) have also S/N \(>3\) in \[\] (\(\sim\)`<!-- -->`{=html}11700 spaxels). The grey shading in the background represents the total sample of MaNGA spaxels as selected by the S/N cut, namely the total BPT (left) and (right) spaxels as previously defined. ### Comparing the two diagnostic diagrams {#subsec:bpt_vs_heii} Similarly to the analysis performed on integrated emission line flux ratios in, in the bottom panels of Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} we compare the spatial distributions of AGN spaxels identified with one diagnostic diagram, within the parameter space of the other diagnostic. More precisely, we show the distribution of BPT-selected AGN spaxels in the diagram (empty/filled blue circles, bottom left panel) and the distribution of -selected AGN spaxels in the BPT diagram (empty/filled red squares, bottom right panel). Except for a very few cases, almost all BPT AGN spaxels fall above the 10% AGN demarcation line of the -diagram, and 82% of the total BPT AGN spaxels is located even above the 50% AGN line in the diagram. Such a fraction increases to 100%, if we consider only BPT AGN spaxels with also S/N() \(>5\) (filled blue circles). Therefore, the -classification overall recovers the BTP-classified AGN. Unlike the BPT AGN spaxels in the diagram, the AGN spaxels are scattered across the BPT plane and extend well into the SF region, with only 29% of them meeting classification criteria for BPT AGN. Hence, Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} clearly demonstrates the power of diagnostic in identifying regions with significant contribution to ionisation from an AGN, which may be erroneously classified as entirely due to star formation. ## Selecting AGN host galaxies in MaNGA {#subsec:AGN_galaxies} We define a criterion to identify AGN host galaxies based on their content of AGN spaxels as classified by the emission line diagnostic diagrams in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}. We identify as BPT () AGN galaxy candidates those objects with at least 20 AGN spaxels meeting the S/N threshold, defined for the BPT () diagram in Sect. [2.2](#subsec:eldiag){reference-type="ref" reference="subsec:eldiag"}. The minimum requirement of 20 AGN spaxels approximately corresponds to one spatial resolution element in MaNGA, in case of adjacent spaxels. We check contiguity among spaxels at a later stage of the analysis (see Sect. [2.3.1](#subsec:fake){reference-type="ref" reference="subsec:fake"} and Appendix [\[appx1\]](#appx1){reference-type="ref" reference="appx1"}). With this definition we obtain populations of BPT and AGN galaxy candidates based on the BPT and spaxels classification, respectively. Among the AGN population, some galaxies are also BPT-selected AGN (hereafter BPT& AGN); while others are identified as AGN only by the diagram (hereafter -only AGN), with the majority of their spaxels located in BPT SF region. Details on the demography of the different AGN populations are provided at the end of Sect. [2.3.3](#subsec:type1_WR){reference-type="ref" reference="subsec:type1_WR"}, after refining our selected sample of AGN galaxies. ### Excluding cases of dubious line emission {#subsec:fake} As already mentioned, the main problem related to the use of the emission line is its faintness, limiting its clear detection to only a small sample of MaNGA galaxies. From our preliminary selection we obtain 178 AGN galaxy candidates, most of which must be rejected because the emission line at S/N\(>\)`<!-- -->`{=html}5 (suspiciously reported by the MaNGA DAP even in spaxels in the galaxy outskirts) does not appear to be real. By visually inspecting single-spaxel spectra using Marvin[^1] and DAP maps, we indeed find, in \(\sim\)`<!-- -->`{=html}50% of the selected galaxies, spaxels with a putative high-S/N , where the emission line is actually resulting from either noise, artifacts or strong residuals after the stellar continuum subtraction. To identify and exclude these galaxies from our AGN sample, for each galaxy we plot the flux map and create the 'subtracted cube', that is the MaNGA datacube after subtracting the DAP stellar continuum model. From the subtracted cube and the DAP emission line model cube[^2] of each galaxy, we then extract the spectrum and the cumulative emission line model using two different central apertures (i.e. \(2.5^{\prime\prime}\) and \(5.5^{\prime\prime}\)). By visually examining both the flux maps and the integrated spectra, we finally exclude cases of non-detection from our preliminary selected sample. We additionally report on five cases of contamination by a foreground galaxy (8084-12701, 8158-1901, 8158-1902, 8987-6101, 9194-6104), producing an emission line (likely H\(\beta\)) around the wavelength in the rest-frame of the target galaxy and, therefore, misinterpreted as the line by the DAP. These five galaxies are therefore excluded. ### Need for re-modelling some MaNGA datacubes {#subsec:refitting} [\[tab1\]]{#tab1 label="tab1"} By visually inspecting extracted spectra along with their cumulative emission line model (as previously explained in Sect. [2.3.1](#subsec:fake){reference-type="ref" reference="subsec:fake"}), we also find 16 MaNGA datacubes of AGN galaxies badly fitted by the MaNGA DAP (Table [1](#tab1){reference-type="ref" reference="tab1"}). We therefore remodelled all them by means of the fitting code for IFU data described in (see also @2021A&A...648A..99T, @2023arXiv230111060C for more details). The fitting procedure by first models the full datacubes with [pPXF]{.smallcaps}, creating dedicated templates for each contributing spectral component: namely, stellar continuum, AGN continuum, and line emission from AGN broad and narrow line regions (BLR and NLR, respectively). The stellar continuum is modelled with the MILES extended stellar population templates, and the AGN continuum through a nth-degree polynomial. Finally, multiple Gaussian components are used to reproduce emission lines, including a broad component to account for any possible BLR contribution to permitted emission lines (such as H\(\alpha\), H\(\beta\) and ). Each set of (broad or narrow) Gaussian components is constrained to have the same kinematics (i.e. same velocity \(v\) and velocity dispersion \(\sigma\)). Then, the fitting code allows the user to subtract from the data the total continuum (i.e. stellar plus AGN continuum) and the unresolved BLR line emission, thus creating a datacube containing only spatially-resolved narrow emission lines. At this point, a refined multiple Gaussian fitting of the narrow emission lines can be performed, with no longer contamination from continuum and/or unresolved emission. Among the 16 badly fitted MaNGA datacubes, six galaxies are likely Seyfert 1.8 since their hydrogen Balmer lines clearly exhibit a very broad component originating in the AGN BLR, which the MaNGA DAP is not designed to reproduce. We therefore fully re-fitted these six datacubes (labelled as 'full' in Table [1](#tab1){reference-type="ref" reference="tab1"}), starting from the modelling of the stellar continuum and including also a very broad Gaussian component (\(\sigma>1000\) km s\(^{-1}\)) in the model of the permitted lines, to account for the BLR contribution to total line emission. In the remaining ten datacubes instead, no BLR line emission is present and the stellar continuum has been correctly modelled by the MaNGA DAP. We hence created the corresponding subtracted datacubes (as defined in Sect. [2.3.1](#subsec:fake){reference-type="ref" reference="subsec:fake"}) and re-modelled emission lines only. In eight datacubes, the inaccurate best-fit model is limited to the faint emission line, while the other brighter emission lines have been correctly modelled. So, in these cases we re-fitted the emission line only (labelled as '' in Table [1](#tab1){reference-type="ref" reference="tab1"}), to obtain a more reliable flux. Two galaxies instead require a re-modelling of their entire subtracted cube, hence we re-fitted all emission lines ('all lines' in Table [1](#tab1){reference-type="ref" reference="tab1"}). In Fig. [\[fignew\]](#fignew){reference-type="ref" reference="fignew"}, we plot as representative example the best-fit results of the main emission lines in the MaNGA galaxy 9026-9101, as obtained from our spectral re-modelling: the subtracted data (black) and the total emission-line model (lightblue) have been extracted from a central 2.5\(^{\prime\prime}\times\)`<!-- -->`{=html}2.5\(^{\prime\prime}\) aperture. To accurately model the narrow-line emission in this galaxy, we used two Gaussian components, one of which is separately shown (orange) in Fig. [\[fignew\]](#fignew){reference-type="ref" reference="fignew"}. After the re-modelling, we create flux maps of the main emission lines and correct line fluxes for dust extinction. At this point, six galaxies (8320-9101, 8341-12705, 8458-3702, 8934-3701, 9026-3701, 9883-3701) are no longer classified as AGN galaxies (but as SF galaxies) according to the diagram, so we exclude them. In Table [1](#tab1){reference-type="ref" reference="tab1"}, we list the refitted MaNGA datacubes along with the type of remodelling. For each galaxy, we also indicate whether the AGN classification is confirmed after the remodelling, and we also identify -only AGN galaxies among such confirmed cases. ### Search for type 1 AGN and due to WR stars {#subsec:type1_WR} We finally look for type 1 AGN in the overall (BPT and/or ) AGN sample, and for central emission due to WR stars among AGN galaxies. To identify both categories we compare the average velocity dispersion \(\left<\sigma\right>\) of different emission lines within a 5.5\(^{\prime\prime}\) central aperture. In particular, we classify as type 1 AGN those with \(\left<\sigma(\text{H}\alpha)\right>>2\left<\sigma([\ion{O}{iii}])\right>\) or \(\left<\sigma(\text{H}\beta)\right>>2\left<\sigma([\ion{O}{iii}])\right>\), resulting in 13 galaxies. We decide to exclude the type 1 AGN since we are primarily interested in AGN likely missed by the BPT classification, while these are very clear, luminous AGN. Furthermore, their datacubes should be refitted (just like the Seyfert 1.8 described in Sect. [2.3.2](#subsec:refitting){reference-type="ref" reference="subsec:refitting"}) and the determination of their SFR and stellar masses would be more uncertain. Regarding emission due to WR stars, we find clear evidence for blue blump in three galaxies -classified as SF after our remodelling (8458-3702, 8934-3701, 9883-3701), identified as WR galaxies by. In Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, we show for comparison two spectra over the wavelength range, respectively extracted from a WR region in the galaxy 8458-3702 (blue), and from the nuclear region of the galaxy 8257-12701 (red), classified as -only AGN in our census. Whereas the former spectrum clearly exhibits a blue bump around the line along with a bright \[\]\(\lambda\)`<!-- -->`{=html}4658Å  typically associated with WR stars (e.g. @2021ApJ...915...21R), the latter distinctly shows a peaked emission line, classified as AGN-like by the diagram. However, we point out that this study does not aim at a rigorous identification of WR regions within MaNGA galaxies but, mostly, at excluding cases of clear WR emission originating in the galaxy centre, but misclassified as AGN-like based on the diagnostic. For this reason, we search in the remaining AGN sample for signatures of blue bump only within a central 5.5\(^{\prime\prime}\) aperture, by requiring \(\left<\sigma(\ion{He}{ii})\right>>2\left<\sigma([\ion{O}{iii}])\right>\) and \(\left<\sigma(\ion{He}{ii})\right>>500\) km s\(^{-1}\) (indicative lower limit taken from @2008A&A...485..657B). We thus tentatively detect one case of blue bump in the central region of the galaxy 8250-6101. Yet, given the marginal detection of such bump compared to the narrow, nebular emission line, we decide to retain this in our AGN sample. Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} summarises the demography of the AGN population (and subpopulations), consisting of 459 AGN host candidates in total (\(\sim\)`<!-- -->`{=html}10% in MaNGA DR15), as resulting from our selection criterion. Out of the total AGN galaxies, 432 are BPT-selected AGN (94%), whereas 81 are -selected AGN (18%). In the AGN subsample, 54 objects are classified as AGN galaxies also by the BPT (12% and 67% of the total and AGN samples, respectively). There are instead 27 -only AGN candidates (6% and 33% of the total and AGN samples, respectively), missed by the BPT AGN classification. There are finally 378 BPT AGN (82%), with no detection (BPT AGN no). In Appendix [\[appx1\]](#appx1){reference-type="ref" reference="appx1"}, we show for every selected -only AGN galaxy spatially resolved maps displaying the spaxel classification according to BPT (Fig. [\[fig_a1\]](#fig_a1){reference-type="ref" reference="fig_a1"}) and (Fig. [\[fig_a2\]](#fig_a2){reference-type="ref" reference="fig_a2"}) classification, respectively, supporting the presence a central AGN and confirming their identification as -only AGN. By crossmatching our sample of -only AGN with the AGN catalogue reported by, we find that 11 out of 27 (41%) -only AGN are identified as AGN based on their radio observations (10 objects) and/or mid-IR colours (2), further supporting the use of to select AGN missed in usual BPT diagnostics. Moreover, classify as AGN also the galaxy 8615-3701, due to the presence of broad emission lines in its SDSS spectrum. This object is indeed one of the MaNGA galaxies we identified as intermediate-type AGN and, for this, fully refitted in Sect. [2.3.2](#subsec:refitting){reference-type="ref" reference="subsec:refitting"}. ## Verifying the AGN-like nature of line emission {#subsec:offset_AGN} We take advantage of the spatially resolved information of MaNGA to verify the AGN-like nature of the detected line emission. In Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}, we plot the radial distribution (distance from the centre in units of effective radius, R\(_\mathrm{e}\)) of the AGN-like BPT (lightblue) and (grey) spaxels. In agreement with the AGN interpretation, the AGN-like emission mainly originates in the galaxy center. In fact, the BPT AGN spaxels are on average at a distance of \(\sim1.2\) R\(_\mathrm{e}\) (dashed lightblue line) and located up to \(\sim3\) R\(_\mathrm{e}\) from the galaxy centre; while the AGN spaxels are exclusively found at shorter distances, where the S/N is high enough to detect , with a mean (dashed grey line) and maximum radial distance of \(\sim0.56\) R\(_\mathrm{e}\) and \(\sim2\) R\(_\mathrm{e}\), respectively. As a final check on the correct selection of AGN-like spaxels, hence on the overall identification of -only AGN galaxies, in Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"} we plot /H\(\beta\) ratio as a function of equivalent width (EW) of H\(\beta\) (left panel), H\(\alpha\) (middle panel) and \[\] (right panel) lines, for all spaxels of -only AGN fulfilling the S/N threshold (they are plotted in Fig. [\[fig_a2\]](#fig_a2){reference-type="ref" reference="fig_a2"}), defined in Sect. [2.2](#subsec:eldiag){reference-type="ref" reference="subsec:eldiag"}. In addition, we show the dependence on the \[\]/H\(\alpha\) ratio using different colourbars according to their -based classification as SF (star symbols, blue colourbar), strong AGN (squares, red colourbar), or weak AGN (triangles, yellow-to-orange colourbar). The /H\(\beta\) ratios overall decrease at increasing line EWs, with the largest EWs associated SF spaxels, consistently with EW(H\(\beta\)) values inferred for SDSS SF galaxies showing WR signatures. Such SF spaxels, featured by low-/H\(\beta\) and high-EW values, also exhibit the lowest \[\]/H\(\alpha\) ratios in agreement with those typical of SF galaxies. In addition to have higher /H\(\beta\) ratios (as already pointed out by the diagram), AGN spaxels are instead overall characterised by larger EW(H\(\alpha\)) (\(>10\) Å) and higher \[\]/H\(\alpha\) ratios (log(\[\]/H\(\alpha\)) \(\gtrsim-0.4\)), consistently with the AGN classification by. The few spaxels with slightly lower \[\]/H\(\alpha\) ratios have however too large /H\(\beta\) ratios to be due to pure star formation. The right panel (i.e. /H\(\beta\) against EW(\[\])) highlights the existence of a separate population of spaxels at lower values of EW(\[\]). Such a spaxel population might either identify cases of AGN nuclear continuum detected to some extent, or central spaxels showing some scattered nuclear emission. We will further investigate this aspect in a future study. # Properties of the AGN host galaxies {#sec:host_properties} In Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} we plot the star formation rate--stellar mass diagram (SFR--M\(_*\); @2007ApJ...660L..43N [@2011A&A...533A.119E; @2015ApJ...801...80L]), separately for the distinct AGN sub-populations, namely the BPT AGN (left panel, blue points), the AGN (middle, red squares) and the -only AGN (right, orange triangles) samples, with the grey shading in the background showing the distribution of SDSS galaxies, whose SFR and M\(_*\) measurements are taken from the MPA/JHU catalogue[^3]. To separate the different regions in the plane of SF, green valley (GV) and quiescent (Q) galaxies, we use the demarcation lines from. Measurements of SFR and M\(_*\) have been taken from the MaNGA Pipe3D catalogue. In particular, we adopt the integrated SFR estimates derived from single stellar population (SSP) models provided by Pipe3D, which are computed from the amount of stellar mass formed in the last 32 Myr. Since we are dealing with AGN galaxies, SSP-based SFRs are indeed expected to be more accurate than H\(\alpha\)-based SFRs (SSP-and H\(\alpha\)-based SFR values are consistent within 50% for SF galaxies), given the likely significant AGN contamination to the H\(\alpha\) emission. Each panel of Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} reports the percentage of AGN galaxies of the respective sub-sample lying in the SF, GV and Q regions. The BPT AGN mainly fall in the GV (45%), with a considerable fraction (29%) of AGN residing in quiescent galaxies. The AGN instead tend to lie on the Main Sequence (MS, 74%), although they are still scattered across the GV (25%). Finally, the -only AGN exhibit the interesting property of being all in MS galaxies. All these findings are consistent with those obtained by using single-fiber SDSS measurements. However, it is important to note that our selected -only AGN reside in massive galaxies (M\(_*\gtrsim10^{10}\) \(M_\odot\)). This gives us confidence that we are not including star forming dwarf galaxies, where the AGN presence may be more questionable and the emission may originate from different sources, such as hot massive stars in metal-poor systems. The discovery through the diagnostics of an AGN sub-population residing in massive MS galaxies and missed by the BPT classification points to the unique ability of the emission line diagnostic to find elusive AGN hosted by actively SF galaxies. This hence highlights the possible relevant role of the line emission in the systematic rest-frame optical search for AGN especially at very high redshift, where the star formation in MS galaxies is expected to be higher. In Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} we compare the line luminosity (\(L_{\text{\ion{He}{ii}}}\)) in -only AGN to that of BPT& AGN selected by both diagnostics. The line luminosity (as well as other commonly used lines, such as \[\]) is indeed known to correlate with the AGN intrinsic luminosity as traced, for instance, by their hard X-ray emission. Therefore, it can be used as a proxy for the AGN luminosity. For this purpose, we compute the total \(L_{\text{\ion{He}{ii}}}\) in a given galaxy from the total line flux obtained by summing the flux contained in S/N()\(>\)`<!-- -->`{=html}5 spaxels. Similarly to what obtained in, we find that the -only AGN are less luminous than the BPT& ones, with a mean luminosity (in units of erg s\(^{-1}\)) of 39.6 (dashed orange line) and 39.9 (dashed black line) in logarithmic scale, respectively. The mean values inferred in for SDSS galaxies are larger (40.7 and 41.1, respectively) but also separated by a comparable difference of a few dex. This difference in \(L_{\text{\ion{He}{ii}}}\) is likely due to the fact that the SDSS survey spans a higher redshift range and therefore includes, on average, more luminous AGN than MaNGA. # Discussion {#sec:discussion} ## Estimating the number of undetected -only AGN {#sec:missing} Previously, we showed that there are 27 galaxies (i.e. 6% of the total selected AGN sample) classified as AGN only on the basis of the diagram. Such -only AGN represent about 33% of the overall -selected AGN population and, interestingly, reside primarily in massive (M\(_*\gtrsim10^{10}\) \(M_\odot\)) MS galaxies, where SF activity plays a primary role as an ionising source. The intense star formation in the galaxy, combined with the low-luminosity nature of these AGN, can therefore lead to the missed BPT-AGN selection. Being more sensitive to AGN emission, the diagnostics allows us to unveil this low-luminosity AGN sub-population and to perform a more complete census of the total AGN population. However, as already mentioned, the main issue related to the use of the diagnostics is the faintness of such an emission line. Among our total selected AGN population (459), we indeed detect the line in 81 objects only. It is therefore important to estimate how many AGN we are missing in the MaNGA sample because of detection limits, so to obtain a less biased estimate of the AGN population. A fraction of this overall undetected AGN population is expected to be among the selected BPT AGN which show no line emission (cerulean region in Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}, 378 objects). Therefore, these are already counted in the total AGN census, whereas we want to estimate the number of missing -only AGN in MaNGA, not yet included in our total selected AGN sample. With this aim, we first estimate the MaNGA detection limit for the line emission by using the brighter \[\] line emission as a proxy. Figure [\[fig8\]](#fig8){reference-type="ref" reference="fig8"} shows the distribution of the \[\] emission line flux contained in single spaxels within a central aperture of radius of 0.6 R\(_\mathrm{e}\), separately, for the AGN subsample selected by both diagrams (filled grey histogram; S/N(\[\])\(>\)`<!-- -->`{=html}3 and S/N()\(>\)`<!-- -->`{=html}5), the -only AGN population (orange contours; S/N(\[\])\(>\)`<!-- -->`{=html}3 and S/N()\(>\)`<!-- -->`{=html}5), and the overall BPT-selected AGN (filled lightblue; S/N(\[\])\(>\)`<!-- -->`{=html}3). By comparing the -only and BPT& distributions, the former is shifted towards smaller values of \[\] flux compared to the latter. This behaviour is expected since in the -only AGN the star formation dominates over the AGN, thus producing smaller \[\]/H\(\beta\) ratios consistent with ionisation due to star formation. The BPT& is coincident with the overall BPT AGN selected population above \(\rm log(F_{\rm [\ion{O}{iii}]}/10^{-17}erg~s^{-1})>1.5\). This is interesting, as it indicates that at high \[\] fluxes all BPT-selected AGN are also -selected. However, at lower fluxes the two populations start to depart and the BPT& distribution shows a clear cutoff at a flux of \(F^{\rm lim}_{\rm [\ion{O}{iii}]}<1.5\times10^{-16}\) erg s\(^{-1}\) cm\(^{-2}\) (vertical black dashed line), while the distribution of all standard BPT AGN keeps increasing to lower \[\] fluxes. The cutoff in the BPT& population is clearly due to the requirement of having detected. Therefore, we take the cutoff value (\(1.5\times10^{-16}\) erg s\(^{-1}\) cm\(^{-2}\)) as the limiting \[\] flux (\(F_{\rm lim}\)) beneath which the emission is likely missed because of the sensitivity of MaNGA observations. [\[tab2\]]{#tab2 label="tab2"} To count how many -only AGN are missed below this limit, we assume that the ratio of -only to all BPT AGN below the \[\] flux limit is the same as above the limit, namely: \[\centering N^{\mathrm{\ion{He}{ii}-only}}_{\mathrm{below-F_{lim}}} = \frac{N^{\mathrm{\ion{He}{ii}-only}}_{\mathrm{above-F_{lim}}}}{N^{\mathrm{BPT}}_{\mathrm{above-F_{lim}}}}\times N^{\mathrm{BPT}}_{\mathrm{below-F_{lim}}}=75,\] where \(N^{\mathrm{\ion{He}{ii}-only}}_{\mathrm{above-F_{lim}}}=5\), \(N^{\mathrm{BPT}}_{\mathrm{above-F_{lim}}}=27\), and \(N^{\mathrm{BPT}}_{\mathrm{below-F_{lim}}}=405\), as inferred by comparing for each object the average \[\] flux in spaxels within 0.6 R\(_{\rm e}\) with \(F_{\rm lim}\). The resulting demography is summarised in Table [2](#tab2){reference-type="ref" reference="tab2"}. Among the 75 -only AGN below our adopted limit, 22 are however detected. Therefore, the undetected -only AGN below \(F_{\rm lim}\) are expected to be 53 in total, thus leading to a final overall AGN sample consisting of 512 galaxies. However, we argue that the estimate of 53 missing -only AGN must be considered as a lower limit to the real number, since such value has been inferred under the assumption that the ratio of -only AGN to BPT AGN is the same at all stellar masses. Instead, we know that BPT AGN mainly reside in GV and Q galaxies, which are typically more massive than SF Main Sequence galaxies, the primary hosts of -only AGN. ## Implications for the AGN census and quenching scenarios As final step of this work, in Fig. [\[fig9\]](#fig9){reference-type="ref" reference="fig9"} we study how the AGN demography in the SFR--M\(_*\) plane changes with respect to the initial BPT-based census (first bar chart), after the inclusion of the -selected AGN (second), and further after the inclusion of the 53 estimated missing -only AGN (third). Each bar chart shows percent fraction of the total AGN population (grey) and, separately, of the SF Main Sequence (blue), GV (green) and Q (red) AGN sub-samples, with respect to the total MaNGA DR15 galaxies (4656). The filled bars represent the same categories but considering only galaxies more massive than M\(_*\gtrsim10^{10}\) \(M_\odot\) (3297 massive galaxies in MaNGA DR15), since almost all (93%) -only AGN are found in this stellar mass range. With respect to our initial BPT AGN census, the diagnostics leads to an increase by 6% and 24% in our total and SF Main Sequence detected AGN samples, respectively, further highlighting the poor BPT efficiency at unveiling AGN in SF galaxies. On the contrary, the GV (4%) and Q (3%) AGN fractions in MaNGA are not affected at all, being all the selected -only AGN located on the MS. We assumed the same to be valid for all the 53 -only AGN expected to lie below the MaNGA detection limit. The inclusion of these 53 objects entails an increase by about 12% in our total (BPT and/or ) AGN sample, and by about 38% in our SF (Main Sequence) AGN population. These expected new figures represent 11% and 4% in the overall MaNGA DR15 sample, respectively, corresponding to 14% and 5% for the more massive case (M\(_*\gtrsim10^{10}\) \(M_\odot\)). Overall our new census implies a modest increase in the overall census of AGN with respect to the classical BPT classification (from 9% to 11%), but the number of AGN on the SF Main Sequence is *doubled*. A substantial number of AGN in SF galaxies compared to GV and Q galaxies points to a picture where significant star formation and black hole accretion are coeval, with a large amount of gas simultaneously available for both star formation and black hole feeding. Such a correlation between star formation and black hole accretion rates is also predicted by cosmological hydrodynamical simulations, such as IllustrisTNG, SIMBA and, to a lesser extent, by EAGLE. To illustrate this in our context, in Fig. [\[fig9\]](#fig9){reference-type="ref" reference="fig9"} we also show the predictions of IllustrisTNG for the redshift z\(=\)`<!-- -->`{=html}0 AGN population (fourth bar chart) for comparison with our observational results. In the simulation, we select as AGN hosts those galaxies with an Eddington ratio \(\lambda_{\rm Edd}\) in the top 20% of all Eddington ratios (\(\lambda_{\rm Edd}>0.016\)), and classify them in SF, GV and Q galaxies according to the same prescription used with MaNGA data (i.e. @2017ApJ...851L..24H). When comparing our observational results (first to third bar charts) with the simulation predictions, we can clearly see that the discrepancy in the fraction of total and SF (MS) AGN starts is mitigated only when adding the -only AGN to the BPT-selected sample. At the same time, we note that the AGN feedback prescription implemented in IllustrisTNG necessarily dictates an association between high AGN accretion rates and their location on the star forming Main Sequence. In this model, AGN feedback is only successful at suppressing global SFR in galaxies when it switches into the 'kinetic' mode at low accretion rates. In consequence, simulated black holes with high \(\lambda_{\rm Edd}\) are expected to almost exclusively reside in galaxies with active star formation. Finally, we highlight that the existence of a considerable AGN population in SF galaxies compared to GV and Q galaxies can have important implications for AGN quenching scenarios. In fact, the remarkably larger fraction of AGN found in GV and Q galaxies in past studies has always favoured either the scenario in which the AGN phase is long-lived enough (\(\sim\)`<!-- -->`{=html}1 Gyr) to suppress the SF, and then evolve towards the green valley and the red sequence, or the scenario in which AGN are able to immediately quench the SF but becoming detectable only with a delay of about 100 Myr. Our discovery of a substantial AGN population in SF galaxies, instead might suggest the scenario in which the AGN does not really quench star formation directly nor instantaneously, but suppress start formation indirectly, on much longer timescales (\(\sim\)`<!-- -->`{=html}1 Gyr) via 'preventative' feedback, namely, by injecting energy into the halo, preventing accretion of fresh gas, and hampering star formation as a consequence of starvation. # Conclusions {#sec:conclusion} We have used the diagnostics to detect AGN activity in galaxies observed in the MaNGA survey, hence with spatially-resolved spectroscopy, and compare the resulting population with the AGN detected via the standard BPT diagrams. The main conclusions from our study are summarised below. 1. By comparing the spatially-resolved BPT and diagrams of MaNGA spaxels (Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}), we find that whereas the diagnostics globally recover the BPT AGN classification, only 29% of the -selected AGN spaxels are classified as AGN by the BPT diagram. 2. Based on the spaxel classification according to the two diagnostics, we obtain a total sample of 459 AGN host galaxies in MaNGA, out of which 432 (94%) are BPT-selected AGN, whereas 81 (18%) are -selected AGN (Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}). In the AGN sample, 27 objects (6%) are classified as AGN by the diagnostics only (-only AGN), out of which 11 (41%) are identified as AGN based on radio observations (10) and/or mid-IR colours (2). This further supports the diagnostic as useful tracer of AGN missed in usual BPT diagnostics. 3. The AGN-like emission is overall centrally located (within 0.6 R\(_\mathrm{e}\)) in all -selected AGN galaxies (Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}), thus supporting the AGN-like ionisation of such emitting regions. Moreover, the high /H\(\beta\) and \[\]/H\(\alpha\) ratios combined with large EWs (EW(H\(\alpha\)) \(>10\) Å) (Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}), compared to pure SF galaxies, further supports our correct identification of the 27 -only AGN as real active galaxies. 4. All -only AGN reside in Star Forming Main Sequence galaxies (Fig. [\[fig6\]](#fig6){reference-type="ref" reference="fig6"}) and are less luminous than the BPT-selected AGN, on average (Fig. [\[fig7\]](#fig7){reference-type="ref" reference="fig7"}). These lower-luminosity AGN are overwhelmed by the intense star formation radiation and, therefore, miss the BPT AGN classification. Furthermore, the -only AGN are mainly found in massive galaxies (M\(_*\gtrsim10^{10}\) \(M_\odot\)). This rules out the possibility that we are including SF dwarf galaxies, whose emission might be due to hot massive stars and X-ray binaries, especially in case of metal-poor environments. 5. Being the an extremely faint emission line, we expect several more AGN to be below the detection limit in MaNGA (Fig. [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}). Under simplifying assumptions, we estimated a lower limit to the missing population of undetected -only AGN (53 objects), all assumed to reside in SF galaxies like the objects we detect. 6. Our inferred, revised AGN census indicates that 11% of galaxies (in MaNGA DR15) host an AGN, of which 4% on the SF Main Sequence, 4% in the Green Valley and 3% in Quiescent galaxies (Fig. [\[fig9\]](#fig9){reference-type="ref" reference="fig9"}). We note that on the SF Main Sequence the number of AGN is doubled with respect to the simple BPT classification. If we restrict the census to galaxies more massive than M\(_*\gtrsim10^{10}\) \(M_\odot\), then we infer that the AGN population rises to 14%, of which 5% on the Star Forming Main Sequence, with the the Green Valley (5%) and Quiescent (4%) AGN populations unaffected. 7. The presence of a substantial number of AGN in SF galaxies compared to GV and Q galaxies is consistent with expectations of cosmological simulations. It may also have important implications on quenching scenarios, possibly supporting the picture where AGN feedback on SF is not instantaneous (i.e. inefficiency of the ejective mode) but, if anything, has a delayed, preventive quenching effect, likely via halo heating. With a view to exploring high redshifts at unprecedented high resolution and sensitivity with JWST and upcoming VLT facilities (ERIS and MOONS), novel emission-line diagnostics will be fundamental to investigate the ionisation conditions of the high-redshift Universe, featured by highly SF and/or metal-poor environments. As photoionisation models predict BPT to fail in correctly identifying the dominant ionisation process in case of metal-poor conditions, our study has demonstrated the crucial role the diagnostic might play in the search for hidden AGN in actively SF galaxies. We finally point out the greater feasibility of the diagnostics compared to other BPT-alternative AGN tracers such as the rarer and fainter optical coronal lines (e.g. \[\]\(\lambda\)`<!-- -->`{=html}5533, \[\]\(\lambda\)`<!-- -->`{=html}7609, \[\]\(\lambda\)`<!-- -->`{=html}7982, \[\]\(\lambda\)`<!-- -->`{=html}5303; @2009MNRAS.397..172G [@2021ApJ...922..155M; @2021ApJ...920...62N; @2023arXiv230113322N; @2022ApJ...936..140R]).

{'timestamp': '2023-02-10T02:00:17', 'yymm': '2302', 'arxiv_id': '2302.04282', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04282'}

# Introduction and the main result ## History of monostatic objects and the gap between upper and lower bounds Static balance points of a given object are points on its surface where, if supported on a horizontal plane, the object could be at rest. The numbers of various types of such balance points are often intuitively clear: for example a (fair) cubic dice has \(S=6\) stable equilibrium positions on its faces, \(U=8\) unstable equilibrium positions at its vertices and also has \(H=12\) saddle-like equilibria on its edges. We call a convex body *monostatic* if it has either one stable or one unstable static equilibrium position. Convex bodies with \(S=1\) stable position are also referred to as *mono-stable* and with \(U=1\) unstable position as *mono-unstable*, whereas convex bodies with \(S=U=1\) (i.e. one stable and one unstable balance position) are called *mono-monostatic*. The geometry of such convex bodies appears to be enigmatic: the existence of a convex, homogeneous mono-monostatic convex body was conjectured by V.I. Arnold in 1995 and proven in 2006. The rich variety of related discrete problems was opened by a brief note by, who asked whether homogeneous, mono-stable polyhedra existed at all and conjectured that homogeneous tetrahedra can not be mono-stable. (Throughout the paper, we use the shorthand *polyhedron* to mean a three-dimensional bounded convex polyhedron.) Both problems have been resolved in, where the authors presented a mono-stable, convex, homogeneous polyhedron with \(F=19\) faces and \(V=34\) vertices and proved that homogeneous polyhedra with \(V=F=4\) vertices and faces (i.e. tetrahedra) can not be mono-stable. I.e. for mono-stable, homogeneous polyhedra we have \(V,F > 4\). It immediately became intuitively clear that the essence of the problem is the rather substantial \(\emph{gap}\) between the respective values of \(F\) and \(V\). Various related problems have been investigated since. In the case of homogeneous mono-unstable polyhedra, the lower bound \(V,F>4\) was established and an example of \(V=F=18\) was provided. For convex mono-unstable \(0\)-skeletons, the lower bound \(F \geq 6, V\geq 8\) has been established in and an example with \(F = 8, V= 11\) was provided in. As we can see, in all investigated problems about monostatic polyhedra the gap between the lower bounds and the best known example exists, in fact, this gap appears to be a characteristic feature of this class of problems. ## The main result: closing the gap for mono-unstable 0-skeletons Our goal in the paper is to close this gap in the case of mono-unstable \(0\)-skeletons by proving the following result: Since in the authors presented examples with \(V=11\) vertices, the essence of our paper is to prove the following: This is an improvement of the lower bound shown in: Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"} is not just a quantitative generalization of Theorem [\[bozoki\]](#bozoki){reference-type="ref" reference="bozoki"}, and for two reasons: first we note that (unlike Theorem [\[bozoki\]](#bozoki){reference-type="ref" reference="bozoki"}), due to the existence of \(V=11\) examples it can not be improved. Second, the tools proving Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"} differ substantially from the tools used in the proof of Theorem [\[bozoki\]](#bozoki){reference-type="ref" reference="bozoki"}: while the latter was proven using a randomized computer search for certificates of infeasibility of certain polynomial systems, those tools have proved to be inefficient at going beyond the case \(V=7\). In the current paper, to resolve the cases \(V=8,9,10\), we combine the techniques introduced in with semidefinite optimization to efficiently generate the hundreds of thousands of infeasibility certificates required to prove Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"}, demonstrating the superior power of these tools. Beyond closing the gap for mono-unstable 0-skeletons in 3 dimensions, our computations also yielded Our proof of Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"} is an easily verifiable computer-assisted proof generated using convex optimization. First, the statement of the theorem is translated to the unsolvability of several systems of polynomial inequalities following the work of; see Theorem [\[thm:BDKR\]](#thm:BDKR){reference-type="ref" reference="thm:BDKR"} below. Then the unsolvability of these systems is proven using a sufficient condition derived from linear algebra (Lemma [\[lem:certificate\]](#lem:certificate){reference-type="ref" reference="lem:certificate"}). The unsolvability certificates take the form of positive integer vectors that are generated using semidefinite optimization. The verification of these certificates can be carried out independently of the method they were generated with, simply by verifying that the generated integer vectors are indeed (strictly) feasible solutions of certain linear matrix inequalities. Proving the infeasibility of systems of polynomial equations and inequalities and the equivalent problem of rigorously certifying lower bounds of polynomials on semialgebraic sets (that is, solution sets of polynomial inequalities) are becoming a fundamental tool in automated system verification and theorem proving, with applications in various areas of engineering, operations research, and statistics, including power systems engineering (optimal power flow), signal processing, and design of experiments. It has also been a particularly popular and successful technique in computer-assisted *geometric* theorem proving. Although computer-assisted proofs in geometry go back at least to the celebrated work of, more recent work combining polynomial optimization and convex optimization techniques have resulted in *easily verifiable* computer-assisted proofs of, for example, lower or upper bounds on optimal packings and other point configurations; see, e.g., to name only a few. Most of these works rely on semidefinite optimization to compute certifiable global lower bounds of polynomials (or trigonometric polynomials) over semialgebraic sets in a manner similar to our approach, and can also be interpreted as applications of Lasserre's moment relaxation of polynomial optimization problems. One major difference in our approach is that instead of formulating the problem as a single large-scale polynomial optimization problem, we work with a large number of small instances of polynomial optimization problems involving only quadratic polynomials whose infeasibility can be proven at the lowest level of the Lasserre hierarchy. In what follows, we shall present the details of our proof without further references to the theory of moment relaxations and polynomial optimization, and derive it instead from basic linear algebraic principles. # Proof of the main result In this section, we prove Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"} (and by extension, Corollary [\[thm:main-theorem-high-dim\]](#thm:main-theorem-high-dim){reference-type="ref" reference="thm:main-theorem-high-dim"}) by certifying the infeasibility of a number of systems of polynomial equations and inequalities--an idea introduced in. We rely on the same necessary condition of the existence of mono-unstable \(0\)-skeletons as in that paper, but improve on the search for infeasibility certificates using semidefinite optimization. The essential results we need from are summarized below in Theorem [\[thm:BDKR\]](#thm:BDKR){reference-type="ref" reference="thm:BDKR"}. Thus, to establish that no mono-unstable 0-skeletons with \(V\) vertices exist, it is sufficient to prove that for all \((V-1)!\) choices of \(j_i\in\{1,\dots,i-1\}\) \((i=2,\dots,V)\), the system of inequalities and equations [\[eq:shadow\]](#eq:shadow){reference-type="eqref" reference="eq:shadow"} has no non-zero solutions. ## Tractable infeasibility certificates Whether a system of polynomial inequalities is solvable over the reals is algorithmically decidable in the real number model using (for example) quantifier elimination methods. However, with their (at least) exponential running time in the number of variables, these exact procedures are prohibitively expensive to apply to our problem. Additionally, they do not produce easily checkable *infeasibility certificates*. This means that if they conclude that the polynomial system in question does not have a solution, it is difficult to independently and efficiently verify that this conclusion was correct, leaving doubts about the validity of the computer-assisted proof. Our approach to verify the unsolvability of all \((V-1)!\) systems [\[eq:shadow\]](#eq:shadow){reference-type="eqref" reference="eq:shadow"} is to look for efficiently computable and efficiently verifiable infeasbility certificates based on sufficient (but not necessary) conditions of infeasibility. The system [\[eq:shadow\]](#eq:shadow){reference-type="eqref" reference="eq:shadow"} can be simplified by expressing, say, each coordinate \(r_{V,k}\,(k=1,\dots,d)\) of the last vertex \(r_V\) as a linear combination of the other variables using [\[eq:shadow-eq\]](#eq:shadow-eq){reference-type="eqref" reference="eq:shadow-eq"} and substituting them back to [\[eq:shadow-ineq\]](#eq:shadow-ineq){reference-type="eqref" reference="eq:shadow-ineq"}, to obtain an equivalent system of \(V-1\) homogeneous quadratic inequalities in \(n := d(V-1)\) variables with integer coefficients. For such systems of inequalities, we can use the following sufficient condition of infeasibility: We can find coefficients \(c_i\) satisfying the condition in Lemma [\[lem:certificate\]](#lem:certificate){reference-type="ref" reference="lem:certificate"} using semidefinite optimization. In the following, we use the common shorthand \(A\succcurlyeq B\) for the relation that the matrix \(A-B\) is positive semidefinite. Semidefinite optimization models such as [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"} are typically solved using numerical methods, which compute solutions that may be only approximately feasible or approximately optimal. This is of no concern for our proof, as we only need to find a componentwise positive feasible solution to [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"}. (The purpose of the norm constraint on \(c\) is to ensure that the problem is bounded, and can safely be violated.) As long as the maximum value of \(z\) is sufficiently positive (compared to the precision of the floating point computation), the approximately feasible and approximately optimal solution returned by a numerical semidefinite optimization method already serves as a rigorous proof of the non-existence of solutions of [\[eq:homquad\]](#eq:homquad){reference-type="eqref" reference="eq:homquad"} by Lemma [\[lem:certificate\]](#lem:certificate){reference-type="ref" reference="lem:certificate"}. Thus, to prove that every 3-dimensional mono-unstable 0-skeleton has at least 11 vertices, we run the following algorithm: for every choice of \((j_2,\dots,j_{10}) \in \{1\} \times \{1,2\} \times \cdots \times \{1,\dots,9\}\), we transform the corresponding system [\[eq:shadow\]](#eq:shadow){reference-type="eqref" reference="eq:shadow"} to an equivalent system of homogeneous quadratic inequalities of the form [\[eq:homquad\]](#eq:homquad){reference-type="eqref" reference="eq:homquad"} by expressing each \(r_{10,k}\; (k=1,2,3)\) as a linear combination of the other variables using [\[eq:shadow-eq\]](#eq:shadow-eq){reference-type="eqref" reference="eq:shadow-eq"} and substituting them back to [\[eq:shadow-ineq\]](#eq:shadow-ineq){reference-type="eqref" reference="eq:shadow-ineq"}, and then we solve the corresponding semidefinite optimization problem [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"} using a numerical method to prove that the system [\[eq:homquad\]](#eq:homquad){reference-type="eqref" reference="eq:homquad"} has no non-zero solutions. The independently verifiable computer-generated proof is the list of positive rational vectors \(c\) (one for each permutation) returned by the semidefinite optimization algorithm. (The \(z\) component of the optimal solution is irrelevant as long as it is positive, and is not part of the infeasibility certificate.) The correctness of these vectors can be verified efficiently in rational arithmetic: it suffices to verify that the matrix \(\sum_{i=1}^m c_i Q_i\) is positive definite, which can be carried out in polynomial time in rational arithmetic, say, using the \(LDL^\mathrm{T}\) form of Cholesky decomposition or by verifying the positivity of the determinant of each leading principal submatrix. ## Implementation The algorithm was implemented using the semidefinite programming solver CSDP, interfaced using Mathematica, on a standard desktop computer. The enumeration and solution of the \(9!\) optimization problems took approximately half an hour. The numerical solutions (specifically, the \(c\) vectors obtained from CSDP) were rounded to nearby positive rational vectors that were confirmed using rational arithmetic to be feasible solutions of [\[eq:SDP\]](#eq:SDP){reference-type="eqref" reference="eq:SDP"}. These vectors were then further scaled up to positive integer vectors for ease of dissemination. Recall that the constraint \(\|c\|_2\leq 1\) need not hold for \(c\) to be a valid certificate. The list of the computed rational \(c\) vectors certifying the unsolvability of the systems [\[eq:shadow\]](#eq:shadow){reference-type="eqref" reference="eq:shadow"} can be found in the public repository <https://github.com/dpapp-github/mono-unstable>. This, along with the proof of Theorem [\[thm:BDKR\]](#thm:BDKR){reference-type="ref" reference="thm:BDKR"}, serves as the independently verifiable computer-assisted proof of Theorem [\[mainlemma\]](#mainlemma){reference-type="ref" reference="mainlemma"}.

{'timestamp': '2023-02-09T02:17:37', 'yymm': '2302', 'arxiv_id': '2302.04252', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04252'}

# Mean-Field Theory in Brownian Circuits {#app:mft} Here, we develop a mean-field theory for the error mitigation threshold in the noisy-mitigated Brownian circuit model. The unitary dynamics in the Brownian quantum circuit model is described by a stochastic Hamiltonian \[H(t) = \sum_{i < j,\mu,\nu} J_{ij\mu\nu}(t) \sigma_i^\mu \sigma_j^\nu,\] where \(\sigma_i^\mu\) are Pauli operators \(\mu \in \{ X,Y,Z\}\) for site \(i\) of \(N\) qubits and \(J_{ij\mu\nu}(t)\) is a white-noise correlated coupling with variance \[\langle J_{ij\mu\nu}(t) J_{k \ell \gamma \delta}(t') \rangle = \frac{J}{2 N} \delta_{ik} \delta_{j\ell} \delta_{\mu \gamma} \delta_{\nu \delta} \delta(t-t').\] The noise and antinoise are treated using a Lindblad master equation \[\dot{\rho} =-i [H(t),\rho] + \sum_i \frac{\gamma_i-\gamma_a}{4}\ \Big(-3 \rho + \sum_\mu \sigma_i^\mu \rho \sigma_i^\mu \Big),\] where \(\gamma_{i}\) is the local random noise rate and \(\gamma_a\) is the antinoise rate. To analyze the dynamics we derive an effective master equation that describes the replicated system \[M_k(\rho) = \mathbb{E} [ U^{\otimes k} \rho U^{\dag \otimes k} ],\] where \(U = \mathcal{T} e^{-i \int_0^t dt' H(t')}\) is the time evolution operator under the Hamiltonian. Expanding to second order in an infinitesimal time-step, we arrive at the equation \(M_k(\rho) = e^{\mathcal{L}_k t}\) for a Lindbladian \(\mathcal{L}_k\) given by \[\mathcal{L}_k(\rho) = \frac{J}{2 N} \sum_{i \ne j, \mu,\nu,r,s} \sigma_{ir}^{\mu }\sigma_{jr}^{\nu} \rho \sigma_{is}^{\mu }\sigma_{js}^{\nu }-\frac{1}{2} \{ \sigma_{ir}^{\mu} \sigma_{is}^{\mu } \sigma_{jr}^{\nu } \sigma_{js}^{\nu }, \rho \},\] where \(r\) and \(s\) are replica indices that run over \(1,\ldots,k\). Thus, we arrive at a master equation describing the average dynamics of the replicated density matrices \[\dot{\rho} = \mathcal{L}_k(\rho) + \sum_i \frac{\gamma_i-\gamma_a}{4}\ \Big(-3 k \rho + \sum_{r,\mu} \sigma_{ir}^\mu \rho \sigma_{ir}^\mu \Big).\] To develop the mean field theory, we study the two-replica problem \(k=2\). Similar to a Haar random circuit, the Lindbladian \(\mathcal{L}_2\) has two steady states \(I^{\otimes N}\) and \(S^{\otimes N}\). As our mean-field ansatz, we therefore use a product state of the form \(\rho = \bigotimes_{i=1}^N \rho_i\). A further simplification arises from the nature of the dynamics that has an effective SU(2) symmetry in the average-replica dynamics. As a result, we can express \[\rho_i = (1/4+\delta_i) |s\rangle \langle s| + (1/4-\delta_i/3) P_T,\] where \(\delta_i\) is the deviation from an infinite temperature state, \(|s\rangle\) is a two-qubit singlet state across the two replicas, and \(P_T\) is the projector onto the two-qubit triplet subspace of the two replicas. The mean-field equations for \(\delta_i\) take the form \[\dot{\delta}_i =-4 \Big[ \Delta_i + \frac{ J}{N} \sum_{j \ne i} (3 + 4\delta_j) \Big] \delta_i,\] where \(\Delta_i = \gamma_i-\gamma_a\). This equation has the two steady-state solutions \(\delta_i =0\) and \(\delta_i =-3/4\), corresponding to the \(I^{\otimes N}\) and \(S^{\otimes N}\) solutions, respectively. In the case of binary disorder, we define two populations of sites \(A_{1/2}\) such that \(\Delta_i = \gamma_{1/2}-\gamma_a\), respectively, for \(i \in A_{1/2}\). We have the zero-mean field condition \(p \Delta_1 +(1-p) \Delta_2 = 0\). Defining \(G_{\pm} = \frac{1}{N} \Big( \sum_{i \in A_2} \delta_i \pm \sum_{i \in A_1} \delta_i \Big)\), we arrive at simple closed set of equations in the large-\(N\) limit \[\begin{aligned} \dot{G}_+ & =-4J (3 + 4 G_+) G_+ + \frac{ 4|\Delta_1|}{2(1-p)} [G_--(2p-1) G_+ ], \\ \dot{G}_-& =-4J (3 + 4 G_+) G_-+ \frac{ 4|\Delta_1|}{2(1-p)} [G_+-(2p-1) G_-]. \end{aligned}\] Setting \(p = 1/2\), these equations reduce to the particularly simple form \[\begin{aligned} \dot{G}_+ & =-4J (3 + 4 G_+) G_+ + 4|\Delta_1| G_-, \\ \dot{G}_-& =-4J (3 + 4 G_+) G_-+ 4|\Delta_1| G_+, \end{aligned}\] For general \(p\), the steady state solutions are \(G_+ = 0,-(3-|\Delta_1|/J)/4,-(3+ \Delta_2/J )/4\). The all identity solution \(G_+ = G_-= 0\) only becomes an unstable fixed point for \(|\Delta_1|/J \ge 3\), whereas it remains stable for weaker disorder. As a result, the phase transition in the mean-field theory occurs at the disorder strength \(|\Delta_1| = 3 J\). Beyond this value of disorder, the mean-field steady-state solution flows to an unphysical state; however, for weaker disorder, the physical mean-field solution remains stable. # Statistical Mechanics Mapping Formalism {#app:statmech} Here, we review the statistical mechanics for the model with noise and antinoise, generalizing the mappings studied in Ref. . In this paper, we focus our attention to calculating second-moment quantities of a quantum state, which includes measures like fidelity, collision probability, and linear cross entropy. The circuit-averaged calculations of such quantities lends itself to a statistical mechanical mapping to an Ising spin model. Consider a second moment measure \(M\), averaged over circuits from ensemble \(\mathcal{U}\). For a state a state \(\rho_C\) of dimension \(2^{2n} \times 2^{2n}\), generated using a circuit \(C\) from an ensemble \(\mathcal{U}\), a circuit-averaged second moment measure \(M\) for can be written as a two-copy expectation of a linear operator \(O_M\) of dimension \(2^{4n} \times 2^{4n}\). \[\begin{gathered} \mathbb{E}_{C \in \mathcal{U}}[M[\rho_C]] = \mathbb{E}_{C \in \mathcal{U}}\left[\tr\left(O_M\rho_C \otimes \rho_C\right) \right] \\ = \tr\left(O_M \mathbb{E}_{C \in \mathcal{U}}\left[\rho_C \otimes \rho_C\right] \right) \end{gathered}\] The circuit-averaged two-copy state \(\mathbb{E}_{C \in \mathcal{U}}[\rho_C \otimes \rho_C]\), therefore, enables us to calculated second-moment measures. We consider circuit models that can be decomposed into a series of elementary two-qubit gates (noisy or noiseless), each drawn independently from the two-qubit Haar ensemble \(\mathcal{U}_2\). The action of the circuit map on two-copies of an initial input state is, thus, given by \[\rho_C{\otimes 2} = {C}_s \circ {C}_{s-1} \circ \cdots \circ {C}_{1}[\rho_0\otimes \rho_0],\] where the elementary single-qubit or two-qubit channels are indexed using integers \([1,s]\). For noiseless gates, \({C}_i[\rho_0\otimes \rho_0] = (C_i\otimes C_i) \rho (C_i^\dagger \otimes C_i^\dagger)\). We can model a noisy circuit by adding error channel \(\mathcal{E}\) after each noiseless gate: \[\rho_C^{\otimes 2} = \mathcal{E}_s \circ {C}_{s-1} \circ \mathcal{E}_{s-1} {C}_{s-1} \circ \cdots \circ \mathcal{E}_1 \circ {C}_{1}[\rho_0\otimes \rho_0].\] In general, the error channel may act on any set of qubits. For our purposes, we assume that the error channel \(\mathcal{E}_i\) acts on the qubit or the pair of qubits acted on by the noiseless gate \(C_i\). Since we draw each gate from the random one-qubit or two-qubit Haar ensemble, we can replace the noiseless maps \(C_i[\rho]\) with the gate-averaged map \(\overline{C}_i[\sigma] = \mathbb{E}_{C \in \mathcal{U}_{1/2}} C[\sigma]\), \[\mathbb{E}_{C \in \mathcal{U}} [\rho_C \otimes \rho_C] = \mathcal{E}_s \circ \overline{C}_{s-1} \circ \mathcal{E}_{s-1} \overline{C}_{s-1} \circ \cdots \circ \mathcal{E}_1 \circ \overline{C}_{1}[\rho_0 \otimes \rho_0].\] The action of a random single-qubit gate \(C\), on a state residing in the two-copy Hilbert space is given by \[\overline{C}^{(1)}[\sigma] = \frac{\tr((1-S/2)\sigma)}{3} I + \frac{\tr((S-1/2)\sigma)}{3} S, \label{eq:single-q-gate}\] where \(I\) and \(S\) are the \(4\times 4\) identity matrix and SWAP matrices, respectively. Similarly, the action of a random two-qubit gate on two copies of a qubit-pair is given by \[\overline{C}^{(2)}[\sigma] = \frac{\tr((1-SS/4)\sigma)}{15} II + \frac{\tr((SS-1/4)\sigma)}{15} SS,\] where we use the shorthand \(SS = S\otimes S\) and \(II = I\otimes I\). If two copies of a quantum state can be represented by a string of \(\rho^{\otimes 2} \in \{I, S\}^{n}\), a single-qubit gate acts on qubit \(k\) by modifying the \(j\)th bit of the I-S string using the transition rules \[I \to I \qquad S \to S,\] while leaving the rest of the bits in the string unchanged. Similarly, a two-qubit gate acting on qubits \(j\) and \(k\) modifies the \(j\)th and \(k\)th bits according to the transition rules: \[II \to II \qquad IS, SI \to \frac{2}{5}(II + SS) \qquad {S}{S}. \to {S}{S}.\] A noiseless random circuit, therefore, can be represented as a linear operator acting on the reduced space spanned by basis elements in \(\{I, S\}^{n}\), with each noiseless single-qubit gate given an identity map, and a two-qubit gate given by the transition matrix \[T[\overline{C}^{(2)}] = \begin{pmatrix}1 & 2/5 & 2/5 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 2/5 & 2/5 & 1 \end{pmatrix}.\] We can similarly, find the transition matrices corresponding to the noise and antinoise channel. The single-qubit depolarization channel with error rate \(q\), given by the following map \[\mathcal{E}^{(1)}(\rho) = (1-q)\rho + q\tr(\rho)\frac{\mathbb{1}}{2},\] acts on two copies of a qubit such that \[\mathcal{E}[I] = \mathcal{E}[I_2 \otimes I_2] = \mathcal{E}^{(1)}[I_2] \otimes \mathcal{E}^{(1)}[I_2] = I_2 \otimes I_2 = I\], where \(I_2\) is a \(2\times 2\) identity matrix. Similarly, \[\begin{aligned} \mathcal{E}&[S] \\ &= (\mathcal{E}^{(1)}\otimes\mathcal{E})[I_2 \otimes I_2 + X\otimes X + Y \otimes Y + Z\otimes Z]/2\\ &= \left[I_2 \otimes I_2 + (1-q)^2\left(X\otimes X + Y \otimes Y + Z\otimes Z\right)\right]/2\\ &= \left[(1-(1-q)^2)/2 I + (1-q)^2S\right] \end{aligned}\] where we have used the fact that \(S = (I_2 \otimes I_2 + X\otimes X + Y \otimes Y + Z \otimes Z)/2\). The transition matrix corresponding to a depolarizing noise, in the statistical mechanical picture is given by, \[T[\mathcal{E}_{q}] = \begin{pmatrix} 1 & (1-(1-q)^2)/2 \\ 0 & (1-q)^2 \end{pmatrix}\] Likewise, the antinoise channel of strength \(q_a\), given by \[\mathcal{A}^{(1)}(\rho) = \frac{1}{1-q_a}\left(\rho-q_a \tr(\rho)\frac{\mathbb{1}}{2} \right),\] acts on two-copies of a qubit such that \[\begin{aligned} \mathcal{A}[I] & = (\mathcal{A}^{(1)}\otimes\mathcal{A}^{(1)})[I_2 \otimes I_2] = I, \text{ and} \\ \mathcal{A}[S] &= \left[\left(1-\frac{1}{(1-q_a)^2}\right)\frac{I}{2} + \frac{1}{(1-q_a)^2}S\right], \end{aligned}\] giving a transition matrix \[T[\mathcal{A}_{q_a}] = \begin{pmatrix} 1 & \left(1-(1-q_a)^{-2}\right)/2 \\ 0 & {(1-q_a)^{-2}}. \end{pmatrix}\] Concatenating the transition matrix for the noise channel and the antinoise channel gives the transition matrix for the composite channel. \(T[\mathcal{A}_{q_a} \circ \mathcal{E}_q] = T[\mathcal{A}_{q_a}].T[\mathcal{E}_{q}]\). In our simulations, we start with an initial state drawn from the Haar ensemble or a random product state. A Haar random state is proportional to \(I^{\otimes n} + S^{\otimes n}\) in the two-copy descrpition, while a random product state is proportional to \((I+S)^{\otimes n}\). We can then use the statistical mechanical formalism discussed above to evolve this state using the respective transition matrices for two-qubit gates, noise and antinoise channels. # Local Probe {#app:crossing} In Fig. [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"}, we presented different correlation metrics between two qubits in the system as a probe of the phase transition. Here, we present an alternative metric based on the entropy of a single qubit in the system. Deep in the below threshold phase, this single-qubit entropy quantity saturates to one bit, while above threshold it diverges to large negative values due to the unphysical density matrix. As a result, we expect a crossing to occur at the phase transition. Numerical simulations of the two-replica stat-mech model for the all-to-all circuit model illustrate this behavior. In Fig. [\[fig:crossing\]](#fig:crossing){reference-type="ref" reference="fig:crossing"}(a), we show the unscaled behavior of the entropy for different system sizes, which shows a crossing near \(\sigma_c/\bar{q} = 0.65(5)\). Collapsing the data with this value of \(\sigma_c\) fixed, we estimate a critical exponent \(\mu = 1.0(2)\). In Fig. [\[fig:data\]](#fig:data){reference-type="ref" reference="fig:data"}(c-d), we fixed \(\mu = 1\) in collapsing the mutual information data based on these scaling results. This single-qubit quantity also has advantages for experimental probes of the transition as it requires minimal tomographic overhead to estimate. # Error Mitigated Fidelity Benchmarks {#app:fid} Here, we introduce mitigated fidelity benchmarks and demonstrate an exponential improvement in a mitigated verison of the linear cross-entropy benchmark below the error mitigation threshold. The task of sampling from the output distribution of a noiseless random circuit is widely conjectured to be intractable with classical computers. This conjecture forms the basis for claims of achievement of quantum computational advantage in recent experiments. The experimental claims remain controversial, however, partly because of the effects of noise on the output that render the signal classically simulatable at high depth. To provide evidence that the output signal still remains close to the ideal case, one can estimate fidelity benchmarks using the samples from the experiment. Verifying the claim of computational advantage in the case of noisy circuits then reduces to the task of achieving a sufficiently high "score" on the benchmark. Recall that the linear XEB is given by the formula \[F_{\rm XEB} = 2^n \sum_x p_{ n}(x) p_{0}(x)-1,\] where \(p_0(x) = |\langle x | U_d \cdots U_1 | 0 \rangle|^2\) is the probability of measuring outcome \(x\) for the noiseless circuit of depth \(d\) and \(p_{ n}(x) = \langle x| \mathcal{E}_d \circ \mathcal{U}_d \circ \cdots \mathcal{E}_1 \circ \mathcal{U}_1( | 0 \rangle \langle 0 |) | x \rangle\) is the analogous probability for the noisy circuit. We now introduce the mitigated linear XEB, which is instead given by the formula \[F_{\rm XEB_M} = 2^n \sum_x p_{n}(x) p_{a}(x)-1,\] where \(p_a(x) = \langle x| \mathcal{A}\circ \mathcal{U}_d \circ \cdots \mathcal{A} \circ \mathcal{U}_1( | 0 \rangle \langle 0 |) | x \rangle\) is a quasi-probability for the circuit with antinoise inserted in place of noise. The quantity \(p_a(x)\) can be computed on a classical computer, which leads to a sampling formula for \(F_{\rm XEB_M}\) using \(M\) samples \(x_i\) obtained from \(p_n(x)\) \[F_{\rm XEB_M} = \frac{2^n}{M} \sum_{i=1}^M p_a(x_i)-1.\] This formula illustrates that the mitigated fidelity can be obtained without directly implementing PEC except in purely classical post-processing. After circuit averaging, one can quickly show that for depolarizing noise and its antinoise partner, the mitigated fidelity is equivalent to the formula \[\bar{F}_{{\rm XEB}_M} = 2^n \sum_x p_{0}(x) p_{an}(x)-1,\] where \(p_{an}(x) = \langle x| \mathcal{A}\circ \mathcal{E}_d \circ \mathcal{U}_d \circ \cdots \mathcal{A} \circ \mathcal{E}_1 \circ \mathcal{U}_1( | 0 \rangle \langle 0 |) | x \rangle\) implements the antinoise on the same copy as the noise. This identity follows because the noise and antinoise on one copy have the identical effect on the \(I\) and \(S\) operators after averaging over circuits. From this expression, we see that, in the case of perfect mitigation, \(\bar{F}_{{\rm XEB}_M}\) reduces to its ideal value. We can also define a mitigated fidelity that takes the form \[\begin{aligned} F_M &= \tr[ \mathcal{A} \circ \mathcal{U}_d \circ \cdots \circ \mathcal{A} \circ \mathcal{U}_1(|0\rangle \langle 0|) \\ &\times \mathcal{E}_d \circ \mathcal{U}_d \circ \cdots \circ \mathcal{E}_1 \circ \mathcal{U}_1(|0\rangle \langle 0|)]. \end{aligned}\] In the case where \(\mathcal{E}_i = \mathcal{A}^{-1}\) for every \(i\), we can see that \(F_M=1\). In Fig. [\[fig:fidelity\]](#fig:fidelity){reference-type="ref" reference="fig:fidelity"}\], we show that the log-fidelity at low noise rates and \(d\) grows as \(\mathcal{O}(Nd)\), whereas using error mitigation this scaling can be improved to \(\mathcal{O}(\sqrt{Nd})\), representing an exponential improvement in the score that brings it closer to the \(\mathcal{O}(d)\) scaling of the log-total variation distance. Moreover, as explained above, the cross-entropy benchmark fidelity can be mitigated entirely in classical post-processing. As a result, the mitigated XEB fidelity can be estimated with existing experimental data from random circuit sampling experiments.

{'timestamp': '2023-02-10T02:00:15', 'yymm': '2302', 'arxiv_id': '2302.04278', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04278'}

# Domain Experts in ML Development Human experts are engaged in various aspects of machine learning development, such as data collection and validation, consultation on the development of knowledge representations, and system evaluation. Moreover, who is identified as 'expert', and thus who partakes in these activities, carries implicit, normative claims about whose knowledge is valid and should be recognized as such. In ML, the label "expert" directs attention to whose knowledge should be upheld as canonical and used to shape how systems behave and how they are validated. In this way, expertise confers power and influence over system development to those identified as possessing it. At the same time, definitions of domain expertise in ML vary, raising questions about who tends to to be recognized as an expert and why. Beyond the general notion of 'expertise' referring to a specialized or standard of knowledge, the bases on which expertise is identified can diverge. For example, in data annotation, "expert" has been used to refer to domain knowledge rooted in annotator lived experience, domain knowledge based on specific training processes (e.g., Wikipedia article editing) or academic study, and even to refer to labels from gold standard datasets, even when the knowledge or experience of gold standard annotators is unspecified. Scholars at the intersection of AI and HCI have expanded focus on the social dimensions of expertise. For example, found that developers failed to recognize domain expertise, treated workers as non-essential, and de-skilled their work by treating them as data collectors. Complementing their investigations, we pursue a systematic review of ML publications to understand the nature and role of expertise in ML development. On the backdrop of recent critiques pointing out that expanded participation in ML does not necessarily improve equitable outcomes, we consider not just *what* kinds of expertise are recognized, but also *how* experts participate. # Systematic Literature Review Using [dblp.org](dblp.org), we conducted a systematic literature review and thematic analysis of machine learning publications that involve domain experts in the development of a machine learning system. [dblp.org](dblp.org) is a bibliography of over 6 million computer science publications from a variety of venues, including prominent machine learning conferences such as NeurIPS, ICML, and AAAI, as well as broadly interdisciplinary venues, such as CHI and FAccT. From an initial pre-search of "expert',' "expertise", and "domain expert", we generated a list of search terms from the top-occurring bigrams, after removing stop-words. The list also included "non-expert", which we used to understand what did *not* constitute expertise. In assessing relevant publications, we focused on papers that involved the development of a machine learning system, including papers that focused on components of ML systems, such as ontologies or causal maps. The publications excluded from our search results included entire books, PhD theses, magazine articles, review papers, and position papers in which no system was developed or evaluated. After removing papers that did not meet our inclusion criteria we were left with 96 papers. For each paper we open coded 1) the bases for the definition of expertise used (e.g., expertise based on educational attainment) or if none was provided, we noted an implicit definition, 2) the phase(s) of development experts or non-experts contributed to (e.g., data collection), and 3) their role or how they participated (e.g., source data provider for ground truth annotations). # Key Findings While the full code book is too extensive to present, we note two high-level findings: #### 1. Expertise is frequently unspecified 38 of the 96 publications did not provide any explicit criteria that motivated the inclusion of the experts or non-experts engaged in system development or evaluation. In cases where the domain suggested a type of expert (e.g., medical experts for precision medicine systems) it was not clear why individual experts were selected or whether they possessed a particular expertise or focus within their field (e.g., cardiologist vs. ophthalmologist; years of practice). Failing to document explicit reasoning for determining expertise introduces issues for scientific reproducibility by under-specifying methodological details, as well as masks the reality of multiple ground truths which may be rooted in different cultural practices or training. This also risks perpetuating representational harms by failing to acknowledge the range of expertise that may be relevant and by presenting an under-specified perspective as canonical. Explicitly naming the individuals or groups sought to influence system development and the basis for doing so is important for understanding which perspectives are represented within ML systems. #### 2. Experts were frequently engaged as data workers Experts often provided or validated data annotations or provided term or concept definitions for ontology development. In 25 publications, experts annotated data or were themselves recorded as data subjects (e.g., using sensors). In only 2 publications was an expert involved in decisions about what a ML system should do or how an algorithm should perform (e.g., active decisions about what is included in an ontology, rather than providing domain definitions for pre-selected components). This points to the use of experts as a "repository of knowledge" that can regurgitate domain-specific facts, ignoring the breadth of skills that constitute expertise. Critiques of automation broadly have pointed out ways that it refocuses workers on menial tasks, rendering traditionally valued components of expertise, such as tacit knowledge, unimportant. # Conclusion Ultimately, expertise should be explicitly documented and justified. In addition to supporting scientific reproducibility and transparency, there is a need to document relevant aspects of expertise to clarify whose knowledge is recognized and used to validate systems. From a responsible AI perspective, should also recognize and evaluate against expertise rooted in different cultural practices and lived experiences--particularly as ML development increasingly turns to participatory approaches. There is an additional need to engage expertise in responsible ways to minimize deskilling and incorporate knowledge beyond that which can be conveniently captured through data labeling. Much of ML development encompasses 'participation as work' leaving a range of expert participation minimally explored.

{'timestamp': '2023-02-10T02:02:13', 'yymm': '2302', 'arxiv_id': '2302.04337', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04337'}

# Introduction {#introduction .unnumbered} A primary goal of emerging quantum computing technologies is to enable the simulation of quantum many-body systems that are challenging for classical computers. Early experimental demonstrations of quantum simulation algorithms have focused on computing ground-and excited-state energies of small molecules or few-site spin and fermionic models. More recently, the scale of quantum simulation experiments has increased in terms of numbers of qubits, diversity of gate sets, and complexity of algorithms, as manifested in simulation of models based on real molecules and materials, various phases of matter such as thermal, topological and many-body localized states, as well as holographic quantum simulation using quantum tensor networks. As quantum advantages in random sampling have been established on quantum hardware, focus has turned to the experimental demonstration of quantum advantages in problems of physical significance. For applications in chemistry and physics, the calculation of the response properties of molecules and materials is of substantial interest. Investigating response properties in the electronic structure theory framework involves calculating quantities such as the one-particle Green's function and density-density response functions, which provide insight into interpreting experimental spectroscopic measurements. Response properties of molecules and materials can be determined either in time domain or in frequency domain. Due to the natural ability of quantum computers to simulate time evolution, near-term algorithms to compute time-domain response properties have been carried out on quantum hardware. However, computing the frequency-domain response from the time-domain response using the typical gate set available on hardware requires a time duration that exceeds the circuit depth limitations of near-term quantum computers. An alternative approach to determine these response properties is by computing them directly in the frequency domain. Frequency-domain algorithms generally involve obtaining the ground-and excited-state energies as well as the transition amplitudes between the ground state and the excited states. Although there are established methods to obtain ground-and excited-state energies on quantum computers, calculating transition amplitudes is less straightforward. Various schemes including variational quantum simulation, quantum subspace expansion and quantum linear algebra to determine frequency-domain response properties have been proposed. However, the accuracy of variational methods depends on the quality of the ansatz, quantum subspace expansion is susceptible to numerical instabilities from basis linear dependence, and quantum linear algebra is out of reach for near-term quantum hardware. Recently, a non-variational scheme amenable to near-term hardware implementation has been proposed. This scheme constructs the electron-added and electron-removed states simultaneously by exploiting the probabilistic nature of the linear combination of unitaries (LCU) algorithm. Nevertheless, this LCU-based algorithm has not yet been demonstrated on quantum hardware due to the lack of efficient implementations of the multi-qubit gates. In this work, we experimentally demonstrate the calculation of frequency-domain response properties of diatomic molecules using a recently reported high-fidelity three-qubit iToffoli gate to implement the LCU circuits on a superconducting quantum processor. The multi-controlled gates present in the original LCU algorithm are decomposed either with iToffoli gates or with CZ gates. We calculate transition amplitudes between the ground state and the \(N\)-electron or \((N\pm 1)\)-electron states of NaH and KH molecules restricted to the highest occupied and lowest unoccupied molecular orbitals (HOMO and LUMO). The transition amplitudes are then used to construct spectral functions and density-density response functions of the diatomic molecules. We apply error mitigation techniques including randomized compiling (RC) during circuit construction, and McWeeny purification during postprocessing, both of which result in marked improvement of the experimental observables. The observables obtained from the reduced-depth circuits with iToffoli decomposition show comparable or better agreement with theory compared to the observables from circuits with CZ decomposition, despite incomplete Pauli twirling in the randomized compiling procedure applied to the iToffoli gate. Our results pave the way for the application of multi-qubit gates for quantum chemistry and related quantum simulation applications on near-term quantum hardware. # Results {#results .unnumbered} ## Quantum Algorithm for Transition Amplitudes of Diatomic Molecules {#quantum-algorithm-for-transition-amplitudes-of-diatomic-molecules .unnumbered} We consider the HOMO-LUMO models of the diatomic molecules NaH and KH as shown in Fig. [\[fig:diatomic_molecules\]](#fig:diatomic_molecules){reference-type="ref" reference="fig:diatomic_molecules"} (see Methods). Such molecular models with reduced active space have been used in benchmarking quantum chemistry methods on quantum computers. The HOMO-LUMO model generates two spatial orbitals or equivalently four spin orbitals, which correspond to four qubits after Jordan-Wigner transformation. To reduce quantum resources, we exploit the number symmetry in each spin sector to reduce the number of qubits from four to two using a qubit-tapering technique (details given in Supplementary Sec. [\[sec:z2_symmetry_transforms\]](#sec:z2_symmetry_transforms){reference-type="ref" reference="sec:z2_symmetry_transforms"}). The observables we aim to determine are the spectral function and density-density response function. Suppose that the molecular Hamiltonian with reduced active space has ground state \(\ket{\Psi_0}\) with energy \(E_0\), and \((N\pm 1)\)-electron eigenstates \(\ket{\Psi_\lambda^{N\pm 1}}\) with energies \(E^{N\pm 1}_\lambda\). Let \(\hat{a}_{p\sigma}^\dagger\) and \(\hat{a}_{p\sigma}\) be the creation and annihilation operators on orbital \(p\) with spin \(\sigma\), respectively. The one-particle Green's function has the expression: \[\begin{aligned} G_{pq}(\omega) &= \sum_{\lambda\sigma} \frac{\langle\Psi_0 |\hat{a}_{p\sigma}| \Psi_{\lambda}^{N+1} \rangle \langle \Psi_{\lambda}^{N+1}|\hat{a}^\dagger_{q\sigma}| \Psi_0\rangle}{\omega + E_0-E_{\lambda}^{N+1} + i\eta} \notag \\ &+ \sum_{\lambda\sigma} \frac{\langle\Psi_0 |\hat{a}_{q\sigma}^\dagger| \Psi_{\lambda}^{N-1} \rangle \langle \Psi_{\lambda}^{N-1}|\hat{a}_{p\sigma}| \Psi_0\rangle}{\omega-E_0 + E_{\lambda}^{N-1} + i\eta} \label{eq:G} \end{aligned}\] where \(\omega\) is the frequency and \(\eta\) is a small broadening factor. The spectral function is related to the Green's function by \(A(\omega) =-\pi^{-1}\Im\, \Tr\, G(\omega)\). For the density-density response function, we consider the charge-neutral \(N\)-electron excited states \(\ket{\Psi_\lambda^N}\) with energies \(E_\lambda^N\) and the number operator \(\hat{n}_{p\sigma}\) on the orbital \(p\) with spin \(\sigma\). The density-density response function has the expression: \[\begin{aligned} R_{pq}(\omega) = &\sum_{\lambda} \frac{\sum_{\sigma\sigma'}\langle\Psi_0|\hat{n}_{p\sigma}|\Psi_\lambda^N\rangle \langle \Psi_\lambda^N|\hat{n}_{q\sigma'}|\Psi_0\rangle}{\omega + E_0-E_\lambda^{N} + i\eta}. \label{eq:R} \end{aligned}\] The operators \(\hat{a}_{p\sigma}, \hat{a}_{p\sigma}^\dagger\) and \(\hat{n}_{p\sigma}\) are not unitary, but they can be written as linear combination of unitary operators as \[\begin{aligned} \hat{a}_{p\sigma} &= (\bar{X}_{p\sigma}-i\bar{Y}_{p\sigma})/2,\\ \hat{a}_{p\sigma}^\dagger &= (\bar{X}_{p\sigma} + i\bar{Y}_{p\sigma})/2,\\ \hat{n}_{p\sigma} &= (I-Z_{p\sigma})/2, \end{aligned}\] where \(I\) is the identity operator, \(Z_{p\sigma}\) is the Pauli \(Z\) operator on orbital \(p\) with spin \(\sigma\), and \(\bar{X}_{p\sigma}\) and \(\bar{Y}_{p\sigma}\) are the Jordan-Wigner transformed Pauli \(X\) and \(Y\) operators on orbital \(p\) with spin \(\sigma\) with a string of \(Z\) operators included to account for the anticommutation relation. The Pauli strings \(\bar{X}_{p\sigma}, \bar{Y}_{p\sigma}\) and \(Z_{p\sigma}\) undergo the same transformation and qubit tapering process as the Hamiltonian (details given in Supplementary Sec. [\[sec:z2_symmetry_transforms\]](#sec:z2_symmetry_transforms){reference-type="ref" reference="sec:z2_symmetry_transforms"}). Except for the identity operator which does not change under the transformation, we label the transformed \(\bar{X}_{p\sigma}, \bar{Y}_{p\sigma}, Z_{p\sigma}\) as \(\tilde{X}_{p\sigma}, \tilde{Y}_{p\sigma}\) and \(\tilde{Z}_{p\sigma}\). The LCU circuits to calculate diagonal and off-diagonal transition amplitudes are given in Figs. [\[fig:diagonal_circuit\]](#fig:diagonal_circuit){reference-type="ref" reference="fig:diagonal_circuit"} and [\[fig:off_diagonal_circuit\]](#fig:off_diagonal_circuit){reference-type="ref" reference="fig:off_diagonal_circuit"}. Each circuit has two system qubits \(s_0\) and \(s_1\), and one ancilla qubit \(a_0\) or two ancilla qubits \(a_0\) and \(a_1\). The unitary \(U_0\) prepares the ground state \(\ket{\Psi_0}\) on the system qubits from the all-zero initial state. The circuit diagrams reflect that the spectral function involving operators \(\tilde{X}_{p\sigma}\) and \(\tilde{Y}_{p\sigma}\) only requires diagonal transition amplitudes, while the density-density response function involving operators \(I\) and \(\tilde{Z}_{p\sigma}\) requires both the diagonal and off-diagonal transition amplitudes. Although the original algorithm proposed performing quantum phase estimation on the system qubits, due to quantum resource constraints we instead apply quantum state tomography to the system qubits while measuring the ancilla qubits in the \(Z\) basis. In the diagonal circuits, we obtain the (unnormalized) system-qubit states \(\frac{1}{2}(\tilde{X}_{p\sigma} \pm i\tilde{Y}_{p\sigma})\ket{\Psi_0}\) or \(\frac{1}{2}(I \pm \tilde{Z}_{p\sigma})\ket{\Psi_0}\) with probabilities \(p_{\pm}\), where the probabilities are specified by the ancilla measurement outcome as \(p_+ = p_{a_0 = 0}\) and \(p_-= p_{a_0 = 1}\); in the off-diagonal circuits, we obtain the (unnormalized) system-qubit states \(\frac{1}{4}[(I-\tilde{Z}_{p\sigma}) \pm e^{i\pi/4}(I-\tilde{Z}_{q\sigma'})]\ket{\Psi_0}\) with probabilities \(p_\pm\), where \(p_+ = p_{(a_0,a_1) = (1, 0)}\) and \(p_-= p_{(a_0, a_1) = (1, 1)}\). We take the overlap of the tomographed system-qubit states with the exact eigenstates, which are then postprocessed according to Eq. 18 in Ref.  or Eq. 25 in Ref.  to yield the transition amplitudes (see Supplementary Sec. [\[sec:transition_amplitudes\]](#sec:transition_amplitudes){reference-type="ref" reference="sec:transition_amplitudes"}). The transition amplitudes are then used to construct the spectral function and density-density response function according to Eqs. [\[eq:G\]](#eq:G){reference-type="ref" reference="eq:G"} and [\[eq:R\]](#eq:R){reference-type="ref" reference="eq:R"}. In the following sections, for simplicity, we will denote the diagonal circuit that applies the operator \(\hat{a}^{(\dagger)}_{p\sigma}\) or \(\hat{n}_{p\sigma}\) as the \(p\sigma\)-circuit, and the off-diagonal circuit that applies the operators \(\hat{n}_{p\sigma}\) and \(\hat{n}_{q\sigma'}\) as the \((p\sigma, q\sigma')\)-circuit. ## iToffoli versus CZ Decomposition in LCU Circuits {#itoffoli-versus-cz-decomposition-in-lcu-circuits .unnumbered} The transformed and tapered operators are two-qubit Pauli strings with multiplicative factors of \(\pm 1\) or \(\pm i\). To apply the single-or double-controlled gates, we follow the standard multi-qubit Pauli gate decomposition with the base gate as CZ or CCZ and use CNOT gates consisting of native CZ gates dressed by Hadamard gates to extend the weights of the Pauli strings. The multiplicative factor \(-1\) or \(\pm i\) can be applied as a single-qubit phase gate on the ancilla in the diagonal circuits, or as the native CS, CS\(^\dagger\) or CZ on the two ancillae in the off-diagonal circuits. Additionally, \(X\) gates are wrapped around the ancilla qubits controlled on the 0 state. Figure [\[fig:doubled_controlled_gate\]](#fig:doubled_controlled_gate){reference-type="ref" reference="fig:doubled_controlled_gate"} shows how a double-controlled gate with ancilla \(a_0\) controlled on 1, ancilla \(a_1\) controlled on 0, and a target operator \(-ZZ\) is applied on the device. We decompose the CCZ gate either with the three-qubit iToffoli gate as shown in Fig. [\[fig:itoffoli_decomposition\]](#fig:itoffoli_decomposition){reference-type="ref" reference="fig:itoffoli_decomposition"} or with the native CZ gates. The iToffoli decomposition starts with a double-controlled \(i\)Z component, followed by a long-range CS\(^\dagger\) gate to cancel the phase factor \(i\). The SWAP gates in the long-range CS\(^\dagger\) part of the circuit are further simplified in the transpilation stage or decomposed into three CZ gates and additional single-qubit gates according to a recent work on the same quantum device. For the CZ decomposition of CCZ, we use a topology-aware quantum circuit synthesis package to obtain the optimal decomposition as eight CZs under linear qubit connectivity, as opposed to the six-CZ decomposition that requires all-to-all qubit connectivity. The spectral function only requires the four diagonal circuits \(0\uparrow, 0\downarrow, 1\uparrow, 1\downarrow\). The density-density response function requires four diagonal circuits \(0\uparrow, 0\downarrow, 1\uparrow, 1\downarrow\) and six off-diagonal circuits \((0\uparrow, 0\downarrow), (0\uparrow, 1\uparrow), (0\uparrow, 1\downarrow), (0\downarrow, 1\uparrow), (0\downarrow, 1\downarrow), (1\uparrow, 1\downarrow)\). We use the same transpilation procedure to optimize the circuits constructed from iToffoli decomposition and CZ decomposition (details given in Methods). The diagonal circuits after transpilation are relatively shallow circuits with maximum circuit depth (excluding virtual \(Z\) gates) of 19, maximum two-qubit gate count of 7 and no iToffoli gates. In the off-diagonal circuits, the circuit depths range from 24 to 29 for iToffoli decomposition and from 54 to 59 for CZ decomposition. As for the two-and multi-qubit gate counts, each iToffoli-decomposed circuit contains two iToffoli gates and 9 to 12 native two-qubit gates, while each CZ-decomposed circuit contains 19 to 21 native two-qubit gates. Hence the iToffoli decomposition results in about half the circuit depth and half the number of two-qubit gates compared to the CZ decomposition. ## Spectral Function and Response Function on Quantum Hardware {#spectral-function-and-response-function-on-quantum-hardware .unnumbered} The spectral function of NaH and KH are shown in Fig. [\[fig:spectral_function\]](#fig:spectral_function){reference-type="ref" reference="fig:spectral_function"}. The density matrices are obtained from quantum state tomography and postprocessed with McWeeny purification. Randomized compiling is not employed for these results. The experimental spectral functions show very good agreement with the exact ones, with maximum peak height deviation of 10.6%. We next turn to the density-density response functions, which are more challenging to determine than the spectral functions because these observables require the deeper off-diagonal circuits containing three-qubit iToffoli gates. We begin by considering a specific off-diagonal circuit needed for the density-density response function, the \((0\uparrow, 0\downarrow)\)-circuit. To understand the influence of the iToffoli gate on the accuracy of the executed circuit, we compute the fidelity of the whole qubit register obtained by quantum state tomography versus circuit depth. The same quantity was computed for a circuit using only CZ gates to decompose the double-controlled gates. The results are shown in Fig. [\[fig:fidelity_trajectory\]](#fig:fidelity_trajectory){reference-type="ref" reference="fig:fidelity_trajectory"}. Although the iToffoli decomposition shows a steeper decrease in fidelity compared to the CZ decomposition, the fidelity at the end of the circuit is higher due to lower circuit depth. The noisy simulation in the inset of Fig. [\[fig:fidelity_trajectory\]](#fig:fidelity_trajectory){reference-type="ref" reference="fig:fidelity_trajectory"} shows a similar trend. The iToffoli gate reported in Ref.  does not consider spectator errors on neighboring qubits, which are cancelled out in the gate calibration in this work (details given in Supplementary Sec. [\[sec:gate_calibration\]](#sec:gate_calibration){reference-type="ref" reference="sec:gate_calibration"}). The cycle benchmarking fidelity of the iToffoli gate accounting for the spectator qubit is 96.1%, lower than the single-qubit gate fidelities which are above 99.5% and the two-qubit gate fidelities which are between 97.6% and 98.8%, which may explain the steeper decay in fidelity with circuit depth in the iToffoli circuit compared to the CZ circuit. Next, we examine the fidelity of the final state in each iToffoli-decomposed circuit used in the calculation of response functions. Figure [\[fig:fidelity_matrix\]](#fig:fidelity_matrix){reference-type="ref" reference="fig:fidelity_matrix"} shows the system-qubit state fidelities on each response function circuit for NaH, where McWeeny purification is applied to the system-qubit density matrix after restricting the full density matrix to each ancilla bitstring sector. Comparing the values in Fig. [\[fig:fidelity_matrix_norc_raw\]](#fig:fidelity_matrix_norc_raw){reference-type="ref" reference="fig:fidelity_matrix_norc_raw"} with those in Fig. [\[fig:fidelity_matrix_rc_raw\]](#fig:fidelity_matrix_rc_raw){reference-type="ref" reference="fig:fidelity_matrix_rc_raw"}, we can see that RC itself only results in a moderate improvement in the fidelities, with the average diagonal fidelities changing from 84.6% to 85.5% and average off-diagonal fidelities changing from 45.2% to 54.8%. However, the results between Fig. [\[fig:fidelity_matrix_rc_raw\]](#fig:fidelity_matrix_rc_raw){reference-type="ref" reference="fig:fidelity_matrix_rc_raw"} and Fig. [\[fig:fidelity_matrix_rc_pur\]](#fig:fidelity_matrix_rc_pur){reference-type="ref" reference="fig:fidelity_matrix_rc_pur"} show that RC combined with purification yield an average diagonal fidelity of 99.9% and an average off-diagonal fidelity of 96.0%, even though purification without RC only leads to a limited improvement in the average diagonal fidelity from 85.6% to 95.7%, and in the average off-diagonal fidelity from 45.2% to 67.4% in Figs. [\[fig:fidelity_matrix_norc_raw\]](#fig:fidelity_matrix_norc_raw){reference-type="ref" reference="fig:fidelity_matrix_norc_raw"} and [\[fig:fidelity_matrix_norc_pur\]](#fig:fidelity_matrix_norc_pur){reference-type="ref" reference="fig:fidelity_matrix_norc_pur"}. We now show the density-density response functions \(\chi_{00}\) and \(\chi_{01}\) of NaH in Fig. [\[fig:response_function\]](#fig:response_function){reference-type="ref" reference="fig:response_function"}. Here \(\chi_{00}\) is obtained from two diagonal circuits \(0\uparrow, 0\downarrow\) and one off-diagonal circuit \((0\uparrow, 0\downarrow)\), while \(\chi_{01}\) is obtained from four off-diagonal circuits \((0\uparrow, 1\uparrow), (0\uparrow, 1\downarrow),(0\downarrow, 1\uparrow),(0\downarrow, 1\downarrow)\). All experimental results are postprocessed with purification after constraining the ancilla qubits to each bitstring subspace. Overall, the iToffoli decomposition yields comparable or better results compared to the CZ decomposition. Without RC, the iToffoli decomposition results in visibly better agreement with the exact response functions in Figs. [\[fig:chi00_norc\]](#fig:chi00_norc){reference-type="ref" reference="fig:chi00_norc"} and [\[fig:chi01_norc\]](#fig:chi01_norc){reference-type="ref" reference="fig:chi01_norc"}, where the root-mean-squared (RMS) errors between the experimental and exact response functions are 0.0086 eV\(^{-1}\) and 0.0372 eV\(^{-1}\) for the iToffoli decomposition and 0.0387 eV\(^{-1}\) and 0.0278 eV\(^{-1}\) for the CZ decomposition for \(\chi_{00}\) and \(\chi_{01}\), respectively. With RC, the two decompositions exhibit comparable RMS errors, which are 0.0081 eV\(^{-1}\) and 0.0143 eV\(^{-1}\) for the iToffoli decomposition, and are 0.0073 eV\(^{-1}\) and 0.0061 eV\(^{-1}\) for the CZ decomposition for \(\chi_{00}\) and \(\chi_{01}\), respectively, as shown in Figs. [\[fig:chi00_rc\]](#fig:chi00_rc){reference-type="ref" reference="fig:chi00_rc"} and [\[fig:chi01_rc\]](#fig:chi01_rc){reference-type="ref" reference="fig:chi01_rc"}. In both \(\chi_{00}\) and \(\chi_{01}\), RC can lead to substantial improvements in the experimental data for both decompositions, which is reflected in the heights of the symmetric peaks at \(\pm 1.4\) eV and \(\pm 24.0\) eV. For \(\chi_{00}\) without RC and using the CZ decomposition, the peak at \(\pm 1.4\) eV is twice the height of the exact peak and the peak at \(\pm 24\) eV is not present in the experimental observable in Fig. [\[fig:chi00_norc\]](#fig:chi00_norc){reference-type="ref" reference="fig:chi00_norc"}. With RC, both features are present, with the deviations of the peak heights being 34.8% and 4.7%, respectively, as in Fig. [\[fig:chi00_rc\]](#fig:chi00_rc){reference-type="ref" reference="fig:chi00_rc"}. For \(\chi_{00}\) with the iToffoli decomposition, the peak height deviations change from 6.1% and 26.6% without RC in Fig. [\[fig:chi00_norc\]](#fig:chi00_norc){reference-type="ref" reference="fig:chi00_norc"} to 11.8% and 24.0% with RC in Fig. [\[fig:chi00_rc\]](#fig:chi00_rc){reference-type="ref" reference="fig:chi00_rc"}. In this case, the results before and after RC are comparable since the experimental observable without RC is already in good agreement with the exact observable. For \(\chi_{01}\), we see a more marked improvement due to RC compared to the corresponding results for \(\chi_{00}\). Without RC, both iToffoli and CZ decompositions produce the wrong sign on the peak at \(\pm 24\) eV in Fig. [\[fig:chi01_norc\]](#fig:chi01_norc){reference-type="ref" reference="fig:chi01_norc"}, but with RC the signs are correctly predicted. The peak height deviations can then be computed as 39.2% and 32.2% for the CZ decompositions and 5.7% and 28.2% for the iToffoli decomposition, as indicated in Fig. [\[fig:chi01_rc\]](#fig:chi01_rc){reference-type="ref" reference="fig:chi01_rc"}. The corresponding results of system-qubit state fidelities and density-density response functions for KH are given in Supplementary Sec. [\[sec:addtional_data_for_kh\]](#sec:addtional_data_for_kh){reference-type="ref" reference="sec:addtional_data_for_kh"}, which follow a similar trend as NaH. Since the iToffoli gate is non-Clifford, our implementation of RC results in incomplete Pauli twirling compared to applying RC to the CZ-decomposed circuits (see Supplementary Sec. [\[sec:randomzied_compiling\]](#sec:randomzied_compiling){reference-type="ref" reference="sec:randomzied_compiling"}). The incompleteness of RC on the iToffoli-decomposed circuits may explain why the two decompositions have similar RMS errors with RC despite the initial advantage for the iToffoli decomposition without RC due to its lower circuit depth. # Discussion {#discussion .unnumbered} We have carried out an LCU-based algorithm to compute the spectral functions and density-density response functions of diatomic molecules from the transition amplitudes determined on a superconducting quantum processor. Using a native high-fidelity iToffoli gate has enabled the required circuit depth to be reduced by around a factor of two. These resulting circuits produced better agreement with the exact results compared to the circuits constructed only from CZ, CS, and CS\(^\dagger\) gates when RC is not employed. We also developed an RC protocol for the non-Clifford iToffoli gate, and have shown that without complete Pauli twirling on the iToffoli gate, the circuits constructed from iToffoli gates gave comparable results as the circuits constructed only from CZ, CS, and CS\(^\dagger\) gates with RC. The quality of the computed quantities was greatly improved by the use of several error mitigation techniques. Specifically, our results highlight the significance of RC combined with McWeeny purification for quantum simulation. McWeeny purification has been widely used in quantum chemistry and started to be exploited in quantum computing for constraining the purity of the output state. Our results have showed that RC or McWeeny purification individually only improves the experimental results to a limited extent, as observed in the change of the average off-diagonal fidelities from 45.8% to 54.2% with only RC, and to 67.4% with only purification in Fig. [\[fig:fidelity_matrix\]](#fig:fidelity_matrix){reference-type="ref" reference="fig:fidelity_matrix"}. However, the combination of RC and purification results in a significant improvement in the quality of the results with the system-qubit state fidelities being 96.0% on average. Moreover, previous works applied purification to the whole qubit register, but we have showed here that the purification scheme can be applied when there is purity constraint on a subset of qubits. Additionally, our work is the first to apply RC to the non-Clifford iToffoli gate. As more native non-Clifford two-qubit and multi-qubit gates become available, our findings may guide future application of RC to non-Clifford gates. Our work is also among the first to demonstrate the practical use of a native multi-qubit gate in quantum simulation via an LCU-based algorithm. LCU as a general algorithmic framework is not limited to determining transition amplitudes in frequency-domain response properties, but has broader applications in areas such as solving linear systems, simulating non-Hermitian dynamics, and preparing quantum Gibbs states. Besides the LCU algorithm, quantum algorithms such as Shor's algorithm and Grover's search algorithm can benefit considerably from native three-qubit gates with reduction in circuit depths and gate counts. Quantum algorithm design and implementation thus far have been mostly restricted to single-and two-qubit gates due to their ease of implementation and demonstrated high fidelity. Meanwhile, early implementations of three-qubit gates were generally slower and more prone to leakage and decoherence compared to the iToffoli gate employed here due to populating higher levels outside the qubit computational space. However, more recent implementations of three-qubit gates have begun to address these challenges yielding fidelities approaching those achieved with two-qubit gates. Further, they have been carried out on quantum devices with tens of qubits, suggesting their utility for larger-scale quantum devices. As such native multi-qubit gates become more prevalent, our work paves the way for using them as native gate components in future quantum algorithm design and implementation. # Methods {#methods .unnumbered} ## Quantum Circuit Construction {#quantum-circuit-construction .unnumbered} The molecular orbitals used in this work are in STO-3G basis with molecular integrals determined from PySCF. We use OpenFermion to map the second-quantized Hamiltonian to qubit operators. The ground-state preparation gate on the system qubits are determined classically by constructing a unitary that maps the all-zero initial states to the ground state and then decomposed into three CZ gates and single-qubit gates using the \(KAK\) decomposition. The LCU circuits are then constructed by applying the gates shown in Figs. [\[fig:diagonal_circuit\]](#fig:diagonal_circuit){reference-type="ref" reference="fig:diagonal_circuit"} and [\[fig:off_diagonal_circuit\]](#fig:off_diagonal_circuit){reference-type="ref" reference="fig:off_diagonal_circuit"}, where the SWAP gates are decomposed according to the scheme in Ref.  and the circuits are transpiled by the functions `MergeInteractions`, `MergeSingleQubitGates` and `DropEmptyMoments` in Cirq. The transition amplitudes are then combined with the classically determined ground-and excited-state energies to calculate the spectral functions and response functions (See Supplementary Sec. [\[sec:transition_amplitudes\]](#sec:transition_amplitudes){reference-type="ref" reference="sec:transition_amplitudes"}). ## Quantum Device {#quantum-device .unnumbered} The quantum device used in this work is a superconducting quantum processor with eight transmon qubits. The algorithm is performed on a four-qubit subset of the device with linear connectivity. Single-qubit gates are performed with resonant microwave pulses. Multiplexed dispersive readout allows for simultaneous state discrimination on all four qubits. CZ gates between all nearest neighbors are performed according to the method in Ref. . The same method allows for a native CS gate on one pair and a native CS\(^{\dagger}\) gate on a different pair, according to the requirements of the algorithm. While single-qubit gates are applied simultaneously, microwave crosstalk requires that all two-and three-qubit gates are applied in separate cycles from each other as well as from any single-qubit gates. TrueQ is used for circuit manipulations in the implementation of RC as well as gate benchmarking. Internal software is used to map the circuits to hardware pulses for implementing the native gate set. # Author contributions {#author-contributions .unnumbered} S.-N.S. and A.J.M. conceptualized the project. S.-N.S., B.M., and J.M.Koh contributed to the codebase. B.M., L.B.N., L.C., and Y.K. performed the device calibration and hardware runs of the circuits. J.M.Kreikebaum fabricated the device. S.-N.S. and B.M. analyzed the data. All authors contributed to discussion of the manuscript.

{'timestamp': '2023-02-09T02:18:09', 'yymm': '2302', 'arxiv_id': '2302.04271', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04271'}

# Abstract {#abstract .unnumbered} Antibodies and humoral memory are key components of the adaptive immune system. We consider and computationally model mechanisms by which humoral memory present at baseline might instead increase infection load; we refer to this effect as EI-HM (enhancement of infection by humoral memory). We first consider antibody dependent enhancement (ADE) in which antibody enhances the growth of the pathogen, typically a virus, and typically at intermediate 'Goldilocks' levels of antibody. Our ADE model reproduces ADE *in vitro* and enhancement of infection *in vivo* from passive antibody transfer. But notably the simplest implementation of our ADE model never results in EI-HM. Adding complexity, by making the cross-reactive antibody much less neutralizing than the *de novo* generated antibody or by including a sufficiently strong non-antibody immune response, allows for ADE-mediated EI-HM. We next consider the possibility that cross-reactive memory causes EI-HM by crowding out a possibly superior *de novo* immune response. We show that, even without ADE, EI-HM can occur when the cross-reactive response is both less potent and 'directly' (i.e. independently of infection load) suppressive with regard to the *de novo* response. In this case adding a non-antibody immune response to our computational model greatly reduces or completely eliminates EI-HM, which suggests that 'crowding out' is unlikely to cause substantial EI-HM. Hence, our results provide examples in which simple models give qualitatively opposite results compared to models with plausible complexity. Our results may be helpful in interpreting and reconciling disparate experimental findings, especially from dengue, and for vaccination. # Author summary {#author-summary .unnumbered} Humoral memory, generated by infection with a pathogen, can protect against subsequent infection by the same and related pathogens. We consider situations in which humoral memory is counterproductive and instead enhances infection by related pathogens. Numerous experiments have shown that the addition of binding antibody makes certain viruses more, rather than less, infectious; typically high levels of antibody are still protective, meaning that infectivity is maximized at intermediate 'Goldilocks' levels of antibody. Additionally, we consider the situation in which cross-reactive humoral memory dominates relative to the *de novo* response. Memory dominance has been documented for influenza infections, but whether it is harmful or not is unclear. We show computationally that both ADE and a certain type of competition between the cross-reactive and *de novo* responses translate into enhancement of infection in certain circumstances but not in others. We discuss the implications of our research for dengue infection, a common mosquito-transmitted viral infection, and vaccination. # Introduction {#introduction .unnumbered} Antibody is a key component of the adaptive immune system. Antibodies bind to viruses, bacteria, and other microbes and thereby inhibit microbe function and aid in microbe clearance. For certain infections, antibody levels are considered the key correlate of protection  , meaning that increasing antibody levels are associated with protection from infection or disease. Paradoxically for certain viral infections, antibody can instead enhance infection-an effect known as antibody dependent enhancement (ADE). ADE has now been reported for many viruses including dengue, West Nile, HIV, influenza A, Ebola, rabies, polio, and HSV1  . It is generally believed that ADE results when antibody bound to virus increases the ability of that virus to infect certain cells, like macrophages, that possess Fc receptors  . In the case of dengue, it has also been reported that maternal antibodies can enhance disease severity in infants    . Consistent with this observation, transfer of dengue antibodies to dengue-infected monkeys can greatly increase viral load    . A related concern to ADE is that cross-reactive humoral memory from previous infections or vaccinations may enhance the subsequent infection    . This is particularly a concern for dengue infection which is caused by four serotypes of virus -- DENV1, DENV2, DENV3, and DENV4 -- that are only  65% similar at the amino acid level  . Certain studies have reported that infection with one serotype of dengue, for example DENV1, may make infection with a second serotype, for example DENV2, more severe probably as a consequence of ADE      . In contrast, the discussion surrounding influenza has focused more on the issue of original antigenic sin  . According to the original antigenic sin model, the immune response remains focused on the first influenza strain encountered even after multiple subsequent influenza infections. Original antigenic sin is not necessarily bad, and certain authors have argued that antigenic seniority is a more apt term  . ## Definitions {#definitions .unnumbered} In this paper we use the following definitions: *Antibody dependent enhancement* (ADE) means that viral growth is enhanced at some level of antibody compared to no antibody. Mechanistically, we model this via higher infectivity of virions with some bound antibody *Enhancement of infection from humoral memory* (EI-HM) means that infection load is greater in the presence of cross-reactive humoral memory than in its absence. While there are many ways to measure infection load, in this paper we consistently use the peak number of infected cells. *Enhancement of infection from passive antibody* (EI-PA) means that infection load is greater when passive antibody -- for example maternal antibody -- is supplied to a host than in its absence. (We assume that this host has an adaptive immune system and will generate a humoral response of its own.) *Memory dominance* means that the cross-reactive memory response dominates the *de novo* immune response. Memory dominance may be viewed as a proxy for original antigenic sin (see Supplement for further explanation). *Suppressive memory* means that cross-reactive memory "directly" suppresses the *de novo* immune response ignoring any "indirect" effects mediated via infection load. See Fig [\[fig1A\]](#fig1A){reference-type="ref" reference="fig1A"} for a conceptual diagram of these differences. # Model and Results {#model-and-results .unnumbered} ## ADE Model {#ade-model .unnumbered} To investigate the mechanisms depicted in Fig [\[fig1A\]](#fig1A){reference-type="ref" reference="fig1A"}, we create a mathematical model of the essential infection components. In this model, infected cells (\(I\)) produce virus. \(V_1\) is virus with little or no bound antibody. \(V_2\) is virus with intermediate amounts of bound antibody. Virus with the most antibody bound, \(V_3\), is neutralized. We assume that \(V_2\) virus is more infectious than \(V_1\) virus (i.e. \(\beta_2 > \beta_1\)). \(A_c\) is cross-reactive memory antibody. \(A_n\) is the *de novo* antibody response. \(A_p\) is passive antibody (e.g. maternal antibody). \(R\) is the non-antibody (e.g. innate or CD8 T cell) immune response. We assume that only a small fraction of target cells are depleted during the course of infection; hence, we do not explicitly model loss of susceptible cells. We show this model set up via a schematic in  . The mechanisms we consider are described by the following set of ordinary differential equations. \[\begin{aligned} \dot{I} &= \underbrace{-bI}_{\text{cell death}}+\underbrace{\beta_1V_1+\beta_2V_2}_{\text{infection}}-\underbrace{IR}_{\text{non-Ab clearance}}\\ \dot{V_1} &= \underbrace{I}_{\text{production}}-V_1(\underbrace{a}_{\text{clearance}}+\underbrace{k_{c1}A_c+k_{n1}A_n+k_{p1}A_p}_{\text{binding by antibodies}})\\ \dot{V_2} &= \underbrace{V_1(k_{c1}A_c+k_{n1}A_n+k_{p1}A_p)}_{\text{newly partially bound}}-V_2(\underbrace{a}_{\text{clearance}}+\underbrace{k_{c2}A_c+k_{n2}A_n+k_{p2}A_p)}_{\text{further binding}}\\ \dot{V_3} &= \underbrace{V_2(k_{c2}A_c+k_{n2}A_n+k_{p2}A_p)}_{\text{newly completely bound}}-\underbrace{aV_3}_{\text{clearance}}\\ \dot{A_c} &= \underbrace{s_c}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}}A_c\\ \dot{A_n} &= \underbrace{s_n}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}}A_n\\ \dot{A_p} &= \underbrace{0}_{\text{no growth}} \\ \dot{R} &= \underbrace{s_R}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}} \end{aligned}\] Here, \(\chi\) is the indicator function. Table [2](#table1){reference-type="ref" reference="table1"} describes the model parameters. Unless otherwise stated we use the default parameter values shown in the table. Model parameters were chosen to give an asymptotic growth of 1.5 natural logs (ln) per day in the absence of antibody and 2.5 ln per day at maximum ADE and peak infection load at 10 days. For comparison the growth of the yellow fever YFV-17D virus was estimated at 1.6 ln per day. These dynamics are shown in Fig S2-7 for each of the models. [\[table1\]]{#table1 label="table1"} ## The model recapitulates *in vitro* ADE {#the-model-recapitulates-in-vitro-ade .unnumbered} In this section, we simulate a 48 hour *in vitro* cell culture experiment. As there is no active immune response, \(A_c(0)=A_n(0)=0\). Fig [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} shows the quantity of infected cells at the end of the simulation for different values of \(A_p\) and compares them to experimental values. An antibody level of 15 produces maximum ADE in the model; this level of antibody produces 10 fold greater infection load after 48 hours as compared to no antibody. Hence the model described above does produce ADE. Higher levels of antibody (\(>74\)) reduce infection growth as compared to no antibody. This pattern of enhanced growth at intermediate levels of antibody and reduced or no growth at higher levels of antibody matches the pattern seen from *in vitro* experiments with dengue   and West Nile   viruses. ## The model produces EI-PA {#the-model-produces-ei-pa .unnumbered} Building upon our previous *in vitro* model to replicate an *in vivo* experiment, we now include additional factors. First, we simulate supply of passive (e.g. maternal) antibody (\(A_p\)) to an infected host. Second, in addition to these passive antibodies, the host's immune system will generate antibodies (\(A_n\)) in response to the infection. Fig [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} shows peak infected cells for different values of \(A_p\). In this case, a passive antibody level of 11 produces maximum enhancement; this level of passive antibody increases infection load 35 fold compared to no passive antibody. Thus, in this model, adding passive antibody can dramatically enhance infection, but higher levels of passive antibody are protective. Hence, the model qualitatively reproduces dengue data from infants     which show that high levels of maternal antibody are protective but intermediate levels are associated with enhanced disease. The model is also consistent with dengue animal studies showing enhanced viral load when passive antibody is supplied    . ## The model does not (necessarily) produce EI-HM {#the-model-does-not-necessarily-produce-ei-hm .unnumbered} In this section, we simulate infection for varying baseline levels of cross-reactive memory antibody. We build upon our original *in vitro* model and now include both *de novo* and cross-reactive antibody response to create an *in vivo* model. Because cross-reactive antibodies may replicate some but not necessarily all the features of *de novo* antibodies such as growth rate or neutralizing activities, we first investigate the simplest scenario where cross-reactive and *de novo* antibodies have all of the same properties. Fig [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} shows peak infected cells for different values of \(A_c(0)\). In this case, we see only protection as \(A_c(0)\) increases. Hence ADE, where partially bound virus is more infectious than free virus, does not necessarily produce EI-HM. ## The model can produce EI-HM in certain circumstances {#the-model-can-produce-ei-hm-in-certain-circumstances .unnumbered} Although the simplest implementation of our *in vivo* model (shown above) does not show enhancement of infection from humoral memory, our model produces EI-HM in certain situations as shown in the following sections. ### Cross-reactive antibody is less neutralizing {#cross-reactive-antibody-is-less-neutralizing .unnumbered} In these simulations the cross-reactive antibody grows at the same rate but is less neutralizing than the *de novo* antibody -- either \(k_{c2}=0.125\) or \(k_{c2}=0.05\). \(k_{c2}=0.125\) can be loosely interpreted as meaning that the cross-reactive antibody is half as neutralizing as the *de novo* antibody, whereas \(k_{c2}=0.05\) can be loosely interpreted as meaning that the cross-reactive antibody is 5 times less neutralizing than the *de novo* antibody. Fig [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} shows peak infected cells for different values of \(A_c(0)\). With \(k_{c2}=0.125\), no EI-HM was observed, but with \(k_{c2}=0.05\), there is some enhancement of infection at low levels of \(A_c(0)\). However, the degree of enhancement is again relatively small -- a 5 fold difference in neutralization ability translates into a maximum enhancement of only 2.7 fold compared to \(A_c(0)=0\). Increasing \(k_{c1}\), the rate at which the cross-reactive antibody enhances virus, has an analogous effect to decreasing \(k_{c2}\). \(k_{c1}=2k_{n1}=2\) does not produce EI-HM whereas \(k_{c1}=5k_{n1}=5\) produces maximum relative enhancement of 2.7 fold. ### Non-antibody immune response is dominant {#non-antibody-immune-response-is-dominant .unnumbered} In these simulations we vary the growth rate, \(s_R\), of the non-antibody immune response. To compare, \(s_R=0\) shows the peak of infected cells when there is no non-antibody immune response and is therefore equivalent to the results in Fig [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. Fig [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} shows peak infected cells for different values of \(A_c(0)\). A growth rate of \(s_R=0.26\) is the highest value of \(s_R\) that shows no EI-HM whereas \(s_R=0.64\) corresponds to primary infection peaking at day 5 (see Fig S7), which is many days before antibody reaches sterilizing levels. With \(s_R=0.64\), enhancement of up to 11 fold is possible. However, the range of \(A_c(0)\) values that lead to enhancement is quite a bit smaller than observed *in vitro*, \(>0\) to 74 compared to \(>0\) to 24. ## Other models {#other-models .unnumbered} We consider three changes to our overarching ADE model: 1) B cells and plasma cells are explicitly included in the model, 2) step functions in the differential equations are replaced with Hill functions, and 3) dissociation between antibody and virus is explicitly modeled. See Supplement for a more detailed description of these changes. None of these changes qualitatively change our results (figures not shown). Explicitly modeling the dissociation between antibody and virus also allows us to consider the situation where the cross-reactive antibody has a higher dissociation rate than the *de novo* antibody. This situation is similar to that described in § Cross-reactive antibody is less neutralizing. When the dissociation rate of cross-reactive antibody is 50/day versus 10/day for the *de novo* antibody, no EI-HM is observed. But when the dissociation rate of cross-reactive antibody is 150/day versus 10/day for the *de novo* antibody, there is some EI-HM at low levels of \(A_c(0)\) but with maximum enhancement of 1.1 fold. See for more details. ## Suppressive memory can produce EI-HM without ADE {#suppressive-memory-can-produce-ei-hm-without-ade .unnumbered} All of our above simulations show memory dominance when \(A_c(0)\) is sufficiently large -- \(A_c(0)>0.036\). This shows that memory dominance does not necessarily mean EI-HM. In fact, with our model ADE (\(\beta_2 >\beta_1\)) is necessary -- but not sufficient -- to produce EI-HM. Our ADE model does not incorporate suppressive memory. In this section we modify the equations for \(A_c\) and \(A_n\) to incorporate this effect. \[\dot{A_c}= s_c\chi(I>1)\frac{\phi A_c}{\phi+A_c+A_n}\] \[\dot{A_n}= s_n\chi(I>1)\frac{\phi A_n}{\phi+A_c+A_n}\] Here \(\chi\) is the indicator function, \(\phi=28\), and \(s_c=s_n=1.5\). In this system increasing \(A_c\) suppresses the growth rate of \(A_n\) and vice versa. We consider the situation where \(A_c\) is a quarter as potent as \(A_n\) (\(k_{c1} =0.25, k_{c2}=0.0625\)) and there is no ADE (\(\beta_1=\beta_2=40\)). In this case there is dramatic EI-HM with fold enhancement of up to 117 fold (Fig [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}). However, adding a non-antibody immune response with \(s_R=0.27\) or \(s_R=0.54\) mostly or completely abolishes EI-HM. Here \(s_R=0.54\) corresponds to primary infection peaking at day 5 (Fig S8). ## Other adverse effects of humoral memory {#other-adverse-effects-of-humoral-memory .unnumbered} EI-HM is one possible adverse effect of humoral memory. Another possible adverse effect is a negative effect of humoral memory at baseline on humoral memory post infection. Such an effect is possible because of the negative feedback between antibody and pathogen and does not require ADE or suppressive memory (see  ). Hence, if a class of antigenically similar pathogens can infect more than once, humoral memory may protect against the immediate infection but enhance later infection. # Discussion and Conclusion {#discussion-and-conclusion .unnumbered} ## Intuitive Explanation {#intuitive-explanation .unnumbered} While our results may seem disparate, they can be explained by two factors: 1) the effect of antibody early in infection and 2) the effect of cross-reactive humoral memory on the rapidity of humoral response. In the case of ADE, antibody may enhance viral growth early in infection, but humoral memory may also accelerate the development of protective levels of antibody. In the case of suppressive memory, viral growth is reduced early in infection, but the development of a potent antibody response may also be delayed. Depending on which of these factors dominates, cross-reactive humoral memory may either enhance or decrease infection. ## Concluding Remarks {#concluding-remarks .unnumbered} Our results are consistent with the hypothesis that ADE and enhancement of infection from passive antibody (EI-PA) can result when virus with intermediate levels of antibody is more infectious than virus with lower or higher levels of antibody  . Hence additional mechanisms, such as suppression of the immune response  , are possible but not necessary to produce the phenomena of ADE and EI-PA. More importantly we show that ADE, EI-PA, and enhancement of infection from cross-reactive humoral memory (EI-HM) are distinct. ADE in our models always implies the possibility of EI-PA. Furthermore, the range of antibody levels that produces ADE is similar to the range of passive antibody levels that produces EI-PA. This is because passive antibody can enhance viral growth early in infection but does very little to accelerate the development of protective levels of antibody. In contrast EI-HM is a double-edged sword, and ADE and EI-PA may not imply EI-HM. In our models, when ADE does produce EI-HM the range of baseline antibody levels that produce EI-HM is very different from the range that produces ADE. This difference in ranges may explain the observation in dengue patients that ADE measurements of pre-infection plasma did not positively correlate with peak viremia  . ADE, EI-PA, and EI-HM are of particular concern for dengue infection. In the case of dengue, there is extensive evidence for ADE *in vitro*          , and EI-PA of up to 100 fold has been demonstrated in monkey experiments    . In contrast, the evidence for greater infection load in secondary infection is more limited. A large monkey experiment involving 118 rhesus macaques showed higher viral load in secondary DENV2 infection as compared to primary DENV2 infection. However, in the same study secondary infections with DENV1, DENV3 and DENV4 showed reduced viremia when compared to the respective primary infections  . A recent epidemiological study showed increased risk of severe dengue in children with low or intermediate levels of dengue antibody at baseline as compared to dengue seronegative children; higher levels of baseline antibody were associated with reduced risk. In contrast, risk of symptomatic dengue decreased essentially monotonically with antibody levels  . Likewise analysis of clinical trials data suggested that vaccination of dengue naive children with the CYD-TDV dengue vaccine increased the risk of severe dengue but reduced the risk of symptomatic dengue. Our results suggest that this pattern could be explained by a scenario in which the majority do not experience EI-HM but a minority -- for example those with very poor quality cross-reactive antibody -- do experience EI-HM and this minority also tends to have more severe disease. However,   and examined disease severity rather than infection load, and in   the degree of enhancement was highly sensitive to the method of classifying disease severity. (Using the 1997 World Health Organization (WHO) dengue severity classification system, the risk of severe dengue was up to 7.6 fold elevated. Using the newer 2009 WHO severity classification, generally regarded as more accurate    , this was reduced to only 1.75 fold.) Despite this limitation, in the case of human dengue infections, our results support the hypothesis that ADE sometimes produces EI-HM but often does not. We show that, in principle, memory dominance plus suppressive memory can also produce EI-HM. In contrast memory dominance alone does not produce EI-HM in any of our models. However, inclusion of a non-antibody immune response in our simulations dramatically reduces EI-HM from suppressive memory, which casts doubt on its real world relevance. It should be emphasized that our simulations only consider the effect of humoral memory present at baseline on that particular infection. A negative effect of humoral memory at baseline on humoral memory post infection can occur as a consequence of the negative feedback between antibody and pathogen and does not require ADE or suppressive memory.  For infections, like influenza, where reinfections are common, some of the apparent adverse effects of humoral memory likely involve not EI-HM but the negative effect of humoral memory at an earlier time point on humoral memory at a latter time point with potentially many infections in between  . Although CD4 T cells can increase pathology  , a mechanism such as ADE is not established for T cells. This implies that activating non-antibody immune responses, especially CD8 T cells, could be an important overlooked component in vaccine development. Although memory dominance and suppressive memory are concerns for antibody as well as T cells, our results suggest that suppressive memory is unlikely to cause EI-HM. Moreover, the ability to strategically direct T cell responses has been demonstrated   , potentially turning memory dominance into a long term advantage. Finally, our results also illustrate potential pitfalls from mathematical modeling, which may be relevant not only for biology but also for other disciplines, such as economics and epidemiology, that use abstract mathematical models. In the simplest version of our models, there is EI-HM from suppressive memory but none from ADE, whereas adding plausible complexity reverses this situation. # Supplement {#supplement .unnumbered} #### S1 Text. {#S1 Text .unnumbered} **Model Dynamics and Alternative Model Results.** #### Fig S1. {#S1_Fig .unnumbered} **Model Schematic.** In our model, infected cells (\(I\)) produce free virus(\(V_1\)) and can be removed via cell death or the non-antibody immune response (\(R\)). Virus can be successively bound with only free (\(V_1\)) and partially bound virus (\(V_2\)) able to infect cells. Binding can come from the *de novo* antibody response (\(A_n\)), the cross-reactive antibody response (\(A_c\)), or the passive antibody response (\(A_p\)), with \(A_n\) and \(A_c\) growing in response to the infected cells. #### S2 Fig. {#S2_Fig .unnumbered} **Dynamics of infection under ADE ***in vitro***.** In this *in vitro* situation, where there is no growth in immune response, the phenomena of ADE is evident as the number of infected cells after 48 hours is greater at intermediate ranges of antibody. We show the differences in dynamics for three different initial antibody concentrations in Panels B and C: low but non-zero (black), intermediate (red), and high (blue). #### S3 Fig. {#S3_Fig .unnumbered} **Dynamics of infection given passive antibodies.** When passive antibodies are given and a *de novo* antibody response is mounted, again ADE appears at intermediate levels (red). #### S4 Fig. {#S4_Fig .unnumbered} **Dynamics of infection given cross-reactive antibodies does not necessarily produce enhancement of infection in these simulations where ***de novo*** and cross-reactive antibodies are identical.** When there are both *de novo* and cross-reactive antibodies present and growing and binding at the same rate, enhancement of infection is not seen, as shown in Panel A; rather, the greater initial cross-reactive antibody load the lower the viral load and number of infected cells as seen in Panels B-C. Here, black represents low initial cross-reactive antibody concentration, red intermediate, and blue high. #### S5 Fig. {#S6_Fig .unnumbered} **Less neutralizing cross-reactive antibodies can sometimes cause EI-HM.** If cross-reactive antibodies are approximately a fifth as good as *de novo* antibodies, some EI-HM may occur as shown in Panel A with slightly higher peak viremia and infected cell counts. While the the line for the low cross-reactive antibody concentration is the same for both \(k_{c2}\) values, as seen in black in Panels B and C, they differ for the intermediate (red) and high (blue) concentrations where \(k_{c2} = 0.125\) is given by the solid lines and \(k_{c2} = 0.05\) by the dashed lines. #### S6 Fig. {#S6_Fig .unnumbered} **Dynamics of infection given cross-reactive antibodies and non-antibody immune response.** Depending on the rate of growth of non-antibody immune response and initial cross-reactive antibody amount, EI-HM can sometimes occur and give a 10-fold increase in peak infected cells. Colors denote level of initial cross-reactive antibody with black, red, and blue representing low, intermediate, and high values. Line types denote non-Ab immune response growth rates with solid, dashed, and x'd lines representing \(s_R=0\), \(s_R= 0.26\), and \(s_R=0.64\) respectively. #### S7 Fig. {#S7_Fig .unnumbered} **Dynamics of infection given suppressive memory.** As shown in Panel A given the scenario where there is no non-antibody immune response, EI-HM can occur with a fold change of \(117\) for \(s_R=0\). With non-antibody immune responses, however, enhancement of infection is greatly reduced with a maximum fold change of \(1.2\) for \(s_R=0.27\) or nonexistent for \(s_R = 0.54\). In Panel B, we consider the viral dynamics for low (black), intermediate (red), and high (blue) points for each of the different growth rates where \(s_R=0\) is a solid line, \(s_R=0.27\) is dashed, and \(s_R=0.54\) is x'd. Note that for \(s_R=0\) in particular if the high point was taken at an even higher antibody level (\(>25\)), it would in fact peak below the low point. #### S8 Fig. {#S9_Fig .unnumbered} **\(\mathbf{k_{\text{off,c}}=100/}\)day produces some EI-HM.** The figure shows simulation results from a modification to the simple ADE model such that dissociation between antibody and virus is explicitly modeled. When the cross reactive antibody has 5 times the dissociation rate of the de novo antibody (\(k_{\text{off,c}}=50/\text{day}\) versus \(k_{\text{off,n}}=10/\text{day}\)), there is no EI-HM. But, when the cross reactive antibody has 10 times or more the dissociation rate of the de novo antibody (\(k_{\text{off,c}}=100/\text{day}\) or \(k_{\text{off,c}}=150/\text{day}\) and \(k_{\text{off,n}}=10/\text{day}\)), there is some enhancement of infection at low levels of baseline cross reactive antibody. #### S9 Fig. {#S10_Fig .unnumbered} **Effect of baseline humoral memory on postinfection antibody levels..** In this model, without ADE or suppressive memory, and with non-antibody immune response we see that higher levels of baseline antibody (\(>\)`<!-- -->`{=html}5.8) negatively affect final (post infection) antibody levels. # Memory Dominance as a Proxy for Original Antigenic Sin {#memory-dominance-as-a-proxy-for-original-antigenic-sin .unnumbered} Original antigenic sin (OAS) is the situation in which the immune response remains stronger towards strains encountered earlier in life compared to strains encountered later in life. This is often seen in the context of influenza where different birth cohorts experience different strains of influenza in their youth as different influenza variants arise, compete, and disappear. Because we wanted to disentangle the nuances of some of the many potential effects of prior immunity, we focus on memory dominance instead of the more complex phenomenon of OAS. However, memory dominance can be viewed as a proxy for OAS. We consider memory dominance to occur when \(A_c > A_n\) with Ac/An indicating the degree of memory dominance. Using our default parameters, if we consider the effective immune response from the antibodies to the *de novo* infection, then we can represent the change as \(\Delta A_c+\Delta A_n\) where \(\Delta A_c\) is the change in \(A_c\) and \(\Delta A_n\) is the change in \(A_n\) both relative to baseline.  We assume the effective immune response from the antibodies to the original infection is boosted by \(a\Delta A_c+b\Delta A_n\) where \(a\geq1\) and \(b\leq1\). If \(A_c >> A_n\) the boost to the old strain will be approximately as large or larger than the boost to the new strain leading to the OAS phenomenon. # Modeling Details {#modeling-details .unnumbered} We simulate various scenarios with ADE that may cause enhancement of infection, including the effects of passive antibodies, cross-reactive antibodies, and non-antibody immune responses. We utilize a system of ordinary differential equations to compare how various initial antibody amounts affect the peak number of infected cells. A schematic of this system is shown below in Fig  [\[SI1\]](#SI1){reference-type="ref" reference="SI1"} and corresponds to Equations 1-8 in the main text. # Behavior of Main Text Models {#behavior-of-main-text-models .unnumbered} For our models we considered relevant biological regimes and matched the basic dynamics of flavivirus infections. In reality the immune response is more complex and dependent on both individual infection history along with underlying frailties, and therefore exactly matching all possible experiments and outcomes is unfeasible. We begin by considering the simple *in vitro* case where we compare different amounts of initial antibody where partially bound virus is more infectious than free virus as shown in Fig [\[SI2\]](#SI2){reference-type="ref" reference="SI2"}. While low antibody levels did not result in as many infected cells after two days, intermediate levels of antibody can lead to about 10-fold enhancement of infection, while high levels limit the infection. It should be noted that this simulation is run over a shorter time frame, as the time course of relevant *in vitro* experiments is generally quite short-no more than a couple of days. In the circumstance where passive antibody is transferred to an individual, there is the possibility of enhancement of infection at intermediate levels of antibody transferred. However, as the individual also mounts a *de novo* response the dynamics of infection are quite different to the previous scenario, as seen in Fig [\[SI3\]](#SI3){reference-type="ref" reference="SI3"}. Here all infections peak at relatively similar timing though at different loads with intermediate antibody levels peaking at the highest amount. When cross-reactive antibodies are introduced, dynamics change yet again. If these cross-reactive antibodies are essentially functionally the same as *de novo* antibodies, then the dynamics follow those shown in Fig [\[SI4\]](#SI4){reference-type="ref" reference="SI4"} where increase in cross-reactive antibody shifts the peak timing earlier and the peak itself down. If this is the case then there is no enhancement of infection. However, if cross-reactive antibodies behave differently from the *de novo* response, there may be some limited enhancement of infection as seen in [\[SI6\]](#SI6){reference-type="ref" reference="SI6"}. When non-antibody immune responses are introduced into the model, the peak of infection can be lowered. However, it can cause EI-HM with viral dynamics peaking higher but clearing quicker at intermediate cross-reactive antibody levels, as seen in Fig [\[SI7\]](#SI7){reference-type="ref" reference="SI7"}. Including suppressive memory alters the results. Here if there is intermediate levels of cross-reactive antibodies the infection is cleared more slowly and with a much higher peak in the absence of a non-antibody immune response. Non-antibody immune growth rate impacts the amount of infection enhancement seen, from 117 fold without it to 1.2 fold with a moderate amount of growth to no enhancement with faster growth, as seen in Fig [\[SI8\]](#SI8){reference-type="ref" reference="SI8"}. # Other ADE Models {#other-ade-models .unnumbered} We consider three changes to the basic ADE model described in §ADE model. None of these changes qualitatively change our results (data not shown). In particular all models show ADE and EI-PA whereas EI-HM is not seen in general but is seen when 1) the cross reactive antibody grows more slowly than the de novo antibody response, 2) the cross reactive antibody is much less neutralizing (or has much higher dissociation rate) than the de novo antibody or 3) the non-antibody immune response is dominant in reducing peak viral load. ### Change 1: B cells and plasma cells are explicitly modeled {#change-1-b-cells-and-plasma-cells-are-explicitly-modeled .unnumbered} When B cells and plasma cells are explicitly modeled, the equations for \(A_c\) and \(A_n\) in the simple model are replaced by the following equations. \[\begin{aligned} \dot{B_c} &=s_c\chi(I>1)B_c\\ \dot{P_c} &= \dot{B_c}\\ \dot{A_c} &= 0.05 P_C-0.05A_c\\ \dot{B_n} &= s_n\chi(I>1) B_n\\ \dot{P_n} &= \dot{B_n}\\ \dot{A_n} &= 0.05P_n-0.05A_n \end{aligned}\] Here, \(B_c\), \(P_c\), and \(A_c\) are the B cells, plasma cells, and antibody level of the cross-reactive memory response, respectively. Analogously, \(B_n\), \(P_n\), and \(A_n\) represent the de novo immune response. Here, \(\chi\) is an indicator function. For the simulations, we used \(B_c(0)=P_c(0)=A_c(0)\), \(B_n(0)=0.036\), and \(P_n(0)=A_n(0)=0\). ### Change 2: Indicator function replaced {#change-2-indicator-function-replaced .unnumbered} Step functions in the differential equations are replaced with Hill functions. Explicitly, \(\chi(I>1)\) is replaced with \(I/(I+1)\). ### Change 3: Dissociation between Ab and virus is explicitly modeled {#change-3-dissociation-between-ab-and-virus-is-explicitly-modeled .unnumbered} For a given virion the proportion of surface proteins bound by antibody is given by the following equations. \[\begin{aligned} \dot{p_c} &= k_{\text{on,c}}A_c(1-p_c-p_n)-k_{\text{off,c}}p_c \\ \dot{p_n} &= k_{\text{on,n}}A_n(1-p_c-p_n)-k_{\text{off,n}}p_n \end{aligned}\] Here, \(p_c\) is the proportion of surface proteins bound by the cross reactive antibody, and \(p_n\) is the proportion bound by the de novo antibody. The association rates for the cross reactive and de novo antibody are \(k_{\text{on,c}}\) and \(k_{\text{on,n}}\) respectively. Similarly, \(k_{\text{off,c}}\) and \(k_{\text{off,n}}\) are the dissociation rates for the cross reactive and de novo antibody respectively. When solving, we discretize the above equations using N=4 binding sites. When the number of sites bound is less than or equal to Floor(\(0.15\cdot N\)), the virion is considered to be in the \(V_1\) category. If the number of sites bound is greater than Floor(\(0.5\cdot N\)), the virion is neutralized (\(V_3\) category). Otherwise the virion is in the highly infectious \(V_2\) category. Similar results were found at higher binding site numbers, though at the cost of computational speed. # Cross-reactive antibody has a higher dissociation rate {#cross-reactive-antibody-has-a-higher-dissociation-rate .unnumbered} For this section, we use change 3 to the basic ADE model where dissociation between antibody and virus is explicitly modeled. In these simulations, the cross reactive antibody has higher dissociation rate than the de novo antibody -- either \(k_{\text{off,c}}=50/\text{day},~ 100/\text{day},~ \text{or}~ 150/\text{day}\) versus \(k_{\text{off,n}}=10/\text{day}\). \(k_{on}\) was held at 0.5 to make antibody units comparable across models. See \[1\] for experimentally measured dissociation rates for dengue antibodies. These measurements approximately span the range of 3/day to 56/day with 13/day being median. Here, values were extracted from Figure 7B in \[1\] using WebPlotDigitizer and converted from per second to per day. Figure [\[SI-9\]](#SI-9){reference-type="ref" reference="SI-9"}, below, shows peak infected cells for different values of \(A_c(0)\). With \(k_{\text{off,c}}=50/\text{day}\), no EI-HM was observed. With \(k_{\text{off,c}}=100/\text{day}\text{ or }150/\text{day}\), there is some enhancement of infection at low levels of \(A_c(0)\). But the degree of enhancement is relatively small--a double or triple in difference in dissociation rate translates into maximum enhancement of only 1.1 fold compared to \(A_c(0)=0\). # Other adverse effects of humoral memory {#other-adverse-effects-of-humoral-memory-1 .unnumbered} We use Main Text Equations 1-8 with the default parameters except that there is no ADE (i.e. \(\beta_1=\beta_2=40\)) and \(s_R=0.64\). In Figure [\[SI9v2\]](#SI9v2){reference-type="ref" reference="SI9v2"}, we see a non-monotonic relationship between baseline antibody and final (post infection) antibody level. Hence for antibody values above 5.8 there is a negative effect of humoral memory at baseline on humoral memory postinfection. Notably, this model has neither ADE nor suppressive memory and hence no EI-HM. # Memory Dominance as a Proxy for Original Antigenic Sin {#memory-dominance-as-a-proxy-for-original-antigenic-sin-1 .unnumbered} Original antigenic sin (OAS) is the situation in which the immune response remains stronger towards strains encountered earlier in life compared to strains encountered later in life. This is often seen in the context of influenza where different birth cohorts experience different strains of influenza in their youth as different influenza variants arise, compete, and disappear. Because we wanted to disentangle the nuances of some of the many potential effects of prior immunity, we focus on memory dominance instead of the more complex phenomenon of OAS. However, memory dominance can be viewed as a proxy for OAS. We consider memory dominance to occur when \(A_c > A_n\) with Ac/An indicating the degree of memory dominance. Using our default parameters, if we consider the effective immune response from the antibodies to the *de novo* infection, then we can represent the change as \(\Delta A_c+\Delta A_n\) where \(\Delta A_c\) is the change in \(A_c\) and \(\Delta A_n\) is the change in \(A_n\) both relative to baseline.  We assume the effective immune response from the antibodies to the original infection is boosted by \(a\Delta A_c+b\Delta A_n\) where \(a\geq1\) and \(b\leq1\). If \(A_c >> A_n\) the boost to the old strain will be approximately as large or larger than the boost to the new strain leading to the OAS phenomenon. # Modeling Details {#modeling-details-1 .unnumbered} We simulate various scenarios with ADE that may cause enhancement of infection, including the effects of passive antibodies, cross-reactive antibodies, and non-antibody immune responses. We utilize a system of ordinary differential equations to compare how various initial antibody amounts affect the peak number of infected cells. A schematic of this system is shown below in Fig  [\[SI1\]](#SI1){reference-type="ref" reference="SI1"} and corresponds to Equations 1-8 in the main text. # Behavior of Main Text Models {#behavior-of-main-text-models-1 .unnumbered} For our models we considered relevant biological regimes and matched the basic dynamics of flavivirus infections. In reality the immune response is more complex and dependent on both individual infection history along with underlying frailties, and therefore exactly matching all possible experiments and outcomes is unfeasible. We begin by considering the simple *in vitro* case where we compare different amounts of initial antibody where partially bound virus is more infectious than free virus as shown in Fig [\[SI2\]](#SI2){reference-type="ref" reference="SI2"}. While low antibody levels did not result in as many infected cells after two days, intermediate levels of antibody can lead to about 10-fold enhancement of infection, while high levels limit the infection. It should be noted that this simulation is run over a shorter time frame, as the time course of relevant *in vitro* experiments is generally quite short-no more than a couple of days. In the circumstance where passive antibody is transferred to an individual, there is the possibility of enhancement of infection at intermediate levels of antibody transferred. However, as the individual also mounts a *de novo* response the dynamics of infection are quite different to the previous scenario, as seen in Fig [\[SI3\]](#SI3){reference-type="ref" reference="SI3"}. Here all infections peak at relatively similar timing though at different loads with intermediate antibody levels peaking at the highest amount. When cross-reactive antibodies are introduced, dynamics change yet again. If these cross-reactive antibodies are essentially functionally the same as *de novo* antibodies, then the dynamics follow those shown in Fig [\[SI4\]](#SI4){reference-type="ref" reference="SI4"} where increase in cross-reactive antibody shifts the peak timing earlier and the peak itself down. If this is the case then there is no enhancement of infection. However, if cross-reactive antibodies behave differently from the *de novo* response, there may be some limited enhancement of infection as seen in [\[SI6\]](#SI6){reference-type="ref" reference="SI6"}. When non-antibody immune responses are introduced into the model, the peak of infection can be lowered. However, it can cause EI-HM with viral dynamics peaking higher but clearing quicker at intermediate cross-reactive antibody levels, as seen in Fig [\[SI7\]](#SI7){reference-type="ref" reference="SI7"}. Including suppressive memory alters the results. Here if there is intermediate levels of cross-reactive antibodies the infection is cleared more slowly and with a much higher peak in the absence of a non-antibody immune response. Non-antibody immune growth rate impacts the amount of infection enhancement seen, from 117 fold without it to 1.2 fold with a moderate amount of growth to no enhancement with faster growth, as seen in Fig [\[SI8\]](#SI8){reference-type="ref" reference="SI8"}. # Other ADE Models {#other-ade-models-1 .unnumbered} We consider three changes to the basic ADE model described in §ADE model. None of these changes qualitatively change our results (data not shown). In particular all models show ADE and EI-PA whereas EI-HM is not seen in general but is seen when 1) the cross reactive antibody grows more slowly than the de novo antibody response, 2) the cross reactive antibody is much less neutralizing (or has much higher dissociation rate) than the de novo antibody or 3) the non-antibody immune response is dominant in reducing peak viral load. ### Change 1: B cells and plasma cells are explicitly modeled {#change-1-b-cells-and-plasma-cells-are-explicitly-modeled-1 .unnumbered} When B cells and plasma cells are explicitly modeled, the equations for \(A_c\) and \(A_n\) in the simple model are replaced by the following equations. \[\begin{aligned} \dot{B_c} &=s_c\chi(I>1)B_c\\ \dot{P_c} &= \dot{B_c}\\ \dot{A_c} &= 0.05 P_C-0.05A_c\\ \dot{B_n} &= s_n\chi(I>1) B_n\\ \dot{P_n} &= \dot{B_n}\\ \dot{A_n} &= 0.05P_n-0.05A_n \end{aligned}\] Here, \(B_c\), \(P_c\), and \(A_c\) are the B cells, plasma cells, and antibody level of the cross-reactive memory response, respectively. Analogously, \(B_n\), \(P_n\), and \(A_n\) represent the de novo immune response. Here, \(\chi\) is an indicator function. For the simulations, we used \(B_c(0)=P_c(0)=A_c(0)\), \(B_n(0)=0.036\), and \(P_n(0)=A_n(0)=0\). ### Change 2: Indicator function replaced {#change-2-indicator-function-replaced-1 .unnumbered} Step functions in the differential equations are replaced with Hill functions. Explicitly, \(\chi(I>1)\) is replaced with \(I/(I+1)\). ### Change 3: Dissociation between Ab and virus is explicitly modeled {#change-3-dissociation-between-ab-and-virus-is-explicitly-modeled-1 .unnumbered} For a given virion the proportion of surface proteins bound by antibody is given by the following equations. \[\begin{aligned} \dot{p_c} &= k_{\text{on,c}}A_c(1-p_c-p_n)-k_{\text{off,c}}p_c \\ \dot{p_n} &= k_{\text{on,n}}A_n(1-p_c-p_n)-k_{\text{off,n}}p_n \end{aligned}\] Here, \(p_c\) is the proportion of surface proteins bound by the cross reactive antibody, and \(p_n\) is the proportion bound by the de novo antibody. The association rates for the cross reactive and de novo antibody are \(k_{\text{on,c}}\) and \(k_{\text{on,n}}\) respectively. Similarly, \(k_{\text{off,c}}\) and \(k_{\text{off,n}}\) are the dissociation rates for the cross reactive and de novo antibody respectively. When solving, we discretize the above equations using N=4 binding sites. When the number of sites bound is less than or equal to Floor(\(0.15\cdot N\)), the virion is considered to be in the \(V_1\) category. If the number of sites bound is greater than Floor(\(0.5\cdot N\)), the virion is neutralized (\(V_3\) category). Otherwise the virion is in the highly infectious \(V_2\) category. Similar results were found at higher binding site numbers, though at the cost of computational speed. # Cross-reactive antibody has a higher dissociation rate {#cross-reactive-antibody-has-a-higher-dissociation-rate-1 .unnumbered} For this section, we use change 3 to the basic ADE model where dissociation between antibody and virus is explicitly modeled. In these simulations, the cross reactive antibody has higher dissociation rate than the de novo antibody -- either \(k_{\text{off,c}}=50/\text{day},~ 100/\text{day},~ \text{or}~ 150/\text{day}\) versus \(k_{\text{off,n}}=10/\text{day}\). \(k_{on}\) was held at 0.5 to make antibody units comparable across models. See \[1\] for experimentally measured dissociation rates for dengue antibodies. These measurements approximately span the range of 3/day to 56/day with 13/day being median. Here, values were extracted from Figure 7B in \[1\] using WebPlotDigitizer and converted from per second to per day. Figure [\[SI-9\]](#SI-9){reference-type="ref" reference="SI-9"}, below, shows peak infected cells for different values of \(A_c(0)\). With \(k_{\text{off,c}}=50/\text{day}\), no EI-HM was observed. With \(k_{\text{off,c}}=100/\text{day}\text{ or }150/\text{day}\), there is some enhancement of infection at low levels of \(A_c(0)\). But the degree of enhancement is relatively small--a double or triple in difference in dissociation rate translates into maximum enhancement of only 1.1 fold compared to \(A_c(0)=0\). # Other adverse effects of humoral memory {#other-adverse-effects-of-humoral-memory-2 .unnumbered} We use Main Text Equations 1-8 with the default parameters except that there is no ADE (i.e. \(\beta_1=\beta_2=40\)) and \(s_R=0.64\). In Figure [\[SI9v2\]](#SI9v2){reference-type="ref" reference="SI9v2"}, we see a non-monotonic relationship between baseline antibody and final (post infection) antibody level. Hence for antibody values above 5.8 there is a negative effect of humoral memory at baseline on humoral memory postinfection. Notably, this model has neither ADE nor suppressive memory and hence no EI-HM. # Abstract {#abstract-1 .unnumbered} Antibodies and humoral memory are key components of the adaptive immune system. We consider and computationally model mechanisms by which humoral memory present at baseline might instead increase infection load; we refer to this effect as EI-HM (enhancement of infection by humoral memory). We first consider antibody dependent enhancement (ADE) in which antibody enhances the growth of the pathogen, typically a virus, and typically at intermediate 'Goldilocks' levels of antibody. Our ADE model reproduces ADE *in vitro* and enhancement of infection *in vivo* from passive antibody transfer. But notably the simplest implementation of our ADE model never results in EI-HM. Adding complexity, by making the cross-reactive antibody much less neutralizing than the *de novo* generated antibody or by including a sufficiently strong non-antibody immune response, allows for ADE-mediated EI-HM. We next consider the possibility that cross-reactive memory causes EI-HM by crowding out a possibly superior *de novo* immune response. We show that, even without ADE, EI-HM can occur when the cross-reactive response is both less potent and 'directly' (i.e. independently of infection load) suppressive with regard to the *de novo* response. In this case adding a non-antibody immune response to our computational model greatly reduces or completely eliminates EI-HM, which suggests that 'crowding out' is unlikely to cause substantial EI-HM. Hence, our results provide examples in which simple models give qualitatively opposite results compared to models with plausible complexity. Our results may be helpful in interpreting and reconciling disparate experimental findings, especially from dengue, and for vaccination. # Author summary {#author-summary-1 .unnumbered} Humoral memory, generated by infection with a pathogen, can protect against subsequent infection by the same and related pathogens. We consider situations in which humoral memory is counterproductive and instead enhances infection by related pathogens. Numerous experiments have shown that the addition of binding antibody makes certain viruses more, rather than less, infectious; typically high levels of antibody are still protective, meaning that infectivity is maximized at intermediate 'Goldilocks' levels of antibody. Additionally, we consider the situation in which cross-reactive humoral memory dominates relative to the *de novo* response. Memory dominance has been documented for influenza infections, but whether it is harmful or not is unclear. We show computationally that both ADE and a certain type of competition between the cross-reactive and *de novo* responses translate into enhancement of infection in certain circumstances but not in others. We discuss the implications of our research for dengue infection, a common mosquito-transmitted viral infection, and vaccination. # Introduction {#introduction-1 .unnumbered} Antibody is a key component of the adaptive immune system. Antibodies bind to viruses, bacteria, and other microbes and thereby inhibit microbe function and aid in microbe clearance. For certain infections, antibody levels are considered the key correlate of protection  , meaning that increasing antibody levels are associated with protection from infection or disease. Paradoxically for certain viral infections, antibody can instead enhance infection-an effect known as antibody dependent enhancement (ADE). ADE has now been reported for many viruses including dengue, West Nile, HIV, influenza A, Ebola, rabies, polio, and HSV1  . It is generally believed that ADE results when antibody bound to virus increases the ability of that virus to infect certain cells, like macrophages, that possess Fc receptors  . In the case of dengue, it has also been reported that maternal antibodies can enhance disease severity in infants    . Consistent with this observation, transfer of dengue antibodies to dengue-infected monkeys can greatly increase viral load    . A related concern to ADE is that cross-reactive humoral memory from previous infections or vaccinations may enhance the subsequent infection    . This is particularly a concern for dengue infection which is caused by four serotypes of virus -- DENV1, DENV2, DENV3, and DENV4 -- that are only  65% similar at the amino acid level  . Certain studies have reported that infection with one serotype of dengue, for example DENV1, may make infection with a second serotype, for example DENV2, more severe probably as a consequence of ADE      . In contrast, the discussion surrounding influenza has focused more on the issue of original antigenic sin  . According to the original antigenic sin model, the immune response remains focused on the first influenza strain encountered even after multiple subsequent influenza infections. Original antigenic sin is not necessarily bad, and certain authors have argued that antigenic seniority is a more apt term  . ## Definitions {#definitions-1 .unnumbered} In this paper we use the following definitions: *Antibody dependent enhancement* (ADE) means that viral growth is enhanced at some level of antibody compared to no antibody. Mechanistically, we model this via higher infectivity of virions with some bound antibody *Enhancement of infection from humoral memory* (EI-HM) means that infection load is greater in the presence of cross-reactive humoral memory than in its absence. While there are many ways to measure infection load, in this paper we consistently use the peak number of infected cells. *Enhancement of infection from passive antibody* (EI-PA) means that infection load is greater when passive antibody -- for example maternal antibody -- is supplied to a host than in its absence. (We assume that this host has an adaptive immune system and will generate a humoral response of its own.) *Memory dominance* means that the cross-reactive memory response dominates the *de novo* immune response. Memory dominance may be viewed as a proxy for original antigenic sin (see Supplement for further explanation). *Suppressive memory* means that cross-reactive memory "directly" suppresses the *de novo* immune response ignoring any "indirect" effects mediated via infection load. See Fig [\[fig1A\]](#fig1A){reference-type="ref" reference="fig1A"} for a conceptual diagram of these differences. # Model and Results {#model-and-results-1 .unnumbered} ## ADE Model {#ade-model-1 .unnumbered} To investigate the mechanisms depicted in Fig [\[fig1A\]](#fig1A){reference-type="ref" reference="fig1A"}, we create a mathematical model of the essential infection components. In this model, infected cells (\(I\)) produce virus. \(V_1\) is virus with little or no bound antibody. \(V_2\) is virus with intermediate amounts of bound antibody. Virus with the most antibody bound, \(V_3\), is neutralized. We assume that \(V_2\) virus is more infectious than \(V_1\) virus (i.e. \(\beta_2 > \beta_1\)). \(A_c\) is cross-reactive memory antibody. \(A_n\) is the *de novo* antibody response. \(A_p\) is passive antibody (e.g. maternal antibody). \(R\) is the non-antibody (e.g. innate or CD8 T cell) immune response. We assume that only a small fraction of target cells are depleted during the course of infection; hence, we do not explicitly model loss of susceptible cells. We show this model set up via a schematic in  . The mechanisms we consider are described by the following set of ordinary differential equations. \[\begin{aligned} \dot{I} &= \underbrace{-bI}_{\text{cell death}}+\underbrace{\beta_1V_1+\beta_2V_2}_{\text{infection}}-\underbrace{IR}_{\text{non-Ab clearance}}\\ \dot{V_1} &= \underbrace{I}_{\text{production}}-V_1(\underbrace{a}_{\text{clearance}}+\underbrace{k_{c1}A_c+k_{n1}A_n+k_{p1}A_p}_{\text{binding by antibodies}})\\ \dot{V_2} &= \underbrace{V_1(k_{c1}A_c+k_{n1}A_n+k_{p1}A_p)}_{\text{newly partially bound}}-V_2(\underbrace{a}_{\text{clearance}}+\underbrace{k_{c2}A_c+k_{n2}A_n+k_{p2}A_p)}_{\text{further binding}}\\ \dot{V_3} &= \underbrace{V_2(k_{c2}A_c+k_{n2}A_n+k_{p2}A_p)}_{\text{newly completely bound}}-\underbrace{aV_3}_{\text{clearance}}\\ \dot{A_c} &= \underbrace{s_c}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}}A_c\\ \dot{A_n} &= \underbrace{s_n}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}}A_n\\ \dot{A_p} &= \underbrace{0}_{\text{no growth}} \\ \dot{R} &= \underbrace{s_R}_{\text{growth}}\underbrace{\chi(I>1)}_{\text{indicator}} \end{aligned}\] Here, \(\chi\) is the indicator function. Table [2](#table1){reference-type="ref" reference="table1"} describes the model parameters. Unless otherwise stated we use the default parameter values shown in the table. Model parameters were chosen to give an asymptotic growth of 1.5 natural logs (ln) per day in the absence of antibody and 2.5 ln per day at maximum ADE and peak infection load at 10 days. For comparison the growth of the yellow fever YFV-17D virus was estimated at 1.6 ln per day. These dynamics are shown in Fig S2-7 for each of the models. [\[table1\]]{#table1 label="table1"} ## The model recapitulates *in vitro* ADE {#the-model-recapitulates-in-vitro-ade-1 .unnumbered} In this section, we simulate a 48 hour *in vitro* cell culture experiment. As there is no active immune response, \(A_c(0)=A_n(0)=0\). Fig [\[fig2\]](#fig2){reference-type="ref" reference="fig2"} shows the quantity of infected cells at the end of the simulation for different values of \(A_p\) and compares them to experimental values. An antibody level of 15 produces maximum ADE in the model; this level of antibody produces 10 fold greater infection load after 48 hours as compared to no antibody. Hence the model described above does produce ADE. Higher levels of antibody (\(>74\)) reduce infection growth as compared to no antibody. This pattern of enhanced growth at intermediate levels of antibody and reduced or no growth at higher levels of antibody matches the pattern seen from *in vitro* experiments with dengue   and West Nile   viruses. ## The model produces EI-PA {#the-model-produces-ei-pa-1 .unnumbered} Building upon our previous *in vitro* model to replicate an *in vivo* experiment, we now include additional factors. First, we simulate supply of passive (e.g. maternal) antibody (\(A_p\)) to an infected host. Second, in addition to these passive antibodies, the host's immune system will generate antibodies (\(A_n\)) in response to the infection. Fig [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} shows peak infected cells for different values of \(A_p\). In this case, a passive antibody level of 11 produces maximum enhancement; this level of passive antibody increases infection load 35 fold compared to no passive antibody. Thus, in this model, adding passive antibody can dramatically enhance infection, but higher levels of passive antibody are protective. Hence, the model qualitatively reproduces dengue data from infants     which show that high levels of maternal antibody are protective but intermediate levels are associated with enhanced disease. The model is also consistent with dengue animal studies showing enhanced viral load when passive antibody is supplied    . ## The model does not (necessarily) produce EI-HM {#the-model-does-not-necessarily-produce-ei-hm-1 .unnumbered} In this section, we simulate infection for varying baseline levels of cross-reactive memory antibody. We build upon our original *in vitro* model and now include both *de novo* and cross-reactive antibody response to create an *in vivo* model. Because cross-reactive antibodies may replicate some but not necessarily all the features of *de novo* antibodies such as growth rate or neutralizing activities, we first investigate the simplest scenario where cross-reactive and *de novo* antibodies have all of the same properties. Fig [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} shows peak infected cells for different values of \(A_c(0)\). In this case, we see only protection as \(A_c(0)\) increases. Hence ADE, where partially bound virus is more infectious than free virus, does not necessarily produce EI-HM. ## The model can produce EI-HM in certain circumstances {#the-model-can-produce-ei-hm-in-certain-circumstances-1 .unnumbered} Although the simplest implementation of our *in vivo* model (shown above) does not show enhancement of infection from humoral memory, our model produces EI-HM in certain situations as shown in the following sections. ### Cross-reactive antibody is less neutralizing {#cross-reactive-antibody-is-less-neutralizing-1 .unnumbered} In these simulations the cross-reactive antibody grows at the same rate but is less neutralizing than the *de novo* antibody -- either \(k_{c2}=0.125\) or \(k_{c2}=0.05\). \(k_{c2}=0.125\) can be loosely interpreted as meaning that the cross-reactive antibody is half as neutralizing as the *de novo* antibody, whereas \(k_{c2}=0.05\) can be loosely interpreted as meaning that the cross-reactive antibody is 5 times less neutralizing than the *de novo* antibody. Fig [\[fig6\]](#fig6){reference-type="ref" reference="fig6"} shows peak infected cells for different values of \(A_c(0)\). With \(k_{c2}=0.125\), no EI-HM was observed, but with \(k_{c2}=0.05\), there is some enhancement of infection at low levels of \(A_c(0)\). However, the degree of enhancement is again relatively small -- a 5 fold difference in neutralization ability translates into a maximum enhancement of only 2.7 fold compared to \(A_c(0)=0\). Increasing \(k_{c1}\), the rate at which the cross-reactive antibody enhances virus, has an analogous effect to decreasing \(k_{c2}\). \(k_{c1}=2k_{n1}=2\) does not produce EI-HM whereas \(k_{c1}=5k_{n1}=5\) produces maximum relative enhancement of 2.7 fold. ### Non-antibody immune response is dominant {#non-antibody-immune-response-is-dominant-1 .unnumbered} In these simulations we vary the growth rate, \(s_R\), of the non-antibody immune response. To compare, \(s_R=0\) shows the peak of infected cells when there is no non-antibody immune response and is therefore equivalent to the results in Fig [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}. Fig [\[fig7\]](#fig7){reference-type="ref" reference="fig7"} shows peak infected cells for different values of \(A_c(0)\). A growth rate of \(s_R=0.26\) is the highest value of \(s_R\) that shows no EI-HM whereas \(s_R=0.64\) corresponds to primary infection peaking at day 5 (see Fig S7), which is many days before antibody reaches sterilizing levels. With \(s_R=0.64\), enhancement of up to 11 fold is possible. However, the range of \(A_c(0)\) values that lead to enhancement is quite a bit smaller than observed *in vitro*, \(>0\) to 74 compared to \(>0\) to 24. ## Other models {#other-models-1 .unnumbered} We consider three changes to our overarching ADE model: 1) B cells and plasma cells are explicitly included in the model, 2) step functions in the differential equations are replaced with Hill functions, and 3) dissociation between antibody and virus is explicitly modeled. See Supplement for a more detailed description of these changes. None of these changes qualitatively change our results (figures not shown). Explicitly modeling the dissociation between antibody and virus also allows us to consider the situation where the cross-reactive antibody has a higher dissociation rate than the *de novo* antibody. This situation is similar to that described in § Cross-reactive antibody is less neutralizing. When the dissociation rate of cross-reactive antibody is 50/day versus 10/day for the *de novo* antibody, no EI-HM is observed. But when the dissociation rate of cross-reactive antibody is 150/day versus 10/day for the *de novo* antibody, there is some EI-HM at low levels of \(A_c(0)\) but with maximum enhancement of 1.1 fold. See for more details. ## Suppressive memory can produce EI-HM without ADE {#suppressive-memory-can-produce-ei-hm-without-ade-1 .unnumbered} All of our above simulations show memory dominance when \(A_c(0)\) is sufficiently large -- \(A_c(0)>0.036\). This shows that memory dominance does not necessarily mean EI-HM. In fact, with our model ADE (\(\beta_2 >\beta_1\)) is necessary -- but not sufficient -- to produce EI-HM. Our ADE model does not incorporate suppressive memory. In this section we modify the equations for \(A_c\) and \(A_n\) to incorporate this effect. \[\dot{A_c}= s_c\chi(I>1)\frac{\phi A_c}{\phi+A_c+A_n}\] \[\dot{A_n}= s_n\chi(I>1)\frac{\phi A_n}{\phi+A_c+A_n}\] Here \(\chi\) is the indicator function, \(\phi=28\), and \(s_c=s_n=1.5\). In this system increasing \(A_c\) suppresses the growth rate of \(A_n\) and vice versa. We consider the situation where \(A_c\) is a quarter as potent as \(A_n\) (\(k_{c1} =0.25, k_{c2}=0.0625\)) and there is no ADE (\(\beta_1=\beta_2=40\)). In this case there is dramatic EI-HM with fold enhancement of up to 117 fold (Fig [\[fig8\]](#fig8){reference-type="ref" reference="fig8"}). However, adding a non-antibody immune response with \(s_R=0.27\) or \(s_R=0.54\) mostly or completely abolishes EI-HM. Here \(s_R=0.54\) corresponds to primary infection peaking at day 5 (Fig S8). ## Other adverse effects of humoral memory {#other-adverse-effects-of-humoral-memory-3 .unnumbered} EI-HM is one possible adverse effect of humoral memory. Another possible adverse effect is a negative effect of humoral memory at baseline on humoral memory post infection. Such an effect is possible because of the negative feedback between antibody and pathogen and does not require ADE or suppressive memory (see  ). Hence, if a class of antigenically similar pathogens can infect more than once, humoral memory may protect against the immediate infection but enhance later infection. # Discussion and Conclusion {#discussion-and-conclusion-1 .unnumbered} ## Intuitive Explanation {#intuitive-explanation-1 .unnumbered} While our results may seem disparate, they can be explained by two factors: 1) the effect of antibody early in infection and 2) the effect of cross-reactive humoral memory on the rapidity of humoral response. In the case of ADE, antibody may enhance viral growth early in infection, but humoral memory may also accelerate the development of protective levels of antibody. In the case of suppressive memory, viral growth is reduced early in infection, but the development of a potent antibody response may also be delayed. Depending on which of these factors dominates, cross-reactive humoral memory may either enhance or decrease infection. ## Concluding Remarks {#concluding-remarks-1 .unnumbered} Our results are consistent with the hypothesis that ADE and enhancement of infection from passive antibody (EI-PA) can result when virus with intermediate levels of antibody is more infectious than virus with lower or higher levels of antibody  . Hence additional mechanisms, such as suppression of the immune response  , are possible but not necessary to produce the phenomena of ADE and EI-PA. More importantly we show that ADE, EI-PA, and enhancement of infection from cross-reactive humoral memory (EI-HM) are distinct. ADE in our models always implies the possibility of EI-PA. Furthermore, the range of antibody levels that produces ADE is similar to the range of passive antibody levels that produces EI-PA. This is because passive antibody can enhance viral growth early in infection but does very little to accelerate the development of protective levels of antibody. In contrast EI-HM is a double-edged sword, and ADE and EI-PA may not imply EI-HM. In our models, when ADE does produce EI-HM the range of baseline antibody levels that produce EI-HM is very different from the range that produces ADE. This difference in ranges may explain the observation in dengue patients that ADE measurements of pre-infection plasma did not positively correlate with peak viremia  . ADE, EI-PA, and EI-HM are of particular concern for dengue infection. In the case of dengue, there is extensive evidence for ADE *in vitro*          , and EI-PA of up to 100 fold has been demonstrated in monkey experiments    . In contrast, the evidence for greater infection load in secondary infection is more limited. A large monkey experiment involving 118 rhesus macaques showed higher viral load in secondary DENV2 infection as compared to primary DENV2 infection. However, in the same study secondary infections with DENV1, DENV3 and DENV4 showed reduced viremia when compared to the respective primary infections  . A recent epidemiological study showed increased risk of severe dengue in children with low or intermediate levels of dengue antibody at baseline as compared to dengue seronegative children; higher levels of baseline antibody were associated with reduced risk. In contrast, risk of symptomatic dengue decreased essentially monotonically with antibody levels  . Likewise analysis of clinical trials data suggested that vaccination of dengue naive children with the CYD-TDV dengue vaccine increased the risk of severe dengue but reduced the risk of symptomatic dengue. Our results suggest that this pattern could be explained by a scenario in which the majority do not experience EI-HM but a minority -- for example those with very poor quality cross-reactive antibody -- do experience EI-HM and this minority also tends to have more severe disease. However,   and examined disease severity rather than infection load, and in   the degree of enhancement was highly sensitive to the method of classifying disease severity. (Using the 1997 World Health Organization (WHO) dengue severity classification system, the risk of severe dengue was up to 7.6 fold elevated. Using the newer 2009 WHO severity classification, generally regarded as more accurate    , this was reduced to only 1.75 fold.) Despite this limitation, in the case of human dengue infections, our results support the hypothesis that ADE sometimes produces EI-HM but often does not. We show that, in principle, memory dominance plus suppressive memory can also produce EI-HM. In contrast memory dominance alone does not produce EI-HM in any of our models. However, inclusion of a non-antibody immune response in our simulations dramatically reduces EI-HM from suppressive memory, which casts doubt on its real world relevance. It should be emphasized that our simulations only consider the effect of humoral memory present at baseline on that particular infection. A negative effect of humoral memory at baseline on humoral memory post infection can occur as a consequence of the negative feedback between antibody and pathogen and does not require ADE or suppressive memory.  For infections, like influenza, where reinfections are common, some of the apparent adverse effects of humoral memory likely involve not EI-HM but the negative effect of humoral memory at an earlier time point on humoral memory at a latter time point with potentially many infections in between  . Although CD4 T cells can increase pathology  , a mechanism such as ADE is not established for T cells. This implies that activating non-antibody immune responses, especially CD8 T cells, could be an important overlooked component in vaccine development. Although memory dominance and suppressive memory are concerns for antibody as well as T cells, our results suggest that suppressive memory is unlikely to cause EI-HM. Moreover, the ability to strategically direct T cell responses has been demonstrated   , potentially turning memory dominance into a long term advantage. Finally, our results also illustrate potential pitfalls from mathematical modeling, which may be relevant not only for biology but also for other disciplines, such as economics and epidemiology, that use abstract mathematical models. In the simplest version of our models, there is EI-HM from suppressive memory but none from ADE, whereas adding plausible complexity reverses this situation. # Supplement {#supplement-1 .unnumbered} #### S1 Text. {#S1 Text .unnumbered} **Model Dynamics and Alternative Model Results.** #### Fig S1. {#S1_Fig .unnumbered} **Model Schematic.** In our model, infected cells (\(I\)) produce free virus(\(V_1\)) and can be removed via cell death or the non-antibody immune response (\(R\)). Virus can be successively bound with only free (\(V_1\)) and partially bound virus (\(V_2\)) able to infect cells. Binding can come from the *de novo* antibody response (\(A_n\)), the cross-reactive antibody response (\(A_c\)), or the passive antibody response (\(A_p\)), with \(A_n\) and \(A_c\) growing in response to the infected cells. #### S2 Fig. {#S2_Fig .unnumbered} **Dynamics of infection under ADE ***in vitro***.** In this *in vitro* situation, where there is no growth in immune response, the phenomena of ADE is evident as the number of infected cells after 48 hours is greater at intermediate ranges of antibody. We show the differences in dynamics for three different initial antibody concentrations in Panels B and C: low but non-zero (black), intermediate (red), and high (blue). #### S3 Fig. {#S3_Fig .unnumbered} **Dynamics of infection given passive antibodies.** When passive antibodies are given and a *de novo* antibody response is mounted, again ADE appears at intermediate levels (red). #### S4 Fig. {#S4_Fig .unnumbered} **Dynamics of infection given cross-reactive antibodies does not necessarily produce enhancement of infection in these simulations where ***de novo*** and cross-reactive antibodies are identical.** When there are both *de novo* and cross-reactive antibodies present and growing and binding at the same rate, enhancement of infection is not seen, as shown in Panel A; rather, the greater initial cross-reactive antibody load the lower the viral load and number of infected cells as seen in Panels B-C. Here, black represents low initial cross-reactive antibody concentration, red intermediate, and blue high. #### S5 Fig. {#S6_Fig .unnumbered} **Less neutralizing cross-reactive antibodies can sometimes cause EI-HM.** If cross-reactive antibodies are approximately a fifth as good as *de novo* antibodies, some EI-HM may occur as shown in Panel A with slightly higher peak viremia and infected cell counts. While the the line for the low cross-reactive antibody concentration is the same for both \(k_{c2}\) values, as seen in black in Panels B and C, they differ for the intermediate (red) and high (blue) concentrations where \(k_{c2} = 0.125\) is given by the solid lines and \(k_{c2} = 0.05\) by the dashed lines. #### S6 Fig. {#S6_Fig .unnumbered} **Dynamics of infection given cross-reactive antibodies and non-antibody immune response.** Depending on the rate of growth of non-antibody immune response and initial cross-reactive antibody amount, EI-HM can sometimes occur and give a 10-fold increase in peak infected cells. Colors denote level of initial cross-reactive antibody with black, red, and blue representing low, intermediate, and high values. Line types denote non-Ab immune response growth rates with solid, dashed, and x'd lines representing \(s_R=0\), \(s_R= 0.26\), and \(s_R=0.64\) respectively. #### S7 Fig. {#S7_Fig .unnumbered} **Dynamics of infection given suppressive memory.** As shown in Panel A given the scenario where there is no non-antibody immune response, EI-HM can occur with a fold change of \(117\) for \(s_R=0\). With non-antibody immune responses, however, enhancement of infection is greatly reduced with a maximum fold change of \(1.2\) for \(s_R=0.27\) or nonexistent for \(s_R = 0.54\). In Panel B, we consider the viral dynamics for low (black), intermediate (red), and high (blue) points for each of the different growth rates where \(s_R=0\) is a solid line, \(s_R=0.27\) is dashed, and \(s_R=0.54\) is x'd. Note that for \(s_R=0\) in particular if the high point was taken at an even higher antibody level (\(>25\)), it would in fact peak below the low point. #### S8 Fig. {#S9_Fig .unnumbered} **\(\mathbf{k_{\text{off,c}}=100/}\)day produces some EI-HM.** The figure shows simulation results from a modification to the simple ADE model such that dissociation between antibody and virus is explicitly modeled. When the cross reactive antibody has 5 times the dissociation rate of the de novo antibody (\(k_{\text{off,c}}=50/\text{day}\) versus \(k_{\text{off,n}}=10/\text{day}\)), there is no EI-HM. But, when the cross reactive antibody has 10 times or more the dissociation rate of the de novo antibody (\(k_{\text{off,c}}=100/\text{day}\) or \(k_{\text{off,c}}=150/\text{day}\) and \(k_{\text{off,n}}=10/\text{day}\)), there is some enhancement of infection at low levels of baseline cross reactive antibody. #### S9 Fig. {#S10_Fig .unnumbered} **Effect of baseline humoral memory on postinfection antibody levels..** In this model, without ADE or suppressive memory, and with non-antibody immune response we see that higher levels of baseline antibody (\(>\)`<!-- -->`{=html}5.8) negatively affect final (post infection) antibody levels.

{'timestamp': '2023-02-10T02:02:15', 'yymm': '2302', 'arxiv_id': '2302.04340', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04340'}

# Introduction We consider, for every \(\alpha>-\frac12\), the probability measure defined on \((0,\infty)\) by \(d\gamma_\alpha(x)=\frac{2}{\Gamma(\alpha+1)}e^{-x^2}x^{2\alpha+1}dx\). This measure has not the doubling property with respect to the usual metric defined by the absolute value \(|\cdot|\) on \((0,\infty)\). Then, the triple \(((0,\infty),|\cdot|,\gamma_\alpha)\) is not homogeneous in the sense of Coifman and Weiss (). Harmonic analysis in the spaces of homogeneous type can be developed following the model of Euclidean spaces \((\mathbb{R}^n,\|\cdot\|,\lambda)\) where \(\|\cdot\|\) denotes a norm and \(\lambda\) is the Lebesgue measure on \(\mathbb{R}^n\). When the measure is not doubling the situation is very different and it is necessary to introduce new ideas (see, for instance,, ,, ,, ,, and ). Tolsa () defined \(\textup{BMO}\)-type spaces, that he named \(\textup{RBMO}\)-spaces, on \((\mathbb{R}^n,\mu)\) when \(\mu\) is a Radon measure on \(\mathbb{R}^n\), which is not necessarily doubling, satisfying that \(\mu(B(x,r))\le Cr^k\), \(x\in \mathbb{R}^k\) and \(r>0\), for some \(k\in \{1,\dots,n\}\) and \(C>0\). He also proved that \(\textup{RBMO}(\mathbb{R}^n,\mu)\) has many of the properties of the classical space \(\textup{BMO}(\mathbb{R}^n)\) of John and Nirenberg. In particular, the integral operators defined by standard Calderón-Zygmund kernels are bounded from \(L^\infty(\mathbb{R}^n,\mu)\) into \(\textup{RBMO}(\mathbb{R}^n,\mu)\). It is clear that, for every \(0<r\le x\), \(\gamma_\alpha((x-r,x+r))\le Cr\). Then, following Tolsa's ideas we can define the space \(\textup{RBMO}((0,\infty),\gamma_\alpha)\) by replacing \(\mathbb{R}^n\) by \((0,\infty)\). However, \(\textup{RBMO}((0,\infty),\gamma_\alpha)\) is not suitable to study harmonic analysis operators associated with Laguerre polynomial expansions because these operators are not defined by standard Calderón-Zygmund kernels (, and ). Motivated by the results in in the Gaussian setting, the authors and R. Scotto () defined a local \(\textup{BMO}\)-type space related to the measure \(\gamma_\alpha\) as follows. We consider the function \(m(x)=\min\{1,1/x\}\), \(x\in (0,\infty)\). Given \(a>0\), we say that an interval \((x-r,x+r)\), with \(0<r\le x\), is *\(a\)-admissible*, or is in the class \(\mathcal{B}_a\), when \(r\le am(x)\). The measure \(\gamma_\alpha\) has the doubling property on \(\mathcal{B}_a\), that is, there exists \(C>0\) such that, for every \(0<r\le x\) being \(r\le am(x)\), we have that \[\gamma_\alpha(I(x,2r))\le C\gamma_\alpha((x-r,x+r)),\] where \(I(x,r):=(x-r,x+r)\cap (0,\infty)\) for \(x,r>0\). A function \(f\in L^1((0,\infty),\gamma_\alpha)\) is said to be in \(\textup{BMO}_a((0,\infty),\gamma_\alpha)\) when \[\|f\|_{*,\alpha,a}:=\sup_{I\in \mathcal{B}_a}\frac{1} {\gamma_\alpha(I)}\int_I|f(y)-f_I|d\gamma_\alpha(y)<\infty,\] where \(f_I=\fint_If(y)d\gamma_\alpha(y)\), for every \(I\in \mathcal{B}_a\). For every \(f\in \textup{BMO}_a((0,\infty),\gamma_\alpha)\), we define \[\|f\|_{\textup{BMO}_a((0,\infty),\gamma_\alpha)}:=\|f\|_{L^1((0,\infty),\gamma_\alpha)}+\|f\|_{*,\alpha,a}.\] The space \(\textup{BMO}_a((0,\infty),\gamma_\alpha)\) actually does not depend on \(a>0\). Then, in the sequel we will write \(\textup{BMO}((0,\infty),\gamma_\alpha)\) and \(\|\cdot\|_{*,\alpha}\) instead of \(\textup{BMO}_a((0,\infty),\gamma_\alpha)\) and \(\|\cdot\|_{*,\alpha,a}\), respectively. This space can be identified with the dual space of the Hardy space \(H^1((0,\infty),\gamma_\alpha)\) studied in (see ). The space \(\textup{BLO}(\mathbb{R}^n)\) of functions of bounded lower oscillation on \(\mathbb{R}^n\) was introduced by Coifman and Rochberg (). Later, Bennett () obtained a characterization of the functions in \(\textup{BLO}(\mathbb{R}^n)\) by using the natural Hardy-Littlewood maximal operators, and Leckband () proved that certain maximal operators associated with singular integrals are bounded from \(L^{p}(\mathbb{R}^n)\cap L^\infty(\mathbb{R}^n)\) into \(\textup{BLO}(\mathbb{R}^n)\) for certain \(1\le p<\infty\). Based on Tolsa's ideas, Jiang () introduced \(\textup{BLO}\)-type spaces in \((\mathbb{R}^n,\mu)\) where \(\mu\) is a positive non-doubling Radon measure with polynomial growth. \(\textup{BLO}\)-spaces in the Gaussian setting were defined by Liu and Yang (). In, Littlewood-Paley functions in non-doubling settings on \(\textup{RBLO}\) spaces were studied. Other results concerning \(\textup{RBLO}\) spaces can be encountered in and. As in happens with \(\textup{RBMO}\)-spaces, \(\textup{RBLO}\)-spaces for \(\gamma_\alpha\) do not work in a correct way in connection with harmonic analysis operators associated to Laguerre polynomial expansions. In this paper we introduce \(\textup{BLO}\)-spaces associated with the measure \(\gamma_\alpha\) on \((0,\infty)\) by using admissible intervals. Let \(a>0\). We say that a function \(f\in L^1((0,\infty),\gamma_\alpha)\) is in \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\) when \[\sup_{I\in \mathcal{B}_a}\frac{1}{\gamma_\alpha(I)}\int_I \left(f(y)-\essinf_{z\in I}f(z)\right)d\gamma_\alpha(y)<\infty.\] For every \(f\in \textup{BLO}_a((0,\infty),\gamma_\alpha)\) we define \[\|f\|_{\textup{BLO}_a((0,\infty),\gamma_\alpha)}:=\|f\|_{L^1((0,\infty),\gamma_\alpha)}+\sup_{I\in \mathcal{B}_a}\frac{1}{\gamma_\alpha(I)}\int_I \left(f(y)-\essinf_{z\in I}f(z)\right)d\gamma_\alpha(y).\] It is not hard to see that \[L^\infty((0,\infty),\gamma_\alpha)\subset \textup{BLO}_a((0,\infty),\gamma_\alpha)\subset \textup{BMO}_a((0,\infty),\gamma_\alpha).\] The main properties of the space \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\) will be established in Section [2](#sec: BLO){reference-type="ref" reference="sec: BLO"}. Our objective is to prove that maximal, variation and oscillation operators defined by singular integrals in the Laguerre settings are bounded from \(L^\infty((0,\infty),\gamma_\alpha)\) to \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\). We now define the operators we are going to consider. Let \(\alpha>-\frac12\). The Laguerre polynomial \(L_k^\alpha\) of order \(\alpha\) and degree \(k\in \mathbb{N}\) (see ) is \[L_k^\alpha(x)=\sqrt{\frac{\Gamma(\alpha+1)}{\Gamma(\alpha+k+1)k!}}e^x x^{-\alpha}\frac{d^k}{dx^k}(e^{-x}x^{\alpha+k}),\quad x\in (0,\infty).\] The Laguerre differential operator \(\widetilde{\Delta_\alpha}\) is given by \[\widetilde{\Delta_\alpha}:=\frac{1}{2}\frac{d^2}{dx^2}+\left(\frac{2\alpha+1}{2x}-x\right)\frac{d}{dx}+\alpha+1, \quad f\in C^2(0,\infty).\] We define, for every \(k\in \mathbb{N}\), \(\mathcal{L}_k^\alpha(x):=L_k^\alpha(x^2)\), \(x\in (0,\infty)\). Then, the sequence \(\{\mathcal{L}_k^\alpha\}_{k\in \mathbb{N}}\) is an orthonormal basis on \(L^2((0,\infty),\gamma_\alpha)\). For every \(k\in \mathbb{N}\), \(\mathcal{L}_k^\alpha\) is an eigenfunction for \(\widetilde{\Delta_\alpha}\) associated with the eigenvalue \(\lambda_k^\alpha=2k+\alpha+1\). For every \(f\in L^1((0,\infty),\gamma_\alpha)\), we define \[c_k^\alpha(f):=\int_0^\infty f(y)\mathcal{L}_k^\alpha(x)d\gamma_\alpha(x),\quad k\in \mathbb{N}.\] We consider the operator \(\Delta_\alpha\) given by \[\Delta_\alpha f=\sum_{k=0}^\infty \lambda_k^\alpha c_k^\alpha(f)\mathcal{L}_k^\alpha,\quad f\in D(\Delta_\alpha),\] being \[D(\Delta_\alpha)=\left\{f\in L^2((0,\infty),\gamma_\alpha):\,\sum_{k=0}^\infty (\lambda_k^\alpha|c_k^\alpha(f)|)^2<\infty\right\}.\] The space \(C_c^\infty(0,\infty)\) of all the smooth functions with compact support in \((0,\infty)\) is contained in \(D(\Delta_\alpha)\) and \(\Delta_\alpha f=\widetilde{\Delta_\alpha}f\), for any \(f\in C_c^\infty(0,\infty)\). The operator \(\Delta_\alpha\) is self-adjoint and positive in \(L^2((0,\infty),\gamma_\alpha)\). Furthermore, the operator \(-\Delta_\alpha\) generates a \(C_0\)-semigroup of operators \(\{W_t^{\alpha }\}_{t>0}\), where, for every \(t>0\), \[W_t^\alpha(f)=\sum_{k=0} ^\infty e^{-\lambda_k^\alpha t}c_k^\alpha(f)\mathcal{L}^\alpha_k,\quad f\in L^2((0,\infty),\gamma_\alpha).\] According to we have that, for every \(x,y,t\in (0,\infty)\), \[\begin{aligned} \label{1.1} \nonumber\sum_{k=0}^\infty e^{-kt}&\mathcal{L}_k^\alpha(x)\mathcal{L}_k^\alpha(y)\\ &=\frac{\Gamma(\alpha+1)}{1-e^{-t}}(e^{-t/2}xy)^{-\alpha}I_\alpha\left(\frac{2e^{-t/2}xy}{1-e^{-t}}\right)\exp\left(-\frac{e^{-t}(x^2+y^2)}{1-e^{-t}}\right), \end{aligned}\] being \(I_\alpha\) the modified Bessel function of the first kind and order \(\alpha\). By using [\[1.1\]](#1.1){reference-type="eqref" reference="1.1"} we can write, for every \(f\in L^2((0,\infty),\gamma_\alpha)\) and \(t>0\), \[\label{1.2} W_t^\alpha(f)(x)=\int_0^\infty W_t^\alpha(x,y)f(y)d\gamma_\alpha(y), \quad x\in (0,\infty),\] where \[W_t^\alpha(x,y)=\frac{\Gamma(\alpha+1)e^{-t(\alpha+1)}}{1-e^{-2t}}(e^{-t}xy)^{-\alpha}I_\alpha\left(\frac{2e^{-t}xy}{1-e^{-2t}}\right)\exp\left(-\frac{e^{-2t}(x^2+y^2)}{1-e^{-2t}}\right),\] for \(x,y,t\in (0,\infty).\) The integral in [\[1.2\]](#1.2){reference-type="eqref" reference="1.2"} is absolutely convergent for every \(f\in L^p((0,\infty),\gamma_\alpha)\), \({1\le p<\infty}\), and for every \(t,x\in (0,\infty)\). By defining \(W_t^\alpha(f)\) by [\[1.2\]](#1.2){reference-type="eqref" reference="1.2"}, for every \(f\in L^p((0,\infty),\gamma_\alpha)\) and \(t>0\), the family \(\{W_t^\alpha\}_{t>0}\) is a \(C_0\)-semigroup in \(L^p((0,\infty),\gamma_\alpha)\), for every \(1\le p<\infty\). Thus \(\{W_t^\alpha\}_{t>0}\) is a symmetric diffusion semigroup in the sense of Stein (). The study of harmonic analysis in Laguerre settings was begun by Muckenhoupt () who proved that the maximal operator \(W_*^\alpha\) defined by \[W_*^\alpha(f)=\sup_{t>0}|W_t^\alpha(f)|\] is bounded from \(L^1((0,\infty),\gamma_\alpha)\) into \(L^{1,\infty}((0,\infty,\gamma_\alpha)\). This property was generalized by Dinger () to higher dimensions. We define the Riesz transform \(R^\alpha\) associated with the Laguerre operator \(\Delta_\alpha\) by \[R^\alpha(f)=\sum_{k=1}^\infty\frac{1}{\sqrt{\lambda_k^\alpha}}c_k^\alpha(f)\frac{d}{dx}\mathcal{L}_k^\alpha,\quad f\in L^2((0,\infty),\gamma_\alpha).\] Thus \(R^\alpha\) defines a bounded operator on \(L^2((0,\infty),\gamma_\alpha)\) (see ). Furthermore, \(R^\alpha\) can be extended from \(L^2((0,\infty),\gamma_\alpha)\cap L^p((0,\infty),\gamma_\alpha)\) as a bounded operator on \(L^p((0,\infty),\gamma_\alpha)\), for every \(1<p<\infty\), and from \(L^1((0,\infty),\gamma_\alpha)\) into \(L^{1,\infty}((0,\infty),\gamma_\alpha)\) (). The authors and R. Scotto () extended the above results by considering variable exponents \(L^{p(\cdot)}\)-spaces. Also, in, endpoint estimates for Riesz transform \(R^\alpha\) were established proving that \(R^\alpha\) defines a bounded operator from \(H^1((0,\infty),\gamma_\alpha)\) into \(L^1((0,\infty),\gamma_\alpha)\) and from \(L^\infty((0,\infty),\gamma_\alpha)\) into \(\textup{BMO}((0,\infty),\gamma_\alpha)\). We can see that \(R^\alpha\) is a principal value integral operator. By proceeding as in the proof of we can see that, for every \(f\in L^p((0,\infty),\gamma_\alpha)\), \(1\le p<\infty\), \[R^\alpha(f)(x)=\lim_{\varepsilon\to 0^+}\int_{|x-y|>\varepsilon,\ y\in (0,\infty)}R^\alpha(x,y)f(y)d\gamma_\alpha(y),\quad \text{a.e. } x\in (0,\infty),\] where \[R^\alpha(x,y)=\frac{1}{\sqrt{\pi}}\int_0^\infty \partial_xW_t^\alpha(x,y)\frac{dt} {\sqrt{t}},\quad x,y\in (0,\infty),\ x\neq y.\] For every \(\epsilon>0\), we define the \(\epsilon\)-truncation of the Riesz transform \(R^\alpha\) by \[R^\alpha_\epsilon(f)(x)=\int_{|x-y|>\varepsilon,\ y\in (0,\infty)}R^\alpha(x,y)f(y)d\gamma_\alpha(y),\quad x\in (0,\infty).\] The maximal Riesz transform \(R^\alpha_*\) is defined by \[R^\alpha_*(f)=\sup_{\epsilon>0}|R^\alpha_\epsilon(f)|.\] From the results given by E. Sasso in we can deduced that the maximal operator \(R^\alpha_*\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\), for every \(1<p<\infty\), and from \(L^1((0,\infty),\gamma_\alpha)\) into \(L^{1,\infty}((0,\infty),\gamma_\alpha)\). We are going to consider the following local maximal Riesz transform operators. For every \(a>0\), we define the maximal operator \(R^\alpha_{*,a}\) by \[R^\alpha_{*,a}(f)(x)= \sup_{0<\epsilon\le am(x)}|R^\alpha_\epsilon(f)(x)|, \quad x\in (0,\infty).\] Let \(\rho>0\). If \(\{c_t\}_{t>0}\) is a subset of complex numbers, we define the \(\rho\)-variation \(\mathcal{V}_\rho(\{c_t\}_{t>0})\) of \(\{c_t\}_{t>0}\) by \[\mathcal{V}_\rho(\{c_t\}_{t>0})=\sup_{0<t_n<t_{n-1}<\dots<t_1,\ n\in\mathbb{N}}\left(\sum_{j=1}^{n-1}|c_{t_j}-c_{t_{j+1}}|^\rho\right)^{1/\rho}.\] If \(\{T_t\}_{t>0}\) is a family of bounded operators in \(L^p((0,\infty),\gamma_\alpha)\), with \(1\le p<\infty\), we define the \(\rho\)-variation operator \(\mathcal{V}_\rho(\{T_t\}_{t>0})\) of \(\{T_t\}_{t>0}\) by \[\mathcal{V}_\rho(\{T_t\}_{t>0})(f)(x)=\mathcal{V}_\rho(\{T_t(f)(x)\}_{t>0}).\] Since Bourgain () studied variational inequalities involving martingales (see also ), \(\rho\)-variation operators has been extensively studied in ergodic theory and harmonic analysis. Campbell, Jones, Reinhold and Wierdl () proved \(L^p\)-boundedness properties for \(\rho\)-variation operators associated to the family of truncations for the Hilbert transform. In those results were extended by considering Riesz transforms in higher dimensions. In order to obtain \(L^p\)-boundedness for \(\rho\)-variation operators it is usual to ask for the condition \(\rho>2\) (see ). For the exponent \(\rho=2\), oscillation operators are commonly considered. Let \(\{t_j\}_{j\in \mathbb{Z}}\) be an increasing sequence of positive real numbers satisfying that \(\lim_{j\to-\infty}t_j=0\) and \(\lim_{j\to +\infty}t_j=+\infty\). If \(\{c_t\}_{t>0}\) is a set of complex numbers, we define the oscillation with respect to \(\{t_j\}_{j\in \mathbb{Z}}\) by \[\mathcal{O}(\{c_t\}_{t>0},\{t_j\}_{j\in \mathbb{Z}})=\left(\sum_{j=-\infty}^{+\infty}\sup_{t_j\le\epsilon_j<\epsilon_{j+1}<t_{j+1}}|c_{\epsilon_j}-c_{\epsilon_{j+1}}|^2\right)^{1/2}.\] If \(\{T_t\}_{t>0}\) is a family of bounded operators in \(L^p((0,\infty),\gamma_\alpha)\), with \(1\le p<\infty\), we define the oscillation operator \(\mathcal{O}(\{T_t\}_{t>0},\{t_j\}_{j\in \mathbb{Z}})\) as follows \[\mathcal{O}(\{T_t\}_{t>0},\{t_j\}_{j\in \mathbb{Z}})(f)(x)=\mathcal{O}(\{T_t(f)(x)\}_{t>0},\{t_j\}_{j\in \mathbb{Z}}).\] \(L^p\)-boundedness properties of the oscillation operators defined by the family of truncations of Hilbert transform and Euclidean Riesz transforms were established in and, respectively. After and, the study of \(\rho\)-variation and oscillation operators defined by singular integrals has been an active working area (see, for instance,, ,, ,, ,, and ). Variation and oscillation operators give information about convergence properties for the family \(\{T_t\}_{t>0}\). Being \(\{T_t\}_{t>0}\) and \(\{t_j\}_{j\in\mathbb{Z}}\) as above, we are going to consider the local \(\rho\)-variation and oscillation operators defined as follows. Let \(a>0\). The \(a\)-local \(\rho\)-variation operator \(\mathcal{V}_{\rho,a}(\{T_t\}_{t>0})\) is given by \[\begin{aligned} \mathcal{V}_{\rho,a}&(\{T_t\}_{t>0})(f)(x)\\ &=\sup_{0<t_n<t_{n-1}<\dots<t_1\le am(x),\ n\in\mathbb{N}}\left(\sum_{j=1}^{n-1}|T_{t_j}(f)(x)-T_{t_{j+1}}(f)(x)|^\rho\right)^{1/\rho}. \end{aligned}\] The \(a\)-local oscillation operator \(\mathcal{O}_a(\{T_t\}_{t>0},\{t_j\}_{j\in \mathbb{Z}})\) is defined by \[\begin{aligned} \mathcal{O}_a&(\{T_t\}_{t>0},\{t_j\}_{j\in \mathbb{Z}})(f)(x)\\ &=\left(\sum_{j\in \mathbb{Z},\,\,t_j\le am(x)}\sup_{t_j\le\epsilon_j<\epsilon_{j+1}<t_{j+1}}|T_{\epsilon_j}(f)(x)-T_{\epsilon_{j+1}}(f)(x)|^2\right)^{1/2}. \end{aligned}\] Our first result is the following. We shall now introduce multiplier operators in the Laguerre setting. A measurable complex function \(M\) defined on \([0,\infty)\) is said to be of Laplace transform type when \[M(x)=x\int_0^\infty \phi(t)e^{-xt}dt,\quad x>0,\] where \(\phi\in L^\infty(0,\infty)\). Suppose that \(M\) is of Laplace transform type. We denote by \(T_M^\alpha\) the spectral multiplier for the Laguerre operator \(\Delta_\alpha\) defined by \(M-M(0)\). For every \({f\in L^2((0,\infty),\gamma_\alpha)}\), \(T_M^\alpha(f)\) is given by \[T_M^\alpha(f)=\sum_{k=1}^\infty M(k)c_k^\alpha(f)\mathcal{L}_k^\alpha.\] Since \(M\) is bounded on \((0,\infty)\), \(T_M^\alpha\) is bounded on \(L^2((0,\infty),\gamma_\alpha)\). Since \(\{W_t^\alpha\}_{t>0}\) is a symmetric diffusion semigroup, \(T_M^\alpha\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\), for every \({1<p<\infty}\) (). The authors and R. Scotto () extended the last result establishing variable \(L^{p(\cdot)}\)-boundedness properties for \(T_M^\alpha\). On the other hand, Sasso () proved that \(T_M^\alpha\) defines a bounded operator from \(L^1((0,\infty),\gamma_\alpha)\) into \(L^{1,\infty}((0,\infty),\gamma_\alpha)\). In, the authors with R. Scotto established the endpoint estimate for \(T_M^\alpha\) from \(L^\infty((0,\infty),\gamma_\alpha)\) into \(\textup{BMO}((0,\infty),\gamma_\alpha)\). From we deduce that there exists a function \(\Lambda\in L^\infty(0,\infty)\) such that, for every \(f\in L^p((0,\infty),\gamma_\alpha)\), \(1\le p<\infty\), \[T_M^\alpha(f)(x)=\lim_{\varepsilon\to 0^+}\left(\Lambda(\varepsilon)f(x)+\int_{|x-y|>\varepsilon,\ y\in (0,\infty)}K_\phi^\alpha(x,y)f(y)d\gamma_\alpha(y)\right),\] for a.e. \(x\in (0,\infty)\), where \[K_\phi^\alpha(x,y)=-\int_0^\infty\phi(t)\partial_tW_t^\alpha(x,y)dt,\quad x,y\in (0,\infty),\ x\neq y.\] A special case of \(T_M^\alpha\) is the imaginary power \(\Delta_\alpha^{i\alpha}\) that appears when \(M_{\eta}(x)=x^{i\eta}\) for \(x\in (0,\infty)\) and \(\eta\in\mathbb{R}\setminus \{0\}\). For these values of \(\eta\), \[M_\eta(x)=x\int_0^\infty \phi_\eta(t)e^{-xt}dt,\quad x\in (0,\infty),\] where \(\phi_\eta(t)=\frac{t^{-i\eta}}{\Gamma(1+i\eta)}\), \(t>0\). Note that \(|\phi'_\eta(t)|\le C/t\), \(t\in (0,\infty)\). We define, for every \(\epsilon>0\), the truncations \[Q_{\phi,\epsilon}^\alpha(f)(x)=\int_{|x-y|>\epsilon,\ y\in (0,\infty)}K_\phi^\alpha(x,y)f(y)d\gamma_\alpha(y),\quad x\in (0,\infty),\] and consider, for every \(a>0\), the \(a\)-local maximal operator \(Q^\alpha_{\phi,*,a}\), which is given by \[Q^\alpha_{\phi,*,a}(f)(x)=\sup_{0<\epsilon\le am(x)}|Q_{\phi,\epsilon}^\alpha(f)(x)|.\] The paper is organized as follows. In Section [2](#sec: BLO){reference-type="ref" reference="sec: BLO"} we state the main properties for the spaces \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\). In the subsequent sections we prove Theorems [\[Th1.1\]](#Th1.1){reference-type="ref" reference="Th1.1"} and [\[Th1.2\]](#Th1.2){reference-type="ref" reference="Th1.2"}. Throughout this paper \(C\) and \(c\) will always denote positive constants than may change in each occurrence. # The spaces \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\) {#sec: BLO} In this section we state the main properties of the spaces \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\). This properties will be useful in the following sections and they can be proved as the corresponding properties for the Gaussian \(\textup{BLO}_a\) space given in (see also for the Euclidean case and for the non-doubling measure case). Let \(a>0\). The local natural maximal operator \(\mathcal{M}^\alpha_a\) associated with the measure \(\gamma_\alpha\) on \((0,\infty)\) is defined by \[\mathcal{M}^\alpha_a (f)(x) = \sup_{I\in \mathcal{B}_a(x)} \frac{1}{\gamma_\alpha(I)} \int_I f(y)d\gamma_\alpha(y),\;x\in(0,\infty),\] for every measurable function \(f\) on \((0,\infty)\) such that \(\int_0^\delta |f(y)| d\gamma_\alpha(y)<\infty\), \(\delta >0\). The space \(\textup{BLO}_a ((0,\infty),\gamma_\alpha)\) can be characterized by using the local natural maximal operator. By combining Proposition [\[propo2.1\]](#propo2.1){reference-type="ref" reference="propo2.1"} and Proposition [\[propo2.2\]](#propo2.2){reference-type="ref" reference="propo2.2"} we can establish the following characterization of \(\textup{BLO}_a ((0,\infty),\gamma_\alpha)\) involving the space \(\textup{BMO} ((0,\infty),\gamma_\alpha)\) and the local natural maximal operator. # Proof of Theorem [\[Th1.1\]](#Th1.1){reference-type="ref" reference="Th1.1"} ## Local variation operators {#subsec: variation} Let \(f\in L^\infty((0,\infty),\gamma_\alpha)\). Since the variation operator \(\mathcal{V}_\rho(\{R^\alpha_\epsilon\}_{\epsilon>0})\) is bounded on \(L^2((0,\infty),\gamma_\alpha)\) (see ) it follows that \[\begin{aligned} \int_0^\infty \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x)d\gamma_\alpha(x) & \leq \left(\int_0^\infty \left( \mathcal{V}_\rho( \{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x)\right)^2d\gamma_\alpha(x) \right)^{1/2} \\ & \leq C \left(\int_0^\infty |f(x)|^2 d\gamma_\alpha(x) \right)^{1/2}\\ &\leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}. \end{aligned}\] According to Proposition [\[propo2.2\]](#propo2.2){reference-type="ref" reference="propo2.2"} the proof will be finished when we see that \[\|\mathcal{M}^\alpha_a (\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f))-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)\|_{L^{\infty}((0,\infty),\gamma_\alpha)} \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}.\] Notice that \[\begin{aligned} 0 & \leq \mathcal{M}^\alpha_a (\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f))(x)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x) \\ & = \sup_{I \in\mathcal{B}_a(x)} \frac{1}{\gamma_\alpha(I)} \int_{I } \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(z) d\gamma_\alpha(z)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x), \end{aligned}\] for almost every \(x\in (0,\infty)\), where \(I \in\mathcal{B}_a(x)\) indicates that \(I\in \mathcal{B}_a\) and \(x\in I\). Let \(x\), \(x_0\), \(r_0\in (0,\infty)\) such that \(I=I(x_0,r_0)\in \mathcal{B}_a(x)\). We decompose \(f\) as follows \[f = f \chi_{4I} + f \chi_{(0,\infty)\setminus 4I} = f_1 + f_2.\] We can write \[\begin{aligned} \frac{1}{\gamma_\alpha(I)} \int_{I } & \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(z) d\gamma_\alpha(z)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x) \\ & \leq \frac{1}{\gamma_\alpha(I)} \int_{I } \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_1)(z) d\gamma_\alpha(z) \\ & \quad + \frac{1}{\gamma_\alpha(I)} \int_{I } \left( \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_2)(z)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_2)(x) \right) d\gamma_\alpha(z) \\ & \quad + \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_2)(x)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)(x) \\ & := J_1 + J_2 + J_3. \end{aligned}\] By using again that the variation \(\mathcal{V}_\rho(\{R^\alpha_\epsilon\}_{\epsilon>0})\) is bounded on \(L^2((0,\infty),\gamma_\alpha)\) we get \[\begin{aligned} J_1 & \leq \left( \frac{1}{\gamma_\alpha(I)} \int_{I } \left( \mathcal{V}_\rho(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_1)(z) \right)^2d\gamma_\alpha(z)\right)^{1/2}\nonumber \\ & \leq C \left( \frac{1}{\gamma_\alpha(I)} \int_{I } \left| f(z) \right|^2 d\gamma_\alpha(z)\right)^{1/2} \leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}. \end{aligned}\] Suppose there exists \(i_0\in \{1,\dots, n-1\}\) such that \(\epsilon_{i_0+1}\leq am(x)<\epsilon_{i_0}\). Thus, for \(z\in I\), \[\begin{aligned} &\left( \sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho} \\ & \leq \left( \sum_{j=1}^{i_0-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho + |R^\alpha_{\epsilon_{i_0}}(f_2)(z)-R^\alpha_{am(x)}(f_2)(z)|^\rho \right)^{1/\rho} \\ &\quad + \left( \sum_{j=i_0+1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho + |R^\alpha_{\epsilon_{i_0+1}}(f_2)(z)-R^\alpha_{am(x)}(f_2)(z)|^\rho \right)^{1/\rho}. \end{aligned}\] Then, recalling that \(m(z)\leq Cm(x)\) for every \(x,z\in I\), where \(C>1\), we obtain \[\begin{aligned} &\mathcal{V}_{\rho,a}( \{R^\alpha_\epsilon\}_{\epsilon>0}) (f_2)(z)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f_2)(x) \\ & \leq \sup_{\substack{ 0<\epsilon_n<\dots<\epsilon_1\leq a m(x) \\n\in\mathbb{N}}} \left( \sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho} \\ &\quad + \sup_{\substack{am(x)\leq \epsilon_n<\dots<\epsilon_1<Cam(x)\\n\in\mathbb{N}}} \left( \sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho}\\ &\quad- \sup_{\substack{ 0<\epsilon_n<\dots<\epsilon_1\leq a m(x) \\n\in\mathbb{N}}} \left( \sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^\rho \right)^{1/\rho} \\ & \leq \sup_{\substack{am(x)\leq \epsilon_n<\dots<\epsilon_1<Cam(x)\\n\in\mathbb{N}}} \sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)| \\ & \quad + \sup_{\substack{0<\epsilon_n<\dots<\epsilon_1\leq am(x)\\n\in\mathbb{N}}} \inf_{\substack{0<\delta_k<\dots<\delta_1\leq am(x)\\k\in\mathbb{N}}} \left[ \left(\sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho} \right. \\ & \qquad-\left.\left(\sum_{j=1}^{k-1} |R^\alpha_{\delta_{j+1}}(f_2)(x)-R^\alpha_{\delta_{j}}(f_2)(x)|^\rho \right)^{1/\rho}\right] \\ & \leq \int_{am(x)<|z-y|<Cam(x)} |R^\alpha(z,y)||f_2(y)| d\gamma_\alpha(y) \\ & \quad + \sup_{\substack{0<\epsilon_n<\dots<\epsilon_1\leq am(x)\\n\in\mathbb{N}}} \inf_{\substack{0<\delta_k<\dots<\delta_1\leq am(x)\\k\in\mathbb{N}}} \left[ \left(\sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho} \right. \\ & \qquad-\left.\left(\sum_{j=1}^{k-1} |R^\alpha_{\delta_{j+1}}(f_2)(x)-R^\alpha_{\delta_{j}}(f_2)(x)|^\rho \right)^{1/\rho}\right]\\ & := J_{2,1}(x,z) + J_{2,2}(x,z). \end{aligned}\] If we write \[R^\alpha(z,y) = e^{\frac{z^2 + y^2}{2}} \mathfrak{R}^\alpha(z,y), \quad z,y\in(0,\infty), \, z\neq y,\] from we know that \[\label{cota-nucleo-potencia} \mathfrak{R}^\alpha(z,y) \leq \frac{C}{\mathfrak{m}_\alpha(I(z,|z-y|))}, \quad z,y\in(0,\infty), \, z\neq y,\] where \(d\mathfrak{m}_\alpha(x)=x^{2\alpha+1} dx\). Therefore, for \(x,z\in I\), since \(m(x)\leq Cm(z)\) and using, we obtain \[\begin{aligned} J_{2,1}(x,z) & \leq C \int_{am(x)<|z-y|\leq Cam(x)} |f_2(y)| e^{\frac{z^2-y^2}{2}} \frac{d\mathfrak{m}_\alpha(y)}{\mathfrak{m}_\alpha(I(z,|z-y|))} \\ & \leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}\int_{am(x)<|z-y|\leq Cam(x)} e^{\frac{(z+y)|z-y|}{2}} \frac{d\mathfrak{m}_\alpha(y)}{\mathfrak{m}_\alpha(I(y,|z-y|))} \\ & \leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}\int_{am(x)<|z-y|\leq Cam(x)} e^{Cam(x)(z+Cam(x))} \frac{dy}{|z-y|} \\ & \leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}\int_{am(x)<|z-y|\leq Cam(x)} \frac{dy}{|z-y|} \\ & \leq C\|f\|_{L^\infty((0,\infty),\gamma_\alpha)}. \end{aligned}\] On the other hand, for every \(x,z\in I\) we have \[\begin{aligned} &\left(\sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)|^\rho \right)^{1/\rho}-\left(\sum_{j=1}^{n-1} |R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^\rho \right)^{1/\rho} \\ & \leq \left(\sum_{j=1}^{n-1} \left|R^\alpha_{\epsilon_{j+1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)-\left(R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)\right)\right|^\rho \right)^{1/\rho} \\ & \leq \left(\sum_{j=1}^{n-1} \left| \int_{\epsilon_{j+1}<|z-y|<\epsilon_j} (R^\alpha(z,y)-R^\alpha(x,y)) f_2(y) d\gamma_\alpha(y) \right.\right. \\ & \quad +\left. \left. \left( \int_{\epsilon_{j+1}<|z-y|<\epsilon_j} R^\alpha(x,y) f_2(y) d\gamma_\alpha(y)\right.\right.\right.\\ & \qquad-\left.\left.\left. \int_{\epsilon_{j+1}<|x-y|<\epsilon_j} R^\alpha(x,y) f_2(y) d\gamma_\alpha(y)\right)\right|^\rho \right)^{1/\rho} \\ & \leq \sum_{j=1}^{n-1} \left| \int_{\epsilon_{j+1}<|z-y|<\epsilon_j} (R^\alpha(z,y)-R^\alpha(x,y)) f_2(y) d\gamma_\alpha(y)\right| \\ & \quad + \left( \sum_{j=1}^{n-1} \left| \int_0^\infty R^\alpha(x,y) \left(\chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \right.\right.\right. \\ & \qquad-\left.\left.\left.\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \right) f_2(y) d\gamma_\alpha(y)\right|^\rho \right)^{1/\rho}. \end{aligned}\] Now, by taking supremum, we get \[\begin{aligned} J_{2,2}(x,z)&\leq \int_{(0,\infty)\setminus 4I} |R^\alpha(z,y)-R^\alpha(x,y) | |f(y)| d\gamma_\alpha(y) \\ & \quad + \sup_{\substack{0<\epsilon_n<\dots<\epsilon_1\leq am(x)\\n\in\mathbb{N}}} \left(\sum_{j=1}^{n-1} \left( \int_0^\infty \left|R^\alpha(x,y)\right| \left|\chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \right.\right.\right. \\ & \qquad-\left.\left.\left.\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \right| |f_2(y)| d\gamma_\alpha(y)\right)^\rho \right)^{1/\rho} \\ & := J_{2,2,1}(x,z) + J_{2,2,2}(x,z). \end{aligned}\] Since (see ) \[\sup_{I \in \mathcal{B}_a} \sup_{x,z\in I } \int_{(0,\infty)\setminus 4I} |R^\alpha(z,y)-R^\alpha(x,y) | |f(y)| d\gamma_\alpha(y) < \infty,\] it follows that \[J_{2,2,1}(x,z) \leq C \|f\|_{L^\infty((0,\infty),\gamma_\alpha)}, \quad x,z\in I.\] In order to estimate \(J_{2,2,2}\) we adapt a procedure developed in . From [\[cota-nucleo-potencia\]](#cota-nucleo-potencia){reference-type="eqref" reference="cota-nucleo-potencia"}, for \(x,z\in I\), we obtain \[\begin{aligned} J_{2,2,2}(x,z) & \leq C \sup_{\substack{0<\epsilon_n<\dots<\epsilon_1\leq am(x)\\n\in\mathbb{N}}} \left(\sum_{j=1}^{n-1} \left( \int_0^\infty \frac{e^{\frac{x^2-y^2}{2}}}{\mathfrak{m}_\alpha(I(y,|x-y|))} \right.\right. \\ & \qquad \times \left|\chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y)-\left.\left.\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \right| |f_2(y)| d\mathfrak{m}_\alpha(y)\right)^\rho \right)^{1/\rho}. \end{aligned}\] Let us observe that, if \(|x-y|\leq am(x)\), then \[x^2-y^2 \leq |x-y||x+y|\leq am(x)(am(x)+2x)\leq C.\] Also, if \(|z-y|\leq am(x)\), then \[|x-y|\leq 2r_0 + am(x) \leq 2am(x_0) + am(x).\] Since \(x\in I \in \mathcal{B}_a\), \(m(x_0)\leq Cm(x)\) so \(|x-y| \leq C am(x)\), and thus \(x^2-y^2 \leq C\), provided that \(|z-y| \leq a m(x)\). This fact together with  lead to \[\begin{aligned} J_{2,2,2}(x,z) & \leq \sup_{\substack{0<\epsilon_n<\dots<\epsilon_1\leq am(x)\\n\in\mathbb{N}}} \left(\sum_{j=1}^{n-1} \left( \int_0^\infty \frac{|f_2(y)|}{|x-y|}\left|\chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \right.\right.\right. \\ & \quad-\left.\left.\left.\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \right| dy\right)^\rho \right)^{1/\rho}, \quad x,z\in I. \end{aligned}\] Let us take \(0<\epsilon_n <\dots < \epsilon_1\leq am(x)\) and \(j\in\{1,\dots,n-1\}\). Then \[\begin{aligned} \int_0^\infty & \frac{|f_2(y)|}{|x-y|} \left| \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) -\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_ j\}}(y) \right|dy \\ & \leq C \left( \int_0^\infty \frac{|f_2(y)|}{|x-y|} \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y)\chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_{j+1} + 2r_0\}}(y) dy\right. \\ & \quad + \int_0^\infty \frac{|f_2(y)|}{|x-y|} \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y)\chi_{\{\epsilon_{j}<|z-y|<\epsilon_{j} + 2r_0\}}(y) dy \\ & \quad + \int_0^\infty \frac{|f_2(y)|}{|z-y|} \chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y)\chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_{j+1} + 2r_0\}}(y) dy \\ & \quad + \left. \int_0^\infty \frac{|f_2(y)|}{|z-y|} \chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y)\chi_{\{\epsilon_{j}<|x-y|<\epsilon_{j} + 2r_0\}}(y) dy \right) \\ & = \sum_{l=1}^4 J^{j,l}_{2,2,2} (x,z), \quad x,z\in I. \end{aligned}\] For the above estimate, we have taken into account that, if \(\chi_{\{\epsilon_{j+1}<|x-y|\leq \epsilon_j\}} (y)-\chi_{\{\epsilon_{j+1}<|z-y|\leq \epsilon_j\}} (y)\neq 0\), then \(\chi_{\{\epsilon_{j+1}<|x-y|\leq \epsilon_j\}} (y)\chi_{\{\epsilon_{j+1}<|z-y|\leq \epsilon_j\}} (y) = 0\), with \({y\in(0,\infty)}\) and \(x\), \(z\in I\). Since \(f_2(y) = 0\) for \(y\in 4I\), it follows that \(J^{j,l}_{2,2,2} = 0\) when \(l=1,3\), \(z\in I\) and \(r_0\geq \epsilon_{j+1}\). Also, \(J^{j,l}_{2,2,2}(x,z) = 0\) when \(l=2,4\), \(z\in I\) and \(r_0\geq \epsilon_j\). If \(z\in I\) and \(y\notin 4I\), then \(2|x-y|\geq |z-y| \geq \frac12 |x-y|\). Hölder inequality leads to \[J^{j,l}_{2,2,2} (x,z) \leq C \left( \int_0^\infty \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \left(\frac{|f_2(y)|}{|x-y|}\right)^2 dy\right)^{1/2} r_0^{1/2}, \quad z\in I, \, l=1,2;\] \[J^{j,l}_{2,2,2} (x,z) \leq C \left( \int_0^\infty \chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \left(\frac{|f_2(y)|}{|z-y|}\right)^2 dy\right)^{1/2} r_0^{1/2}, \quad z\in I, \, l=3,4.\] We obtain \[\begin{aligned} &\left(\sum_{j=1}^{n-1}\left| \sum_{l=1}^{4} J^{j,l}_{2,2,2} (x,z) \right|^\rho\right)^{1/\rho} \\ &\quad \leq C r_0^{1/2} \left[ \left( \sum_{j=1}^{n-1} \left( \int_0^\infty \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \left(\frac{|f_2(y)|}{|x-y|}\right)^2 dy\right)^{\rho/2}\right)^{1/\rho} \right. \\ &\qquad + \left. \left( \sum_{j=1}^{n-1} \left( \int_0^\infty \chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \left(\frac{|f_2(y)|}{|z-y|}\right)^2 dy\right)^{\rho/2} \right)^{1/\rho} \right] \\ & \quad \leq C r_0^{1/2} \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)} \left[ \left( \sum_{j=1}^{n-1} \int_{(0,\infty)\setminus 4I} \chi_{\{\epsilon_{j+1}<|x-y|<\epsilon_j\}}(y) \frac{dy}{|x-y|^2} \right)^{1/2} \right. \\ & \qquad + \left. \left( \sum_{j=1}^{n-1} \int_{(0,\infty)\setminus 4I} \chi_{\{\epsilon_{j+1}<|z-y|<\epsilon_j\}}(y) \frac{dy}{|z-y|^2}\right)^{1/2} \right] \\ & \quad\leq C r_0^{1/2} \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)} \left( \int_{(0,\infty)\setminus 4I} \frac{dy}{|x-y|^2} + \int_{(0,\infty)\setminus 4I} \frac{dy}{|z-y|^2}\right)^{1/2} \\ & \quad \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}. \end{aligned}\] We conclude that \(J_{2,2,2}(x,z) \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}\). By putting together the above estimates we obtain \[\mathcal{V}_{\rho,a} \left(\{R^\alpha_\epsilon\}_{\epsilon>0}\right) (f_2)(z)- \mathcal{V}_{\rho,a} \left(\{R^\alpha_\epsilon\}_{\epsilon>0}\right) (f_2)(x) \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}, \quad x,z\in I.\] Here, \(C>0\) does not depend on \(x\), \(z\in I\), so it follows that \[J_2 \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}.\] We now estimate \(J_3\). Note first that if \(|x-y|<r_0\), then \(|y-x_0|<2r_0\), so it is clear that \[\int_{\epsilon_1<|x-y|<\epsilon_2} R^\alpha(x,y)f_2(y) d\gamma_\alpha(y) = 0, \quad 0<\epsilon_1<\epsilon_2\leq r_0.\] Suppose that \(r_0\leq a m(x)\). We have, for any \(x\in I\), that \[\begin{aligned} \mathcal{V}_{\rho,a}&\left(\{ R^\alpha_\epsilon\}_{\epsilon>0}\right)(f_2)(x)\\& \leq \sup_{0<\epsilon_n<\dots<\epsilon_1 \leq r_0, \ n\in \mathbb{N}} \left(\sum_{j=1}^{n-1} | R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^\rho\right)^{1/\rho} \\ & \quad + \sup_{r_0<\epsilon_n<\dots<\epsilon_1 \leq am(x), \ n\in \mathbb{N}} \left(\sum_{j=1}^{n-1} | R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^\rho\right)^{1/\rho} \\ & = \sup_{r_0<\epsilon_n<\dots<\epsilon_1 \leq am(x), \ n\in \mathbb{N}} \left(\sum_{j=1}^{n-1} | R^\alpha_{\epsilon_{j+1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^\rho\right)^{1/\rho} \\ & \leq \mathcal{V}_{\rho,a}\left(\{ R^\alpha_\epsilon\}_{\epsilon>0}\right)(f)(x) \\ & \quad + \sup_{r_0<\epsilon_n<\dots<\epsilon_1 \leq am(x), \ n\in \mathbb{N}} \left(\sum_{j=1}^{n-1} | R^\alpha_{\epsilon_{j+1}}(f_1)(x)-R^\alpha_{\epsilon_{j}}(f_1)(x)|^\rho\right)^{1/\rho}. \end{aligned}\] Since, for every \(y\in 4I\), \(|x-y|\leq 5r_0\leq 5 am(x)\), by using again [\[cota-nucleo-potencia\]](#cota-nucleo-potencia){reference-type="eqref" reference="cota-nucleo-potencia"} and, we deduce that \[\begin{aligned} \sup_{r_0<\epsilon_n<\dots<\epsilon_1 \leq am(x), \ n\in \mathbb{N}} & \left(\sum_{j=1}^{n-1} | R^\alpha_{\epsilon_{j+1}}(f_1)(x)-R^\alpha_{\epsilon_{j}}(f_1)(x)|^\rho\right)^{1/\rho} \\ & \leq C \int_{|x-y|>r_0,\ y\in 4I} \frac{e^{\frac{x^2+y^2}{2}}|f(y)|}{\mathfrak{m}_\alpha(I(y,|x-y|))} d\gamma_\alpha(y) \\ & \leq C \int_{|x-y|>r_0,\ y\in 4I}\frac{e^{\frac{x^2-y^2}{2}}|f(y)|}{|x-y|} dy \\ & \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}\int_{|x-y|>r_0,\ y\in 4I}\frac{dy}{|x-y|} \\ & \leq C \frac{\|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}}{r_0}\int_{4I} dy\\ & \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}, \end{aligned}\] that is, \(J_3\leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}\) for the case \(r_0\leq am(x)\). When \(r_0>am(x)\), \[\mathcal{V}_{\rho,a}\left(\{ R^\alpha_\epsilon\}_{\epsilon>0}\right)(f_2)(x) = 0,\] so \(J_3\leq 0\) in this case. We conclude that \[J_3 \leq C \|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}.\] By combining the above estimates, since the constant \(C>0\) does not depend on \(x\in (0,\infty)\) or \(I\in \mathcal{B}_a(x)\), we get \[\|\mathcal{M}^\alpha_a \left( \mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f)\right)-\mathcal{V}_{\rho,a}(\{R^\alpha_\epsilon\}_{\epsilon>0}) (f) \|_{L^{\infty}(0,\infty),\gamma_\alpha} \leq C \|f\|_{L^{\infty}(0,\infty),\gamma_\alpha}.\] Thus the proof is finished. ## Local oscillation operators Theorem [\[Th1.1\]](#Th1.1){reference-type="ref" reference="Th1.1"} for oscillation operators can be proved by using the procedure developed in the previous section for the variation operator, so we give a sketch of the proof. According to, the oscillation operator \(\mathcal{O}(\{R_\epsilon^\alpha\}_{\epsilon>0}, \{t_j\}_{j\in \mathbb Z})\) is bounded on \(L^2((0,\infty),\gamma_\alpha)\). This property implies that \(\mathcal{O}_a(\{R_\epsilon^\alpha\}_{\epsilon>0}, \{t_j\}_{j\in \mathbb Z})(f)\in L^1((0,\infty),\gamma_\alpha)\) for every \(f\in L^\infty((0,\infty),\gamma_\alpha)\). In order to prove our result, it is sufficient to find a positive constant \(C\) such that, for every \(f\in L^\infty((0,\infty),\gamma_\alpha)\), \[\begin{aligned} \label{norm_M-O} \nonumber \|\mathcal{M}^\alpha_a (\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f))-& \mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f)\|_{L^{\infty}((0,\infty),\gamma_\alpha)}\\ &\leq C\|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}. \end{aligned}\] Fix \(f\in L^\infty((0,\infty),\gamma_\alpha)\) and let \(x,x_0,r_0\in (0,\infty)\) such that \(I=I(x_0,r_0)\in \mathcal{B}_a(x)\). We write \(f=f\chi_{4I}+f\chi_{(0,\infty)\setminus 4I}:=f_1+f_2\). \[\begin{aligned} &\frac{1}{\gamma_\alpha(I)} \int_{I } \mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f) (z) d\gamma_\alpha(z)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f) (x)\\ &\leq \frac{1}{\gamma_\alpha(I)} \int_{I } \mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_1) (z) d\gamma_\alpha(z)\\ &\quad + \frac{1}{\gamma_\alpha(I)} \int_{I } \left[\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (z)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)\right] d\gamma_\alpha(z)\\ &\quad +\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f) (x)\\ &:=J_1+J_2+J_3. \end{aligned}\] It is immediate from the \(L^2\)-boundedness of \(\mathcal{O}(\{R_\epsilon^\alpha\}_{\epsilon>0}, \{t_j\}_{j\in \mathbb Z})\) () that \[J_1\leq C\|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}.\] We now estimate the integrand of \(J_2\). For certain \(C>1\), we have \[\begin{aligned} \mathcal{O}_a (\{R^\alpha_\epsilon\}_{\epsilon>0}, &\{t_j\}_{j\in \mathbb Z}) (f_2) (z)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)\\ &\leq \left(\sum_{\substack{j\in \mathbb Z\\ t_j\leq C am(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|R^\alpha_{\epsilon_{j-1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)\right|^2 \right)^{1/2}\\ &\quad-\left(\sum_{\substack{j\in \mathbb Z\\ t_j\leq am(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|R^\alpha_{\epsilon_{j-1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)\right|^2 \right)^{1/2}. \end{aligned}\] We define \[\label{j0} j_0(x)=\max\{j\in \mathbb Z: t_j\leq am(x)\}\] and also, provided that \(t_{j_0(x)+1}\leq Cam(x)\), we consider \[\label{j1} j_1(x)=\max\{j\in \mathbb Z: j>j_0(x), t_j\leq Cam(x)\}.\] Thus, when \(t_{j_0(x)+1}> Cam(x)\), we can write \[\begin{aligned} \mathcal{O}_a &(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (z)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)\\ &\leq \left(\sum_{\substack{j\in \mathbb Z\\ j\leq j_0(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|R^\alpha_{\epsilon_{j-1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)\right|^2 \right)^{1/2}\\ &\quad-\left(\sum_{\substack{j\in \mathbb Z\\ j\leq j_0(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|R^\alpha_{\epsilon_{j-1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)\right|^2 \right)^{1/2}\\ &\leq \left(\sum_{\substack{j\in \mathbb Z\\ j\leq j_0(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|D(x,z)\right|^2 \right)^{1/2}:=\tilde{J_2}(x,z), \end{aligned}\] where \[D(x,z):=R^\alpha_{\epsilon_{j-1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)-\left(R^\alpha_{\epsilon_{j-1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)\right).\] On the other hand, if \(t_{j_0(x)+1}\leq Cam(x)\), we get \[\begin{aligned} \mathcal{O}_a &(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (z)-\mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)\\ &\leq \tilde{J_2}(x,z)+\left(\sum_{\substack{j\in \mathbb Z\\j_0(x)< j\leq j_1(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|R^\alpha_{\epsilon_{j-1}}(f_2)(z)-R^\alpha_{\epsilon_{j}}(f_2)(z)\right|^2 \right)^{1/2}\\ &\leq \tilde{J_2}(x,z)+\int_{\frac a\rho m(x)\leq |z-y|\leq Cam(x)} |R^\alpha(z,y)||f_2(y)| d\gamma_\alpha(y), \end{aligned}\] where in the last inequality we have used that \(t_{j_0(x)}\leq am(x)\leq t_{j_0(x)+1}\leq \rho t_{j_0(x)}\) with \(\rho>1\). Notice that we can estimate \(\tilde{J_2}(x,z)\) in the following form \[\begin{aligned} &\tilde{J_2}(x,z)\\ &\leq \left(\sum_{\substack{j\in \mathbb Z\\ j\leq j_0(x)}} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\left|\int_{\epsilon_{j-1}<|z-y|<\epsilon_j} (R^\alpha(z,y)-R^\alpha(x,y))f_2(y) d\gamma_\alpha(y)\right.\right.\\ &\quad +\left.\left. \int_0^\infty \left(\chi_{\{\epsilon_{j-1}<|z-y|<\epsilon_j\}}(y)-\chi_{\{\epsilon_{j-1}<|x-y|<\epsilon_j\}}(y)\right) R^\alpha(x,y) f_2(y) d\gamma_\alpha(y)\right|^2\right)^{1/2}\\ &\leq \int_{(0,\infty)\setminus 4I} |R^\alpha(z,y)-R^\alpha(x,y)||f_2(y)| d\gamma_\alpha(y)\\ &\quad +\left(\sum_{\substack{j\in \mathbb Z\\ j\leq j_0(x)}} \left(\sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j}\int_0^\infty \left|\chi_{\{\epsilon_{j-1}<|z-y|<\epsilon_j\}}(y)-\chi_{\{\epsilon_{j-1}<|x-y|<\epsilon_j\}}(y)\right|\right.\right.\\ &\qquad \times \left.\left.|R^\alpha(x,y)| |f_2(y)| d\gamma_\alpha(y) \right)^2\right)^{1/2}. \end{aligned}\] At this point, we can proceed as in the proof of the corresponding result for variation operators \(\mathcal{V}_{\rho,a}\), by using Hölder's inequality with an exponent \(s\in (1,2)\) instead of applying it with exponent 2. In this way, we deduce that \[J_2\leq C \|f\|_{L^\infty((0,\infty), \gamma_\alpha)}.\] In order to study \(J_3\), we first recall that \[\int_{\epsilon_1<|x-y|<\epsilon_2} R^\alpha(x,y) f_2(y)d\gamma_\alpha(y), \quad 0<\epsilon_1<\epsilon_2\leq r_0.\] Then, if \(r_0\geq am(x)\), we obtain \[\label{osc-zero} \mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f_2) (x)=0.\] Suppose now that \(r_0<am(x)\) and define \(j_0(x)\) as in [\[j0\]](#j0){reference-type="eqref" reference="j0"}. If \(t_{j_0(x)}\leq r_0\), we again have [\[osc-zero\]](#osc-zero){reference-type="eqref" reference="osc-zero"}. If not, we define \(j_1=\max\{j\in \mathbb Z: t_{j_1}\leq r_0\}\). Then \[\begin{aligned} \mathcal{O}_a (\{R^\alpha_\epsilon\}_{\epsilon>0},& \{t_j\}_{j\in \mathbb Z}) (f_2) (x)\\ &=\left(\sum_{j=j_1+1}^{j_0(x)} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j} |R^\alpha_{\epsilon_{j-1}}(f_2)(x)-R^\alpha_{\epsilon_{j}}(f_2)(x)|^2\right)^{1/2}\\ &\leq \mathcal{O}_a(\{R^\alpha_\epsilon\}_{\epsilon>0},\{t_j\}_{j\in \mathbb Z}) (f) (x)\\ &\quad +\left(\sum_{j=j_1+1}^{j_0(x)} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j} |R^\alpha_{\epsilon_{j-1}}(f_1)(x)-R^\alpha_{\epsilon_{j}}(f_1)(x)|^2\right)^{1/2}. \end{aligned}\] Since \(t_{j_1}\leq r_0\leq t_{j+1}\leq \rho t_{j_1}\), it follows that \[\begin{aligned} &\left(\sum_{j=j_1+1}^{j_0(x)} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j} |R^\alpha_{\epsilon_{j-1}}(f_1)(x)-R^\alpha_{\epsilon_{j}}(f_1)(x)|^2\right)^{1/2}\\ &\quad \leq \sum_{j=j_1+1}^{j_0(x)} \sup_{t_{j-1}\leq \epsilon_{j-1}<\epsilon_j\leq t_j} |R^\alpha_{\epsilon_{j-1}}(f_1)(x)-R^\alpha_{\epsilon_{j}}(f_1)(x)|\\ &\quad \leq C \int_{|x-y|>r_0/\rho} \frac{e^{\frac{x^2+y^2}{2}}|f(y)|}{\mathfrak{m}_\alpha (I(y,|x-y|))} d\gamma_\alpha(y)\\ &\quad\leq C\|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}, \end{aligned}\] where we have used again the bound given in [\[cota-nucleo-potencia\]](#cota-nucleo-potencia){reference-type="eqref" reference="cota-nucleo-potencia"} and. We conclude that \[J_3\leq C\|f\|_{L^{\infty}((0,\infty),\gamma_\alpha)}.\] By putting together all of the above estimates, we get [\[norm_M-O\]](#norm_M-O){reference-type="eqref" reference="norm_M-O"} and the proof of Theorem [\[Th1.1\]](#Th1.1){reference-type="ref" reference="Th1.1"} for local oscillation operators is completed. ## Local maximal Riesz transform We firstly prove that \(R^\alpha_{*,a}\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\). In order to do so, we need to decompose, for every \(\epsilon>0\), the truncated integral \(R^\alpha_\epsilon\) into two parts, called local and global parts (see ). For every \(\tau>0\), we consider the sets \[L_\tau=\left\{(x,y,s)\in (0,\infty)\times (0,\infty)\times (-1,1): \sqrt{q_-(x,y,s)}\leq \frac{a(1+\alpha)\tau}{1+x+y}\right\}\] and \[G_\tau=((0,\infty)\times (0,\infty)\times (-1,1) )\setminus L_\tau.\] Here, and in the sequel, we denote \(q_{\pm}(x,y,s)=x^2+y^2\pm 2xys\), for \(x,y\in (0,\infty)\) and \(s\in (-1,1)\). We choose a function \(\varphi\in C^\infty ((0,\infty)\times (0,\infty)\times (-1,1))\) such that \(0\leq \varphi\leq 1\), \[\varphi(x,y,s)=\begin{cases}1, & (x,y,s)\in L_1,\\ 0, & (x,y,s)\in G_2, \end{cases}\] and \[|\partial_x \varphi(x,y,s)|+|\partial_y \varphi(x,y,s)|\leq \frac{C}{\sqrt{q_-(x,y,s)}}, \quad x,y\in (0,\infty), s\in (-1,1).\] We define, for each \(\epsilon>0\) and \(x\in (0,\infty)\) \[\begin{aligned} R^{\alpha,\textup{loc}}_\epsilon (f)(x)&=\int_{|x-y|>\varepsilon,\ y\in (0,\infty)}R^{\alpha,\textup{loc}}(x,y)f(y)d\gamma_\alpha(y),\\ R^{\alpha,\textup{glob}}_\epsilon (f)(x)&=R^\alpha_\epsilon (f)(x)-R^{\alpha,\textup{loc}}_\epsilon (f)(x),\ \end{aligned}\] where \[R^{\alpha,\textup{loc}} (x,y)=\int_{-1}^1 R^{\alpha,\textup{loc}} (x,y,s) \Pi_\alpha(s)ds,\quad x,y\in (0,\infty)\] and, for \(x,y\in (0,\infty)\), \(s\in (-1,1)\), \[R^{\alpha,\textup{loc}} (x,y,s)=-\frac{2}{\sqrt{\pi}}\int_0^\infty \frac{e^{-t(\alpha+2)}\left(e^{-t}x-ys\right)}{(1-e^{-2t})^{\alpha+2}} e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2} \varphi(x,y,s) \frac{dt}{\sqrt{t}}.\] We also consider the maximal operators associated with the above, \[R^{\alpha,\textup{loc}}_* (f)=\sup_{\epsilon>0} \left|R^{\alpha,\textup{loc}}_\epsilon (f)\right|, \quad R^{\alpha,\textup{glob}}_* (f)=\sup_{\epsilon>0} \left|R^{\alpha,\textup{glob}}_\epsilon (f)\right|,\] which clearly verify \[R^\alpha_* (f)\leq R^{\alpha,\textup{loc}}_* (f)+R^{\alpha,\textup{glob}}_* (f).\] According to (see also ), we have that \[R^{\alpha,\textup{glob}}_* (f)(x)\leq C\int_0^\infty K^\alpha(x,y)f(y) d\gamma_\alpha(y), \quad x\in (0,\infty),\] where \[K^\alpha(x,y)=\int_{-1}^1 K^\alpha(x,y,s) \chi_{G_1}(x,y,s) \Pi_\alpha(s) ds, \quad x,y\in (0,\infty),\] and, for \(x,y\in (0,\infty)\) and \(s\in (-1,1)\), \[\label{Kalfa} K^\alpha(x,y,s)=\begin{cases} 1, & s<0, \\ \left(\frac{q_+(x,y,s)}{q_-(x,y,s)}\right)^{\frac{\alpha+1}{2}}\exp\left(\frac{x^2+y^2-\sqrt{q_-(x,y,s)q_+(x,y,s)}}{2}\right), & s\geq 0. \end{cases}\] It follows that \(R^{\alpha,\textup{glob}}_*\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\) (see ). We recall that the measure \(\mathfrak{m}_\alpha\) defined in Section [3.1](#subsec: variation){reference-type="ref" reference="subsec: variation"} has the doubling property on \((0,\infty)\). Therefore, by, \(e^{-y^2}R_{\alpha}^\textup{loc}(x,y)\), for \(x,y\in (0,\infty)\), is an \(\mathfrak{m}_\alpha\)-standard Calderón-Zygmund kernel, that is, for every \(x,y\in (0,\infty)\), \(x\neq y\), \[\left|e^{-y^2}R_{\alpha}^\textup{loc}(x,y)\right|\leq \frac{C}{\mathfrak{m}_\alpha(I(x,|x-y|))},\] and \[\left|\partial_x\left[e^{-y^2}R_{\alpha}^\textup{loc}(x,y)\right]\right|+\left|\partial_y\left[e^{-y^2}R_{\alpha}^\textup{loc}(x,y)\right]\right|\leq \frac{C}{|x-y|\mathfrak{m}_\alpha(I(x,|x-y|))}.\] If we define the operators \(R^{\alpha,\textup{loc}}\) and \(R^{\alpha,\textup{glob}}\) in the obvious way, we can see as above that the later is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\). Since \(R^\alpha\) is also bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\) (see ), we conclude that \(R^{\alpha,\textup{loc}}\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\). By proceeding as in, we deduce that \(R^{\alpha,\textup{loc}}\) is bounded on \(L^p((0,\infty),\mathfrak{m}_\alpha)\) for every \(1<p<\infty\). Moreover, since \(R^{\alpha,\textup{loc}}\) is an \(\mathfrak{m}_\alpha\)-Calderón-Zygmund operator, \(R^{\alpha,\textup{loc}}_*\) is bounded on \(L^p((0,\infty),\mathfrak{m}_\alpha)\) for every \(1<p<\infty\). By using again the arguments given in, we get that \(R^{\alpha,\textup{loc}}_*\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\). It follows now that \(R^\alpha_*\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \(1<p<\infty\). Particularly, using this property for \(p=2\), for any \(f\in L^\infty((0,\infty),\gamma_\alpha)\), \[\|R^\alpha_{*,a}\|_{L^1((0,\infty),\gamma_\alpha)}\leq C\|f\|_{L^\infty((0,\infty),\gamma_\alpha)}.\] We recall that, from [\[cota-nucleo-potencia\]](#cota-nucleo-potencia){reference-type="eqref" reference="cota-nucleo-potencia"}, \[|R^\alpha(x,y)|\leq C\frac{e^{\frac{x^2+y^2}{2}}}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y,\] and also we can see (as in ) that \[\sup_{I \in \mathcal{B}_a} \sup_{x\in I } r_0 \int_{(0,\infty)\setminus 2I} |\partial_x R^\alpha(x,y)| d\gamma_\alpha(y)<\infty.\] By proceeding as in the proof of, it yields \[\sup_{I \in \mathcal{B}_a}\left\|\mathcal{M}^\alpha_a\left(R^\alpha_{*,a}(f)\right)-R^\alpha_{*,a}(f)\right\|_{L^\infty((0,\infty),\gamma_\alpha)}\leq C\|f\|_{L^\infty((0,\infty),\gamma_\alpha)},\] meaning that \(R^\alpha_{*,a}\) is bounded from \(L^\infty((0,\infty),\gamma_\alpha)\) into \(\textup{BLO}_a((0,\infty),\gamma_\alpha)\). # Proof of Theorem [\[Th1.2\]](#Th1.2){reference-type="ref" reference="Th1.2"} In this section, we will study \(L^\infty((0,\infty),\gamma_\alpha)\)-\(\textup{BLO}_a((0,\infty),\gamma_\alpha)\) estimates for the \(a\)-local maximal operator \[\begin{aligned} Q^\alpha_{\phi,*,a}(f)(x)&=\sup_{0<\epsilon\le am(x)}|Q_{\phi,\epsilon}^\alpha(f)(x)|\\ &=\sup_{0<\epsilon\le am(x)}\left|\int_{|x-y|>\epsilon, \ y\in (0,\infty)}K_\phi^\alpha(x,y)f(y)d\gamma_\alpha(y)\right|, \end{aligned}\] for \(x\in (0,\infty)\) and \(a>0\). We recall that \[K_\phi^\alpha(x,y)=-\int_0^\infty \phi(t) \partial_t W_t^\alpha(x,y) dy, \quad x,t,\in (0,\infty),\ x\neq y,\] being \[W_t^\alpha(x,y)=\left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2}\Pi_\alpha(s) ds, \quad x,y,t\in (0,\infty).\] Firstly, we shall see that \(Q^\alpha_{\phi,*,a}\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \({1<p<\infty}\). We define, for \(x,y,t\in (0,\infty)\), \[W_t^{\alpha,\textup{loc}}(x,y)=\left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2}\varphi(x,y,s)\Pi_\alpha(s) ds,\] and \[W_t^{\alpha,\textup{glob}}(x,y)=W_t^\alpha(x,y)-W_t^{\alpha,\textup{loc}}(x,y).\] In terms of these, we consider \(K_\phi^{\alpha,\textup{loc}}\) and \(K_\phi^{\alpha,\textup{glob}}\) given as \(K_\phi^\alpha\) but with \(W_t^\alpha\) replaced by \(W_t^{\alpha,\textup{loc}}\) and \(W_t^{\alpha,\textup{glob}}\), respectively. Similarly, we define \(Q_{\phi,*,a}^{\alpha,\textup{loc}}\) and \(Q_{\phi,*,a}^{\alpha,\textup{glob}}\) by putting \(K_\phi^{\alpha,\textup{loc}}\) and \(K_\phi^{\alpha,\textup{glob}}\) instead of \(K_\phi^\alpha\), respectively. We will first deal with \(Q_{\phi,*,a}^{\alpha,\textup{glob}}\). Notice that, for \(x,y,t\in (0,\infty)\) and \(s\in(-1,1)\) \[\begin{aligned} \partial_t&\left[\left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \exp\left(-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}\right)\right]\\ &=P_{x,y,s}\left(e^{-t}\right) \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1}\exp\left(-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}\right), \end{aligned}\] where, for every \(x,y\in (0,\infty)\) and \(s\in (-1,1)\), \(P_{x,y,s}\) is a polynomial whose degree is at most four. Hence, \[\begin{aligned} |K_\phi^{\alpha,\textup{glob}}(x,y)|&\leq C \int_{-1}^1 \sup_{t>0} \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2}\chi_{L_1^c}(x,y,s)\Pi_\alpha(s) ds\\ &\leq C \int_{-1}^1 K^\alpha(x,y,s) \chi_{L_1^c}(x,y,s)\Pi_\alpha(s) ds, \quad x,y\in (0,\infty), \end{aligned}\] being \(K^\alpha(x,y,s)\) as in [\[Kalfa\]](#Kalfa){reference-type="eqref" reference="Kalfa"}, for \((x,y,s)\in L_1^c\). From, it follows that the operator whose kernel is the one on the right-hand side is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \({1<p<\infty}\), and so will be \(Q_{\phi,*,a}^{\alpha,\textup{glob}}\). Furthermore, for every \(f\in L^p((0,\infty),\gamma_\alpha)\), \(1<p<\infty\), \[\lim_{\varepsilon\to 0^+} \int_{|x-y|>\varepsilon,\ y\in (0,\infty)} K_\phi^{\alpha,\textup{glob}}(x,y) f(y) d\gamma_\alpha(y)=\int_0^\infty K_\phi^{\alpha,\textup{glob}}(x,y) f(y) d\gamma_\alpha(y),\] for a.e. \(x\in (0,\infty)\). We now consider the operators \[T_M^{\alpha,\textup{loc}}(f)(x)=\lim_{\varepsilon\to 0^+}\left(\Lambda(\varepsilon)f(x)+\int_{|x-y|>\varepsilon,\ y\in (0,\infty)}K_\phi^{\alpha,\textup{loc}}(x,y)f(y)d\gamma_\alpha(y)\right),\] and \[T_M^{\alpha,\textup{glob}}(f)(x)=\int_0^\infty K_\phi^{\alpha,\textup{glob}}(x,y) f(y) d\gamma_\alpha(y),\] for a.e. \(x\in (0,\infty)\). Since \(T_M^\alpha\) and \(T_M^{\alpha,\textup{glob}}\) are both bounded on \(L^2((0,\infty),\gamma_\alpha)\) (), also \(T_M^{\alpha,\textup{loc}}\) is bounded on \(L^2((0,\infty),\gamma_\alpha)\). Moreover, for every \(f\in L^\infty((0,\infty),\gamma_\alpha)\) \[T_M^{\alpha,\textup{loc}}(f)(x)=\int_0^\infty K_\phi^{\alpha,\textup{loc}}(x,y)f(y)d\gamma_\alpha(y),\quad x\notin \supp(f).\] Let us now consider \(\mathbb{K}_\phi^\alpha(x,y):=e^{-y^2} K_\phi^{\alpha,\textup{loc}}(x,y)\), for \(x,y\in (0,\infty)\). We have \[\begin{aligned} \mathbb{K}_\phi^\alpha(x,y)&=(\alpha+1)\int_0^\infty \varphi(t) \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}\varphi(x,y,s) \Pi_\alpha(s) ds dt\\ &\quad-\int_0^\infty \varphi(t) e^{-t(\alpha+1)} \int_{-1}^1 \partial_t \left[\frac{e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}}{(1-e^{-2t})^{\alpha+1}}\right]\varphi(x,y,s)\Pi_\alpha(s) ds dt\\ &:=\mathbb{K}_{\phi,1}^\alpha(x,y)+\mathbb{K}_{\phi,2}^\alpha(x,y), \quad x,y\in (0,\infty). \end{aligned}\] As in, we can prove that \[|\mathbb{K}_{\phi,2}^\alpha(x,y)|\leq \frac{C}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y,\] and \[|\partial_x\mathbb{K}_{\phi,2}^\alpha(x,y)|+|\partial_y\mathbb{K}_{\phi,2}^\alpha(x,y)|\leq \frac{C}{|x-y|\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y.\] On the other hand, using, i.e., \(q_-(e^{-t}x,y,s)\geq q_-(x,y,s)-2(1-e^{-2t})\), for every \((x,y,s)\in N_1\), and the estimates obtained in, we get 1. \[\begin{aligned} |\mathbb{K}_{\phi,1}^\alpha(x,y)|&\leq C \int_0^\infty |\varphi(t)| \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(x,y,s)}{1-e^{-2t}}} \Pi_\alpha(s) dsdt\\ & \leq C \int_0^\infty |\varphi(t)| e^{-t(\alpha+1)} dt\int_{-1}^1 \frac{\Pi_\alpha(s)}{q_-(x,y,s)^{\alpha+1}} ds\\ &\leq \frac{C}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y; \end{aligned}\] 2. by, \[\begin{aligned} |\partial_x \mathbb{K}_{\phi,1}^\alpha(x,y)|&\leq C \int_0^\infty |\varphi(t)| \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}\\ &\qquad \times \left[\frac{e^{-t}|e^{-t}x-ys|}{1-e^{-2t}}\varphi(x,y,s)+|\partial_x \varphi(x,y,s)|\right] \Pi_\alpha(s) dsdt\\ &\leq C \int_0^\infty |\varphi(t)| \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} e^{-t}\int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}\\ &\qquad \times \left[\frac{\sqrt{q_-(e^{-t}x,y,s)}}{1-e^{-2t}}+\frac{1}{\sqrt{q_-(x,y,s)}}\right] \Pi_\alpha(s) dsdt\\ &\leq C \int_0^\infty |\varphi(t)| e^{-t(\alpha+2)} \int_{-1}^1 \left[\frac{e^{-\frac{q_-(x,y,s)}{2(1-e^{-2t})}}}{(1-e^{-2t})^{\alpha+3/2}}\right.\\ &\qquad +\left.\frac{e^{-\frac{q_-(x,y,s)}{1-e^{-2t}}}}{(1-e^{-2t})^{\alpha+1} \sqrt{q_-(x,y,s)}}\right]\Pi_\alpha(s) dsdt\\ &\leq C \int_0^\infty |\varphi(t)| e^{-t(\alpha+2)} dt\int_{-1}^1 \frac{\Pi_\alpha(s)}{q_-(x,y,s)^{\alpha+3/2}} ds\\ &\leq \frac{C}{|x-y|\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y; \end{aligned}\] 3. by (with \(x\) and \(y\) interchanged), and proceeding like before, \[\begin{aligned} |\partial_y \mathbb{K}_{\phi,1}^\alpha(x,y)|&\leq C \int_0^\infty |\varphi(t)| \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}\\ &\qquad \times \left[\frac{|y-e^{-t}xs|}{1-e^{-2t}}\varphi(x,y,s)+|\partial_y \varphi(x,y,s)|\right] \Pi_\alpha(s) ds dt\\ &\leq C \int_0^\infty |\varphi(t)| \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1} \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}}\\ &\qquad \times \left[\frac{\sqrt{q_-(e^{-t}x,y,s)}}{1-e^{-2t}}+\frac{1}{\sqrt{q_-(x,y,s)}}\right] \Pi_\alpha(s) ds dt\\ &\leq C \int_0^\infty |\varphi(t)| e^{-t(\alpha+1)} dt\int_{-1}^1 \frac{\Pi_\alpha(s)}{q_-(x,y,s)^{\alpha+3/2}} ds \\ &\leq \frac{C}{|x-y|\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y. \end{aligned}\] All of the above proves that \(T_M^{\alpha,\textup{loc}}\) is an \(\mathfrak{m}_\alpha\)-Calderón-Zygmund operator. Therefore, \(T_M^{\alpha,\textup{loc}}\) is bounded on \(L^p((0,\infty),\mathfrak{m}_\alpha)\) for every \(1<p<\infty\), which yields \(Q_{\phi,*,a}^{\alpha,\textup{loc}}\) is also bounded on \(L^p((0,\infty),\mathfrak{m}_\alpha)\) for every \({1<p<\infty}\). The arguments in allow us to deduce that \(Q_{\phi,*,a}^{\alpha,\textup{loc}}\) is also bounded on \(L^p((0,\infty),\gamma_\alpha)\) for every \({1<p<\infty}\). Finally, we conclude that \(Q^\alpha_{\phi,*,a}\) is bounded on \(L^p((0,\infty),\gamma_\alpha)\) for any \(1<p<\infty\). We have proved above that \[|K_\phi^{\alpha,\textup{loc}}(x,y)|\leq C \frac{e^{y^2}}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty),\ x\neq y.\] We also saw that \[|K_\phi^{\alpha,\textup{glob}}(x,y)|\leq C \int_{-1}^1 K^\alpha(x,y,s) \chi_{L_1^c}(x,y,s)\Pi_\alpha(s) ds, \quad x,y\in (0,\infty),\] where \(K^\alpha(x,y,s)\) was defined in [\[Kalfa\]](#Kalfa){reference-type="eqref" reference="Kalfa"}. It is easy to see that, for any \((x,y,s)\in L_1^c\), \[|K^\alpha(x,y,s)|\leq C\begin{cases} 1, & s\in (-1,0), \\ \frac{\exp\left(\frac{x^2+y^2}{2}\right)}{q_-(x,y,s)^{\alpha+1}}, & s\in [0,1). \end{cases}\] Moreover, for any fixed constant \(c>0\), if \(x,y\in (0,\infty)\) with \(|x-y|\leq cam(x)\) and \(s\in (-1,1)\), \[\begin{aligned} q_-(x,y,s)&=(x-y)^2+2xy(1-s)\leq (x-y)^2+4y(|x-y|+y)\\ &\leq 5(x-y)^2+4y^2+4|x-y|x\leq C(1+y^2), \end{aligned}\] which yields \[|K^\alpha(x,y,s)|\leq \frac{C}{q_-(x,y,s)^{\alpha+1}}\begin{cases} (1+y^2)^{\alpha+1}, & s\in (-1,0), \\ \exp\left(\frac{x^2+y^2}{2}\right), & s\in [0,1), \end{cases}\] for any \((x,y,s)\in L_1^c\) with \(|x-y|\leq cam(x)\). According to, we obtain \[|K_\phi^{\alpha,\textup{glob}}(x,y)|\leq C \frac{e^{\frac{x^2+y^2}{2}}}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty), 0<|x-y|\leq cam(x).\] Hence, we conclude that \[|K_\phi^\alpha(x,y)|\leq C \frac{e^{\frac{x^2+y^2}{2}}}{\mathfrak{m}_\alpha(I(x,|x-y|))}, \quad x,y\in (0,\infty), 0<|x-y|\leq cam(x).\] We are going to prove now that \[\sup_{I \in \mathcal{B}_1} \sup_{x\in I } \int_{(0,\infty)\setminus 2I} |\partial_x K_\phi^\alpha(x,y)|d\gamma_\alpha(y) <\infty.\] By partial integration, we have that \[K_\phi^\alpha(x,y)=\int_0^\infty \phi'(t) W_t^\alpha(x,y) dt, \quad x,y,\in (0,\infty),\] and thus, \[\begin{aligned} \partial_x K_\phi^\alpha(x,y) & =\int_0^\infty \phi'(t) \partial_x W_t^\alpha(x,y) dt\\ &=-2\int_0^\infty \phi'(t) \left(\frac{e^{-t}}{1-e^{-2t}}\right)^{\alpha+1}\\ &\quad \times \int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2} \frac{e^{-t}(e^{-t}x-ys)}{1-e^{-2t}} \Pi_\alpha(s) ds dt, \quad x,y\in (0,\infty). \end{aligned}\] By, \(|e^{-t}x-ys|\leq \sqrt{q_-(e^{-t}x,y,s)}\) for every \(x,y\in (0,\infty)\) and \(s\in (-1,1)\). Then, using the hypothesis on \(\phi'\), \[\begin{aligned} |\partial_x K_\phi^\alpha(x,y)|&\leq C \int_0^\infty |\phi'(t)| \frac{e^{-t(\alpha+2)}}{(1-e^{-2t})^{\alpha+1}} \\ &\quad \times\int_{-1}^1 e^{-\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2} \sqrt{q_-(e^{-t}x,y,s)}\Pi_\alpha(s) ds dt\\ &\leq C\int_0^\infty \frac1t \frac{e^{-t(\alpha+2)}}{(1-e^{-2t})^{\alpha+3/2}}\int_{-1}^1 e^{-c\frac{q_-(e^{-t}x,y,s)}{1-e^{-2t}}+y^2}\Pi_\alpha(s) ds dt \end{aligned}\] for any \(x,y\in (0,\infty)\). Therefore, by, there exists \(C>0\) such that \[\sup_{x\in I } \int_{(0,\infty)\setminus 2I} |\partial_x K_\phi^\alpha(x,y)|d\gamma_\alpha(y)\leq C\] for every \(I \in \mathcal{B}_1\), as claimed.

{'timestamp': '2023-02-10T02:02:44', 'yymm': '2302', 'arxiv_id': '2302.04356', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04356'}

# Introduction Human social interactions are based on a complex exchange of a variety of signals, including facial expressions, gaze, gestures, and posture. Gaze plays a special part as it serves to perceive objects or other humans and at the same time it communicates to others where one focuses the attention. For instance, our gaze helps us indicate social interest, understand other people's mental and emotional state, and see what they are attending to. This process of using someone else's eye movement as information of what they are attending to and shifting one's own attention accordingly is called gaze cueing and is discussed as a prerequisite for joint attention, the case in which both persons visually attend the same object.\ Social gaze has been mostly studied in human and non-human primates, and with the emergence of social robots has become introduced into robotic systems and has become a growing branch of research ever since. A great amount of research on gaze cueing in human-robot interaction (HRI) uses virtual agents or pictures of robots instead of physically copresent robots. From other work on human-robot interaction, we know however, that the physical presence of a robot fundamentally shapes the way we perceive and interact with it. What we do not know, however, is how the physical presence of an agent and gaze cueing relate. This paper thus explores whether the physical presence of an agent affects the strength of the gaze cueing effect in human-robot interaction.\ # Related Work The common paradigm used to study gaze cueing is a variation of the Posner paradigm, in which participants are asked to localize a target stimulus while that stimulus is consistently and inconsistently cued by a facial stimulus. In several studies using schematic and human faces as stimuli, a gaze cueing effect (GCE) has been observed, evidenced by faster reaction times in responding to congruent (gazed-at) target stimuli compared to incongruent (non-gazed-at) target stimuli. In contrast to these common findings, studies using robotic stimuli show mixed results. While Admoni and colleagues find a gaze cueing effect with images of human, schematic, and no GCE for the robot faces in one study, Wiese and colleagues report a more pronounced GCE for robotic face stimuli compared to human face stimuli in a study with individuals with autism spectrum disorder (ASD). Notably, most studies on gaze cueing in HRI used only static images of gazing robots. For instance, in a study investigating the effect of human-likeness of a robot on the strength of the gaze cueing effect, Martini et al. used morphed images of human and robot faces as stimuli . To date, however, it is still unclear to what extent these results can be generalized to copresent robots in real human-robot interaction. Physical presence appears to be an important factor shaping our perception of and interaction with robots. For example, in an early study conducted by Lee et al., in which participants were either introduced to a physical Aibo robot dog, or its virtual equivalent, a positive effect of embodiment on ratings of the interaction with the robot and the robot's social presence was found. These results indicate the importance of physical embodiment in HRI, even if it is not necessary to complete the interaction successfully. Wiese and colleagues employed a gaze cueing paradigm using a copresent humanoid robot. Participants were instructed to indicate the appearance of a light stimulus on their left or right side, which was cued by a Meka robot's gaze shift. Even though the researchers informed the participants that the robot's gaze is uninformative of the appearance of the light stimulus, participants seemed to follow the robot's gaze, as a congruency effect could still be found. Using a similar setup, Kompatsiari et al. were able to replicate these results with an iCub robot when the robot established mutual gaze with the observer before turning to the target stimuli, as in the previously reported experiment, but not when no mutual gaze was established. Further controlled investigations will be needed, in which participants are faced with scenarios and robots that are constant except for the way they are presented, to better grasp the role of presence on gaze cueing effects in HRI. To our knowledge, there is only one systematic study conducted to date that explores the relation of embodiment, presence and facial cueing. Mollahosseini et al.  investigated the effect of robot embodiment and presence on interaction tasks that involved typical measures of communication, and found that these factors affect recognition of facial expressions, and especially eye gaze. While further analysis of the data revealed a significant effect of embodiment, unlike other studies examining social interaction in HRI, no main effect of presence was found. This could be due to the nature of the task, which involved only still representations of the gaze rather than actual movements, or to the fact that it generally did not rely on social attributions that could relate to the presence of the robot, but on purely geometric cues. # Methods and Material We chose a mixed experimental design with two independent variables: (1) type of robot presence (between-subject) with two levels: a physical robot and a virtual version of the same robot as depicted in Fig. [\[figure2\]](#figure2){reference-type="ref" reference="figure2"} (2) gaze cue congruency (within-subject) with two levels: congruent, and incongruent as depicted in Fig. [\[figure3\]](#figure3){reference-type="ref" reference="figure3"}. Participants were randomly assigned to one type of robot and observed it gazeing at one of two light stimuli situated on either side of the table. The robot shifted its gaze in 80 trials, 40 congruent trials in which the robot looked at the lamp that was to change color and 40 incongruent trials in which the robot looked at the opposite lamp (see Fig. [\[figure3\]](#figure3){reference-type="ref" reference="figure3"}). In half of the conditions the robot looked at the left side and in the other half at the right side. All conditions were pseudo-randomly shuffled. Most of the features detailed in this section are also illustrated in this video <https://youtu.be/n_rU9XNE-bI>. ## Participants A total of 42 participants were recruited via posters in the university building, Facebook advertisement, and email. Three participants were excluded due to technical problems with the setup. The final sample consisted of 42 participants (16 females; mean age = 29.98). All participants provided written informed consent in line with the ethical approval of the study granted by the Committee for Research Ethics at the Czech Technical University in Prague. When asked about their experience with robots on a scale from 1 (very poor) to 5 (very good), the mean score was 2.33 (SD = 1.2) and only one participant answered 5. Data was stored and analysed anonymously. Testing time was about fifteen minutes. ## Measures Similar to previous studies on gaze cueing, the influence of gaze cueing on participants' gaze following behavior was determined by measures of mean correct reaction time. A response was considered incorrect if it was made with the wrong key press, and considered correct if the correct key was pressed. Responses that were given in a response time that was more than 2.5 standard deviations away from the individual mean response time of a participant were excluded from further analyses. To further check for possible effects of a robot's physical presence on gaze following behavior, average correct response times (RTs) were calculated for each participant and each experimental condition. ## Experimental Set-Up and Procedure The present experiment was designed to examine both self-reported and behavioral effects of a robot's presence and gaze cues in a human-robot interaction task. During the experiment, participants were instructed to indicate the appearance of a target stimulus that was either congruent or incongruent with the position being gazed at by an iCub robot that was either physically present in the same room with the participants (copresence condition) or presented via a monitor (virtual agent condition) by pressing a corresponding key on a keyboard. After completion of the task, participants were asked to indicate the way they perceived the robot by completing the Czech translation of the three subscales "Anthropomorphism", "Animacy" and "Likeability" of the Godspeed series [^1]. The experiment was conducted in April 2022 in the laboratories of the Department of Cybernetics of the Czech Technical University in Prague. At the beginning of the experiment, participants received written and oral instructions and gave informed consent. They were informed that the task was to respond as fast as possible to the color change of one of two light stimuli. Responses had to be given by pressing the appropriate key on a keyboard. Participants were also informed that iCub might move randomly during the time of the experiment. Reaction times were measured as a dependent variable. After receiving the instructions participants had the opportunity to ask questions about the task. Each trial began with iCub making eye contact by looking straight ahead in the direction of the observer. After 250 ms, iCub shifted his gaze either toward the lamp that was on his left side or toward the lamp that was on his right side. Subsequently, after 200 ms, one of the two lamps changed color, either on the corresponding side of the gaze cue or on the non-corresponding side of the gaze cue. When the target stimulus was presented, participants responded as quickly and as accurately as possible to the position of the target stimulus by pressing the \"x\" or \"m\" key on a standard keyboard. The target stimulus remained unchanged until a response was made or a time-out criterion (5000 ms) was reached. Then the light was turned off again and iCub looked straight ahead again to signal readiness to begin the next trial. Figure [\[figure3\]](#figure3){reference-type="ref" reference="figure3"} shows an exemplary trial sequence. ## Hardware and Software The iCub is a small humanoid robot that resembles a 4-year old child. It is one meter tall and has 53 Degrees of Freedom (DoF). Most relevant to gaze, it has 3 DoF in the neck and 3 coupled DoF for the two eyes (tilt, version, and vergence) in an anthropomorphic arrangement. For the virtual iCub we used the freely available simulator. The iCub gaze controller is used to command where the robot should look. Our custom made C++ program that controls the movement of the robot using the YARP middleware and records the reaction times in the experiment is publicly available[^2]. For our experiment we designed custom lamps which consisted of mate covered red led lamps controlled via Arduino Nano. ## Analysis Statistical analyses were conducted using R (version 4.1.2). The average correct RTs for congruent and incongruent trials were compared for each robot presence condition individually to check for consistency with previous studies regarding the strength of the gaze cueing effect. To test for the effect of cue-target congruence, for the virtual robot condition a t-test was calculated comparing the mean reactions times of congruent trials and incongruent trials. For the copresent robot condition a Wilcoxon test was calculated instead, to account for non-normally distributed data. A gaze cueing effect is evidenced by significant differences in reaction times of congruent and incongruent trials. Moreover, a mixed model analysis of variance (ANOVA) with a between-subject factor of robot presence (2: copresence vs. virtual presence), and within-subject factors of cue-target congruency (2: congruent vs. incongruent) was calculated. This analysis was used to assess the individual effect of cue-target congruency and the effect of presence specificity of gaze cueing. Robot presence specific gaze-cueing effects would be evidenced by a significant interaction between presence condition and cue-target congruency (over and above a main effect of congruency). By contrast, presence nonspecific gaze cueing would manifest in terms of a main effect of cue-target congruency (not accompanied by a presence \(\times\) congruency interaction), with equal facilitation for all robot presence conditions of interest. # Results ## The Effect of Presence and Gaze Mean reaction times were subjected to a two-way analysis of variance with two levels of robot presence (copresent robot, virtual agent) and two levels of cue-target congruence (congruent, incongruent) to test the effect of robot physical presence and gaze cueing behavior on participants' reaction times in a localization task (see Table [2](#table1){reference-type="ref" reference="table1"}). It is important to note that the data was not normally distributed in each group. An ANOVA was conducted either way, as F-Tests have been reported to be relatively robust to violations of normality when homogeneity of variance is given. \ *Note. \*\*\* p\(<\).001, \*\* p\(<\).01, \* p\(<\).05* [\[table1\]]{#table1 label="table1"} The main effect of gaze congruency yielded an F-ratio of F(1, 40) = 18.71, p \(<\) .001, indicating that mean reaction times differed significantly between congruent and incongruent trials, with faster reaction times on congruent cue-target trials (M = 288 ms, SD = 37.2 ms) than on incongruent cue-target trials (M = 298 ms, SD = 43.2 ms). To illustrate the size of the GCE by robot presence group, individual analyses were performed for both robot conditions as displayed in Figure [\[figure4\]](#figure4){reference-type="ref" reference="figure4"}. For the comparison of mean reaction times in congruent and incongruent trials in the copresent robot group a Wilcoxon signed-rank test was calculated to account for non-normally distributed data as indicated by a significant Shapiro-Wilk test (W = .9, p = .009). On average, participants in the copresent robot condition responded faster on congruent trials (M = 291 ms) than on incongruent trials (M = 305 ms). Results of the Wilcoxon signed-rank test showed that this difference was statistically significant (p = .001), with a large effect size, r = 0.7. An additional t-test conducted on the mean reaction times of participants in the virtual agent group revealed a significant difference between congruent and incongruent trials (t (20) =-2, p =.03) with shorter reaction times for congruent (M = 284 ms) than incongruent trials (M = 291 ms). The effect size was at a moderate level, r = 0.5. While participants in the virtual agent group (M = 287 ms) on average reacted faster than participants in the copresent group (M = 298 ms), the main effect of robot presence on participants' mean reaction times was non-significant F(1, 40) = 0.64, p \(>\).05. Yet, it is worth noting that the variances of the reaction times of the two robot conditions were significantly different, p =.012. Moreover, no significant interaction effect of gaze congruency and presence could be found, F(1, 40) = 2.22, p \(>\).05. ## Analysis of Godspeed Indices To further test how robot presence influences the way participants perceive the robot, responses of the participants to the Godspeed subscales Anthropomorphism, Animacy and Likeability were taken into account. A standard t-test was used to examine the influence of robot presence on animacy ratings. Significant results in a Shapiro-Wilk test with mean ratings of anthropomorphism and likability as outcome variables indicated non-normally distributed data, so an unpaired two-samples Wilcoxon test was computed to test the influence of robot presence on perceived anthropomorphism and likability. Mean ratings for all three Godspeed subscales by robot presence condition can be found in Figure [\[figure5\]](#figure5){reference-type="ref" reference="figure5"}. ### Anthropomorphism For ratings of the robot's perceived anthropomorphism, participants in the virtual agent condition assessed the robot's anthropomorphism slightly higher (M = 12.29, SD = 2.83) than people in the copresent robot condition (M = 12.05, SD = 4.03). Results of the independent samples Wilcoxon test, however, were not significant; W = 208.5, p =.77, r =-0.05. A linear model including sex as an additional predictor was tested due to the significant correlation of anthropomorphism ratings and sex. The model revealed no significant difference between presence groups, whereas ratings between sexes significantly differed (p \(<\).05), with females (M = 13.56, SD = 3.44) rating the robot as more anthropomorphic on average than males (M = 11.31, SD = 3.21). ### Animacy On average, participants in the copresent robot condition rated the robot's perceived animacy higher (M = 15.19, SD= 4.50) than participants in the virtual agent group (M = 13.95, SD = 3.67). This difference was not significant; t(38.43) = 0.98, p =.33. ### Likeability Participants in the copresent robot condition assessed the robot's likeability slightly higher (M = 20.19, SD = 4.25) than people in the copresent robot condition (M = 18.10, SD = 3.94). Results of the independent samples Wilcoxon test indicate that this difference was not significant; W = 289, p =.09, r = 0.26. A linear model including sex as an additional predictor was tested due to the significant correlation of likeability ratings and sex. The model revealed no significant difference between presence groups, whereas ratings between sexes significantly differed (p \(<\).01). On average, females (M = 21.23, SD = 3.41) rated the robot as more likeable than males (M = 17.84, SD = 4.13). # Discussion Our results are consistent with previous research on gaze following with copresent robots: Participants consistently exhibited gaze-following behavior, as evidenced by slower reaction times on trials in which the robot cued the wrong target compared to trials in which the cued location and target location matched. Hence, our results replicate the well-known finding that participants locate a target that is congruent with the cued direction more quickly than a target that is incongruent with the cued direction. In particular, our results are in line with findings by Wiese and by Kompatsiari showing that a gaze cueing effect can be found when the target stimulus is predicted by the gaze of an embodied robot. This finding is particularly valuable given the ongoing replication crisis in psychological research, which has highlighted the problem of replicating the results of many scientific studies and allows for the generalization of the gaze cueing effect across different robotic platforms. Crucially, with an effect of 14 ms, the gaze cueing effect in our study is smaller than the 25 ms found by but in line with findings of other studies using more controlled settings. Differences in the extent of the effect might be explained by the design of the robot---possibly, the Meka robot used in offers more social cues or other affordances than the iCub robot used in the present study or other robotic platforms. Further studies will be needed to better understand the relationship of robot design and gaze following behavior. Moreover, our results show that even the virtual version of our robot consistently triggered a gaze cueing effect as indicated by slower reaction times in incongruent compared to congruent trials, showing that gaze following in HRI is not limited to physically present versions of embodied robots. Importantly, however, the same stimuli did not elicit varying degrees of gaze-cueing when comparing the different ways the robot was presented to the participants, as evidenced by a non-significant interaction of robot presence and cue-target congruency. Across conditions participants consistently followed the gaze, independently of whether they were confronted with a copresent iCub or a virtual iCub. As the novelty of this study lies in comparing the way the different ways of presenting the robot influence simple social attention mechanisms, as evidenced by the gaze cueing effect, there exists hardly any literature indicating similar or contradictory results. However, these results add to findings of Mollahosseini who were comparing the effect of embodiment and presence of four different types of agents on similar outcome variables. Results showed that embodiment but not physical presence was the factor that accounts for the significant difference in the participants' response, as indicated by no significant difference in results when presenting the participants with a copresent robot compared to a telepresent robot. However, the outcomes differed significantly for the comparisons of virtual agents and both forms of embodied robots. In contrast, our study found no significant difference in gaze following behavior between the presence conditions, as participants' reaction times were not significantly different when interacting with a virtual agent compared to a copresent robot. There are multiple possible reasons why physical presence might not additionally influence the gaze cueing effect. One could be that our results are based on the fact that robots in both conditions moved in a similar, human-like manner. Previous research has shown that (natural) movement is linked to mind attributions---famously for example in the Heider and Simmel illusion , in which participants attribute mental states to three moving geometrical figures (two triangles that seem to "hunt" a circle). Studies examining how mental state attributions alter gaze following behavior in trials with photographs of robots have shown that this manipulation led to the occurrence of a gaze cueing effect that was not otherwise present. Differences between our results and those of studies using only photos of agents or robots or non-moving agents, as in the research of Mollahosseini, might be due to the association of motion and mind attributions leading to typical social interaction phenomena, such as gaze cueing. Moreover, in contrast to previous findings, the copresent robot did not generate more positive ratings than the virtual agent, as indicated by participants' ratings of the robot's anthropomorphism, animacy, and likability. Similar results were reported in a study, in which the ratings of four types of agents that differed in terms of their embodiment and physical presence were compared after a simple conversational interaction. Notably, neither the interaction reported by nor the interaction reported in our study involved physical touch or a particular focus on spatial relations. The differences between the results of this and other studies could be explained by the advantages of the physical presence of a robot in more complex interaction scenarios. # Conclusion Social gaze is a fundamental part of human interactions and becomes more relevant in the scope of social human-robot interaction. Gaze cueing, the event in which we observe our interaction partner's gaze and shift our own attention accordingly, is broadly studied using images or virtual versions of robots. As previous research has pointed to the broad range of effects physical presence has in human-robot interaction, the question arises whether it affects the strength of the gaze cueing effect and hence, whether results from studies using images or virtual agents can be generalized to copresent robots. We designed a study to investigate the relationship of physical presence and social gaze by adapting a gaze cueing paradigm with two types of agents (1) a copresent robot and (2) a virtual version of the same robot. The results of our study indicate that gaze cueing is a stable effect in basal human-robot interaction across different robot presence conditions. Thereby, we add to the understanding of possibilities to generalize results from studies using virtual agents and pictorial stimuli to real life human-robot interaction. # Introduction Human social interactions are based on the complex exchange of a variety of social signals, including pointing, gestures, and posture. Of these social signals, gaze in particular appears to play a crucial role as it has a specific dual function in perceiving information from and signaling information to others. For instance, our gaze helps us indicate social interest, understand other people's mental and emotional state, and see what they are attending to. This process of using someone else's eye movement as information of what they are attending and shifting one's own attention accordingly is called gaze cueing and is discussed as a prerequisite for joint attention, the case in which both persons visually attend the same object.\ Social gaze has been mostly studied in human and non-human primates, and with the emergence of social robots become introduced into into robotic systems-and has become a growing branch of research ever since. A great amount of research on gaze cueing in human-robot interaction (HRI) uses virtual agents or pictures of robots instead of physically copresent robots. \[For instance, in a study investigating the effect of human-likeness of a robot on the strength of the gaze cueing effect, Martini et al. (2015) used morphed images of human and robot faces as stimuli.\] From other work on human-robot interaction, we know however, that the physical presence of a robot fundamentally shapes the way we perceive and interact with it. What we do not know, however, is how the physical presence of an agent and gaze cueing relate. This paper thus explores whether the physical presence of an agent affect the strength of the gaze cueing effect in social human-robot interaction.\ # Related Work ## Gaze Cueing in Human-Human Interaction. The common paradigm used to study gaze cueing is a variation of the Posner paradigm, in which participants are asked to localize a target stimulus while that stimulus is inconsistently cued by a facial stimulus. In several studies using schematic and human faces as stimuli, a gaze cueing effect (GCE) has been observed, evidenced by faster reaction times in responding to congruent stimuli compared to incongruent stimuli-. ## Gaze Cueing in Human-Robot Interaction In contrast to these common findings, studies using robotic stimuli show mixed results. While Admoni and colleagues (2011) find a gaze cueing effect with images of human, schematic, and no GCE for the robot faces in one study, Wiese and colleagues (2014) report in a study with individuals with autism spectrum disorder (ASD) even a more pronounced GCE for robotic face stimuli compared to human face stimuli. Importantly, most studies on gaze cueing in HRI used only static images of gazing robots, therefore it is still unclear to what extent these results can be generalized to copresent robots in real human-robot interaction. ## The Role of Physical Presence This is particularly important in light of previous research showing that physical presence appears to be an important factor shaping our perception of and interaction with robots. For example in an early study conducted by Lee et al. (2006), in which participants were either introduced to a physical Aibo robot dog, or its screen equivalent, a positive effect of embodiment on ratings of the interaction with the robot and the robot's social presence could be found. After a first introduction, participants in both experimental groups were instructed to try out the robot's various sensors, which functioned the same in both versions, and to interact freely with the robot for 10 minutes. Ratings of the pleasantness of the interaction and the social presence of the robot were positively correlated with physical embodiment, indicating the importance of physical embodiment in HRI, even if it is not necessary to complete the interaction successfully. Wiese and colleagues employed a gaze cueing paradigm using an embodied humanoid robot. Participants were instructed to indicate the appearance of a light stimulus on their left or right side, which was cued by a Meka robot's gaze shift (Figure [\[figure1\]](#figure1){reference-type="ref" reference="figure1"}). Even though the researchers informed the participants that the robot's gaze is uninformative of the appearance of the light stimulus, participants seemed to follow the robot's gaze, as a congruency effect could still be found. Using a similar setup, were able to replicate these results with an iCub robot when the robot established mutual gaze with the observer before turning to the target stimuli-as in the previously reported experiment ()-but not when no mutual gaze was established. Importantly, further controlled investigations will be needed, in which participants are faced with scenarios and robots that are constant except for the way they are presented, to better grasp the role of presence on gaze cueing effects in HRI. To our knowledge, there is only one systematic study conducted to date that explores the relation of embodiment, presence and facial cueing. Results of a study investigating the effect of robot embodiment and presence on interaction tasks that involved typical measures of communication-visual language, facial expressions and gaze, indicate that while visual speech does not seem to be influenced by the way the interaction partner is embodied and present, recognition of facial expressions, and especially eye gaze seem to be related to these factors. While further analysis of the data revealed a significant effect of embodiment, unlike other studies examining social interaction in HRI, no main effect of presence was found. This could be due to the nature of the task, which involved only still representations of gaze rather than actual movements, or the fact that it generally did not rely on social attributions that could relate to the presence of the robot, but on purely geometric cues. # Methods and Material We chose a mixed experimental design with two independent variables: (1) type of robot presence (between-subject) with two levels: a physical robot, a simulated version of the same robot (Figure [\[figure2\]](#figure2){reference-type="ref" reference="figure2"}) (2) gaze cue congruency (within-subject) with two levels: congruent, and incongruent (Fig. 3). Participants were randomly assigned to one type of robot and observed it gazeing at one of two light stimuli situated on either side of the table. The robot shifted its gaze in 80 trials, 40 congruent trials in which the robot looked at the lamp that was to change color and 40 incongruent trials in which the robot looked at the opposite lamp. In half of the conditions the robot looked at the left side and in the other half at the right side. All conditions were pseudo-randomly shuffled. ## Participants A total of 42 participants were recruited via posters in the university building, Facebook advertisement, and email. Three participants were excluded and retested due to technical problems with the set up. The final sample consisted of 42 participants (16 females; mean age = 29.98). All participants provided written informed consent in line with the ethical approval of the study granted by the Committee for Research Ethics at the Czech Technical University in Prague. When asked about their experience with robots on a scale from 1 (very poor) to 5 (very good), the mean score was 2.33 (SD = 1.2) and only one participant answered 5. Data was stored and analysed anonymously. Testing time was about fifteen minutes. ## Measures Similar to previous studies on gaze cueing (; ), the influence of gaze cueing on participants' gaze following behavior was determined by measures of mean correct reaction time. A response was considered incorrect if it was made with the wrong key press, and considered correct if the correct key was pressed. Responses that were given in a response time that was more than 2.5 standard deviations away from the individual mean response time of a participant were excluded from further analyses. To further check for possible effects of a robot's physical presence on gaze following behavior, average correct response times (RTs) were calculated for each participant and each experimental condition. ## Experimental Set-Up and Procedure The present experiment was designed to examine both self-reported and behavioral effects of a robot's presence and gaze cues in a human-robot interaction task. During the experiment, participants were instructed to indicate the appearance of a target stimulus that was either congruent or incongruent with the position being gazed at by an iCub robot that was either physically present in the same room with the participants (head presence condition) or presented via a monitor (telepresence condition) by pressing a corresponding key on a keyboard. After completion of the task, participants were asked to indicate the way they perceived the robot by completing the Czech translation of the three subscales \"Anthropomorphism\", \"Animacy\" and \"Likeability\" of the Godspeed series (Bartneck et al., 2009). The experiment was conducted in April 2022 in the laboratories of the Department of Cybernetics of the Czech Technical University in Prague. The study was approved by the Committee for Research Ethics at the Czech Technical University in Prague under the reference number 00000-07/21/51903/EKCVUT. At the beginning of the experiment, participants received written and oral instructions, and gave informed consent. They were informed that the task was to respond as fast as possible to the color change of one of two light stimuli. Responses had to be given by pressing the appropriate key on a keyboard. Participants were also informed that iCub might move randomly during the time of the experiment. Reaction times were measured as dependent variable. After receiving the instructions participants had the opportunity to ask questions about the task. The experiment consisted of 80 trials, 40 congruent trials in which the robot looked at the lamp that was to change color and 40 incongruent trials in which the robot looked at the opposite lamp. In half of the conditions the robot looked at the left side and in the other half at the right side. All conditions were pseudo-randomly shuffled. Each trial began with iCub making eye contact by looking straight ahead in the direction of the observer. After 250 ms, iCub shifted his gaze either toward the lamp that was on his left side or toward the lamp that was on his right side. Subsequently, after 200 ms, one of the two lamps changed color, either on the corresponding side of the gaze cue or on the non-corresponding side of the gaze cue. When the target stimulus was presented, participants responded as quickly and as accurately as possible to the position of the target stimulus by pressing the \"x\" or \"m\" key on a standard keyboard. The target stimulus remained unchanged until a response was made or a time-out criterion (5000 ms) was reached. Then the light was turned off again and iCub looked straight ahead again to signal readiness to begin the next trial. Figure [\[figure3\]](#figure3){reference-type="ref" reference="figure3"} shows an exemplary trial sequence. ## Hardware and Software The iCub is a small humanoid robot that resembles a 3-year old child. It is one meter tall and has 53 degrees of freedom (DoF) and is considered one of the most detailed humanoids within the field of cognitive robotics. It can move his head and eyes and is endowed with a sophisticated programming for natural gaze behavior. For the virtual iCub we used the freely available simulator. Our custom made C++ program that controls the movement of the robot using the YARP platform and records the reaction times in the experiment is publicly available[^3]. Basically the same program is used for both physical and simulated robots with the small changes in some constants (speed, acceleration and ). For our experiment we designed custom lamps which consisted of mate covered red led lamps controlled via Arduino Nano. ## Analysis Statistical analyses were conducted using R version 4.1.2. The average correct RTs for congruent and incongruent trials were compared for each robot presence condition individually to check for consistency with previous studies regarding the strength of the gaze cueing effect. To test for the effect of cue-target congruence, for the virtual robot condition a t-test was calculated comparing the mean reactions times of congruent trials and incongruent trials. For the copresent robot condition a Wilcoxon test was calculated instead, to account for non-normally distributed data. Significant differences in reaction times between congruent and incongruent trials evidence the presence of a gaze cueing effect. Moreover, a mixed model analysis of variance (ANOVA) with a between-subject factor of robot presence (2: copresence vs. virtual presence), and within-subject factors of cue-target congruency (2: congruent vs. incongruent) was calculated. This analysis was used to assess the individual effect of cue-target congruency and the effect of presence specificity of gaze cueing. Robot presence specific gaze-cueing effects would be evidenced by a significant interaction between presence condition and cue-target congruency (over and above a main effect of congruency). By contrast, presence nonspecific gaze cueing would manifest in terms of a main effect of cue-target congruency (not accompanied by a presence x congruency interaction), with equal facilitation for all robot presence conditions of interest. All follow-up comparisons between conditions were carried out with paired t-tests. # Results Mean reaction times were subjected to a two-way analysis of variance with two levels of robot presence (copresent robot, virtual agent) and two levels of cue-target congruence (congruent, incongruent) to test the effect of robot physical presence and gaze cueing behavior on participants' reaction times in a localization task (see Table [2](#table1){reference-type="ref" reference="table1"}). It is important to note that the data was not normally distributed in each group. An ANOVA was conducted either way, as F-Tests have been reported to be relatively robust to violations of normality when homogeneity of variance is given (Blanca Mena et al., 2017). \ *Note. \*\*\* p\(<\).001, \*\* p\(<\).01, \* p\(<\).05* [\[table1\]]{#table1 label="table1"} The main effect of gaze congruency yielded an F-ratio of F(1, 40) = 18.71, p \(<\).001, indicating that mean reaction times differed significantly between congruent and incongruent trials. This result was followed by a simple comparison between the congruent and incongruent trials using a paired t-test. The results of the paired t-test showed that, on average, participants had faster reaction times on congruent cue-target trials (M = 288 ms, SD = 37.2 ms) than on incongruent cue-target trials (M = 298 ms, SD = 43.2 ms). This difference was significant t(80) =-1, p \(<\).01, with a moderate effect size of r = 0.7. To illustrate the size of the GCE by robot presence group, individual analyses were performed for both robot conditions (see Figure [\[figure4\]](#figure4){reference-type="ref" reference="figure4"}). For the comparison of mean reaction times in congruent and incongruent trials in the copresent robot group a Wilcoxon signed-rank test was calculated to account for non-normally distributed data as indicated by a significant Shapiro-Wilk test (W =.9, p =.009). On average, participants in the copresent robot condition responded faster on congruent trials (M = 291 ms) than on incongruent trials (M = 305 ms). Results of the Wilcoxon signed-rank test showed that this difference was statistically significant (p =.001), with a large effect size, r = 0.7. An additional t-test conducted on the mean reaction times of participants in the virtual agent group revealed a significant difference between congruent and incongruent trials (t (20) =-2, p =.03) with shorter reaction times for congruent (M = 284 ms) than incongruent trials (M = 291 ms). The effect size was at a moderate level, r = 0.5. While participants in the virtual agent group (M = 287 ms) on average reacted faster than participants in the copresent group (M = 298 ms), the main effect of robot presence on participants' mean reaction times was non-significant F(1, 40) = 0.64, p \(<\).05. Moreover, no significant interaction effect of gaze congruency and presence could be found, F(1, 40) = 2.22, p \(<\).05. # Discussion Our results are consistent with previous research on gaze following with copresent robots (): Participants consistently exhibited gaze-following behavior, as evidenced by slower reaction times on trials in which the robot cued the wrong target compared to trials in which the cued location and target location matched. Hence, our results replicate the well-known finding that participants locate a target that is congruent with the cued direction more quickly than a target that is incongruent with the cued direction (; ). In particular, our results are in line with findings by and, that showed that a gaze cueing effect can be found when the target stimulus is directed by the gaze of an embodied robot. This finding is particularly valuable given the ongoing replication crisis in psychological research, which has highlighted the problem of replicating the results of many scientific studies () and allows for the generalization of the gaze cueing effect across different robotic platforms. Importantly, with an effect of 14 ms, the gaze cueing effect in our study is smaller than the 25 ms found by but in line with findings of other studies using more controlled settings (; ). Differences in the extent of the effect might be explained by the design of the robot-possibly, the Meka robot used by offers more social cues or other affordances than the iCub robot used in the present study or other robotic platforms. Further studies will be needed to better understand the relationship of robot design and gaze following behavior. Moreover, our results show that even the screen versions of our robot consistently triggered a gaze cueing effect as indicated by slower reaction times in incongruent compared to congruent trials, showing that gaze following in HRI is not limited to physically present versions of embodied robots. Importantly, however, the same stimuli did not elicit varying degrees of gaze-cueing when comparing the different ways the robot was presented to the participants, as evidenced by a non-significant interaction of robot presence and cue-target congruency. Across conditions participants consistently followed the gaze, independently of whether they were confronted with a copresent iCub or a simulated iCub. As the novelty of this study lies in comparing the way the different ways of presenting the robot influence simple social attention mechanisms, as evidenced by the gaze cueing effect, there exists hardly any literature indicating similar or contradictory results. Importantly, however, these results add to findings of who were comparing the effect of embodiment and presence of four different types of agents on similar outcome variables. Results showed that embodiment but not physical presence was the factor that accounts for the significant difference in the participants' response, as indicated by no significant difference in results when presenting the participants with a copresent robot compared to a telepresent robot. However, the outcomes differed significantly for the comparisons of virtual agents and both forms of embodied robots. In contrast, our study found no difference in gaze following behavior between the presence conditions, as participants' reaction times were not significantly different when interacting with a virtual agent compared to a copresent robot. There are multiple possible reasons why physical presence might not additionally influence the gaze cueing effect. One could be that our results are based on the fact that robots in both conditions moved in a similar, human-like manner. Previous research has shown that (natural) movement is linked to mind attributions-famously for example in the Heider and Simmel illusion (), in which participants attribute mental states to three moving geometrical figures (two triangles that seem to "hunt" a circle). Studies examining how mental state attributions alter gaze following behavior in trials with photographs of robots have shown that this manipulation led to the occurrence of a gaze cueing effect that was not otherwise present (). Differences between our results and those of studies using only photos of agents/robots () or non-moving agents, as in the research of, might be due to the association of motion and mind attributions leading to typical social interaction phenomena, such as gaze cueing. # Conclusion Social gaze is a fundamental part of human interactions and becomes more relevant in the scope of social human-robot interaction. Gaze cueing, the event in which we observe our interaction partner's gaze and shift our own attention accordingly, is broadly studied using images or virtual versions of robots. As previous research has pointed to the broad range of effects physical presence has in human-robot interaction the question arises whether it affects the strength of the gaze cueing effect and hence, whether results from studies using images or virtual agents can be generalized to copresent robots. \[\...\] The results of our study indicate that gaze cueing is a stable effect in basal human-robot interaction across different robot presence conditions. # First Section ## A Subsection Sample Please note that the first paragraph of a section or subsection is not indented. The first paragraph that follows a table, figure, equation etc. does not need an indent, either. Subsequent paragraphs, however, are indented. ### Sample Heading (Third Level) Only two levels of headings should be numbered. Lower level headings remain unnumbered; they are formatted as run-in headings. #### Sample Heading (Fourth Level) The contribution should contain no more than four levels of headings. Table [3](#tab1){reference-type="ref" reference="tab1"} gives a summary of all heading levels. Displayed equations are centered and set on a separate line. \[x + y = z\] Please try to avoid rasterized images for line-art diagrams and schemas. Whenever possible, use vector graphics instead (see Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"}). For citations of references, we prefer the use of square brackets and consecutive numbers. Citations using labels or the author/year convention are also acceptable. The following bibliography provides a sample reference list with entries for journal articles , an LNCS chapter , a book , proceedings without editors , and a homepage . Multiple citations are grouped ,. [^1]: <https://www.bartneck.de/2008/03/11/the-godspeed-questionnaire-series/> [^2]: [github.com/Sabka/icub-hri-cuing](github.com/Sabka/icub-hri-cuing) [^3]: [github.com/Sabka/icub-hri-cuing](github.com/Sabka/icub-hri-cuing)

{'timestamp': '2023-02-10T02:02:01', 'yymm': '2302', 'arxiv_id': '2302.04329', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04329'}

# Introduction The notion of disentangled learning representations was introduced by-it is meant to be a robust approach to feature learning when trying to learn more about a distribution of data \(X\) or when downstream tasks for learned features are unknown. Since then, disentangled learning representations have been proven to be extremely useful in the applications of natural language processing, content and style separation, drug discovery, fairness, and more. Density estimation of learned representations is an important ingredient to competitive disentanglement methods. state that representations \({\mathbf{z}} \sim Z\) which are disentangled should maintain as much information of the input as possible while having components which are mutually *invariant* to one another. Mutual invariance motivates seeking representations of \(Z\) which have *independent* components extracted from the data, necessitating some notion of \({p_Z({\mathbf{z}})}\). Leading unsupervised disentanglement methods, namely \(\beta\)-VAE, FactorVAE, and \(\beta\)-TCVAE all learn \({p_Z({\mathbf{z}})}\) via the same variational Bayesian framework, but they approach making \({p_Z({\mathbf{z}})}\) independent with different angles. \(\beta\)-VAE indirectly promotes independence in \({p_Z({\mathbf{z}})}\) via enforcing low \(D_{\mathrm{KL}}\) between the representation and a factorized Gaussian prior, \(\beta\)-TCVAE encourages representations to have low Total Correlation (TC) via an ELBO decomposition and importance weighted sampling technique, and FactorVAE reduces TC with help from a monolithic neural network estimate. Other well-known unsupervised methods are Annealed \(\beta\)-VAE, which imposes careful relaxation of the information bottleneck through the VAE \(D_{\mathrm{KL}}\) term during training, and DIP-VAE I & II, which directly regularize the covariance of the learned representation. For a more in-depth description of related work, please see Appendix [\[app:related_work\]](#app:related_work){reference-type="ref" reference="app:related_work"}. While these VAE-based disentanglement methods have been the most successful in the field, point out serious reliability issues shared by all. In particular, increasing disentanglement pressure during training doesn't tend to lead to more independent representations, there currently aren't good unsupervised indicators of disentanglement, and no method consistently dominates the others across all datasets. stress the need to find the right inductive biases in order for unsupervised disentanglement to truly deliver. We seek to make disentanglement more reliable and high-performing by incorporating new inductive biases into our proposed method, Gaussian Channel Autoencoder (GCAE). We shall explain them in more detail in the following sections, but to summarize: GCAE avoids the challenge of representing high-dimensional \({p_Z({\mathbf{z}})}\) via disentanglement with Dual Total Correlation (rather than TC) and the DTC criterion is augmented with a scale-dependent latent variable arbitration mechanism. This work makes the following contributions: - Analysis of the TC and DTC metrics with regard to the curse of dimensionality which motivates use of DTC and a new feature-stabilizing arbitration mechanism - GCAE, a new form of noisy autoencoder (AE) inspired by the Gaussian Channel problem, which permits application of flexible density estimation methods in the latent space - Experiments[^1] which demonstrate competitive performance of GCAE against leading disentanglement baselines on multiple datasets using existing metrics # Background and Initial Findings To estimate \({p_Z({\mathbf{z}})}\), we introduce a discriminator-based method which applies the density-ratio trick and the Radon-Nikodym theorem to estimate density of samples from an unknown distribution. We shall demonstrate in this section the curse of dimensionality in density estimation and the necessity for representing \({p_Z({\mathbf{z}})}\) as a collection of conditional distributions. The optimal discriminator neural network introduced by satisfies: \[\argmax_D \mathbb{E}_{{\mathbf{x}}_r \sim X_{real}}\left[ \log D({\mathbf{x}}) \right] + \mathbb{E}_{{\mathbf{x}}_f \sim X_{fake}} \left[ \log \left(1-D({\mathbf{x}}_f) \right) \right] \triangleq D^*({\mathbf{x}}) = \frac{p_{real}({\mathbf{x}})}{p_{real}({\mathbf{x}}) + p_{fake}({\mathbf{x}})}\] where \(D({\mathbf{x}})\) is a discriminator network trained to differentiate between "real" samples \({\mathbf{x}}_r\) and "fake" samples \({\mathbf{x}}_f\). Given the optimal discriminator \(D^*({\mathbf{x}})\), the density-ratio trick can be applied to yield \(\frac{p_{real}({\mathbf{x}})}{p_{fake}({\mathbf{x}})} = \frac{D^*({\mathbf{x}})}{1-D^*({\mathbf{x}})}\). Furthermore, the discriminator can be supplied conditioning variables to represent a ratio of conditional distributions. Consider the case where the "real" samples come from an *unknown* distribution \({\mathbf{z}} \sim Z\) and the "fake" samples come from a *known* distribution \({\mathbf{u}} \sim U\). Permitted that both \({p_Z({\mathbf{z}})}\) and \({p_U({\mathbf{u}})}\) are finite and \({p_U({\mathbf{u}})}\) is nonzero on the sample space of \({p_Z({\mathbf{z}})}\), the optimal discriminator can be used to retrieve the unknown density \({p_Z({\mathbf{z}})}=\frac{D^*({\mathbf{z}})}{1-D^*({\mathbf{z}})}p_U({\mathbf{z}})\). In our case where \({\mathbf{u}}\) is a uniformly distributed variable, this "transfer" of density through the optimal discriminator can be seen as an application of the Radon-Nikodym derivative of \({p_Z({\mathbf{z}})}\) with reference to the Lebesgue measure. Throughout the rest of this work, we employ discriminators with uniform noise and the density-ratio trick in this way to recover unknown distributions. This technique can be employed to recover the probability density of an \(m\)-dimensional isotropic Gaussian distribution. While it works well in low dimensions (\(m\leq8\)), the method inevitably fails as \(m\) increases. Figure [\[fig:kl_div_joint\]](#fig:kl_div_joint){reference-type="ref" reference="fig:kl_div_joint"} depicts several experiments of increasing \(m\) in which the KL-divergence of the true and estimated distributions are plotted with training iteration. When number of data samples is finite and the dimension \(m\) exceeds a certain threshold, the probability of there being any uniform samples in the neighborhood of the Gaussian samples swiftly approaches zero, causing the density-ratio trick to fail. This is a well-known phenomenon called the *curse of dimensionality* of density estimation. In essence, as the dimensionality of a joint distribution increases, concentrated joint data quickly become isolated within an extremely large space. The limit \(m \leq 8\) is consistent with the limits of other methods such as kernel density estimation (Parzen-Rosenblatt window). Fortunately, the same limitation does not apply to conditional distributions of many jointly distributed variables. Figure [\[fig:kl_div_cond\]](#fig:kl_div_cond){reference-type="ref" reference="fig:kl_div_cond"} depicts a similar experiment of the first in which \(m-1\) variables are independent Gaussian distributed, but the last variable \({\mathbf{z}}_m\) follows the distribution \({\mathbf{z}}_m \sim \mathcal{N}(\mu=(m-1)^{-\frac{1}{2}}\sum_{i=1}^{m-1}{\mathbf{z}}_i,\ \sigma^2=\frac{1}{m})\) (i.e., the last variable is Gaussian distributed with its mean as the sum of observations of the other variables). The marginal distribution of each component is Gaussian, just like the previous example. While it takes more iterations to bring the KL-divergence between the true and estimated conditional distribution to zero, it is not limited by the curse of dimensionality. Hence, we assert that conditional distributions can capture complex relationships between subsets of many jointly distributed variables while avoiding the curse of dimensionality. # Methodology ## Analysis of Dual Total Correlation {#analysis-of-dual-total-correlation .unnumbered} Recent works encourage disentanglement of the latent space by enhancing the Total Correlation (TC) either indirectly or explicitly. TC is a metric of multivariate statistical independence that is non-negative and zero if and only if all elements of \({\mathbf{z}}\) are independent. \[{\text{TC}(Z)} = \mathbb{E}_{\mathbf{z}} \log \frac{{p_Z({\mathbf{z}})}}{\prod_i {p_{Z_i}({\mathbf{z}}_i)}} = \sum_i {h(Z_i)}-{h(Z)}\] evaluate many TC-based methods and conclude that minimizing their measures of TC during training often does not lead to VAE \(\mu\) (used for representation) with low TC. We note that computing \({\text{TC}(Z)}\) requires knowledge of the joint distribution \({p_Z({\mathbf{z}})}\), which can be very challenging to model in high dimensions. We hypothesize that the need for a model of \({p_Z({\mathbf{z}})}\) is what leads to the observed reliability issues of these TC-based methods. Consider another metric for multivariate statistical independence, Dual Total Correlation (DTC). Like TC, DTC is non-negative and zero if and only if all elements of \({\mathbf{z}}\) are independent. \[\text{DTC}({\mathbf{z}}) = \mathbb{E}_{\mathbf{z}} \log \frac{\prod_i {p_{Z_i}({\mathbf{z}}_i|{\mathbf{z}}_{\setminus i})}}{{p_Z({\mathbf{z}})}} = {h(Z)}-\sum_i {h(Z_i|Z_{\setminus i})}\] We use \({\mathbf{z}}_{\setminus i}\) to denote all elements of \({\mathbf{z}}\) except the \(i\)-th element. At first glance, it appears that \(\text{DTC}({\mathbf{z}})\) also requires knowledge of the joint density \(p({\mathbf{z}})\). However, observe an equivalent form of DTC manipulated for the \(i\)-th variable: \[{\text{DTC}(Z)} = {h(Z)}-{h(Z_i|Z_{\setminus i})}-\sum_{j \neq i} {h(Z_j|Z_{\setminus j})} = {h(Z_{\setminus i})}-\sum_{j \neq i} {h(Z_j|Z_{\setminus j})}.\] Here, the \(i\)-th variable only contributes to DTC through each set of conditioning variables \({\mathbf{z}}_{\setminus j}\). Hence, when computing the derivative \(\partial\,{\text{DTC}(Z)}/\partial {\mathbf{z}}_i\), no representation of \({p_Z({\mathbf{z}})}\) is required-only the conditional entropies \({h(Z_j|Z_{\setminus j})}\) are necessary. Hence, we observe that the curse of dimensionality can be avoided through gradient descent on the DTC metric, making it more attractive for disentanglement than TC. However, while one only needs the conditional entropies to compute gradient for DTC, the conditional entropies alone don't measure how close \({\mathbf{z}}\) is to having independent elements. To overcome this, we define the summed information loss \({\mathcal{L}_{\Sigma I}}\): \[{\mathcal{L}_{\Sigma I}}(Z) \triangleq \sum_i {I(Z_i;Z_{\setminus i})} = \left[\sum_i {h(Z_i)}-{h(Z_i|Z_{\setminus i})} \right] + {h(Z)}-{h(Z)} = {\text{TC}(Z)} + {\text{DTC}(Z)}.\] If gradients of each \({I(Z_i;Z_{\setminus i})}\) are taken only with respect to \({\mathbf{z}}_{\setminus i}\), then the gradients are equal to \(\frac{\partial {\text{DTC}(Z)}}{\partial {\mathbf{z}}}\), avoiding use of any derivatives of estimates of \({p_Z({\mathbf{z}})}\). Furthermore, minimizing one metric is equivalent to minimizing the other: \({\text{DTC}(Z)} = 0 \Leftrightarrow {\text{TC}(Z)} = 0 \Leftrightarrow {\mathcal{L}_{\Sigma I}} = 0\). In our experiments, we estimate \({h(Z_i)}\) with batch estimates \(\mathbb{E}_{{{\mathbf{z}}_{\setminus i}}}{p_{Z_i}({\mathbf{z}}_i|{\mathbf{z}}_{\setminus i})}\), requiring no further hyperparameters. Details on the information functional implementation are available in Appendix [\[app:impl:info_func\]](#app:impl:info_func){reference-type="ref" reference="app:impl:info_func"}. ### Excess Entropy Power Loss {#excess-entropy-power-loss .unnumbered} We found it very helpful to "stabilize" disentangled features by attaching a feature-scale dependent term to each \({I(Z_i;Z_{\setminus i})}\). The entropy power of a latent variable \({\mathbf{z}}_i\) is non-negative and grows analogously with the variance of \({\mathbf{z}}_i\). Hence, we define the Excess Entropy Power loss: \[{\mathcal{L}_{\text{EEP}}}(Z) \triangleq \frac{1}{2 \pi e} \sum_i \left[ {I(Z_i;Z_{\setminus i})} \cdot e^{2 {h(Z_i)}} \right],\] which weighs each component of the \({\mathcal{L}_{\Sigma I}}\) loss with the marginal entropy power of each \(i\)-th latent variable. Partial derivatives are taken with respect to the \({{\mathbf{z}}_{\setminus i}}\) subset only, so the marginal entropy power only weighs each component. While \(\nabla_\phi{\mathcal{L}_{\text{EEP}}} \neq \nabla_\phi{\mathcal{L}_{\Sigma I}}\) in most situations (\(\phi\) is the set of encoder parameters), this inductive bias has been extremely helpful in consistently yielding high disentanglement scores. An ablation study with \({\mathcal{L}_{\text{EEP}}}\) can be found in Appendix [\[app:ablation_study\]](#app:ablation_study){reference-type="ref" reference="app:ablation_study"}. The name "Excess Entropy Power" is inspired by DTC's alternative name, excess entropy. ## Gaussian Channel Autoencoder {#gaussian-channel-autoencoder .unnumbered} We propose Gaussian Channel Autoencoder (GCAE), composed of a coupled encoder \(\phi: X \rightarrow Z_\phi\) and decoder \(\psi: Z \rightarrow \hat{X}\), which extracts a representation of the data \({\mathbf{x}} \in \mathbb{R}^n\) in the latent space \({\mathbf{z}} \in \mathbb{R}^m\). We assume \(m \ll n\), as is typical with autoencoder models. The output of the encoder has a bounded activation function, restricting \({\mathbf{z}}_\phi \in (-3, 3)^m\) in our experiments. The latent space is subjected to Gaussian noise of the form \({\mathbf{z}} = {\mathbf{z}}_\phi + \nu_\sigma\), where each \(\nu_\sigma \sim \mathcal{N}(0, \sigma^2I)\) and \(\sigma\) is a controllable hyperparameter. The Gaussian noise has the effect of "smoothing" the latent space, ensuring that \({p_Z({\mathbf{z}})}\) is continuous and finite, and it guarantees the existence of the Radon-Nikodym derivative. Our reference noise for all experiments is \({\mathbf{u}} \sim \text{Unif}(-4, 4)\). The loss function for training GCAE is: \[\mathcal{L}_{\text{GCAE}} = \mathbb{E}_{{\mathbf{x}}, \nu_\sigma} \left[ \frac{1}{n}\|\hat{{\mathbf{x}}}-{\mathbf{x}}\|^2_2\right] + \lambda\,{\mathcal{L}_{\text{EEP}}}(Z),\] where \(\lambda\) is a hyperparameter to control the strength of regularization, and \(\nu_\sigma\) is the Gaussian noise injected in the latent space with the scale hyperparameter \(\sigma\). The two terms have the following intuitions: the mean squared error (MSE) of reconstructions ensures \({\mathbf{z}}\) captures information of the input while \({\mathcal{L}_{\text{EEP}}}\) encourages representations to be mutually independent. # Experiments We evaluate the performance of GCAE against the leading unsupervised disentanglement baselines \(\beta\)-VAE, FactorVAE, \(\beta\)-TCVAE, and DIP-VAE-II. We measure disentanglement using four popular supervised disentanglement metrics: Mutual Information Gap (MIG), Factor Score, DCI Disentanglement, and Separated Attribute Predictability (SAP). The four metrics cover the three major types of disentanglement metrics identified by in order to provide a complete comparison of the quantitative disentanglement capabilities of the latest methods. We consider two datasets which cover different data modalities. The **Beamsynthesis** dataset is a collection of \(360\) timeseries data from a linear particle accelerator beamforming simulation. The waveforms are \(1000\) values long and are made of two independent data generating factors: *duty cycle* (continuous) and *frequency* (categorical). The **dSprites** dataset is a collection of \(737280\) synthetic images of simple white shapes on a black background. Each \(64 \times 64\) pixel image consists of a single shape generated from the following independent factors: *shape* (categorical), *scale* (continuous), *orientation* (continuous), *x-position* (continuous), and *y-position* (continuous). All experiments are run using the PyTorch framework using 4 NVIDIA Tesla V100 GPUs, and all methods are trained with the same number of iterations. Hyperparameters such as network architecture and optimizer are held constant across all models in each experiment (with the exception of the dual latent parameters required by VAE models). Latent space dimension is fixed at \(m=10\) for all experiments, unless otherwise noted. More details are in Appendix [\[app:exp_details\]](#app:exp_details){reference-type="ref" reference="app:exp_details"}. In general, increasing \(\lambda\) and \(\sigma\) led to lower \({\mathcal{L}_{\Sigma I}}\) but higher MSE at the end of training. Figure [\[fig:mse_scatter\]](#fig:mse_scatter){reference-type="ref" reference="fig:mse_scatter"} depicts this relationship for Beamsynthesis and dSprites. Increasing \(\sigma\) shifts final loss values towards increased independence (according to \({\mathcal{L}_{\Sigma I}}\)) but slightly worse reconstruction error. This is consistent with the well-known Gaussian channel-as the relative noise level increases, the information capacity of a power-constrained channel decreases. The tightly grouped samples in the lower right of the plot correspond with \(\lambda=0\) and incorporating any \(\lambda>0\) leads to a decrease in \({\mathcal{L}_{\Sigma I}}\) and increase in MSE. As \(\lambda\) is increased further the MSE increases slightly as the average \({\mathcal{L}_{\Sigma I}}\) decreases. Figure [\[fig:mig_scatter\]](#fig:mig_scatter){reference-type="ref" reference="fig:mig_scatter"} plots the relationship between final \({\mathcal{L}_{\Sigma I}}\) values with MIG evaluation scores for both Beamsynthesis and dSprites. Our experiments depict a moderate negative relationship with correlation coefficient \(-0.823\). These results suggest that \({\mathcal{L}_{\Sigma I}}\) is a promising unsupervised indicator of successful disentanglement, which is helpful if one does not have access to the ground truth data factors. ## Effect of \(\lambda\) and \(\sigma\) on Disentanglement {#effect-of-lambda-and-sigma-on-disentanglement .unnumbered} In this experiment, we plot the disentanglement scores (average and standard deviation) of GCAE as the latent space noise level \(\sigma\) and disentanglement strength \(\lambda\) vary on Beamsynthesis and dSprites. In each figure, each dark line plots the average disentanglement score while the shaded area fills one standard deviation of reported scores around the average. Figure [\[fig:hyperparams_bs\]](#fig:hyperparams_bs){reference-type="ref" reference="fig:hyperparams_bs"} depicts the disentanglement scores of GCAE on the Beamsynthesis dataset. All \(\sigma\) levels exhibit relatively low scores when \(\lambda\) is set to zero (with the exception of FactorScore). In this situation, the model is well-fit to the data, but the representation is highly redundant and entangled, causing the "gap" or "separatedness" (in SAP) for each factor to be low. However, whenever \(\lambda > 0\) the disentanglement performance increases significantly, especially for MIG, DCI Disentanglement, and SAP with \(\lambda \in [0.1, 0.2]\). There is a clear preference for higher noise levels, as \(\sigma=0.1\) generally has higher variance and lower disentanglement scores. FactorScore starts out very high on Beamsynthesis because there are just two factors of variation, making the task easy. Figure [\[fig:hyperparams_dsprites\]](#fig:hyperparams_dsprites){reference-type="ref" reference="fig:hyperparams_dsprites"} depicts the disentanglement scores of GCAE on the dSprites dataset. Similar to the previous experiment with Beamsynthesis, no disentanglement pressure leads to relatively low scores on all considered metrics (\(\sim 0.03\) MIG, \(\sim 0.47\) FactorScore, \(\sim 0.03\) DCI, \(\sim 0.08\) SAP), but introducing \(\lambda > 0\) signficantly boosts performance on a range of scores \(\sim 0.35\) MIG, \(\sim 0.6\) FactorScore, \(\sim 0.37\) SAP, and \(\sim 0.45\) DCI (for \(\sigma=\{0.2, 0.3\}\)). Here, there is a clear preference for larger \(\sigma\); \(\sigma=\{0.2, 0.3\}\) reliably lead to high scores with little variance. ## Comparison of GCAE with Leading Disentanglement Methods {#comparison-of-gcae-with-leading-disentanglement-methods .unnumbered} We incorporate experiments with leading VAE-based baselines and compare them with GCAE \(\sigma=0.2\). Each solid line represents the average disentanglement scores for each method and the shaded areas represent one standard deviation around the mean. Figure [\[fig:beamsynthesis_comparison\]](#fig:beamsynthesis_comparison){reference-type="ref" reference="fig:beamsynthesis_comparison"} depicts the distributional performance of all considered methods and metrics on Beamsynthesis. When no disentanglement pressure is applied, disentanglement scores for all methods are relatively low. When disentanglement pressure is applied (\(\lambda,\beta > 0\)), the scores of all methods increase. GCAE scores highest or second-highest on each metric, with low relative variance over a large range of \(\lambda\). \(\beta\)-TCVAE consistently scores second-highest on average, with moderate variance. FactorVAE and \(\beta\)-VAE tend to perform relatively similarly, but the performance of \(\beta\)-VAE appears highly sensitive to hyperparameter selection. DIP-VAE-II performs the worst on average. Figure [\[fig:dsprites_comparison\]](#fig:dsprites_comparison){reference-type="ref" reference="fig:dsprites_comparison"} shows a similar experiment for dSprites. Applying disentanglement pressure significantly increases disentanglement scores, and GCAE performs very well with relatively little variance when \(\lambda \in [0.1, 0.5]\). \(\beta\)-VAE achieves high top scores with extremely little variance but only for a very narrow range of \(\beta\). \(\beta\)-TCVAE scores very high on average for a wide range of \(\beta\) but with large variance in scores. FactorVAE consistently scores highest on FactorScore and it is competitive on SAP. DIP-VAE-II tends to underperform compared to the other methods. ## Disentanglement Performance as Z Dimensionality Increases {#disentanglement-performance-as-z-dimensionality-increases .unnumbered} We report the disentanglement performance of GCAE and FactorVAE on the dSprites dataset as \(m\) is increased. FactorVAE is the closest TC-based method: it uses a single monolithic discriminator and the density-ratio trick to explicitly approximate \({\text{TC}(Z)}\). Computing \({\text{TC}(Z)}\) requires knowledge of the joint density \({p_Z({\mathbf{z}})}\), which is challenging to compute as \(m\) increases. Figure [\[fig:m_comparison\]](#fig:m_comparison){reference-type="ref" reference="fig:m_comparison"} depicts an experiment comparing GCAE and FactorVAE when \(m=20\). The results for \(m=10\) are included for comparison. The average disentanglement scores for GCAE \(m=10\) and \(m=20\) are very close, indicating that its performance is robust in \(m\). This is not the case for FactorVAE-it performs worse on all metrics when \(m\) increases. Interestingly, FactorVAE \(m=20\) seems to recover its performance on most metrics with higher \(\beta\) than is beneficial for FactorVAE \(m=10\). Despite this, the difference suggests that FactorVAE is not robust to changes in \(m\). # Discussion Overall, the results indicate that GCAE is a highly competitive disentanglement method. It achieves the highest average disentanglement scores on the Beamsynthesis and dSprites datasets, and it has relatively low variance in its scores when \(\sigma=\{0.2, 0.3\}\), indicating it is reliable. The hyperparameters are highly transferable, as \(\lambda \in [0.1, 0.5]\) works well on multiple datasets and metrics, and the performance does not change with \(m\), contrary to the TC-based method FactorVAE. GCAE also used the same data preprocessing (mean and standard deviation normalization) across the two datasets. We also find that \({\mathcal{L}_{\Sigma I}}\) is a promising indicator of disentanglement performance. While GCAE performs well, it has several limitations. In contrast to the VAE optimization process which is very robust, the optimization of \(m\) discriminators is sensitive to choices of learning rate and optimizer. Training \(m\) discriminators requires a lot of computation, and the quality of the learned representation depends heavily on the quality of the conditional densities stored in the discriminators. Increasing the latent space noise \(\sigma\) seems to make learning more robust and generally leads to improved disentanglement outcomes, but it limits the corresponding information capacity of the latent space. # Conclusion We have presented Gaussian Channel Autoencoder (GCAE), a new disentanglement method which employs Gaussian noise and flexible density estimation in the latent space to achieve reliable, high-performing disentanglement scores. GCAE avoids the curse of dimensionality of density estimation by minimizing the Dual Total Correlation (DTC) metric with a weighted information functional to capture disentangled data generating factors. The method is shown to consistently outcompete existing SOTA baselines on many popular disentanglement metrics on Beamsynthesis and dSprites.

{'timestamp': '2023-02-10T02:03:05', 'yymm': '2302', 'arxiv_id': '2302.04362', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04362'}

# Introduction and Main result {#sec:intro} We are interested in mixing for the continuous time dynamics of the \(\mathbb Z^d\)-periodic Lorentz gas (\(d\in\{1,2\}\)). This model has been introduced by Lorentz in to model the diffusion of electrons in a low conductive metal. It describes the behaviour of a point particle moving at unit speed in the plane \(\mathcal D_2:=\mathbb R^2\) (when \(d=2\)) or on the tube \(\mathcal D_1:=\mathbb R\times \mathbb T\) (when \(d=1\)) between a \(\mathbb Z^d\)-periodic locally finite configuration of convex obstacles with disjoint closures and \(\mathcal C^3\) boundary (with non null curvature), with elastic collisions on them (pre-collisional and post-collisional angles being equal). We write \(\Omega_d\) for the set of possible positions, that is the set of positions in \(\mathcal D_d\) that are not inside an obstacle. The set of configurations is the set \(\widetilde{\mathcal M}\) of couples of position and unit velocity \((q,\vec v)\in \Omega_d\times \mathbb S^1\), identifying pre-collisional and post-collisional vectors at a collision time (rigorously, \(\widetilde{\mathcal M}\) is the quotient of \(\Omega_d\times \mathbb S^1\) by the equivalence relation identifying pre-and post-collisional vectors). The Lorentz gas flow \((\Phi_t)_t\) maps a configuration \((q,\vec v)\) (corresponding to a couple position and velocity at time 0) to the configuration \(\Phi_t(q,\vec v)=(q_t,\vec v_t)\) corresponding to the couple position and velocity at time \(t\) of a particle that was at time \(0\) at position \(q\) with velocity \(\vec v\). This flow \((\Phi_t)_t\) preserves the infinite Lebesgue measure \(\widetilde\nu\) on \(\Omega_d\times\mathbb S^1\), normalized so that \(\widetilde\nu\left((\Omega_2\cap[0,1[^2)\times\mathbb S^1\right)=1\) if \(d=2\) and so that \(\widetilde\nu\left((\Omega_1\cap([0,1[\times\mathbb T))\times\mathbb S^1\right)=1\) if \(d=1\). It is natural to consider also the dynamics at collision times. The space \(\widetilde M\) for this dynamics is the set of configurations \((q,\vec v)\in\widetilde{\mathcal M}\) with \(q\in\partial \Omega_d\). The collision map \(\widetilde T:\widetilde M\to\widetilde M\), that maps a configuration at a collision time to the configuration at the next collision time, is referred to as the Lorentz gas map and preserves an infinite measure \(\widetilde\mu\) absolutely continuous with respect to the Lebesgue measure. Let us write \(\widetilde W_t:\widetilde{\mathcal M}\rightarrow \mathcal D_d\) for the map corresponding to the displacement up to time \(t\) : \[\forall (q,\vec v)\in\widetilde{\mathcal M},\quad \Phi_t(q,\vec v)=(q_t,\vec v_t)\quad\Rightarrow\quad \widetilde W_t(q,\vec v)=q_t-q\,.\] If \(d=1\) we set \(\widetilde W'_t:\widetilde{\mathcal M}\rightarrow \mathbb R\) for the first coordinate of \(\widetilde W_t\). If \(d=2\), \(\widetilde W_t\) takes its values in \(\mathbb R^2\), we then just set \(\widetilde W'_t=\widetilde W_t\). In both cases, \(\widetilde W'_t\) is the natural projection of \(\widetilde W_t\) on \(\mathbb R^d\). When every trajectory touches eventually at least one obstacle, we speak of *finite horizon* Lorentz gas. In the finite horizon case, it follows from that \(\widetilde W_t\) satisfies a standard central Limit Theorem meaning that \((\widetilde W'_t/\sqrt{t})_t\) converges strongly in distribution[^1], as \(t\rightarrow +\infty\), to a centered Gaussian random variable with non degenerate variance matrix given by an infinite sum.\ When there exists at least a trajectory that never touches an obstacle, we speak of *infinite horizon* Lorentz gas. In this article, we focus on the \"fully dimensional\" infinite horizon case, meaning that there exist at least \(d\) non parallel unbounded trajectories touching no obstacle. In this case it follows from by Szász and Varjú (see Section [2.1](#sec:obstacles){reference-type="ref" reference="sec:obstacles"}) that[^2] \[\label{eq:Sigma} \frac{\widetilde W'_t}{\sqrt{t\log t}}\Longrightarrow \mathcal N(0,\Sigma)\, ,\quad \mbox{as }t\rightarrow +\infty\,,\] where \(\Sigma\) is a \(d\)-dimensional definite positive symmetric matrix which, furthermore, is given by an explicit formula in terms of the configuration of obstacles (recalled at the beginning of Section [6](#sec:proofjCLT){reference-type="ref" reference="sec:proofjCLT"}). We are interested here in the question of strong mixing. We recall that an infinite measure preserving system \((\hat X,\hat T,\hat\mu)\) is said to be strongly mixing if there exist a sequence \(\widehat{a}_n\to\infty\) and a class of integrable functions \(f,g\) so that \[\label{def:mix} \widehat{a}_n\int_{\hat M}f.g\circ \hat T^n\, d\hat\mu \to\int_{\hat M}f\, d\hat\mu \int_{\hat M}g\, d\hat\mu.\] The sequence \(\widehat{a}_n\) gives the speed of convergence to \(0\) of \(\int_{\hat M}f.g\circ \hat T^n\, d\hat\mu\). The first such rate was obtained in  for a very restrictive class of intermittent maps preserving an infinite measure. This was later generalized to larger classes of such maps in  and . For other notions of mixing in the infinite measure set up (such as local-global and global-global) introduced in  (see also  and the reference therein). In the set up of the *discrete* time Lorentz gas \((\widetilde M,\widetilde T,\widetilde\mu)\), mixing in the sense of [\[def:mix\]](#def:mix){reference-type="eqref" reference="def:mix"} is well understood in both finite and infinite horizon case and it is a direct consequence of a mixing local limit theorem (MLLT) for the cell change (see e.g. ). For the finite horizon case, we refer to  for the key LLT (which implies MLLT and thus, mixing) and to  for expansions of any order. In the much more difficult set up of infinite horizon case, we refer to  for LLT (which again, implies MLLT and mixing) and to  for error terms. There is a pletora of limit theorems known in the discrete time set up with finite horizon case. Some results are also known for the discrete time Lorentz gas with infinite horizon case (in particular,  and more recently, ). Mixing for *continuous* time Lorentz gas \((\widetilde{\mathcal M},(\Phi_t)_t,\widetilde\nu)\) is seriously more challenging. Even in the set up of the *finite* horizon Lorentz gas flow, mixing in the sense of [\[def:mix\]](#def:mix){reference-type="eqref" reference="def:mix"} was open until the work of  and very recently, expansions of any order have been obtained in. Strictly speaking, the work  focused on a mixing local limit theorem (MLLT) for the Sinai billiard flow with *finite* horizon, but as in, mixing in the sense of [\[def:mix\]](#def:mix){reference-type="eqref" reference="def:mix"} and MLLT are equivalent. For related, but weaker, results on MLLT for group extensions of suspension flows with bounded roof function, not applicable as such to Sinai billiards we refer to . Nothing is known about the mixing for the Lorentz gas flow with *infinite* horizon. In this paper we address this open question and establish Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} gives mixing for observables with support that may contain configuration with infinite free flights. In the set up of the Lorentz gas flow with infinite horizon this class of observables is the natural one. Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} can be rephrased in terms of vague convergence (see comments after Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"}). The main ingredients, which are new and important results on their own, used in the proof of Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} are 1. Strong mixing for observables \(f,g\) with supports uniformly 'close' to a collision time (Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"}), i.e. the supports of \(f,g\) are at a bounded time of the closest collision time (either in the past or in the future). This mixing result is an easy consequence of a MLLT for the Sinai billiard flow with infinite horizon. The present MLLT, Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, is a first main result and is established via two joint local limit results for the Sinai billiard map on the cell change function and flight time together: a) Joint LLT, Lemma [\[lem:jointllt\]](#lem:jointllt){reference-type="ref" reference="lem:jointllt"}; b) Joint Local Large Deviation, Lemma [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"}. 2. A tightness type result, Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}, that allows \(f,g\) to have any compact support in \(\widetilde{\mathcal M}\). In particular, the supports of \(f,g\) can contain configurations of particles that will never hit an obtstacle. The proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} exploits a very delicate decomposition of the type of possible free flights along with Joint MLLT with error terms as in Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"} and Corollary [\[coro:llt0cor\]](#coro:llt0cor){reference-type="ref" reference="coro:llt0cor"}. We emphasize that the MLLT with *error terms*, namely Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"} and Lemma [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"} are required ingredients for the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}. A strategy of the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} is provided in Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"}. The proof of this result exploits several new ideas, and several new technical estimates are obtained throughout the proof. We conclude the introductory section with a very brief summary of the various results along with an outline of the paper. In Section [2](#sec:MLLT){reference-type="ref" reference="sec:MLLT"}, we introduce most of the required notation, and state MLLTs for the Sinai billiard flow as in Proposition [\[prop:MLLTkappa0\]](#prop:MLLTkappa0){reference-type="ref" reference="prop:MLLTkappa0"} for the cell change function, and as in Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} for the flight function. In Section [2](#sec:MLLT){reference-type="ref" reference="sec:MLLT"} we also record a consequence of Proposition [\[prop:MLLTkappa0\]](#prop:MLLTkappa0){reference-type="ref" reference="prop:MLLTkappa0"}, namely Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"} that proves mixing for continuous observables with compact support consisting of configurations at a bounded time from the closest collision; in short, this gives mixing for continuous observables with 'non-infinite' free flights. In Section [3](#sec:MLLT+LLD){reference-type="ref" reference="sec:MLLT+LLD"}, we state the joint limit results (Joint CLT, Joint MLLT with *error terms*, Joint LLD) for the Sinai billiard map, for the couple formed by the cell change function with the flight time. Using the statement of these key technical ingredients, in Sections [4](#sec:proofMLLT){reference-type="ref" reference="sec:proofMLLT"} and [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"} we prove Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}, and Corollary [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"}. The proofs of the technical key results stated in Section [3](#sec:MLLT+LLD){reference-type="ref" reference="sec:MLLT+LLD"} are included in Sections [6](#sec:proofjCLT){reference-type="ref" reference="sec:proofjCLT"}, [7](#sec:proofjMLLT){reference-type="ref" reference="sec:proofjMLLT"} and [\[sec:jLLD\]](#sec:jLLD){reference-type="ref" reference="sec:jLLD"}. While the joint LLD, Lemma [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"}, follows by slightly modifying the proof of LLD for the cell change function obtained in , all the other technical results obtained in this paper, namely, the Joint CLT and the Joint MLLT with *error terms* stated in Section [3](#sec:MLLT+LLD){reference-type="ref" reference="sec:MLLT+LLD"} are new and require serious new ideas and work. In Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"} we state and prove the tightness result Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}. At the beginning of Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"}, we use the statement of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} to complete the proof of Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"}. The role and the novelty of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} has been already summarized in item 2. above. Finally, we mention that in Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"}, as a by product of certain technical lemmas we obtain a large deviation result, namely Proposition [\[prop:L\]](#prop:L){reference-type="ref" reference="prop:L"}, which is of independent interest. #### Notation We use "big O" and \(\ll\) notation interchangeably, writing \(b_n=O(c_n)\) or \(b_n\ll c_n\) if there are constants \(C>0\), \(n_0\ge1\) such that \(b_n\le Cc_n\) for all \(n\ge n_0\). As usual, \(b_n=o(c_n)\) means that there exists \(\varepsilon_n\) such that, for all \(n\) large enough, \(b_n=c_n\varepsilon_n\) and \(\lim_{n\rightarrow +\infty}\varepsilon_n=0\) and \(b_n\sim c_n\) means that \(b_n=c_n+o(c_n)\). Unless otherwise specified, given \(x\in\mathbb{R}^d\), we let \(|x|\) be the usual Euclidean norm of \(x\). Thoroughout this article, when \(d=1\), we identify \(\mathbb Z^1\) (resp. \(\mathbb R^1\)) with \(\mathbb Z\times\{0\}\) (resp. \(\mathbb R\times\{0\}\)). In particular, any \((q,z)\in\Omega_1\times\mathbb R\), \(q+z\) means \(q+(z,0)\). # MLLT for the Sinai billiard flow and mixing for the \(\mathbb Z^d\)-extension flow {#sec:MLLT} ## Notations and previous results {#sec:obstacles} Let \(d\in\{1,2\}\). The domain \(\Omega_d\) of the \(\mathbb Z^d\)-periodic Lorentz gas is given by \(\Omega_d:=\mathcal D_d\setminus \bigcup_{i=1}^I\bigcup_{\ell\in\mathbb Z^d}(\mathcal O_i+\ell)\) where \(\mathcal O_1...,\mathcal O_I\) is a nonempty finite family of convex open sets with \(\mathcal C^3\) boundary of non null curvature such that the obstacles \(\mathcal O_i+\ell\) have pairwise disjoint closures. We recall that we are interested in the fully dimensional infinite horizon and so assume throughout that the interior of the billiard domain \(\Omega\) contains at least \(d\) unbounded corridors (made of unbounded parallel lines) the direction of which are not parallel to each other. ### Sinai billiard {#sinai-billiard .unnumbered} Quotienting the system \((\widetilde{\mathcal M},(\Phi_t)_t)\) by \(\mathbb Z^d\) (for the position), we obtain the Sinai billiard flow \((\mathcal M,(\phi_t)_t)\) (see ) which describes the evolution of point particles moving at unit speed in \(\Omega:=\Omega_d/\mathbb Z^d=\mathbb T^2\setminus \bigcup_{i=1}^I\overline{\mathcal O_i}\) with elastic reflection off \(\partial\Omega\) (where \(\overline{\mathcal O_i}\) is the image of \(\mathcal O_i\) by the canonical projection \(p_d:\mathcal D_d\rightarrow \mathbb T^2\)). The flow \((\phi_t)_t\) preserves the probability measure \(\nu\) on \(\mathcal M=(\Omega\times\mathbb S^1)/\equiv\) that is proportional to the Lebesgue measure, where \(\equiv\) is the equivalence relation identifying pre-and post-collisional vectors. The Poincaré map of \(\phi_t\) with Poincaré's section \(\partial\Omega\times\mathbb S^1\) is the Sinai billiard map \((M,T,\mu)\), where the two-dimensional phase space \(M=\{(q,\vec v)\in\mathcal M:q\in\partial\Omega\}\) (position in \(\partial\Omega\) and unit post-collisional velocity vector) is identified with \(\partial\Omega\times (-\pi/2,\pi/2)\) (we parametrize here the post-collisional velocity vector by its angle with the normal to \(\partial\Omega\)). This map \(T\) sends a post-collisional vector to the post-collisional vector corresponding to the next collision. This map preserves the probability measure \(\mu\) with density \(\cos\varphi/(2|\partial\Omega|)\) at the point \((q,\varphi)\in\partial\Omega\times[-\frac \pi 2,\frac\pi 2]\). The flight time between consecutive collisions is the return time of \((\phi_t)_t\) to \(M\) and we denote it by \(\tau:M\to\mathbb{R}_{+}\). In this notation, we have the following identification \[\begin{aligned} T(x)&=(\phi_\tau)(x)=\phi_{\tau(x)}(x)\,. \end{aligned}\] We set \(\tau_n:=\sum_{k=0}^{n-1}\tau\circ T^k\), with the usual convention \(\tau_0:=0\). For any \(x\in \mathcal M\), we set \(N_t(x)\in\mathbb N_0\) for the *collisions number* in the time interval \((0,t]\) starting from the configuration \(x\). We observe that, for \(x\in M\), this quantity satisfies \[\label{eq:lapn} \tau_{N_t(x)}(x)\le t<\tau_{N_t(x)+1}(x)\,.\] Furthermore, for all \(x\in M\) and all \(u\in[0,\tau(x))\) and any \(t\in [0,+\infty)\), \(N_{t}(\phi_u(x))=N_{t+u}(x)\). With these notations, the Sinai billiard flow \((\mathcal M,(\phi_t)_t,\nu)\) is isomorphic to the suspension flow \((\widehat{\mathcal M},(\widehat{\phi}_t)_t,\widehat\nu)\), given by \[\begin{aligned} \widehat{\mathcal M}&=\{(x,u)\in M\times[0,+\infty):0\le u< \tau(x)\}\\ \widehat\phi_s(x,u)&=(T^n(x),s+u-\tau_{N_{s}(\phi_u(x))}(x))\\ \widehat\nu&=(\mu\times \operatorname{Leb})/{\mu(\tau)},\quad\mbox{where }\mu(\tau):=\int_M\tau\,d\mu \,, \end{aligned}\] via the isomorphism \((x,u)\in\widehat{\mathcal M}\mapsto \phi_u(x)\in\mathcal M\) (this map is injective, its image is the set of configurations in \(\mathcal M\) that do not belong to an infinite free flight). ### \(\mathbb Z^d\)-extension and cell change function {#mathbb-zd-extension-and-cell-change-function .unnumbered} We recall that the \(\mathbb Z^d\)-periodic Lorentz gas map \((\widetilde M,\widetilde T,\widetilde\mu)\) can be represented by the \(\mathbb Z^d\)-extension of the Sinai billiard map \((M,T,\mu)\) by the *cell change* function \(\kappa\) that can be defined as follows. For any \(\ell\in\mathbb Z^d\), we call \(\ell\)-cell the set \(\mathcal C_\ell\) of configurations \((q,v)\in\widetilde M\) such that \(q\in\bigcup_{i=1}^I(\partial O_i+\ell)\). Because of the \(\mathbb Z^d\)-periodicity of the model, there exists \(\kappa:M\rightarrow\mathbb Z^d\), called the *cell change function*, such that \[\label{eq:defkappa} \widetilde x=(q,\vec v)\in\mathcal C_\ell\quad\Rightarrow\quad \widetilde T(x)\in\mathcal C_{\ell+\kappa(p_d(q),\vec v)}\] Note, for any \(\widetilde x\in\mathcal M\), there exists a unique \((x=(q,\vec v),\ell)\in M\times\mathbb Z^d\) such that [^3] \(\widetilde x=(p_{d,0}^{-1}(q)+\ell,\vec v)\) (\(\vec v\) is the velocity of \(\widetilde x\), setting \(\widetilde q\) for the position of \(\widetilde x\), \((q,\ell)\) is such that \(q=p_{d}(\widetilde q)\) and \(\widetilde x\in\mathcal C_\ell\)). Formula [\[eq:defkappa\]](#eq:defkappa){reference-type="eqref" reference="eq:defkappa"} can be rewritten under the form \[\label{Zdextension} \forall ((q,\vec v),\ell)\in M\times\mathbb Z^d,\ T(q,\vec v)=(q',\vec v')\ \Rightarrow\ \widetilde T(p_{d,0}^{-1}(q)+\ell,\vec v) = \left(p_{d,0}^{-1}(q')+\ell+\kappa(q,\vec v),\vec v'\right)\,.\] This gives the identification of \((\widetilde M,\widetilde T,\widetilde\mu)\) by the \(\mathbb Z^d\)-extension of \((M,T,\mu)\) by \(\kappa:M\rightarrow \mathbb Z^d\). A direct and classical induction ensures that, for any \(((q,\vec v),\ell)\in M\times\mathbb Z^d\), \[\label{Zdextensionforn} T^n(q,\vec v)=(q'_n,\vec v'_n)\ \Rightarrow\ \widetilde T^n(p_{d,0}^{-1}(q)+\ell,\vec v)= \left(p_{d,0}^{-1}(q'_n)+\ell+\kappa_n(q,\vec v),\vec v'_n\right)\,,\] where we set \(\kappa_n:=\sum_{j=0}^{n-1}\kappa\circ T^j\). ## Mixing for the Lorentz gas seen as a suspension flow We will use crucially the fact established in the previous section that \((\widetilde{\mathcal M},(\Phi_t)_t,\widetilde\nu)\) can be represented as a suspension flow by \((x,\ell)\mapsto \tau(x)\) over \((\widetilde M,\widetilde T,\widetilde\mu)\) which itself can be represented as a \(\mathbb Z^d\) extension of \((M,T,\mu)\) by \(\kappa\). Thus, we can represent \(\widetilde {\mathcal M}\) by \(\widehat{\mathcal M}\times\mathbb Z^d\). In this part, we state a mixing local limit theorem for \(\kappa_n\) and see how we can use it to easily derive Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} in the case of functions \(f,g\) supported at a bounded time from a collision, i.e. for functions that are compactly supported in \(\widehat{\mathcal M}\times\mathbb Z^d\). As detailed in Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"}, these functions form a much more restrictive class than the ones of Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"}. To state these results, we shall introduce two classes of sets \(\mathcal F\) (resp. \(\widetilde{\mathcal F}\)) that will correspond to the set of sets of configurations in \(\mathcal M\) (resp. \(\widetilde{\mathcal M}\)) with previous collision in some fixed subset of \(M\) (resp. some fixed cell of \(\widetilde M\)), at some time in a fixed bounded time interval. We state now a MLLT for \(\kappa_{N_t}\) defined on \(\mathcal M\) by \[\forall (x,u)\in\widehat{\mathcal M},\ \forall t\in[0,+\infty),\quad\kappa_{N_t}(\phi_u(x)):=\kappa_{N_t(\phi_u(x))}(x)=\kappa_{N_{t+u}(x)}(x)\,.\] This observable \(\kappa_{N_t}\) will be understood as the cell change during the time interval \((0,t]\). This result is contained in a more general MLLT stated in Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"} (applied with \(w_t=\ell\), \(w=0\), \(K=\{0\}\)). An immediate consequence of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is the following light version of Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} for compactly supported observables in the 'extended suspension' \(\widehat{\mathcal M}\times\mathbb{Z}^d\); in particular, the supports of these functions only contain configurations that have hit or will hit an obstacle in a bounded time. The mixing result in Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"} can be rephrased in terms of the vague convergence of the family of \(\mu_t\) to \(\mu\otimes\mu\) where \(\mu_t\) is the measure on \((\widehat{\mathcal M})^2\) defined by \(\mu_t(A'\times B')=\mu(A'\cap \Phi_{-t}B')\) for \(A',B'\in\widetilde{\mathcal F}\) (this is a consequence of the Portmanteau theorem as in, for instance, ), and the same applies for Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"}. ## MLLT for the infinite horizon Sinai flow In this section we state the MLLT for a natural cocycle of the Sinai billiard flow, which corresponds to the displacement. ### Free flight {#free-flight .unnumbered} Due to the \(\mathbb Z^d\)-periodicity the *free flight* \(\widetilde V: \widetilde M\rightarrow \mathcal D_d\) which is defined by \[\label{deftildeV} \forall (q,\vec v)\in \widetilde M,\quad \widetilde T(q,\vec v)=(\widetilde q,\vec v_1)\quad \Rightarrow\quad \widetilde V(q,\vec v)=\widetilde q-q\][\[defV\]]{#defV label="defV"} goes to the quotient by \(\mathbb Z^d\), i.e. there exists \(V:M\rightarrow \mathcal D_d\) such that \[\widetilde V(q,\vec v)=V\left(p_d(q),\vec v\right)\,.\] When \(d=2\), this quantity is related to the flight time \(\tau\) via the following identity \[\label{linktauV} \mbox{if }d=2\, ,\quad \tau=|V|\,.\] Let us show that the free flight \(V\) is cohomologous to the cell change \(\kappa\). It follows from [\[defV\]](#defV){reference-type="eqref" reference="defV"}, [\[deftildeV\]](#deftildeV){reference-type="eqref" reference="deftildeV"} and [\[Zdextension\]](#Zdextension){reference-type="eqref" reference="Zdextension"} that, for all \(x=(q,\vec v)\in M\), if \(T(q,\vec v)=(q',\vec v')\), then \[\begin{aligned} \nonumber V(x)&= \widetilde V\left(p_{d,0}^{-1}(q),\vec v\right)\\ &=p_{d,0}^{-1}(q')+\kappa(q,\vec v)-p_{d,0}^{-1}(q)=\kappa(x)+H_0(T(x))-H_0(x)\, ,\label{coboundPsi} \end{aligned}\] with \(H_0(q,\vec v)=p_{d,0}^{-1}(q)\). Proceeding as for \(\widetilde W_t\) in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"}, if \(d=1\) we set \(V':\mathcal M\rightarrow \mathbb R\) for the first coordinate of \(V\), and if \(d=2\), \(V\) takes its values in \(\mathbb R^2\), we then just set \(V'=V\). The following nonstandard CLT was proved in  for \(V'\): \[\label{eq:cltV} a_n^{-1}\sum_{j=0}^{n-1} V'\circ T^j\implies \mathcal N(0,\Sigma_0)\,,\] where \(a_n=\sqrt{n\log n}\) and where \(\Sigma_0\in\mathbb{R}^{d\times d}\) is a positive-definite symmetric \(d\)-dimensional matrix (see [\[defSigma0dim1\]](#defSigma0dim1){reference-type="eqref" reference="defSigma0dim1"} and [\[defSigma0dim1\]](#defSigma0dim1){reference-type="eqref" reference="defSigma0dim1"}) for precise formulas). An important ingredient of  is that \(V\) lies in the domain of a nonstandard CLT; that is, there exists \(c > 0\) such that \[\label{tailV} \mu(|V | > t)\sim ct^{-2}\,.\] ### Displacement function \(W_t\) {#displacement-function-w_t .unnumbered} We have already defined in Section [1](#sec:intro){reference-type="ref" reference="sec:intro"} the displacement function \(\widetilde W_t:\widetilde{\mathcal M}\rightarrow\mathcal D_d\) and \(\widetilde W'_t:\widetilde{\mathcal M}\rightarrow\mathbb R^d\) its projection on \(\mathbb R^d\). Due to the \(\mathbb Z^d\)-periodicity of our model, both displacement functions go the quotient by \(\mathbb Z^d\), i.e. there exists \(W_t:\mathcal M\rightarrow \mathcal D_d\) and \(W'_t:\mathcal M\rightarrow\mathbb R^d\) such that \[\forall (q,\vec v)\in\widetilde{\mathcal M},\quad \widetilde W_t(q,\vec v)=W_t(p_d(q),\vec v)\quad\mbox{and}\quad \widetilde W'_t(q,\vec v)=W'_t(p_d(q),\vec v)\,.\] Observe that \(W_t\) is a cocycle: \[\label{Wt_cocycle} \forall x\in \mathcal M,\ \forall t,s\ge 0,\quad W_{t+s}(x)=W_s(x)+W_t(\phi_s(x))\,\] and that \[\label{eq:xi} \forall x=(q,\vec v)\in M, \quad V(x)=W_{\tau}(x):=W_{\tau(x)}(x)\,.\] Thus the nonstandard CLT for \(V'\) stated in [\[eq:cltV\]](#eq:cltV){reference-type="eqref" reference="eq:cltV"} implies a nonstandard CLT for \(W'_t\) via the relation [\[eq:xi\]](#eq:xi){reference-type="eqref" reference="eq:xi"} together with the classical scheme of lifting limit theorems from the induced map to the original system (map or flow)  . This leads to the following result where we use the notation \(a_t:=\sqrt{t\log t}\). Let \(v_0:\mathcal M\to\mathbb S^1\) be the velocity map which is given by \(v_0(q,\vec v)=\vec v\). Note that \[\label{eq:Wt} \mbox{if }d=2,\quad W_t:=\int_0^tv_0\circ\phi_s\, ds \,.\] If \(d=1\), then \(W_t\) is the equivalent class in \(\mathcal D_1\) (that is, \(W_t\) is the canonical projection) of the ergodic integral \(\int_0^t v_0\circ\phi_s\,ds\). ### MLLT for the displacement function {#mllt-for-the-displacement-function .unnumbered} Let us see that, due to [\[eq:xi\]](#eq:xi){reference-type="eqref" reference="eq:xi"}, the coboundary equation [\[coboundPsi\]](#coboundPsi){reference-type="eqref" reference="coboundPsi"} for \(V-\kappa\) leads to a similar equation involving \(W\). We consider the function \(H_1:\mathcal M_0\rightarrow \mathcal D_d\) mapping \(\mathbf x\in\mathcal M\) to the position of its representant in \(\mathcal D_d\) with previous collision in \(\mathcal C_0\), that is \[\label{def:H1} \forall ((q,\vec v),u)\in\widehat{\mathcal M},\quad H_1(\phi_u(q,\vec v))=\mathfrak p\left(\Phi_u\left(p_{d,0}^{-1}(q),\vec v\right)\right) =p_{d,0}^{-1}(q)+W_u(q,\vec v)\,,\] where \(\mathfrak p:\widetilde{\mathcal M}\rightarrow\Omega_d\) is the natural projection. In other words, if \(d=2\), then \[\label{eq:H1} H_1(\phi_u(q,\vec v))=p_{d,0}^{-1}(q)+u\vec v\,;\] if \(d=1\), \(H_1(\phi_u(q,\vec v))\) is the class of \(p_{d,0}^{-1}(q)+u\vec v\) in \(\mathcal D_1\). Recall that we set \(N_t:\mathcal M\rightarrow \mathbb N_0\) for the *collisions number* in the time interval \((0,t]\) (see in particular [\[eq:lapn\]](#eq:lapn){reference-type="eqref" reference="eq:lapn"}). The above defined function \(H_1\) satisfies the following important property: \[\label{eq:Wcobound} \forall (x=(q,\vec v),u)\in \widehat {\mathcal M}\, ,\quad W_t(\phi_u(x))=\kappa_{N_{t+u}(x)}(x)+H_1(\phi_t(\phi_u(x)))-H_1(\phi_u(x))\,.\] Indeed, setting \(N:=N_{t+u}(x)\), we notice that \[\label{phitphiu} \phi_t(\phi_u(x))=\phi_{u'}(q', \vec w),\quad \mbox{with }u':=u+t-\tau_{N}(x),\ (q',\vec w):=T^{N}(x)=\phi_{\tau_N}(x)\,,\] and \((T^N(x),u')\) is in \(\widehat{\mathcal M}\). Therefore, it follows from [\[Wt_cocycle\]](#Wt_cocycle){reference-type="eqref" reference="Wt_cocycle"} and [\[eq:xi\]](#eq:xi){reference-type="eqref" reference="eq:xi"} that \[\begin{aligned} W_t(\phi_u(x))&=W_{t+u}(x)-W_u(x)=W_{u'}(\phi_{\tau_N}(x))+W_{\tau_N}(x)-W_u(x)\\ &=W_{u'}(T^N(x))+W_{\tau_N}(x)-W_u(x)\\ &=W_{u'}(T^N(x))+V_{N}(x)-W_u(x)\,. \end{aligned}\] Finally, using [\[coboundPsi\]](#coboundPsi){reference-type="eqref" reference="coboundPsi"} and [\[eq:H1\]](#eq:H1){reference-type="eqref" reference="eq:H1"}, we obtain that \[\begin{aligned} W_t(\phi_u(x)) &=W_{u'}(T^N(x))+\kappa_{N}(x)+H_0(T^N(x))-H_0(x)-W_u(x)\\ &=H_1(\phi_t(\phi_u(x)))+\kappa_{N}(x)-H_1(\phi_u(x))\,, \end{aligned}\] as announced. The proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} is provided in Section [4](#sec:proofMLLT){reference-type="ref" reference="sec:proofMLLT"}, and will appear as a consequence of an analogous result (Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"}) stated for \(\kappa_{N_t}\) instead of \(W_t\) # Statements of the Joint LLT with error term and the joint LLD for the billiard map {#sec:MLLT+LLD} Let \(d\in\{0,1,2\}\). In this section we state the main technical results that will be used in the proofs of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} (MLLT for the Sinai flow) and Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} (mixing for the Lorentz gas), including those used in the proof of the key tightness-type result Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} (stated in Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"}). We are interested in joint MLLT and LLD for the pair \[\widehat{\Psi}=\widehat{\Psi}^{(d)}:=(\kappa,\widetilde\tau):M\to\mathcal D_d\times\mathbb R\, ,\quad\mbox{with }\widetilde\tau:=\tau-\mu(\tau)\, ,\quad\mbox{if }d\in\{1,2\}\] or for \[\widehat \Psi=\widehat\Psi^{(0)}:=\widetilde\tau\, ,\quad\mbox{if }d=0\,.\] Note that \(\int_M \widehat\Psi\,d\mu{=0}\). When \(d\in\{1,2\}\), our joint limit results are related to the fact that that the sums of \((\kappa\circ T^k,\widetilde\tau\circ T^k)_k\) satisfies a CLT with nonstandard normalization \(a_n^{-1}\). In particular, as clarified in the proof of Sublemma [\[sub:asl\]](#sub:asl){reference-type="ref" reference="sub:asl"} below, the vector \(\widehat\Psi\) is so that \(\mu(|\widehat\Psi|>t)\sim ct^{-2}\). As usual, we write \(\widehat\Psi_n=\sum_{j=0}^{n-1}\widehat\Psi\circ T^j\) and similarly for \(V_n,\,\tau_n, \widetilde\tau_n\). We start with a nondegenerate CLT with nonstandard scaling for \(\widehat\Psi_n\). This result is proved in Section [6](#sec:proofjCLT){reference-type="ref" reference="sec:proofjCLT"} by adapting the proof of the CLT for \(V_n\) established in  via , writing \(\widehat\Psi_n\) as a function of the two dimensional cell change plus a Lipschitz function. Let us state a MLLT for \(\widehat\Psi_n\) with a uniform error term uniform. We write \(\Lambda_{d+1}\) for the Haar measure on \(\mathbb Z^d\times\mathbb R\) given by the product of the counting measure on \(\mathbb Z^d\) and of the Lebesgue measure on \(\mathbb R\). This result is proved in Section [7](#sec:proofjMLLT){reference-type="ref" reference="sec:proofjMLLT"}. The scheme of this proof follows the one of the MLLT established in  but there are at least two main differences. First, here we need to obtain a Joint LLT, which is different from  , which obtains LLT with error terms for the cell change function. The new ingredients needed to deal with the Joint LLT (with error) are summarized in Section [6](#sec:proofjCLT){reference-type="ref" reference="sec:proofjCLT"}. Moreover, the error term obtained therein is not sharp enough for the present purposes. To establish Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"}, we need to be much more careful with the error terms all throughout the proof and this requires entirely new estimates, all obtained in Section [7](#sec:proofjMLLT){reference-type="ref" reference="sec:proofjMLLT"}. A consequence of Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"} is The following MLLT for \(\widehat\Psi_n\) will be shown (in Section [7](#sec:proofjMLLT){reference-type="ref" reference="sec:proofjMLLT"}) from Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"}. The joint LLD we shall need is The proof of this result, given in Section [\[sec:pfkey\]](#sec:pfkey){reference-type="ref" reference="sec:pfkey"} is a more or less obvious adaptation of the proof of  with the additional complication that \(\widehat\Psi_n,\tau_n\) are nonlattice valued. As already mentioned in the introduction, this is the only result of the current paper that does not require any novelty. # Proof of MLLT for the Sinai flow (Proposition [\[prop:MLLTkappa0\]](#prop:MLLTkappa0){reference-type="ref" reference="prop:MLLTkappa0"} and Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"}) {#sec:proofMLLT} In this section we assume \(d=1\) or \(d=2\). In this section, we complete the proof of Theorem [\[thm:main\]](#thm:main){reference-type="ref" reference="thm:main"} by stating and proving the following result (assuming, for the moment, the statement of the results stated in Section [3](#sec:MLLT+LLD){reference-type="ref" reference="sec:MLLT+LLD"}). Using Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"} we complete ## Proof of Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"} Recall that \(A=\phi_I(A_0)\) and \(B=\phi_J(B_0)\) with \(A_0,B_0\subset M\) such that \(\mu(\partial A_0)=\mu(\partial B_0)=0\) and \(I,J\subset\mathbb R\) two bounded intervals. We start by proving the lemma for \(w_t\in\mathbb Z^d\) and \(K=0\). We follow a decomposition somewhat similar of , see also  and , with the obvious difference that one needs to figure out how to exploit the Joint LLT [\[lem:jointllt\]](#lem:jointllt){reference-type="ref" reference="lem:jointllt"} and the Joint LLD [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"}. Writing \(\mathbf{x}=\phi_u(x)\) with \((x,u)\in\widehat{\mathcal M}\), we use the product structure of the measure \(\widehat \nu\) and partition the set considering the different values taken by \(N_t\): \[\begin{aligned} \label{sumQn} \nu\left( A\cap\{\phi_t \in B,\, \kappa_{N_t}=w_t\} \right) = \frac 1{\mu(\tau)}\sum_{n\ge 0} \int_I Q_n(t,u)\, du\,, \end{aligned}\] where \[\begin{aligned} Q_n(t,u)&:=\mu\left(A_0\cap T^{-n} B_0\cap\{ \kappa_n =w_t,\, \widetilde\tau_n\in u+t-n{\mu(\tau)}-J\}\right)\\ &=\mu\left(A_0\cap T^{-n}(B_0)\cap\left\{ \widehat\Psi_n \in (w _t, t-n{\mu(\tau)})+ \{0\}\times J_u\right\}\right)\,, \end{aligned}\] with \(J_u=u-J\), recalling that \(\widehat\Psi_n=(\kappa_n,\widetilde\tau_n)\). For \(L\) large, we split the sum as \[\nu\left(A\cap\{\phi_t \in B,\, \kappa_{N_t}=w_t\} \right)= S_1(t,L)+S_2(t,L)\,,\] where \[S_1(t,L):=\frac 1{\mu(\tau)}\sum_{n\,:\,|n-t/{\mu(\tau)}|\le La_t}\int_I Q_n(t,u)\,du\,,\] \[S_2(t,L):=\frac 1{\mu(\tau)}\sum_{n\,:\,|n-t/{\mu(\tau)}|> La_t}\int_I Q_n(t,u)\,du\,.\] The main ingredient needed for the Proof of Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"} is The proof of Lemma [\[lem:S1S2\]](#lem:S1S2){reference-type="ref" reference="lem:S1S2"} is provided in the paragraph [4.1.1](#sub:llllllll){reference-type="ref" reference="sub:llllllll"} below. Equipped with the statement of Lemma [\[lem:S1S2\]](#lem:S1S2){reference-type="ref" reference="lem:S1S2"} we can complete ### Proof of Lemma [\[lem:S1S2\]](#lem:S1S2){reference-type="ref" reference="lem:S1S2"} {#sub:llllllll} For the proof of Lemma [\[lem:S1S2\]](#lem:S1S2){reference-type="ref" reference="lem:S1S2"}(b), note that \(Q_n(t,u)\le |I|{\widetilde Q}_n(t)\) where \[{\widetilde Q}_n(t)=\sup_{u\in I} \mu\left(\widehat\Psi_n\in (w_t,t-n{\mu(\tau)})+\{0\}\times J_u\right).\] Lemma [\[lem:S1S2\]](#lem:S1S2){reference-type="ref" reference="lem:S1S2"}(b) is an immediate consequence of the next two sublemmas. # Proof of mixing for the Lorentz gas (Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"}) {#sec:proofmixing} In this section we assume \(d=1\) or \(d=2\). Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"} states mixing for functions with compact support in the suspension \(\widehat{\mathcal{M}}\times\mathbb Z^d\). To deal with the natural class of functions (with compact support in the manifold) in Theorem [\[cor:mixingrate\]](#cor:mixingrate){reference-type="ref" reference="cor:mixingrate"} we crucially rely on the following tightness-type result, which is the most delicate part of this work. Before proceeding to the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}, let us see how Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} follows from Corollary [\[cor:easymix\]](#cor:easymix){reference-type="ref" reference="cor:easymix"} and Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}. The rest of this section is devoted to the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}. Recall that \(K_0\) is fixed and that we have to estimate \(\widetilde\nu(B_{R_0}\cap \Phi_{-t}(B_0))\). The strategy of our proof is divided in two steps. In a first step (corresponding to Subsection [5.2](#Step1){reference-type="ref" reference="Step1"}), we explain how we can neglect \"bad\" configurations with first or last long free flights (Lemma [\[lem:B0\]](#lem:B0){reference-type="ref" reference="lem:B0"}) or having a small number of collisions within the time interval \([0,t]\) (Lemmas [\[lem:B0\'\]](#lem:B0'){reference-type="ref" reference="lem:B0'"} and [\[B1second\]](#B1second){reference-type="ref" reference="B1second"}). In a second step (corresponding to Subsection [5.3](#Step2){reference-type="ref" reference="Step2"}), we will estimate the probability of the set of \"good\" configurations belonging to \(B_{R_0}\cap\Phi_{-t}(B_0)\) by writing (as in the proof of Proposition [\[LLTkappaNt\]](#LLTkappaNt){reference-type="ref" reference="LLTkappaNt"} but with additional sums and complications) this set as a union of sets the measure of which corresponds to the measure of a set that can be in terms of the Sinai billiard map \(T\). In the process of proving Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"}, we obtain the following large deviation result, which is interesting in its own right. The bound \(\mathcal O(\log t/t)\) seems to be optimal because \(\tau\) is in the domain of non standard CLT with normalization \(\sqrt{t\log t}\). Although, a proof of the optimal bound seems to be lacking in *the i.i.d. scenario* with \(\sqrt{t\log t}\) normalization, one could form the idea about the optimal bound by redoing several steps in . ## Notations and recalls for the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} {#notationtightness} Before entering deeper in the proof, let us introduce some needed notations. Recall that \(\kappa\) stands for the cell change (with values in \(\mathbb Z^d\)). It will be useful to consider \(\widetilde\kappa:M\rightarrow\mathbb Z^2\) for the cell-change for the \(\mathbb Z^2\)-periodic Lorentz gas; so that \(\kappa=\pi_d(\widetilde\kappa)\) where \(\pi_2=Id\) and \(\pi_1:\mathbb R^2\rightarrow \mathbb R\) is the canonical projection on the first coordinate. Note that, when \(d=2\), \(\widetilde\kappa=\kappa\). We extend the definition of \(\widetilde \kappa\) to \(\widetilde{\mathcal M}\) by setting \(\widetilde \kappa(\Phi_u(q,\vec v))=\widetilde \kappa(p_d(q),\vec v)\) for every \((q,\vec v)\in \widetilde M\) and every \(u\in[0,\tau(x))\). Let us write \(\widetilde N_t(\mathbf x)\) for the number of collisions in the time interval \((0,t]\) for a trajectory starting from \(\mathbf x\in\widetilde{\mathcal M}\). Recall that \(H_0\) is the bounded coboundary defined in [\[coboundPsi\]](#coboundPsi){reference-type="eqref" reference="coboundPsi"}. Throughout the rest of this section we fix \(c_1\) so that \[\label{eq:defc1} c_1\in(0,1/(1000\mu(\tau)))\text{ and } 2c_1\Vert H_0\Vert_\infty<1/100.\] We consider the constant \(a_0\) appearing in Lemma [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"}. Up to decreasing if necessary its value, it follows from e.g. that there exists \(C'_0>0\) such that \[\begin{aligned} \label{Deco2} \forall n'\in\mathbb N_0,\quad Cov\left(f((\kappa\circ T^m)_{m\ge n'}),g((\kappa\circ T^{m'})_{m'\le 0})\right)\le C'_0\Vert f\Vert_\infty\Vert g\Vert_\infty e^{-a_0n'}\,, \end{aligned}\] for any bounded measurable functions \(f,g\). by noticing that \(f((\kappa\circ T^m)_{m\ge 0})\) is constant on stable curves and that \(g((\kappa\circ T^{m'})_{m'\le-1})\) is constant on unstable curves and as such, these functions are bounded by their infinite norms in the respective spaces \(\mathcal H^-\) and \(\mathcal H^+\) considered in.\ We fix \(K>0\) be so that \[\begin{aligned} \label{Deco1} C_0'e^{-a_0 K\log t}\le t^{-100} \end{aligned}\] We recall that it is proved in  that \[\label{tailkappa} \mu(\widetilde\kappa=z)=\mathcal O(|z|^{-3})\,,\] and that the set \(\mathfrak C\) of unit vectors of \(\mathbb R^2\) corresponding to the corridor directions in \(\mathcal D_2\) (i.e. the direction of a line in \(\mathbb R^2\) touching no obstacle) is finite. Finally, recall that by , \[\label{controltimelog} \forall V>0,\quad \mu\left(\widetilde\kappa=z, \exists |j|\le V\log (|z|+2),\ j\ne 0,\ |\widetilde\kappa|\circ T^j>|z|^{4/5}\right)=\mathcal O\left(|z|^{-3-\frac 2{45}}\right)\.\] ## Control of \"bad\" configurations {#Step1} The first next lemma allows us to neglect trajectories with no collision before time \(t\). The next lemma ensures that we can neglect trajectories with long first or long last free flight. The lemma below deals with the reamining range, namely \(n=N_t\le c_1 t\) with \(c_1\) as in [\[eq:defc1\]](#eq:defc1){reference-type="eqref" reference="eq:defc1"}. To prove this lemma, we will deal separately with the cases \(d=1\) and \(d=2\). The main ingredients of the proofs below in these two cases come down to a very delicate decomposition of the involved sum along with fine estimates via the use of [\[controltimelog\]](#controltimelog){reference-type="eqref" reference="controltimelog"} and of Lemma [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"}. The proof of Lemma [\[B1second\]](#B1second){reference-type="ref" reference="B1second"} for \(d=1\) will use the following intermediate results. Lemma [\[B1second\]](#B1second){reference-type="ref" reference="B1second"} in the case \(d=1\) follows from the following result. We take the line below to quickly complete We continue with We are back to the proof of Lemma [\[B1second\]](#B1second){reference-type="ref" reference="B1second"} assuming that \(d=2\). We will decompose the quantity we have to estimate in \(p_{t,1}+p_{t,2}\), distinguishing the case where the free flight at time \(t/2\) is larger or smaller than \(t/4\). **Estimate when the free flight at time \(t/2\) is larger than \(t/4\).** In this part, we study \[p_{t,1}:=\widetilde\nu\left(B_0\cap\Phi_{-t}(B_0)\cap\{\widetilde N_t\le c_1t,\tau\circ \Phi_{\frac t2}>t/4\}\right)\,.\] We will use the fact that we can neglect the trajectories such that \(\tau\circ \Phi_{t/2}>a_t^3\) since, due to Lemma [\[lem:B0\]](#lem:B0){reference-type="ref" reference="lem:B0"}, \[\label{B0t/2} \widetilde\nu(B_0\cap\{\tau\circ \Phi_{t/2}>a_t^3\})\le (2K_0+1)^2\nu\left(\tau\circ \phi_{t/2}>a_t^3\right)\ll a_t^{-3}\,.\] It follows from [\[B0t/2\]](#B0t/2){reference-type="eqref" reference="B0t/2"} and from Sublemma [\[twolongff\]](#twolongff){reference-type="ref" reference="twolongff"} that \[p_{t,1}=\widetilde\nu\left(B_0\cap\Phi_{-t}(B_0)\cap\left\{\widetilde N_t\le c_1t\right\}\cap\Phi_{-\frac t2}\left(\left\{\tau\in\left[\frac t4,a_t^3\right]\right\}\setminus\widetilde D_t\right)\right)+o(a_t^{-2})\,.\] Let us study the event appearing in the above formula.\ The fact that \(\tau\circ \Phi_{\frac t2}>t/4\) and that \(\Phi_{t/2}\not\in D_t\) implies that the \(K\log t\) free flights just before and just after the one occuring at time \(t/2\) have all length smaller than \(t/(100(K\log t)^2)\) and thus that \[\label{tauKlogt}\tau_{K\log t}\circ \widetilde T\circ\Phi_{t/2}<t/(100(K\log t))\quad\mbox{and}\quad |\tau_{-K\log t}|\circ\Phi_{t/2}<t/(100(K\log t))\,.\] Recall that the configuration is in \(B_0\cap\Phi_{-t}(B_0)\) and satisfies \(\tau\circ\Phi_{t/2}>t/4\). Since \(\widetilde N_t\le c_1t\) implies and since we are in dimension 2, the free flight of length \(t/4\) made at time \(t/2\) has to be canceled by the sum of the other \((\lfloor c_1t\rfloor-1)\) free flights made during the time interval \([0;t]\). So, \[\label{tauc1t} \tau_{-c_1t}\circ\Phi_{t/2}>t/8\quad\mbox{or}\quad (\tau_{c_1t}-\tau)\circ\Phi_{t/2}>t/8\,.\] The combination of Conditions [\[tauKlogt\]](#tauKlogt){reference-type="eqref" reference="tauKlogt"} and [\[tauc1t\]](#tauc1t){reference-type="eqref" reference="tauc1t"} implies that at least one of the two next conditions should holds true \[\tau_{-c_1t}\circ \widetilde T^{-K\log t}\circ\Phi_{t/2}> \tau_{-c_1t}\circ\Phi_{t/2}-\tau_{K\log t}\circ\Phi_{t/2}>11t/100\] or \[\tau_{c_1t}\circ \widetilde T^{K\log t}\circ\Phi_{t/2}> (\tau_{c_1t}-\tau)\circ\Phi_{t/2}-\tau_{K\log t}\circ \widetilde T\circ\Phi_{t/2}>11t/100\,.\] Note that the second condition above corresponds to the first one above one up to composing by \(\Phi_t\) and up to using time reversal. Therefore, \[\begin{aligned} p_{t,1}&\le 2\widetilde\nu\left(B_0\cap\left\{\tau\circ \Phi_{\frac t2}\in\left[\frac t4;a_t^3\right],\ \tau_{c_1t}\circ \widetilde T^{K\log t}\circ\Phi_{t/2}>\frac{11t}{100} \right\}\right) +o(a_t^{-2})\\ &\le 2(2K_0+1)^2\nu\left(\tau\circ \phi_{\frac t2}\in\left[\frac t4;a_t^3\right],\ \tau_{c_1t}\circ T^{K\log t}\circ\phi_{t/2}>\frac{11t}{100} \right) +o(a_t^{-2}) \end{aligned}\] using again the fact that \(B_0\) is made of at most \((2K_0+1)^2\) copies of \(\mathcal M\). Now using \(\phi_{t/2}\) invariance of \(\nu\), we obtain \[\begin{aligned} p_{t,1}&\le 2(2K_0+1)^2\nu\left(\tau\in\left[\frac t4;a_t^3\right],\ \tau_{c_1t}\circ T^{K\log t}>\frac{11t}{100} \right) +o(a_t^{-2})\\ \nonumber&\ll o(a_t^{-2})+ \sum_{a'=t/4}^{a_t^3}\mu\left(\tau>a' \tau_{c_1t}\circ T^{K\log t}>11t/100\right) \\ \nonumber&\ll o(a_t^{-2}) +\sum_{a'=t/4}^{a_t^3}\mu\left(|\kappa|>a'-2\Vert H_0\Vert_\infty,|\kappa|_{c_1t}\circ T^{K\log t}>10t/100 \right)\\ \nonumber&\ll o(a_t^{-2})+ \sum_{a'=t/4}^{a_t^3}\left(\mu\left(|\kappa|>a'-2\Vert H_0\Vert_\infty\right)\mu\left(|\kappa|_{c_1t}>10t/100\right)+\mathcal O(t^{-100})\right)\\ \label{AAA}&\ll o(a_t^{-2})+\sum_{a'=t/4}^{a_t^3}(a')^{-2} \mu\left(|\kappa|_{c_1t}>t/10\right). \end{aligned}\] But, using the second part of Lemma [\[lem:lld\]](#lem:lld){reference-type="ref" reference="lem:lld"}, \[\begin{aligned} \nonumber\mu\left( |\widetilde\kappa|_{c_1t}>t/10\right) &\le\mu\left( \tau_{c_1t}>9t/100\right)\\ \nonumber&\le\mu\left( \tau_{c_1t}>t^3\right)+\mu\left( 9t/100<\tau_{c_1t}\le t^2\right)\\ \nonumber &\ll \frac{\mathbb E_{\mu}[\tau_{c_1 t}]}{t^3}+\sum_{k= 9t/100}^{t^3}\frac{t}{\sqrt{t\log t}}\frac{\log (k-c_1t\mu(\tau))}{1+(k-c_1t\mu(\tau))^2}\\ \nonumber &\ll t^{-2}+\sum_{k\ge t(9/100-c_1\mu(\tau))}\frac{t}{\sqrt{t\log t}}\frac{\log t}{1+k^2}\\ &\ll t^{-2}+ \frac{t}{\sqrt{t\log t}}\frac{\log t}t=\sqrt{\frac{\log t}t}\,. \label{LD} \end{aligned}\] This together with [\[AAA\]](#AAA){reference-type="eqref" reference="AAA"} implies that \[\begin{aligned} \label{BBB} p_{t,1} =o(a_t^{-2})\,. \end{aligned}\] **Estimate when the free flight at time \(t/2\) is smaller that \(t/4\).** Let \[p_{t,2}:=\widetilde\nu\left(B_0\cap\Phi_{-t}(B_0)\cap\{\widetilde N_t\le c_1t,\tau\circ \Phi_{\frac t2}\le t/4\}\right)\,.\] Using again the fact that \(B_0\) is made of at most \((2K_0+1)^2\) copies of \(\mathcal M\) and that \(\phi_{t/2}\) preserves \(\nu\), \[\begin{aligned} p_{t,2}&\le (2K_0+1)^2\nu(N_t\le c_1t,\tau\circ\phi_{t/2}\le t/4)\label{controlpt2}\\ &\le (2K_0+1)^2\nu\left( N_{t/2}\le c_1t,N_{-t/2}\le c_1t, \tau\le t/4\right)\,.\nonumber \end{aligned}\] This together with Sublemma [\[twolongff\]](#twolongff){reference-type="ref" reference="twolongff"} and time reversibility gives that \[\begin{aligned} &\nu(N_t<c_1t,\tau\circ\phi_{t/2}\le t/4)\\ &\le o(a_t^{-2})+ \nu\left( N_{t/2}\le c_1t,N_{-t/2}\le c_1t, \tau\le t/4 ,\min(\tau_{K\log t}-\tau,\tau_{-K\log t})\le t/100\right)\\ &\le o(a_t^{-2})+2 \nu\left( N_{t/2}\le c_1t,\, N_{-t/2}\le c_1t, \, \tau\le t/4, \, \tau_{-K\log t}<t/100\right). \end{aligned}\] Since we also know that \[\tau_{-c_1t}\circ T^{-K\log t}=\tau_{-K\log t-c_1t}-\tau_{-K\log t} >\tau_{-c_1t}+\tau-\tau-\tau_{-K\log t}>\frac t2-\frac t4-\frac t{100}\,.\] we obtain \[\begin{aligned} &\nu(N_t\le c_1t,\tau\circ\phi_{t/2}\le t/4)\le o(a_t^{-2})+2 \nu( N_{t/2}\le c_1t,\, \tau_{-c_1t}\circ T^{-K\log t}>24t/100,\tau\le t/4). \end{aligned}\] Let \[A_{a',t'}= \{\tau\ge a'-4\Vert H_0\Vert_\infty, \tau_{c_1t}-a' > 48t/100\}\,.\] Using again the representation by a suspension flow and the decorrelation estimate [\[Deco2\]](#Deco2){reference-type="eqref" reference="Deco2"} combined with [\[Deco1\]](#Deco1){reference-type="eqref" reference="Deco1"}, \[\begin{aligned} &\nu(N_t\le c_1t,\tau\circ\phi_{t/2}\le t/4)\\ &\le \sum_{a'=0}^{\lfloor t/4\rfloor}\mu\left(\tau\ge a',\tau_{c_1t}-a'\ge t/2,\tau_{-c_1t}\circ T^{-K\log t}> 24t/100\right)\\ &\le \sum_{a'=0}^{\lfloor t/4\rfloor}\mu\left(|\kappa|\ge a'-2\Vert H_0\Vert_\infty,|\kappa|_{c_1t}-a'\ge 49t/100,|\kappa|_{-c_1t}\circ T^{-K\log t}> 23t/100\right)\\ &\le \sum_{a'=0}^{\lfloor t/4\rfloor}\left(\mu\left(A_{a',t}\right)\mu\left(\tau_{-c_1t}> 22t/100\right)+\mathcal O(t^{-100})\right)\\ &\le \sum_{a'=0}^{\lfloor t/4\rfloor}\mu\left(A_{a',t}\right)\sqrt{\frac{\log t}t}\,, \end{aligned}\] where in the last line we have used [\[LD\]](#LD){reference-type="eqref" reference="LD"}. The previous displayed estimate together with the argument used in the proof of [\[intermediatedim1b\]](#intermediatedim1b){reference-type="eqref" reference="intermediatedim1b"} gives that \[\begin{aligned} \nu(N_t\le c_1t,\tau\circ\phi_{t/2}\le t/4)&\le o(a_t^{-2})+\nu(N_{48t/100+4\Vert H_0\Vert_\infty}\le c_1 t)\sqrt{\frac{\log t}t}\\ &= o(a_t^{-2})\,, \end{aligned}\] which together with [\[controlpt2\]](#controlpt2){reference-type="eqref" reference="controlpt2"} ensures that \(p_{t,2}=o(a_t^{-2})\), which combined with [\[BBB\]](#BBB){reference-type="eqref" reference="BBB"} ends the proof of Lemma [\[B1second\]](#B1second){reference-type="eqref" reference="B1second"} in the case \(d=2\).  Since \(\mathfrak C\) is finite, it is enough to fix \(\vec w_1,\vec w_2\) and to prove that \[\begin{aligned} \label{eq:sh0} \lim_{R_0\rightarrow +\infty}\limsup_{t\rightarrow +\infty}a_t^d\sum_{n=\lfloor c_1t\rfloor }^{\lfloor t/\min\tau\rfloor}\sum_{a= R_0}^{\lfloor a_t^d\log R_0\rfloor}\sum_{b= 0}^{\lfloor a_t^d\log R_0\rfloor} \mu(A'_{a,b,n,t}(\vec w_1,\vec w_2,K'))=0\,. \end{aligned}\] The aimed result [\[eq:sh0\]](#eq:sh0){reference-type="eqref" reference="eq:sh0"} will be proved via the next series of technical lemmas, splitting the summation over \(n,a, b\) in smaller ranges. ### Concluding the proof of Theorem [\[lem:tight\]](#lem:tight){reference-type="ref" reference="lem:tight"} The conclusion follows from Lemmas [\[lem:B0\]](#lem:B0){reference-type="ref" reference="lem:B0"} (and the comment thereafter), [\[B1second\]](#B1second){reference-type="ref" reference="B1second"}, [\[lem:decomp\]](#lem:decomp){reference-type="ref" reference="lem:decomp"} and [\[B1\'\]](#B1'){reference-type="ref" reference="B1'"}. # Proof of joint CLT (Lemma [\[lem:clt\]](#lem:clt){reference-type="ref" reference="lem:clt"}) {#sec:proofjCLT} Let \(d\in\{0,1,2\}\). In this section we show that arguments established in  and  can be adapted to the study of \(\widehat\Psi\) instead of \(\kappa\). The main idea comes down to a basic observation, namely that \(\widehat\Psi\) can be written as the sum of a vector in \(\mathbb{Z}^{d+1}\) that 'behaves like' \(\kappa\) and a bounded function. The mentioned vector in \(\mathbb{Z}^{d+1}\) is precisely \(\left(\kappa,|\widetilde\kappa|-\mathbb E_{\mu}[|\widetilde\kappa|]\right)\) which, as \(\kappa\) is constant on good set and has a similar tail probability. In particular, \(\widehat\Psi\) is in the domain of a nonstandard CLT with normalization \(\sqrt{n\log n}\). For details are provided around equation [\[eq:rewr\]](#eq:rewr){reference-type="eqref" reference="eq:rewr"}. We will prove the convergence in distribution of \((\widehat \Psi_n/\sqrt{n})_n\) by establishing the pointwise convergence of its characteristic function, with the use of Fourier perturbed operator on the quotient tower constructed by Young in (see ) as Szász and Varjú did in  to establish the CLT and LLT for \(\kappa\). It follows from [\[coboundPsi\]](#coboundPsi){reference-type="eqref" reference="coboundPsi"} that \(\tau=|V|=|\widetilde \kappa|+\mathcal O(1)\), where \(\widetilde\kappa:M\rightarrow \mathbb Z^2\) is the cell change in the \(\mathbb Z^2\)-periodic Lorentz gas (see Subsection [5.1](#notationtightness){reference-type="ref" reference="notationtightness"}).\ We have already recalled several properties of \(\kappa\). Let us recall, in particular, the precise tail of \(\kappa\) (this is partially recalled in [\[tailkappa\]](#tailkappa){reference-type="eqref" reference="tailkappa"}). By completed by, there exists \(L_0>0\) and a finite set \(\mathcal E\) made of \((L,w)\in(\mathbb Z^d)^2\) with \(w\) prime such that \[\begin{aligned} \label{eq:t1} |\widetilde\kappa|>L_0\quad\Rightarrow \quad \exists (L,w)\in\mathcal E,\ \exists N\in\mathbb N^*\ \widetilde\kappa=L+Nw \end{aligned}\] and \[\begin{aligned} \label{eq:tail} \mu(\widetilde\kappa=L+Nw)=c_{L,w}N^{-3}+\mathcal O(N^{-4})\, ,\quad\mbox{as }N\rightarrow +\infty\,, \end{aligned}\] with \(c_{L,w}>0\). This set \(\mathcal E\) parametrizes the set of corridors mentionned in Section [5](#sec:proofmixing){reference-type="ref" reference="sec:proofmixing"} (the set \(\mathfrak C\) therein corresponds to the set of unit vectors proportional to some \(w\) such that there exists \(L\in\mathbb Z^2\) such that \((L,w)\in\mathcal E\)). Then, when \(d=2\), the variance matrix \(\Sigma_0\) (for the Sinai billiard map) appearing in the Central Limit Theorem for the displacement given by [\[eq:cltV\]](#eq:cltV){reference-type="eqref" reference="eq:cltV"} corresponds to the following quadratic form \[\label{defSigma0dim2} \mbox{If }d=2,\quad \forall t\in\mathbb R^2,\quad \langle\Sigma_0 t,t\rangle:=\frac 12\sum_{(L,w)\in\mathcal E}c_{L,w}\langle t,w\rangle^2\,.\] It is not degenerate since, when \(d=2\), we assume the existence of at least two non parallel corridors, and so of two non parallel \(w,w'\) such that there exists \(L,L'\in\mathbb Z^2\) such that \((L,w),(L',w')\in\mathcal E\). When \(d=1\), setting \(\pi_1(w_1,w_2)=w_1\), \(\Sigma_0\) is given by the formula \[\label{defSigma0dim1} \mbox{If }d=1,\quad \Sigma_0:=\frac 12 \sum_{(L,w)\in\mathcal E}c_{L,w}(\pi_1(w))^2\] which is non null since we assumed the existence of at least an unbounded line touching no obstacle. We recall that the variance matrix \(\Sigma\) for the flow appearing in [\[eq:Sigma\]](#eq:Sigma){reference-type="eqref" reference="eq:Sigma"} is given by \(\Sigma=\Sigma_0/\sqrt{\mu(\tau)}\).\ The variance matrix \(\Sigma_{d+1}\) of the limit of \(a_n^{-1}\widehat\Psi_n\) will appear to be given by the following pretty similar formula: \[\label{Sigmad+1} \forall t\in\mathbb R^{d+1},\quad \langle\Sigma_{d+1} t,t\rangle:=\frac 12\sum_{(L,w)\in\mathcal E}c_{L,w}\langle t,(\pi_d(w),|w|)\rangle^2\,,\] with, as in Section [5.1](#notationtightness){reference-type="ref" reference="notationtightness"}, \[\forall w\in\mathbb Z^2,\quad \pi_2(w)=w\quad \mbox{and}\quad \pi_1(w_1,w_2)=w_1\, ,\] and with the convention \[\forall (w,z)\in\mathbb Z^2\times\mathbb R,\quad(\pi_0(w),|w|)=w\quad\mbox{and more generally}\quad (\pi_0(w),z)=z.\] Throughout this section, we fix some (arbitrary) \(q\in [1,2)\), and some \(b_q>2\) so that \[\label{defbq} \frac{1}{b_q}+\frac 1q<1\,.\] This choice will determine the choice of the Banach space on the Young tower. ## Expression of the characteristic functions via Fourier Perturbed operator We observe that \[|{\widehat \Psi}(x)-{\widehat \Psi}(y)|\le d(x,y)+d(T(x),T(y))\, ,\] for any \(x,y\) in the same connected component of \(M\setminus (\mathcal S_0\cup T^{-1}(\mathcal S_0))\), where \(\mathcal S_0\) is the set of post-collisional vectors tangent to \(\partial\Omega\). We recall that the diameter of the connected components of \(M\setminus \bigcup_{k=-n}^nT^{-k}(\mathcal S_0)\) is \(\mathcal O(\beta_1^n)\) for some \(\beta_1\in(0,1)\). As in , we consider the towers constructed by Young in (see also ). We recall some facts on Young towers and introduce some notation that we shall use in the reminder of this paper. We let \((\Delta,f_\Delta,\mu_{\Delta})\) be the hyperbolic tower, which is an extension of \((M,T,\mu)\) by \(\pi:\Delta\mapsto M\) (with \(\pi(x,\ell)=T^\ell(x)\)) and write \(\left(\overline\Delta,f_{\overline\Delta},\mu_{\overline\Delta}\right)\) for the quotient tower (obtained from \(\Delta\) by quotienting out the stable manifolds). The quotient tower is identified with \(\overline\Delta:=\{(x,\ell)\in\Delta\, :\, x\in\overline Y\}\), where \(\overline Y\) is an unstable curve of a well chosen set \(Y\subset M\), and write \(\overline\pi:\Delta\rightarrow\overline\Delta\) for the projection corresponding to the the holonomy along the stable curves of \(\Delta\). The dynamical system \((\Delta,f_\Delta,\mu_{\Delta})\) is given by - The space \(\Delta\) is the set of couples \((x,\ell)\in Y\times\mathbb N_0\) such that \(\ell<R(x)\), where \(R\) is a return time to \(Y\). - The map \(f_\Delta\) is given by \(f_\Delta(x,\ell)=(x,\ell+1)\) if \(\ell<R(x)-1\) and \(f_\Delta(x,R(x)-1)=(f^{R(x)}(x),0)\). - The measure \(\mu_{\overline\Delta}\) is given by \(\mu(A\times\{\ell\})=\mu(A\cap \{R>\ell\})/\mathbb E_\mu[R.1_Y]\), for any measurable set \(A\subset Y\). We assume that the greatest common divisor (g.c.d.) of \(R\) is \(1\), which can be done because of total ergodicity[^6] of \(T\); this assumption is not essential, since one can also deal directly with \(g.c.d.(R)\ne 1\), but it helps simplifying the proofs and notation throughout the reminder of this paper. The partition \(\mathcal P\) on \(\Delta\) consists of a union of partitions of the different levels which become finer and finer as one goes up in the tower. The partition \(\mathcal P\) is used to define a separation time \(s(\cdot,\cdot)\) on \(\Delta\): \[s(x,y):=\inf\{n\ge-1\, :\, \mathcal P(f_\Delta^{n+1}(x))\ne \mathcal P(f_\Delta^{n+1}(y))\}\,,\] The separation time \(s(x,y)\) satisfies the following property: \(\pi(x)\) and \(\pi(y)\) are in the same connected component of \(M\setminus (\mathcal S_0\cup T^{-1}(\mathcal S_0))\) if \(s(x,y)\ge 0\). In particular if \(s(x,y)>2n\), then \(\pi(f_\Delta^n(x))\) and \(\pi(f_\Delta^n(y))\) are in the same connected component of \(M\setminus\bigcup_{k=-n}^nT^{-k}(\mathcal S_0)\). Since the atoms of the partition \(\mathcal P\) are unions of stable curves, this separation time has a direct correspondent \(\overline s(\cdot,\cdot)\) on the quotient tower \(\overline\Delta\). Let \(P\) be the transfer operator of \(\left(\overline\Delta,f_{\overline\Delta},\mu_{\overline\Delta}\right)\), i.e. \(P\) is defined on \(L^1(\mu_{\overline\Delta})\) by \[\int_{\overline\Delta}H.P(G)\, d\mu_{\overline\Delta}=\int_{\overline\Delta}H\circ f_{\overline\Delta}. G\, d\mu_{\overline\Delta}\,.\] Let \(\beta\in(\beta_1^{\frac 14},1)\) and close enough to 1. It follows from  and  that there exists \(\varepsilon'>0\) such that, for all \(\varepsilon\in]0,\varepsilon'[\), \(P\) is quasicompact on the Banach space \(\mathcal B=\mathcal B_{\varepsilon}\) corresponding to the set of functions of the form \(e^{\varepsilon\omega}H\), with \(H\in\mathcal B_0\), where \(\omega(x,\ell)=\ell\) and where \(\mathcal B_0\) is the Banach space of bounded functions \(H:\Delta\rightarrow \mathbb C\) that are Lipschitz continuous with respect to the ultrametric \(\beta^{\overline s(\cdot,\cdot)}\) (the space \(\mathcal B_0\) corresponds to the space \(\mathcal B_\varepsilon\) when \(\varepsilon=0\)). The space \(\mathcal B\) is then endowed with the norm \(\Vert\cdot\Vert_{\mathcal B}\) given by \[\label{eq:Bo} \Vert H\Vert_{\mathcal B}=\Vert e^{-\varepsilon\omega}H\Vert_{\mathcal B_0}\,.\] Recall \(b_q\) satisfies [\[defbq\]](#defbq){reference-type="eqref" reference="defbq"}. Choose \(\varepsilon\) so that \(e^{\varepsilon \omega}\in \mathbb L^{b_q}(\mu_{\overline\Delta})\) which implies that \(\mathcal B\) is continuously embedded in \({\mathbb L}^{b_q}(\mu_{\overline\Delta})\) since \[\label{fixb0} \Vert H\Vert_{L^{b_q}}\le \Vert e^{\varepsilon\omega}\Vert_{L^{b_q}}\Vert e^{-\varepsilon\omega}H\Vert_{\infty}\le\Vert e^{\varepsilon\omega}\Vert_{L^{b_q}}\Vert H\Vert_{\mathcal B} \,.\] (This particular choice of \(b_q\) will be used in the proof of Sublemma [\[sunl:pi\]](#sunl:pi){reference-type="ref" reference="sunl:pi"} below.) Since we assume that \(g.c.d.(R)=1\), 1 is the only (dominating) eigenvalue of modulus 1 of \(P\), and it is simple and isolated in the spectrum of \(P\). In particular, there exists \(\theta\in(0,1)\) (depending on \((\beta,\varepsilon)\)) such that \[\label{quasicompactness} \Vert P^n-\mathbb E_{\mu_{\overline\Delta}}[\cdot]1_{\overline\Delta}\Vert_{\mathcal L(\mathcal B)}=\mathcal O(\theta^n)\, ,\quad\mbox{as }n\rightarrow +\infty\,.\] Let \(t\in\mathbb R^{d+1}\). Recall that via , \[\begin{aligned} \label{coboundPsibis} {\widehat \Psi}\circ\pi=\overline\Psi\circ\overline\pi+\chi\circ f_\Delta-\chi\,, \end{aligned}\] with \(\overline\Psi=(\overline\kappa,\overline\tau):\overline\Delta\rightarrow \mathbb Z^d\times\mathbb R\), where \(\overline\tau\) is the version of \(\widetilde\tau=\tau-\mu(\tau)\) on \(\overline\Delta\) given by \[\overline\tau:=\widetilde\tau\circ\pi +\sum_{n\ge 1}\left(\tau\circ\pi\circ f_{\Delta}^n-\tau\circ\pi\circ f_{\Delta}^{n-1}\circ f_{\overline\Delta}\right)\] and where \(\chi=(\mathbf{0},\chi_0)\) with \(\mathbf 0\) the null element of \(\mathbb Z^d\) and with \[\chi_0:=\sum_{n\ge 0}\left(\tau\circ\pi\circ f_\Delta^n-\tau\circ\pi\circ f_\Delta^{n}\circ \overline\pi\right)\,.\] By  (see also Section [\[sec:app\]](#sec:app){reference-type="ref" reference="sec:app"}) \(\overline\tau\) is locally Lipschitz continuous (on each atom of Young's partition) with respect to the ultrametric \(\beta^{s(\cdot,\cdot)}\), with \(\chi:\Delta\rightarrow \{0\}^d\times\mathbb R\) bounded and Lipschitz in the following sense: \[\label{regbartau} \sup_{k\ge 1} \sup_{x,y:s(x,y)>2k}\frac{|\chi(f_\Delta^k(x))-\chi(f_\Delta^k(y))|}{\beta^k}< \infty\,,\] It follows from the coboundary equation [\[coboundPsibis\]](#coboundPsibis){reference-type="eqref" reference="coboundPsibis"} that \[\mathbb E_\mu[e^{i\langle t,\frac{\widehat \Psi_n}{\sqrt{n}}\rangle}] = \mathbb E_{\mu_{\Delta}}\left[e^{-i\langle \frac{t}{\sqrt{n}},\chi\rangle} e^{i\langle \frac{t}{\sqrt{n}},\overline\Psi_n\rangle\circ\overline\pi}e^{i\langle \frac{t}{\sqrt{n}},\chi\circ f_\Delta^n\rangle}\right]\,.\] Let \(\overline K:\overline\Delta\rightarrow\mathbb Z^2\) be the version of \(\widetilde\kappa\) on \(\overline\Delta\), i.e. the function such that \(\overline K\circ\overline\pi=\widetilde\kappa\circ\pi\). It follows from [\[coboundPsi\]](#coboundPsi){reference-type="eqref" reference="coboundPsi"} and [\[coboundPsibis\]](#coboundPsibis){reference-type="eqref" reference="coboundPsibis"} that the function \(\Theta:\overline\Delta\to\mathbb R^{d+1}\) defined by \[\begin{aligned} \label{eq:rewr} \Theta:=\overline\Psi-\Upsilon ,\quad\mbox{with } \Upsilon:=\left(\pi_d(\overline K),|\overline K|-\mathbb E_{\mu_{\overline \Delta}}[|\overline K|]\right)\,, \end{aligned}\] is bounded and Lipschitz, and \(\Upsilon\) is constant on partition elements (as \(\overline\kappa=\pi_d(\overline{K})\) corresponding to the cell change). Since both \(\overline\Psi\) and \(\Upsilon\) have mean zero, \(\mathbb E_{\mu_{\overline \Delta}}(\Theta)=0\). We define the Fourier-perturbed operators \(P_t,\widetilde P_t\in\mathcal L(\mathcal B)\) by \(P_tv=P(e^{i\langle t,\overline\Psi\rangle} v),\widetilde P_tv=P(e^{i\langle t,\overline\Upsilon\rangle} v)\) for \(t\in\mathbb{R}^{d+1}\). By  (which exploits ), up to enlarging the value of \(\theta\in(0,1)\) appearing in [\[quasicompactness\]](#quasicompactness){reference-type="eqref" reference="quasicompactness"}, there exist \(\beta_0\in(0,\pi]\), a continuous function \(t\mapsto \lambda_t\in\mathbb C\) and two families of operators \((\Pi_t)_t\) and \((U_t)_t\) acting on \(\mathcal B\) such that \(t\mapsto \Pi_t\in\mathcal L(\mathcal B,L^1(\overline\Delta))\) is continuous and such that, for every \(t\in[-\beta_0,\beta_0]^{d+1}\) and every positive integer \(n\), \[\label{spgap-Sz} P_t^n=\lambda_t^n\Pi_t+U_t^n\, ,\quad \widetilde P_t^n=\widetilde \lambda_t^n\widetilde\Pi_t+\widetilde U_t^n\,,\] \[\label{spgap-Sz-bis} \mbox{with}\quad\quad\quad\quad\sup_{t\in[-\beta_0,\beta_0]^d}\left(\Vert U_t^n\Vert_{\mathcal B}+\Vert \widetilde U_t^n\Vert_{\mathcal B}\right) =\mathcal O\left(\theta^n\right)\, ,\quad\mbox{as }n\rightarrow +\infty\,.\] Set \(k=k_n:=(\log n)^2\) and \(F_{k,u}(x):=e^{i\langle u,\overline\Psi_k(\overline\pi(x))+\chi\circ f_\Delta^k(x)\rangle}\). Observe that \[\begin{aligned} \mathbb E_\mu[e^{i\langle t,\frac{\widehat \Psi_n}{\sqrt{n}}\rangle}] &=\mathbb E_{\mu_{\Delta}}\left[e^{-i\langle \frac{t}{\sqrt{n}},\chi\circ f_\Delta^k\rangle} e^{i\langle t,\overline\Psi_n\rangle\circ\overline\pi\circ f_\Delta^k}e^{i\langle \frac{t}{\sqrt{n}},\chi\circ f_\Delta^{n+k}\rangle}\right]\\ &=\mathbb E_{\mu_{\Delta}}\left[F_{k,-\frac{t}{\sqrt{n}}} e^{i\langle \frac{t}{\sqrt{n}},\overline\Psi_n\rangle\circ\overline\pi} F_ {k,\frac{t}{\sqrt{n}}}\circ f_\Delta^n\right]\,. \end{aligned}\] We approximate \(F_{k,u}(x)\) by its conditional expectation \(\widehat F_{k,u}(x)=\overline F_{k,u}(\overline\pi(x))\) on the set \(\{y\in\Delta\, :\, s(x,y)>2k\}\), where \(s\) is the separation time on \(\Delta\) as recalled earlier in this section. Since \(e^{i\langle u,\overline\Psi\rangle}\) is bounded and Lipschitz on \(\overline\Delta\) and since \(e^{i\langle t,\chi\rangle}\) is bounded and Lipschitz on \(\Delta\) (in the sense of [\[regbartau\]](#regbartau){reference-type="eqref" reference="regbartau"}), it follows that \[\begin{aligned} \mathbb E_\mu[e^{i\langle t,\frac{\widehat \Psi_n}{\sqrt{n}}\rangle}] &=\mathbb E_{\mu_{\overline\Delta}}\left[\overline F_{k,-\frac{t}{\sqrt{n}}} e^{i\langle \frac{t}{\sqrt{n}},\overline\Psi_n\rangle} \overline F_ {k,\frac{t}{\sqrt{n}}}\circ f_{\overline\Delta}^n\right]+\mathcal O(\beta^k)\\ &=\mathbb E_{\mu_{\overline\Delta}}\left[\overline F_{k,\frac{t}{\sqrt{n}}} P_{\frac{t}{\sqrt{n}}}^n( \overline F_ {k,-\frac{t}{\sqrt{n}}})\right]+\mathcal O(\beta^k)\\ &=\lambda_t^{n-2k}\mathbb E_{\mu_{\overline\Delta}}\left[\overline F_{k,t} \Pi_{\frac{t}{\sqrt{n}}}( P_{\frac{t}{\sqrt{n}}}^{2k}(\overline F_ {k,-\frac{t}{\sqrt{n}}}))\right]+\mathcal O(\beta^k+\theta^n)\,, \end{aligned}\] as \(k,n\rightarrow +\infty\). Furthermore \(\Vert \overline F_ {k,u}\Vert_{\infty}\le 1\) and \(P_{\frac{t}{\sqrt{n}}}^{2k}(\overline F_ {k,-\frac{t}{\sqrt{n}}})\) are uniformly (in \(k,n\)) Lipschitz with respect to Young's ultrametric \(\beta^{s(\cdot,\cdot)}\). Thus by continuity of \(t\mapsto \Pi_t\in\mathcal L(\mathcal B,L^1(\overline\Delta))\), \[\begin{aligned} \mathbb E_\mu[e^{i\langle t,\frac{\widehat \Psi_n}{\sqrt{n}}\rangle}] =\lambda_{\frac{t}{\sqrt{n}}}^{n-2k}\left(\mathbb E_{\mu_{\overline\Delta}}\left[\overline F_{k,\frac{t}{\sqrt{n}}}\right] \mathbb E_{\mu_{\overline\Delta}} \left[P_{\frac{t}{\sqrt{n}}}^{2k}(\overline F_ {k,-\frac{t}{\sqrt{n}}}))\right]+o(1)\right)\,. \end{aligned}\] Observe that, for every \(t\in\mathbb R^{d+1}\), \[\begin{aligned} \mathbb E_{\mu_{\overline\Delta}}\left[\overline F_{k,\frac{t}{\sqrt{n}}}\right] &=\mathbb E_{\mu_\Delta}\left[e^{i\langle \frac{t}{\sqrt{n}},\overline\Psi_k\circ\overline\pi+\chi\circ f_\Delta^k\rangle}\right]=1+ o(1)\quad \mbox{as }n\rightarrow +\infty\,, \end{aligned}\] due to the dominated convergence theorem (using the fact that \(\lim_{k\rightarrow +\infty}\overline\Psi_k/k=0\) \(\mu_{\overline\Delta}\)-almost-surely). Analogously \[\forall t\in\mathbb R^{d+1},\quad \mathbb E_{\mu_{\overline\Delta}} \left[P_{\frac{t}{\sqrt{n}}}^{2k}(\overline F_ {k,-\frac{t}{\sqrt{n}}}))\right] = \mathbb E_{\mu_\Delta}\left[e^{i\langle \frac{t}{\sqrt{n}},\overline\Psi_k\circ\overline\pi-\chi\rangle\circ f_\Delta^k}\right]=1+ o(1)\, ,\quad\mbox{as }n\rightarrow +\infty\,.\] Thus \[\begin{aligned} \label{eq:Four} \mathbb E_\mu[e^{i\langle t,\frac{\widehat \Psi_n}{\sqrt{n}}\rangle}] =\lambda_{\frac{t}{\sqrt{n}}}^{n-2k}+o(1)\,. \end{aligned}\] An important observation that will allow us to adapt the results of  to the present context is that \[P_t-\widetilde P_t=P\left(\left(e^{i\langle t,\overline\Psi\rangle}-e^{i\langle t,\overline\Upsilon\rangle}\right)\cdot\right) =P_t\left(\left(e^{i\langle t,\overline\Theta\rangle}-1\right)\cdot\right)\,.\] ## Regularity of the dominating eigenvalues and its spectral projector In this part, we prove that for \(q\in[1,2)\) chosen before [\[defbq\]](#defbq){reference-type="eqref" reference="defbq"}, \[\Vert \Pi_t-\Pi_0\Vert_{\mathcal B\rightarrow L^q(\mu_{\overline\Delta})}\quad \mbox{and}\quad \lambda_t=1-\log(1/|t|)\langle t,\Sigma_{d+1} t\rangle+\mathcal O(t^2)\,,\] as \(t\rightarrow 0\). We do so, via the following several steps: we first establish in Sublemma [\[sub:asl\]](#sub:asl){reference-type="ref" reference="sub:asl"} an equivalent of \(\lambda_t-1\), and we use it to establish in Sublemma [\[sunl:pi\]](#sunl:pi){reference-type="ref" reference="sunl:pi"} the announced estimate of \(\Vert \Pi_t-\Pi_0\Vert_{\mathcal B\rightarrow L^q(\mu_{\overline\Delta})}\) that we finally use to establish in Sublemma [\[sunl:lamb\]](#sunl:lamb){reference-type="ref" reference="sunl:lamb"} the announced expansion of \(\lambda_t\). To obtain such estimates, we will control the error between \(\lambda_t\) and \(\widetilde\lambda_t\), and between \(\Pi_t\) and \(\widetilde\Pi_t\). Here we crucially exploit that \(\Upsilon\) and \(\kappa\)satisfy similar properties. This allows us to adapt some results obtained for \(\kappa\) in  with the use of . We start by studying \(P_t-\widetilde P_t\). Observe that \((e^{i\langle t,\overline\Theta\rangle}-1)\cdot\in\mathcal L(\mathcal{B})\) is dominated by \(\left\Vert e^{i\langle t,\overline\Theta\rangle}-1\right\Vert_{\mathcal B_0}\). This implies that \(\left\Vert\widetilde P_t-P_t\right\Vert_{\mathcal L(\mathcal{B})}=\mathcal O(|t|)\), and thus that \[\label{Pit-tildePit} \left\Vert \Pi_t-\widetilde\Pi_t\right\Vert_{\mathcal B}=\mathcal O(t)\,.\] In particular, the announced estimate on \(\Vert \Pi_t-\Pi_0\Vert_{\mathcal B\rightarrow L^q(\mu_{\overline\Delta})}\) will follow from the same estimate for \(\Vert \widetilde\Pi_t-\widetilde\Pi_0\Vert_{\mathcal B\rightarrow L^q(\mu_{\overline\Delta})}\). Lemma [\[lem:clt\]](#lem:clt){reference-type="ref" reference="lem:clt"} follows immediately from [\[eq:Four\]](#eq:Four){reference-type="eqref" reference="eq:Four"}, combined with the continuity of \(t\mapsto\Pi_t\in\mathcal L(\mathcal B\rightarrow L^1(\mu_{\overline\Delta}))\) and from the first sublemma below. In the reminder of this section, we prove a stronger version of Sublemma [\[sub:asl\]](#sub:asl){reference-type="ref" reference="sub:asl"}, along with a strong continuity estimate on \(\Pi_t\), that will be essential in the proofs of Lemmas [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"} and [\[lem:jointllt\]](#lem:jointllt){reference-type="ref" reference="lem:jointllt"} carried out in Section [7](#sec:proofjMLLT){reference-type="ref" reference="sec:proofjMLLT"}. Recall that \(q\in[1,2)\) has been fixed at the beginning of the present section. # Proofs of joint MLLT (Lemmas [\[lem:jointllt0\]](#lem:jointllt0){reference-type="ref" reference="lem:jointllt0"} and [\[lem:jointllt\]](#lem:jointllt){reference-type="ref" reference="lem:jointllt"}) {#sec:proofjMLLT} Let \(d\in\{0,1,2\}\). Compared to CLT, a specific property required to prove the MLLT is the non-arithmeticity (or minimality), which is treated in the next lemma. [^1]: In this article, the strong convergence in distribution means the convergence in distribution with respect to any probability measure absolutely continuous with respect to the Lebesgue measure. [^2]: The notation \(\Longrightarrow \mathcal N(0,C)\) means the strong convergence in distribution to a gaussian random variable of distribution \(\mathcal N(0,C)\), that is centered with variance matrix \(C\). [^3]: where, if \(d=1\), we identify \(\mathbb Z^1\) with \(\mathbb Z\times\{0\}\), meaning that for any \(q'\in\mathcal D_1\) and any \(\ell\in\mathbb Z^1\), the notation \(q'+\ell\) means \(q''+(\ell,0)\) [^4]: Again, in this formula, if \(d=1\), the notation \(w_t+K\) means \((w_t,0)+K\) and \(\mathbb Z^d\) means \(\mathbb Z\times\{0\}\). [^5]: We use here the classical formula \(\mathbb E_\mu[X]=\int_0^{+\infty}\mu(X>z)\, dz\) valid for any positive measurable \(X:M\rightarrow[,+\infty)\). [^6]: The idea of using the total ergodicity of \(T\) for constructing a new tower with \(g.c.d.(R)=1\) was suggested in  and used in  for ensuring aperiodicity of the version of \(\kappa\) on \(\overline\Delta\). The details of such a tower construction are contained in. [^7]: More precisely, see *Estimate of \(\lambda_t\) (there) in the proof of* . In particular, see

{'timestamp': '2023-02-09T02:17:45', 'yymm': '2302', 'arxiv_id': '2302.04254', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04254'}

# Introduction Over the last few years, several new results have addressed and attempted to resolve a long-standing paradox about information and black holes. One important tool utilised in these recent developments is the Ryu-Takayanagi prescription. In essence, it exploits the AdS/CFT correspondence enabling to perform complicated entropy computations in the strongly conformal field theory by identifying them with area computations on the \(AdS\) bulk. Boundary conformal field theories or BCFTs are \(d\)-dimensional strongly coupled conformal field theories on manifolds with boundary. The boundary, a \(d-1\) manifold, is usually called the defect. Using holography once, we arrive at the so-called "intermediate picture\", in which the boundary is dual to \(AdS_d\) gravity localised in a Karch-Randall (KR) brane coupled to a non-gravitating bath where our CFT lives in. A second application of holography yields the full picture in which the brane is an end-of-the-world brane embedded in an \(AdS_{d+1}\) space. The CFT now lives in the boundary of this bulk. These double descriptions are usually called "doubly holographic" and throughout this note we will adhere to this convention. In the past we have used this doubly holographic model to study the phase structure of the entropy in a subregion of a non-gravitating bath. This was itself an extension to the case where the bath was gravitating, forming a wedge-holographic model. In this note, however, we will keep the bath non-gravitating as in the previous study. In other words, we will consider a black string state in an \(AdS_{d+1}\), in which we embed a KR brane. The advantage of this model is that it allows us to place a black hole in the background of the CFT, because the bulk theory is dual to a CFT thermal state in a non-gravitating eternal black hole background. On the other hand, the black string horizon is connected to the horizon of the black hole in the brane and the bath. The main motivation behind this note is to continue analysing the time evolution of the entropy of a subregion of the bath in relation to the Page curve. The Page curve describes the unitary evolution one should expect entropy to follow according to our understanding of quantum mechanics. In particular, we have previously observed the emergence of a Page curve from the bulk perspective by comparing two competing minimal area surfaces or *RT surfaces* at a classical level. The early time entropy is generally controlled by a horizon penetrating surface, namely the *HM surface* which under some circumstances is then replaced by a lower area surface called the *island surface* whose area is constant. When this happens at some finite time, we say that the evolution of the entropy follows a Page curve. The latter type of surfaces anchor on the brane, enclosing the so-called *islands* on it. The whole surface is the union of the island surface and its partner on the thermofield double side. These have been previously studied in, and reviewed in It is clear that whenever these island surfaces exist and depending on their relative area to the HM surface, we can either have constant for all times or initially rising Page curves. However, when such surfaces don't exist this story breaks and we lose any hope of describing evolution as unitary. This is for example the case of empty \(AdS_{d+1}\) where the existence of such surfaces depends on a single parameter called the *critical angle*. This problem with unitarity is resolved if we include a black string in the bulk. As explained above, in this case we found that an island surface always exists. This existence is independent of the two parameters in the theory, which are the anchoring point in the CFT, namely \(\Gamma\) and the brane angle \(\theta_b\). Therefore, one could classify the two types of entropy curves (i.e. constant-for-all-times or initially rising Page curve) in a phase diagram by studying the area difference between the island surface and the HM surface. This phase diagram is driven by the Page angle, which is defined as the angle for which the island surface and the HM surface have the same area. The critical angle still appears in this picture, since below the critical angle, island surfaces exist only above some anchoring point in the brane, called the critical anchor. Similar entropy phase structures have also been studied analytically in lower dimensions. Following these results, one may wonder if they are still true in the case where one includes dynamical gravity on the brane. In lower dimensions this was achieved by including JT gravity on the brane. In higher dimensions we can mimic this by including an Einstein-Hilbert term in the brane action similar to Dvali-Gabadadze-Porrati (DGP) gravity in an \(AdS\) background; this was already done for empty \(AdS\), for topological black holes and for the wedge black string model. This opens a new parameter that our theory depends on, extending our previous result. This begs the question of whether the parameter space in this family of theories is limited. These limitations may arise in the same way as we observed in empty \(AdS\), where the lack of RT-surfaces meant that entropy couldn't possibly evolve unitarily for some brane angles. Therefore, our first task will be to classify the structure of RT surfaces and compute their areas. In doing this we will not only address the unitarity problem, but also check when we have a Page curve. In particular, assert that for some part of parameter space, the wedge holographic model does reproduce the Page curve, while maintaining the massless graviton. Although we do not study this model here, we refer the reader to future work. Before we summarise these results let us observe that in all our previous studies, the natural choice of boundary conditions on the brane were Neumann boundary conditions. Without a DGP term these conditions translate to right-angle anchoring of the RT surface on the brane. An important consequence of the boundary conditions in the two-brane scenario is that the only surface which satisfied them on both branes was the horizon. This leads to the trivial result that the entropy between the defect and its thermofield double does not evolve with time (for empty \(AdS\)). Although we do not study the two-gravitating-brane picture in this note, we observe that relaxing the Neumann boundary conditions can enable new RT surfaces for this case which were not previously studied. In this former paper, as well as in the case of a non-gravitating bath, time-dependence came from those surfaces anchoring in the non-gravitating region, which is the case we are going to study in this note. #### Overview of the Paper. In section [2](#Section2){reference-type="ref" reference="Section2"} we present the setup of the system studied here. We remind the reader about double holography and the one-brane black string model. We then proceed to introduce dynamical gravity on the brane at the action level and derive the corresponding equations of motion to find the RT surfaces in section [3](#Section3){reference-type="ref" reference="Section3"}. In section [4](#Section4){reference-type="ref" reference="Section4"} we compute and present the corresponding areas and the phase structure. The note concludes in section [5](#Section5){reference-type="ref" reference="Section5"} with some remarks. # Double Holography and Dynamical Gravity {#Section2} In this section we give a brief overview of the holographic model and the black string setup we will use in this paper, depicted in figure [\[fig:Setup\]](#fig:Setup){reference-type="ref" reference="fig:Setup"}. Since the setup is mostly the same as we have been using in the past few papers, we refer the unfamiliar reader to those papers. The new ingredients in this study will be covered in the latter parts of this section, so the familiar reader may want to skip to that part. ## Doubly holographic model and the black string The main setup we want to study in this note is the KR braneworld model. This model has a good description in terms of three equivalent pictures: - **Boundary:** \(d\)-dimensional CFT with a \((d-1)\)-dimensional defect (\(\text{BCFT}_d\) ). - **Bulk:** Einstein gravity in an asymptotically \(\text{AdS}_{d+1}\) spacetime containing an \(\text{AdS}_d\) Karch-Randall (KR) brane. This KR brane also has dynamical gravity. - **Intermediate:** \(d\)-dimensional CFT coupled to "dynamical\" gravity on the \(\text{AdS}_d\) brane, with transparent boundary conditions between the brane at infinity and a nongravitating \(\text{CFT}_d\) (the *bath*) on half space. One of the most important consequences of the transparent boundary conditions between the brane and the non-gravitating bath is that the graviton on the brane acquires a mass. This can be seen from the boundary description, since these boundary conditions lead to the non-conservation of the stress tensor. This in turn picks up an anomaly dimension, which translates to the mass of the graviton [^1]. The mass of the graviton can be related to the brane angle (i.e. it is roughly quadratic in the brane angle) for small angles and to the \(AdS_{d+1}\) length scale for large angles. This means that most computations of the Page curve in \(d>2\) at a semiclassical level have been performed in models with a massive graviton . Since we are using the same model, this note and all computations and results presented here have the same feature. We remark that in the authors claim they're able to reproduce the Page curve evolution in massless gravity. Now that we have seen the general idea of coupling a gravitating region to a non-gravitating bath and the three equivalent descriptions that holography provides, it is time to include the black string. From the bulk perspective, we start with a \((d+1)\)-dimensional black string environment in which we embed a \(d\)-dimensional KR brane. Setting the \(AdS_{d+1}\) radius to \(1\) we can describe the bulk geometry with the metric, \[\begin{split} ds^2 &= \frac{1}{u^2 \sin^2\mu}\left[-h(u) dt^2 + \frac{du^2}{h(u)} + d\vec{x}^2 + u^2 d\mu^2\right],\\ h(u) &= 1-\frac{u^{d-1}}{u_h^{d-1}}, \end{split}\label{eq:BS}\] where \(t \in \mathbb{R}\), \(u > 0\), \(0 < \mu < \pi\), and \(\vec{x} \in \mathbb{R}^{d-2}\). These coordinates slice the \(AdS_{d+1}\) space into \(AdS_d\) subspaces, one for each constant \(\mu\). The black string induces a black hole in each slice, one of which is the KR brane, i.e. the \(\mu = \theta_b\) slice. In particular, by the \(AdS/CFT\) correspondence, this setup is a gravitating black-hole coupled to a non-gravitating black-hole background in the bath. Typically, the black string features a thermodynamic instability, called the Gregory-Laflamme instability. In global \(AdS\) this instability is controlled by a combination of two of the parameters in the theory, the \(AdS\) radius and the horizon distance \(u_h\). Considering large enough black strings (compared to the horizon distance) we avoid this instability. For the case at hand, by considering the Poincaré patch we automatically avoid this, i.e. we only consider large black holes. Small black holes will be in considered in. We are interested in finding the entanglement entropy between a subregion \(\mathcal{R}\) (the radiation region) in the bath and its complement \(\bar{\mathcal{R}}\) at a fixed time slice (which we choose to be \(t=0\)). We will use the usual RT prescription in which we first find all extremal surfaces \(\Sigma\) satisfying the homology constraint \[\partial\Sigma = \partial \mathcal{R} \cup \partial \mathcal{I}.\] The second boundary term in the homology constraint, written in terms of the island \(\mathcal{I}\), allows for surfaces to cross the brane into the other KR braneworld. ## Area functional Now we wish to explicitly find the surfaces satisfying the constraint above. As we will see shortly, and at the level of finding the RTs, the DGP term will affect the boundary condition so we will start by not writing it explicitly. The final step will be to extremise the set of these surfaces and pick the minimal area one. Looking at the metric, [\[eq:BS\]](#eq:BS){reference-type="eqref" reference="eq:BS"}, we can exploit the translational symmetry along \(x_i\) to parameterise the potential RT surfaces as \(u(\mu)\). Substituting this into our metric, for a constant \(t=0\) slice, we can write an area density functional \[\mathcal{A} = \int_{\theta_b}^{\pi} \frac{d\mu}{(u\sin\mu)^{d-1}} \sqrt{u^2 + \frac{u'(\mu)^2}{h(u)}}, \label{eq:AF}\] We observe that \(u_h\) can be written as an overall factor of \(u_h^{2-d}\) in the integral above, meaning that it scales out of the problem. Thus, we can without loss of generality set \(u_h=1\). Furthermore, we remark that we will be, mostly, omitting the word "density\" throughout this note Our objective now is to extremise this functional and if the set of such extremising surfaces contains more than one element, to pick the minimal area one. This is done by solving the Euler-Lagrange equation (i.e. the equation of motion) and imposing appropriate boundary conditions. This equation can be written explicitly as \[\begin{split} u'' =-(d-2) u\,h(u) + (d-1) u' \cot\mu \left(1-\frac{\tan\mu}{2} \frac{u'}{u\,h(u)} + \frac{u'^2}{u^2\,h(u)}\right)-\left(\frac{d-5}{2}\right) \frac{u'^2}{u}. \end{split} \label{eq:eomODE}\] Until now, this is apparently the same as we have done in the previous study, so let's see what changes now. ## DGP term The area of the RT surface to compute entanglement entropy arises from evaluating the Ricci scalar on a solution with a conical defect, required by the replica trick. Therefore, following, by introducing a Ricci scalar (via an Einstein-Hilbert term) on the brane, namely induced by the bulk gravity, we can get an additional contribution to the entanglement entropy, called the DGP term. The full action on the brane will thus be, \[S =-T \int d^d x \sqrt{-\tilde g} + \frac{1}{16 \pi G_b} \int d^d x \sqrt{-\tilde g} \tilde R\] where the \(\tilde g\) is the induced metric on the brane, \(\tilde R\) is the corresponding Ricci scalar, \(T\) is the brane's tension [^2] and \(G_b\) is the Newton's constant on the brane which parameterises the strength of gravity there (i.e. the DGP term). With this, the entanglement entropy computed from the area of the RT surface is \[S_{EE} = \frac{1}{2} \left( \frac{2}{4G} \mathcal{A} + \frac{1}{4G_b} \mathcal{A}_b \right) = \frac{1}{8G} (2\mathcal{A} + \lambda_b \mathcal{A}_b) \label{eq:SEE}\] where \(\mathcal{A}\) is calculated in [\[eq:AF\]](#eq:AF){reference-type="eqref" reference="eq:AF"} and \(\lambda_b = \frac{G}{G_b}\) [^3] [^4]. There is also a factor of \(2\), which comes from the fact that we have two copies of spacetime in our KR braneworld. The factor of \(\frac{1}{2}\) comes from the \(\mathbb{Z}_2\) orbifolding, although this is a conventional choice. The latter is the independent parameter that we are going to be free to tune and it is the main new ingredient in this note. The term \(\mathcal{A}_b\) is the area density of the point where the RT surface intersects the brane, it is thus given by \[\mathcal{A}_b = \left. \frac{1}{(u \sin \mu)^{d-2}} \right|_{u=u_b,\mu = \theta_b} \label{eq:AF2}\] where \(u_b\) is the \(u\)-coordinate where the brane intersects the RT surface. This is the last piece that we need in order to derive the correct boundary conditions for the problem at hand, so let's proceed to this now. The equation of motion [\[eq:eomODE\]](#eq:eomODE){reference-type="eqref" reference="eq:eomODE"} will not be affected by the boundary term \(\mathcal{A}_b\). In this way, abbreviating the equation of motion by \(EOM\) and extremising \[\begin{split} 0 \equiv & 2 \delta \mathcal{A} + \lambda_b \delta \mathcal{A}_b=-\int_{\theta_b}^\pi d\mu (\delta u) (EOM) \\ & +\left.\frac{2 \delta u}{(u \sin \mu)^{d-1}}\frac{u'}{h(u) \sqrt{u^2 + \frac{(u')^2}{h(u)}}}\right |_{\theta_b}^{\pi} + \left. \frac{(2-d)\lambda_b \delta u}{u^{(d-1)} (\sin \mu)^{d-2}} \right|_{u=u_b,\mu = \theta_b} \end{split}\] For Neumann boundary conditions on the brane we must set the second line to zero, evaluated at the boundary on the brane, namely \[u' = \left. \pm (2-d) \frac{u \, \lambda_b \, h(u) \, \sin \mu}{\sqrt{4-(d-2)^2 h(u) \, \sin^2\mu \, \lambda^2}} \right|_{u=u_b,\mu = \theta_b} \label{eq:DGPBC}\] If we further assume \(d>2\), and since \(0<\mu< \pi/2\) then the sign of \(u'\) is fixed to be the same as the sign of \(-\lambda_b\), by requiring that it indeed solves the Neumann boundary conditions. We observe that the square root in the denominator of [\[eq:DGPBC\]](#eq:DGPBC){reference-type="eqref" reference="eq:DGPBC"} means that \(\lambda_b\) can only take values in \((-\frac{2}{(d-2)\sin \theta_b \sqrt{h(u_b)}},\frac{2}{(d-2)\sin \theta_b \sqrt{h(u_b)}})\), for which \(u'\) can take any value, for fixed \(\theta_b\) and \(u_b\). For the rest of this note we will fix \(d=4\) [^5]. Following the discussion in we can also write the induced Newton's constant on the brane as \[\frac{1}{G_{eff}} = \frac{1}{G} (1+\lambda_b)\] This means that at the special value of \(\lambda_b =-1\) gravity on the brane becomes infinitely strong and we should be careful when analysing the results. Furthermore, although we will analyse all values of \(\lambda_b\), if \(\lambda_b \leq-1\), the effective Newton's constant \(G_{eff}\) is negative and so the theory is non-physical. As we can see from the equation of motion [\[eomODE\]]{#eomODE label="eomODE"}, solving it analytically is quite difficult so we are going to follow the same approach as in and solve it numerically, using a shooting method. We will now proceed to present some results and therefore (in addition to fixing \(d=4\)) we fix \(u_h=1\). The latter will not affect any result, since this is an overall scale in the functional which doesn't affect the equations of motion. ### The \(\lambda\)-parameter Geometrically we can think of the boundary condition given by \(u’\), as the angle at which the RT surface intersects the brane. For instance, when \(u’(\theta_b)=0\) we recover our old right-angle condition, for example if \(\lambda_b=0\). However for other values of \(\lambda_b\) the situation is more complicated since, by looking at [\[eq:DGPBC\]](#eq:DGPBC){reference-type="eqref" reference="eq:DGPBC"}, we also have to consider \(u_b\) and \(h(u_b)\). For example, if we fix the shooting point to \(u_b=0\), then \(u'=0\) independently of the value of \(\lambda_b\). We observe that this will not always be possible because the critical anchor \(u_{crit}\) also plays a role in this story. We remind the reader that the *critical anchor* is the lowest point on the brane from which it is not possible to reach the bath, by shooting from it. In other words, RT candidates are defined in the region of the brane \((u_{crit}, u_h)\), which receives the name of *atoll*. For \(\lambda_b = 0\) the critical anchor is zero for a brane angle \(\theta_b \geq \theta_c\), where \(\theta_c\) is the critical angle. For general \(\lambda_b\) the critical anchor is non-zero even above the critical angle and hence in general we are not always able to shoot from \(u_b=0\). On the other hand if we fix \(u_b = u_h\) then equation [\[eq:DGPBC\]](#eq:DGPBC){reference-type="eqref" reference="eq:DGPBC"} implies that \(u'(\theta_b)=0\), for any value of \(\lambda_b\). This means that if we shoot to the bath from the horizon, we will stay at the horizon and we will always reach the bath. If the surface penetrates the horizon and does not reach the bath on the TFD side from which it was launched, its area will represent increasing entropy over time and will never result in a Page curve. Therefore, for the parameter combinations where this occurs, the entropy evolution in our theory will not be unitary and we will reject this theory. In other words, in addition to the reality condition we imposed a priori on \(\lambda_b\), we will further impose that the corresponding RT surfaces reach the bath. Doing this numerically gives us a critical anchor and an atoll for each angle and \(\lambda_b\). Surprisingly the bounds imposed by this are lower than the theoretical reality bounds, as seen in figure [\[fig:Lvsu01\]](#fig:Lvsu01){reference-type="ref" reference="fig:Lvsu01"}. This figure shows how for fixed values of \(\lambda_b\) we get a critical anchor and a corresponding atoll which changes as we change \(\lambda_b\). This generalises our old story of critical anchors, which is now both \(\lambda_b\) and \(\theta_b\) dependent (see figure [\[fig:Lvsu02\]](#fig:Lvsu02){reference-type="ref" reference="fig:Lvsu02"}). Furthermore, we observe that there is a range of \(\lambda_b\) values, for which shooting anywhere on the brane, we can reach the bath. In other words, there is a range where the critical anchor shrinks to zero. As an example, for the case shown in figure [\[fig:Lvsu01\]](#fig:Lvsu01){reference-type="ref" reference="fig:Lvsu01"}, this is \(\lambda_b \in (-0.884,-0.999)\) By performing the computation for multiple values of the brane angle, we can see how the range of values gets wider as the \(\theta_b\) increases, shown in figure [\[fig:Lvsu02\]](#fig:Lvsu02){reference-type="ref" reference="fig:Lvsu02"}. We also notice that as the brane angle increases, the lower bound curves decrease, while the upper bound curves increase. For small values of the anchoring point u, when \(\abs{\lambda_{b}}<1\) this increase (or decrease) is not substantial and so all curves accumulate near \(\lambda_b= \pm 1\). These are somewhat special values, at \(\lambda_b =-1\) the effective Newton's constant on the brane becomes infinite. On the other hand, at \(\lambda_b=1\), the value of the brane and the bulk Newton's constants match numerically, \(G_b =G\) [^6]. In particular, for the angles considered here, we will see that \(\lambda_b=-1\) is also special when we study the RT structure. For \(\theta_b = \theta_{c} \approx 0.98687\) the lower bound curve of the range crosses entirely the \(\lambda_b=0\) axis. This is consistent with previous results in since for angles above the critical angle the atoll should cover the whole brane. Additionally, for angles below the critical angle, the critical anchor monotonically decreases as the angle increases for any fixed \(\lambda_b \geq-0.884\) and in particular for \(\lambda_b=0\). # RT Structure {#Section3} In this section we are going to explore the different types of Ryu-Takayanagi (RT) surfaces we can have for a fixed set of parameters, that is for a fixed anchoring point in the bath \(\Gamma\), a fixed value for \(\lambda_b\) and a fixed brane angle \(\theta_b\). We begin by describing the Hartman-Maldacena (HM) surface, which penetrates the Einstein-Rosen bridge and extends into the thermofield-double side. The other surfaces are going to be different versions of island surfaces, which anchor on the brane and cross into the second copy of spacetime in our KR braneworld. ## HM surface To study the Hartman-Maldacena surface we are going to change the parametrisation of our surface, such that \(\mu = \mu(u)\). In this way the area functional becomes: \[\mathcal{A} = \int \frac{du}{(u \sin\mu)^3} \sqrt{\frac{1}{h(u)}+u^2 \mu'(u)^2} \.\label{eq:AFmu}\] where, since we know the surface doesn't intersect the brane, the DGP term doesn't affect the solutions to the Euler-Lagrange equations of this functional. To determine the boundary conditions at the horizon we can, in general, consider the Hartman-Maldacena surface anchoring at \(\Gamma\) on one side of the thermofield double and \(\tilde \Gamma\) at the other side. For simplicity and following, we will choose the symmetric case \(\Gamma = \tilde \Gamma\). This imposes Dirichlet boundary conditions on the bath as with any other RT surface. The other boundary is at the horizon, so we need to determine what the conditions are here. In order to do so we are going to assume this surface is smooth across the Einstein-Rosen bridge, in other words, and since we assumed the HM anchors at the same distance from the defect on both sides of the TFD, that it has no "kinks\". In adapted (tortoise-like) coordinates, \(dr = \frac{du}{u\sqrt{h}}\), this can be mathematically written as: \(\mu'(r=0)=0\) Following, we can either expand the solution to the equation of motion in this coordinates, or translate to our old parametrisation \(\mu(u)\). If we do the latter, we note that the expansion will contain half integer powers of \((u_h-u)\) [^7]. After doing this, we uniquely fix the solution by choosing the \(\Gamma\) and the angle at which the HM surface crosses the horizon \(\theta_{HM}\). This is equivalent (using our power expansion) as choosing the boundary conditions: \[\mu(u_h \Gamma)=\pi, \hspace{1cm} \mu(u_{h}) = \theta_{HM}, \hspace{1cm} \mu'(u_{h}) = \frac{2 \cot \theta_{HM}}{u_{h}},\label{eq:bdryHM}\] ## Island surfaces In this section we are going to analyse the RT surfaces giving rise to islands, i.e. surfaces which anchor on the bath and the brane. In the doubly holographic models, islands are generally considered to be the regions on the brane going into the horizon, depicted in figure [\[fig:Setup\]](#fig:Setup){reference-type="ref" reference="fig:Setup"}. In previous work we found that, when there was no DGP term, the existence of RT surfaces was driven by a specific angle value, which we called the critical angle \(\theta_c\). For angles \(\theta<\theta_c\), there is a specific value of anchoring point, the critical anchor, on the brane below which we can never reach the bath. Something similar will occur in our setup; for a fixed angle and a fixed DGP coupling \(\lambda_b\), there is a finite range of anchoring points from which we can reach the bath. In other words, we generalise the critical anchor slightly, where now this depends on these two parameters. We already noted this in figure [\[fig:Lvsu02\]](#fig:Lvsu02){reference-type="ref" reference="fig:Lvsu02"}, where we gave the minimum range of values for which we can reach the bath by shooting from anywhere in the brane. As we increase or decrease \(\lambda_b\), we can reach the whole bath only by shooting from a section of the brane, namely the atoll. We can extract the critical anchor information and plot them as a function of \(\theta_b\) for each \(\lambda_b\), like we show in figure [\[fig:CritAnch\]](#fig:CritAnch){reference-type="ref" reference="fig:CritAnch"} for \(\lambda_b>-1\) and figure [\[fig:CritAnch2\]](#fig:CritAnch2){reference-type="ref" reference="fig:CritAnch2"} for \(\lambda_b<-1\). As we can see in figure [\[fig:CritAnch\]](#fig:CritAnch){reference-type="ref" reference="fig:CritAnch"} the critical anchors are monotonically decreasing with increasing brane angle or decreasing \(\lambda_b\). There are two main behaviours we observe in this figure. When \(1<\lambda_b<2\) the curves don't decrease with decreasing slope, but show a small bump. The \(\lambda_b=1\) value is somewhat significant because it is where the Newton's constants on the brane and the bulk numerically match. When \(-1<\lambda_b<1\), we see that the critical anchors, all follow similar curves to our previous results, i.e. \(\lambda_b=0\). In fact, if we think of the critical angle as the angle at which the critical anchor shrinks to the defect \(u_b=0\), then by following the curves in figure [\[fig:CritAnch\]](#fig:CritAnch){reference-type="ref" reference="fig:CritAnch"} to their intersection with the \(\theta_b\)-axis, we can also define a generalised critical angle. As we can see this decreases with increasing \(\lambda_b\). Of course, one may think of the critical angle as determined in empty \(AdS\) and hence there is no possible generalisation in this case. In the next section we are going to take this latter interpretation. When \(\lambda_b \leq-1\), the behaviour of the critical anchors in relation to \(\lambda_b\) is reversed. As seen in figure [\[fig:CritAnch2\]](#fig:CritAnch2){reference-type="ref" reference="fig:CritAnch2"}, decreasing \(\lambda_b\) increases the critical anchor, while the behaviour in relation to the brane angle remains unchanged. In this case, we also notice that the critical anchors seem to have a similar behaviour as when \(1<\lambda_b<2\) with a slope that appears to be small (near zero) for some angles. This produces generalised critical angles above \(\pi/2\) for \(\abs{\lambda_b}>1\). When \(\lambda_b=-1\), i.e. when \(G_{eff}\) becomes infinite making the theory on the brane non-physical, the critical anchors approach \(u_b=0\) asymptotically. Another way of putting the above observations is that the size of the atoll increases when we increase the brane angle or when the value of \(\lambda_b\) approaches \(-1\), from above in figure [\[fig:CritAnch\]](#fig:CritAnch){reference-type="ref" reference="fig:CritAnch"} or from below in figure [\[fig:CritAnch2\]](#fig:CritAnch2){reference-type="ref" reference="fig:CritAnch2"}. If we view the degrees of freedom on the brane as redundant with those on the bath, in the limit when the effective Newton's constant is infinite, i.e. \(G_{eff} \rightarrow \infty\) information on the brane gets delocalised. On the other hand when this ratio has a large absolute value information on the brane is localised near the horizon, i.e. the resulting size of the island is small and is located near the horizon. Furthermore we can study in more detail how the anchoring point on the brane affects the anchoring point on the bath. When doing so we discover an interesting behaviour at \(\lambda_b \leq-1\), which we depict in figure [\[fig:AnchvsGamma\]](#fig:AnchvsGamma){reference-type="ref" reference="fig:AnchvsGamma"}. We distinguish two behaviours. When \(\theta_b \leq 0.825\) there is a one-to-one correspondence between shooting points on the brane, above some critical anchor, and an anchoring point \(\Gamma\) on the bath. This is similar to the situation we observed for other values of \(\lambda_b\). However when we increase the angle past this value, the critical anchor shrinks to the defect and there is a region close to the horizon, for which a single point in the bath may have up to three corresponding RT-candidates. This region is characterised by some \(\Gamma_{min}\) above which these three surfaces exist and below which we have a unique RT candidate. This unique RT candidate is the one which has a smallest shooting point \(u_b\) and which we call *RT-1*. The next RT-candidate, ordering them by shooting point is what we call *RT-2*. Finally the RT-candidate which is shot farthest from the defect receives the name of *RT-3*. This situation is depicted in figure [\[fig:PolarPlot\]](#fig:PolarPlot){reference-type="ref" reference="fig:PolarPlot"}. Moreover, the value of \(\Gamma_{min}\) monotonically decreases with increasing brane angle \(\theta_b\)[^8]. Associated to this value there is a shooting point \(u_{min}\), such that the point \((u_{min}, \Gamma_{min})\) is in the curves shown in figure [\[fig:AnchvsGamma\]](#fig:AnchvsGamma){reference-type="ref" reference="fig:AnchvsGamma"}. These \(u_{min}\) points also decrease with increasing \(\theta_b\). The important thing for the case at hand, is that following the RT prescription, when there is a competition between RT-candidates we need to choose the minimal area one. Thus, if there is some transition in the RT surface between the RT-candidates listed, this occurs for decreasing values of \(\Gamma\). This is in fact what will happen as we will see shortly, once we compute the areas of the RT-candidates. When \(\lambda_b=-1.1\), the same three competing surfaces and the same behaviour is found for angles \(\theta_b \geq 1.225\). When \(\lambda_b =-1.2\) this behaviour is found for angles \(\theta_b \geq 1.4\) and when \(\lambda_b =-1.3\) it is found for \(\theta_b \geq 1.525\). This poses a technical difficulty when computing areas for \(\lambda_b \leq-1\) and angles above \(\pi/2\). Furthermore, as we will see in the next section, the areas generally decrease with \(\lambda_b\) for any given brane angle. Therefore, with the aim of characterising when the areas vanish and producing a phase diagram, we will decide to compute the areas up to a brane angle of \(\pi/2\). # Numerical Results {#Section4} Now that we have given a detailed account of what the possible RT-candidates are we can proceed to compute their areas. As we mentioned above we will have two competing RT surfaces: the island surface and the HM surface. The former has a constant area in time while the latter has an area which grows in time, because of the growing Einstein-Rosen bridge. In this section we compute the area differences between the island surfaces and the Hartman-Maldacena surface \[\Delta A (t) = A_{IS} – A_{HM} (t) \label{eq:DeltaA}\] which is automatically finite in the UV. We perform the computation at \(t=0\), since we know what the time-evolution of each term. We use the results from our previous section about the existence of RT surfaces to find the areas [^9]. Finally, from this data we obtain the entropy phase structure for the present case. The situation where \(\lambda \leq-1\) brings a new challenge, since for certain angles we have several competing surfaces, so we will also need to be careful here to compute the smaller area surface. ## Area difference for fixed parameters Before proceeding we make a small philosophical distinction about the parameters. Although there is no a priori hierarchy between the three parameters and we are free to choose, \(\lambda_b\), \(\theta_b\) and \(\Gamma\), we would like to think about the first two as in some sense fixed before \(\Gamma\). What we mean is that the brane angle and the gravity on the brane are fixed in the particular universe we study. However, the location on the bath where we measure the entropy depends on the location of the observer. In other words, an observer in a particular universe can always move further from the defect to choose another region on the bath to measure the entropy. That observer cannot change the brane angle nor the strength of the gravity on the brane unless that observer changes universes. Although we'll vary all of the parameters in our computations, we are going to do so by fixing \(\lambda_b\) and \(\theta_b\) and plotting the areas as a function of \(\Gamma\). Methodologically, we have decided to compute the areas for angles \(\theta \in (0.4,\pi/2)\) (in steps of \(\theta = 0.025\)) [^10]. The parameter \(\lambda_b\) will run in the range \((-2,2)\), in steps of \(\lambda_b = 0.05\). It's worth noting that this range includes values that correspond to theories that are considered non-physical. However, for the sake of completeness, we will still consider these values in our analysis. As we noted above, the particular value of \(\lambda_b\) may limit the range of shooting points on the brane, i.e. the size of the atoll, but in terms of the position on the bath \(\Gamma\), we will still cover the whole bath, so \(\Gamma \in (0,1)\). ### \(\lambda=-1\) We now present some of the technical difficulties, in particular for the \(\lambda=-1\) case, where the situation is a bit more intricate. We have noticed above that for a fixed \(\Gamma\) we can have up to three competing RT surfaces. We have ordered these according to the shooting point on the brane: RT-1, the one closer to the defect, RT-2, the one in the middle and RT-3, the one closer to the horizon (see figure [\[fig:PolarPlot\]](#fig:PolarPlot){reference-type="ref" reference="fig:PolarPlot"}). In figure [\[fig:AnchvsGamma\]](#fig:AnchvsGamma){reference-type="ref" reference="fig:AnchvsGamma"} we can see that when the angle is \(\theta \leq 0.825\), there is a unique surface for each shooting point on the brane and hence there is no competition. Furthermore, when \(\Gamma<\Gamma_{min}\), there is also a unique surface, so again, this is the smallest area surface. However, when we increase the angle above \(0.85\) two surfaces, RT-2 and RT-3 appear for \(\Gamma>\Gamma_{min}\). Let us analyse each of the areas independently to explain this situation. For RT-1, the area is positive near \(\Gamma=0\) and diverges as \(\Gamma \rightarrow 1\). The shooting point for RT-1 is located on the brane, close to the \(AdS\) boundary, and it is therefore not regulated. This might lead one to expect a positive divergence for this area, but this is compensated by subtracting a similar DGP contribution, in [\[eq:SEE\]](#eq:SEE){reference-type="eqref" reference="eq:SEE"}. When \(\Gamma=1\) the area blows up again, because here the HM has large negative renormalised area. This behaviour was noted in. This competition between the non-boundary area and the DGP term, produces a nearly constant area for every angle, for this surface. For RT-2 and RT-3, we see a similar behaviour, reaching a smallest area at the point where we shoot directly to \(\Gamma_{min}\), since this shooting point is closer to the defect. More precisely, RT-2 and RT-3 have monotonically increasing area for \(\Gamma>\Gamma_{min}\). However, in this range of \(\Gamma\)s, RT-3 has slightly smaller area, only equating RT-2 precisely at \(\Gamma_{min}\), where these surface are degenerate [^11]. This rules out RT-2 as a competing surface. Nevertheless, these surfaces have decreasing area with increasing angle and hence, immediately after appearing they will dominate over RT-1. Both surfaces have the same divergence when \(\Gamma \rightarrow 1\), which we observed in RT-1 and is what we would expect for any of these. Collecting these results, we have the following behaviour in the area for a fixed angle as a function of \(\Gamma\). Below \(\theta_b=0.85\), the area (determined by the area of RT-1) is slightly positive close to the defect and then increases with \(\Gamma\). It diverges again for \(\Gamma\) close to the horizon because of the Hartman-Maldacena surface. When the angle is at or above \(0.85\) this behaviour is maintained at \(\Gamma<\Gamma_{min}\), being RT-1 the dominant surface. At \(\Gamma_{min}\), there is a transition in which RT-3 becomes dominant, with a smaller area and keeps dominating all the way up to the stage where \(\Gamma\) reaches the horizon and the area diverges. This produces a jump from a larger area RT-1, to a smaller area, RT-3 which looks discontinuous. Although here we won't resolve this discontinuity, we suspect that sampling more points in the bath can show a steep but smooth transition between these surfaces. This would produce two points for which \(\Delta \mathcal{A} =0\). However the conservative approach we'll take here is to think of this jump as discontinuous. This is similar to the discontinuity observed in, when the tiny islands take over. As we can see in figure [\[fig:AnchvsGamma\]](#fig:AnchvsGamma){reference-type="ref" reference="fig:AnchvsGamma"}, increasing the angle means that \(\Gamma_{min}\) is decreased and appears for lower shooting points \(u_b\). In other words RT-2 and RT-3 appear "sooner\" and thus this jump also occurs sooner. ### Area difference as a function of \(\lambda_b\) Here we look closer at the numerical results, depicting only a representative subset of the areas computed. These show the main features we observed in our computations. In particular, we can plot the area difference \(\Delta \mathcal{A}\) for \(t=0\) as a function of \(\Gamma\) for several \(\lambda_b\)-values. This produces a 2-D surface for each angle. Below, in figures [\[fig:AvsGvsL1\]](#fig:AvsGvsL1){reference-type="ref" reference="fig:AvsGvsL1"} and [\[fig:AvsGvsL2\]](#fig:AvsGvsL2){reference-type="ref" reference="fig:AvsGvsL2"} we just show a few of these surfaces. As we can see in figures [\[fig:AvsGvsL1\]](#fig:AvsGvsL1){reference-type="ref" reference="fig:AvsGvsL1"} and [\[fig:AvsGvsL2\]](#fig:AvsGvsL2){reference-type="ref" reference="fig:AvsGvsL2"}, the area increases monotonically with \(\lambda_b\) until we reach \(\lambda_b=-1\). Here we observe that the areas near the defect, \(\Gamma=0\), are positive, as we described above and which will be more evident when we present these areas for fixed \(\lambda_b\)s. In figure [\[fig:AvsGvsL1\]](#fig:AvsGvsL1){reference-type="ref" reference="fig:AvsGvsL1"}, for \(\lambda_b\) between \(-1\) and \(0\) we see how the near defect area becomes increasingly negative as we increase the angle. This behaviour is exacerbated until we reach \(\theta_b=1\), for which the negative areas also appear closer to \(\lambda_b=0\). In figure [\[fig:AvsGvsL2\]](#fig:AvsGvsL2){reference-type="ref" reference="fig:AvsGvsL2"}, this trend is maintained even for positive \(\lambda\)s. To explain this we recall that for \(\lambda_b=0\) we recover our results in. Here we observed this divergence, which was caused by infinitesimal surfaces which wrapped the defect, namely the tiny islands. For non-zero values of \(\lambda_b\), we can now have a non-zero critical anchor, even above the critical angle. Hence we don't obtain these tiny islands as limits of any surface, namely we cannot continuously shoot to the defect. However, the critical anchor gets smaller as \(\lambda_b \rightarrow-1\) from below (see figure [\[fig:CritAnch2\]](#fig:CritAnch2){reference-type="ref" reference="fig:CritAnch2"}) and so, when we have large negative \(\lambda_b\) (\(\lambda_b \leq-1\)) the RT-surface has a finite large shooting point \(u_b\) and doesn't diverge to negative infinity. As we increase \(\lambda_b\) the critical anchor sinks into the defect and the tiny islands take over. For small negative values of \(\lambda_b\), we are subtracting a large negative area from an already negative term (tiny islands) by the DGP contribution as given in equation [\[eq:AF2\]](#eq:AF2){reference-type="eqref" reference="eq:AF2"}. Since the range of \(\lambda_b\)s for which the critical anchor shrinks into the defect increases with angle, this produces an increasingly pronounced dip in figures [\[fig:AvsGvsL1\]](#fig:AvsGvsL1){reference-type="ref" reference="fig:AvsGvsL1"} and [\[fig:AvsGvsL2\]](#fig:AvsGvsL2){reference-type="ref" reference="fig:AvsGvsL2"} near \(\Gamma=0\). Something similar occurs for larger values of \(\lambda_b\), where again the appearance of a critical anchor truncates the existence of tiny islands, preventing the area difference to diverge to negative infinity. This behaviour suggests that any definition of the critical angle that depends on the existence of tiny islands, or as mentioned before, on the critical anchor sinking to the defect, will also be affected by the inclusion of the DGP term. However, our focus is on comparing our parameters to the critical angle of empty \(AdS\) without a DGP term, and thus we will not define the critical angle in this manner. Furthermore, away from the \(\Gamma=0\) defect the areas are monotonically increasing with one exception. Although it is less noticeable in the other angles, in figure [\[fig:AvsGvsL2\]](#fig:AvsGvsL2){reference-type="ref" reference="fig:AvsGvsL2"}, for \(\theta_b=1.4\) and \(\lambda_b=-1\), the area monotonically increases until it reaches the \(\Gamma_{min}\) we described before. At this point we have a discontinuity corresponding to the transition between RT-1 and RT-3. Away from the \(\lambda_b=-1\) value the area is finite and monotonically increases with \(\Gamma\) for every angle. In all cases, the area diverges to infinity as \(\Gamma \rightarrow 1\). Above we mentioned that this divergence was caused by the fact that the HM surface has a renormalised area which diverges to \(-\infty\). In other words, the explanation of this divergence comes from the fact that in [\[eq:DeltaA\]](#eq:DeltaA){reference-type="eqref" reference="eq:DeltaA"} we are subtracting a \(-\infty\) term. Another way to see this divergence is using the fact that when \(\Gamma\) gets close to the horizon, the corresponding HM shrinks to a point. This means that it "cuts off\" less of the RT surface which still picks up a divergence. We recall that that a negative area difference \(\Delta \mathcal{A}\), means that the island surface is smaller than the HM surface, which generally grows with time. Therefore, we can summarise these observation by: an observer sitting close to the defect for \(\lambda_b \geq-1\), will (depending on the angle) measure a constant entropy curve. Furthermore, because of the competing surfaces at larger angles, this is also true for an observer sitting closer to the horizon. We remind the reader that exactly at \(\lambda_b=-1\), the effective Newton's constant on the brane, \(G_{eff}\), becomes infinite and hence the theory is considered non-physical. As we increase \(\lambda_b\) the area differences will become positive and the observer will be able to measure a non-constant entropy curve. ### Area difference as a function of \(\theta_b\) To get some more intuition and to show some of the things we described above, we can similarly fix the \(\lambda_b\) value and plot \(\Delta \mathcal{A}\) vs \(\Gamma\) for several angles. In this way each 2D surface corresponds to one \(\lambda_b\)-value. We can see these surfaces in figures [\[fig:AvsvsGvsTh1\]](#fig:AvsvsGvsTh1){reference-type="ref" reference="fig:AvsvsGvsTh1"} and [\[fig:AvsvsGvsTh2\]](#fig:AvsvsGvsTh2){reference-type="ref" reference="fig:AvsvsGvsTh2"} In figure [\[fig:AvsvsGvsTh1\]](#fig:AvsvsGvsTh1){reference-type="ref" reference="fig:AvsvsGvsTh1"} we show some of the surfaces obtained for \(\lambda_b \leq 0\). When \(\lambda_b\) is sufficiently large and negative (panel on the left of the figure) we see that the area is finite independently of the angle [^12]. This is what we observed above and it is because there is a non-zero critical anchor which, even when \(\Gamma=0\), keeps our surface anchored away from the defect. The area monotonically increases with the brane angle \(\theta_b\) and \(\lambda_b\). As before, the area is divergent when \(\Gamma \rightarrow 1\). In the centre figure, the areas are approximately constant with brane angle, but increasing with \(\Gamma\) for most of parameter space. However for larger angles we observe the transition between RT-1 and RT-3. The latter surface has monotonically increasing area with \(\Gamma\), which is also appreciated in this panel. Furthermore, we observe how this transition occurs at decreasing \(\Gamma_{min}\) as we increase the brane angle. We remind the reader that thus far, these theories have \(G_{eff}<0\) and hence can be considered as non-physical. On the right panel of figure [\[fig:AvsvsGvsTh1\]](#fig:AvsvsGvsTh1){reference-type="ref" reference="fig:AvsvsGvsTh1"}, the areas still increase with \(\Gamma\), but the trend with respect to \(\theta_b\) is reversed, namely now we have decreasing areas with brane angle. This behaviour with respect to the two parameters will be maintained for all \(\lambda_b >-1\) in figure [\[fig:AvsvsGvsTh2\]](#fig:AvsvsGvsTh2){reference-type="ref" reference="fig:AvsvsGvsTh2"}. The other new ingredient is the tiny island effect which appears when \(\Gamma \rightarrow 0\) and \(\theta>\theta_{crit}\). The dependence of the critical anchor on \(\lambda_b\) can clearly be seen by the fact that the areas diverge faster to \(-\infty\) for angles larger than \(\theta_c\) as \(\lambda_b\) approaches zero from below. The leftmost panel of figure [\[fig:AvsvsGvsTh2\]](#fig:AvsvsGvsTh2){reference-type="ref" reference="fig:AvsvsGvsTh2"} shows the case where the DGP coupling is turned off, \(\lambda_b=0\) and this matches the results in. Comparing these figures we see how the tiny islands are somewhat tamed by the DGP coupling. However as we can see by comparing the first and third panel of [\[fig:AvsvsGvsTh1\]](#fig:AvsvsGvsTh1){reference-type="ref" reference="fig:AvsvsGvsTh1"} the DGP term has a greater effect on the smaller angles, reversing the monotonic behaviour of the area as a function of brane angle. This is evident from equation [\[eq:AF2\]](#eq:AF2){reference-type="eqref" reference="eq:AF2"} which is proportional to \(1/\sin{\theta_b}^2\) and since the anchors are roughly fixed in a small range, as we would expect for small angles. We will comment on the signs of the areas below. In general, the areas for any \(\lambda_b\) and \(\theta_b\) can be positive or negative so we can ask the question about when the areas are precisely zero, this is what we proceed to do now. ## Phase structure We have presented a detailed account for the area differences as a function of the different parameters in our theory; \(\Gamma\), \(\theta_b\) and \(\lambda_b\). Now that we have observed the different trends we can analyse the entropy phase structure given in this parameter space. We recall that the sign of \(\Delta \mathcal{A}\) will determine what surface dominates at \(t=0\). Since the HM has a linear area evolution with time, a positive \(\Delta \mathcal{A}\) means that the entropy follows a Page curve evolution. However, when \(\Delta \mathcal{A}\) is negative the constant area island surface is the RT-surface whose area is proportional to the entropy and hence the entropy is just constant. Therefore, the question of when \(\Delta \mathcal{A}\) vanishes is a question about when we have an entropy evolution following a Page curve. We remark that both situations satisfy unitarity, since we already discarded the non-unitary theories, that is how we limited the parameter space in the first place. Here we present the main result in this note which is the entropy phase structures diagrams. These are curves showing the values of \(\Gamma\) and \(\theta_b\) for which \(\Delta \mathcal{A} = 0\) for fixed \(\lambda_b\) values. From the plot in figure [\[fig:PhaseStructure\]](#fig:PhaseStructure){reference-type="ref" reference="fig:PhaseStructure"} we observe several different behaviours. When \(\lambda_b\) is between \(-2\) and \(-1\) we observe that the curves in the phase diagram never intersect the \(\theta_b\)-axis. This means that the area never vanishes at the defect (for this range of angles), or in other words there is no proper definition of the Page angle, the angle for which the vanishing area surface has \(\Gamma=0\). In fact, for these few first curves, the region where \(\Delta \mathcal{A}>0\) is a smaller region of parameter space which is also close to the horizon. Since \(\Delta \mathcal{A}<0\) means that there is a flat entropy curve (because the island surface starts dominant at \(t=0\)), this means that for most of the parameter space there is no Page curve. In particular when \(\lambda_b =-2\) or \(\lambda_b =-1.5\), we see the tendency of the curve to drop, but it is still mostly close to the horizon. Again we remark that these theories (when \(\lambda_b \leq-1\)) are unphysical from the start. However in the limit where \(G_{eff} \rightarrow \infty\), or in other words when \(\lambda_b \rightarrow 1\), we see that the competing surfaces described above only allows for vanishing areas in a small region of angles. Therefore, the corresponding \(\lambda_b =-1\) curve (in yellow) is shorter than the other curves. In this case we can imagine a vertical line from the left end of this curve delimiting the \(\Delta \mathcal{A} >0\) region from the \(\Delta \mathcal{A}<0\) region. When \(\lambda_b>-1\), the curves intersect the \(\theta_b\)-axis and roughly half of parameter space will give us a page curve, \(\Delta \mathcal{A}>0\) and half will have a constant entropy evolution. As we increase \(\lambda_b\) the constant entropy region gets smaller. In fact when \(\lambda_b>1\), we didn't observe negative areas and hence there is no vanishing area curves, although we suspect this will not be the case if we look at larger angles. This can also be seen in figure [\[fig:AvsvsGvsTh2\]](#fig:AvsvsGvsTh2){reference-type="ref" reference="fig:AvsvsGvsTh2"}. This means that the corresponding curves (which are not shown in the figure), would just be flat lines sitting on the \(\theta_b\)-axis. We can also reformulate these observations in terms of what, in previous works, we called the constant entropy belt. This is the region of the bath containing anchoring points \(\Gamma\), for which the area is \(\Delta \mathcal{A}<0\). In the past,, this belt had a size which was monotonic with angle and this is what we observe here for \(\lambda_b >-1\). In particular we observe this for what would reproduce our previous results \(\lambda_b=0\). In this way we remark that this figure, directly extends figure 10 in However, when \(\lambda_b=-2\) or \(\lambda_b=-1.5\), this monotonicity is broken. We see the belt first decreases and then increases in both cases. Finally, since some of the curves appear to intersect the \(\theta_b\)-axis, we can extract, by extrapolating these curves, a Page angle for each value of \(\lambda_b\). We now proceed to do this extrapolation. This extrapolation will carry some intrinsic error, but to show the robustness of our method, we observe that for \(\lambda_b=0\) the Page angle we obtain is \(0.972\). This is only marginally different from the value we had in our previous study of \(\theta_P \approx 0.975\). ## Page angles Extracting the Page angle, namely the angle for which the area of the curve shooting to the defect vanishes, for each value of \(\lambda_b\) in figure [\[fig:PhaseStructure\]](#fig:PhaseStructure){reference-type="ref" reference="fig:PhaseStructure"} we obtain the figure [\[fig:Pageangles\]](#fig:Pageangles){reference-type="ref" reference="fig:Pageangles"}. As we noted before, we can check our method by comparing the \(\theta_P\) we obtain now compared to what we obtained before and we observe that this is roughly the same, namely approximately \(0.972\). We can see that the Page angle monotonically increases with the value of \(\lambda_b\). Therefore, there is another interesting quantity we can extract from this plot and this is the value of the DGP coupling for which the Page angle is the critical angle. We recall that in empty \(AdS\) the Page angle is the same as the critical angle. The reason why this is not true for the black string, is because the HM has a non-zero renormalised area and hence subtracts a small number from the island surface area at the critical angle. Thus, the area at the critical angle is slightly negative and hence the value of \(\theta_b\) for which the area vanishes lies slightly below \(\theta_c\). In the case at hand, by fine tuning the strength of the gravity in the brane, namely the DGP coupling parameter, \(\lambda_b\), we can lift the effect of the black string and make the Page angle match the critical angle again. A similar thing has been done in, where the fine tuned parameter is the black hole size. This happens for approximately \(\lambda_b \approx 0.0364\). # Conclusions {#Section5} We have extended the study in analysing the subregion entropy for a doubly holographic black string model. The system is dual to a \(BCFT_d\) with a black-hole background and to a non-gravitating bath coupled to an \(AdS_d\) brane in which dynamical gravity can been turned on. This dynamical gravity is parameterised by a coupling \(\lambda_b\) which measures the strength of gravity on the brane. In this way we are testing the parameter space of the black string model with respect to the anchoring point on the bath, the \(AdS_d\) brane angle and the strength of the DGP term, \((\Gamma, \theta_b, \lambda_b)\). In empty \(AdS\) we previously observed that island surfaces did not exist below the critical angle. This issue was resolved at finite temperature, namely for the \(\lambda_b=0\) system giving us a rich phase structure. However this phase structure could be modified by the inclusion of dynamical gravity. Therefore our aim here was to limit the parameter space for well-behaved theories supporting islands and to study the subregion entropy phases when the strength of gravity on the brane was modified. In particular, throughout this note we focused on unitary evolution of entropy and the existence of a Page curve. In order to do this, we have first proceeded to study the RT structure of the system, namely the existence and number of competing RT-candidates. For most of the parameter space there was a single island RT-candidate for a shooting point on the brane. This shooting point was limited to only a fraction of the brane, characterising the size of the island. This extends our critical anchor story, since now, changing the \(\lambda_b\) parameter allows for non-zero critical anchors even above the critical angle. For the particular case of \(\lambda_b=-1\) we observed 3 competing RT-candidates and steep phase transition between two of these. This phase transition was captured by a special value of anchoring point on the bath \(\Gamma_{min}\). All of this shows a richer RT structure which we needed to take into account when computing their areas. With this in mind we computed the areas and presented them as 2D surfaces. Some of the behaviour they exhibited matched our previous work but some of it was new and needed to be explained. In particular, as mentioned, we highlight the case of \(\lambda=-1\) and how the tiny island effects were lifted by large absolute value strength of DGP coupling. All of this is summarised in figure [\[fig:PhaseStructure\]](#fig:PhaseStructure){reference-type="ref" reference="fig:PhaseStructure"}, which constitutes the main result in this note. Here we distinguish the parts of parameter space which give us a Page curve from those that give us an eternally constant entropy. In any case it seems both situations present a unitary evolution and hence means that the theories in the parameter space are stable against unitarity tests. We can use this information to obtain the value of \(\lambda_b\) for which the critical angle matches the Page angle, which we obtained to be \(\lambda_b \approx 0.0364\). In other words, one can fine tune the strength of gravity on the brane so that the effect of the black string on the renormalised area of HM is lifted and we obtain what we had in the case of empty \(AdS\). The difference is that now, unlike in empty \(AdS\) we do have island surfaces below the critical angle. In recent work it was discussed that the DGP terms may solve the issue of massive gravity in island models. In this note we have seen how the parameter space is confined in order for islands to exist. Something similar might be expected for wedge holographic models were we can also impose other constraints. This is going to be analysed in more detail in future work. This study has focused on the entropy phase structure and RT surfaces for angles \({\theta_b< \pi/2}\). However, the relationship between brane tension and angle indicates that for larger \(\lambda_b\), there may be a range of subcritical brane angles that are still physical above \(\pi/2\). It would be interesting to explore this further for larger angles and values of DGP coupling, and determine how the competition between three RT surfaces is resolved and if any phase transitions occur. # Supporting Plots {#app:A} Here we show some of the plots which were used in part of the computations in the main text. They illustrate some of the descriptions and claims above. Figure [\[fig:Supp1\]](#fig:Supp1){reference-type="ref" reference="fig:Supp1"} shows what we claimed for \(\lambda_b=-1\). This is that \(\Gamma_{min}\) decreases as we increase the brane angle. Figure [\[fig:Supp2\]](#fig:Supp2){reference-type="ref" reference="fig:Supp2"} shows the areas for two of the competing surfaces when \(\lambda_b=-1\) and \(\theta_b=1.35\). The surfaces RT-2 and RT-3 become degenerate at \(\Gamma_{min}\), while the surfaces RT-1 and RT-2 become degenerate at \(\Gamma=1\). Above this value we have three competing RT surfaces: RT-1, RT-2 and RT-3. We can see how RT-2 and RT-3 have increasing area as \(\Gamma \rightarrow 1\), but overall, RT-3 has a slightly smaller area than RT-2 ::: [^1]: It is a standard fact in \(AdS/CFT\) that the anomalous dimension of operators in the boundary appears as a mass term in the \(AdS\) equations. [^2]: This is related to the brane angle via \(T =\frac{(d-1)}{4 \pi G} [\cos \theta_b + \frac{\lambda_b}{2} (d-2) \sin^2 \theta_b]\). The author would like to thank Hao Geng for highlighting this important point. [^3]: We will occasionally omit the subindex and write \(\lambda\) instead of \(\lambda_b\) to simplify notation. [^4]: G is the bulk's Newton's constant. [^5]: This means, when comparing with, that the induced Newton's constant \(G_{RS} = (d-2) G/2\) is the same as the bulk Newton's constant \(G\) and hence their \(\lambda_b\) is the same as our \(\lambda_b\). This is not true in general dimensions (or general radius curvature). [^6]: We remark that these do not have the same dimensions and hence this equality is just a numerical one. [^7]: This is because \(r \sim \sqrt{u_{h}-u}\) [^8]: see figure [\[fig:Supp1\]](#fig:Supp1){reference-type="ref" reference="fig:Supp1"} in appendix [6](#app:A){reference-type="ref" reference="app:A"}. [^9]: Formally we are computing area differences given by [\[eq:DeltaA\]](#eq:DeltaA){reference-type="eqref" reference="eq:DeltaA"}, but in the text we will loosely refer to "areas\". [^10]: We remark that the tensionless branes are generally not at \(\theta_b=\pi/2\), given that the tension vanishes at some angle depending on \(\lambda_b\), which can be obtained from the relation between brane tension and angle quoted above. The choice of this range of angles is based on practical reasons, namely that the region of interest for this note lies in this range, as we will see form the phase diagram. [^11]: See figure [\[fig:Supp2\]](#fig:Supp2){reference-type="ref" reference="fig:Supp2"} in appendix [6](#app:A){reference-type="ref" reference="app:A"} which illustrates this explanation. [^12]: We expect the areas to diverge as the brane angle approaches \(\theta_b \rightarrow 0\), from it becoming the \(AdS\) boundary, but this is outside the range of values computed in this note

{'timestamp': '2023-02-27T02:02:51', 'yymm': '2302', 'arxiv_id': '2302.04279', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04279'}

# Introduction and the Main Results Throughout this paper \(\grp{G}\) denotes a finite group, not necessarily abelian but written additively, and \(\ff{q}\) denotes the finite field of order \(q\). For a finite set \(\set{S}\), \(\# \set{S}\) denotes the cardinality of \(\set{S}\). Let \(f:\grp{G}\to \grp{G}\). The set of distinct images of \(f\) is denoted by \(\im{f}:=\{f(x)\,:\,x\in\grp{G}\}\), and we write \(V(f):=\# \im{f}\). We call \(f\) a *permutation* over \(\grp{G}\) if \(\im{f}=\grp{G}\), i.e. when \(f\) is a bijection over \(\grp{G}\). For \(b\in \grp{G}\), the set of preimages of \(b\) under \(f\) is denoted by \(\preim(f,b):=\{x\in \grp{G}\,:\, f(x)=b\}\). The *uniformity of \(f\)* is defined by \[u(f):=\max_{b\in \grp{G}} \#\preim(f,b).\] For a nonzero \(a\in \grp{G}\), the *differential operator of \(f\) in the direction of \(a\)* is defined by \[\Delta_{f,a}(x):=f(x+a)-f(x),\] and the *differential uniformity (DU) of \(f\)* is defined by \[\delta_{f}:=\max_{a\in \grp{G}\setminus\{0\}}u(\Delta_{f,a}).\] The concept of DU was first suggested by Nyberg. The lower the DU, the more resistant \(f\) is to differential attacks when used in a cryptosystem. Functions with optimal DU are called *planar* when \(\# \grp{G}\) is odd (with \(1\)-DU). A classic example is \(x^2\), which is planar over every field of odd characteristic. When \(\# \grp{G}\) is even, the optimal functions are called *almost perfect nonlinear (APN)* (with \(2\)-DU). One of the most important problems related to DU is the construction of permutations over finite fields with optimal or low DU. These functions are the most desired for the construction of S-boxes in cryptosystems. Planar functions cannot be permutations, so over \(\ff{q}\) when \(q\) is odd the problem becomes that of finding permutations with low DU. For extensions of \(\ff{2}\), while there are several examples of APN permutations over odd dimensions, there is only one known APN permutation in even dimensions, an example in dimension \(6\) discovered by Browning, Dillon, McQuistan, and Wolfe in 2010. For more than a decade, the existence of APN permutations over higher even extensions of \(\ff{2}\) remains open and is considered one of the most important open problems in the theory of APN functions. Motivated by the construction of low DU permutations, the authors introduced the notions of permutation resemblance in. The concept provides a new way to measure the "distance" of a function from being a permutation. For two functions \(f,h:\grp{G}\to \grp{G}\), we first define the *resemblance of \(f\) to \(h\)* by \[\res(f,h)=V(f-h).\] Observe that for any functions \(f,h:\grp{G}\to \grp{G}\), we have \(\res(f,f)=\# \{0\}=1\), \(\res(f,h)=\res(h,f)\), and \(\res(f,h+c)=\res(f,h)\) for any constant \(c\in \grp{G}\). The central concept of this paper is the following. It is important to note that there could be many functions (\(h\) or \(g\) in the above expressions) that give the \(\pres(f)\). Given a non-permutation \(f\), \(\pres(f)\) can be understood as the "minimum" changes required in order to modify \(f\) into a permutation. Unlike other methods for measuring the distance a function is from being a permutation (\(V(f)\) is a common one), permutation resemblance can be used to construct low DU permutations if we start with function \(f\) with low or optimal DU. In, it was proved that if \(f,g\) are two functions over a finite abelian group \(\grp{G}\), then \[\delta_{g+f}\le \delta_{f}\cdot \big(V(g)^{2}-V(g)+1\big).\] In particular, if \(V(g)=\pres(f)\), then \[\delta_{g+f}\le \delta_{f}\cdot \big(\pres(f)^{2}-\pres(f)+1\big).\] Thus, by computing \(\pres(f)\) for an optimal DU function \(f\), and finding a set of such \(g\), we obtain a set of permutations of the form \(g+f\) whose DU is controlled by \(\delta_{f}\) and \(\pres(f)\). This observation motivates the present article. The aim of this article is to present algorithms for computing \(\pres(f)\) for an arbitrary \(f\) and for constructing permutations \(g+f\) that satisfy \(V(g)=\pres(f)\). The key idea is based on rephrasing the problem of determining \(\pres(f)\) into the problem of searching a certain family of subtables in the *subtraction table* indexed by \(\grp{G}\) and \(\im{f}\), see Section [2](#sc:sub_tb){reference-type="ref" reference="sc:sub_tb"} for details. In Section [3](#sc:alg_IP){reference-type="ref" reference="sc:alg_IP"}, we further formulate this new problem as a linear integer programming problem (IP). Though it is not often used in the research of permutations over finite groups, linear programming is known to be a useful tool in extremal combinatorics, a recent example being the paper by Wagner which disproved a number of open conjectures in extremal combinatorics with linear programming methods. By solving the IP, one obtains the exact value of \(\pres(f)\) for any \(f:\grp{G}\to \grp{G}\), and a set of permutations \(g+f\) such that \(V(g)=\pres(f)\). In Section [4](#sc:alg_avg){reference-type="ref" reference="sc:alg_avg"}, we present an algorithm for constructing a feasible solution of this IP. Using this algorithm, we then prove the following upper bounds for \(\pres(f)\) when \(f\) is a two-to-one function. In Section [6](#sc:alg_ex){reference-type="ref" reference="sc:alg_ex"}, we discuss some conditions when a \(k\)-subset of \(\grp{G}\) could be a candidate of the image of \(g:\grp{G}\to \grp{G}\) such that \(g+f\) is a permutation, and give an algorithm to test this. Finally, we close the article by generalizing the IP in Section [3](#sc:alg_IP){reference-type="ref" reference="sc:alg_IP"} into a formulation that can optimize DU. This IP has a large number of variables and constraints, but it can combine both requirements of \(\pres(f)\) and DU, and we believe it has significant potential to be used/adapted to create examples under other optimal measurements. We also give some computational results concerning \(\pres(x^d)\) over \(\ff{q}\) with \(\gcd(d,q-1)>1\) in an appendix. # The subtraction table {#sc:sub_tb} Let \(\# \grp{G}=q\) and \(f:\grp{G}\to \grp{G}\). The *subtraction table of \(f\)* is a table \(M_{f}\) with \(q\) rows and \(V(f)\) columns. The rows of \(M_{f}\) are indexed by the \(q\) elements of \(\grp{G}\), and the columns are indexed by the \(V(f)\) elements of \(\im{f}\). For \(r\in \grp{G}\) and \(c\in \im{f}\), the entry \(m_{r,c}\) of \(M_{f}\) at row \(r\), column \(c\), is defined by \(m_{r,c}:=r-c\). For \(t\in\itr\), \(0\le t\le u(f)\), denote the set of elements of \(\grp{G}\) with exactly \(t\) preimages under \(f\) by \[P_{t}:=\{b\in \grp{G} \,:\, \# \preim(f,b)=t\}.\] In particular, \(P_{0}\) is the set of all non-images of \(f\), and \(\im{f}=\bigcup_{t=1}^{u(f)}P_{t}\). We order the columns of \(M_f\) by \(P_{1},P_{2},\dots,P_{u(f)}\), and the rows by \(P_{1},P_{2},\dots,P_{u(f)}, P_{0}\), where elements within the same set \(P_{t}\) can be listed in arbitrary order but in the same way for both rows and columns. With this ordering, the diagonal of the upper part of \(M_f\) indexed by \(P_{1},P_{2},\dots,P_{u(f)}\) is all \(0\). This table is useful when working with the sum \(g+f\) for some \(g:\grp{G}\to \grp{G}\), since if \(g(x)+f(x)=b\) for some \(x,b\in \grp{G}\), then the value of \(g(x)\) is the value of the corresponding entry \(m_{b,f(x)}\) of \(M_{f}\). For the remaining, we call a collection of entries of \(M_{f}\) a *subtable of \(M_{f}\)*. We do not assume that a subtable is a block, unless otherwise specified. Next, we define a family of subtables of \(M_{f}\) important for our algorithms. Let \(S\) be a \(k\)-subset of \(\grp{G}\), and \(\mathcal{A}\) be a collection of entries of \(M_{f}\). We call \(\mathcal{A}\) an *admissible subtable of \(M_{f}\) with value set \(S\)* if it satisfies the following conditions: - \(S\) is the set of values of the entries in \(\mathcal{A}\). - For every \(c\in \im{f}\), there are exactly \(\# \preim(f,c)\) distinct \(r\) such that \(m_{r,c}\in\mathcal{A}\), i.e., every column \(c\) of \(M_f\) has exactly \(\#\preim(f,c)\) distinct entries in \(\mathcal{A}\). We may simply use the term "admissible subtable" when there is no danger of confusion. Condition (A2) implies that \(\mathcal{A}\) has exactly \(\sum_{c\in\im{f}} \# \preim(f,c)=q\) entries. Note that for (A1) and (A2) to be true, we must have \(k\ge u(f)\) since every element of \(\grp{G}\) appears exactly once in each column of \(M_f\). Admissible subtables are crucial for our algorithms since each of them corresponds to a set of functions with known image sets. This correspondence is presented in the following lemma and its proof. For an admissible subtable \(\mathcal{A}\), we define \[\range(\mathcal{A}):=\{r\in \grp{G}\,:\, m_{r,c}\in\mathcal{A}\}.\] That is, \(\range(\mathcal{A})\) is the set of indices of the rows of \(M_{f}\) which have at least one entry in \(\mathcal{A}\). As shown in the proof of Lemma [\[lm:subtb_correspondence\]](#lm:subtb_correspondence){reference-type="ref" reference="lm:subtb_correspondence"}, \(\range(\mathcal{A})\) is equal to \(\im{g+f}\) for any corresponding \(g\). In particular, if \(\range(\mathcal{A})=\grp{G}\), then any corresponding \(g\) satisfies \(\im{g+f}=\grp{G}\) and therefore \(g+f\) is a permutation. This gives the following useful corollary. By Corollary [\[co:ub_ppres_subtb_correspondence\]](#co:ub_ppres_subtb_correspondence){reference-type="ref" reference="co:ub_ppres_subtb_correspondence"}, the problem of determining \(\pres(f)\) can be rephrased as follows. We close this section by considering \(f(x)=x^2\) over \(\ff{9}\), giving \(M_{f}\), an example of admissible subtable, and the correspondence described in Lemma [\[lm:subtb_correspondence\]](#lm:subtb_correspondence){reference-type="ref" reference="lm:subtb_correspondence"}. # Linear Integer Programming Approach {#sc:alg_IP} As stated in Problem [\[pb:pres_table\]](#pb:pres_table){reference-type="ref" reference="pb:pres_table"}, determining \(\pres(f)\) of a function \(f:\grp{G}\to\grp{G}\) is equivalent to finding an admissible subtable \(\mathcal{A}\) of \(M_f\) whose value set has the minimum cardinality. This problem can be phrased as the following binary linear integer program: \[\begin{aligned} &\text{minimize:}\qquad \sum_{v\in \grp{G}}y_{v} \notag\\ &\text{subject to:} \notag\\ & \sum_{c\in\im{f}}x_{r,c}=1, \text{ for all }r\in \grp{G},\label{eq:IP_row}\\ & \sum_{r\in \grp{G}}x_{r,c}=\# \preim(f,c), \text{ for all }c\in\im{f},\label{eq:IP_col}\\ & x_{r,c}\le y_{v}, \text{ for all }r\in \grp{G},\,c\in\im{f},\,v\in \grp{G}\text{ such that }r-c=v,\label{eq:IP_countxy}\\ & x_{r,c},y_{v}\in\{0,1\}, \text{ for all }r\in \grp{G},\,c\in\im{f},\,v\in \grp{G}. \notag \end{aligned}\] We associate every entry \(m_{r,c}\) with a \(\{0,1\}\)-valued variable \(x_{r,c}\). These variables record the coordinates of the entries of \(\mathcal{A}\). If \(m_{r,c}\in \mathcal{A}\), then \(x_{r,c}=1\); otherwise, \(x_{r,c}=0\). Note that by (A2), an admissible subtable must have exactly \(q\) entries. So \(\range(\mathcal{A})=\grp{G}\) if and only if each row has exactly one entry in \(\mathcal{A}\). Thus, we add constraints ([\[eq:IP_row\]](#eq:IP_row){reference-type="ref" reference="eq:IP_row"}) to make sure that \(\range(\mathcal{A})=\grp{G}\). Similarly, for each column \(c\in\im{f}\), we require ([\[eq:IP_col\]](#eq:IP_col){reference-type="ref" reference="eq:IP_col"}) so that each column \(c\) has exactly \(\# \preim(f,c)\) entries in \(\mathcal{A}\). To count \(\# S\), we associate a \(\{0,1\}\)-valued variable \(y_{v}\) for every \(v\in \grp{G}\). If \(v\in S\), then \(y_{v}=1\); otherwise, \(y_{v}=0\). Hence, the objective is to minimize \(\sum_{v}y_{v}\). Finally, we add the constraints ([\[eq:IP_countxy\]](#eq:IP_countxy){reference-type="ref" reference="eq:IP_countxy"}), which force \(y_{v}=1\) whenever any entry \(m_{r,c}\) with value \(v\) is chosen in \(\mathcal{A}\). This IP has a total of \(q(V(f)+1)\) binary variables, \(q+V(f)\) equality constraints, and \(qV(f)\) inequality constraints. For functions over a finite field of order \(q\), we use Magma to generate \(M_f\), and use Gurobi via Python interface to solve the IP. Some computational results using the algorithm are given in an appendix. # Algorithm for Constructing an Admissible Subtable {#sc:alg_avg} We now present an algorithm for constructing an admissible subtable \(\mathcal{A}\) such that \(\range(\mathcal{A})=\grp{G}\). This gives a feasible solution of the IP in Section [3](#sc:alg_IP){reference-type="ref" reference="sc:alg_IP"} and thus an upper bound for \(\pres(f)\). We first describe the algorithm for a general \(f\), and then focus on the special case when \(f\) is two-to-one, a specific class of functions with strong connections to functions with optimal DU. ## General case {#ssb:alg_a_general} Let \(\# \grp{G}=q\), \(f:\grp{G}\to \grp{G}\) and \(u(f)=u\). The main idea of the algorithm is to iteratively choose a value \(v\) that appears at least the average number of times in the subtraction table, and then append those entries with value \(v\) to \(\mathcal{A}\). The algorithm is described below, followed by an explanation. We use \(()\) to describe a table or subtable, and \(\{\}\) to describe a set. For example, \(M_f=(m_{r,c} \,:\, r\in \grp{G},c\in\im{f})\) is the subtraction table that has \(qV(f)\) entries, and \(S=\{m_{r,c} \,:\, r\in \grp{G},c\in\im{f}\}\) is the set of all distinct values of the entries \(m_{r,c}\) that has \(q\) elements. To simplify the arguments, we say a row or column of \(M_f\) is in \(P_j\) for some \(0\le j\le u\), if it is indexed by an element in \(P_j\). To construct an admissible subtable \(\mathcal{A}\), we need to choose \(j\) entries from each column of \(M_f\) that is in \(P_{j}\) for some \(1\le j\le u\) to satisfy condition (A2). To make sure that \(\range(\mathcal{A})=\grp{G}\), each row of \(M_f\) has to be chosen exactly once. We first choose the entry with value \(0\) from each column. By our definition of \(M_f\), this is the diagonal of the upper part of \(M_f\) whose rows and columns are indexed by elements of \(\im{f}\). Thus, our initial subtable is \(\mathcal{A}_0=(m_{c,c}\,:\, c\in\im{f})\) with value set \(S_{0}=\{0\}\). Now, each row in \(P_1,\dots,P_u\) in \(M_f\) has been chosen once, so the remaining entries should be chosen from the rows in \(P_0\). Also, each column has been chosen once, so \(j-1\) more entries should be chosen from every column in \(P_j\) for all \(2\le j\le u\). This is equivalent to making a selection from each row and each column exactly once from the table \(M_0\) that is shown in Table [2](#tb:alg_a_m0){reference-type="ref" reference="tb:alg_a_m0"}. The table \(M_0\) is constructed by gluing \(j-1\) copies of the block of the subtable of \(M_{f}\) whose rows are in \(P_{0}\) and columns are in \(P_{j}\) for all \(2\le j\le u\). We assign each block a second index for the columns. For example, \(P_{j,1},\dots,P_{j,j-2},P_{j,j-1}\) are the \(j-1\) copies of \(P_j\). \ **Step 1:** \(\mu_0=\lceil 16/8\rceil=2\), so we can take \(S_{1}=\{0,a_{4}\}\), and obtain \(\mathcal{A}_{1}\) by adding the shaded entries to \(\mathcal{A}_{0}\). \(M_{1}\) is the table formed by deleting row \(a_{4},a_{6}\) and column \(0,a_{3}\), and the side length of \(M_{1}\) is \(n_{1}=n_0-\mu_0=2\). & \ **Step 2:** \(\mu_1=\lceil 4/8\rceil=1\), so we can take \(S_{2}=\{0,a_{4},a_{3}\}\), and obtain \(\mathcal{A}_{2}\) by adding the shaded entries to \(\mathcal{A}_{1}\). \(M_{2}\) is the table formed by deleting row \(a_{5}\) and column \(a_{2}\), and the side length of \(M_{2}\) is \(n_{2}=n_1-\mu_1=1\). & \ **Stop (\(k=2\)):** Now we stop and return \(S=\{0,a_{4},a_{3}\}\), \(\mathcal{A}=\mathcal{A}_{2}\cup \{\text{the shaded entry}\}\). & \ **Output:** The shaded entries present the output \(\mathcal{A}\). This gives an upper bound \(\pres(f)\le \# S=3\). Note that the worst case scenario in ([\[eq:pres_bound_nk\]](#eq:pres_bound_nk){reference-type="ref" reference="eq:pres_bound_nk"}) gives \(1+2+n_{2}=4\). & \ **Alternative stop (\(k=1\)):** We may also stop right after Step 1 instead, and return \(S=\{0,a_{4},a_{6},a_{7}\}\), \(\mathcal{A}=\mathcal{A}_{1}\cup\{\text{the shaded entries}\}\). & \ **Alternative output:** The shaded entries present the alternative output \(\mathcal{A}\) if we stop right after Step 1. This gives an upper bound \(\pres(f)\le \# S=4\). In this case, \(1+1+n_{1}=4\) so the worst case scenario in ([\[eq:pres_bound_nk\]](#eq:pres_bound_nk){reference-type="ref" reference="eq:pres_bound_nk"}) is met. & \ We remark that the algorithm may be improved by using a different method to choose \(v_t\) in each step \(t\). For example, we could use a greedy strategy to select a value \(v_t\) with the maximum appearance and append all relevant entries to \(\mathcal{A}_t\), instead of taking the average appearance only. See Table [13](#tb:alg1_ex_greedy){reference-type="ref" reference="tb:alg1_ex_greedy"} for an example. We use the average number here in order to estimate \(\# S\) in the next section for the special case of two-to-one functions. \ **Stop (\(k=1\)):** Now we stop and return \(S=\{0,a_{6},a_{7}\}\), \(\mathcal{A}=\mathcal{A}_{1}\cup\{\text{the shaded entries}\}\). & \ **Output:** The shaded entries present the output \(\mathcal{A}\). This gives an upper bound \(\pres(f)\le \# S=3\). In this case, \(1+1+n_{1}=3\) so the worst case scenario in ([\[eq:pres_bound_nk\]](#eq:pres_bound_nk){reference-type="ref" reference="eq:pres_bound_nk"}) is met. & \ # The special case when \(f\) is two-to-one We now focus on the special case when \(f\) is a two-to-one function. Our motivation for doing so is directly related to the problem of finding permutations with low DU in that a substantial class of optimal DU functions, that is planar functions, are two-to-one. First, consider the case when \(\# \grp{G}=q\) is even and \(V(f)=q/2\). The uniformity is \(u(f)=2\) and \(\im{f}=P_2\). So the initial square table \(M_0\) is simply the lower half of \(M_f\), i.e., the block whose rows are in \(P_0\) and columns are in \(P_2\). Since there is no need to repeat any blocks, we omit the second indices of the columns. Using Algorithm [\[alg:average\]](#alg:average){reference-type="ref" reference="alg:average"}, we prove the following upper bound for \(\pres(f)\). A slight modification of the previous proof yields the following. An immediate application of this theorem is the following. # The Existence of Admissible Subtables for Given Value Sets {#sc:alg_ex} We have seen that an upper bound \(\pres(f)\le k\) can be proved by finding an admissible subtable of \(M_{f}\) whose value set is a \(k\)-subset of \(\grp{G}\) and the range is \(\grp{G}\). In this section, we discuss the existence of such admissible subtables when the value set is given. Let \(f :\grp{G}\to \grp{G}\) and \(S\) be a \(k\)-subset of \(\grp{G}\). Clearly, any admissible subtable of \(M_{f}\) with value set contained in \(S\) must be a subtable of \(M_{S}:=(m_{r,c}\in S\,:\, m_{r,c}\in M_f)\), the subtable of all entries with values in \(S\). Define the range of \(M_{S}\) similarly as \[\label{eq:def_rangeMS} \range(M_{S}):=\{r\in \grp{G}\,:\, m_{r,c}\in S\},\] i.e., the indices of the rows of \(M_f\) that has an entry in \(S\). Hence, if we want to show \(\pres(f)\le k\) by finding an admissible subtable with value set \(S\), we should only proceed when \(\range(M_{S})=\grp{G}\). Indeed, if \(\range(M_{S})\neq \grp{G}\) for every \(k\)-subset \(S\) of \(\grp{G}\), then there is no admissible subtable with value set a \(k\)-subset and range \(\grp{G}\). So by Corollary [\[co:ub_ppres_subtb_correspondence\]](#co:ub_ppres_subtb_correspondence){reference-type="ref" reference="co:ub_ppres_subtb_correspondence"}, \(\pres(f)>k\). If we view each row of \(M_f\) as a set of entries, then \(M_f\) can be seen as a hypergraph \(H_{f}\) with vertex set \(\grp{G}\), and edge set the set of the rows of \(M_f\), which are all of cardinality \(V(f)\). For a hypergraph \(H\), a subset of the vertex set of \(H\) is called a *vertex cover* (also known as a *transversal*) if it intersects every edge of \(H\). From this point of view, \(S\) is a vertex cover of \(H_{f}\) if and only if every row of \(M_f\) has at least one entry in \(S\), i.e., \(\range(M_{S})=\grp{G}\). Therefore, it is possible that tools from the theory of hypergraphs may be used to study \(\pres(f)\). We call a subset \(S\subseteq \grp{G}\) a *cover of \(\grp{G}\) associated with \(f\)* if \(\range(M_{S})=\grp{G}\), or simply a *cover* when \(\grp{G}\) and \(f\) are clear from the context. In the next theorem, we give a sufficient condition for the existence of a cover with cardinality \(k\).\ When a cover \(S\) is found, the following algorithm searches for an admissible subtable \(\mathcal{A}\) of \(M_f\) such that \(\range(\mathcal{A})=\grp{G}\) and its value set is contained in \(S\). The algorithm is based on the fact that the desired \(\mathcal{A}\) must be a subtable of \(M_S=(m_{r,c}\in S\,:\, m_{r,c}\in M_f)\). For each column \(c\), let \(R_c\) be the set of indices of rows \(r\) such that \(m_{r,c}\in S\). Similarly, for each row \(r\), let \(Q_r\) collects the indices of columns \(c\) such that \(m_{r,c}\in S\). So the set of all \(R_c\) and \(Q_r\) carry the information of \(M_S\) where the search of admissible subtables should be limited in. The process starts by going over every row \(r\in \grp{G}\) and appending one entry from \(Q_r\) to \(\mathcal{A}\). More precisely, for \(c\in\im{f}\), we let \(A_{c}\) be the set of indices of rows \(r\) such that \(m_{r,c}\) is chosen in \(\mathcal{A}\). For every row \(r\), find exactly one \(c\in Q_r\) such that \(\# A_{c}\le\#\preim(f,c)\). This means that \(m_{r,c}\in M_S\) and column \(c\) still has "openings". Assign \(r\) to a \(A_{c}\) and update \(a_{r}=\mathrm{TRUE}\) to mark that \(r\) has been assigned to some \(A_{c}\). If no such \(c\) exists, then we leave \(r\) unassigned. After going over every row \(r\) once in the first step, let \(B\) be the set of indices of unassigned rows. If \(B\neq \emptyset\), then we work through every row \(r\in B\) and assign \(r\) to some \(A_{c}\) such that \(c\in Q_{r}\), and remove \(r\) from \(B\). If this makes \(\# A_{c}>\#\preim(f,c)\), then we move another element \(r'\) from \(A_{c}\) to \(B\). Repeat this step until there are no more unassigned rows (\(B=\emptyset\)), or if the algorithm fails to assign every row within some large number \(N\) of iterations. If the process succeeds with \(B=\emptyset\), then \(\bigcup_{c\in\im{f}}A_{c}=\grp{G}\), \(\#A_{c}=\#\preim(f,c)\) for each \(c\in\im{f}\), and the output \(\mathcal{A}=(m_{r,c}\,:\, c\in\im{f},r\in A_{c})\) is an admissible subtable with \(\range(\mathcal{A})=\grp{G}\) and value set contained in \(S\). This also verifies that \(\pres(f)\le \# S\). On the other hand, if \(B\neq \emptyset\) when the algorithm terminates, then no conclusion can be drawn for \(\pres(f)\) and one needs to restart with a different choice of \(S\). # A Generalization of the Linear Integer Program for Optimizing DU {#sc:gen_IP} In this final section, we generalize the IP in Section [3](#sc:alg_IP){reference-type="ref" reference="sc:alg_IP"} to find a permutation of optimal DU in a given finite group. As should be clear by now, one of the central motivations for studying permutation resemblance is to better understand how to construct low DU bijections, so we feel the generalization to be a natural next question. The generalized IP is the following: \[\begin{aligned} &\text{minimize:} \qquad DU \notag\\ &\text{subject to:} \notag\\ & \sum_{c\in\grp{G}}x_{r,c}=1, \text{ for all }r\in \grp{G},\label{eq:gen_IP_row}\\ & \sum_{r\in \grp{G}}x_{r,c}=1, \text{ for all }c\in\grp{G},\label{eq:gen_IP_col}\\ & \sum_{c\in \grp{G}}\sum_{r\in \grp{G}}z_{a,b,r,c}=\delta_{a,b}, \text{ for all }(a,b)\in\grp{G}^2,\,a\neq 0\label{eq:gen_IP_sum_delta}\\ & 2z_{a,b,r,c}\le x_{r+b,c+a}+x_{r,c}\le z_{a,b,r,c}+1, \text{ for all }(a,b,r,c)\in\grp{G}^4,\,a\neq 0\label{eq:gen_IP_z}\\ & 0\le\delta_{a,b}\le DU, \text{ for all }(a,b)\in\grp{G}^2,\,a\neq 0\label{eq:gen_IP_delta_DU}\\ & x_{r,c},z_{a,b,r,c}\in\{0,1\}, \text{ for all }(a,b,r,c)\in\grp{G}^4,\,a\neq 0 \notag \\ & DU,\delta_{a,b}\in\itr,\, \text{ for all }(a,b)\in \grp{G}^2,\,a\neq 0 \notag. \end{aligned}\] Let \(f:\grp{G}\to\grp{G}\). We first need to redefine the subtraction table of \(f\). For \(c\in\im{f}\), recall that in the previous setting, the table \(M_f\) only has the information of \(\#\preim(f,c)\), but does not keep track of the elements of \(\preim(f,c)\). Also, each column \(c\in\im{f}\) of an admissible subtable only records \(g(\preim(f,c))\) and \((g+f)(\preim(f,c))\) as whole sets, but does not specify which element in \(\preim(f,c)\) is to be mapped to which element. In order to compute DU directly in the IP, we need to keep more information on the preimages. We do this by "decompressing" each column. More precisely, for \(1\le t\le u(f)\), the new table will now have \(t\) copies of every column \(c\in\im{f}\) in \(P_t\), and each copy will be indexed by exactly one element of \(\preim(f,c)\) instead. Then the columns of the decompressed subtraction table \(M'_f\) are indexed by the whole group \(\grp{G}\), and the indices of the columns are now to be viewed as the elements of the domain, instead of \(\im{f}\). The entry of \(M'_f\), using the new indices, is defined by \(m_{r,c}:=r-f(c)\) for \(r,c\in\grp{G}\). With this new setting, an admissible subtable \(\mathcal{A}'\) should have exactly one entry in each column. Table [14](#tb:sub_x2_adm_decomp){reference-type="ref" reference="tb:sub_x2_adm_decomp"} is an example of \(M'_f\) and \(\mathcal{A}'\) obtained by decompressing Table [1](#tb:sub_x2_adm){reference-type="ref" reference="tb:sub_x2_adm"}. We keep the values \(f(c)\) for reference, but the actual column indices are now the \(c\)'s. This new admissible subtable \(\mathcal{A}'\) represents exactly one \(g\) and \(g+f\): for every grey entry \(m_{r,c}\in\mathcal{A}'\), the corresponding \(g\) sends \(c\) to the entry value of \(m_{r,c}\), and \(g+f\) sends \(c\) to \(r\). For example, \(g(0)=0\) and \((g+f)(0)=0\), \(g(1)=\alpha\) and \((g+f)(1)=\alpha^2\), \(g(-1)=1\) and \((g+f)(-1)=-1\), etc. The definition of \(\range(\mathcal{A}')\) still makes sense and remains the same as before. To construct a permutation \(g+f\), we still need to choose every row exactly once for an admissible subtable. ::: To construct a permutation \(F:=g+f\) by IP, we once again associate every entry \(m_{r,c}\) of \(M'_f\) with a \(\{0,1\}\)-valued variable \(x_{r,c}\). As before, \(x_{r,c}=1\) if and only if \(m_{r,c}\in\mathcal{A}'\), but in terms of functions, this now means \(F(c)=r\). As mentioned, we require every row and every column to be chosen exactly once in \(\mathcal{A}'\), in order to make sure that \(F\) is injective and well-defined, respectively. These correspond to the constraints ([\[eq:gen_IP_row\]](#eq:gen_IP_row){reference-type="ref" reference="eq:gen_IP_row"}) and ([\[eq:gen_IP_col\]](#eq:gen_IP_col){reference-type="ref" reference="eq:gen_IP_col"}). Next, define an integer-value variable "\(DU\)" to record the DU of \(F\). Clearly, the objective is to minimize \(DU\). To compute \(DU\), for every \((a,b)\in \grp{G}^{\star}\times\grp{G}\), define an integer-valued variable \(\delta_{a,b}\) recording the number of solutions of \(\Delta_{F,a}(x)=b\). Since by definition of DU, \(DU\) is the maximum among all \(\delta_{a,b}\), we add constraints ([\[eq:gen_IP_delta_DU\]](#eq:gen_IP_delta_DU){reference-type="ref" reference="eq:gen_IP_delta_DU"}). To compute each \(\delta_{a,b}\), observe that \(x\) satisfies \(F(x+a)-F(x)=b\) if and only if there is another input \(y\) such that \[y=x+a,\text{ and }F(y)=F(x)+b.\] In other words, we have both \(x_{F(x),x}=1\) and \(x_{F(y),y}=x_{F(x)+b,x+a}=1\). Therefore, the definition of \(\delta_{a,b}\) can be written into the following sum of products of binary variables. For a logical statement \(S\), the indicator function \(\mathbbm{1}_{(S)}=1\) if \(S\) is true, and \(\mathbbm{1}_{(S)}=0\) otherwise. \[\label{eq:ext_IP_delta} \begin{aligned} \delta_{a,b}&=\#\{c\in\grp{G} \,:\, F(c+a)-F(c)=b\}\\ &=\sum_{c\in\grp{G}}\mathbbm{1}_{(F(c+a)-F(c)=b)}\\ &=\sum_{c\in\grp{G}}\mathbbm{1}_{(x_{F(c)+b,c+a}=1\text{ and }x_{F(c),c}=1)}\\ &=\sum_{c\in\grp{G}}\sum_{r\in\grp{G}}\mathbbm{1}_{(x_{r+b,c+a}=1\text{ and }x_{r,c}=1)}\\ &=\sum_{c\in\grp{G}}\sum_{r\in\grp{G}}x_{r+b,c+a}x_{r,c}. \end{aligned}\] To make the last expression of ([\[eq:ext_IP_delta\]](#eq:ext_IP_delta){reference-type="ref" reference="eq:ext_IP_delta"}) linear, we replace each \(x_{r+b,c+a}x_{r,c}\) by a \(\{0,1\}\)-valued variable \(z_{a,b,r,c}\), which yields ([\[eq:gen_IP_sum_delta\]](#eq:gen_IP_sum_delta){reference-type="ref" reference="eq:gen_IP_sum_delta"}), and add two inequalities ([\[eq:gen_IP_z\]](#eq:gen_IP_z){reference-type="ref" reference="eq:gen_IP_z"}) for each choice of \((a,b,r,c)\in\grp{G}^{4},a\neq 0\). For transforming polynomial constraints into linear ones, we refer the reader to and references therein. From the final formulation of this generalized IP, we can see that the optimal solutions do not rely on the choice of \(f\). In fact, one can just take \(f=0\) if the only requirements are the constructed function being a permutation and having lowest DU. Compare to the IP in Section [3](#sc:alg_IP){reference-type="ref" reference="sc:alg_IP"}, this generalized IP is more costly to solve since it has both a larger number of variables and a larger number of constraints. That said, the generalized IP can do more combinations of different measurements of a function. For example, we can still compute \(V(g)\) by adding \(\{0,1\}\)-valued variables \(y_v\) for all \(v\in\grp{G}\), an integer-value variable \(V(g)=\sum_{v\in\grp{G}}y_v\), and constraints \(x_{r,c}\le y_v\) for every \((r,c)\in\grp{G}^2\) and \(r-f(c)=v\). Hence, we can restrict the problem further by setting the value of \(V(g)\) to \(\pres(f)\) if it is known, or require \(V(g)\) in an interval of possible values for \(\pres(f)\). Alternatively, instead of minimizing \(DU\), we could also replace the objective by minimizing \(V(g)\), and restricting \(DU\) to being equal to, or bounded above by, a certain value.

{'timestamp': '2023-02-10T02:00:45', 'yymm': '2302', 'arxiv_id': '2302.04300', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04300'}

# Introduction Nonparametric density estimation is a statistical task whose mathematical properties are very well-studied. The universal consistency of popular nonparametric density estimators, like the histogram and kernel density estimator, has been known for some time. In addition, there exist finite sample bounds and rates of convergence for nonparametric density estimators when the target density is known to come from a smooth class of densities. It is well-known that nonparametric density estimation suffers strongly from the curse of dimensionality. In theoretic works this typically manifests as a dimensionality exponent somewhere in rates or bounds (see Theorem 1 in, for example). Recently it has been proven that combining smoothness assumptions with a low-rank/multiview assumption can drastically improve the rate of convergence of a nonparametric density estimator, obviating the curse of dimensionality. In particular, it has been proven that there exist universally consistent nonparametric density estimators that converge at rate \(\tilde{O}_n\left(1/\sqrt[3]{n}\right)\) whenever the target density satisfies a *multiview model* assumption, \[\label{eqn:mv-finite} p\left(x_1,\ldots,x_d\right) = \sum_{i=1}^k w_i \prod_{j=1}^d p_{i,j}\left(x_j\right),\] where \(p_{i,j}\) are Lipschitz continuous probability density functions (pdfs) and \(w\) lies in the probability simplex. Remarkably this rate is independent of dimension, \(d\), the number of components, \(k\), or the Lipschitz constants of the component marginal densities. On the same class of densities it was also shown that the standard histogram estimator converges at rate \(\omega\left(1/\sqrt[d]{n}\right)\) regardless of choice of rate on bin width, which is a clear instance of the curse of dimensionality. While the result above demonstrates the potential benefits of incorporating multiview structure into density estimation, the assumption [\[eqn:mv-finite\]](#eqn:mv-finite){reference-type="eqref" reference="eqn:mv-finite"} is quite strong. This work extends the analysis of so that it is applicable to virtually any pdf. To do this, a new characterization of density complexity, the *non-negative Lipschitz spectrum* (NL-spectrum) is introduced. The NL-spectrum is an infinite sum (\(k=\infty\)) extension of [\[eqn:mv-finite\]](#eqn:mv-finite){reference-type="eqref" reference="eqn:mv-finite"}, that characterizes how fast \(w_i\) decays and the Lipschitz constants grow. The NL-spectrum is then shown to be applicable to an extremely general class of pdfs: every (Lebesgue) almost everywhere (a.e.) continuous pdf has an NL-spectrum. This enables the extension of the analysis in to a much larger class of pdfs. A finite sample bound depending on NL-spectra is then derived for the low-rank histogram estimators introduced in. Finally this bound is used to show rates of convergence of estimators based on NL-spectra rates of growth and decay. In particular, in the infinite component version of [\[eqn:mv-finite\]](#eqn:mv-finite){reference-type="eqref" reference="eqn:mv-finite"}, if it is known that \(\sum_{i=k+1}^\infty w_i \in O_k\left(k^{-\alpha}\right)\) and the Lipschitz constants of \(p_{k,j}\), \(L_k\) satisfy, \(L_k \in O_k\left(k^\beta\right)\) then there exists a universally consistent estimator that converges at rate \(\tilde{O}_n\left(n^{-\alpha/(3\alpha + \beta +1)}\right)\). ## Related Work The first work to investigate nonparametric density estimators for multiview models was, which proposed a method for factorizing a kernel density estimate to recover a multiview model. While contained theoretical guarantees for the rate of recovery of the multiview components, it did not demonstrate that the multiview assumption could be leveraged to improve estimator convergence. Instead it was proposed an approach to nonparametric mixture modeling. Other approaches to nonparametric mixture modeling with strong theoretical guarantees include assuming the data is grouped with respect to mixture components or that the are mixture components satisfy concentration assumptions. Like, these approaches did not produce rates of convergence that beat typical nonparametric rates. Other investigations into low-rank density estimation have focused on identifiability and recoverability, while other works have proposed low-rank methods that are empirically shown to improve nonparametric estimation without rate guarantees. Low-rank approaches to matrix estimation have also been studied extensively. Non-negative matrix or tensor factorization is a task similar to multiview density estimation since it can be used to recover a low-rank probability matrix/tensor. Low-rank matrix methods have been studied extensively in the field of *compressed sensing* which has produced improved matrix estimators with strong theoretical guarantees using the restricted isometry property or restricted strong convexity. Although a bit different than the methods presented so far, enforcing low non-negative rank for coupling measures in Wasserstein distance estimation has also been investigated. This has been shown to yield improved statistical estimation with computational benefits. While distance estimation is fairly different from density estimation, this method is noteworthy since it optimizes over a class of low-rank probability measures and has strong theoretical analysis demonstrating improved estimator convergence. The works are the first to show, via strong theoretic guarantees, that a multiview assumption can be used to improve the rate of convergence of nonparametric density estimators. Those works also show that a non-negative Tucker factorization can also be used to this effect and found that Tucker factorization produced better estimators in practice. Similarly to the proofs in those works, the results in this paper are *not* adaptations of techniques developed for compressed sensing or non-negative matrix factorization. # Results {#sec:results} This section presents and discusses the main results of this paper. Proofs of all results can be found in Section [3](#sec:proofs){reference-type="ref" reference="sec:proofs"}. Before introducing the results, some notation and terminology needs to be introduced. Other notation will be introduced intermittently through this work and a table summarizing notation in this work can be found in Appendix [\[appx:notation\]](#appx:notation){reference-type="ref" reference="appx:notation"}. For a pair of sets, \(A\) and \(B\), \(A\times B\) denotes the Cartesian product. For a pair of real-valued functions, \(f:A \to \mathbb{R}\) and \(g:B \to \mathbb{R}\), their product is defined as \(f\times g:A\times B\to \mathbb{R}\): \(\left(a, b\right) \mapsto f(a)g(b)\). Note that if \(A=B\) in this case then \(f\times g\) is a function on \(A\times A\), *not* a function on \(A\); \(f\cdot g: x\mapsto f(x)g(x)\). \(\mathbb{N}\) denotes all integers greater than 0. For a set \(A\), \(\mathbbm{1}_A\) denotes the indicator function on \(A\). For \(n \in \mathbb{N}\), \(\left[n\right] = \left\{1,2,\ldots,n\right\}\). For sets and functions, the product operator \(\prod\) and power operator \(\cdot^{\times n}\) will always mean the products, \(\times\), defined above. The term *almost everywhere* (a.e.) will always refer to the Lebesuge measure. A *pdf* is an a.e non-negative function with Lebesuge integral equal to one. Equalities (or inequalities) of functions will always mean equality (or inequality) almost everywhere. ## The Non-negative Lipschitz Spectrum In density estimation the smoothness of a pdf is often used as a measure of its complexity or how difficult the density is to estimate. The following is a characterization of a density's complexity in manner similar to the spectrum of a linear operator and is the focus of the rest of this work. Intuitively, an NL-spectrum where \(w\) decays slowly and the \(L_i\)'s are large indicates a more complex pdf. Later results will describe this precisely. The "smooth" descriptor of an NL-spectrum will not play any role for any of the results in the rest of this work. It is simply included because it is an additional regularity property that was simple to include in the proofs and may perhaps be useful in future works. It's worth noting that the NL-spectrum of a pdf is not unique. This lack of uniqueness is not only due to trivial modifications of the spectrum, e.g. reordering or repeated summands, but may also occur since, unlike the singular value decomposition of a matrix, minimal non-negative factorizations and factorizations of tensors are not necessarily unique up to scaling and reordering. There are many works investigating the uniqueness or lack of uniqueness of (non-negative) matrix/tensor/measure factorizations. This lack of uniqueness might be considered a potential disadvantage of the NL-spectrum compared smoothness-based characterizations of complexity such as Hölder, Sobolev, and Nikol'ski smoothness that are characterized by one or two scalar values rather than a pair of infinite series. The following theorem shows that NL-spectra describe a very rich class of pdfs. This set of pdfs arguably contains *all* pdfs that are of practical interest; it is difficult to imagine a real-life situation where one would be interested in estimating a pdf that is essentially discontinuous on a set of positive measure. In comparison the Hölder, Sobolev, and Nikol'ski smoothness classes are not applicable to pdfs that contain a single discontinuity. Decompositions or approximations reminiscent of the NL-spectrum exist elsewhere in analysis and probability theory. For example, a multivariate Riemann sum has a form similar to [\[eqn:mv-finite\]](#eqn:mv-finite){reference-type="eqref" reference="eqn:mv-finite"}, \[\sum_{i=1}^k w_i \mathbbm{1}_{\left(a_{i,1},b_{i,1}\right) \times \cdots \times \left(a_{i,d},b_{i,d}\right)} = \sum_{i=1}^k w_i \mathbbm{1}_{\left(a_{i,1},b_{i,1}\right)} \times \cdots \times \mathbbm{1}_{\left(a_{i,d},b_{i,d}\right)}= \sum_{i=1}^k w_i \prod_{j=1}^d \mathbbm{1}_{\left(a_{i,j},b_{i,j}\right)}.\] A similar decomposition has been mentioned in works on stochastic processes. The following quote is from: Theorem [\[thm:decomp\]](#thm:decomp){reference-type="ref" reference="thm:decomp"} stands apart from previous results because the NL-spectrum decomposition is exact, not an approximation, it is a mixture of pdfs, the component marginals have strong regularity properties, and it describes a very rich and practically useful class of pdfs. ## Nonparametric Estimator Results The following theorems show the existence of nonparametric density estimators whose performance depends on the NL-spectrum of the target density. Let \(\pazocal{D}_d\) be the set of all pdfs supported on the \(d\)-dimensional unit cube, \(\left[0,1\right]^{\times d}\). The following theorem is derived using a low-rank histogram estimator introduced in. While the estimators in this work are technically implementable, they are computationally intractable. It will be helpful to introduce a slightly different way of characterizing NL-spectra. There exist estimators with with following behavior on NL-classes. One can apply this to get estimators that are adapted to NL-classes with polynomial rates. Universal consistency holds on \(\pazocal{D}_d\); if the sampling density \(p\in \pazocal{D}_d\) is fixed then \(\left\|p-V_n\right\|_1 \overset{p}{\rightarrow} 0\). Here the decay in \(w\) works against growth in \(L\) with regards to convergence rate. Leaving \(\beta\) fixed and letting \(\alpha \to \infty\) this estimator approaches a rate of \(\tilde{O}_n\left(1/\sqrt[3]{n}\right)\) which matches the rate of convergence in for finite rank densities. This approximately matches the optimal rate of convergence for one-dimensional histograms in, "for smooth densities, the average \(L^1\) error for the histogram estimate must vary at least as \(n^{-1/3}\)." # Proofs {#sec:proofs} This section is broken up into two subsections. The fist subsection builds towards and then proves Theorem [\[thm:decomp\]](#thm:decomp){reference-type="ref" reference="thm:decomp"} and the second subsection proves estimator bounds and rates. Before proving the results of this paper, more notation must be introduced. Let \(\pazocal{F}\) be the set of smooth, Lipschitz continuous pdfs on \(\mathbb{R}\). Let \(\pazocal{F}^d \triangleq \left\{f_1 \times \cdots \times f_d \mid f_i \in \pazocal{F} \right\}\), note that these are also pdfs. Recall that the *conical hull* of a set \(S\) in a real-valued vector space is \(\operatorname{cone}\left(S\right) \triangleq \left\{ \sum_{i=1}^n w_i s_i\mid n\in \mathbb{N}, w_i\ge 0,s_i\in S\right\}\). The \(\operatorname{cone}\) operator will only be applied to \(\pazocal{F}^d\). Though standard notation, the reader is reminded that, for a set \(S\), \(S^{\mathbb{N}}\) is the set of all infinite sequences of elements of \(S\): \(S^{\mathbb{N}} \triangleq \left\{\left(s_i\right)_{i=1}^\infty \mid s_i \in S \right\}\). Finally \(\lambda\) denotes the Lebesgue measure where dimension will always be clear from context. ## Towards a Proof of Theorem [\[thm:decomp\]](#thm:decomp){reference-type="ref" reference="thm:decomp"} The following lemma states that under an indicator function on a multivariate interval, \(\prod_{i=1}^d\mathbbm{1}_{\left(a_i,b_i\right)}\), one can fit a positively scaled element of \(\pazocal{F}^d\) that approximates it arbitrarily well in \(L^1\) distance. The following lemma shows that, for any function in a class of sufficiently regular non-negative functions, one can find an element in the conical hull of \(\pazocal{F}^d\) that fits under the function and approximates that function arbitrarily well in \(L^1\) distance. Lemma [\[lem:compact-decomposition\]](#lem:compact-decomposition){reference-type="ref" reference="lem:compact-decomposition"} can now be generalized to prove Theorem [\[thm:decomp\]](#thm:decomp){reference-type="ref" reference="thm:decomp"}. ## Proofs of Estimator Bounds and Rates The estimator results build upon results in that analyzed histogram-style estimators of densities on \(d\)-dimensional unit cubes. Before presenting these results a bit more notation is required. In the authors define a notion of a low-rank histogram. For this context a "histogram" refers to a pdf that is constant on a partition containing equally spaced cubes. In particular they define \(\pazocal{H}_{1,b}\) to be the set of all histograms on the unit interval with \(b\) evenly spaced bins i.e. \[\pazocal{H}_{1,b} \triangleq \left\{\sum_{i=1}^b w_i \mathbbm{1}_{[(i-1)/b,i/b]}\mid w_i\ge 0, \int \sum_{i=1}^b w_i \mathbbm{1}_{[(i-1)/b,i/b]}d\lambda = 1 \right\}.\] A low-rank histogram on \(d\)-dimensional space, with \(b\) bins per dimension, is then defined as \[\pazocal{H}_{d,b}^k \triangleq \left\{\sum_{i=1}^k w_i \prod_{j=1}^d h_{i,j}\mid 0 \le w_i, \sum_{i=1}^k w_i = 1, h_{i,j} \in \pazocal{H}_{1,b} \right\}.\] Let \(\operatorname{Lip}_L\) be the set of \(L\)-Lipschitz continuous from \(\mathbb{R}\) to \(\mathbb{R}\). The following proposition was proven in and gives a bound on how well one can select estimators from \(\pazocal{H}_{d,b}^k\). To begin proving Theorem [\[thm:nl-finite\]](#thm:nl-finite){reference-type="ref" reference="thm:nl-finite"}, the bias term, \(\min_{q\in\pazocal{H}_{d,b}^k}\|p-q\|_1\), is analyzed via NL-spectrum. The following is a bound on rank-one bias that simplifies the analysis in (c.f. Theorem 2.5 in that paper). Note that a minimizer exists due to the compactness of the set \(\pazocal{H}_{d,b}^1\) (see Appendix A.1 in ). In the authors investigated pdfs whose domains were \([0,1]^{\times d}\), not \(\mathbb{R}^d\). This is a rather subtle point, but this means that the density \(p(x) =2x\mathbbm{1}_{[0,1]}(x)\) would be a valid 2-Lipschitz continuous density in that paper, but it is not in this work. The results from that work are applicable to densities in \(\pazocal{D}_d\cap \operatorname{Lip}_L\). The results in this work could be adjusted to be slightly more general on the domain \([0,1]^{\times d}\), but this was omitted for simplicity. The next lemma characterizes the bias in terms of NL-spectrum.  ◻ ::: [^1]: As a technical convenience it will always be assumed that \(w_1 >0\). [^2]: The facts used here are common in single variable analysis texts. For a complete treatment of the multivariate versions of these facts see Sections 2 and 3 in Chapter IV in. Note that the terminology in that work is somewhat nonstandard.

{'timestamp': '2023-02-10T02:00:27', 'yymm': '2302', 'arxiv_id': '2302.04292', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04292'}

# Introduction Precise measurements of magnetic fields are an important tool in understanding the spin properties and spatial distribution of magnetic fields in low dimensional materials. Over the past decade, intensive technique development in magnetic field sensing showcased the importance of coupling spatial resolution with minute sensitivity. Where techniques, such as fluxgate and hall effect sensors, are limited to micrometer scale sensing resolution, recent advances in quantum sensing have focused on optically detected magnetic resonance (ODMR) with color centers in semiconductors, e.g., diamond-NV centers. These defect-based color centers feature two key advantages---optical read-out and a high sensitivity to local magnetic fields. In ODMR, the ground state spin is initialized with optical excitation, probed with microwave pulses to coherently drive the magnetic sublevels of a triplet ground state via a singlet excited state, and then optical read-out is used to determine the populations of the spin sublevels. These color centers allow for nanoscale resolution of magnetic fields , but do not allow for intimate sensing of fields near the analyte, because defect qubits, like diamond-NV centers, operating as magnetic field sensors are spatially limited; they are difficult to bring in close proximity to the analyte as the defect is embedded in a crystal, e.g., the smallest analyte-qubit distance reported is 9 nm. This limitation is functionally related to the nature of these defects, many of which would not be thermodynamically stable in isolation; indeed, placing defects closer to the surface leads to their removal or lower coherence times. Molecules are zero-dimensional and functionally all surface. They may feature lower coherence times, but they are tunable and solution processable. This cross platform portability, in particular, enables post-synthetic processing of molecular color centers in thin film geometries to probe magnetic fields with distances ranging from angstroms to micrometers. Recent work has even highlighted this functionality with self-assembled monolayers of radical-based spins. The molecular color center Cr(*o*-tolyl)\(_{4}\) has a paramagnetic triplet \(S=1\) ground state (\(^{3}\)A symmetry [^1]) owing to the Cr\(^{4+}\) center and a zero-field splitting that allows for optical addressability analogous to color centers in solids. Here, light is used to resonantly excite an electron between the \(M_{s}=0,\pm 1\) sublevels via a singlet excited state. When exposed to even tiny magnetic fields, these electronic states undergo a Zeeman shift in energy (\(E'\)) proportional to the intensity of the magnetic field (\(B\)), the free electron \(g\)-factor (\(g_{e}\)), the Bohr magneton (\(\mu_{B}\)), and the Planck constant (\(h\)) as \(E^\prime = g_{e} \frac{\mu_{B}}{h}SB\,,\) where \(E^\prime\) is the difference in energy of the first singlet excited state (\(^{1}\)E symmetry) in the presence of a finite magnetic field (\(E_{B, ^{1}\mathrm{E}}\)) from that at zero field (\(E_{B=0, ^{1}\mathrm{E}}\)): \[\label{eq.energy} E^\prime = E_{B, ^{1}\mathrm{E}}-E_{B=0,^{1}\mathrm{E}}\.\] While precise sensing of magnetic fields is then obtained by experimentally measuring shifts in the energy levels \(M_{s}=0,\pm 1\) sublevels, we use shifts in the \(^{1}\)E excited state as a surrogate in our model [^2]. Recently, many correlated 2D magnets such as ferromagnetic CrGeTe\(_{3}\) and antiferromagnetic CrI\(_{3}\), which exhibits layer dependent magnetic ordering below 45 K, have been studied. To fully characterize these materials, high resolution sensors are needed. CrI\(_{3}\) is an ideal analyte for quantum magnetic field sensing at the nanoscale, as prior studies have shown large discrepancies (five orders of magnitude) in magnetic fields with distance away from the surface of the monolayer (a). Zhong et al. used a CrI\(_{3}\)/WSe\(_{2}\) heterostructure to calculate an effective magnetic field of 13 T from the valley splitting in WSe\(_{2}\) at 3.50 Å from CrI\(_{3}\). Single spin microscopy with a diamond-NV center has measured maximum magnetic fields of 0.31 mT at a distance of 62 nm from CrI\(_{3}\). Additional studies on CrI\(_{3}\)-based heterostructures have shown that such large magnetic fields are due to proximity exchange rather than the typical magnetostatics associated with stray magnetic fields. This raises the question of how the 2D magnet distributes the magnetic flux, and this knowledge is needed to guide codesign of spin-based logic and memory devices. One approach to resolve the distance dependent magnetic field strengths is to layer thin films of Cr(*o*-tolyl)\(_{4}\), the sensor, of varying thicknesses on CrI\(_3\), the analyte and perform thickness dependent ODMR measurements. This quantum sensing geometry in the single-molecule limit is shown in a and it is the configuration we consider. (We discuss thin film effects later.) We expect two important magnetic interactions, proximity exchange and direct magnetic fields, to occur between Cr(*o*-tolyl)\(_4\) and the CrI\(_{3}\) substrate. Proximity exchange is a local effect and should dominate at short distances \(d\) where electron wavefunctions directly overlap; in contrast, direct magnetic fields from diploar contributions should dominate at longer distances. Notably, it is very difficult to measure proximity exchange through conventional magnetic sensing techniques outside of NMR spectroscopy. Here, we use density functional theory (DFT) simulations with magnetostatic calculations to show how the first excited state of Cr(*o*-tolyl)\(_{4}\) is affected by these interactions as a function of distance, demonstrating molecular color centers as a quantum magnetic sensing platform. We first performed electronic structure calculations on monolayer CrI\(_{3}\) at the DFT-PBE level without spin-orbit coupling and in a ferromagnetic spin configuration [^1]: While the isolated molecule has \(C_{1}\) symmetry, we use the symmetry labels of the idealized tetrahedral structure. [^2]: Accurate quantitative changes in zero field splitting values are difficult to obtain with single-particle orbitals from DFT. Post-Hartree--Fock methods (i.e., complete active space self-consistent field methods) give more accurate results, but are not tractable for the sensor-analyte system explored. This necessitates a surrogate value for zero field splitting. Here we use the \(^{1}E\) singlet excited state that is involved in the excitation to the magnetic sublevels as a proxy.

{'timestamp': '2023-02-09T02:17:34', 'yymm': '2302', 'arxiv_id': '2302.04248', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04248'}

# Introduction {#sec:intro} Magnetic fields play a crucial role in explaining diverse processes throughout the solar atmosphere. However, the physical conditions in the chromosphere hinder the study of magnetic fields in this layer, which is key for linking the photospheric and coronal field topologies. Polarization measurements in the photosphere of the quiet Sun show magnetic elements of diverse sizes and lifetimes. These structures are divided into a network and an internetwork (@1975BAAS....7..346L; @1975BAAS....7R.346S). The former consists of nearly vertical magnetic flux tubes expanding with height due to force imbalances. Network structures are bright due to their lower opacity and harbor kG magnetic fields. In contrast, the internetwork is formed by small and short-lived structures randomly wandering due to convective motions until they merge with the network (e.g., @1985AuJPh..38..961Z; @2007ApJ...666L.137C; @2009ApJ...700.1391M; @2014ApJ...797...49G). These structures often appear as bipolar features linked by magnetic loops (e.g., @2007ApJ...666L.137C; @2009ApJ...700.1391M). Internetwork fields have strengths of the order of hG (e.g., @2003A&A...408.1115K; @2007ApJ...659..829A; @2007A&A...469L..39M; @2016A&A...593A..93D) and are crucial to our understanding of the quiet Sun magnetism. The coupling of the quiet Sun magnetic field between the photosphere and the outer atmosphere is often described by umbrella-shaped canopies rising from network structures and expanding higher up. However, the inference of hG fields in the internetwork led to more complex schemes, such as the multiscale carpet model where low-lying canopies link network and internetwork. In addition, the omnipresence of the internetwork suggests that it may contain an important fraction of the photospheric magnetic flux, thereby contributing greatly to the energy balance of the solar atmosphere. In this regard, found internetwork magnetic flux values resembling those in the network and active regions and highlighted the importance of the internetwork as a contributor to the network flux. Moreover, internetwork fields are able to travel higher up across the photosphere. Despite great efforts made over the years, the magnetism of the quiet Sun photosphere is not totally understood due to the small scale and weakness of the signals of the observed magnetic elements (see the reviews by [@2006ASPC..358..269T], @2011ASPC..437..451S, and @2019LRSP...16....1B). These drawbacks are exacerbated in the chromosphere, where the particle density drops. This decrease benefits the expansion of magnetic fields leading to weak polarization signals. For instance, Stokes \(V\) amplitudes in the 854.2 nm line in the quiet Sun are of about 10\(\mathrm{^{-2}}\)--10\(\mathrm{^{-3}}\) relative to the continuum intensity. However, noise levels below 10\(\mathrm{^{-3.5}}\) are required to detect linear polarization induced by the Zeeman effect in the \(\lambda\)`<!-- -->`{=html}8542 line. For field strengths of between 10 and 100 G, the Hanle effect also contributes to the linear polarization of this line. In this last case, the expected Stokes \(Q\) and \(U\) signals at disk center are of the order of 10\(^{-4}\). In addition, a low collisional rate causes a weak coupling of the radiation to the local physical conditions, which invalidates the assumption of local thermodynamic equilibrium (LTE). Fortunately, the use of circular polarization measurements in 854.2 nm has shed light on the magnetism of the quiet-Sun chromosphere. For example, the analysis of these measurements, combined with information in other spectral lines has allowed the study of the interaction of rising small-scale internetwork loops with network loops and the role of cancellations of opposite-polarity internetwork photospheric fields in the local chromospheric heating. In fact, the chromospheric heating is another key aspect related to the magnetic field in the quiet-Sun chromosphere. Diverse mechanisms can heat the chromosphere, but it is the Ohmic current dissipation that greatly contributes in the low-mid chromosphere of bright points and plages. This predominance in the chromospheric heating makes the inference of the chromospheric magnetic field in quiet-Sun magnetic elements an appealing task as part of further exploration. Despite the availability of diverse nonLTE inversion codes, the weak-field approximation (WFA) is commonly used to estimate the chromospheric magnetic field because it requires little time and is relatively straightforward (see Sect. [3.1](#subsec:wfa){reference-type="ref" reference="subsec:wfa"}). Specifically, this technique works well for measurements in the \(\lambda\)`<!-- -->`{=html}8542 line. According to, the longitudinal field estimates are reliable for field strengths up to 1.2 kG, but are underestimated for noise levels of 10\(\mathrm{^{-3}}\) in units of the continuum intensity. Given its importance, we aim to characterize the magnetic field of quiet-Sun structures depending on their size using high-spatial-resolution data taken in the 854.2 nm line with the Swedish 1m Solar Telescope. In this study, we determined the longitudinal magnetic field by applying the WFA to different spectral ranges across the line. In addition, we analyzed the estimates obtained from data at the original cadence and temporally integrated signals that mimic longer scanning times. # Observation and data reduction {#sec:observation} On August 1, 2019, we acquired four time series of quiet-Sun regions at different positions over the center of the solar disk with the CRisp Imaging SpectroPolarimeter attached to the SST. Table [1](#tab:description_regions){reference-type="ref" reference="tab:description_regions"} shows some observing characteristics of each sequence. Seeing conditions were good during the observing run except in the late morning. We therefore only analyzed the first 35 scans of the last time series. Each time series consists of full-Stokes measurements in the 854.2 nm line at high spatial resolution. The line was sampled at 21 nonequidistant wavelength positions between \(\pm\)`<!-- -->`{=html}1.040 Å  from the line center using steps of 65 mÅ  in the core and wider in the wings, plus two extra points at \(\pm\)`<!-- -->`{=html}1.755 Å. We acquired 12 accumulations per wavelength position. As the exposure time of each accumulation is 17 ms (with four modulation states), each sampling was completed every 31.5 s. The field of view (FOV) is \(\sim\)`<!-- -->`{=html}55\(\times\)`<!-- -->`{=html}55and the pixel size 0.057. We performed the data reduction using the SSTRED pipeline (@2015A&A...573A..40D; [@2021A&A...653A..68L]). Images were restored with the Multi-Object-Multi-Frame-Blind-Deconvolution technique. Finally, we performed the polarimetric calibration on a pixel-by-pixel basis as proposed by. The noise level in the resulting Stokes \(Q\), \(U\), and \(V\) images is 2.3--2.8\(\times\)`<!-- -->`{=html}10\(\mathrm{^{-3}}\), 2.4--2.6\(\times\)`<!-- -->`{=html}10\(\mathrm{^{-3}}\), and 2.2--2.5\(\times\)`<!-- -->`{=html}10\(\mathrm{^{-3}}\) as measured on the \(Q\)/\(I_{0,QS}\), \(U\)/\(I_{0,QS}\), and \(V\)/\(I_{0,QS}\) maps at the first observed wavelength of each series. \(I_{0,QS}\) refers to the first observed wavelength intensity averaged over a very quiet area. Given the weak linear polarization signals expected in the quiet-Sun chromosphere (see Sect. [1](#sec:intro){reference-type="ref" reference="sec:intro"}), the polarimetric sensitivity of our spectropolarimetric data is not large enough to detect them in all the studied elements. # Analysis {#sec:analysis} Considering their brightness and conspicuous Stokes \(V\) signals, we search for field concentrations in the 854.2 nm intensity filtergrams at \(-\)`<!-- -->`{=html}1.755 Å  and magnetograms at \(\pm\)`<!-- -->`{=html}0.39 Å  with the CRISPEX tool. We found 31 concentrations of different sizes, which we delimited by adopting thresholds of intensity at--1.755 Å  and circular polarization at--0.39 Å  of 0.95 \(I_{0,QS}\) and \(\pm\)`<!-- -->`{=html}4\(\times\)`<!-- -->`{=html}10\(\mathrm{^{-3}}\) \(I_{0,QS}\). The red rectangles overplotted in Fig. [\[fig:data1\]](#fig:data1){reference-type="ref" reference="fig:data1"} display the location of these structures. Most of the concentrations are visible during the entire run and often undergo fragmentations and mergings with other nearby concentrations. Such interactions are also included in this study. While static plane-parallel mediums show antisymmetric Stokes \(V\) profiles, asymmetries appear in realistic model atmospheres. These asymmetries are mainly due to velocity gradients along the line of sight (LOS) and can be enhanced by magnetic-field gradients with height (@1975A&A....41..183I; @1978A&A....64...67A). Additionally, the presence of different isotopes in the solar atmosphere also induces asymmetries in the \(\lambda\)`<!-- -->`{=html}8542 line. We used an automated code to identify pixels with two-lobed \(V\) profiles. Before that, we corrected the circular polarization signals for residual crosstalk from Stokes \(I\) similarly to, resulting in noise levels of 2.1--2.4\(\times\)`<!-- -->`{=html}10\(\mathrm{^{-3}}\) as measured on the \(V\)/\(I_{0,QS}\) maps at the first observed wavelength of each sequence. We also applied a principal component analysis denoising technique (PCA, @1901Pearson, @1933Hotelling) to the original circular polarization signals, leading to better-defined \(V\) profiles that prevent detection errors. Specifically, we used four eigenvectors for the Stokes \(V\) profiles in the first and third series and five in the second and fourth ones. We also smoothed the \(V\) profiles with a boxcar average of a width of 2 pixels to ease the detection of the desired pixels. Finally, the subsequent analysis was performed using the signals resulting from the application of the PCA technique. In addition to using the data at the original cadence, we analyzed the results from different scanning times (\(\mathrm{\Delta t}\)) to inspect the variation of the longitudinal magnetic field (see Sect. [3.1](#subsec:wfa){reference-type="ref" reference="subsec:wfa"}) with the signal addition, in the hope of increasing the detectability of weaker fields. For this aim, we first identified the scans showing each structure and grouped them as sets of two, four, and ten scans. Then, we combined the two-lobed Stokes \(V\) signals and the corresponding Stokes \(I\) profiles at each pixel according to each set of scans to imitate observations with \(\mathrm{\Delta t}\) of 1.05, 2.10, and 5.25 minutes. We also performed these combinations using all scans displaying each structure. We ruled out pixels with complex \(V\) profiles by considering that their signals are zero. To avoid signal cancellation, we also discarded pixels with reversed polarities compared to that prevailing in each structure. ## Application of the WFA {#subsec:wfa} The WFA can be applied when the Zeeman splitting is much smaller than the Doppler width of the spectral line under consideration and the longitudinal component of the magnetic field is constant along the LOS. Under these circumstances, the observed Stokes \(V\) profile is related to the observed Stokes \(I\) profile as \[V(\lambda) \ = \-C \ f \ B \ \cos\gamma \ \frac{\partial I(\lambda)}{\partial\lambda}, \label{eq:wfa_1}\] with \(C\,=\,4.6686\times10^{-13}\,\lambda_{0}^{2}\,g_{\mathrm{eff}}\), where \(\lambda_{0}\) and \(g_\mathrm{eff}\) are the central wavelength (in Å) and the effective Landé factor of the line, respectively. The angle \(\gamma\) is the inclination of the magnetic field with respect to the LOS direction. The longitudinal component of the magnetic field vector is \(B_{\|}\) = \(B \,\cos\gamma\), with \(B\) the field strength (in G). The observed profile \(I(\lambda)\) is assumed to be the intensity that would emerge with zero magnetic field. The filling factor \(f\) is the fractional area occupied by the magnetic component within the resolution element. For simplicity, we assume \(f\) = 1. Thus, the shape of Stokes \(V\) is proportional to the derivative of Stokes \(I\) scaled by the \(B_{\|}\) estimate. We compute \(B_{\|}\) as the maximum-likelihood estimation from Eq. ([\[eq:wfa_1\]](#eq:wfa_1){reference-type="ref" reference="eq:wfa_1"}) assuming Gaussian noise with zero mean and variance \(\sigma^2\), which leads to \[B_{\|} \ = \-\ \dfrac{\sum\limits_{i} \ \frac{\partial I(\lambda_{i})}{\partial\lambda_{i}} \ V(\lambda_{i})}{C \ \sum\limits_{i}\left(\frac{\partial I(\lambda_{i})}{\partial\lambda_{i}}\right)^{2}}. \label{eq:wfa_2}\] ## Definition of the analyzed spectral ranges The 854.2 nm line covers a large range of formation heights. Its core is formed in the low-to mid-chromosphere, while its wings have a photospheric contribution. The variation from the LTE wings to the nonLTE chromospheric core occurs at \(\pm\)`<!-- -->`{=html}0.03--0.04 nm from the line center. Therefore, the estimation of \(B_{\|}\) values from the low-to mid-chromosphere is possible if the spectral range is restricted to the line core. Given its wide formation region, delimiting spectral ranges in the \(\lambda\)`<!-- -->`{=html}8542 line is a common method to infer \(B_{\|}\) at certain heights. Recently, this approach was also applied to synthetic data in the  h and k lines. In this paper, we applied Eq. ([\[eq:wfa_2\]](#eq:wfa_2){reference-type="ref" reference="eq:wfa_2"}) to different spectral ranges across the line to inspect how the \(B_{\|}\) estimates differ. Specifically, we selected almost all the spectral sampling (\(\Delta \lambda = \pm1.040\) Å), a limited range including both lobes of the \(V\) profile (\(\pm0.325\) Å), and the inner-core positions (\(\pm0.130\) Å). Hereafter, we denote these ranges as \(\mathrm{\Delta \lambda_{1}}\), \(\mathrm{\Delta \lambda_{2}}\), and \(\mathrm{\Delta \lambda_{3}}\), respectively. # Results {#sec:results} The analyzed quiet-Sun structures evolve differently with time. While some of them undergo fragmentations, others either grow as they merge with nearby chunks or expand before decreasing. All cases show two-lobed \(V\) profiles coinciding with Stokes \(I\) signals that sometimes show asymmetries. We divided the structures we found into three groups: small, medium-sized, and large field concentrations based on their aspect and size at the original cadence. Small field concentrations are elements with lengths of 0.45--1.2wandering around granules. Medium-sized ones seem to be chains of magnetic elements expanding 3.5--5.5 between granules. Finally, large concentrations are fixed structures spreading over 10--20. Figure [\[fig:hist1\]](#fig:hist1){reference-type="ref" reference="fig:hist1"} shows histograms of the number of pixels with two-lobed \(V\) profiles and \(B_{\|}\) estimates for each \(\Delta \lambda\) per group, which are normalized to the number of scans and pixels throughout the time sequences in each group (\(\sim\)`<!-- -->`{=html}96\(\times\)`<!-- -->`{=html}10\(^{3}\), 266\(\times\)`<!-- -->`{=html}10\(^{3}\), and 682\(\times\)`<!-- -->`{=html}10\(^{3}\) pixels for small, medium-sized, and large concentrations), respectively. On average, pixels inside small, medium-sized, and large concentrations represent 0.06%, 0.13%, and 2.39% of the FOV per frame, respectively. The distribution of the number of pixels with two-lobed \(V\) signals in large concentrations diverges from those in small and medium-sized ones. The former has two compact peaks standing for the two cases classified as large concentrations, while the latter two groups show broader size variability. Small concentrations tend to show negative polarity and often have \(B_{\|}\) values below 250 G for all \(\Delta \lambda\). Medium-sized concentrations usually have positive polarity and harbor estimates below 400 G. The distributions of the large concentrations are smoother and reach \(B_\|\) values above 400 G. In addition, most of the pixels in large concentrations have negative polarity, which is the polarity of the largest structure. Generally, all distributions appear shifted to weaker estimates as \(\Delta \lambda\) narrows down. All cases of each group share similar properties. We therefore describe them by showing one case for each group. ## Small field concentrations {#sub:small} We identified 19 cases as small concentrations. They appear in intergranular lanes and are highly influenced by neighboring granules. We focus on the case labeled '3' in the first row of Fig. [\[fig:data1\]](#fig:data1){reference-type="ref" reference="fig:data1"}. During its temporal evolution (see the movie attached to Fig. [\[fig:gr1_2\]](#fig:gr1_2){reference-type="ref" reference="fig:gr1_2"}), small positive-polarity field concentrations appear between granules and move closer to each other until they form a single structure. After that, the structure changes its shape in reaction to the motion of the surrounding granules. Figure [\[fig:gr1_2\]](#fig:gr1_2){reference-type="ref" reference="fig:gr1_2"} also displays the variation of the blue-wing magnetogram (at--0.39 Å) and \(B_{\|}\) with the scanning time \(\mathrm{\Delta t}\). At the original cadence, estimates usually range from 100 to 360 G for \(\mathrm{\Delta \lambda_{1}}\), and decay to 80--250 G and to 40--150 G for \(\mathrm{\Delta \lambda_{2}}\) and \(\mathrm{\Delta \lambda_{3}}\). Due to its dynamic behavior, the structure appears larger in maps at long \(\mathrm{\Delta t}\). Furthermore, \(V\) signals and \(B_{\|}\) estimates vary with \(\mathrm{\Delta t}\) depending on the individual \(I\) and \(V\) profiles that are summed and the number of scans involved. Despite pixel-by-pixel differences, the estimates obtained for each \(\mathrm{\Delta \lambda}\) do not differ significantly with \(\mathrm{\Delta t}\). We remind the reader that \(\mathrm{\Delta t}\) refers to the time used to calculate the averaged Stokes \(I\) and \(V\) profiles (see Sect. [3](#sec:analysis){reference-type="ref" reference="sec:analysis"}). We assessed the validity of the WFA conditions by comparing the observed Stokes \(V\) profiles to the \(I\) derivatives scaled by the estimates for each \(\Delta \lambda\). Figure [\[fig:gr1_fit\]](#fig:gr1_fit){reference-type="ref" reference="fig:gr1_fit"} displays these comparisons at two different locations as \(\mathrm{\Delta t}\) increases. The two upper and lower rows show the \(I\) and \(V\) profiles from a pixel at the center and from another closer to the edge of the concentration, respectively. The plus symbols (\(+\)) and crosses (\(\times\)) in Fig. [\[fig:gr1_2\]](#fig:gr1_2){reference-type="ref" reference="fig:gr1_2"} indicate these two specific locations. Although these locations do not share the same features, their \(I\) and \(V\) signals rely on the imprints left at these pixels during the instants involved in the signal addition. Consequently, all the estimates vary with the scanning time. We note that the gradient between the estimates obtained for each \(\mathrm{\Delta \lambda}\) changes very little with \(\mathrm{\Delta t}\). From a qualitative comparison, we observe a systematic divergence between the Stokes \(V\) profiles and the scaled \(I\) derivatives at the position of the blue lobe. Both curves only match in some pixels for \(\mathrm{\Delta \lambda_{3}}\) (as in the fourth row of Fig. [\[fig:gr1_fit\]](#fig:gr1_fit){reference-type="ref" reference="fig:gr1_fit"}), but they slightly differ in others due to a relative shift between Stokes \(I\) and \(V\) at the line core (see the second row). This relative shift may be due to the coexistence of multiple components in the same resolution element. Using wider spectral ranges, the divergences between the \(V\) profiles and the scaled \(I\) derivatives remain regardless of \(\mathrm{\Delta t}\). At this point, we note that we only describe the goodness of the fits from a qualitative point of view. A quantitative description is also possible. In that case, we would conclude that the quality worsens with \(\mathrm{\Delta t}\) because the noise level for \(V\)/\(I\) decreases with the square root of the number of scans involved in each sum. This is mostly a consequence of the WFA not being able to explain the observations independently of \(\mathrm{\Delta t}\). As such a conclusion seems counterintuitive to what we observe from a naked-eye comparison, we opted for a qualitative description of the quality of the fits. In addition, we observe that the fit quality between Stokes \(V\) and the scaled \(I\) derivatives at the outer-wing positions improves close to the concentration edge (fourth row in Fig. [\[fig:gr1_fit\]](#fig:gr1_fit){reference-type="ref" reference="fig:gr1_fit"}). To examine this further, we compared the synthetic 854.2 nm \(V\) profiles resulting from two model atmospheres with the STiC code. The model atmospheres are based on the FALC model to which we added \(B_{\|}\) values monotonically decreasing between log(\(\mathrm{\tau_{500}}\)) = 1.3 and \(-\)`<!-- -->`{=html}5.3: (1) from 1 000 to 100 G, and (2) from 300 to 100 G. We observe similarities in the outer-wing positions of the synthetic \(V\) profiles (shown in Fig. [\[fig:syn\]](#fig:syn){reference-type="ref" reference="fig:syn"}) and the observed ones, which indicates the high sensitivity of the \(V\) signals of the outer-wings to \(B_\|\) in the photosphere. Therefore, the fit quality in the outer-wing positions relies not only on the spectral range but also on the pixel location, as the field concentrations fan out with height. Other features may involve gradients in other parameters or the presence of an unresolved magnetic component not included here for simplicity. In view of the above findings, we compared the estimates for \(\mathrm{\Delta \lambda_{3}}\) and the outer-wing positions (between \(\pm\)`<!-- -->`{=html}1.040 and \(\pm\)`<!-- -->`{=html}0.520 Å), from which we expect mainly chromospheric and photospheric contributions, respectively. We denote the latter range as \(\mathrm{\Delta \lambda_{4}}\). As it may change with height, we delimited the structure again using intensity maps at--0.520 Å, where we still discern the structure despite the presence of fibrils. However, Fig. [\[fig:gr1_flux\]](#fig:gr1_flux){reference-type="ref" reference="fig:gr1_flux"}a--b shows that the structure barely varies with height, whereas \(B_{\|}\) varies conspicuously. Central and outer regions of the concentration show significant height gradients in \(B_{\|}\) that resemble those used in the synthesis with STiC. Moreover, these gradients are consistent with time and are independent of the profile smoothing and of the noise decrease, as they appear regardless of \(\mathrm{\Delta t}\) (see Fig. [\[fig:gr1_fit\]](#fig:gr1_fit){reference-type="ref" reference="fig:gr1_fit"}). Therefore, WFA conditions cannot be satisfied by adding signal. Finally, we computed the magnetic flux from the chromospheric and photospheric contributions using the expression \[\Phi(B_{\|}) \ = \ \sum_{i,j}B_{\|}(i, j) \cdot \Delta A, \label{eq:flux}\] where \(B_{\|}(i, j)\) is the estimate for \(\mathrm{\Delta \lambda_{3}}\) and \(\mathrm{\Delta \lambda_{4}}\) at a pixel with coordinates \((i, j),\) and \(\Delta A\) is the pixel area. Specifically, the \(B_{\|}\) values used in Eq. ([\[eq:flux\]](#eq:flux){reference-type="ref" reference="eq:flux"}) are those estimated at the original cadence. Figure [\[fig:gr1_flux\]](#fig:gr1_flux){reference-type="ref" reference="fig:gr1_flux"}c shows the resulting magnetic flux values, which vary with time depending on the temporal evolution of the structure. In addition, given the height gradient in \(B_\|\), the photospheric and chromospheric magnetic flux values differ by one order of magnitude during the time interval showing this case (see also Table [2](#tab:flux){reference-type="ref" reference="tab:flux"}). [\[tab:flux\]]{#tab:flux label="tab:flux"} ## Medium-sized field concentrations {#sub:gr2} We classified ten cases as medium-sized field concentrations; they have a practically fixed position and are observed as stable structures. In this subsection, we show an analysis of the case labeled '3' in the second series (see Fig. [\[fig:data1\]](#fig:data1){reference-type="ref" reference="fig:data1"}), which is portrayed in Fig. [\[fig:gr2_2\]](#fig:gr2_2){reference-type="ref" reference="fig:gr2_2"}. This positive-polarity case undergoes fragmentations and mergings (see the animation attached to Fig. [\[fig:gr2_2\]](#fig:gr2_2){reference-type="ref" reference="fig:gr2_2"}). At the original cadence, the \(B_{\|}\) estimates for \(\mathrm{\Delta \lambda_{1}}\) and \(\mathrm{\Delta \lambda_{2}}\) are usually above 200 G and exceed 350 G at the center of the structure. These values are weaker for \(\mathrm{\Delta \lambda_{3}}\) overall, reaching \(\sim\)`<!-- -->`{=html}250--300 G. Because of its stable behavior, \(B_{\|}\) maps barely change for \(\mathrm{\Delta t}\) \(<\) 5.25 minutes. The intensity profiles at the center of the concentration are different from those emerging closer to the edges, as shown in the upper and lower rows of Fig. [\[fig:gr2_fit\]](#fig:gr2_fit){reference-type="ref" reference="fig:gr2_fit"}. Locations close to the edge usually show line asymmetries and shifts, which may indicate that the physical conditions at the center of the structure are firmer than in the outer regions. In particular, the intensity profiles represented in the lower panels display slight asymmetries compared to those in the upper rows. Stokes \(V\) profiles also differ as amplitudes weaken away from the center. Aside from the \(V\) amplitude variations, we do not find changes with \(\mathrm{\Delta t}\) worth mentioning. Regarding the \(B_{\|}\) estimates, the central and outer locations of the concentration show values of \(\sim\)`<!-- -->`{=html}420 and 260 G for \(\mathrm{\Delta \lambda_{1}}\) at the original cadence, respectively. These \(B_{\|}\) weaken by about 60 G and 100--200 G for \(\mathrm{\Delta \lambda_{2}}\) and \(\mathrm{\Delta \lambda_{3}}\), respectively. Moreover, the gradient between the estimates for each spectral range is stable with \(\mathrm{\Delta t}\) and comparable to the individual values. We also compared the mostly chromospheric and photospheric \(B_{\|}\) estimates in this structure. The inferred values diverge significantly (see Fig. [\[fig:gr2_flux\]](#fig:gr2_flux){reference-type="ref" reference="fig:gr2_flux"}a--b), having median values of 105 and 524 G, respectively. Thus, the magnetic flux values for \(\mathrm{\Delta \lambda_{3}}\) and \(\mathrm{\Delta \lambda_{4}}\) differ again by one order of magnitude (Fig. [\[fig:gr2_flux\]](#fig:gr2_flux){reference-type="ref" reference="fig:gr2_flux"}c and Table [2](#tab:flux){reference-type="ref" reference="tab:flux"}) but, unlike the small case, they are stable with time. The height gradient in \(B_{\|}\) also explains the poor fit quality between the \(V\) profiles and the scaled \(I\) derivatives for \(\mathrm{\Delta \lambda_{1}}\) and \(\mathrm{\Delta \lambda_{2}}\). In contrast, fits are usually better for \(\mathrm{\Delta \lambda_{3}}\), although they are affected by relative shifts between Stokes \(I\) and \(V\) at some positions. All in all, the estimates for wider spectral ranges are unsafe as the WFA conditions are unfulfilled. ## Large field concentrations {#sub:gr3} We identified two cases of large field concentrations; they are labeled '1' and '2' in the last row of Fig. [\[fig:data1\]](#fig:data1){reference-type="ref" reference="fig:data1"} and may be an example of *enhanced network*. We focus on case 2, shown in Fig. [\[fig:gr3_2\]](#fig:gr3_2){reference-type="ref" reference="fig:gr3_2"}, the polarity of which is negative. The movie attached to Fig. [\[fig:gr3_2\]](#fig:gr3_2){reference-type="ref" reference="fig:gr3_2"} shows that the \(B_{\|}\) values seem to oscillate with time, mostly at the center of the structure. As these oscillations coincide with overall intensity fluctuations, we think they are related to seeing instabilities instead of actual changes in the structure. At the original cadence, we infer estimates that progressively increase from the edges to the center of the structure, where they are of about \(-\)`<!-- -->`{=html}750 G for \(\mathrm{\Delta \lambda_{1}}\). The \(B_{\|}\) maps for \(\mathrm{\Delta \lambda_{2}}\) and \(\mathrm{\Delta \lambda_{3}}\) are similar to that obtained for \(\mathrm{\Delta \lambda_{1}}\), though estimates are weaker reaching up to \(-\)`<!-- -->`{=html}650 and \(-\)`<!-- -->`{=html}550 G, respectively. In addition, the concentration grows with \(\mathrm{\Delta t}\) due to interactions with nearby structures and to the variable seeing conditions, whereas \(B_{\|}\) values weaken globally. At the original cadence, asymmetries and blueshifts are apparent in the intensity profiles from the center of this concentration and from a pixel closer to the edge (see upper and lower rows of Fig. [\[fig:gr3_fit\]](#fig:gr3_fit){reference-type="ref" reference="fig:gr3_fit"}, respectively). These features persist with the scanning time until they soften at the longest \(\mathrm{\Delta t}\) due to the profile smoothing. At the same time, \(V\) amplitudes decrease with the distance to the center of the structure and barely change with \(\mathrm{\Delta t}\). As a result of these changes along the concentration, \(B_{\|}\) values differ significantly depending on the pixel location. We find that, at the original cadence, estimates range from \(-\)`<!-- -->`{=html}500 to \(-\)`<!-- -->`{=html}700 G and from \(-\)`<!-- -->`{=html}180 to \(-\)`<!-- -->`{=html}260 G for the different \(\mathrm{\Delta \lambda}\) at the center and closer to the edge, respectively. These \(B_\|\) gradients are stable with \(\mathrm{\Delta t}\), and so the conditions along the formation region related to the used spectral ranges hardly vary with time. Although not displayed, the difference between the mostly chromospheric and photospheric estimates in this concentration is not as significant as in smaller structures. The median values in this case for \(\mathrm{\Delta \lambda_{3}}\) and \(\mathrm{\Delta \lambda_4}\) are \(-\)`<!-- -->`{=html}160.7 and \(-\)`<!-- -->`{=html}360.3 G, respectively. Consequently, the photospheric magnetic flux is three to four times greater than the chromospheric flux (see Table [2](#tab:flux){reference-type="ref" reference="tab:flux"}). Similarly to the medium-sized case, the magnetic flux values inferred in this concentration have a stable temporal evolution. The reduced difference between the photospheric and chromospheric estimates may explain a better match between Stokes \(V\) and the scaled \(I\) derivatives regardless of the spectral range, scanning time, and location. # Discussion {#sec:discussion} Assuming WFA conditions, the reliability of the estimates for wide spectral ranges seems to rely on the concentration size. Small and medium-sized field concentrations have steeper time-independent height gradients in \(B_{\|}\) compared to their chromospheric \(B_{\|}\) values than the large ones, which leads to unreliable estimates of the former as the WFA requisites are unfulfilled. On the contrary, estimates for inner-core ranges are usually reliable in all concentrations. Despite its assets, some observers may employ other restoration methods on the basis of the liabilities of MOMFBD. In this section, we compare results from data restored with different methods in order to decipher whether or not the restoration technique is related to the fulfillment of WFA conditions. In addition, as the WFA is widely used, we also discuss our results by comparing them to previous studies that used similar data. ## Results from differently restored data MOMFBD is a successful image-correction technique. However, its application comes with high computational cost and may lead to amplified noise levels. Alternatively, images can be restored by adding the accumulations acquired per wavelength position. In this case, most of the reduction process is carried out using the CRISPRED pipeline and the effects of seeing in observations are partially corrected by applying the destretching technique. During this process, differential motions in small subfields of the wide-band images, which are previously derotated and aligned, are measured using cross-correlation. After combining accumulations, the polarimetric calibration is performed as in. Finally, each pair of images with orthogonal polarization states, recorded simultaneously by the narrow-band cameras, are combined to remove the seeing-induced polarization, and the residual crosstalk is corrected. We also corrected each monochromatic polarization image for high-frequency polarimetric interference fringes in a similar way to. As each image is apodized in this process, we clipped the blurred edges of the output polarization maps, obtaining smaller images. The noise level of the resulting Stokes \(V\) images is in the range of 1.8--2.5\(\times\)`<!-- -->`{=html}10\(^{-3}\) measured on the \(V\)/\(I_{0}\) maps at the first observed wavelength for each series. We refer to the datasets restored with and without the MOMFBD technique as M and NM datasets, respectively. We examined 27 of the cases shown in Fig. [\[fig:data1\]](#fig:data1){reference-type="ref" reference="fig:data1"} because some of them are partially or completely beyond the image boundaries of the NM datasets. Each case has been delimited using thresholds of intensity at \(-\)`<!-- -->`{=html}1.755 Å  and \(V(\lambda)/I(\lambda)\) at \(-\)`<!-- -->`{=html}0.39 Å  of 0.95 \(I_{0,QS}\) and 4\(\times\)`<!-- -->`{=html}10\(^{-3}\), respectively. After that, we estimated \(B_{\|}\) by applying the WFA in pixels with two-lobed \(V\) profiles and analyzed their variation with the scanning time, \(\mathrm{\Delta t}\) (see Sect. [3](#sec:analysis){reference-type="ref" reference="sec:analysis"}). Figure [\[fig:gr123_nm\]](#fig:gr123_nm){reference-type="ref" reference="fig:gr123_nm"} shows the \(B_{\|}\) estimates obtained for the cases depicted in Sect. [3](#sec:analysis){reference-type="ref" reference="sec:analysis"} using the NM datasets. The physical quantities in this figure appear smeared in comparison to those in Figs. [\[fig:gr1_2\]](#fig:gr1_2){reference-type="ref" reference="fig:gr1_2"}, [\[fig:gr2_2\]](#fig:gr2_2){reference-type="ref" reference="fig:gr2_2"}, and [\[fig:gr3_2\]](#fig:gr3_2){reference-type="ref" reference="fig:gr3_2"}. We observe not only blurrier intensity images but also smoother \(B_\|\) estimates. Indeed, the latter are generally 50--100 G weaker for all \(\mathrm{\Delta t}\). These differences between both datasets appear because the intensity signal in each pixel is spread out over its surroundings in the NM datasets, as neither the blurring effect of the telescope point spread function (PSF) nor the low-altitude seeing is corrected. Nonetheless, the smearing relies on the concentration size. This correlation is clear in Fig. [\[fig:gr123_nm_fit\]](#fig:gr123_nm_fit){reference-type="ref" reference="fig:gr123_nm_fit"}, where the divergence of the \(I\) and \(V\) profiles from the NM and M datasets increases as the concentration size decreases. Furthermore, in Fig. [\[fig:gr123_nm_fit\]](#fig:gr123_nm_fit){reference-type="ref" reference="fig:gr123_nm_fit"}, we observe that the fit quality between the Stokes \(V\) profiles and the scaled \(I\) derivatives from the NM datasets is similar to that described for the M datasets. However, the smearing in the NM datasets leads to better fits in all concentrations, particularly for \(\mathrm{\Delta \lambda_3}\). In the case of wider spectral ranges, the fit quality again relies on the concentration size, because height gradients in \(B_\|\) are more abrupt in small and medium-sized concentrations. Magnetic flux variations between the photosphere and chromosphere are therefore of the same order as in the M datasets (Table [3](#tab:fluxnm){reference-type="ref" reference="tab:fluxnm"}). Specifically, the median values of the estimates for the photospheric \(\mathrm{\Delta \lambda_4}\) range (and the chromospheric \(\mathrm{\Delta \lambda_3}\) one) in the small, medium-sized, and large concentrations are 461.3 (33.7), 475.3 (174.3), and \(-\)`<!-- -->`{=html}404.9 (\(-\)`<!-- -->`{=html}248.2) G, respectively. [\[tab:fluxnm\]]{#tab:fluxnm label="tab:fluxnm"} Finally, Fig. [\[fig:hist2\]](#fig:hist2){reference-type="ref" reference="fig:hist2"} shows the histograms of the number of pixels with two-lobed \(V\) profiles in the concentrations analyzed using the NM datasets and their \(B_{\|}\) values for each spectral range. The former distributions are normalized to the number of scans showing each group whereas the latter are normalized to the number of pixels per group in all sequences (\(\sim\)`<!-- -->`{=html}46\(\times\)`<!-- -->`{=html}10\(^{3}\), 150\(\times\)`<!-- -->`{=html}10\(^{3}\), and 510\(\times\)`<!-- -->`{=html}10\(^{3}\) pixels for small, medium-sized, and large structures), respectively. Although all distributions resemble those for the M datasets (Fig. [\[fig:hist1\]](#fig:hist1){reference-type="ref" reference="fig:hist1"}), they are affected by the value smearing, as fewer pixels meet the threshold used to delimit each case (Fig. [\[fig:hist2\]](#fig:hist2){reference-type="ref" reference="fig:hist2"}a) and the peaks of some distributions are shifted to stronger estimates (Fig. [\[fig:hist2\]](#fig:hist2){reference-type="ref" reference="fig:hist2"}b--d). ## Comparison with previous studies Over the past few years, investigations based on high-spatial-resolution data have reported a weakening of the quiet-Sun magnetic fields with height (e.g., @2018ApJ...857...48G; @2019A&A...621A...1R; @2020A&A...642A.210M). We also observe a transition to weaker longitudinal fields as the spectral range narrows down around the line core. Close to the limb, found longitudinal magnetic fields of \(\sim\)`<!-- -->`{=html}400 G in the photosphere decreasing to 200 G in the chromosphere of quiet-Sun areas. To estimate the latter, these authors applied the WFA to a spectral range of \(\pm\)`<!-- -->`{=html}1 Å  from the 854.2 nm line center. At a heliocentric angle of 37\(^{\circ}\), inferred \(B_{\|}\) values of about 250 G outside a plage by performing STiC inversions of full-Stokes data in the \(\lambda\)`<!-- -->`{=html}8542 line (sampled within \(\pm\)`<!-- -->`{=html}880 mÅ  from the line center). Also outside a plage but closer to the disk center, reported a weakening of \(B_{\|}\) from \(\sim-\)`<!-- -->`{=html}600 G in the photosphere to \(-\)`<!-- -->`{=html}200 G in the low-mid chromosphere by applying a spatially constrained WFA method to the wings of the 589.6 nm line and the 854.2 nm core (\(\pm\)`<!-- -->`{=html}110 mÅ). Recently, found maximum \(B_{\|}\) values of 400--800 G in internetwork clusters in 617.3 nm data that weaken by more than 200 G when applying the WFA to a spectral range of \(\pm\)`<!-- -->`{=html}200 mÅ  in the 854.2 nm line core. We report \(B_{\|}\) values of hG in quiet-Sun magnetic concentrations. Our mostly photospheric and chromospheric estimates in small and medium-sized structures are compatible with those previously found. However, our results for wider ranges across the line are stronger. Large concentrations also show stronger estimates, which is likely due to a departure of their physical properties from those typically found in the quiet Sun. Finally, we note that differences in the heliocentric angle may also lead to divergence of the results from those of other studies. Furthermore, an interesting aspect of our study is the comparison of the photospheric and chromospheric estimates in concentrations of different size. Small concentrations usually host compact height variations of \(B_{\|}\) exceeding 400---500 G, where the maximum can be stronger than 800 G. Medium-sized and large structures show similar height variations but more extended over the concentrations. However, inspection of the average height variation of \(B_{\|}\) reveals that small and medium-sized concentrations have stronger average height variation (of about 400 and 425 G, respectively) than large concentrations (of \(\sim\)`<!-- -->`{=html}300 G). Corroborating this result by analyzing other large concentrations will be of great interest because our sample only has two large structures, which were observed under poor seeing conditions. In addition, the height variations of \(B_{\|}\) can be considered weaker or stronger than the chromospheric \(B_{\|}\) estimated in each structure. In this regard, when compared to large concentrations, small and medium-sized concentrations display stronger height gradients in \(B_{\|}\) than their chromospheric estimates. Consequently, large structures usually show better fits between the observed \(V\) profiles and the scaled \(I\) derivatives for wider spectral ranges. Investigations of the stratification of the magnetic field with height in sunspots have revealed discrepancies depending on the approach used to determine the magnetic gradients (see the review of @2018SoPh..293..120B). Undoubtedly, similar studies in the quiet Sun are also of great importance as they can shed light on the topology of its magnetic field. Nonetheless, the elusive detection of the weak polarization signals emerging from the quiet-Sun chromosphere hinders these investigations. This impasse will be overcome by using observations acquired with the next-generation solar telescopes, such as DKIST and EST. # Conclusions and summary Characterization of the magnetic field of the quiet-Sun chromosphere is essential for comprehending the chromospheric heating and the coupling of the magnetic field between the photosphere and the outer atmosphere. For this reason, we determined the longitudinal component of the magnetic field inferred in quiet-Sun magnetic concentrations depending on their size. Specifically, we applied the WFA to high-spatial resolution 854.2 nm data acquired close to the disk center with SST/CRISP using different spectral ranges and temporally combined signals. We find a correlation between the amplitude of \(B_{\|}\) estimates and concentration size, with large structures hosting the largest values of \(B_{\|}\). By restricting the spectral range to the line core, we infer chromospheric estimates that smoothly change from \(\sim\)`<!-- -->`{=html}50 at the edges to 150--500 G at the center of the concentrations. These values increase as the spectral range widens due to the influence of the photospheric contribution, which is especially strong compared to the chromospheric estimates in small and medium-sized concentrations. Consequently, the photospheric magnetic flux is three to four times greater than the chromospheric one in the large concentration and this difference can be as large as one order of magnitude in the small and medium-sized cases. The effect of the height gradients in \(B_{\|}\) on the reliability of the estimates given by the WFA depends on the concentration size and spectral range. Estimates for the inner-core range are usually reliable regardless of the concentration size, though they are affected by relative shifts between the \(I\) and \(V\) signals at some locations. In contrast, estimates for wider spectral ranges are generally unreliable in small and medium-sized structures. Although differences in \(B_{\|}\) and size are seen in small concentrations due to their dynamic behavior, signal addition usually leads to estimates akin to those at the original cadence for all spectral ranges. Therefore, we infer that the conditions along the considered formation regions may be stable with time in the studied structures and that signal addition has not warranted the detection of weaker signals. Moreover, signal addition reduces the noise of the chromospheric estimates and is not related to a better fulfillment of the WFA requisites when using wider ranges. Finally, although deconvolving processes may enhance noise levels, data restored by destretching show similar results to those obtained when using the MOMFBD technique. However, results from the former are affected by the smearing caused by the partial correction of the data.

{'timestamp': '2023-02-09T02:17:54', 'yymm': '2302', 'arxiv_id': '2302.04258', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04258'}

# Introduction Text classification is a classical problem in [NLP]{acronym-label="NLP" acronym-form="singular+short"} that aims to assign a label to textual units like words, sentences, paragraphs, or documents. Text classification has a wide range of applications. It is used in question answering, spam detection, sentiment analysis, news categorization, user intent detection, and many others. [NLP]{acronym-label="NLP" acronym-form="singular+short"} applications are becoming more popular daily due to the advances in various computational linguistics and the abundance of training data from websites, personal communications (emails, text messages), social media, tickets, insurance claims, user reviews, and questions/answers from customer services. The insurance industry collects a large amount of textual data as part of their day to day processes. The vast majority of this textual data is read and interpreted by human agents resulting in higher costs and slower processes. This creates an opportunity to automate some text processing tasks and make the insurance processes faster and cheaper. Traditional [NLP]{acronym-label="NLP" acronym-form="singular+short"} techniques will assist insurance companies to personalize insurance products and better respond to existing and future clients' needs. More accurate underwriting models and constant learning from new data will be reflected in premiums that are often conservatively overestimated [due to]{style="color: black"} a lack of non-traditional underwriting data. [Using [NLP]{acronym-label="NLP" acronym-form="singular+short"} solutions in the insurance industry yields]{style="color: black"} better customer satisfaction and profit. The use cases include claims classification, optimizing payment processes, monitoring policy changes, personalized product offerings, improved risk assessment, enhanced fraud detection, and business process automation. Documents and textual data are rich sources of information that can be used to solve various problems. However, extracting information from this type of data requires more complex techniques due to the unstructured nature of textual data, which can be time-consuming and challenging. Text classification is an important step in text processing. This can be done manually by domain experts or automatically. [Recently]{style="color: black"}, due to the increasing number of textual documents, automatic text classification has become popular and important. Automatic text classification approaches are grouped into rule-based and data-driven methods. Rule-based methods categorize text samples using a set of pre-defined rules that requires deep expert knowledge, which is context-based and hard to acquire. On the other hand, data-driven methods use [ML]{acronym-label="ML" acronym-form="singular+short"} approaches to find patterns to classify textual samples. However, most of the ML-based text classifiers are built from labeled training [samples]{style="color: black"}. Manual labeling of a large set of training documents is time-consuming. In the past few years, researchers investigated various forms of semi-supervised learning to reduce the burden of manual labeling by using a small labeled set for every class and a large unlabeled set for classifier building. Semi-supervised learning is a hybrid approach that combines supervised and unsupervised learning elements to train models with a small amount of labeled data, and a large amount of unlabeled data. [Moreover, word embedding techniques were used to represent the words of a text using representation learning methods in a way that words that have the same meaning have a similar representation.]{style="color: black"} Similar to word embeddings, distributed representations for sentences can also be learned in an unsupervised fashion by optimizing some auxiliary objectives, such as the reconstruction loss of an autoencoder. Such unsupervised learning results in sentence encoders that can map sentences with similar semantic and syntactic properties to similar fixed-size vector representations. In this paper, a novel text classification model, CRL+, that combines [[CRL]{acronym-label="CRL" acronym-form="singular+short"}]{style="color: black"} and [Active Learning]{style="color: black"} is proposed to handle the challenge of using semi-supervised textual data for text classification in the insurance industry. In this method, supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} will be used to train a RoBERTa transformer model to encode the textual data into the representational vector and then classify using a classification layer. This [CRL]{acronym-label="CRL" acronym-form="singular+short"}-based transformer model will be used as the base model of a modified Active Learning mechanism to classify all the data in an iterative manner. The remainder of this paper is organized as follows. Section [9](#sec:rw){reference-type="ref" reference="sec:rw"} provides a summary of [the related works in the literature]{style="color: black"}. Section [10](#sec:bg){reference-type="ref" reference="sec:bg"} provides some background about the [methods used in this paper]{style="color: black"}. Section [11](#sec:pm){reference-type="ref" reference="sec:pm"} describes the proposed approach [for classifying]{style="color: black"} semi-supervised textual data. Section [12](#sec:em){reference-type="ref" reference="sec:em"} explains the experimental setup, including the dataset and evaluation metrics. Section [13](#sec:res){reference-type="ref" reference="sec:res"} shows the experimental results. Finally, Section [14](#sec:con){reference-type="ref" reference="sec:con"} concludes the paper. # Related Works {#sec:rw} It is estimated that around 80% of all information is unstructured, with text being one of the most common unstructured data types. The unstructured data structure is irregular or incomplete, and there is no predefined data model. Compared to structured data, [this data is]{style="color: black"} still difficult to retrieve, analyze and store. This is where text classification with ML comes in. ML-based techniques can automatically classify all manner of relevant text, from legal documents, social media, surveys, and more, in a fast and cost-effective way. Some of the most popular machine learning algorithms for creating text classification models include the [NB]{acronym-label="NB" acronym-form="singular+short"} family of algorithms, [SVM]{acronym-label="SVM" acronym-form="singular+short"}, and [DNN]{acronym-label="DNN" acronym-form="singular+short"}. Recent research shows that it is effective to cast many [NLP]{acronym-label="NLP" acronym-form="singular+short"} tasks as text classification by allowing [DNN]{acronym-label="DNN" acronym-form="singular+short"} to take a pair of texts as input. Compared to [traditional]{style="color: black"} [ML]{acronym-label="ML" acronym-form="singular+short"}, [DNN]{acronym-label="DNN" acronym-form="singular+short"} algorithms need more training data. However, they [do not]{style="color: black"} have a threshold for learning from training data like traditional machine learning algorithms, such as [SVM]{acronym-label="SVM" acronym-form="singular+short"} and [NB]{acronym-label="NB" acronym-form="singular+short"}. The two main [DNN]{acronym-label="DNN" acronym-form="singular+short"} architectures for text classification are [CNN]{acronym-label="CNN" acronym-form="singular+short"} and [RNN]{acronym-label="RNN" acronym-form="singular+short"}. [RNN]{acronym-label="RNN" acronym-form="singular+short"}s are trained to recognize patterns across time, whereas [CNN]{acronym-label="CNN" acronym-form="singular+short"}s learn to recognize patterns across space. [RNN]{acronym-label="RNN" acronym-form="singular+short"}s work well for the [NLP]{acronym-label="NLP" acronym-form="singular+short"} tasks such as [QA]{acronym-label="QA" acronym-form="singular+short"}, where the comprehension of long-range semantics is required. In contrast, [CNN]{acronym-label="CNN" acronym-form="singular+short"}s work well where detecting local and position-invariant patterns are essential. Thus, [RNN]{acronym-label="RNN" acronym-form="singular+short"}s have become one of [NLP]{acronym-label="NLP" acronym-form="singular+short"} 's most popular model architectures. [LSTM]{acronym-label="LSTM" acronym-form="singular+short"} is a popular architecture, which addresses the gradient vanishing problems in [RNN]{acronym-label="RNN" acronym-form="singular+short"}s by introducing a memory cell to remember values over arbitrary time Intervals. There have been works improving [RNN]{acronym-label="RNN" acronym-form="singular+short"}s and [LSTM]{acronym-label="LSTM" acronym-form="singular+short"} models for [NLP]{acronym-label="NLP" acronym-form="singular+short"} applications by capturing richer information, such as tree structures of natural language, long-span word relations in text, and document topics. Character-level [CNN]{acronym-label="CNN" acronym-form="singular+short"}s have also been explored for text classification. Studies investigate the impact of word embeddings and [CNN]{acronym-label="CNN" acronym-form="singular+short"} architectures on model performance. presented a [VDCNN]{acronym-label="VDCNN" acronym-form="singular+short"} model for text processing. It operates directly at the character level and uses only small convolutions and pooling operations. This study shows that the performance of [VDCNN]{acronym-label="VDCNN" acronym-form="singular+short"} improves in deeper models. [DNN]{acronym-label="DNN" acronym-form="singular+short"} algorithms, like word2vec or Glove , are also used to obtain better vector representations for words and improve the accuracy of classifiers trained with traditional machine learning algorithms. [Recently, transformer models were introduced to improve the performance of [LSTM]{acronym-label="LSTM" acronym-form="singular+short"} models. Unlike [LSTM]{acronym-label="LSTM" acronym-form="singular+short"}, transformer models can be fully parallelized on GPUs, which makes the training step faster. Moreover, transformers can memorize longer sentences better than the [LSTM]{acronym-label="LSTM" acronym-form="singular+short"}s due to their attention mechanism. AS one of the most popular transformer models, the BERT model was introduced by Google in 2019. BERT model outperforms its predecessors, Word2Vec and ELMo, exceeding state-of-the-art by a margin in multiple natural language understanding tasks. Section [10.5](#sec:tran){reference-type="ref" reference="sec:tran"} will discuss the transformers in more detail.]{style="color: black"} # Background {#sec:bg} In this chapter, background materials related to this paper will be introduced. ## Semi-Supervised Learning Semi-supervised learning is a type of [ML]{acronym-label="ML" acronym-form="singular+short"} in which [only a portion of the training samples are labeled.]{style="color: black"} Most of the real-world problems in the fintech and insurance industries are semi-supervised. [Therefore]{style="color: black"}, handling this type of problem [has recently gained momentum]{style="color: black"}. Active Learning and [CRL]{acronym-label="CRL" acronym-form="singular+short"} are two popular solutions for these problems. ## Active Learning Active Learning is a mechanism to label unlabeled data iteratively using an [ML]{acronym-label="ML" acronym-form="singular+short"} model. This mechanism [increases the performance of [ML]{acronym-label="ML" acronym-form="singular+short"} models on datasets with limited labeled samples]{style="color: black"}. In this mechanism, first, an [ML]{acronym-label="ML" acronym-form="singular+short"} model is trained using the labeled data. Then, the model is applied to unlabeled data to classify them. Based on the classification result, some samples are selected to pass to an expert person to label them manually and add them to the labeled data. This process is [repeated]{style="color: black"} until the model meets [predefined performance criterion]{style="color: black"}. ## Contrastive Representation Learning [CRL]{acronym-label="CRL" acronym-form="singular+short"} is a [ML]{acronym-label="ML" acronym-form="singular+short"} technique used for unsupervised and semi-supervised problems as a pre-training to enhance the performance of [ML]{acronym-label="ML" acronym-form="singular+short"} models. This model is used to train a representation space in which similar sample points stay close, while dissimilar ones are far apart. [CRL]{acronym-label="CRL" acronym-form="singular+short"} was first developed as a self-supervised method for unsupervised image classification problems (or as an unsupervised pre-training for supervised problems). In this method, a self-supervised pre-training was done on the data using augmentation techniques to map the samples from the original space to the contrastive space. Then, Khosa et al. proposed a supervised version of [CRL]{acronym-label="CRL" acronym-form="singular+short"} and showed its advantages compared to the self-supervised version on image classification problems. In this method, instead of considering the anchor sample and its augmented one as the positive samples, the samples with similar labels are also considered as positives. Moreover, several [NLP]{acronym-label="NLP" acronym-form="singular+short"} versions of [CRL]{acronym-label="CRL" acronym-form="singular+short"} were proposed that include both self-supervised and supervised [versions]{style="color: black"} of these models. In this paper, a supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} model, introduced in, is used as the base model to pre-train the RoBERTa model using the Active Learning mechanism. ## [RNN]{acronym-label="RNN" acronym-form="singular+short"} [RNN]{acronym-label="RNN" acronym-form="singular+short"} is a type of neural network that predicts new situations based on previous ones. [RNN]{acronym-label="RNN" acronym-form="singular+short"}s can handle sequential problems like [NLP]{acronym-label="NLP" acronym-form="singular+short"}. Several [RNN]{acronym-label="RNN" acronym-form="singular+short"} models are proposed in the literature. One of the most popular [RNN]{acronym-label="RNN" acronym-form="singular+short"} models is [LSTM]{acronym-label="LSTM" acronym-form="singular+short"}, [which suffers less from the vanishing gradient problem compared to [RNN]{acronym-label="RNN" acronym-form="singular+short"}]{style="color: black"}. ## Transformers {#sec:tran} One of the computational bottlenecks [of training [RNN]{acronym-label="RNN" acronym-form="singular+short"}s on GPUs]{style="color: black"} is the sequential processing of text. Transformers overcome this limitation by applying self-attention to compute [an attention score]{style="color: black"} in parallel for every word in a sentence or document an \"attention score\" to model each word's influence on another. Due to this feature, Transformers allow for much more parallelization than [CNN]{acronym-label="CNN" acronym-form="singular+short"}s and [RNN]{acronym-label="RNN" acronym-form="singular+short"}s, which makes it possible to efficiently train huge models on large amounts of data on GPUs. Since 2018, we have seen the rise of a set of large-scale Transformer-based [PLM]{acronym-label="PLM" acronym-form="singular+short"}s. Compared to earlier contextualized embedding models based on [CNN]{acronym-label="CNN" acronym-form="singular+short"}s or [LSTM]{acronym-label="LSTM" acronym-form="singular+short"}s, Transformer-based [PLM]{acronym-label="PLM" acronym-form="singular+short"}s use much deeper network architectures (e.g., 48-layer Transformers ), and are pre-trained on much larger amounts of text corpora to learn contextual text representations by predicting words conditioned on their context. This [PLM]{acronym-label="PLM" acronym-form="singular+short"}s are fine-tuned using task-specific labels and have created a new state of the art in many downstream [NLP]{acronym-label="NLP" acronym-form="singular+short"} tasks, including text classification. Although pre-training is unsupervised (or self-supervised), fine-tuning is supervised learning. A recent survey by Qiu et al. categorizes popular [PLM]{acronym-label="PLM" acronym-form="singular+short"}s by their representation types, model architectures, pre-training tasks, and downstream tasks. [PLM]{acronym-label="PLM" acronym-form="singular+short"}s can be grouped into two categories, autoregressive and autoencoding [PLM]{acronym-label="PLM" acronym-form="singular+short"}s. One of the earliest autoregressive [PLM]{acronym-label="PLM" acronym-form="singular+short"}s is OpenGPT, a unidirectional model that predicts a text sequence word by word from left to right (or right to left), with each word prediction depending on previous predictions. It consists of 12 layers of Transformer blocks, each consisting of a masked multi-head attention module, followed by a layer normalization and a position-wise feed-forward layer. OpenGPT can be adapted to [NLP]{acronym-label="NLP" acronym-form="singular+short"} applications such as text classification by adding task-specific linear classifiers and fine-tuning using task-specific labels. BERT is one of the most widely used autoencoding [PLM]{acronym-label="PLM" acronym-form="singular+short"}s. Unlike OpenGPT, which predicts words based on previous predictions, BERT is trained using the [MLM]{acronym-label="MLM" acronym-form="singular+short"} task that randomly masks some tokens in a text sequence and then independently covers the masked tokens by conditioning on the encoding vectors obtained by a bidirectional Transformer. There have been numerous works on improving BERT. RoBERTa is more robust than BERT, mainly because its pre-training method focuses on [MLM]{acronym-label="MLM" acronym-form="singular+short"} with changing mask tokens per epoch, and is trained using much more training data. ALBERT lowers the memory consumption of the BERT model and increases its training speed. DistillBERT utilizes knowledge distillation during pre-training to reduce the size of BERT by 40%, while retaining 99% of its original capabilities and making the inference 60% faster. SpanBERT extends BERT to better represent and predict text spans. Electra uses a more sample-efficient pre-training task than [MLM]{acronym-label="MLM" acronym-form="singular+short"}, called replaced token detection. Instead of masking the input, it corrupts it by replacing some tokens with plausible alternatives sampled from a small generator network. ERNIE incorporates domain knowledge from external knowledge bases, such as named entities, for model pre-training. ALUM introduces an adversarial loss for model pretraining that improves the model's generalization to new tasks and robustness to adversarial attacks. BERT and its variants have been fine-tuned for various [NLP]{acronym-label="NLP" acronym-form="singular+short"} tasks, including [QA]{acronym-label="QA" acronym-form="singular+short"}, text classification \[154\], and [NLI]{acronym-label="NLI" acronym-form="singular+short"}. # Proposed Method {#sec:pm} [T]{style="color: black"}his paper [combines]{style="color: black"} supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} with an Active Learning mechanism to enhance the performance of these methods to classify textual data. In this paper, a modified version of SupCL-Seq is used as the [CRL]{acronym-label="CRL" acronym-form="singular+short"} model. SupCL-Seq extends the self-supervised Contrastive Learning for textual data to a supervised setting. In the developed model, several dropouts are used to make augmented samples from the anchor by changing sample embeddings. Then, the anchor sample, its augmented samples, and other samples with the same label in the dataset are used to train the representation using the contrastive loss function. The representation learning task consists of an encoder part of a transformer (RoBERTa) to make the augmented samples and train the [CRL]{acronym-label="CRL" acronym-form="singular+short"} part of the model. Figure [\[fig:crl\]](#fig:crl){reference-type="ref" reference="fig:crl"} shows the training of the developed model. Then, a classification layer is added to the trained representation learning model to do the final classification. Finally, all the [CRL]{acronym-label="CRL" acronym-form="singular+short"} encoder parameters are frozen, and the classification model is trained. Equation [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"} shows the supervised contrastive loss function. \[\label{eq:1} \mathcal{L}^{sup}_i = \sum_{i\in I} \frac{-1}{|p(i)|} \sum_{\textit{p} \in p(i)} log \frac{e^{cosine(\Tilde{x_i},\Tilde{x_p})/\tau}}{\sum_{b\in B(i)} e^{cosine(\Tilde{x_i}, \Tilde{x_b})/\tau}}\] Where [\(p(i)\)]{style="color: black"} contains positive samples (samples with a similar label with the anchor), \(B(i)\) has negative samples (samples with a different label with the anchor), [\(\tau\)]{style="color: black"} is the scaling term, and \(cosine(a,b)\) is the similarity function between \(a\) and \(b\) (see equation [\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}). \[\label{eq:2} cosine(A, B) =\frac{A.B}{||A||||B||}=\frac{\sum^{n}_{i=1}A_iB_i}{\sqrt{\sum_{i=1}^{n}A_i^2}\sqrt{\sum_{i=1}^{n}B_i^2}}\] This method is used to pre-train the encoder of the RoBERTa transformer to encode the textual data into a contrastive representation before passing it to the classification layer. To enhance the performance of this model on semi-supervised applications, this model is used as the base model in the Active Learning mechanism. Figure [\[fig:ac\]](#fig:ac){reference-type="ref" reference="fig:ac"} shows the proposed model. To make the proposed method [free from human intervention]{style="color: black"}, the Active Learning mechanism is changed in a way to remove the expert person from it. In the proposed method, in each iteration of the Active Learning mechanism, the classified samples with reasonable confidence (based on the softmax layer) are added to the labeled samples and used to train the [CRL]{acronym-label="CRL" acronym-form="singular+short"}-based RoBERTa for the next iteration. This process continues until the model gets a pre-defined performance or a specific number of iterations is passed. # Experimental Setup {#sec:em} ## Dataset To evaluate the proposed method for labeling and classification of the insurance-based textual data, a dataset consisting of obituary texts is used to predict the cause of death. However, among more than 2,500,000 samples, only 3% of them were labeled. Table [4](#tbl:labeled_data){reference-type="ref" reference="tbl:labeled_data"} shows the number of labeled data for each class in the mentioned dataset. [\[tbl:labeled_data\]]{#tbl:labeled_data label="tbl:labeled_data"} ## Evaluation Metrics The basic [ML]{acronym-label="ML" acronym-form="singular+short"} metrics are [TP]{acronym-label="TP" acronym-form="singular+short"}, [TN]{acronym-label="TN" acronym-form="singular+short"}, [FP]{acronym-label="FP" acronym-form="singular+short"}, and [FN]{acronym-label="FN" acronym-form="singular+short"}, which represent the number of samples correctly classified as positive, correctly classified as negative, wrongly classified as positive, and wrongly classified as negative, respectively. Using these basic metrics, more complex metrics, including Accuracy, Precision, Recall, and F-measure, are defined and used to quantify the performance of [ML]{acronym-label="ML" acronym-form="singular+short"} algorithms. \[Accuracy=\dfrac{TP+TN}{TP+TN+FP+FN} \label{eq:det_acc}\] \[Precision=\dfrac{TP}{TP+FP} \label{eq:det_pre}\] \[Recall=\dfrac{TP}{TP+FN} \label{eq:det_rec}\] \[F-measure=\dfrac{2\times Precision\times Recall}{Precision+Recall} \label{eq:det_fm}\] - Accuracy indicates the number of correctly classified samples over the entire dataset (see equation [\[eq:det_acc\]](#eq:det_acc){reference-type="ref" reference="eq:det_acc"}). - Precision indicates the number of samples classified correctly as each class label over total samples classified for that class (see equation [\[eq:det_pre\]](#eq:det_pre){reference-type="ref" reference="eq:det_pre"}). - Recall indicates the number of samples classified as each class label correctly over the total instances of the dataset for that class label (see equation [\[eq:det_rec\]](#eq:det_rec){reference-type="ref" reference="eq:det_rec"}). - F-measure is the harmonic value of precision and recall (see equation [\[eq:det_fm\]](#eq:det_fm){reference-type="ref" reference="eq:det_fm"}). # Results {#sec:res} The proposed combination of the supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"}, RoBERTa transformer, and Active Learning is evaluated on an obituary dataset introduced in Section [12](#sec:em){reference-type="ref" reference="sec:em"}. Table [5](#tbl:pmres){reference-type="ref" reference="tbl:pmres"} shows the performance of the proposed method. This table illustrates that the proposed method can accurately classify textual insurance data for all class labels. Moreover, the proposed method is compared with the RoBERTa transformer with supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} and Active Learning with the RoBERTa base model. As illustrated in Table [6](#tbl:compare){reference-type="ref" reference="tbl:compare"}, the proposed method outperformed both these models. This table shows that the supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} model significantly outperformed the Active Learning model. However, adding Active Learning to this model empowers it to detect the cause of death more accurately. [\[tbl:labeled_data\]]{#tbl:labeled_data label="tbl:labeled_data"} ## Evaluation Metrics The basic [ML]{acronym-label="ML" acronym-form="singular+short"} metrics are [TP]{acronym-label="TP" acronym-form="singular+short"}, [TN]{acronym-label="TN" acronym-form="singular+short"}, [FP]{acronym-label="FP" acronym-form="singular+short"}, and [FN]{acronym-label="FN" acronym-form="singular+short"}, which represent the number of samples correctly classified as positive, correctly classified as negative, wrongly classified as positive, and wrongly classified as negative, respectively. Using these basic metrics, more complex metrics, including Accuracy, Precision, Recall, and F-measure, are defined and used to quantify the performance of [ML]{acronym-label="ML" acronym-form="singular+short"} algorithms. \[Accuracy=\dfrac{TP+TN}{TP+TN+FP+FN} \label{eq:det_acc}\] \[Precision=\dfrac{TP}{TP+FP} \label{eq:det_pre}\] \[Recall=\dfrac{TP}{TP+FN} \label{eq:det_rec}\] \[F-measure=\dfrac{2\times Precision\times Recall}{Precision+Recall} \label{eq:det_fm}\] - Accuracy indicates the number of correctly classified samples over the entire dataset (see equation [\[eq:det_acc\]](#eq:det_acc){reference-type="ref" reference="eq:det_acc"}). - Precision indicates the number of samples classified correctly as each class label over total samples classified for that class (see equation [\[eq:det_pre\]](#eq:det_pre){reference-type="ref" reference="eq:det_pre"}). - Recall indicates the number of samples classified as each class label correctly over the total instances of the dataset for that class label (see equation [\[eq:det_rec\]](#eq:det_rec){reference-type="ref" reference="eq:det_rec"}). - F-measure is the harmonic value of precision and recall (see equation [\[eq:det_fm\]](#eq:det_fm){reference-type="ref" reference="eq:det_fm"}). # Results {#sec:res} The proposed combination of the supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"}, RoBERTa transformer, and Active Learning is evaluated on an obituary dataset introduced in Section [12](#sec:em){reference-type="ref" reference="sec:em"}. Table [5](#tbl:pmres){reference-type="ref" reference="tbl:pmres"} shows the performance of the proposed method. This table illustrates that the proposed method can accurately classify textual insurance data for all class labels. Moreover, the proposed method is compared with the RoBERTa transformer with supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} and Active Learning with the RoBERTa base model. As illustrated in Table [6](#tbl:compare){reference-type="ref" reference="tbl:compare"}, the proposed method outperformed both these models. This table shows that the supervised [CRL]{acronym-label="CRL" acronym-form="singular+short"} model significantly outperformed the Active Learning model. However, adding Active Learning to this model empowers it to detect the cause of death more accurately. ::: [\[tbl:compare\]]{#tbl:compare label="tbl:compare"} # Conclusion {#sec:con} Insurance companies gathered enormous amounts of textual data through different channels. This information can help insurance companies to [perform highly complex advanced analytics]{style="color: black"} using data mining and machine learning. In this [paper]{style="color: black"}, CRL+, a semi-supervised active contrastive representation learning model is proposed to map the semi-supervised data into a contrastive space. This learned representation can be used for classification or regression models and also sequence-to-sequence applications. The proposed method is evaluated using an obituary dataset to classify the cause of death based on the document's content and compared with a modifier Active Learning and contrastive representation learning methods using RoBERTa base models. The experiments show that the proposed method outperforms the others in all metrics. The proposed method could be improved by using a two-step pre-processing and adding self-supervised contrastive learning. Moreover, the Active Learning part of the algorithm could be modified to make it more efficient.

{'timestamp': '2023-02-10T02:02:27', 'yymm': '2302', 'arxiv_id': '2302.04343', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04343'}

# Introduction Electrons accelerated in the forward shocks of young supernova remnants (SNR) can emit synchrotron radiation. This emission is mostly seen at radio wavelengths, but can also be seen in X-rays thanks to the fast shocks of young SNRs. In some cases, the X-ray brightness fades downstream and produces a bright shell-like structure of rims and filaments. A better understanding of these rims could help constrain the downstream magnetic field structure. Current models use three physical effects to account for the existence of the synchrotron filaments in X-ray: advection (bulk motion of the plasma), diffusion (random motion of electrons at small scales), and magnetic damping (where the X-ray emission reflects the magnetic field morphology). The relative importance of these effects has an influence on the evolution of the filaments widths with energy. ## The evolution of rim widths with energy The models describing the formation of X-ray synchrotron rims can mostly be divided into two main categories, depending on how they take magnetic damping effects into account. In the "loss-limited" models, the magnetic field is considered constant over the width of the rim. Electrons only travel a certain distance until they lose enough energy through diffusion and advection for their radiation to drop below the X-ray band. If the magnetic field is constant downstream, most energetic electrons will radiate and cool more quickly through advection, which results in a narrowing of the synchrotron rims with energy. However, diffusion will dilute this effect: more energetic electrons may diffuse further than would be expected from pure advection, weakening the energy dependence of the widths at higher energies. Hence, \"loss-limited\" models predict a narrowing of the X-ray synchrotron filaments, with a weakening of the energy dependence with energy. "Loss-limited" models require a strong magnetic field amplification to successfully account for the formation of thin rims. proposed that the rim profiles could reflect the magnetic field morphology, adding a magnetic damping mechanism to the effects of diffusion and advection. The magnetic field could be damped downstream of the shock and prevent electrons from radiating efficiently, thus reducing synchrotron flux at all wavelengths. Thus, a damping mechanism is supposed to produce relatively energy-independent rim widths below a threshold energy and may decrease or increase once advection and/or diffusion controls rim widths. Hence, the evolution of these widths with energy is a constraint on the magnetic field amplification and the amount of damping. However, there are various physically possible damping mechanisms, so the relationship between the filaments widths and the actual magnetic field variations is not straightforward. observed synchrotron narrowing in SN\(1006\) and concluded that it was too strongly energy-dependent to be well described by the damping mechanism only: the damping lengths would need to be larger than the synchrotron-loss lengths. conducted a similar study in Tycho's SNR. They proposed a range of radio and X-ray rim profiles showing the influences of both the magnetic field and the damping length. Their work highlights the similarity between the effects of a strong magnetic field and a small damping length on lowering the dependence in energy of the rim widths, showing that it is difficult to probe a model with no further information on the magnetic field. In the same paper, a moderate narrowing of the rim widths was found in Tycho's SNR in X-rays, but this was insufficient to constrain the model. However, "loss-limited" models alone cannot account for the formation of thin radio filaments: hydrodynamic models with diffusive shock acceleration cannot produce radio profiles with narrow rims in a purely advected magnetic field. The observation of thin radio filaments hence proved that there is a damping effect in Tycho, which is not necessarily sufficient to account for the narrowing in X-rays. ## Thin X-ray rims in Cassiopeia A Cassiopeia A (hereafter, Cas A) is among the most studied astronomical objects at X-ray wavelengths. It benefits from extensive observations (about \(3\) Ms in total with *Chandra*) and it is surrounded by a synchrotron shell showing filamentary structures, making it an ideal laboratory to investigate the potential narrowing of synchrotron rims with energy. investigated some properties of the synchrotron rims visible at the forward shock in the \(1\) Ms *Chandra* observation of Cas A, including a comparison between the linear profiles of some filaments in the \(0.3\)-\(2.0\) keV, \(3.0\)-\(6.0\) keV and \(6.0-10.0\) keV energy bands. They only found a slight difference in the widths between the \(0.3\)-\(2.0\) keV and the \(3.0\)-\(6.0\) keV images, but none between the \(3.0\)-\(6.0\) keV and the \(6.0\)-\(10.0\) keV images, as shown in Table \(1\) of their paper. They also found that the shape of the decline in emission downstream of the shock was similar to that upstream, contrary to model predictions. Here, we obtained more detailed images by using a new method to retrieve accurate maps of the synchrotron emission around different energy bands. This method is based on the General Morphological Components Analysis, a blind source separation algorithm that was introduced for X-ray observations by. It can disentangle spectrally and spatially mixed components from an X-ray data cube of the form \((x,y,E)\). The new images thus obtained suffer less contamination by other components, such as thermal emission from lines or continuum. An updated version of this algorithm, the pGMCA, has been developed to take into account the Poissonian nature of X-ray data. It was first used on Cas A data to probe the three-dimensional morphological asymmetries in the ejecta distribution, and proved perfectly suited for producing clear, detailed and unpolluted images of both the ejecta and the synchrotron at different energies. Thanks to these new images, we are able to study the profiles of some filamentary structures associated with the forward shock, as well as find some associated with the reverse shock. We labelled "upstream" and "downstream" the sides of the filament profiles according to their location relative to the shock they are associated with. However, the widths of the profiles do not necessarily correspond to the actual width of the upstream and downstream shock as projection effects might have an effect. This paper is structured as follows. In Section [2](#sect:define-prof){reference-type="ref" reference="sect:define-prof"}, we will show the images of the synchrotron we used, and the way the filaments profiles are defined. In Section [3](#sect:gaussian){reference-type="ref" reference="sect:gaussian"}, we will present a way to quantify the narrowing of the filaments, and discuss our results. # Using pGMCA to probe the synchrotron rims widths in Cas A {#sect:define-prof} Being one of the brightest sources in the X-ray sky, Cas A is the perfect extended source to showcase the capabilities of pGMCA. Cas A benefits from years of extensive observations, and offers high statistics and strongly overlapping components that the algorithm is well suited to disentangle. Here, we used it to obtain three images of the synchrotron at different wavelengths, from which we will derive the linear profiles of some wisely chosen synchrotron rims. ## Images definition For our study, we used *Chandra* observations of the Cas A SNR, which was observed with the ACIS-S instrument in 2004 for a total of 980 ks. We used only the 2004 data set to avoid the need to correct for proper motion across epochs. The event lists from all observations were merged in a single data cube. The spatial bin size is the native *Chandra* bin size of \(0.5\) arcseconds, in order to produce images as detailed as possible. As we are not interested in the spectral lines, we chose a spectral bin size of \(58.4\) eV to keep a good number of counts in every pixels. We applied the pGMCA algorithm on three bandwidths: between \(0.4\) and \(1.7\) keV, between \(2.5\) and \(4\) keV and between \(5\) and \(8\) keV. We chose these bandwidths on different criteria: large enough for pGMCA to work properly, not too large in order not to affect a possible dependency in energy of the filament widths, and we tried as much as possible to avoid any line emission in the same bandwidths. This last criterion could not be fulfilled at lower energies, and some pollution from other emission can be seen in the first image displayed in Fig. [\[fig:images_sync\]](#fig:images_sync){reference-type="ref" reference="fig:images_sync"}, particularly in the southwest. The two other images seem clear and unpolluted, even with a square root scaling. They probably constitute the most detailed and accurate maps of the synchrotron in X-rays in Cas A to this day, especially at low energy. These images present filamentary structures throughout the ejecta, including some following the known layout of the reverse shock. In order to assess the non-thermal nature of these filaments, we extracted the spectrum of one and fitted a `phabs*powerlaw` model in *Xspec*. The results are shown in Fig. [\[fig:images_filament\]](#fig:images_filament){reference-type="ref" reference="fig:images_filament"}, together with a spectrum extracted from a bright region with low synchrotron level fitted with the same model. The filament presents a spectrum that can be well-fit with a simple power-law nonthermal model, while the other region obviously cannot. This is consistent with the results from, where synchrotron emission from filaments inside of the remnant was attributed to the reverse shock. ## The filaments linear profiles {#sect:profilesFS} Thanks to the highly detailed images of the synchrotron emission we were able to find using pGMCA, we could investigate the narrowing with energy of the filaments both at the forward shock and, for the first time, at the reverse shock. In order to compare the widths of the rims, we defined boxes surrounding small regions crossing a filament at the forward shock, and did likewise for the reverse shock. The boxes are shown in Fig. [\[fig:images_sync\]](#fig:images_sync){reference-type="ref" reference="fig:images_sync"}, and the normalized linear profiles obtained perpendicularly to the rims are presented in Fig. [\[fig:profilesFS\]](#fig:profilesFS){reference-type="ref" reference="fig:profilesFS"} for the forward shock, and in Fig. [\[fig:profilesRS\]](#fig:profilesRS){reference-type="ref" reference="fig:profilesRS"} for the reverse shock. We also looked at some filaments whose positions could not allow us to label clearly. A line-of-sight effect is likely at stake, and these filaments could either be attached to the forward or to the reverse shock. The boxes are also shown in Fig. [\[fig:images_sync\]](#fig:images_sync){reference-type="ref" reference="fig:images_sync"}, in blue, and the linear profiles obtained perpendicularly to the rim are presented in Fig. [\[fig:profilesother\]](#fig:profilesother){reference-type="ref" reference="fig:profilesother"}. # Quantifying the narrowing of the synchrotron rims {#sect:gaussian} ## Modeling the filaments linear profiles In order to measure the widths of the filament profiles we extracted, we fitted them with a piecewise two-exponential model, as in  : \[h(r)= \left\{ \begin{array}{ll} A_d \exp\Big(\frac{r-r_0}{w_d}\Big) + C_d, \quad r < r_0 \\ A_u \exp\Big(\frac{r-r_0}{w_u}\Big) + C_u, \quad r \geq r_0 \end{array} \right.\] All parameters are free, except for \(A_d\) that is fixed to ensure continuity at \(r=r_0\) , with \(A_d=A_u+(C_u-C_d)\). This model is well adapted to describe the sharp peak displayed by most profiles, but some filaments (such as the \(6\)th or \(7\)th of the forward shock) present plateaus or gaussian-like features that are not accounted for in this model. However, this model remains overall a good way to estimate the widths of the profiles we extracted without neglecting the asymmetry around the peak, between the upstream and downstream media. The resulting fitted models are displayed in Fig. [\[fig:profilesFS\]](#fig:profilesFS){reference-type="ref" reference="fig:profilesFS"} (forward shock), Fig. [\[fig:profilesRS\]](#fig:profilesRS){reference-type="ref" reference="fig:profilesRS"} (reverse shock) and Fig. [\[fig:profilesother\]](#fig:profilesother){reference-type="ref" reference="fig:profilesother"} (other profiles). In some cases, we had to remove secondary peaks in the profiles to focus on the main one when fitting our model. From the parameters of our model we can derive two Full Widths at Half Maximum (FWHM) for each profile, FWHM\(_u=2\ln({2})w_u\) and FWHM\(_d=2\ln({2})w_d\), describing the \"sharpness\" of the profile on both sides, a larger FWHM meaning a wider profile. The mean between these FWHMs gives an estimation of the actual width of the profile. We also define \(m_E\) : \[m_E=\frac{\ln\big({\text{FWHM}_2/\text{FWHM}_1}\big)}{\ln\big({E_2/E_1}\big)}\] Where FWHM\(_1\) and FWHM\(_2\) are the mean FWHMs of a same filament in two energy bands, and \(E_1\) and \(E_2\) are the lower energy values for each energy band. For each filament, we calculate the \(m_E\) between the \(0.4\)-\(1.7\) keV and \(2.5\)-\(4.0\) keV energy bands (\(E_1=0.4\), \(E_2=2.5\)), and between the \(2.5\)-\(4.0\) keV and \(5.0\)-\(8.0\) keV energy bands (\(E_1=2.5\), \(E_2=5.0\)). This parameter aims to evaluate the narrowing of the filaments widths and quantify its dependence on energy: positive values mean widening, negative means narrowing, larger values mean higher energy-dependence while weaker values mean weaker energy-dependence. The FWHM\(_u\), FWHM\(_d\), mean FWHM and \(m_E\) values derived from our fitted models are shown in Table [\[tab:profilesFS\]](#tab:profilesFS){reference-type="ref" reference="tab:profilesFS"} (forward shock) Table [\[tab:profilesRS\]](#tab:profilesRS){reference-type="ref" reference="tab:profilesRS"} (reverse shock) and Table [\[tab:profilesother\]](#tab:profilesother){reference-type="ref" reference="tab:profilesother"} (other profiles). As the "upstream" and "downstream" labels do not make sense for the profiles that were not clearly identified, we named both sides "right" and "left," the orientations corresponding to the plots of Fig. [\[fig:profilesother\]](#fig:profilesother){reference-type="ref" reference="fig:profilesother"}. As there is no straightforward way to estimate the pGMCA algorithm errors, the errors shown in our Tables are the model fitting errors and their propagation in the calculation of the mean FWHM and \(m_E\) values. ## Discussion A quick look at the mean FWHMs or at the \(m_E\) signs shows that there is indeed a narrowing of the filaments profiles with energy. In all of the seven forward shock profiles, five reverse shock profiles, and five unidentified profiles, only two present a widening between \(0.4\)-\(1.7\) keV and \(5.0\)-\(8.0\) keV: FS \(7\) and unidentified profile \(2\). In both cases, this widening is so low that the fitting errors allow for the possibility of a narrowing as well. Hence, we can reasonably conclude that the narrowing of the filaments with energy in Cas A is a global effect, that can be observed on several filaments both at the forward and at the reverse shock. In Sect. [3.3](#sect:PSF){reference-type="ref" reference="sect:PSF"} we will see that this narrowing is not due to the evolution of Chandra's point spread function (PSF) with energy. Future studies could also take into account the possible effects of dust scattering on the observed widths of the filaments. However, in first approximation, scattering effects have an energy dependence in \(E^{-2}\), which is not consistent with our observations. Hence, the narrowing we observe is likely not primarily due to dust scattering. We can see in the results a common trend regarding the evolution of \(m_E\) values: the mean \(m_E\) between \(2.5\)-\(4.0\) keV and \(5.0\)-\(8.0\) keV are significantly larger, in absolute, than the mean \(m_E\) between \(0.4\)-\(1.7\) keV and \(2.5\)-\(4.0\) keV. While this result is to handle cautiously, given the importance of the errors, the low statistics on which the means are calculated and the non-negligible number of outliers, it would mean that the observed narrowing of the synchrotron rims has a stronger energy-dependence at higher energies than at lower energies. Following, this result would be characteristic of a damping mechanism: the rim widths are relatively energy-independent below a threshold energy and increase once advection controls rim widths. This is in apparent contradiction with, where a moderate narrowing was observed between \(0.3\)-\(2.0\) keV and \(3.0\)-\(6.0\) keV, but not between \(3.0\)-\(6.0\) keV and \(6.0\)-\(10.0\) keV. However, the energy bands we compared are not the same, and our "high energy \(m_E\)" is defined between \(2.5\) and \(5.0\) keV, while theirs is between \(3.0\) and \(6.0\) keV. The highly detailed synchrotron images we obtained with pGMCA may also have an influence, allowing for a more precise profile definition and width measurement with less thermal emission contamination. In, the model used to describe the forward shock filament profiles predicts a sharp decline upstream, after the peak, that they did not observe. We can see from the FWHM\(_u\) and FWHM\(_d\) values from Table [\[tab:profilesFS\]](#tab:profilesFS){reference-type="ref" reference="tab:profilesFS"} that we did not observe it either. The FWHM\(_u\) values even tend to be larger than FWHM\(_d\) values, meaning that the profiles are sharper downstream than upstream of the forward shock. On the contrary, we can see in Table [\[tab:profilesRS\]](#tab:profilesRS){reference-type="ref" reference="tab:profilesRS"} that at the reverse shock, FWHM\(_u\) values are mainly smaller than FWHM\(_d\) values, meaning that the profiles are sharper upstream (even though it is not apparent in the mean FWHM\(_u\) and FWHM\(_d\) values, that are driven by a few outliers). However, we will see in Sect. [3.3](#sect:PSF){reference-type="ref" reference="sect:PSF"} that this could be linked to the PSF. Not much can be said to identify our "other" profiles as belonging to the forward or to the reverse shock. Although it could be tempting to attempt an identification based on their FWHM on each side, line of sight effects are likely involved and for a filament facing the observer, "left" and "right" of the peak are not trivially equivalent to "upstream" or "downstream". Nonetheless, Table [\[tab:profilesFS\]](#tab:profilesFS){reference-type="ref" reference="tab:profilesFS"} mostly follows the same trend of narrowing, with stronger energy-dependence at high energies, that we observed on both forward and reverse shock rim profiles. ## Influence of the PSF {#sect:PSF} Previous studies by, and did not take into account the PSF in their filaments widths measurements. Yet, the PSF evolves with the energy, and could have an impact on our widths comparison between energy ranges. In a first attempt to take it into account, we generated the 1-sigma PSF maps of the merged 2004 observations of Cas A using the `merge_obs` routine from CIAO around the \(0.4\)-\(1.7\) keV, \(2.5\)-\(4.0\) keV and \(5.0\)-\(8.0\) keV energy bands. We then intended to convolve our piecewise two-exponential model to the PSF profiles along each box for each energy range while fitting. However, the results we obtained were endowed with disproportionate errors due to the additional uncertainty brought by the spreading. Hence, we renounced taking the PSF directly into account in our fitting, and decided to present a "worst case scenario" to give an idea of the possible effects of the PSF on the results shown in Tables [\[tab:profilesFS\]](#tab:profilesFS){reference-type="ref" reference="tab:profilesFS"}, [\[tab:profilesRS\]](#tab:profilesRS){reference-type="ref" reference="tab:profilesRS"} and [\[tab:profilesother\]](#tab:profilesother){reference-type="ref" reference="tab:profilesother"}. To do so, we searched the box along which the PSF evolved the most, both spatially and with energy, which was the third box of our FS profiles, the furthest from the optical center. We then convolved the profiles of the three energy ranges PSF maps along this box with an infinitely thin filament, i.e. a Dirac function. The results were then fitted with our model, and the differences between the retrieved FWHMs for each energy range give an idea of the influence of the PSF in the worst scenario, with infinitely thin filaments in our worst box as regards to the PSF. The results are shown in Fig. [\[fig:PSF_test\]](#fig:PSF_test){reference-type="ref" reference="fig:PSF_test"}, and it appears that the PSF can potentially have a significant influence. The FWHMs upstream are consistently larger upstream than downstream, and the FWHMs on both sides increase with energy. It is important to note that the PSF profiles all behave in a similar way: the spreading radius values increase both with energy and with the distance from the center (downstream to upstream for the forward shock, upstream to downstream for the reverse shock). Hence, the narrowing of the filaments with energy we observed might be underestimated because of the PSF, and our observations regarding the sharper decline downstream than upstream of the forward shock might be due to the PSF. The \(m_E\) between the \(0.4\)-\(1.7\) keV and \(2.5\)-\(4.0\) keV profiles is \(0.05 \pm 0.03\) and the \(m_E\) between the \(2.5\)-\(4.0\) keV and \(5.0\)-\(8.0\) keV profiles is \(0.22 \pm 0.08\), indicating that the PSF would tend to widen the filaments more at higher energies. ## Interpretation The calculated \(m_E\) values tend to indicate a stronger energy-dependence of the narrowing at high energies, which is likely underestimated by the PSF. According to, this would be characteristic of a damping mechanism, where the rims widths are supposed to be energy-independent below a threshold energy, and decrease or increase once advection and/or diffusion controls the widths. Following the same paper, the action of a damping mechanism can be more confidently assessed by the observation of thin synchrotron filaments in radio. In the radio VLA observations of Cas A, such filaments can be seen, which also suggests that damping effects are at stake in this SNR's synchrotron emission. A thorough joint study of both radio and X-ray filaments could allow for an estimation of the damping lengths. # Conclusions From our study of the filament profiles in Cas A, we can conclude the following : 1. Our blind source separation method was able to produce clear, detailed and unpolluted maps of the synchrotron emission around three energy bands: \(0.4\)-\(1.7\) keV, \(2.5\)-\(4.0\) keV, and \(5.0\)-\(8.0\) keV. These images clearly show filamentary structures all around the remnant, some associated to the forward shock, and some to the reverse shock. Some profiles can also be seen that cannot be clearly attributed to the forward or to the reverse shock, possibly because of line of sight effects. 2. There is a noticeable narrowing of the synchrotron rims from lower to higher energies. Contrary to a previous study by, we find that this effect is also visible between \(2.5\)-\(4.0\) keV and \(5.0\)-\(8.0\) keV, not only at lower energies. It is observed for filaments both at the forward and at the reverse shock, and for non-identified filaments. The evolution of the PSF with energy cannot account for this narrowing; on the contrary, the PSF would tend to make filaments appear wider with energy. 3. There seems to be a stronger energy-dependence of this narrowing at higher energies, which would be a sign that there is a damping mechanism at stake. The observation of thin synchrotron profiles in the radio 6cm VLA observation of Cas A is another indicator of this damping mechanism. However, it does not exclude the possible influence of the loss-limited effect as well. 4. The filament profiles at the forward shock tend to be sharper downstream than upstream. At the reverse shock, the profiles mainly tend to be sharper upstream. However, this could be due to the radial degradation of the PSF away from the aim point. It could also be caused by projection effects that do not have the same impact on the forward and the reverse shock as they propagate in opposite directions.

{'timestamp': '2023-02-10T02:02:42', 'yymm': '2302', 'arxiv_id': '2302.04352', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04352'}

# 3D Numerical Solution of Active Polar Fluids {#simdetails} In Einstein summation notation, the incompressible viscous active polar fluid equations are : with constants as described in the main text. We decompose the molecular field \(\mathbf{h}\) into parallel and perpendicular components, The vector \(\mathbf{h_\perp}\) is computed from the variational derivative of the Frank free energy density \[F_{3D}=\frac{K_s}{2} (\nabla\cdot \mathbf{p})^2+ \frac{K_t}{2} (\mathbf{p}\cdot \nabla\times \mathbf{p}){}^2 + \frac{K_b}{2} (\mathbf{p}\times(\nabla \times \mathbf{p}))^2-\frac{1}{2}h^0_{\Vert}\Vert \mathbf{p}\Vert^2 \label{eq:freeEnergy3d}\] with respect to \(\mathbf{p}\). The total stress \(\sigma^{(\mathrm{tot})}_{\alpha\beta}=\sigma^{(\mathrm{s})}_{\alpha\beta}+\sigma_{\alpha \beta}^{(\mathrm{ant})}+\sigma^{(\mathrm{e})}_{\alpha\beta}\) is decomposed as the sum of the symmetric (\(\mathrm{s}\)), antisymmetric (\(\mathrm{ant}\)), and equilibrium (\(\mathrm{e}\)) stresses. The equilibrium stress, also called the Ericksen stress, is given by \[\sigma_{\alpha \beta}^{(\mathrm{e})}=-\frac{\partial F_{3D}}{\partial\left(\partial_{\beta} p_{\gamma}\right)} \partial_{\alpha} p_{\gamma},\] with \(F_{3D}\) from Eq. ([\[eq:freeEnergy3d\]](#eq:freeEnergy3d){reference-type="ref" reference="eq:freeEnergy3d"}). The anti-symmetric stress is \[\sigma_{\alpha \beta}^{(\mathrm{ant})}=\frac{1}{2}\left(p_{\alpha} h_{\beta}-p_{\beta} h_{\alpha}\right).\] Setting \(p_\gamma \frac{Dp_\gamma}{Dt}=0\) to maintain constant polarity magnitude \(p_\gamma p_\gamma\), we derive the Lagrange multiplier \[h_\Vert=-\gamma\Big[\lambda \Delta \mu-\frac{\nu}{p_x^2+p_y^2+p_z^2}\Big(u_{xx}p_x^2+u_{yy}p_y^2+u_{zz}p_z^2+ 2u_{xy}p_xp_y+2u_{yz}p_yp_z+2u_{xz}p_xp_z \Big)\Big]. \label{eq:lag}\] Substituting the decomposition of \(\mathbf{h}\) with the Lagrange multiplier from Eq. ([\[eq:lag\]](#eq:lag){reference-type="ref" reference="eq:lag"}) and combining it with the force-balance Eq. ([\[eqP2\]](#eqP2){reference-type="ref" reference="eqP2"}), we derive the steady-state component-wise Stokes flow equations that are implemented in computer code using a custom C++ expression system in the scalable scientific computing library OpenFPM . At time 0, the polarity is homogeneously aligned with the anchoring boundary condition except a point perturbation of 0.001 radians in both positive Y and Z directions at \(x=L/2\) to break the symmetry. The time evolution of the polarity is computed using Adams-Bashforth-Moulton predictor-corrector time integration with a time step of 0.01 and renormalization of the slopes. The steady state is detected with a tolerance of \(10^{-8}\). Note that near the critical activity, the transition can be very slow and difficult to catch numerically. Hence, for the simulations shown in the main text, we used a more accurate direct solver for the velocity from the MUMPS library, which is based on LU-decomposition, and we increased the spatial resolution from \(18\times19\times5\) to \(64\times65\times5\) grid points for higher accuracy. Further, a smaller absolute tolerance of \(10^{-11}\) and relative tolerance of \(10^{-9}\) were used for adaptive time stepping of the same stepper. The velocity field is computed by iteratively correcting pressure and solving the implicit system of incompressible Ericksen-Leslie Stokes equations with hydrodynamic stress-free boundary conditions and the constraint of no flow at \(x=L/2\), \(y=L/2\). At each time step, the resulting linear system of equations is solved numerically using the iterative GMRES solver as implemented in the PetSc software library . We checked that using a higher resolution yields the same results, confirming grid convergence. # Derivation of Critical Activity Follwing, we derive \(\mathbf{h_\perp}\) from the model equations at steady state and then equate it to the \(h_\perp\) based on the Frank free energy to analyze the response to a perturbation in 3D. We consider a thick film that is infinitely extended along the X and Z directions and has a thickness of \(L\) in the Y direction. The surface of the film at \(y = L\) and \(y=0\) is stress-free (\(\sigma_{xy} = 0\) and \(\sigma_{yz} = 0\)), and impenetrable \((v_y (x, y, z, t) = 0)\). The polarity is fixed on the top (\(y=L\)) and bottom (\(y=0\)) such that \((p_x, p_y, p_z) = (\cos (\theta_0) \cos (\phi_0) ,\, \sin (\theta_0) \cos (\phi_0) ,\, \sin (\phi_0) )\). Under these conditions, \(v_y = 0\) everywhere due to incompressibility and translation invariance in X and Z directions. Further, \(~u_{xy}=\partial_yv_x,~u_{yz}=\partial_yv_z\) and \(u_{xx}=u_{yy}=u_{zz}=u_{xz}=0\). Fixing polarity to be perpendicular to the boundary wall, i.e., \((\theta_0,\phi_0)=(\pi/2,0)\), and assuming small perturbations \(\epsilon(y),\kappa(y)\), the restoring force up to linear order of tilt is \(\mathbf{h_\perp}=(K\frac{\partial \kappa(y)}{\partial y^2},0,K\frac{\partial \epsilon(y)}{\partial y^2})\), where \(K=K_s=K_t=K_b\) is the elastic constant in the single-constant approximation of the Frank free energy. Using Eq. ([\[eqP4\]](#eqP4){reference-type="ref" reference="eqP4"}) and imposing \(\sigma_{xy}^{(tot)}=\sigma_{zy}^{(tot)}=0\), we obtain the strain rates \(u_{xy}, u_{yz}\) and substitute them in Eq. ([\[eqP1\]](#eqP1){reference-type="ref" reference="eqP1"}) to obtain the force associated with \(\mathbf{h_\perp}\). The so-obtained non-linear equation is decoupled from the flow and only depends on \((\theta,\phi)\). We do not reproduce this equation here due to its excessive length. Substituting into this equation a small perturbation \((\epsilon(y),\kappa(y))\) and linearizing around \((\theta_0,\phi_0)=(\frac{\pi}{2},0)\), we obtain the dynamical equation of the perturbation as described in the main text. This leads to a spontaneous flow transition under extensile active stress. Up to linear orders of tilt, the strain rates in this case are: \[\frac{\partial^2}{\partial y}\begin{bmatrix}v_x(y) \\ v_z(y) \end{bmatrix}=\frac{2 \Delta\mu (\gamma \lambda \nu +\zeta )}{\gamma (\nu-1)^2+4 \eta } \begin{bmatrix} -\epsilon (y) \\ \kappa(y) \end{bmatrix}. \label{eq:mu1v}\] We repeat the analysis for \((\theta_0,\phi_0)=(0,0)\) and obtain the critical activity of a 3D spontaneous flow transition under contractile active stress. In this case, the restoring force up to linear order of tilt is \(\mathbf{h_\perp}=(0,-K\frac{\partial \kappa(y)}{\partial y^2},K\frac{\partial \epsilon(y)}{\partial y^2})\). Up to linear order of tilt, the strain rate in this case is: \[\frac{\partial v_x(y)}{\partial y}=\frac{2 \Delta\mu (\gamma \lambda \nu +\zeta )}{\gamma (\nu +1)^2+4 \eta }\epsilon(y).\] Assuming \(u_{xz}\neq0\), \(u_{yz}\neq0\), but \(u_{xy}=0\), thus allowing polarity to vary in both the X and Y directions, \((\theta(x,y),\phi(x,y))\), and using \(\mathbf{h_\perp}=(0,-K\Laplace_{\{x,y\}}\kappa(x,y),0)\), we find the two-dimensional perturbation mode that corresponds to the out-of-plane wrinkling transition under extensile active stress as described in the main text. In this case, the strain rate up to linear order of tilts is \[\frac{\partial v_z(x,y)}{\partial x}=\frac{2 \Delta \mu (\gamma \lambda \nu +\zeta )}{4 \eta +\gamma \left(\nu-1\right)^2}\kappa(x,y).\]

{'timestamp': '2023-02-09T02:17:56', 'yymm': '2302', 'arxiv_id': '2302.04259', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04259'}

# Introduction {#sect:introduction} Unraveling the quantum nature of black holes is among the most important objectives of research in quantum gravity. Despite the impressive achievements of Einstein's general relativity, the singularities and instabilities characterizing classical black holes indicate that a more fundamental description ought to take over the classical framework. How such a fundamental theory of gravity looks like is an outstanding open question. Over the years, several proposals have been put forth. Insofar as quantum gravity lives at extremely high energies---above the Planck scale---testing and discriminating between theories is challenging, and theoretical consistency has become a fundamental guidance in constraining different theories. Among the variety of consistency constraints, recovering a gravitational effective field theory in the infrared (IR) starting from the deep ultraviolet (UV) is arguably one of the most important requirements, and only a few theories have managed to pass this test so far. Among them, asymptotically safe gravity  has emerged as a minimal while promising proposal, conjecturing that quantum gravity be described by a quantum field theory (QFT) whose UV behavior is controlled by an interacting fixed point of the gravitational renormalization group (RG) flow. The fixed point acts as an attractor for a subset of RG trajectories, providing a UV completion for the theory and rendering it renormalizable à la Wilson. The presence of an asymptotically safe fixed point for quantum gravity, akin to the asymptotically-free one in quantum chromodynamics, is responsible for a distinctive hallmark of asymptotic safety: its anti-screening character . On the formal side, gravitational anti-screening is tied to the attractivity properties of the fixed point; on the phenomenological side, this hallmark can be encoded in an effective Newton coupling which vanishes in the regimes where quantum-gravity effects are expected to be important. This intuitive picture has been extensively exploited in the literature to model deviations from general relativity induced by an asymptotically safe UV completion of gravity. The corresponding asymptotic-safety-inspired models are typically obtained by replacing the observed Newton constant with an effective, coordinate dependent one, which smoothly interpolates between the observed value in the IR and its fixed-point scaling in the UV. This procedure is better known as RG improvement and the resulting gravitational models are dubbed RG-improved spacetimes. The RG improvement has been a valuable instrument to explore possible implications of asymptotically safe gravity in astrophysics and cosmology, particularly at the dawn of asymptotically safe phenomenology. The method has even inspired an entirely new program, which is by now detached from asymptotically safe gravity, and goes under the name of "scale-dependent gravity" . Yet, more rigorous derivations and arguments, grounded either on the functional integral or on the effective action, are in order to unravel asymptotic-safety-induced modifications of classical black holes and early-universe cosmology. Such fundamental approaches have been the focus of the asymptotic safety program in the past few years. In this chapter we review of the state-of-the-art of black holes in asymptotically safe gravity (see also ), from the early works on RG-improved black holes to the most recent developments involving computations and considerations based on the gravitational effective action. In our narrative we will mostly be following a chronological order. The chapter is organized as follows. In Sect. [2](#sect:as-basics){reference-type="ref" reference="sect:as-basics"} we will briefly review the asymptotic safety scenario for quantum gravity, providing all basic ingredients required for the understanding of the subsequent sections on asymptotically safe black holes. The ideas behind the RG improvement, its recipe, and its most pressing issues will be the topic of Sect. [3](#sect:RG-improv-key-idea){reference-type="ref" reference="sect:RG-improv-key-idea"}. We will discuss the resulting RG-improved black holes, in all their facets, in Sect. [\[sect:rg-improved-black-holes\]](#sect:rg-improved-black-holes){reference-type="ref" reference="sect:rg-improved-black-holes"}. Attempts to ameliorating the method and removing its ambiguities will be the focus of Sect. [\[sect:refined-rg-imp\]](#sect:refined-rg-imp){reference-type="ref" reference="sect:refined-rg-imp"}, where we will also comment on the physical implications of these improvements and how they compare with past results. Sect. [\[sect:bh-from-frg\]](#sect:bh-from-frg){reference-type="ref" reference="sect:bh-from-frg"} will be devoted to the subject of black holes from the effective action: we will discuss the most important developments of the past years, which unlocked intriguing aspects of black holes within and beyond asymptotic safety, grounded on first-principle calculations in quantum gravity. We will summarize all these points in our conclusions, Sect. [\[sect:conclu\]](#sect:conclu){reference-type="ref" reference="sect:conclu"}. # Asymptotic safety in a nutshell {#sect:as-basics} Asymptotically safe gravity  is one of the most conservative approaches to quantum gravity. It relies on the framework of QFT, and conjectures that the high-energy behavior of the gravitational RG trajectory realized by Nature be controlled by a UV fixed point, where all (essential) running couplings approach a finite, non-zero value. This condition is known as "asymptotic safety", and can be regarded as a non-perturbative generalization of the well-known concept of asymptotic freedom, whereby couplings vanish in the UV. Asymptotic freedom or safety guarantee that a theory be renormalizable, as well as UV complete with respect to a free or interacting fixed point, respectively. The first case can be related to perturbative renormalizability. The second one corresponds to a generalized notion of renormalizability, often regarded as "non-perturbative renormalizability". Power-counting arguments only hold for the former, while the existence of the latter cannot be determined a priori, and must be investigated by appropriate RG techniques, typically beyond perturbation theory. One of their analytical realizations is the framework of the FRG , while corresponding lattice approaches are employed within the eucliden and causal dynamical triangulation programs . Thanks to these powerful methods, the asymptotic safety conjecture has been tested within a large number of approximations and against a variety of different starting assumptions (see  and references therein). Notably, the asymptotic safety approach to quantum gravity does not need new physics, at least in principle, as long as its introduction is not required by compelling experimental evidence. The FRG combines together two ingredients: on the one hand, the Wilsonian idea of renormalization  and, on the other hand, the functional approach introduced in QFT to perform a path integral quantization and to study scattering amplitudes. The result of this combination is a functional approach to renormalization, by which one can - Handle with any QFT, including those that are (perturbatively or power-counting) non-renormalizable but non-perturbatively renormalizable. - Determine the bare action from first principles, i.e., as an RG fixed point. - Derive the quantum effective action stemming from an RG fixed point and an appropriate number of initial conditions. To this end, one needs to introduce a scale-dependent version of the effective action, \(\Gamma_k\), dubbed the "effective average action" (EAA). Here \(k\) is an artificial RG scale and not a physical momentum. It can be shown that the (Euclidean[^1]) path integral can be translated into the following functional integro-differential equation for the effective average action, \[\label{eq:flow-eq} k\partial_k \Gamma_k = \frac{1}{2} \textrm{STr} \left(\left(\Gamma_k^{(2)}+\mathcal{R}_k \right)^{-1}k\partial_k\mathcal{R}_k\right)\,,\] better known as the Wetterich equation . Here \(\mathcal{R}_k\propto k^2\) is a RG-scale dependent mass term which is added to the original Lagrangian to regularize the path integral and to integrate quantum fluctuations with momenta \(p^2 \gtrsim k^2\). The symbol "STr" stands for a supertrace, summing over internal indices and integrating over the volume in coordinate or momentum space. Finally, \(\Gamma_{k}^{(2)}\) denotes the second functional derivative of the EAA with respect to all fields appearing in \(\Gamma_{k}\). The RG fixed points are identified by the conditions \(k\partial_k g_i(k) = 0\), where the index \(i\) labels the running couplings in \(\Gamma_k\), and \(g_i(k)\) denotes their dimensionless counterpart. A fixed point can be reached by some RG trajectories either in the IR or in the UV. Accordingly, a given fixed point provides a UV completion for all RG trajectories belonging to its basin of attraction. The dimension of the latter determines the number \(N\) of relevant directions, and thus the number of free parameters of the theory. Note that this number depends on the specific fixed point considered. Once one identifies the set of fixed points, the next question to ask is whether any of the RG trajectories departing from a fixed point in the UV can reach an IR that is compatible with experimental and observational data. This can be verified by integrating the flow equation [\[eq:flow-eq\]](#eq:flow-eq){reference-type="eqref" reference="eq:flow-eq"} complemented by a sufficient number of initial conditions dictated by observations. If a solution compatible with these initial conditions exists, and if it reaches a fixed point in the UV, then - The RG trajectory \(\Gamma_k^{sol}\) realized by the particular system analyzed is consistent at all scales and in particular it is UV complete. - The theory associated with the RG trajectory \(\Gamma_k^{sol}\) is renormalizable. - Its observables, including scattering amplitudes, can be computed using the standard effective action, which is obtained as the limit \(\Gamma_0\equiv \lim_{k\to0} \Gamma_k^{sol}\). - If the theory \(\Gamma_k^{sol}\) is UV completed by a fixed point with \(N\) relevant directions, all infinitely many couplings in the corresponding effective action \(\Gamma_0\) will be written in terms of \(N\) free parameters only. The next ingredient to discuss is the practical resolution of the Wetterich equation. Albeit this is exact and provides a clear recipe to compute the functional integral, solving it is involved and one has to resort to approximations. In particular, independent of the specific theory---identified by a set of fields and the symmetries of their interactions---the functional \(\Gamma_k\) contains infinitely many interaction terms. In order to make computations doable, one possibility is to project the RG flow onto a managable sub-space of couplings. The calculation can then be improved in a step-by-step fashion, exploring larger and larger sub-spaces. In practice, this projection is performed by expanding \(\Gamma_k\) according to a certain criterion (e.g., a derivative or vertex expansion) and by "truncating" the resulting series to a certain order. Obtaining the same results independent of the type of expansion and truncation order is considered as evidence of their stability . In the case of gravity a particularly convenient way to express \(\Gamma_k\) is via a curvature expansion of the form  \[\label{eq:EAA-quadratic} \Gamma_k=\int \dd[4]x \sqrt{-g}\left(\frac{1}{16\pi G_k}\qty(R-2\Lambda_k)+R\, g_{R,k}(\Box)\,R+ C_{\mu\nu\sigma\rho}\,g_{C,k}(\Box)\,C^{\mu\nu\sigma\rho}\right) \.\] Here \(\Box=g_{\mu\nu}D^\mu D^\nu\) is the d' Alembertian operator, and \(G_k\), \(\Lambda_k\), \(g_{R,k}(\Box)\) and \(g_{C,k}(\Box)\) are the running Newton, cosmological and quartic couplings, respectively. Only the latter two couplings can depend on the d' Alembertian operator, since in the volume term \(\Box\) cannot act on any field, while a term \(F_k(\Box)R/G_k\) is equivalent to \(R/G_k\) modulo a total derivative. Provided that a well-definite UV completion exists, the limit \(k\to0\) defines an effective action resembling closely the structure in Eq. [\[eq:EAA-quadratic\]](#eq:EAA-quadratic){reference-type="eqref" reference="eq:EAA-quadratic"}  \[\label{eq:eff-action} \Gamma_0=\int \dd[4]x \sqrt{-g}\left(\frac{1}{16\pi G_N}\qty(R-2\Lambda)+R\,g_R(\Box)\,R+ C_{\mu\nu\sigma\rho}\,g_C(\Box)\,C^{\mu\nu\sigma\rho}\right) \,,\] where \(G_N\equiv G_0\) and \(\Lambda\equiv\Lambda_0\) are the observed Newton and cosmological constants. The "form factors" \(g_{R}(\Box)\) and \(g_{C}(\Box)\) encode instead the physical momentum dependence of the quartic couplings , generalizing the relation \(\partial^2\sim-p^2\) to curved spacetimes . As the effective action [\[eq:eff-action\]](#eq:eff-action){reference-type="eqref" reference="eq:eff-action"} is the result of integrating over all quantum fluctuations, quantities computed using it already encode all quantum effects. In particular, quantum black holes and cosmologies ought to be derived as solutions to the dressed field equations \[\label{eq:full-eom} \frac{\delta \Gamma_0[g_{\mu\nu}^{sol}]}{\delta g_{\mu\nu}^{sol}}=0\.\] We now have all ingredients to introduce the so-called "RG improvement" and then to step into the main topic of this chapter. # RG improvement: key idea {#sect:RG-improv-key-idea} The RG improvement was introduced in the context of gauge and matter field theories as a short-cut to access some of terms in the effective action and, in some cases, determine leading-order modifications to the solutions to the corresponding effective field equations . The RG improvement procedure consists of the following steps. First, starting from a classical system, e.g., an action or a solution, one replaces its couplings with their running counterparts, \(g_i\to g_i(k)\). At the level of the action, this step is akin to promoting an ansatz for the action to an EAA \(\Gamma_k\). Second, the running couplings \(g_i(k)\) are replaced with the solutions to the corresponding RG equations, \(k\partial_k g_i(k)=\beta_i[g_j(k)]\), complemented by suitable physical initial conditions. Finally, \(k\) is identified with a scale of the system which could act as a physical IR cutoff. The reason why this procedure is supposed to work, at least in simple cases, lies in the so-called *decoupling mechanism* : if in the flow of \(\Gamma_k\) there are physical IR scales (e.g., masses, curvature, or interactions terms) that prevail over the unphysical regulator \(\mathcal{R}_k\) below a certain threshold scale \(k_{dec}\)---dubbed the decoupling scale, then the right-hand side of Eq. [\[eq:flow-eq\]](#eq:flow-eq){reference-type="eqref" reference="eq:flow-eq"} gets smaller, thus slowing down the RG flow; as a result, the EAA at the decoupling scale approximates the full effective action \(\Gamma_0\). This idea is illustrated in Fig. [\[fig:decoupl\]](#fig:decoupl){reference-type="ref" reference="fig:decoupl"}. [^1]: FRG computations are typically performed in Euclidean. First steps towards Lorentzian calculations have been taken in , while Lorentzian computations based on the spectral FRG have been developed and applied in .

{'timestamp': '2023-02-10T02:00:12', 'yymm': '2302', 'arxiv_id': '2302.04272', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04272'}

# Introduction Understanding how the composition of the Solar System's protoplanetary disc varied as a function of the heliocentric distance has always been a key problem of planetary science with implications on the formation of planetesimals, planets, meteorites, and the delivery of organics, water and prebiotic materials to planets. While it is now understood that the contemporary asteroid main belt consists of objects that originally accreted in the terrestrial planet and Jupiter formation regions, as well as objects that were implanted from the primordial Kuiper belt, the compositional structure of the original trans-Neptunian disc (TND) is not yet understood. It has been theoretically demonstrated that the original configuration of the TND was a low-inclination formation, starting from 23 au with a drop in density past 30 au. The TND may have had a colour gradient of trans-Neptunian objects (TNOs) increasing in redness with heliocentric distance due to the sublimation of surface volatiles such as ammonia, methanol and hydrogen sulfide. The present-day TND consists of the Hot Classical objects (HCs), comprised of bodies that may have formed within 30 au but were scattered outwards by the migration of Neptune, and the Cold Classical objects (CCs), comprised of bodies that probably formed outside of 30 au and had much-reduced interactions with Neptune. The resonant population consists of TNOs that are in mean motion resonances with Neptune located at \(>\)`<!-- -->`{=html}30 au from the Sun such as the 5:4, 4:3 and 5:3 mean motion resonances at \(\sim\)`<!-- -->`{=html}34.7 au, \(\sim\)`<!-- -->`{=html}36.2 au and \(\sim\)`<!-- -->`{=html}42.4 au. The ratio p:q denotes the resonance of p orbital periods of Neptune to q periods of the TNO. Scattered disc objects are TNOs that are on orbits which are currently scattering off Neptune such that their semi-major axes, \(a\) change by more than 1.5 au in 10 myrs. In addition to the HCs, CCs, resonant objects and scattered disc objects, the Neptunian Trojans (NTs) located at \(\approx\)`<!-- -->`{=html}30 au in the Sun-Neptune L4 and L5 Lagrange points, are hypothesized to have been captured from the TND into co-orbital resonances with Neptune during its outward migration. The colours of TNOs are known to be bimodally distributed between "red" (R) and 'very-red'' (VR) object colours where R objects are defined as having an optical spectral slope of \(\lesssim\)`<!-- -->`{=html}20 \(\%\) / 100 nm corresponding to a \(g-i\) colour index of \(<\)`<!-- -->`{=html}1.2 in the SDSS \(g\)and \(i\) bandpasses. The "very-red" (VR) category of TNOs are defined as having an optical spectral slope \(\gtrsim\)`<!-- -->`{=html}20 \(\%\) / 100 nm corresponding to a \(g-i\) colour index of \(>\)`<!-- -->`{=html}1.2. The HCs are a more equal mixture of "red" (R) and 'very-red'' (VR) objects while the CCs have a higher ratio in the number of VR objects to the number of R objects. One of the explanations for the colour dichotomy between R and VR objects is that the original TND had a colour transition boundary from R to VR objects occurring in the primordial disc between \(\approx\)`<!-- -->`{=html}30 and \(\approx\)`<!-- -->`{=html}40 au. Out of 32 known L4 and L5 Lagrange point NTs, 16 have optical colours which cover a wide range in optical slope with the majority having an optical spectral slope of \(<\)`<!-- -->`{=html}20 \(\%\) / 100 nm or \(g-i\) \(<\) 1.2. Presently, only one NT is known to have colours that place it in the VR category, 2013 VX\(_{30}\) with \(g-i\) = 1.52 \(\pm\) 0.06. The dearth of VR category objects is surprising because the NTs were captured at a similar heliocentric distance as HCs, but HCs have a higher VR to R colour ratio. This suggests that the transition boundary between R and VR objects was actually much further out from where the NTs were captured, more than 30 au from the Sun and possibly as far out as 40 au near the formation region of the CCs. In this work we expand on the previous work available on the visible colours of NTs, with observations of 18 objects, 15 of which are new, which increases the number of NTs with known visible colours to 31. # Observations We obtained optical \(g\), \(r\)/R, and \(i\)/I photometry of 18 NTs with the Hale 5.1 m telescope (P200 hereafter) at Palomar Observatory, the Gemini North 8.1 m telescope (Gemini N hereafter), and the Keck I 10 m telescope at Maunakea Observatory, and the Gemini South 8.1m telescope (Gemini S hereafter) at Cerro Pachón. Observations of our 18 NT targets were divided between the P200, Keck I, Gemini N, and Gemini S during 2020-2022. Five NTs were observed with the P200 using the Wafer-Scale Imager for Prime focus (WaSP) instrument. Three NTs were observed with Gemini N using the Gemini-North Multi-Object Spectrograph (GMOS-N). Six NTs were observed with Gemini S using the Gemini-South Multi-Object Spectrograph (GMOS-S) Four NTs were observed with Keck I using the Low Resolution Imaging Spectrometer (LRIS). NTs on orbits which have been demonstrated with numerical calculations to be likely temporary captures from the background trans-Neptunian population were not observed. Photometry of NTs was obtained with Sloan Digital Sky Survey (SDSS)-equivalent \(g\), \(r\), and \(i\) filters with the P200, Gemini N, and Gemini S. Observations of NTs with Keck I used the SDSS-equivalent \(g\) filter and the, Cousins R and I filters. Images were obtained in alternating \(g\), \(r\)/R, and \(i\)/I sequences to minimize the effect on colour measurements caused by variations in the brightness of the NTs due to their rotation. Exposure times ranged between 30 s and 300 s depending on conditions and the faintness of targets and the number of exposures per filter ranged between 3 and 15. The NTs were tracked at their non-sidereal rates, typically \(\approx\)`<!-- -->`{=html}0.05/s. A complete technical description of the facility and instrumental set for our NT observations is provided in. Observations of NTs at all four sites occurred when the targets were as close to opposition as possible at at minimum airmass for maximum throughput and image quality. Seeing varied between 0.8-1.3as measured in the WaSP images taken with the P200, between 0.8-1.1as measured in the LRIS images taken with Keck I, was \(\approx\)`<!-- -->`{=html}0.5as measured in GMOS-N images taken with Gemini N, and varied between 0.6-0.9as measured in GMOS-S images taken with Gemini S. Standard stars from the Panoramic Survey Telescope and Rapid Response System survey were identified in the same fields as the science observations. A complete list of our observations can be found in Table S1 of. Taking a similar approach as, data from each facility were detrended and flattened using bias frames and flat-field images obtained with an inside-dome flat field panel. Cosmic rays and other blemishes were removed from individual images using the L.A.Cosmic Laplacian cosmic ray identification algorithm. The data were stacked in each photometric filter to enhance the signal of the NT detections. A complete description of the data reduction for each facility and instrument combination is available in the Supplemental Material, along with Fig. S1 showing examples of the P200, Gemini N/S, and Keck I NT detections. The photometry of our NT targets and standard stars was performed using an aperture centred on the NT detections with a radius of 1.0-2.5that was 1.5-2 times the seeing measured in the images. Sky subtraction was completed by taking the median pixel value within an annulus centred on the NT detection that had an inner radius of 3.0-7.5and an outer radius of 6-11. The NT photometry obtained with the P200, Keck I, and Gemini N/S was calibrated using the Pan-STARRS photometric catalog. # Results The photometric measurements of our 18 NT targets are summarized in Table 1. We have plotted the optical colours of the NTs observed by us and by and in \(g-i\) vs \(g-r\) colour space in Fig. 1. The average \(g-i\) value of the 18 NTs is \(\approx\)`<!-- -->`{=html}0.84, equivalent to a spectral gradient of 8\(\%\) / 100 nm normalized to 550 nm and significantly redder than the Sun which has \(g-i\) = 0.58. Out of the 18 NTs that we observed, four have \(g-i\) colours \(\gtrsim\)`<!-- -->`{=html}1.2, the rough boundary separating the R and VR groups: 2013 VX\(_{30}\), 2011 HM\(_{102}\), 2013 TZ\(_{187}\), 2015 VV\(_{165}\). Our measurements of the optical colours of the VR NT 2013 VX\(_{30}\) with \(g-i\) =1.15 \(\pm\) 0.17 is broadly consistent with the \(g-i\) \(\approx\) 1.5 by, though their \(g\)-\(i\) colour measurement more robustly places it past the 1.2 \(g-i\) VR colour boundary. ::: One of the NTs we observed, 2011 HM\(_{102}\), has \(g-r\) = 0.91 \(\pm\) 0.07 and \(r-i\) = 0.34 \(\pm\) 0.06 placing it into the VR category with \(g-i\) = 1.25 \(\pm\) 0.06. observed 2011 HM\(_{102}\) and found \(r-i\) = 0.31 \(\pm\) 0.04, consistent with our measurements, but found \(g-r\) = 0.51 \(\pm\) 0.04, significantly bluer than our measured \(g-r\) = 0.91 \(\pm\) 0.07. The difference could be due to underestimated uncertainties in the \(g-r\) colour measurement by ourselves or by, or due to lightcurve variations. The SNR of our composite 2011 HM\(_{102}\) \(g\) and \(r\) observations was \(\sim\)`<!-- -->`{=html}20 each and is consistent with the SNR expected from integration time calculations simulating the conditions of our observations[^1]. To test the latter hypothesis, we measured the \(g\) magnitude of 2011 HM\(_{102}\) in the \(g\) images taken at Gemini S taken in the \(g\), \(r\), and \(i\) sequence on 2022 Jul 18 UTC and found that there were no significant variations in the brightness of 2011 HM\(_{102}\) in excess of \(\approx\)`<!-- -->`{=html}0.05 over the roughly half-hour observing sequence. In addition, our \(g-i\) measurements of 2014 QO\(_{441}\) of \(g-i\) = 0.84 \(\pm\) 0.08 and of 2014 UU\(_{240}\) of \(g-i\) = 0.52 \(\pm\) 0.07 are generally consistent with the values measured by. # Discussion and conclusion An initial impression of the NT colours in Fig. 1 is an apparent lack of a bimodal colour distribution. Following the example of, we have compared the \(g-i\) colours of NTs with those of other dynamical classes. Using the optical colours compiled by the Minor Bodies in the Outer Solar System (MBOSS) database, we have plotted the cumulative \(g-i\) colour distribution of the Jovian Trojans (JTs), Centaurs, Scattered Disc objects, Plutinos, Resonant objects, HCs, CCs and Detached Objects with the colours of NTs in Fig. 2. By visual inspection, the cumulative \(g-i\) distribution of the JTs is distinct in lacking any VR objects compared to objects of the other TNO dynamical classes. However, it must be noted that while they lack VR objects, the Jupiter Trojans are bimodal in colour, albeit with bluer mean colours compared to the TNO population. The cumulative distribution of the \(g-i\) colours of the NTs is located between these two groups, containing more VR objects than JTs, but disproportionately fewer VR objects compared to the other TNO classes, especially the CCs. To quantify the differences between the \(g-i\) distribution of the NTs and the \(g-i\) distribution of other dynamical classes, we apply the Kolmogorov-Smirnov (KS) test which measures the maximum difference between two cumulative distributions. The KS method tests the null hypothesis that two cumulative distributions are drawn from the same parent distribution. We have also applied the Kuiper variant of the KS test which is more sensitive to differences between distributions at their edges. Table 2 shows the associated statistical score and p-value of the KS and Kuiper tests between the NTs and other TNO dynamical classes. The statistical score is a quantified measure of the maximum difference between two cumulative distributions with a larger statistical score corresponding to a lower p-value for datasets of similar size. The comparison between the NTs and the CCs results in the largest statistical score of 0.74 for both the KS and Kuiper test corresponding to a p-value of \(<\) 0.0001. The comparison between the NTs and JTs results in the smallest statistical score of 0.22 corresponding to a p-value of 0.2138 for the KS test and a statistical score of 0.28 and a p-value of 0.0882 for the Kuiper test. The KS and Kuiper tests between the NTs and the Centaurs, Scattered Disc objects, Plutinos, Resonant objects, HCs and Detached objects have p-values \(<\)`<!-- -->`{=html}0.005. ------------------ ---- ---------- ---------------------------- --------- ---------------------------- Class N KS Score KS P K Score K P NTs 31 0.0 1.0 0.0 1.0 JTs 76 0.2186 0.2138 0.2830 0.0882 Centaur 43 0.4449 0.0010 0.3946 0.0012 Scattered Disc 27 0.4182 0.0086 0.4053 0.0008 Plutinos 35 0.4203 0.0039 0.4389 \(<\)`<!-- -->`{=html}0.0001 Resonant 16 0.7137 \(<\)`<!-- -->`{=html}0.0001 0.6992 \(<\)`<!-- -->`{=html}0.0001 HCs 39 0.4797 0.0004 0.5833 \(<\)`<!-- -->`{=html}0.0001 CCs 31 0.7419 \(<\)`<!-- -->`{=html}0.0001 0.7419 \(<\)`<!-- -->`{=html}0.0001 Detached objects 19 0.5195 0.0019 0.4873 \(<\)`<!-- -->`{=html}0.0001 ------------------ ---- ---------- ---------------------------- --------- ---------------------------- : Kolmogorov-Smirnov (KS) and Kuiper (K) variant NT statistical score and p-values (P). The \(g-i\) colour distribution of the NTs is distinct compared to other TNO classes. Previous studies show a much larger p-value for the tests between the cumulative optical colour distribution of the NTs and JTs. Our expanded sample with additional VR NTs implies dissimilarity in the \(g-i\) colours of NTs and JTs, although it only rules out the null hypothesis at the 1-2-\(\sigma\) level with a p-value of \(\approx\)`<!-- -->`{=html}0.08-0.21. The small number of the VR NTs and the large error bars on the \(g-i\) colours may make drawing a strong conclusion about the differences between the optical colours of NTs and JTs difficult. TNO evolutionary models predict that the observed proportion of VR and R objects in different TNO classes is a result of the separation between VR and R objects in the original TND located at a radial distance from the Sun denoted as \(r_*\). The combined sample of our observed NTs with those from the literature results in a VR to R ratio of \(\approx\) 1:8. Although the location of \(r_*\) may also be affected by the density profile of the original TND, a higher proportion of VR objects to R objects may imply a closer in value of \(r_*\) compared to a lower proportion. In the case of a disc with an exponential density profile, an NT VR to R ratio of 1:8 may imply a \(r_*\) interior to 35 au whereas a truncated profile may imply a \(r_*\) interior to 30 au. In either case, the discovery of additional NTs and measurements of their optical colours will provide additional constraints on the compositional gradient of the original TND. In addition to the location of the transition boundary between R and VR objects in the original TBD, the colours of TNOs could also be affected by post formation evolutionary effects such as collisions and thermal processing. # Supplemental Material {#supplemental-material .unnumbered} The supplemental material for this manuscript is available online. [^1]: <https://www.gemini.edu/instrumentation/gmos/exposure-time-estimation>

{'timestamp': '2023-02-10T02:00:16', 'yymm': '2302', 'arxiv_id': '2302.04280', 'language': 'en', 'url': 'https://arxiv.org/abs/2302.04280'}